Manhattan Scientist 2017 by CETheodosiou

The Manhattan Scientist Series B Volume 4 Fall 2017

A journal dedicated to the work of science students

The Manhattan Scientist Series B

Volume 4

Fall 2017

ISSN 2380-0372

Student Editors Mia Bertoli Daisuke Kuroshima Hannah Mabey Lilliana McHale Tyler Reese Faculty Advisor and Editor in Chief

Constantine E. Theodosiou, Ph.D.

Manhattan College

4513 Manhattan College Parkway, Riverdale, NY 10471 Student Handbook manhattan.edu MANHATTAN COLLEGE â&#x20AC;˘ RIVERDALE, NY 10471

Series B

The Manhattan Scientist

Volume 4 (2017)

A note from the dean The present Volume is the seventieth anniversary issue of the Manhattan Scientist. Science research has dramatically evolved, since 1947, the year of Scientist’s first issue, particularly with the development of high quality instrumentation. Humankind has witnessed a monumental growth in its scientific knowledge in all – and newly developed – disciplines. Humans have been to the moon, tamed nuclear fission, developed the semiconductors (transistors), harnessed solar energy, discovered the elements of life through DNA analyses, and have been steadily progressing in their efforts to modify basic food sources to be free from disease and to cure human diseases at the cellular and molecular level. As we have moved from the twentieth century, the “century of physics,” to the twenty-first century, the “century of biology,” we are at the threshold of new discoveries. While at the end of the ninetieth century scientists believed they knew all there was to be known about nature, the twenty-first one brings promise of unlimited growth in our knowledge and fantastic benefits from responsibly harnessing it. The present volume includes thirty-six papers, covering all disciplinary subjects of the School of Science at Manhattan College. These activities continue to be aligned with the College’s mission, to “provide a contemporary, person-centered educational experience that prepares graduates for lives of personal development, professional success, civic engagement, and service to their fellow human beings.” The quality of the projects continues to be exceptional, considering that these students were in high school only two to three years ago. It also reflects the quality of our faculty and the level of their commitment to research as part of the students’ educational experience. The participants ranged from High School summer interns, to primarily undergraduate students encompassing all our majors, to graduate students in Mathematics, and guest students from our sister institution, Bethlehem University in Palestine. I would like to express our gratitude to the faculty who willingly provided critical mentoring to our students and future colleagues, with minimal or no compensation for these efforts. This work continued to receive critical financial support for our students from a variety of sources (in no particular order): the School of Science Research Scholars Program, the Jasper Scholars Program of the Office of the Provost, the Catherine and Robert Fenton endowed chair in Biology, the Linda and Dennis Fenton ’73 endowed biology research fund, the Michael J. ’58 and Aimee Rusinko Kakos endowed chair in Chemistry, Jim Boyle, ’61, Kenneth G. Mann ’63, the Camille and Henry Dreyfus Foundation Senior Scientist Mentor Program, and a National Science Foundation research grant.

Series B

The Manhattan Scientist

Volume 4 (2017)

This anniversary issue, sadly coincides with the passing of our dear colleague Edward Brown, Physics ’61. Professor Brown had been a student, professor of physics, chair of the department of physics, and founding dean of the School of Science at Manhattan College. While he was a student he published in the Manhattan College Journal of Physics. His paper, reprinted in this volume, providing an explanation of the Ice Ages, is an amazing display of his intellect at the young age of 23. We are proud to dedicate this Volume to the work and memory of Professor Brown. A biographical note was contributed by the faculty of the department of physics alongside with a note of appreciation from Dr. Roy Helmy, Biochemistry ’99. On the cover of this issue is an image from the highly publicized full solar eclipse observed in the Unites States in 2017. We hope that its nature of temporary blocking the lifegiving light may parallel the eventual passing of the present climate of science denial and ignoring the scientific facts that seems to prevail among many of our political leaders, at the peril of the humankind’s future survival. I would like to express my deep appreciation to the students for their efforts and their persistence in the road of scientific discovery. We are all honored to showcase our students’ and colleagues’ continuing contributions to the body of scientific knowledge. It is with great pleasure that the editors present the publication of Series B, Volume 4, of The Manhattan Scientist.

Constantine Theodosiou Dean of Science and Professor of Physics

ISSN 2380-0372

Series B

The Manhattan Scientist

In memory of Edward Brown Colleague and Teacher

Volume 4 (2017)

Series B

The Manhattan Scientist

Volume 4 (2017)

Edward Brown† Ed Brown joined Manhattan College in 1957 as a student majoring in physics. Since then more than a half a century has gone by and throughout this entire period he distinguished himself as among the best and the most dedicated members of our Manhattan community. After graduation he joined our physics faculty while working on his doctorate at NYU. Even in these early years as a junior faculty he had achieved a leadership position within the physics department, so he naturally became the chairman of his department in 1986 and continued in this capacity until September of 1993 when he was appointed as the first Acting Dean of our, then new, School of Science. The eventful years 1992-93 culminating in the restructuring of Manhattan College and the creation of the School of Science owe much to the leadership qualities of Ed Brown who showed unique leadership in bringing the diverse departments of Science and Mathematics together to be able to speak with one voice in what seemed to be a struggle at the time. It is the rare individual who can do all of this while maintaining good relations with departments in the Arts and Humanities, where significant opposition to the restructuring remained. Ed Brown was just such an individual, so much so that he was appointed the Dean of Science in 1994 and remained in that position for 16 years. Under his leadership the School of Science experienced significant renewal in both teaching, curricular developments and scholarship in a new atmosphere of openness and support. His success and long tenure as the Dean of Science can be attributed to his knack in building consensus among faculty in diverse fields and his ability to guide them in new directions. His style was always direct and seemingly effortless as he made an issue crystal clear and convinced others his was the best solution to a problem. Ed Brown’s clear intellect, ability to grasp and connect a wide range of knowledge, combined with a charisma rooted in a genuine respect and enthusiasm for faculty’s efforts in teaching and scholarship, his curricular leadership in all the scientific fields, both mathematical and non-mathematical, both experimental and theoretical were crucial in bringing the School of Science into the 21st century. In addition to his lifelong service to our college, Ed Brown was a theoretical physicist par excellence with a large number of publications to his name in major journals including Physical Review. Most notable among these are his papers on Green’s Functions as applied to magnetic systems. In 2010 Ed Brown rejoined the physics faculty until he passed away unexpectedly this last May. His absence will remain an irreplaceable loss for Manhattan College. The physics faculty at Manhattan College

Series B

The Manhattan Scientist

Volume 4 (2017)

In appreciation ... It was the fall semester of August 1995 and I was entering Manhattan College as a freshman. One of our first science courses was with Dean Brown on the first floor of Hayden Hall. The course was to familiarize us with the School of Science and the various opportunities we would encounter during our 4 year tenure in Manhattan College. Dean Brown advised us to not pick a career that we didn’t find interesting or inspiring. He lectured us to work hard and enjoy our time to experience the liberal learning. Over the next 3 years, I would do research with Dr. Wasacz, Dr. Kerrigan, and Dr. Fan. Manhattan College provided me with a sound scientific background and taught me how to perform research. I published my first scientific paper at the College as a junior in Dr. Fan’s lab. Dean Brown encouraged me to apply for a NYPEW foundation grant and I was terrified because only a select few students were capable of receiving the grant. I also had to explain to my father that I could not work in our family deli store for two summers straight as the grant required a full time commitment for two summer semesters. Dean Brown insisted that this is a once in a lifetime opportunity that would allow me to explore my passions for academic research. I ultimately had the courage to tell my father that I couldn’t work in the store and for two summers I did research in the field of Biochemistry and Inorganic Chemistry (I did agree to work in our family store late nights and on the weekends). It would be no surprise to learn that I now lead a research group at Merck & Company developing lifesaving medicines. If you look at one’s career, you often find out that one comes to a crossroad and needs to make a decision. For me, Dean Brown and the faculty of chemistry were there to help and guide the way. They were like a catalyst that helped lower the activation energy to enable me to be the scientific leader I am today. Dean Brown was not only a professor of Physics, but also a mentor who would listen, not judge, and guide you to the correct path. Dean Brown will always be in my heart and I owe him my gratitude for his support. Roy Helmy, Manhattan College 1999

[ Reprinted from the: Manhattan College Journal of Physics, Vol. 2 No. 1, 1962 ]

Series B

The Manhattan Scientist

Volume 4 (2017)

Table of Contents Biochemistry Characterizing in vivo chromatin remodeler interactions at the yeast nucleosome core Brian Evans . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 Exploring chromatin dynamics within the DNA damage response pathway in living cells Bright Shi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9 Examining nucleosomal dynamics in living Saccharomyces cerevisiae (yeast) cells Amanda Zimnoch . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

Biology Predicting rates of bark formation on saguaro cacti (Carnegiea gigantea) Mia Bertoli . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 Leaf venation patterns and interveinal area in tree leaves Jorge Gonzalez . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35 Predicting rates of mortality for saguaro cactus plants (Carnegiea gigantea) Cole Johnson . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43 Bark formation and death progression in saguaro cacti (Carnegiea gigantea) Marissa Locastro . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51 Automated quantitative analysis of tree branch similarity using 3D point cloud registration Matthew Maniscalco . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 Evidence of Giardia lamblia oocyst stage in bivalves collected in New York City Monique Ng and Joseph Annabi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 65 Xylem conductivities from stems to leaves for species of grass plants Humberto Ortega . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71 Dynamics of flowering branches of Artemisia tridentata Ismael PeË&#x153;na . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79 Quantification of water conductivity in xylem cells of Artemisia tridentata Claudia Ramirez . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93 Effects of chemical exposure on tadpole behavior and personality Cassidy Stranzl . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 101 Generating a 3D CAD model of tree branches using a 3D scanner Michael Volgende . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107

Chemistry Methylammonium lead iodide nanowires for perovskite solar cells Jacqueline DeLorenzo . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117 Remarkably rapid reduction of toxic Cr(VI) levels Patsy Griffin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125 Synthesis, isolation,and characterization of cyclic organic Cr(VI) molecules Christopher Kim . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 135 The synthesis of antimony sulfide nanowires in aqueous solution James Ksander . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141 Determination of the crystal structure of aluminosilicate zeolite ZSM-18 Daisuke Kuroshima . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 147 Magnetic/Titanium dioxide hybrid nanomaterials as photocatalysts for water purification Hannah Mabey . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 153

Series B

The Manhattan Scientist

Volume 4 (2017)

From solution to absorption: Innovative method for removing toxic chromate Dominick Rendina . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 159 Determining the framework topology of a microporous aluminosilicate zeolite, SUZ-9 Christine E. Schmidt . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167 Variations of the A cation in organic/inorganic perovskite materials Melissa Skuriat . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177 Study of the interaction between mefenamic acid, fludarabine and human serum albumin by spectroscopic methods Ewa Swiechowska . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .. 181

Computer Science Image recognition using autoencoding in multilayer neural networks and multi-value neurons Niko Colon and Alexander Gonzalez . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 191 Image recognition using analysis of discrete Fourier transform by multilayer neural networks with multi-valued neurons Alexander Gonzalez and Niko Colon . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199 Monte Carlo computer investigation of ideal dendrimers in two and three dimensions Timothy Hamling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 209

Mathematics Statistical binary classification of MRI data Sana Altaf and Melissa Brenner . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 215 Whatâ&#x20AC;&#x2122;s in the air? Using mathematical models to predict Boston air quality Anthony DePinho, Tara Ippolito, Biyonka Liang, Kaela Nelson, Annamira Oâ&#x20AC;&#x2122;Toole . . . . . . . . . . . . . . . . . . . . . . . . . . . . 225 Using kurtosis for mathematical classification Tenzin Kalden . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 235 Classification of magnetic resonance imaging (MRI) data using small sample sizes Hope Miedema . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 243 Building a mathematical model for Lacrosse analytics Samantha Morrison . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251

Physics Search for Lorentz invariance violation from gamma ray bursts Linh Nguyen . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 261 Study of simulated particle data and practical applications Danielle Rabadi . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267 Reconstruction of the Higgs boson using computational methods Tyler Reese . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 273

International Genetically Engineered Machine (iGEM) competition E.(lectro) coli and the GO(x) lden nAnode B. Evans, F. Begum, A. Lazkani, S. Rithu, D. Abdur-Rashid, G. Sanossian, and S. Coby . . . . . . . . . . . . . . . . . . . . . . . 279

On the cover: NASA image from the 2017 total solar eclipse, observed in the United States

Characterizing in vivo chromatin remodeler interactions at the yeast nucleosome core Brian Evansâ&#x2C6;&#x2014; Department of Chemistry and Biochemistry, Manhattan College Abstract. In humans and yeast, genetic information is stored as chromatin. Within the chromatin, DNA is wound around octameric units of histone proteins. The histone-DNA complex forms the nucleosome and represents the most basic level of DNA compaction. A class of proteins known as chromatin remodelers physically repositions the nucleosome which directly affects gene regulation. Due to their large sizes and numerous auxiliary subunits, chromatin remodeler interactions at the nucleosome have yet to be fully characterized and remain relatively obscure. In this work, we use synthetic biology to trap histone-protein interactions in living cells using the unnatural amino acid pBPA. With the addition of a short peptide fusion tag, the technology allows for the identification of crosslinks formed with the histone-protein complex at any of the particular residue at which pBPA is encoded. This system has the potential to make detailed crosslinking maps of any chromatin remodeler complex which can then be used to identify and assign biologically important structures. By properly implementing the system, we exemplify the potential of this technology for creating crosslinking maps for the chromatin remodeler RSC. This conclusion is supported by Western blot analysis and a clearly identified crosslink between a Myc tagged subunit of RSC (sth1 protein) with histone H2A harboring pBPA at position A61.

Introduction Humans and yeast are separated by a billion years of evolution. Even still, their core histones have retained essential and fundamental roles in gene regulation. In humans and yeast, genetic information is stored as chromatin. Within the chromatin, DNA is wound around octameric units of histone proteins. The histone-DNA complex forms the nucleosome and represents the most basic level of DNA compaction [1]. Depending on the cellular needs, this complex is directly modified and reorganized in order to control transcriptional activation or repression. The nucleosome is chemically modified at its N-terminal tails which protrude out of the nucleosome interface. These chemical modifications are known as post translational modifications (PTMs). Different PTMs have altered and defined effects on the chromatin structure [2]. For example, methylation is generally associated with cellular signals that promote the chromatin complex to become more compact and less active. PTMs recruit proteins such as chromatin remodeler (CR) complexes that physically reposition nucleosomes [3]. The CR family is comprised of large multisubunit complexes. They all contain an ATPase active subunit for the direct repositioning of the nucleosome as well as numerous additional subunits that are crucial for stabilizing the protein histone interactions. Each member of the CR family has unique function in the dynamic process of chromatin regulation. The remodeler, RSC, contains over 15 subunits [4]. The enzymatic subunit, Sth1, is an ATPase responsible for catalytic control of nucleosomal translocation. Previous research studies on RSC â&#x2C6;&#x2014;

Research mentored by Bryan Wilkins, Ph.D.

The Manhattan Scientist, Series B, Volume 4 (2017)

Evans

often rely on the reconstitution of nucleosomal arrays in solution. These studies inherently fail to properly recapitulate the true nuclear environment. As a result, to date, there are no high-resolution structural data on CRs in complex with the nucleosome. Therefore, to enhance the understanding of CR behavior a technique is required that illuminates the molecular contacts that occur between the nucleosome and CR proteins inside the living nucleus. This work uses a system utilizing the unnatural amino acid p-benzoylphenylalanine (pBPA) to illuminate the interactions between the active subunit of RSC, the ATPase sth1, and the yeast nucleosome core. As the name implies, unnatural amino acids are not found naturally in nature. These synthetic amino acids can be used to produce unnatural proteins that often confer new traits and expand the number possibilities to examine these proteins [5, 6]. For instance, pBPA possesses the ability to form a covalent bond between two interacting proteins. When excited by UV light, this unnatural amino acid forms a diradical which can subsequently abstract a hydrogen from a nearby protein. The remaining two radicals combine to form a covalent bond between the two proteins, trapping the interactions between the proteins of interest [7]. The multistep mechanism is illustrated below in Fig. 1. This crosslinking can then be identified through Western blot and immunodetection methods.

Figure 1. Multistep mechanism for the UV induced formation pBPA crosslinking [8].

We utilize this system by inserting pBPA into histone proteins. Once the system is in place, we can then monitor crosslinks that form from the nucleosome in the living nucleus. Since RSC complex interactions with the nucleosome are so difficult to characterize, this system provides a means to more accurately map the interactions in living cells. To properly integrate the system in vivo we use a dual plasmid system. The first plasmid contains the histone gene and a stop codon inserted at the particular amino acid of interest [9]. The second plasmid contains additional machinery needed to â&#x20AC;&#x153;trickâ&#x20AC;? the endogenous system to

Evans

The Manhattan Scientist, Series B, Volume 4 (2017)

incorporate the unnatural residue. This includes the genes necessary for an amber suppressor tRNA and the aminoacyl-tRNA synthetase specific for the pBPA [6]. By changing the position of the pBPA, we can essentially scan the nucleosome interactions to assay for crosslink formations. The histone protein of interest can be identified after crosslinking via immunoblotting against a small fusion peptide tag. For this work, a Myc tag was genetically incorporated using site directed PCR mutagenesis [10]. Using this information, we can identify biologically important residues on the RSC complex that would potentially be targets for therapeutic research.

Materials and Methods Yeast Culture To gain better structural and functional insight into RSC and its enzymatic subunit, sth1, yeast strains were produced harboring site-specifically encoded pBPA on histone H2A at position A61. The BY4741 cellular yeast strain (genotype: MATa his3∆1 leu2∆0 met15∆0 ura3∆0) was used as the wild type strain. The histone proteins of interest were tagged with a C-terminal HA-tag for antibody detection. This strain was transformed with the dual plasmid system to generate the proper system for pBPA insertion. The histone plasmid obtains a selectable marker for uracil and the synthetase plasmid obtains one for leucine. In effort to maintain these plasmids, the yeast cells were cultured in dropout media lacking leucine and uracil with a final concentration of 1 mM pBPA and 2% glucose. The cells were grown in a shaker overnight at 30◦ C. The following day, the optical density (OD) of the cells were measured using a spectrometer at absorption of 600 nm. The OD600 was used to normalized the cells so that they had equal densities. 1 OD is equivalent to a 1 mL culture when the OD600 is equal to one. Each culture was split into a set of 12 ODs, one to be used as negative control and the other for UV induced crosslinking sample. Generation of Sth1-myc tagged strain The plasmid pYM5 [10] was used for the amplification of the 3-myc-HIS3 cassette. The PCR reaction mixture was as follows: approximately 300 ng pYM5 template plasmid, 200 µM dNTP mixture, 0.5 µL Phusion polymerase (1 U, Thermo Scientific F530S), 10 µL HF Phusion buffer, 0.2 µM S2 primer, 0.2 µM S3 primer and water to a final volume of 50 µL. The reactions were performed on a BioRad Thermal Cycler under the following cycling conditions: 30 cycles, denaturing at 98◦ C for 20 s, annealing at 62◦ C for 30 s, and extension at 72◦ C for 1 min. Initial denaturing was set to 98◦ C for 1min and the final extension was set at 98◦ C for 5 min. To check the success of reactions 2µL of sample was mixed with 2µL of DNA running dye and then loaded R safe) and subjected into a 1% agarose gel (50ml 1X TBE Buffer, 0.5g Agarose, 5µL CYBR to electrophoresis for 1 hour at 125V. Agarose gel images we captured using an Omega LumTM G Imaging System on CYBR setting. The expected PCR product for each PCR was ∼1700 base pairs. This strain was then transformed with the dual plasmid system for pBPA and cultured as detailed above for the wild type cells.

The Manhattan Scientist, Series B, Volume 4 (2017)

Evans

Crosslinking 12 ODs of cells were resuspended in 100 µL of water. A set of control samples were then kept to compare to UV induced samples. The UV induced samples were placed in depressed wells of R UVL-225D 50 Watt, 365 nm, an aluminum plate exposed to UV light from a UVP Mineralight 1.25 Amps, 115V∼60Hz lamp for a total of 20 minutes. The plate was kept on ice to maintain a temperature of 4◦ C. After 10 minutes of exposure, the cells on the plate were mixed. Protein Isolation After the crosslinking procedure the cells were then collected and lysed. This was accomplished via a trichloroacetic acid (TCA) precipitation assay. 1 mL of lysis buffer (1.2% betamercaptoethanol, 300 mM NaOH, 1 mM PMSF in water) was added to each sample. Each sample was then kept on ice for 10 minutes. After the allotted time, 160 µL of 50% TCA was added. The samples were then kept on ice for at least 20 to 60 minutes. The whole cell lysates were centrifuged at 4◦ C at 15000 r.p.m. for at least 10 min. The supernatant was discarded and the cells were washed with 1 mL of cold acetone. This washing step was repeated and the samples were centrifuged at 4◦ C at 15000 r.p.m. for at least 10 minutes. The supernatant is discarded a final time and the samples are left to dry for approximately 30 minutes. The pellets are resuspended in 100µL 1X SDS loading buffer and boiled at 95◦ C for 20 minutes. The samples were then centrifuged at 15000 r.p.m. and stored at -20◦ C for later use. Protein Electrophoresis Proteins were separated via electrophoresis on an 8% SDS-PAGE gels in standard tris-glycine buffer (25 mM Tris, 192 mM glycine, 0.1 % SDS). 5µL of pre-stained Rec protein ladder and 10 µL of each yeast protein sample were added. Electrophoresis was performed at 100V until resolved the appropriate distance. Western Blotting The protein in the polyacrylamide gel was transferred to a PVDF membrane via western blotting in Towbin buffer with 20% methanol. The transfer was performed at 250 mA for two hours. The PVDF membrane was subsequently washed three times with water and then blocked with 3% BSA solution in TBS for at least 20 minutes. The BSA was removed and the membrane and the primary antibody was added. The membrane was kept at 4◦ C on a rocker overnight. The membrane was then washed with TBS three times and two additional 5 minute washes of TBS. The secondary antibody was a conjugate to HRP and the solution was anti-rabbit at 1:10000 in 5% milk. After the secondary antibody was added, the membrane was kept on rocker for at least 45 minutes. The membrane was then washed with TBS three times and two additional 5 minute washes of TBS. Then the membrane was washed with 30 mL of 0.1% TBS for 10 minutes. The membrane was then finally washed 10 times with water.

Evans

The Manhattan Scientist, Series B, Volume 4 (2017)

Antibodies used in this work: Primary: Anti c-Myc Sc-40, produced in mouse, Santa Cruz Biotechnology (9E10) Anti-HA, produced in rabbit, Abcam (ab9110) Secondary: Anti-mouse IgG peroxidase conjugated, produced in goat, Sigma (A4416) Anti-rabbit IgG peroxidase conjugated, produced in goat, Sigma (A0545) Western Blot Imaging A 300 µL of peroxide/luminol solution SuperSignalTM West Femto Maximum Sensitivity Substrate (ThermoFisher product number: 34095) was mixed and added to the surface of the PVDF membrane. The normal “Chemi” setting on Azure c600 imager was used to obtain images of the membrane with exposure times ranging from 5-15 minutes.

Results PCR verification To ensure one of the genetically modified strains produced was successfully mutated to express the desired myc-tag, PCR verification of the Sth1 gene myc insertion was performed. DNA was isolated from cell lines that grew in the absence of histidine because these strains successfully incorporated the myc-histidine cassette. However, growth of these cells in the absence of histidine did not ensure that the cassette was incorporated at the proper genomic locus. PCR primers were designed to anneal to the 3’-end of the genomic Sth1 gene (∼350 bp upstream of the terminal end) and a segment of the genomic DNA just downstream of the Sth1gene (∼350 bp upstream). The primers were designed to produce a PCR product of ∼700 bp when primed to wild type genomic DNA. If the myc-histidine cassette was incorporated at the proper gene location then the primer that anneals to the genomic region would be repositioned ∼1500 bp further downstream due to the insertion of the coding sequences for both the myc-tag and the histidine gene. PCR verification reactions were performed on DNA isolated from the modified strains as well as from wild type cells. The products of these PCR reactions were compared on a 1% agarose gel. The verification result is illustrated clearly in Fig. 2. The PCR verification product for the Sth1-myc strain #2 had a similar electrophoretic migration rate as the control Wt strain. This strongly indicates that strain #2 was unsuccessfully mutated. On the other hand, strain #1 exhibited a clear electrophoretic shift. This shift indicates strain #1 successfully incorporated the expected genetic mutation to express the myc-tag due to a larger PCR product from the insertion of the tag and a selectable marker. For future experiments we used strain #1 as the Sth1-3myc cellular line. This modified strain was transformed with the dual plasmid system for the incorporation of pBPA. Western Blot Analysis Western blot analysis was first used to verify the strain by blotting against the myc-tag. After successfully verifying this, crosslinks to Sth1 from histone H2A were examined in Wt and Sth1

The Manhattan Scientist, Series B, Volume 4 (2017)

Evans

Figure 2. PCR verification of sth1 gene illustrating electrophoretic shift in strain #1 due to insertion of myc-tag compared to Wt strain without myc-tag.

tagged yeast strains by blotting against the HA-tag on H2A. The western blot result is visualized in Fig. 3. As expected, no crosslinks are observable in the UV negative samples. In the UV positive samples, a band shift was observed that was clearly UV dependent in response to a histone H2A interaction. The Sth1-histone H2A interaction was also identified by an electrophoretic size shift relative to the myc-tagged Sth1 protein when compared to crosslinks in Wt cells.

Figure 3. Western blot illustrating successful cross linking identification on H2A at A61.

Discussion By properly manipulating the genome of a yeast strain we were able to modify a cell line that harbored the coding sequence for the Sth1 protein fused with the DNA coding sequence for a 3Myc tag. The translation of this modified gene produces full-length Sth1 protein with a Cterminal 3myc tag. Utilizing an expanded genetic code in yeast we then transformed the Sth1-3myc strain, and a Wt strain, with plasmids that harbor the machinery essential for the site-directed incorporation of pBPA into histone proteins. Previous crosslinking and mass spectral data (not shown) suggested that the Sth1 protein forms a covalently trapped interaction with histone H2A

Evans

The Manhattan Scientist, Series B, Volume 4 (2017)

when pBPA is inserted at protein position alanine 61 (A61). We expressed mutant histone H2AA61pBPA (containing a fused HA tag) in living yeast cell lines (both Sth1-3myc and Wt) and crosslinked these proteins when cells were in their exponential growth phase. Protein-protein interactions between Sth1 and H2A were expected to be easily identifiable when crosslinks in the mutant and control strains were compared through western blot analysis. Due to the additional protein size of the fused Sth1-3myc protein we hypothesized that a crosslink to this protein would be electrophoretically retarded as compared to the same crosslink to the Wt protein. Our results verified this shift in protein electrophoretic mobility when proteins were analyzed by anti-HA antibodies (Fig. 3). There is a clear size shift for the Sth1-3myc proteins in the expected size range of an Sth1-H2A crosslinked protein. While this size difference is not very large ( 3.5 KDa) it is enough to resolve the difference in size as compared to the Wt crosslink. The myc fusion tag combined with the pBPA crosslinking technology represents a powerful and salient methodology to assess structural and dynamic mapping of a chromosomal remodeler, in vivo. By scanning each of the histone proteins with pBPA we can now illuminate the Sth1nucleosomal contacts and derive an interactome map for these proteins. Most importantly, since this work is performed in vivo it provides the first known mechanistic details of these protein contacts under true physiological conditions. While Sth1 is only a single subunit of the RSC complex, we can now shift our attention to other subunits to paint a complete structural image of the contacts this remodeler makes with the nucleosome. Therefore, this technique has the power to bridge the gap between in vitro versus in vivo experimentation of chromatin remodelers. An interaction map will help researchers discover and assign biologically relevant structures within the RSC remodeler and its many subunits. With this new valuable structural insight, RSC has the potential to become an important medicinal target for disease and developmental disorders.

Acknowledgments This work was funded by the Manhattan College Jasper Scholars Program. The author expresses special thanks to his mentor, Dr. Bryan Wilkins, for his continued guidance and support.

References [1] Luger, K., M¨ader, A. W., Richmond, R. K., Sargent, D. F. and Richmond, T. J. Crystal structure of the nucleosome core particle at 2.8 A resolution. Nature 389, 251–260 (1997). [2] Jenuwein, T. and Allis, C. D. Translating the histone code. Science 293, 1074–1080 (2001). [3] Bartholomew, B. Regulating the chromatin landscape: Structural and mechanistic perspectives. Annu. Rev. Biochem. 83, 671–696 (2014). [4] Cairns, B. R., Lorch, Y., Li, Y., Zhang, M., Lacomis, L., Erdjument-Bromage, H., Tempst, P., Du, J., Laurent, B. and Kornberg, R. D. RSC, an essential, abundant chromatin-remodeling complex. Cell 87, 1249–1260 (1996).

The Manhattan Scientist, Series B, Volume 4 (2017)

Evans

[5] Liu, C. C. and Schultz, P. G. Adding new chemistries to the genetic code. Annu. Rev. Biochem. 79, 413–444 (2010). [6] Chin, J. W., Cropp, T. A., Anderson, J. C., Mukherji, M., Zhang, Z. and Schultz, P. G. An expanded eukaryotic genetic code. Science 301, 964–967 (2003). [7] Dorman, G. and Prestwich, G.D. Benzophenone photophores in biochemistry. Biochemistry 33, 5661–5673 (1994). [8] Shi, B. Exploring chromatin dynamics within the DNA damage response pathway in living cells. An expanded eukaryotic genetic code. Manhattan Scientist B3, 19-28 (2016). [9] Wilkins, B. J., Rall, N. A., Ostwal, Y., Kruitwagen, T., Hiragami-Hamada, K., Winkler, M., Barral, Y., Fischle, W. and Neumann, H. A cascade of histone modifications induces chromatin condensation in mitosis. Science 343, 77–80 (2014). [10] Knop, M., Siegers, K., Pereira, G., Zachariae, W., Winsor, B., Nasmyth, K. and Schiebel, E. Epitope tagging of yeast genes using a PCR-based strategy: More tags and improved practical routines. Yeast 15, 963–972 (1999).

Exploring chromatin dynamics within the DNA damage response pathway in living cells Bright Shi∗ Department of Chemistry and Biochemistry, Manhattan College Abstract. Genetic information is stored in the form of chromatin, consisting of DNA, histones and other essential proteins. Histone proteins mediate all aspects of chromatin function and are regulated by sets of posttranslational modifications (PTMs). Modification patterns dictate differential pathways dependent upon cellular queues. This dynamic behavior is at the heart of all chromatin related processes, such as replication, transcription and repair. Unfortunately, DNA is inherently susceptible to damage. There are numerous forms of damaging factors, where several DNA damage pathways collectively protect the genome from life-threatening mutations that have direct links to both cancer and aging. Therefore, it is crucial that methods be developed that enable us to study the chromatin processes in order to better understand DNA damage pathways. We are using a synthetic biology approach that can trap histone-protein interactions in living cells, using unnatural amino acids. Comparing histone-protein interactions that are altered, due to DNA damage, will help us resolve the mechanisms that reshape chromatin structure under damaging stress. Many factors recognize and repair different types of damage but the orchestration of their function is still largely unknown. DNA damage signaling promotes broad changes in histone PTMs, and how the modifications control interactions at the nucleosomal interface during the response pathway is elusive. We can monitor histone PTMs across the cell cycle and correlate their influence on histone-protein interactions during damage pathways. We aim to expose nucleosomal repair protein-protein interactions and the mechanistic details of repair dynamics in yeast.

Introduction Chromatin is a nucleoprotein macromolecule that stores genetic information in the form of DNA, histones, and other essential proteins. Chromatin functions in both DNA packaging and processing. Chromatin’s most basic repeating unit is called the nucleosome. It consists of DNA wrapped around a histone octameric complex made of two histone H2A-H2B dimers and one histone H3-H4 tetramer [1]. Each of the histones possess N-terminal tails that protrude out of the nucleosome and interface with the solvent front. The tails can be highly modified by posttranslational modifications (PTMs) that alter the chemical properties of the protein. Through these modifications histone proteins mediate all chromatin functions. Different PTMs result in different outcomes. For example, when acetylation marks are upregulated, in general the chromatin becomes more active and less compact. Histone interacting proteins can then associate with the chromatin fiber to help modulate the cell’s functioning. A position at the N-terminal tails can be modified by different types of PTMs at different times to yield various effects. In general, chromatin regulations were not determined by a single PTM. Sets of PTMs at different position had been identified to collectively lead to a cellular function. By enhancing our understanding of chromatin behavior in living cells, one may lay the footstone that can develop into new strategies to fight against cancers and other chromatin related ∗

Research mentored by Bryan Wilkins, Ph.D.

The Manhattan Scientist, Series B, Volume 4 (2017)

Shi

diseases. Studying chromatin dynamics in vivo is very difficult and requires a technique that can reveal the interactions that occur between chromatin and chromatin-associated proteins. This work utilizes a synthetic biological approach that uses unnatural amino acids to capture histone-protein interactions in living cells to identify and monitor how PTMs influence chromatin dynamics during DNA damage repair pathways. DNA is inherently susceptible to damages, ranging from internal cellular stress to external environmental factors. So far, many factors are known that recognize and repair DNA damage, but the precise regulation of each step is still unknown. Several DNA damage pathways collectively protect the genome to counteract potentially life-threatening mutations [2]. DNA damage signaling promotes changes in histone PTMs as well as the recruitment of nucleosomal remodeling complexes. How these modifications control nucleosomal interface interactions are unknown to us. This work aims to expose nucleosomal repair protein-protein interactions and the mechanistic details of repair dynamics in yeast. Yeast is an ideal model organism for these studies because much of the repair pathway has been conserved from yeast to humans [3]. We aimed to identify changes in protein-histone interactions in response to DNA damage caused by hydrogen peroxide (H2 O2 ) and methyl methanesulfonate (MMS). MMS is an alkylating drug that specifically methylates guanine and adenine DNA bases. It causes mutations that break DNA double-strands and lead to replication problems. H2 O2 causes oxidative stress that can lead to mutations in DNA by creating an abasic site, i.e. loss of the base from the nucleotide. By comparing changes in protein-histone interactions, we can further reveal the mechanisms that shape chromatin structure. We can monitor how changes in interactions correlate with specific PTMs that occur during the change. To study the changes in protein-nucleosome interactions we used an amino acid suppression technique [4]. Synthetic amino acids that do not exist in nature are used to produce “unnatural” proteins that can be analyzed for protein function. One specific unnatural amino acid that suits our purpose is p-benzoylphenylalanine (pBPA) [5]. It possesses the ability to create covalent bonds, or chemical “bridges,” between two interacting proteins when activated with UV-light (365 nm). Thus, by tagging the histone at our desired positions, we can capture any protein that directly contact pBPA tagged histone. When pBPA is in its excited state it forms a diradical that can easily abstract a hydrogen from a neighboring protein that is in direct contact with the pBpa-containing protein (Fig. 1). This causes two radicals that readily recombine to form a covalent bond between the two proteins. Radical formation does not happen by chance, the proteins must be in interacting distance for this covalent bond to be possible, ensuring the accuracy of this method. This bridging allows for the identification of the crosslinked protein by mass spectral or immunodetection methods. The techniques being used have been previously developed for the study of chromatin [6]. We can take advantage of this technique to insert pBpa into histone proteins and monitor crosslinks from the nucleosome in the living nucleus.

Shi

The Manhattan Scientist, Series B, Volume 4 (2017)

Figure 1. pBPA structure and benzophenone chemistry. Under UV-light (â&#x2C6;ź 350 nm) pBPA is excited, but no energy is released at equilibrium to damage the cell. In its excited form, pBPA can easily form a covalent bond with another protein through radical recombination, following hydrogen abstraction [8].

In order to form the histone crosslinks within living yeast cells, we use a dual plasmid system for the suppression of a genetically installed stop codon with pBpa. One plasmid contains the histone gene with a TAG (stop codon) inserted at an amino acid position of interest and the other contains genes for an amber suppressor tRNA and an evolved aminoacyl-tRNA synthetase (aaRS) specific for the unnatural amino acid, pBpa [7]. The tRNA/aaRS pair has been engineered to work with the endogenous translational machinery but are orthogonal to the host system [4]. This means that the tRNA/aaRS pair are specific for only the pBpa and not any of the naturally occurring amino acids. The TAG codon is a natural ribosomal stop signal; however, this system â&#x20AC;&#x153;tricksâ&#x20AC;? the cellular translational machinery to read through the stop as a sense codon (suppression of the stop codon). We site-specifically mutate codons in the histone gene to TAG in order to genetically place pBpa at desired positions in the fully translated protein. The mutant pBpa-histones are expressed in vivo and will incorporate into the native chromatin landscape. Once the pBpa is distributed through the chromatin fiber the cells are exposed to UV light and the protein is activated to crosslink to binding partners (Fig. 2). Following histone crosslinking, protein was extracted out of the cells through TCA precipitation. The crosslinked products were analyzed by gel electrophoresis and western blotting to identify pBpa positions in the histone proteins that crosslinked to other proteins. The histone protein of interest was identified via immunoblotting against a small fusion peptide tag that was introduced to the gene coding sequence on the histone plasmid (the human influenza hemagglutinin, HA-tag, was used). Finally, using this approach we induced DNA damage to understand how the crosslinking would be altered during the response. Alterations in crosslinked patterns insinuate that at the position being probed there is a histone-protein interaction change in response to the DNA damage pathway. This leads to the ability to understand how the change might correlate with PTMs on the histone that could possibly regulate the protein binding.

The Manhattan Scientist, Series B, Volume 4 (2017)

Shi

Methods Yeast culture growth and DNA damage induction The yeast cells used in this study were the BY4741 cellular strain (genotype: MATa his3∆1 leu2∆0 met15∆0 ura3∆0). The plasmids transformed into the cells contains the histone, the system synthetase and its specific tRNA [6]. The histone gene was genetically modified with a fusion tag coding for the HA peptide protein tag. The histone plasmid confers an uracil gene selectable marker and the synthetase plasmid confers a leucine gene selectable marker. In order to maintain all plasmids during the experiment, all cells were cultured in synthetic drop out media minus uracil and leucine, supplemented with a final concentration of 2% glucose (w/v) and 1 mM pBpa. DNA damaging agents were added as described below. A single colony was cultured in 25 mL of media in a 100 mL flask, and grown overnight at 30 ◦ C. The next morning the optical density of the cells was measured by spectroscopy at an absorption of 600 nm (optical density 600, OD600 ). The cell cultures will be diluted back to ∼ 0.8A and grown for designated time with DNA damaging agents, thus the cells will be in healthy growing condition when extracted. Cells were grown in sets so that there were flasks for the control, as well as those that received the DNA damaging drug. The OD600 was measured for each culture and they were normalized so that they had equal densities. This was used to set the equivalent number of cells in each assay equal to each other. The control received no drug. DNA damaging agent was added to produce 0.025% MMS (v/v), and 0.05% MMS (v/v). When the drug was added, the cells were grown two hours at 30 ◦ C, with shaking. The cells were then split into equal portions equivalent to 12 ODs (1 OD is equivalent to a 1 mL culture at A600 = 1.0). The remainder cells grew for two more hours. The exact amount of DNA damaging agent was calculated as such: 25 mL (Total Solution) * 0.00025 (0.025% MMS) = 6.25 µL of MMS.

Figure 2. General scheme for pBpa incorporation into the landscape of chromatin. Using a two plasmid system the pBpa can be genetically inserted site-specifically into the desired histone at the installed stop codon [8].

Shi

The Manhattan Scientist, Series B, Volume 4 (2017)

Crosslinking and protein isolation 12 ODs of cells was collected and then resuspended in 100 muL of water. The cells were then placed on an aluminum plate and exposed to 365 nm of light for 20 min under an ice-cold environment. The cells were mixed every 10 min. Following crosslinking, the cells were collected and then subjected to cell lysis to isolate the cellular proteins. Whole cell lysates were prepared via TCA precipitation. To each sample 1 mL of lysis buffer (1.2% beta-mercaptoethanol, 300 mM NaOH, 1 mM PMSF in water) was added. The sample was incubated on ice for 10 min, and then 160 µL of 50% trichloroacetic acid (TCA) was added, and incubated on ice for 1 more hour. The lysates were centrifuged at 4◦ C at 15000 r.p.m. for 10 min, the supernatant was discarded and the cells were washed with acetone. The wash was repeated one more time and then allowed to air dry at 37◦ C for 30 min. Each culture was resuspended with 100 µL of SDS-PAGE loading buffer and boiled at 95◦ C for 20 min, while shaking at 1500 rpm. The samples were stored at -20◦ C until used. Protein electrophoresis and western blotting Proteins were separated via electrophoresis on homemade 15/12% SDS-PAGE gels in standard tris-glycine buffer (25 mM Tris, 192 mM glycine, 0.1% SDS). Wells were loaded with 5 µL of marker solution (pre-stained Rec protein ladder) or 10 µL of yeast protein samples that were used for immunodetection with anti-HA antibodies. Electrophoresis was performed at 150 V until the denatured proteins resolved the appropriate distance for histone size protein strains to be visible through western blotting. Proteins in the agarose gels were transferred to a PVDF membrane via western blotting in Towbin buffer with 20% ethanol. The transfer was performed at 250 mA for 1 hr. The membrane was washed with water 3 times and then blocked in 3% BSA-TBS solution (for HA and H3S10 ph antibody) for 20 - 30 minutes on a rocker platform. For H4K16 Ac, a 3% milkTBS blocking solution was used. The blocking solution was removed and then the primary (1◦ C) antibody was added and incubated on the membrane overnight at 4◦ C, while shaking at 40 rpm. The primary solutions were as follows: anti-HA at a 1:10000 dilution in 3% BSA-TBS; and anti-H4K16 ac at a 1:3000 dilution in 3% milk-TBS 3% milk. All primary antibodies were raised in rabbit. The next day, the membrane was washed with TBS 3 times, each time for 2-3 min on a rocker platform. The secondary (2◦ C) antibody was then added for 45 min. The secondary antibodies were conjugate to HRP and the solutions were as follows: anti-rabbit at 1:10000 in 5% milk for each of the primary antibodies used. The membrane was then washed with TBS 3 times, each time for 2-3 min on a rocker platform. Then the membrane was washed with 20 mL of 0.1% TBST (0.1% tween in TBS) for 20 min. Finally the membrane was washed with 10 quick rinses of water. Antibodies used: anti-HA (abcam, ab9110), rabbit; anti-H4K16ac (active motif, 39167), rabbit; peroxidase conjugated (sigma, A0545), goat.

The Manhattan Scientist, Series B, Volume 4 (2017)

Shi

Western blot imaging To the membrane, 300 ÂľL of peroxide/luminol solution ECL Select (Amersham) was added, covering the membrane. Images were obtained digitally using Azure C600 western blotting system for about 10 minutes for HA antibody.

Results Multiple sites were assayed in this report thus far, but only a few yield trust worthy data, they are H2A A61 and H3 S22. These histones had the indicated amino acid replaced by pBPA using the described suppression system and were assayed for their crosslinking efficiencies and patterns during the DNA damage pathways. These were visualized by western blotting and immuno-detection (Figs. 3, 4 and 5). We first analyzed the H4K16ac signal for crosslinks from histone H2A A61, H3 T11, and H3 R52 that contained pBpa at its designated position. Crosslinked samples were assayed for histone-protein interaction by detecting bulk histone via antibody detection against the HA-fusion tag on the protein. Fig. 3 shows how the levels of H4K16 acetylation, a post translational modification, were affected by 0.05% MMS. For all three H2A A61, H3 T11, and H3 R52, the wild type cell cultures expressed stronger signals than cultures grown with MMS. The densitometry graph on the right side of the figure was an average of 3 experiment, showing 40 - 50% signal decrease after adding MMS for 2 or 4 hours.

Figure 3. H4K16ac signal test of H2A A61, H3 T11, and H3 R52. Shown on the left, H4K16ac signal decreases in all position two hours after adding MMS. After normalizing the signal strength to the Wt type signal, growing under 0.05% MMS for two hours reduce H4K16 signal by slightly more than 40% while after growing for four hours under the same condition, the signal increase by a few percent.

Row A of Figs. 4 and 5 represents the concentrated bulk of the histone molecules that did not have crosslinking partners. Rows B and C represents signals that are from crosslinks that have occurred to the histone protein of interest. When the histone is covalently bound to a crosslinking partner it migrates slower during electrophoresis due to the increased size of the complexed

Shi

The Manhattan Scientist, Series B, Volume 4 (2017)

interaction. A denser band refers to a stronger signal from the antibody, meaning that more histone or the histone-protein complex is present. The negative UV lane represents cells that were not exposed to crosslinking activation. Crosslinked proteins appear only in the lanes that contain proteins from cells that have been exposed to UV light. As the DNA damaging agent is increased

H2A A61 pBPA WB:HA B

Figure 4. Western blot for yeast cells with H2A A61pBpa. The first two lanes detail the signal of wild type (Wt) yeast cells with and without UV light (negative and positive controls). The second two lanes were yeast cells treated with 0.025%/0.05% of MMS for two hours. The last two lanes were yeast cells treated with 0.025%/0.05% MMS for four hours.

in concentration there appears to be a decrease in crosslinking efficiencies. This is most apparent in row B and C where the strongest signal dominates in the wild type (Wt) cultures versus the damaged cells. In Fig. 4, Wt cell cultures with H2A A61 pBPA that undergo UV light express a crosslink that is missing for the negative control Wt cell culture that did not undergo UV light. At row B crosslink signals significantly decrease for cell cultures growing with MMS for four hours compare to the two-hour cell cultures. There was relatively no change in row B crosslinking between cell cultures growing for 2 hours with MMS and without MMS (the positive control). Fig. 5 shows the crosslink of yeast cell cultures with H3 S22 pBPA. H3 S22 bind to a protein that expresses stronger signal 4 hours growing in 0.025% MMS, while generally signal strength decreases as cell culture grew longer in DNA damaging agent. This band noted with asterisk only exist in the cell culture with lower MMS concentration. In both row B and C, cell cultures grew in 0.025% MMS for 4 hours express significant increase of crosslinking signal.

Discussion H2A results were significant because it resides in a surface exposed region of the nucleosome. This region is referred to as the H2A acidic patch and is essential as a docking site for nucleosomal proteins. This domain has been shown to be a key point of regulation at the nucleosomal

The Manhattan Scientist, Series B, Volume 4 (2017)

Shi

C B A

Figure 5. Western blot for yeast cells with H2A S22 pBpa. The blot was visualized with primary HA antibody. The first two lanes show the signal of wild type (Wt) yeast cells with and without UV light. The second two lanes were yeast cells treated with 0.025%/0.05% of MMS for two hours. The last two lanes were yeast cells treated with 0.025%/0.05% MMS for four hours.

surface. Protein interactions with the H2A acidic patch regulates gene expression in response to environmental change. Our results suggest that under levels of increased DNA damaging agent (stressed conditions) the ability of the histone to interact with nuclear proteins is reduced. There is a decrease in contacts between histone and the protein it crosslinks to, at least for some of the contacts we detect. H4K16 acetylation blotting was utilized to access the opened versus closed structure of chromatin. Increased acetylation means increased accessibility of chromatin for binding proteins. DNA damage pathways are thought to quickly alter chromatin structure so that the DNA becomes easily accessible for repair enzymes to act on the damaged loci. H4K16 ac western blot for H2A A61, H3 T11 and H3 R52 shows that 0.05% of DNA damaging agent significantly decrease acetylation, suggesting that normal cell functions were minimized to prevent synthesizing proteins with damaged DNA. Comparing to the 2-hour culture, H4 K16 acetylation increased in the 4-hour culture. This suggests that some DNA damage had been repaired, so some cell’s normal function could have resumed. There are other logical explanations for this phenomenal, such that the normal cell functions had paused for too long and the chromatin must become more active for the cell to survive. So, it is possible that the yeast cells’ self-repairing process may not be sufficient to repair the damaged done and therefore the cells can no longer regulate the H4K16ac properly. Generally, as DNA gets damaged, most cell functions are reduced. This was expressed by a decrease in the signal strength of the band. In Fig. 4, the results follow this pattern. Crosslinks in

Shi

The Manhattan Scientist, Series B, Volume 4 (2017)

row B significantly drops after the cells grew with MMS for 4 hours. It is interesting that even the cell cultures grew with 0.05% MMS, double the amount of low MMS cultures, also took 4 hours for the crosslink to decrease. This suggests that this amount of DNA damage is tolerable for the normal day functions that this crosslink represents. This suggestion became more interesting when comparing to H3 S22 western blot. In Fig. 5, H3 S22 pBPA western blot shows that there was a specific protein that begin to interact at the position only 4 hours after cells grew in 0.025% MMS solution. It suggests that this specific protein interacts with H3 S22 to repair the DNA damage, while 0.05% MMS does not express the same crosslink. It is possible that 0.05% MMS deals too much damage for the cells to repair with that specific protein. Comparing H2A A61 and H3 S22 blots, they suggest that although the cell cultures were all damaged by MMS, different amount of damage require different DNA repairing mechanism. Although this is more likely, other logical explanations are possible. Such that cells undergo too much damage, that even though yeast cells could repair 0.025% MMS damage, it cannot repair the damage done by 0.05% MMS. To proceed toward the initial goal of this experiment, i.e. to reveal the mechanisms that shape chromatin structure under DNA damage, this experiment must be repeated at other significant positions surrounding the histone octamer. These positions are typically the sites that are most likely to come in contact with other proteins that bind onto the histone. We will begin to isolate the protein complex marked with asterisk in figure 5 and do a mass spectrometry to identify it. Mass spectrometry is necessary to reveal the actual proteins that crosslink with the histone on the western blots. Only when the structures and the functions of these proteins are known, the mechanisms can be understood.

Acknowledgement This work was funded by the School of Science Research Scholars Program. The author would like to thank his advisor, Dr. Bryan Wilkins, for his guidance and support through this project.

References [1] Luger, K., Mader, A. W., Richmond, R. K., Sargent, D. F. and Richmond, T. J. Crystal structure of the nucleosome core particle at 2.8 A resolution. Nature 389, 251-260 (1997). [2] Tsabar, M. and Haber, J. E. Chromatin modifications and chromatin remodeling during DNA repair in budding yeast. Curr. Opin. Genetics Dev. 23, 166-173 (2013). [3] Fontana, L., Partridge, L. and Longo, V. D. Extending healthy life span - From yeast to humans. Science 328, 321-326 (2010). [4] Liu, C. C. and Schultz, P. G. Adding new chemistries to the genetic code. Annu. Rev. Biochem. 79, 413-444 (2010). [5] Dorman, G. and Prestwich, G. D. Benzophenone photophores in biochemistry. Biochemistry 33, 5661-5673 (1994).

The Manhattan Scientist, Series B, Volume 4 (2017)

Shi

[6] Wilkins, B. J., Rall, N. A., Ostwal, Y., Kruitwagen, T., Hiragami-Hamada, K., Winkler, M., Barral, Y., Fischle, W. and Neumann, H. A cascade of histone modifications induces chromatin condensation in mitosis. Science 343, 77-80 (2014) [7] Chin, J. W., Cropp, T. A., Anderson, J. C., Mukherji, M., Zhang, Z. and Schultz, P. G. An expanded eukaryotic genetic code. Science 301, 964-967 (2003). [8] Shi, B. Exploring chromatin dynamics within the DNA damage response pathway in living cells. An expanded eukaryotic genetic code. Manhattan Scientist B3, 19-28 (2016).

Nucleosomal dynamics in living Saccharomyces cerevisiae (yeast) cells Amanda Zimnochâ&#x2C6;&#x2014; Department of Chemistry and Biochemistry, Manhattan College Abstract. There has been extensive research on chromatin and chromatin related proteins, however, there has been limited research on how chromatin behaves and interacts within living cells. The goal of this research was to examine the behavior of nucleosomes inside the living yeast nucleus through an approach that monitors protein-protein interactions. We aimed to genomically tag genes known to have chromatin functionality with short protein fusions to better understand the spatiotemporal contacts these proteins have at the nucleosomal interface. Successfully tagged genes were expressed and assayed for nucleosomal interaction in vivo by crosslinking from histone proteins that were site-specifically mutated with the photo-crosslinking unnatural amino acid, p-Benzoylphenylalanine (pBpa), through a genetic expansion technique in Saccharomyces cerevisiae (yeast). We successfully tagged multiple chromatin related proteins and are currently working to identify histone-protein interactions.

Introduction Histones are small basic proteins that bind into octomeric units around which DNA is associated to form nucleosomes. This is the core mechanism of eukaryotic DNA compaction. There are 4 core histones (H2A, H2B, H3, and H4) all of which possess net positive charges that allows negatively charged double-stranded DNA to wrap around them. The DNA wraps around a histone octamer â&#x2C6;ź1.7 times. The nucleosome represents the most basic repeating unit for chromosomes wherein they resemble beads on a string in the lowest higher-ordered structuring, the 10 nm fiber. Further wrapping and condensing occurs to facilitate successive levels of higher ordered structuring until they ultimately condense into the classic X-shaped chromosomes that most people are familiar with. While ground-breaking work has provided crystal structures for the core nucleosome, there is little experimental evidence for an accepted model of higher level chromatin organization [1]. Structurally, the nucleosome creates a barrier in the cells that must be regulated in order to provide access to proteins and enzymes required in DNA related mechanisms. While the wrapping of DNA around the nucleosome is essential for DNA compaction it also establishes that DNA as inaccessible. The DNA that contacts the nucleosome is not free to be accessed by enzymes such as polymerases that are critical for replication and translation. It is not very well understood how chromatin dynamics are regulated and how the cell controls compaction versus accessibility on the microsecond time scale. In order to fully understand these dynamics a technique must be developed that allows for the spatiotemporal monitoring of chromatin related proteins that allow for mechanistic insights into their actions. â&#x2C6;&#x2014;

Research mentored by Bryan Wilkins, Ph.D.

The Manhattan Scientist, Series B, Volume 4 (2017)

Zimnoch

Our work utilizes a method that allows for protein-protein interaction monitoring in living cells by introducing unnatural amino acids (UAA) into proteins. This process, referred to as an expanded genetic code, hijacks the transcriptional/translational pathways of the cell to direct UAAs sitespecifically into a protein of choice [2]. Since histone proteins are at the core of the nucleosome we direct a photo-crosslinking UAA into histone proteins to trap histone-protein interactions at the nucleosomal surface. We utilized the UAA p-benzoylphenylalanine (pBPA) because, when it is activated with 365 nm light, it will covalently bind to neighboring protein contacts within 0-0.4 nm [3]. Once the crosslinks are created they are stable and irreversible, allowing for isolation and characterization under harsh conditions. This technique becomes a powerful tool for identifying protein contacts when the protein of interest is labeled with a protein tag. We selected proteins which have been known to have chromatin functionality and used a polymerase chain reaction (PCR) method to install tags at the genomic level [4, 5]. We chose to PCR amplify a selection cassette containing a MYC-tag and the HIS3-M6X gene (Fig. 1). The primers tagging resulted in the ability of the cells to grow on agar plates which lacked histidine. The proper tagging of these genes was verified via PCR and protein expression of the target gene. Proteins were visualized by western blotting, and antibody recognition. Strains containing the tagged proteins were then assayed for nucleosomal interaction in vivo by crosslinking from histone proteins that contained pBpa [6]. Our protein targets included the proteins Snf2 and Rsc1. Each of them has a role in a chromatin remodeler complex. These large complexes act to move, slide and alter nucleosomal architecture. They have essential roles in the release and manipulation of DNA accessibility and compaction. While we understand much about their functions we know very little about how they contact the nucleosome itself. We targeted these two proteins as a first attempt at establishing a crosslinking technique to visualize chromatin remodeler nucleosome contacts. Research on nucleosomes and their dynamics is extremely valuable because of their significance in cellular integrity and health. Nucleosomes regulate various chromosomal roles such as chromatin remodeling, gene silencing, transcriptional activity, and replication, all of which are essential components of life.

Materials and Methods

The generalized procedure for the genomic mutation experiment was as follows: (1) Polymerase Chain Reaction of the myc-HIS3M6X cassette (2) verification of PCR results on agarose gels (3) ethanol precipitation and isolation of PCR products (4) yeast cell transformations with PCR products (5) cell culturing (6) genomic DNA purification (7) genomic mutation PCR purification [5]. After cells were verified for the tagging of interest, we then probed the expressed mutant proteins by western blotting and antibody detection of the myc tag. First, cells were grown and the cellular proteins were isolated via TCA precipitation and separated by an 8% SDS-PAGE protein gel. The proteins were then transferred to a solid membrane and antibodies were used to bind to

Zimnoch

The Manhattan Scientist, Series B, Volume 4 (2017)

Figure 1. General scheme for homologous for this PCR were also specifically designed to contain 5’-extensions that contained homology to the 3’-end of the gene target. PCR products were then transformed into yeast cells and proper recombination. The S2 and S3 primers were specially designed to contain homology regions specific to a gene of interest (YFG). The S3 primer was designed to anneal to the 3’-end of a target gene and the S2 primer was designed to anneal to the 3’-genomic region of the gene. Each of the primers also contained regions that annealed to a reporter cassette that amplified the sequences of the 3-myc tag and the HIS3M6X gene.

the myc tag that was introduced to the protein. The antibodies used horseradish peroxidase and ECL substrate to visualize the protein of interest. When we verified a mutant yeast strain, we then performed yeast transformations with two plasmids carrying genes needed for the proper insertion of pBPA into histone proteins. Cells were grown to exponential phase and then exposed to UV-light in an attempt to crosslink histones to the tagged protein. Polymerase chain reaction Polymerase chain reaction is a procedure which can produce millions of copies of a targeted segment of DNA from just one original [4]. We modified the polymerase chain reaction by using primers (S2 and S3) that were specially designed so that they contained 3’-homology to the amplification target from a plasmid (tag and marker) and also contained 5’-homology regions specific

The Manhattan Scientist, Series B, Volume 4 (2017)

Zimnoch

for the genomic sequences that flanked the intended genomic insertion site of the tag. The PCR product therefore yields the tag and selected marker with 5’-homologous regions that will bind to the genome of yeast. The myc-tag that was inserted would allow for the unnatural amino acid to bind with it later on, and the marker chosen was the HIS3M6X gene, which would allow cells to produce their own histidine. The samples were then run on an agarose gel at 100V for 45 minutes. The PCR products were transformed into cells to initiate genomic tagging. The cells were then plated on agar plates deficient of histidine because proper recombination also inserts the HIS3 gene, allowing the cells to produce their own histidine. We then looked for our tagged proteins in whole cell lysates via western blotting and antibody visualization. After preparing the cellular lysates, each of the proteins were separated by size through gel electrophoresis on an 8% Trisacetate gel. The proteins were transferred from the gel onto a membrane to provide solid support for antibody binding. The membrane was washed and then incubated in antibodies specific to the myc-tag of interest. A photo was then taken using a CCD imager. Ethanol precipitation of DNA The DNA from three PCR reactions per gene of interest was pooled together and then the total DNA was precipitated and concentrated into 10 µl of water. The entirety of the PCR reactions was ethanol precipitated by adding 1/10 volume of 3M sodium acetate (pH 5.2), 2-2.5 volumes of 100% ethanol, and placed on ice for 20 minutes. After the incubation, the samples were centrifuged for 10 minutes at maximum speed after which the supernatant was discarded. 1 mL of ethanol was added, and the samples were centrifuged once again and decanted. The samples were left to air dry for about 30 minutes, until they were visibly dry. Yeast transformation of homologous PCR products and selection Yeast cells were transformed using a standard heat shock lithium acetate protocol. Wild type yeast cells, BY4741 (genotype: MATa his31 leu20 met150 ura30), were grown to an OD600 = 1.0, 50 mL total. The cells were collected by centrifugation and then washed in sterile water. Following another centrifugation round, the water was discarded and the cells were re-suspended in 1 mL of competent cell buffer (100mM lithium acetate, 10 nM Tris-HCl, pH 7.5, and 1 mM EDTA in water). The transformation reactions were as follows: 10µL concentrated DNA from PCR, 10 µL single stranded (carrier) DNA, 100 µL competent cell solution, and 700 µL PEG solution (100 mM lithium acetate, 10 mM Tri-HCl, pH 7.5, and 1 mM EDTA in 40% PEG 3350). The transformations were incubated at 30◦ C for 30 min and then 80 µL DMSO was added. The cells were heat shocked at 42◦ C for 10 min and then washed with sterile water. Following the wash, the cells were re-suspended in 100 µL water and plated on SC agar plates minus histidine. The plates were incubated at 30◦ C for 2-3 days. A negative control was also performed in which no PCR DNA was added to the reaction. Individual cell colonies that grew in the absence of histidine were streaked on new agar plates to screen for potential false positives. Colonies that survived the second round of screening were

Zimnoch

The Manhattan Scientist, Series B, Volume 4 (2017)

grown in full medium (YPDA, glucose 2%) to an OD600 >1.0, 25 mL total. 12 ODs were collected and the cells lysates were prepared from whole cells via TCA precipitation. TCA precipitation of whole cells Precipitations were performed using 1 mL of lysis buffer (1.2% beta-mercaptoethanol, 300 mM NaOH, 1 mM PMSF in water) per sample. All samples were incubated on ice for 10 minutes, then 160 µL of 50% (w/v) TCA was added and samples were once again incubated on ice for 20 minutes. The samples were centrifuged at full speed for 10 minutes, the supernatant was discarded, and cells were washed with acetone. The centrifugation step, decantation and washing of cells with acetone were repeated once again. The samples were then left to air dry for 10 minutes or until visibly dry. The pellets were re-suspended using 100 µl of 2X SDS buffer and placed in a sonicator on with heat and left in for 10 minutes. Samples were placed in a boiling apparatus at 92◦ C and 1500 rpm for 10 minutes and then centrifuged at max speed for 5 minutes. Electrophoresis and Western blotting In general, protein samples were analyzed on NUPAGE 8% SDS-PAGE gels and run in MOPS buffer (50 mM MOPS, 50 mM Tris, pH 7.5, 3% SDS, 1 mM EDTA) for 2-3 hours at 150 V. Western blots were performed in transfer buffer containing 10% methanol (chemical composition of transfer buffer), using PVDF membrane, and transferred at a constant mA of 250 for 2-3 hours. Following the transfer, the membrane was washed and then incubated in a blocking solution (3% milk and 0.05% Tween-20 in TBS). Primary antibody incubation (1◦ C anti-myc, mouse, 1:1000 dilution in blocking solution) was performed overnight at 4◦ C, and the next day the membrane was washed twice in TBS. The membrane was then incubated in the secondary antibody solution (2◦ C anti-myc rabbit HRP conjugated, goat, 1:1000 dilution in 5% milk, 0.05% Tween-20 in TBS) for 45 minutes, then washed twice in TBS, and finally washed in freshly made TBS with 0.1% Tween-20 for 10 minutes. The wash solution was removed with ten quick rinses with deionized water. Imaging of the membrane was performed using The Omega LumTM G Imaging System set to Chemi for 15-45 minutes per image (dependent on the intensity of the signal). Amersham ECL Select substrate was used to visualize the proteins (∼400 µL substrate per membrane). The solution was evenly distributed on top of the membrane prior to imaging.

Results and Discussion The initial PCR for RSC1 appears to have worked, because the PCR product using the S2/S3 primer pair reveals bands at approximately 1500 b.p. in length, which is the expected size of the amplified cassette (Fig. 2). Following PCR purification and transformation into yeast, we checked for the proper insertion of the PCR product into the genome by plating the transformations on growth agar that lacked histidine. From the colonies that survived, we isolated the genomic DNA to PCR verify that

The Manhattan Scientist, Series B, Volume 4 (2017)

Zimnoch

the mutation was inserted at the proper chromosomal loci. If the PCR product were inserted into the genome at the wrong loci, then cells would still confer selection on (-)Histidine plates; however, these would be false positives. Verification primers were designed with binding targets inside the gene of interest and in the genomic region just downstream of the gene. A properly mutated gene has the 1500 bp cassette added to its 3â&#x20AC;&#x2122;-end, and therefore will result in a larger PCR product as compared to the wild type gene. Fig. 2 shows that there is a large increase in bp size from the wild type signal (750 b.p.) and all the RSC1 signals. We then grew cells from the yeast strains that yielded positive mutational verification and performed western blotting analysis on their cell lysates with an anti-myc antibody (Fig. 5). The initial S2/S3 PCR for SNF2 was previously performed by Gabriela Bukanowska. The verification PCR for SNF2 +1 appears to have worked due to the large jump beFigure 2. Agarose separation tween the signal compared to the wild type (Fig. 4). The negof S2/S3 PCR for amplification ative and positive signs on the clone counts refer to the fact of the 3Myc-HIS3MX6 cassette that we observed cells that grew on a negative control. The of RSC1. The clone number is an arbitrary number assigned to negative clones showed no positive insertion of the mutation, the sample for identification post as expected. When we ran Snf2 +1 and two Rsc1 samples on analysis. a western blot, along with positive control Bdf1-3myc, there is was a positive signal for Snf2, and there was a very weak signal for Rsc1 #12. We believe that the Rsc1 #12 signal is a true signal because the size of the protein shift, as compared to Snf2 and Bdf1 is correct. Snf2 has a molecular weight of approximately 195 kDa, Rsc1 is 107 kDa and Bdf1 is 90 kDa. While the marker is not depicted in the figure it is clear that the sizes of the proteins are of these relative size separations. We plan to perform the Rsc1 western blot again to better verify this signal and clearly establish the protein was indeed tagged. We successfully genomically tagged the genes for SNF2 and RSC1 with the coding sequence for a myc tag. These genomic mutations then ultimately were expressed in living cells that produced full-length protein containing the short protein fusion. These short tags were visualized by western blot analysis using antibodies that recognized the tag. The strains that were successfully mutated were then transformed with plasmids that harbored the genes needed to produce histone proteins with the unnatural amino acid, pBPA. The goal was to then scan the crosslinking probe through the N-terminal region of histone H3 and determine if either Snf2 or Rsc1 protein could be captured in a histone-protein interaction. We were unsuccessful at transforming theses strains with the UAA machinery, however we will continue to do this in an effort to get the crosslinking to work. Once this is established the

Zimnoch

The Manhattan Scientist, Series B, Volume 4 (2017)

Figure 3. Agarose separation of S2/S3 verification PCR for amplification of the 3MycHIS3MX6 cassette of RSC1. The clone number is an arbitrary number assigned to the sample for identification post analysis. Wt indicates wild type cells with no recombination attempts.

Figure 4. Agarose separation of S2/S3 verification PCR for amplification of the 3Myc-HIS3MX6 cassette of SNF2. The negative signs refer to the fact that we observed cells that grew on a negative control. Wt indicates wild type cells with no recombination attempts.

Figure 5. Western blot analysis of successfully tagged genes. Bdf1 was used as a positive control to aid in the identification of the samples. Snf2 and Rsc1 indicate the genes that were targeted.

next step in this research is to attempt crosslinking these proteins with the photoreceptive unnatural amino acid, p-Benzophenylalanine (pBpa). If we can identify the binding of these proteins to the histone we can then scan the entire nucleosomal surface to provide a low resolution mapping of the Snf2/Rsc1 chromatin interactome, in living cells. In the future, this experiment can be duplicated with any other genes or proteins in order to expand the knowledge in this field.

Acknowledgment This work was funded by the School of Science Research Scholars Program. The author would like to thank Dr. Bryan Wilkins for his mentoring of this work.

References [1] Luger, K., M¨ader, A. W., Richmond, R. K., Sargent, D. F. & Richmond, T. J. Crystal structure of the nucleosome core particle at 2.8 A resolution. Nature 389, 251–260 (1997) [2] Liu, C. C. and Schultz, P. G. Adding new chemistries to the genetic code. Annu. Rev. Biochem. 79, 413–444 (2010) [3] Dorman, G. and Prestwich, G. D. Benzophenone photophores in biochemistry. Biochemistry 33, 5661–5673 (1994)

The Manhattan Scientist, Series B, Volume 4 (2017)

Zimnoch

[4] Schochetman, G., Ou, C.-Y., and Jones, W. K. “Polymerase Chain Reaction.” The Journal of Infectious Diseases, vol. 158, no. 6, 1154–1157 (1988) [5] Knop, M., Siegers, K., Pereira, G., Zachariae, W., Winsor, B., Nasmyth, K., and Schiebel, E. Epitope tagging of yeast genes using a PCR-based strategy: More tags and improved practical routines. Yeast 15, 963–972 (1999) [6] Wilkins, B. J., Rall, N. A., Ostwal, Y., Kruitwagen, T., and Hiragami-Hamada, K. A cascade of histone modifications induces chromatin condensation in mitosis. Science 343, 77-80 (2014)

Predicting rates of bark formation on saguaro cacti (Carnegiea gigantea) Mia Bertoliâ&#x2C6;&#x2014; Laboratory of Plant Morphogenesis, Department of Biology, Manhattan College Abstract. Saguaro cactus (Carnegiea gigantea) is a species native to the Sonoran Desert that exhibits epidermal browning, or barking, on all twelve cactus surfaces as a result of sun exposure. Bark formation causes premature death from the average 200 to 300-year life span. Species were sampled in 1994, 2002, 2010, and 2017 to examine bark progression over time. The hypotheses for this research were (1) the sum of bark formation on all twelve surfaces predicts cactus death, (2) bark formation on south-facing surfaces predicts cactus death, and (3) cactus outliers with bark formation rates slower than predicted are shaded. The machine learning program WEKA was used to create a decision tree of surfaces that were the best predictors of cactus death in 2017. The computer program MATLAB determined the outliers, or the surfaces that barked slower or faster than expected. These outliers were compared to images showing the amount of shading surrounding each surface. The results of this research indicate that the sum of all twelve surfaces is a main predictor of cactus death with 100% and 72% accuracy. When the sum is not used, the south-facing surfaces are a main predictor of cactus death with 72% and 64% accuracy. The cactus outliers that barked slower than predicted are more shaded than the outliers that barked faster than predicted. Computer programs WEKA and MATLAB provide excellent tools to predict cactus death and bark formation rates over time.

Introduction Saguaro cactus (Carnegiea gigantea) is a keynote species in the Sonoran Desert of Arizona and Mexico (MacMahon,1985). Saguaro is a dominant type of cacti found in areas of diverse species. The Sonoran Desert is a unique desert since it is the only desert in the world that receives precipitation in both summer and winter seasons. Although a wide range of vegetation occurs in this desert, saguaros, tall long-lived cacti, can live for hundreds of years (Steenbergh and Lowe, 1997). Recent research has shown that saguaros no longer live to be above 150 years (Evans et al., 1994a; 1994b). Recent evidence shows that epidermal browning, the presence of bark on the surfaces of saguaros, leads to premature death of the plants (Evans et al., 2005). Bark formation is not a normal process for most cactus plants (Anderson, 2001) so the presence of bark on saguaros has occurred relatively recently (Evans et al., 1992). Based on previous research, UV-B exposure is the cause of epidermal browning on cactus plants (Evans et al., 2001). For cactus plants to survive, they must have functioning stomata that perform gas exchange. Stomata are found on cactus surfaces, therefore extensive barking blocks the stomata, which prevents photosynthesis and respiration. Bark formation causes premature death from the average 200 to 300-year life expectancy. Saguaro cacti contain multiple surfaces that bark starting on south-facing surfaces and ending on north-facing surfaces, since south-facing surfaces are exposed to more sunlight. Bark formation occurs at a rate of 2.3% per year, and is more prevalent now than in the past. â&#x2C6;&#x2014;

Research mentored by Lance Evans, Ph.D.

The Manhattan Scientist, Series B, Volume 4 (2017)

Bertoli

The purpose of the current project is to determine rates of bark formation over time. Data for this research spans over a 23-year period, starting in 1994 to present. Cactus mortality and rates of bark formation over time were predicted for cacti in 2017. These analyses provide information for how individual cactus environment affects barking rates over time and if individual cacti are barking faster or slower than expected. The hypotheses for this research were (1) the sum of bark formation on all twelve surfaces predicts cactus death, (2) bark formation on south-facing surfaces predict cactus death, and (3) cactus outliers with bark formation rates slower than predicted are shaded. These hypotheses allow for analysis of the surrounding characteristics of cactus plants that influence bark formation rates over time.

Materials and Methods Field conditions Data were taken from Tucson Mountain Park in Tucson, Arizona over a 23-year period, starting in 1994. In 1994, 1149 saguaro cacti (Carnegiea gigantea) (Engelm) in 50 plots were selected for analysis (Evans et al., 1995). The same cacti in the same plots were evaluated again in 2002, 2010, and 2017. These data sets were used to predict cactus death and rates of bark formation over the 23-year period. Data sets used Saguaro cacti have multiple surfaces, however, only twelve individual surfaces were studied. The twelve surfaces evaluated were of crests and troughs facing north, south, east, and west that barked at various rates. Each crest, right trough, and left trough for each cardinal direction were evaluated for the percentage of green. The percentage of green, or healthy areas, on each of the twelve individual surfaces was determined in the field. These data were entered into an Excel file, which were converted to percent bark, which was to be entered into various programs for analysis. Bark is a thick outer covering on surfaces as a result of sun exposure. Bark prevents adequate gas exchange resulting in premature death. Entering these data into various programs provides information for bark progression rates over time. Thus each line of data had the plot number, the cactus number, and year of data and then bark percentages for the twelve surfaces. WEKA WEKA 3.8 is a Machine Learning program that is used to make predictions with large data sets (Witten et al., 1999). WEKA uses data sets that can be put into categories to make predictions. WEKA decides if each individual datum should be put into one category or a second category, and then determines the accuracy of that prediction. Data entered into WEKA consisted of cactus bark percentages for each cactus in 1994, 2002, 2010, and 2017, and the category “dead” or “alive.” For the current study, data of bark percentages for each cactus surface from 1994, 2002 and 2010 were used to predict cactus death in 2017. Data for 1160 cacti for three samples in 1994, 2002, and 2010 gave a total of 3,480 used to predict death in 2017. For each cactus, the program predicted alive or

Bertoli

The Manhattan Scientist, Series B, Volume 4 (2017)

dead and provided the accuracy of prediction of alive or dead vs. actual alive and dead from 2017 observations. The first aspect of this research was to determine if the sum of the amounts of bark on all twelve surfaces from 1994, 2002, and 2010 would accurately predict cactus death in 2017. WEKA analyzed these data and determined the best course of action to most accurately predict cactus death. WEKA can use just one surface or multiple surfaces in succession to determine cactus death. First, bark percentages of all twelve surfaces for each individual cactus were added to obtain the sum of bark percentages. This information, along with bark percentages for individual surfaces, was entered into WEKA. WEKA determined the surface threshold to predict cactus death. The second aspect of this research was to determine if the amount of bark on south-facing surfaces from 1994, 2002, and 2010 would accurately predict cactus death in 2017. WEKA was used as the predictive entity for cactus death. Bark percentages of all twelve individual surfaces from 1994, 2002, 2010, and 2017 were entered into WEKA. WEKA determined the best surface or surfaces to predict cactus death. MATLAB MATLAB is a computer program that uses large data sets to generate figures or make predictions (Gilat, 2004). In this research MATLAB determined if amounts of bark on all twelve surfaces from 1994, 2002, and 2010 would accurately predict bark percentages in 2017. Data entered into MATLAB was cactus plot number, cactus number, year, and bark percentages on each surface. MATLAB used this data to generate a histogram using the program “validate model” created by DeBonis (2015). “Validate model” generated a histogram with the number of samples vs. the percent error. The percent error was the percent difference between the value of bark percentage predicted by MATLAB and the actual value of bark percentage observed in the field. The MATLAB program “leave one out” created by DeBonis (2015) provided a numerical representation of the histogram. “Leave one out” provided values of the percent error for each individual cactus. These values were negative or positive, and the values 2 times the standard deviation were considered outliers. The negative outliers meant that the observed values were less than predicted values, the cacti barked slower than expected. The positive outliers meant that the observed values were greater than predicted values, the cacti barked faster than expected. It was hypothesized that cactus outliers that barked slower than predicted would be more shaded and cactus outliers that barked faster than predicted would be less shaded.

Results The first aspect of this research was to determine if the sum of bark percentages on all twelve surfaces could predict cactus death. The WEKA program determined the best surface or surfaces to predict cactus death. WEKA determined that if the sum of all twelve surfaces was greater than 1193%, cacti were deemed to be dead with 100% accuracy (Fig. 1).

The Manhattan Scientist, Series B, Volume 4 (2017)

Bertoli

SUM Figure 1. The sum of rates of bark formation on all twelve surfaces predicts cactus death with 100% accuracy (WEKA)

>1193

≤1193

Dead

Alive

The second aspect was to lower the sum threshold to SUM 772%. At this sum threshold, WEKA determined that the sum was the best predictor of cactus death (Fig. 2). Cacti with sums less than 772 SUM selected east left trough as were deemed to be alive. When the sum threshold≤1193 was > 772,>1193 WEKA the best predictor. If the east left trough was ≤ 95, cacti were alive. If the east left trough values Dead percentages ≤ 98, cacti were Alive crest had bark were > 95, WEKA selected east crests. If the east >772 ≤772 alive while if east crests were > 98, cacti were dead. Overall, WEKA was 72% accurate (Fig. 2). EL

Alive SUM >772

≤772

Alive

Figure 2. The sum of rates of bark formation on all twelve surfaces predicts cactus death with 72% accuracy (WEKA)

>95

≤95

Alive >95

Alive

≤98 ≤98

Alive

>98

Alive

Dead

The third aspect was to predict cactus death using only data from individual surfaces of the twelve surfaces available. WEKA determined that south right troughs were the best single predictor of cactus death. If south right troughs had bark percentages > 70, cacti were dead. If south right troughs had bark percentages ≤ 70, then WEKA selected north right troughs as the next best candidate. If north right troughs had bark percentages ≤ 15, cacti were alive. If north right troughs had bark percentages > 15, then WEKA selected west crests. If west crests had bark percentages ≤ 35, cacti were deemed alive. If west crests had bark percentages ≥ 35, WEKA chose west right troughs. If west right trough had bark percentages ≤ 7, cacti were alive, but if bark percentages

Bertoli

The Manhattan Scientist, Series B, Volume 4 (2017)

> 7, cacti were dead. Overall, this WEKA procedure was 64% accurate (Fig. 3). SR

>70

≤70

Dead

Figure 3. Bark percentages on the south right trough predicts cactus death with 64% accuracy (WEKA)

>15

≤15

Alive

WC ≤35

>35

Alive

WR >7

≤7

Alive

Dead

The fourth aspect was to predict cactus death from individual surfaces; south-facing surfaces were again determined to be the main predictor of cactus death. WEKA determined that south left troughs were the best single predictor of cactus death. If south left troughs had bark percentages of ≤ 95 WEKA selected north left troughs. If north left troughs had bark percentages ≤ 45, WEKA selected west left troughs, if north left troughs had bark percentages > 45, WEKA selected north crests. If south left troughs had bark percentages > 95, WEKA proceeded to north crests. If north crests had bark percentages ≤ 85, WEKA selected north left troughs. If north crests had bark percentages > 85, WEKA selected west right troughs. This continued until cacti were deemed dead or alive. Overall, this WEKA procedure was 72% accurate (Fig. 4).

>95

≤95

≤45

>45

NL ≤85 NL

>85

Figure 4. Bark percentages on the south left trough predicts cactus death with 72% accuracy (WEKA)

The fifth aspect was using the program MATLAB, which predicted expected rates of bark formation using data from previous years. The program “Validate model” created a histogram, as shown in Fig. 5, of the percent error between predicted and observed rates of bark formation. The

The Manhattan Scientist, Series B, Volume 4 (2017)

Bertoli

program â&#x20AC;&#x153;leave one outâ&#x20AC;? provided numerical data of the histogram. Negative values indicated that a cactus barked slower than predicted. Positive values indicated that a cactus barked faster than predicted. Cactus outliers that barked slower than predicted were shaded two times more often than cactus outliers that barked faster than average (Table 1).

Figure 5. Histogram with number of cactus samples vs. percent error

Table 1. Comparison of Mean Error Values Among the Cactus Groups

Shaded Slow Unshaded Slow Shaded Fast Unshaded Fast Average

Number of samples

Mean error value

Percent shaded

40 70 20 118 3950

-47.0 -46.6 +48.6 +47.4 1.2

36* 64* 15* 85*

*comparing slow with fast cacti all differences were significant at p < 0.001

Discussion Saguaro cacti exhibit epidermal browning as a result of sun exposure, which causes premature death. The purpose of this research was to determine how barking on specific surfaces predicts cactus mortality and if surrounding vegetation characteristics influenced bark formation rates. The sum of bark percentages on all twelve surfaces was the most accurate predictor of cactus death. There is a threshold of bark percentages surrounding the surfaces evaluated and that area cannot perform gas exchange if it exceeds that threshold. The sum of bark percentages of 1200 means that the entire cactus has 100% bark, however, WEKA determined that if the sum was greater 1193, all cacti were deemed dead. Therefore, the sum of the cactus surfaces needs to be less than

Bertoli

The Manhattan Scientist, Series B, Volume 4 (2017)

1193 for that area to perform adequate gas exchange. When that threshold is lowered to 772, other surfaces were viewed to obtain the absolute threshold of bark percentages for cactus death. Once that absolute threshold is reached, the surfaces evaluated are unable to perform gas exchange. South-facing surfaces are the first surfaces to exhibit epidermal browning because they are exposed to more sunlight. Once the south-facing surfaces start barking, the other surfaces follow. If the sum of all surfaces was not used for evaluations, WEKA determined that south-facing surfaces were the main predictor of cactus death. The threshold of bark percentages on south right troughs needs to be greater than 72% to be determined dead. If bark percentages are less than that threshold, then subsequent surfaces need to have specific bark percentages in order for individual cacti to be deemed dead. Once the predicted surfaces are barked to a certain extent, the plant cannot perform photosynthesis or respiration. These analyses provide information on how the orientation of the cactus in relation to the sun affects the barking order on cactus surfaces and eventual cactus death. Overall, WEKA was an accurate tool to predict bark thresholds for cactus mortality. The computer program MATLAB provided analysis of predicted bark percentages in 2017. The percent error analyzed by â&#x20AC;&#x153;leave one outâ&#x20AC;? provided the outliers of the cactus surface being predicted. The negative and positive values of double the standard deviation were compared to pictures of cacti in the field to observe the surrounding characteristics of the individual cactus. The negative cactus outliers that barked slower than predicted were shaded two times more often than positive cactus outliers that barked faster than predicted. This further provides evidence that the sun is the cause of epidermal browning and those cacti that are shaded by surrounding vegetation will live longer due to less sun exposure. MATLAB was an adequate tool to predict expected bark percentages in 2017. Potential future research is to predict cactus death and bark formation rates over time in previous years and compare this data to the results found in 2017. This can provide analysis of how barking has changed per individual cactus over the years. If the cacti that were outliers in 2017 were average in the previous years, it is important to determine what changed leading up to 2017 that caused them to bark at different rates. Surrounding characteristics should continue to be analyzed to determine how it affects bark formation rates over time.

Acknowledgment This work was funded by the Catherine and Robert Fenton Endowed Chair to Dr. Lance Evans.

References Anderson, E.F. 2001. The Cactus Family. Timber Press, Portland, OR, p. 776. DeBonis, M. 2015. Unpublished - private communication Evans, L.S., K.A. Howard and E. Stolze. 1992. Epidermal Browning of Saguaro Cacti (Carnegiea gigantea) J. Torrey Bot. Soc. 142: 231-239.

The Manhattan Scientist, Series B, Volume 4 (2017)

Bertoli

Evans, L.S., V.A. Cantarella, K.W. Stolte and K.H. Thompson. 1994a. Phenological changes associated with epidermal browning of saguaro cacti at Saguaro National Monument. Eviron. Exp. Bot. 34: 9-17. Evans, L.S., V.A. Cantarella, L. Kaszczak, S.M. Krempasky, and K.H. Thompson. 1994b. Epidermal browning of saguaro cacti (Carnegiea gigantea). Physiological effects, rates of browning and relation to sun/shade conditions. Eviron. Exp. Bpt. 34: 107-115. Evans, L.S., J. Sullivan and M. Lim.2001. Initial effects of UV-B radiation on stem surfaces of Stenocerus thurberi(organ pipe cacti). Eviron. Exp. Bot. 46: 181-187. Evans, L.S., A. J. Young, and Sr. J. Harnett. 2005. Changes in scale and bark stem surface injuries and mortality rates of a saguaro (Carnegiea gigantea) cacti population in Tucson Mountain Park. Can. J. Bot. 83: 311-319. Gilat, Amos. 2004. “MATLAB: An Introduction with Applications,” 2nd Edition. John Wiley & Sons. MacMahon, J. A. 1985. Deserts. The Audubon Society Nature Guides. Alfred A., Steenbergh, W. and C. Lowe. 1977. Ecology of the Saguaro II. Scientific Monograph Series 8. National Park Service, Washington, DC. P. 242. Witten, I.H., E. Frank, L. Trigg, M. Hall, G. Holmes, and S.J. Cunningham. 1999. http://www.cs. waikato.ac.nz/ ml/publications/1999/99IHW-EF-LT-MH-GH-SJC-Tools-Java.pdf Weka: Practical Machine Learning Tools and Techniques with Java Implementations. Proceedings of the ICONIP/ANZIIS/ANNES’99 Workshop on Emerging Knowledge Engineering and Connectionist-Based Information Systems. pp. 192–196. Retrieved 2007-06-26.

Leaf venation patterns and interveinal areas in tree leaves Jorge Gonzalez∗ Department of Biology, Manhattan College

Abstract. The purpose of this study was to determine leaf venation patterns and interveinal areas in tree leaves of seventeen species. Leaf samples were obtained from the Manhattan College campus and Van Corlandt Park. Two leaves of each species were photographed for analysis. Leaf midribs were called primary veins, veins that lead from primary veins were called secondary veins, veins that lead from secondary veins were called tertiary veins, and veins that lead away from tertiary veins were called quaternary veins. The leaf areas of entire leaves were called primary leaf areas. Secondary leaf areas consisted of bisecting the area between two secondary veins on each side of each secondary vein. Entire leaf areas were well correlated with secondary areas for all seventeen species (y = 0.032x + 0.678, r2 = 0.40). Tertiary areas were well scaled with entire leaf areas (y = 0.0024x + 0.043, r2 = 0.50), with an area of 0.15 cm2 . Mean quaternary areas were not well scaled to entire leaf areas. The number of quaternary areas among all seventeen species were well scaled to entire leaf area (y = 41.0x + 1045, r2 = 0.52). On average there were 23 , 316 and 2930 secondary, tertiary and quaternary areas, respectively, per leaf. Relatively small coefficients of variation indicate that the leaf areas were very similar among the species tested. Overall, there were similar characteristics among all species of this study.

Introduction Leaves are the photosynthetic ‘engines’ of plants. Leaves fix carbon dioxide to make carbohydrates during the day and transport these carbohydrates at night (Hopkins et al., 2009). Thus, carbon fixing plants are the primary producers for the Earth and produce energy for all other organisms (Lambers et al., 1998). For plant leaves, stomata openings are required for carbon dioxide uptake. At the same time that carbon dioxide is being taken up by leaves, water vapor is being lost to the atmosphere. Thus, water use efficiency is the relationship between water loss (transpiration) to carbon gain (photosynthesis) by plant leaves (Hopkins et al., 2009). Water released during leaf transpiration is derived from water in stems that is transported via veins to all parts of the plant leaves (Hopkins et al., 2009). Plant leaves are very diverse, but with a casual view leaves may appear similar (Krupnick, 2001). Upon close inspection leaf venations and interveinal areas vary markedly among species (Fig. 1). Leaves show a wide variety of venation patterns and shapes of interveinal patterns (Fig. 2). Leaf veins transport water and nutrients to leaf cells from stems. In addition, veins transport sugars from leaves to stems. The normal pattern for water flow is from stems to leaf petioles and then to smaller and smaller veins so that all cells of the leaf obtain the water needed for normal function. ∗

Research mentored by Lance Evans, Ph.D.

The Manhattan Scientist, Series B, Volume 4 (2017)

Gonzalez

Figure 1. Leaf images of four tree species to show per current leaf venation among leaves. A = Ulmus Pumila B = Tilia platyphyllos C = Morus rubra D = Celtis occidentalis

Figure 2. Image of varies venation patterns for tertiary areas (e.g. percurrent, reticulate)

Leaf venation patterns are not similar among primary and secondary veins (Fig. 3) but venation patterns may be similar for tertiary and quaternary veins and their associated interveinal areas. Transport from the smallest veins to their associated interveinal areas may be similar since water in these smallest areas must be moved among cell walls of these cells and this movement of water via cell walls from cell to cell may be limiting to plant metabolism. In this study, interveinal areas associated with secondary, tertiary, and quaternary veins (Figs. 3,4,5) were investigated.

Gonzalez

The Manhattan Scientist, Series B, Volume 4 (2017)

Figure 3. Diagram of veins and areas of leaves that were used to bisect leaves into separate areas. Note that each secondary area is served by one secondary vein. Each tertiary area is served by one tertiary vein (Gonzalez, 2016)

Figure 4. Image of a leaf with a variety of veins and a secondary area (Gonzalez, 2016)

Figure 5. Image of a leaf with a variety of veins and two tertiary areas (Gonzalez, 2016)

The following hypotheses were investigated: 1. 2. 3. 4. 5.

Secondary areas are scaled to entire leaf area. Tertiary vein areas are scaled to entire leaf area. Quaternary vein areas are similar sizes. The number of tertiary areas remains relatively the same among all leaves. Quaternary areas increase as leaf area increases.

Materials and Methods Leaf sampling The leaves were chosen from plants on the Manhattan College campus and Van Cortlandt Park during the summer-fall of 2016 and summer of 2017. Several species were chosen for analysis

The Manhattan Scientist, Series B, Volume 4 (2017)

Gonzalez

shown in Tables 1 and 2. Each leaf species was identified using three sources (Kershner et al., 2008, ww.tropicos.org and dendro.cnre.vt.edu). Table 1. Percurrent leaf areas in cm2 of various species. Species

Leaf

Mean secondary

Mean tertiary

Mean quaternary

Salix nigra Morus rubra Hydrangea arboescens Tilia platyphyllos Betula alleghaniensis Malus pumila Viburnum lentago Tilia americana Betula papyrifera Magnolia x soulangena Ulmus americana Celtis occidentalis Carpinus caroliniana Cornus kousa Ostrya virginiana Latana camara Ulmus thomasii

84.5 71.3 62.6 61.2 60.7 59.6 57.5 49.5 41.9 40.9 40.2 38.1 36.6 30.6 18.7 16.1 11.8

1.97 3.56 3.87 3.04 1.68 3.97 1.44 2.47 1.90 2.05 1.44 2.16 2.37 2.55 1.10 0.81 0.38

0.26 0.19 0.25 0.18 0.16 0.29 0.11 0.12 0.19 0.14 0.13 0.07 0.17 0.07 0.10 0.15 0.04

0.018 0.019 0.034 0.018 0.015 0.021 0.015 0.015 0.019 0.011 0.011 0.009 0.013 0.014 0.016 0.020 0.008

Mean Standard Deviation Variation

46.0 20.0 43.6

2.20 1.02 47.3

0.15 0.06 44.6

0.016 0.005 36.3

For each plant species chosen, images of whole plants and individual leaves were saved for analysis. For each species, two of the same trees were chosen and two leaves were obtained from each tree. For each leaf it was placed on a light box with a ruler for measurement purposes and photographic images were taken of each leaf to identify secondary areas, tertiary areas, and quaternary areas (Figs. 4, 5, 6). Tissue Processing In order to view the leafâ&#x20AC;&#x2122;s veins, certain species required to be decolorized with 95% ethanol for three days. If additional decolorizing was necessary leaves were placed in 10% NaOH for two hours (Jensen, 1962). Next, the leaves were washed with water bleach was added to remove any pigments in the leaf. The leaf was poured into a dye called safranin (Jensen, 1962) that stained all veins, most importantly, tertiary and quaternary veins that were not visible with the green pigments. The staining processed was used on leaves whose quaternary areas were not visible under the light box. For other leaves, the smallest areas were visible and therefore did not need staining.

Gonzalez

The Manhattan Scientist, Series B, Volume 4 (2017)

Table 2. Characteristics of percurrent leaves of various species Species Salix nigra Viburnum lentago Betula allegnaniensis Ulmus thomasii Ulmus americana Ostrya virginiana Latana camara Magnolia x soulangena Morus rubra Tilia americana Tilia platyphyllos Celtis occidentalis Betula papyrifera Hydrangea arboescens Carpinus caroliniana Malus pumila Cornus kousa Mean Standard Deviation Variation

Secondary areas

Tertiary areas

Quaternary areas

43 40 36 31 28 22 20 20 20 20 20 18 17 16 15 15 12

325 523 375 320 298 215 108 288 357 412 331 544 176 250 215 206 438

499 3830 3920 1420 3650 2200 791 3720 3760 3300 3400 4230 1170 1840 2810 2910 2190

23.1 9.19 39.7

316 118 37.5

2930 1150 39.1

Average number of tertiary in a secondary Average number of quaternary in a tertiary

15 9.5

Measurements For this analysis, leaf midribs were called primary veins. Veins that lead from primary veins were called secondary veins. Veins that lead from secondary veins were called tertiary veins. Veins that lead away from tertiary veins were called quaternary veins. After leaves were photographed images were downloaded and areas were determined with ImageJ (National Institutes of Health: https://imagej.nih.gov/ij/). Secondary leaf areas consisted of bisecting the area between two secondary veins on each side of each secondary vein. Leaf areas associated with secondary (herein termed secondary areas), tertiary (herein were termed tertiary areas), and quaternary (herein termed quaternary areas) veins were traced (Fig. 3). All areas were traced (Fig. 6) so that individual secondary, tertiary and quaternary areas were measured. For the leaves in which tertiary and quaternary veins were not visible by using the light box the leaf was put under a dissecting microscope and viewed at 16 x magnification to get a closer and accurate view of the veins. (Fig. 5). In order, to see quaternary areas the magnification was put to its max at 34Ă&#x2014; in which quaternary areas were clearly seen (Fig. 6).

The Manhattan Scientist, Series B, Volume 4 (2017)

Gonzalez

Figure 6. Image of a leaf with tertiary veins marked with numerous quaternary areas (Gonzalez, 2016)

Results The purpose of this study was to determine the characteristics of the venation patterns of tree leaves. The leaves of this study was mainly focused on leaves with percurrent pattern. The plant species had a wide range of leaf areas (11.8 to 84.5 cm2 ) and secondary (0.38 to 1.97 cm2 ) areas (Table 1). Secondary areas were well scaled with entire leaf areas (y = 0.032x + 0.678, r2 = 0.40; Fig. 7), and there is a variation of 47.3 among all species. Mean percurrent leaf areas were 46.0 cm2 , respectively, and secondary areas were 2.20 cm2 , respectively.

Figure 7. Relationship between secondary areas and leaf areas for seventeen species with percurrent venation pattern. The equation of the line is y = 0.032x + 0.678, r2 = 0.40

Figure 8. Relationship between mean tertiary areas and leaf areas for seventeen species for percurrent species. The equation of the line is y = 0.0024x + 0.043, r2 = 0.50

In contrast, tertiary areas were not similar while quaternary areas were similar among all percurrent species. Mean percurrent tertiary areas were 0.15 cm2 , and quaternary areas were

Gonzalez

The Manhattan Scientist, Series B, Volume 4 (2017)

0.016 cm2 . Since tertiary areas were not similar among the leaf groups, mean tertiary areas (y = 0.0024x+0.043, r2 = 0.50; Fig. 8) are well scaled to leaf area. On the other hand, quaternary areas were similar among all percurrent species so the mean quaternary areas were not well scaled to leaf area. Furthermore, quaternary areas were similar among the leaf groups, so numbers of quaternary areas (y = 41.0x + 1045, r2 = 0.52; Fig. 9) were well scaled with leaf areas. The number of tertiary areas were not well scaled with leaf areas among species. There was an average number of 15 tertiary areas per secondary area, while there was an average number of 9.5 quaternary areas per tertiary area. Additional data showed that as the tertiary areas increased among all species the number of quaternary areas increased (y = 30.7x + 4.75, r2 = 0.48; Fig. 10). Overall, there were between 100 to 500 hundred tertiary areas per leaf and between 500 to 4200 quaternary areas per leaf (Table 2).

Figure 9. Relationship between number of quaternary areas and leaf area for seventeen percurrent species. The equation of the line is y = 41.0x + 1045, r2 = 0.52

Figure 10. Relationship between the number of quaternary areas per tertiary and mean tertiary area. The equation of line y = 30.7x+4.75, r2 = 0.48.

Discussion The overall purpose of this study was to understand water movement from leaf veins to areas of leaves between leaf veins. After leaf initiation at shoot apices, leaves enlarge via leaf marginal meristems, or cells that divide along the margins of leaves. In addition, as leaves enlarge cell enlargement occurs producing smaller veins to feed all areas among the leaf. Leaf veins will have xylem cells that have obtained water from stems. Primary veins (midribs) distribute water to secondary veins, to tertiary veins and then to quaternary veins. Eventually, water is released from xylem cells to leaf cells. Movement of water from cell to cell within areas between veins is by adsorption to the cellulose cell walls and then absorption of water into cells. Of course, water is lost from leaves via stomatal openings. The purpose of this study was to understand relationships among leaf veins and leaf areas. By definition, secondary veins are connected directly to the main vein (midrib) so they are fixed in space. Secondary areas should increase as leaves enlarge via margin meristems and as cells within

The Manhattan Scientist, Series B, Volume 4 (2017)

Gonzalez

leaves enlarge; our data support that outcome. Secondary leaf areas were larger for larger leaves than for smaller leaves. In contrast, the situation with tertiary areas and quaternary areas is more complicated. As stated above, as leaves enlarge near margins, new tertiary and quaternary areas arise. It is probable that tertiary veins arise initially from secondary veins, but as leaf cells within tertiary areas enlarge, the smaller quaternary veins arise to provide water to these enlarging cells and enlarging tertiary areas. The data of this study are consistent with the idea that tertiary areas will enlarge as the leaf enlarges and quaternary areas will be constant as leaves enlarge. Based on the data, he number of tertiary areas will remain constant and new quaternary veins will arise to provide water to the tertiary and quaternary areas, respectively. The current study only involved analyses of a limited number of species. Data from additional species will be obtained to examine these relationships further. In addition, we have processed anatomical tissues sections of veins with xylem cells to determine numbers of xylem cells in veins of the species tested. Xylem conductivity values (McCulloh et al., 2003) will also be determined in the future. Future research will include obtaining leaves of various leaf areas in the spring from species sampled in this study to determine the pattern of leaf growth relative to the lengths of secondary veins. Another objective will be to determine the origins (locations and enlargements) of tertiary and quaternary veins and the relationships between veins and enlargement of tertiary and quaternary areas associated with these veins.

Acknowledgment The author is indebted to the Catherine and Robert Fenton Endowed Chair to Dr. Lance Evans for financial support of this research.

References Gonzalez, J. 2016. Leaf venation patterns and the distribution of water within leaves. Manhattan Scientist Ser. B, Vol. 3, 57-62 Hopkins, G. William, Huner P. Norman. 2009. Plant Physiology. John Wiley & Sons, Inc. Jensen, W.A. 1962. Botanical Histochemistry. Principles and Practice. W.H. Freeman. San Francisco, CA. Kershner, B., Matthews, D. Nelson, G. 2008. Field Guide to Trees of North America. Sterling Publ. Co., New York. Krupnick, G. “Centres of Plant Diversity: Introduction.” 2001. Centres of Plant Diversity: Introduction. Department of Botany, National Museum of Natural History, Web. 13 Sept. 2016. http://botany.si.edu/projects/cpd/introduction.htm. Lambers, H., F. S. Chapin, and T. L. Pons. 1998. Plant Physiological Ecology. New York: Springer McCulloh, K.A, J. S. Sperry, and F.R. Alder. 2003. Water transport in plants obeys Murray’s law. Nature 421: 939-942.

Predicting rates of mortality for saguaro cactus plants (Carnegiea gigantea) Cole Johnson∗ Department of Biology, Manhattan College Abstract. Saguaro cacti (Carnegiea gigantea) are one of more than twenty other columnar cactus species in the Americas subject to bark formation, or epidermal browning. Bark formation is known to be the result of UV-B light coming from the sun. Extensive bark formation can lead to premature death of saguaro cactus plants that are known to live for hundreds of years. This study uses WEKA machine learning programming to predict bark accumulation and death progression for a population of saguaro cacti located in Tucson Mountain Park, Tucson, AZ from 1994-2017 over 8-year intervals. It was hypothesized that south-facing surfaces would be the major predictors of bark accumulation among healthy cacti while north-facing surfaces would be the major predictors of death among cacti of varying health degrees. Each individual cactus was assigned a health category based on the percentage of bark contained on its surfaces from 1994-2002, 2002-2010, and 2010-2017. The data was then run though WEKA programming to determine which cactus surfaces (north, south, east, and west) were the major predictors of bark accumulation and cactus mortality. It was found that bark percentages on south-facing surfaces are best at predicting bark accumulation. It was also found that both north-facing and south-facing surfaces can predict cactus mortality at high accuracies. It was concluded that WEKA machine learning techniques can predict cactus morbidity and mortality at high accuracies for several health classes of cacti.

Introduction The saguaro cactus (Carnegiea gigantea) is a columnar plant species native to Tucson, Arizona. This species, along with more than 24 other columnar cacti species, suffers from an accumulation of epidermal barking on its stem surfaces resulting in premature death (Evans and DeBonis, 2015). Bark formation is caused by a build-up of epicuticular waxes on the stem surfaces of the cactus plant (Evans et al., 1994a; 1994b). As the amount of wax on the stem surfaces accumulates, the stomata on the surface of the plant are blocked therefore preventing gas exchange from occurring (Evans et al., 1994a; 1994b). The inability to exchange gases deprives the cactus of necessary elements needed to carry out photosynthesis and respiration, resulting in premature death (Evans et al. 1994a, 1994b, Evans and Macri 2008). A previous study shows that bark formation causes cacti to reach mortality at a rate of 2.3% per year (Evans et al., 2005; 2013). This rate is extremely high given that saguaro cacti have a life expectancy of several hundred years (Steenbergh and Lowe, 1997). Studies have shown the cause of bark formation to be UV-B light exposure coming from the sun (Evans et al., 2001). Tucson Arizona has a latitude of 32.2◦ N which is above the Tropic of Cancer. Therefore, due to the location of the sun, south-facing surfaces of saguaro cacti in Tucson, AZ receive four times more sunlight than north-facing surfaces (Evans and Macri, 2008). It is for this reason that south-facing surfaces have been found to be the first to accumulate bark ∗

Research mentored by Lance Evans, Ph.D.

The Manhattan Scientist, Series B, Volume 4 (2017)

Johnson

whereas north-facing surfaces are last to acquire bark before cactus death occurs (Evans et al., 1992; 1994a). This study examined the progression of bark accumulation to predict injury and death for cacti of various health statuses from 1994-2017 over eight-year time intervals. The following were hypothesized: 1. South-facing surfaces are the best predictors of bark accumulation for healthy cacti 2. North-facing surfaces are the best predictors of cactus death for cacti of various health statuses 3. A WEKA decision tree can accurately predict cactus death using bark percentages on 12 surfaces as criteria

Materials and Methods Field and survey conditions A sample of 1,149 saguaro cacti (Carnegiea gigantea) were studied over a 24-year time period. All cacti were located in 50 field plots within Tucson Mountain Park near Tucson, AZ (32.2238â&#x2014;Ś N, 111.1443â&#x2014;Ś W) (Evans et al., 2005). Cacti were first selected in 1993. All selected cacti were taller than 4 m and assumed to be more than 80 years old (Steenburgh and Lowe, 1977; Pierson and Turner, 1998). Physical characteristics, nearby vegetation and topographical features, GPS coordinates, as well as human-designated markers were used to distinguish each cactus plant within a plot (Evans et al., 2005). The same cacti have been evaluated throughout the duration of the study. Cacti were only evaluated if they could be positively identified (Evans et al., 2005). Field evaluations took place in 1993, 2002, 2010, and 2017. Cactus surface evaluations During each field evaluation, the percent green area on the surface of each cactus was visually determined and recorded (Evans et al., 2013). The amount of green area on the cactus was taken as opposed to the amount of bark in order to eliminate bias created by different intensities of barking (Evans et al., 2013). The evaluation took place at an elevation of 1.75 meters from the ground (Evans et al., 2013). A height of 1.75m was used due to the fact that it is comparable breast height of the evaluator. This made the visual evaluations more convenient to execute. Also, evaluations occurred at a range of 8 cm in length along each of the surfaces examined (Evans et al., 1994a; 1994b; 1995). The width of the evaluation was dependent on the width of the specific cactus rib. A total of 12 surfaces were evaluated on each cactus (Evans et al., 1995; 2003; 2005). These surfaces consisted of the crests, right troughs, and left troughs on the cactus ribs residing closest to each of the four compass directions on the circle of azimuth (Evans et al., 1994a; 1994b; 1995). Crests are the segment of a cactus rib that protrudes outward towards the environment where troughs are concave indentations located to either side of the crest (Geller and Nobel, 1984; Gibson and Nobel, 1986). The position (right or left) of each trough on either side of the adjacent crest was assigned based on how the evaluator viewed them. The percentage of green area on each surface was converted to a percentage of bark and placed into a Microsoft Excel workbook. The analysis was done for three time periods (1994-2002, 2002-2010, 2010-2017).

Johnson

The Manhattan Scientist, Series B, Volume 4 (2017)

Health categories of cacti Each cactus was assigned a health status corresponding to the formation of bark on the southfacing crest only (Table 1) (Evans et al., 2013), and was assigned two health categories to represent bark accumulation over each 8-year period. The initial status was assigned at the beginning of the evaluation period while the second status was assigned at the end of each evaluation period. Table 1. Criteria used to assign classes of barking injury to individual saguaro cacti (Carnegiea gigantea). Each criterion pertains to barking percentage on the south crest only. Healthy

Criteria

Healthy Bark Slight bark Medium bark Severe bark Dead

Less than 20% bark during entire evaluation period Less than 20% bark during initial evaluation, but more than 20% bark during final evaluation Between 20% and 49% bark during initial evaluation Between 50% and 80% bark during initial evaluation Greater than 80% bark during entire evaluation period Dead

Theory behind WEKA 3.8 To unveil which surfaces were prominent in predicting morbidity and mortality, a Machine Learning Program known as WEKA 3.8 was used. WEKA uses criteria known as classifiers to predict certain outcomes from one measurement period to the next (Hall et al. 2009). For our purposes, WEKA generated these predictions in the form of J48 decision trees. This format of decision tree is produced through the implementation of the ID3 algorithm and allows one to predict target values within a set of data (i.e., White et al., 1941; Hitchcock, 1971). The variables examined were each of the 12 surfaces. The decision trees produced by WEKA expose which surfaces play crucial roles in predicting cactus morbidity and mortality within a given health category of cacti. The decision trees produced can be easily interpreted by those who are unfamiliar with Machine Learning programs and the methods they utilize (Hyafil and Rivest, 1976). WEKA created the J48 decision trees by generating a high accuracy classifier that is produced using data from every cactus in the data set. The data is then divided into several subsets. The program then creates its own unique classifiers relevant to each of the individual subsets of data. This allows WEKA to determine which classifiers were most crucial in generating the high accuracy classifier. WEKA then performed a cross validation on the data set to ensure the classifier generated was relevant to the entire data set and not just the subsets formed (Hyafil and Rivest, 1976). The cross validation implemented was by tenfold. Cross validation by tenfold splits the entire data base used to generate the classifier into ten equal subsets. The ten subsets are then split into two groups. Nine of the subsets are used to train the program and produce a classifier while the remaining subset is used to test the accuracy of the classifier produced. WEKA will continue to train and test each of the classifiers with the remaining subsets of data until every subset has been analyzed (Hyafil and Rivest, 1976). A confusion matrix

The Manhattan Scientist, Series B, Volume 4 (2017)

Johnson

is then produced to demonstrate how many of the samples analyzed were predicted correctly by WEKA. The accuracies of the trees were determined using the confusion matrices. Use of WEKA Cacti belonging to the first health categories at the beginning of each time interval were separated into files. “Bark” is used to represent increased injury of a healthy cactus and can only be assigned to a cactus at the end of the evaluation period. Therefore, morbidity and mortality cannot be predicted within cacti belonging to the “bark” category. The fate of each cactus at the end of each 8-year evaluation was noted in the file in a format recognizable by WEKA. Slightly barked to severely barked cacti could either remain in the same health category or reach mortality. Healthy cacti at the beginning of the evaluation could either remain healthy, become barked, or die by the end of the evaluation. WEKA used the data from each of the separate files to generate decision trees that predict morbidity and mortality in each of the health categories. The trees produced use the 12 surfaces to predict cactus injury and death at the end of each 8-year period. The different trees from each of the classes expose which surfaces play a prominent role in making such predictions.

Results Predicting bark accumulation of healthy cacti The first aspect of this research sought to examine which surfaces could predict bark accumulation among healthy cacti over each of the evaluation periods (Table 2). WEKA determined the south-facing surfaces to be the most crucial for making this prediction during each interval. Table 2. Predicting bark of healthy saguaro cacti (Carnegiea gigantea) that became unhealthy using WEKA with south facing surface data.

†

Time interval

n†

Accuracy (%)

1994 - 2002 2002 - 2010 2010 - 2017

386 85 66

63.5 75.8 63.4

n refers to the number of cactus plants analyzed.

From 1994-2002, south-facing surfaces predicted bark accumulation of healthy cacti at an accuracy of 63.5%. Bark accumulation among healthy cacti was predicted with 75.8% accuracy from 2002-2010 and with 63.4% accuracy from 2010-2017 by south surfaces. These results suggest that south-facing surfaces are accurate and consistent in their abilities to predict bark accumulation over several time intervals. Predicting death among cacti of various health statuses The purpose of this study was to unveil which surfaces play prominent roles in predicting death for cacti of various health statuses over each time interval. The WEKA analyses revealed

Johnson

The Manhattan Scientist, Series B, Volume 4 (2017)

that north-facing and south-facing surfaces can predict cactus death at accuracies ranging from 72.6% to 92.5% (Table 3). From 1994-2002, north-facing surfaces were found to be the major predictors of cactus death among both healthy cacti and slightly barked cacti. Death of healthy cacti Table 3. Predicting death of saguaro cacti (Carnegiea gigantea) with various predictive surfaces during each time interval using WEKA. Time Interval

Health Status

Major Predictor

nâ&#x20AC;

Accuracy (%)

1994-2002 1994-2002 2002-2010 2002-2010 2010-2017 2010-2017

Healthy Slight bark Healthy Severe bark Medium bark Severe bark

North North South North North South

231 112 191 448 107 427

89.2 84.7 83.5 76.1 92.5 72.6

â&#x20AC;

n refers to the number of cactus plants analyzed.

during this time interval could be predicted by WEKA with 89.2% accuracy while death of slightly barked cacti were predicted with 84.7% accuracy. From 2002-2010, WEKA determined the major predictor of death for healthy cactus plants to be south-facing surfaces. South-facing surfaces during this time interval could predict death of healthy cacti with 83.5% accuracy. Also from 20022010, WEKA determined that north-facing surfaces were the major predictors of death for severely barked cacti. These surfaces could predict death with an accuracy of 76.1% accuracy. From 20102017, death of medium barked cacti could be best predicted by south-facing surfaces with an accuracy of 92.5%. The death of severely barked cacti could be predicted with 72.6% accuracy. The results demonstrate that north-facing and south-facing surfaces are the major predictors of cactus death and also that WEKA can predict cactus death at high accuracies. Using WEKA decision trees to predict cactus death The goal of this study was to produce a decision tree model capable of predicting cactus death at high accuracies using WEKA programming. Fig. 1 is an example of a decision tree produced by WEKA. This tree uses the 12 cactus surfaces to predict death among slightly barked cacti from 1994-2002. The decision tree model can predict the death of slightly barked cacti with an accuracy of 84.7% from 1994-2002. Fig. 2 was produced by WEKA to predict cactus death among healthy cacti from 1994-2002. It does so with an 89.2% accuracy.

Discussion The purpose of this study was to closely examine bark accumulation and death progression of a population of saguaro cacti. Bark formation on saguaro cacti was a rare phenomenon prior to the 1950s. Before this time, saguaros experienced limited barking and reached very large heights on the growth spectrum. Recently, saguaro cacti and other columnar cacti species have been experiencing bark formation at rapid rates. Bark formation is a precursor to death and it can provide

Figure 6. WEKA decision tree model predicting death for healthy cacti from 1994-2002 using bark The Manhattan Scientist, Series Volume 4 (2017) Johnson percentages B, on each of the 12 designated cactus surfaces as data.

Figure 5. WEKA decision tree model predicting death for slightly barked cacti from 1994-2002 using bark percentages on each of the 12 designated cactus surfaces as data.

≤ 3

> 3

Alive

> 35

≤ 35

Alive

> 95

≤ 95

NC < 5

≥ 5

SC ≤ 7

≥ 3

Alive

≤ 3 Alive

> 3 Dead

< 3

Alive

Dead

≤ 10

> 10

≤ 2

> 2

Alive

Dead

Alive

Dead

Accuracy: 84.7 %

Figure 1. WEKA decision tree model predicting death for slightly barked cacti from 1994-2002 using bark percentages on each of the 12 designated cactus surfaces as data. [Accuracy: 84.7%]

Figure 2. WEKA decision tree model predicting death for healthy cacti from 1994-2002 using bark percentages on each of the 12 designated cactus surfaces as data. [Accuracy: 89.2%]

Accuracy: 89.2 %

great insight as to when a cactus will reach mortality. As previously stated, bark accumulation on cactus plants typically follows a certain pattern. The south-facing surfaces are the first to exhibit epidermal barking whereas north-facing surfaces are normally last (DeBonis et al., 2015). For this reason, it was hypothesized that south-facing surfaces would be the major predictors of bark accumulation while north-facing surfaces would be the major predictors of cactus death. The obtained results confirm these hypotheses. Table 2 demonstrates that south-facing surfaces are the major predictors of bark accumulation among healthy cacti. This result reinforces that claim that UV-B light exposure coming from the sun is a major cause of bark formation. Also, the accuracy of which south-facing surfaces can predict bark accumulation remains more or less constant over each time interval. Therefore, further investigation of this study could include determining whether or not the population of cacti is in steady state equilibrium. If the population is in steady state equilibrium, the percentage of cacti reaching death at the end of the evaluation for each health category would be consistent over time. To do this, one would have to calculate the average rate of bark accumulation for each health class over each time interval. A statistical analysis including a T-test on data from each time interval will reveal whether or not the population is in steady-state equilibrium. Previous research has shown that once initial barking on south-facing surfaces has begun, east surfaces are the next to accumulate bark. East surfaces accumulate bark three years following the

Johnson

The Manhattan Scientist, Series B, Volume 4 (2017)

south, while west surfaces begin showing bark formation eight years following the south surfaces. Once the south surfaces have begun showing bark formation, north surfaces will exhibit barking fifteen years after (DeBonis et al., 2015). Table 3 shows that both north-facing surfaces and southfacing surfaces are the major predictors of cactus death. Barking on east and west surfaces plays a more prominent role during the intermediate steps in determining cactus death as demonstrated by the WEKA decision tree models (Figs. 1 and 2). Both the results of this study as well as previous ones could potentially provide substantial evidence in unveiling factors that influence the rate of bark accumulation and mortality for individual cactus plants. The saguaro cactus was known to live for several hundreds of years (Steenbergh and Lowe, 1977); however, recent studies show the expected lifespan of these organisms has decreased substantially since bark formation became more prominent in the 1950s (Oâ&#x20AC;&#x2122;Brien et al., 2011). The saguaro cactus and other plant species subject to premature death caused by epidermal barking are delicacies to several groups of people across North America. If factors directly related to influencing the rate of bark accumulation can be determined, then preventative action may be taken to increase the longevity of these once prosperous plants.

Acknowledgments This work was supported by the Catherine and Robert Fenton Endowed Chair to Dr. Lance Evans. The author is grateful to Dr. Lance Evans for his mentorship and extensive help with this study.

References DeBonis et al. 2015. Unpublished data - personal communication. Evans, L. S. and A. Macri. 2008. Stem surface injuries of several species of columnar cacti of Ecuador. J. Torrey Bot. Soc. 135: 475-482. Evans, L. S., and M. DeBonis. 2015. Predicting morbidity and mortality of Saguaro cacti (Carnegiea gigantea) J. Torrey Bot. Soc. 142: 231-239. Evans, L. S., A. J. Young, and Sr. J. Harnett. 2005. Changes in scale and bark stem surface injuries and mortality rates of saguaro (Carnegiea gigantea) cacti population in Tucson Mountain Park. Can. J. Bot. 83: 311-319. Evans, L. S., K. A. Howard and E. Stolze. 1992. Epidermal Browning of Saguaro Cacti (Carnegiea gigantea): Is it new or related to direction? Environ. Exp. Bot. 32: 357-363. Evans, L. S., P. Boothe and A. Baez. 2013. Predicting morbidity and mortality for a saguaro cactus (Carnegiea gigantea) population. J. Torrey Bot. Soc. 140: 247-255. Evans, L. S., J. Sullivan and M. Lim. 2001. Initial effects of UV-B radiation on stem surfaces of Stenocereus thurberi (organ pipe cacti). Environ. Exp. Bot. 46: 181-187. Evans, L. S., M. Zugermayr, and A. J. Young. 2003. Changes in surface and mortality rates of saguaro (Carnegiea gigantea) cacti over a twelve-year paper. J. Torrey Bot. Soc. 130: 238-243.

The Manhattan Scientist, Series B, Volume 4 (2017)

Johnson

Evans, L. S.; V. A. Cantarella, K. W. Stolte and K. H. Thompson. 1994a. Phenological changes associated with epidermal browning of saguaro cacti at Saguaro National Monument. Environ. Exp. Bot. 34: 9-17. Evans, L. S., V. A. Cantarella, L. Kaszczak, S. M. Krempasky, and K. H. Thompson. 1994b. Epidermal browning of saguaro cacti (Carnegiea gigantea). Physiological effects, rates of browning and relation to sun/shade conditions. Environ. Exp. Bot. 34: 107-115. Evans, L. S., V. Sahi, and S. Ghersini. 1995. Epidermal browning of saguaro cacti (Carnegiea gigantea): Relative health and rates of surficial injuries of a population. Environ. Exp. Bot. 35: 557-562. Geller, G. and P. Nobel. 1984. Cactus ribs: Influence of PAR inception and CO2 uptake. Photosynthetica 18: 482-494. Gibson, A. and P. Nobel. 1986. The Cactus Primer. Harvard University Press, Cambridge, MA. 286 p. Hitchcock, A. 1971. Manual of the Grasses of the United States. Dover Publishers, New York, NY. 1051 p. Hyafil, L. and R. Rivest. 1976. Constructing optimal binary decision trees in NP-complete. Inform. Process. Lett. 5: 15-17. Oâ&#x20AC;&#x2122;Brien, K., D. Swann, and A. Springer. 2011. Results of the 2010 saguaro consensus at Saguaro National Park. National Park Service, U.S. Department of Interior, Tucson, AZ. 49 p. Pierson, E. and R. Turner. 1998. An 85-year study of saguaro (Carnegiea gigantea) demography. Ecology 79: 2676-2693. Steenbergh, W. F. and C. H. Lowe. 1977. Ecology of the Saguaro II reproduction, germination, establishment, growth, and survival of the young plant. National Park Service Monograph Series Eight. White, A., R. A. Dyer, and B. L. Sloane. 1941. The Succulent Euphorbieae (Southern Africa). Abbey Garden Press, Pasadena, CA. 937 p.

Bark formation and death progression in saguaro cacti (Carnegiea gigantea) Marissa LoCastro∗ Department of Biology, Manhattan College Abstract. Cactus injury from bark accumulation has been shown to occur on stem surfaces of tall, long-lived cactus species throughout the Americas. The amount of sun exposure directly correlates to the rate of bark accumulation on cactus plants. The build up of bark causes morbidity of these succulent plants, and inevitable premature mortality. The effects of epidermal browning have been analyzed on saguaro cacti (Carnegeia gigantea (Engelm.) Britt and Rose) native to Tucson Mountain Park in Tucson, Arizona. The latitude of the Tucson Mountain Park desert causes four times more sunlight on south-facing surface than north-facing surfaces. This study utilized data from 50 desert field plots, collecting data for 1100 cactus plants over four sampling periods. The cacti data for bark covered was collected in 1994, 2002, 2010, and 2017. The bark percentage data was run through a standard machine-learning program, WEKA 3.8, to determine characteristics of cacti in a given sampling interval, that can be used to predict bark formation in the sequential time interval. Decision trees were created by the J48 tree classifier and confusion matrices were produces to determine the accuracy of the tree predictions. A cactus’ bark accumulation can be predicted by previous bark formation rates at a 91% accuracy. Bark percentage data was statistically analyzed by t-test to produce histograms showing the effects of south surface shading by nearby vegetation on the rate of bark formation. The presence of shading on a cactus plant slows bark formation by 15% confirming the powerful influence of sun exposure on bark accumulation and cactus injury.

Introduction Epidermal browning has been shown to take place on saguaro cacti (Carnegiea gigantea) native to Tucson, Arizona (Engelm.) (Britt and Rose). Sun exposure and injury cause bark formation on the stem of the cactus plant (Durisco and Graban, 1992; Evans et al., 1992, 1994a, 1994c, 1995; Turner and Funicelli, 2000). Bark formation begins on the south-facing surfaces of the cactus and forms around the stem to the north facing surfaces (Evans et al., 2001) since south-facing surfaces receive more sunlight than other surfaces over the year (Evans et al., 2001). Epidermal browning enhances morbidity and as the browning increases, a given cactus will be more susceptible to premature mortality (Evans et al., 2003, 2005). Increased overall mortality of the cacti population is accounted for by sunlight. Based on previous analysis of UV-B exposure on natural stands, it is conclusive that UV-B is also the cause of the epidermal browning symptoms observed in cactus plants (Evans et al., 2001). In order for a cactus to fulfill its expected life time, it must have functioning stomata for gas exchange. The stomata are located on cactus surfaces that consist of a thick cuticle, an epidermis, and several hypodermal cell layers. Bark accumulation is essentially the buildup of epicuticle waxes followed by proliferation of epidermal cells to form a scale and bark (Evans et al., 1994a, 1994b; Evans and Macri, 2008). The bark inhibits the functioning of the stomata on the surfaces, and ∗

Research mentored by Lance Evans, Ph.D.

The Manhattan Scientist, Series B, Volume 4 (2017)

LoCastro

inevitably the functioning of gas exchange. Gas exchange includes photosynthesis and respiration, vital to cactus life (Evans and Macri, 2008). This is a study of interest because saguaro cacti are dying before their expected 200-year lifetime (Turner 1992; Evans et al., 1995, 2003, 2005). The experiments of this study span from 1993 to 2017 to determine factors and influences that are contributing to evident premature mortality of saguaros. The specific hypotheses of this study were (1) previous bark formation data from an eight-year sampling interval can be used to predict bark accumulation in a sequential eightyear sampling interval and (2) bark formation rates will be slowed in the presence of shading on south-facing surfaces by nearby vegetation.

Materials and Methods Field conditions Saguaro cacti (Carnegiea gigantea) (Engelm.) native to Tucson Mountain Park (32.2238â&#x2014;Ś N, 111.1443â&#x2014;Ś W) were analyzed in this study. In 1994, 50 permanent plots with 1149 cacti were randomly selected (Evans et al., 2005). The cacti were evaluated in 1994, 2002, 2010, and 2017. The selected cacti were all taller than 4 meters. Physical features of cacti, nearby vegetation, topographical features, GPS points, and man-made rock piles were used to identify and distinguish cacti for each study evaluation (Evans et al., 2005). Data from cacti in 1994, 2002, 2010, and 2017 were used for this study. Data sets used Saguaro cacti are ribbed around the entire stem of their body. The protrusion of each rib is termed the crest and the indentations are termed as troughs (Geller and Nobel, 1984; Gibson and Nobel, 1986). For the four field evaluations, the single rib closest to each of the cardinal directions was evaluated; South, East, North and West. Each cactus was sampled at the crest, right trough, and left trough for each cardinal direction (Evans et al. 1994a, 1994b, 1995). The sample was 8 cm long at an elevation of 1.75 meters from the ground (Evans et al., 1994a, 1994b, 1995). Data was recorded in percent green of the surfaces of interest. Percent green was converted to percent bark for analysis. Data of percent bark areas include crests, right troughs and left troughs for south, east, north, and west-facing surfaces, making a total of twelve surfaces of data for each cactus (Evans et al., 1995, 2003, 2005). Data were transferred into a Microsoft Excel file. Nine photographs were taken for each cactus plant to accompany the bark percentage data. Photographs of the crest, right trough, and left trough for each cardinal direction (S, E, W, and N) were archived. At least one photograph of the entire cactus with the surrounding vegetation and terrain were archived. Health Evaluation of cacti Data entered into data sheets consisted of (1) plot number, (2) cactus number, (3) year, and (4) bark percentage for crest, right trough, and left trough of each cardinal direction (12 data points).

LoCastro

The Manhattan Scientist, Series B, Volume 4 (2017)

Field data of percent bark from 1994, 2002, 2010, and 2017 were entered into a Microsoft Excel file. Each cactus was assigned a relative degree of bark coverage based on its amount of bark on the south-facing crest (Evans et al., 2005). A category of health based on bark accumulation was assigned at the end of each eight-year sampling interval. For each period of study, a cactus was assigned one level of health to show bark accumulation and health degradation over the 8-year period. This was done for all three evaluation intervals. Cacti with bark less than 20% on the south crest were termed as healthy. If a healthy cactus plant exceeded 20% bark on the south facing crest in a given time interval, it was termed unhealthy. If the cactus exceeded 50% bark on the south facing crest in a given time interval, it was termed unhealthier, and if it exceeded 80% bark on the south facing surface it was termed extremely unhealthy. If a healthy cactus plant died within the interval, its description was termed as dead. Theory of WEKA 3.8 WEKA 3.8 is a Machine Learning program (Hall et al., 2009) employed to analyze the data set for cacti. WEKA generates decision trees to make prediction regarding bark formation. Classifiers, or criteria, generated from a given data set, predict bark formation for the next time period. This study specifically utilized the classifier J48 tree. The J48 tree is the type of decision tree produced. It is a specific format that implements the algorithm ID3 to predict the target variable in a set of data (i.e., White et al., 1941; Hitchcock 1971) The J48 tree is a high accuracy classifier (Hyafil and Rivest, 1976) that uses data from every cactus in the study. WEKA uses subsets of data, acting as classifiers, to set criteria that divides data into two groups. Subsets investigates the high accuracy classifier to determine which data was most prominent in making accurate predictions. Measurements from subsets of cacti yield the best classifier from the least amount of data, therefore the classifier must be cross validated to ensure relevance and accuracy to the entire data base, not just the subset. The J48 tree classifier was accompanied with the cross-validation test. The cross-validation test was used because it ensures high accuracy decision trees, understandable by those unfamiliar with Machine Learning Programs (Hyafil and Rivest, 1976). Cross-validation was set at tenfold, splitting the entire data base into 10 equal sets. Majority of each set is used to train the decision tree, and the minority of data is used to test of decision tree. Classifiers are produced with the large subset of data and accuracy is tested with the small subset of data. This is repeated for all ten groups of data. The 10 classifiersâ&#x20AC;&#x2122; performances are then averaged together to generate the most accurate decision tree. Health Degradation Analysis Each cactus was assigned a health status based on bark accumulation on the south crest over a sampling interval of eight years. To understand how the cacti progressively degrade in health, data from an eight year interval was compared to the sequential eight year sampling period. A Microsoft Excel sheet was generated with each cactus and its bark percentage for 2002 from the

The Manhattan Scientist, Series B, Volume 4 (2017)

LoCastro

1994 to 2002 interval, its health status for 2010 from the 2002 to 2010 interval, and its health status for 2017 from the 2010 to 2017 interval. The file was then separated into individual spread sheets consisting of bark accumulation percentages. This included movement such as healthy to unhealthy, unhealthy to unhealthier, unhealthy to dead, etc. These files were run through the WEKA 3.8 program to generate decision trees that predict bark accumulation on the south crest of a cactus plant using data from a previous sampling period. Since more morbid health categories resulted from more bark formation, a cactus plant could only either degrade in health or remain in its initial health status. Surrounding vegetation analysis A landscape photo of each cactus was taken in the 2017 evaluation. These photos were sorted into two groups: cacti with surrounding vegetation and cacti without surrounding vegetation. The cacti with surrounding vegetation were further grouped based on whether the vegetation around the cactus provides shade. The photos with shading were grouped based on which cardinal direction(s) was/were shaded. The photos were statistically analyzed, by t-test, with bark percentages covering the entire cactus plant to determine correlation between vegetation and bark formation on cacti from 2010 to 2017.

Results Predicting morbidity in saguaro cacti The first hypothesis addressed the ability of bark formation to be predicted by previous bark formation data. The ability to predict bark formation on a saguaro cactus is shown in Table 1. The data was run through WEKA 3.8 for all levels of bark accumulation for all sampling intervals, but this table shows a sampling of the results. Table 1 utilizes changes in health for cacti over the interval from 1994 to 2002, to predict how the same cacti will behave over the interval from 2002 to 2010. It does the same thing using data from 2002 to 2010, to predict bark accumulation from 2010 to 2017. Data on bark accumulation from 1994 to 2002 can predict bark accumulation from 2002 to 2010 at a 90.5% accuracy. Data on bark accumulation from 2002 to 2010 can predict bark accumulation from 2010 to 2015 at a 90.9% accuracy. These results support the hypothesis. WEKA 3.8 produces decision trees in order to produce accuracies for the predictability that data from one interval can predict data in the next interval. These decision trees determine defining characteristics of a data set that determine the likelihood a given cactus will change to two possible conditions. A decision tree was produced for each possible health change for each interval comparison. Figs. 1 and 2 are decision trees that match the data presented in Table 1.

LoCastro

The Manhattan Scientist, Series B, Volume 4 (2017)

≤ 40

> 40

Dead

≤ 60

> 60

Unhealthier

Dead

Figure 1. Decision tree predicting the likelihood that a cactus that was healthy over the 1994 to 2002 interval, will become unhealthy or die in the 2002 to 2010 interval.

SC > 80

≤ 80 Dead

WC ≤ 95

Unhealthier

> 95 Dead

Figure 2. Decision tree predicting the likelihood that a cactus that was unhealthy over the 1994 to 2002 interval, will become even more unhealthy or die in the 2002 to 2010 interval.

From these decision trees, cacti are separated out based on characteristics that determine bark formation in a sequential interval. It was determined that the South crest is the major predictor of bark accumulation. The south crest is the predominant surface that begins the decision trees to determine the likelihood of bark accumulation, and to what degree. This was concluded because the south crest is the first surface analyzed onthe majority of the trees generated by this data set. The numbers next to the lines leading to the bark accumulation of the cacti are the percent of barking on the respective surface labeled in the circles.

The Manhattan Scientist, Series B, Volume 4 (2017)

LoCastro

Table 1. Predicting bark formation of saguaro cacti (Carnegiea gigantea) with various predictive surfaces over several interval periods using WEKA. Interval Period 1994-2002 to predict 2002-2010 Bark Formation Healthy to Unhealthier or Dead Unhealthy to Unhealthier or Dead

2002-2010 to predict 2010-2017

Predictive Surface

Cutoff2

Accuracy3

Predictive Surface

Cuttoff

Acurracy

South Crest (52)1

88.5

South Crest (40)

92.5

South Crest (37)

83.8

South Crest (38)

84.2

Numbers in parenthesis labeled with a 1 means that is the number of cactus plants analyzed by WEKA to produce the decision tree for the bark coverage change listed in the bark formation column. 2 Cuttoff refers to the percentage of bark coverage on the predictive surface that was used to discriminate between classes. 3 Accuracy is a value generate by WEKA is based upon the results of the Confusion Matrix.

Surrounding vegetation analysis The second hypothesis addressed the effects of surrounding vegetation hypothesizing that shading by surrounding vegetation will slow bark formation rates. The portrait photos of each cactus in 2017 was separated into files based on surrounding vegetation present or absent around the cactus. Therefore, this portion of the experiment only utilized data from 2017. The data for percent bark covering the entire cactus stem was then averaged for each cactus in each of the groups. These groups were cacti with no surrounding vegetation, cacti with vegetation that do not shade the cactus plant, and cacti with vegetation that shade the cactus plant. The average bark coverage of these groups was analyzed with a t-test to determine whether there is a statistical significance between the groups of cacti. The results are shown in Fig. 3. Statistical analysis of this data concluded that cacti with surrounding vegetation that provide shade over the cactus plant are significantly different from cacti with non-shading surrounding vegetation and from cacti with no surrounding vegetation at all. The non-shading surrounding vegetation was not statistically different from cacti that had no surrounding vegetation. Vegetation surrounding cactus plant can only affect the amount of bark accumulation if the surrounding vegetation provides shade to the plant. By comparing the average amount of bark accumulated by each group, it was determined that shading vegetation slows bark formation by 13% on south crests. These results support the hypothesis. South crest surface was chosen because this was the surface of interest when analyzing cacti health at the initial start of this evaluation. Directional vegetation analysis The result from the surrounding vegetation analysis was furthered in this portion of the research. This analysis hypothesized that because the south is exposed to the most sunlight, shading

The Manhattan Scientist, Series B, Volume 4 (2017)

Bark Coverage (%)

LoCastro

56.6

51.3

45.7

50 40 30 20 10 0 No surrounding vegetation

Surrounding vegetation ‐ no shade

Surrounding vegetation ‐ shaded

Bark Coverage (%)

Figure 3. Histogram of the average amount of bark coverage for cacti with no surrounding vegetation, cacti with surrounding vegetation that do not provide shade, and cacti with surrounding vegetation that do provide shade. 60 50

50.2 37.5

40 30 20 10 0

Not shaded

Shaded

Figure 4. Histogram of the average amount of bark coverage on the entire cactus for cacti with south facing surfaces shading and for cacti with any surfaces besides the south facing surfaces shaded (i.e. north surface shading, east surface shading, and west surface shading).

on south-facing surfaces will greatly reduce bark formation as compared to the other cardinal directions. The group of photos for cacti with surrounding vegetation were further grouped based on the cardinal direction to which the shading vegetation was located. It was hypothesized that south side shading would have a significant influence on bark formation, therefore the photos of vegetation shading the south surfaces were isolated. The average amount of bark over the entire cactus for south side shading was compared to the average amount of bark over the entire cactus for shading anywhere besides the south side. The results of this statistical analysis by t-test are shown in Fig. 4. Statistical analysis of this data concluded that cacti with south side shading are significantly different from cacti with shading at the other three cardinal directions. By comparing the average amount of bark accumulated by each group, it was determined that south side shading by neighboring vegetation slows bark formation by 15% on south crests. This result supports the hypothesis. South crests were chosen because the south facing surfaces are most exposed to sunlight and the south was the major predictive surface for the analysis of bark formation.

The Manhattan Scientist, Series B, Volume 4 (2017)

LoCastro

Discussion Predicting health changes in saguaro cacti Based on data analysis, it was concluded that the health change of a saguaro cacti over an 8-year period can be predicted by south crest data from a previous 8-year interval at 91% accuracy. The average accuracy of the predictions exceeding 90% greatly supports the hypothesis that previous data over a time interval can be used to predict data in a sequential interval. This study can be furthered to find the major predictors that cause the change from one health status to next. Ultimately, this understanding can be used to isolate characteristics of extremely unhealthy cacti right before mortality to fully understand the premature death of saguaros. Surrounding vegetation analysis Based on data analysis, it was concluded that shading of cactus plants by neighboring vegetation decreased bark percentages on south crests by 13% (probability 0.003) from 2010 to 2017. Non-shading neighboring vegetation had no effect. This results supports the hypothesis that shading on a cactus plant will reduce bark formation. It is evident because when surrounding vegetation does not provide shade, the bark formation rates are the same as when there is no vegetation present around the cactus at all. These data can be used as an additional reference to the effect of sun exposure on these cactus plants. It is known that sunlight is the cause of epidermal browning, and by having data proving that shade from the sun slows cactus injury from epidermal browning, the understanding of sun exposure on cactus plants is further understood. It is known that sunlight is the cause of epidermal browning on cactus plants. This study is furthered in this research by investigating the effects of shading on cardinal directions. Directional Vegetation Analysis Based on data analysis, it was concluded that between 2010 and 2017, the presence of south surface shading by neighboring vegetation decreased changes in bark percentages by 15% on south crests (probability 0.01). It was hypothesized that shading on south facing surfaces will cause a decrease in bark formation rates because the latitude of Tucson Mountain Park results in the most sun exposed on south facing surfaces of cactus plants. The results of this analysis support the hypothesis and further support the effects of sunlight on cactus plants and the way that sunlight behaves in the particular desert of study. The south facing surfaces are subject to the most sunlight, therefore shading of these surfaces results in the greatest decrease in bark formation rates. This study can be furthered by investigating the effects of shading on north facing surfaces. The north facing surface receive the least amount of sunlight and form bark last as compared to the other cardinal directions. It can be hypothesized that because the north accumulates bark last and is least subject to injury, that shading can slightly prolong premature death by slowing bark formation on the north and maintaining the gas exchange that is vital for cactus life.

LoCastro

The Manhattan Scientist, Series B, Volume 4 (2017)

Acknowledgment This work was funded by the Catherine and Robert Fenton Endowed Chair to Dr. Lance Evans.

References Durisco, M. and S. J. Graban. 1992. Epidermal browning and population dynamics of giant saguaros in long-term monitoring plots, p 237-262. In C. Stone and S. Bellantoni [eds.], Proceedings of the Symposium of Research in Saguaro National Monument. Southwest Parks and Mountain Association, Tucson, AZ. Evans, L. S. 2005. Stem surface injuries of Neobuxmaumia tetetzo and Neobuxbaumia mezcalaensis of the Tehuacan Valley of Central Mexico. J. Torrey Bot. Soc. 132: 33-37. Evans, L. S. and B. J. Fehling. 1994. Surficial injuries of several long-lived columnar cacti of the Sonoran Desert, Environ. Exp. Bot. 34: 19-23. Evans, L. S. and A. Macri. 2008. Stem surface injuries of several species of columnar cacti of Ecuador. J. Torrey Bot. Soc. 135: 475-481. Evans, L. S., V. Cantarella, K. Stolte, and K. Thompson, 1994a. Epidermal browning of saguaro cacti (Carnegeia gigantea): Surface and internal characteristics associated with browning. Environ. Exp, Bot, 34: 9-17. Evans, L. S., V. Cantarella, L. Kaszczak, S. Krempasku, and K. Thompson. 1994b. Epidermal browning of saguaro (Carnegiea gigantea): Physiological effects, rates of browning, and relation to sun/shade. Environ. Exp. Bot. 34: 107-115. Evans, L. S., C. Mckenna, C. Ginocchio, G. Montengro, and R. Kiesling. 1994c. Surficial injuries of several cacti of South America. Environ. Exp. Bot. 34: 285-292. Evans, L. S., V. Sahi, and S. Ghersini. 1995. Epidermal browning of saguaro cacti (Carnegiea gigantea): Relative health and rates of surficial injuries of a population. Environ. Exp. Bot. 35: 557-562. Evans L. S., J. H. Sullivan, and M. Lim. 2001. Initial effects of UV-B radiation on stem surfaces of Stenocereus thurberi (Organ Pipe Cacti). Environ. Exp. Bot. 46: 181-87. Evans, L. S., M. Zugermayr, and A. J. Young. 2003. Changes in surface and mortality rates of saguaro (Carnegiea gigantea) cacti over a twelve-year period. J. Torrey Bot. Soc. 130: 238243. Evans, L. S., A. J. Young, and J. Harnett. 2005. Changes in scale and bark stem surface injuries and mortality rates of saguaro cacti (Carnegiea gigantea, Cactaceae) population in Tucson Mountain Park, Can. J. Bot. 83: 311-319. Geller, G. and P. Nobel. 1984. Cactus ribs: Influence of PAR inception and CO2 uptake. Photosynthesis 18: 482-494. Gibson, A. and P. Nobel. 1984. The Cactus Primer. Harvard University Press, Cambridge, MA. 286 p.

The Manhattan Scientist, Series B, Volume 4 (2017)

LoCastro

Hall, M., E. Frank, G. Holmes, B. Pfrahringer, P. Reutemann, and I. Wittten. 2009. The WEKA Data Mining Software: An Update, SIGKEE Explorations, Volume 11, Issue 1. Hitchlock, A. 1971. Manual of the Grasses of the United States. Doyer Publishers, New York, NY. 1050 p. Hyafil, L. and R. Rivest. 1976. Constructing optimal binary decision trees in NP-complete. Inform. Process. Lett. 5: 15-17. Turner, D. and C. Funicelli. 2000. Ten-year resurvey of epidermal browning and population structure of saguaro cactus (Carnegiea gigantea) in Saguaro National Park. United States Geological Survey Technical Reports No. 69, Tucson, AZ. 30 p. Turner, R. M. 1992. Long term saguaro population studies at Saguaro National Monument, p. 263267. In C. P. Stone and E. S. Bellantoni [eds.], Proceedings of the symposium on research in Saguaro National Monument. Southwest Parks and Monuments Association, Tucson, AZ. White, A., R. A. Dyer, and B. L. Sloane. 1941. The Succulent Euphorbieae (Southern Africa). Abbey Garden Press, Pasadena, CA. 937 p.

Automated quantitative analysis of tree branch similarity using 3D point cloud registration Matthew Maniscalco∗ Department of Mechanical Engineering, Manhattan College Abstract. Over the last decade commercial three-dimensional scanners have been becoming more obtainable by the general public. Three-dimensional scanners are now being produced at a lower cost than in previous years as well as providing high accuracy and ease of use. These scanners are also being used in many fields such as in the manufacturing industry, the medical field, and in engineering fields. It allows for faster production and comparison as the object that is in question can be easily digitized and analyzed. For this research a 3D scanner was selected, as it would provide high accuracy as well as quick analysis in the comparison of the similarity of tree branches. Branches were gathered rom around the Manhattan College campus. For each tree two terminal branches were removed from the tree and then scanned and compared using a program developed in MATLAB[1].

Introduction Tree branches may seem to grow in unpredictable geometries, but they are not as erratic as they seem. Leonardo Da Vinci [2] noticed that there is a correlation between the area of a tree trunk and the branches along any arced section encompassing the tree. It was thought that if the branch areas of one layer were summed, they would be equal to the trunk area. The purpose of this research is to use the newest methods available to develop a computer program that can compute the similarity between terminal branches across different species of trees. If the geometries are consistently similar it may be possible to model tree branch growth with a formula and be able to predict how different tree species may grow over time. This would be a revolutionary discovery as something like this has never been accomplished before. In a previous study [3], the tree branches had to be analyzed manually, They were input to an Excel spreadsheet, and then analyzed using “3D Simquant,” a program developed using MATLAB. This program was developed to compare the coordinates of the branches using approximately 90 points; they are are not enough points to yield an accurate comparison. To increase efficiency and accuracy, a commercial 3-dimensional scanner was used and 5000 points on each branch will be considered for the comparison. The scanner will be able to obtain the thickness and curves of each branch with a 0.1 mm accuracy. This method will also greatly decrease the time it takes to compare the similarity between branches. The hypothesis of this study is: Can 3D scanned images, coupled with the use of the MATLAB computer vision toolbox and Cloud Compare, produce an accurate comparison of tree branches with various thicknesses and geometries? ∗ Research mentored by Zhara Shahbazi, Ph.D., and Lance Evans, Ph.D.

The Manhattan Scientist, Series B, Volume 4 (2017)

Maniscalco

Materials and Methods For one analysis two terminal branches, both of which are from tree species found in the vicinity of Manhattan College, were scanned with a commercial scanner [4]. After successful scans, stereolithography (STL) files [5] were constructed for analysis. A STL file represents each stem surface as a mesh of triangles that are joined together. All portions of the surface are covered with triangles (Fig. 1). The STL files were then imported into CloudCompare [6]. CloudCompare

Figure 1. Stereolithography of scanned tree branches.

is a software that allows the creation, editing and analyzing of point clouds. For this research, only the creation and editing tools were needed. For most stem pairs in CloudCompare the number of points used was 5000 per branch when creating point clouds. Increasing the amount of points to more than 5000 had little effect on the RMSE value obtained. Using 5000 points provided the best compromise between computational time and accuracy, as shown in Fig. 2. Using a large number 5

RMSE Value

4 3 2 1 0

Number of Sample Points (×104)

Figure 2. RMSE value as a function of the number of points used per branch

of points optimizes the image to produce the lowest possible RSME values. A point cloud is a group of points on a 3D coordinate system that, when viewed, create an image of a surface. An example of a point cloud can be seen in Fig. 3. Both branches from Fig. 1, are converted into a point cloud in CloudCompare. Next, branch pairs are oriented such that their ZY planes are aligned and the bottom ends are also aligned, as shown in Fig. 3. This alignment prior to exporting as PCD files is required to prevent errors when the image is run through “3D Register” MATLAB code. A PCD file is a point cloud database file. The file stores the coordinates of all the points generated in CloudCompare. The PCD files are then imported into the “3D Register” through the GUI Fig. 4. The purpose of 3D Register is to align the two images of each species so that the images can be compared analytically

Maniscalco

The Manhattan Scientist, Series B, Volume 4 (2017)

Figure 3. Point cloud

Figure 4. “3D Register”

to determine their similarities. The larger branch is inputted as “PointCloud1” and the other branch goes into “PointCloud2”, then “RUN” is pressed. The program then computes the RMSE value of the two branches and displays a visual of how the branches were matched up. A sample of what the program would display for the alignment of the branches is shown in Fig. 5. The result from

Figure 5. sample output of 3D Register

3D Register displays three images, the one on the left represents the image that is imported into the program by the user. The center image is the two branches scaled to approximately the same size. The third image is the part in which the program tries to align both branches and computes the RMSE value based on how well the points align; the RMSE value is then displayed to the user. The lower the RMSE value, the more similar the branches. For example, an RMSE value of zero to one represents almost identical branches, an RMSE value from one to two represents very similar branches, two to three represents not so similar, and anything higher than three are two distinct branches with very little similarity in geometry.

Results and Conclusion The use of a 3D scanner streamlined the branch comparison process in comparison to the previous method used. Combining 3D scanning with MATLAB’s Vision toolbox produced somewhat

The Manhattan Scientist, Series B, Volume 4 (2017)

Maniscalco

consistent results, although there are many areas in which it can be improved. The branches that were scanned showed very interesting results. The table shows the four-different species of trees that were compared using 3D Register, and their corresponding RMSE values. The RMSE value obtained for species one shows that there is some similarity between the two terminal branches, for species two there Species RMSE Values is very high similarity, so they have almost identical geometry. For Species 1: 3.80 species three and four the RMSE value indicates that these pairs Species 2: 0.18 are not very similar. Species 3: 9.14 Species 4:

9.62

The main outcome of this research was to see if 3D scanned images coupled with the use of the MATLAB computer vision toolbox and Cloud Compare can produce an accurate comparison between terminal tree branches. This research proves that it is possible to compare the geometries of various tree branches. Now that this method has been developed, more terminal branches need to be compared and analyzed. For each tree species a minimum of five terminal branches will be compared to see how the RMSE values change depending on location of the branches on the tree species. Once this is accomplished than conclusions can be drawn. This research may also be of use to other fields such as the medical field. If a patient is in need of a new bone to be constructed and transplanted into their body a 3D scan of the old bone can be compared with that of the fabricated one. This will then tell doctors if the fabricated bone is close enough in geometry to be a good replacement, which can be very helpful. It could also be used in the engineering field to compare the similarity of gears or other mechanical devices.

Acknowledgments This work was funded by the Catherine and Robert Fenton Endowed Chair to Dr. Lance Evans. The author thanks Drs. Z. Shahbazi and L. S. Evans for valuable advice.

References [1] MathWorks (2017). Computer Vision System Toolbox [2] Aratsu, R. (1998). â&#x20AC;&#x153;Leonardo Was Wise - Trees Conserve Cross- Sectional Area Despite Vessel Structure.â&#x20AC;? J. Young Investigators, 1. [3] Brucculeri, J. A., Evans, L., and Shahbazi, Z. (2017). Automated Quantitative Analysis of Terminal Tree Branch Similarity by 3D Registration, ASME 2017 International Design Engineering Technical Conferences and Computers and Information in Engineering Conference (58110), pp. V001T02A004; 12 pages. doi:10.1115/DETC2017-67831 [4] https://www.einscan.com/einscan-pro [5] StereoLithography Interface Specification, 3D Systems, Inc., October 1989 [6] CloudCompare (version 2.8.1) [GPL software]. (2017). http://www. cloudcompare.org

Evidence of Giardia lamblia oocyst stage in bivalves collected in Bronx, NY Monique Ng∗ Department of Biology, Manhattan College Joseph Annabi∗ Fordham Preparatory School, Bronx, NY 10458 Abstract. Bivalves are important bio-indicators due to their sensitivity to pollution, their potential in trapping pollutants, and their widespread distribution. Bivalves are filter feeder organisms, and may retain parasitic oocysts, such as Giardia lamblia (G. lamblia) oocysts in their tissue. Infections by G. lamblia lead to symptoms such as diarrhea, and cause giardiasis. This project focuses on three bivalve species that were collected at Orchard Beach (OB) and Clason Point Soundview (SV): Mya arenaria (55 at OB), Crassostrea virginica (17 at OB; 26 at SV), and Geukensia demissa (39 at OB). The bivalves were previously tested for the presence of β-giardin DNA, which determines exposure to G. lamblia. The goal of this project was to find evidence of hard-shelled-, weather-, and chemical-resistant oocysts in bivalves collected from Orchard Beach and Clason Point Soundview. This was achieved by detecting the gene for oocyst cell wall proteins, CWP2, using the polymerase chain reaction. We found that CWP2 was not detected in Geukensia demissa samples collected at Orchard Beach. In contrast, we observed a prevalence of 60% in Mya arenaria collected from Orchard Beach. A prevalence of 24% was observed in Crassostrea virginica collected from Orchard Beach, while a prevalence of 11.5% was observed in Crassostrea virginica collected from Clason Point Soundview; the two sites are only ten miles away from each other.

Introduction The intestinal parasite, Giardia intestinalis, which is synonymously known as Giardia lamblia or Giardia duodenalis, causes giardiasis. G. lamblia is a parasitic organism that causes diarrhea, but is also known to stunt cognitive development in children of developing countries due to persistent malnutrition (Berkman et al., 2002). The Center for Disease Control (CDC) reports 33% of individuals in developing countries have contracted giardiasis; it is also the most common gastrointestinal disease in the United States (CDC, 2015). According to Collier et al. (2012), giardiasis has cost the U.S. healthcare system roughly $35 million; therefore monitoring the intestinal parasite is crucial to the health of the population. Giardiasis is spread by waterborne transmission of the oocysts, which contaminate water and swimming pools. The parasite is spread through other mechanisms, such as foodborne transmission. It has been reported that fruits and vegetables (Robertson et al., 2001), contaminated school milk (Gelletlie et al., 1997), and eating undercooked food contaminated with the parasite are all modes of transmission (CDC, 2015). When the parasite encounters harsh environmental factors, it begins encystment, and releases the oocysts in feces. Giardia is spread through its infectious oocysts, and it has been determined ∗

Research mentored by Ghislaine Mayer, Ph.D.

The Manhattan Scientist, Series B, Volume 4 (2017)

that as few as 10 oocysts can cause an infection (Adam, 2001). The hard-shelled oocysts persist in unfavorable conditions, such as low temperature water, and they are resistant to chemical disinfection such as chlorine (CDC, 2015). One method of tracking the parasite’s effect on the environment is based on utilizing filter feeder aquatic invertebrates such as bivalves, because they can act as reservoirs to infectious diseases (Golberg, 1986). Bivalves have previously been used as bio-sentinels, and are now being used to track fecal contamination because of their precise detection (O’Connor, 2002). It is crucial to monitor hard-shelled oocysts in aquatic environments, because poor water quality may affect bivalves for human consumption. The use of bivalves as bio-indicators is ubiquitous. Blue mussels (Mytilus edulis) have been previously used to track fecal contamination from sewage overflow due to heavy rainfall (Tryland et al., 2014). A correlation was found between heavy rainfall and sewage overflow, along with an increase in parasite in the water and an increase in parasites in mussels (Tryland et al., 2014). In 2012, there was a 66.7% prevalence of β-giardin DNA in the Mytilus galloprovincialis collected from Italy (Giangaspero et al., 2014). Previous studies in the U.S. have also used bivalves as bioindicators for infectious oocysts. G. lamblia oocysts along with other waterborne pathogens such as Cryptosporidium parvum oocysts were detected in the Chesapeake Bay in oysters (Graczyk et al., 2006). The only report of intestinal parasites in bivalves from New York City were samples collected at Orchard Beach, Bronx, NY (Tei et al., 2016). In the fall of 2014, from the bivalves collected at Orchard Beach, the prevalence of G. lamblia was 37.5% in Mya arenaria, 4.5% in Geukensia demissa, 60% in Crassostrea virginica, and 20.6% in Mytilus edulis, respectively (Tei et al., 2016). In the fall of 2015, the prevalence of G. lamblia was found to be 33% in C. virginica (Kowalyk, 2015). The following year, the prevalence of G. lamblia was 60% in Mya arenaria, 23.53% in Crassostrea virginica found at Orchard Beach, and 11.5% in Crassostrea virginica found at Clason Point Soundview. There were no oocysts detected in Geukensia demissa samples collected at Orchard Beach. The goal of this study is to assess whether the bivalve tissues infected with G. lamblia have the potential to be infectious.

Materials and Methods Sample collection and DNA isolation On September 15, 2016, the bivalves collected at low tide at Orchard Beach (OB) and Clason Point Soundview (SV) were as follows: Mya arenaria (55 at OB; 1 at SV), Crassostrea virginica (17 at OB; 26 at SV), Geukensia demissa (39 at OB; 37 at SV), Mytilus edulis (61 at OB; 0 at SV). The mantle, the foot, the digestive tract, and the gill were dissected from each large specimen. The DNA and RNA of the tissue were extracted using the Qiagen DNeasy Blood & Tissue Kit (Qiagen, Valencia, CA). DNA was extracted according to the manufacturer’s protocol. The elution was repeated twice (200 µL the first elution and 150 µL the second elution) for maximum DNA yield.

The Manhattan Scientist, Series B, Volume 4 (2017)

Figure 1. Detection of CWP2 amplicon in Mya arenaria collected at Orchard Beach Top lane: Lane 1: 100 bp marker ; Lane 3-19: DNA positive for CWP2. Middle lane: Lane 1: 100 bp marker; Lane 3-10 11-15: DNA positive for CWP2. Bottom lane: Lane 1: 100 bp marker ; Lane 3-9: DNA positive for CWP2 ; Lane 12: positive control *Negative control not shown

Polymerase chain reaction and agarose gel electrophoresis Oocysts in G. lamblia were detected using oligonucleotide primers targeting the cyst wall protein 2 gene (CWP2). The forward primer sequence was: 5’-CTCTTCGACCTGCCTTACATGAT3’; the reverse primer sequence was: 5’-CAAACGAGATCGGTGTTGCA-3’ (Eligio-Garcia et al., 2010). The PCR conditions were as follows: 95◦ C for 5 min, 95◦ C for 30 s, 55◦ C for 30 s, 72◦ C for 1 min, 72◦ C for 5 min, for 40 times, and 4◦ C until the gel electrophoresis. The amplicon is 58 bp, therefore a 2.5% agarose stained with ethidium bromide was used for visualization.

Results Prevalence of Giardia lamblia oocyst in Mya arenaria Fifty-five Mya arenaria specimens were collected in 2016 and were tested using PCR to detect CWP2 DNA. Thirty-three out of 55 samples tested positive for G. lamblia, resulting in a 60% prevalence (Table 1). Prevalence of Giardia lamblia oocyst in Geukensia demissa Thirty-nine samples of Geukensia demissa were collected at Orchard Beach, and were previously tested by James Limonta for the presence of β-giardin. None of the eleven β-giardin positive

The Manhattan Scientist, Series B, Volume 4 (2017)

Table 1. Detection of CWP2 in various species collected from Orchard Beach Species Name

Prevalence of CWP2 DNA

Mya arenaria

60% (33/35)

Geukensia demissa

0% (0/39)

Crassostrea virginica

23.53% (4/17)

samples were positive for the CWP2 gene, thus there were no oocyst DNA detected (Table 1). Prevalence of Giardia lamblia Oocyst in Crassostrea virginica collected from OB Seventeen samples of Crassostrea virginica were collected at Orchard Beach, and were previously tested by James Limonta for the presence of β-giardin, which detects the presence of G. lamblia. Nine samples tested positive for β-giardin (James Limonta unpublished observation). Four samples were positive for the CWP2 gene, thus resulting in a 23.53% prevalence (Table 1). Prevalence of Giardia lamblia oocyst in Crassostrea virginica collected from SV Twenty-six samples of Crassostrea virginica were collected from Clason Point Soundview, and were previously tested by James Limonta for the presence of β-giardin. Of the four individual samples, three were positive for the CWP2 gene, thus resulting in a 11.5% prevalence (Table 2). Table 2. Detection of CWP2 in Crassostrea virginica collected from Soundview Species Name

Prevalence of CWP2 DNA

Crassostrea virginica

11.5% (3/26)

Prevalence of Giardia lamblia oocyst in bivalve tissues The tissue distribution of oocysts was determined by dissecting the mantle, digestive glands, gills, abductor, hemolymph, foot, and siphon of each bivalve specimen. The oocyst DNA was predominately found in the siphon, foot, digestive glands, and mantle. In contrast, G. lamblia was found in low levels in the hemolymph (Table 3). Oocyst genes were not detected in any tissues of Geukensia demissa.

Discussion In 2016, we found G. lamblia was detected in Mya arenaria at 60%, whereas in 2015, G. lamblia was detected at a prevalence of 37.5% (Tei et al., 2016). Our findings in Mya arenaria show an increase in G. lamblia detection among bivalves collected at the same location. In 2016 at Orchard Beach, we found G. lamblia was detected in C. virginica at a prevalence of 23.53%, whereas in 2015, G. lamblia was detected at 60% (Freda Tei, unpublished observation). Our findings in Crassostrea virginica show a decrease in G. lamblia detection, thus indicating a healthier

The Manhattan Scientist, Series B, Volume 4 (2017)

Table 3. Prevalence of Giardia lamblia distribution in bivalve tissues Mollusks

Mantle

Digestive

Gills

Abductor

Hemolymph

Foot

Siphon

Mya arenaria

21.8% (12/55)

23.6% (13/55)

14.5% (8/55)

18.2% (10/55)

20% (11/55)

11.76% (2/17) 3.85% (1/26)

5.88% (1/17) 3.85% (1/26)

11.22%

4.65%

18.18%

20%

Geukensia demissa Crassostrea virginica- OB Crassostrea virginica- SV

11.76% (2/17) 11.54% (3/26)

11.76% (2/17)

7.69% (2/26)

Total

17.35%

18.52%

13.89%

aquatic environment with less G. lamblia contamination. On the contrary, in 2016 at Clason Point Soundview, we found G. lamblia was detected in Crassostrea virginica at 11.5%. Bivalve collection did not take place at Clason Point Soundview prior to 2016, therefore there was no data for comparison. Despite the fact that Mya arenaria, Crassostrea virginica, and Geukensia demissa samples were all collected a few feet away from each other at Orchard Beach, they varied in levels of prevalence. Out of 39 samples of Geukensia demissa, no samples were detected with oocyst DNA. Based on the results, this suggests that Geukensia demissa may be a poor model system for detecting G. lamblia oocysts. This could be due to the differential ability of these bivalves to retain G. lamblia oooycts. It would be interesting to compare the susceptibility of Geukensia demissa to oocysts of other protozoans. Crassostrea virginica bivalves were collected at two locations, Orchard Beach and Clason Point Soundview. The Crassostrea virginica samples collected at Orchard Beach had a 23.53% prevalence, whereas the prevalence was 11.5% at the Soundview location. It is intriguing that there is a drastic difference in G. lamblia oocyst prevalence when the beaches are very close. The distribution of G. lamblia oocysts among the Crassostrea virginica at both locations showed that the mantle had a higher prevalence. This may be because the mantle is part of the feeding structure, which captures infectious oocysts. The varying levels in prevalence among Crassostrea virginica may be attributed to collection at separate beaches, which may differ in levels of G. lamblia oocysts. The distribution of oocyst DNA was found at a relatively similar prevalence in tissues of Mya arenaria and Crassostrea virginica. Therefore, it seems that each tissue was relatively equally affected. There was an overall prevalence of 60% in Mya arenaria tissue. Based on the results, the high prevalence for oocyst DNA detected in Mya arenaria indicate that this bivalve may be an excellent bio-indicator for G. lamblia contamination in the aquatic environment. In summary, we have found evidence of G. lamblia oocysts in the bivalves collected from two beaches in the Bronx, NY. This strongly suggests that these bivalves have the potential of being infectious if consumed raw.

The Manhattan Scientist, Series B, Volume 4 (2017)

Acknowledgment This work was funded by the Linda and Dennis Fenton ’73 endowed biology research fund. The authors would like to express their gratitude to Dr. Ghislaine Mayer for mentoring this research, and thank Jordan DeCade for her work on the β-giardin detection.

References Adam, R. D. (2001). Biology of Giardia lamblia. Clin Microbiol Rev 14, 447–475 Berkman, D. S., Lescano, A. G., Gilman, P. H., Lopez, S. L., and Black, M. M. (2002). Effects of stunting, diarrhoeal disease, and parasitic infection during infancy on cognition in late childhood: a follow-up study. The Lancet, 359, 564-571. doi:10.3410/f.1004822.55404 CDC (Centers for Decease Control) (2015). Parasites - Giardia. (2015, July 21). Retrieved July 31, 2017, from https://www.cdc.gov/parasites/ giardia/infection-sources.html#one Collier, S. A., Stockman, L. J., Hicks, L. A., Garrison, L. E., Zhou, F. J., and Beach, M. J. (2012). Direct healthcare costs of selected diseases primarily or partially transmitted by water. Epidemiology and Infection ,140, 2003-2013. doi:10.1017/s0950268811002858 Eligio-Garc´ıa, L., Pilar, C. M., Andr´es, F., Apolinar, C., Adri´an, C., and Enedina, J. (2011). Giardia intestinalis: Expression of ubiquitin, glucosamine-6-phosphate and cyst wall protein genes during the encystment process. Experimental Parasitology, 127(2), 382- 386. doi:10.1016/j.exppara.2010.08.017 Gelletlie, R., Stuart, J., Soltanpoor, N., Armstrong, R., and Nichols, G. (1997). Cryptosporidiosis associated with school milk. The Lancet, 350, 1005-1006. doi:10.1016/s0140-6736(05)64071-8 Goldberg, E. D. 1986. “The mussel watch concept.” Envir. Monit. Assess. 7: 91-103 Graczyk, T. K., Girouard, A. S., Tamang, L., Nappier, S. P., and Schwab, K. J. (2006). Recovery, Bioaccumulation, and Inactivation of Human Waterborne Pathogens by the Chesapeake Bay Nonnative Oyster, Crassostrea ariakensis. , Appl Environ Microbiol 72, 3390-3395. doi:10.1128/aem.72.5.3390-3395.2006 Kowalyk, S. M. (2015). Prevalence of Human Intestinal Parasites in Atlantic Oysters. Manhattan Scientist, 2, B, 47-52. Retrieved September 10, 2017. O’Connor, T. P. (2002). National distribution of chemical concentrations in mussels and oysters in the USA. Mar Environ Res, 53, 117-143. doi:10.1016/s0141-1136(01)00116-7 Robertson, L. J., and Gjerde, B. (2001). Occurrence of Parasites on Fruits and Vegetables in Norway. Journal of Food Protection, 64, 1793-1798. doi:10.4315/0362-028x-64.11.1793 Tei, F., Kowalyk, S., Reid, J., Presta, M., Yesudas, R., and Mayer, D. (2016). Assessment and Molecular Characterization of Human Intestinal Parasites in Bivalves from Orchard Beach, NY, USA. Int J Environ Res Public Health , 13, 381. doi:10.3390/ijerph13040381 Tryland, I., Myrmel, M., Østensvik, Ø, Wennberg, A. C., and Robertson, L. J. (2014). Impact of rainfall on the hygienic quality of blue mussels and water in urban areas in the Inner Oslofjord, Norway. Marine Pollution Bulletin, 85, 42-49. doi:10.1016/j.marpolbul.2014.06.028

Xylem conductivities from stems to leaves Humberto Ortega∗ Laboratory of Plant Morphogenesis, Department of Biology, Manhattan College Abstract. Grasses are the main sources of carbohydrates for human populations worldwide. Grasses grow in a wide range of environments from wet rice paddies in Asia to deserts with restricted water. Thus, research regarding water distribution characteristics from grass stems to leaves are important. The purpose of the present study is to understand the characteristics of xylem cells, the cells that transport water, their characteristics within vascular bundles and characteristics of xylem conductivity. The latter is a parameter used to estimate the ability of cells to transport water, from stems to leaves. Data from fourteen species of C3 grasses, species mostly confined to temperate climates, and thirteen species of C4 grasses, species found predominately in tropical climates, were obtained with standard histological procedures. As expected, stems with large stem diameters had more vascular bundles. For both grass groups, the number of vascular bundles were linearly related with stems diameters (y = 25.1x − 17.6; r2 = 0.92). The number of bundles in leaves were also linearly related with number of bundles in stems (y = 0.097x + 7.00; r2 = 0.85). As a consequence, xylem conductivity values per bundle in leaves were linearly related with number of bundles in stems (y = 0.052x–0.004, r2 = 0.80). Moreover, xylem conductivity values of entire leaves were linearly related with of entire stems (y = 0.060x + 0.036, r2 = 0.89). Overall, leaf conductivity values were about nine percent of stem conductivities.

Introduction Grass species are among the most ubiquitous group of plants on Earth. They occur in various ecosystems and they form a biome in both temperate (grasslands) and tropical (savannah) environments. Grasses serve as a food source for a large number of species. For example, humans rely on many crops and grains that are derived from grass plants (Table 1). Table 1. Worldwide production of food grains from cultivated grasses for 2013 (Ortega, 2016) Grain Zea mays (corn) Triticum aestivum (wheat) Oryza sativum (rice) Hordeum vulgare (barley) Sorghum Oats

Grain Production (metric tons) 817,000,000 681,000,000 678,000,000 123,000,000 61,000,000 20,000,00

www.nueokstate.edu www.geohive.com www.agrostats.com www.spectrumcommodities.com www.nationmaster.com

Grass plants serve as a major source of food in the diets of most human societies. In grassland and savannah ecosystems, grasses provide food for migrating and stationary animal species. American buffalo (Bison bison) followed the growth and maturity of grasslands in the Midwest ∗

Research mentored by Lance Evans, Ph.D.

The Manhattan Scientist, Series B, Volume 4 (2017)

Ortega

from Mississippi to Minnesota throughout the year. Wildebeest (Connochaetes taurinus) migration on the Serengeti plain is highly influenced by the timing of grass growth when the annual rains occur. In order for grasses to reach maturity and produce grains, adequate amounts of water must be transferred from stems to leaves. Water in grass plants is transported from roots to stems to leaves (Fig. 1) while some water is transported to flowers and seeds.

Figure 1. Diagram of water flow throughout the tissues up to transpiration of water vapor from the leaf (Ortega, 2016).

Figure 2. Diagram of vessel distribution in the grass stem and leaf blade(Ortega, 2016).

In vascular bundles, water is only transported in xylem vessel cells. Thus, vascular bundles and xylem cells in vascular bundles form a continuum from the roots to the leaves, flowers, or fruits. One aspect of this study was to determine if water transport characteristics were similar among several grass species. Other stem and leaf properties such as stem diameter, leaf width, bundle count, average vessels per bundle, conductivity per bundle, and conductivity per xylem vessel were measured in order to understand the water transport phenomena in grasses. We hypothesized the following: 1. 2. 3. 4.

Species with larger stem diameters have more bundles per stem. Species that have more bundles per stem have more bundles in leaves. Species with larger stem conductivities per bundle have larger leaf conductivities per bundle. Species with larger stem xylem conductivities have larger leaf xylem conductivities.

Ortega

The Manhattan Scientist, Series B, Volume 4 (2017)

Materials and Methods Plant sampling Grass plants were sampled from numerous locations in tropical and temperate climates (Table 2). A large number of photographs were processed of the geographic locations, the growth habit, and plant parts for each species. Plants taken as samples had no morphological abnormalities and appeared normal for other plants in the area. All plants had flowers or seeds that were used for species identification. Wagner et al. (1999) and Hitchcock (1950) were used for species identification. All species names were verified with http://www.tropicos.orgwww.tropicos.org and http://www.ipni.orgwww.ipni.org. Tissue sampling Sampled grass plants were subdivided into stem tissues below nodes, leaf sheath tissues, and leaf blade tissues. Each tissue sampled was 0.5 cm to 3.5 cm in length. In some cases, the leaf blade had two samples—one near the leaf sheath and the other near the center of the leaf. In addition to these tissues, stem diameters, leaf widths, number of bundles, and vessel radii were measured. Histological procedures Tissue samples were fixed in FAA (Jensen, 1962) for 24 hours. After fixation, samples were put through a series of tertiary butanol (Fisher Scientific, Pittsburgh, PA). Tissues were placed in liquid Paraplast-Xtrawax (McCormick Scientific, Richmond, IL) in an oven at 56◦ C. After a second change of wax, tissue samples were embedded in Paraplast. Tissues were sectioned with a microtome at 35 µm. Sections were stained in Safranin (Jensen, 1962) and later made permanent with Canada balsam (CAS 8007-47-4, Acros, Fisher Scientific, Pittsburgh, PA). Microscopic analysis The number of vascular bundles in each tissue were determined. The mean number of vessel conduits per vascular bundle was determined from at least 10% of all vascular bundles in a tissue. In 5% of all bundles, diameters of individual vessels were determined using ImageJ (National Institutes of Health, http://rsb.info.nih.gov/ij). Two diameter measurements at right angles were determined for each conduit. Xylem conductivities were determined using the Hagen-Poiseuillle equation (McCulloh et al., 2009), π × number of conduits × average radius of conduits(cm)4 . 8 × viscosity of water

The units of xylem conductivity are g cm MPa−1 s−1 .

Statistical analysis Individual conductivity values were used to find average conductivities in each tissue. The average tissue conductivity was multiplied by the number of bundles in a tissue to provide the average conductivity of the tissue. Mean values were obtained from two samples of each species.

The Manhattan Scientist, Series B, Volume 4 (2017)

Ortega

Table 2. Grass species sampled for this study. Species C3 species ´ ove& D.L¨ove Agropyron junceiforme A.L¨ Alopecurus pratensis fo. brachyglossus (Peterm.) S¸erb. & Ny´ar. Arundo donax ssp. plinii (Turra) Mateo & Figuerola Bromus rigidus ssp. ambigens (Jord.) Pignatti Calamagrostis × acutiflora (Schrad.) DC. Manhattan College, Bronx, NY (40.89 N, 73.90 W), August 2013 Dactylis glomerata ssp. nestorii Rossello & L. Saez Festuca rubra ssp. villosa (Mert. ex Koch) S.L. Lu Hordeum vulgare var. abdulbasirovii Omarov Koelaria glauca ssp. pohleana (Domin) Tzvelev Lolium multiflorum fo. submuticum (Mutel) Anghel & Beldie Phragmites australis ssp. berlandieri (E. Fourn.) Saltonstall & Hauber Poa nemoralis var. popovii Tzvelev Poa pratensis ssp. stenachyra (Keng ex Keng f. & G.Q. Song) Soreng & G.H. Zhu Sphenopholis intermedia var. macrantha B. Boivin C4 species Andropogon virginicus var. decipiens C.S. Campb. Axonopus fissifolius var. polystachyus (G.A. Black) L.B. Sm. & Wassh. Cenchrus agrimonioides var. laysanensis F. Br Chloris gayana fo. oligostachys (Barratte & Murb.) Maire & Weiller Digitaria fuscescens (J. Presl) Henrard Digitaria insularis (L.) Fedde Digitaria setigera var. calliblepharata (Henrard)Veldkamp Miscanthus sinensis ssp. condensatus (Hack.) T. Koyama Pennisetum alopecuroides ssp. sordidum (Koidz.) T. Koyama Saccharum officinarum ssp. sinense (Roxb.) Burkill Sacciolepis indica ssp. oryzetorum (Makino) T. Koyama Sorghum halepense var. propinquum (Kunth) Ohwi Zea mays ssp. huehuetenangensis (Iltis & Doebley) Doebley

Location, sample date Ring of Kerry, County Kerry, Ireland (52.15 N, 9.57 W), July 2016 Tarrytown, NY (41.08 N, 73.86 W), September 2014 Hanga Roa, Isla de Pascua (Easter Island) (27.09 S, 109.26 W), March 2016 University of California, Riverside, CA (33.97 N, 117.32 W), August 2015

Manhattan College, Bronx, NY (40.89 N, 73.90 W), August 2013 Ring of Kerry, County Kerry, Ireland (52.15 N, 9.57 W), July 2016 Manhattan College, Bronx, NY (40.89 N, 73.90 W), July 2013 Ring of Kerry, County Kerry, Ireland (52.15 N, 9.57 W), July 2016 Cashel, County Tipperary, Ireland (52.52 N, 7.89 W), July 2016 Van Cortlandt Park, Bronx, NY (40.90 N, 73.89 W), July 2013 Cashel, County Tipperary, Ireland (52.52 N, 7.89 W), July 2016 Manhattan College, Bronx, NY (40.89 N, 73.90 W), July 2013 Manhattan College, Bronx, NY (40.89 N, 73.90 W), July 2013 Hanavave, Fatu Hiva, Marquesas Islands (10.47 S, 138.66 W,) March 2016 Atuona, Hiva Oa, Marquesas Islands (9.80 S, 139.04 W), March 2016 Kihei, Maui, Hawaii (20.75 N, 156.46 W), July 2016 University of California, Riverside, CA (33.97 N, 117.32 W), August 2015 Hakata ‘O, Ua Pou, Marquesas Islands (9.61 S, 40.09 W), March 2016 Kihei, Maui, Hawaii (20.75 N, 156.46 W), July 2016 Hakata ‘O, Ua Pou, Marquesas Islands (9.61 S, 40.09 W), March 2016 Manhattan College, Bronx, NY (40.89 N, 73.90 W), July 2013 Manhattan College, Bronx, NY (40.89 N, 73.90 W), July 2013 Kihei, Maui, Hawaii (20.75 N, 156.46 W), July 2016 Hanga Roa, Isla de Pascua (Easter Island) (27.09 S, 109.26 W), March 2016 Tarrytown, NY (41.08 N, 73.86 W), July 2013 Tarrytown, NY (41.08 N, 73.86 W), July 2013

IPNI: Plant Name Search. International Plant Names Index, 23 April 2015. http://www.ipni.org/ipni/plantnamesearchpage.do; http://www.tropicos.org. August 2017.

Ortega

The Manhattan Scientist, Series B, Volume 4 (2017)

Results The purpose of this study was to determine relationships between water transport in stems and leaves of several grass species. Overall, there were no statistically significant differences between the xylem characteristics for C3 and C4 grass species (Tables 3 and 4). Since there were no differences between species of both groups, data from all species were pooled for further analyses. Table 3. Characteristics of stems of C3 and C4 grass species pertaining to water transport. Stem diameter (mm)

Vessel radii (µm)

Number of bundles

Bundle conductivity (g cm MPa−1 s−1 )

Stem conductivity (g cm MPa−1 s−1 )

C3 species Largest value Smallest value Mean S.D.

11.9 1.04 2.66 2.78

25.3 6.5 13.7 6.1

343 15.0 49.6 85.7

0.109 0.0008 0.0233 0.039

37.5 0.015 3.32 9.93

C4 species Largest value Smallest value Mean S.D.

19.2 1.28 4.48 4.76

30.9 8.7 16.7 7.25

444 20 94.7 115.5

0.208 0.0014 0.0398 0.0604

92.4 0.053 9.32 25.2

Species

Results of t-tests indicated no statistically significant differences between data for the C3 and C4 species tested for all parameters tested.

Table 4. Characteristics of leaves of C3 and C4 grass species pertaining to water transport. Width (mm)

Number of bundles

Bundle conductivity (g cm MPa−1 s−1 )

Leaf conductivity (g cm MPa−1 s−1 )

C3 species Largest value Smallest value Mean S.D.

15.6 2.0 6.53 4.39

63 5.0 14.99 16.8

0.0257 0.00015 0.00475 0.0077

1.08 0.00129 0.13 0.00771

C4 species Largest value Smallest value Mean S.D.

15.1 3.07 7.91 4.07

45 5 16.7 11.6

0.131 0.00029 0.0205 0.0402

5.88 0.0028 0.709 1.67

Species

Results of t-tests indicated no statistically significant differences between data for the C3 and C4 species tested for all parameters tested.

In this study, stem diameters ranged from 1.04 mm to 19.2 mm and the number of bundles

The Manhattan Scientist, Series B, Volume 4 (2017)

Ortega

in these stems ranged from 15 to 444 bundles per stem. Diameters of stems among these species were positively scaled with numbers of vascular bundles [y = 25.1x − 17.6, r2 = 0.92] (Fig. 3). This close scaling indicates that stems are well designed to conduct water towards leaves and reproductive structures.

Figure 3. Bundle count vs. stem diameter. Circles, C3 species; triangles, C4 species.

Figure 4. Leaf bundle count vs. stem bundle count. Circles, C3 species; triangles, C4 species.

Relationships between water conduction characteristics from stems to leaves were well scaled. Numbers of xylem bundles in leaves were well scaled with numbers of bundles in stems among the species tested [y = 0.097x + 7.00; r2 = 0.85] ( Fig. 4). Numbers of bundles in stems ranged from 15 to 444 bundles per stem, while numbers of bundles in leaves ranged from 5 to 63 bundles per leaf. Overall, larger stems had more bundles than smaller stems (number of bundles vs stem diameter; y = 25.1x − 17.6; r2 = 0.92) and the numbers of bundles in leaves was well scaled with numbers of bundles in stems (y = 0.097x + 7.00; r2 = 0.85). Among all species tested, the number of xylem vessels per bundle was between two and four. Xylem conductivities were calculated to determine water conductivity. Xylem conductivities per bundle in leaves were well scaled with xylem conductivities per bundle in stems among the species tested [y = 0.052x − 0.004, r2 = 0.80] (Fig. 5). Xylem conductivities in stem bundles ranged from 0.0008 g cm MPa−1 s−1 to 0.208 g cm MPa−1 s−1 . Xylem conductivities in leaf bundles ranged from 0.00015 g cm MPa−1 s−1 to 0.131 g cm MPa−1 s−1 . Overall stem tissue conductivities were compared to overall leaf tissue conductivities. Xylem conductivities of overall leaf tissues were well scaled with xylem conductivities of overall stem tissues for the 27 species of this test [y = 0.060x + 0.036, r2 = 0.89] (Figs. 6 and 7). Xylem conductivities in leaves ranged from 0.00129 g cm MPa−1 s−1 to 5.88 g cm MPa−1 s−1 . Xylem conductivities in stems ranged from 0.015 g cm MPa−1 s−1 to 92.4 g cm MPa−1 s−1 .

Ortega

The Manhattan Scientist, Series B, Volume 4 (2017)

Figure 5. Leaf conductivity per bundle vs. stem conductivity per bundle. Circles, C3 species; triangles, C4 species.

Figure 6. Leaf conductivity vs. stem conductivity. Circles, C3 species; triangles, C4 species.

Figure 7. Leaf conductivity vs stem conductivity (outliers excluded). Circles, C3 species; triangles, C4 species.

Discussion Grasses provide humans with grains for consumption worldwide due to their availability in most parts of the world. Grasses are found in a large variety of habitats that have sufficient water availability that meets the grassesâ&#x20AC;&#x2122; demands for water. Grasses are found in various parts of the world; they receive different amounts of sunlight which correspond to the type of carbon fixation used to produce food used by the plants. When examined, vascular bundle and xylem characteristics between C3 and C4 grasses indicated that there were no differences in how water is transported in C3 and C4 grasses. In other words, C3 and C4 grasses conduct water in the same way despite which habitat they are situated in or which habitat they belong to. There were strong relationships between stem and xylem characteristics and xylem conductivities. Grasses with larger stem diameters tended to have a larger amount of vascular bundles

The Manhattan Scientist, Series B, Volume 4 (2017)

Ortega

present in those stems. When looking at the distribution of vascular bundles in stems and leaves, the number of bundles in leaves was equal to about 10% of that found in stems. Bundles play a large role in xylem conductivity since they conduct water throughout plants. Our data suggests that xylem conductivity per bundle in leaves would be equivalent to about 5% of the xylem conductivity per bundle in stems. In all species, stem bundle conductivity was larger than the conductivity found in leaf bundles. In all species examined, conductivity in stem tissues were generally larger than conductivity found in leaf blade tissues. The overall stem conductivity of leaf blades was 6% of the xylem conductivity of stems. Stems must supply water to many leaves as well as reproductive structures like seeds and flowers. Certain species had extremely large conductivity values compared to the majority of the grass, so when these outliers were taken out, the same trend was observed. In grasses with relatively low conductivity values, overall stem conductivity of leaf blades was 6% of the xylem conductivity of the stems, suggesting that this value is consistent throughout grasses regardless of plant size. Six percent of water is being transported into leaves. These data suggest that most water is retained within stems to provide water for additional leaves and reproductive structures. The results of this study show that the amount of water retention in stems and contribution of water to each leaf is consistent among all 27 species irrespective of habitat.

Acknowledgment This work was funded by the Catherine and Robert Fenton Endowed Chair to Dr. Lance Evans.

References Cope, T., Gray, A. 2009. Grasses of the British Isles B.S.B.I. Handbook No. 13. Botanical Society of the British Isles. London. Hitchcock, A. S. 1950. Manual of the Grasses of the United States. United States Government Printing Office, Washington D.C. IPNI: Plant Name Search. IPNI: Plant Name Search. International Plant Names Index, 23 Apr. 2015. Web. 22 Aug. 2017. http://www.ipni.org/ipni/plantnamesearchpage.do. Jensen, W. A. 1962. Botanical Histochemistry. W.H. Freeman, San Francisco, CA. 408 p. McCulloh, K. A., J. S. Sperry, and F. R. Adler. 2003. Water transport in plants obeys Murrayâ&#x20AC;&#x2122;s law. Nature 421: 939-942. National Institutes of Health. Image J. https://imagej.nih.gov/ij/ Ortega, H. 2016. Xylem conductivity from stems to leaves of grass plants. Manhattan Scientist. Series B, 3: 77-84. Tropicos.org. Missouri Botanical Garden. http://www.tropicos.org Wagner, W. L., Herbst D. R., and Sohmer S. H. 1999. Manual of the Flowering Plants of Hawaiâ&#x20AC;&#x2122;i. Bishop Museum Press, Honolulu, HI.

Quantification of eccentric growth in stems of Artemisia tridentata ˜ ∗ Ismael Pena Laboratory of Plant Morphogenesis, Department of Biology, Manhattan College Abstract. Artemisia tridentata, Big Sagebrush, is a shrub that dominates the western United States. A. tridentata exhibits eccentric growth along its stem due to the death of vascular cambium affecting future stem growth. In the study concerning rates of eccentricity between subspecies, stems of A. tridentata spp. Wyomingensis and A. tridentata spp. tridentata were analyzed and compared. All stem samples were of a similar diameter, 8 mm, which allowed for comparable results. Samples from both subspecies exhibited eccentric growth, leading to 95% of all samples showing eccentric growth. In the study concerning quantification of eccentricity along stems, three stems of A.tridentata spp. wyomingensis were examined. The stems grew over a period of 27 to 32 years, with distances of 33 to 50cm. Stems which had more than 10 growth rings had eccentricity present. The results support the idea that death of vascular cambium is (1) localized, (2) not maintained well from segment to adjacent segment, and (3) ubiquitous in stems of Artemisia. To our knowledge, this is the first study to focus on the quantification of eccentricity along stems for A. tridentata.

Introduction Sagebrush is a dominant plant species found in the Western United States, especially in the Great Basin Desert. The Great Basin Desert is a climatic culmination of desert areas, having an annual precipitation of upwards 175mm annually (Cronquist et al., 1972). This sagebrush community can cover vast areas, more specifically about 60 million hectares, and is especially dominant at elevation above 1370m in the north locations and 1700m in the south location. The extent of the community once reached contained more area than any other community in the region (Hironaka and Tisdale, 1963; Daubenmire, 1970; Doescher et al., 1990; Bilbrough and Richards, 1991; West, 1999; Welsh, 2005; Soulard, 2012). The Great Basin Desert has been a well-studied and analyzed region (Fowler and Koch, 1982; MacMahon, 1985). Secondary xylem (wood) is a prominent structure found in perennial land plants (Schweingruber et al., 2006). Secondary xylem gives the plant support and stability allowing for it to grow tall, eventually leading to the support of secondary branches and finally the leaves used in photosynthesis. Secondary xylem production has been thoroughly studied in numerous plants by Schweingruber et al. (2006), who also discuss eccentric characteristics of reaction woods (tension and compression woods) in response to gravity. Stems of Artemisia tridentata and of a variety of its subspecies exhibit unusual growth patterns. They display twisting and turning, as well as a specific type of eccentric growth (Diettert, 1938; Moss, 1940; Ferguson and Humphrey, 1959; Ferguson, 1964; Miller and Shultz, 1987). Big sagebrush (Artemisia tridentata spp. tridentata) is described by Ferguson (1964)when he analyzes the ages and growth status of plants at various locations in the Western United States. The pres∗

Research mentored by Lance Evans, Ph.D.

The Manhattan Scientist, Series B, Volume 4 (2017)

PeË&#x153;na

ence of interxylary cork and eccentric growth in Artemisia stems led Wang (2004) to postulate that Artemisia may have descended from ancestors which were not woody in nature. Artemisia tridentata, and other subspecies of Artemisia, produce eccentric growth in their stem because of the death of vascular cambium, which has been shown by recent evidence (Evans et al., 2012). The eccentric growth occurs naturally in the main stem at nodal areas where flowering branches grow. When eccentric growth occurs and the vascular cambium dies the damage done to the stem is irreversible, leading to no production of secondary xylem in that certain area. This death of the vascular cambium can be seen by the absence of xylem and phloem cells in areas where annual rings should be present. In our study we try to identify (1) the twisting nature of stem samples and (2) describe eccentric growth both along stem samples and between two subspecies. The focus of this study was to quantify the characteristics of annual rings in stem segment along 0.5 m stem samples from three individual stems of Artemisia tridentata spp. wyomingensis. The study also compared the eccentricity values measured between samples of Artemisia tridentata spp. tridentata and Artemisia tridentata spp. wyomingensis in samples of around 9-10 mm in length.

Figure 1. Image of side by side comparison of a sample of Artemisia tridentata spp. wyomingensis (left) and Artemisia tridentata spp. tridentata (right). Straight black lines with numbers represent the angle at that position from the origin angle (0o ). Black lines on the rings each represent and outline the rings on the samples.

Methods and Materials Samples to quantify eccentric growth of two subspecies To quantify the eccentricity in stems of Artemisia tridentata, individual stem segments from 30 plants of Artemisia tridentata spp. tridentata were sent from near Thistle, Utah (40.0o N, 111.5â&#x2014;Ś W). Individual stem segments from 30 plants of Artemisia tridentata spp. wyomingensis were sent from near Milford, Utah (38.4204o N, -113.0937o W). Stem segments were cut, packaged, and shipped to Manhattan College. Once received, the samples were chosen based on their average diameters. Average diameters were determined by the use of an electronic caliper. Similar stem diameter segments were chosen, around 8mm. One cross sectional segment was chosen from each

PeË&#x153;na

The Manhattan Scientist, Series B, Volume 4 (2017)

Figure 2. Three images of stem segment 34 from stem #1 of Artemisia tridentata spp. wyomingensis of this study. Image A shows the original picture taken right after sawing, no alterations. Image B shows angle lines drawn in and numbered at 36 degree intervals, creating ten sectors for the segment. Image C shows the same image as image B with the annual rings highlighted. The distances between the scale lines in each figure is 1.6 mm. For simplicity, sectors 0 to 36, 36 to 72, 72 to 108, 108 to 144, 144 to 180, 180 to 216, 216 to 252, 252 to 288, 288 to 324, and 324 to 360 degrees were referred to as sectors #1 through #10 respectively

stem sample for analysis. The stem segments were cut with a saw (Stanley Fatmax Flush Cut Pull Saw; www.Stanleytools.com). A black line was also drawn on the outside bark from the top to the bottom, to designate a zero-degree mark. Sample processing to determine eccentricity along stems To determine the eccentricity along stems, we chose stems of Artemisia tridentata spp. wyomingensis (Beetle and Young). These samples were taken from plants near Milford, Utah (38.4204o N, -113.0937o W) in June 2014 and sent to Manhattan College. The stems chosen were characteristic of the plants in this region. Three stems were chosen, each ranging from 334 to 550 mm in length, from three separate plants. These stems were chosen for their similarity in diameters, with the smallest values being 4 mm and the largest being 27 mm. The bark around the stem was then removed to allow a black straight line to be drawn vertically along the stem. This line was used to determine the orientation of the stem segments once the stem was sawed into individual pieces. A total of 62 segments were obtained from stem #1, 37 from stem #2, and 44 from #3, The stem segments were cut using a saw (Dewalt DWGT20541 flush cut pull saw (www.dewalt.com). The segments were also measured using a digital caliper which is accurate to 0.01 mm. Processing stem segments Samples were viewed under a Leica EZ4 microscope (www.leicaâ&#x20AC;&#x201C;microsystems.com). A small drop of water was added to most of the samples for a better view of growth rings. Multiple picturesof the cross section of the sample were taken with a Canon PowerShot ELPH100HS camera (www.canon.com). Each image was then uploaded and categorized according to a stem sample number. All of the measurements taken from the images were done using a computer

The Manhattan Scientist, Series B, Volume 4 (2017)

Pe˜na

program called ImageJ (imagej.nih.gov/ij). Once uploaded to ImageJ, the angle tool was used to measure and highlight angles each 36 degrees apart. The image was then saved and lines were drawn in to create 10 areas using Microsoft Paint Program (www.microsoft.com). The degree number was then written on a specific line. A sector was then classified by which two degree lines it was between, i.e. sector 0 − 36◦ , 36 − 72◦ , etc. The image was then uploaded once again to ImageJ to measure the areas of each sector. Once the areas were measured they were entered into a Microsoft Excel spreadsheet to calculate averages and eccentricity. Samples were considered to be eccentric if standard deviations divided by means of measurements were greater than 25.

Results Variability of eccentric growth for stem segments of two subspecies To understand the variability of eccentric growth between Artemisia tridentata spp. tridentata and Artemisia tridentata spp. wyomingensis, multiple samples of each were analyzed (Table 1). The mean diameter between the two species is very similar, with A. tridentata spp. tridentata having a mean diameter of 8.3 mm2 and A. tridentata spp. wyomingensis having a mean diameter of 8.4 mm2 , wyomingensis having an average of 14 rings per sample, and spp. tridentata having 5.1. When looking at the eccentricity we can see differences in the degree of eccentricity, while both still have considerable eccentricity values present. When looking at A. tridentata spp. wyomingensis we can see that the mean eccentricity value was 110% with a range from 16.8 to 184. A tridentata spp. tridentata shows a mean eccentricity value of 50.7% with a range from 14.5 to 131. The standard deviations for each measurement are shown in parentheses. Overall, the data show that more than 95% of all samples showed eccentric growth. Table 1. Stem characteristics of Artemisia tridentata spp. wyomingensis and Artemisia tridentata spp. tridentata. Standard deviations are in parentheses. Number of segment samples

Mean diameter (mm)

Mean number of rings

Artemisia tridentata spp. wyomingensis

8.35 (2.4)

13.9 (4.0)

111 (50.8)

184

16.8

Artemisia tridentata spp. tridentata

8.29 (0.73)

5.1 (2.4)

50.7 (28.6)

131

14.5

Subspecies

Eccentricity ———————————Mean Largest Smallest

Quantification of eccentricity along stems of A. tridentata spp. wyomingensis The purpose of this study was to quantify the characteristics of annual rings within individual stem segments along stems from three stems (0.3 to 0.5 m) of Artemisia tridentata spp. wyomingensis. The amount of xylem rings in all sectors for some stem segments for the three stems are shown in the Appendix. In all of the three stems the smallest segments towards the tip, which

Pe˜na

The Manhattan Scientist, Series B, Volume 4 (2017)

had a 4 mm diameter, had six or eight annual rings while the largest segments, which had a 27 mm diameter, had between 35 and 38 annual rings (Table 2). Along each stem we then see that the mean age difference between the smallest and largest segments was thirty years. What we see from these results is that overall the three stem samples showed similar data (Table 2). Table 2. Comparisons of stem areas, and largest number of rings for tip-most and base-most segments and total branch lengths for the three branch samples of Artemisia tridentata spp. wyomingensis of this study. For each species, the tip-most and base-most stem segments had stem diameters of approximately 4 and 27 mm, respectively. Stem sample number

Total branch length (mm) Number of stem sections Tip-most sample Stem area (mm2 ) Number of rings Base-most sample Stem area (mm2 ) Number of rings

550 62

334 37

520 44

9.6 8

3.4 8

6.4 6

760 35

603 38

727 38

Diversity of ring numbers-illustrated example In each stem of the three stems we see that all of the segments with ten or more rings showed eccentric growth (Appendix: Tables 4 - 6). This eccentricity is shown by the differences in ring numbers and sector areas within sectors for stem segments (Fig. 3, Appendix: Table 4). An increase in sector area going up the stem showed for more ring growth (Table 3). If we look at stem #1, sector #7 (216◦ to 252◦ ) segment 34 - 36 (young to old) the number of rings counted were 16, 9, and 20, illustrated in Figs. 3C - 3A. Comparing it to sector #9 (288◦ to 324◦ ) for the same segments the number of rings were 9, 16, and 10. We see that older sections mostly show eccentric growth. Moving down the stem towards the base, for sectors #3 (72◦ to 108◦ ) the number of rings counter were 28, 7 and 30 for segments 56, 57, and 58 (young to old), illustrated in Figs. 3F 3D. The drastic change in rings in some sectors (i.e. sector 2 - 36 to 72◦ ) from segment 56 to segment 57 can be seen as being caused by the death of vascular cambium after the seventh year in sector #3 (72◦ to 108◦ ). The growth seemed to be most exacerbated in segment 56 for sector #3. We see from viewing the above ring numbers that the number of rings might not be the same in a particular sector for two adjacent stem segment. Among all stems, all segment showed eccentricity when more than ten annual rings were present (Appendix). Collectively, for all stems, 17 (28%) segments out of all segments contained only one annual ring in one sector, and 43 (66%) had fewer than five rings in one sector (Fig. 4). We can see that death of the vascular cambium is most present in younger stems (less than five rings), but the occurrence of eccentric growth in later years shows it is not confined to specific years.

The Manhattan Scientist, Series B, Volume 4 (2017)

PeË&#x153;na

Table 3. Maximum number of annual rings per segment as a function of stem area for three stem samples of Artemisia tridentata spp. wyomingensis of this study. Samples of stems were taken from 110 to 285 mm2 . For each stem sample, a line was drawn from point to point to show the patterns. Slope values for the three stems ranged from 0.021 to 0.025. Stem segment area (mm2 ) 100 110 130 140 150 160 170 180 185 189 190 200 205 210 215 220 230 270

Stem 1

21 22

23 23 24

21 20 20 19

25 26 27 22 20 20

27 27 28

23 25 24

Diversity of ring numbers-entire stems Further analyses were done to determine the difference in the numbers of rings for sectors of successive segments. For sectors that showed a change of 0 to 2 rings, 3 to 10 rings, and more than 10 rings between successive segments, stem #1 showed percentages of 42.9, 42.4 and 14.7% and stem #2 showed percentages of 61.9, 25.9, and 12.2%. Additional analyses were done to determine if low numbers of rings were maintained from segment to segment. If this was the case, the results would suggest that the effect of eccentricity is not localized. A low number of rings from segment to segment rarely occurred (25 times out of 620 sector-segments for stem #1). These data would then suggest that effects of eccentricity were localized and is not perpetuated along the stems of the samples. Diversity of ring areas In addition to the amount of rings present, the use of measuring segment areas among the sectors and segments can be used to analyze the eccentricity of each stem. From the results we can see that segment and sector areas differed considerably among segments (Appendix). Data from sector #3 (72â&#x2014;Ś -108â&#x2014;Ś ) show that area enlargement of rings were relatively uniform from segment 38 to 39 (Fig. 5). Data from segment 39 to 40, in contrast, shows rings enlarged

Pe˜na

The Manhattan Scientist, Series B, Volume 4 (2017)

Figure 3. Examples of eccentric growth among stem segments from stem #1 of Artemisia tridentata spp. wyomingensis of this study. The black mark (0o ) was rotated to be in the upper, left-hand area for each image. All successive segments were aligned using the locations of the biological center and the sector lines every 36o . Thus, with these points of reference, each image was rotated and aligned as necessary to compare changes among successive segments. Images A through C are segments 36, 35, and 34, respectively. The distances between scale lines in images A through C is 1.6 mm. Images D through F are segments 58, 57, and 56, respectively. The distances between scale lines in images D through F is 1 mm. The distances between consecutive segments are 7.8 mm. In all cases the black mark is in the upper portion of the image so the orientation among segments is maintained. For simplicity, sectors 0 to 36, 36 to 72, 72 to 108, 108 to 144, 144 to 180, 180 to 216, 216 to 252, 252 to 288, 288 to 324, and 324 to 360 degrees were referred to as sectors #1 through #10 respectively.

markedly, especially when looking above ring 12. This non-uniform pattern in different in annual areas among segments was also seen amongst sectors within segments. Data of three adjacent sectors of segment 40 in stem #1 were compared and analyzed (Fig. 6). Data showed sector #5 (144◦ - 180◦ ) contained only nine, small rings with sector #4 (108◦ - 144◦ ) being similar with 17 rings that had relatively small annual rings. Diversity is seen in sector #3 (72◦ - 108◦ ) which had 23 large rings.

Discussion Various desert plants can be seen to have unusual stem characteristics (Jones and Lord, 1982). Unusual stem characteristics include things such as Axis splitting (Ginsburg, 1963). The splitting appears in the axils of branches of the smaller shrubs and then move outwardly. This pattern continues eventually creating two stems from the one main stem (Jones and Lord, 1982). This

The Manhattan Scientist, Series B, Volume 4 (2017)

Pe˜na

Figure 4. Largest (diamonds) and smallest (circles) number of rings for each stem segment for stem #1 of Artemisia tridentata spp. wyomingensis. The equation of the line for the largest number of rings was y = .81x0.40 with an r2 of 0.78 while the equation of the line for the smallest number of rings was y = 0.0007x + 4.29 with an r2 of 0.041.

Number of Rings Segment‐1

25 20 15 10 5 0 0

100

150

200

250

300

Stem Segment Area (mm2)

Figure 5. Comparison of sector areas of annual rings for sector 72-108 for stem segments 33 (circles), 32 (squares) and 31 (diamonds) of stem #1 of Artemisia tridentata spp. wyomingensis.

Sector Areas (mm2)

6 5 4 3 2 1 0 0

10 15 Ring Number

splitting is not shown in in the production of new stems of Artemisia species of our samples. In Artemisia tridentata we see that there is an eccentric growth pattern present in the stems of various subspecies such as A. tridentata spp. tridentata, A. tridentata spp. wyomingensis, A. nova, A. filifolia, A. bigelovii, A. tripartita (Evans et al., 2012). Overall showing us that the eccentric growth patterns described occur in various species of Artemisia. The pattern of eccentric growth in A. tridentata is unique due to various factors such as the localization of vascular cambium death, which does not affect adjacent vascular cambium. The adjacent vascular cambium can continue its growth and function to produce annual rings which can spread out to partially fill in areas left vacant by the previous death of vascular cambium. This growth of annual rings can lead to a disproportionate xylem growth pattern (Diettert, 1938; Fig. 3 of this study). The eccentricity pattern of growth for Artemisia is not the same as reaction woods (tension and compression woods) in plants, which are not a result of the death of the vascular cambium. Reaction wood is instead a product of differential cambium productivity which is dependent on external stressors such as gravity (Mauseth, 1998). Overall showing how eccentric

Pe˜na

The Manhattan Scientist, Series B, Volume 4 (2017)

Sector Areas (mm2)

7 6 5 4

Figure 6. Comparison of sector areas of annual rings for sectors #3 [72-108o ] (diamonds), #2 [36-72o ] (squares), and #3 [0-36o ] (circles) for stem segment 27 of stem #1 of Artemisia tridentata spp. wyomingensis.

3 2 1 0 0

10 Ring Number

growth pattern is not similar to reaction wood, and is unique to Artemisia (Diettert, 1938; Ferguson and Humphrey, 1959; Ferguson, 1964) The quantification of eccentric growth was one by subdividing the stem segments of the three stems into ten, 36◦ radial sectors. The criteria for determining eccentricity in stems was if the standard deviation of the number of rings or segment areas were 25% or more of the mean value for the ten sectors of each segment, then the sample was considered eccentric (Evans et al., 2012). This criterion is similar to that used by Love et al. (2009) in the stems of Populus. This criterion was important for our study because it allowed for comparisons of sectors in each segment and among sectors of successive segment along a stem, as well as when comparing samples between subspecies of Artemisia tridentata. Data from the comparison of Artemisia tridentata spp. tridentata and A. tridentata spp. wyomingensis suggest that levels of eccentricity between subspecies can be quite different but still both show high amounts of eccentricity present between subspecies. The purpose of this study was to see if there will be a considerable amount of eccentricity in both subspecies of sagebrush. The focus of the study was on taking multiple samples from each subspecies to look into the nature and degree of eccentricity in each. The samples between the subspecies were of similar diameter. 8 mm, and the number of rings was greater in A. tridentata spp. wyomingensis. Data shows that eccentricity values in A. tridentata spp. wyomingensis have a higher mean value, 111%, than A. tridentata spp. tridentata, 50.7% (Table 1). The standard deviation of the mean of A.tridentata spp. wyomingensis was 50.8 while the standard deviation of the mean of A. tridentata spp. tridentata was 28.6. Overall, it can be concluded from our results that for morethan 40 stem samples of the two subspecies of sagebrush, more than 95% of all samples showed eccentric growth. From the data gathered from the three branches, all of the stem segments that had more than ten annual rings showed eccentric growth. The three stem samples used were of three separate plants which had a similar diameter and length in order to be considerable as replicate branches. The eccentric characteristics between branches were similar, such as their changes in number of rings among sectors and segments. Though this was the case, it is probable that two individual stems of Artemisia would not have exactly the same pattern of eccentric growth from one segment to

The Manhattan Scientist, Series B, Volume 4 (2017)

PeË&#x153;na

another due to the localization of death of the vascular cambium being random. To our knowledge, this is the first publication to quantify and document the frequency of eccentricity along a stem and its variety of manifestation which come as a result. As stated previously, an eccentric pattern of growth is widespread in species of Artemisia. The purpose of the study was to sample many small stem segments from each of the three stems, allowing the characterization of eccentricity along stems. In total, 143 stem segments were analyzed, each segment containing ten sectors. Thin segments of the stem were important to best determine changes from one wood segment to the following. Analysis of the change in each sector resulted in a more efficient method to determine these changes. The demarcations of the sectors were arbitrary but could still be tracked through each stem, allowing for changes to be accurately determined. To our knowledge, there has been no previous or subsequent studies quantifying segment to segment changes of the eccentric growth behavior in stems of Artemisia. In Artemisia tridentata, terminal stems have clearly shown to produce defined indeterminate and determinate branches during their current-year growth. Determinate branches will produce leaves and flowers, while indeterminate branches only produce leaves (Harris and Harris, 2009). The production of indeterminate and determinate branches produced per year is variable between individual plants Artemisia tridentata (Evans et al., 2012). In this study, each tissue segment with more than ten annual rings showed eccentricity. Approximately ten percent of all the eccentric growth started with a loss of vascular cambium towards the end of the first year and the end of the second. This meaning, eccentric growth also originated after the first/second year. Currently, no specific cause(s) are established for the death of this cambium after the first/second year but it is possible that this death is due to the localized freezing of liquid water of the vascular cambium as a result of the exfoliating bark of Artemsia plants (Provenza et al., 1987). Taking into consideration each of the ten sectors of every one of the 62 segment of stem #1, there was a decrease of more than ten rings occurring more than 13% among segments. A difference of 10 or more rings between adjacent segments is significant due to the oldest stem segment, base most segments, only having 35 rings. Additional data showed large decreases in the amount of rings in a sector were maintained from one segment to another in 25 cases among 620 sector-segments of this study. This results support the idea that death of vascular cambium is (1) localized, (2) not maintained well from segment to adjacent segment, and (3) ubiquitous in stems of Artemisia. In conclusion the data also showed that eccentric occurs in more than 90% of stem samples from the three stems. The collaboration of extensive centric stem growth, large number of flowering stems, and large number of vegetative branches on individual stems seem probable reasons to speculate that they contribute to the shrubby nature of Artemisia species. Though this is tempting a definite answer has yet to be concluded. Would the characteristics described above be a limitation to the overall vertical growth of sagebrush? Is there possibly different factors of nutrition which could lead to them growing taller?

PeË&#x153;na

The Manhattan Scientist, Series B, Volume 4 (2017)

Acknowledgment This work was funded by the Catherine and Robert Fenton Endowed Chair to Dr. Lance Evans.

References Bilbrough, C. J., and R. H. Richards. 1991. Branch architecture of sagebrush and bitterbrush: Useof a branch complex to describe and compare patterns of growth. Can. J. Bot. 69:1288-1295. Cronquist, A., A. H. Holmgren, N. H. Holmgren, and J. L. Reveal. 1972. Intermountain Flora, Volume 1: Geological and Botanical History of the Region, its Plant Geography and a Glossary. The Vascular Cryptogams and the Gymnosperms. The New York Botanical Garden, New York. Daubenmire, R. 1970. Steppe vegetation of Washington. Technical Bulletin 62. Washington State Agricultural Station, College of Agriculture. Washington State University. Pullman, WA. Diettert, R. A. 1938. The morphology of Artemisia tridentata Nutt. Lloydia 1:3-14. Doescher, P. S., R. Miller, J. Wang, and J. Rose. 1990. Effects of nitrogen availability on growth and photosynthesis on Artemisia tridentata ssp. wyomingensis. Great Basin Naturalist 50: 9-19. Evans, L. S., A. Citta, and S. Sanderson. 2012. Flowering branches cause injuries to secondyear main stems of Artemisia tridentata Nutt. subspecies tridentata. Western North American Naturalist 72: 447-456. Ferguson, C. W. and R. R. Humphrey. 1959. Growth rings of sagebrush reveal rainfall records. Progressive Agriculture in Arizona 11:3. Fowler, D. and D. Hoch. 1982. The Great Basin. Pages 7 to 63 in Reference Handbook of the deserts of North America. Ed., G.L. Bender. Greenwood Press Westport, CT. Ginsburg, C. 1963. Some anatomic features of splitting of desert shrubs. Phytomorphology 13: 92-97. Harris, J. G., and M. W. Harris. 2009. Plant identification terminology. 2nd edition. Spring Lake Publishing, Spring Lake, UT. Hironaka, M. and J. Tisdale. 1963. Secondary succession in annual vegetation in southern Idaho. Ecology 44: 810-812. Jones, C. S., and E. M. Lord. 1982. The development of split axes in Ambrosia dumosa (Gray) Payne (Asteraceae). Bot. Gaz. 143: 446-453. Love, J. S., J. Borklund, M. Vahala, J. Hertzberg, J. Kangasjarvi and B. Sundberg. 2009. Ethylene is an endogenous stimulator of cell division in Populus. Proceedings of The National Academy of Science. 106: 5984-5989. MacMahon, J. A. 1985. Deserts. The Audubon Society Nature Guides. Alfred A. Knopf. New York, 638 p. Mauseth, J. D. 1998. Botany: an introduction to plant biology. 2nd edition. Jones and Bartlett Publishers. Sudbury, MA Miller, R. F. and L. M. Shultz. 1987. Development and longevity of ephemeral and perennial leaves on Artemisia tridentata Nutt. ssp. wyomingensis. Great Basin Naturalist 47: 227-230.

The Manhattan Scientist, Series B, Volume 4 (2017)

Pe˜na

Moss, E. H. 1940. Interxylary cork in Artemisia with a reference to its taxonomic significance. American Journal of Botany 27: 762-768. Provenza, F. D., J. T. Flinders and E. D. McArthur. 1987. Proceedings: Symposium on Plant Herbivore Interactions. Snowbird, Utah. August 1985. Intermountain Research Station, Forest Service, U.S. Dept. of Agriculture, 1987 – Biotic communities - 179 pages Schweingruber, F. H., A. Borner, and E.-D. Schulze. 2006. Atlas of Woody Plant Stems: Evolution, Structure, and Environmental Modifications. Springer-Verlag, Berlin. Soulard, C. E. 2012. Central Basin and Range Ecoregion 2012. in Status and Trends in the Western United States – 1973 to 2000 (Edited by B. M. Sleeter, T. S. Wilson and W. Acevecb) US Geological Survey Professional Paper 1794-A. Wang, W. M. 2004. On the origin and development of Artemisia (Asteraceae) in the geological Botanical Journal of the Linnean Society. 145: 331-336. Welsh, B. 2005. Big sagebrush: a sea fragmented into lakes, ponds, and puddles. General Technical Report RMRS-GTR-144. Fort Collins, CO. West, N. E. 1999. Managing for biodiversity of rangelands. Pages 101-126 in W. W. Collins and C. O. Qualset, editors. Biodiversity in agroecosystems. CRC Press, Boca Raton, FL.

PeË&#x153;na

The Manhattan Scientist, Series B, Volume 4 (2017)

Appendix Table 4. Number of xylem rings for the 10 sectors for the 62 segments of Stem #1 of Artemisia tridentata ssp. wyomingensis. Segment 1 is the smallest diameter segment and segment 62 is the largest diameter segment sampled. Sector (degrees) Segment

0 -36

36 -72

72 -108

108 -144

144 -180

180 -216

216 -252

252 -288

288 -324

324 -360

Eccentricity1 (%)

1 3 6 9 12 15 18 21 24 27 30 33 36 39 42 45 48 51 54 57 60 63

8 2 4 9 10 3 21 19 16 18 20 13 18 20 20 26 25 28 30 35 33 34

8 2 8 9 10 14 16 9 16 19 20 19 18 20 20 7 25 16 30 24 33 21

5 3 8 9 9 8 7 3 4 2 20 18 6 20 6 4 16 5 8 28 28 4

5 6 8 9 11 1 6 4 2 17 15 2 14 20 1 2 1 3 16 7 13 1

5 6 8 6 1 8 6 6 8 10 11 7 13 20 1 3 1 4 1 7 10 1

6 6 7 6 9 5 6 1 8 11 11 7 20 16 1 4 1 6 1 14 10 1

8 6 8 5 10 4 5 1 8 8 8 7 9 14 1 4 1 16 1 17 15 8

8 3 8 5 10 3 6 1 7 13 6 13 7 11 1 5 1 24 1 13 12 7

8 4 8 9 10 3 10 1 14 2 11 17 16 5 1 13 1 24 1 13 24 16

8 4 8 9 10 3 21 1 16 5 6 17 3 5 2 26 1 28 7 35 33 12

21 40 17 4 32 73 62 124 53 60 44 49 37 41 145 99 143 67 108 56 48 101

If the percentage is 25 or above, the segment is considered eccentric. If the percentage is less than 25, the segment is considered not to be eccentric.

Table 5. Number of xylem rings for the 10 sectors for the 37 segments of Stem #2 of Artemisia tridentata ssp. wyomingensis. Only odd numbered segments are shown. Segment 1 is the smallest diameter and segment 37 is the largest diameter segment sampled. Sector (degrees) Segment

0 -36

36 -72

72 -108

108 -144

144 -180

180 -216

216 -252

252 -288

288 -324

1 3 5 7 9 11

8 13 11 13 13 16

6 13 12 13 5 8

4 7 7 2 2 5

3 3 3 2 2 4

4 2 1 2 2 3

5 2 1 2 2 3

5 3 1 2 3 3

8 6 1 2 4 4

8 3 1 4 8 8

324 Eccentricity1 -360 (%) 8 3 9 12 8 16

30 58 61 67 58 65

The Manhattan Scientist, Series B, Volume 4 (2017)

PeË&#x153;na

Table 5. (continued) 13 15 17 19 21 23 25 27 29 31 33 35 37

21 19 4 24 12 26 24 25 27 34 37 38 1

21 14 19 22 26 23 22 1 4 22 24 27 7

3 1 9 1 1 1 5 1 11 2 11 12 20

2 1 3 1 1 1 1 1 2 1 3 2 30

2 1 2 1 1 1 1 1 2 1 1 3 2

2 1 2 1 1 1 1 1 2 1 1 3 1

3 1 2 1 1 1 1 1 2 1 1 4 1

3 1 2 1 1 1 1 5 2 1 2 4 1

10 1 3 1 1 1 1 11 17 1 11 2 3

15 19 7 6 5 7 21 1 31 11 32 2 4

101 102 68 114 98 113 113 87 123 142 142 142 140

If the percentage is 25 or above, the segment is considered eccentric. If the percentage is less than 25, the segment is considered not to be eccentric.

Table 6. Number of xylem rings for the 10 sectors for segments of Stem #3 of Artemisia tridentata ssp. wyomingensis. Segment 1 was the smallest diameter segment sampled and segment 44 was the largest diameter segment sampled. Not all segments are shown. Sector (degrees) Segment

0 -36

36 -72

72 -108

108 -144

144 -180

180 -216

216 -252

252 -288

288 -324

324 -360

Eccentricity1 (%)

1 5 9 13 17 21 25 29 33 37 41 44

6 5 2 5 16 21 22 1 25 28 34 38

6 6 2 18 20 21 22 37 17 30 1 4

6 6 2 15 5 2 13 30 16 2 1 4

6 6 2 6 2 2 9 6 1 3 1 5

6 6 3 4 2 2 6 5 1 3 1 1

3 6 4 4 3 6 6 1 1 3 1 1

6 6 4 4 3 2 6 1 1 3 1 1

6 5 2 2 3 2 8 1 1 6 1 25

6 5 16 2 8 3 8 1 21 10 1 38

6 5 17 3 8 12 22 11 25 28 21 38

15 8.7 103.8 83.7 85 102.3 55 52 94.1 98.2 174.5 104.3

If the percentage is 25 or above, the segment is considered eccentric. If the percentage is less than 25, the segment is considered not to be eccentric.

Growth dynamics of Artemisia tridentata Claudia S. Ramirezâ&#x2C6;&#x2014; Department of Biology, Manhattan College Abstract. Stems of Artemisia tridentata Nutt. ssp. tridentata exhibit a cyclic growth pattern that starts in June and ends in November every year. Such plants display a period of linear stem elongation, followed by branch development and extension, which increases over time. Typical growth rates of branches (vegetative) were less than two mm per day. Around the 260th day of year, sagebrush branches shifted from vegetative to flowering branches. During the vegetative state of the stem, the number of branches per stem increased linearly. There is an initial increase in stem length but over time, it ceases to grow as it begins to flower. At this time, the stem is no longer lengthening but the diameter for each terminal stem increases 0.017 per day with an r2 value of 0.82. This occurs because once the stem becomes flowering, branch length increases markedly during reproductive growth, leading to the production of hundreds of seeds per branch. For current-year growth, the number of branches per stem increase linearly during the vegetative portion of the growth cycle. By the end of the growth period, about thirty branches with an average length of 45 mm were produced.

Introduction Artemisia tridentata ssp. tridentata, Big Sagebrush, is a tall, native shrub. Individual branches are usually short and most stems have woody trunks (MacMahon, 1992: USDA plant guide). Depending on soil conditions, the various subspecies of Artemisia tridentata may grow to be up to four meters tall (Fig. 1). Prior to immigration of Anglo-Americans, species of sagebrush occupied most non-saline portions of Montana, Wyoming, Colorado, Utah, Idaho, and Nevada below 3000 m elevation. In addition, sagebrush subspecies occupy eastern Washington, eastern Oregon, throughout California, and into Baja California and Mexico above 1000 m elevation (total area = 1.5 million km2 ; Daubenmire, 1970; Welch, 2005). Within these areas, Artemisia tridentata is well suited to the environments of these regions and is the dominant species in undisturbed areas (Welch, 2005). Of course, these areas have become agricultural and are currently used for other practices (MacMahon, 1992; Welch, 2005). Sagebrush plants may live for hundreds of years and produce 10 to 40 terminal shoots each year. Each stem terminal may produce 20 to 30 flowering branches every year (Fig. 2; Evans et al., 2012). Early in the growing season, many indeterminate, vegetative branches are produced and many become determinate, flowering branches (Fig. 3) at the end of the growing season (Figs. 4 and 5; Evans et al., 2012). The purpose of this study is to focus on growth dynamics and patterns of terminal stems of Artemisia tridentata from June until November 2015 to understand changes in terminal stem growth and branches through the period. â&#x2C6;&#x2014;

Research mentored by Lance Evans, Ph.D.

The Manhattan Scientist, Series B, Volume 4 (2017)

Ramirez

Figure 1. Image of a sagebrush plant (Artemisia tridentata) in its vegetative state in the wild. Sagebrush plants grow as individual plants and not in a spread fashion in deserts.

Figure 2. Image of Sagebrush (Artemisia tridentata) with many flowering stem terminals late in the growing season.

Ramirez

The Manhattan Scientist, Series B, Volume 4 (2017)

Figure 3. Image of a flowering terminal stem from Artemisia tridentata from November 2, 2015 (DOY=306). Image shows the junction, the junction of previous yearâ&#x20AC;&#x2122;s growth and current yearâ&#x20AC;&#x2122;s growth (pencil closest to the bottom).

Figure 4. Image of one terminal stem sample of Artemisia tridentata. The vegetative stem sample only has leaves. The sample was harvested on July 9th , 2015 (DOY= 190). This sample was not dissected. Image shows the junction, the junction of previous year growth and current year growth (pencil closest to the bottom).

Figure 5. Image of a young vegetative terminal stem sample from Artemisia tridentata from June 4, 2015 (DOY 155). Branches have been removed from the stem, assigned a number and placed near its corresponding node. Previous yearâ&#x20AC;&#x2122;s growth has been removed from the sample.

It was hypothesized that for current-year growth: 1. Terminal stem lengths increase linearly during the vegetative portion of growth. 2. The number of branches per stem increase linearly during the vegetative portion of the growth cycle. 3. The number of branches per stem increase linearly during the vegetative portion of the growth cycle 4. Branch lengths increase markedly during vegetative growth

The Manhattan Scientist, Series B, Volume 4 (2017)

Ramirez

Focusing on data from both the branches and the main stem over time allowed us to propose a model for stem growth over the development period. The understanding of the growth patterns of Artemisia tridentata is very important since each branch produces hundreds of seeds, which play a large ecological role to the surrounding wildlife and inhabitants situated around sagebrush growth.

Materials and Methods The terminal stems analyzed in this study were shipped from Thistle, Utah (40.00◦ N, 111.49◦ W). Stem samples were randomly selected once a week from June to November 2015 and shipped to Manhattan College for processing. Each shipment consisted of six separate terminal stem samples. After receiving the boxes, the samples were organized by date to ensure correct analysis. From each mailing box, random terminal stems were selected for examination for each date. The junction of previous year’s growth from current year’s growth was determined for the samples. All branches from current year were removed. Each detached branch was laid near its corresponding node and marked with “V” for vegetative or “F” for flowering (Figs. 5 and 6). All branches were numbered starting at the junction. Images were taken to document the features of each stem

Figure 6. Image of a flowering terminal stem of Artemisia tridentata after dissection from September 17, 2015 (DOY 260). Branches have been removed from the stem, assigned a number and placed near its corresponding node. Most branches have elongated compared with samples from earlier in the season. Previous year’s growth has been removed from the sample.

sample. A ruler was placed in all pictures to ensure the scale was correct for subsequent measurements. Images were placed on ImageJ (National Institutes of Health, http://rsb.info.nih.gov/ij) for measurements of (1) stem lengths above junction, (2) stem diameter at junction, (3) the number of branches, and (4) branch lengths. Data were placed in Excel (Microsoft Inc.) for analysis.

Ramirez

The Manhattan Scientist, Series B, Volume 4 (2017)

Results For this study, ten branches of Artemisia tridentata spp. tridentata were processed to study growth and development of stems and branches from June through November 2015 (DOY 155 to 327). Fig. 4 shows a vegetative branch from early June before dissection. Fig. 5 shows the latter after dissection. From June through mid-August (DOY 155 to 219), stems elongated for the vegetative portion of the growth phase at a rate of 1.39 mm per day (Fig. 7). Stem elongation was sporadic and did not show a linear relationship after mid-August (DOY 219). Over the same period, stem diameters at junctions of 2014 growth to 2015 growth lengthened at a rate of 0.040 mm per day (Fig. 8). Thus, although stem elongation did not increase markedly after DOY 219, stem diameters increased throughout the entire period.

Figure 7. Relationship between stem lengths and day of year (DOY) for eight terminal stem samples of Artemisia tridentata from June 4 (DOY 155) through September 10, 2015 (DOY 253). Samples with only vegetative growth are shown (y = 1.39x â&#x2C6;&#x2019; 127.63; r2 = 0.7734).

Figure 8. Relationship between the stem diameters at the junction (base) of the current-year growth and day of year (DOY) for samples of Artemisia tridentata (y = 0.04x + 183.62; r2 = 0.0075).

Number of branches per plant for sagebrush plants increased linearly at a rate of 0.20 per day over the period from June to mid-August 2015. Thus, during the vegetative period stem lengths

The Manhattan Scientist, Series B, Volume 4 (2017)

Ramirez

increased and the number of branches increased linearly. Fig. 9 shows the relationship between the number of branches on terminal stems and day of year. There is a linear relationship through

Figure 9. Relationship between number of branches on terminal stems and DOY for current year growth samples of Artemisia tridentata. (y = 0.2005x − 14.102; r2 = 0.7704)

the vegetative period of the growth cycle up to day 240. As stated above, stem lengths increased and more branches were produced during the vegetative growth. However, individual branch lengths did not increase markedly (rates of 0.65 and 2.47 mm per day; Fig. 10). Fig. 3 shows a flowering branch from late October of the growth year. During this period, stem lengths did not increase, and individual number of branches did not increase. However, cumulative branch lengths increased rapidly at a rate of 6.3 to 9.56 mm per day (Fig. 10) during the flowering period.

Figure 10. Relationship between cumulative branch lengths for four terminal stem samples of Artemisia tridentata. Four samples were chosen to clearly distinguish the pattern seen. Vegetative branches are indicated color coded in green and flowering branches in orange. Samples include Jun-4 (y = 2.47x + 17.4; r2 = 0.98), Jul-9 (y = 0.65x − 2.32; r2 = 0.99), Nov-2 (y = 9.56x + 67.76; r2 = 0.97), and Nov-23 (y = 6.31x − 47.8; r2 = 0.99).

Ramirez

The Manhattan Scientist, Series B, Volume 4 (2017)

Discussion This paper documents the seasonal activity of sagebrush plants from early spring through late fall. Since sagebrush is a perennial plant, the terminal stems become dormant the previous winter and begin to grow in spring. During early spring, we expected that initial terminal stem growth would be rapid. The data collected shows how it increased linearly at a rate of 1.38 mm per day. Plant growth is concentrated on the main stem vertical growth. In addition, we predicted that these fast growing stems would produce a large number of side branches during the early part of the growing season (late May to July). All the energy and water is directed toward this enlargement until late July. After this initial rapid terminal stem growth, we predicted that the terminal stem growth would cease and that the side branches would elongate rapidly and eventually produce many flowers and seeds. The number of branches per stem during Artemisisa tridentata development increase linearly during vegetative growth (June through August). The data suggests that the growth pattern of Artemisisa tridentata takes a turn after late July from vegetative, vertical stem growth and number of branches to strictly branch growth. The branches begin extending around mid-August and become flowering around late-September. Data shows how there is a switch between vegetative, vertical growth to flowering and production of seeds. As the season progresses, the first branch of the new growth gains distance from the junction. Aside from stem elongation during the vegetative period, there may be possible branch death near the junction near the end of the growth period. This creates a gap between the beginning of the new growth and the first branch. Stem Diameters of sagebrush plants increase from June until November each year. Although the growth of the main stem ceases to develop late July, stem diameters continue to grow in order to nurture the plant during branch elongation successfully. More water and energy needs to be conducted in order to effectively progress into the flowering stage. Individual branch length increases markedly during reproductive growth. During this period, the stem is no longer growing and all the energy is concentrated toward branch growth and development of flowers and seeds. This occurs over the period of August through November. Overall the results show that terminal stems grew and produced a large number of branches during the vegetative period (June through August), while individual branches grew rapidly and produced hundreds of seeds per branch during the reproductive period as stem diameters continued to grow.

Acknowledgment This work was funded by the Catherine and Robert Fenton Endowed Chair to Dr. Lance Evans.

References Diettert, R. R. 1938. The morphology of Artemisia tridentata. Nutt. Lloydia 1:3-74

100

The Manhattan Scientist, Series B, Volume 4 (2017)

Ramirez

Evans, L. S., A. Citta, and S. C. Sanderson. 2012. Flowering branches cause injuries to secondyear main stems of Artemisia tridentata Nutt. Subspecies tridentata. Western North American Naturalist. 72 :447-456. National Institutes of Health. Image J. https://imagej.nih.gov/ij/ USDA plant guide, United States Department of Agriculture, Natural Resources Conservation Service, Plant Guide. http://plants.usda.gov/plantguide/pdf/pg artr2.pdf Welch, B. 2005. Big Sagebrush: a sea fragmented into lakes, ponds, and puddles. General Technical Report RMRS-GTR-144, Fort Collins, CO MacMahon, J. A. 1992. Deserts. New York: Knopf.

Effects of chemical exposure on tadpole behavior and personality Cassidy Stranzl∗ Department of Biology, Manhattan College Abstract. Individual animal personality is a relatively new field with very few studies on amphibian species. We studied whether a population of Bullfrog tadpoles had individual personality and if this individual behavioral consistency is affected by the widely-used insecticide carbaryl and the less common eco-friendly insecticide lemon grass oil. We measured the personality traits of activity and exploration by quantifying movements and proportion of time active from 15 minute video trials. Tadpoles were observed in their familiar “home” environment and then in a new “novel” environment on two consecutive days. We found that populations of tadpoles do exhibit evidence of personality and that carbaryl causes an increase in both activity and exploration behaviors. These findings could suggest a possible reason for recent amphibian species decline and also provide a tadpole-friendly lemon grass oil alternative to the widely-used insecticide carbaryl. To our knowledge, this work is the first to study the effect of chemical exposure on individual amphibian behavioral types.

Introduction The concept that individual animals have personality is a relatively new concept that has recently begun receiving attention from the scientific community. Individual animals differ in how they feed, cope with predation, interact with their environment, mate, and act overall. Personality is defined as any differences in intraspecies individual behavior that is consistent in varying situations and repeatable over time. Carlson and Langkilde (2013) suggest that there are five different personality axes that we can study: activity, boldness, exploration, aggression, and sociability. When an individual behaves in a certain way repeatedly, that behavior is considered to be a personality trait. A behavioral type is an individual’s average behavior, placing it somewhere along each of the five axes. For example, if a tadpole is repeatedly more active and exploratory in different situations, that tadpole is considered to be of the behavioral types highly active and highly exploratory. The behavior of activity is generally how much time the individual spends translocating relative to being inactive. The behavior of exploration is how much an individual is willing to explore a new, novel environment when taken from a familiar home environment. These behaviors have fitness consequences. For example, more exploratory individuals are more likely to disperse to a new metapopulation, while more active individuals may be more likely to encounter a predator (Conrad et al., 2011). The insecticide carbaryl’s mode of action is by the inhibition of synaptic enzyme receptors, leading to acetylcholine accumulation, and consequentially nervous system failure (EPA, 2012). In 2005 and 2007, it was found to be the third most commonly used insecticide and the second most frequently detected in water, being present in 50% of streams in urban settings (EPA, 2012). ∗

Research mentored by Gerardo Carfagno, Ph.D.

102

The Manhattan Scientist, Series B, Volume 4 (2017)

Stranzl

While laws regulate that concentrations of carbaryl present in water remain at sub-lethal doses (EPA, 2012), it has not been studied whether these sub-lethal levels affect how aquatic organisms behave. Lemon grass oil is another type of insecticide that is less commonly used and is naturally derived from the Lemongrass plant. This insecticide is considered to be more “eco-friendly” when compared to carbaryl. Amphibian species have recently shown significant population decline and higher extinction rates relative to other species groups (Stuart et al., 2004). Amphibians play vital roles in their ecosystems, so it is imperative to investigate the possible causes of their decline. We studied tadpole population behavior and whether or not tadpoles exposed to carbaryl levels similar to urban environmental levels had altered behavior when compared to lemon grass oil exposed tadpoles and control tadpoles. We hypothesized that tadpoles would exhibit evidence of individual personalities and that certain chemicals can have an effect on these individual personalities.

Materials and Methods The study consisted of 15 tadpoles: 5 control, 5 tadpoles that were exposed to carbaryl in a previous experiment, and 5 tadpoles exposed to lemon grass oil in a previous experiment. Tadpoles were housed individually in their dechlorinated home environment tank prior to videotaping. All tadpoles were similar in developmental stage near Gosner stages 38-41 (Gosner, 1960). Insecticide exposed groups were subjected to chemicals at concentrations similar to those that wild urban tadpoles would be exposed to 1 week prior to videotaping. We assumed that this time frame allowed for chemicals to take effect on tadpole behavior and still remain in body systems. Tadpoles were videotaped in their “home” environment for 15 minutes and then immediately transferred to a “novel” environment for an additional 15 minutes. Disturbance during transfer was minimized by using a mesh net. All trials were run on two consecutive days to test for repeatability. A 2×4 (6.25 cm) square grid was placed on the bottom of the home tank and a 4×8 (6.25 cm) grid on the bottom of the novel tank to measure horizontal movement. A 2×8 (6.25 cm) grid was placed on the back of both tanks to measure vertical movement. Blinds were placed in between tanks to prevent visualization of other tadpoles altering behavior. A GoPro camera was positioned above the tanks to record horizontal movement and a Canon Vixia HF300 camera was positioned on the side of tanks to measure vertical movement. After a trial was complete, tadpoles were transferred back to their home tanks. Novel tanks were washed, dried, and refilled with dechlorinated water in between all sessions. Home tank water was not changed until after both sessions had been completed. Videos were analyzed over the course of a summer and each individual tadpole’s behavior and position were tracked in Excel. Individual proportion of time spent active, number of boxes visited relative to total boxes, surface proportion of time, and center box proportion of time (novel tank), were averaged for both day 1 and day 2 of data. Averages of all tadpoles, and of each group were calculated, as well as averages of day 1 and day 2. T-tests were used to determine the significance of differences between the same tadpole on different days (correlated) and between

Stranzl

The Manhattan Scientist, Series B, Volume 4 (2017)

103

groups (independent).

Results Fig. 1 shows the average proportions of time that all tadpoles spent in the home and novel tanks on day 1 and day 2. The proportion of time spent active in the novel tank on day 2 was significantly larger than on day 1 (P < 0.05, t = 0.003, df = 14). However, home activity did not differ significantly from day 1 to day 2. 40%

Day 1

Day 2

30% Figure 1. Average proportion of time out of 15 minutes all tadpoles spent active in Home and Novel tanks during Day 1 and Day 2.

20%

10%

Fig. 2 shows the average relative number of boxes visited for each of the three treatment groups. Carbaryl exposed tadpoles visited significantly more boxes in the home tank than the lemon grass oil exposed group (P < 0.05, t = 0.049, df = 8). The difference in number of boxes visited between carbaryl and control groups and the difference between control and lemon grass oil groups were both not significant. 20 15 Figure 2. Average number of boxes visited within 15 minutes for each of the three treatment groups in their home tanks.

10 5 0

104

The Manhattan Scientist, Series B, Volume 4 (2017)

Stranzl

Fig. 3 shows the average proportion of time spent at the surface for each of the three treatment groups. Carbaryl exposed tadpoles spent significantly more time at the surface than both the control group (P < 0.05, t = 0.015, df = 8) and the lemon grass oil exposed group (P < 0.05, t = 0.017, df = 8). The difference between control and lemon grass oil surface usage was not significant. 8% 6% Figure 3. Average proportion of time out of 15 minutes each of the three treatment groups spent at the surface in their home tanks.

4% 2% 0%

Discussion As shown in Fig. 1, there was a significant increase in average tadpole activity, in novel environments as compared to activity in their home tank. This shows an increase in active behavior in the novel tank from day 1 to day 2. There was also significantly more novel tank boxes visited on day 2 than there was on day 1 (P < 0.05, t = 0.004, df = 14), showing an increase in exploratory behavior. A reason for this could be that on day 1, when transferred to a new and larger environment, the tadpoles were fearful and hesitant to readily explore the novel tank. However, on day 2, after having already been conditioned to experience the novel environment on day 1, on average tadpoles were much more exploratory and active and less afraid of the novel tank. This provides us with evidence of personality among the population of tadpoles because as they grew less afraid of novel environment exposure, they increased activity and exploration behaviors. Tadpoles on average also spent a majority of time avoiding center boxes in the novel tank, indicating a hesitancy to become fully exposed while exploring novel environments. Fig. 2 shows that in the home tank, there were significantly more boxes crossed by the carbaryl exposed treatment group than the lemon grass oil exposed group. While on average, carbaryl exposed tadpoles did visit more boxes (average = 10 boxes) than the control group (average = 4.14 boxes), this number was not significant. This indicates that tadpoles subjected to carbaryl were significantly more exploratory than the lemon grass oil exposed group and were on average more exploratory than the control group. The carbaryl exposed group also spent significantly more

Stranzl

The Manhattan Scientist, Series B, Volume 4 (2017)

105

time at the surface (as seen in Fig. 3) than both the control and lemon grass oil exposed groups. This also indicates that carbaryl subjected tadpoles are more active than control or lemon grass oil exposed tadpoles. These results provide evidence that the insecticide carbaryl has an effect on tadpole personality, causing them to become more active and exploratory than normal. There were no significant differences between the control group and the lemon grass oil exposed group, suggesting that lemon grass oil has no effect on tadpole behavior and personality. Our results add to the small but growing bank of evidence that point to amphibian personality. The results that tadpoles exposed to carbaryl insecticide are significantly more active and exploratory than control tadpoles is important because this would be similar to the concentration of carbaryl to which wild tadpoles are exposed. Therefore, tadpoles in the wild are likely experiencing the same shift in active and exploratory personality axes. Tadpoles belonging to the behavioral types of highly active and highly exploratory are far more likely to be out in the open and consequentially, possibly more often exposed to predation. So while the concentration of carbaryl in aquatic habitats may be sub-lethal, our evidence suggests that this pollutant alters tadpole personality in ways that could be harmful, and may ultimately lead to death anyway via a predator. This is significant, because carbaryl is one of the most widely used insecticides, and is detected in approximately half of urban streams (EPA, 2012). Our evidence also suggests that lemon grass oil insecticide has no significant effect on tadpole personality because no differences were detected between the lemon grass oil exposed group and the controls. For this reason, lemon grass oil insecticide could be a better alternative to carbaryl because it does not alter individual personality in ways that may eventually lead to species decline. Our results were consistent with the results of previous studies in that tadpoles do exhibit evidence of individual personalities just like other organisms, and this can be determined by studying the personality axes of activity and exploration (Carlson et al., 2013). In addition, these personalities are consistent through differing situations and repeatable over time (Wilson, et al., 2012; Urszan et al., 2015). Our results also agree with previous studies in that different groups or populations of tadpoles may exhibit differing personalities, in our case caused by exposure to an environmental pollutant (Brodin et al., 2013). The next step in this study is to determine if individuals within groups show evidence of unique personalities. This will be achieved by reanalyzing the data obtained from the original video recordings. In order to do this, we will have to provide evidence for consistent and repeatable variations between individual tadpoles within specific groups.

Acknowledgements This work was funded by the Linda and Dennis Fenton â&#x20AC;&#x2122;73 Endowed Biology Research Fund. The author would like to thank Dr. Maria Maust-Mohl and Taylor Maher for help with video analyses, and Dr. Gerardo Carfagno for being the best advisor!

106

The Manhattan Scientist, Series B, Volume 4 (2017)

Stranzl

References Brodin, Tomas, Martin I. Lind, Mira Kaltiala Wiberg, and Frank Johansson. 2013. Personality Trait Differences Between Mainland and Island Populations in the Common Frog (Rana temporaria). Behavioral Ecology and Sociobiology. 67:135-143. Carlson, Bardley E. and Tracy Langkilde. 2013. Personality Traits Are Expressed in Bullfrog Tadpoles During Open-Field Trials. Journal of Herpetology. 47.2:378-383. Conrad, J. L., K. L. Wienersmith, T. Brodin, J. B. Saltz, and A. Sih. 2011. Behavioural Syndromes in Fishes: A Review with Implications for Ecology and Fisheries Management. Journal of Fish Biology. 78:395-435. EPA (Environmental Protection Agency). 2012. Aquatic Life Ambient Water Quality Criteria for Carbaryl â&#x20AC;&#x201C; 2012. (EPA Publication No. 820-R-12-007). Washington, D. C.: U. S. Environmental Protection Agency. Gosner, Kenneth L. 1960. A simplified Table for Staging Anuran Embryos and Larvae with Notes on Identification. Herpetologica. 16(3):183-190. Reale, Denis, Simon M. Reader, Daniel Sol, Peter T. McDougall, and Nick J. Dingemanse. 2007. Integrating Animal Temperament Within Ecology and Evolution. Biological Reviews. 82:291318. Sih, Andrew, Julien Cote, Mara Evans, Sean Fogarty, and Jonathan Pruitt. 2012. Ecological Implications of Behavioural Syndromes. Ecology Letters. Stamps, Judy A. and Ton G. G. Groothuis. 2010. Developmental Perspectives on Personality: Implications for Ecological and Evolutionary Studies of Individual Differences. 2010. Philosophical Transactions of the Royal Society. 365: 4029-4041. Stuart, Simon N., Janice S. Chanson, Neil A. Cox, Bruce E. Young, Ana S. L. Rodrigues, Debra L. Fischman, and Robert W. Waller. 2004. Science. 306(5702):1783-1786. Urszan, Tamas Janos, Janos Yorok, Attila Hettyey, Laszlo Zsolt Garamszegi, and Gabor Herezeg. 2015. Behavioural Consistency and Life History of Rana dalmatina Tadpoles. Oecologia. 178:129-140. Wilson, Alexander D. M. and Jens Krause. 2012. Personality and Metamorphosis: is Behavioral Variation Consistent Across Ontogenetic Niche Shifts? Behavioral Ecology.

Generating a 3D CAD model of tree branches using a 3D scanner Michael Volgende∗ Department of Mechanical Engineering, Manhattan College Abstract. With the advancement of technology and the growth of new industries, some research experiments that were not possible a decade ago can now be conducted. Three-dimensional (3D) scanning is a relatively new field and has a wide variety of uses. For example, scanners can be used for industry, healthcare, and science. For this research a 3D scanner was used to scan tree branches taken off trees from around the Manhattan College area. Trees vary both biologically and mechanically. Biologically, trees differ due to their gnome and the location of the tree. Mechanically, trees are different based on the geometry of each individual branch. This project continues previous research that focused on one dimensional wire models. Ten pairs of branches were taken to produce a 3D model. Using the 3D models, we generated a comparison between each branch and each tree.

Introduction Understanding the mechanical properties of tree branches will allow for a better understanding of the nature of trees [1]. Analyzing these stresses will allow for a better understanding of how trees are impacted by their environment and how the tree adapts. Why trees place their branches where they are or why a particular species produces the “pattern” they do. For example, various samples can be taken off different sections of a tree and then compared. Some of the samples may break after a certain load is applied, while other samples may be able to endure the load. The results of the stresses provide how certain “situations/actions” impact the branch, and due to these results how the branch impacts the tree can also be analyzed. The results of these mechanical stresses can simulate how the branch would react in a natural disaster and this can lead to the prevention of damage to the tree. We can see what parts of the tree will break under certain forces and prevent any damage if the tree is in a residential area. Analyzing how mechanical stresses impact a tree can also lead to other real world examples. For example, this technique will be used to eventually compare mechanical stresses of tree branchesof a variety of sizes and geometries. Using a 3D scanner is a more efficient method to analyze the mechanical stresses on tree branches. The scanner produces more accurate results and greatly reduces the time it takes analyze the stresses. Not all branches are perfectly straight or curved, so the scanner is able to get the true shape and curvature of the branch. Scanning the tree branch is 10 times faster than modeling the whole tree branch in Abaqus. This allows for more accurate results than measuring the length and angle of each branch. Results with a 3D scanner will be compared with results of wire models. A wire model is depicted as a line in Abaqus/CAE and is used to idealize a solid in which both its thickness and depth are considered small compared to its length [2]. ∗

Research mentored by Zahra Shahbazi, Ph.D., and Lance Evans, Ph.D.

108

The Manhattan Scientist, Series B, Volume 4 (2017)

The hypotheses of this study are: Images from 3D scanners produce an accurate representation of tree branches. Mechanical stresses of tree branches determined with a 3D scanner are similar to stresses determined from wire models for simple cases. 3D models of tree branches derived from 3D scanners coupled with analysis with Computer Aided Design (CAD) and Finite-Element Analysis (FEA) provide accurate estimates of mechanical stresses, and mechanical properties of tree branches with various sizes and geometries.

Materials and Methods Wire model calculations Before purchasing the 3D Scanner, it was a necessary step to ensure that a simple wire and 3D model produce similar results. This was also used to ensure that the Finite software Abaqus gave accurate data and did not differ too much. Three simple cases were tested to prove the theory that 3D models can give accurate mechanical stresses. A simple cantilever beam with a rectangular and another with a circular cross section. Both cases were under a concentrated load a one end and the other end being fixed. The third case was a rectangular cantilever beam under the force of gravity. The results from the models were compared with an exact solution that was hand calculated. Table 1 provides all the assumed variables that were used to find the stress. Table 1. 1: Given data for wire and 3D model comparison Density (ρ)

657

kg m3

Force (P )

1 Newton

Young’s Modulus (E)

10,000 Pa

Poisson’s Ratio (v)

0.3

Length (L)

Thickness (t)

For all cases the first step to calculate the stress on the beam is to find the moment of inertia. This can be found using the formula in Eq. (1) for the rectangular cross section and formula in Eq. (2) for the circular cross section. 1 I = bh3 (1) 12 π I = D4 (2) 64 where b is the length of the base and h is the height (both are one in this situation). This simplifies formula is used in the general formula, Eq. (3). σ xx =

P z(L − x) I

(3)

The Manhattan Scientist, Series B, Volume 4 (2017)

109

where P is the force applied, z is half the length of the beam and L − x is the distance of the branch. Using all the given data from Table 1, we get σ xx = 60 MPa. For case 3 a different method was used because the load for Case C was different than Cases A and B. For Case C a distributed load was applied, while Case A and B had an applied force. Using the same assumptions as in Table 1, the volume and the mass were both needed, V = AL

(4)

m = ρV.

(5)

where A is the area of the beam, L is the length of the beam and ρ is the density of a tree branch used to simulate the beam. Applying the summation of moment equation gives us: X MA = 0

(6)

L =0 (7) 2 where MA is the moment about point A (Fig. 1) and FYD is the distributed force along the beam. This simplifies to L L L (8) MA = FYD = mg = ρ V . 2 2 2 Inserting the given values into Eq. (8) yields the moment about the beam. This, in turn is used in the following equation to find the stress MA − FYD

σ xx =

MZ . I

(9)

Figure 1. Sample Branch (A) and STL file of Branch (B)

As seen from Tables 2 through 4, the smaller the element size the more accurate the data. This is true for all problems in Abaqus not just wire. The smaller the element size, the less accurate both models were; however, once the element size increased, the percent difference decreased

110

The Manhattan Scientist, Series B, Volume 4 (2017)

substantially. The percent difference for the wire model are all less than 1 percent which proves that Abaqus produced accurate data, and that 3D models can be used to get the mechanical stresses of tree branches. Table 2. Results for case A of wire and 3D comparison with percent difference Element size

3D (Pa)

Wire (Pa)

% difference for 3D model

% difference for wire model

1 0.5 0.25 0.125 0.0625

37.87 38.86 48.47 58.21 70.24

57 58.5 59.25 59.63 59.81

36.88 35.23 19.21 2.98 -17.07

5 2.5 1.25 .616 .317

Table 3. Results for case B of wire and 3D comparison with percent difference Element size

3D (Pa)

Wire (Pa)

% difference for 3D model

% difference for wire model

1 0.5 0.25 0.125 0.0625

60.73 68.61 87.40 95.8 106.8

96.77 99.31 100.6 101.2 101.5

40.38 32.64 14.19 5.94 -4.84

5 2.5 1.23 .648 .353

Table 4. Results for wire and 3D comparison with percent difference Element size 1 .5 .25 .125 .0625

3D (×106 Pa) 5.75×10 1.22 1.54 1.87 2.28

−13

Wire (×106 Pa)

% difference for 3D model

% difference for wire model

1.75 1.84 1.89 1.91 1.92

100 36.78 20.2 3.10 -18.1

9.32 4.66 2.07 1.03 .518

Choice of a commercial scanner Three dimensional imaging is a relatively new area of science and new models and applications are produced every year. Tree branches have varied and complex geometries. The branch terminals used in this study had thin branches. An important aspect of the scanner for the project involved detection of thin branches. After researching numerous scanners, the decision was made to purchase the Ein-Scan Pro (Ein-Scan Pro Basic Version; Intel Core i5 with NVIDIA GTX660 display card (Dynamism, Inc. Chicago, IL). The Ein-Scan Pro is a versatile scanner and can scan objects from about 3 cm to 40 m and possesses an accuracy of 0.1 mm [3, 4]. See Appendix for the application of the scanner with terminal branches.

The Manhattan Scientist, Series B, Volume 4 (2017)

111

Image transfer to a usable Abaqus file Once the scanner produced an image, the image was placed in aSTL (“Standard Triangle Language”) file, as seen in Fig. 2. This is a 3D surface file that allows the user to edit the image as well as import it into various programs. The STL file was transferred to NX. NX is a CAD design software that enabled imaging edits. For this research, NX was used to “trace and fill” (fill in empty gaps along the branch) in the branch and “smooth” the branch for easy use in Abaqus. Diameter measurements along branches were determined with a caliper and branch lengths were determined with a ruler. Once in NX, points are placed along the branch to present the form and shape of branches especially in curved or branched areas since all points were connected to form curves on an image Fig. 2. Then all curves were joined (Fig. 2). After all points and curves were connected, a datum plane is placed on each curve. A datum plane is a plane with direction coordinates (X, Y or Z) that can be altered and used to sketch on. For branches, a circle was drawn on the datum plane with the proper diameter. The last step is to use a function in NX called “swept” this function creates a body by sweeping the section. The body of the branch is traced along these curves. After all portions of the branch have been swept, the resulting product can be seen below (Fig. 2).

Figure 2. Example pictures of procedure developed during this experiment.

Once the image was complete, it was exported from NX into Abaqus. Abaqus is a finiteelement analysis software that enables the branch to be analyzed for mechanical properties. In Abacus, the applied external forces and constraints to the branch can be visualized. Mechanical stresses were determined on the branches in Abaqus. A composite image of this process can be seen in Fig. 3.

Results A 3D scanner produces an accurate representation of tree branches. This task was accomplished using the method mentioned in this paper. Once the sample has been chosen, the leafs must be taken off the branch Fig. 4. The next step in the process was to scan the front side of the

112

The Manhattan Scientist, Series B, Volume 4 (2017)

Figure 3. Displacement and Stress on two sample branches.

branch. The resulting image is a three dimensional figure of the top of the branch. This file is imported as an STL file which can be altered and imported into the auto cad software, Siemens NX. Since the imported file is only the top section of the branch, the remaining portion of the branch must be â&#x20AC;&#x153;traced and filled in.â&#x20AC;? The diameter of each section of the branch was taken using a caliper. Points were place along the branch at changes of direction or angle. The file acted as a path for this step. A function in NX was used to connect all the point and each diameter. Once the 3D model was complete, it was imported into Abaqus where boundary conditions and forces were applied and the mechanical stresses were analyzed. The second aim of the research was to determine if mechanical stresses determined with 3D models of tree branches are similar to stresses determined from wire models for simple cases. Wire and 3D models are similar for simple cases. As seen in the tables above, the simple wire and 3D models produce accurate results within 5 percent. Different scenarios were conducted using circular and rectangular cross sections and gravity as the force. Case A involved a rectangular cross section with a concentrated load. Case B was similar to case A except the cross-section for this case was circular rather than rectangular. The last case was for a rectangular cross-section with gravity applied along the whole body. These results ensured that the wire and the 3D model were compatible and the scanner would provide accurate results. Images from 3D scanners coupled with analysis with Computer Aided Design (CAD) and Finite-Element Analysis (FEA) provide mechanical properties of tree branches that have various shapes, sizes, and geometries. This was accomplished using the method stated in the report. In

The Manhattan Scientist, Series B, Volume 4 (2017)

113

Figure 4. Branch before scanning (A-C) with Condensed Procedure (D)

Abaqus, boundary conditions and forces are applied to the 3D model. In this study, the end of the branch is held and cannot be rotated or moved, as it would not in real life. The forces that can be applied vary, for example a concentrated load can be applied to simulate a bird standing on the branch, or an area forces (like pressure) can be applied to simulate wind, snow or rain. The sample is also meshed. A mesh is an approximation of the geometry of the physical part. Essentially, it connects all the small areas generated by the software [2]. This can be altered by choosing a different element size, i.e. how far each area is spaced, or the element type, as different shapes impact the part differently. After being meshed, the mechanical stresses of the branch can be visualized.

Discusion Using a 3D Scanner produced an accurate 3D model of a tree branch that is ten times faster than methods used in previous years and also gives more accurate data. There is less error this way compared to measuring the branch by hand and creating the profile based off that. The scanner provides a 3D image with accurate points that are used to develop the model. No previous knowl-

114

The Manhattan Scientist, Series B, Volume 4 (2017)

edge of NX or Abaqus is needed, the steps are very basic and any person once shown will be able to develop an accurate model of a tree branch. Wire models and 3D models are similar for simple cases. This was a necessary step in determining if a 3D scanner could be used. As the element size of the model decreases so does the force and accuracy. The smaller the element size the more accurate the mechanical stresses are. This makes sense because the smaller the meshed areas are, the more accurate Abaqus will be. There are numerous future applications that can be done using this research. One of those is to try and automate the procedure. This will be done to try and decrease the time it takes to produce the 3D model of the branch. A second application is to compare each tree species and analyze how each tree is impacted by various factors. The 3D model will break if a large enough force is applied and this can be visualized in Abaqus. Since we can see how each branch will be impacted by various factors, this can be used to aid the community, e.g. by labeling which trees are at risk to break during certain storms. Another application is to add leaves to the 3D models to analyze their effect on branches under different environmental factors (e.g. snow or wind). This would allow researchers to get a better understanding of how trees are impacted by their environment.

Acknowledgments This work was funded by the Catherine and Robert Fenton Endowed Chair to Dr. Lance Evans. The author thanks Drs. Zahra Shahbazi and Lance S. Evans for valuable advice.

References [1] Mattheck, Claus. 1998. “Design in Nature: Learning from Trees.” Springer [2] D S Simulia. Abaqus/CAE User’s Guide. Retrieved from: http://abaqus.software.polimi.it /v6.14/books/usi/default.htm?startat=pt03ch11s09s03.html [3] Shining 3D, en.shining3d.com, “EinScan Series.” [4] https://store-0jms5nq.mybigcommerce.com/content/EinScan%20Pro%20and%20ProPlus.pdf

APPENDIX Use of Ein-Scan Pro for our application The scanner is composed of two sensors at the top and bottom with a light in the middle. The scanner has four settings, Handheld HD Scan, Handheld Rapid Scan, Automatic Scan, and Free Scan. The scanner uses a white LED light to pick up the location of the branch. Therefore, any fluorescent or natural light would disrupt the scan and produce inaccurate results. To compensate for this the scanner was used in a completely (as possible) dark room. The scanner produces three types of output files: an STL (“Standard Triangle Language”), an OBJ, and a PLY (Polygon File

The Manhattan Scientist, Series B, Volume 4 (2017)

115

Format). Each of these files are usually used for 3D CAD modeling, but for this research we used an STL file. Once the scan was finished, the software that came with the scanner was able to mesh the object, as seen in Fig. 3. Due to the complexity of the shape of each branch, it was determined that the best mode this research was the Handheld Rapid Scan. Process of Scanning The process of scanning was extremely difficult due to the complex geometry of each tree branch. Many of the branches used in this research are very small in diameter, ranging from 2 to 8 millimeters. The larger the diameter, the more surface area the branches contains and therefore the easier the scanning process is. Before scanning, the leaves of the branch were removed and the diameter of the branch was recorded using a caliper (Fig. 4). The scanner works by determining the shape of the object by using “point cloud” recognition. This “point cloud” data is the location of points in any dimension and in our research, three dimensions X, Y and Z are used. In other terms, the scanner shines a light on the object and the two sensors determine the location of each point that is recognized. For example, in our study the light is beamed onto the branch and the surrounding area, the scanner then picks up the location of each point on the branch (up to .1 millimeter) and from these points the shape of the object is formed. However, if the object is too small or too thin in our case, the scanner had a difficult time picking up the points of the object and essentially the scanner gets “confused” and loses track of the object. This was one of the major issues when scanning a very thin tree branch, the scanner had no problem picking up the front face of the object, but when the object was rotated the scanner would lose track of the branch very easily. The scanner also had a difficult time with branches that contained many side branches that were in all directions. The scanner could not determine where one side branch ended and another branch began. The first method used to scan each branch, involved clamping the sample to a yellow apparatus, as seen below. This was then placed on a turntable so it could be rotated. Starting with either the front or back face the object was then rotated slowly so the scanner could pick up the sides of the branch. The yellow apparatus was not just used to hold the branch in place; it was also used as a reference point for the scanner. This ensured the scanner always had something to pick up and would not lose its position. Before rotating the object, the yellow apparatus was scanned and then the scan proceeded upwards from that location. This worked well for the thicker branches; however, as the branches became thinner this process became significantly harder. Other methods were attempted to try and work around this issue, however it was to no avail. For example, since the light was white, and white objects reflect more, some of the samples were spray painted white so that the scanner could pick up more of the points on the branch. Another technique that was attempted was using the Handheld HD scan. This involved placing stickers with special lightreflective material around the branch. During the scanning process the light would reflect off the stickers and back to the sensors on the scanner. Unfortunately, this did not work either, as the stickers had little to no effect on producing a better image.

116

The Manhattan Scientist, Series B, Volume 4 (2017)

The final solution to the problem of the procedure mentioned in this report was to scan only the front face of the branch, instead of scanning the entire branch. This produced a 3D image of just the top half of the branch, that was then imported into NX CAD Design where it was to be â&#x20AC;&#x153;traced.â&#x20AC;?

Methylammonium lead iodide nanowires for perovskite solar cells Jacqueline DeLorenzoâ&#x2C6;&#x2014; Department of Chemistry and Biochemistry, Manhattan College Abstract. Methylammonium lead iodide (MALI) is an organic-inorganic lead halide that arranges itself as a tetrahedral perovskite structure and is an optimal material for converting photons into electricity. The movement of electrons through the band gap of MALI generates electricity. The focus of this research is to produce MALI nanowires to be incorporated into a photovoltaic device. These 1D nanomaterials are obtained by conducting a precipitation reaction through a polycarbonate template to produce lead iodide nanowires. Immersion of the template into a solution containing the cation, methylammonium results in the production of the desired MALI nanowires. The sizes of the template pores range from 15, 50, 100, and 200 nm. We explore how time, concentration, pore size, and temperature affect the production of these 1D nanomaterials. X-ray diffraction and SEM imaging are used to characterize the nature of these synthesized nanowires.

Introduction With the ever-increasing world population comes the demand for more and more energy to be produced. Due to evidence of climate change and the dangers of global warming, the need for energy to be produced from renewable resources has significantly increased. Solar energy has rose in popularity over time due to its ease and renewability. However, traditional solar cells are composed of silicon, which can prove to be rather expensive. In order to lower the cost of solar cells and make this energy more widely available to the public, silicon may be replaced with an organometal halide, such as methylammonium lead iodide (Fig. 1). Methylammonium lead iodide is an ideal alternative because it absorbs a large portion of the light spectrum and has a band gap of approximately 1.6 eV [1]. In addition, the synthesis of methylammonium lead iodide (MALI) is relatively easy in comparison to silicon solar cells. The synthesis for MALI can be done under mild heating conditions (150â&#x2014;Ś C) and conducted in relatively safe solvents, such as isopropyl alcohol and dimethylformamide [2, 3]. Recent work has shown that this material is capable of producing solar cells with efficiencies over 22%, which exceeds commercially available silicon based solar cells [4]. However, there is a lack of application of nanotechnology to these types of solar cells. This research aims to understand the effect 1D nanomaterials have on the efficiency of perovskite solar cells. By creating methylammonium lead iodide based solar cells, solar energy may become more competitive with traditional carbon-based fuels. Another attractive property of MALI is that it arranges at an atomic level into a perovskite structure. Perovskite materials are a crystal lattice structure based on pyramid-like tetrahedral arrangements of atoms or molecules [5]. At the center of the MALI molecule there is the central inorganic cation, lead (Pb2+ ). This ion is surrounded by eight organic methylammonium cations and six iodide anions. This repeating arrangement allows for replacement of each of the individual â&#x2C6;&#x2014;

Research mentored by Alexander Santulli, Ph.D.

118

The Manhattan Scientist, Series B, Volume 4 (2017)

DeLorenzo

Figure 1. Atomic structure of methylammonium lead iodide. [1], [2]

components separately, thus allowing for optimization of the various cations and anions to produce the most stable and efficient material for use in solar cells. The final photovoltaic device is composed of several layers. A thin layer of glass is coated with fluorine-doped tin oxide and titanium dioxide. Atop these layers a layer of MALI, and a layer of spiro-MeOTAD are spray coated onto the glass, respectively. Finally, a layer of Gold is deposited by sputter coating to complete the device (Fig. 2). The purpose of the methylammonium lead iodide is to absorb photons of light. This excites electrons within the MALI to move from the lower energy valence band to the high energy conduction band [6]. Titanium dioxide conducts electrons and acts as the driving force to move electrons from the conduction band to lower energy levels. Spiro-MeOTAD acts as a hole conductor as electrons leave spiro-MeOTAD to jump to the unoccupied lower MALI orbital. Gold acts as a metal conductor and donates electrons to move to the spiro-MeOTAD orbital, filling in the hole left by light excitation. This creates a cycle that continuously refills the valence band with electrons that move across the band gap to the conduction band to generate electricity (Fig. 3). Our objective was to create methylammonium lead iodide nanowires, understand the experimental parameters that influence the synthesis of the nanowires and, ultimately, determine the effect 1D nanostructures have on the performance of perovskite solar cells.

Materials and Methods Synthesis of lead (II) iodide nanowires In order to create the methylammonium lead iodide nanowires, lead iodide nanowires were first synthesized. This was done through a U-tube reaction. The half-cells of the U-tubes were clamped together with a polycarbonate template separating the two half-cells of the reaction. The pores of the template ranged in size from 15-200 nm (Fig. 4). Simultaneously, solutions of potassium iodide and lead (II) nitrate were poured into each end of the U-tube system. The reaction was then allowed to proceed for various lengths of time before the reaction was stopped by pouring

DeLorenzo

The Manhattan Scientist, Series B, Volume 4 (2017)

119

the solutions out of the U-tube. The temperature of the reaction was controlled by immersing the entire U-tube into an ice-bath.

Gold Spiro-MeOTAD Solution methylammonium lead iodide Titanium Dioxide Fluorine-Doped Tin Oxide Glass Figure 2. 1-D layers of a completed photovoltaic device

Figure 3. The movement of electrons within MALI. Spiro-MeOTAD takes the role of PANI in this case [3]

Figure 4. Basic U-tube system in ice bath.

After a given reaction time, the U-tube was dismantled and the template was removed. The excess product was removed from the surface of the template by rubbing the template on an Arkansas stone to grind off any debris. To isolate the nanowires from the template, the polycarbonate was dissolved using methylene chloride and the solid product was isolated by centrifugation. To ensure complete removal of polycarbonate, the product was washed and centrifuged a total of 4 times. For electron microscopy analysis and X-Ray diffraction investigations, the product was dispersed into methylene chloride and deposited onto silicon substrates.

120

The Manhattan Scientist, Series B, Volume 4 (2017)

DeLorenzo

Conversion to MALI To convert the synthesized nanowires to MALI nanowires, the template containing the lead (II) iodide nanowires was immersed into a 0.1 M solution of methylammonium iodide in isopropyl alcohol. The template was left in the MALI solution for several minutes before being removed and rinsed with isopropyl alcohol. Following this conversion process, the nanowires were isolated using the same methods mention above of dissolving in methylene chloride and isolated through centrifugation. Making the photovoltaic device Glass coated with a layer of Fluorine-doped Titanium Oxide was used to make the photovoltaic device. An airbrush was used to spray coat the methylammonium lead iodide nanowires onto a small section of the glass. A solution of chlorobenzene containing 68 mM spiro-OMeTAD, 55 mM tert-butylpyridine, and 9 mM of a lithium salt was then spray coated atop the MALI nanowires [7]. The chlorobenzene evaporated during spray coating, leaving the important hole-conductor spiro-MeOTAD compound. Lastly, the glass was placed in a gold-sputtering machine to layer the photovoltaic device with a thin layer of conductive gold (Fig. 5).

Figure 5. Final photovoltaic device.

Results and Discussion It was found that the concentration of the precursors is crucial to creating a predominantly 1D motif. Initially the concentrations of the Pb(NO3 )2 and KI were set to 0.1 M to ensure precipitation would occur. However, analysis of these samples by Scanning Electron Microscopy (SEM) revealed that there was a significant contamination of the sample with â&#x20AC;&#x153;bulkâ&#x20AC;? PbI2 (Fig. 6). A concentration gradient is created in order to force lead iodide wires to form in the pores of the template rather than on the surface. The system was placed in an ice bath to lower the temperature and force the nanowires out of solution. The reaction was run for time periods ranging from 1 hour to 24 hours. It was found that the reaction needed at least 2 hours to form a significant amount of nanowires. Four hours was the optimal amount of time for the reaction to produce the most nanowires. Past 4 hours the amount of nanowires produced remained constant.

DeLorenzo

The Manhattan Scientist, Series B, Volume 4 (2017)

121

Figure 6. Contaminated MALI.

After the reaction was run for an allotted amount of time, the u-tube system was broken down and the template was removed. To remove excess lead iodide that formed on the surface, the template was rubbed against a polishing stone, leaving only the nanowires inside the pores. Mineral oil was used as a lubricant to avoid the template ripping while polishing. The template was then consecutively immersed in isopropyl alcohol and then methylammonium. Once the methylammonium lead iodide wires were formed the template was dissolved in methylene chloride to isolate the wires. This solution was centrifuged and decanted several times until only the wires remained. X-Ray diffraction In order to test if it was in fact lead iodide and methylammonium lead iodide that was being synthesized, these samples were run through an X-ray Diffraction machine to test if the samples were pure. First, our synthesized lead iodide was run and compared to a bulk sample of standard lead iodide. The peaks between the two sets of data are accurately aligned showing that the two samples are composed of the same material, proving that lead iodide was successfully created (Fig. 7). Lead iodide was then compared to the synthesized methylammonium lead iodide to test if the lead iodide was being completely transformed into the new product. According to the graph, the peaks do not align between the two samples meaning that they are not composed of the same materials. This indicates that lead iodide was successfully transformed into a new substance, methylammonium lead iodide (Fig. 8). SEM Imaging Scanning electron microscope photos were taken of the lead iodide and methylammonium lead iodide nanowires at the Fordham Laboratory and at Brookhaven National Laboratory. Nanowires were synthesized in pores of 15, 50, 100, and 200 nm. 100 and 200 nm nanowires were synthesized with the most precision resulting in the greatest amount of unbroken nanowires. The nanowires shown in Fig. 9 were created under the optimal concentration, temperature, and time conditions.

122

The Manhattan Scientist, Series B, Volume 4 (2017)

Figure 7. X-ray diffraction between synthesized and standard lead iodide.

DeLorenzo

Figure 8. X-ray diffraction between synthesized lead iodide and methylammonium lead iodide.

Figure 9. (A) 100 nm lead iodide nanowires. (B) 200 nm lead iodide nanowires. (C) 100 nm methylammonium lead iodide nanowires. (D) 200 nm methylammonium lead iodide nanowires.

DeLorenzo

The Manhattan Scientist, Series B, Volume 4 (2017)

123

The synthesized nanowires ranged in about 1.0 â&#x2C6;&#x2019; 7.0 microns in length and 0.1 â&#x2C6;&#x2019; 0.2 nm in width. 15 and 50 nm wires were also created, but not with much success (Fig. 10).

Figure 10. (A) 15 nm MALI nanowires. (B) 50 nm MALI nanowires. (C) 15 nm MALI nanowires from supernatant liquid. (D) 50 nm nanowires from supernatant liquid.

Due to the relatively small size of the template pores, the lead iodide diffused too quickly and resulted in incomplete nanowires. The supernatant liquids were collected from the 15 nm and 50 nm templates and centrifuged at a higher speed to isolate any remaining wires that were too small to come out of solution. This centrifugation revealed many small nanowires; however, they were broken and incomplete. In order to create 15 and 50 nm wires in greater amounts, the concentration of the reagents and the separation techniques must be fine-tuned to discover the optimal conditions.

Conclusion In conclusion, both lead iodide and methylammonium lead iodide nanowires were successfully synthesized and characterized using X-ray diffraction and SEM imaging. When a voltmeter

124

The Manhattan Scientist, Series B, Volume 4 (2017)

DeLorenzo

is attached to the photovoltaic device and placed in the light an electrical current can be measured meaning that these wires can indeed be used to create a functioning solar cell. These nanowires offer a less expensive alternative to traditional solar cells and open a door to the potential creation of nanowires made from other inexpensive perovskite materials, such as tin. With further investigation, optimal photovoltaic devices may be created to be both high functioning and inexpensive.

Acknowledgments This work was funded by the Manhattan College School of Science Summer Research Scholars Program. The author thanks the Department of Chemistry and Biochemistry and the Dean of Science for this opportunity, and especially her research faculty advisor Dr. Alexander Santulli for the advisement and support.

References [1] Kim, H.-S., Lee, C.-R., Im, J.-H., Lee, K.-B., Moehl,T., Marchioro, A., Moon, S.-J., HumphryBaker, R., Yum, J.-H., Moser, J. E., Gr¨atzel, M., and Park, N.-G. Lead iodide perovskite sensitized all-solid-state submicron thin film mesoscopic solar cell with efficiency exceeding 9%. Sci. Rep. 2, 591 (2012). [2] Manser, J. S., Saidaminov, M. I., Christians, J. A., Bakr, O. M. and Kamat, P. V. Making and breaking of lead halide perovskites. Acc. Chem. Res. 49, 330–338 (2016). [3] Yamanaka, T., Masumori, K., Ishikawa, R., Ueno, K. and Shirai, H. Role of isopropyl alcohol Solvent in the Synthesis of Organic–inorganic halide CH(NH2)2PbIxBr3–x perovskite thin films by a two-step method. J. Phys. Chem. C 120, 25371–25377 (2016). [4] Yang, W. S., Park, B.-W., Jung, E. H., Jeon, N. J., Kim, Y. C., and Lee, D. U. Iodide management in formamidinium-lead-halide–based perovskite layers for efficient solar cells. Science 356, 1376-1379 (2017). [5] Lee, M. M., Teuscher, J., Miyasaka, T., Murakami, T. N. and Snaith, H. J. Efficient hybrid solar cells based on meso-superstructured organometal halide perovskites. Science 338, 643647 (2012). [6] Boix, P. P., Agarwala, S., Koh, T. M., Mathews, N. and Mhaisalkar, S. G. Perovskite solar cells: Beyond methylammonium lead iodide. J. Phys. Chem. Lett. 6, 898–907 (2015). [7] Sevhab. (No Title) wikicommons (2014). Available at: https://commons.wikimedia.org/wiki /File:Perovskite unit cell.png.

Remarkably rapid reduction of toxic Cr(VI) levels Patsy Griffin∗ Department of Chemistry and Biochemistry, Manhattan College Abstract. Chromium(VI) (Cr(VI)) is found in the environment both organically by being a product of the erosion of natural chromium deposits, as well as industrially through inadequate industrial waste disposal practices and runoff from dye or textile factories. When consumed its effects range from certain cancers to skin damage as well as respiratory problems. With this contaminated water being drank by Americans across 42 states, the growing problem of Cr(VI) in drinking water has become a top priority. Chromium’s other common form, Cr(III), is an essential human dietary element and is found in many vegetables, fruits, meats, grains, and yeast; for this reason, this project aimed to chemically reduce toxic Cr(VI) into beneficial Cr(III) through the use of an ascorbic acid ketal (5,6-isopropylidene-Lascorbic acid).

Introduction Hexavalent chromium, otherwise known as chromium(VI), is a priority pollutant due to its known carcinogen properties, according to the US Environmental Protection Agency (EPA, 2017; Wielinga et al., 2001; Yurkow et al., 2002). While its distribution amongst the environment is mostly due to its applications in industries such as leather tanning, stainless steel production, dye or textiles factories, it can also be found organically in areas with natural chromium deposits. Due to its strong toxicity the national primary drinking water regulation established the MCL (maximum contaminant level) for total chromium to be 0.1 ppm (0.1 mg/L or 0.862 µM) in order for drinking water to be deemed safe for consumption (EPA). Otherwise Cr(VI) has the ability to cause severe health problems such as respiratory damage, sinonasal cancer, dermatitis, DNA defects, and skin ulcers (Dayan and Paine, 2001). Contrastingly, the reduced form of toxic hexavalent chromium, Cr(III), is an essential nutrient for both humans and animals. Due to the fact that the oxidation state of chromium is what determines human and ecotoxicological effects, as well as physicochemical properties, such as mobility and transport behavior, the reduction of Cr(VI) to Cr(III) is very important in the remediation of contaminated chromium sites (Xu et al., 2004). With the escalated levels of chromium in ground water becoming a pressing issue, the water treatment industry has been in heavy pursuit of the most effective method of Cr(VI) removal (Owlad et al., 2008; Ponder et al., 2000; Hawley et al., 2004). One of the three most popular methods, which is employed in this research, is known as chemical oxidation-reduction. In this process the reaction takes place between Cr(VI) and a reducing agent to produce Cr(III) as well as the oxidized form of the given reducing agent. For this project specifically, chemicals such as an ascorbic acid derivative and granular activated carbon were used in order to reduce the concentration of Cr(VI) in solution. ∗

Research mentored by John Regan, Ph.D.

126

The Manhattan Scientist, Series B, Volume 4 (2017)

Griffin

Background In addition to lead or iron metal, another commonly used Cr(VI) reducing agent is ascorbic acid and its derivatives. Scheme 1, shown below, displays the reduction of Cr(VI) using ascorbic acid, or vitamin C. While there are some definite benefits to using ascorbic acid for a reducing agent, there are also deterring factors. For instance, ascorbic acid has several positive health attributes. But, a negative characteristic is that due to its very polar and soluble nature, the long-term consequences of its uncontrolled release into the environment are not well known.

Scheme 1

Due to the dangers of having untethered ascorbic acid in the Cr(VI) contaminated water, this project employed methods of both sorption via the use of granular activated carbon, commonly referred to as GAC and oxidation-reduction This solid adsorbent is commonly used by water treatment facilities because of its low cost and ability to maintain its solid phase when acting as a hexavalent chromium adsorbent (Weber, 1967; Park et al., 2016). GACâ&#x20AC;&#x2122;s adsorbent abilities arise from its porous physical properties which contains both hydrophobic and hydrophilic binding domains on the surface. Scheme 2 demonstrates the binding phenomena between GAC and the ascorbic acid ketal (5,6-isopropylidene-L-ascorbic acid) used in this project.

Scheme 2

This specific ascorbic acid ketal was chosen in hopes of improving the results found by my fellow researcher, Analisse Rosario, who reported that of the approximate 67 mg of ascorbic acid that was able to be loaded onto 1 g of GAC, only about 3 mg was responsible for reducing chromate

Griffin

The Manhattan Scientist, Series B, Volume 4 (2017)

127

(Rosario, 2016). This result suggested that there was competitive binding of ascorbic acid and GAC such that the C2-C3 oxygen atoms that are responsible for the 3 electron reduction of Cr(VI) to Cr(III) (Scheme 1) are not always accessible to chromate when bound to the hydrophobic domains within GAC (Scheme 3). Our hypothesis was that using the ascorbic acid ketal derivative would improve chromate reduction by producing more favorable binding orientations. This would likely occur by allowing the ketal group to bind in the hydrophobic binding domains, thereby exposing the C2-C3 groups to the chromate as shown in Scheme 2.

Scheme 3

Materials and Methods: What are GAC and AAK? GAC is an abbreviation for granulated activated carbon which is commonly used to adsorb natural organic compounds, taste and odor-producing compounds, and synthetic organic chemicals in drinking water treatment processes. Ascorbic acid ketal (AAK), is commercially available or can be readily prepared in the laboratory. Using ascorbic acid, acetone, and trifluoroacetic acid (TFA) with heating (Scheme 4) AAK can be synthesized in high yield and purity. Specifically, the ketal is made by mixing 10 g of ascorbic acid in 180 mL reagent grade acetone and 18 mL trifluoroacetic acid (TFA) in a 500 mL round bottom flask. After heating the mixture for 18 hours at reflux, the mixture is cooled to room temperature, filtered and washed with a minimum amount of MTBE.

Scheme 4

128

The Manhattan Scientist, Series B, Volume 4 (2017)

Griffin

Loading GAC with the ascorbic acid ketal. Preparation of GAC-AAK The optimal procedure for loading ascorbic acid ketal on GAC begins by first dissolving the ketal compound in distilled water while stirring. Specifically, 7 g of AAK was added to 600 mL distilled water and vigorously stirred until complete dissolution had occurred in approximately 20 minutes. A mixture of 7 g GAC was soaked in 200 mL distilled water for 10-15 minutes and the water decanted. The aqueous solution of AAK was added to the GAC and kept, without stirring, for 2 hours. The GAC-AAK was filtered and washed with copious amounts of water and dried overnight at 60◦ C. Potassium chromate reduction All chromium reduction trials were conducted using dry 0.5 g samples of the GAC-AAK product discussed above. With these samples chromate reduction was executed two different ways, one by a static reaction and the second by rotating via the Tek-Tator Variable Rotator. Chromate levels after exposure to GAC or GAC-AAK were determined by using EPA method 7196A from UV/Vis peak heights at 540 nm and correlated to a standardized calibration curve. In each case, UV/Vis spectrographic data was collected on an Agilent 8453 Spectrophotometer in 1 cm quartz cuvettes. Reported data values are averages of 2 trials each. A. Potassium chromate reduction under static conditions For the static trials each 0.5 g GAC-AAK sample was transferred into 250 mL beakers which contained various volumes of 200 µM potassium chromate solution. After adding the GAC-AAK sample to these solutions, the reaction vessel was covered with parafilm and left for 18 hrs. A sample of the water was collected and the concentrations of Cr(VI) were determined as described above. B. Potassium chromate reduction under dynamic conditions Cr(VI) reduction was measured in a dynamic stirring environment with the GAC-AAK complex. Each trial was conducted using a 0.5 g sample of GAC-AAK added to various volumes of 200 µM potassium chromate solution and the mixtures rotated on a Tek-Tator Variable Rotator for the specified time and speed. A sample of the water was collected and the concentrations of Cr(VI) were determined as described above.

Results and Discussion Ascorbic acid ketal loaded onto GAC (GAC-AAK) proved to be more effective at reducing Cr(VI) than when ascorbic acid is adsorbed onto GAC. That is, AAK is able to load 99.5 mg onto GAC and 8.6 mg is responsible for reducing chromium as compared to 67 mg of loading and 3.1 mg for reduction in Rosario’s research (Rosario, 2016). The maximum loading occurred at 2 hours, soaking the GAC for any longer than this proved to have no greater effect on reduction. The low loading results of Rosario suggested that there was ineffective binding occurring in GAC at the C2-C3 site, which is ideally the epicenter of chromate reduction. Theoretically reduction is

Griffin

The Manhattan Scientist, Series B, Volume 4 (2017)

129

primarily done by both carbons 5 and 6 binding onto GAC and therefore exposing carbons two and three, but it is possible that the unreactive carbons were instead exposed. By using this ascorbic acid ketal derivative, it was hypothesized that there would be improvement in chromate reduction by producing more productive binding orientations. This would happen by eliminating the chances of hydrogen bonding at that site and therefore allow the hydrophobic component (ketal) to produce hydrophobic binding. In order to reduce the chromate using the GAC-AAK complex, a method of static reaction was first employed. To 0.5 g of GAC-AAK in 150 mL beakers was added various amounts of 200 µM potassium chromate solution and were left undisturbed for 1 hour. After the reaction time an aliquot was tested using the EPA guidelines mentioned above and their concentrations were calculated. The results concluded that after an hour of static reacting, a 0.5 g sample of the GACAAK complex was able to successfully reduce 5 mL of 200 µM chromate solution to 12.5 µM, which is about 15 times over the EPA’s recommended limit. This was then compared to the ability to reduce Cr(VI) for GAC alone, where the GAC-AAK complex reduces Cr(VI) about 2 times more effectively than solely GAC. These results are compiled in Table 1 and Fig. 1. Table 1. Concentration of Cr(VI) using GAC and GAC-AAK after one hour reaction time Material

Volume of 200 µM Cr(VI)

Concentration of Cr(VI)

S.D.

GAC GAC GAC GAC-AAK GAC-AAK

2.5 mL 5 mL 7 mL 5 mL 10 mL

9.8 µM 24.2 µM 38.6 µM 12.5 µM 66.2 µM

±0.127 ±5.23 ±2.79 ±2.55 ±1.95

Concentration of Cr(VI) solution (μM)

70 60

GAC

GAC‐AAK

50 40 30 20 10 0

2.5

5 7 Volume of Cr(VI) solution (mL)

Figure 1. One hour static Cr(VI) reductions of GAC and GAC-AAK complex

Allowing the GAC-AAK to be exposed to Cr(VI) solutions for longer periods of time increases

130

The Manhattan Scientist, Series B, Volume 4 (2017)

Griffin

the amount of chromate that can be reduced. When the complex was reacted statically for 18 hours with potassium chromate solution volumes of 15, 20, and 25 mL, their Cr(VI) concentrations were all lower than that of the 1-hour trial samples. Both the 15 mL and 20 mL samples successfully reduced levels to be below EPA standards. Although the 25 mL sample was not below the required toxicity standard at 6.77 µM it was still a success in that its concentration was half that of the 5 mL sample tested for one hour even at 5 times the volume. Again this was compared to the ability of GAC alone to reduce Cr(VI), proving that the GAC-AAK is more effective in chromate reduction as seen in Table 2 and Fig. . Table 2. Concentrations of Cr(VI) using GAC and GAC-AAK after an 18-hour reaction time Volume of 200 µM Cr(VI)

Concentration of Cr(VI)

S.D.

GAC GAC GAC GAC-AAK GAC-AAK GAC-AAK

15 mL 20 mL 25 mL 15 mL 20 mL 25 mL

28.7 µM 54.9 µM 71.3 µM ≤1 µM ≤1 µM 6.8 µM

±4.14 ±2.26 ±1.09 ±0.092 ±0.035 ±1.39

Concentration of Cr(VI) solution (μM)

Matrerial

GAC‐AAK

GAC

25 20 15 10 5 0 ‐5

Volume of Cr(VI) solution (mL)

Figure 2. Eighteen hour static Cr(VI) reductions of GAC and GAC-AAK complex

These data suggest that the longer the chromate is in contact with GAC-AAK, the better able chromate is at traveling deeper into the GAC and therefore can more effectively interact with the AAK. Another conclusion drawn from this data is that the reduction of Cr(VI) does not solely occur on the surface of the GAC-AAK complex, but instead is dependent on how long (and as a result how deep) the chromate takes to travel within the pores of GAC itself. Support for this hypothesis is found when the reduction levels improve from the 1 hour to 18 hour trials. Another method that was employed for effective chromate reduction using the GAC-AAK complex was one that involved dynamic motion during the reaction. Each trial was conducted much like that of the static reduction trials, using a 0.5g sample of the AAK loaded GAC added

Griffin

The Manhattan Scientist, Series B, Volume 4 (2017)

131

Concentration of Cr(VI) Solution (μM)

to various volumes of 200 µM potassium chromate solution in 250 mL beakers placed on the TekTator Variable Rotator for particular times. The first variable that was tested was which stirring speed was optimal for chromate reduction. To do this GAC-AAK samples were combined with 30 mL of chromate solution and stirred for 30 minutes. At the end of the stirring period the EPA standard tests were conducted in order to determine the concentration of Cr(VI) in the water sample. It was found that the sample which stirred at 50 rpm was reduced to 98 µM, the 100 rpm sample 90 µM, and lastly the 150 rpm sample was best at 9 µM. Therefore, the Cr(VI) sample which stirred with GAC-AAK at 150 rpm was reduced by 95% (Fig. 3). 100 80 60 40 20 0 50

100

150

Stirring speed (rpm)

Figure 3. Optimal stirring speeds for Cr(VI) reduction

After the optimal stirring speed for Cr(VI) reduction was determined, the next variable that was focused on was how quickly the reaction was actually taking place. To be able to analyze this a kinetics experiment was conducted in which 5 mL aliquots were removed from a larger overall volume of chromate solution. A 1.5 g sample of GAC-AAK and 75 mL of the 200 µM K2 CrO4 were combined in a 400 mL beaker and placed on the Tek-Tator Variable Rotator. At different time intervals up until 30 minutes 5 mL aliquots were removed and tested following the same EPA standards as discussed in other sections of this paper. The results concluded that the most rapid reduction occurs within the first 3 minutes of the reaction. The average Cr(VI) concentration after 3 minutes was 70.3 µM (Table 3, Fig. 4). The initial velocity of the reaction is equated to approximately 11 mg of chromate being reduced within the first 3 minutes. At the end of 30 minutes stirring the solution was deemed safe to consume with its concentration being ≤ 1 µM. Due to the fact that the velocity of the reaction was so fast, the remainder of all stirring reduction trials were conducted with 30 minutes as the reaction time. Besides determining the kinetics of dynamic chromate reduction, the breakthrough volume (maximum amount of Cr(VI) solution that can be treated before exceeding EPA guidelines) after 30 minutes of stirring was also found. The procedure began with 0.5 g of GAC-AAK in 250 mL beakers in various amounts of 200 µM potassium chromate solution that were stirred at 150 rpm for 30 minutes. After the reaction time concluded an aliquot was then tested using the EPA guide-

132

The Manhattan Scientist, Series B, Volume 4 (2017)

Griffin

Concentration of solution (μM)

Table 3. Kinetic trials for the reduction of Cr(VI) using GAC-AAK complex Material

Time [minutes]

Concentration of Cr(VI) [µM]

S.D.

GAC-AAK GAC-AAK GAC-AAK GAC-AAK GAC-AAK GAC-AAK

3 6 9 12 15 30

70.3 38.9 20.2 9.0 3.4 ≤1

±2.26 ±3.12 ±1.88 ±0.509 ±0.127 ±1.42

80 60 40 20 0 3

Time (minutes) Figure 4. Kinetic trials for Cr(VI) reduction with GAC-AAK complex

lines mentioned in the earlier section and their concentrations calculated. The results concluded that after half an hour of stirring, a 0.5 g sample of the GAC-AAK complex was able to successfully reduce up to 30 mL of 200 µM chromate solution to below 1 µM. The official breakthrough value was 35mL having its concentration be 20.72 µM. This was then compared to the ability to reduce Cr(VI) for GAC alone, where the GAC-AAK complex reduces Cr(VI) about 20 times more effectively than solely GAC. These results are compiled in Table 4 and Fig. 5. Table 4. Concentrations of Cr(VI) using GAC and GAC-AAK after a 30-minute reaction time Material

Volume of 200µM Cr(VI)

Concentration of Cr(VI)

S.D.

GAC GAC-AAK GAC-AAK GAC-AAK GAC-AAK GAC-AAK

15 mL 15 mL 20 mL 25 mL 30 mL 35 mL

27.2 µM ≤1 µM ≤1 µM ≤1 µM ≤1 µM 20.7 µM

±0.778 ±0.615 ±0.064 ±0.092 ±0.674 ±0.375

Griffin

The Manhattan Scientist, Series B, Volume 4 (2017)

133

Concentration of solution (μM)

GAC

GAC‐AAK

20 25 30 Volume of Cr(VI) solution (mL)

Figure 5. Thirty-minute stirring Cr(VI) reductions of GAC and GAC-AAK complex

Conclusions Ascorbic acid ketal loaded onto GAC is more effective at reducing Cr(VI) than ascorbic acid bound to GAC. AAK is able to load 99.5 mg onto GAC and 8.6 mg of that is responsible for reducing chromium as compared to 67 mg of ascorbic acid loaded and 3.1 mg used in the reduction in Rosario’s research (Rosario, 2016). Eighteen hour static reduction conditions are much more efficient than 1 hour static reductions. Within an hour time frame GAC-AAK is not capable of reducing 5 mL of chromate to sub-micro molar while overnight 20 mL is successfully reduced to EPA standards. Dynamic Cr(VI) reduction is completed in 30 minutes with the most rapid reduction taking place within the first 6 minutes. The optimal stirring speed for the reduction of Cr(VI) is 150 rpm. Dynamic Cr(VI) reduction conditions are more effective than static. After 30 minutes of stirring 35 mL of chromate solution is able to be brought to safe consumable concentration levels.

Acknowledgments This work was funded by a gift from Kenneth G. Mann ’63. The author would like to thank Dr. Mann for his financial support, and Dr. John Regan for his mentorship, patience, and valuable advice.

References Dayan, A. D. and A. J. Paine. “Mechanisms of Cr toxicity, carcinogenicity and allergenicity: Review of the literature from 1985 to 2000.” Human and Experimental Toxicology 20, 439-451 (2001) EPA (Environmental Protection Agency). “Chromium in drinking water.” www.epa.gov/ dwstandardsregulations/chromium-drinking-water (2017)

134

The Manhattan Scientist, Series B, Volume 4 (2017)

Griffin

Owlad, M., M. K. Aroua, W. A. W. Daud, and S. Baroutian. “Removal of hexavalent Cr-contaminated water and wastewater: A review.” Springer Science + Business Media B.V. (2008) Rosario, Analisse. “Eliminating aqueous chromium(VI) with renewable technology.” The Manhattan Scientist, 3, 175-183 (2016) Weber, Jr., W. J. “Sorption from solution by porous carbon.” Principles and Applications of Water Chemistry. John Wiley and Sons, New York, NY 89-126 (1967) Wielinga, B., M. M. Mizuba, C. M. Hansel, and S. Fendorf. “Iron Promoted Reduction of chromate by dissimilatory iron-reducing bacteria.” Environ. Sci. Technol. 35, 522–527 (2001) Xu, Xiang-Rang, Hua-Bin Li, Xiao-Yan Li, and Ji-Dong Gu , “Reduction of hexavalent chromium by ascorbic acid in aqueous solutions,” Chemosphere 57, 609-613 (2004) Yurkow, E. J., J. Hong, S. Min, and S. Wang. “Photochemical reduction of hexavalent chromium in glycerol-containing solutions.” Environ. Pollut. 117, 1–3 (2002)

Synthesis, isolation, and characterization of cyclic organic Cr(VI) molecules Christopher Kim∗ Department of Chemistry and Biochemistry, Manhattan College Abstract. Methods for removal of chromium remain a prevalent field of research. Removal of chromium ions from water sources is feasible through adsorption onto granulated activated charcoal (GAC) [1]. A study had shown that the proposed formation of a chromate ester complex was more readily absorbed by GAC in comparison to chromium ions alone [2]. A recent research also suggested that formations of hypothetical cyclic organic chromium complexes contributed to higher absorption and removal rates [3]. Current methods to synthesis, isolate, and characterize the hypothetical cyclic organic chromium complexes have identified the narrow opportunities for success. Observations from this research provide promising hypothetical ideas for further study.

Introduction Hexavalent chromium, also known as Cr(VI), is a naturally occurring substance from chromeiron deposits. Hexavalent chromium and other chromium ions from industrial waste sites have also leaked into public water sources. Two prominent forms of hexavalent chromium anions in water 2− sources are chromate (CrO2− 4 ) and dichromate (Cr2 O7 ). Hexavalent chromium has been labeled as toxic due to its strongly oxidizing property. Exposure to chromium in chromium production workers has been documented to show an increase of lung cancer [4]. A study of villagers living near a chromium alloy plant with an intake of 0.57 mg/kg-day of Cr(VI) from a local water well elucidated the correlation between increased intake of hexavalent chromium with the following health effects: oral ulcers, diarrhea, abdominal pain, indigestion, vomiting, leukocytosis, and immature neutrophils [5]. The Safe Drinking Water Act requires the United States Environmental Protection Agency to set maximum contaminant levels in public drinking waters in an effort to deter adverse health effects. This maximum contaminant level for chromium has been established as 0.1 mg/L [6]. Removal of chromate through modified and unmodified activated carbon remains a prevalent field of research. A recent study demonstrated the facilitation of chromate removal through the use of activated charcoal and formation of cyclic organic-Cr(VI) carbonates and ureas from diols and diamines respectively [3]. The focus of this research was to synthesize, isolate, and characterize said novel molecules using various diols.

Materials and Methods The organic reagents used were ethylene glycol, trans-1,2-cyclohexanediol, cis-1,2-cyclohexanediol, pinacol, 1,2-dihydroxybenzene, cis-2-butene-1,4-diol, and 2-phenyl-1,3-propanediol. ∗

Research mentored by Richard Kirchner, Ph.D., and John Regan, Ph.D.

136

The Manhattan Scientist, Series B, Volume 4 (2017)

Kim

Figure 1. Scheme of diol reacting with chromic acid to form cyclic organic Cr(VI) molecule.

Chromium sources included potassium chromate and sodium chromate. Solvents used were ethyl alcohol, isopropanol, acetone, diethyl ether, n-hexane, methyl tert-butyl ether (MTBE), dichloromethane (DCM), and dimethylformamide (DMF). Hydrochloric acid (15%) was also used. All chemicals were ACS grade. Stock solutions of dilute and saturated potassium chromate (2.0×10−2 M and 140.4 M, respectively) were prepared. Organic reagents (ethylene glycol or trans-1,2-cyclohexanediol) were added to the solution maintaining a 1:1 or 2:1 molar ratio. Hydrochloric acid was added dropwise until pH of 3 was reached. The reaction solution containing the respective organic compound dissolved in the potassium chromate stock solution was stirred for 30 minutes and separated into five equal components. A series of diluted ethyl alcohol, isopropanol, or heptanol was prepared (1, 1/2, 1/4, 1/8, and 1/16) and added onto the separated solutions. Liquid-liquid extraction methods were conducted using MTBE or DCM as solvents. Products’ identities were assessed using UV/Visible, infrared, and NMR spectroscopy. Organic reagents (pinacol, trans-1,2-cyclohexanediol, cis-1,2-cyclohexanediol, 1,2-dihydroxybenzene, cis-2-butene-1,4-diol, and 2-phenyl-1,3-propanediol) and sodium chromate were dissolved in DMF or anhydrous acetone using a 4:1, 3:1, 2:1, or 1:1 molar ratio. Solutions were left for various periods of time (less than three days) while capped to prohibit introduction of water via the atmosphere. The pinacol reaction solution was centrifuged and the resultant supernatant liquid was decanted. The supernatant liquid was extracted and rid of DMF with a mixture of diethyl ether and n-hexane. The solution was then left to dry. The 1,2-dihydroxybenzene-reaction solution was subject to liquid-liquid extraction methods via a separatory funnel. Due to the high polarity of DMF and its subsequent difficulty for removal, DCM was first used to extract the organic layer. The organic layer was then washed several times with water and brine and then dried.

Results All methods conducted failed to produce any crystals of pure or impure products usable for analytical purposes. Furthermore, analytical and qualitative data for all experimental methods demonstrated that no expected hexavalent chromium compounds were formed.

Kim

The Manhattan Scientist, Series B, Volume 4 (2017)

137

Table 1. Starting organic reagents with respective cyclic formation with addition of chromate. Name

Structure

Hypothetical cyclic organic-Cr(VI) structure

Ethylene glycol

Pinacol

Trans-1,2-cyclohexanediol

Cis-1,2-cyclohexanediol

Cis-butene-1,4-diol

2-Phenyl-1,3-propanediol

1,2-Dihydroxybenzene

UV spectroscopic data of dilute and saturated potassium chromate solutions with organic reagents (ethylene glycol or trans-1,2-cyclohexanediol) treated with hydrochloric acid and added onto ethyl alcohol, isopropanol, or heptanol demonstrated peaks at 350 nm, indicating presence of Cr(VI). Infrared and NMR spectroscopic data of product solutions demonstrated exact properties of starter organic reagents and no dissimilarities. We were not able to isolate the products of the organic reagents (pinacol, trans-1,2-cyclohexanediol, 1,2-dihydroxybenzene, cis-2-butene-1,4-diol, and 2-phenyl-1,3-propanediol) brown tacky

138

The Manhattan Scientist, Series B, Volume 4 (2017)

Kim

material with imbedded, clear crystals were found. Infrared and NMR data analysis indicated that the clear crystals were pinacol. The experiment using 1,2-dihydroxybenzene was also promising. The reaction solution proceeded from yellow to orange and then from orange to purple. After the work-up to remove DMF was completed and the solution was left to dry, a sludgy, tarry substance was formed instead of crystals.

Discussion The purpose behind treating the potassium chromate solutions and dissolved organic reagents with hydrochloric acid was to form chromic acid and potassium chloride salt. K2 CrO4 + 2 HCl â&#x2020;&#x2019; H2 CrO4 + 2 KCl Figure 2. Chemical equation of potassium chromate and hydrochloric acid.

The chromic acid would then be the oxidizing agent for the hypothetical cyclic organic chromium compounds.

Figure 3. Proposed reaction between 1,2-cyclohexanediol and chromic acid to form hypothetical cyclic organic Cr(VI) molecule.

The cyclic organic chromium compounds could then be isolated using liquid-liquid extraction methods while the potassium chloride salts could easily be removed. However, the addition of hydrochloric acid was instead forming dichromate since chromate and dichromate are in equilibrium in an aqueous solution. Primary and secondary alcohols are typically oxidized to aldehydes/carboxylic acids or ketones respectively. However, vicinal diols typically undergo an intermediary cyclic ester formation and cleaved to form two separate carbonyl compounds when periodic acid or lead tetraacetate are used. The focus point of this research project was to use an oxidizing agent, which would be chromic acid in this context, on promising diols that would not undergo vicinal cleavage and to isolate a stable and crystalline intermediary. When dilute chromate solutions with organic reagents treated with hydrochloric acid did not produce any viable crystals due to the formation of more dichromate ions and equilibrium rates not favoring product formation, the idea to shift equilibrium rates to favor product formation and not form dichromate ions was incorporated. Such a shift required an increase in the concentration of potassium chromate and a saturated potassium chromate solution to be prepared. However, when the organic reagents were added onto the saturated potassium chromate solution without addition

Kim

The Manhattan Scientist, Series B, Volume 4 (2017)

139

of hydrochloric acid, only the potassium chromate recrystallized. Infrared and NMR analyses failed to indicate any conclusive results of desirable products being formed. Literature research of synthesizing cyclic organic chromium compounds yielded feasible experimental methods [7, 8]. Such methods included dissolving sodium dichromate and organic starter reagents in polar aprotic solvents, specifically DMF, on a micro-scale level [7, 8]. One published article demonstrated synthesizable and viable crystals for analytical purposes using cis-1,2cyclohexanediol [7]. The article also showed useful qualitative information; the reaction solution with cis-1,2-cyclohexanediol, disodium dichromate, and DMF turned from yellow to orange and finally to dark green, and at three days, green, needle-like crystals were formed [7]. This reaction was repeated to verify feasibility with sodium chromate instead of sodium dichromate. The reaction was successful in producing the same green, needle-like crystals. Other organic reagents were chosen due to their hypothetical viabilities to form stable chromium compounds. Such properties that were thought to facilitate stability and crystal formation were bulky aromatic rings, stereochemistry, no free rotations of hydroxyl groups, and energetically favored products, as seen in the cis-1,2-cyclohexanediol. However, all products formed were not isolatable in crystalline form.

Conclusion We were not able to isolate the hypothetical cyclic organic chromium compounds, except the cyclic organic chromium compound using cis-1,2-cyclohexanediol. However, we have identified the properties akin to cis-1,2-cyclohexanediol that suggests viability for future research.

Acknowledgments This work was funded by the Camille and Henry Dreyfus Foundation Senior Scientists Mentor Program and by the Michael J. ’58 and Aimee Rusinko Kakos Endowed Chair in Science. The author would like to thank Dr. Richard Kirchner and Dr. John Regan for providing unlimited guidance, support, and advice throughout the entirety of this research project.

References [1] Frank DeSilva. “Activated Carbon Filtration.” Water Quality Products Magazine (January 2000) Web. [2] Nakajima, Akira, and Yoshinari Baba. “Mechanism of Hexavalent Chromium Adsorption by Persimmon Tannin Gel.” Water Research 38.12 (2004): 2859-864. Web. [3] John Regan, Douglas Huntington, and Joseph F. Capitani. “Bidentate Reagents Form Cyclic Organic-Cr(VI) Molecules For Aiding in the Removal of Cr(VI) from Water: Density Functional Theory and Experimental Results” Structural Chemistry (2017). Web. [4] H.J. Gibb, P.S. Lee, P.F. Pinsky, and B.C. Rooney “Lung Cancer Among Workers in Chromium Chemical Production.” American Journal of Industrial Medicine (AJIM) 38.2 (July 7, 2000): 115-126. Web.

140

The Manhattan Scientist, Series B, Volume 4 (2017)

Kim

[5] US Environmental Protection Agency. (1998). Toxicological Review of Hexavalent Chromium: In Support of Summary Information on the Integrated Risk Information System (IRIS). Washington D.C. [6] “Chromium In Drinking Water.” US Environmental Protection Agency, 2017, https://www. epa.gov/dwstandardsregulations/chromium-drinking-water. [7] Ruben Bartholom¨aus, Klaus Harms, Aviva Levina, and Peter A. Lay. “Synthesis and Characterization of a Chromium(V) cis-1,2-Cyclohexanediolato Complex: A Model of Reactive Intermediates in Chromium-Induced Cancers.” Inorganic Chemistry, 51(21), 11238-11240 (2012). Web. [8] Swetlana Gez, Robert Luxenhofer, Aviva Levina, Rachel Codd, Peter A. Lay. “Chromium(V) Complexes of Hydroxamic Acids: Formation, Structures, and Reactivities.” Inorganic Chemistry, 44(8), 2934-2943 (2005). Web.

Synthesis of antimony sulfide (Sb2 S3 ) in aqueous solution James L. Ksander∗ Department of Chemistry and Biochemistry, Manhattan College Abstract. As fossil fuel production peaks, the world’s population continues to grow and demand more energy. Humanity must then turn to renewable energy to avoid an imminent energy crisis. This work hopes to provide more stable, efficient, and cost effective solar panels by making Antimony Sulfide nanowires, a substance shown to be more efficient and less toxic than its lead iodide predecessors, in a simple aqueous solution at room temperature. The goal of this research project was to devise a process for producing Antimony Sulfide nanowires from solution using a polycarbonate template as an interface. After producing the nanowires in solution the hope was to construct a solar cell and test the photoelectric properties of the cell to determine its efficiency to other varieties of solar cells. During this research only the production of Antimony Sulfide nanowires in aqueous solution was completed and verified.

Introduction Antimony sulfide, called stibnite in its mineral form, is similar to other heavy metal compounds used in solar cells such as lead iodide but has certain characteristics that make it a more enticing option. First among these aspects is its incredibly low solubility, Ksp = 1.6 × 10−93 , versus lead iodide, Ksp = 4.4 × 10−9 , which means that the possibility of contamination from broken solar arrays using antimony sulfide is much lower than with lead iodide [1]. Antimony sulfide based solar cells have been shown to have efficiencies between five and six percent which is lower than other perovskite cells but less research has been done on antimony sulfide cells and the perovskite cells are much less stable [2]. In addition to use in solar cells, antimony sulfide nanowires have also been shown to have use as an anode in lithium ion batteries further demonstrating its unique ability to use and store energy [3]. To facilitate the formation of nanowires in a uniform and consistent manner experiments were run using polycarbonate and AlO3 templates with pores of varying size. The hope in using these templates was that the reaction between reagent solutions would occur in the pores, which could then either be annealed to crystal state in the template or the template could be removed and the wires then annealed. The polycarbonate templates would be removed using methyl chloride while the AlO3 templates could be removed using acid. Polycarbonate templates were the first choice in this research because of their resistance to dissociation in acid-base conditions as well as their flexible physical characteristics. To synthesize antimony sulfide in aqueous solution most experiments were run using the combination of antimony potassium tartrate, Sb2 K2 (C4 H2 O6 )2 , and sodium sulfide, Na2 S, solutions. Both solutions were decently soluble in water in concentrations from 0.01M - 0.1M. Antimony acetate, Sb(CH3 CO2 )3 , and sodium thiosulfate, Na2 S2 O3 , solutions were also used. ∗

Research mentored by Alexander Santulli, Ph.D.

142

The Manhattan Scientist, Series B, Volume 4 (2017)

Ksander

A tube furnace was used to anneal amorphous samples to their crystalline structure in atmospheres of both regular air and nitrogen gas. Antimony sulfide demonstrates very different physical properties. Most apparent is the change from bright orange while amorphous to dark grey when crystalline, however, these physical properties are very similar to those of antimony oxide. To determine purity and crystal structure of samples collected X-ray diffraction was used, the results of which were compared to the results of samples with previously confirmed composition. Scanning electron microscopes were also used to confirm the presence of nanowires in both amorphous and annealed samples.

Methods and Materials

The first attempts to produce antimony sulfide nanowires were conducted in a 200 mL U-tube with 0.1M solutions of antimony tartrate and sodium sulfide. 200 nm polycarbonate templates were used throughout all U-tube experiments as the reaction interface, and the experiment was left to react for twenty hours overnight. Once over, the template was removed from the U-tube ground down on a grinding stone to remove excess precipitate and then dissolved in methylene chloride. A centrifuge was then used to separate the precipitate from the methylene chloride from which the precipitate was pipetted out onto a glass slide. Both light microscopes and SEMs were used to confirm the presence of nanowires in the precipitate. Attempts were made to duplicate the results using aluminum oxide 200 nm templates under the same conditions. Due to either osmotic pressure or reaction with the solutions all aluminum oxide templates could not survive the experiment due to breaking apart in the U-tube. The solutions would then immediately react with each other to produce a dark orange to brown solution. Precipitates from these reactions were vacuum filtered then annealed in tube furnace for testing in X-ray diffractometer. Differing concentrations of reactant solutions were run, ranging from 0.01M to 0.1M for both solutions. In order to prevent the osmotic pressure from pushing solutes too strongly to either side, uneven concentrations and uneven volumes of reagent solutions were used. The results of these changes did not appear to affect the overall product of the reaction. Both bulk and template sample of product from the reactions would be annealed in a tube furnace from 100◦ C - 500◦ C from between 5−60 minutes. To reach these temperatures the furnace would heat up at roughly five to ten degrees per minute. After annealing the samples would then be collected in a test tube where they were prepared for the X-ray diffractometer by crushing the samples into a powder that was subsequently suspended in methylene chloride for ease of transfer to a silicone disk that would then be inserted into the XRD. The XRD would be run from 10◦ and 60◦ theta with between three and eight seconds for every hundredth of a degree. Such tests would range from three to twelve hours. After initial test runs revealed the presence of antimony oxide efforts were made to restrict the oxidizing elements of the experiment. Sodium thiosulfate’s use was attempted in a bulk reaction but was quickly ruled out due to solubility, reaction rate, and product issues. Antimony acetate was

Ksander

The Manhattan Scientist, Series B, Volume 4 (2017)

143

similarly examined for use but was ruled out due to being insoluble in water. All solutions after the initial experiments were run using pure water deaerated by nitrogen gas aspiration for thirty minutes. All further annealing was done under an atmosphere of nitrogen gas. In order to pursue parameters where a dark brown precipitate was coming out of solution a bulk sample was prepared using three to one concentrations of sodium sulfide to antimony tartrate. The precipitate was then vacuum filtered and annealed in the tube furnace for thirty minutes at 300◦ C, after which XRD data indicated the presence of antimony hydroxide, Sb(OH)3 . Evidence of antimony hydroxide complexes, Sb(OH)4 − , forming and subsequently pushing the equilibrium towards the dissolution of antimony hydroxide also presented itself during these experiments. After solutions of sodium sulfide showed the ability to dissolve antimony hydroxide the solutions were tested for their pH, which ranged from ten to twelve. “Used” solutions of sodium sulfide, named so because they were collected after use from previous U-tube experiments, were then subject to several tests under alkaline and acidic conditions. No change to the solution was shown when alkaline conditions were increased. When concentrated hydrochloric acid was added to the solution a bright orange solid immediately precipitated out of the solution. Concentrated HCl was added dropwise until no more precipitate formed. The precipitate was then collected via vacuum filtration and annealed under nitrogen atmosphere for thirty minutes at 300◦ C. After annealing the sample, now a dark grey color, was transferred to the XRD the results of which confirmed the sample as very pure antimony sulfide.

Results and Discussion

This research began with the initial hope of being able to incorporate antimony sulfide nanowires into a solar cell but that goal proved quite lofty as even producing pure antimony sulfide proved elusive. The original experiments conducted with simple solutions of antimony tartrate and sodium sulfide initially showed promise as an orange precipitate was collected within the polycarbonate template and the subsequent annealing would turn the sample a dark grey or black color that was indicative of crystallized antimony sulfide. After subjecting the samples to XRD tests the composition was shown to be almost pure antimony oxide, as shown by the most prominent peak at 2θ = 27.7 in Fig. 1. Measures were then taken to remove oxidizing elements from the experiments we assumed we were moving more towards antimony sulfide. Through more or less trial and error, in response to changes in the XRD data received from samples, we were able to zero in on several peaks in the 20-30◦ 2θ region which began as merely a slight bump on the data graphs as shown in Fig. 2. The peaks between 20 and 30 became clearer as, in an attempt to remove as much moisture from the samples as possible, acetone was used to dry the samples before annealing. These peaks, shown most clearly in Fig. 3 as a grouping of peaks around 2θ = 25, were consistent with antimony hydroxide which increased the attention we gave to the dark brown precipitate that would be found alongside the orange precipitate of antimony sulfide when running previous experiments. Once the solutions were tested for their pH and sodium sulfide solutions were found

144

The Manhattan Scientist, Series B, Volume 4 (2017)

Figure 1. XRD data of a sample made of predominantly Sb2 O3

Figure 2. XRD data of sample dried with acetone beginning to show anomalous peaks between 20◦ and 30◦ 2θ

Figure 3. XRD data showing the anomalous peaks between 20◦ and 30◦ 2θ indicative of Sb(OH)3

Ksander

The Manhattan Scientist, Series B, Volume 4 (2017)

145

to be extremely alkaline the chemical equations in Fig. 4 were theorized. Sb3+ (aq) + 3OH− (aq) Sb(OH)3(s) Sb(OH)3(s) + OH− (aq) Sb(OH)4 − (aq) H2 O(l) + Na2 S(s) HS− (aq) + 2Na+ (aq) +S2+ (aq) + OH− (aq) HS− (aq) + H+ (aq) H2 S(aq) 2H2+ (aq) + S2− (aq) 3H2 S(aq) + 2Sb3+ (aq) Sb2 S3(s) + 6H+ (aq) Figure 4. Equilibria theorized to show the species of solutes within the experiments performed.

The creation of antimony hydroxide and subsequent complex ion formation which dissolves the antimony hydroxide back into the solution were the primary factors behind our inability to produce antimony sulfide in aqueous solution. The XRD data shown in Fig. 5 show the greatly reduced peak at 27.7 as well as many new peaks the most prominent of which appear at 25◦ and 29◦ clearly demonstrating the purity of the antimony sulfide sample.

Figure 5. XRD data of the bulk sample of antimony sulfide (Sb2 S3 )

Conclusion and Future Work Though the goals of this research might have seemed somewhat ambitious for the short amount of lab time available these goals of producing antimony sulfide nanowires within a polycarbonate template from aqueous solution now seem quite within reach. The only further issues that were unveiled during this research were whether or not to anneal the nanowires within the template and perfecting the optimal conditions under which antimony sulfide nanowires form. Future research would decide whether the antimony sulfide can be annealed at a low enough temperature that the template does not burn off and oxidize the nanowires or whether the nanowires can maintain their uniform nature, gained from being produced in the template pores, after the template has been

146

The Manhattan Scientist, Series B, Volume 4 (2017)

Ksander

dissolved and the wires themselves annealed. Once that is accomplished then transferring the nanowires and polycarbonate template to a solar cell would be more of a practical assembly issue than a chemical one.

Acknowledgements This work was funded by a gift from Kenneth G. Mann ’63, and by the Michael J. ’58 and Aimee Rusinko Kakos Endowed Chair in Science. The author would like to thank Dr. Alexander Santulli for his help and mentorship.

References [1] John A. Olmsted, Gregory M. Williams, and Robert C. Burk. Chemistry. John Wiley & Sons Canada, Ltd., 2016. [2] Karl C. G¨odel, Yong Chan Choi, Bart Roose, Aditya Sadhanala, Henry J. Snaith, Sang Il Seok, Ullrich Steiner and Sandeep K. Pathakad, “Efficient Room Temperature Aqueous Sb2S3 Synthesis for Inorganic-Organic Sensitized Solar Cells with 5.1% Efficiencies.” Chemical Communications, The Royal Society of Chemistry, 14 Apr. 2015, pubs.rsc.org/en/Content /ArticleLanding/2015/CC/C5CC01966D#! [3] Denis Y. W. Yu, Harry E. Hoster and Sudip K. Batabyal “Bulk antimony sulfide with excellent cycle stability as next-generation anode for lithium-ion batteries.” Nature News, Nature Publishing Group, 2 Apr. 2014, www.nature.com/articles/srep04562.

Determination of the crystal structure of aluminosilicate zeolite ZSM-18 Daisuke Kuroshima∗ Department of Chemistry and Biochemistry, Manhattan College Abstract. ZSM-18 the first aluminosilicate zeolite reported that has three-membered rings. The focus of this research was to use the sophisticated crystallographic computer program SuperFlip to determine the structure of ZSM-18 from synchrotron X-ray powder diffraction data. The graphics program ATOMS is used to visualize and determine bond connections between ATOMS generated in a SuperFlip solution. Various incomplete solutions were derived and are discussed.

Introduction Zeolites are hydrated aluminosilicate minerals that are microporous. In nature pores commonly contain water. When the zeolite is heated, loosely bonded water boils out from it. This is the origin of its name, meaning ”boiling-stone” in Greek. Zeolites are used for molecular sieving, ion exchange beds, and purification of water. ZSM-18 is one of hundreds of zeolites that have been synthesized in laboratories [1]. Zeolites are synthesized from heating aqueous mixtures of alkali metal oxide, aluminum oxide, silicon oxide, and often an organic template molecule that determines pore size and shape. ZSM-18 was first synthesized at ExxonMobil. The published structure (Fig. 1) was determined by Model Building but never confirmed by an x-ray study.

Figure 1. Framework toplogy of literature model [1]. Lines connect Si atoms. Oxygen atoms, approximately located in the middle of each line, are not shown. Pores are defined by the number of Si atoms that connect to make up the pore. The large 12-ring pore is visible in the center. Three-, four- and five-rings are also visible. The seven-ring pore is not visible.

∗

Research mentored by Richard Kirchner, Ph.D.

148

The Manhattan Scientist, Series B, Volume 4 (2017)

Kuroshima

The published structure of ZSM-18 was controversial because it contains the first example of a zeolite with three-membered-rings as well as other odd-numbered rings. It also has very long and questionable Si-O-Si connections between layers. Alternative models for the structure of ZSM-18 have been suggested. The goal of the present research is to completely determine or verify the correct structure of ZSM–18 by an x-ray crystal structure analysis refinement.

Background on ZSM-18 ZSM-18 stands for Zeolite-Socony-Mobil preparation 18. Socony is an older name for the Standard Oil Company of New York. Julius Ciric at Mobil Oil Company laboratory synthesized ZSM-18 [2] but the structure was not confirmed by an X-ray study. Synchrotron X-ray powder data was collected by J. M Bennett. Absorption experiments show that the largest pore in ZSM-18 is a 12-membered ring [3]. Since silicon and oxygen ATOMS usually have tetrahedral bond angles it was unusual that the proposed model contained Si in a three-membered ring (Fig. 2). In addition, the literature model has one extremely long Si-O-Si connection that might have resulted from an incorrect space group.

Figure 2. A three-ring.Tetrahedral silicon atoms (white) are connected by oxygen atoms (red). The O-Si-O bonds are approximately tetrahedral (109.5◦ ) but the Si-O-Si bonds (approximately 131◦ ) are greatly distorted from tetrahedral values.

An organic template Triquat, N[CH2 CH2 N(CH3 )3 ]3 [4], was used in the synthesis of ZSM18. The zeolite mineral crystalizes around the organic Triquat template (Fig. 3). However the crystalline product was not calcined to remove the template from the zeolite large cavity. Previous work in our group determined from an analysis of the electron density in the large pore that the Triquat molecule is disordered. However, the literature model containing a correctly disordered Triquat in the large cavity could not be refined, suggesting something is wrong or unknown about the ZSM-18 structure.

Kuroshima

The Manhattan Scientist, Series B, Volume 4 (2017)

149

Figure 3. Structure of the Triquat template molecule. Each arm is Nâ&#x20AC;&#x201C;CH2 CH2 â&#x20AC;&#x201C;NH3 + . Carbon and hydrogen atoms are not shown for clarity.

Experimental Methods Synchrotron X-ray powder diffraction data (Fig. 4) was provided by J. M. Bennett. It was input into SuperFlip [5], currently the most powerful crystallographic program used to determine crystal structures. For powder data the structure is correct when its calculated powder pattern matches the experimental one.

Figure 4. X-ray powder diffraction data for SUZ-9 [6]

SuperFlip is a computer program for application of the charge-flipping algorithm for structure solution of crystal structures from diffraction data [5]. Synchrotron X-ray powder data was put into S UPER F LIP initially using the cell dimensions and space group, P63 /m, from the literature model of ZSM-18 [1]. SuperFlip calculates 10 solutions each with their own space group. These solutions were then input into the graphics program ATOMS [7] to display all the atom positions. The graphic program ATOMS allows manipulation of the result from SuperFlip. False atoms can be deleted, the symmetry can be changed, and distance limits can be set to reveal bonds between atoms. The color and size of each atom can be changed to help identify structural features.

150

The Manhattan Scientist, Series B, Volume 4 (2017)

Kuroshima

Rotating the display in three dimensions also helps reveal the possible structure. ATOMS is used to modify coordinates for all atoms and better visualize and improve the proposed structure.

Results In 60 trials, SuperFlip never suggested space group P63 /m. Most commonly, SuperFlip suggested PÂŻ6 and P11m. PÂŻ6 is somewhat similar to P63 /m, but P11m has much less symmetry. Presumably P11m shows up because ZSM-18 has the organic template Triquat in its pores. Triquat has symmetry lower than P63 /m. Therefore, a symmetry lower than P63 /m is possible. However when P63 /m is used in the ATOMS program (even when SuperFlip suggested P11m), the results were more organized. Typical results (Figs. 5, 6 and 7) show mainly 3- and 4-membered rings.

Figure 5. SuperFlip result displayed in ATOMS. Silicon atoms are in green and oxygen atoms in red.

Figure 6. Second best SuperFlip result displayed by ATOMS. Silicon atoms are in green and oxygen atoms in red.

Kuroshima

The Manhattan Scientist, Series B, Volume 4 (2017)

151

Figure 7. Another result from SuperFlip displayed by ATOMS. Silicon atoms are in green and oxygen atoms in red.

The small picture on the lower right side of Fig. 5 suggests a structure made up from Si in 4-membered rings (shown above as three connecting squares). Behind the 4-rings are 3-membered rings (shown at far right). Because silicon is usually tetrahedral its presence in 3-rings is unusual. This model is one of the best SuperFlip solutions. Since most atoms are shown as oxygen, I interpret these atoms to be silicon and consider the green atoms as error. This result has two huge 12 membered rings and 3-, 4-, and 6-member rings. This result also has disorganized squares. This solution has two 12-membered rings that overlap, which is not reasonable since 12-membered ring pores must be open to accommodate Triquat. The solution in Fig. 7 has 3-, 4-, and 6-membered rings, but does not have 12-membered rings. Most oxygen atoms are between two silicon atoms, as expected. This solution also shows that the framework contains cubes.

Conclusion

The structure of the ZSM-18 has not yet been completely revealed from ab initio calculations in SuperFlip. More effort extracting a complete framework topology (structure) from SuperFlip results is needed.

Acknowledgments

This project was funded by the Michael J. â&#x20AC;&#x2122;58 and Aimee Rusinko Kakos Endowed Chair in Science. The author expresses special thanks to Dr. Richard Kirchner, for acting as his faculty advisor for this research.

References

[1] Ch. Baerlocher and L. B. McCusker, Database of Zeolite Strucutres: http://www.iza-structure. org/databases

152

[2] [3] [4] [5] [6] [7]

The Manhattan Scientist, Series B, Volume 4 (2017)

J. Ciric, US. Pat. 3,950,496 A (April 13, 1976) S. L. Lawton, W. J. Rohrbaugh, Science. 247(4948):1319-22. (March 1990) A. W. Burton and H. B. Vroman, US. Pat. 9,452,424(September 27, 2016) L. Palatinus and G. Chapuis, SuperFlip, J. Appl. Cryst. 40, 786-790 (2007) K. D. Schmitt and D. J. Kennedy, Zeolites, 14:635-642 (1994) E. Dowty, ATOMS, Shape Software, 521 Hidden Valley Road, Kingsport TN 37663

Kuroshima

Magnetic/TiO2 core-shell nanoparticles as photocatalysts for water purification Hannah Mabeyâ&#x2C6;&#x2014; Department of Chemistry and Biochemistry, Manhattan College Abstract. The goal of this research was to synthesize various core-shell nanoparticles through decomposition and precipitation reactions. The synthesized nanoparticles were used as photocatalysts. The synthesized nanoparticles have a magnetic core and are coated in a TiO2 shell. The magnetic core can serve to minimize toxicity concerns because the nanoparticles can easily be removed from solution by use of a strong permanent magnet. The magnetic core could also allow for easy reuse of the catalyst. By exposing the synthesized magnetic nanoparticles to an alternating current magnetic field, we plan to see if hyperthermic effects are induced on microorganisms, which would result in the disinfection of contaminated water. Through a generalized coating method that was developed, the magnetic core of the nanoparticles was coated in a TiO2 shell. When TiO2 is exposed to UV light (365 nm) reactive oxygen species (ROS) are formed and these ROS cause oxidative damage to contaminants in water. The advanced oxidation process (AOP) relies on the formation of ROS and the reaction of these ROS with pollutants as a method to purify water. The efficacy of our synthesized nanoparticles was tested by how methylene blue (MB) and E.coli were degraded when over time in the presences of the synthesized nanoparticles, exposed to UV light and when exposed to an induction heater.

Introduction Water pollution is a growing environmental concern, millions of people do not have direct access to clean water [1]. 1.2 trillion gallons of untreated sewage and industrial waste is dumped into U.S. waters annually [2]. The disposal of untreated wastewater is the main cause of water pollution. The pollutants in the wastewater from industrial and domestic sources are predominantly pathogens and organic chemicals [3, 4]. In order to minimize pollution concerns, these contaminants must be removed from the effluents before being discharged into water bodies. There are several methods in use to purify water but can be high in cost and generate harmful byproducts [4]. Efficient, environmentally friendly, and cost effective methods to purify water would greatly improve this situation. The advanced oxidation process is an alternative method used to purify water. Oxidation is the transfer of one or more electrons from the electron donor to the electron acceptor, which causes chemical transformation of both the electron donor and the electron acceptor. This can sometimes result in chemical species having an odd number of valence electrons, referred to as radicals [5]. Because these species have an unpaired electron they are very unstable and therefore very reactive. The advanced oxidation process involves two stages of oxidation, the formation of reactive oxygen species (ROS), particularly â&#x20AC;˘OH, and the reaction of these ROS with organic pollutants in water, which oxidize and degrade pollutants. The benefit of using the advanced oxidation process is that it is environmentally friendly and does not generate unwanted byproducts [5]. â&#x2C6;&#x2014;

Research mentored by Hossain Azam, Ph.D., and Alexander Santulli, Ph.D.

154

The Manhattan Scientist, Series B, Volume 4 (2017)

Mabey

TiO2 is a semiconductor; this means that it has a filled valence band and an empty conduction band which allows TiO2 to have a small band gap. When TiO2 is illuminated by UV light (365 nm) the photon with energy that is greater than or equal to the band gap excites valence band electrons to the conduction band. This results in positive valence band holes and excited-state conduction band electrons. The resulting positive valence band holes and excited conduction band electrons initiate •OH formation [5], shown in Fig. 1 below. Oxidation intermediates

Excitation

VB h+

Organic pollutants

Recombination

Energy

− CB e

Mineralization products



Photo‐reduction

hν



Photo‐oxidation OH; R Oxidation intermediates

Mineralization products Figure 1. TiO2 mechanism of degradation of organic pollutants

Materials and Methods Synthesis of Fe3 O4 Nanoparticles: Ferric chloride and ferrous chloride were mixed in a 2:1 molar ratio. This solution was heated to 50◦ C for 10 minutes. Ammonia solution was then added and black iron oxide particles precipitated from solution. The particles were separated from solution by use of a strong permanent magnet. The particles were washed several times to remove any surface impurities and left in an oven at 100◦ C to dry overnight [6]. Overall reaction: Fe2+ + 2Fe3+ + 8OH− → Fe3 O4 + 4H2 O Synthesis of nickel and cobalt nanoparticles 4.55 g of Na2 C2 O4 was added to 80 mL of water and heated for about 10 min until dissolved. In a separate beaker 4.05 g of metal chloride hexahydrate was dissolved in 10 mL of water. The solution of metal salt was added slowly to the oxalate solution. For nickel, light green precipitate

Mabey

The Manhattan Scientist, Series B, Volume 4 (2017)

155

formed and for cobalt, light pink precipitate formed. The precipitate was collected by filtration then dried by heating at 100â&#x2014;Ś C for 30 minutes. The metal oxalate was then annealed at 500â&#x2014;Ś C for 30 min [7]. Coating nanoparticles A method was generalized to coat the synthesized magnetic nanoparticles. a.) TiO2 coating A mixture of 10 mL hexane and 2 mL tetrabutyl orthotitanate (TBOT) was made. 600 mg of Fe3 O4 nanoparticles were added to the TBOT and hexane mixture and sonicated for 5 minutes. The Fe3 O4 nanoparticles were left to sit in the solution for 20-30 minutes. The supernatant was decanted; the nanoparticles were washed 3-4 times with hexane. The coated core-shell nanoparticles were collected using a strong permanent magnet. The nanoparticles were then annealed for 30 minutes at 500â&#x2014;Ś C. b.) SiO2 coating Same procedure was followed, with the substitution of tetraethyl orthosilicate (TEOS) Methylene Blue (MB) degradation A 0.025 mM MB solution was made. A 100 mL MB solution was added to each of two evaporating dishes. 5 mg of the nanocatalyst being tested were sonicated in one evaporating dish containing the MB solution; the other evaporating dish with MB solution was left as the control. The samples were then illuminated by UV light (365 nm). UV/Visible spectroscopy was used to analyze how the concentration of MB was degraded over time to determine the efficacy of the synthesized photocatalysts. E. coli disinfection A 102 CFU/mL stock solution of E. coli was made. 5 mg of the synthesized photocatalyst being tested was dispersed in 100 mL of the E. coli stock solution. 100 mL of the E. coli stock solution was used as the control. 50 mL of the control solution and 50 mL of the solution containing the synthesized core-shell nanoparticles were added to two separate petri dishes and illuminated by UV light. Samples were taken at different time points the samples were diluted and filtered through a membrane. The filtered samples were placed on agar plates and incubated for 24 hours.

Results and Discussion From analyzing the SEM of the synthesized nanoparticles, it can be seen in Fig. 2 that the uncoated Fe3 O4 nanoparticles are uniform in structure and are on the nanoscale. The coated nanoparticles in Fig. 3 are also on the nanoscale. It can be seen that there is not a separation of different species within the image, which shows a uniform spherical structure throughout, and this indicates that there is a uniform coating over the magnetic nanoparticles. Fig. 4 shows that the SiO2 coating is on the nanoparticles but is thicker than the TiO2 coating in Fig. 3. In Fig. 5, an XRD was taken

156

The Manhattan Scientist, Series B, Volume 4 (2017)

Mabey

of the synthesized Fe3 O4 nanoparticles in order to verify that they were in fact Fe3 O4 not another species such as Fe2 O3 or iron by itself. These peaks do in fact confirm that only Fe3 O4 is present.

Figure 2. Scanning electron microscope images of the synthesized Fe3 O4 nanoparticles

Figure 3. Scanning electron microscope images of the synthesized TiO2 coated Fe3 O4 nanoparticles

200 180 160 140 120 100 20

2Î¸ (degrees) Figure 4. Scanning electron microscope image of the synthesized SiO2 coated Fe3 O4 nanoparticles

Figure 5. X-Ray diffraction of the synthesized Fe3 O4 nanoparticles

The measurements also show that the magnetic nanoparticles were successfully coated in a shell. When looking at the mean length of the Fe3 O4 nanoparticles, taking into account the standard deviation, it is smaller than the mean length of both the TiO2 coated and SiO2 coated Fe3 O4 nanoparticles (Table 1). This proves what was determined from the SEM images, that the magnetic nanoparticles were successfully coated in a shell, confirming that core-shell nanoparticles were synthesized.

Mabey

The Manhattan Scientist, Series B, Volume 4 (2017)

157

Table 1. Measurements taken of the synthesized nanoparticles using ImageJ software Nanoparticle

Mean Length (nm)

Standard Deviation (nm)

Fe3 O4 TiO2 coated Fe3 O4 SiO2 coated Fe3 O4

10.991 16.883 23.00

1.871 2.493 4.00

Fig. 6 shows the results from experiments done to test the efficacy of the synthesized nanoparticles on MB degradation when illuminated by 365 nm UV light. This data shows that TiO2 nanoparticles were the most efficient in the photo degradation of MB, followed by the synthesized TiO2 coated Fe3 O4 nanoparticles. It was found that SiO2 coated Fe3 O4 nanoparticles had no Fe O TiO coated Fe O SiO coated Fe O TiO Control photocatalyst effect. Column1 2

Slope of Trendline

-0.0002

-0.0014

-0.0002

-0.0006

-0.0002

I / I0

1.0

0.9 Control TiO2 TiO2

0.8

Fe3O4 Fe3O4 TiO2 coated Fe3O4 TiO2 coated Fe3O4 SiO2 coated Fe3O4 SiO2 coated Fe3O4

0.7 0

100

150

200

Time (min) Figure 6. MB degradation in the presence of different nanoparticles

Fig. 7 shows the disinfection of E. coli from contaminated water in the presence of various nanoparticles. Similar results were seen in the disinfection of E. coli and the degradation of MB, where TiO2 nanoparticles showed the highest efficacy as a photocatalyst. Our experimental results show that we successfully synthesized magnetic/TiO2 core-shell nanoparticles. The synthesized nanoparticles showed good results in the degradation of common pollutants in wastewater through the advanced oxidation process. Preliminary results using an induction heater with the synthesized magnetic/TiO2 core-shell nanoparticles showed a temperature increase of up to 4.5â&#x2014;Ś C in 3 hours. These results are promising and we plan to look further into how an induction heater can cause hyperthermic effects to microorganisms. We found that SiO2 -coated Fe3 O4 nanoparticles showed no photocatalyst ability.

0 0 0 0

500000 300000 100000 0

30000000000

0 0 0 0

22000000000

1.08696E‐05 6.52174E‐06 2.17391E‐06 0

The Manhattan Scientist, Series B, Volume 4 (2017)

Log (N/N0)

158

70 85 110 120

1.E+00

TiO 1 2 coated Fe3O4

1.E-01

TiO 2 2

1.E-02

Fe 3 3 O4

1.E-03

Control

Mabey

1.E-04 1.E-05 1.E-06 1.E-07 0

Time (min)

100

120

Figure 7. E. coli disinfection in the presence of different nanoparticles

Acknowledgements This work was funded by the Jasper Summer Scholars program and the Korean Institute of Science and Technology (KIST). The author would like to thank Arvind Kannan from the Environmental Engineering department for his guidance, and Dr. Hossain Azam and Dr. Alexander Santulli for their support and mentorship.

References [1] WHO/UNICEF, Progress on sanitation and drinking water. 90 (2015). [2] “Rivers and Streams.” Environmental Protection Agency (EPA), 13 Mar. 2013. [3] Bethi, Bhaskar, et al. “Nanomaterials-based advanced oxidation processes for wastewater treatment: A review.” Chemical Engineering and Processing, 109, 178–189 (2016). [4] Bai Gajbhiye, Susheela. “Photocatalytic degradation study of methylene blue solutions and its application to dye industry effluent.” International Journal of Modern Engineering Research (IJMER), 2(3), 1204-1208 (2012) [5] Andreozzi, Roberto, et al. “Advanced oxidation processes (AOP) for water purification and recovery.” Catalysis Today, 53, 51–59 (1999) [6] Wei, Yan, et al. “Synthesis of Fe3 O4 nanoparticles and their magnetic properties.” Procedia Engineering, 27, 632-637 (2012). [7] Zhilin, D. M. and R. M. Peetz. “Pyrolysis of oxalates, a convenient method to synthesize powdered cobalt and nickel with catalytic activity: A laboratory for general or inorganic chemistry laboratory.” Journal of Chemical Education 91, 119-122 (2014).

From solution to adsorption: innovative method for removing toxic chromate Dominick Rendinaâ&#x2C6;&#x2014; Department of Chemistry and Biochemistry, Manhattan College Abstract. Hexavalent chromium is a well-known pollutant that leads to multiple cancers in humans in the lungs, nose, and nasal areas (Kumar et al., 2017). Granulated activated carbon (GAC) is utilized extensively as a method to remove toxic chromate from solution; however it has limitations when used in non-acidic conditions with low levels of salinity (Owlad et al., 2009). Hydroxamic acids are organic molecules which have an ability to form complexes with chromium. In an effort to increase the efficiency of GAC hydroxamic acids has been adsorbed onto GAC to assist in the removal of chromate at neutral pH values. Properties of the hydroxamic acid vary from more polar samples to less polar and symmetrical to asymmetrical molecules, in conditions that vary with time and kinetics.

Introduction Metal contaminants are a global issue ranging from developing states to modern first world societies, Hexavalent chromium is one of many and is a very well-known toxic metal considered as a priority pollutant (Mohan and Pittman, 2006). Industries which utilize the usage of chromium such as leather tanning, cement, mining, dyeing, fertilizer, and photography discharge wastewater that contains toxic metals into rivers (Kumar et al., 2017). The World Health Organization (WHO) states that the maximum allowable levels of chromate in drinking water must be below a level of 0,05mg/L (Owlad et al., 2009). The carcinogenic properties of chromium (VI) which have been studied to cause cancer in lungs, noses, and nasal sinuses (Kumar et al., 2017) makes removal of the metal a top priority issue. The use of activated carbon is the most accepted way to effectively remove Cr(VI), and alternatives have been studied for the removal of such heavy metals. Variations of activated carbon include granular-activated carbon (GAC), powder-activated carbon (PAC), activated carbon fibrous (ACF), and activated carbon clothe (ACC) (Owlad et al., 2009). Despite the widespread use of activated carbon, it was found that issues arise in terms of adsorption capacity and the efficacy of Cr(VI) is dependent on pH and salinity of a solution, which was reduced at neutral pH and low levels of salinity (Owlad et al., 2009). This suggests the lack of ability of activated carbon to remove Cr(VI) is reduced in the presence of ions, meaning solutions of chromate need to be acidified for optimal performance which can be expensive. To combat the cost of commercial activated carbon, biosorbents have been studied as a possible replacement, which are derived from low-cost agricultural waste (Owlad et al., 2009). One promising option which has been studied is the use of coconut shell charcoal (CSC) for its abundance provided from the coconut industry and the innate presence of functional groups such as carboxylic acids, hydroxyls, and lactones (Babel and Kurniawan, 2003). â&#x2C6;&#x2014;

Research mentored by John Regan, Ph.D.

160

The Manhattan Scientist, Series B, Volume 4 (2017)

Rendina

In comparison to commercial activated carbon, little is known about CSC properties. While activated carbon is useful in chromate removal, its ability to remove Cr(VI) at higher pH values becomes a problem. The objective of this project was to enhance the removal of Cr(VI) by adsorption of organic molecules onto granular activated carbon and granular coconut shell carbon. Organic molecules containing functional groups known as hydroxamic acids were used, ranging from polar samples to non-polar samples to understand the ability of activated carbon to adsorb organic molecules to remove Cr(VI) from solution.

Materials and Methods Hydroxamic acids listed in Fig. 1 were utilized; benzohydroxamic acid (BHA), N-phenylbenzohydroxamic acid (NPBHA), acetohydroxamic acid (AHA), and dihydroxamic succinate (SHA).

Figure 1. Hydroxamic acid samples; 1. BHA, 2. NPBHA, 3. AHA, 4. SHA

All data collected were performed in duplicate throughout: 1. To determine the effectiveness of hydroxamic acids to capture chromate, samples were dissolved in water, and volumes of a 200 ÂľM Cr(VI) stock solution were added to the mixture. GAC/CSC was then added to the solution and soaked for an hour to adsorb the hydroxamic acid-chromate complex. A UV/VIS spectra was taken of the solution at 372 nm to determine whether the hydroxamic acid sample had the desired efficacy. 2. Successful samples of hydroxamic acids were then loaded onto GAC or CSC depending on the polarity of the substance, less polar molecules were loaded onto GAC whereas more polar molecules were loaded onto CSC with the knowledge that CSC naturally possesses more polar molecules such as carboxylic acids and hydroxyls to aid in adsorption (Babel and Kurniawan, 2003). The adsorption of hydroxamic acids onto activated carbon typically occurred after one hour in a 1:1 ratio of activated carbon to hydroxamic acid. Loaded samples of activated carbon with hydroxamic acid were filtered and washed with water to remove excess material. Samples

Rendina

The Manhattan Scientist, Series B, Volume 4 (2017)

161

prepared through these methods were used to determine all other empirical data. Hydroxamic acid samples which successfully removed Cr(VI) were soaked for 19 hours statically with hexavalent chromium to determine optimal breakthrough values. 3. Stirring datum was gathered using a Tek-Tator V Variable Rotator. Statically loaded activated carbon samples were soaked in Cr(VI) stock solution for one hour rotating at 50 rpm, 100 rpm, and 150 rpm. Hydroxamic acid was also loaded onto activated carbon rotating at 100rpm and was compared to the statically loaded activated carbon which stirred in chromate solution for one hour. The pH of chromate solution was adjusted from neutral solution to a pH of 4 and 2 respectively by adding concentrated HCl dropwise. 4. The removal of chromate was determined through the use of the Environmental Protection Agency’s (EPA) Method 7196A, a colorimetric method for determining hexavalent chromium in solution. A UV/VIS spectrum using glass cuvettes and recording peak values at 540 nm was utilized and the concentrations of chromium were determined from a calibration curve.

Results BHA and SHA were the only two hydroxamic samples out of the four which worked in solution (data not shown). 1 g of SHA and BHA were loaded onto 1 g of CSC and GAC respectively and kept in 20 mL of 200 µM chromate overnight. GAC and CSC not containing any hydroxamic acid was exposed to Cr(VI) as well to show the effectiveness of hydroxamic acid. The results in Table 1 show that GAC-BHA is somewhat more effective than GAC at removing chromate. However, CSC-SHA is approximately 100-fold more efficient at chromate removal than CSC alone and, also, tenfold more capable of lowering Cr(VI) levels than GAC. Table 1. SHA vs BHA in Hexavalent Chromium removal Material

Concentration Cr(VI) (averaged)

GAC GAC-BHA 1:1 CSC CSC-SHA 1:1

22 ±2 µM 14 ±3 µM 197 ±3 µM 2.00 ±0.01 µM

The breakthrough value of an adsorbent is defined as the maximum amount of chromate that can be removed from solution per gram of adsorbent to achieve levels of chromate that are deemed safe by regulatory agencies. The data in Table 2 show 1 g of CSC-SHA within chromate solution under static conditions for 19 hours, the breakthrough value is between 20 - 40 mL of a 200 µM solution. Rotational speeds of mixing of the solid CSC-SHA and chromate solution were examined. The purpose was to determine if an increase in the mobility of chromate ions in solution affects the rate and amount chromate capture. Toward this goal, 0.5 g of CSC-SHA 1:1 was added to 35 mL of 200 µM chromate solution and rotated at either 50, 100, or 150 rpm. The data in Fig. 2

162

The Manhattan Scientist, Series B, Volume 4 (2017)

Rendina

Table 2. Breakthrough values of CSC-SHA 1:1 in 200 µM Cr(VI) solution Volume of Cr(VI)

Concentration Cr(VI) (averaged)

20 mL 40 mL 50 mL 60 mL 70 mL

<1.0 µM 6.5 µM 8.3 µM 21.6 µM 31.4 µM

shows that 150 rpm captured Cr(VI) the best while 50 rpm did not perform as well. Furthermore, the 150 rpm rotation experiment was able to remove chromate after 30 minutes as opposed to reactions conducted at 100 rpm which required 60 minutes. These results suggest the mobility of the solution is more effective than a static environment. And the speed of the rotation will also affect the outcome due to an increase diffusion of chromate ions into the porous cavities of CSC-SHA.

Concentration (µM)

200 50RPM 150

100RPM 150RPM

100

0 0

Time (Minutes) Figure 2. Concentration vs. time, statically loaded SHA on CSC in stirred Cr(VI) solutions

Loading SHA onto CSC while stirring at 100 rpm gave better results for chromium removal in comparison to statically loading SHA onto CSC. One-half gram of SHA was dissolved in water and 0.5 g of CSC was added. Stirring was continued at 100 rpm for one hour, whereas the static sample underwent the same conditions but without stirring. Both materials were stirred at 100 rpm in 35 mL of 200 µM Cr(VI) for one hour. Within 30 minutes the sample prepared through stirring was able to remove the same amount of chromate as the static sample at 60 minutes. Fig. 3 summarizes these results: CSC-SHA 1:1 loaded at 100 rpm was able to reduce the chromate concentration to 11 µM in 30 minutes, whereas statically loaded CSC-SHA 1:1 lowered Cr(VI) levels to 33 µM during the same time, a 3-fold improvement for Cr(VI) capture. The adsorption characteristics of CSC-SHA at different pH values were determined and com-

Rendina

The Manhattan Scientist, Series B, Volume 4 (2017)

163

Concentration (µM)

250 Static loading

200 150

100rpm loading

100 50 0 0

Time (Minutes)

Figure 3. Static vs. rotational loading of SHA onto CSC, Cr(VI) removal comparison

pared to CSC alone. The pH of the solution drastically affects the removal capabilities as seen in Table 3. Samples of statically loaded CSC-SHA 1:1 were compared to CSC alone in 35mL of 200 µM in pH conditions of 6, 4, and 2. At pH 6, CSC was unable to remove any chromate from solution which proves that SHA is contributing to the removal process. However at pH 2 CSC-SHA 1:1 can reduce Cr(VI) levels to a concentration of < 1 µM within 30 minutes. Under these conditions, CSC-SHA 1:1 is 80 times more effective than CSC alone. Table 3. Concentrations (µM) of chromate vs time in variable acidic conditions pH 6

pH 4

pH 2

Time (minutes)

CSC

CSC-SHA 1:1

CSC

CSC-SHA 1:1

CSC

CSC-SHA 1:1

0 30 60

200 200 200

200 33 ±1 11 ±1

200 87 ±1 32 ±2

200 45 ±21 4±3

200 80 ±8 11 ±1

200 <1 <1

Discussion The underlying principle of adsorbing organic molecules onto activated carbon was the belief that the surface has a set amount hydrophilic and hydrophobic sites in which the molecule could adsorb onto the surface, but in different orientations. Using BHA as an example, the hydrophilic sites would adsorb the more polar end of the molecule and adhere to the surface with hydrogen bonding, whereas the hydrophobic sites would bind the aromatic ring and attach the molecule by van der Waal’s forces. These principles are seen in Fig. 4. This leads to complications when dealing with asymmetric molecules since the nature of their bonding ability to activated carbon is more favored towards the hydrophilic orientation and it remains uncontrollable. The desired orientation for BHA is to have most of the molecules attach

164

The Manhattan Scientist, Series B, Volume 4 (2017)

Rendina

Figure 4. Hydrophilic Binding vs Hydrophobic Binding

onto activated carbon in the hydrophobic orientation, however, this desired orientation relies on the strength of van der Waalâ&#x20AC;&#x2122;s forces which are significantly weaker than hydrogen bonding forces. This leads to the belief that many of these molecules will bind in the hydrophilic orientation, leading to a counterproductive process with no significant benefits. As seen in Table 1, the difference in chromate removal of GAC-BHA in comparison to GAC alone is small while CSC-SHA preforms significantly greater than that of CSC alone, suggesting that molecules which are hydrophobic and asymmetric are less optimal. The polarity and symmetry of SHA provide a more reliable source of chromate removal since the binding orientation will remain the same (Fig. 5), and the use of a more polar molecule allows the usage of the less expensive CSC for its polar affinity (Babel and Kurniawan, 2003).

Figure 5. Symmetry Benefits of SHA on hydrophilic sites in carbon

The two molecules that were unsuccessful lacked the benefits of both SHA and BHA. AHA is believed to not possess a large enough carbon group to provide supportive van der Waalâ&#x20AC;&#x2122;s, causing a complete preference for hydrophilic binding. NPBHA did not produce the desired effect in solution where BHA was able to; this suggests that the loss of the hydrogen atom on the nitrogen atom affects the chelating effect of the molecule negatively. Many of the experiments of CSC-SHA 1:1 showed the rotation changes the kinetics of migration of the Cr(VI) from solution onto the sample, which produces a better result in shorter amounts of time. Solutions of chromate that rotated at 150 rpm provided a lower concentration of chromate in shorter amounts of time as seen in Fig. 2.

Rendina

The Manhattan Scientist, Series B, Volume 4 (2017)

165

Loading SHA onto CSC was much better when the solution stirred at 100 rpm as opposed to remaining static as seen in Fig. 3, which gives a suggestion to how much of the material can actually be loaded onto CSC. Loading SHA onto CSC at 100 rpm and stirring the CSC-SHA at 100 rpm was able to produce a Cr(VI) concentration of 11 µM which is a 3-fold improvement from statically loaded SHA on CSC rotating at 100 rpm. This suggests that the amount of sites that CSC can adsorb also lies within the pores of the granulated carbon and not just the surface. By rotating the liquid, CSC is able to adsorb SHA deeper into the granulated coconut carbon framework. More experimentation is needed to determine what rotational speed increases the efficiency of the sample. The pH of the solution drastically affects the removal capabilities of the sample. Table 3 demonstrates that CSC provides no removal benefits at pH 6 but a substantial improvement at a pH of 2. With SHA loaded onto CSC, the sample works at pH 6, and this can be optimized by lowering the pH. It is unclear to what extent a pH decrease assists the CSC-SHA 1:1 sample as the concentration of chromate was less than 1 µM before 30 minutes. It is nevertheless more optimal in such conditions. The amount of potassium chromate removed was optimized by rotating the mixture at 150 rpm and at a pH of 2. One gram of CSC-SHA is able to remove 2.72 mg of chromate. It is believed that this number can increase and the full potential of the material has not been examined thoroughly. Fig. 3 shows that samples that were loaded with SHA while rotating preformed more optimal than a sample loaded statically. Despite the success of the sample at a pH of 2 it is also worth noting that CSC performs better at lower pH as well.

Conclusion The success of chromate removal with CSC-SHA 1:1 has been proven for polar hydroxamic acid molecules which possess a plane of symmetry to bypass the problem of hydrophilic versus hydrophobic binding sites. Less polar asymmetric hydroxamic acids were inefficient in comparison, but more experimentation may be done on other similar types of hydroxamic acids with larger carbon groups. Under rotating conditions for both the loading of hydroxamic acid and removal of chromate yielded better results. Future experimentation can be done for hydroxamic molecules similar to SHA but with more hydroxamic substituents attached onto the molecule.

Acknowledgments This work was funded by the Manhattan College Jasper Scholars Program. The author would like to thank the Department of Chemistry and Biochemistry for the opportunity, and Dr. John Regan for the advisement and support in producing these results.

References Babel, S. and T. A. Kurniawan. “Cr(VI) removal from synthetic wastewater using coconut shell charcoal and commercial activated carbon modified with oxidizing agents and/or chitosan”. Chemosphere 54, 951-967, 06 October 2003.

166

The Manhattan Scientist, Series B, Volume 4 (2017)

Rendina

Environmental Protection Agency (EPA). “Method 7196A: Chromium, Hexavalent.” Web 27 Jan. 2017. www.epa.gov/hw-sw846/sw-846-test-method-7196a-chromium-hexavalent-colorimetric Kumar, A., A. Balouch, A. A. Pathan, A. M. M. Abdullah, M. S. Jagirani, F. A. Mustafai, M. Zubair, B. Laghari, and P. Panah. “Remediation techniques applied for aqueous system contaminated by toxic chromium and nickel ion.” Geology, Ecology and Landscapes. 1(2), 143-153, 03 May 2017. Mohan, D. and C. U. Pittman Jr. “Activated Carbons and low cost adsorbents for remediation of tri- and hexavalent chromium from water.” Journal of Hazardous Materials B137, 762-811, 29 June 2006. Owlad, M., M. K. Aroua, W. A. W. Daud, and S. Baroutian. “Removal of Hexavalent ChromiumContaminated Water and Wastewater: A Review.” SpringerLink. Springer Science, 20 Nov. 2008. Web 02 Sept. 2016.

Determining the framework topology of a microporous aluminosilicate zeolite, SUZ-9 Christine E. Schmidt∗ Department of Chemistry and Biochemistry, Manhattan College Abstract. SUZ-9 was independently synthesized and characterized in the laboratories of BP and Exxon Mobil [1, 2]. This article describes the research done to solve the structure of SUZ-9 because neither company determined its structure. Using synchrotron x-ray powder diffraction data, the structure of SUZ-9 was solved utilizing knowledge of crystallographic concepts, hints from various SuperFlip and Zefsa-II solutions, and physical model building techniques using known building units found in other members of a 12-ring family. A systematic and exhaustive examination of possible models using only a small data set of 99 reflections allowed the framework topology of SUZ-9 to be determined in P¯ 62m symmetry. The calculated powder pattern matches the experimental powder pattern, indicating a correct topology.

Introduction A zeolite is a stable, microporous aluminosilicate mineral formed by interlinking tetrahedra of SiO4 and AlO4 [3]. The sharing of oxygen atoms between tetrahedra makes a three-dimensional structure. The microporous structures of zeolites vary in shapes and sizes, allowing them to be used as detergents, sieves, and catalysts. Zeolites are found with many different topologies, which contain large pores in a uniform arrangement. SUZ-9 was thought to be the largest member of a family of zeolites with 12-ring pore openings (Table 1).

Methods The synchrotron x-ray diffraction data set (5◦ − 25◦ 2θ) collected by J. M. Bennett on calcined microcrystalline samples of SUZ-9 was entered into sophisticated computer programs that determine structures from powder diffraction data. Theoretical calculations in SuperFlip [4], a charge-flipping algorithm for solving crystal structures, provided a number of trial solutions for the topology of SUZ-9, none of which were completely interpretable. SuperFlip revealed the best hints, including suggested space groups P/63 /mcm, P6/mmm, and P¯6m2. Fig. 1 shows a typical SuperFlip result obtained by Daisuke Kuroshima, which suggested 12-rings at the vertices and 6-rings across the long diagonal of the unit cell. Building units of gme, can, and d6r could be recognized. However, it is impossible to determine all up-down [001] connections. Zefsa-II [5], a simulated annealing program, was run on the college’s most capable computer by J. F. Capitani. The results also suggested 12-ring pores at the vertices (Fig. 2). ∗

Research mentored by Richard Kirchner, Ph.D.

168

The Manhattan Scientist, Series B, Volume 4 (2017)

Schmidt

Table 1. Family of 12-Ring Porous Zeolites. Each member has 12-ring pore openings (a ring of 12 Si atoms sharing ˚ and similar adsorption O atoms). All members have hexagonal cell dimensions with the same value for c (7.5 A) properties. Framework Type

Material Name

˚ a (A) (a = b)

˚ c (A)

Building units

Space Group

OFF

Offretite

13.291

7.5

d6r can gme

P¯ 6m2

MAZ

Mazzite

18.102

7.5

gme

P63 /mmc

LTL

LZ-212

18.13

7.5

can d6r ltl

P6/mmm

LTF

LZ-135

31.3830

7.5

gme

P63 /mmc

MOZ

ZSM-10

31.575

7.5

d6r can pau ltl

P6/mmm

SUZ-9

36.14

7.5

gme

P¯ 62m

Figure 1. A typical SuperFlip result. The [001] projection of atom positions are shown in black. Atoms that reveal 12-ring pores or 6-rings are colored pink. The main ”hints” are the presence and location of the rings shown in pink. The final topology of SUZ-9 is superimposed (black lines).

Using hints from SuperFlip, models were constructed using the building units listed in Table 1 and assembling them into new configurations to produce the cell constants of SUZ-9. The physical ˚ models were made using metal tetrahedra connected by plastic tubing cut to scale (1 cm ↔ 1 A).

Schmidt

The Manhattan Scientist, Series B, Volume 4 (2017)

Figure 2. A Zefsa-II Simulated Annealing/ Parallel Tempering result. The [001] projection shows Si atom positions. The green lines are bonds ˚ between two Si atoms. A 12-ring pore ( 3.1 A) is clearly shown centered around each vertex.

169

Figure 3. One of the SUZ-9 models built using the ltl cage building unit. A sketch of the unit cell is drawn, which shows the model fits the cell dimensions. The plastic tubes represent the connection between the tetrahedral Si atoms. Oxygen atoms are not shown.

Silicon atoms were represented by the metal tetrahedra and oxygen atoms were at the centers of each plastic tube. My first promising model (Fig. 3) was formed by connecting 12-ring ltl cages (black) with 6and 8-ring bridges (yellow). This simple and elegant model, in P6/mmm symmetry, was consistent with the unit cell and chemical characterization known about SUZ-9. The atom positions in this model were input into the ATOMS graphics program [6] to get crystallographic coordinates for all atoms. These were imported into GSAS [7], where a Rietveld refinement was done to calculate and compare the calculated and experimental powder patterns. This model’s calculated powder pattern was different than the experimental, indicating this model is incorrect. Model building was continued to systematically construct all models from possible combinations of 12-ring pores and building units in the 12-ring family. Five research students, Jessi Dolores, Christopher Kim, Daisuke Kuroshima, Anthony O’Mara, and Dominick Rendina, built stacks of gmelinite (gme) units (Fig. 4). These stacks were assembled with alternating heights to give 12-ring pores and their connections. One of these physical models had all of the known characterizations, including a similar calculated powder pattern, of SUZ-9.

Results and Discussion Atom positions taken from the physical model were input into the graphics program ATOMS. The framework topology of SUZ-9 (Fig. 5) consists of gme stacks joined together to form 12-ring pores. By inspecting the physical model, P. W. R. Corfield determined the space group is P¯62m. The five-student model has 12-ring pores at the vertices constructed from stacks of gme build-

170

The Manhattan Scientist, Series B, Volume 4 (2017)

Schmidt

˚ which is equal to the c cell dimension ing units. The gme building unit has a height of 7.5 A, in SUZ-9. The 12-ring pores at the vertices were made by connecting six stacks with alternating heights of ± 0.25 in z. The interior space between the 12-ring pores at the vertices was filled by a different 12-ring pore in the middle of the a and b axes and the center of the unit cell. These are made by joining four gme stacks that did not alternate in height, but had one high and three low stacks. Fig. 4 shows that the connections between high and low stacks create bridging 5-rings, while connections between equal height stacks create 4- and 6-rings.

7.5 Å

Figure 4. The gme building block (left). Gme stacks can be made with similar height (middle)

and alternating stacks.unit The stacks arbitrarily “low,” whereas Figure 4. Theheight gme(right) building (left).in green Gmeare stacks cannamed be combined withthesimilar height (middle) and alternating stacks in red are named “high.” height (right) stacks. The stacks in green are arbitrarily named “low,” whereas the stacks in red are named “high.”

Results and Discussion positions taken located from the physical were input the graphics program to those found in the MOZ, The Atom 12-ring pores at themodel vertices in into SUZ-9 are similar ATOMS. The framework topology of SUZ-9 (Figure 5) consists of gme stacks joined together to MAZ, topologies [8].P.W.R. They aredetermined all composed form and 12-ringLTF pores. (LZ-135) By inspecting the physical model, Corfield the space of alternating high/low gme group is P-62m. a pore labeled as 12chan1 in Fig. 5. The 12chan1 pore in MOZ, MAZ, and LTF stacks creating

Figure 5. The [001] projection of SUZ-9 drawn in ATOMS. The bonds represent the connections between the tetrahedral Si atoms. Oxygen atoms have been omitted for clarity. The gme stacks are shown as red (high) and green (low). The pores are labeled for reference.

The five-student model has 12-ring pores at the vertices constructed from stacks of gme building The gme projection building unit a height of in 7.5ATOMS. Å, which The is equal to represent the c cell the connections between the tetraFigure units. 5. The [001] of has SUZ-9 drawn bonds dimension SUZ-9.Oxygen pores at the vertices werefor made by connecting stacksare with The 12-ring hedral Siinatoms. atoms have been omitted clarity. The gmesix stacks shown as red (high) and green (low). alternating heights of ± 0.25 in z. The interior space between the 12-ring pores at the vertices was The pores are labeled for reference. filled by a different 12-ring pore in the middle of the a and b axes and the center of the unit cell.

Schmidt

The Manhattan Scientist, Series B, Volume 4 (2017)

171

was rotated by 30â&#x2014;Ś about [001] in order to construct the SUZ-9 topology. The slightly elliptical 12-ring pores, labeled 12chan3, in the middle of the a and b axis and in the center of the cell are different from those at the vertices. They are constructed from gme stacks with heights in a high-low-low-low pattern producing 4-, 5-, and 8- ring pore walls, unique in the 12-ring family. There are two types of 8-rings present in the structure. The ones around the 12chan1 pore are labeled as 8chan1 and the ones around the 12chan3 pore are designated as 8chan4, as shown in Fig. 6. The 8chan1 pore is found in MAZ and LTF. The 12chan1 pore is surrounded by 8rings made from combining alternating high-low gme stacks with inner 5- and 8-ring walls. The 12chan3 pores are surrounded by 8-rings (8chan4) made from combining high-low-low-low gme stacks producing 4-, 5-, 6-, and 8-ring inner walls. The 8chan4 pore is also unique to SUZ-9.

Figure 6. Projection [001] and perspective view (middle) of the 8chan1 and the unique 8chan4 pore drawn in ATOMS. Side view [010] of 8chan4 (right).

A SiO2 framework topology was assumed to refine all atom coordinates of the physical model in DLS-76 [9]. A Distance Least Squares (DLS) calculation presuming random distribution of Al in tetrahedral sites at a 1:3 Al/Si ratio produced final atom coordinates insignificantly different than just presuming Si tetrahedral sites. Atom positions off of symmetry sites were moved by hand to the proper sites. All atoms were imported to GSAS-II [10], a crystallographic program used for refining structures. A second phase taken directly from the Zeolite Database [11] was imported, accounting for the Offretite (OFF) impurity in the Mobil sample [2]. All background, instrument, histogram scale, and phase fraction parameters for both phases were refined. However, all attempts to refine SUZ-9 atom parameters failed because the number of variables was greater than the number of reflections. A Difference Fourier (DF) calculation done by Daisuke Kuroshima on the SUZ-9 phase (Fig. 7) was performed to produce an electron density map of the topology. Results showed the highest peaks were seen near Si atoms and the formed bonds between DLS refined atom positions. The highest peaks were about half of a typical Si intensity. At this point, no significant DF peaks appeared in any pore.

172

The Manhattan Scientist, Series B, Volume 4 (2017)

Schmidt

Figure 7. Difference Fourier map for SUZ-9. The [001] projection showing the framework topology superimposed upon the DF peaks (black dots).

A Difference Fourier calculation on the OFF phase (Fig. 8) has the highest two peaks slightly displaced from framework oxygen atoms. Two very weak peaks are seen in the 12-ring pore, too weak to be considered Na+ or K+ ions.

Figure 8. Difference Fourier maps for the OFF phase. The [001] projection showing DF peaks (black dots) and the OFF framework topology. The left shows a DF with OFF atoms. The right shows the DF after moving two O positions to their DF peaks.

To further refine the SUZ-9 topology, the restraints function in GSAS-II was experimented with. LZ-135 (LTF), a member of the 12-ring family, has a similar Al/Si ratio in T sites and similar combinations of gme stacks to SUZ-9. After an atom-by-atom comparison between SUZ-9 and LTF done by Christopher Kim, all identical stereochemical positions in LTF were transferred as restraints for SUZ-9. In other words, the bond (T-O) distance and bond (O-T-O) angle connections found in LTF which are similar to connections found in SUZ-9 were used as ideal values for restraint calculations in GSAS-II. When atoms in SUZ-9 were associated with the unique pores (12chan3 and 8chan4), bond and angle values were obtained from DLS-76 and 6 APID cycles. Refining the SUZ-9 phase in GSAS-II with these restraints resulted in some of the refined bond and angle values to be out of range. The agreement between the calculated and experimental powder pattern looked worse. Increasing restraint weight factor values improved the fit of the model to the data (Fig. 9), but did not significantly decrease the R(F) values for the SUZ-9 phase. Ë&#x161; distances from the framework oxygen atoms suggesting DF peak #7 and peak #8 have 2.4 - 2.7A

Schmidt

The Manhattan Scientist, Series B, Volume 4 (2017)

173

these peaks could be Na+ or K+ ions in the framework pores. A Rietveld refinement was not possible because the number of variables needing refinement in SUZ-9 was larger than the number of observations in the small synchrotron data set.

Figure 9. The [001] projection with superimposed Difference Fourier peaks after 28 cycles of refinement with restraints weight factor of 10,000. The nine highest DF peaks are shown in black, smaller DF peaks are shown in purple. DF peaks that might be Na+ or K+ were recolored for clarity. DF peak 7 is orange, while 8 is yellow.

The agreement at this stage between calculated and experimental powder patterns is shown in Fig. 10. The SUZ-9 pattern is calculated from very restrained Si and O atom positions, while the OFF pattern is calculated from atom coordinates taken from the zeolite database, and slightly improved as described in Fig. 8.

Figure 10. GSAS-II plot for the SUZ-9 model. The dark blue crosses are the experimental data points collected to 25â&#x2014;Ś 2Î¸. The green line is the calculated powder pattern (combined SUZ-9 and OFF phases). The light blue line represents the difference between the calculated and experimental powder pattern. The red line is the background correction. Below the difference line are blue hash marks, representing SUZ-9 reflection positions, and red hash marks, representing OFF reflection positions.

174

The Manhattan Scientist, Series B, Volume 4 (2017)

Schmidt

To identify the peaks in the pores Dr. Christopher Koenigsmann and his research student Brett Musialowicz used the Fordham University Zeiss EVO MA10 Lab6 Scanning Electron Microscope (SEM) to characterize SUZ-9, in particular to determine the cell composition by EDX (Energy Dispersive X-ray) spectra. Two SEM images of SUZ-9 are shown below (Fig. 11).

Figure 11. SEM images of SUZ-9

The SEM images are not capable of revealing any actual pores in the zeolite. Neither image seems to reveal much about the crystal morphology of SUZ-9. However, the higher resolution SEM image (on the right) is curious, looking like a pile of rice. The EDX spectrum (Fig. 12) and data (Table 2) unambiguously give the cell composition of SUZ-9.

Figure 12. The EDX spectrum of SUZ-9

Schmidt

The Manhattan Scientist, Series B, Volume 4 (2017)

175

Table 2. The composition of SUZ-9. The mean value combines 18 runs on a carbon coated sample and 10 runs on an untreated sample of SUZ-9. There was no significant difference between carbon coated and untreated samples.

mean value estimated error

Si atomic %

Al atomic %

K atomic %

Na atomic %

O atomic %

21.97 3.32

6.40 0.86

5.89 1.57

0.29 0.03

65.44 5.69

The mean atomic % values for Si and Al indicate that, presuming a random distribution of Al in the zeolite framework topology, a tetrahedral site is approximately 75% Si and 25% Al This is consistent with what we estimated earlier based upon information in the patent. It is also consistent with what is found in LZ-135. The atomic % of Al and K are within experimental error the same value. When Al substitutes for Si, the framework acquires a negative charge. The presence of K+ in pores maintains charge neutrality. The atomic % oxygen is larger than expected from the number of oxygen atoms connecting tetrahedral sites in the TO2 framework. Presumably, the pores and cavities contain water.

Conclusion The very good fit between the experimental and calculated powder patterns shown in Figure 10 was obtained before the DF peaks in the pores were determined to be K+ . The agreement should improve to excellent after K+ is included in the refinement of SUZ-9. There can be no doubt that the correct topology, and perhaps even the almost complete structure will be determined!

Acknowledgements This research was funded by the Camille and Henry Dreyfus Foundation Senior Scientist Mentor Program. The author would like to thank her mentor, Dr. Richard Kirchner for guidance, Nur Rafidah Mohd Faisal for help preparing the figures, and Jessi Dolores, Christopher Kim, Daisuke Kuroshima, Anthony O’Mara, and Dominick Rendina for research collaborations.

References [1] K. D. Schmitt, Zeolites, 15, 315-7 (1995) [2] K. D. Schmitt, US. Pat. 5,399,337 (May 21, 1995) [3] G. L. Price, “What is a Zeolite?” http://personal.utulsa.edu/∼geoffrey-price/Research/index. html [4] L. Palatinus and G. Chapuis, SuperFlip, J. Appl. Cryst. 40, 786-90 (2007) [5] M. Falcioni, M. W. Deem, J. Chem. Phys. 110, 1754-66 (1999) [6] E. Dowty, ATOMS, Shape Software (2016) [7] A. C. Larson, R. B. Von Dreele, General Structure Analysis System, Los Alamos Natl. Lab., LAUR 86-748 (2000)

176

The Manhattan Scientist, Series B, Volume 4 (2017)

Schmidt

[8] L. B. McCusker, C. Baerlocher, S. T. Wilson, R. W. Broach, J. Phys. Chem. 113, 9838-44 (2009) [9] C. Baerlocher, A. Hepp, W. M. Meier, DLS-76, Institut f¨ur Kristallographie: ETH Z¨urich, Switzerland (1976). [10] B. Toby, R. B. Von Dreele, GSAS-II, J. Appl. Cryst. 46, 544 (2013 ) [11] C. Baerlocher, L.B. McCusker, Database of Zeolite Structures: http://www.iza-structure.org /databases/

Appendix Table 3. Atomic coordinates describing the topology of SUZ-9 Name

Name

Si1 Si2 Si3 Si13 Si14 Si4 Si5 Si15 Si7 Si11 Si12 Si9 Si8 Si10 Si16 Si6 O11 O12 O22 O23 O34 O44 O45 O56 O36 O513

0.17860 0.23400 0.24570 0.54668 0.24380 0.34000 0.54260 0.54800 0.18100 0.45030 0.39210 0.09560 0.23400 0.39700 0.25000 0.39970 0.14590 0.22370 0.21000 0.28600 0.34190 0.38100 0.41880 0.44600 0.35900 0.50180

0.12620 0.18210 0.33520 0.36730 0.67143 0.39040 0.33500 0.17930 0.04970 0.11630 0.04500 0.33820 0.05265 0.15450 0.58010 0.24230 0.14587 0.15210 0.21024 0.21470 0.29580 0.38066 0.34400 0.28680 0.24100 0.36740

0.28720 0 0 0.21000 0.5 0 0.28870 0.21180 0.20950 0.21000 0.5 0.5 0.5 0.20470 0.5 0.29170 0.24200 0.17690 0 0 0 0 0.17380 0.26400 0.18370 0.22420

O610 O78 O89 O910 O912 O1011 O1112 O1314 O1315 O1416 O1516 O46 O1115 O17 O771 O88 O1515 O1010 O1111 O1313 O77 O66 O1 O55 O1212

0.40800 0.22600 0.28450 0.35400 0.35120 0.44000 0.42410 0.54810 0.58360 0.62450 0.54910 0.58100 0.50170 0.15430 0.14900 0.20660 0.55560 0.38400 0.43800 0.55450 0.19440 0.38460 0.19090 0.44810 0.37200

0.20430 0.07390 0.06760 0.12400 0.05800 0.15420 0.07210 0.32900 0.41500 0.37450 0.21870 0.29010 0.13540 0.07520 0 0 0.19420 0.13720 0.10330 0.36100 0.05070 0.23390 0.13200 0.34600 0

0.21950 0.32330 0.5 0.31920 0.5 0.27370 0.32590 0.32020 0.27980 0.5 0.32330 0.5 0.23500 0.23300 0.27700 0.5 0 0 0 0 0 0.5 0.5 0.5 0.5

Variations of the A cation in organic/inorganic perovskite materials Melissa Skuriatâ&#x2C6;&#x2014; Department of Chemistry and Biochemistry, Manhattan College Abstract. With the increasing evidence and recognition of the effects of climate change, there is a pressing need for sources of energy that are less harsh on the environment. This demand has prompted the use of solar cells and the continuous research on how to improve their efficiency. Perovskite solar cells have recently been studied and proven to perform well in solar cells. They hold power conversion efficiencies that are comparable to those of photovoltaic materials that have been studied for about forty years. The goal of this research project was to determine what combinations and ratios of organic cations within the perovskite structure would be optimal for use within a solar cell. These materials, such as the standard methylammonium lead iodide, are tested for absorbance and stability using a spectrophotometer. Once optimal combinations have been determined, the materials will be tested in an X-ray diffractometer to determine crystal structures, synthesized in the form of nanowires, and then placed in a solar cell to test electrical conductivity. The use of perovskite solar cells is a cleaner energy alternative, and it is cheaper and more efficient than other photovoltaic devices.

Introduction Perovskite is the mineral CaTiO3 , and it was originally discovered in the Ural Mountains of Russia. A perovskite material is any material that imitates the same ABX3 crystalline structure of the perovskite mineral. In this general structure, A and B represent positively charged atoms, or cations, and X represents a halogen anion, or a negatively charged atom. It has been found that perovskite materials used within solar cells are most effective when the A cation is organic, the B cation is a large, inorganic cation, and X is a small halogen anion [1]. The organic cations tested throughout this research project were acetamidinium iodide, formamidinium iodide, guanidinium iodide, imidazolium iodide, and the standard methylammonium iodide. These organic cations are hydrocarbons with a minimum of one nitrogen, which is active in acid-base reactions. It is imperative that the organic cations are not too large so as to prevent any disruption of the perovskite structure. The organic cation is located at the core of the structure, and is surrounded by the inorganic cations and halogen anions. Lead served as the inorganic cation, while iodide and chloride worked as the halogen anion component. However, the chloride samples were not able to be optimized. Compared to traditional silicon solar cells, perovskite solar cells are much cheaper but just as efficient for use within a solar cell. This is due to the fact that silicon solar cells are synthesized using a multi-step process, whereas perovskite materials for solar cells are synthesized using a one-step precipitation reaction [2]. Perovskite solar cells that have been studied over the last five years have reached power conversion efficiency rates as high as 22.1%, while cadmium telluride, a photovoltaic material that has been studied for almost forty years, has reached a power conversion â&#x2C6;&#x2014;

Research mentored by Alexander Santulli, Ph.D.

178

The Manhattan Scientist, Series B, Volume 4 (2017)

Skuriat

efficiency of 22.3%. Perovskite materials are also capable of converting photons into electricity with minimal energy loss, which is beneficial for use within a solar cell [1]. Although lead is toxic, it is readily available and inexpensive. It has also shown to be an optimal component within perovskite solar cells, especially when paired with methylammonium iodide. When perovskite materials come into contact with moisture, the characteristic structure that allows for photovoltaic energy conversion degenerates, and it hinders performance. During this research project, samples were kept isolated from moisture to prevent this setback. It has not yet been established how stable perovskite materials are over time [3]. In addition to determining which organic cation samples absorbed light best, the stability of each sample was studied.

Materials and Methods Preparation of Samples Samples of perovskite materials were made using 0.88 M concentrations of organic cation/lead halide materials in dimethylformamide solvent. Samples included: methylammonium lead iodide, acetamidinium lead iodide, formamidinium lead iodide, guanidinium lead iodide, and imidazolium lead iodide. Methylammonium lead iodide was also mixed with the four other organic cation/lead iodide samples in 25% - 75%, 50% - 50%, and 75% - 25% mixtures. Preparation of Sample Slides Microscope slides were cut into squares using a glass cutter. Slides were then sonicated in acetone for five minutes, isopropyl alcohol for five minutes, and then dried with nitrogen gas. After each slide was dry, it was adhered to a Teflon base on a spin coater computer fan. The system was covered with a desiccator to contain any spraying of each sample. The spin coater was attached to a Variac to control spin speed; spinning at 2000 rpm for 30 seconds was the standard time used for all samples. After spinning, each sample was placed on a hot plate and covered with a crystallizing dish. Samples were heated at 150â&#x2014;Ś C for 5 minutes each in order to activate the perovskite material. An Agilent Cary Series UV-Vis-NIR Spectrophotometer tested the absorbance capability. The samples were left out exposed to air for about 24 hours, and they were each placed in the spectrophotometer once more to observe how air affected the absorbance and stability of each sample.

Results Fig. 1a displays the absorbance capabilities of the lead iodide samples with the various organic cations. The samples were tested immediately after being heated and cooled. As shown in Fig. 1a, methylammonium lead iodide and formamidinium lead iodide hold peaks in the blue and red regions, showing that each can absorb a broad range of light. Guanidinium lead iodide holds a strong peak in the blue region, and it strongly absorbs blue light. Acetamidinium lead iodide and imidazolium lead iodide hold broad peaks in the near-IR region, and each absorbs light in that region. The sudden increase in absorption at â&#x2C6;ź750 nm for methylammonium lead iodide, and

Skuriat

The Manhattan Scientist, Series B, Volume 4 (2017)

179

1.5

Acetamidinium Lead Iodide Formamidinium Lead Iodide

1.0

Absorbance (a.u.)

Guanidinium Lead Iodide

0.5

Imidazolium Lead Iodide

Methylammonium Lead Iodide

0.0 300

500

700

900

1100

300

500

700

900

1100

Wavelength (nm)

Figure 1. UV-visible spectra of 0.88 M organic cations with lead iodide (a) after sample heating (left) and (b) after exposure to air for 23 hours (right).

∼800 nm for formamidinium lead iodide reveal the ability for these compounds to absorb a large portion of the visible spectrum. It should be noted that any separation of data in the graphs along the vertical axis could be due to light scattering within the thin films of the material. Fig. 2 shows each sample slide after heating and cooling. Methylammonium lead iodide, the standard sample, has a dark black color resulting from the absorption of the majority of visible light. Similarly, formamidinium lead iodide is a dark color and also absorbs over that broad range. Although acetamidinium lead iodide and imidazolium lead iodide are paler and yellow, their absorption spectra suggest that each absorbs light well in the near-IR region. The guanidinium lead iodide is a bright yellow, and it absorbs light in the blue region, as expected.

ALI

FLI

GLI

ILI

MALI

Figure 2. Sample slides of each 0.88 M organic cation/lead iodide solution after heating and cooling. From left to right: acetamidinium lead iodide, formamidinium lead iodide, guanidinium lead iodide, imidazolium lead iodide, and methylammonium lead iodide.

ALI

FLI

GLI

ILI

MALI

Figure 3. Sample slides of each 0.88 M organic cation/lead iodide solution after exposure to air for 23 hours. From left to right: acetamidinium lead iodide, formamidinium lead iodide, guanidinium lead iodide, imidazolium lead iodide, and methylammonium lead iodide.

Fig. 1b displays the absorbance capabilities of the same samples of lead iodide with various organic cations after they were exposed to air for 23 hours. The absorption peaks that had been present in Fig. 1a have now significantly decreased, implying that samples do not absorb light as well after prolonged exposure to air. In comparison to Fig. 1a, the spectra of acetamidinium lead iodide and imidazolium lead iodide changed the least, suggesting that those samples are the

180

The Manhattan Scientist, Series B, Volume 4 (2017)

Skuriat

most stable. Fig. 3 shows the sample slides after being exposed to air for 23 hours. All samples have faded, and they have become yellow and pale. This observation corresponds to the decreased ability of each sample to absorb a large portion of visible light. To explore synergistic effects between mixtures of the various cations, the standard methylammonium lead iodide was mixed in 25% - 75%, 50% - 50%, and 75% - 25% ratios with the four other organic cation/lead iodide solutions. The methylammonium lead iodide/acetamidinium lead iodide (MALI - ALI) mixes, particularly the 50% MALI - 50% ALI and 75% MALI - 25% ALI mixtures, have peaks in the blue and red regions of light. The methylammonium lead iodide/formamidinium lead iodide (MALI - FLI) mixes, specifically the 50% MALI - 50% FLI and 75% MALI - 25% ALI mixes also displayed that broad range of absorption. This indicates that the mixtures of the materials have absorption properties of both types of materials. The methylammonium lead iodide/guanidinium lead iodide mixes primarily absorbed light in the blue region, and the methylammonium lead iodide/imidazolium lead iodide mixes were generally stable.

Discussion Methylammonium lead iodide and formamidinium lead iodide samples provide the best range of light absorption; however, they suffer from stability issues. From the mixes, the methylammonium lead iodide/acetamidinium lead iodide and methylammonium lead iodide/ formamidinium lead iodide mixes displayed a broad range of absorption. Imidazolium lead iodide, as well as the methylammonium lead iodide/imidazolium lead iodide mixes, were the most stable. An optimal solar cell can be created by combining materials that can absorb light well and/or remain stable over time. The next step of this research is to use an X-ray diffractometer to test all samples created. This will determine the crystal structure of each sample and determine if we have a new mixture crystal lattice, or a mixture of two individual components. Once this step is complete, the different samples will be placed in a solar cell and tested for photovoltaic response.

Acknowledgements This work was funded by the Manhattan College Jasper Summer Scholars Program. The Department of Chemistry and Biochemistry provided the equipment used in this research project. The author would like to thank Dr. Alexander Santulli for his guidance and mentorship.

References [1] “Perovskites and Perovskite Solar Cells: An Introduction.” Ossila, www.ossila.com/pages/ perovskites-and-perovskite-solar-cells-an-introduction [2] Atomic Movies May Help Explain Why Perovskite Solar Cells Are More Efficient. 26 July 2017, phys.org/news/2017-07-atomic-movies-perovskite-solar-cells.html. [3] “Perovskite Solar.” Perovskite-Info, www.perovskite-info.com/perovskite-solar.

Study of the interaction between mefenamic acid, fludarabine and human serum albumin by spectroscopic methods Ewa Swiechowskaâ&#x2C6;&#x2014; Department of Chemistry and Biochemistry, Manhattan College Abstract. This study examines interactions between human serum albumin (HSA) and two drugs, fludarabine (anticancer drug) and mefenamic acid (anti-inflammatory). The interactions were studied in two buffers (tris-HCl and hepes) of the physiological pH and at different temperatures using UV/visible absorption and fluorescence spectroscopy. The interaction between drugs and plasma proteins is an important pharmacological parameter which influences the effectiveness of many drugs. HSA is a principal extracellular protein with a high concentration in blood plasma and the carrier for many drugs to different molecular targets. This study investigates the parameters of binding such as binding constant, number of binding sites, and nature of binding force as well as thermodynamic parameters associated with the binding process. The data indicates that one drug does not bind with HSA at the Sudlow site I located in the subdomain IIA, while another drug binds strongly.

Introduction Human serum albumin (HSA) is the most abundant protein found in the human blood plasma [1]. It reversibly binds to drugs and transports them to their target site. Most drugs are transported to the target cell in the circulatory system with an albumin; therefore, the drug binding makes an important factor in the pharmacokinetic behavior of many drugs, affecting their active concentration and rate of delivery. In general, drugs have two forms in circulation, bound or unbound to plasma proteins. In the blood stream, drugs are transported partly in the solution as free or unbound and partly reversibly bounded to blood components such as HSA. Only, the unbounded drugs can passively diffuse through the barriers into organs where they can reach the target site and where the pharmacological effects of the drug occur. Some drugs bind too strongly to proteins circulating in the blood and cannot be released to their target sites. Therefore, the tight binding of to HSA makes it difficult for the drugs to leave the bloodstream and reach the target site, affecting the active concentration of the drug. Human serum albumin is a single chain globular protein of 585 amino acid residue including tryptophan [2, 3]. HSA contains three homologous domains (named I, II and III), where each domain is made up by two separate helical subdomains (named A and B) connected by random coil. According to Sudlow nomenclature, there are two specific drug binding sites on HSA, namely site I and site II. Sudlow site I is located in subdomain IIA and the Sudlow site II is located in subdomain IIIA of HSA. With the presence of tryptophan, a non-polar amino acid with florescence properties located in the IIA subdomain, the drug binding with HSA can be studied using fluorescent spectroscopy. Tryptophan is highly sensitive to its environment, and the changes of its emission spectra often occur in response to the substrate binding [2]. â&#x2C6;&#x2014;

Research mentored by Jianwei Fan, Ph.D.

182

The Manhattan Scientist, Series B, Volume 4 (2017)

Swiechowska

In this study, interactions between two drugs, fludarabine and mefenamic acid with HSA were investigated. Fludarabine, an anti-cancer drug along with its analogues effectively interfere with the DNA synthesis and stops the cancer cell growth. However, its binding to HSA has not been yet studied. Mefenamic acid and many other anti-inflammatory drugs are widely used and studied due to their general pain-reliving characteristics, where the interactions between drug and plasma proteins have a strong influence on their pharmacodynamics behavior. This study aims to explore the binding of two drugs with HSA using UV/visible absorption and fluorescence spectroscopy. The characteristics of the binding, i.e. binding constant, number of binding sites, and the nature of the binding force are determined. In addition, the thermodynamic parameters associated with the binding process are also calculated. Through this study it was found that only mefenamic acid binds with HSA and itâ&#x20AC;&#x2122;s binding constant at different temperatures was determined.

Figure 1. Fludarabine

Figure 2. Mefenamic acid

Experimental Materials HSA and mefenamic acid were purchased from Sigma Aldrich and Fisher, respectively. The fludarabine was purchased from Sigma Aldrich. The 0.05 M tris-HCl buffer (pH = 7.4) and 0.05 M hepes buffer (pH = 7.4) were prepared using analytical reagent grade. All solutions were prepared with ultrapure deionized water. Physical measurements The fluorescence spectra of all solutions were measured with a Photon Technology International (PTI) spectrofluometer equipped with a 1.0 cm quartz cell connected to the thermostat bath. The binding parameters were measured at three different temperatures (15C, 25C and 35 C). The exaction wavelength of HSA was set to 279 nm, and the emission spectra were taken in the range of 290 nm to 450 nm. The slit width of exaction of monochromator was set to 1.50 mm, whereas the emission of monochromator was set to 1.00 mm. The absorption spectrum was recorded with an Agilent 8453 UV/visible photodiode array spectrophotometer.

Swiechowska

The Manhattan Scientist, Series B, Volume 4 (2017)

183

Results and Discussion Part 1: HSA with fludarabine UV/visible absorption spectra The spectra of 10 µM HSA and 1 µM fludarabine are shown in Fig. 3. The absorption maximum of HSA is at 279 nm and the absorption maximum of fludarabine is at 261 nm. Fig. 3 shows that the drug does not interfere with the absorption of HSA.

Figure 3. Absorption spectra of a) 10 µM HSA (λmax = 279 nm); b) 1 µM fludarabine (λmax = 261 nm).

Fluorescence quenching spectra The fluorescence intensity of HSA with an addition of fludarabine shows no quenching due to the drug (Figs. 4 and 5). Both drug and water were added to HSA in the increasing amounts resulting in the same decrease. Therefore, it was concluded that the fludarabine does not interact with HSA at the Sudlow site I located in the subdomain IIA and the steady quenching is caused by the dilution. Part 2: HSA with mefenamic acid UV/visible absorption spectra The spectra of 10µM HSA and 127µM mefenamic is in Fig. 6. The absorption maximum of HSA is at 279 nm and the absorption maximum of mefenamic acid is at 286 nm. Fig. 6 shows the interference of drug with HSA. Due to the closeness of the absorption maxima of HSA and mefenamic acid when the excitations photons are set to 279 nm, the inner filter effect by mefenamic acid is present. Inner filter effect is the competition for excitation photons between HSA and mefenamic acid, which results in the apparent decrease of fluorescence intensity of HSA [4]. Therefore, the exact fluorescence intensity was calculated to correct for the co-absorption of excitation photons by mefenamic acid.

184

The Manhattan Scientist, Series B, Volume 4 (2017)

Swiechowska

Figure 4. Emission spectra of HSA with increasing fludarabine concentration in tris-HCl (left) and hepes (right) buffer.

Figure 5. Emission spectra of HSA with increasing water concentration in tris-HCl buffer.

Swiechowska

The Manhattan Scientist, Series B, Volume 4 (2017)

185

Figure 6. The absorption spectra of 10 µM HSA (λmax = 279 nm) and 127 µM mefenamic acid (λmax = 286 nm).

Fluorescence quenching spectra The emission spectra of 10µM of HSA in Tris-HCl in the absence and presence of mefenamic acid are shown in Fig. 5. The emission spectrum of HSA was run at the 279 nm and the changes in the fluoresce intensity were observed. The addition of increasing amounts of 127 µM mefenamic acid from 50 µL to 120 µL resulted in decrease of the fluorescent intensity at 333 nm. The emission intensity of HSA decreased, but the extent of the quenching gradually slowed down. The largest quenching was achieved when 50 µL of mefenamic acid was added corresponding to 127 µM mefenamic acid added and the molar concentration ratio of HSA to mefenamic acid reached around 1 to 1 (Fig. 7). The peak at 305 nm was due to the water scattering. Stern-Volmer quenching constant The observed emission intensity of fluorescence was corrected for inner filter effect using the following equation [4, 5] F corr = Fobs × eA/2 (1)

where F corr and F obs are the fluorescence intensities corrected and observed, respectively. A is the absorbance of mefenamic acid at the excitation wavelength (279 nm) at the same concentration as it is in the mixture. The equation holds true when the absorbance is less than 0.3. The corrected intensity was then fitted with the Stern-Volmer equation [4] F 0 /F corr = 1 + Ksv [Q]

(2)

where Ksv is the Stern-Volmer quenching constant, and [Q] is the concentration of the quencher. F 0 is the emission intensity in the absence of the quencher and F corr is the corrected emission intensity in the presence of the quencher. Fig. 8 displays F 0 /F corr of HSA against the concentration of mefenamic acid in two different buffers at three different temperatures. The slope of the line gives Ksv values, which are summarized in Table 1. It can be seen from the Fig. 8 that the Ksv

186

The Manhattan Scientist, Series B, Volume 4 (2017)

Swiechowska

Figure 7. Emission spectra of HSA with increasing mefenamic acid concentration: a) tris-HCl buffer, b) hepes buffer

values decrease as the temperature increases in hepes buffer, and that the Ksv are higher in the hepes buffer than in tris-HCl buffer at the same temperatures (Table 1). Overall, the Ksv values decrease with an increase in temperature, however in tris buffer there is a slight increase in Ksv 288 K

298 K

308 K

F0 / Fcorr

3 2 1

298 K

308 K

2 1 0

0 0

288 K

Mefenamic acid (μM)

Figure 8. Stern-Volmer plots of HSA at three different temperatures: a) in tris buffer and b) in hepes buffer.

from 25◦ to 35◦ C due to the environmental factors such as temperature control. A future study will

Swiechowska

The Manhattan Scientist, Series B, Volume 4 (2017)

187

reexamine the temperature dependence on the quenching at higher temperatures (25◦ C and 35◦ C). Table 1. Ksv , Ka and n at three temperatures in tris and in hepes buffer. Tris buffer Temp (K) 288 298 308

Ka × 105 M−1 373 38.7 5.16

Hepes buffer

Ksv × 105 M−1 3.06 1.65 2.75

n 1.33 1.21 1.002

Ka × 106 M−1 4.04 1.96 1.11

Ksv × 105 M−1 4.59 4.13 4.10

n 1.15 1.12 1.07

Quenching mechanism The Stern-Volmer equation models what is called dynamic quenching, quenching which occurs when the quencher diffuses through solution and interacts with the luminescent species, resulting in the deactivation of the excited state. The Stern-Volmer constant is Ksv = Kq × τ , where Kq is the bimolecular quenching rate constant, and τ is the lifetime of the excited state of HSA, which equals 1 × 10−8 s [4, 9]. The maximum value of kq is kd , the diffusion rate constant in aqueous solutions, which is 1010 M−1 s−1 , or kq ≤ kd = 1010 M−1 s−1 [10]. However, the value of kq calculated from the obtained Ksv value and the t of HSA is much larger than kd , i.e. kd = Ksv /τ = 1 × 105 /1 × 10−8 = 1013 1010 . This indicates that the dynamic quenching mechanism in unlikely the dominate process [4]. Binding parameters and thermal dynamic functions The quenching mechanism is believed to be static since both Ksv and Ka decrease with an increase in temperature. For pure static quenching, the Stern-Volmer constant, Ksv equals Ka , the binding constant of the mefenamic acid with HSA. The Ka is derived from the following binding process [6] P + nD ⇔ Dn P (3) where P is the free protein, D is the drug or the quencher, n is the number of binding sites on the protein, and Dn P is the protein drug complex. The binding constant is obtained through the following equation using the mass action law Ka =

[DnP ] [P ][D]n

(4)

Since the free protein is the only fluorescent species, [P ] Fcorr = [P ]0 F0

(5)

188

The Manhattan Scientist, Series B, Volume 4 (2017)

Swiechowska

where the [P ]0 is the initial concentration of free protein, Fcorr and F0 is the emission intensities of protein with and without the presence of the drug. The following double log equation can be derived by substituting Eq. (5) into Eq. (4) (6)

log[(F0 − Fcorr )/Fcorr ] = log(Ka ) + n log[D].

Eq. (6) is used (see Fig. 9) to derive the binning constant (Ka ) and the number of binding sites (n) on the protein. 288 K

298 K

0.4

308 K

0.0 ‐0.2 ‐0.4 ‐0.6 ‐0.8

288 K

308 K

298 K

0.2

Log [(F0‐Fcorr)/Fcorr]

0.2

‐6

‐5.9

‐5.8

‐5.7

‐5.6

Log [Mefenamic acid]

‐5.5

‐5.4

0.0 ‐0.2 ‐0.4 ‐0.6 ‐0.8 ‐1.0 ‐6.4 ‐6.3 ‐6.2 ‐6.1

‐6

‐5.9 ‐5.8 ‐5.7 ‐5.6 ‐5.5

Log [Mefenamic acid]

Figure 9. log[(F0 − Fcorr )/Fcorr ] vs. log[mefenamic acid] in a) in tris-HCl buffer and b) in hepes buffer at three temperatures: 288K, 298K, and 308K.

Non-steroidal anti-inflammatory drugs are known for their binding capacity for HSA. The binding site for mefenamic acid was determined to be one. This study found that the binding site for mefenamic acid is one located in the subdomain IIA. HSA contains several residues including tryptophan, an amino acid with fluorescence properties located in the subdomain IIA. Thermodynamic functions In order to explain the nature of the interaction between mefenamic acid and HSA, the binding constant was determined at three temperatures (288, 298 and 303K). The thermodynamic parameters of binding, entropy changes (∆S ◦ ) and enthalpy changes (∆H ◦ ), were obtained from Van ’t Hoff equation 1 log(Ka ) = (−∆H ◦ + T ∆S ◦ ) (7) 2.303 RT where Ka is the binding constant at the given temperature and R is the universal gas constant. The values of ∆S ◦ and ∆H ◦ were obtained from the intercept and the slope of the Van ’t Hoff plot (Fig. 10). The free energy change (∆G◦ ) was calculated using Gibbs equation, ∆G◦ = ∆H ◦ − T ∆S ◦ . All calculated parameters are listed in the Table 2. Force contributing to binding The type of binding force can be deduced from the sign of the thermodynamic functions [7, 8, 11]. According to Ross and Subramanian [11], the hydrophobic interactions occur if ∆H > 0 and ∆S > 0 since the process includes the reorganization of the solvent structure around the protein

Swiechowska

The Manhattan Scientist, Series B, Volume 4 (2017) 8

189

tris‐HCl buffer

Log (Ka)

Hepes buffer 7

5 320

330 1/T * 10‐5

340

350

(K‐1)

Figure 10. log (Ka ) vs. 1/T at three different temperatures: 288K, 298K, and 308K. The ∆S and ∆H were determined from the y-intercept and the slope, respectively. Table 2. Thermodynamic parameters in tris and in hepes buffer. Tris Buffer

Hepes Buffer

Temp (K)

∆H (kJ/mol)

∆S (J/molK)

∆G (kJ/mol)

∆H (kJ/mol)

∆S (J/molK)

∆G (kJ/mol)

288 298 308

-158.1

-403.7

-41.8 -37.8 -33.8

-47.7

-39.2 -36.0 -35.6

-36.4

and the ligand species, however if ∆H < 0 and ∆S > 0 the force is electrostatic and if ∆H < 0 and ∆S < 0, then the forces are mainly van der Walls and hydrogen bonding. In both buffers, both ∆H and ∆S are negative which indicates that the binding interaction between HSA and mefenamic acid is mainly van der Waals forces and hydrogen bonding. The negative ∆H values are indicative of the presence of H-bonding, especially in the tris buffer.

Conclusion The accumulated data indicates that only mefenamic acid does quench the fluorescence of HSA and affects the microenvironment of the tryptophan in the domain IIA of HSA. The quenching mechanism is believed to be static by forming a complex since both Ka and Ksv values decrease with an increase in temperature and are similar in value. Mefenamic acid contains one binding site located in the subdomain IIA of HSA and the binding force is mainly van der Walls and hydrogen bonding. The reaction between HSA and mefenamic acid is exothermic and spontaneous. This study also investigated the effects of buffers on the interaction between HSA and mefenamic acid. In tris buffer, the thermodynamic parameters obtained from temperature-dependent measurements indicate that the binding between HSA and mefenamic acid is mainly van der Walls and hydrogen bonding, however the electrostatic forces are not excluded. This study also indicated that HSA is more stable in the tris buffer, which might contribute to selection of buffers in the fu-

190

The Manhattan Scientist, Series B, Volume 4 (2017)

Swiechowska

ture. Furthermore, the methods established in this study will serve as a basis for the future study of different drugs.

Acknowledgment This research was funded by the School of Science Research Scholars Program. The author would like to express her gratitude to her research mentor, Dr. Jianwei Fan, and to Dr. John Regan for his valuable feedback at different stages of the research, as well as providing chemicals.

References [1] Motoharu I, Hideo N, Nishigorikazuo M. The Classification of Drugs on the Basis of the Drugbinding Site on Human Serum Albumin. Chemical and Pharmaceutical Bulletin. 1982; 30(12): 4489-93. [2] Siemiarczuk A, CE Petersen, CE Ha, J Yang, and NV Bhagavan. Analysis of tryptophan fluorescence lifetimes in a series of human serum albumin mutants with substitutions in subdomain 2A. Cell Biochem Biophys. 2004;40(2):115-22. [3] Taverna M, Maire AL, Mira JP, Guidet B. Specific antioxidant properties of human serum albumin. Ann Intensive Care. 2013; 3:4. [4] Lakowicz JR. Principles of Fluorescence Spectroscopy, 3rd Edition, Springer, New York, NY 2006. [5] Anbazhagan V, Renganathan R. Study on the binding of 2,3-diazabicyclo [2.2.2] oct-2-ene with bovine serum albumin by fluorescence spectroscopy. J Lumin. 2008; 128(9):1353-8. [6] Singh TS, Mitra S. Interaction of cinnamic acid derivatives with serum albumins: a fluorescence spectroscopic study. Spectrochim Acta Part A: Mol Biomol Spectros. 2011; 78(3):942-8. [7] Amidon GL, Anik ST. Hydrophobicity of polycyclic aromatic compounds. Thermodynamic partitioning analysis. J Phys Chem. 1980; 84(9): 970-4. [8] Leckband D. Measuring the forces that control protein interactions. Annu Rev Biophys Biomol Struct. 2000; 29:1-26. [9] Feliciano M, Kroger M, Irizarry J, Prentzas S, Fan J and Wang E. Spectroscopic Study of the Interaction between Dipicolinic Acid and Human Serum Albumin. J Res Anal. 2016; 2(4): 102-107. [10] Islam MM, Sonu VK, Gashnga PM. Caffeine and sulfadiazine interact differently with human serum albumin: A combined fluorescence and molecular docking study. Spectroshim Acta Part A Mol Biomol Spectrosc. 2016; 153:22-33. [11] Ross PD, Subramanian S. Thermodynamics of protein association reactions: forces contributing to stability. Biochemistry.1981; 20(11):2096-102.

Image recognition using autoencoding in multilayer neural networks and multi-value neurons Niko Colon* and Alexander Gonzalez* Department of Computer Science, Manhattan College Abstract. Deep learning neural networks are often used in image and pattern recognition, because of their ability to extract features from patterns and then recognize patterns based on those features. With a deep learning Multilayer Neural Network with Multi-Valued Neurons (MLMVN) we should be able to learn images or other patterns better than with a real-valued neural network. This paper will present our experiment on image recognition using the MLMVN network and autoencoding as one of the utilizations of the deep learning concept. In this work, we employ the commonly used benchmark, the MNIST dataset, consisting of 60,000 images of handwritten digits ranging from 0 to 9, which are used for learning. We will also test the MLMVN network with a test set of 10,000 images suggested by the creators of the MNIST dataset for testing.

Introduction A neural network is an intelligent tool that is inspired by the way the brain processes information. The key element of a neural network is the large number of highly interconnected processing elements (neurons) working in unison to solve specific problems. The network gathers experimental knowledge acquired through a learning process that is not pre-defined. Instead, the network learns from the environment formed by learnings samples that are united in a learning set [1]. A neuron learns from its environment by taking samples from the learning set and adjusting its weights according to the learning algorithm, in order to implement a corresponding input/output mapping (Fig. 1). The neuron then processes its inputs by taking the weighted summation of all the

Figure 1. Model of an artificial neuron and itâ&#x20AC;&#x2122;s learning process (upload.wikimedia.org/wikipedia/commons/6/60/ Artif icialN euronM odel english.png)

input values using the weights containing knowledge acquired as a result of the learning process. â&#x2C6;&#x2014;

Research mentored by Igor Aizenberg, Ph.D., and Lawrence Udeigwe, Ph.D.

192

The Manhattan Scientist, Series B, Volume 4 (2017)

Colon + Gonzalez

The learning process is a procedure where weights should be adapted using a specific learning algorithm to capture the knowledge. The aim of the learning process is to map a given relation between inputs and outputs of a neuron. A nonlinear activation function is used to transform a weighted sum of inputs into output. A neural network learns in the same way. The only difference is that in the network all its neurons contribute to the implementation of the corresponding input/output mapping. A feedforward neural network, also referred to as a Multilayer Perceptron (MLP), is a neural network where neurons are grouped in layers, and outputs of all neurons from the preceding layer are connected to the corresponding inputs of neurons from the current layer (Figs. 2 and 3) [2]. In a feedforward neural network, its input layer does not contain neurons, only the inputs distributed among the first hidden layer neurons. All layers of neurons except the last one are called “hidden layers,” because their desired output is not known in advance.The last layer producing the output of the network is called an “output layer.”

Figure 2. Structure of a Multi-Layer Perceptron (MLP)

Figure 3. Structure of Multi-Layer Multi-Valued Neural Network (MLMVN). All network mentioned in testing have this structure, the only differences being the number of neurons in the hidden layer and the input layer containing the input features of the samples and not neurons.

One of the main applications of artificial neural networks is solving classification problems. It is difficult or even impossible to find a decision rule because of the numerous possible ways to determine membership of objects belonging to different classes. When solving well defined classi-

Colon + Gonzalez

The Manhattan Scientist, Series B, Volume 4 (2017)

193

fication problems, typically we know the features based on which certain objects can be classified. This is completely different from image recognition problems, because a set of formal features based on which certain images should be recognized can be difficult to find. Image recognition problems are among the most challenging due to the inability to formalize features, such as shapes, in terms of some measurable units. Deep learning is a machine learning technique that is helpful for solving challenging problems such as image recognition, speech recognition, and computer translation [3, 4, 5]. Deep learning is based on learning pattern (data) representations, as opposed to pattern descriptions based on some formalized features. This approach allows specific patterns to be extracted during the learning process that is suitable for machine learning, instead of creating features based on the researcherâ&#x20AC;&#x2122;s point of view. Deep learning is utilized using MLP containing multiple layers. This makes it possible to extract useful features from the data to be learned by processing them in a layer by layer flow. In such a case, it should be expected that first layers extract features from the data, while the following layers learn from these features. The most popular deep learning architectures are a convolutional multilayer neural network and autoencoding using MLP. Autoencoding is an approach where MLP shall first be trained to reproduce patterns to be recognized. After this process is complete, it should be expected that a part of the multilayer neural network has extracted some specific features from the data The next part of the network can reproduce the data, composing them from these features. Since these extracted features can be used to reproduce the data, they should also be used to recognize patterns represented by the data. To perform recognition, that second part of the network that was used for reproduction of the data, should be substituted with a new part, which can be trained to recognize patterns representing by these data. We will employ autoencoding here. At the same time we will implement autoencoding not using MLP, but a different tool. A Multilayer Neural Network with Multi-Valued Neurons (MLMVN) (Fig. 3) is a type of neural network that has multiple layers where each neuron (multi-valued neuron (MVN)) operates with complex-valued inputs, outputs, activation function, and weights [6]. The advantages of using multivalued neurons is that MVN makes it possible to examine real world problems, like signal processing in the frequency domain, where complex numbers result from a Fourier transform. MVN is also more functional than any real-valued neuron with regard to real-valued input/output mappings. Respectively, MLMVN is a more functional network than MLP; this was proven by many applications. It allows faster learning, and also generalizes better than a real-valued network (neuron), giving it an advantage when it comes to real world applications [6, 7, 8]. Image recognition is the process of identifying and detecting an object or feature in a digital image or video. In this work, we employ the MNIST dataset (Fig. 4), the commonly used benchmark consisting of 60,000 images of handwritten digits ranging from 0 to 9, which are used for learning. We will also test the MLMVN network with a set of 10,000 images, suggested by the creators of the MNIST dataset [9].

194

The Manhattan Scientist, Series B, Volume 4 (2017)

Colon + Gonzalez

Figure 4. Some of the images that are used in the MNIST learning dataset [9]

Autoencoding and its implementation One of the most popular applications of neural networks is pattern recognition. Particularly image recognition is really a pressing problem. Recently, a number of researches employed two deep learning techniques in their image recognition experiments with the MNIST dataset. The best result for the MNIST dataset so far is 99.65% classification rate obtained using MLP 784-25002000-1500-1000-500-10, meaning, 784 inputs corresponding to the number of pixels in 28Ă&#x2014;28 images, 2500 neurons in the first hidden layer, 2000 neurons in the second hidden layer, 1500 neurons in the third hidden layer, 1000 neurons in the fourth hidden layer, 500 neurons in the fifth hidden layer, and 10 output neurons corresponding to 10 different images, which a network had to recognize [10]. We would like to implement autoencoding followed by the recognition of the corresponding images using MLMVN. We wanted to attempt to recognize images from the MNIST dataset using MLMVNs with different topologies, but with a significantly smaller number of neurons than the one used in [10]. From the recognition standpoint the goal was the same: to train a network using the same 60,000 learning samples from the MNIST dataset and test the results using 10,000 test samples from the same dataset. We are expecting that the MLMVN should show better results and learn faster, because of its proven advantages over real-valued neural networks. If this expectation holds true, then this will show a more efficient way for pattern recognition for all images. It is important to note that our initial plan was to use a supercomputer with powerful GPUs, where a big neural network can be utilized by employing parallel GPU-based processing using CUDA. Unfortunately, a supercomputer was not available during the summer and instead we had to use a conventional desktop computer. This drastically complicated our task because serial implementation of big neural networks is computationally very inefficient and the most desirable experiments simply cannot be performed with 60,000 learning samples because of the very limited resources of any conventional computer, even one with a powerful CPU.

Colon + Gonzalez

The Manhattan Scientist, Series B, Volume 4 (2017)

195

Methodology

Our experiment can be broken down into three main phases. The first phase was to create two MLMVNs and autoencode both networks. The first network had topology 784-1024-1024-1024784 (thus 784 inputs, three hidden layers with 1024 neurons each, and an output layer with 784 neurons). The second network has the same amount of input and output neurons with the only difference being three hidden layers with 2048 neurons each. Both networks were initialized using random weights for their neurons and learned 1200 samples from the MNIST learning set, due to the limitation of conventional computer memory. Autoencoding is an unsupervised learning algorithm that reproduces the inputs as the outputs, which shows why the input and output layers have the same amount of neurons (Fig. 5). This should make it possible for a network to extract features from each sample that will help it recognize the digit in the next phase of our experiment.

Figure 5. The learning process of autoencoding a MLMVN network

The second phase of our experiment was to split half the neural network, preserving half the weights from the old networks and creating new layers with random weights. This should create new weights that will be more efficient for image recognition than the previous weights. By also changing the output layer, it will now have 10 neurons instead of 784 neurons in order to signify one of the possible 10 digits it can be. Then, the networks learned the same 1200 samples again in order to recognize images based on the results of autoencoding.

Experimental Results

The last phase of our experiment was to test how networks can recognize the test images from the MNIST dataset of 10,000 samples that the networks have not seen. Unfortunately, the results of both networks tested were not at the hoped for level. Both networks were only tested with 200 samples out of the 10,000 test set sample, because of the limitation of the amount of samples it could learn from the learning set. The network with 3 hidden layers of 1024 neurons was able to recognize about 51% of the digits or 102 samples (Fig. 6). The network with 3 hidden layers of 2048 neurons was able to recognize 42% of the test sample digits or 85 samples (Fig. 6). Thus, our experiment was changed in order to achieve better results. We created a new network 784-1024-1024-1024-784, i.e. input layer of 784 neurons, three hidden layers with 1024 neurons

196

The Manhattan Scientist, Series B, Volume 4 (2017)

Colon + Gonzalez

Figure 6. The results of recognition when testing the 1024 MLMVN and the 2048 MLMVN networks with the test set.

each, and an output layer with 784 neurons. Instead of only 1200 learning samples we decided to use the batch learning approach in order to bypass the computer memory limitation. This approach allows the network to learn the entire 60,000 learning set in increments of 500 samples. Once the network learned the samples, we saved the weights in a matrix and then, using the weights gained from the previous learn samples, we learned the next increment of samples. The next phase was mostly the same: split the network, generate random weights for new output layer containing 10 neurons preserving the weights resulting from autoencoding for other neurons. Our learning experiment for this network is still ongoing. After 8 passes over the entire learning set containing 60,000 samples this network was able to recognize correctly 75.13% of the test samples from the entire test set of 10,000 samples (Fig. 7). This recognition accuracy is increasing by about 1.2%

Figure 7. The result of recognition when testing the 1024 MLMVN network with the batch learning approach with the test set.

Colon + Gonzalez

The Manhattan Scientist, Series B, Volume 4 (2017)

197

after each pass through the entire learning set. However, each pass with a 500-sample batch takes about 70 hours on the conventional desktop computer.

Conclusions and Future Work Looking at our results for batch learning one concludes that it can recognize 75.13% of the test set. This significant improvement in results is due to the fact that this network was able to learn the entire learning set. The learning process is still running in a loop over the entire learning set. Our goal is to achieve the recognition rate of about 99%. When examining all the networks together one can see that, in general, the network had trouble recognizing the digits 3, 5, and 8. This was most likely due to the fact that these digits have loops that make it harder for the network to tell the difference between each other. One can also infer, based on the result of the 1024-neuron network that went through the batch learning approach, that if one were able to use a computer with a higher memory and with CUDA-enabled GPU, one should have significantly better results than the ones already achieved. In the future it is possible to use the School of Science Dionysos computer with 24-thread CPU, 64GB of DDR3 RAM, and 4×Tesla k20 CUDA GPUs. This will make it possible to drastically speed up the learning process.

Acknowledgments This work was funded by the School of Science Research Scholars Program. The authors would like to thank Dr. Igor Aizenberg for the support and guidance throughout the research, and Dr. Lawrence Udeigwe for mentoring and aiding in the understanding of the mathematics behind the components of the networks.

References [1] I. Azenberg. “Complex-valued neural networks with multi-valued neurons.” Springer (2011) [2] Architecture of Neural Networks - Stanford Computer Science, https://cs.stanford.edu/people /eroberts/courses/soco/projects/neural-networks/Architecture/feedforward.html [3] F. Gomez and J. Schmidhuber. “Co-evolving recurrent neurons learn deep memory POMDPs.” Proc. GECCO, Washington, D. C., pp. 1795-1802, ACM Press, New York, NY, USA (2005) [4] G. E. Hinton, S. Osindero, and Y. W. Teh. “A Fast Learning Algorithm for Deep Belief Nets” (PDF). Neural Computation. 18 (7): 1527–1554. PMID 16764513. doi:10.1162/neco. 2006.18.7.1527 (2006) [5] “Autoencoders.” Unsupervised Feature Learning and Deep Learning Tutorial, ufdl.standford. edu/tutorial.unsupervised/Autoencoders/. [6] I. Aizenberg and C. Moraga, “Multilayer Feedforward Neural Network based on Multi-Valued Neurons and a Backpropagation Learning Algorithm,” Soft Computing, vol. 11, No 2, pp. 169-83 (2007)

198

The Manhattan Scientist, Series B, Volume 4 (2017)

Colon + Gonzalez

[7] I. Aizenberg, D. Paliy, J. Zurada, and J. Astola, “Blur Identification by Multilayer Neural Network based on Multi-Valued Neurons,” IEEE Transactions on Neural Networks, vol. 19, No 5, pp. 883-898 (2008) [8] I. Aizenberg, “MLMVN with Soft Margins Learning,” IEEE Transactions on Neural Networks and Learning Systems, vol. 25, No 9, pp. 1632-1644 (2014) [9] THE MNIST DATABASE. (n.d.). Retrieved October 03, 2017, from http://yann.lecun.com /exdb/mnist/ [10] Claudiu Dan Ciresan, Ueli Meier, Luca Maria Gambardella, and Juergen Schmidhuber. “Deep Big Simple Neural Nets Excel on Handwritten Digit Recognition.” Neural Computation, 22 (12) (2010)

Image recognition using analysis of discrete Fourier transform by multilayer neural networks with multi-valued neurons Alexander Gonzalez* and Niko Colon* Department of Computer Science, Manhattan College Abstract. This work investigates the use of a complex-valued neural network for the recognition of images represented using the low frequency phases of their Fourier transforms [1]. The networks with different topologies were trained using a learning set and then the results of learning were tested using a test set for their recognition rates.

Introduction A neural network is a massively parallel distributed processor that has a natural propensity for storing experimental knowledge and making it available for use (Fig. 1). This means that: 1) knowledge is acquired by the network through a learning process; 2) the strength of the interconnections between neurons is implemented by means of the synaptic weights used to store the knowledge [2]. The inception of the artificial neural network was inspired by the realization that the human brain and the conventional digital computer have an entirely different computing process [3]. While a conventional computer processes any information based on a program designed according to some algorithm and loaded by the user, a brain, as well as an artificial neural network, operate based on the knowledge they acquire as a result of the learning process.

Figure 1. Model of an artificial neuron

The most basic and fundamental element of a neural network is an artificial neuron, which is an information processing unit that is an abstract model of a biological neuron [2, 3]. Artificial neurons interact with each other similarly to the way biological neurons in an animal brain interact with each other, which sets the foundation for the learning process. The artificial neuron organizes a set of weights corresponding to its inputs, and the knowledge obtained through the learning process is stored in these weights [2, 3]. The weights are formatted as real-valued numbers, and â&#x2C6;&#x2014;

Research mentored by Igor Aizenberg, Ph.D., and Lawrence Udeigwe, Ph.D.

200

The Manhattan Scientist, Series B, Volume 4 (2017)

Gonzalez + Colon

they are used to process the real-valued inputs. Each input is multiplied by its corresponding weight. The sum of these products is called the weighted sum of the neuron. To produce the overall output of the neuron, the weighted sum is passed through a nonlinear activation function, which transforms the weighted sum limiting the range of output and adapting it to a specific nonlinear input/output mapping implemented by the neuron (Fig. 2).

Figure 2. Structure of a feedforward neural network. All MLMVN networks mentioned in testing have this structure, however with differences in the number of neurons in the hidden layer. NOTE: The input layer does not consist of neurons, rather just the input features of the samples that get directly input to the hidden layer.

We would like to use an artificial neural network for solving image recognition problems. These problems are difficult to formalize because it is very difficult to find any formal set of features based on which a neural network can learn the problem to be able to generalize, and thus recognize images not participating in the learning set. In our attempt to recognize images, we use a Multilayer Neural Network that is composed of Multi-Valued Neurons (MLMVN) [2, 4]. This type of network can be classified under a larger category of Complex-Valued Neural Networks (CVNN), networks that use complex-valued numbers as their weights, inputs and outputs [2]. This is different than most other networks, which use real-valued numbers as their weights. The Multi-Valued Neuron (MVN) uses complex numbers as its weights and implements a phase-dependent activation function. This activation function is a function of the argument of a weighted sum. MVN classifies its inputs using multi-valued threshold logic over the field of the complex numbers. The activation function takes the argument of the weighted sum and maps it onto the unit circle. The unit circle is divided up into k-sectors where, for example in classification problems, k is the number of different classes of objects. Traditional neural networks usually utilize a sigmoidal activation function, which can be treated as a generalization of the threshold activation function implementing Boolean logic. In the latter case, there are two â&#x20AC;&#x153;truth valuesâ&#x20AC;? that the output of the neuron can be mapped to. This means that the set C2 = {0,1}, encompasses all the possible outputs of the neuron, under k-valued logic, where k = 2. A sigmoidal activation makes

Gonzalez + Colon

The Manhattan Scientist, Series B, Volume 4 (2017)

201

it also possible to learn continuous input/output mappings whose range is reduced to the interval [-1, 1]. However, the MVN operates using a phase-dependent activation function, which allows the neurons to learn and implement discrete and continuous input/output mappings. In the discrete case, the set Ck = { 0, 1, . . . , k-1} contains all the possible outputs of the neuron and represents the number of classes k, where k is a finite integer. In the continuous case, k = â&#x2C6;&#x17E;, meaning that the neuron can produce a continuous output located on the unit circle. The use of multi-valued logic makes classification easier when there are more than 2 classes of objects to be classified. The phase-dependent activation function that the multi-valued neuron utilizes allows it to be more flexible and more efficient when a neuron and a network learn. Another important advantage of this approach is its ability to work with phase and treat it properly, without breaking its circularity. The phase-dependent activation function maps the weighted sum onto the unit circle using its argument. By mapping the weighted sum onto the unit circle, we can take advantage of the kth roots of unity to create k sectors on the complex plane in which the weighted sums can fall on.

More on neural networks Artificial neural networks are computing systems used to solve problems using an intelligent approach that employs knowledge gained from an iterative learning process. They have been used to perform tasks such as pattern recognition, image recognition, speech recognition, social network filtering, medical diagnostics, etc. They are designed to loosely resemble the way neurons in the brain interact with each other and how animals such as humans learn. They are trained using a learning dataset, which is passed through the network where information about the data is extracted by the neurons. In image recognition, artificial neural networks have been used for several decades and some of the more traditional networks used are the multilayer perceptron (MLP) and the convolutional neural network. To train a neural network to recognize certain images, we must provide it with a large learning dataset. This dataset should usually be composed of many examples of images showing objects that we want to be recognized by the network. The image information is usually the intensity values of the pixels that make up the image. Basically, each image is represented by a set of numbers, and this information is what gets inputted into the network. Every image dataset is also accompanied by the desired output information, which tells the network what should be outputted when a certain image is passed through the network. The combination of an imagesâ&#x20AC;&#x2122; input data with its desired output makes up what is called a learning sample. To a network, each image passed is a sample, and that sample contains the information the network can extract from. The purpose of training a network with certain learning data, is so that it can generalize the information it has learned, and can recognize other data samples that are similar to the learning data. This way of training a neural network is essentially how animal brains learn; however in a much simpler sense. It usually takes many iterations of learning samples being passed through the network for it to learn enough information about those samples so that it can recognize samples

202

The Manhattan Scientist, Series B, Volume 4 (2017)

Gonzalez + Colon

similar to the ones that participated in the learning process. The knowledge of a network and its ability to generalize is usually tested by a set of testing samples; the latter are simply samples of images that are somewhat similar to the learning sample images, but at the same time did not participate in the learning process. If the network is successful in outputting the correct desired output for a testing sample, then that means it recognizes the sample, i.e. it recognizes the image. If the network can mostly recognize all the samples, that means it has successfully learned and has a “generalized” understanding of the images it is supposed to recognize. However, if it does not recognize enough of the testing images, then it means that the network needs to be trained by a larger set of learning samples, or that the network simply just “memorized” the information and did not really learn from it how to generalize. One can think of it as the difference between critical thinking towards solving a problem and memorization of the steps to solve a problem.

Fourier transform and the role of phase The Fourier transform of a signal breaks the signal down into frequencies present in this signal. Thus the Fourier transform is a complex-valued function of a continuous variable which, when written in polar form, is expressed in terms of its magnitude and phase [5]. Magnitude and phase are both independent functions in the sense that knowledge of one is not sufficient to deduce the other; however both are necessary to uniquely define a signal [5]. This means that both components contain certain “information” about a signal [5], and can be used to reconstruct them. In the case of images, we can utilize the knowledge of the Fourier transform magnitude and phase to reconstruct them. However, two questions arise: which one is more important, and is it possible to recognize an image using information obtained from only one of them? These ideas have been mentioned in the past in numerous papers, and we aim to employ them in our use of neural networks. Oppenheim and Lim [6] examined a series of experiments where images and spectrograms of speech patterns were stripped down to their Fourier Transform magnitude and phase, then reconstructed using only one of the two. They observed was those images and spectrograms constructed by using only phase information were more recognizable and resembled the original signals more than those constructed by using information obtained from the magnitude only. These authors’ assessment of these experiments describes how the Fourier transform phase (phase spectrum) of an image contains more of the important information than the Fourier transform magnitude (power spectrum). Furthermore, they concluded the following about images and signals: 1) the magnitude contains information about all intensities of the image information, intensities of possible noise, and information that affects the sharpness of an image (blur, contrast, brightness, etc.); 2) the phase contains information about the “location” of events in a signal [6], as well as the edges, shapes, and orientation of all objects in an image. These properties make phase much more important in the recognition of images. We explored these concepts, inspired by [6], in our own experiments, where we analyzed the reconstruction of images. In experiment A (Fig. 3), we took an image and passed it through a

Gonzalez + Colon

The Manhattan Scientist, Series B, Volume 4 (2017)

(a)

(b)

(c)

(d)

203

Figure 3. (a) Original image; (b) Phase spectrum (Fourier transform phase); (c) Power spectrum (Fourier transform magnitude); (d) The image reconstructed using the inverse Fourier transform from its original phase and unitary magnitude instead of the original one

discrete Fourier transform, which was then separated into its Fourier transform phase and magnitude. Each component was then used to construct its own image, one magnitude-only image and one phase-only image. We also constructed an image using only phase information, however, we set all the values to unitary magnitude. This experiment visualized the difference in information of phase and magnitude, while showing that some participation by both components in the reconstruction of an image can aid significantly in its recognition. In experiment B (Fig. 4), we did something similar: we used two images and instead of constructing images from their broken-down Fourier Transform components, we just reconstructed images by swapping the phase of one image with the phase of the other. We observed that the synthesized images resembled more closely the image from which its phase component was extracted. The magnitude of the other image did play a role in the construction of the synthesized image; however, it was not enough to influence the synthesized image to resemble more the image from which the magnitude was extracted. This experiment elucidates further how and why the phase is more important than the magnitude in the reconstruction of images, and also most important in their recognition.

Method and learning approach To prepare the data for learning, we first put it through a Fast Fourier Transform (FFT) algorithm that computes the Discrete Fourier Transform (DFT). This implementation of the FFT con-

204

The Manhattan Scientist, Series B, Volume 4 (2017)

(e)

(f)

(g)

(h)

Gonzalez + Colon

Figure 4. (e) Original image 1; (f) Image with magnitude of image (h) and phase of image (e); (g) Image with magnitude of image (e) and phase of image (h); (h) original image 2

verts the data from the time/spatial domain to the frequency domain. Since the Fourier transform resulted in complex-valued coefficients and since, as observed above, the phase is significantly more important than the magnitude in image recognition, we will use the phases of the Fourier spectral coefficients as the features for recognition. The main advantage of using the MLMVN with this dataset is due to its complex-valued neuron’s ability to treat the phase properly (Fig. 5). Since all our data are now represented by phases, we can take a portion of the data which holds the most significant information about the images. We used only 24 phases, corresponding to the lowest three frequencies (#1, 2, and 3), i.e. 4 phases corresponding to the first frequency, 8 phases corresponding to the second frequency, and 12 phases corresponding to the third frequency. This means we can commence learning using the MLMVN with a reduced learning set, which may decrease the network size and the computing time for learning. We take the phases of the lowest frequencies (as corresponding to the Fourier coefficients making the highest contribution to the signal energy and containing the information about the biggest shapes in images), input them into the MLMVN along with the desired output, and train the network. This type of training the network is called “supervised” learning, because we guide the network by letting it know what it is supposed to output given a certain input. Then, through many iterations of correcting the weights based on the error of the network for a given sample, the network accumulates knowledge about the training set. Using this “generalized” knowledge, the network can recognize those samples that

Gonzalez + Colon

The Manhattan Scientist, Series B, Volume 4 (2017)

205

Figure 5. Importance of proper treatment of phase as an angle. For example, if we do not care of the nature of phase, we may treat numbers φ = 0.001 and ψ = 2π − 0.001 = 6.282 as such located in the opposite ends of the interval [0, 2π]. The formal difference of these two numbers is 6.281, however the difference of the angles (phases) represented by these numbers is only 0.002 radians. Thus, we must treat them as arguments of complex numbers for the proper treatment of phase [2].

are somewhat like those in the learning set, but did not participate in learning. Our approach to training the networks was facilitated using the MLMVN-LLS batch learning algorithm [7] written in MATLAB code. This function created a structured MLMVN with two layers, one hidden and one output. The size of the hidden layer varied from test-to-test because the optimal size of a network for a specific learning set is unknown. This knowledge was gained through trial and error. The size of the output layer however always stayed the same, because we knew the desired output for each sample and the number of classifications for learning.

Results / Number analysis The data we used was obtained from the Modified National Institute of Standards and Technology (MNIST) database downloaded from http://yann.lecun.com/exdb/mnist/, which is a standard database for experimenting with and testing neural networks and other image processing systems. The data is two sets of images of handwritten digits, each image having a size of 28 × 28 pixels. The learning dataset is 60,000 images, while the testing set consists of 10,000 images. Testing and experimenting consisted of many run-throughs of the learning code with different size of networks and learning sets, initially starting small and increasingly getting larger. We started each network with a specific size learning set, then increased it after each experiment until the error of the network no longer dropped during learning. This meant that the network could no longer learn from the learning set and that it had “maxed out” the amount of information it could extract from one learning session at that size. After each experiment, we tested the network for the following parameters: 1) how much of

206

The Manhattan Scientist, Series B, Volume 4 (2017)

Gonzalez + Colon

the entire testing set it was able to recognize, 2) how many iterations of learning it took, and 3) the images for which numbers were most recognizable. Table 1 shows the results of the different size networks at the highest learning set tested in which they were able to converge. Table 1. Results of the different successful networks with the most sample participation Network Size

Results of Testing

Inputs

Hidden Layer

Output Layer

Samples Learned

Number of Iterations

Recognition

24 24 24 24 24 24

512 768 1024 1536 2048 3072

10 10 10 10 10 10

4500 5000 5000 10000 10000 17000

1188 558 456 1666 673 1785

83.81% 83.00% 81.87% 84.89% 84.50% 86.35%

The MATLAB testing code was modified to also output the statistics of the recognition of each digit. This was based on the number of occurrences of an image of that digit in the learning set, compared with how many times an image of that digit (in the testing set) were successfully recognized. We took these statistics only from the networks that had seen the most sample participation, and those networks are mentioned in Table 1. The percentage of each digit recognition is presented in Table 2. There you can see that recognition rates for the digits 3, 5, and 8 are the lowest out of all the numbers. This means that they were the hardest numbers for the networks to learn. There can be many factors contributing to the low recognition rates for those numbers; however, it is clear from the simple analysis of these digits, that these rates can be attributed to their shapes. Such similar shapes to other digits can cause the networks to misinterpret the image, and classify it as a different digit. Although the analysis of this is not our main objective, further investigation of these relationships would be very interesting and useful in improving the recognition rates of the images in this dataset. Table 2. Statistics of recognition for each digit, based on networks in Table 1 Digit

Percent Recognition

0 1 2 3 4 5 6 7 8 9

91.35% 91.12% 82.38% 78.05% 88.48% 74.18% 89.88% 88.59% 73.14% 81.65%

Gonzalez + Colon

The Manhattan Scientist, Series B, Volume 4 (2017)

207

Since we used conventional computers, we also utilized a different approach that was conservative on memory and processing cost. We trained the network using a batch-learning method [7], where we divided the learning set into small, similar size batches of samples. The network would then learn and adjust its weights based on that small batch. After it learned one batch, it would move on to the next batch, and learn and adjust its weights based on that new batch. This was repeated until the network had learned all the batches, thus having seen the entire learning set in one global iteration and gained some slight information of each batch. Although these batch learning training sessions worked well on our limited memory conventional computers, they took a considerable amount of time to complete. Nevertheless, they yielded good classification rates, and the result of the testing of two different networks successfully trained using batch learning are listed in Table 3. Table 3. Results for the networks trained with batch learning Batch Size

Network Size

Results of Testing

# of Samples

Inputs

Hidden Layer

Output Layer

Number of Global Iterations

Recognition

100 1000

24 24

512 1024

10 10

345 60

83.81% 93.24%

Conclusion and future work

The results of this work support that the Fourier Transform phase of an image carries enough information about an image, such that we can train a neural network to recognize images only using information extracted from the phase. Also, that the Multilayer Neural Network with MultiValued Neurons is an efficient network for recognizing images in this way since its neurons use a phase-dependent activation function for classification. It is clear that this work is far from done. This approach to recognizing images has a lot of potential, considering that all neural network training sessions were done on ordinary desktop computers, while in the literature neural network training is executed on high-powered computational machines. Nevertheless, even with this large constraint, we were able to get networks to recognize a considerable amount of images, even though they were not even able to learn all of the learning set. Further research into this approach, on larger and faster machines, has the potential to yield outstanding results. We hope to continue working on this approach in the future, seeking better results for the recognition rates of the MLMVN networks. To do this, we need to use a computer with a powerful GPU for parallel processing, increasing the number of samples able to be implemented in a training session and also decreasing the time it takes to complete a training session. A computer with these capabilities was unavailable during the summer when we conducted this research, but we will use one in the future. We also hope to test these networks on more datasets with varioust kinds of images, analyzing and comparing their recognition rates.

208

The Manhattan Scientist, Series B, Volume 4 (2017)

Gonzalez + Colon

Acknowledgements This research was supported by the School of Science Research Scholars Program. The authors would like to thank Dr. Igor Aizenberg for the support and guidance throughout the research, and Dr. Lawrence Udeigwe for mentoring and aiding them in the understanding of the mathematics behind the components of the networks.

References [1] Wikipedia Contributors. “Fast Fourier transform.” Wikipedia, The Free Encyclopedia, 15 Sept. 2017. Web. https://en.wikipedia.org/wiki/Fast Fourier transform [2] I. Aizenberg, Complex-valued neural networks with multi-valued neurons. Berlin: SpringerVerlag Publishers (2011) [3] S. Haykin, Neural networks: A comprehensive foundation, 2nd edn. Prentice Hall, Englewood Cliffs (1998) [4] I. Aizenberg and C. Moraga, “Multilayer feedforward neuranNetwork based on multi-valued neurons (MLMVN) and a backpropagation learning algorithm,” Soft Computing, 11(2), 169183 (2007) [5] M. H. Hayes, J. S. Lim, and A. V. Oppenheim, “Signal reconstruction from phase or magnitude,” IEEE Transactions on Acoustics, Speech, Signal Processing, 28, 672-680 (1980) [6] A. V. Oppenheim and J. S. Lim, “The importance of phase in signals.” IEEE Proceedings, 69, 529–541 (1981) [7] E. Aizenberg and I. Aizenberg, “Batch linear least squares-based learning algorithm for MLMVN with soft margins,” Proceedings of the 2014 IEEE Symposium on Computational Intelligence and Data Mining (CIDM), 48-55 (2014)

Monte Carlo computer investigation of ideal dendrimers in two and three dimensions Timothy Hamling∗ Department of Computer Science, Manhattan College Abstract. A dendrimer is a type of symmetrical tree structure polymer. It can be classified by the number of branches connecting each junction. This study uses Monte Carlo growth simulations to look at properties of different types of dendrimers. Excellent agreement is found with theoretical predictions.

Introduction Monte Carlo (MC) simulations are one way to investigate polymeric materials and they can generate random configurations. Zajac and Bishop [1] used a MC growth algorithm to create linear chains of different lengths. The MC method works by using random numbers to decide in which direction the chain should grow. A linear chain grown with this method starts at an origin point in a simple square lattice in two dimensions or a simple cubic lattice in three dimensions. Then, a random growth direction is selected, and a new bond, of magnitude one, is grown. This new location serves as the next point to grow from, and the process is continued until the desired number of bonds have been placed. From these configurations, many properties can be calculated, such as the mean-square radius of gyration hS 2 i, the asphericity hAi, and shape indicators such as hδi in two dimensions and hP i in three dimensions. Bishop et al. [2] expanded upon this procedure and used a similar MC growth algorithm to study various comb-shaped polymers. In both studies, excellent agreement was found with theoretical values. In this investigation, the MC growth algorithm is modified to create and examine dendrimers with different numbers of branches. Dendrimers are molecules shaped like branching trees, and can fold up to form a molecular cage [3]. They have attracted attention because they can serve as cargo carriers of drugs into hard to reach areas of the body. A perfectly ideal dendrimer is shaped like a fractal tree, and would thus have rotational symmetry. They can be classified by the number of branches and functionality of the junctions. This research examines four dendrimer polymers, each with a different structure. A variety of properties are calculated and compared to theoretical predictions. Fig. 1 displays 9-branch, 12-branch, 21-branch, and 39-branch dendrimers. In Fig. 1, the junction beads are represented by circles, while the branches are shown connecting the junctions. The branches can be composed of any number of beads as long as each branch has the same number, n. In all the dendrimer structures the total number of beads, N , is given by the equation, N = NB ∗ (n − 1) + 1, where NB is the total number of branches. The 9-branch and 21-branch dendrimers are different generations, first and second respectively, of a similar family of dendrimers. These have a central junction with three branches. Each ∗

Research mentored by Marvin Bishop, Ph.D.

210

The Manhattan Scientist, Series B, Volume 4 (2017)

Hamling

Figure 1. Representations of the 9-branch (A), 21-branch (B), 12-branch (C), and 39-branch (D) dendrimers

outer junction also has three branches. The 9-branch dendrimer has one layer of outer branches, while the 21-branch dendrimer has two layers of outer branches. The 12-branch and 39-branch dendrimers are the first and second generation of a different family. These structures have three branches at the center junction, and four branches emanating from the outer junctions. The 12branch structure is similar to the 9-branch structure in that it has one layer of outer branches, while the 39-branch structure is similar to the 21-branch structure in that they both have two layers of outer branches.

Method The current growth algorithm uses portions of the polymer growth algorithm described in Zajac and Bishop [1] and Bishop et al. [2]. The growth algorithm for all the dendrimers was generally the same. At the start of a simulation, a random number is chosen to indicate direction. For the two-dimensional structures, since the dendrimers are being grown on a regular grid, there are four directions which could be chosen, whereas in three dimensions, there are six possible directions. Once the direction is selected, a new bead is placed in that direction. For the ideal systems considered here, overlapping beads are allowed. To construct a dendrimer with the correct number of beads, this process is repeated N − 1 times, since the same origin is always used as the placement for the first bead in a branch. The algorithm starts by growing n beads to create a branch. Then, all the connecting branches are grown from the junction at the end of that branch. In the case of the second-generation 21-branch and 39-branch structures, additional branches are then grown from the end junctions of the inner branches. Once a complete multi-branch component is grown, the algorithm cycles back to the central junction to repeat the process two more times. After constructing a dendrimer, various properties are calculated. To measure the total size of a dendrimer, the mean-square radius of gyration, hS 2 i, is used. The braces hi represent the average over multiple dendrimer simulations. For large dendrimers, hS 2 i, follows the well-known [4] scaling law, hS 2 i = C(N − 1)2ν . (1)

In this equation, the value of C depends on what polymer model is used, and the exponent 2ν should be equal to 1 for the ideal dendrimers examined here.

Another property investigated was the g-ratio. This measures how compact the dendrimer structure is. It uses the radius of gyration of a dendrimer, hS 2 i, and the radius of gyration of a linear

Hamling

The Manhattan Scientist, Series B, Volume 4 (2017)

211

polymer chain, hS 2 il , which contains the same number of beads as the dendrimer, g = hS 2 i/hS 2 il .

The overall shape of the dendrimer can be found using the radius of gyration tensor. This tensor is composed of different eigenvalues. In two dimensions, they are λ2 ≤ λ1 , while in three dimensions, they are λ3 ≤ λ2 ≤ λ1 . These eigenvalues are the principle moments of gyration, and there is one for each direction along the principle orthogonal axes [5]. hS 2 i is equal to the average trace of the radius of gyration tensor, or λ1 + λ2 in two dimensions and λ1 + λ2 + λ3 in three dimensions. The average asphericity hAi of a dendrimer in dimension d is defined by Rudnick and Gaspari [6, 7] as + * Pd 2 (λ − λ ) i j i>j (2) hAi = P (d − 1)( di=1 λi )2 In addition to the asphericity, other properties can be calculated to identify the overall shape of a dendrimer. In two dimensions, one uses hδi = hλ1 i/hS 2 i while in three dimensions * + 27(λ1 − λ)(λ2 − λ)(λ3 − λ) hP i = (3) P ( 3i=1 λi )3 is used, where λ = (λ1 + λ2 + λ3 )/3.

Another important property is the scattering function, S(k)q.F ormally, thisisdef inedasS(k) = P N PN ik • (Rm −Rl ) e .(4)Here, N is the number of beads in the polymer, k is the scattering l m vector, and Rl and Rm are the respective positions of the l-th and m-th beads. The values given by this function depend upon the spatial dimension occupied by the dendrimer. After averaging over the angles in two dimensions, the scattering function is defined as * N N + 1 XX S(k) = 2 J0 (x) (5) N m l 1 N2

where J0 (x) is the 0-th order Bessel function, whereas in three dimensions the scattering function is defined as * N N + 1 XX S(k) = 2 sin(x)/x , (6) N m l

where x2 = k 2 hS 2 i. In the case of linear polymers, the theoretical values for the scattering function follow the Debye equation [4], 2(x − 1 + e−x ) S(k) = . (7) x2 Dendrimers are not linear and therefore do not follow the Debye equation. Howeever, the theoretical linear values have been included as a comparison in some of the graphs. The scattering

212

The Manhattan Scientist, Series B, Volume 4 (2017)

Hamling

function results for each of the four dendrimer structures matched well with the theoretical equations. Each structure has a different exact equation for S(k). The 9-branch dendrimer S(k) follows the equation: 2(3 + x − 3e−x/9 − 12e−x/3 + 12e−4x/9 ) S(k) = (8) x2 which has a Taylor expansion of S(k) = 1 − 49x/243 + ...

This equation was tested against the MC results by comparing the g-ratio from the samples against the theoretical g-ratio obtained from this expansion. The g-ratio is found by dividing the x term of the expansion of S(k) of the 9-branch dendrimer by the corresponding term of a linear chain with the same number of beads. The Taylor expansion for a linear chain yields S(k) = 1 − x/3 + ... This gives a g-ratio of 49/81 = 0.60494 for the 9-branch dendrimer. Similar results are obtained from the exact equations of the other structures. The 12-branch, 21-branch and 39-branch scattering functions also closely follow their exact predictions, and the calculated g-ratios are in fine agreement with the Taylor expansions of these equations. The exact prediction for the 12-branch g-ratio is 35/72 = 0.48611, the 21-branch g-ratio is 1363/3087 = 0.44153, and the exact prediction for the 39-branch g-ratio is 5257/19773 = 0.26587. These g-ratio values do not vary with dimension for a given dendrimer structure.

Results All the codes were written using the C language, and were compiled and run on a Linux machine using the GCC Compiler. Table 1 shows the values found for 9-branch dendrimers in three dimensions. In these simulations 100,000 independent samples were generated for each of four different total bead amounts. The number of beads N is shown at the top of the table. hλi values for each of the eigenvalues are included, as well as, hAi. Since this dendrimer was constructed in a three-dimensional space, hP i values were also calculated. The radius of gyration is presented for both the dendrimer, and for a corresponding linear polymer chain. The error deviation is shown as a number in parenthesis (one standard deviation from the mean). Thus, a value of 3.76(1) would range from 3.75 to 3.77 (i.e., 3.76 ± 0.01). Table 1. Effect of the number of beads, N , for 9-branch dendrimers in three dimensions. Property N = hλ1 i hλ2 i hλ3 i hAi hS 2 i hP i hS 2 i`

181

442

892

1342

3.76(1) 1.39(1) 0.57(1) 0.255(1) 5.73(1) 0.215(1) 9.16(1)

12.22(2) 4.40(1) 1.81(1) 0.262(1) 18.43(2) 0.226(1) 30.16(5)

29.73(5) 10.64(2) 4.33(1) 0.264(1) 44.71(5) 0.230(1) 73.60(12)

60.07(9) 21.38(3) 8.71(1) 0.266(1) 90.17(11) 0.233(1) 148.32(24)

90.27(14) 32.14(5) 13.08(2) 0.266(1) 135.50(16) 0.233(1) 223.33(37)

Hamling

The Manhattan Scientist, Series B, Volume 4 (2017)

213

The data tables show some interesting results. For any dendrimer, hAi and hP i level off as the number of total beads in the dendrimer increases. One other observation is that hS 2 i was always higher for the linear chains than it was for the dendrimers, which is to be expected because the dendrimer is less spread out than the linear chain and would therefore have a shorter distance from its center to an outer point. To check that the radii of gyration matched the scaling law, Eq.1, we used a C program to fit our hS 2 i results to a power function. The theoretical power for the radii of gyration should be one. For three-dimensional dendrimers, the powers are as follows: 0.992(1), 0.986(1), 0.992(1), and 0.992(1) for 9, 12, 21 and 39 branches, respectively. The results for the two-dimensional dendrimers were also in good agreement with the theoretical value of one, being 0.991(1), 0.985(1), 0.991(1), and 0.993(1) for 9, 12, 21, and 39 branches, respectively. The data in Table 1 have been extrapolated using a linear function in 1/N . This fits the data as N becomes infinite since 1/N → 0. Table 2 shows the resulting fits for the 9-branch Table 2. Results for 9-branch dendrimers in two and three dimensions. Property

2D Extrapolated

Wei

3D Extrapolated

Wei

g − ratio hAi hδi hP i

0.606(1) 0.281(1) 0.765(1)

0.604939(0) 0.2812(8) 0.765121(0)

0.606(1) 0.266(1)

0.604939(0) 0.26583(9)

0.233(1)

0.116134(0)a

Note that the Wei method defines hP i with an additional factor of 1/2.

structures in two and three dimensions. The properties fit were the g-ratio, hAi, hδi values for twodimensions, and hP i values for three-dimensions. Each of the extrapolations are compared to the exact calculations, determined using the Wei Method [8]. Excellent agreement is shown between the extrapolated MC results and the theoretical predictions. The values of the asphericity for each dendrimer reflect how spherical and symmetric the dendrimer is. For comparison, an hAi value of 0.0 represents a perfect circle in two dimensions and a perfect sphere in three dimensions. The data showed that, as the number of branches increased, the values for hAi decreased. The 9-branch dendrimer had an extrapolated hAi value of 0.266(1) in three dimensions, while the 39-branch dendrimer in three dimensions had an extrapolated hAi value of 0.165(1). This indicates that dendrimers with more branches have a more symmetric and round shape. Finally, Fig. 2 shows results for the three-dimensional dendrimer S(k) MC calculations compared to the exact equations. This graph indicates that dendrimers with more branches have smaller and smaller S(k) values as x increases. This indicates that beads are more closely packed and crowded together as the number of branches increases, which is to be expected. All S(k) results for these graphs were calculated using 100,000 sample runs. There is excellent agreement between the MC results and the exact predictions. Similar agreement was obtained in the 2D cases.

214

The Manhattan Scientist, Series B, Volume 4 (2017)

Hamling

Figure 2. The reciprocal of the three-dimensional dendrimer S(k) vs. x. The solid line is the exact Debye equation for linear polymer chains. The squares represent the MC linear results when N = 892 beads. The long-dashed line is the exact equation for 9-branch dendrimers. The circles represent the MC 9-branch results when N = 892. The small-dotted line is the exact equation for 12-branch dendrimers. The up-triangles represent the MC 12-branch results when N = 889. The dash-dot line is the exact equation for 21-branch dendrimers. The down-triangles represent the MC 21-branch results when N = 883. The dash-dot-dot line is the exact equation for 39-branch dendrimers. The diamonds represent the MC 39-branch results when N = 898.

Conclusions Four different dendrimers in both two and three dimensions were studied. This was done using a Monte Carlo growth algorithm. Multiple properties were computed, and excellent agreement was found between the Monte Carlo simulations and the available theoretical predictions.

Acknowledgements This work was funded by the School of Science Research Scholars Program. The author would like to thank his mentor, Dr. Marvin Bishop, for all his help throughout this project, as well as Dr. Robin de Regt for obtaining the analytical S(k) predictions, the Manhattan College Computer Center and the Kakos Center for Scientific Computing for generous grants of computer time. He would also like to thank the departments of Mathematics and Computer Science, as well as his friends and family for their support throughout the entire research process.

References [1] [2] [3] [4] [5] [6] [7] [8]

G. Zajac and M. Bishop, Comp. Educ. J., 6(1),44 (2015) M. Bishop, J. Stone, C. von Ferber, and R. de Regt, Physica A, 484, 56 (2017) A.W. Bosman, H.M. Hanssen and E.W. Meijer, Chem. Rev., 99, 1665 (1999). P.G. de Gennes, Scaling Concepts in Polymer Physics, (Cornell University Press, Ithaca, 1979) K. Solc and H. Stockmeyer, J. Chem. Phys., 54, 2756 (1971) J. Rudnick and G. Gaspari, Science, 237, 384 (1987) and references therein J. Rudnick and G. Gaspari, J. Phys. A, 19, L191 (1986) G. Wei, Physica A, 222, 152 (1995); ibid., 222, 155 (1995).

Statistical binary classification of MRI data Sana Altafâ&#x2C6;&#x2014; and Melissa Brennerâ&#x2C6;&#x2014; Department of Mathematics, Manhattan College Abstract. Magnetic Resonance Imaging (MRI) is a versatile medical imaging modality that limited by its slowness. performance of methods to accelerate MRI, we will evaluate statistical learning methods to detect the presence of a tumor. Current approaches for accelerating MRI are not linear and have different noise for different parts of the image. The new methods, e.g., compressed sensing, are non-linear and create unpredictable artifacts in the images. The method of optimization we used specified the clinical task, e.g., detecting a tumor. We used logistic regression, linear discriminant analysis (LDA) and support vector machines (SVM) to classify both simulated data and data from MRI images. We concluded that all methods performed well in classifying both the simulated data and the data from MRI images.

Introduction An important part of treating patients with tumors is being able to detect these tumors in an MRI scan. These scans use a computer along with strong magnetic fields and radio waves to generate detailed images of the brain. Essentially, an MRI creates a three-dimensional picture of the body by using a magnet to align the nuclei of water molecule atoms. Early detection is an essential part of treating malignant tumors successfully. A metastatic tumor (a tumor that has spread to other areas of the body from its primary location) is far more complex to treat and has a poor prognosis. Timely diagnosis of cancer makes it possible to limit the treatment to a relatively small procedure, thereby preserving the affected organ and preventing side effects of a systemic treatment [1]. During an MRI session, a patient slides into a scanner shaped like a tunnel. It is a painless process, which takes between 30 and 60 minutes. By allowing medical professionals to distinguish between tissues that is normal vs. diseased, MRI plays a significant role in performing cancer diagnosis. A classification method known as linear discriminant analysis (LDA) is commonly used in optimizing imaging systems [2]. Our goal is to determine whether there are other classification methods that would better classify MRI data. We performed logistic regression and support vector machine (SVM) classification on our data in order to compare it to the results we found after performing LDA classification.

Data sets We generated simulated data using two-dimensional normal random vectors to represent data with and without a tumor. The data without a tumor had mean = (2,1), where the (x, y) notation â&#x2C6;&#x2014;

Research mentored by Angel R. Pineda, Ph.D.

216

The Manhattan Scientist, Series B, Volume 4 (2017)

Altaf + Brenner

denotes that the x-coordinate is one feature and the y-coordinate is the second feature, and co 2 1 variance matrix = and the data with a tumor was simulated as having a mean = (5,5) and 1 2 10 2 covariance matrix = . Fig. 1 shows a scatter plot of the simulated data. 2 1 Simulated Data

Figure 1. Scatter plot of simulated data generated using multivariate normal in testing the classification methods.

MRI data with two features The MRI data were generated placing a ring signal (shown in Fig. 2, Diagram C) in multiple locations of an MRI of a brain. These images were generated using the total variation algorithm [3]. Each image was compressed to a vector of features which are the coefficients of an expansion in the Laguerre-Gauss (LG) basis which gives an orthonormal basis for radially symmetric images [4]. The basis images are radially symmetric, so we show slices through the middle (cross-sections) of the first three LG basis are shown in Fig. 2. Fig. 3 shows a scatter plot of the first two features from the MRI data.

Figure 2. Data generated by inserting lesions in MRI scans and representing those images using a basis expansion. A, Brain with lesion; B, Lesion zoomed in; C, Ring signal modeling tumor growth; D, Cross section of the first three basis functions used to generate the features (Laguerre-Gauss basis).

Altaf + Brenner

The Manhattan Scientist, Series B, Volume 4 (2017)

217

Figure 3. Scatter plot of MRI observations. First two features obtained from MRI Images with and without a ring lesion.

Methods Classification methods Linear discriminant analysis (LDA) Linear discriminant analysis is a classification method that models the predictors separately in each of the response classes. In our case, the classes are “tumor” and “no tumor.” LDA assumes that the data has a multivariate normal distribution and uses Bayes’ theorem to get estimates for the probability of a tumor (Y = y) given the data (X = x) [5], πk fk (x) pk (x) = P r(Y = k|X = x) = PK l=1 πl fl (x)

(1)

where pk (x), the posterior probability, can be calculated by estimating πk , the prior probability that a random chosen observation comes from the kth class, and fk (x), the probability that the data X = x is observed given the class Y = y. In our case, the posterior probability is the probability that there is a tumor present in an MRI scan. For the LDA approach, we began by fitting the model on a subset of our data. From there, we ran a prediction on just the features in the test sample. This helped us to visualize how well the model predicts “tumor” versus “no tumor.” when it does not know the correct diagnosis. Once this was completed, we were able to create a confusion matrix which indicated to us the true positive and false positive results of the model shown in Fig. 4.

Figure 4. An example of a confusion matrix that resulted from running LDA on our MRI data. The correct results are indicated in green and the false results in red. The true positive fraction is 43/50 and the false positive fraction is 4/50.

218

The Manhattan Scientist, Series B, Volume 4 (2017)

Altaf + Brenner

In other words, the confusion matrix told us how many times the model predicted “tumor” when there was a tumor and how many times it predicted “no tumor” when there was a tumor. We were then able to calculate the true positive and the false positive fractions from the model. Logistic regression Logistic regression is a method where a data set is analyzed based on one or more independent variables determining an outcome. These outcomes can be measured with multiple variables. Used to find the best fitting model to express the relationship between the binary outcome and the set of independent variables, logistic regression uses the conditional distribution of y given x. In logistic regression, this distribution is a Bernoulli distribution since the dependent variable is binary. Also, the predicted values are probabilities bounded within (0, 1) through the logistic function since this type of regression predicts the probability of specific outcomes. Typically, a particular outcome for the dependent variable that is considered a “success” is labeled as 1 and a “failure” as 0. Logistic regression is used to predict the odds of being a favorable outcome based on the values of the independent variables, also known as predictors. “Odds” are defined as the probability that a particular outcome is favorable divided by the probability that it is a failure. If we let p be the probability of success, after algebraic manipulation of the logistic function and a logarithmic transformation, the model becomes linear in the features, p = a1 x 1 + a2 x 2 + . . . + ax x k . (2) log 1−p Support vector machine (SVM) After performing LDA and logistic regression on our simulated data, we sought a different method to determine which of the three would work best on our data. The support vector machine is a generalization of a classifier that finds the plane that best separates the data. The main idea is that it uses a measure of distance (kernel) to identify the points nearest to the boundary between the two classes and finds a potentially nonlinear boundary to make the classification [5]. In our research we saw how SVM separated our data based on a tumor being present or absent. In our project, we used a radial kernel. The radial kernel successfully separates two classes that result in nonlinear boundaries. It has a local behavior, which means that only nearby training observations have an effect on the class label of a test observation. Evaluation methods Receiver operating characteristic (ROC) curves ROC curves are helpful to display the perfomance of a classifier with different thresholds. In our case, the thresholds in the ROC curves represent what determines whether there is a tumor

Altaf + Brenner

The Manhattan Scientist, Series B, Volume 4 (2017)

219

present or absent in our data. We vary that threshold to help identify which one provides us with the best true positive fraction. The ROC curve helps the viewer visualize how well the model is working with various thresholds and helps determine the optimal threshold for their data [5]. In our ROC curves, the true positive fraction (TPF) was represented by the y-axis and the false positive fraction (FPF) was represented by the x-axis. A true positive fraction indicates the fraction of times the model predicted a tumor was present and when the tumor was actually present. Similarly, a false positive fraction indicates the fraction of times the model predicted a tumor was present when there was no tumor present. For LDA, logistic regression, and SVM, we were able to create ROC curves to help visualize the rate of success and failure for each of the methods at different thresholds. Ideally, a perfect ROC curve has as high a true positive fraction as possible for any false positive fraction. In order to create our ROC curves, we began by incrementing our threshold, t. We separated the training set and for each variation of t, we created a confusion matrix. This matrix allowed us to then calculate both the true positive fraction and the false positive fraction. After we finished varying t, we plotted all of the true positive fractions and the false positive fractions for all the possible thresholds. This created our ROC curve which allowed us to visualize how each model performed at various thresholds. Once this was completed for LDA, logistic regression, and SVM, we calculated the area under each of the ROC curves (AUC). Since we have a curve from 0 to 1 we know that the area under our ROC curves actually indicates the average of our function; i.e. the area under the curve is the average true positive fraction. Therefore, we know that the graph with the greater respective area under its curve would help indicate which model will serve as a better classifier for our data. k-fold cross validation We began the method of cross validation by implementing the process on our simulated data and then applying it to the MRI data. Cross validation allows one to estimate the test error associated with a given statistical learning method in order to evaluate its performance [5]. Cross validation will then allow us to compare the performance of both logistic regression, LDA and SVM on our data and determine which is a better fit for our purposes. k-fold cross validation works by randomly dividing the data into k groups of approximately equal size [5]. k-fold then uses one set as the validation set and the rest of the sets to fit the method. This process continues throughout all k subsets. For our purposes, we sliced the data into 10 different folds, where, in each subset of data, we determined the AUC of the ROC curves for LDA, logistic regression, and SVM. Once we had calculated the AUC for each fold, we could then determine the average AUC for each method. Then we were able to compare the average AUC for LDA, logistic regression, and SVM and had a better insight into which method is a better classifier.

220

The Manhattan Scientist, Series B, Volume 4 (2017)

Altaf + Brenner

Results Linear discriminant analysis (LDA) Simulated data After performing 10-fold cross validation, we found that the average area under the ROC curve was 0.99 for LDA for the simulated data. Fig. 5 shows a sample ROC curve for a single fold used in cross validation.

Figure 5. LDA ROC curve for simulated data with an AUC of 0.98.

MRI data using two features After performing 10-fold cross validation, we found that the average area under the ROC curve was 0.89 for LDA using two features. Fig. 6 shows a sample ROC curve for a single fold used in cross validation.

Figure 6. LDA ROC curve for MRI data set using two features, with an AUC of 0.91.

Figure 7. LDA ROC curve for MRI data set using four features, with an AUC of 0.98.

MRI data using four features After performing 10-fold cross validation, we found that the average area under the ROC curve was 0.98 for LDA using four features. Fig. 7 shows a sample ROC curve for a single fold used in cross validation.

Altaf + Brenner

The Manhattan Scientist, Series B, Volume 4 (2017)

221

Logistic regression Simulated data After performing 10-fold cross validation, we found that the average area under the ROC curve was 0.99 for logistic regression for the simulated data. Fig. 8 shows a sample ROC curve for a single fold used in cross validation.

Figure 8. Logistic regression ROC curve for simulated data, with an AUC of 0.98.

MRI data using two features After performing 10-fold cross validation, we found that the average area under the ROC curve was 0.89 for logistic regression with two features. Fig. 9 shows a sample ROC curve for a single fold used in cross validation.

Figure 9. Logistic regression ROC curve for MRI data set using two features, with an AUC of 0.91.

Figure 10. Logistic regression ROC curve for MRI data set using four features, with an AUC of 0.98.

MRI data using four features After performing 10-fold cross validation, we found that the average area under the ROC curve was 0.98 for logistic regression with four features. Fig. 10 shows a sample ROC curve for a single fold used in cross validation.

222

The Manhattan Scientist, Series B, Volume 4 (2017)

Altaf + Brenner

Support vector machine (SVM) Simulated data After performing 10-fold cross validation, we found that the average area under the ROC curve was 0.98 for SVM. Fig. 11 shows a sample ROC curve for a single fold used in cross validation.

Figure 11. Support vector machine ROC curve for simulated data with an AUC of 0.97

MRI data using two features After performing 10-fold cross validation, we found that the average area under the ROC curve was 0.88 for SVM with two features. Fig. 12 shows a sample ROC curve for a single fold used in cross validation.

Figure 12. Support vector machine ROC curve for MRI data set using two features, with an AUC of 0.93.

Figure 13. Support vector machine ROC curve for MRI data set using four features, with an AUC of 1.00.

MRI data using four features After performing 10-fold cross validation, we found that the average area under the ROC curve was 1.00 for SVM with four features. Fig. 13 shows a sample ROC curve for a single fold used in cross validation.

Altaf + Brenner

The Manhattan Scientist, Series B, Volume 4 (2017)

223

Discussion Simulated data Since we know that the simulated data follows a normal distribution, we expected LDA to be the best classifier. After performing 10-fold cross validation, we found that LDA and logistic regression have average AUCs of 0.99. We can also note that SVM has an average AUC of 0.98 as shown in Table 1. Once again, these areas are very close to 1 and extremely close to one another, showing that all methods are classifying the simulated data well. In the case of SVM, further optimization of the parameters for this training data could have improved the AUC. Table 1. Cross-validated AUCs for simulated data Method LDA LR SVM

Cross-Validated AUC 0.99 0.99 0.98

MRI data Two features After performing 10-fold cross validation on all three methods, we found that the average AUC of LDA is 0.89, 0.89 for logistic regression, and 0.88 for SVM as shown in Table 2. We see that all average areas are very close to one another, indicating that all classification methods perform similarly. Four features After performing 10-fold cross validation on the methods, we found that LDA has an average AUC of 0.98, that of logistic regression is 0.98, and 1.00 for SVM. Since these values are so close to one another and to 1, we can again come to the conclusion that all classification methods are performing extremely well on our MRI data. We do note that the non-linear classifier in SVM may have led to improved performance when using four features. This will be the subject of future research. Table 2. Cross-validated AUCs for MRI data Method LDA LR SVM

two features 0.89 0.89 0.88

four features 0.98 0.98 1.00

224

The Manhattan Scientist, Series B, Volume 4 (2017)

Altaf + Brenner

Conclusions Detecting tumors before they spread is an important factor in curing cancer. Several studies show that an MRI is one of the most effective techniques for identifying tumors. It is imperative that we use the tools at hand, such as statistical detection to discover efficient ways of optimizing imaging systems. Overall, we have been able to work with LDA, logistic regression, and SVM to classify our data. We have also been able to apply 10-fold cross validation on all three methods to determine if, on average, one of those methods is a better fit for our data. In general, we have determined that all methods perform better when using four features as compared to two features with our MRI data. We also determined that LDA, logistic regression, and SVM all perform very well on all of our data sets and they perform very similarly to one another. At this point we can conclude that all three methods serve as good classifiers for our data. The non-linear boundary of SVM may be useful when using more features. In the future we would like to determine whether the slight differences in the AUCs between the three methods are statistically significant. This will help us to determine further if any of the three methods are performing better than the other two. Other future endeavours include optimization of image reconstruction based on these methods of classification and looking at other types of tumors to determine if these methods would still perform well. We will continue our research to work towards finding new methods of improving the detection of tumors in MRI scans.

Acknowledgments The authors would like to thank Dr. Angel Pineda and the entire mathematics department at Manhattan College for their help throughout this research.

References [1] Berger A. Magnetic Resonance Imaging. British Medical Journal 2002; 324:35-35. [2] Barrett HH, Yao J, Rolland JP, Myers KJ. Model Observers for Assessment of Image Quality. Proceedings of the National Academy of Sciences USA 1993; 90: 9758-9765. [3] Lustig, M, Donoho D, Pauly JM. Sparse MRI: The Application of Compressed Sensing for Rapid MR Imaging. Magnetic Resonance in Medicine 2007; 58: 1182-1195. [4] Pineda AR, Yoon S, Paik DS, Fahrig R, Optimization of a Tomosynthesis System for the Detection of Lung Nodules, Medical Physics 2006; 33: 1372-1379. [5] James G, Witten D, Hastie T, Tibshirani R. An Introduction to Statistical Learning with Applications in R. New York: Springer; 2015.

What’s in the air? Using mathematical models to predict Boston air quality Anthony DePinho+ *, Tara Ippolito*, Biyonka Liang*, Kaela Nelson*, Annamira O’Toole* + Department of Mathematics, Manhattan College * Team Research in Computational and Applied Mathematics, Institute for Applied Computational Science, Harvard John A. Paulson School of Engineering and Applied Sciences National Science Foundation, Research Experience for Undergraduates Abstract. Exposure to pollutants such as NO2 , SO2 , and PM2.5 are a significant concern, especially for those living in large cities. However, most major cities have five or fewer active air quality sensors. Various studies have shown that geostatistical models using traffic count, elevation, and land cover as variables can predict pollutant levels with high accuracy. However, collecting training data containing sufficient geospatial variation often involves large scale deployment of sensors over the area of interest. In this study, we trained geospatial and spatio-temproal models for three EPA criteria pollutants - NO2 , SO2 , and PM2.5 - using data collected from 398 counties across the US and applied the models to produce intra-urban pollution concentration levels for a 107.495 square mile region covering the Greater Boston area. The performance of the geospatial model (Land Use Regression) and spatio-temporal model (Guassian Process) were found to be comparable to similar models in literature. Our study also addresses the public health challenge of effectively and meaningfully communicating scientific findings in environmental science to the general public. Specifically, we designed an interactive web interface for visualizing our Boston air pollution predictions. This interface serves as a proof-of-concept for an accessible, educational, and scientific tool for urban residents to understand the impact of air quality.

Introduction Air pollutants originate from multiple sources, some anthropogenic and others from reactions in the atmosphere itself. Criteria pollutants are those whose levels are monitored and regulated by the Environmental Protection Agency (EPA). These pollutants are amongst those that can significantly impact the environment and on human health. For example, Particulate Matter 2.5 (PM2.5 ) is an especially dangerous pollutant in long-term human exposure. Findings from the American Heart Association indicate that PM2.5 air pollution contributes to worsened cardiovascular health and (to a lesser extent) pulmonary health [1, 2]. Stroke, arrhythmia, and heart failure exacerbation are some of the more serious consequences of exposure in individuals with heightened risk of cardiovascular problems. Currently, PM2.5 exposure is ranked as the 13th leading cause of worldwide mortality with approximately 800,000 premature deaths per year [1]. Both NO2 and SO2 are particularly harmful for those with respiratory illnesses, they can also contribute to particulate matter concentration when they react with other chemicals in the atmosphere [3, 4]. Literature suggests that NO2 and SO2 also impact cardiovascular health, especially in conjunction with particulate matter [2]. The EPA reported that multiple studies showed evidence of “increased risk of susceptibility to both viral and bacterial infections after NO2 exposures” [4]. The same EPA report also stated that airway inflammation and hyper-responsiveness were seen in human clinical studies [4]. The EPA report on SO2 exposure finds that studies indicate that SO2 is “associated with

226

The Manhattan Scientist, Series B, Volume 4 (2017)

DePinho et al.

episodic respiratory illness and aggravation of respiratory diseasesâ&#x20AC;? [3]. Based on widespread assessment and study of the dangers of these pollutants, modeling their concentrations and educating the public is vital to improve human health. Among the main sources of particulate matter are combustion of fossil fuels (i.e. traffic and power plants) [1], construction and demolition leading to particle suspension, as well as physiochemical transformation of gases already existing in the atmosphere [5]. The main sources of NO2 are the combustion of fossil fuels in industrial processes as well as traffic pollutants [5]. SO2 concentration is mostly affected by combustion of fossil fuels that contain sulfur, this can be from cars as well as power plants [5].

Problem statement There are two main challenges that motivate this project. Firstly, there are four Environmental Protection Agency (EPA) air quality sensors located in the Greater Boston area that are responsible for recording all air quality data for all the significant pollutants in the entire city (Fig. 1). As large portions of the city are outside the range of existing sensors, there are not enough EPA sensors to provide an adequate assessment of intra-urban spatial variations in air quality conditions. Since the installation of new sensors to provide more widespread coverage of land is a difficult task, our main objective is to implement statistical modeling techniques to model air pollution in areas not covered by existing sensors. The independent variables impacting pollutant concentration that we will consider include land use and weather.

Figure 1. Locations of EPA Air Quality Sensors in Greater Boston

The second challenge that this project addresses is the difficulty of making scientific data and findings accessible to the general public, especially when such information affects the health and welfare of urban residents. While environmental monitoring organizations like the EPA makes all of its air quality data publicly available, the data used in this study consist of numerical databases or scientific reports. In these formats, the air quality information can be of limited use to laymen with

DePinho et al.

The Manhattan Scientist, Series B, Volume 4 (2017)

227

little scientific training. Furthermore, air pollution data are often collected and reported by agencies for regulation purposes, not for health or educational purposes. Thus there is a further gap to be filled in connecting the results of air quality monitoring to public health concerns. A major goal of this project is to design an easy-to-read, interactive interface which provides a more intuitive visualization of Bostonâ&#x20AC;&#x2122;s air quality. The interface will provide users with a way to understand air quality conditions on both a city wide and granular (neighborhood) level. Our design for the interface aims to communicate the results of our statistical models meaningfully to anyone who uses the interface. In the end, we hope that our interface serves as an educational tool for the public, as a tool for scientists, and potentially as an aid in city policy and zoning decisions.

Procedures and Methods Data collection In our models, we include static or geospatial data as well as dynamic data. Geospatial data consists of land use, topography, and bus routes while dynamic data consists of traffic and weather. The training data consists of geospatial predictors and air pollutant readings collected from 398 counties throughout 16 US states, the majority of which were on the East Coast. The trained models are applied to a 107.495 square mile region covering Greater Boston. The latter is divided into a 50 by 50 grid for which geospatial predictors were extracted per grid cell. Dynamic data include traffic and weather over the grid and air quality from the four Boston sensor sites that were collected hourly over several weeks. Data sources and formats are shown in Table 1. Table 1. Data sources and formats Data

Source

Method

Format

Land Use, Green Space Air Quality Weather Topology Traffic

MassGIS Oliver, USGS USGS, EPA NOAA, Weather Underground Harvard Center for Geographical Analysis MassDOT, MBTA, MassGIS

Direct download Direct download Direct download, web scraping Direct download Direct download, web scraping

.shp,.shx, .dbf, .prj .shp,.shx, .dbf, .prj JSON, CSV CSV JSON

Data rastering The land use data was downloaded as GIS shapefiles, wherein the areas used for the training, testing and forecasting are covered by polygons (and multigons - nested polygons) each describing a single land use type. Average daily traffic volume and real-time traffic volume data was downloaded by state as point-wise estimates, i.e. reported for a network of sensor locations across each state. Elevations (from sea level) are provided by Google Maps. Weather data were provided by the EPA for each sensor location, in addition, real-time weather data for stations throughout Boston were collected for a period of one week. We divided a 107.495 square mile region encompassing the Greater Boston area into a 50Ă&#x2014;50 grid (Fig. 2). The grid cells are uniformly sized squares with a length of approximately 1/16th we

228

The Manhattan Scientist, Series B, Volume 4 (2017)

DePinho et al.

Figure 2. Visual representation of the Greater Boston area grid system

uniformly sampled 100 random points. Each point is assigned a land use type corresponding to the land use polygon within which it fell. The land use types of the sampled points within each cell are then categorized into 11 land use categories and totaled for each category producing a land use proportion or percentage break down for each grid cell. The distribution of land use types in our data is visualized in Fig. 3. Next, the elevation is measured at each sampled point and the elevations are averaged to produce an average elevation per grid cell. Then, counts from traffic sensors falling within each grid are averaged producing a single indicator per cell. Finally, for the centroid of each cell, the are averaged.

Figure 3. Distribution of Land Use in Site Data

In the training set, 100 random points were uniformly sampled from a 1/16 squared mile grid centered at each EPA sensor site. Geospatial descriptors were extracted for each site through the same process as described above, with the exception that weather data was taken from EPA provided yearly averages rather than real-time data. For sites where these data were missing, we

DePinho et al.

The Manhattan Scientist, Series B, Volume 4 (2017)

229

imputed the missing value with the average value for the state to which the site belonged. The land use data used to characterize the EPA sites in the training set were collected by the US Geological Survey during the period from 1970 to 1980. The training data consists of data from 1,948 EPA monitoring sites in counties across 16 US states. The data were collected and averaged over the year of 1980. Since each site monitors different conditions and pollutants, we created subset datasets for each pollutant. We first subset out the sites that monitored our pollutants of interest: NO2 , SO2 and PM2.5 . The component PM2.5 readings (Silicon PM2.5 , Titanium PM2.5 , etc.) were added for each site to obtain an overall PM2.5 reading for each site. The training set contained 1949 PM2.5 , 311 NO2 , and 585 SO2 observations. Computational resources Our primary computing environment is Jupyter Notebook, using an IPython3 kernel. For statistical modeling we used SciPy and scikit-learn libraries, and Matplotlib for visualization. For manipulating GIS data, we use the Shapely and PyShp libraries. The web application was developed in D3, Leaflet, CSS, and HTML. Amazon Web Services were used for large scale data collection and processing. Finally, GitHub was used for project collaboration and organization.

Mathematical modeling Land use regression Land Use Regression (LUR) is a linear regression model commonly used to predict air pollutant concentration based on geospatial variation, using predictors such as land use and average (static) weather conditions like wind speed and air pressure. A separate land use regression model was fitted for each EPA criteria pollutant in our study. The LUR model was trained and tested on data collected from US sites outside of Boston; then it was used to predict concentration levels for each grid cell covering the Greater Boston Area. In addition, we performed variable selection and analysis to reason about the impact of dynamic variables on the concentration of atmospheric pollutants. The form of the utilized LUR models is as follows y = Î˛0 + Î˛1 X1 + Î˛2 X2 + ... + Î˛n Xn + , where the dependent variable y is the pollutant concentration of a given area; and X1 ...Xn represent the set or subset of geospatial predictor variables. Prior to variable selection, these predictors consisted of a total of 14 variables from two categories: land use and weather. Land use types were categorized into industrial, commercial, medium density residential area, open space, crop land, water, wetland, transitional, forest, transportation, and urban public space. The weather data consisted of 3 metrics: outdoor temperature, solar radiation and wind speed. These metrics were chosen as they were collected at the most sites that also collected the pollutants of interest.

230

The Manhattan Scientist, Series B, Volume 4 (2017)

DePinho et al.

Variable Selection and Validation for LUR Three different metrics were used for variable selection: p-value, R2 and AIC, and the results were compared. For R2 variable selection, an 8-fold cross validation was implemented. Using backwards stepwise elimination, the set of predictor variables from the model with the worst metric –highest mean p-value, the lowest validation R2 and the highest AIC– was eliminated at each step. For all metrics, variable selection reduced the predictor set to 8 − 12 variables.

Table 2 details the results of our variable selection on all three LUR models. Testing R2 was in the range of 0.203 − 0.234 for PM2.5 , 0.562 − 0.593 for NO2 and 0.368 − 0.403 for SO2 . By these metrics, we believe that levels of NO2 are most correlated with geospatial variations. Table 2. Variable selection for PM2.5 , NO2 , and SO2 Variable selection metric

MSE Training R2 Test R2 ————————————————————————————————PM2.5 NO2 SO2 PM2.5 NO2 SO2 PM2.5 NO2 SO2

P-value AIC R2

9.910 21.970 16.955

0.569 3.015 2.107

0.273 1.074 0.443

0.211 0.231 0.209

0.635 0.613 0.639

0.431 0.445 0.439

0.222 0.234 0.203

0.593 0.573 0.562

0.368 0.403 0.374

Using results from our variable selection, we analyzed predictors that appear most often in the final subsets (Fig. 4).

Figure 4. Land use type frequency in variable selection

Since a primary source of all three pollutants is the combustion of fossil fuel, we expected to see certain land use variables, such as “industrial”, “transportation”, “commercial” and “medium density residential”, to be preserved by variable selection. This was validated by the variable selection results. However, in the selected subsets, we often observed variables that did not have

DePinho et al.

The Manhattan Scientist, Series B, Volume 4 (2017)

231

known scientific relationships to the pollutants. “Wetlands” for PM2.5 is one example. We suspect these unexpected inclusions may be the result of unknown confounding variables. Gaussian process Though LUR models are effective when we suspect strong correlation between geospatial variation and concentration levels of pollution, they lack flexibility to describe more complex, nonlinear, relationships between geospatial predictors and pollution levels. In addition, such models cannot be naturally extended to include dynamic, temporal, data. Thus, one way to introduce non-linearity and temporal dependence into our model is by using a Gaussian Process model. A Gaussian Process model is a statistical model that uses a non-parametric representation of the underlying function relating predictor and response. Specifically, we assumed that any subset of their pollution concentration levels (both observed, y, and unknown, y∗) have a joint Gaussian distribution, (y, y∗) ∼ N (µ, Σ)

wherein the covariance matrix Σ is determined by some metric of similarity of the geospatial and dynamic characteristics of the locations of the corresponding observation sites. That is, each entry in the covariance matrix is computed by a function in terms of the predictors, this is the kernel function. In this study, the standard radial basis function (RBF) kernel is used K(x, x*) = σf2 exp(

−(x − x*)2 ) + σn2 δ(x, x*), 2`2

where σf2 is the amplitude of the air quality approximation, ` is the length scale, and σn2 is the noise variance. In our models, we used a constant noise level of 0.001. Lastly, our Gaussian Process model builds on the results of our LUR models by incorporating the predicted pollution values as the mean of the Gaussian Process model. The evaluation of our Gaussian Process model are detailed in Table 3. Table 3. Evaluation of Gaussian Process Models for three pollutants Pollutant

MSE

Training R2

Test R2

NO2 SO2 heightPM2.5

0.000 0.000 0.000

0.497 0.314 0.206

0.453 0.332 0.199

The temporal dependence of the pollution levels found by our Gaussian Process models are visualized in Figs. 5 through 7.

Statistical interpretation Examining Figs. 5 and 6 for NO2 and PM2.5 , respectively, one can see that the Gaussian Processes are not fitting to higher concentration levels. Though we attempted to increase the

232

The Manhattan Scientist, Series B, Volume 4 (2017)

DePinho et al.

amplitude of the Gaussian Process, the fit maintained its clipped appearance. The predicted levels of NO2 seems to experience significantly less clipping, hence its superior R2 values.

Figure 5. Gaussian fit of NO2

Figure 6. Gaussian fit of PM2.5

Figure 7. Gaussian fit of SO2

DePinho et al.

The Manhattan Scientist, Series B, Volume 4 (2017)

233

A common theme in all three of our Gaussian Process models is that they did not perform significantly better than our LUR models, as we initially hypothesized (and as existing studies in literature would indicate). One possible explanation is that further feature engineering is required - in a number of studies in literature, significant features include variables absent from our model, e.g. distance from grid center or monitoring site to city center, total length of road segments contained in grid etc. Another compelling possibility explaining the poor performance of the Gaussian Process model is that it mixes geospatial predictors extracted from the 1970â&#x20AC;&#x2122;s and 80â&#x20AC;&#x2122;s with dynamic predictors gathered from 2017. Given the limited time and computing resources, coupled with the difficulty of obtaining model-ready data we were unable to gather sufficient geospatial and dynamic data from the same time period. Further data collection is our immediate future goal.

Interface The primary goal for our web interface was to make data exploration and analysis an interactive and educational experience for all users. We wanted to implement a tour that would educate the user on the effects of air pollution and our research findings. The other setting in our interface is an interactive advanced view setting, which would include air pollutants in relation to health effects, visuals of the different geographical layers, and a time bar to show how pollution varies throughout the day. While we have successfully implemented the framework for most of the features in the advanced view button, the take a tour option still needs to be completed. We also need to add all of the desired layers and data to the advanced view, and change the way the data is visualized. To date, our progress is as follows: We have used HTML as a tool to set up the format and outline of our interface. This includes assigning space allocated to the map, sidebar, and footer. We used CSS as our style-sheet for our interface. Here, we implemented the color scheme, the translucent sidebar and footer, and the two buttons (tour and advanced view). We have a fully functioning drop down menu, as well as a placeholder area for a small graph to show pollution over time when specific cells in the grid are hovered over. Lastly, we used D3.js, a Javascript library, to attach an interactive map that is drag-able and zoom-able, to implement an on-click reaction to the buttons, and to create an interactive drop-down menu on the advanced view button. Some of the layers have on-click functionality, while others still need to be implemented. Further steps will include adding in more functional data layers, changing the way pollutant concentrations are visualized, making graphs to show pollutant concentrations over time when cells are hovered over, add in the take a tour feature, and add in health concern layers.

Conclusions Air quality measurements, while not widely available or understandable, are crucial for understanding public health. Given that the average person is unaware of the air quality in the area they live, our main goal was to model the intra-urban pollution variations in the Boston area and present this information to the public in a intuitive and engaging way. While there were time and

234

The Manhattan Scientist, Series B, Volume 4 (2017)

DePinho et al.

resource constraints, which limited the quality of the data collected, our models can be expanded upon to create better predictions in a variety of places. Our interface could also be generalized to fit a number of other cities and their data. Since Sulfur Dioxide, Nitrogen Dioxide, and Particulate Matter 2.5 can exacerbate cardiovascular and respiratory issues, it is crucial that the public have knowledge of areas to avoid and city-specific issues to be addressed. Ideally, our models, with up to date data and more parameters, would accurately predict criteria pollutant concentrations in each grid cell; thus areas with problematic concentrations could be appropriately researched in an effort to reduce concentrations. Most importantly, with health layers the public can reduce their risk of cardiovascular or respiratory issue flare ups by assessing the temporal and spatial elements of the map.

Acknowledgement The authors would like to thank Weiwei Pan of Harvard Institute for Applied Computational Science, Pavlos Protopapas of Harvard Institute for Applied Computational Science, Gary Adamkiewicz of Harvard T. H. Chan School of Public Health, and Jaime Hart at Harvard T. H. Chan School of Public Health.

References [1] The American Heart Association. Circulation, vol. 121, no. 21, Oct. 2010, pp. 2331–2378. doi:10.1161/cir.0b013e3181dbece1. [2] US Environmental Protection Agency. Air Quality Criteria for Particulate Matter (October 2004). Available at: https://cfpub.epa.gov/ncea/risk/recordisplay.cfm?deid=87903. Accessed July 26, 2017. [3] US Environmental Protection Agency. Risk and Exposure Assessment to Support the Review of the SO2 Primary National Ambient Air Quality Standards: Final Report (July 2009). Available at: https://www3.epa.gov/ttn/naaqs/standards/so2/data/200908SO2REAFinalReport.pdf. Accessed July 26, 2017. [4] US Environmental Protection Agency. Risk and Exposure Assessment to Support the Review of the NO2 Primary National Ambient Air Quality Standard (November 2008). Available at: https://www3.epa.gov/ttn/naaqs/standards/nox/data/20081121 NO2 REA final.pdf. Accessed July 26, 2017. [5] Brook, R.D. “Air Pollution and Cardiovascular Disease: A Statement for Healthcare Professionals From the Expert Panel on Population and Prevention Science of the American Heart Association.” Circulation, vol. 109, no. 21, Jan. 2004, pp. 2655–2671., doi:10.1161/01.cir.0000128587.30041.c8.

Using kurtosis for mathematical classification Tenzin Kalden∗ Department of Mathematics, Manhattan College Abstract. Classification is an important topic in applied mathematics and artificial intelligence. It is the framework for which biometrics, word and character recognition, and many other new and important applications are run “under the hood.” One classic example of a binary classifier is one which employs the mean and standard deviation of the data set as a mechanism for classification in mathematics. Principal Component Analysis has played a major role in this effort. This paper proposes that one should also include kurtosis in order to make this method of classification more precise.

Introduction Principal Component Analysis (PCA) is an integral and wide-spread technique used in classification. One need only search for PCA in scholarly journals to realize the huge number of citations (as an example, see Ref. [1]). Mahalanobis distance [2] goes hand-in-hand with PCA and oftentimes is used as a criterion to determine to which class a particular data point belongs. In view of the fact that PCA finds the directions of maximal variance for projections of the data onto lines passing through its mean, PCA relies heavily on the mean and variance of the data in order to make decisions regarding classification. Indeed, PCA is using these two statistics in order to model the shape of each class. DeBonis [3] introduced skew into the classifier in order to obtain a more accurate model of each class and thus obtain a more accurate method of classification. In this paper it is proposed to add one more statistic, namely kurtosis, in order to obtain a more accurate model of each class and thus obtain an even more accurate method of classification. PCA in a sense is assuming that the data is Gaussian and this, in and of itself, may be a debilitating assumption. We have found that by introducing kurtosis into our classification the error rate reduces significantly and as a classifier is comparable in accuracy with known ones.

Towards classification Kurtosis distance One classic binary classifier uses Mahalanobis distances to make a decision as to which class a test point belongs to. The distance from a point to the mean of a collection of points is defined as follows: Definition: Let ~x ∈ R and X ⊆ Rd be a collection of points with mean µ ~ = [µ1 µ2 · · · µd ]. Let A be the matrix whose rows are the points in X. Let E be a square matrix with columns forming the ∗

Research mentored by Mark J. DeBonis, Ph.D.

236

The Manhattan Scientist, Series B, Volume 4 (2017)

Kalden

PCA basis for A. Let B = EA and set ~y = [y1 y2 · · · yd ] = E~xT . Consider the standard deviation in each of the PCA directions, σi (i = 1, 2, . . . , d). The squared distance from ~x to X is defined to be d X (yi − µi )2 d(~x, X) = . σi2 i=1 We will now generalize this distance to the case where we take into consideration the kurtosis of the data. We came up with six different variations of a distance formula which involves kurtosis. The intuition behind the definitions is the following: should a distribution have low kurtosis (i.e. less than three), then the distribution should have fewer extreme outliers and so such outliers should be considered very far from the mean of the distribution, whereas if the distribution has high kurtosis (i.e. more than 3) then the distribution has more extreme outliers and so such outliers should be considered not so far away. Furthermore, we require that the distance reduce to Mahalanobis distance in the case that the kurtosis equals three. Definition: Let ~x ∈ R and X ⊆ Rd be a collection of points which we may assume has mean ~0. Let A be the matrix A whose rows are the points in X. Let E be a square matrix with columns forming the PCA basis for A. Let B = EA and set ~y = [y1 y2 · · · yd ] = E~xT . Let κi (i = 1, 2, . . . , d) be the kurtosis of B in each of the directions of PCA. The squared distance from ~x to the mean of X is defined to be 1/κi d X (yi − µi )6 (1) d(~x, X) = σi6 i=1 d(~x, X) =

2 d X 3(yi − µi ) i=1

κi σ

2 d X 3 y i − µi d(~x, X) = κ σ i=1 i

" # 2 d 6 1/κi X y i − µi (yi − µi ) + d(~x, X) = σ σi6 i=1 " 2 2 # d X y i − µi 3(yi − µi ) d(~x, X) = + σ κi σ i=1

" 2 2 # d X y i − µi 3 yi − µi d(~x, X) = + . σ κi σ i=1

(2)

(3)

(4)

(5)

(6)

Kalden

The Manhattan Scientist, Series B, Volume 4 (2017)

237

Kurtosis error ellipse The data shown in Fig. 1 have high kurtosis in both dimensions, while in Fig. 2 they have low kurtosis in both dimensions. Error ellipses are shown for both Mahalanobis and kurtosis distances.

Figure 1. Error ellipses for Mahalanobis and kurtosis distances on data with high kurtosis in both dimensions.

Figure 2. Error ellipses for Mahalanobis and kurtosis distances on data with low kurtosis in both dimensions.

A classifier based on kurtosis distance Given two classes C1 , C2 â&#x160;&#x2020; Rd of training data, we now give a top-down outline of the algorithm for training the kurtosis binary classifier:

238

The Manhattan Scientist, Series B, Volume 4 (2017)

Kalden

1. Perform the kurtosis distance transformation from C1 ∪ C2 into R2 . In order to perform this transformation, several things must be done for each class Ci (i = 1, 2): (a) Find the mean of class Ci and shift so that its mean is zero. (b) Compute the principal components of Ci and perform the corresponding change of basis. Call the resulting set Pi . (c) Find the kurtosis of Pi . (d) For each point ~x ∈ Pi associate a point in (d(~x, P1 ), d(~x, P2 )) ∈ R2 according to one of the definition provided in Eqs. (1) through (6). Set Ti = { (d(~x, P1 ), d(~x, P2 )) | ~x ∈ Pi }. 2. Compute a linear discriminant function (for instance, Fisher, LSVM or logistic) on the transformed classes T1 and T2

Experimental Results The classifier was applied to thirteen data sets from the UCI Machine Learning Repository [6]. We chose numerical data sets with no missing values consisting of two classes (Table 1). For consistency, we employed a logistic linear discriminant function. Table 1. UCI Machine Learning Repository data sets implemented. UCI data

Instances

Attributes

Classes

appendicitis australian breast cancer german haberman heart ionosphere indian diabetes indian liver liver disorder magic sonar transfusion

106 690 569 1000 306 120 351 768 579 341 19020 208 748

7 14 30 24 3 44 32 8 9 6 10 60 4

2 2 2 2 2 2 2 2 2 2 2 2 2

First, we wanted to verify that using kurtosis distance can indeed improve classification versus Mahalanobis distance. Therefore, we implemented our classifier algorithm described in the previous section but varied the distance function over different distance metrics. We performed 10-fold validation 100 times (thus, 1000 instances) and computed the mean and standard deviation of the error rate. For nearly every data set tested at least one of the versions of kurtosis distance improved the classifier versus Mahalanobis distance. Table 2 summarizes our results. In all but two cases at least one of our kurtosis distances outperformed Mahalanobis distances. Consider the case of the liver disorder data set; a case in which the kurtosis distance performed considerably better than Mahalanobis distance. After shifting the mean to zero, rotating the data to

Kalden

The Manhattan Scientist, Series B, Volume 4 (2017)

239

Table 2. Percent (%) error rate results when training on UCI data using our kurtosis distances versus Mahalanobis distances. Values in bold are the most desirable results. UCI data

Kurt1

Kurt2

Kurt3

Kurt4

Kurt5

Kurt6

Mahal

appendicitis australian breast cancer german haberman heart ionosphere indian diabetes indian liver liver disorder magic sonar transfusion

15.9 0.0 4.5 25.7 26.9 7.2 11.5 26.9 29.1 36.0 20.4 26.7 21.5

16.8 0.0 4.5 26.6 26.6 7.2 11.0 29.0 29.2 40.3 23.3 26.5 22.3

14.9 0.0 4.6 24.2 27.1 7.2 10.6 25.2 28.6 37.5 18.6 25.9 22.4

13.8 0.0 4.4 24.9 26.0 7.1 9.0 23.9 28.9 31.4 18.0 26.0 22.1

13.5 0.0 4.5 24.9 25.7 7.1 9.0 23.9 28.8 31.4 18.0 26.2 22.2

13.5 0.0 4.5 24.9 25.7 7.2 9.2 23.9 28.9 31.7 18.0 26.2 22.2

15.0 0.0 4.4 24.3 27.1 6.9 10.7 25.0 28.8 37.4 18.6 26.0 22.4

its principal components, we computed the kurtosis in the components of class one and class two. The results are displayed in Table 3. Table 3. Kurtosis for the six components of each class of the liver disorder data set. Class

c1 Kurt

c2 Kurt

c3 Kurt

c4 Kurt

c5 Kurt

c6 Kurt

1 2

3.9 2.9

3.2 5.7

3.0 14.1

7.3 4.1

3.9 12.6

10.9 10.1

One sees a number of the components in both classes exhibit a high amount of kurtosis. For instance, the sixth component in both classes exhibits high kurtosis. This fact contributes to the success of our classifier. One other experiment was run regarding the comparison of the kurtosis distance to the Mahalanobis distance. There were two datasets where the Mahalanobis distance performed better: breast cancer and heart. In each case we considered the features which exhibited high kurtosis (i.e. â&#x2030;Ľ 4) and ran the classifier on only those features. For the breast cancer data this amounted to 14 features, while in the case of the heart data were reduced to three features. In both cases the kurtosis distance now performed better than the Mahalanobis distance. Table 4 summarizes our results. We now list our results using our kurtosis classifier and show that they are comparable to results in other articles citing these data sets using other popular methods. In each case we selected the kurtosis distance which performed the best on each of the data sets. Maszczyk and Wlodzislaw [7], using 10-fold cross-validation, compared their proposed clas-

240

The Manhattan Scientist, Series B, Volume 4 (2017)

Kalden

Table 4. Percent (%) error rate results when training on UCI data using our Kurtosis distances versus Mahalanobis distance on the subset of features with kurtosis ≥ 4. UCI data

Kurt1

Kurt2

Kurt3

Kurt4

Kurt5

Kurt6

Mahal

breast cancer heart

5.2 27.8

5.2 27.7

5.2 28.9

4.9 29.2

5.0 29.2

5.2 28.9

sifier (SFM) to several popular classification methods which include support vector machines. Menze et al. [8], also using 10-fold cross-validation ten times, compared their proposed classifier (oRF) to several popular classification methods which include random forest algorithms. Table 5 displays the comparison of our results, using the same experimental method, to some of the results listed in Refs. [7] and [8]. Our results are comparable and in two cases better than the other methods.

UCI data

Kurtosis

SVML[7]

SVMG[7]

SFM[7]

kNN[8]

CART[8]

adaboost[8]

oRF-rnd[8]

oRF-lda[8]

oRF-ridge[8]

Table 5. Percent (%) error rate results: A comparison of our proposed method versus methods found in [7] and [8] on UCI data.

ionosphere diabetes heart liver disorders appendicitis sonar australian

9.0 23.9 7.1 31.4 13.5 25.9 0.0

10.5 23.1 17.5 12.4 24.5 14.5

5.4 23.8 17.2 13.3 13.4 14.4

5.4 22.4 18.8 13.2 12.0 15.8

14 31.7 35.5 32 18.2 -

10.5 34.2 21.3 34.1 27.2 -

6.1 26.6 18.3 26.8 13.1 -

5.6 27.7 20 27.8 14.4 -

5.4 26.1 18 25.8 18 -

5.6 26.2 17.9 26.2 18.2 -

Jeatrakul and Wong [9], using ten iterations of 80% training and 20% testing, compared their proposed classifier (CMTNN) to several popular neural network classification methods. Table 6 shows the comparison of our results, using the same experimental method, to the results listed in Ref. [9]. Our results are comparable and in one instance better than the other methods. Table 6. Percent (%) accuracy results (mean ± standard deviation): A comparison of our proposed method versus methods found in [9] on UCI data. UCI data

Kurtosis

BPNN[9]

GRNN[9]

RBFNN[9]

PNN[9]

CMTNN[9]

ionosphere diabetes liver

91.1±3.2 76.2±3.2 68.7±6.0

90.3±4.2 76.2±4.4 70.0±6.3

93.1±1.8 75.3±3.9 64.1±6.7

90.1±4.0 76.6±2.5 67.5±4.5

85.6±4.0 75.3±3.9 64.1±6.7

93.4±2.9 76.5±3.4 70.7±7.3

Kalden

The Manhattan Scientist, Series B, Volume 4 (2017)

241

Conclusion We have shown in this paper that one might benefit from incorporating kurtosis into a classifier. Table 2 suggests that this might be the case, for when we added kurtosis to the mean and standard deviation of data, the resulting classifier demonstrated in most cases improvement over Mahalanobis distance. Tables 5 and 6 illustrate that a classifier based on the kurtosis distance alone can have results comparable and sometimes better than known classifiers. Our classifier works well on data sets which exhibit ample enough kurtosis. We believe that a classifier which makes use of a distance metric could benefit from considering the use of this kurtosis distance metric. Incorporating other techniques into our classifier could only make it better and we shall do this in future work, e.g. a distance metric that takes into account the clusters of each class (distance to a class could be defined as distance to the closest cluster in the class). However our intention for performing this research was to simply propose the use of kurtosis in classification. The results show that this just might be worth considering.

Acknowledgment The author would like to thank everyone in the mathematics department at Manhattan College for their guidance throughout this research, and especially his mentor Dr. Mark DeBonis.

References [1] Xu S, Zhou Z, Lu H, Luo X, Lan Y (2014). “Improved Algorithms for the Classification of Rough Rice Using a Bionic Electronic Nose Based on PCA and the Wilks Distribution.” Sensors 2014, 14(3), 5486-5501 [2] Mahalanobis, PC (1936). “On the generalised distance in statistics.” Proceedings of the National Institute of Sciences of India. 2 (1): 49–55. [3] DeBonis M (2015). “Using Skew for Classifiers.” Int. J. of Pattern Recognition and Artificial Intelligence, 2015, Vol. 29, No. 03. [4] Li Y, Xu LQ, Morphett J and Jacobs R (2003). “An Integrated Algorithm of Incremental and Robust PCA.” In Proc. IEEE International Conference on Image Processing (ICIP2003), Barcelona, Spain, September 2003. [5] Sebe N, Lew MS, Cohen I, Garg A, Huang TS (2002). “Emotion Recognition Using a Cauchy Naive Bayes Classifier.” 16th International Conference on Pattern Recognition: Proceedings, 11-15 August 2002, Qu´ebec City, Canada, pp. 17-20, IEEE Computer Society Press, Los Alamitos, CA [6] Blake C, Keogh E, Merz C (1998). UCI repository of machine learning databases. Department of Information and Computer Science, UC, Irvine, http://www.ics.uci.edu/∼mlearn/MLRepo sitory.html [7] Maszczyk T, Wlodzislaw D (2010). “Support Feature Machines: Support Vectors are not enough.” WCCI 2010 IEEE World Congress on Computational Intelligence July, 18-23, 2010 - CCIB, Barcelona, Spain, 3852-9

242

The Manhattan Scientist, Series B, Volume 4 (2017)

Kalden

[8] Menze B, Kelm M, Splitthoff D, Koethe U, Hamprecht F (2011). “On Oblique Random Forests.” Machine Learning and Knowledge Discovery in Databases. Springer 6912: 435-469 [9] Jeatrakul P, Wong, KW (2009). “Comparing the performance of different neural networks for binary classification problems.” Eighth International Symposium on Natural Language Processing, 2009, 111-5

Classification of magnetic resonance imaging (MRI) data using small sample sizes Hope Miedema∗ Department of Mathematics, Manhattan College Abstract. While MRI images are critical for the detection of tumors, their use is limited because they are slow to generate. As an effort to accelerate the imaging process, it is useful to know which classifiers work best for the detection of tumors. It is especially important for current research in MRI to analyze the accuracy of prediction of classifiers given a small sample size because observations can take a long time to generate. We used logistic regression, linear discriminant analysis, and Firth’s bias-reduced logistic regression on samples of various sizes to see if one outperforms the others on images with and without total variation regularization. We have found that total variation regularization to reconstruct images results in decreased detectability of tumors. We have also found that there is bias for small sample sizes for LDA and LogistF.

Introduction Magnetic Resonance Imaging (MRI) uses the magnetic properties found in the nuclei of hydrogen atoms. A limitation of MRI is that it is slow. Research is currently being conducted on new methods to accelerate the process. The research presented in this paper is working to determine which classifier is most accurate in tumor detection. The most accurate classifier will then be used to assess which method of acceleration is best. Since MRI images take a long time to generate, scientists need to use small sample sizes. Therefore, the binary classifier determining whether or not an image has a tumor needs to function at high accuracy for small samples. In this paper, we will analyze three classifiers: logistic regression, linear discriminant analysis (LDA), and Firth’s bias-reduced logistic regression (logistF) on various tumor types (small, large, ring), as well as two different imaging methods (with and without total variation (TV) regularization) [1]. The work presented in this paper follows related work which studied the detection of a ring lesion using total variation [2].

Lesion types and imaging methods Three types of lesions were studied, two of which represented tumors and one which represented the growth of a tumor using a ring. The images were approximated using the LaguerreGauss basis [3]. Images and their cross-sections can be seen in Fig. 1. The lesions were placed in multiple locations of an MRI volume. The features used for the classification are the first ten coefficients of the expansion of the images in the Laguerre-Gauss basis set. Sample images with and without total variation regularization are shown in Fig. 2. ∗

Research mentored by Angel R. Pineda, Ph.D.

244

The Manhattan Scientist, Series B, Volume 4 (2017)

Miedema

Figure 1. Lesions representing a small tumor, a large tumor and the growth of a tumor (ring) with cross-sections shown along with their approximations using the first 10 elements of the Laguerre Gauss basis.

Figure 2. Brain scan with ring lesion reconstructed with and without total variation. Note that total variation reduces noise but also blurs the image.

Linear discriminant functions Logistic regression Given a set of data points grouped between two classes, the intention of a linear classifier is to find a line that best divides the data so that any new data tested will be classified correctly. One way of doing this is with an ordinary least squares fit. However, the least squares fit creates a line that allows for probabilities to be less than 0 or greater than 1. Thus, the ordinary least squares fit

Miedema

The Manhattan Scientist, Series B, Volume 4 (2017)

245

is not an acceptable model. A better fit for binary data is a logistic model which creates a logistic curve bounded between 0 and 1 [4]. Binary logistic regression is a method used to classify a data point into two discrete classes. With binary logistic regression, one can obtain the probabilities of a data point belonging to either class. To do this, a dummy variable of 0 or 1 is used for each class. Classification uses characteristics of the data called features to predict which class a data point belongs to. For a logistic curve with one feature, the function is f (x) =

1 ea+bx = , a+bx 1+e 1 + e−a+bx

(1)

where f (x) represents the probability of success and is bounded within (0, 1); the variable x represents the feature. The variable a controls the center of the curve and b controls the growth rate. If the feature is categorical, dummy variables are used. If the feature is continuous, the value is plugged in directly. Using algebra, we can rewrite the above equation called the “log odds” as f (x) = ea+bx , 1 − f (x) h f (x) i ln = a + bx. 1 − f (x)

(2) (3)

At this point, we use the maximum likelihood estimate (MLE) to find estimates for a and b by maximizing the likelihood (using a binomial model) that the training data was generated using those parameters [4]. For our MRI data, the features are the coefficients of the basis of our image using Laguerre-Gauss functions [3], therefore, our features are continuous. Once a and b are found, the probability of being in a certain class is found by plugging in the feature value for x. Linear discriminant analysis Another popular method for classification is linear discriminant analysis (LDA). Reasons to use LDA rather than logistic regression include [4]: 1. When classes are well-separated, since logistic regression is more unstable with well-separated classes, and LDA is not. 2. When the distributions of the predictors are close to normal on a small amount of data. We want to see if reasons 1 and 2 cause LDA to outperform logistic regression. Let k represent a given class and Y represent the response variable (a dummy variable taking on the values of 0 or 1). We want to know the probability of a data point being correctly predicted. Therefore, we want to know the probability that the prediction response Y is in a given class k, given an observation X, equivalently P (Y = k|X). Bayes’ theorem states: P (X|Y = k)P (Y = k) . P (Y = k|X) = Pk i=1 P (X|Y = i)P (Y = i)

(4)

246

The Manhattan Scientist, Series B, Volume 4 (2017)

Miedema

Since we have estimates for P (Y = k) from the training data (prior probabilities), if we can find estimates for P (X|Y = k), we can enter what we know into Bayes’ Theorem to find P (Y = k|X). Therefore, we must find a way to estimate P (X|Y = k). To do this, we assume the probability has a normal distribution. After using the normal distribution function for P (X|Y = k) and our estimates for the posterior probabilities, we can estimate the posterior probability P (Y = k|X). We use the posterior probability to classify the image. Firth’s bias-reduced logistic regression Logistic regression is subject to perfect separation when being used to classify small samples. Perfect separation causes one of the parameters to diverge to infinity. Firth’s approach to reduce bias was proposed as a solution to this problem in logistic regression [5]. Firth’s bias-reduced classifier maximizes the penalized likelihood function which ensures all parameters will converge.

Receiver operating characteristic (ROC) curves ROC curves are a visual tool used to see how well the classifier is performing over various thresholds. A false positive fraction (FPF) is the calculated fraction of times images without a tumor were incorrectly classified as having a tumor. A true positive fraction (TPF) is the fraction of instances when images with a tumor were diagnosed correctly. ROC curves graph the false positive fractions on the x-axis and true positive fractions on the y-axis as thresholds vary from 0 to 1. Visually, a good classifier would cling to the upper left corner indicating a low fraction of incorrect diagnosis and a high fraction of correct diagnosis. In order to tell which classifier is working best, we calculate the area under the curve (AUC) to find which is closest to 1. Fig. 3 shows a sample ROC curve for LDA for ring signal data with TV. K -fold cross validation In an effort to know how our classifier will perform on future data, we used k-fold cross validation. By training the classifier on some of the data and testing on the rest, we can estimate

Figure 3. Sample ROC curve for LDA with a sample size of 200. The AUC for this figure is 0.84.

Miedema

The Manhattan Scientist, Series B, Volume 4 (2017)

247

the performance of the classifier on new data. K-fold cross validation starts by randomly dividing the data into k sections or folds. The first section becomes the first testing set while the other k − 1 sections are used to train the classifier. Once the classifier is trained with the training data, the testing data is then passed through the classifier and we calculate the AUC, which represents how accurately the classifier predicted the classes of the testing data. This occurs k times for each fold and then we take the mean of the k AUC values in order to get a cross-validated estimate the performance of the classifier. Fig. 4 represents tenfold cross validation.

Figure 4. Graphical display of tenfold cross validation. For each fold we use the other 9/10 of the data as a training set.

Since we are analyzing data with small samples, it is possible for an entire fold to be all “No Tumor” or all “Tumor” data. If the entire fold only contains one class, we cannot get both a true positive fraction and a false positive fraction. If this occurs, the fold is disregarded from the average. Therefore, if this occurs many times during tenfold cross validation, there are not enough points to graph a curve and thus, not enough points to calculate an AUC. For our research, to ensure enough points to plot a curve, we made half of the test data “No Tumor” and the other half “Tumor.” Fig. 5 represents tenfold cross validation with the testing data coming from half of “No Tumor” and another half coming from “Tumor” data.

Figure 5. Tenfold cross validation with the balanced method where each training and testing set has half tumor and half no-tumor data.

Results We computed the half and half tenfold cross validation 100 times (reshuffled each time) for each signal type and each imaging method; then, we took the mean of the 100 iterations (Table 1).

248

The Manhattan Scientist, Series B, Volume 4 (2017)

Miedema

Table 1. Mean of 100 tenfold cross validations for all three signals and all three classifiers for both 50 and 200 sample sizes and images with and without total variation regularization.

Discussion Table 1 shows how classifiers performed differently based on different signal types, different imaging methods, or different sample sizes. We see that all classifiers for all tumor types are performing with higher accuracy on data with No TV with the smallest difference (seen in the results with 200 samples) for the large signal. Fig. 6 gives a visualization of the decrease TV causes in accuracy of the small signal using logistic regression and LDA. The small signal was the tumor where there was the biggest difference between with and without TV.

Figure 6. Box plot for logistic regression and LDA for 200 observations of the small signal data. Images with No TV cause both classifiers to perform more accurately than images with TV.

There was high bias occurring in small samples as seen in Table 1. The most accurate crossvalidated estimates occurred when the sample size was large; therefore, the high AUC at small samples sizes must be because of bias. Fig. 7 displays a box plot of all 3 classifiers for small sample sizes 50, 60, 70, and 80. One can see that at sample size 50 there is high bias for LDA and LogistF. LogistF modifies logistic regression to behave like LDA. As sample sizes increase, the bias decreases. With even bigger sample sizes, one sees not only a decrease in bias, but also, a decrease in variance. Fig. 8 displays a box plot for logistic regression and LDA for sample sizes of 50, 100, 150, and 200. There is a significant decrease in bias from 50 to 100 and as sample sizes increase, the variance decreases. To quantify this change, Table 2 displays the standard deviations

Miedema

The Manhattan Scientist, Series B, Volume 4 (2017)

249

for both classifiers for large signal data with No TV. We can see that standard deviation decreases as sample sizes increase.

Figure 7. Box plot of logistic regression, LDA, and Firthâ&#x20AC;&#x2122;s Bias Reduced LR (LogistF) for samples sizes of 50, 60, 70, and 80 for large signal data with No TV.

Figure 8. Box plot of logistic regression and LDA for sample sizes of 50, 100, 150, and 200 for large signal data with No TV.

Table 2. Standard deviations for large signal data with No TV for sample sizes 100, 200, and 300 for logistic regression and LDA.

250

The Manhattan Scientist, Series B, Volume 4 (2017)

Miedema

In order to find the smallest sample size that still gives accurate results, we need to find the sample sizes with low bias, as well as low variance. This is something that we will continue to work on in the future as we explore standard deviation as a function of sample size to examine whether the differences are statistically significant. We will also study the use of averaged k-fold estimates for evaluation of performance, variance through Monte Carlo studies, feature selection methods and non-linear classifiers.

Conclusions We have found that logistic regression had lower bias than LDA, and logistF for 50 sample images. There is bias in small sample sizes for logistic regression, LDA, and logistF and as sample size increases variance and standard deviation decrease. Images with “No TV” cause classifiers to perform more accurately than images with “TV” for our 3 classifiers. We have also seen that the large signal was the least affected by “TV.”

Acknowledgment The author would like to thank everyone in the mathematics department at Manhattan College for their guidance throughout this research, especially Dr. Angel Pineda.

References [1] Lustig, M, Donoho D, Pauly JM. Sparse MRI: The Application of Compressed Sensing for Rapid MR Imaging. Magnetic Resonance in Medicine 2007; 58: 1182-1195. [2] Altaf S, Brenner M. Statistical Binary Classification of MRI Data, Manhattan Scientist. 2017. Series B 4: 215-224. [3] Pineda AR, Yoon S, Paik DS, Fahrig R. Optimization of a Tomosynthesis System for the Detection of Lung Nodules, Medical Physics 2006; 33: 1372-1379. [4] James G, Witten D, Hastie T, Tibshirani R. An Introduction to Statistical Learning with Applications in R. New York: Springer; 2015. [5] Heinze H, Schemper M. A Solution to the problem of separation in logistic regression, Statistics in Medicine 2002; 21: 2409-2419.

Building a mathematical model for Lacrosse analytics Samantha Morrison∗ Department of Mathematics, Manhattan College Abstract. We created a mathematical model to describe the playstyle of the Manhattan College Women’s Lacrosse Team during the 2017 season, focusing on questions posed by student athletes.We examined shots on goal and clears, aiming to evaluate the team’s midfield transitions. This analysis employed a variety of mathematical tools in order to identify, quantify, and compare patterns in players’ behavior during games.

Introduction This is an overview of the project goals and basic understanding of Lacrosse. Project goals The goal of our project was to use mathematics to analyze the performance and style of the Manhattan College Women’s Lacrosse team. After interviews with students athletes Nicole Quivelli and Jordyn DiCostanzo, we identified some concerns of the team and posed four main questions along with hypotheses based on the experiences and inputs of the team members. 1. Does the number of passes in a “clear” affect its likelihood for success? It was hypothesized that a higher number of passes in a clear would make it more likely that the clear was successful. 2. Who were the last four players involved in a shot on goal most of the time? It was hypothesized that the player who we identified as Attacker 2 would be involved in the majority of the shots on goal. 3. Are the midfield transitions effective? It was hypothesized that the midfielders were not involved or not effective in plays. 4. Did the team’s style differ more at the beginning or at the end of the season when compared to their average game? We hypothesized that the team’s style in each game would be different at the beginning of the season, and would become more similar toward the end of the season. Lacrosse terms The following are terms we use to understand the game of lacrosse. Some definitions have been modified for the purposes of our computations, and may differ from professional lacrosse definitions. Fig. 1 shows the field location of the players and the definitions of their positions. 1. We define an attempted clear as an event where possession of the ball is gained in the Defensive Third of the field, and the team attempts to move the ball into the Offensive Third either by running or by passing (Fig. 1). We say a clear is successful if the team is able to get the ball into the Offensive Third. ∗ Research mentored by Helene Tyler, Ph.D.

252

The Manhattan Scientist, Series B, Volume 4 (2017)

Morrison

Figure 1. Lacrosse player field positioning. The field is divided into thirds with the twelve player positions represented by green “X”s. The upper third shows the Jaspers’ goal, with the goalkeeper, G, inside the goal, and the four defenders, D1, D2, D3, and D4, also in the upper third. We call this the Defensive Third. In the middle third we see the three midfielders, M1, M2, M3. We call this the Midfield. Finally, in the lower third we see the opponent’s goal and the four attackers, A1, A2, A3, A4. We call this the Offensive Third.

2. We define the length of a clear to be the number of players involved in that clear. Note that every clear must have length of at least 1 because a clear begins when a player gains possession of the ball. 3. For every shot on goal, we define a shot on goal sequence to be an ordered sequence of four players such that the player in the fourth position made the shot, the player in the third position made the assist, and so on. Note, a shot on goal sequence may contain fewer than four players if once the Jaspers gained possession of the ball, a shot on goal was made before four passes were completed. In these cases, the remaining positions are called empty. 4. We define a midfield transition as a sequence of passes that involves a midfielder. An effective midfield transition transfers possession of the ball from the defenders to the attackers.

Data management Our data set came from game film of the Manhattan College Women’s Lacrosse game film from the 2016 season. This film was made available online and analyzed by Jordyn DiCostanzo and Nicole Quivelli. In this section we will discuss data collection, entry, and processing. Data collection Over the course of the season, the team played sixteen games. The Jasper’s played against a different opponent in each game. These games were filmed, and the videos were made available to the student athletes. DiCostanzo and Quivelli developed a key with a symbol for each event that could occur during the game, for example a ground ball was represented by • and a completed pass was represented by >. Next they watched each of the sixteen games and recorded every play using the key.

Morrison

The Manhattan Scientist, Series B, Volume 4 (2017)

253

Data entry and organization After the games were recorded we specified three parts of the game records to focus on; the shot on goal sequences, the attempted clears, and the completed passes. We listed the four types of players, goalie, defender, midfielder, and attacker, and assigned every starter in each of those types to a number arbitrarily. This is where we get the twelve positions listed in Fig. 1. When other players substituted for a starter, the plays of the starter as well as the substitutes were recorded together under the name of the position. Hence, the passes involving position A1, involve the first of the four starting attackers as well as all players who substituted for her. From here, we entered the number of completed passes between each possible pair of positions into Excel spreadsheets for each of the sixteen games. These served as our adjacency matrices, which we will discuss in a subsequent section. We were then able to compute each of the entries of an “average game” by adding all corresponding entries and dividing by sixteen. We entered both the clears and shot on goal sequences into text files by writing the player’s number followed by a comma until the end of the sequence. Spaces in the shot on goal sequences that were empty were marked by a zero. When entering both the clears and the shot on goal sequences we differentiated between successful clears and shots and unsuccessful or missed clears and shots. Data processing We wrote code in order to calculate the conditional probabilities and the network centrality measures using Python, specifically the NetworkX package. In order to calculate the matrix norms we used MatLab. For our graphics we used Excel.

Probability considerations In order to determine whether the length of a clear affects its likelihood for success, as well as to determine which four players were involved in a shot on goal most of the time, we used the concept of conditional probabilities [1]. We determined whether the length of a clear affects its likelihood for success by calculating the conditional probability of success given length. That is, we divided the number of successful passes less than or equal to a given length by the total number of passes less than or equal to that length, Pr{Success ∩ Length} . (1) Pr{Success|Length} = Pr{Length} We addressed the question of which four players were involved in a shot on goal similarly. As discussed above,we recorded plays by position rather than individual players. Therefore, we were able to determine the positions of the players who were most likely to be involved in shot on goal sequences, but not the individual players. We also considered empty spots in the shot on goal sequence as their own position. We did this by determining the conditional probability of a

254

The Manhattan Scientist, Series B, Volume 4 (2017)

Morrison

player’s position given her placement in the shot on goal sequence. That is, we divide the number of players of a given position in a given place during the shot on goal sequence by the total number of players in the same place in the shot on goal sequence, Pr{Position|Placement} =

Pr{Position ∩ Placement} . Pr{Placement}

(2)

Network theory In order to determine whether the midfield transitions were effective, as well as how the team’s play style changed throughout the season, we used concepts from network theory. We considered the Jaspers as a weighted, directed network where each position is represented by a node and arrows between nodes are weighted according to the number of passes between positions. An example of a weighted, directed network with three nodes can be seen in the graph in Fig. 2.

Figure 2. From the directed graph on the left, we see that player 1 passes to player 2 three times. Because of this, there is a 3 in the entry of the adjacency matrix, A12 , on the right.

Adjacency matrices and network centrality measures The following are definitions of commonly used network centrality measures, along with our interpretation of these measures when applied to the game of Lacrosse [2, 3, 4, 5]. 1. We define an adjacency matrix, or passing matrix, Aij = [Aij ] as an n×n matrix, containing the number of passes from player i to player j, or the weight of the arrow from i to j, for each entry aij , where n is equal to the number of positions. 2. We define the length of an arrow in a graph to be the reciprocal of its weight. For example, in Fig. 2, the weight of the arrow from player 1 to player 3 is 13 . 3. We define the geodesic distance from player i to player j, denoted dij , as the length of a shortest path from player i to player j. If there is no path from player i to player j, then dij is infinite. In Fig. 2, the distance from player 3 to player 2 is infinite. Also in Fig. 2, there are two paths from player 1 to player 3, one of length 1 and one of length 65 . Since 56 is less than 1, the distance from player 1 to player 3 is 65 . 4. We define the betweenness score of a player, CB (i), as the extent to which a player, i, lies on the shortest path between two other players, j and k. We calculate betweenness by dividing

Morrison

The Manhattan Scientist, Series B, Volume 4 (2017)

255

all shortest paths between all pairs of j and k through i by the total number of shortest paths between j and k, 1 X nijk . (3) CB (i) = 121 j6=k6=i gjk

5. The PageRank score of player i, xi , is a recursive notion of “popularity” or importance. The PageRank algorithm was originally developed by Google in order to rank search results. Because of its recursive notion, it is calculated using the eigenvector of the adjacency matrix, xi = p

X Aji x + q, out j L j j6=i

(4)

P where Lout j = k Ajk are the total passes made by player j. 6. We define the eccentricity score of player i, CE (i), as the inverse of the distance from the player who is farthest from her, 1 CE (i) = , (5) max{dji } where dji is the distance between players i and j.

Matrix norms In order to determine whether the play style of the team has changed significantly over the season, we use the concept of matrix norms. Matrix norms are a notion of size of a matrix. While there are many different matrix norms, below are the three matrix norms that we use in our analysis, as well as their definitions and interpretations of their applications to the game of lacrosse [4, 6]. 1. We define the infinity norm of a matrix, kAk∞ , as the maximum row sum of the matrix. In our adjacency matrices we can interpret this to be the maximum number of passes that a player sent during a game. 2. We define the 1-norm of a matrix, kAk1 , as the maximum column sum of the matrix. In our adjacency matrices we can interpret this to be the maximum number of passes that a player received during a game. 3. We define the Euclidean norm of a matrix as the square root of the sum of the squares of all entries in the matrix, qX kAkF = A2ij . (6) Compared to the 1-norm and infinity norm which look at an extreme, the maximum column or row sum, the Euclidean norm looks at more of an average size of the game.

Results and Interpretation In this section we discuss the results of our analysis and their interpretations. We also discuss possible conclusions from these results and whether they support or contradict our hypotheses.

256

The Manhattan Scientist, Series B, Volume 4 (2017)

Morrison

Analysis of Clears Our analysis of the seasonâ&#x20AC;&#x2122;s aggregate clears can be seen in Fig. 3. We find that over the entire season, the team made 296 attempted clears, and they were successful 80.74% of the time. The Jasperâ&#x20AC;&#x2122;s highest successful clear rate per game was 89%, which they achieved against both Marist and Sacred Heart. The game with the lowest successful clear rate was Manhattan vs. Rutgers, with the likelihood that a clear was successful at 33%. The greatest number of attempted clears in a game was 22, which occurred against both Monmouth and Niagara, and fewest number of attempted clears in a game was 14, which occurred against Quinnipiac.

Figure 3. The frequency and success rates of clears by length

Excluding clears with length 7, the success rates of the clears increased as the l engths of the clears increased; this was only true in the aggregate count. It was not true of most individual games. As we know from the definition, clears must have at a length of at least 1, and the longest clears of the season had a length of 8. We see that clears of longer lengths were much less frequent. In fact, only 17.9% of clears were of length 5 or greater, and 0.34% of clears were of length 8. So, while it is true that longer clears had a higher success rate, we do not find enough evidence to conclude that our hypothesis was correct because there were so few clears of greater lengths. Analysis of Shots on Goal Our analysis of the aggregate shots on goal can be seen in Fig. 4. We separate the shots on goal by scored goals and missed shots. We find in our analysis that midfielders are playing a big role in making shots on goal. Midfielders made 37.58% of successful shots and 41.88% of missed shots. Because of the way that we looked at position rather than player, we can not support nor contradict our hypothesis that Attacker 2 was the most involved player in shots on goal. We made this hypothesis under the assumption that a shot on goal made by someone other than an attacker would be rare. However, because of the unexpectedly large role that midfielders were playing in

Morrison

The Manhattan Scientist, Series B, Volume 4 (2017)

257

Figure 4. Top: The make up of the shot on goal sequence for scored goals. We see that 61.74% of goals were scored by attackers and 37.58% of goals were scored by midfielders. We also find more empty spots than expected. We see that 14% of goals had no assist, 19% had no third, and 23% had no fourth. Bottom: The make up of the shot on goal sequences for missed shots, we see that the empty spots are even more prevalent. We see that 20% of missed shots had no assist, 26% had no third, and 30% had no fourth.

shots on goal, we believe that it is unlikely that our hypothesis was correct. Another important finding from the shot on goal analysis was that more shots on goal than we expected had less than four players involved. Only 80% of missed shots and 86% of scored goals had a complete four person sequence. We also see that defenders were involved in shot on goal sequences, and in fact they were more involved in successful shots on goal than missed shots. This is interesting because it suggests that either defenders are in the midfield or offensive zone, making them available to be included in the sequences, or that the team is not making more than four passes between the defensive zone and making a shot on goal. The second option is supported by the clears analysis which we discussed above.

258

The Manhattan Scientist, Series B, Volume 4 (2017)

Morrison

Network centrality measures Our analysis of the team’s network centrality measures can be seen in Table 1. Our hypothesis was that the midfielders were not involved or effective in plays. We are able to conclude that our hypothesis was false. We find that the position with the highest betweenness score is M2, and a midfielder also has the second highest betweenness score. However, the betweenness score of M2 is much greater than the betweenness score of any other position. We also see that four out of the twelve positions have a betweenness score of 0, and the goalie has the third highest betweenness score. We also see that with the exception of A2, who tied for first, the midfielders had the highest PageRank scores. We see that the PageRank scores of the defenders are much smaller than the scores of the attackers and midfielders. When looking at the eccentricity scores we see that M2 has the highest eccentricity scores, but the attackers and the midfielders’ scores all within the range of 0.632 and 0.75 while the defenders and goalie’s scores were between 0.247 and 0.461. Finally, we see that while D1 does not have the lowest eccentricity score, it is the position that is farthest from all of the attackers and midfielders, and similarly, although A1 has an eccentricity score of 0.632, it is farthest from all of the defenders and the goalie. Table 1. Network centrality measures of each position of the team based on the aggregate game. The left column gives the betweenness scores with the highest score in bold. The middle column shows the PageRank scores with the highest score, which was achieved by two positions, in bold. Finally, the right column gives the eccentricity scores, with the highest score in bold; the position that was farthest away, is shown next to the score.

Player A1 A2 A3 A4 M1 M2 M3 D1 D2 D3 D4 G

CB (i) 0 0.058 0.017 0 0.099 0.322 0.066 0 0 0.083 0.033 0.088

xi 0.10 0.15 0.11 0.11 0.13 0.12 0.15 0.03 0.03 0.02 0.03 0.02

CE (i) 0.632 D1 0.705 D1 0.637 D1 0.632 D1 0.682 D1 0.718 D1 0.750 D1 0.374 A1 0.461 A1 0.335 A1 0.419 A1 0.247 A1

Distance from the Average Game Our analysis of each game’s distance from the average game can be seen in Fig. 5. We find the distance of each game from the average game by taking the difference of the adjacency matrices and comparing the norm of the resulting matrix to the norm of the adjacency matrix of the average

Morrison

The Manhattan Scientist, Series B, Volume 4 (2017)

259

Figure 5. The distance of each game from the average game by the Euclidean norm, 1-norm, and infinity norm. The games are plotted in chronological order from left to right. The units on the vertical axis are in passes.

game. Our hypothesis was that the games in the beginning of the season would be farther from the average game, while the games toward the end of the season would be closer. None of the matrix norms that we used supported this hypothesis. In fact, the first game of the season, Manhattan vs. Army, and the last game of the season, Manhattan vs. Fairfield were the most similar games to each other.

Future Work We hope to evaluate more network centrality measures and matrix norms, and explore the ways in which they can be applied to the game of lacrosse. We also hope to analyze the team by player rather than by position, and to analyze the game data of the Jaspers’ opponents.

Acknowledgement This work was funded by the Jasper Summer Fellows program. The author would like to thank Dr. Helene Tyler for advisement and guidance and Andr´e Oliveira for his help with Python. She also acknowledges Nicole Quivelli and Jordyn DiCostanzo for their help in collecting the Jaspers’ game data, and their expertise in lacrosse.

References [1] Devore, J. N., Berk, K. N. (2012). Modern Mathematical Statistics with Applications. Springer Texts in Statistics.(ISBN 978-1-4614-0390-6) [2] Oliveira, A. (2014). Pitch perfect: Re-analyzing the passing networks of Manhattan College women’s soccer. The Manhattan Scientist, Ser. B, Vol. 1, p 109

260

The Manhattan Scientist, Series B, Volume 4 (2017)

Morrison

[3] Oliveira, A. P., Tyler, H. R. (2015). Measurement and Comparison of Passing Networks in Collegiate Soccer. The Minnesota Journal of Undergraduate Mathematics, 1, 1. [4] Clemente, F. M., Martins, F. M. L., Mendes, R. S. (2016). Social Network Analysis Applied to Team Sports Analysis. SpringerBriefs in Applied Sciences and Technology. (ISBN 2191-5318) [5] Newman, M. E. J. Networks: An introduction. Oxford New York: Oxford University Press, 2010. [6] Trefethen, L. N., Bau, D. (1997). Numerical Linear Algebra. SIAM. (ISBN 0898713617)

Search for Lorentz invariance violation from Fermi LAT gamma ray bursts Linh Nguyen∗ Department of Physics, Manhattan College Abstract. Lorentz invariance is such an important principle of fundamental physics that it should constantly be subjected to experimental scrutiny as well as theoretical questioning. It has been suggested that the interactions of energetic particles with the foamy structure of space-time, thought to be generated by quantum-gravitational (QG) effects, might violate Lorentz invariance, so that they do not propagate at a universal speed of light [1]. Distant astrophysical sources of energetic photons with rapid time variations, such as gamma ray bursts (GRBs), provide ideal experimental opportunities for testing Lorentz invariance. The Fermi Large Area Telescope (LAT) is an excellent experimental tool for making such tests with sensitivities probably exceeding those possible using other detectors.

Introduction According to relativistic kinematics, a photon’s velocity in vacuum does not depend on its energy [2]. However, vacuum quantum fluctuations can influence light propagation as it allows the creation of continuously appearing and disappearing charged fermion pairs [3]. High-energy astrophysics and related aspects of cosmology are the essential science issues for the Fermi Large Area Telescope (LAT) [4]. There is some information that may be provided by measurements related to fundamental physics. One such possibility is to use transient high energy emissions from distant astrophysical objects observed by the LAT to probe the validity of Lorentz invariance. From the quantum-mechanical point of view, the vacuum is the lowest-energy state of a physical system. It should be regarded as a medium that may have a virtual structure, even if it is devoid of physical particles. As such, it may have non-trivial effects on particle propagation, even if Lorentz invariance is an underlying principle. This effect is, of course, familiar in the cases of photons propagating through plasmas at high temperatures or superconductors at low temperatures. Might high-energy photons of astrophysical origin exhibit analogous effects? Lorentz invariance expresses the proposition that the laws of physics are the same for different observers. For example, an observer at rest on Earth or one who is rotated through some angle, or traveling at a constant speed relative to the observer at rest.Gamma ray bursts comprise a unique phenomenon in astronomy. They are not coming from the surface of the Earth nor are they due to the Earth’s atmosphere [5]. They are originating randomly in the sky, and are happening in deep space. Long GRBs come from hypernovae. Short GRBs come from merging neutron starts. In the supernova, the energy gets flung out in all directions, expanding as a sphere. If that energy could be collected and send out as a beam that would explain the bursts. When the core of a very massive star collapses, the material outside the core falls down, forming an accretion disk. The magnetic ∗

Research mentored by Rostislav Konoplich, Ph.D.

262

The Manhattan Scientist, Series B, Volume 4 (2017)

Nguyen

field of that material and from the black hole coil around, pointing up and down out of the disk and away from the black hole. This launches twin beams that have the energy equal to the total energy of the supernova. The aim of this project is to study NASA software and extract data that lead to the energy and time information of photons, in order to prepare a basis for the study of deviations of the photon velocity from c. In this project, we propose to perform a study of the timing properties of gamma radiation emitted by distant gamma ray bursts (GRBs) in high-energy spectral gamma ray band. The light curves of GRBs exhibit a large variation in time scale, ranging from dozens of milliseconds to minutes. Therefore, looking for correlations with distance of the time lags between the arrival times of photons of different energies in GRBsâ&#x20AC;&#x2122; light curves might shed light on the refractive ability of the vacuum and thus to probe a possible violation of Lorentz invariance a quantum gravity and other relevant theoretical frameworks.

Extract data from Fermi LAT The analysis should be performed on a selection of Fermi LAT GRBs with measured redshifts and relatively bright GeV emission. We downloaded an important table information which contains the time, right ascension (RA) and declination (DEC) to locate the event that we wanted to retrieve data (Fig. 1). In order to retrieve LAT photon event information, we used the Fermi Science Tools, Query, FV and XQuartz from NASA.

Figure 1. LAT photon, event, and spacecraft data query

Figure 2. Data query

To select all photons within a circular region around the source: The FSSCâ&#x20AC;&#x2122;s website data server (FERMI Science Support Center) provide an online tool to search for particular event information. In this thread we would use this set of parameters (Fig. 2):

Nguyen

• • • • • • • •

The Manhattan Scientist, Series B, Volume 4 (2017)

263

Object name or coordinates: 139.34, -25.67 Coordinate system: “J2000” Search radius (degrees): 12 Observation dates: 517160738.28, 517160938.28 Time system: “MET” Energy range (MeV): 100, 500000 LAT data type: “Extended” Spacecraft data: “checked”

The “Query Submitted” webpage will be displayed and provide an estimate of the time to complete the query, as well as a link to the results webpage. We downloaded the events data file (EV00) to ourworking directory. Data preparation: We would like to prepare the GRBs’ data for analysis by executing the Science Tool from a Linux terminal. We uploaded the events data file to FermiTools directory to preparing these data for analysis. We used gtselect command to make cuts based in columns in the event data file such as time, energy, position, zenith angle, instrument coordinates, event class, and event type. We directed the input file to the events data file ( EV00) that we downloaded from Query. Then inserted the parameter from Fermi Table (Fig. 1) to the FermiTools (Fig. 3). We used ScienceToolsv10r0p5-fssc-20150518-x86 64-unknown-linux-gnu-libc2.12 version which can be directly download from NASA website (Installation steps can also be found on the website). In this thread we would use this set of parameters:

Figure 3. Example of data preparation by Science Tool

We downloaded the filtered data back into our working directory and opened FV and XQuart to produce a text file based on the filtered file. This step is straightforward as the two programs are designed to open the filtered file and provide selection to modify the data. For the purpose of this research project, we only needed the photon energy and time information (Fig. 4). We also make adjustments to the text file in the terminal to delete the header and set the initial time to 0 for further examination (Fig. 5).

264

The Manhattan Scientist, Series B, Volume 4 (2017)

Nguyen

Figure 4. Modify data file through FV

Figure 5. (Top) The command to adjust the text file. (Bottom left) The Energy and Time text file before adjustment. (Bottom right) The Energy and Time text file after adjustment.

Nguyen

The Manhattan Scientist, Series B, Volume 4 (2017)

265

The data file contains important information about the photon energy and time for approximately 35 high energy gamma ray bursts, which serve as the main sources for the data analysis step. The sources are used to create histogram of the distribution of number of photons emitted by using ROOT, the European Organization for Nuclear Research (CERN) data analysis framework (Fig. 6). The structure of each histogram is based on computational models composed in C++. We compared the histogram of distribution of number of photons detected by Fermi LAT to the expected value and saw that there is a time difference in arrival of photons. We used this information as the base line to build different estimator on the data of GRBs to compress the distribution graph as much as possible to get the same features as the ideal case.

Figure 6. Distribution of photons built from GRBs Fermi-LAT data

Conclusion The purpose of this project was to study NASA software and process sources to extract energy vs. time information of photons in Gamma Ray Bursts and to prepare a basis for the study of deviations of photon velocity from the speed of light c. The method we used which is apply multiple software to analyze raw data is consider efficient and effective. The modified files were use as main sources for building estimators to look for correlations with distance of the time lags between the arrival times of photons of different energies in GRBsâ&#x20AC;&#x2122; light curves.

Acknowledgments This work was funded by the National Science Foundation Grant PHY-1402964. The author would like to thank Dr. Rostislav Konoplich for giving him the opportunity to join his team at CERN in Switzerland and providing guidance throughout the entirely of this research project. He

266

The Manhattan Scientist, Series B, Volume 4 (2017)

Nguyen

would also like to thank Drs. Alexandre Sakharov and Konstantin Zhukov, and also Nikita Belyaev for helping and enabling him to learn from the amazing experience at CERN.

References [1] J. Ellis, N. Harries, A. Meregaglia, A. Rubbia, and A. S. Sakharov, Physical Review D 78, 033013 (2008) doi:10.1103/PhysRevD.78.033013 [arXiv:0805.0253 [hep-ph]]. [2] V. Gharibyan, Physics Letters B 611, 231 (2005) doi:10.1016/j.physletb.2005.02.053. [3] M. Urban, F. Couchot, X. Sarazin, and A. Djannati-Atai, The European Physical Journal D 67, 58 (2013) doi 10.1140/epjd/e2013-30578-7. [4] W. B. Atwood et al. [Fermi-LAT Collaboration], Astrophys. J. 697, 1071 (2009) doi:10.1088/0004-637X/697/2/1071 [arXiv:0902.1089 [astro-ph.IM]]. [5] J. Ellis, N. E. Mavromatos, D. V. Nanopoulos, and A. S. Sakharov, Astronomy & Astrophysics 402, 409 (2003) doi:10.1051/0004-6361:20030263 [arXiv:astro-ph/0210124]

Study of simulated particle data and practical applications Danielle Rabadi∗ Department of Physics, Manhattan College Abstract. This research focuses on the advantages of using ATLAS Outreach Open Data in high school and undergraduate classrooms. CERN provides particle collision datasets, which the public can analyze and explore. This data was collected by the ATLAS detector, one of the four detectors of the Large Hadron Collider. Users can easily access the necessary software by downloading each from their respective sources. Software includes: Virtual Machine, ROOT, MadGraph, Pythia, and PGS. ATLAS also provides ROOTbooks documents that describe the background of the Higgs boson and other particles, how particles interact with each other, the outcomes of those interactions and how to set up the environment to explore the datasets. The point of using all this software is to simulate collision events in frameworks of different theoretical models. The user may analyze use the simulated events to predict effects expected with real data and to gain a better understanding of the basic concepts of particle physics. With knowledgeable guidance, this platform has the potential to be a beneficial learning tool for students to use when studying particle physics. Further research includes using this software to analyze large quantities of collected statistical data.

Introduction Scientists in every corner of the world continue to develop and create technology to make arduous tasks easier and less time consuming. From computational models to precision instruments, technology over the past few decades has proven to be an essential asset for the science community. Scientists at CERN use a variety of hardware and software to cut down on time spent doing calculations and focus more on analyzing the data gathered by instruments respective to their studies. CERN is the world’s largest nuclear research facility. It is home to the Large Hadron Collider (LHC), a high energy particle accelerator completed in 2008. Particle accelerators are used to test out elementary particle and nuclear research hypotheses [1]. The LHC uses a combination of hardware and software to test, measure, and record different variables present during a particle collision. Accelerated almost to the speed of light, two high energy particle beams are directed by the strong magnetic field created by superconducting electromagnets. When those particle beams collide, they interact to form a heavier particle for a brief moment in time. Since this particle is so unstable, it decays into lighter stable particles [2]. There are many different combinations of particle interactions to test, as each produce variable outcomes, such as the type of particle formed during interaction and the types of particles, which decayed from that particle. One such experiment began with the collision of two protons. Those protons had interacted to form what scientists discovered to be the Higgs Boson [3]. Theorized in 1964, the Higgs Boson explained how matter acquired mass, and became a basis for the Standard Model. ∗

Research mentored by Rostislav Konoplich, Ph.D., and Josif Zhitnitskiy, Ph.D.

268

The Manhattan Scientist, Series B, Volume 4 (2017)

Rabadi

Digging deeper into the research Since 2008, scientists at CERN have been collecting and analyzing particle data with the hope of making new discoveries in particle and nuclear physics fields. In 2014, scientists from CERN released particle collision data collected from the ATLAS detector in the LHC to the public called ATLAS Outreach Open Data [4]. The purpose was to encourage anyone with a computer and an interest in particle physics to explore the datasets provided. Not only does this platform give people the opportunity to find and discover new particles, but also doubles as an educational tool that students in high school and undergraduate physics classrooms can benefit from. Particle physics is not an easy field to learn; it is even more difficult to grasp conceptually. This program helps beginners in the field to understand those basic concepts. As a result, students become more knowledgeable about particle physics and are encouraged to pursue physics as a field of interest. The platform provided by CERN and the software necessary to run analysis are easily accessible in classrooms provided there is a functioning computer. To confirm that the programs are a beneficial learning tool, this research focused on studying how the programs help the user visualize, use, and interpret open-source data. In order to graph simulated data, we used programs and software provided by the ATLAS Outreach, as well as external software. ATLAS Outreach provided ROOT, Scientific Linux, datasets from the ATLAS detector, and ATLAS Virtual Machine ROOTbooks. MadGraph, Pythia, and PGS software were necessary programs obtained outside the programs provided by ATLAS. These external programs were used for the purpose of creating theoretical models of collision events. The ATLAS Outreach provided some computational tools, such as ROOT for data analysis and Scientific Linux as an operating system as well as a complete set of experimental collision data from the ATLAS Detector in Geneva. The experimental data set comes from the 2012 run of the LHC, the same data set used in the original discovery of the Higgs. In conjunction with ROOT and Linux other software used included: MadGraph5, Pythia8, and PGS. These were used to create theoretical models of collision events which were then analyzed in ROOT. In order to analyze the datasets, the user must read through the ATLAS VM ROOTbooks as this provides a guide to download, use, and explore the data (Figs. 1-4). There are different links to documents that explain specific steps to set up and explore the datasets, and background information, including: background information surrounding the Higgs Boson, explanation of Python ROOTbook, and setting up the environment for a Virtual Machine to run in [4]. A Virtual Machine (VM) is a host operating system that acts as a virtual computer on which files of the datasets can be accessed and viewed, using ROOT software. VM serves the same function as a regular computer, yet is portable and manageable, and accesses data on a secure network. Being able to save progress on datasets and transfer it to different computers via flash drive is very convenient, especially if a particular computer is too slow to run events or experiences technical difficulties.

Rabadi

The Manhattan Scientist, Series B, Volume 4 (2017)

Figure 1. Background information on the Higgs and its interaction with other particles.

Figure 3. Details pertaining to the different types of datasets provided by ATLAS.

269

Figure 2. Step by step guide on how to setup the computer to access datasets.

Figure 4. Describes how to produce histograms with Executable ROOTbook.

Once the environment is setup and ready to go, MadGraph, Pythia, and PGS need to be downloaded from their respective web sources to simulate processes of specific particle collisions and to create distributions of final state particles. For this research, the Higgs boson production and its decay into four leptons were simulated and analyzed using the external software combined with VM and ROOT software. MadGraph5, a Fortran-based software that analyses every possible simulated particle collision outcome, was used to generate all possible outcomes for a Higgs boson

270

The Manhattan Scientist, Series B, Volume 4 (2017)

Rabadi

production. Through this software, users can view all the different possible collision outcomes in the form of a Feynman Diagram. This is essential when understanding particle interactions. Since the documents describe this process in words, the diagrams help users visualize and conceptually grasp particle interactions. Along with MadGraph, Pythia8 and PGS were implemented onto the platform (Fig. 5). Pythia8 is a program that simulates events of the states of selected particles to the extent at which

Figure 5. Order of software used to produce simulated data.

the simulated outcome is most likely to occur in real life. Each event records the initial and final states of particles, the types of resulting particles, and their momenta. Pythia8 was applied to shower MadGraph generated events (Figs. 5, 6, and 7). PGS, a detector simulator used to record

Figure 6. MadGraph â&#x20AC;&#x201C; MadEvent file to generate selected events.

Figure 7. Pythia File

Rabadi

The Manhattan Scientist, Series B, Volume 4 (2017)

271

and analyze the final states of observed variables of particle collision, was also used in concurrence with MadGraph. After all the events were successfully compiled into a file, the file can be opened in ROOT. Once ROOT Object Browser is accessed, histograms can be opened and edited to look 9). Rabadipresentable (Figs. 8 andThe Manhattan Scientist, 263 Series B, Volume 4 (2017)

Figure 8. An example of combining two different

Figure 8. An example of combining two files onto one histogram to make comparison bedifferenttween filesmodels onto easier. one histogram to make comparison between models easier.

Figure 9. Different possibilities to create a visually

Figure 9. Different possibilities to create a visually appealing graph. appealing graph.

For the present work, this process of creating events and producing simulated data was done

For the present process of creating events datawas wasapplydone three times for thework, samethis particle collision interaction; theand onlyproducing differencesimulated between each three times for the parameters same particle only difference between each was ing different of collision Standard interaction; Model (SM),the Beyond Standard Model (BSM), and applying Mixture, different parameters of Standard Model (SM), Beyond Standard Model (BSM), and Mixture, between the SM and BSM, to each dataset. The resulting histograms with showering applied to between the SM and BSM, toand each dataset. The resulting histograms with applied to datasets of the SM, BSM, Mixture were provided to Tyler Reese for hisshowering research project. datasets of the SM, BSM, and Mixture were provided to Tyler Reese for his research project. Summary and Conclusion

Summary and Conclusion

Using this software individually and jointly for the first time was simple and educational. ATLAS VM ROOTbooks takes the user step by step through the process of setting up computer Using this software individually and jointly for the first time was simple and educational. environment and producing graphs of selected particles and comparing multiple variables from ATLAS VM ROOTbooks takes the user step by step through the process of setting up computer one or more datasets. MadGraph generates all possible results for a collision, making it easy to environment andthe producing graphs combinations. of selected particles and comparing multiple variables from understand different particle

one or more datasets. MadGraph generates all possible results for a collision, making it easy to This platform is easy to use as there is enough information for someone to set up and explore understand the different particle combinations. the datasets. Even though there is enough information to visualize “how” particles collide and

This platform is easy to background use as thereknowledge is enough is information someone“why” to set particles up and explore “what” they decay into, essential tofor understand behave the datasets. though there is enough information tostudents visualize “how” particles collide a certainEven way when colliding at high energies. As such, in physics classrooms shouldand be “what” theyparticle decay into, background knowledge is essential to understand “why” to particles behave taught physics by their teachers/professors beforehand and this platform visualize those a certain way when colliding at high energies. As such, students in physics classrooms should be concepts taught in classrooms. Using ATLAS VM as a tool in high school and undergraduate taught particle physics by their teachers/professors beforehand and this platform to visualize those concepts taught in classrooms. Using ATLAS VM as a tool in high school and undergraduate

272

The Manhattan Scientist, Series B, Volume 4 (2017)

Rabadi

physics classes will help students understand to concepts of particle physics at a basic level without going into the advanced aspects of the field. Implementing the program in classes would also encourage students to consider looking into physics as a potential field of study. Furthermore, this software could be used in different subjects that require analysis of large data sets, as long as the correct code is executed to compute mathematical equations necessary for study.

Acknowledgments This work was partially funded by the National Science Foundation Grant PHY-1402964. The author would like to thank the School of Science Research Scholars Program for additional support and the Department of Physics at Manhattan College for providing materials for the research. She would also like to thank Dr. Rostislav Konoplich and Dr. Josif Zhitnitskiy for their guidance and support, and Tyler Reese for his collaboration.

References [1] “CERN Accelerating science.” The Large Hadron Collider | CERN, https://home.CERN/topics /large-hadron-collider. [2] Ben Dotson, “How Particle Accelerators Work.” Energy.gov, https://energy.gov/articles/howparticle-accelerators-work. [3] Konrad Jende, “International Masterclass: Hands on Particle Physics.” International Physics Masterclasses, https://kjende.web.CERN.ch/kjende/en/zpath hboson.htm. [4] “Access Open Data from the ATLAS Experiment at CERN.” ATLAS | Open Data and Tools for Education, http://opendata.atlas.CERN/index.php. [5] Bryan Webber, “Parton shower Monte Carlo event generators.” Scholarpedia, www.scholarpedia.org/article/Parton shower Monte Carlo event generators.

Reconstruction of the Higgs boson using computational methods Tyler Reese∗ Department of Physics, Manhattan College Abstract. In high energy particle physics, especially at the Large Hadron Collider, computational models are often created using the Monte Carlo method. These theoretical simulations are used to compare with experimental data in order to draw conclusions about particle detection. The Higgs boson was found in this fashion and this study’s purpose is to replicate those results. It is demonstrated that, of the two Z Bosons, one Z is on the experimental mass shelf while the other Z is virtual if they originated along the Higgs’ decay path. Different Higgs models can be generated with different CP-Parities: even, odd, or mixed. These models can be distinguished by the symmetries of their angular distributions.

Background Five years ago, in the summer of 2012, the ATLAS and CMS Collaborations at the Large Hadron Collider (LHC) at CERN, in Switzerland, announced the observation of the Higgs boson. The Higgs boson is a neutral particle with spin 0 and a mass of 125 GeV [1, 2] initially predicted by the Standard Model. The most frequent detectable signals can be found via the gluon-fusion process and the Higgs decay into γγ, ZZ, or W W gauge boson pairs. Though, the photon is massless it still results from a Higgs decay process. Z and W bosons do have mass, breaking the electroweak symmetry. From the detection and measurement of fermions, the mass of the Higgs Boson can be reconstructed as a result of further decay along the path. In this experiment the process investigated was: gg → H → ZZ → 4 leptons. The purpose of this experiment was to replicate the results of the ATLAS Collaboration using the data set and tools provided by ATLAS Outreach program.

Methodology Three simulations were created with theoretical physical effects applied. Each simulation was done with a different CP parity for the Higgs. The Standard Model was simulated where the Higgs had spin 0 and positive CP parity; this is the CP-even (0+) state [3]. CP parity is a combination of Charge Conjugation and Particle parities. C parity transformation converts a particle into its antiparticle and vice versa. P parity transformation creates something akin to a mirror image of the physical system. Next, for the Beyond the Standard Model simulation, the Higgs was again given spin 0 and negative CP parity, the CP-odd (0-) state. In the third model the Higgs could be a mixture of states 0+ and 0-. After the simulations were completely generated, analysis was done in ROOT. The purpose of the analysis was the actual reconstruction of the Higgs mass. This was done by first taking the four ∗

Research mentored by Rostislav Konoplich, Ph.D., and Josif Zhitnitskiy, Ph.D.

274

The Manhattan Scientist, Series B, Volume 4 (2017) e-

g Z

μ-

g Z

μ-

Figure 1. gg → H → ZZ → 4e

μ+

Reese

μ+

Figure 2. gg → H → ZZ → 2e 2µ

μ+

Figure 3. gg → H → ZZ → 4µ

components of the momentum vectors of the four leptons which originate along the HZZ path, as seen in Figs. 1-3. Events were created along the paths shown by these Feynman diagrams. Only the transverse momentum can be detected so the vector is found by p~(px , py , pz , |~p |) = (pT cos φ, pT sin φ, pT sinh φ, pT cosh φ).

(1)

The angle φ is the azimuthal angle between the transverse plane and the beam axis. In general, the invariant mass of a particle can be reconstructed from its decay constituents in the same way as if it were a two-particle collision, MH2 = m21 + m22 + 2(E1 ·E2 − |~p1 ||~p2 | cos θ).

(2)

Here subscripts 1 and 2 denote values for the ZZ pairs and subscript H is the invariant mass of the Higgs (mother) particle. Charged leptons are directly detected, so once their momentum has been compiled, Eq. 2 can be used to reconstruct masses backwards along the decay path. The invariant mass of a Z Boson can be found by summing the magnitudes of the momentums of its daughter lepton pairs. After invariant masses for the ZZ pair have been calculated, those masses are then used to reconstruct the Higgs mass. Before any combinations could be done though some events had to be excluded based on their final state energy. Events with too low momentum were unlikely to have come from Higgs production, meaning they were resulting from the showering of the beam collision. Also, leptons, once chosen to be included in analysis, were sorted by particle ID and charge: electron positron muon anti-muon groups, electron positron electron positron groups, and muon anti-muon muon anti-muon groups. This allowed for simplicity in the calculations and particle pairings to be easily identifiable by only knowing their grouping.

Theoretical models The ATLAS Outreach program provided some computational tools, such as ROOT for data analysis, Scientific Linux as an operating system, and a complete set of experimental collision data

Reese

The Manhattan Scientist, Series B, Volume 4 (2017)

275

from the ATLAS Detector in Geneva. The experimental data set comes from the 2012 run of the LHC, the same data set used in the original discovery of the Higgs. In conjunction with ROOT and Linux, we used additional software, including MadGraph5, Pythia6, and PGS. These were used to create theoretical models of collision events which were then analyzed in ROOT. The simulations were done in such a way as to provide a clear theoretical basis from which accurate comparisons could be made with any experimental analysis. Also, the models would hopefully provide details on the process for use in creating a more precise analysis of the experimental data set. The procedure for generating the simulations went as follows. First, a generator level group of events was created as a baseline in the Monte Carlo generator MadGraph for a total energy of colliding protons at 8 TeV. Secondly, particle showering effects were applied using Pythia. Then, the geometry and physical effects of the ATLAS detector were simulated and applied to the events, using PGS simulator. The models analyzed were those which had undergone all three generation effects. This should have created an accurate theoretical model matching the actual nature of the collisions.

Figure 4. Mass distribution of Higgs, CP-even

Figure 5. Mass distribution of ZZ pair, CP-even

In all three simulations with different CP the Higgs mass was taken to be as 125 GeV, which reflects the current experimental value. All models were generated having 100,000 event entries. After applying all standard exclusions about 10% of events remained in the 0+ and Mixed simulations and 20% in the 0- case. All analyses showed the peak position of distribution for the mass of the on-shelf Z Boson to be 90 Âą 10 GeV and the virtual Z Boson to be 25 Âą 10 GeV (Figs. 5-10).

Angular symmetries In terms of the Higgs signal, the three invariant mass simulations with different CP look indistinguishable save differences in distribution tails. The simplest way to show the differences between them is by plotting various angular distributions within the tensor structure of the system. Apparently, altering the CP parities of the Higgs boson will affect the distributions in the final state. Histograms were created for all angles defined in Fig. 10. The cosine of each angle was plotted

276

The Manhattan Scientist, Series B, Volume 4 (2017)

Reese

Figure 6. Mass distribution of Higgs, CP-odd

Figure 7. Mass distribution of ZZ pair, CP-odd

Figure 8. Mass distribution of Higgs, mixed parity

Figure 9. Mass distribution of ZZ pair, mixed parity

Figure 10. Definitions of observable structure along gg → H → ZZ → 4 lepton decay [4]

Reese

The Manhattan Scientist, Series B, Volume 4 (2017)

277

from -1 to +1. The distributions of cos θ1 and cos θ2 (Figs. 11-13) make for a good example. For the CP-even and CP-odd distributions the shape of the histograms were inverted relative to each other, with

Figure 11. cos θ1 and cos θ2 distributions in the Standard Model simulation

Figure 12. cos θ1 and cos θ2 distributions in the Beyond the Standard Model simulation

Figure 13. cos θ1 and cos θ2 distributions in the Mixed Model simulation

278

The Manhattan Scientist, Series B, Volume 4 (2017)

Reese

different amplitudes. In the Mixed Model the shape of the distribution appeared to correspond to that of the Standard Model simulation. Despite equal contributions from both parities of the Standard Model simulation. Despite equal contributions from both parities the CP-even signal seems to dominate over that of the odd. If a Higgs boson of such mixed parity could be produced experimentally it would be difficult to distinguish from the Standard Model Higgs which has already been seen by ATLAS.

Conclusion A major takeaway from this research is that the results of the LHC experiment can be replicated in a theoretical model. These simulations can be used as a comparison for any experimental results done under the same theoretical conditions. The existence of the Higgs particle and the validity of the Standard Model are once again confirmed. Most importantly, for the Manhattan College Physics Department, a functioning computational system is now in place for undergraduate research. Further experiments and simulations now can be done by students in order to provide them with an elementary understanding of the physics at play regardless of their level of mathematical understanding. Possibly even an ambitious high schooler could be introduced to and guided through a small scale project. As long as the student is willing to work hard there is an excellent opportunity to learn.

Acknowledgments This work was funded by the National Science Foundation Grant PHY-1402964. The author would like to thank Dr. Rostislav Konoplich and Dr. Josif Zhitnitskiy for giving him this opportunity and for their help throughout. He would further like to thank Danielle Rabadi for her collaboration over the summer.

References [1] The ATLAS Collaboration. “Observation of a new particle in the search for the standard model Higgs boson with the ATLAS detector at the LHC.” Phys. Letters B 716.1, 1-29 (2012) [2] The CMS Collaboration. “Observation of a new boson at a mass of 125 GeV with the CMS experiment at the LHC.” Phys. Letters B 716.1, 30-61 (2012) [3] The ATLAS Collaboration. “Evidence for the spin-0 nature of the Higgs boson using ATLAS data.” Phys. Letters B 726.1-3, 120-44 (2013) [4] N. Belyaev, R. Konoplich, L. Egholm Pedersen, and K. Prokofiev. “Angular asymmetries as a probe for anomalous contributions to HZZ vertex at the LHC.” Physical Review D 91, 115014 (2015)

E.(lectro) coli and the GO(x) lden nAnode∗ Brian Evans Department of Chemistry and Biochemistry, Manhattan College Farzana Begum, Amanda Lazkani, Syeda Rithu Department of Chemical Engineering, Manhattan College Dawud Abdur-Rashid Department of Biology, Manhattan College Gregory Sanossian, Samuel Coby Department of Mechanical Engineering, Manhattan College Abstract. We worked toward designing a biofuel cell that maximizes electron shuttling by utilizing variants of the Aspergillus niger glucose oxidase (GOx) that has been engineered for increased stability, activity, and affinity to gold nanowires. GOx possesses high specificity towards glucose, oxidizing it, and producing electrons. We envisioned a system that includes gold nanowires as the anode to increase the surface area for acceptance of electrons. Our enzyme contains a cysteine tag to provide thiol groups, known to have a high affinity for gold. The cysteine tags will sequester the enzymes to the gold wires, increasing electron transfer efficiency. Additionally, we are recombinantly expressing the MtrCAB operon from Shewanella oneidensis, an electric bacterium, in E. coli. The MtrCAB is responsible for the production of membrane bound cytochromes and known to generate bacterial nanowires. We anticipate that the MtrCAB system will allow for an electric E. coli that produces nanowire connections between the anode and bacterium for direct electron transfer.

Introduction Energy has become a necessity to sustain our society and to further its advancement. The depletion of fossil fuels and the need for clean electricity production has called attention to biofuel cells, which convert chemical energy into electrical energy via enzymatic reactions. This source of energy is sustainable, renewable, and does not emit CO2 . Conventional fuel cells are generally cost-ineffective in regard to energy production. In addition, once one of the active masses in a conventional fuel cell is fully consumed, the current-producing reaction ceases. As an alternative, glucose powered biofuel cells hold much promise as a viable clean energy resource. Glucose is energy dense, cost-efficient, and readily abundant. The redox enzymes used to power biofuel cells are renewable and less expensive compared to the precious metal catalysts used in conventional fuel cells. In addition, these enzymes are optimized in neutral pH buffers, making them an attractive candidate to power ultralow power consuming implantable medical devices. Glucose oxidase (GOx) is one of the most well studied enzymes for use in biofuel cells because they catalyze an oxidation reaction that generates electrons at the anode. GOx is a flavoprotein oxoreductase that oxidizes glucose by using oxygen as the electron acceptor to produce gluconoβ-lactone (glucanolactone) and hydrogen peroxide [1]. This enzyme is a homodimeric protein, ∗

Research mentored by Bryan Wilkins, Ph.D., and Alexander Santulli, Ph.D.

280

The Manhattan Scientist, Series B, Volume 4 (2017)

Evans et al.

with a flavin adenine dinucleotide (FAD) molecule noncovalently bound at the active site of each subunit. Glucose oxidase is a relatively large enzyme, with an average diameter of about 8 nm. This enzyme comes with both advantages and disadvantages. The enzyme has highly stable catalytic activity, most likely due to the fact that its redox center is insulated by a protein shell. The shell effectively buries the active site, FAD, in a deeply embedded protein matrix. As a result, glucose oxidase generally requires mediators to achieve successful electron transfer to an electrode because of the long electron tunneling distances and the steric constraints. Two such mediators that have been given the most attention are carbon nanotubes and gold nanoparticles because of their large active surface area and exceptional electrical properties [2]. Carbon nanotubes; however, are toxic to the human body. Gold nanoparticles (GNP) are not poisonous to the human body, and can aid long-term stability of GOx molecules. Interestingly, electrical bacterial cells exist in nature. One such organism is the bacterium Shewanella oneidensis [3]. This organism contains an mtrCAB operon and CymA gene that are responsible for electron transport across the cell membrane of the electric bacterium [4, 5]. The mtr operon and CymA gene express a system of cytochromes that work together to pass electrons to the extracellular matrix to reduce extracellular metals. Conceivably, instead of a metal, these cells could transfer electrons directly to an anode. CymA is an inner membrane tetraheme cytochrome c. MtrA is a periplasmic decaheme cytochrome c and MtrC is an extracellular decaheme cytochrome c. Based on previous literature expressing CymA in E. coli, in addition to the mtrCAB operon increases the reduction rate of extracellular acceptors [5]. We hypothesized that if the mtrCAB operon and CymA gene were stably expressed in E. coli that we could create electrical E. coli cells (“E. lectro coli”) that could directly supply electrons in a biological fuel cell. We aimed to create a more efficient bioanode through a three-tiered approach. First, we wanted to improve the activity and efficiency of the electron producing enzyme, GOx. We utilized mutant variants of the Aspergillus niger GOx that were reported to have increased stability and activity [6]. We then proceeded to genetically introduce a cysteine tag to the GOx variants as a way to sequester them to our gold anode (thiols have high affinity for gold). Second, we set to increase the anode surface area as a means of amplified affinity for GOx and decrease the distance that an electron must travel to reach the anode (direct transfer of electrons to the anode through enzymatic sequestering). Our anode was designed, not as a flat plate electrode, but as nanowires. We envisioned that the nanowire anode would produce a “nanograss” [7] like structure allowing for a direct correlation between the amount of enzyme in close proximity to the anode and the performance of the anode in the biofuel cell. GOx was hypothesized to essentially bury itself within the nanograss increasing its binding to the anode due to improved positioning of the electron export channel because the orientation of the enzyme was no longer limited at the nanowire anode (“nAnode”), as it would be at a flat surface. Lastly, we sought to create an electric E. coli strain that harbors the cytochrome electron transport system of Shewanella oneidensis. We intended to recombinately express the S. oneidensis

Evans et al.

The Manhattan Scientist, Series B, Volume 4 (2017)

281

cytochrome operon to create an E. coli strain that can directly transfer electrons to an anode, instead of metal ions. We believed that a combination of electric bacterium and localized GOx would increase the efficiency of a classical bioanode.

Project design Glucose Oxidase (GOx) Variant design: The sequence for A. niger GOx was derived from the NCBI Genebank number AAA32695 (E.C 1.1.3.4). This sequence was utilized by Holland et. al. [6] for the rational design of increased stability in GOx mutants. Each of our GOx genes originated from this wild type (Wt) sequence. Taking advantage of IDTâ&#x20AC;&#x2122;s (http://www.idtdna.com) generous support of iGEM team projects we ordered gBlock fragments for each of our gene segments for their direct cloning into the expression vector constructs. These were designed for cloning into the expression plasmid pCDF-duet (Novagen). GOx variant gene design (Fig. 1) The Wt GOx amino acid sequence was codon optimized for E. coli. To this sequence the coding region for an N-terminal 6x histidine tag followed by a TEV protease site was introduced. These coding regions were added to install an affinity tag for the purpose of protein isolation on a nickel-resin column that could then be released following purification. Next, a short C-terminal glycine linker was added followed by a SacI site just prior to the stop codon. We rationalized that this linker would allow for enzymatic addition of future protein fusions or tags that proved beneficial for specific assays. Each of the terminal sequences were also codon optimized for E. coli. Each GOx mutant was then designed starting from the Wt gBlock unit. The quadruple GOx mutant

Figure 1. Schematic representation of the GOx gBlock designs used for cloning in this work.

(GOx-4mut) was derived from research by Holland et al. [8] that was found to have increased stability, activity, and possess the ability to directly transfer electrons to a gold nanoparticle. The points of mutation in the original research were T56V, T132S, H447C, and C521V. Each of the mutants were designed into our gBlock by using the codon optimized sequence for Val, Ser, and Cys. The final gBlock sequence for the quadruple mutant matched that of the Wt gBlock, except for the point mutations.

282

The Manhattan Scientist, Series B, Volume 4 (2017)

Evans et al.

Two additional GOx enzyme derivatives of the GOx-Wt and the GOx-4mut gBlocks were designed. One coding region produced GOx-Wt with a C-terminal 3x cysteine tag (GOx-cys). This sequence was added as an affinity tag for our anode due to reports suggesting that gold nanoparticles are attracted to thiol groups [7, 9]. The last GOx derivative was GOx-4mut with the addition of a C-terminal 3x cysteine tag (GOx-4mut-cys). These two tagged mutants were designed in an effort to create a sequestering mechanism for our gold anode. As stated above, each construct was codon optimized for E. coli codons. Anode design To design the nAnode, we applied a template based synthesis of gold nanowires (Au-NW). The rationale behind using nanowires as opposed to a flat Au surface is two fold. First, the nanowires have an increased surface area in comparison to the flat anode. The second reason for using nanowires as opposed to a flat anode or nanoparticles is the directionality of the nanowires. The 1 dimensional (1D) nature of the nanowires will greatly enhance the removal of charge from the anode, which is suspected to enhance the efficiency of the overall device. To achieve our goals set forth in our design, we utilized a template based â&#x20AC;&#x153;U-Tubeâ&#x20AC;? (Fig. 2) synthetic method to create Au-NW.

Figure 2. Cartoon scheme of template based U-tube.

This method was selected for several reasons. First, by using a template we can physically control the size of the Au-NW, allowing us to tailor the design of the anode for optimal performance. Secondly, this reaction occurs at room temperature, in a relatively benign solvent of ethanol. Thirdly, we can easily remove the template and fully characterize the properties of the Au-NWs. Finally, by carefully controlling the reaction conditions, a free-standing array of AuNWs can be produced. This final reason is very exciting for our application because it has the potential to make nanograss. We reason that the vertical alignment within the nanograss sample will greatly enhance the removal of charge from the enzyme to the anode, whilst maintaining a large surface area.

Evans et al.

The Manhattan Scientist, Series B, Volume 4 (2017)

283

Experimental Cloning, transformation and verifications of GOx variants into pCDF-duet Each gblock was designed using the prefix EcoRI and suffix PstI sites for insertion into both the pCDF-duet vector. 100-200 ng of gBlock DNA was digest in a total volume of 10-15 µL, with digested DNA concentrations in the range of 5 - 10 ng/µL (final concentration). NEB buffer 2 was used at a 1× final concentration with 0.5 µL of EcoRI (NEB) and 0.5 µL of PstI (NEB) in each reaction. The reactions were incubated at 37 ◦ C for 1 hr and then the enzymes were heat inactivated at 80 ◦ C for 20 min. pCDF-duet was digested (∼1000 ng) and then the linearized backbone was separated via electrophoresis using a 0.8% agarose gel. The DNA was removed from the gel and purified using the Thermo Scientific GeneJET Gel Extraction Kit. Ligations were then performed to insert the gBlocks into the vector backbone. Approximately 25 - 50 ng of digested vector DNA was used for each ligation and a 3-fold molar amount of digested insert (gBlocks) was used. Negative controls contained water in place of the insert. T4 DNA ligase buffer was used at a 1× final concentration, with 0.5 ◦ C of T4 DNA ligase. Water was added to adjust the final volume to 10 µL. The reaction was incubated at 16 ◦ C for 30 min and then heat inactivated at 80 ◦ C for 20 min. Cellular transformations were performed using 2 µL of the ligation reaction into 50 µL DH5alpha cells and performed as follows. Transformation reactions were incubated on ice for 30 min, heat shocked at 42 ◦ C for 90 s and then placed back on ice for 2 min. 500 µL of SOC was added and then the cells were incubated at 37 ◦ C for 1 hr. Cells were then plated on LB agar medium plates that were supplemented with the appropriate antibiotic (pCDF - spectinomycin). Ligation verifications were performed by digest reaction of plasmid DNA that was isolated from cell colonies that grew in the presence of the appropriate antibiotics. The digestion protocol was identical to that stated above and then the DNA fragment products were then separated by electrophoresis on a 1% agarose gel. Digestions that produced 2 fragments (the size of vector and insert) verified that the ligation worked in those clones and the DNA was then sequenced. GOx expression, verification and purification The pCDF plasmid utilizes a T7lac promoter that is IPTG inducible. This vector requires an E. coli expression host (DE3) containing a chromosomal copy of the gene for T7 RNA polymerase. We transformed BL21(DE3) cells with DNA from successful ligation reactions. The transformation protocol was the same as detailed above. Cells carrying the appropriate vector were then grown in LB medium, at 37 ◦ C, supplemented with spectinomycin. From an overnight starter culture, we inoculated a 500 ml culture for protein expression and purification. Cells were allowed to grow to an OD600 = 0.8-1.0 and then IPTG was added to a final concentration of 1 mM. Cells were then allowed to express protein for ∼16 h. Cell aliquots were taken for protein production analysis at several time points. At 16 h the cells were collected by centrifugation and then resuspended in 10 mL phosphate buffer (PB, pH 7.0). Lysozyme (∼10 mg) was added to the cell solution, mixed

284

The Manhattan Scientist, Series B, Volume 4 (2017)

Evans et al.

well and then incubated on ice for 30 min. The lysis was then vortexed and clarified at high speed centrifugation (∼18K rpm, 20 min). The supernatant was collected and used as the load sample for a Ni2+ -NTA column. The GOx genes were designed with a 6x HIS-tag that was utilized as a tag for detecting the protein through Western blot analysis. The fractions that were taken at time points following IPTG induction were lysed by boiling 100 µL of cells (pelleted) in 100 µL SDS running buffer. The whole cell lysate was analyzed by SDS-PAGE electrophoresis on a 15% gel. Separated proteins were blotted to PVDF membrane in Towbin buffer (25 mM Tris, 192 mM glycine, 20% (v/v) methanol (pH 8.3)). The blotted proteins were decorated with anti-HIS primary antibody and then secondary antibody conjugated with HRP. We then visualized our proteins using ECL substrate. The GOx proteins were purified via affinity column chromatography. The GOx expression lysates were passed over a 500 µL bed volume of BioRad, ProfinityTM IMAC Resin, Ni-charged (#1560131), pre-equilibrated with PB buffer. The flow through was collected and the resin was washed with excess PB buffer. GOx was eluted using PB buffer supplemented with 250 µM imidazole. Elutions were performed as 5×1 mL fractions. Each fraction was tested for presence of protein as well as the activity of GOx. GOx functional verification GOx catalyzes the oxidation of D-glucose and produces hydrogen peroxide. Enzymatic acR Red Glucose/Glucose Oxidase Assay Kit tivity using the Invitrogen Molecular Probes Amplex (A22189) which allows for a one-step detection of glucose oxidase activity by coupling the proR Red duction of H2 O2 to the activity of horseradish peroxidase (HRP). H2 O2 reacts with Amplex (a colorless molecule), in the presence of HRP, to yield resorufin (red-fluorescent product). This colorimetric assay directly couples the activity of glucose oxidase to the production of resorufin. This kit was used according to the Amplex Red manufacturer’s protocol. To 50 µL of the assay reaction mixture (PB buffer, HRP, glucose, and Amplex Red) 50 µL of the elution fractions was added. The kit provided Wt GOx, which was used as the positive control (0.1 U/mL final working concentration). A negative control was used which was elution buffer only. Nanowire synthesis The nanowires were synthesized using polycarbonate filter membranes (templates) with track etched pores with a diameter of 200 nm. These templates were then sonicated in a bath sonicator in pure ethanol for approximately 10 s to fill the pores with ethanol. Following the sonication step, the template was then mounted into a U-Tube reactor to separate the two half-cell compartments (Fig. 3). A 0.005 M solution of HAuCl [4] was used to fill one side of the U-Tube apparatus. Simultaneously, a 0.005 M solution of NaBH4 was used to fill the other side. It was critical to fill both sides at the same time and at the same rate to ensure that the delicate template was not broken. Once filled, the U-Tube was left to react for a particular amount of time. The specific amount of time depended on several things, concentration of reactants, pore size, and temperature.

Evans et al.

The Manhattan Scientist, Series B, Volume 4 (2017)

285

Figure 3. Set up of the U-tube reaction for the Au-NW.

After allowing the reaction to occur, the solutions were carefully poured out of the U-Tube, the Au solution was saved for later use, while the NaBH4 solution was discarded. The U-tube apparatus was dismantled and the template carefully removed and dried with a Kim-wipe. To isolate the nanowires from the template for imaging in a Scanning Electron Microscope, first the excess gold was removed from the surface of the template by grinding with a grinding stone. The template was then placed into a centrifuge tube and dissolved using methylene chloride. The AuNWs were then removed by centrifugation at 4000 RPM for 7 minutes, and the supernate was discarded. This process was repeated 3 more times to ensure all of the polycarbonate had been removed. At this point, the AuNWs were ready for characterization with an SEM or to be used in electrochemical experiments.

Results Following each gBlock ligation, single clones were selected and grown in media supplemented with the appropriate antibiotic. The DNA from each clone was then isolated and test digested with the same enzymes used prior to ligation. DNA that yielded the correct insert and vector DNA fragments were then retained, sequenced, and stored for further use. These correct sized fragments verified that the ligations worked properly. We used EcoRI and PstI sites for the creation of our vectors and therefore we used the same enzymes to test the ligation products for the proper insertions. We first tested the ligations of GOx-Wt and GOx-4mut (Fig. 4). Following enzymatic digest and electrophoretic separation it was clear that the pCDF-GOx-4mut plasmid had been properly constructed. However, our results indicated that the clones for the Wt construct were false positives

286

The Manhattan Scientist, Series B, Volume 4 (2017)

Evans et al.

because digests of the DNA from these clones yielded only a single linearized band at the same size as the plasmid without insert.

Figure 4. gBlocks of GOx-Wt and GOx-4mut were cloned into the vector pCDF. Clones that conferred resistance to spectinomycin were cultured and then the DNA isolated and test digested with EcoRI and PstI. In our first cloning attempt we were only successful with the GOx-4mut and not the Wt. We sequenced pCDF-GOx-4mut clone 1 and used this vector for the remainder of our experiments.

We successfully cloned the Wt variant after retrying the ligation. As previously, The DNA was isolated and test digested with EcoRI and PstI. Our second attempt yielded positive clones (Fig. 5). We sequenced pCDF-GOx-Wt clone 11 and used this vector for the remainder of our experiments.

Figure 5. gBlock of GOx-Wt was successfully cloned into pCDF. DNA from positive clones were digested with EcoRI and PstI and then separated by electrophoresis. We sequenced pCDF-GOx-Wt clone 11 and used this vector for the remainder of our experiments.

We then cloned gBlock GOx-cys and GOx-4mut-cys into the pCDF vector. We were unable to successfully acquire clones from our GOx-cys ligations but we did manage to succeed in making pCDF-GOx-4mut-cys (Fig. 6). DNA from those ligations were test digested with EcoRI and PstI. Each of our pCDF-GOx constructs was transformed into BL21(DE3) cells for IPTG inducible expression. Cells were collected at several time points following induction to test for the presence of GOx protein. Whole cell lysates were separated by SDS-PAGE and then blotted. Our constructs express proteins that contain 6x HIS tags and we decorated our blots against that HIS tag (Fig. 7).

Evans et al.

The Manhattan Scientist, Series B, Volume 4 (2017)

287

Figure 6. gBlock of GOx-4mut-cys was successfully cloned into pCDF. DNA from positive clones were digested with EcoRI and PstI and then separated by electrophoresis. We sequenced clone 1 and used this vector for the remainder of our experiments.

Our data suggests that each of our test expressions produced full length GOx protein.

Figure 7. Western blot analysis of test expressions of GOx Wt, 4mut, and 4mut-cys. In the presence of IPTG there is protein present.

We could clearly demonstrate that GOx protein was produced in the presence of IPTG (Fig. 7). We then chose to express protein overnight in 500 ml of cells and then isolate GOx from those expressions. Using the BioRad Ni2+ -NTA system described in the experimental we isolated each of our GOx proteins 1 ml elutions. Each of those elutions were then tested for GOx activity using the Amplex Red activity kit (Fig. 8). Elutions that contained active GOx convert Amplex Red into a visible red dye in this colorimetric assay. It is clear that each of our GOx isolations yielded full-length and enzymatically active protein. However, our results did indicate that the rate of enzymatic activity for our isolations was much slower than that for the positive control.

288

The Manhattan Scientist, Series B, Volume 4 (2017)

Evans et al.

Figure 8. Amplex Red assay. GOx positive control is performed with GOx from the Amplex Red kit. The negative elution buffer row are elutions from an expression performed from an empty pCDF vector. The relative colorimetric changes are after 3 hours reaction time.

Discussion Unfortunately, by the time of the competition we were still not at our intended goal for this project. We were unable to test our enzymes on the anode because we are still working toward quantifying our GOx protein and using the Amplex Red kit to determine its actual kinetic activity. From our latest Au-NW synthesis we feel that our anode wires look as expected on the template (Fig. 9).

Figure 9. Picture of our Au-NW synthesis. The gold colored portion is where the reaction occurred within the pores. We are unsure at this point why the reaction did not occur evenly over the entire template, but the entire area should have a gold color. This was performed as indicated in the experimental section.

While we still need to dissolve our template, rescue the wires and scan them by microscopy, we strongly believe that our gold anode will soon be ready to test with our enzymes. We are in the process of designing a prototype for our fuel cell and are working to have that data as we continue

Evans et al.

The Manhattan Scientist, Series B, Volume 4 (2017)

289

this project. We fully expect that sequestering GOx modified enzymes to a nanowire anode will increase the efficiency of the oxidative half of a glucose powered fuel cell. This work is important because increasing the efficiency of a biofuel cell will allow for many advances in the field of medicine and technology. Biofuel cells can be used as portable power sources for miniaturize electronics as well as self-power implanted medical devices to improve health. For example, biofuel cells have promise to power continuous glucose monitors (CGMs). Current blood glucose monitoring is very cumbersome and relies heavily on an external power source. The promise of a self-powered glucose biosensor implant would tremendously help patients monitor blood glucose levels without using any invasive techniques. Another potentially important application for biofuel cells may be their ability to power cardiac pacemakers. Conventional pacemakers are powered by lithium batteries, which only last between five to eight years. The replacement of these devices requires open heart surgery. This highly invasive surgery has significant surgical costs and may pose a risk to the patient. Hypothetically, a glucose fuel cell could be implanted in vivo. The fuel cell would reside within the heart and would utilize the glucose present in the blood stream to power the pacemaker. As long as there is a continuous supply of glucose, the fuel cell would provide a theoretically limitless supply of electricity. Overall, glucose biofuel cells are a promising alternative power source that may be able to power ultra-low power devices and decrease our reliance on conventional batteries. Implemented correctly in medical devices, glucose biofuel cells may also be able to provide energy for long periods of time without the need for surgical replacements.

Acknowledgments This work was generously supported by the School of Science and the office of the Provost at Manhattan College. The authors express special thanks to theirs mentors, Drs. Bryan Wilkins and Alexander Santulli, for their continued guidance and support. Editorâ&#x20AC;&#x2122;s note: This research was performed for the 2017 iGEM competition (iGEM.org), an international competition in synthetic biology aimed at building a biological machine with a beneficial output. This work was awarded a bronze medal at the iGEM Jamboree in November 2017. The judges were impressed with the overall concept and lauded this teamâ&#x20AC;&#x2122;s efforts; they were especially impressed with the fact that they competed against teams that had more resources, more training and more time, yet still managed to perform at a competitive level.

References [1] Wohlfahrt, G., Witt, S., Hendle, J., Schomburg, D., Kalisz, H. M. , and Hecht, H.-J. 1.8 and Ë&#x161; resolution structures of the Penicillium amagasakiense and Aspergillus niger glucose 1.9 A oxidases as a basis for modelling substrate complexes. Acta Cryst. D 55, 969-977 (1999). doi:10.1107/S0907444999003431

290

The Manhattan Scientist, Series B, Volume 4 (2017)

Evans et al.

[2] Cosnier, S., Gross, A. J., Le Goff, A., and Holzinger, M. Recent advances on enzymatic glucose/oxygen and hydrogen/oxygen biofuel cells: Achievements and limitations. Journal of Power Sources 325, 252–263 (2016). [3] Barchinger, S. E. et al. Regulation of Gene Expression in Shewanella oneidensis MR-1 during Electron Acceptor Limitation and Bacterial Nanowire Formation. Appl. Environ. Microbiol. 82, 5428–5443 (2016). [4] Shi, L., Rosso, K. M., Clarke, T. A., Richardson, D.J., Zachara, J.M., and Fredrickson, J.K. 2012. Molecular underpinnings of Fe (III) oxide reduction by Shewanella oneidensis MR-1. Front. Microbiol. 3, 50 (2012). ncbi.nlm.nih.gov. doi:10.3389/fmicb.2012.00050 [5] Jensen, H. M., TerAvest, M. A., Kokish, M. G., and Ajo-Franklin, C. M. CymA and Exogenous Flavins Improve Extracellular Electron Transfer and Couple It to Cell Growth in Mtr-Expressing Escherichia coli. ACS Synth Biol 5, 679–688 (2016). [6] Holland, J. T., Harper, J. C., Dolan, P. L., Manginell, M. M, Arango, D. C., Rawlings, J. A., Apblett, C. A., and Brozik, S. M. Rational Redesign of Glucose Oxidase for Improved Catalytic Function and Stability. PLoS ONE 7, e37924–10 (2012). doi:10.1371/journal.pone. 0037924 [7] Koenigsmann, C., Santulli, A. C., Sutter, E. and Wong, S. S. Ambient Surfactantless Synthesis, Growth Mechanism, and Size-Dependent Electrocatalytic Behavior of High-Quality, Single Crystalline Palladium Nanowires. ACS Nano 5, 7471–7487 (2011). [8] Holland, J. T., Lau, C., Brozik, S., Atanassov, P., and Banta, S. Engineering of Glucose Oxidase for Direct Electron Transfer via Site-Specific Gold Nanoparticle Conjugation. J. Am. Chem. Soc. 133, 19262–19265 (2011). [9] Nuzzo, R.G., and Allara, D.L. Adsorption of bifunctional organic disulfides on gold surfaces. 1983, J. Am. Chem. Soc. 105, pp. 4481–4483.

4513 Manhattan College Parkway, Riverdale, NY 10471 manhattan.edu