Manhattan Scientist 2019 by CETheodosiou

The Manhattan Scientist Series B Volume 6 Fall 2019

A journal dedicated to the work of science students

The Manhattan Scientist Series B

Volume 6

Fall 2019

ISSN 2380-0372

Faculty Advisor and Editor

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 6 (2019)

A note from the dean Reaching the sixth volume of the new series of the Manhattan Scientist is a pivotal point of success and endurance. What characterizes this volume is the breadth of subjects and the balanced participation from all fields of study in the School of Science. Twenty-nine papers cover research in experimental, theoretical, and computational science performed by sophomores through first year graduate students. On the cover of this issue is an image from a paper by J. Piekarewicz and F. Fattoyev in Physics Today, the membership magazine of the American Institute of Physics with over 134,000 subscribers. Dr. Fattoyev is a member of our physics department faculty. This work has continued to receive critical financial support for our students from a variety of sources: 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 science, Jim Boyle ’61, Kenneth G. Mann ’63, a National Science Foundation research grant, and a National Institutes of Health research grant. Given the recent concerns about academic integrity and incidents of plagiarism among authors at all levels, it is important to state that the Manhattan Scientist papers are authored only by the students, cover the students’ work, and are considered, in spite of their quality and depth, as preliminary reports of work performed with their mentor faculty member(s). As such, it is expected that several of these papers will be expanded, adapted, and submitted for publication (with the faculty as co-authors) to peer reviewed journals in their respective field of research. This year the editorial process did not include student editors. The faculty mentors reviewed and edited their students’ work, as did fellow students within some research groups. The overall management of this project has been carried out by the Editor in Chief, who read and edited all the submitted papers, especially their language style, consistency of figure and table styles, as well as that of references. As in the previous volumes, the task of typesetting so many articles in LaTeX has been arduous, but well worth the effort to recognize and best present the researchers’ work.

Series B

The Manhattan Scientist

Volume 6 (2019)

I would like to express my deep appreciation to the students for their efforts and their persistence on the road of scientific discovery. We are all honored to showcase our studentsâ&#x20AC;&#x2122; and colleaguesâ&#x20AC;&#x2122; continuing contributions to the body of scientific knowledge. It is with great pleasure that we present the publication of Series B, Volume 6, of The Manhattan Scientist.

Constantine Theodosiou Dean of Science and Professor of Physics

ISSN 2380-0372

Series B

The Manhattan Scientist

Volume 6 (2019)

Table of Contents Biochemistry Biofilm production stimulated by plant shoots Khaitlyn Figueroa . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 Understanding histone DNA interactions via crosslinking and chromatin immunoprecipitation (ChIP) Kimberly Heller and Alon Brown . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 Role of tyrosine kinases in Bacillus subtilis biofilm formation Jalah Jarvis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 Enhancing enzymatics fuel cells Seth Serrano . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29 Genotype-phenotype characterization of biofilm produced by Bacillus subtilis wild isolates Ryan Torres . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35

Biology Xylem conductivities in leaf veins Maya Carvalho-Evans . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49 Molecular characterization of Giardia lamblia in oysters ( Crassostrea virginica) collected from two sites in NYC Fatimatou Diallo . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61 Bark formation for Cephalocereus columna-trajani, Neobuxbaumia macrocephala, and Neobuxbaumia mezcalaensis Phillip Dombrovskiy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 Mechanical stresses of tree branches: Primary and secondary stems Deidre Franks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83 Reiterative engineering properties within tree branches Hasan Hamid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87 Quantification of eccentric growth in stems of Purshia tridentata Zemima Kashroo . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91 Molecular analysis and prevalence of Giardia Iamblia in Geukensia demissa collected from two beaches in NYC Jailinne Lopez . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97 Mexican columnar cacti spines with regard to bark coverage Catherine Anne McDonough . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105 Predicting bark coverage on saguaro cacti (Carnegiea gigantea) Olivia Printy . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113

Chemistry The interaction of amphotericin B: An ab initio study Alon Brown . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123 Synthesizing aromatics using iridium catalyzed ketone alkylations and classical methods Sharron Fernandez and Jennifer LaPoff . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131 Synthesis of commercial fragrance compounds possessing an aromatic ring linked to a polar functional group by a 2 C spacer Jennifer LaPoff and Sharron Fernandez . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 141

Series B

The Manhattan Scientist

Volume 6 (2019)

Computer Science Reinforced learning approach to reordering sentences in extractive text summarization William Kulp . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 151 Computing direction of maximal skew in a dataset Michael Rozycki . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165 Cloud-based automated sound synthesis using machine learning Kyle Watkins . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 177

Environmental Science Air pollutants and childhood asthma in the Bronx Jovan Gonzalez . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 185 Determination of heavy metals in Tibbetts Brook Tatianna Peralta . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 193

Mathematics Cluster detection in a dataset using the BCM and Oja synaptic learning rules Michael Campiglia . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 199 Using laterally inhibiting neurons to detect clusters Sebastian Peña . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 211 Monte Carlo computer investigations of higher generation ideal dendrimers Brandon Thrope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 223 Optimizing data acquisition for deep learning in magnetic resonance imaging (MRI) Quinn Torres and Marcus Wong . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 233 Optimizing network architecture for deep learning in magnetic resonance imaging (MRI) Marcus Wong and Quinn Torres . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 241

Physics CP studies on VH produced Higgs boson decay processes and background estimates Sarah Reese . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 249 Studies of Higgs boson properties and search for new physics with ATLAS Joseph Thomas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 255

On the cover: Image of the explosive merging of two neutron stars. (Credit: NASA’s Goddard Space Flight Center/CI Lab.; taken from J. Piekarewicz and F. Fattoyev, “Neutron-rich matter in heaven and on Earth,” Physics Today 72, 7, 30 (2019); https://doi.org/10.1063/PT.3.4247)

Biofilm production stimulated by plant shoots Khaitlyn Figueroa∗ Department of Chemistry and Biochemistry, Manhattan College Abstract. Bacillus subtilis is a strain of bacteria that aids in the protection of plants. B. subtilis is a non-virulent, soildwelling bacteria that is able to form biofilms through a symbiotic relationship with plant roots. B. subtilis biofilms grow over roots, providing a barrier to protect the plant from exogenous factors such as pathogens, as well as triggering plant defense systems. We performed an experiment to test if plant shoots, which include the stems and leaves of the plant, induce biofilm formation like plant root samples do. Our results showed that shoots do stimulate biofilm formation. B. subtilis biofilms are regulated by five histidine kinases (KinA-E) that activate, through phosphorylation, a master regulator called Spo0A. The activation of Spo0A allows for the transcription of genes that make up the biofilm components. An experiment was done comparing a wild type strain expressing all of the kinases and mutant strains lacking one of the kinases. Our results showed that the mutant strain lacking KinC was unable to create a strong biofilm in the presence of plant samples, suggesting it is essential for biofilm formation. Further work includes looking into KinC and how it senses stimulatory molecules.

Introduction With a continuously growing population, the demand for food escalates and this drives the need for agricultural practices that reduce crop loss, such as from disease. One way to protect growing crops from harm, without the use of chemical pesticides, is by utilizing beneficial bacteria to promote plant health. These beneficial bacteria have been shown to work by creating biofilms on the roots of plants. Biofilms consist of a group of microorganisms that are held together via a sugary matrix, which can attach to a variety of surfaces. Some biofilms can be detrimental to organisms around it, but Bacillus subtilis is a strain of bacteria that aids in the protection of plants. B. subtilis is a non-virulent, soil-dwelling bacteria that is able to form biofilms through a symbiotic relationship with plant roots. B. subtilis biofilms grow over roots, providing a barrier to protect the plant from exogenous factors such as pathogens, as well as triggering plant defense systems (Vlamakis et al., 2013). B. subtilis biofilms are regulated by five histidine kinases (KinA-E) that activate, through phosphorylation, a master regulator called Spo0A. Spo0A is responsible for the transcription of the matrix components essential for biofilm formation. One kinase that has been shown to have a role in plant protection is KinD (Chen, 2012). In particular, KinD has an extracellular CACHE domain that is known to bind to small molecules, including pyruvate (Beauregard, 2013), an intermediate in several metabolic pathways. The mechanism of activation of initiation of the five histidine kinases and the biofilm pathway is still unclear and remains a big challenge in the field of B. subtilis biofilm formation. Some chemicals that are known to stimulate biofilms of B. subtilis are malic acid, which is secreted by plant roots (Rudrappa, 2008), and plant polysaccharides (specifically the plant’s ∗

Research mentored by Sarah Wacker, Ph.D.

The Manhattan Scientist, Series B, Volume 6 (2019)

Figueroa

cell wall), which can be used as a carbon source for the extracellular matrix of the biofilm (Wu, 2013). Previous research from Alexis Brown, another research student at Manhattan College, suggests that a variety of plant root samples, such as tomato and spinach, stimulate B. subtilis to produce biofilms. A question worth answering is whether or not plant shoots are able to stimulate biofilm production as well. The shoots of plants, which include leaves and stems, consists of many metabolites that are also located in the roots. Determining if the shoots stimulate biofilms just as well as the roots gives insight on whether the stimulatory molecule is activating the biofilm pathway as a result of general metabolism by B. subtilis or if it is from a specific molecule that has been generated due to an evolutionary relationship between B. subtilis and plant roots. Here we report our research findings that plant shoots are able to induce biofilm formation. Our work also characterizes the chemical nature of the shoot samples, as well as attempting to quantify transcription rates of the biofilm pathway via a luciferase assay. Finally, we examined bacterial strains lacking the histidine kinases (KinA-E) to determine if any of these kinases has a role in plant-stimulated biofilm formation.

Purpose The purpose of this research project was to determine if plant shoot extract stimulates biofilm formation like root extract. Other goals include the chemical characterization of plant samples as well as quantifying the transcriptional output of the biofilm pathway that results from these samples. I would also like to characterize the role of the five kinases, especially KinD, in sensing these plant samples.

Methods Collecting plant samples Ginger root, spinach leaves, parsley leaves, cilantro roots and leaves, and celery roots and leaves were purchased from a local farmer’s market. Each sample was washed with MilliQ water and processed in a juicer. Juice samples were centrifuged for 25 minutes at 4◦ C at 20,000 rpm in an SS34 rotor. To sterilize the samples, the juice extracts were filtered in steps through smaller pore samples (P8, 1, 40, 5, 0.45 µm) finishing with 0.22 micron filters. Treatments C-18 SepPak column. 1 mL of a juice sample was run through two connected SepPak C18 Plus short cartridges (Waters Corporation) using a syringe. This sample was labeled as “flowthrough.” Next, the column was washed with 1 mL of sterile water, labeled as the “water wash.” Samples were eluted with 1ml volumes of 20%, 30%, 40%, 50%, 60%, and 80% methanol. Seperation with 10kDa filter. 12 mL of the desired sample was added to a 10kDa Centrifugal Filter Unit. The 10kDa Centrifugal Filter Unit was spun at 4000 rpm for 15 minutes. Approximately 5 mL passed through the filter, which was labeled “less than 10 kDa.” The sample was diluted by

Figueroa

The Manhattan Scientist, Series B, Volume 6 (2019)

adding water to reach a volume of 12 mL. It was centrifuged at 4000 rpm for 5 minutes again. Steps were repeated until 12 mL had been collected for both greater and less than 10kDa. PD-10 column. The cap of the PD-10 column was removed and the column storage buffer was poured out. The column was equilibrated by filling it with equilibration buffer which entered the packed bed completely. This was repeated four times and all flowthrough was discarded. A maximum of 2.5 mL of the sample entered the packed bed completely and the flowthrough was discarded. A tube was placed under the column for collection of sample. The sample was eluted with 3.5 mL water and collected as elution. Heat denaturation. The sample was heated at 90◦ C for 30 minutes and centrifuged. The supernatant was transferred to a new tube. Proteinase K. 1ml of plant sample was combined with 10ul of Proteinase K (Thermo Scientific, 20 mg/mL), and incubated at 50◦ C for 1 hour. Experiments Streaking out bacterial cells on plates. Plates containing LB media were dried by a flame. Desired strain cells were struck from a glycerol stock onto plate and grown at 30◦ C for 24 hours. Pellicle assay. 3 mL of media (LB or MSn) was added to each well. 3 µL of B. subtilis cells were added to each well at a dilution of 1:1000 (Beauregard et al., 2013). Extract sample (150 µL) was added to its respective well to reach a final concentration of 5%. The cells were grown at 30◦ C for 48 hours and imaged. Luciferase assay. 200 µL of LB and cells were added to each well of a 96-well plate. This required the use of a plate reader for the luciferase assay and B. subtilis that harbored a PsdpA .lux reporter. This was placed in a plate reader and data was collected for analysis.

Data/Results The following samples were prepared and tested in biofilm pellicle assays in MSn media: tomato root extract, ginger root extract, spinach leaf extract, celery leaf extract, celery root extract, cilantro leaf extract, cilantro root extract, parsley leaf extract, parsley root extract. Examples of the strong biofilms produced by extracts of tomato root, cilantro leaf, and spinach leaf are shown in Table 1. The most stimulating samples were the spinach leaves and tomato root. As spinach leaves produced a strong biofilm and there was a lot of this sample created, this sample was chosen for further treatments. A C-18 SepPak column was used to separate hydrophilic molecules from hydrophobic molecules. The results show hydrophilic molecules including the flowthrough, water wash, and 10% methanol wash stimulate biofilm formation more than hydrophobic molecules such as the 40% and 80% methanol washes (Table 2).

The Manhattan Scientist, Series B, Volume 6 (2019)

Figueroa

Table 1. Shoots stimulate biofilm production. 5% extracts of cilantro leaves and spinach leaves stimulate biofilm formation at comparative levels as 5% tomato root extract. All assays were run in MSn media.

Control in biofilm assay

Cilantro leaves

Tomato roots

Spinach leaves

Table 2. Results of pellicle assay with samples from C-18 SepPak column.

Control

Flowthrough

10% Methanol wash

40% Methanol wash

Water wash

80% Methanol wash

Two different treatments were used to separate molecules based on size. A 10kDa filter was used to separate molecules that are larger than 10kDa from molecules that are smaller than 10kDa. Table 3 shows the molecules greater than 10 kDa stimulate more than molecules less than 10kDa. A PD-10 column separates high molecular weight molecules (> 5kDa) such as proteins from low molecular weight molecules (< 1kDa), such as salts. Table 4 shows the results that large molecules again stimulate more than small molecules like salts.

Figueroa

The Manhattan Scientist, Series B, Volume 6 (2019)

Table 3. Results of pellicle assay with samples from separation with 10kDa filter

Greater than 10 kDa

Control

Less than 10 kDa

Table 4. Results of pellicle assay with samples from a PD-10 column

Control

PD-10 1 fraction (proteins)

PD-10 2 fraction (salts)

PD-10 3 fraction

In order to investigate whether proteins are responsible for the stimulation by high molecular weight samples, we tested two methods to disrupt the protein portions of the plant extract. Proteinase K was used to break down proteins into amino acids. The results show stimulation (Table 5), suggesting either that proteins are not the stimulating molecules or that the resulting amino acids can trigger biofilm formation by B. subtilis. Samples were separately treated with heat, which denatures functioning proteins and other heat sensitive molecules. The results show that heating the sample does not affect stimulation of biofilms (Table 6). Table 5. Results of pellicle assay with plant extract treated with Proteinase K

Control

Proteinase K

The Manhattan Scientist, Series B, Volume 6 (2019)

Figueroa

Table 6. Results of pellicle assay with extract exposed to heat denaturation treatment

Heated

Control

No Heat control

Tables 2-6 show the results of the treatments using spinach juice extract and are considered representative for shoot stimulation. These results are consistent with other plant samples that underwent the same treatments. The chemical nature of the stimulating molecules are hydrophilic and large (> 5kDa). The most stimulating treatments are the flowthrough from the C18-SepPak column and treatment of the sample with Proteinase K. Not all samples were able to stimulate biofilm formation by B. subtilis and the results of the observed stimulation over our number of replicates are shown in Tables 7 and 8. Table 7. Most stimulating molecules that create strong biofilms Sample Tomato Root Juice Tomato Root Proteinase K Tomato Root Wash Spinach Leaf Juice Spinach Leaf Flowthrough Spinach Leaf Wash Spinach Leaf Greater than 10 kDa Spinach Leaf Proteinase K Parsley Leaf Juice Parsley Leaf Flowthrough Parsley Leaf Wash Parsley Leaf Proteinase K Cilantro Leaf Juice Cilantro Leaf Juice Flowthrough Cilantro Leaf Juice Wash Spinach Leaf Juice Heated Spinach Leaf Juice Control

Stimulation / Attempts

% Successful Stimulation

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

100 100 100 100 100 66 75 100 100 100 33 100 66 66 50 100 100

Luciferase assay The luciferase assay was meant to serve as a quantitative measurement for which molecules stimulated the biofilm pathway best. This can be measured by the production of light in a strain

Figueroa

The Manhattan Scientist, Series B, Volume 6 (2019)

Table 8. Least stimulating molecules that create weak biofilms Sample Spinach Leaf 10% Methanol Wash Spinach Leaf 20% Methanol Wash Spinach Leaf 40% Methanol Wash Spinach Leaf 80% Methanol Wash Ginger Root 10% Methanol Wash Ginger Root 40% Methanol Wash Ginger Root 60% Methanol Wash Parsley Leaf 10% Methanol Wash Parsley Leaf 40% Methanol Wash Parsley Leaf 60% Methanol Wash Spinach Leaf Less Than 10 kDa Ginger Root Less Than 10 kDa Ginger Root PD-10 2 Fraction Ginger Root PD-103 Fraction

Stimulation / Attempts

% Successful Stimulation

0/3 0/2 0/2 0/2 0/2 0/2 0/2 0/1 0/1 0/1 0/2 0/2 0/1 0/1

0 0 0 0 0 0 0 0 0 0 0 0 0 0

of bacteria that carries the luciferase gene under the control of a promoter that is activated during biofilm formation, such as PsdpA -lux. High luciferase readings should indicate high biofilm stimulation. The quantification of the transcription of the biofilm pathway via a luciferase reporter was not as reliable as thought to be; however, it did offer insight on which treatments stimulated better than others. It was also determined that later peaks in the luciferase assay may be more representative of the biofilm assay compared to earlier peaks. Luciferase assays were run with the samples on three different days. The luciferase assay from July 17 shows high rates of transcription for treatments like celery root, spinach juice and spinach flowthrough (Fig. 1). It also shows low rates of transcription for parsley juice and parsley

Figure 1. Results of luciferase assay with stimulating molecules on date 07/19/19.

flowthrough. The luciferase assay on July 23 shows high rates of transcription for treatments like proteinase K treated parsley, celery root and proteinase K treated ginger (Fig. 2). It also shows

The Manhattan Scientist, Series B, Volume 6 (2019)

Figueroa

low rates of transcription for parsley juice. The luciferase assay on July 31 shows high rates of transcription for proteinase K treated spinach juice (Fig. 3). These results indicate the luciferase assay was not very consistent from day to day, and was also not representative of the results of the pellicle assays. When luciferase assays were completed with different kinds of media and reporters (data not shown), the results were also inconsistent.

Figure 2. Results of luciferase assay with stimulating molecules on date 07/23/19.

Figure 3. Results of luciferase assay with stimulating molecules on date 07/31/19.

Mutant strains versus Wild Type Table 9 shows the results of comparing the effect of a mutant strain missing one of the five kinases and a wild type strain expressing all kinases. The results show that all mutant strains were able to form biofilms with plant samples except the strain lacking KinC. This suggests that KinC is important for biofilm formation that is stimulated by plant samples.

Figueroa

The Manhattan Scientist, Series B, Volume 6 (2019)

Table 9. Results of the biofilm assay comparing wild type B. subtilis and mutant strains that lack a kinase. All mutant strains were able to form a biofilm except the ∆KinC strain, suggesting it is essential in biofilm formation by plant samples. Control

Cilantro Leaf Juice

Spinach Juice

Wild Type

∆KinA

∆KinB

∆KinC

∆KinD

∆KinE

Discussion

Bacillus subtilis is a strain of bacteria that aids in the protection of plants. Previous research suggested that plant roots stimulate the biofilm pathway (Beauregard, 2013). Our results of the pellicle assay stimulated by plant shoots showed that the shoots do stimulate biofilm formation. This finding suggests the symbiotic relationship between B. subtilis and plants is a result of recognition of general molecules, such as those used for metabolism, by B. subtilis and not due to specific molecules that are only excreted by particular plants or plant tissues. This gives us more insight into how biofilms may be activated. Our attempt to categorize the stimulatory molecules showed

The Manhattan Scientist, Series B, Volume 6 (2019)

Figueroa

that large hydrophilic molecules stimulate the biofilm pathway best. Future work includes further characterization of these molecules. The luciferase assay was an attempt to visualize the transcription rates of the biofilm pathway. This was unsuccessful, as the results were inconclusive. The results varied day to day, and were not representative of the results of the biofilm assays. The luciferase assay was completed with various reporters and different kinds of media, but these experiments need to be more rigorously pursued in order to determine whether there is a right combination of media/reporter that best represents the results of the biofilm assay. B. subtilis biofilms are regulated by five histidine kinases (KinA-E) that are activated by a master regulator called Spo0A. A biofilm assay comparing a wild type strain that expressed all the kinases was compared to mutant strains of B. subtilis that lack at least one of the kinases. The results showed that a mutant strain lacking KinC was unable to form a biofilm under these conditions. This means that KinC is essential in biofilm formation. Future work involves studying KinC in more depth.

Acknowledgments This work was supported by the Manhattan College Jasper Summer Scholars Program. The author would like to extend gratitude towards Dr. Wacker, for giving her the opportunity to work with her.

References [1] Vlamakis, H., Chai, Y., Beauregard, P., Losick, R., and Kolter, R. (2013). “Sticking together: building a biofilm the Bacillus subtilis way.” Nat. Rev. Microbiol. 2013 Mar;11(3):157-68. doi: 10.1038/nrmicro2960 [2] Chen Y., Cao S., Chai Y., Clardy J., Kolter R., Guo J. H., Losick R. (2012) “A Bacillus subtilis sensor kinase involved in triggering biofilm formation on the roots of tomato plants.” Mol. Microbiol. 85(3):418-30. doi: 10.1111/j.1365-2958.2012.08109.x. [3] Rudrappa, T., Czymmek, K. J., Paré, P. W., and Bais, H. P. (2008). “Root-secreted malic acid recruits beneficial soil bacteria.” Plant physiology, 148(3), 1547-1556. doi:10.1104/pp.108. 127613 [4] Beauregard, P. B., Chai Y., Vlamakis H., Losick R., Kolter R. (2013). “Bacillus subtilis biofilm induction by plant polysaccharides.” Proc. Natl. Acad. Sci. USA.110(17):E1621-30. doi: 10.1073/pnas.1218984110. [5] Wu, R., Gu, M., Wilton, R., Babnigg, G., Kim, Y., Pokkuluri, P. R., Szurmant, H., Joachimiak, A., Schiffer, M. (2013). “Insight into the sporulation phosphorelay: crystal structure of the sensor domain of Bacillus subtilis histidine kinase, KinD.” Protein Sci. 2013 May;22(5):56476. doi: 10.1002/pro.2237. Epub 2013 Mar 18.

Understanding histone DNA interactions via crosslinking and chromatin immunoprecipitation (ChIP) Kimberly Carmen Heller∗ and Alon Brown∗ Department of Chemistry and Biochemistry, Manhattan College Abstract. Chromatin Immunoprecipitation (ChIP) is a technique that is used to help identify and understand interactions between protein and DNA. Using formaldehyde chemical crosslinking in living cells, protein is covalently trapped along the chromosome, and then the chromatin fiber is sheared into small DNA-protein fragments. Following the precipitation of a target protein, the DNA that was associated with the target is isolated and used for either qPCR or sequencing. The isolated DNA gives chromosomal occupancy to the protein target. Specifically, our research focused on developing a ChIP protocol that would allow for proper chromosomal occupancy of the H2A and H#histone protein harboring an unnatural amino acid. Model organism budding yeast cells, S. cerevisiae was used, along with unnatural amino acid UAA and an HA tag that was attached to the H2A and H3 histone, requiring optimization of a double cross-link. Once optimizing the double cross-link, which aided in the determination of nucleosomal occupancy of the H2A H3 histones, we then worked towards configuring a successful double immunoprecipitation step. Since optimizing the ChIP protocol, moving forward we hope to identify and use the primers necessary to complete a ChIP analysis via qPCR.

Introduction DNA is a fundamental component of all forms of life on earth, and it can be coiled and packaged into dense units of sequenced nucleic acids known as chromatin. In the organization of DNA into chromatin is a complex but vital process which keeps DNA packaged within the nucleus tightly, and it aids in the regulation of gene expression. Nucleosomes are the fundamental, basic polymeric unit of DNA that is responsible for the efficient compaction of DNA. Histone proteins are the octameric structure of which nucleosomes are comprised. Approximately one hundred forty-seven base pairs of DNA tightly coil and wrap around the histone proteins distributed along the chromatin, forming a nucleosomal unit [1]. These DNA-histone protein interactions, and successive nucleosomal structures through the genome, are responsible for the effective packaging of DNA in the nucleus. The storage of DNA in this manner is quite efficient at maintaining the reduced spatial occupancy of the macromolecule but creates an intriguing problem. When DNA is wrapped in a nucleosome, it is essentially inaccessibly by other enzymes that require the genetic material template for processes such as replication and transcription. The intimate relationship between histones and DNA suggest that histone proteins play an important role in regulating DNA accessibility. If DNA was not packaged in such an intricate way it would be nearly impossible to regulate gene transcription because the DNA would be unorganized and unidentifiable within the nucleus. Two sets of four core histones assemble to create an octameric unit, around which DNA winds. These histones are H3, H4, H2A and H2B. Histones H3 and H4 form dimers which then assemble ∗

Research mentored by Bryan Wilkins, Ph.D.

The Manhattan Scientist, Series B, Volume 6 (2019)

Heller + Brown

into a tetramer, and histones H2A and H2B form two distinct dimers. With the aid of chaperones and other non-nucleosomal proteins, a nucleosome is formed by first, the association of the H3-H4 tetramer [2]. DNA begins to associate with the tetramer and then two H2A-H2B dimers are brought in to finish the nucleosomal structure. Once the native chromatin structure is established it must then be regulated by, not only histones, but non-histone proteins that can bind to the nucleosome and help stabilize or destabilize its structure to promote or suppress access to the DNA. While many of these proteins have been identified our understanding of their mechanistic actions is limited. One very important aspect of chromatin biology is to determine what is called chromosomal occupancy. This refers to the identification of specific loci, along the genome, where proteins bind and influence accessibility to the DNA. Dependent on the nucleosomal positioning associated with the regulator’s interaction (i.e. at a promotor, enhancer or insulator region) informs about its potential action in chromosomal signaling. The specific loci (occupancy) of protein-DNA interactions can be defined using a technique call chromatin immunoprecipitation (ChIP) [3, 4]. Chromatin immunoprecipitation protocols involve formaldehyde crosslinking between protein and DNA to generate identifiable spatiotemporal interactions/relationships. The target protein of interest is isolated and co-precipitates its crosslinked DNA. ChIP protocols yield purified DNA fragments that have direct interactions with a protein of interest. These DNA fragments can be quantified via qPCR and/or sequenced to reveal genomic locations for the specific interactions. We envision a novel alteration to the classical ChIP technique that involves an additional chemical crosslinking step, prior to formaldehyde. Using an expanded genetic code, it is now possible to site-specifically encode unnatural amino acids into proteins [5]. We take advantage of the unnatural amino acid p-benzoylphenylalanine (pBPA), a photo-crosslinker that is activated with ∼365 nm light [6]. This amino acid forms chemical covalent bonds with neighboring proteins when activated. We encode pBPA into histone proteins and allow them to saturate the native chromatin landscape of yeast (S. cerevisiae) in order to study histone-protein interactions, in vivo [7]. Since histone proteins are highly efficient targets for ChIP, due to their direct interactions with DNA, we suggest that crosslinking with pBPA would enhance the efficiency of ChIP for proteins that bind to DNA more transiently. For instance, proteins that act at the nucleosome, but interact with histones, instead of the DNA itself, are more difficult to assess via ChIP techniques because they do no precipitate high levels of associated DNA fragments. However, if that protein binds to histones during its mode of action, then it could first be captured in a bridging crosslink, prior to ChIP. Once it is “stuck” to the histone, ChIP can be performed as usual, where the histone protein is precipitated, carrying with it the formaldehyde crosslinked DNA and the pBPA crosslinked protein. The precipitated pool would then be precipitated again against the captured protein, allowing for the DNA in the bridging interaction to be purified and sequenced (scheme outlined in Fig. 1). This double IP approach will clarify how proteins that bind the nucleosome, but not directly to the DNA are located along the chromatin fiber.

Heller + Brown

The Manhattan Scientist, Series B, Volume 6 (2019)

Figure 1. Schematic of double crosslinking and double IP for use of UV-activating unnatural amino acids during chromatin immunoprecipitation techniques.

Materials and Methods Histone expression and cell growth This research was performed in Saccharomyces cerevisiae (yeast) and used HA-tagged H2A and H3 histones that were expressed via plasmid. These procedures were followed as reported [7]. In brief, the coding sequences of the H2A and H3 histone, flanked by 450 base pairs of DNA up/downstream from the gene fragment, was cloned into a pRS426 vector. All wild type (Wt) cell cultures were grown in a minus uracil standard synthetic complete dropout medium, supplemented with 2% final glucose concentration. All pBPA-histone expressions were cultured in minus uracil minus leucine drop out media. The double selection maintained a second plasmid, pESC-pBPA [8], which harbored the genes of the pBPA-evolved tRNA aminoacyl synthetase and suppressor tRNA. These cultures were also supplemented with 2% final glucose concentration, plus 1 mM pBPA final concentration. All cultures were grown overnight at 30◦ C and 215 rpm. Approximately 5-10×109 cells was collected from 250 mL cultures for each ChIP assay. Once the cells reached the optimal OD600 they were then spun down at 2,000 rpm at 4◦ C for five minutes until a pellet was formed. Once the pellet was formed, the cells were washed with PBS (140 mM NaCl, 3 mM KCl, 10 mM Na2 HPO4 , 2 mM KH2 PO4 , pH 7.4). UV and formaldehyde crosslinking pBPA crosslinking was performed by resuspending the cell pellet in ∼2 mL PBS and exposing the cells to 365 nm light, at a distance of ∼10 cm, for 30 min, on ice. Following UV-crosslinking, the cells were recollected. At this point, the Wt and pBPA crosslinked cells were treated the same throughout the ChIP protocol. ChIP preparation was performed as previously reported [9]. The cellular pellet was resuspended in 9 mL of PBS and was then formaldehyde crosslinked via the addition of 1 mL of an 11% formaldehyde solution (0.1 M NaCl, 1 mM EDTA at a pH of 8.0, 0.5 mM EGTA at a pH of 8.0, 50 mM HEPES at a pH of 8.0, and 11% formaldehyde). Cells were left

The Manhattan Scientist, Series B, Volume 6 (2019)

Heller + Brown

to fix in the formaldehyde for 10 minutes at room temperature while sitting on a tabletop shaker at a constant speed of roughly 80 rpm. Next, the cells were quenched with glycine (125 mM) for five minutes. The cells were spun down again at 2,000 rpm for five minutes, the glycine was removed and then 25 mL of 1 M sorbitol solution was used to wash the cells. Cell lysis The cell pellet was resuspended in 14 mL of a freshly made zymolyase buffer (1 M sorbitol, 50 mM Tris-HCl, pH 7.4, 10.5 µL of -mercaptoethanol was added immediately before use). 7 mg of zymolyase 20T enzyme was added to an additional 1 mL of zymolyase buffer. The zymolyase buffer/enzyme solution was added to the resuspended pellet and incubated on a tabletop shaker for 45 minutes at room temperature. After the zymolyase enzyme solution was incubated and spheroplasts were formed, the entire solution was spun down in a centrifuge at 3,000 rpm for 5 minutes. Another 25 mL of 1 M sorbitol was used to wash the cells and centrifuged at 2,000 rpm for 5 minutes. The final pellet of spheroplasts were resuspended in 1 mL of NP-S buffer (5 mM sperimidine, 0.075% NP-40, 10 mM Tris-HCl at a pH of 7.4, 50 mM NaCl, 5 mM MgCl2 , 1 mM BME) with protease inhibitor (Thermo Fisher HALT protease). This solution was centrifuged at 4 ◦ C for ten minutes at 15,000 rpm. The pellet was isolated, after removing the supernatant, and resuspended in 600 µL of the NP-S buffer/protease inhibitor solution. It is imperative to note that the chromatin necessary to continue with the ChIP protocol is mostly in the cloudy solution of resuspended spheroplasts. Enzymatic chromatin fragmentation Micrococcal nuclease, also known as MNase, was the enzyme specified to shear the chromatin, randomly, until it reached the desired fragmented size of between 100-800 bp. We used New England Bio Labs MNase (M0247S). Following some optimization, we concluded that 25 units (1µg/µL) of MNase would be appropriate for successfully carrying out the shearing process. Chromatin was incubated at 37◦ C, after the addition of MNase, for 20-minute intervals for two hours. As a means to verify the MNase digest worked properly, a 1.5% agarose gel was prepared. Images were taken to ensure that the DNA was sheared to desired sizes. Once it was determined the chromatin was sheared to the correct size, the samples were diluted with elution buffer (10 mM Tris-HCl, 1mM EDTA, and 1% SDS, pH 8.0) to a final volume of 300 µL. Approximately 10 % (30 µL) of the chromatin was saved for a future input sample of the IP step of the ChIP protocol. Preparation of antibody beads and chromatin immunoprecipitation Performing the double immunoprecipitation was not possible without optimizing the first IP of the HA tag first. The double IP is not discussed as part of the methods because we have not successfully conducted this part of the ChIP protocol; it is addressed in the future research section. Rather than using magnetic beads preconjugated with HA antibodies, we used Thermo Fisher ChIP Grade Pierce Protein A/G Magnetic Beads (26162) and manually bound them to an HA

Heller + Brown

The Manhattan Scientist, Series B, Volume 6 (2019)

antibody (Abcam rabbit polyclonal HA antibody, 9110). For each IP reaction, approximately 30 µL of magnetic beads were resuspended and placed on a magnetic stand to separate the magnetic beads from the solution they were stored in. Using 1 mL of a BSA/PBS buffer (0.5% BSA), the magnetic beads were washed three times. After the final wash, the beads were resuspended in 250 µL of the BSA/PBS and then 3 µL of the antibody was added to the BSA/PBS and magnetic bead solution. This complex was incubated at 4◦ C for 2 hours, while gently mixing on a rotating platform shaker. The beads were additionally washed three more times with 1 mL of BSA/PBS and separated by a magnet. Once more, they were resuspended in 100 µL BSA/PBS until the MNase and IP samples were prepared and ready for precipitation. The beads are placed on a magnetic stand, separated from the buffer solution, and the MNase treatments prepared for the IP were added to the freshly antibody bound magnetic beads. The MNase IP samples were left mixing on a rotating platform shaker overnight (4 ◦ C) to ensure that the tagged chromatin was properly bound to the antibodies on the magnetic beads. After incubation, the beads were precipitated out via a magnetic stand, the supernatant from this action was saved as the flow thru. The beads were washed a total of seven times with 250 µL of each wash solution, as follows: 3x with low salt buffer (0.1% Triton x-100, 2 mM EDTA, 0.1% SDS, 150 mM NaCl, 20 mM HEPES - pH 8.0), 1x with high salt buffer (0.1% Triton x-100, 2 mM EDTA, 500 mM NaCl, 0.1% SDS, 20 mM HEPES - pH 8.0), 1x with LiCl buffer (0.5 M LiCl, 1% NP-40, 1% Na-deoxycholate, 100 mM Tris-HCl - pH 7.5), and 2x with TE. In order to track the protein throughout the IP, the flow through and washes were all saved for a future western blot. Following the final TE wash, the beads were precipitated out and centrifuged at 1,000 rpm for 3 minutes. Note, in order to avoid deforming of the beads it is imperative to spin them at this low speed. Residual TE was pipetted out after centrifugation. Once the beads were isolated, 250 µL of elution buffer (10 mM Tris HCl, 1 mM EDTA, 1% SD, pH 8.0) was added. The beads were eluted at 65 ◦ C, shaking at 1,000 rpm, for 30 minutes on a thermal mixer. Upon conclusion of the thermal mixing, beads were spun down at 20,000 rpm for one minute (4◦ C); the elution buffer supernatant was extracted and kept in a separate tube and a small portion of the elution was also saved for a future western blot. DNA purification and DNA analysis Remaining IP samples were reverse crosslinked overnight at 65◦ C. The same was done to the 10% input samples after diluting them with elution buffer to a total volume of 250 µL, and from this point on was continuously treated as one of the IP samples. One volume of TE was added to each IP sample/input. RNase enzyme was added to a final concentration of 0.2 µg/µL and incubated for two hours (37◦ C) to ensure RNA activity is halted when conducting the purification. Proteinase K, at a final concentration of 0.2 µg/µL was added and incubated at 55◦ C for an additional two hours. The addition of proteinase K is meant to prevent the continuous activity of certain proteins that might degrade desired proteins for purification. To purify the chromatin, we followed Thermo Fisher Scientific’s protocol for phenol chlo-

The Manhattan Scientist, Series B, Volume 6 (2019)

Heller + Brown

roform and ethanol precipitation. One volume of (25:24:1) phenol:chloroform:isoamyl alcohol was added to each IP sample and vortexed for roughly 20 seconds, until the solution is white and cloudy. All samples were sequentially centrifuged for five minutes, at 16,000 rpm, and the top aqueous layers were carefully extracted and placed into a separate tube. Glycogen, NH4 OAc (7.5 M), and 100% ethanol were added in these respective quantities: 1 µL, half the total volume of the sample, and 2.5× the total volume of the solution. DNA was precipitated out via overnight freezing at 20◦ C. The DNA fragments were centrifuged at maximum speed for thirty minutes (4◦ C), forming a small pellet of the remaining purified DNA. Next, 150 µL of 70 % ethanol was added to the pellet of DNA. Again, the sample was centrifuged at 15,000 rpm (4◦ C) for 2 minutes. Any ethanol supernatant was removed and discarded, and the DNA pellet was centrifuged at 15,000 rpm (4◦ C) for an additional thirty minutes. The samples were placed in an incubator at 37◦ C, until all remaining ethanol was dried from the pellet. TEN buffer (40 mM Tris-HCl, 1 mM EDTA, 150 mM NaCl, pH 7.5) was used to resuspend the DNA in each sample, and was then centrifuged once more (two minutes at 15,000 rpm, 4◦ C) for isolate the purified DNA.

Results and Discussion Growing the H2A and H3 cells with the HA tag was straight forward and did not require any optimization. In regards to the formaldehyde crosslinking step, no troubleshooting was involved and for the entirety of the research experience. Letting the 11% formaldehyde crosslink for 10 minutes was sufficient enough because the HA protein interacts directly with the DNA. Yeast cells contain very ridged cell walls. They are difficult to lyse and efficient lysis requires enzymatic assistance with degrading the glycoprotein cellular coat. This process results in the removal of most of the yeast cellular membrane structure, producing a spheroplast. Yeast spheroplasts are viable when produced in a high osmotic buffer, such as 1 M sorbitol. We utilize the enzyme zymolase for this function and tested for its efficiency in the presence of SDS. Healthy yeast cells will not lyse in presence of 1% SDS, yet a yeast spheroplast will. Spheroplasting of the cells can be visualized under 400× magnification with a microscope. With the addition of the detergent to spheroplasated versus normal yeast cells there was visible cell lysis when visualized by microscopy (Fig. 2). It is clear that yeast cells in SDS (right-top) are still visualized however spheroplasts (bottom-right) are no longer apparent. These results indicate that we achieved efficient spheroplasting. Lysing the cells was achieved in NP-S buffer. This was also checked for efficiency via microscopy and found to be fully lysed under our conditions (data not shown). In efforts to optimize the chromatin shearing step of the protocol, we first examined different time intervals in a timed assay. H2A and H3 Wt cells were subjected to formaldehyde crosslinking, spheroplasting and cell lyse as described. The isolated chromatin fraction was then treated with MNase in a time trial assay to determine the proper time needed for the most efficient sheared products. Downstream analysis of ChIP DNA suggests that fragmentation should result in DNA sizes between 100-800 bp. Each MNase assay was performed on 25 µg of chromatin, using 3 U of MNase per reaction. The reactions were stopped in excess EDTA at the indicated time points.

Heller + Brown

The Manhattan Scientist, Series B, Volume 6 (2019)

Figure 2. Images of yeast cells and yeast spheroplasts in the presence of SDS. Images taken at 400Ă&#x2014; magnification by confocal microscopy.

Samples of each assay were then analyzed via 1.5% agarose gel (Fig. 3). We achieved efficient shearing of DNA for both histone H2A and histone H3. This is highlighted by the dashed boxes in Fig. 3, indicating the lanes and approximate sizes of DNA observed in the sheared sample. While

Figure 3. Electrophoretic analysis of chromatin shearing from the expression of wild type histones on 1.5% agarose gels. Each lane consists of approximately 1000 ng of fragmented chromosomal DNA from MNase digested at the indicated time point. The boxed lanes represent effective shearing in the range of 100-600 bp.

The Manhattan Scientist, Series B, Volume 6 (2019)

Heller + Brown

the shearing efficiency was not achieved in the same time frame, this is not considered a problem moving forward. The amount of DNA that is recovered from each timed assay is sufficient enough for several IP reactions, therefore, testing sheared DNA efficiency is recommended for each new ChIP experiment. Next, we set to precipitate the target histone-DNA conjugates in an immunoprecipitation reaction. Our histone proteins were genetically modified to carry an HA-epitope and we used ChIP grade HA antibodies against this tag. Following the conjugation of the antibody to magnetic protein A/G beads, we allowed the precipitation reaction to incubate at 4â&#x2014;Ś C, overnight. The beads were then washed excessively in varying salt conditions and then eluted for purification. We checked the progression of the IP over each step to follow the protein and ensure that it moved to the correct fraction. The progression was monitored and visualized via western blotting using HA antibody detection (Fig. 4). In each IP, for both H2A and H3, we observed that the beads were sufficiently saturated with protein, evidenced by a small percentage of target protein in the flow through as compared to the load (lane F and L, Fig. 4, respectively). There appeared to be little, to no, protein

Figure 4. Western blot analysis of HA immunoprecipitations of fragmented chromatin samples for both H2A and H3. The protein was monitored across the procedure and the individual steps are labled: L, load; F, flow through; W1, low salt wash; W2, high salt wash; W3 LiCl wash; W4, TE wash; E, elution. Arrows point to protein of interest in the load and elution fractions

lost during the washes (lanes W1-4), and the bulk of the histones appeared in the elution fractions as expected (lane E). Interestingly, there appears to be purified protein that co-migrates with the antibody fragments themselves, seen as the upper band in the western blot elution of H3. In this IP there is a fraction that migrates at the proper size of a histone (â&#x2C6;ź15 kDa) but a second band is â&#x2C6;ź70 kDa (antibody heavy chain plus histone). It is not unusual for proteins to co-migrate especially given the high affinity of the antibody to the epitope. Although, this was unexpected due to the use of denaturing conditions, we conclude that we have successfully precipitated histone protein because the intensity of the signal for load matched that of the elution for Histone H2A and the

Heller + Brown

The Manhattan Scientist, Series B, Volume 6 (2019)

two elution signals are about half of the total input, suggesting the protein is half migrating with antibody. Conveniently, the fact that they comigrate with the antibody is not of concern because subsequent steps require the proteolytic digestion of the proteins in the sample. We then followed the same protocol as outlined above but expressed histone containing pBPA at either positions H3 T6, H3 S22 or H2A A61. These two sites are of importance because they are sites of histone-protein interactions with chromatin associated proteins (publication in preparation from this lab). We aimed to perform ChIP-qPCR analysis of different chromatin loci to establish that pBPA-histones distribute into the chromatin landscape similar to that of Wt. It is important to provide definitive proof that nucleosomal occupancy is not affect by the mutation in the histone proteins. To do this, we performed ChIP in cell lines expressing the above named mutant histones, but we did not UV-crosslink. We set out to establish the mutant histones do not influence huge changes in positioning before we begin the process of the double IP. Following formaldehyde crosslinking, cell lysis and chromatin fragmentation we obtained sampled of chromatin containing unnatural histones. These samples were all verified by electrophoresis (both agarose and SDSPAGE western blotting â&#x20AC;&#x201C; data not shown) to conclude that we successfully obtained fragmented fractions of both Wt and mutant chromatin fragments in the correct size range. In the final steps of the ChIP assay we isolated the DNA crosslinked fragments from our proteins via proteinase K digestion and DNA precipitation. This yielded pure protein but in very low quantity, however with sufficient product to perform qPCR. We designed primers to test four separate loci along the yeast chromosome to quantify the DNA precipitated from crosslinked histone. We expect that these occupancies will be unchanged as compared to Wt histones. These analyses are currently being examined and our first qPCR results suggest that histone occupancy is identified at each locus, as expected. However, these results are very preliminary, and we cannot fully report on them here. We are working toward analyzing replicates of our ChIP-qPCR data to verify our initial results and then we will begin optimization of the UV-crosslinking, double IP procedure. This new protocol will provide a unique biochemical approach to studying chromatin protein interactions in the living nucleus revealing spatiotemporal understanding of their dynamics.

Acknowledgments The authors would like to extend gratitude towards Dr. Bryan Wilkins, for giving them the opportunity to work with him. The work of Kimberly Heller was supported by the Michael J. â&#x20AC;&#x2122;58 and Aimee Rusinko Kakos endowed chair in science. The research of Alon Brown was supported by an NIH R15 grant to Dr. Wilkins.

References [1] Luger, K., MaĚ&#x2C6;der, A. W., Richmond, R. K., Sargent, D. F. and Richmond, T. J. Crystal structure of the nucleosome core particle at 2.8 resolution. Nature 389, 251-260 (1997). doi: 10.1038/38444

The Manhattan Scientist, Series B, Volume 6 (2019)

Heller + Brown

[2] Winkler, D. D. and Luger, K. The Histone Chaperone FACT: Structural Insights and Mechanisms for Nucleosome Reorganization. Journal of Biological Chemistry 286, 18369-18374 (2011). doi: 10.1074/jbc.R110.180778 [3] Lieleg, C., Krietenstein, N., Walker, M. and Korber, P. Nucleosome positioning in yeasts: methods, maps, and mechanisms. Chromosoma 124, 131â&#x20AC;&#x201C;151 (2014). doi: 10.1007/s00412014-0501-x [4] Milne, T. A., Zhao, K. and Hess, J. L. Chromatin immunoprecipitation (ChIP) for analysis of histone modifications and chromatin-associated proteins. Methods Mol. Biol. 538, 409-423 (2009). doi: 10.1007/978-1-59745-418-6 21 [5] Xie, J. and Schultz, P. G. An expanding genetic code. Methods. 36, 227-238 (2005). doi: 10.1016/j.ymeth.2005.04.010 [6] Chin, J. W., Martin, A. B., King, D. S., Wang, L. and Schultz, P. G. Addition of a photocrosslinking amino acid to the genetic code of Escherichia coli. Proceedings of the National Academy of Sciences 99, 11020-11024 (2002). doi: 10.1073/pnas.172226299 [7] Wilkins B.J., Rall N.A., Ostwal Y., Kruitwagen T., Hiragami-Hamada K., Winkler M., Barral Y., Fischle W., Neumann H. A cascade of histone modifications induces chromatin condensation in mitosis. Science 343, 77-80 (2014). doi: 10.1126/science.1244508. [8] Chin J.W., Cropp T.A., Anderson J.C., Mukherji M., Zhang Z., Schultz P.G. An expanded eukaryotic genetic code. Science 301, 964-967 (2003). doi: 10.1126/science.1084772 [9] Lee, J. B. and Keung, A. J. Chromatin Immunoprecipitation in Human and Yeast Cells. Methods Mol. Biol. 1767, 257-269 (2018).

Role of tyrosine kinases in Bacillus subtilis biofilm formation Jalah Jarvisâ&#x2C6;&#x2014; Department of Biology, Manhattan College Abstract. The Gram-positive bacterium Bacillus subtilis is a model organism for studying the way biofilms form. This means research with B. subtilis has useful applications to biofilms formed on medical devices. A better understanding of how biofilms form can lead to developing ways to disassemble biofilms and get them out of medical devices. Biofilm formation in B. subtilis has been studied extensively and tyrosine kinases have a specific role in their formation. The tyrosine kinase pair EpsA and EpsB have been well characterized in previous research. EpsB has been found to autophosphorylate in the absence of the biofilm matrix. The tyrosine kinase pair TkmA and PtkA is not as well characterized, however. In this research the main goal was to characterize the role of TkmA and PtkA in B. subtilis biofilm formation.

Introduction Bacteria are able to adhere to surfaces by creating these multicellular communities called biofilms. In a biofilm, bacteria are surrounded by a self-produced matrix of proteins and carbohydrates. In nature, biofilms are traditionally composed of multiple different species of bacteria [1]. In the case of this research, the biofilms studied were of Bacillus subtilis alone because it is a good model organism to study the molecular mechanisms of how biofilms form [1]. The Gram-positive bacteria B. subtilis is also useful because it has a well-characterized biofilm pathway. In the traditional biofilm pathway in B. subtilis KinA-KinD activate Spo0A which results in the inhibition of SinR and the de-repression of matrix operons, epsA-O and tapA-sipW-tasA (Fig. 1). This pathway which controls production of matrix components is well characterized, but its regulation is still not fully elucidated. The research presented here examines whether biofilms are also regulated by tyrosine kinase proteins. In B. subtilis, there are two tyrosine kinases, each consisting of a kinase protein (EpsB or PtkA) and modulator (EpsA or TkmA). In tyrosine kinases, the modulator protein is required for kinase activity and contains an extracelluar domain. Previous research found that deletions of EpsA, EpsB, TkmA and PtkA all result in decreased biofilm formation [2]. Additionally, it was found that deletion of SinR, which typically results in the increased formation of biofilm matrix and a strong biofilm phenotype, also increase biofilm severity in tyrosine mutants by a small factor [2]. These results suggested the possibility that tyrosine kinases are acting within the canonical biofilm pathway, upstream of SinR. They also yielded the possibility of cross-talk between the two tyrosine kinase complexes [2]. EpsA and EpsB are the first two genes of the fifteen gene eps operon, which is under the control of Spo0A and is generally responsible for synthesizing the exopolysaccharide component, EPS, of the B. subtilis biofilm matrix. EpsA is capable of sensing EPS in the matrix â&#x2C6;&#x2014;

Research mentored by Sarah Wacker, Ph.D.

The Manhattan Scientist, Series B, Volume 6 (2019)

Jarvis

Figure 1. Simplified schematic of the canonical biofilm pathway of B. subtilis. Histidine kinases (KinA-KinD) feed phosphate groups into a phosphorelay which results in activation of the master regulator Spo0A. Activated Spo0A causes expression of transcription of the biofilm matrix operons, epsA-O and tapA-sipW-tasA, through inhibition of the transcriptional repressor SinR.

and responds to the EPS by activating the tyrosine kinase EpsB. EpsB, whose activity is regulated by autophosphorylation in addition to EpsA activity, then phosphorylates the glycosyltransferase EpsE [3]. Thus EpsA/EpsB are required for biofilm formation and regulated by an environmental signal (EPS), and act downstream of the canonical Spo0A pathway. The kinase pair TkmA/PtkA also has a role in biofilm formation, as indicated by deletion mutants of both genes [4, 5]. While TmkA/PtkA have several known targets during normal cell physiology [6], the specific function and targets of this BY-kinase in biofilm formation are unclear (Fig. 2).

Purpose The goal of this research was to determine if the tyrosine kinase proteins PtkA and EpsB have functions upstream of the traditional biofilm pathway. This would clarify what potential biofilm-relevant targets PtkA might have. As EpsB has been shown to be regulated by autophosphorylation during biofilm formation, I was also interested in whether PtkA is similarly regulated by autophosphorylation during biofilm formation. Finally, I wanted to purify PtkA for in vitro studies.

Figure 2. Image of B. subtilis kinase protein (EpsB or PtkA) and modulator (EpsA or TkmA). Image adapted from [7]

Jarvis

The Manhattan Scientist, Series B, Volume 6 (2019)

Materials and Methods Luciferase assay Markerless, gene deletion mutants of tkmA, ptkA, and tkmA-ptkA together were introduced into B. subtilis 3610 carrying luciferase reporters through SPP1 phage transduction [8]. Previously, luciferase reporters were created for the promoters PsdpA -lux and Peps -lux with deletions in epsA and epsB [2]. For the luciferase assay, bacteria were grown, shaking, in LB at 37◦ C for 3 hours to reach mid-log phase. Cells were diluted 1:100 in LB and plated in a 96-well clear bottom plate. Luminescence and absorbance (at 600 nm) readings were taken every 10 minutes for 14 hours in a FilterMax F5 multi-mode plate reader (Molecular Devices) at 37◦ C, with shaking in between readings. During data analysis, luminescence values were normalized to cell growth (absorbance at 600 nm) for each data point and readings from the same sample on the same day were averaged. Protein purification for phosphorylation assay B. subtilis containing either FLAG-His-tagged PtkA or FLAG-His-tagged-EpsB under an IPTG-inducible promoter were streaked out on LB plates and left at room temperature for two days. From these plates, 5ml LB cultures of each strain were grown at 37◦ C overnight. 4 mL of each culture were transferred to 100ml of LB media, shaking at 37◦ C until OD600 is 0.5. Then IPTG (100 µL of 1M) was added to all cultures, and the cultures continued to shake at 37◦ C for 2 hours. Each sample was transferred to two 50ml tubes and centrifuged at 4000 rpm for 15 minutes to collect the cells. Each sample was then resuspended in 2.5 mL of His-lysis buffer (50 mM Tris, pH8, 250 mM NaCl, 10 mM Imidazole, 10 mM Na-fluoride, 1mM PMSF, and 1 mM Na-orthovanadate). The samples were lysed using sonication (six times 45 seconds on ice with a one-minute break in between). Samples were spun at 15000 rpm for 30 minutes at 4◦ C, to separate out membranes. The membrane pellet was resuspended in 750 µl of His lysis buffer with 2% tween-20 and triton-X and rotated at 4◦ C overnight. The resuspended membrane samples were spun at 14000 rpm for 20 minutes at 4◦ C. His dynabeads (75 µL/sample) were washed with His lysis buffer and combined with the supernatant from before and after membrane prep with enough lysis buffer for a final volume of 13ml. Samples were incubated at 4◦ C for 2½ hours while rotating, before the resin was collected on a magnet and washed with His lysis buffer three times. Proteins were eluted by adding 2× 65 µL of elution buffer (37 mM Tris, pH8, 185 mM NaCl, 250 mM Imidazole, 7.5 mM Na-fluoride, 0.75 mM PMSF, and 0.75 mM Na-orthovanadate). SDS-PAGE Strains that had tagged PtkA or tagged EpsB were purified and analyzed on a 12% polyacrylamide gel. The samples were combined with 5× loading dye and then denatured at 97◦ C for 10 min. 5 µL of the ladder and 20 µL of each sample were loaded into the gel. The gel was run at 120V in Tris-Glycine SDS buffer for 15 minutes then moved to 180V until the dye ran to the bottom of the gel.

The Manhattan Scientist, Series B, Volume 6 (2019)

Jarvis

Western blot Western blot analysis was conducted using Anti-FLAG and Anti-phospho-tyrosine antibodies. After SDS-PAGE the gel was transferred to PVDF membrane at 250 mA for 1 hour in Tris-glycine transfer buffer. After the western blot transfer, the membrane with the protein was blocked in TBST (TBS with 0.1% Tween-20) with 5% BSA for 1hr at room temperature. The TBS-T with 5% BSA solution was exchanged for TBS-T with 2.5% BSA solution and 1:5000 Anti-FLAG antibody or 1:1250 Anti-phospho-tyrosine antibody and incubated at 4◦ C overnight. The membranes were washed 3 times in TBS-T before incubation in secondary antibody (1:20,000 Anti-rabbit for AntiFLAG or Anti-mouse for Anti-phospho-tyrosine) for one hour at room temperature. The membranes were washed in TBS-T again 3 times before reacting with chemiluminescence reagents. The membranes were visualized on BioRad imaging board using chemiluminescence setting. Coomassie stain The method for the Coomassie stain was the same as the SDS-PAGE. After the gel was run it was placed in Coomassie stain overnight at room temperature. The gel was washed in de-staining dye for 2 hours and then imaged.

Results Luciferase assay In order to investigate whether EpsB and PtkA, as well as their associated modulator proteins, have an effect on the traditional biofilm transcriptional pathway (Fig. 1), I conducted a luciferase reporter assay. I examined two reporters, Peps -lux and PsdpA -lux, in wildtype bacteria and strains missing EpsA, EpsB, TkmA, PtkA, or both TkmA and PtkA. The luciferase assay shows that ∆epsB and ∆ptkA have minimal effects on the biofilm transcriptional pathway. In Figs. 3 and 4 the wildtype in all the graphs is about the same height as the tyrosine kinase complex strains. If the wildtype had a dramatic difference in height compared with the mutant strains this would have indicated that ∆epsB and ∆ptkA have a significant effect on the traditional biofilm pathway. Based on the graphs for the eps and sdpA luciferase reporter, the tyrosine kinases do not appear to have targets in the Spo0A transcriptional biofilm pathway (Fig. 1). Purification of recombinant His PtkA The purification of recombinant His-PtkA from E. coli was unsuccessful. Fig. 5 shows the elution (lane 7) that should contain the purified protein, but this was not accomplished. Therefore, future work with the purification of PtkA in E. coli should focus on determining which proteins interact with one another through purifying the tyrosine kinases and their modular proteins and conducting binding studies.

Jarvis

The Manhattan Scientist, Series B, Volume 6 (2019) Peps-lux Luminescence / Absorbance

Peps-lux Luminescence / Absorbance

2000

1000

100

200

300

400

500

600

700

800

2000

1000

100

200

300

400

500

600

700

800

Time (mins)

10000

5000

100

200

300

400

500

Time (mins)

600

700

800

PsdpA-lux Luminescence / Absorbance

Figure 3. Graph of eps luciferase reporter. (Left) Wildtype (blue line), ∆epsA (red line), and ∆epsB (black line) cells were grown for 14 hours to observe transcription of the biofilm matrix operon, eps. (Right) Transcriptional analysis of wildtype ( blue line), ∆tkmA (red line), ∆ptkA (black line), and ∆tkmA∆ptkA (green line).

10000

5000

100

200

300

400

500

600

700

800

Time (mins)

Figure 4. Graph of sdpA luciferase reporter. (Left) Wildtype (blue line), ∆epsA (red line), ∆epsB (black line), and ∆epsAB (yellow line) cells were grown for 14 hours to observe transcription of the biofilm matrix operon, eps. (Right) Transcriptional analysis of wildtype (blue line), ∆tkmA (red line), ∆ptkA (black line), and ∆tkmA∆ptkA (green line).

Figure 5. His-affinity purification of PtkA. Samples are: 1. Lysate; 2. Pellet; 3. Supernatant; 4. Flowthrough; 5. Wash 1; 6. Wash 2; 7. Elution; 8. Pellet from resolubilized membrane fraction; 9. Supernatant from resolubilized membrane fraction.

The Manhattan Scientist, Series B, Volume 6 (2019)

Jarvis

Biofilm time course Published research shows EpsB become autophosphorylated as a biofilm matures, with no autophosphorylation at early timepoints [3]. A typical pellicle biofilm is fully developed between 60-72 hours. In order to determine if PtkA is autophosphorylated during biofilm formation, samples of tagged-PtkA and tagged-EpsB were purified from pellicle biofilms that had been allowed to develop for 44 and 68 hours. After purification these samples were probed by western blot with an anti-phospho-tyrosine antibody to determine if the proteins are differentially phosphorylated at these time points. Fig. Fig. 6 shows that EpsB has a strong signal indicating phosphorylation of EpsB at 44 hours, but not at 68 hours. There does not appear to be any phosphorylated protein in the lanes of tagged PtkA at 44 or 68 hours (though this is hard to determine because of the very intense signal in lane 3). When I used the same samples and probed for the presence of the tagged protein or examined total protein, protein was not visible in any lane at the correct molecular weight (data not shown). Without information on how much tyrosine kinase is in each lane, it is difficult to interpret Fig. 6.

Figure 6. Anti-phospho-tyrosine western Blot of tagged PtkA and EpsB biofilm time course. Strains are in the following order: 1. Tagged PtkA 44hrs 2. Tagged PtkA 68hrs 3. Tagged EpsB 44hrs 4. Tagged EpsB 68hrs

Phosphorylation assay with mutants of ∆epsH Considering the inconclusive results from the time course (Fig. 6), to help answer the question of whether PtkA is regulated by autophosphorylation in a biofilm, I conducted a phosphorylation assay of PtkA and EpsB in a ∆epsH background. As a ∆epsH mutant is unable to produce the exopolysaccharide that signals biofilm state to the tyrosine kinase EpsA/EpsB, I expect this mutation to provide insight into whether PtkA is similarly regulated. In the anti-phospho-tyrosine western blot (Fig. 7) there are bands at the expected size of 35 kDa for tagged EpsB and tagged PtkA in wildtype, but not ∆epsH cells. These bands indicate autophosphorylation of tagged EpsB and tagged PtkA. The anti-FLAG western blot with the same samples (Fig. 8) shows there is protein present at the size of 35 for tagged EpsB and tagged PtkA, but only in the wild type cells. The band at size 10 in both western blots is too small to be full-length PtkA or EpsB, but might be a degraded protein.

Jarvis

The Manhattan Scientist, Series B, Volume 6 (2019)

Figure 7. Anti-phospho-tyrosine western blot of tagged PtkA and EpsB in ∆epsH mutants. Strains in following order: 1. Tagged EpsB ∆epsH; 2. Tagged EpsB; 3. Tagged PtkA ∆epsH; 4. Tagged PtkA.

Figure 8. Anti-FLAG western blot of tagged PtkA and EpsB in ∆epsH mutants. Strains in following order: 1. Tagged EpsB ∆epsH, 2. Tagged EpsB, 3. Tagged PtkA ∆epsH, 4. Tagged PtkA.

Discussion One of the goals of this research was to determine if tyrosine kinases have a function upstream of the traditional biofilm pathway. Figs. 3 and 4 show the Peps and PsdpA luciferase reporter assays and the blue line indicates the wildtype. The tyrosine kinase strains compared to the wildtype look to be about the same shape and height which indicates that the tyrosine kinases do not have targets in the canonical biofilm pathway of B. subtilis and therefore have a minimal effect on the canonical biofilm pathway. This indicates that the B. subtilis tyrosine kinases most likely do not have a function upstream of the traditional biofilm pathway. Another one of the goals of this research was to determine if the tyrosine kinase PtkA was regulated by autophosphorylation during biofilm formation, similar to the regulation that has been reported for EpsB [3]. In order to determine if PtkA is regulated by autophosphorylation during biofilm formation I ran a phosphorylation assay. The time course pellicles results (Fig. 6) indicate EpsB is phosphorylated at 44 hours, but not at 66 hours. This is contradictory to the results from previous research [3], which shows that there is no initial phosphorylation in EpsB at 24 hours and as time goes on phosphorylation of EpsB gets stronger. However, multiple western blots of the samples probing for the protein tag did not show any protein present in any of the samples (data not shown). Also, in a repeat of this western blot with anti-phospho-tyrosine antibody, the band seen in lane 3 of Fig. 6 was no longer present (data not shown). Thus, I cannot conclude anything from the results of our biofilm time course except that this assay needs to be optimized. As the biofilm timecourse is time intensive and was not working well, I examined auto-phosphorylation of EpsB and PtkA in samples that are missing the synthetic protein, EpsH. As EpsH is required to form the polysaccharide component of the biofilm matrix, ∆epsH bacteria are unable to form a biofilm and previous research shows EpsB is phosphorylated in this condition [3]. My results from phosphorylation assays in ∆epsH (Figs. 7 and 8) are also contrary to previous research [3]. The results of this research shows that the strains of PtkA and EpsB that are capable of forming biofilms autophosphorylate. However, in previous research [3] the wildtype strain of EpsB, the one

The Manhattan Scientist, Series B, Volume 6 (2019)

Jarvis

able to form biofilms, did not show autophosphorylation and the ∆epsH mutant, not able to for biofilms, showed autophosphorylation. As with the timecourse, though, there is a big caveat. For this assay I was able to identify my tagged protein in a western blot with an anti-FLAG antibody (Fig. 8). In the tagged-protein blot, I also only saw bands in lanes 2 and 4, suggesting the reason there aren’t bands in the other lanes of the anti-phospho-tyrosine western blot (Fig. 7) is because there isn’t protein in these samples. The final goal of this research study was to purify PtkA for in vitro studies. The purification of recombinant His PtkA in E. coli was unsuccessful. Therefore, future work would be to determine which proteins interact with one another through purifying the tyrosine kinases and their modulator proteins and conducting binding studies. First step should be to purify proteins in E. coli. Overall this research highlights the challenges of studying bacterial tyrosine kinases and their roles in biofilm formation. These kinases though have substantial biofilm phenotypes [2, 5] and thus represent new poorly characterized mechanisms for regulating biofilm formation.

Acknowledgements This work was suported by the Linda and Dennis Fenton ’73 endowed biology research fund. Thank you to Dr. Sarah Wacker for advising and supporting during this research study.

References [1] Mielich-Süss, B., & Lopez, D. (2015). Molecular mechanisms involved in Bacillus subtilis biofilm formation. Environmental microbiology,17(3), 555–565. doi:10.1111/14622920.12527 [2] Huffman, T. (2018). “Characterization of the role of bacterial tyrosine kinase in Bacillus subtilis biofilms.” Manhattan Scientist, 5: 27-36. [3] Elsholz, A. K., Wacker, S. A., & Losick, R. (2014). Self-regulation of exopolysaccharide production in Bacillus subtilis by a tyrosine kinase. Genes & development, 28(15), 1710–1720. doi:10.1101/gad.246397.11 [4] Gao, T., Greenwich, J., Li, Y., Wang, Q. & Chai, Y. (2015). “The Bacterial Tyrosine Kinase Activator TkmA Contributes to Biofilm Formation Largely Independently of the Cognate Kinase PtkA in Bacillus subtilis”. Journal of Bacteriology, 197, 3421-3432 [5] Gerwig, J., Kiley, T. B., Gunka, K., Stanley-Wall, N. & Stülke, J. (2014). “The protein tyrosine kinases EpsB and PtkA differentially affect biofilm formation in Bacillus subtilis.” Microbiol. Read. Engl. 160, 682–691. [6] Grangeasse, C; Cozzone, AJ; Deutscher, J; Mijakovic, I. (2007). “Tyrosine phosphorylation: an emerging requlatory device of bacterial physiology.” Trends in Biological Sciences, 32(2): 86-94. [7] Olivares-Illana, V, et. al. (2008). Structural Basis for the Regulation Mechanism of the Tyrosine Kinase CapB from Staphylococcus aureus. PLoS biology. 6. e143. [8] Yasbin RE, Young FE. Transduction in Bacillus subtilis by bacteriophage SPP1. J Virol. 1974;14(6):1343–1348.

Enhancing enzymatic fuel cells Seth Serrano∗ Department of Biology, Manhattan College Abstract. Enzymatic fuel cells are a renewable energy source that do not produce high enough current densities to be viable alternatives for nonrenewable energy sources. Our ultimate goal was to increase the current density of the anode, using modified glucose oxidase enzymes. We expressed the glucose oxidase gene in yeast cells and attempted to extract the enzyme from the soluble and insoluble fraction. Unfortunately, we were not able to isolate high enough concentrations of the protein to allow for assays on productivity to be performed.

Introduction Research shows that enzymatic fuel cells not only produce harnessable energy, but can, “harvest energy out of [the] extracellular fluids of a mammal” (Cosnier, et al., 2016). These fuel cells are an ecofriendly alternative to traditional energy sources that can even be used to power bionic implants such as glucose sensors and pacemakers. Impeding the full-scale incorporation of enzymatic fuel cells into our daily lives, and no doubt to the chagrin of the environmentally conscious, are several things. For their uses in bionic implants there are sterilization and biocompatibility issues, as well as issues with long-term operational stability (Cosnier, et al., 2016). One issue, however, stands above the rest, as is normally the case: the current densities produced by enzymatic fuel cells is unimpressively modest. Enzymatic fuel cells, like galvanic cells, have an anode and a cathode. The difference is the mechanism for oxidation and reduction which in enzymatic fuel cells, is an enzyme. For our project, we focused on glucose oxidase, the enzyme for the anode. Glucose oxidase is a homodimer native to Aspergillus niger approximately 65 kDa in size (Fig. 1). Each of its subunits have FAD as a cofactor which is involved in the shuttling of electrons – electrons that need shuttling as glucose oxidase carries out its job: oxidizing glucose. Glucose oxidase, “catalyzes the oxidation of β-D-glucose to glucono-δ-lactone and the concomitant reduction of molecular oxygen to hydrogen peroxide” (Frederick et al., 1989) (Fig. 2). Glucose oxidase has been utilized for glucose sensors, glucose detection kits and, because a byproduct of its consumption of glucose is hydrogen peroxide, it has been used for food preservation (Frederick et al., 1989). For the role that it plays in enzymatic fuel cells, it would be ideal if glucose oxidase was stable, as catalytically active as possible, and had a high affinity for glucose. Fortunately, mutants have already been made that have increased stability, catalytic activity and affinity for glucose, mutant (Holland et al., 2012). We hypothesize that these mutants would generate a higher current density, increasing the efficiency of enzymatic fuel cells. ∗

Research mentored by Bryan Wilkins, Ph.D.

The Manhattan Scientist, Series B, Volume 6 (2019)

Serrano

Figure 1. Crystal structure of glucose oxidase. (http://www-jmg.ch.cam.ac.uk/stuff/go/go1.html)

Figure 2. Glucose oxidase oxidizes glucose. Electron shuttling is facilitated by FAD.

Unfortunately, glucose oxidase is a bulky protein which inhibits electron transfer to the electrode, making mediators necessary (Cosnier et al., 2016). Using biological thiol groups (cysteine residues), we believe that we can sequester the protein to a gold electrode, eliminating the need for any mediators, reducing the distance the electrons need to travel and, once again, increasing the current density by reducing the electron travel distance (Fig. 3).

Figure 3. Scheme of thiol sequester to gold anode.

Project design Glucose oxidase mutants The four glucose oxidase variants that were engineered were GOx-Wt (Wt, wild type), GOxcys (cys, cysteine tagged), GOx-4mut (mut, mutant: T56V, T132S, H469C, C543V), GOx-4mut-

Serrano

The Manhattan Scientist, Series B, Volume 6 (2019)

cys. The sequences for the variants were engineered from A. niger glucose oxidase. The genetic framework for the wild type and mutant sequences was provided by the sequence from NCBI Genebank number AAA32695 (E.C 1.1.3.4). Each variant has an N-terminal 6x histidine tag and a proximal TEV protease site. GOx-Cys has a C-terminal 3× cysteine tag to provide an increased affinity to the gold anode. GOx-4mut-cys was designed for increased stability, efficiency, substrate affinity and adherence to the gold anode for direct electron shuttling. Glucose oxidase expression and lysis Our goal was to express GOx-WT and GOx-Cys in yeast, which is eukaryotic just as A. niger is eukaryotic, since previous efforts to express it in E. coli failed due to the protein always being in the insoluble fraction. We hypothesized that expressing the GOx protein in a system similar to its own would increase the solubility of the protein. Glucose oxidase was expressed from the yeast vector, pYRS-GAL1, which was under control of galactose and carries a tryptophan selection marker. Yeast strains were transformed with this expression vector and grown on agar plates lacking tryptophan. Expressions were performed in 50 mL volumes of complete medium lacking tryptophan and supplemented with 2% galactose sugar. Cells were grown to the appropriate densities in media lacking galactose, prior to induction. The GAL1 promoter is inhibited by the presence of glucose (the standard growth metabolite for yeast). As a result, we tested the effectiveness of using raffinose as an alternative to glucose for the initial growth phase. After incubating for 24 hours, with a cell density of OD600 ∼1.0, the cells were pelleted, and washed thrice with water before being resuspended in complete medium lacking tryptophan and 2% galactose and incubated for 24 h, 48 h, and 72 h to determine the optimal time. Whole cell lysates were used to determine if raffinose or glucose worked more efficiently as the intial growth sugar. YeastBusterTM Protein Extraction Reagent (Millipore 71186) was used for this lysis, following the manufacture’s protocol. The whole cell lysates were then analyzed by SDS-PAGE and western blotting. Glucose oxidase isolation Glucose oxidase has a 22 amino acid leader sequence that signals its secretion from the cell. We expected the protein to be concentrated in the media and isolating it from this fraction was our primary goal since it would yield the most glucose oxidase. To do this, we followed the procedure of Holland et al. (2012) with slight modifications. We first expressed protein in 50 mL cultures as described above. Next, we clarified the cell culture by centrifugation and saved the supernatant, which was then dialyzed against 50 mM NaH2 PO4 buffer, pH 8. The dialysis was then subjected to a Ni-NTA column using a 500 µL bed volume. The Ni beads were washed in 50 mM NaH2 PO4 buffer, pH 8 with 20 mM imidazole. Following the protein binding, the column was washed with the same buffer used for equilibrating the column. Finally, elutions were performed in 50 mM NaH2 PO4 buffer, pH 5 with 250 mM imidazole. The elutions were checked for purity using SDSPAGE and western blotting.

The Manhattan Scientist, Series B, Volume 6 (2019)

Serrano

Protein extraction from the cell wall and pellet were also performed according to Rocha et al., 2010. This was done to get any residual glucose oxidase that could be extracted from the cell wall and whole cell lysate that would otherwise have been insoluble. For the cell wall extraction, the pelleted cells from an expression were resuspended and incubated in 5mL of enzyme release buffer which consisted of 50mM potassium phosphate, pH7, 10 mM betamercaptoethanol, 10 mM dithiotheitol, and 2 mM MgSO4 for one hour before being spun down and the supernatant collected. The pellet extraction was performed on the pellet left after the previous step. It was resuspended in 100 mM phosphate buffer containing 2 mM MgCl2 , 2% DTT and protease inhibitor. Total cell disruption was performed by adding glass beads and vortexing five times for one minute, with one-minute intervals, keeping them on ice in between vortexing. The only modification we made to this procedure was that instead of using 10 mL of enzyme release buffer in 100 mL samples we used 5 mL of enzyme release buffer in 50 mL samples. SDS-PAGE electrophoresis and western blot 12% acrylamaide gels were used for all the SDS-PAGE electrophoresis which were performed in standard running buffer (25 mM Tris, 192 mM glycine, 0.1% SDS). The samples were denatured by heat in SDS loading buffer and staining was achieved using commassie. PVDF membranes were used for western blots in standard transfer buffer (25 mM Tris, 192 mM glycine, 20% methanol) to transfer the proteins to for western blotting. It was run for 1 h at a constant 100 V. Using Ponceau-S the membranes were checked for protein transfer before being washed. The membranes were then blocked in 5% milk-TBS (w/v, 50 mM Tris, pH 7.5), for 1 h with shaking. After removing the blocking solution, they were incubated in a 1:5000 dilution of anti-His antibodies (5% milk-TBS), overnight, with shaking at 4â&#x2014;Ś C. The membrane was washed with TBS and incubated in a 1:10000 dilution of anti-mouse HRP-conjugated secondary antibody (5% milk-TBS), for at least an hour at 4â&#x2014;Ś C before being excessively washed and activated using Amersham ECL select substrate and finally imaged.

Results A timed expression of GOx-WT was performed to determine whether glucose or raffinose would yield higher concentration of glucose oxidase after growing in galactose for 24 h, 48 h, and 72 h. Glucose oxidase is roughly 65 kDa. Western blotting of whole cell lysates revealed that cells grown in raffinose, after 48 h of growing in galactose, had the highest concentration of glucose oxidase, as can be seen in Fig. 4. It is possible that after 72 h, the protein was overexpressed resulting in its aggregation or degradation.

Figure 4. Daily test expresssion for GOx from cells grown in raffinose or glucose

Serrano

The Manhattan Scientist, Series B, Volume 6 (2019)

A western blot of the pellet extraction and cell wall extraction for glucose oxidase, as well as media, revealed that there was indeed glucose oxidase that could be extracted from the cell wall and pellet (Fig. 5). At the 24 h time point the most glucose oxidase could be seen in the cell wall and pellet extraction. As expected, however, most of the glucose oxidase was seen in the media. The 48 h and 72 h samples were purified and concentrated via dialysis and a Ni-NTA purification column whereas the 24 h sample was only purified via dialysis.

Figure 5. Presence and relative concentrations of GOx in cell wall extraction, pellet extraction, and media.

These results were replicated but only with limited success. Comassie stains of cell wall and pellet extraction after 24 h, 48 h, and 72 h, purified through a Ni-NTA column, show trace amounts of glucose oxidase (Fig. 6).

Figure 6. Commasie stain of GOx from cell wall extraction (above) and pellet extraction (below).

A western blot of the media after glucose oxidase was expressed for 48 h, purified by the dialysis and Ni-NTA confirmed our initial findings (Fig. 7). Next, we tried to replicate the results with GOx-cys and began attempts to concentrate the protein using 30 kDa molecular weight cut off Centricon Plus-70 centrifugal fiber.

Discussion We succeeded in identifying that glucose oxidase was indeed being expressed by the yeast cells. It is secreted so the bulk of it can be isolated from the media, and we can also extract it from

The Manhattan Scientist, Series B, Volume 6 (2019)

Serrano

Figure 7. Western blot of glucose oxidase 48 h expression after purification through Ni-NTA column.

the cell wall and pellet. Increasing yield without having to drastically increase our volume would be greatly beneficial. We have begun to grow the mutants and next steps include activity assays for the wild type and the mutants to display that the mutants are indeed more efficient and also affinity assays for the GOx-cys to assess if it will, in fact, adhere more to the gold anode. If the mutants do have increase affinity and the GOx-cys variant does adhere to the gold anode, we would like to combine this research with research I performed with Dr. Alexander Santulli where we made gold nanowires to increase the surface area of the anode in hopes of increasing the area that glucose oxidase has to adhere to the anode and, as a result, the current density of the enzymatic fuel cell.

Acknowledgements This work was supported by the Linda and Dennis Fenton â&#x20AC;&#x2122;73 endowed biology research fund. The author would like to thank Dr. Bryan Wilkins for his insight, guidance, and patience.

References Cosnier, S., Gross, A. J., Goff, A. L., and Holzinger, M. Recent advances on enzymatic glucose/oxygen and hydrogen/oxygen biofuel cells: Achievements and limitations. Journal of Power Sources 2016, 325, 252-263 Frederick, F. R.; Tung, J.; Emerick, R. S.; Masiarz, F. R.; Chamberlain, S. H.; Vasavada, A.; Rosenberg, S.; Chakraborty, S.; Schopter, L. M.; Massey, V. Glucose Oxidase from Asperigillus niger: cloning, gene sequence, secretion from Saccharomyces cerevisae, and kinetic analysis of a yeast derived enzyme. Journal of Biological Chemistry. 1989, 3793-3802 Holland, J. T., Harper, J. C., Dolan, P. L., Manginell, M. M., Arango, D. C., et al. (2012) Rational Redesign of Glucose Oxidase for Improved Catalytic Function and Stability. PLoS ONE 7(6): e37924. doi:10.1371/journal.pone.0037924 Rocha, S.N.; Abrahao-Neto, J.; Cerdan, M.E.; Gonzalez-Siso, M.I.; Gombert, A.K. Heterologous expression of glucose oxidase in the yeast Kluyveromyces marxianus. Microbial Cell Factories 2010, 9:4

Genotype-phenotype characterization of biofilms produced by Bacillus subtilis wild isolates Ryan Torres∗ Department of Chemistry and Biochemistry, Manhattan College Abstract. Biofilms are communities of microbes that form on surfaces. The bacteria Bacillus subtilis forms biofilms on the roots of plants, which results in the protection of the plant from pathogens, a process that is referred to as biocontrol. As it is a factor in biocontrol efficacy, understanding biofilm formation on a molecular level will allow researchers to develop better methods to protect plants, resulting in agricultural benefits. This research characterizes the biofilm phenotypes and genetic backgrounds of 20 wild isolates of Bacillus that have previously been characterized in regard to their biocontrol efficacy. My results show that environmental conditions differently affect biofilm formation in the 20 bacteria and open the door to finding specific genetic mutations that cause differences in biofilm formation.

Introduction Bacillus subtilis is a bacterium commonly found in the soil as well as associated with other organisms such as humans and plants. This gram-positive bacterium is aerobic, motile, and is also able to produce spores. This bacterium, in particular, has been shown to produce antibiotic molecules that inhibit virulent strains of bacteria. This mechanism of defense has been shown to work well when the bacteria is associated with a plant in a symbiotic relationship. When associated with plants, the bacteria become non-motile and forms a community that is surrounded by a matrix of self-produced molecules; this community is called a biofilm. These properties of B. subtilis synergistically enhance the plants’ defense against pathogens. The biofilm acts as a physical barrier against pathogens while antibiotics inhibit them. The biofilm formation process begins by bacteria adhering to surface and reproducing. While this fission is occurring, they produce an exopolysaccharide barrier around the colony being formed. The formation of the biofilm is dependent on many variables, including growth conditions and the bacteria’s genetic composition. When associated with plants B. subtilis are able to induce systemic resistance against pathogens. Biocontrol efficacy is the measure of how much a bacteria protects a plant against a virulent pathogen. Two key variables of biocontrol efficacy are biofilm formation and antimicrobial production. Studies have shown that Bacillus subtilis isolates have been able to induce systemic resistance and under the same conditions, biofilm formation is dependent on conserved genes [1]. The goal of this research is to characterize the phenotypes and genotypes of biofilm formation for 20 wild isolates of B. subtilis. These wild isolates were found in soil and extracted for study by Chen et al. [1]. Their 16S RNA was sequenced and used to characterize them as B. subtilis [1]. We hypothesize that the wild isolates have specific genetic changes compared to 3610, the common ∗

Research mentored by Sarah Wacker, Ph.D.

The Manhattan Scientist, Series B, Volume 6 (2019)

Torres

laboratory strain of B. subtilis used in biofilm studies, that influence their biofilm production. Our knowledge of their genes will help us form better strategies to protect plants. There is a number of genes that have been discovered to play a role in biofilm formation in B. subtilis through studies using the model organism 3610. Known genes such as sinR and sinI which are found in the 3610-biofilm formation pathway will be focused on, specifically variations between genes in these strains and others known to be involved in biofilm signaling. Due to differences in phenotypes between the wild isolates and model organism one can predict that there will be mutations within these genes or new genes pertaining to their biofilm formation pathway.

Materials and Methods Strains, media and growth conditions The model B. subtilis strain 3610 as well as wild isolates Bs1-Bs20 where struck out from glycerol stocks on LB agar. A representative colony was chosen and inoculated into 5 mL of LB media which was incubated at 37, shaking at 200 rpm for 5 hours. Biofilm assays were grown from 3µL of LB cell culture on MSn (5 mM potassium phosphate buffer pH7, 0.1 M Mops pH7, 2 mM MgCl2 , 0.05 mM MnCl2 , 1µM ZnCl2 , 2 µM thiamine, 700 µM CaCl2 , 0.5% NH4 Cl, 0.5% glycerol) and LBGM (LB with 1% glycerol and 0.1 mM MnSO4 ) agar plates which were incubated at 30 for 48hrs. At the 48-hour time point, the colony is photographed and rated on how severe their biofilm appears. Genomic DNA extraction 3 mL of LB cell culture was centrifuged at 14,000 rpm for 1 min. The cell pellet was resuspended in 400 µL of H2 O. To each vessel 50 µL of EDTA (0.5M) and 60 µL of lysozyme (10 mg/mL) were added. The mixture was incubated at 57◦ C for 30 minutes. After incubation 650 µL of nuclei Lysis buffer (Promega) was added, followed by 250 µL of protein precipitation solution (Promega); the mixture was then vortexed for 20 seconds. The mixture was centrifuged at 14,000 rpm, the supernatant was transferred to a new tube and mixed with 600 µL of isopropanol and centrifuged again for 5 minutes. After centrifugation, the isopropanol supernatant was removed making sure not to disturb the pellet. The pellet was washed with 1 mL of 70% ethanol making sure to dislodge the pellet from the side of the vessel by centrifuging at 14,000 rpm for 5 minutes. The ethanol supernatant is removed, and the vessel is air-dried. The remaining pellet is resuspended in 100 µL of H2 O. DNA concentrations were quantified using a Qbit assay. These initial DNA concentrations were used in determining volumes for reactions for the Tn5 DNA library prep. Before library prep DNA was run on an agarose gel in order to check for possible degradation of DNA. Tn5 DNA library prep 5.4 µL of purified Tn5 (1.3 mg/mL) and 3.3 µL of both adaptor 1 & 2 (40 µM) was prepared on ice (Final concentrations: Adaptor 1 and Adaptor 2 = 11 µM; Tn5 = 11.25 µM). After each adaptor was added the mixture was vortexed and centrifuged. The now pre-charged transposase

Torres

The Manhattan Scientist, Series B, Volume 6 (2019)

is diluted 1:4. A total volume of 50 µL was made by combining 10 µL of pre-charged Tn5 with 40 µL of H2 O making sure to keep any vessel with Tn5 on ice due to its high rate of degradation when exposed to heat. Tagmentation reactions proceeded with the following volumes, 2 µL of diluted pre-charged Tn5, the appropriate volume of DNA for a 50 ng input, 4 µL 5× TD Buffer, and the appropriate amount of H2 O for a total sample volume of 20 µL. The mixture is mixed by pipetting up and down before incubation at 55◦ C for 7 minutes, held at 4◦ C (Thermal cycler). Since the prior reaction is complete the Tn5 protein is killed with a detergent, 2.5 µL of 0.2% SDS is added to each reaction vessel and set for the same thermal cycler program that was used for the tagmentation reaction. After the Tn5 ss killed 2µL of each tagmentation reaction is added to a new PCR tube on ice R along with 10 µL of One Taq 2X Master Mix with Standard Buffer (NEB), 4 µL of H2 O, 2 µL of 10 µM Primer 1 (i7), and 2 µL of 10 µM Primer 2 (i5) mixed by pipetting. Each i7 primer is different in order to identify which piece of DNA belongs to which strain during sequencing. PCR reaction took place in a thermal cycler with the corresponding program lasting 8 cycles. The last step of the Library prep was to perform an SPRI bead size selection. In a new vessel, 81 µL of H2 O and 19 µL of PCR sample was added. While this was being done XP beads came to room temperature and resuspended by pipetting and inverting. 55 µL of XP beads was added and mixed by pipetting up and down, the vessel was allowed to incubate at room Figure 1. Schematic of the procedure for the preparation temperature for 10 minutes. After the inof Tn5 DNA libraries. The process begins by combincubation period /magnets were used to aging transposomes with adapter DNA and genomic DNA. The transposomes then cleave the genomic DNA and liggregate the magnetic beads to the side of ate adaptors to the DNA in a reaction called tagmentation. the tube. Approximately 155 µL of superLimited-cycle PCR is then used to add index adaptors to natant was transferred to a new tube makeach strand of DNA. This image comes from the Nextera XT DNA Library Prep Kit [2]. ing sure not remove any beads. 12 µL of XP beads were added to the tubes with the supernatant and incubated at room temperature for 10 minutes. Magnets are used once again to aggregate beads to the side of the vessel, but this time the supernatant is disposed of. The beads are washed with 70% ethanol twice and then resuspended with 13 µL of Tris and incubated at room

The Manhattan Scientist, Series B, Volume 6 (2019)

Torres

temperature for 10 minutes. The remaining DNA has a stronger attractive force to the Tris then the beads and thus magnets can be used in order to transfer the supernatant to new PCR tubes without disturbing them. Prepared library samples were then sent for sequencing. EPS purification The purification of biofilm-associated EPS began with growing pellicles in large beakers. For each strain, 250 mL of LBGM was added to four 600 mL beakers and bacteria were added at a dilution of 1:1000. Pellicles were grown-up for 48 hrs before EPS purification. Two strains were tested, one that was ranked low in biofilm severity (Bs1) as well as one that was ranked high (Bs16). The purification began by transferring each of the pellicles to a 50 mL tube. Centrifuging was necessary in order to remove all of the biofilm from the liquid culture. After, 1×PBS is added to obtain a volume of 50 mL; cells were then centrifuged at 790×g for 10 minutes. The media was removed from the tube, making sure not to disturb the pellet. The pellet was then rewashed two more times. The PBS wash was removed and 15 mL of PBS was added, pipetting up and down in order to fully resuspend the pellicles. The suspension was sonicated at 20% duty for 10x1 second pulses and then mixed. This was repeated three times for a 30 second total sonication time. The mixtures were then centrifuged at 3400×g at 4◦ C for 10 minutes. The supernatant was then transferred to a new 50 mL tube. 5× isopropanol was added to the supernatant and polysaccharide was precipitated by rotating at 4◦ C for 24 hours. After 24 hours, the precipitated biofilms were collected by centrifugation at 10,500×g for 10 minutes at 4◦ C. Each pellet was then resuspended in a solution of MgCl2 (0.1 M), DNase (0.1 mg/mL), and RNase (0.1 mg/mL). Resuspended pellets were then sonicated at 20% duty for two 5 s intervals. The samples were statically incubated at 37◦ C for 1 hour. Using equal volumes of phenol-chloroform extractions were performed on each sample, centrifuging at 4,000 rpm for 5 min and then removing the top layer and performing an extraction once again, three times total. Samples were then transferred to a 3500 MWCO dialysis membrane (3 mL slide-a-lyzer) and allowed to dialyze in 1 L of MilliQ water for 3 hours at 4◦ C. The water is switched and allowed to dialyze under the same conditions for 1 hour. The sample is then removed from the dialyzer and placed in a new tube kept at 4◦ C until quantification by sulfuric acid absorbance or freeze-dried and weighed. Congo red absorbance readings LBGM agar with Congo red (0.8 g/L) was used as the media for bacterial growth. PVDF membranes were washed with methanol and then several times with water, before being placed on the plate and inoculated with 5 µL of LB cell culture. Next, the plates were incubated at 30◦ C for 48hrs. After incubation, the membranes along with the biofilm were washed with 1× PBS in order to remove any loosely adhered bacterial cells. The biofilm itself is then removed from the membrane and put into a tube of 10 mL of 1× PBS. The mixture was then sonicated in order to get the cells into suspensions so that they may be quantified. Using an absorbance reader, we measured the absorbance at the specific wavelength associated with Congo red (496 nm) as well as the optical density (600 nm).

Torres

The Manhattan Scientist, Series B, Volume 6 (2019)

Figure 2. Schematic depiction of the experimental system used to grow biofilms. The biofilm (A) was developed on a 3 cm × 3 cm sheet of PVDF (B) resting on an agar culture medium (C). The entire unit, components A through B, was transferred to a new vial containing 10 mL of 1×PBS. (Adapted from [3])

Mutation identification Sequencing results were aligned with the genome of Bacillus subtilis, strain 168, by Dr. Andrés Bendesky at Columbia University in order for the reads to be analyzed using computer software (Integrative genomic viewer). Mutations were identified in sinI and sinR by analyzing the code and comparing it to that of the known genome of strain 168.

Results Biofilm scale results Producing a scale based on the difference of biofilm phenotype was needed for future genetic laboratory work. Without a working biofilm severity scale, any phenotypic changes that may occur after the introduction of genetic mutations would be harder to evaluate. With an unbiased scale, we are able to determine changes by visual appearance rather than having to resort to more mathematical methods. This will increase the rate at which genetic experiments are completed. Biofilm assays on MSn and LBGM were imaged 48 hrs after inoculation. The scale was from 1-10 and was based on visual characteristics such as wrinkles, amount of EPS that look to be produced, height, and diameter. Assays were done on two different media, both giving rise to different phenotypes. Assays were repeated and documented 30 times for each strain in order to build a robust library of phenotype variability. The most common number score was used for each of the strain(mode). MSn (Table 1) is a media with minimal nutrients needed for growth, due to this many phenotypic differences in biofilm formation are observed between wild isolates and a scale is very easily producible. LBGM (Table 2) is a very rich media and has glycerol and manganese additives that promote biofilm formation [4]. The laboratory strain 3610 is not able to form a wrinkly biofilm on MSn (Table 3), but 3610 has the ability to easily grow and form a biofilm in LBGM. Many of the wild strains also readily form very robust biofilms on LBGM; thus, this assay shows us what we would call an “Ultra” robust biofilm. Due to phenotypic differences between plates a separate scale was made for each media.

The Manhattan Scientist, Series B, Volume 6 (2019)

Torres

Table 1. Wild isolate biofilm severity results MSn

Bs1

Bs2

Bs3

Bs4

Bs5

Score:3 Bs6

Score:6 Bs7

Score:1 Bs8

Score:6

Score:7 Bs10

Score:4 Bs11

Score:8 Bs12

Score:3 Bs13

Score:1 Bs14

Score:8 Bs15

Score:1

Score:2

Bs16

Bs17

Score:3 Bs18

Score:5 Bs19

Score:7 Bs20

Score:7

Score:2

Score:4

Score:6

Table 1: Wild Isolate biofilm severity results on MSn

Bs9

Torres

The Manhattan Scientist, Series B, Volume 6 (2019)

Table 2. Wild isolate biofilm severity results LBGM

Bs1

Bs2

Bs3

Bs4

Bs5

Score:3

Score:4 Bs8

Score:4 Bs9

Score:4

Bs6

Score:5 Bs7

Score:3

Score:5

Score:2

Score:7

Score:10

Bs11

Bs12

Bs13

Bs14

Bs15

Score:9 Bs16

Score:4 Bs17

Score:3 Bs18

Score:5

Bs19

Bs20

Score: 8

Score:4

Score:3

Score:8

Score:6

Table 3. 3610 the model organism on MSn and LBGM Table 2: Wild Isolate biofilm severity 3610 results on MSnLBGM

Score:2 Table 3: 3610 the model organism on MSn and LBGM

3610 on LBGM

Score:2

Bs10

The Manhattan Scientist, Series B, Volume 6 (2019)

Torres

EPS purification The EPS purified from samples BS1 and Bs16 were not able to be quantified by absorbance due to time limits. Though the procedure itself gave insight into the structure and amount of EPS purified. Bs1 is rated as 3 on the biofilm severity scale while Bs16 is rated as 8 (on LBGM media, Table 2). During the purification, the amount of biofilm and the appearance of the sample was much more severe in Bs16 than it was for Bs1. This is consistent with the biofilm scale. Congo red absorbance data In order to have a more objective and quantitative assay of biofilm matrix production, biofilms from wild isolates Bs1, Bs5, Bs9, and Bs16 were grown on media containing the protein-dye Congo red. Congo red has been shown to bind preferentially to proteins rich in Î˛-sheets, as it is found in the amyloid-like protein of the B. subtilis biofilm matrix, TasA [5]. These strains were chosen as they represent low biofilm production on LBGM (Bs1 and Bs5) as well as high biofilm production on LBGM (Bs9 and Bs10, Table 2). Biofilms from each strain were solubilized and analyzed by absorbance at 496 nm, the wavelength corresponding to Congo red (Fig. 3). Bs5 had the greatest incorporation of Congo red, although it only had a biofilm severity score of 4. Bs9

Figure 3. Measurement of incorporation of Congo red dye into the biofilm matrix.

and Bs10 had medium amounts of Congo red incorporation although they had high scores on the biofilm severity scale (Bs9 = 7 and Bs10 = 8). Bs1, which had the lowest biofilm score of 3, also had the lowest Congo red incorporation. This experiment was only conducted once and needs to be repeated. Purification of DNA and preparation for sequencing The genomic DNA for all 20 wild isolates was extracted and purified with minimal degradation. Extracted DNA was run on an agarose gel with a ladder in order to determine the presence of genomic DNA, and to make sure there is minimal degradation from the extraction process itself (Fig. 5). A successful library was prepared from each of the extracted genomic DNA samples of the twenty wild isolates, as well as laboratory strain 3610. These libraries were analyzed via a bioanalyzer that determined the amount of prepped DNA by base-pair concentration (Fig. 4).

Torres

The Manhattan Scientist, Series B, Volume 6 (2019)

Figure 4. Tn5 prepped library data. Each strainâ&#x20AC;&#x2122;s DNA was not tested due to the ability to extrapolate, since they underwent the same procedure their data must be similar.

The Manhattan Scientist, Series B, Volume 6 (2019)

Torres

Figure 5. Agarose gel of purified genomic DNA from wild isolates. A) DNA from wild isolates Bs1-Bs10. B) DNA from wild isolates Bs11-Bs19. DNA from wild isolate Bs20 and laboratory strain 3610 is not shown.

Genome alignment and identification of mutations Upon sequencing, the number of reads per strain was high for each sample (Fig. 6), which indicates there should be high coverage of each genome. However, some of the strains have low coverage (Fig. 7) and these strains also had a large number of reads which could not be aligned. This indicates that some of the sequenced strains have genomes that substantially diverge from the genome of B. subtilis 168, which was used for the alignment. These wild isolates that did not align well are likely not members of the Bacillus subtilis species. In support of this, we sequenced the genome of Bacillus amyloliquefaciens GB03, an ingredient in commercial biocontrol products. The genome of GB03 has previously been determined [6] and, like our strains, was previously classified as Bacillus subtilis, but does not align well to the 168 genome (Fig. 7).

Figure 6. Reads per sample for each wild isolate was high. There was a high number of reads Data received from sequencing most be accurate. Differences in coverage must be due to the filter alignment.

In order to identify possible mutations that affect biofilm formation, I examined the sequences of the wild isolates in two genes known to have a major role in the regulation of biofilm formation,

Torres

The Manhattan Scientist, Series B, Volume 6 (2019)

Figure 7. Coverage of wild isolates in relation to the genome of B. subtilis 168. Strains with low coverage have genomes that have large regions which do not align to the genome of 168. Strains with higher coverage have genomes that match very well and are believed to part of the Bacillus subtills species.

sinI, and sinR. My analysis revealed that sinR had high coverage in all strains, but it had the same variations in the sequences of all of the â&#x20AC;&#x153;low coverageâ&#x20AC;? strains. Of these variations, only one was coding, converting a serine to an alanine at the end of the sinR protein (Fig. 8). In all of the low coverage strains, much of sinI was not covered by sequencing reads. Interestingly, GB03 has the same pattern of variations as the low coverage wild isolates. There was one coding variation between genes sinI and sinR that was found in the strains that had high coverage. In both strains Bs12 and Bs20, a coding mutation was found in sinI that converts a valine to methionine (Fig. 9). Valine is a very hydrophobic amino acid when in a protein, but methionine has an alkyl sulfur group at the end of its chain. This sulfur atom has two lone pairs of electrons that are open for interaction. In terms of reactions alkyl sulfur atoms are more nucleophilic then its oxygen analog thus it can react with many electron-deficient systems.

Discussion In this study, twenty wild isolates that had been previously defined as Bacillus subtilis were characterized in terms of their biofilm phenotypes and their genetic sequences. In order to characterize the strength of the biofilm produced by each strain I set up biofilm assays on two separate media and then scored each biofilm based on the number of wrinkles and height of the colony. Interestingly, I found that there was a substantial media dependence on my biofilm scores. While some bacteria had consistent scores on both media (for example, Bs1 was scored a 3 on both MSn and LBGM, though it should be noted that the scales for the two media were different), others were much higher on one media than the other (for example, Bs11 scored a 1 on MSn and a 9 on LBGM). There was also no obvious correlation between the biofilm phenotypes that I scored and the biocontrol efficacy scores that were previously published [1]. We would like to develop a less subjective scoring system for biofilm formation. To this end, I tried to quantify both of the major components of the B. subtilis biofilm matrix, tasA protein, and exopolysaccharide. While these

The Manhattan Scientist, Series B, Volume 6 (2019)

Torres

studies are only at the initial phase, they are promising methods for creating more quantitative and rigorous measures of biofilm formation.

Figure 8. Mutation found within sinR of all low coverage strains, converting TCCâ&#x2020;&#x2019;GCG. This is a coding mutation and when coded into its corresponding protein there is an amino acid mutation Serineâ&#x2020;&#x2019;Alanine.

Figure 9. Visual representation of a noncoding and coding mutation found in both strains Bs12 and Bs20. Mutations that start on the first base pair are more likely to change the entire codon leading to a coding mutation. The second mutated codon is an example of this GTGâ&#x2020;&#x2019;ATG a valine to a methionine when coded to the protein. The mutation before is noncoding.

Torres

The Manhattan Scientist, Series B, Volume 6 (2019)

In terms of the genetic characterization of the wild isolates, we now have full genome sequences for all twenty strains of Bacillus. With the help of our collaborators, we determined that nine of the twenty strains are unlikely to be Bacillus subtilis, based on their alignment to the genome of B. subtilis 168. Interestingly, the wild isolates that do not align well with 168 represent both high biofilm strains (for example, Bs10 which was rated as an 8 on MSn and a 10 on LBGM) and low biofilm strains (for example, Bs1 which was rated a 3 on both MSn and LBGM). Likewise, the wild isolates which have very similar genomes to 168 and are definitely B. subtilis also represent a diverse mixture of biofilms. These non-B. subtilis strains are likely to still be in the genus Bacillus and all seem to be highly related to one another and to Bacillus amyloliquefaciens GB03. These strains must undergo further analysis to determine which species of Bacilli they belong to. I found that there were several variations within the biofilm regulatory genes sinI and sinR. Some of these mutations are coding, meaning there are also changes in the amino acid sequence of the protein the gene codes for. Due to these amino acid changes being very different from each other in terms of their side chain, it is hypothesized that these mutations affect the structure of the protein formed as well as the structure of the biofilm produced. In order to test this, the mutations must be further studied via in vitro binding studies as well as genetic transformations into 3610, the model organism. Our developments in phenotypic analysis will allow us to determine if any changes in phenotype occur during these transformations. These changes in phenotype, as well as DNA, will be documented, and further studies will be conducted involving its interactions with plant protection. Overall, understanding the genetic and phenotypic differences among wild isolates of Bacillus has the potential to allow for the development of better methods of plant protection as well as biofilm disruption.

Acknowledgment This work was supported by the School of Science Research Scholars Program.

References [1] Chen Y., Yan F., Chai Y., Liu H., Kolter R., Losick R, Guo J. H. “Biocontrol of Tomato Wilt Disease by Bacillus Subtilis Isolates from Natural Environments Depends on Conserved Genes Mediating Biofilm Formation.” Environmental Microbiology, U.S. National Library of Medicine, Mar. 2013, www.ncbi.nlm.nih.gov/pubmed/22934631. [2] Nextera XT DNA Library Prep Reference Guide by Illumina. May 2019. Document # 15031942 v05. [3] Anderl J. N., Franklin M. J., Stewart P. S. “Role of Antibiotic Penetration Limitation in Klebsiella Pneumoniae Biofilm Resistance to Ampicillin and Ciprofloxacin.” Antimicrobial Agents and Chemotherapy, American Society for Microbiology, July 2000, www.ncbi.nlm.nih.gov /pubmed/10858336. [4] Shemesh, M. and Chai, Y. “A Combination of Glycerol and ManganesePromotes Biofilm Formation in Bacillus subtilis via Histidine Kinase KinD Signalling” Journal of Bacteriology, May 2013, 195 (12) 2747-2754.

The Manhattan Scientist, Series B, Volume 6 (2019)

Torres

[5] Diehl A., Roske Y., Ball .L, Chowdhury A., Hiller M., Molière N., Kramer R., Stöppler D., Worth C. L., Schlegel B., Leidert M., Cremer N., Erdmann N., Lopez D., Stephanowitz H., Krause E., van Rossum B. J., Schmieder P., Heinemann U., Turgay K., Akbey Ü., Oschkinat H. “Structural changes of TasA in biofilm formation of Bacillus subtilis” Proceedings of the National Academy of Sciences March 2018, 115 (13) 3237-3242; DOI: 10.1073/pnas.1718102115 [6] Choi S. K., Jeong H., Kloepper J. W., Ryu C. M. “Genome Sequence of Bacillus amyloliquefaciens GB03, an Active Ingredient of the First Commercial Biological Control Product.” Genome Announc. Oct 2014 2(5):e01092-14. doi:10.1128/genomeA.01092-14

Xylem conductivities in leaf veins Maya Carvalho-Evans∗ Laboratory of Plant Morphogenesis, Biological Sciences Research Laboratories, Department of Biology, Manhattan College Abstract. Leaves are the main photosynthetic organ of plants, and they come in a variety of shapes and venation patterns. The purpose of this study was to determine if leaf areas, number of xylem vessels, xylem radii and xylem conductivities are well scaled among vein orders in leaves with percurrent leaf venation. Data from thirty-two fullyenlarged leaves were analyzed, and had lamina areas ranging from 7.25 to 887 cm2 . Xylem conductivities were determined for primary, secondary and tertiary veins, and shown to be well scaled with their respected areas. Xylem radii and total numbers of xylem vessels well scaled among vein orders. This data illustrates a leaf’s ability to distribute resources, specifically water, throughout lamina areas.

Introduction Leaf veins have many purposes: transport of water and photosynthates, formation of a framework for the laminas mechanical support, and provisions of mechanical support for mesophyll cells among veins (Roth-Nebelsick et al., 2001). Dicotyledonous plants have many types of venation patterns (Banavar et al., 1999; Durand, 2007; Corson, 2010; Dodds, 2010). Plants have several types of leaf venations (Banavar et al., 1999; Durand 2007; Corson 2010; Dodds 2010). Dicotyledonous species with simple leaves often possess pinnate venation characterized by a primary vein (midrib or midvein) that extends between leaf tips (Runions et al., 2005). Secondary veins extend from the primary vein to leaf margins. Leaf veins, large and small, distribute nutrients and they may act as cantilevered beams to provide mechanical support for leaves (Niklas 1992; Dengler and Kang, 2001; Blonder et al., 2011; Runions et al., 2005; Sack and Scoffoni, 2013). Dicotyledonous species have three basic leaf venation patterns: brochidodromous, craspedodromous, and eucamptodoromous venation. Secondary veins join together near leaf margins for leaves with brochidodromous venation. Secondary veins terminate at leaf margins for leaves with craspedodromous venation. Secondary veins terminate within leaf lamina near leaf margins for leaves with eucamptodoromous venation (Hickey 1979; Roth-Nebelsick et al., 2001). Brochidodromous venation has be suggested to be the most primitive (Hickey, 1971; Hickey and Doyle, 1972) however, there are no conclusive evidence for evolutionary patterns. This suggests that venation patterns may change (Takhtajan, 1980). All biological tissues, including leaves, aim to maximize their resources, while minimizing the coasts of tissue construction (West et al., 1997; Banavar et al., 2000). Some but not all leaf venation patters have veins that reconnect. Vein “reconnections” means that the vein is connected on each end (Blonder et al., 2011). Non-reconnecting veins provide the highest supply rates of nutrients ∗

Research mentored by Lance Evans, Ph.D.

The Manhattan Scientist, Series B, Volume 6 (2019)

Carvalho-Evans

and water per unit leaf mass (Banavar et al., 1999; Durand, 2007; Corson, 2010; Dodds, 2010). In contrast, recent research has found that reconnecting networks are selected for if a high risk of herbivory or other lamina damage occurs or when resources are at a higher demand (Durand, 2006; Katifoi et al., 2010). Primary, secondary, and tertiary veins (Fig. 1) exhibit a high degree of reconnections in percurrent leaf venation (Blonder et al., 2011). Reconnected veins create areas (Fig. 2) within them which can be quantitatively characterized and xylem conductivities can be determined for each vein order, lending itself well to study.

Figure 1. Image of a lamina (Magnolia x soulangeana Soul.-Boud) illustrating primary (pink), secondary (blue), and tertiary veins (white).

Figure 2. Image of a lamina (Magnolia x soulangeana) illustrating primary (surrounded by pink line), secondary (surrounded by blue line), and tertiary areas (surrounded by white line).

The following hypothesis was investigated: For fully-enlarged leaves, leaf areas, number of xylem vessels, xylem radii, and xylem conductivities are well scaled among vein orders.

Materials and Methods Species and tissue sampling All leaf samples were collected from the Manhattan College campus, Bronx, NY and the Van Cortlandt Park, Bronx, NY. Thirty-two species with percurrent simple leaves, non-lobed and a single primary vein were selected. All species had either craspedodromous or eucamptodoromous venation patterns. All leaves were fully-enlarged. Species were identified using the Kershner et al. (2008) methodology (www.tropicos.org and dendro.cnre.vt.edu).

Carvalho-Evans

The Manhattan Scientist, Series B, Volume 6 (2019)

Leaf area measurements Photographic images of leaves were taken and uploaded to a computer. Lamina areas were measured using ImageJ (National Institutes of Health, http://rsb.info.nih.gov/ij). Lamina areas were measured by tracing the perimeter the entire leaf (Fig. 2). Secondary lamina areas were measured by tracing the perimeter around two secondary veins adjacent to one another. Tertiary lamina areas were measured using the same process as secondary lamina areas. Six secondary and tertiary lamina areas were randomly selected and measured for each leaf. A mean was determined for the six measurements. Xylem conductivity measurements Tissue samples were taken from primary, secondary and tertiary veins to determine xylem conductivity measurement. Tissue samples of primary veins (Fig. 3) were cut in the middle, at one-half the distance from the petiole to the leaf tip. Secondary (Fig. 4) and tertiary (Fig. 5) tissue samples were cut near the primary vein sample. All tissues were fixed in FAA (Jensen, 1962) for 24 hours. Samples were then dehydrated through multiple tertiary butanol solutions (Fisher Scientific, Pittsburgh, PA). Once dehydrated, samples were submerged in liquid ParaplastXtra wax (McCormick Scientific, Richmond, IL) at 56◦ C for 24 hours. The wax was changed once and then tissues were embedded in Paraplast. Sections between 15 and 35 µm were cut using a microtone, and put on microscope slides. Tissues were then stained using a 2% safranin solution (Jensen, 1962) and made permanent with Canada balsam (CAS 8007-47-4, Acros; Fisher Scientific, Pittsburgh, PA).

Figure 3. Image of a lamaina cross section (Cornus kousa Bürger ex Miq.) showing a primary vein. Note parenchyma cells on each side of the primary vein. Blue arrows represent the movement of water through xylem vessels. The vascular bundle (yellow circle) has 146 xylem vessels.

The Manhattan Scientist, Series B, Volume 6 (2019)

Figure 4. Image of a lamina cross section (Ulmacea celtis occidentalis L) showing a secondary vein. The vascular bundle has 49 xylem vessels.

Carvalho-Evans

Figure 5. Image of a lamina cross section (Ulmacea celtis occidentalis L) showing a tertiary vein. The vascular bundle has five xylem vessels

Only and all xylem vessels were counted an evaluated for each vein sample. Photographs (at 200 magnification) and coordinates of each vein sample were taken. Photographic images were uploaded to ImageJ (National Institutes of Health, http://rsb.info.nih.gov/ij) to determine xylem vessel radii. Two diameter measurements were obtained for each xylem vessel (Fig. 6). At least thirteen xylem vessel diameters were measured in primary veins, at least ten diameters

Figure 6. Image of a cross section of a portion of a vascular bundle (Hyrandgea arborescens L.) showing xylem vessels (X). Vessel cells are round with thick cell walls. Non-vessel cells are also shown.

in secondary veins, and at least seven or all vessels (if less than seven) were measured in tertiary veins. Diameters were converted to radii for each xylem vessel, and a mean radii was determined. Mean radii were raised to the fourth power and used to calculate xylem conductivities (McCulloh et al., 2009). That value combined with the number of xylem vessels in a vein were used to determine conductivity. Since xylem cells are rarely perfectly circular (Tyree and Zimmermann, 2002), the derived conductivities are approximations. Xylem conductivities of each primary, secondary, and

Carvalho-Evans

The Manhattan Scientist, Series B, Volume 6 (2019)

tertiary veins were calculated using the Hagen-Poiseuille equation: π · number of conduits · average radius of conduits (cm)4 . 8 · viscosity of water

(K. McCulloh, personal communication). The units for Hagen-Poiseuille equation are g cm MPa−1 s−1 . The slope of the regression line was analyzed and significance was determined.

Results Comparison of laminar areas Entire lamina areas ranged from 7.25 cm2 to 887 cm2 , with a mean of 139 cm2 for the thirtytwo species (Table 1). Secondary areas ranged from 0.26 cm2 to 55.1 cm2 , with a mean of 6.65 cm2 and tertiary areas ranged from 0.039 cm2 to 6.16 cm2 , with a mean of 0.909 cm2 . A wide range of areas was provided by the thirty-two species for analysis. Mean secondary lamina areas were well scaled with entire lamina areas (y = 0.038x + 0.55; r2 = 0.86; p = 0.01). Secondary lamina areas were well scaled to tertiary areas (y = 0.22x − 0.31; r2 = 0.77; p = 0.012). Table 1. Leaf (laminar) areas (in cm2 ) for thirty-two plant species with percurrent leaves Species Amelanchier arborea (F. Michx.) Fernald Arctium lappa L. Asclepias syriaca Betula alleghaniensis Britton Betula papyrifera Carpinus caroliniana Walter Carya tomentosa Catalpa bignoniodies Walter Catalpa speciosa (Warder) Engelm Celtis occidentalis L. Cornus drummodii C.A. Mey Cornus kousa Bürger ex Miq. Cynanchum auriculatum Royle ex Wight Euphorbia pulcherrima Wild. Ex Klotzsch Hamamelis mollis Oliv. Hibiscus rosa-sinensis L. Hydrangea arborescens L. Lantana camara L. Liriodendron tulipifera L. Magnolia x soulangeana Soul.-Boud Magnolia kobus DC. Malus pumila Mill. Morus rubra Lour. Ostrya virginiana Britton, Sterns & Poggen b. Oxydendrum arboretum (L.) DC. Phytolacca americana

Entire leaves

Secondary areas

Tertiary areas

66.7 488 235 67.4 44.9 43.5 113 887 311 123 59.6 29 195 26.4 105 82.6 60.8 16.4 59.8 126 313 58.7 80.5 26.4 117 197

2.35 17.0 6.61 1.53 2.14 2.304 3.37 55.1 13.9 6.93 4.05 2.62 10.9 1.3 7.47 6.29 3.35 1.025 4.48 5.27 8.26 4.12 2.88 1.37 4.15 9.2

0.335 2.79 0.417 0.135 0.224 0.189 0.503 6.16 2.77 0.877 0.401 0.195 1.97 0.189 0.348 0.767 0.255 0.119 0.433 1.16 1.09 0.273 0.247 0.144 0.276 2.900

(continued on next page)

The Manhattan Scientist, Series B, Volume 6 (2019)

Carvalho-Evans

Table 1. (continued from previous page) Species

Entire leaves

Secondary areas

Tertiary areas

Platanus occidentalis L. Salix nigra Marshall Tilia americana L. Tilia platyphyllos Scop Ulmus pumila L. Viburnum lentago L.

158 83.4 66.8 74.2 7.25 120

10.0 2.11 2.81 3.62 0.262 5.79

2.81 0.243 0.234 0.225 0.0395 0.368

Mean Standard Deviation

139 171

6.65 9.65

0.909 1.31

Table 2. Xylem conductivities (g·cm·MPa−1 ·s−1 ) for thirty-nine plant species with percurrent leaves Species Amelanchier arborea Arctium lappa Asclepias syriaca Betula alleghaniensis Betula papyrifera Carpinus caroliniana Carya tomentosa Catalpa bignoniodies Catalpa speciosa Celtis occidentalis Cornus kousa Cynanchum auriculatum Euphorbia pulcherrima Hamamelis mollis Hibiscus rosa-sinensis Hydrangea arborescens Lantana camara Liriodendron tulipifera Magnolia kobus Magnolia x soulangeana Malus pumila Morus rubra Ostrya virginiana Oxydendrum arboreum Phytolacca americana Platanus occidentalis Salix nigra Tilia americana Tilia platyphyllos Ulmus pumila Viburnum lentago Mean Standard Deviation

Primary veins

Secondary veins

0.437 18.2 12.2 1.22 0.553 0.476 2.68 20.4 12.9 0.776 0.134 1.06 0.0605 0.311 0.339 0.467 0.153 1.123 3.49 1.24 0.222 1.52 0.973 4.67 5.08 3.61 0.0906 1.33 0.123 0.178 1.39

0.0690 4.75 0.311 0.00876 0.0116 0.00801 0.0452 10.4 0.48 0.312 0.0164 0.100 0.00828 0.0574 0.0106 0.00853 0.0221 0.155 0.294 0.0333 0.00823 0.12 0.0264 0.197 0.165 0.237 0.00963 0.0594 0.00367 0.00598 0.199

3.09 5.25

0.568 1.98

Tertiary veins 0.000588 0.0428 0.000494 0.000234 0.000466 0.000195 0.000132 0.103 0.0255 0.00314 0.00017 0.0223 0.000839 0.00114 0.000996 0.00079 0.00159 0.0063 0.00897 0.0015 0.000127 0.000156 0.000157 0.00189 0.00238 0.0375 0.000309 0.000219 0.0000297 0.0000513 0.000706 0.00830 0.0204

Carvalho-Evans

The Manhattan Scientist, Series B, Volume 6 (2019)

Xylem conductivities vs. leaf area

Xylem conductivity secondary vein (g·cm·MPa-1·s-1)

Xylem conductivity primary vein (g·cm·MPa-1·s-1)

Xylem conductivities were determined for primary, secondary and tertiary veins (Table 2). Xylem conductivities for primary veins ranged from 0.0906 to 20.4 with a mean of 3.09 g·cm·MPa−1 ·s−1 . Xylem conductivities for secondary veins ranged from 0.00367 to 10.4 and tertiary veins ranged from 0.0000513 to 0.103 g·cm·MPa−1 ·s−1 . Primary vein xylem conductivities were well scaled with entire lamina areas (y = 0.028x − 0.74; r2 = 0.81; p = 0.012; Fig 7). Similarly, secondary and tertiary vein xylem conductivities were well scaled with their respective areas (Fig. 8 and Fig. 9).

0 0

500

0 0

1000

Fig. 7.

Figure 7. Relationship between entire lamina area and xylem conductivities of primary veins for thirty-two herbaceous plants (y = 0.028x − 0.74; r2 = 0.81; p = 0.012).

Xylem conductivity tertiary veins (g·cm·MPa-1·s-1)

Secondary lamina area (cm²)

Entire lamina area (cm²) Fig. 8.

Figure 8. Relationship between secondary lamina area and xylem conductivities of secondary veins for thirty-two herbaceous plants (y = 0.19x − 0.72; r2 = 0.89; p = 0.01).

0.10

0.05

0.00 0

Tertiary lamina area (cm²)

Fig. 9.tertiary lamina area and xylem conductivities of tertiary veins for thirty-two herbaFigure 9. Relationship between ceous plants (y = 0.014x − 0.005; r2 = 0.84; p = 0.011).

The Manhattan Scientist, Series B, Volume 6 (2019)

Carvalho-Evans

Comparisons of xylem conductivities among vein orders

Xylem conductivity tertiary veins (g·cm·MPa-1·s )

0.10

-1

Xylem conductivity secondary veins (g·cm·MPa-1·s-1)

Primary, secondary and tertiary vein xylem conductivities were compared to one another. Among the thirty-two species, xylem conductivities for secondary veins were about 30% of primary veins (y = 0.30x − 0.36; r2 = 0.64; p = 0.043; Fig. 10). In contrast, xylem conductivities of tertiary veins were less than one percent of secondary veins (y = 0.0095x+0.0027; r2 = 0.85; p = 0.021; Fig. 11).

0 0

Xylem conductivity primary veins (g·cm·MPa-1·s-1) Fig. 10.

Figure 10. Relationship between xylem conductivities of primary veins and xylem conductivities of secondary veins for thirty-two herbaceous plants (y = 0.30x − 0.36; r2 = 0.64; p = 0.043).

0.05

0.00 0

5 10 Xylem conductivity secondary veins (g·cm·MPa-1·s-1)

Fig. 11.

Figure 11. Relationship between xylem conductivities of secondary veins and xylem conductivities of tertiary veins for thirty-two herbaceous plants (y = 0.0095x + 0.0027; r2 = 0.85; p = 0.021).

Xylem numbers and radii among vein orders Among the thirty-two species 12,301 xylem cells were counted and 830 xylem cell radii were measured. Vessel radii ranged from 4.4 to 19.0, 7.4 to 32.0, and 12.1 to 39.1 µm for tertiary, secondary and primary veins, respectively. Total numbers of vessels per vein ranged from 2.5 to 30, 13 to 109, and 35 to 352 for tertiary, secondary and primary veins, respectively. The ratio of numbers of xylem cells in primary to secondary veins was 3.76:1. The ratio of numbers of xylem cells in secondary to tertiary veins was 3.30:1. For the thirty-two herbaceous species, the number of xylem vessels in secondary to primary veins were well scaled (y = 0.559x − 26.8; r2 = 0.72; p = 0.032; Fig. 12). Likewise, the number of xylem vessels in tertiary to secondary veins were well scaled (y = 0.047x + 6.58; r2 = 0.28; p = 0.12; Fig. 13). Similarly, xylem radii in secondary to primary veins were well scaled (y = 0.73x − 0.43; r2 = 0.73; p = 0.023; Fig. 14) and tertiary to secondary veins xylem radii (y = 0.46x + 0.65; r2 = 0.74; p = 0.024; Fig. 15).

Carvalho-Evans

The Manhattan Scientist, Series B, Volume 6 (2019) 30 Number of xylem vessels tertiary veins

Number of xylem vessels secondary veins

300

200

100

0 0

100

200

300

400

Number of xylem vessels primary veins

100

200

300

Number of xylem vessels secondary veins

Fig. 12.

Figure 12. Relationship between number of xylem vessels in secondary veins and number of xylem vessels in primary veins for thirty-two species of herbaceous plants (y = 0.559x − −26.8; r2 = 0.72; p = 0.032).

Fig. 13.

Figure 13. Relationship between number of xylem vessels in tertiary veins and number of xylem vessels in secondary veins for thirty-two species of herbaceous plants (y = 0.047x + 6.58; r2 = 0.28; p = 0.12).

Tertiary radii (m)

Secondary radii (mm)

Fig.Figure 14.

20 30 Primary radii (mm)

14. Relationship between measurements of secondary vein xylem radii and primary vein xylem radii of thirty-two herbaceous plants (y = 0.73x − 0.43; r2 = 0.73; p = 0.023).

10 20 30 Secondary radii (m)

Fig. 15.

Figure 15. Relationship between measurements of tertiary vein xylem radii and secondary vein xylem radii of thirtytwo herbaceous plants (y = 0.46x + 0.65; r2 = 0.74; p = 0.024).

In addition, midribs and petioles were compared. Mean vessel diameters of midribs were 92% of petioles and mean xylem conductivities of midribs were 52% of petioles.

Discussion Research has shown that leaves are the central component of a plants growth ad hydraulic system (Hickey, 1973; Fisher and Evert, 1982; Russin and Evert, 1985; Gottlieb, 1993; Roth-

The Manhattan Scientist, Series B, Volume 6 (2019)

Carvalho-Evans

Nebelsick, 2001; Coen et al., 2004; Dengler and Kang, 2001; Coomes et al., 2008; Taneda and Tersashima 2012; Sack and Scoffini, 2013; Bar and Ori, 2014; Gleason et al., 2018). This is due to a leafs photosynthetic abilities, and stomata which regulate water loss. Therefore, trees produce leaves that spread laterally for greater exposure to sunlight, and develop efficient veins with xylem and phloem to move water and solutes. Final leaf shape and venation patterns are well scaled (Dengler and Kang, 2001). In this study, only percurrent species with a single primary vein were used. The purpose is to observe the scaling between primary, secondary, and tertiary lamina areas and the efficiency of hydraulic conductivity. The suitable scaling of hydraulic conductivities and lamina areas among the three tiers was determined by xylem radii and numbers of xylem cells. A study by Gleason et al. (2018) has some results that can be compared with the results of the current study. Gleason et al. (2018) studied evergreen angiosperm species’ leaves that are native to Eastern Australia. The purpose of Gleason et al. study was to compare hydraulic conductivity with other characteristics of plants that were specific to cool-dry versus hot-dry climates. All of those species had leaf areas less than 50 cm2 . The leaves of the current study ranged between 7.3 to 887 cm2 and were taken from trees and shrubs of a single location, therefore having similar water availabilities. However, there are differences between to the studies in anatomical data. For example, the ratio between average radii of secondary and primary veins for Gleason et al. (2018) was 0.83 while the same statistics for the current study was 0.73. The number of vessels between secondary and primary veins were about 2.31 times for Gleason et al. (2018), while the same statistic for the current study was 5.2.

Acknowledgments This work was supported by the Linda and Dennis Fenton ’73 endowed biology research fund. The author is also indebted to the Catherine and Robert Fenton endowed chair in biology to Lance S. Evans for financial support for this research. Thanks is given to Glenn Albert for help in species identifications.

References Banavar, J., Colaiori, F., Flammini, A., Maritan, A., Rinaldo, A., 2000. Topology of the fittest transportation network. Phys. Rev. Lett. 84, 4745-4748. Banavar, J., Maritan, A., Rinaldo, A., 1999. Size and form in efficient transportation networks. Nature 399, 130-132. Bar, M., Ori, N., 2014. Leaf development and morphogenesis. Development 141, 4219-4230. Blonder, B., Violle, C., Bentley, L., Enquist, B., 2011. Venation networks and the origin of the leaf economics spectrum. Ecol. Lett. 14, 91-100. Coen, E., Rolland-Lagan, A. G., Matthews, M., Bangham, J. A., Prusinkiewicz, P., 2004. The genetics of geometry. Proceedings of the National Academy of Sciences 101, 4728–4735.

Carvalho-Evans

The Manhattan Scientist, Series B, Volume 6 (2019)

Coomes, D. A., Heathcote, S., Godfrey, E. R., Shepherd, J. J., Sack, L., 2008. Scaling of xylem vessels and veins within the leaves of oak species. Biology Letters 4, 302–306. Corson, F., 2010. Fluctuations and redundancy in optimal transport networks. Phys. Rev. Lett. 104, 048703. Dengler, N., Kang, J., 2001. Vascular patterning and leaf shape. Current Opinion in Plant Biology 4, 50–56. Dodds, P. S., 2010. Optimal form of branching supply and collection networks. Phys. Rev. Lett. 104, 048702. Durand, M., 2006. Architecture of optimal transport networks. Phys. Rev. E. 73, 016116. Durand, M., 2007. Structure of optimal transport networks subject to a global constraint. Phys. Rev. Lett. 98, 88701. Fisher, D. G., Evert, R. F., 1982. Studies on the leaf of Amaranthus retroflexus (Amaranthaceae): quantitative aspects, and solute concentration in the phloem. American Journal of Botany 69, 1375–1388. Gleason, S. M., Blackman, C. J., Gleason, S. T., McCulloch, K. A., Ocheltree, T. W., Westoby, M., 2018. Vessel scaling in evergreen angiosperm leaves conforms with Murray’s law and areafilling assumptions: Implications for plant size, leaf size and cold tolerance. New Phytologist 218, 1360-1370. Gottlieb, M. E., 1993. Angiogenesis and vascular networks: complex anatomies from deterministic non-linear physiologies. In: Garcia-Ruiz, J. M., Louis, E., Meakin, P., Sander, L. M., (Eds.), Growth patterns in physical sciences and biology. Plenum Press, New York, pp. 267–276. Hickey, L., 1971. Evolutionary significance of leaf architectural features in the Woody dicots. Amer. J. Bot. 58, 469 Hickey, L., 1979. A revised classification of the architecture of dicotyledonous leaves. In: Metcalfe, C R., Chalk, L., (Eds.), Anatomy of the Dicotyledons 2nd Ed. Vol. I. Systematic anatomy of the leaf and stem. Oxford, Clarendon Press, pp. 25-39. Hickey, L., Doyle, J., 1972. Fossil evidence on evolution of angiosperm leaf venation. Amer. J. Bot. 59, 661. Jensen, W. A., 1962. Botanical Histochemistry. Principles and Practice. W.H. Freeman. San Francisco, CA. Katifori, E., Szollosi, G. J., Magnasco, M. O., 2010. Damage and fluctuations induce loops in optimal transport networks. Phys. Rev. Lett. 104, 048704. Kershner, B., Matthews, D., Nelson, G., 2008. Field Guide to Trees of North America. Sterling Publ. Co., New York. McCulloh, K. A., Sperry, J. S., Meinzer, F. C., Lachenbruch, B., Arala, C., 2009. Murray’s law, the ‘Yarrum’ optimum, and the hydraulic architecture of compound leaves. New Phytologist 184, 234–244. Niklas, K. J., 1992. Plant Biomechanics. Chicago, The University of Chicago Press.

The Manhattan Scientist, Series B, Volume 6 (2019)

Carvalho-Evans

Roth-Nebelsick, A., Uhl, D., Mosbugger, V., Kerp, H., 2001. Evolution and function of leaf venation architecture: a review. Annals of Botany 87, 553–566. Runions, A., Fuhrer M., Lane, B., Federl, P., Gaëlle, A., Lagan, R., Przemyslaw, Prusinkiewicz., 2005. Modeling and visualization of leaf venation patterns. ACM Transactions on Graphics 24(3), pp. 702-711. Russin, W. A., Evert, R. F., 1985. Studies on the leaf of Populus deltoids (Salicaceae): quantitative aspects, and solute concentrations of the sieve tube members. American Journal of Botany 72, 487–500. Sack, L., Scoffoni, C., 2013. Leaf venation: structure, function, development, evolution, ecology and applications in the past, present and future. New Phytol, 198, 983-1000. doi:10.1111/nph.12253 Takhtajan, A., 1980. Outline of the classification of flowering plants (Magnoliophyta). Bot. Rev. 46, 225-359. Taneda, H., Terashima, I., 2012. Co-ordinated development of the leaf midrib xylem with the lamina in Nicotiana tabacum. Annals of Botany 110, 35–45. Tyree, M. T., Zimmermann, MH., 2002. Xylem structure and the ascent of sap. Berlin, Germany: Springer. West, G., Brown, J., Enquist, B., 1997. A general model for the origin of algometric scaling laws in biology. Science, 276, 122-126

Molecular characterization of Giardia lamblia in oysters (Crassostrea virginica) collected from two sites in New York City Fatimatou Dialloâ&#x2C6;&#x2014; Department of Biology, Manhattan College Abstract. Giardia lamblia has been shown to be prevalent in bivalves including oysters (Crassostrea virginica). It is usually found in food, water, surfaces, and soil that has been contaminated with the feces of infected humans or animals. In humans, G. lamblia is responsible for giardiasis, a gastrointestinal disease which causes diarrhea. The objectives of this research are to determine the prevalence of G. lamblia in C. virginica collected in the fall of 2018 from two different marine environments in NYC, Orchard Beach and Clason Point, and to identify the genotype of G. lamblia using molecular techniques. A total of thirty-nine (39) specimens of C. virginica were collected from Clason Point (23) and Orchard Beach (16) on September 24, 2018, during low tide. Each specimen was dissected to isolate the digestive gland, adductor muscle, mantle, gills, hemolymph, and foot. These tissues were assayed for the presence of G. lamblia using primers that target the Î˛-giardin gene. G. lamblia was found to have a higher prevalence in C. virginica collected from Orchard Beach than from Clason Point, at 12.5% and 8.7%, respectively. Additionally, G. lamblia was detected only in the digestive gland, adductor muscle, and mantle tissues of the oysters. From Clason Point, G. lamblia had a 4.35% prevalence in both the digestive glands and the adductor muscles. From Orchard Beach, G. lamblia had a prevalence of 12.5% in the mantle. Lastly, all of the positive specimens were identified to be of the G. lamblia assemblage A genotype.

Introduction Giardia lamblia also known as Giardia duodenalis and Giardia intestinalis is known as a zoonotic protozoon parasite (Prystajecky et al., 2015). It is an endoparasite that inhabits the upper small intestines of mammals (Cernikova et al., 2018). G. lamblia is known to be released in the environment by contaminated human and animal feces (Adell et al., 2014). It is also carried by land surface water into coastal waters (Hogan et al., 2013). There are 8 genotypes of G. lamblia which are named from A to H. However, only genotypes A and B are known to infect humans (Lopez-Romero et al., 2015). In humans, it is responsible for giardiasis, a gastrointestinal disease which causes diarrhea (Lopez-Romero et al., 2015). It is also associated with the development of irritable bowel syndrome and chronic fatigue in humans (Cernikova et al., 2018). G. lamblia is known to be the third most common diarrhea-causing parasite throughout the world. Previous work has shown that giardiasis is 30% prevalent in third world countries, whereas it is about 3% prevalent in developed countries (Cernikova et al., 2018; Lopez-Romero et al., 2015). In 2003, 9100 cases of giardiasis were reported by the United States Centers for Disease Control and Prevention (Miller et al., 2005). Infection of G. lamblia is usually obtained through direct contact with infected hosts or through consumption of contaminated drinking water and food (Cernikova et al., 2018; Feng and Xiao, 2011; Tei et al., 2016). â&#x2C6;&#x2014;

Research mentored by Ghislaine Mayer, Ph.D.

The Manhattan Scientist, Series B, Volume 6 (2019)

Diallo

G. lambliaâ&#x20AC;&#x2122;s life cycle consists of two stages, the trophozoite (vegetative form) and the cyst (infective form) (Cernikova et al., 2018; Lopez-Romero et al., 2015). The cysts can survive for a long period in water and can become concentrated in shellfish (Adell et al., 2014). The trophozoite is what colonizes the host intestines and the cyst is resistant to environmental conditions. After an individual consumes food or water contaminated by cysts, a process called excystation is induced where one cyst generate two trophozoites. The trophozoites then colonize the proximal small intestine without invading its epithelia. As the trophozoites migrate down the gastrointestinal tract, they turn into cysts, a process known as encystation. This is due to changes in pH and levels of bile and cholesterol in the lower gastrointestinal tract (Cernikova et al., 2018; Lopez-Romero et al., 2015). The cysts are removed from the host through the feces and the cycle is completed when a new host is infected by the cyst (Cernikova et al., 2018; Lopez-Romero et al., 2015) (Fig. 1).

Figure 1. The Life Cycle of Giardia lamblia (Pathogen and Environment | Giardia | Parasites | CDC. Retrieved from https://www.cdc.gov/parasites/giardia/pathogen.html)

Although the quality of drinking water is not of concern in the United States, it is important to investigate the prevalence of G. lamblia in bivalves such as oysters (Crassostrea virginica) consumed raw on a day to day basis and in the oceans frequented by individuals during the hot summer months. These factors make this research relevant to public health. Also, in the discipline of microbiology, mussels, such as oysters may be used as biosensors of intestinal parasites such as G. lamblia in marine environments (Miller et al., 2005). This is because they are filter feeders. Thus, parasite cysts can be detected in their tissues indicating the contamination of the marine environment by parasites (Miller et al., 2005). Previous studies have shown the prevalence of G. lamblia in bivalves from Orchard Beach in

Diallo

The Manhattan Scientist, Series B, Volume 6 (2019)

NYC (Tei et al., 2016). It was found that G. lamblia DNA was present in Mya arenaria, Geukensia demissa, Crassostrea virginica, and Mytilis edulis (Tei et al., 2016). The objectives of this research are to do a molecular characterization of G. lamblia to determine the genotype and of prevalence of G. lamblia in C. virginica collected in the fall of 2018 from Orchard Beach and Clason Point, NYC.

Materials and Methods Oyster collection A total of thirty-nine (39) samples of C. virginica were collected from Clason Point (40.8160◦ N, 73.8564◦ W) and Orchard Beach (43.5178◦ N, 70.3773◦ W) on September 24, 2018 during low tide (Fig. 2). Of the total, twenty-three (23) samples were collected from Clason Point and sixteen (16) from Orchard Beach. Before dissection, these samples were kept in seawater (4◦ C). Each sample was dissected to isolate tissues such as the digestive gland, adductor muscle, mantle, gills, hemolymph, and foot. DNA was extracted using the DNeasy Blood and Tissue extraction kit from Qiagen (Germantown, MD, USA).

Figure 2. Collection sites: Clason Point (40.8160◦ N, 73.8564◦ W) and Orchard Beach (43.5178◦ N, 70.3773◦ W) (maps.google.com)

Molecular analysis To detect G. lamblia in the tissues, Nested-PCR was performed. This was done in two steps in which the final volume of each PCR tube was 25 µL. In both steps, a volume of 24 µL of cocktail with parasite-specific primers that target the β-giardin gene were used for each sample

The Manhattan Scientist, Series B, Volume 6 (2019)

Diallo

(Tei et al., 2016). In the first step, 100 ng purified G. lamblia DNA was used as positive control and water was used as negative control. The primers used for this step were the Gia7 Forward Primer (50 -AAGCCCGACGACCTCACCCGCAGTGC-30 ) and the Gia759 Reverse Primer (50 -GAGGCCGCCCTGGATCTTCGAGACGAC-30 ) (Hong et al., 2014). In the second step, 1 µL of the first step PCR product was used for all samples and controls. The primers used in this step were the Gia7 Nested Forward Primer (50 -GAACGAACGAGATCGA GGTCCG-30 ) and the Gia759 Nested Reverse Primer (50 -CTCGACGAGCTTCGTGTT-30 ) (Hong et al., 2014). The PCR conditions used were 94◦ C for 3 min, 94◦ C for 45 s, 52◦ C for 45 s, and 72◦ C for 1 min, a final extension at 72◦ C for 7 min (Hong et al., 2014). PCR products were visualized by ultraviolet light using agarose gel electrophoresis stained with ethidium bromide. Lastly, positive samples were sequenced and the genotypes were analyzed using BLAST (“Nucleotide BLAST: Search nucleotide databases using a nucleotide query”).

Results Prevalence og G. lamblia in oysters Specimens of C. virginica from both Clason Point and Orchard Beach tested positive for G. lamblia (Figs. 3 and 4).

Figure 3. Representative agarose gel electrophoresis showing a positive G. lamblia sample from Clason Point. Lane 1: 100 bp marker; Lane 3: positive control (511 bp); Lane 8: positive for β-giardin gene; Lanes 5-7, and 9-18: negative for β-giardin gene; Lane 20: negative control.

Figure 4. Representative agarose gel electrophoresis showing a positive G. lamblia sample from Orchard Beach. Lane 1: 100 bp marker; Lane 8: positive for β-giardin gene; Lane 11: negative control.

Diallo

The Manhattan Scientist, Series B, Volume 6 (2019)

It was found that out of the 23 specimens collected from Clason Point, two specimens tested positive for G. lamblia resulting in a prevalence of 8.7%. On the other hand, out of the 16 specimens collected from Orchard beach, two specimens tested positive for G. lamblia resulting in a prevalence of 12.5% (Fig. 5).

Figure 5. Prevalence of Giardia lamblia in oyster samples from Clason Point and Orchard Beach. Giardia lamblia was detected in oyster sample numbers 7 and 16 from Clason Point and in oyster sample numbers 7 and 8 from Orchard Beach. The prevalence of G. lamblia was 8.70% from Clason Point and 12.50% from Orchard Beach.

Figure 6. Giardia lamblia Tissue distribution prevalence in oysters from Clason Point and Orchard Beach. Giardia lamblia was detected in the digestive gland, abductor muscle, and mantle tissues of the oysters. From Clason Point, G. lamblia was 4.35% prevalent in the digestive glands and 4.35% prevalent in the abductor muscles. From Orchard Beach G. lamblia was 12.5% prevalent in the mantle. For all other tissues, G. lamblia was not detected.

From the tissues tested from both sites, G. lamblia was only detected in the digestive gland, adductor muscle, and mantle tissues of oysters. From Clason Point, G. lamblia was detected only in the digestive gland at a prevalence of 4.35% and adductor muscle at a prevalence of 4.35% (Fig. 6). Conversely, from Orchard Beach G. lamblia was only detected in the mantle tissue which resulted in 12.5% prevalence (Fig. 6). As for the gills, hemolymph, and foot, G. lamblia was not detected in specimens from both sites resulting in 0% prevalence (Fig. 6). Genotype of Giardia lamblia In summary, positive samples of G. lamblia from both sites were of the assemblage A genotype. From Clason point, sample number 7 was identified as 100% assemblage A and isolate 81f, and sample number 16 was identified as 99.78% assemblage A and isolate 14JBG (Table 1). From Orchard Beach, sample number 7 was identified as 100% assemblage A and isolate 14JBG and sample number 8 was identified as 100% assemblage A and isolate 81f.

The Manhattan Scientist, Series B, Volume 6 (2019)

Diallo

Table 1. Giardia lamblia Genotype distribution in oysters from Clason Point and Orchard Beach. Site

Sample ID

Isolate

Assemblage

% Identity

Clason Point

7ACOYF 16DCOYF

81f 14JBG

A A

100 99.78

Orchard Beach

7MOOYF 8MOOYY

14JBG 81f

A A

100 100

Discussion We have shown that there is a higher prevalence of G. lamblia in C. virginica collected from Orchard Beach (12.5%) than Clason Point (8.7%) in the Fall of 2018 (Fig. 5). It is uncertain why there is a higher prevalence in Orchard Beach. This is probably due to the number of wastewater treatment plants around each site. As seen in Fig. 7, there are more wastewater treatment plants

Figure 7. Map of wastewater treatment plants near Clason Point and Orchard Beach (maps.google.com)

around Clason Point than Orchard beach. This could be due to the topography of the two sites; Orchard Beach is more sandy than Clason Point. Moreover, G. lamblia was only detected in the mantle of specimens collected from Orchard Beach. However, previous studies have shown G. lamblia detection in other tissues such as the digestive glands, gills, and adductor muscles (Tei et al., 2016). On the other hand, G. lamblia was only detected in the digestive gland and adductor muscle.

Diallo

The Manhattan Scientist, Series B, Volume 6 (2019)

Although G. lamblia has multiple genotype assemblages, ranging from A to H (Lopez-Romero et al., 2015), in this study, only assemblage A was identified (Table 1). All of the positive specimens for G. lamblia were identified as assemblage A genotype. These are important findings because humans are susceptible to infection from consumption of oysters contaminated with G. lamblia of assemblage A. This also implies that humans might have played a role in contaminating these marine environments through feces. Additionally, because oysters can be bioindicators (Adell et al., 2014), the prevalence of G. lamblia in them indicates that the water is also contaminated. G. lamblia can be carried into coastal water through sewage water (Graczyk et al., 2008). According to Hogan et al. (2013), G. lamblia can also be carried into coastal water through runoff surface water. Therefore, when humans swim in these marine environments, they might accidentally ingest G. lamblia cysts and become infected. Additionally, this suggests that humans might have been involved in the contamination of these two marine environments, perhaps through runoff and sewage water.

Conclusions The screening of Oysters (Crassostrea virginica) has shown the prevalence of G. lamblia in both Clason Point and Orchard Beach. Through molecular analysis, it was found that all positive specimens for G. lamblia were of assemblage A genotype which can potentially infect humans. It was also found that there is a higher prevalence of G. lamblia in C. virginica collected from Orchard Beach than Clason Point. Additionally, G. lamblia was detected in the mantle at a higher percentage than other tissues. Analysis of the other bivalve species collected at both sites will give a more accurate assessment of the prevalence of G. lamblia at both sites.

Acknowledgement This work was supported by the Linda and Dennis Fenton â&#x20AC;&#x2122;73 endowed biology research fund. The author thanks Dr. Michael Judge for directing and supervising the collection of the oysters. She especially thanks her advisor, Dr. Ghislaine Mayer, for her support, knowledge, guidance, and understanding throughout this research.

References Adell A. D., Smith W. A., Shapiro K., Melli A., Conrad P. A. (2014). Molecular epidemiology of Cryptosporidium spp. and Giardia spp. in mussels (Mytilus californianus) and California sea lions (Zalophus californianus) from central California. Applied and Environmental Microbiology, (24), 7732-40. doi: 10.1128/AEM.02922-14. Cernikova, L., Faso, C., and Hehl, A. B. (2018). Five facts about Giardia lamblia. PLoS Pathogens, 14(9), 1 - 5. https://doi.org/10.1371/journal.ppat.1007250 Feng, Y., and Xiao, L. (2011). Zoonotic Potential and Molecular Epidemiology of Giardia Species and Giardiasis. Clinical Microbiology Reviews, 24(1), 110. doi: 10.1128/CMR.00033-10.

The Manhattan Scientist, Series B, Volume 6 (2019)

Diallo

Graczyk, T. K., Kacprzak, M., Neczaj, E., Tamang, L., Graczyk, H., Lucy, F. E., and Girouard, A. S. (2008). Occurrence of Cryptosporidium and Giardia in sewage sludge and solid waste landfill leachate and quantitative comparative analysis of sanitization treatments on pathogen inactivation. Environmental Research, 106(1), 27-33. https://doi.org/10.1016/j.envres.2007.05.005 Hogan, J. N., Daniels, M. E., Watson, F. G., Oates, S. C., Miller, M. A., Conrad, P. A., Shapiro, K., Hardin, D., Dominik, C., Melli, A., Jessup, D. A., Miller, W. A. (2013). Hydrologic and vegetative removal of Cryptosporidium parvum, Giardia lamblia, and toxoplasma gondii surrogate microspheres in coastal wetlands. Applied and Environmental Microbiology, (6), 1859-1865. doi: 10.1128/AEM.03251-12 PMCID: PMC3592235 Hong S. H., Anu D., Jeong Y. I., Abmed D., Cho S. H., Lee W. J., Lee S. E. Molecular characterization of Giardia duodenalis and Cryptosporidium parvum in fecal samples of individuals in Mongolia. Am J Trop Med Hyg. 2014 Jan; 90(1):43-47. doi: 10.4269/ajtmh.13-0271. Lopez-Romero, G., Quintero, J., Astiazaran-Garcia, H., & Velazquez, C. (2015). Host defences against Giardia lamblia. Parasite Immunology, (8), 394. https://doi.org/10.1111/pim.12210 Miller, W. A., Atwill, E. R., Gardner, I. A., Miller, M. A., Fritz, H. M., Hedrick, R. P., Melli, A. C., Barnes, N. M., Conrad, P. A. (2005). Clams (Corbicula fluminea) as bioindicators of fecal contamination with Cryptosporidium and Giardia spp. in freshwater ecosystems in California. International Journal for Parasitology, 35(6), 673-684. https://doi.org/10.1016 /j.ijpara.2005.01.002 Nucleotide BLAST: Search nucleotide databases using a nucleotide query. (n.d.). https://blast. ncbi.nlm.nih.gov/Blast.cgi?PROGRAM=blastn&PAGE TYPE=BlastSearch&LINK LOC= blasthome Prystajecky, N., Tsui, C. K.-M., Hsiao, W. W. L., Uyaguari-Diaz, M. I., Ho, J., Tang, P., and IsaacRenton, J. (2015). Giardia spp. Are Commonly Found in Mixed Assemblages in Surface Water, as Revealed by Molecular and Whole-Genome Characterization. Applied and Environmental Microbiology, 81(14), 4827- 4834. doi: 10.1128/aem.00524-15 Tei, F. F., Kowalyk, S., Reid, J. A., Presta, M. A., Yesudas, R., and Mayer, D. C. (2016). Assessment and Molecular Characterization of Human Intestinal Parasites in Bivalves from Orchard Beach, NY, USA. International journal of environmental research and public health, 13(4), 381. doi:10.3390/ijerph13040381

Bark formation for Cephalocereus columna-trajani, Neobuxbaumia macrocephala, and Neobuxbaumia mezcalaensis Phillip Dombrovskiy∗ Laboratory of Plant Morphogenesis, Department of Biology, Manhattan College Abstract. Past studies have observed the formation of bark on tall, long-lived columnar cacti between the 32◦ N and 32◦ S latitude lines. This dense layer of bark has certain negative implications upon the cactus in that early morbidity is experienced. As opposed to the cacti thriving for hundreds of years at a time, premature death is expected. The formation of such bark stems from the epidermal cells dividing uncontrollably without a feedback or inhibition system. The current study is observational and focused on three cactus species of the Tehuacán Valley, Puebla, Mexico (18◦ N, 97◦ W). The species include: Cephalocereus columna-trajani, Neobuxbaumia macrocephala, and Neobuxbaumia mezcalaensis. The purpose of this study was two-fold: (a) examining the effects of bark formation on internal cactus tissues, and (b) observing the relative percentage of bark formation on different surfaces of the cactus stem. The three cactus species reacted differently to sunlight-induced bark formation and did not follow a hypothesized 2:1 ratio of percent bark composition between the south-facing and north-facing ribs that compose the cactus stem.

Introduction Cactus plants are a group of succulents that grow in the western hemisphere, in particular North and South America. They possess unique adaptations that allow them to thrive in areas of increased aridity, or areas with little to no annual rainfall amounts. Tall columnar cacti, such as the Carnegiea gigantea, are capable of living for up to hundreds of years at a time (National Park Service, 2016). The longevity of these columnar cacti, however, has experienced a decrease that was first observed in the early 1940s (Lajtha et al., 1997). Research was conducted to propose a cause for declining saguaro cactus populations, and has pointed to the formation of dense bark layers atop the surface of the cactus plants (Steenbergh and Lowe, 1977;, Gibson et al., 1986; Stolte, 1988). The development of such bark has been noted for a multitude of cactus species stretching across the 32◦ N and 32◦ S latitudinal lines (Evans et al., 1994; 2015). Evans et al. (1994) have shown that the cause for such formation of bark is exposure of the cactus surface to UV-B radiation from the sun. The location of the bark formation on the surface of cactus plants is related to direction (Evans et al., 1992). Sunlight-induced bark formation occurs most prominently on the equatorial-most-facing surface, before spreading to other surfaces of the cactus. Evans et al. (2001) performed experiments to confirm that UV-B radiation is the significant contributing factor for the formation of bark. Healthy, green surface tissues of cacti develop epicuticular waxes as the primary symptom of bark formation (Evans and Macri, 2008). Afterwards, the epidermis layer of the cactus’ skin begins to undergo mitosis, producing new cells to the exterior. The formation of bark may be considered a wound-healing mechanism; the epidermal cells damaged by enhanced levels of UV-B radiation ∗

Research mentored by Lance Evans, Ph.D.

The Manhattan Scientist, Series B, Volume 6 (2019)

Dombrovskiy

produce polyphenolic compounds, which contribute to the darker coloration denoted as true bark (Evans et al., 1994; 2001). With the advent bark formation, functional consequences follow for the cactus plant at the interphase of the cactus stem and surrounding environment: the important metabolic processes of photosynthesis and cellular respiration are inhibited. A dense layer of bark Fig. 1.enveloping the cactus stem results in poor gas exchange (Lajtha et al., 1997). The current study examined three endemic cactus species, namely Cephalocereus columnatrajani, Neobuxbumia macrocephala, and Neobuxbaumia mezcalaensis, of the Tehuacán Valley, Mexico (Fig. 1). The research focused on histology, examining the effects of the formation of bark

Figure 1. Cactus plants at Tehuacán-Cuicatlán Biosphere Reserve, San Juan Raya, Puebla, Mexico 18◦ N, 97◦ W.

on once healthy cactus tissues on the underlying, more interior tissues, and to determine the relative amounts of bark formation present on the cardinal-facing surfaces for the cacti species. A previous study (Geller, 1984) helped establish the relative amounts of sunlight projected onto an upright pole (along the North, East, South, and West directions) at all latitudesand at any particular date of the year. The compiled data were examined and averaged to conclude that a 2:1 ratio persists between the amount of sunlight hitting a south-facing and north-facing surface, respectively, at the latitude where the research took place.

Dombrovskiy

The Manhattan Scientist, Series B, Volume 6 (2019)

A qualitative and quantitative approach was used via the aid of technology and microscopy. It was hypothesized that the three cactus species of the Tehuacán Valley, Mexico, would exhibit similar changes to interior tissues with the presence of bark formation. It was also hypothesized that roughly a 2:1 ratio of percent bark composition would be observed between the south-facing and north-facing cactus surfaces, respectively.

Materials and Methods Site selection The research took place at two location of the Tehuacán Valley, Mexico, between the months of May and June in 2018. A population of Neobuxbaumia macrocephala and Neobuxbaumia Fig. 2. mezcalaensis cacti (Fig. 2) was observed in the Tehuacán-Cuicatlán Biosphere Reserve, San Juan

Figure 2. Cacti observed in field settings: A. Cephalocecerus columna-trajani; B. Neobuxbaumia macrocephala; C. Neobuxbaumia mezcalensis

Raya, Puebla, Mexico (18.10◦ N, 97.21◦ W). A population of Cephalocereus columna-trjani cacti (Fig. 2) was observed at a region of the Sierra Madre del Sur near Highway 125, Puebla, Mexico (18.20◦ N, 97.30◦ W). The Tehuacán Valley is home to a rich biodiversity of cactus species, which may be attributed to the rain shadow effect of the Sierra Madre Oriental mountains (UNESCO; Smith, 1965). The vegetation type that dominates the region is a thorn scrub forest with especially

The Manhattan Scientist, Series B, Volume 6 (2019)

Dombrovskiy

rocky soils (Valiente-Banuet and Ezcurra, 1991). The species of interest were identified with the help of descriptions by Anderson (2001) and were confirmed by local guides. A total of 54 C. columna-trajani plants, 77 N. macrocephala plants, and 124 N. mezcalaensis plants were studied. Bark surface coverage study methods and analysis Cacti were selected from areas that were not densely populated. Plants living in densely populated areas are exposed to less direct sunlight due to shade by other plants. The cactus plants were selected from both flat and sloped landscapes. At a height of 1.7 meters above the ground, the cactus plants were examined for the percent composition of bark along the cardinal-facing surfaces. Efforts were made to sample a wide variety of cactus surfaces, ranging from those with healthy tissues to those with complete bark coverage. About the cactus stem run cactus ribs that protrude outwards toward the surrounding environment. Each cactus rib consists of a central protrusion known as a crest, as well as adjacent intrusions known as troughs, situated both to the right and left of the crest (Fig. 3). Fig. 3. A

Figure 3. Image displaying the rib of a young, healthy N. mezcalaensis cactus. A: Crest of the cactus rib; a protrusion of tissue. B: Adjacent trough of the cactus rib; an indentation of tissue.

The percent composition of bark was evaluated using past standard procedures (Evans et al., 1995; 2005). Data were collected for each cactus rib closest to the azimuthal directions of South, East, North, and West. Consequently, values were recorded for the percent composition of bark on 12 surfaces; each cardinal-facing cactus surface possessed a central crest, a right trough, and a left trough. The sampled cacti were then grouped into classes based upon the percentage of bark on the south-facing crest: Class I possessed 0-24% bark coverage, Class II 25-49% , Class III 5074%, and Class IV 75-100%. Fig. 4 shows the progression of bark formation for Cephalocereus

Dombrovskiy

The Manhattan Scientist, Series B, Volume 6 (2019)

columna-trajani cacti.Fig. 4. A

Figure 4. Pictures showing the progression of bark coverage on the south-facing crest of the stem surfaces of C. columna-trajani. A: Class I, 0-24% coverage. B: Class II, 25-49% coverage. C: Class III, 50-74% coverage. D: Class IV, 75-100% coverage.

Histology excision procedures Images of the three cactus species were taken prior to the excision of histological samples. Four categories of cactus health were established, i.e. (1) cacti with completely healthy surfaces, (2) cacti with initial bark formation, (3) cacti with extensive bark formation, and (4) cacti with complete bark coverage. Rectangular shaped samples were removed from the surface of the cacti species at a height of approximately 1.7 meters above the ground, and were nearly 2 cm × 1 cm × 1 cm in dimensions. The flat troughs were used for the sample excisions because of the presence of areoles along the crests of the cactus’ ribs. Tissues were treated with FAA (formalin–acetic acid–alcohol) for preservation before arrival at the research laboratory (Jensen, 1962). Histological analysis The collected tissue samples were exposed to variable alcohol concentrations to become dehydrated. Five solutions of increasing tertiary butanol concentrations were used. Each sample spent 24 hours at a time within the solutions. Samples were stirred and spun at constant rates. The samples used were embedded in a melted paraffin solution (Paraplast X-tra Tissue Embedding Medium; McCormick Scientific, St. Louis, MO). The tissue samples were then prepared to be examined via microscopy with the help of a microtome. Tissue sections of 25 to 45 µm were placed on to microscope slides. Xylene solution deparaffinized the tissues on the microscope. A

The Manhattan Scientist, Series B, Volume 6 (2019)

Dombrovskiy

5% safranin solution was used to stain the cactus tissue samples. Canada balsam was used as the mounting medium for the addition of a coverslip. The histology slides were viewed under an optical microscope at 40× and 100× magnification. Photos were taken via an iPhone camera during the observational study. Photographs were uploaded to a computer and the software package ImageJ (ImageJ, National Institutes of Health) was used to analyze the measurements of the different tissue types within the excised cactus samples. Microsoft Excel was used for the organization of the data. Statistical t-test analysis, at an alpha-level of 0.05, was conducted for characteristic comparisons of different cells before and after bark formation (Excel, Microsoft 2007).

Results C. columna-trajani cacti possessed few significant differences upon comparing the crest surfaces to each other, as well as the trough surfaces to each other (Table 1). Of the sampled cactus plants for C. columna-trajani, Class III cacti were the only ones for which the percent bark compoTable 1. Bark percentages on surfaces of Cephalocereus columna-trajani, Neobuxbaumia macrocephala, and Neobuxbaumia mezcalaensis. Upper case letters: Statistical analyses of crest comparisons. Lower case letters: Statistical analyses of crest to trough comparisons. Roman numerals: Statistical analyses of trough to trough comparisons. If letters are different, bark percentages were different at p < 0.05. ∗ Denotes that both right trough and left trough bark percentage values have been averaged to arrive at a single value. South Cactus Class

East

North

West

Crest

Troughs*

Crest

Troughs

Crest

Troughs

Crest

Troughs

Cephalocereus columna-trajani I ( 0 – 24% bark) II (25 – 49% bark) III (50 – 74% bark) IV (75 – 100% bark)

11A 35A 67A 98B

5a 3a 13a 17a

56B 90B 60A 100A

4a 23b 15a 25a

53B 48A 55A 93B

6a 5a 13a 12b

39B 63A 63A 94B

3a 14b 16a 17b

Cephalocereus columna-trajani I ( 0 – 24% bark) II (25 – 49% bark) III (50 – 74% bark) IV (75 – 100% bark)

11A 31A 63A 96A

6a 14a 16a 59a

16A 35A 48A 84B,D

6a,c 12a 12a 51a,c

11A 21B 22B 65C

3b 7b 4b 442b,c

11A 23B 47A 76C,D

4b,c 11b 11a 42b,c

Neobuxbaumia mezcalaensis I ( 0 – 24% bark) II (25 – 49% bark) III (50 – 74% bark) IV (75 – 100% bark)

10A 33A 66A 93A

9a 7c 24a 31a

10A 30A 59A,B 88A

7a 19a,c 18a 27a,b

9A 60A 57A,B 77B

6a 33a 18a 21b

10A 34A 40B 66C

7a 7b,c 17a 18b

sitions were similar across all surfaces. Cacti of Class I had significant differences in bark composition between south-facing and north-facing surfaces. The contrary was true for cacti of Class II, III, and IV for C. columna-trajani, in which the percent bark coverage between the southern and northern surfaces of the cacti were statistically similar. Upon inspection of N. macrocephala, bark

Dombrovskiy

The Manhattan Scientist, Series B, Volume 6 (2019)

coverage for crest surfaces of Class I cacti were statistically similar (Table 1). Such was not the case for cacti of Class IV, where significant differences were observed across all cardinal-facing crests and troughs. The south-facing crest of the sampled N. macrocephala cacti on average possessed higher amounts of bark formation than on remaining surfaces for both Class III and IV. Bark formation patterns for N. mezcalaensis (Table 1) followed a relationship similar to that of N. macrocephala. Class I cacti had significantly similar bark percentages on both crest and trough surfaces, while significant differences were noted for Class IV cacti across both crests and troughs. With greater formation of bark on the south-facing crest, greater variation in bark percentage composition among all remaining cactus surfaces appeared. Examination of the histology slides in combination with ImageJ yielded comparisons of cell dimensions both before and after bark formation (Table 2). In all three species studied, the cuticle layer disappeared entirely even when only a few epidermal cell proliferations had occurred, or when initial bark formation had occurred. A diverTable 2. Statistics for surface tissue characteristics of Cephalocereus columna-trajani, Neobuxbaumia macrocephala, and Neobuxbaumia mezcalaensis. ∗ Denotes that no epidermal cell width values were taken for C. columna-trajani because cells were globular, and width values would be relatively proportional to depth values. Only cell depth values were recorded. C. columna-trajani

N. macrocephala

N. mezcalaensis

Depth of cuticle (µm)

Before bark After bark

35.5 ± 6.79 Not present

50.0 ± 10.2 Not present

38.3 ± 14.6 Not present

Epidermal cell shape

Before bark

Globular

Columnar

Rectangular

Epidermal cell depth (µm)

Before bark

Epidermal cell width (µm)

84.8 ± 14.0

107 ± 15.4

59.2 ± 6.8

Before bark

-∗

54.1 ± 9.55

103 ± 17.5

No. of hypodermal cell layers

Before bark After Bark P

6.60 ± 0.49 6.65 ± 1.61 0.910

8.26 ± 2.37 7.31 ± 1.37 0.076

7.47 ± 2.26 8.53 ± 2.09 0.180

Hypodermal cell shape

Before bark After bark

Rectangular/Cuboidal Deformed

Globular/Rectangular Deformed

Globular/Cuboidal Deformed

Hypodermal cell depth (µm)

Before bark After bark P

116.7 ± 27.1 66.3 ± 18.8 <0.01

59.6 ± 15.0 65.0 ± 18.8 <0.01

93.8 ± 25.0 81.7 ± 30.9 0.510

Hypodermal cell width (µm)

Before bark After bark P

126.7 ± 35.0 66.2 ± 19.0 <0.01

179 ± 45.0 214 ± 30.9 0.075

89.5 ± 23.4 81.2 ± 31.0 0.284

Chlorenchyma cell shape

Before bark

Columnar/Globular

Columnar

Chlorenchyma cell depth (µm)

Before bark After bark P

239 ± 109 142 ± 43.3 <0.01

260 ± 109 262 ± 100 0.952

332 ± 128 172 ± 74.5 <0.01

The Manhattan Scientist, Series B, Volume 6 (2019)

Dombrovskiy

sity of cell shapes and sizes was present among all three cactus species of the study. Epidermal cells appeared in globular, columnar, and rectangular shapes for C. columnatrajani, N. macrocephala, and N. mezcalaensis, respectively. Significant differences to the number of hypodermal cell layers before and after bark formation were only noted for N. macrocephala. Deformation of hypodermal cells was common to all three species. For C. columna-trajani, hypodermal cells shrunk significantly in both depth and width. For N. macrocephala, the hypodermal cell lumens only significantly increased in depth. For N. mezcalaensis, no significant changes to either hypodermal cell depth or width were observed. Only N. macrocephala retained the columnar shape of its healthy chlorenchyma cells throughout bark formation, with average cell depths of 260 Âľm and 262 Âľm before and after bark formation, respectively. The effects of bark formation on the interior tissues were also studied. Qualitative examination of the prepared histological slides for C. columna-trajani displayed a hypodermis layer that remained intact even with complete bark coverage of the crest surface (Fig. 5). With dividing epidermal cells procuring the dense bark layer, the hypodermal cell lumens appeared to shrink and shrivel. Similarly, the chlorenchyma cells closest to the hypodermis layer became altered in shape and size. Observation of N. macrocephala similarly displayed an intact hypodermis at first with the epidermal cells being responsible for the production a dense bark layer (Fig. 6). With the advent of having an extensive amount of bark present, the hypodermal cells changed in both shape and size, to the degree that deformation of the entire layer was noted. Depressions in the hypodermal tissues could be observed, as well as the thinning of the cell walls composing the hypodermal cells themselves. With complete bark formation, it was difficult to distinguish a hypodermal cell from the more superficial, dividing epidermal cells that produces bark. The internal tissues of N. mezcalaensis behaved similarly in response to the formation of bark. Once cuboidal and globular shaped hypodermal cells assumed shriveled shapes with initial bark formation, and eventually became flattened discs under a dense layer of deceased epidermal cells, also referred to as bark (Fig. 7). Deformation of the hypodermis layer was most prominent for sampled cacti with complete bark formation upon their surfaces.

Discussion The initial hypothesis that the three cactus species of the TehuacaĚ n Valley, Mexico, would exhibit similar internal responses to the formation of bark is not supported by the data presented in the study. The three species did have commonalities in the components of their healthy, green anatomies: there exists a cuticle layer, an epidermis, a hypodermis layer, and a chlorenchyma layer. The dimensions, specifically depth and width, of the cells that compose these anatomical features were dissimilar (Table 2). Much variation existed among C. columna-trajani, N. macrocephala, and N. mezcalaensis when the process of bark formation was examined, even though the mechanism by which bark formed was similar to descriptions of past research studies and literature (Gibson, 1986; Evans et al., 2015). While the internal most tissues remained whole, the epidermal

Dombrovskiy

The Manhattan Scientist, Series B, Volume 6 (2019) a Fig. 5.

C b

Figure 5. Surface tissues, removed tissue samples, and prepared histological slide for C. columna-trajani. A-C: Healthy, green surface tissues. A: The cactus surface possesses little to no bark formation. B: The excised stem tissue shows the green color due to photosynthetic cells possessing the required pigments for necessary metabolic processes. C: Microscope slide of healthy, green surface tissues. The cuticle is especially thin. A single-celled epidermis layer (a) is present; the cells are globular shaped. A multi-cellular hypodermis layer (b) is visible made up of cuboidal and globular cells, along with a more interior chlorenchyma layer (c) of globular shaped cells. D-F: Images of a cactus surface showing initial bark formation. D: Surface of a trough with bark. E: Excised stem sample showing the appearance of a yellowish color and loss of green pigment. F: The cuticle is not retained when bark begins formation. The epidermal cells have begun to divide to produce the bark layer, while he hypodermal cell lumens have begun to shrink. G-I: Images of a surface with extensive bark coverage. G: Surface of a trough with bark. H: Excised tissue sample where a bark layer can be noted, as well as the progression of the yellow hue to more internal tissues; the development of polyphenolic compounds can be attributed to the change in color of the tissues. I: The hypodermis remains intact, with the cell lumens continuing to shrink in size and become deformed in structure. Dense striations of flattened epidermal cells create the bark layer. The chlorenchyma cells remain intact and are not affected in size or form. J-K: Surface tissue showing complete bark coverage. J: A rough texture has developed to the cactus surface. K: The green pigment has almost entirely disappeared from the surface tissues; decay of the interior tissues is prevalent. L: The hypodermis remains distinguishable with a similar number of hypodermal cell layers, and the chlorenchyma cells retain their size and form. Many more striations of deceased epidermal cells make up the dense layer of bark.

The Manhattan Scientist, Series B, Volume 6 (2019) A

Dombrovskiy

a b

Fig. 6.

Figure 6. Surface tissues, removed tissue samples, and prepared histological slide for N. macrocephala. A-C: Healthy, green surface tissues. A: The cactus surface possesses little to no bark formation. B: The excised stem tissue shows the green color due to photosynthetic cells possessing the required pigments for necessary metabolic processes. C: Microscope slide of healthy, green surface tissues. The cuticle is especially thin. Columnar shaped epidermal cells line the external most tissue (a). More interior is a thick hypodermis layer (b). Most interior is a layer of chlorenchyma cells responsible for photosynthesis (c). D-F: Images of a surface with initial bark. D: Surface with initial development of grey colored bark. E: Removed stem sample showing that the green color of the underlying chlorenchyma cells is retained. F: The cuticle fractures with the advent of bark formation. The epidermal cells have begun to proliferate and to develop polyphenolic compounds, displaying brown hue under the microscope. The hypodermal cell lumens have begun to alter in size and form. G-I: Samples of a surface with extensive bark formation. G: bark along the surface develops via patches that come together with time. H: Excised tissue sample displaying the disappearance of the green color from internal tissues, with replacement by yellowish hues. I: The bark is developed out of crippled epidermal cells that flatten and coalesce. The hypodermal cells are now deformed in shape. The chlorencyhma cells closest to the hypodermis layer begin to deform. J-L: Images of a surface with complete bark. J: Image of a cactus rib that has complete bark coverage. K: The breakdown of internal tissues is evident in the excised tissue. L: An intense brown coloration develops as part of the bark layer. The hypodermis layer is distinguishable, although it has become deformed. The chlorenchyma cells change in both size and form.

Dombrovskiy

The Manhattan Scientist, Series B, Volume 6 (2019)

Fig. 7

a b

c D

Figure 7. Surface tissues, removed tissue samples, and prepared histological slide for N. mezcalaensis. A-C: Images of a relatively healthy surface. A: Surface of a trough with no bark present. B: Excised stem samples showing green chlorenchyma (interior). C: Histological slide showing a single-celled layer of globular-shaped cells for the epidermis (a) with a thing cuticle. The hypodermis layer (b) consists of multiple layers of cuboidal and globular cells. Columnar chlorenchyma cells (c) are situated most interiorly. D-F: Images for a cactus surface with initial development of bark in the form of small speckles. D: Sample of a surface that appears to have speckling with bark. E: The internal tissues have begun to develop yellow to brown colors due to the buildup of polyphenolic compounds. I: The cuticle layer disappears even with initial bark formation. Epidermal cells proliferate to produce the bark layer while the hypodermis and chlorenchyma layer remains intact. G-I: Images of a surface with extensive bark coverage. G: Newly forming bark is lighter in color than older bark patches. H: Excised tissue with bark layer most superficial and interior tissues that continue to change in color towards yellow hues. I: The hypodermis remains intact and distinguishable. The hypodermal cells have lost their secondary cell wall characteristics, and appear to have thinner cell walls. Deformation of the chlorenchyma cells continues. J-L: Surface with complete bark coverage. J: Surface image for a cactus with complete bark coverage. K: Removed cactus tissues with a dense layer of bark that appears to resemble a cork material. Internal tissues appear to be discolored and especially dry. L: With complete bark coverage, the hypodermis deforms. The once columnar-shaped chlorenchyma cells have also undergone deformation and became altered in size and shape.

The Manhattan Scientist, Series B, Volume 6 (2019)

Dombrovskiy

cells were the source for bark layer formation. The single layer epidermis goes on to produce a new layer of epidermal cells to the exterior of the cactus stem. Subsequently, it is the internal most epidermal cell that continues to undergo mitosis, pushing the oldest of epidermal cells further away from the cactus interior. Eventually, due to inhibition of the oldest epidermal cells from acquiring much needed energy and water, these cells cripple and desiccate, acquire polyphenolic compounds attributing to color, and afterwards die. The variation among the species lies in the way the internal tissues react to the presence of bark atop the cactus stem surface. Three different responses were documented among the three species of the study. For C. columna-trajani, the cuticle layer disappeared entirely with initial bark formation. Visual inspection of the tissues under microscopy showed that the hypodermis layer remained intact throughout bark formation and experienced no noticeable deformations. The cells that compose the hypodermis layer were significantly altered in size when compared before and after bark formation. The parenchyma cells significantly decreased in size as well. For N. macrocephala, the cuticle likewise disappeared entirely with the advent of bark formation. Only significant differences in the hypodermal cell depth were noted. Unlike as with C. columna-trajani, N. macrocephala experienced deformation of the hypodermis layer, as well as a loss of secondary cell wall characteristics. N. macrocephala cacti were the only sampled ones of the study to not experience statistically significant deformation in dimensions of the chlorenchyma cells. The third cactus species of the research, N. mezcalaensis, similarly lost the cuticle layer to bark formation. It did not possess significant differences in hypodermal cell size, but deformation of the layer itself occurred. The chlorenchyma cells of N. mezcalaensis experienced significant decreases in size. The second hypothesis of the research was that approximately a 2:1 ratio of percent bark composition would be present on the south-facing and north-facing crest surfaces of the cacti, respectively, based on the results of earlier research (Geller, 1984). The acquired data of this study refutes that hypothesis. On average, the south-facing crest for both N. macrocephala and N. mezcalensis did possess greater amounts of bark formation than did the north-facing crest. C. columna-trajani cacti maintained statistically similar bark percentages between south and northfacing surfaces for Class II, III, and IV, with only Class I cacti maintaining significantly greater bark formation on the north-pointing surface as compared to the south-pointing. It is evident that anatomical and physiological changes occur to the underlying tissues of the cactus stem with the formation of bark. The change in color to the internal tissues can be attributed to the production of polyphenolic compounds, as well as potential decay of tissues. The loss of the green color may be a loss of green pigment, responsible for photosynthesis. With inhibition of photosynthesis due to inadequate light reaching photosynthesizing cells because of the thick layer of bark cells, deterioration of the healthy green color and loss of function could thus be a possibility.

Dombrovskiy

The Manhattan Scientist, Series B, Volume 6 (2019)

Acknowledgments This work was supported by the Linda and Dennis Fenton ’73 endowed biology research fund. The author is indebted to the Catherine and Robert Fenton Endowed Chair in Biology to Dr. Lance Evans for additional financial support of this research.

References Anderson, E.F. 2001. The Cactus Family. Timber Press. Portland, OR. Evans L.S., Howard K., Stolze E. Epidermal browning of saguaro cacti (Carnegiea gigantea); is it new or related to direction? Environmental and Experimental Botany, 32 (1992), pp. 357-362. Evans, L.S., V. Cantarella, K. Stolte, and K.H. Thompson. 1994. Epidermal browning of saguaro cacti (Carnegiea gigantea): Surface and internal characteristics associated with browning. Environ. Exper. Bot. 34: 9-17. 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. Exper. 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. Ex per. Bot. 46: 181-187. Evans L.S., M. Zugermayer, and Young A. J. B. 2003. Changes in surface injuries and mortality rates of Saguaro (Carnegiea gigantea) cacti over a twelve-year period. J. Torrey Bot. Soc 130:238–243. Evans, L.S., A. Macri. 2008. Stem Surface Injuries of Several Species of Columnar Cacti of Ecuador. J. Torrey Bot. Soc. 135(4): 475-482. Evans, L.S., M. Cooney. 2015. Sunlight-induced bark formation in long-lived South American columnar cacti. Flora - Morphology, Distribution, Functional Ecology of Plants. 217. 33-40. Geller, G. and P. Nobel. 1984. Cactus ribs: influence of PAR interception and CO2 uptake. Photosynthetica 18: 482-494. Gibson, A.C. and P.S. Nobel. 1986. The cactus primer. Harvard University Press, Cambridge, MA. Jensen, W.A., 1962. Botanical Histochemistry. W.H. Freeman and Co. University of California, Berkeley. Lajtha, K., K. Kolberg, and J. Getz. 1997. Ecophysiology of the saguaro cactus (Carnegiea gigantea) in the Saguaro National Monument: relationship to symptoms of decline. Journal of Arid Environments 36: 579–590. National Park Service. 2016. Saguaro cactus. https://www.nps.gov/orpi/learn/nature/saguarocactus.htm San Juan Raya Centro Ecoturistico. San Juan Raya homepage. http://www.sanjuanraya.com. Smith, C.E. 1965. Flora Tehuacán Valley. Fieldiana Botany 31: 107–143. Steenbergh WF, Lowe CH. 1977. Ecology of the saguaro: 2. Reproduction, germination, establishment, growth, and survival of the young plant. Sci. Monogr. Ser. 8.Washington, DC: USDI National Park Service. 248 p.

The Manhattan Scientist, Series B, Volume 6 (2019)

Dombrovskiy

Stolte, K.W., 1988. Trip report: Saguaro and Chiricahua National Monuments, November 6-14, 1987. Memo, National Park Service Air Quality Division, Lakewood, Colorado. Valiente-Banuet, A. and E. Ezcurra. 1991. Shade as a Cause of the Association Between the Cactus Neobuxbaumia tetetzo and the Nurse Plant Mimosa luisana in the Tehuacan Valley, Mexico UNESCO. TehuacaĚ n-CuicatlaĚ n Valley: Origin habitat of Mesoamerica. United Nations Educational, Scientific and Cultural Organization; http://whc.unesco.org/en/list/1534/.

Mechanical stresses of tree branches: Primary and secondary stems Deirdre Franksâ&#x2C6;&#x2014; Laboratory of Plant Morphogenesis, Department of Biology, Manhattan College Abstract. Each tree species displays a unique morphology. However, all trees have some similarities such as main stems, branches and leaves. Are the mechanical stresses similar among tree species? This research focused upon understanding tree branch mechanical stresses to determine if tree show a uniform of characteristics. Branch samples of forty species were collected from around Manhattan College campus. Data of the two species, Celtis occidentalis and Acer plantanoides were analyzed. Bending stress values for primary branches of C. occidentalis (0.0077 Pa) and A. plantanoides (0.0076 Pa) almost identical. Bending stresses for secondary branches of C. occidentalis (0.0075 Pa) and A. platanoides (0.0085 Pa) were like bending stresses of primary branches. These similarities suggest that bending stresses may be a unifying principle among tree species.

Introduction Evolution has allowed for tree species to display unique morphologies found amongst each individual species. These diversities are observed in each treeâ&#x20AC;&#x2122;s branch dispersion as well as their leaf morphologies. Despite the obvious physical differences, trees can maintain similar structural functions. Each of these morphologies has allowed for leaves to achieve their maximum development, photosynthetic yield, and reproduction for the trees (Evans et al., 2007). Structurally, a vast majority of trees can be seen with primary branch stems laterally projecting from the main trunk (Evans et al., 2007). These primary branches have secondary branches and higher tier branches that produce leaves for direct access to the sun. Another structural similarity may be found in trees mechanical stresses of a species primary and secondary branches. The trees are under stress brought upon by either self-weight or external loads via environmental factors, wind or snow (Wilson and Archer, 1979). Previous research has been done for mechanical stress of the main vertical stem of the tree and its resistance to these self and external loads (Evans et al., 2007). There has been little in-depth analysis on how branch morphology may give further unifying properties among all tree species. The objective of the research was to provide an analysis of mechanical stress values and relationships for primary and secondary stems. This study will look to find a unifying factor for these branches overlooking their distinct morphologies and may display trees undergoing similar behavioral patterns when expending energy for branch production. Two hypotheses will be tested: 1. Bending stresses of primary stems will be similar for all plant species tested. 2. Bending stresses of primary and secondary stems will be similar for each plant species tested. Data of only two species are presented. The ongoing research will consider data from forty tree species. â&#x2C6;&#x2014;

Research mentored by Lance Evans, Ph.D.

The Manhattan Scientist, Series B, Volume 6 (2019)

Franks

Materials and Methods All branches for this research were obtained around the Manhattan College campus, Brust Park, and Broadway in Riverdale, New York from June through August 2019. Primary stems were analyzed starting from stem terminals to stem bases. Secondary branches, simple or complex, were removed from primary stem for separate analysis. Branch measurements included length, mass, and diameter. Primary stem analysis included its junctions, portions of the stem located between two secondary branches. Complex secondary branches, having at least three tertiary branches, underwent similar measurements to determine individual bending stress. Plants from two species, Celtis occidentalis (Fig. 1) and Acer platanoides (Fig. 2) were ana-

Figure 1. Terminal branch of Celtis occidentalis with and without leaves present.

Figure 2. Terminal branch of Acer platanoides with and without leaves present.

Franks

The Manhattan Scientist, Series B, Volume 6 (2019)

lyzed. All stem segments were photographed. Image J (imagej.nih.gov/ij) was used to determine stem diameters. Rulers were placed in each photograph to set scale for measurements. Data were placed into Microsoft Excel 2016 (www.microsoft.com) for analysis. Equations for section modulus and bending moment were entered in each spreadsheet allowing for bending stress to be calculated. The equation used for section modulus was S = π × (segment diameter)3 /32

and the equation used for the bending moment was

M = 1/2(main segment length/1000) × {(main segment length/1000) × g} + (main segment length/1000) × {[(cumulative weight - main segment weight)/1000] × g} + (bending moment of secondary branch). Bending stresses were calculated the plots of bending moment (y-axis) versus second modulus (x-axis) for all samples.

Results

Bending stresses for primary branches of C. occidentalis (Fig. 3) and A. platanoides (Fig. 4) were 0.0077 and 0.0076 Pa, respectively. These slopes were nearly identical considering that C. occidentalis (Fig. 1) had an alternate phyllotaxy and A. platanoides (Fig. 2) had an opposite phyllotaxy. In addition to similar slopes, the data are consistent from terminals to stem bases with r2 values of 0.95 and 0.99, respectively. Celtis occidentalis - Main Stem Bending Moment (N-m)

Bending Moment (N - m)

4 3 2

Acer platanoides - Main Stem

y = 0.0077x - 0.0964 R² = 0.9518

6 5 4 3

y = 0.0076x - 0.1398 R² = 0.9949

2 1

0 0

100

200

300

400

500

600

Section Modulus (10-9 m3)

Figure 3. Data of a primary stem of Celtis occidentalis

0 0

200

400

600

800

Section Modulus (10-9 m3)

Figure 4. Data of a primary stem of Acer platanoides

For both species, bending stresses of secondary stems were like bending stresses of main stems. Bending stresses for primary (Fig. 3) and secondary (Fig. 5) branches of C. occidentalis were 0.0077 and 0.0075 Pa, respectively. Bending stresses for primary (Fig. 4) and secondary (Fig. 6) branches of A. platanoides (Fig. 4) were 0.0076 and 0.0085 Pa, respectively. In a manner like primary stems, slopes of the data from terminals to stem bases for the secondary stems had r2 values of 0.96 and 0.93, respectively.

The Manhattan Scientist, Series B, Volume 6 (2019) Celtis occidentalis - Side Branch 2 0.6

Bending Moment (N-m)

0.5

Bending Moment (N-m)

Franks

Acer platanoides - Secondary 3

0.4 0.3

y = 0.0075x - 0.3209 RÂ˛ = 0.9613

0.2 0.1 0 0

100

0.5 0.4 0.3 0.2

y = 0.0085x - 0.1541 RÂ˛ = 0.9285

0.1 0 0

Section Modulus (10-9 m3)

Figure 5. Data of a secondary stem of Celtis occidentalis

Figure 6. Data of a secondary stem of Acer platanoides

Discussion Although, trees species have a wide variety of morphologies, the overall objective of this study is to determine if the mechanical properties of tree branches are similar. Both similar among species and similar for branches within each tree species. Tree species have distinct morphological traits due to continuous evolution. Each of these distinctions provides for adequate photosynthesis for adequate reproduction and species maintenance. he data of this preliminary study shows that the stems of these two species have similar bending stresses for their primary and secondary branches and that both species have similar stresses. One of the aims of the overall experiment with the forty species that will eventuality be analyzed is do even the smallest tertiary or higher order branches show identical or nearly identical bending stresses from stem tips to stem bases as the join the lower order stems, eventually to primary stems. The current data set indicates that trees can act as uniform units of one another. The growth and energy expenditure required to produce primary and secondary branches is done to evenly distribute stress amongst each of the branches. If done correctly, one branch should not be put under more stress than the others within a tree. This is a principle being seen within these two species with slight deviations when comparing the secondary and primary stress values. The deviations could be a result of premature growth or fewer tertiary branches present on the secondary stem.

Acknowledgements This research was funded by the Catherine and Robert Fenton Endowed Chair to Dr. Lance S. Evans.

References Evans, L.S., Kahn-Jetter, Z., Martinez, M., Tarsia, P., Torres, J. Trees (2007) 22: 283. https://doi.org/10.1007/s00468-007-0182-7 Wilson B, Archer R (1979) Tree Design: some biological solutions to mechanical problems. Bioscience 29:293-298

Reiterative engineering properties of tree branches Hasan Hamid∗ Department of Mechanical Engineering, Manhattan College Abstract. The purpose of this report is to investigate the reiterative properties of various tree branches. The mechanical stresses of 40 diverse species of trees were analyzed and utilizing spreadsheet software, the data was accumulated onto graphs. Two of the species being analyzed in this report are Betula nigra and Catalpa speciosa. When data was plotted onto graphs, the slopes gave values of mechanical stress. Primary branches of the two species were 0.0039 and 0.0040 Pa, respectively. For secondary branches of Betula nigra, the values were 0.0045, 0.0036 and 0.0026 Pa, similar to primary branches. Also, for secondary branches of Catalpa speciosa, the values were 0.0025, 0.0030 and 0.0035 Pa, once again analogous to primary branches. For both species, values of secondary branches were close to primary branches. These data suggest that branch bending stresses are similar for all branches of a species.

Introduction Tree branches come in varying morphologies and those of the same species are diverse in some distinct visual aspect. Indeed, it can be thought that there can be no similarities in the random placement of branches but when examining the branches, a general sequential increase in thickness and length can be identified visually (Wilson and Archer, 1977). Thus, a correlation must exist, and certain properties are analyzed in order to showcase such correlations. Two properties are then introduced to begin the analysis – bending stress and section modulus (Beer and Johnson, 1992). Bending stress, in engineering, is known to act on a beam when a load or force is applied to a horizontal member. In the case of branches, the only load characterized on the beam is its weight (Beer and Johnston, 1992). The branch can also be assumed as a beam due to the notion that the length is considerably longer than its thickness. The following formula depicts how bending moment is calculated: M = F L, where M is the bending moment; F is the load being applied and L is length. To calculate bending moment for the branches, it is important to take into consideration the stems from which the branches diverge out of. The stems, called nodes, contribute to the bending moment of the branches and themselves have a weight to be accounted for. To take this into account, the length of the node from one branch to next is considered and its weight measured – its weight is considered to act at the center point. The branch sprouts from the end of each node i.e., if a branch grows at point A and the next branch at point B, the node length is from A to B and the weight of the branch at A is added to weight of node A to B. The bending moment value for each preceding node is added to the following one as the weight is distributed along the entire tree ∗

Research mentored by Lance Evans, Ph.D.

The Manhattan Scientist, Series B, Volume 6 (2019)

Hamid

branch from tip to base. Thus, the formula of bending moment can be expanded to: W n Ln + Wb Ln , 2 where Wn is the weight of the node, Ln is the length of node and Wb is the weight of the branch. M=

The second engineering property to be analyzed is the section modulus (Beer and Johnston, 1992). This is a geometric property of the cross-sectional area of the beam to be analyzed. For the purpose of this research, the branch is considered a beam and the cross-section is considered to that of a circle. The branches are assumed to be of cylindrical shape and have a uniform diameter throughout its entirety. The section modulus is calculated with the equation as depicted below: πD3 , 32 where I is the section modulus and D is the diameter of the node. I=

The mechanical stress can then be obtained by noting that it is simply the ratio of bending moment and section modulus (Beer and Johnston, 1992). By plotting the values of bending moment against section modulus, the slope gives the average value of mechanical stress for the specific branch being analyzed. The hypothesis of this study is that all tree branches (large to small) of all tree species will have the same bending stresses (Shahbazi, et al.,2015). Data of only two species are shown here. The ongoing project will analyze data from forty tree species. The formula below shows how mechanical stress is calculated: Mc σ= , I where σ is the mechanical stress; M is the bending moment; c is the distance from the neutral axis to the edge of the cross-section and I is the section modulus.

Materials and Methods Data of primary (main) and secondary branches of Betula nigra and Catalpa speciosa were analyzed. Tree branches sampled were within 15% of parallel to the ground allowing for simplification of the calculations for bending stresses. Branches were cut from trees on the Manhattan College campus (Bronx, New York 10471). Leaves were removed so that the dimensions of all stems could be measured. Photographs were taken of all stem sections from node to node along the main stems and all branches from the main stem. Photographs were downloaded into computer files. Each photograph was imported into ImageJ (imagej.nih.gov/ij) and diameters were obtained. Stem segments were weighed with a mass balance to the nearest 0.01 g - all stem segment lengths were determined from ImageJ. An Excel file (www.microsoft.com) was constructed to calculate the bending moment and section modulus of each branch segment. The components of all stem segments were constructed in an Excel file. A plot of bending moment on the y-axis and sectional modulus on the x-axis gave an estimate of the mechanical stress, σ.

Hamid

The Manhattan Scientist, Series B, Volume 6 (2019)

Results The hypothesis of this study is that all tree branches (large to small) of all tree species will have concordant bending stresses (Shahbazi et al., 2015) - data of only two species are shown here. Data of Betula nigra and Catalpa speciosa show that mechanical stresses of main stems were similar (Figs. 1 and 2). The slopes of Betula nigra and Catalpa speciosa were 0.0039 and 0.0040 Pa, respectively. The results support the hypothesis that main stem branch stresses were similar. The subservient hypothesis states that mechanical stresses are constant from branch bases to branch tips of main stems. The results for B. nigra were poor since the R2 value was 0.75 (Fig. 1) but the results for C. speciosa were exceptional since the R2 value was 0.95 (Fig. 2).

Figure 1. Relationship to determine bending stress for primary branches of Betula nigra.

Figure 2. Relationship to determine bending stress for primary branches of Catalpa speciosa.

Data for B. nigra demonstrated a strong relation in slopes for primary (0.0039; Fig. 1) and secondary (0.0036; Fig. 3) branches. Data for the C. speciosa species demonstrated a strong relation in slopes for primary (0.0040; Fig. 1) and secondary (0.0035; Fig. 4) branches. For the limited number of comparisons, all branches had similar slope values.

Figure 3. Relationship to determine bending stress for secondary branches of Betula nigra.

Figure 4. Relationship to determine bending stress for secondary branches of Catalpa speciosa.

The Manhattan Scientist, Series B, Volume 6 (2019)

Hamid

Discussion Upon investigation of the reiterative processes within the primary and secondary branches of both species – Betula nigra and Catalpa speciosa – the parallels between the two are deemed to be substantial. Data of Betula nigra and Catalpa speciosa demonstrate strong correlations in all secondary branches exhibiting the notion of constant mechanical stresses ensuring structural rigidity and integrity.

Bending Moment (N-m)

A brief examination of primary, secondary and tertiary branches of Prunus serrulata (Fig. 5.) shows that these branches all had slopes of between 0.0027 and 0.0035. The R2 values for the three slopes were 0.92, 0.98 and 0.99 which show that each branch had constant stresses. From the 5. few samplesFig. analyzed, the results are encouraging. The ongoing objective is to examine branches from 40 tree species. Prunus - Primary, Secondary and Tertiary Branches

y = 0.0027 x - 0.0235, R² = 0.9906 y = 0.0035 x - 0.0926, R² = 0.9132 y = 0.0031 x - 0.0293, R² = 0.9778

5 4 3 2 1 0 0

500 1000 Section Modulus (10-9 m3)

1500

Figure 5. Relationship to determine bending stress for primary, secondary and tertiary branches of Prunus serrulata

Acknowledgement This research was funded by the Catherine and Robert Fenton Endowed Chair to Dr. Lance S. Evans.

References Beer F. P. and Johnston Jr., E. R. (1992). Mechanics of materials. 2nd Edition. McGraw-Hill, Inc. New York Shahbazi, Z., Kaminski, A.,and Evans, L. (2015) “Mechanical stress analysis of tree branches.” American Journal of Mechanical Engineering, 3(2) 32-40. doi: 10.12691/ajme-3-2-1. Wilson B., Archer R. (1977) “Reaction wood; Induction and mechanical action.” Ann. Rev. Plant Physiol. 28:23-43

Quantification of eccentric growth in stems of Purshia tridentata Zemima Khasrooâ&#x2C6;&#x2014; Laboratory of Plant Morphogenesis, Department of Biology, Manhattan College Abstract. The purpose of this research was to determine the frequencies of eccentric growth in stems of Purshia tridentata in order to get a better understanding of their uneven growth patterns. Stems of P. tridentata were obtained from the plants grown in Central Utah. Two of the stems (536 mm and 616 mm) were cut and analyzed for this research. Individual stem samples were photographed and their xylary rings were traced in Microsoft paint. Additionally, 10 transect lines were drawn that were each 36 degrees apart from each other. This allowed to measure and see the changes along each sector. Each xylary ring area was measured in a computer program known as ImageJ and the measurements were recorded in Microsoft excel spreadsheet where % eccentricities were calculated. The results for the first hypothesis which compared two adjacent sectors of the same wood sample, showed that eccentricity was localized in the first ring and the outer rings. For the second hypothesis, which compared two sectors of consecutive wood samples, also showed that eccentricity occurred for rings 1 and the outer rings. Overall, the preliminary data shown in this research determined that eccentricity occurred during the first year of growth as well as in the outer rings.

Introduction The development of wood in plants was a major episode in plant evolution (Schweingruber et al., 2008). Primitive plants that did not have wood were small and low growing. The formation of wood in ferns and conifers allowed these species to be the dominant plant groups during the Jurassic Period (Schweingruber et al., 2008). Wood holds stems upright, supports branches to hold leaf canopies, transport water and minerals to all plant parts within vessels and tracheids (Mauseth, 2014; Hopkins and Huner, 2004). Wood or secondary xylem is produced by the vascular cambium. The vascular cambium is initially produced as parenchyma cells (a primary tissue). Later these cells become meristematic and becomes the vascular cambium, which is a layer of cells between the primary xylem and primary phloem (Taylor, 2009). The vascular cambium becomes a secondary tissue and produces both secondary xylem (wood) and secondary phloem (Mauseth, 2014; Hopkins and Huner, 2004). Plants that grow in the arid regions of the world are particularly vulnerable to growth abnormalities (Evans et al., 2019). Periodic and/or sustained droughts affect plant growth of perennial plants that must withstand the rigors of extended drought (MacMahon, 1985). Annuals and biennials may be less affected. Big sagebrush (Artemisia tridentata) is an example of a perennial that must survive the harsh winters with high snow loads in the Great Basin Desert of North America. Big sagebrush is vulnerable to uneven growth patterns of xylary rings called its eccentric growth (Diettert, 1938). Eccentric growth is defined as uneven growth of tissues in localized areas because of the death of the vascular cambium (Evans et al., 2019). Eccentric growth can hence cause structurally weakened stems as well as death of stem terminals ( Scarinci et al., 2017). This eccentric â&#x2C6;&#x2014;

Research mentored by Lance Evans, Ph.D.

The Manhattan Scientist, Series B, Volume 6 (2019)

Khasroo

growth with poor wood development may be the reason for sagebrush to be a shrub and not growth tree-like (Evans et al., 2019) . In preliminary observations, plants of Purshia tridentata (Fig. 1) shows eccentric growth (Fig. 2). The purpose of the current study is to describe this eccentric growth and compare the characteristics of eccentric growth with those of A. tridentata.

Figure 1. Picture of Purshia tridentata showing its uneven growth morphology (Bilbrough and Richards, 1991).

Figure 2. Eccentric growth of a sample of stem in Purshia tridentata.

Khasroo

The Manhattan Scientist, Series B, Volume 6 (2019)

The overall goal of this research is to determine amounts of eccentric growth in stems of P. tridentata. The first hypothesis is that eccentric growth occurs between adjacent stem sectors within stem samples. The second hypothesis is that eccentric growth occurs among consecutive stem samples. This report will provide some examples. The study is ongoing, so a detailed analysis will not be presented.

Materials and Methods Stem samples of Purshia tridentata were obtained from the plants grown in Central Utah. Two stems of about 536 mm and 616 mm long were selected for this study. Most of the bark was removed. A straight black line was drawn along the length of the stem with a sharpie (Fig.3). This straight line provided a guide to maintain the orientation among consecutive stem samples

Figure 3. Stem of Purshia tridentata with a black line along the stem.

once stem samples were cut. Eight mm long samples were cut along stems with a pull saw (Stanley Fatmax Flush Cut Pull Saw; (www.Stanleytools.com). Water was added to the cut surfaces to provide a good contrast to view the xylary rings. A Canon PowerShot ELPH100HS camera (www.canon.com) was used to obtain photographs of wood surfaces. A millimeter ruler was included in each photograph to ensure accurate distance measurements (Evans et al., 2019). Images of wood surfaces were uploaded in Microsoft Paint Program (www.microsoft.com) on the computer. Each xylem ring was traced and transect lines were drawn from the stem center to the outer edge of the sample at 36-degree intervals (0◦ , 36◦ , 72◦ , 108◦ , 144◦ , 180◦ , 216◦ , 252◦ , 288◦ , and 324◦ ) (Evans et al., 2019). Each wood sample had ten sectors of equal radial proportions around the center shown in Fig. 4 (Evans et al., 2019).

Figure 4. Xylary rings (black lines) and sector lines (red) of a stem segment of Purshia tridentata. The black line along the intact stem is shown at the 0◦ location on the segment. All lines and degree designations were drawn in Microsoft paint.

The Manhattan Scientist, Series B, Volume 6 (2019)

Khasroo

To determine areas of each sector in each ring in each wood sample, images were uploaded into ImageJ (imagej.nih.gov/ij). Areas of each sector, in all rings in all wood samples were entered into Microsoft excel (www.microsoft.com) spreadsheet and the % eccentricities were calculated. Eccentricity is a measure of variability among the samples, and were determined by dividing the standard deviation by the mean and multiplying by 100% (Snedecor and Cochran, 1967). An arbitrary threshold greater than 25% was determined to be eccentric (Love et al., 2009, Evans et al., 2012).

Results The first hypothesis is that eccentric growth occurs between adjacent stem sectors within stem samples. Two examples were shown. A comparison of two sectors in wood sample #7 of stem #1 showed eccentricity values of 41.1, 38.4, 57.6 and 25.8% for rings 1, 6, 7 and 8, respectively (Fig. 5). A comparison of two sectors in wood sample #9 of stem #2 showed eccentricity values of 47.1 and 28.7% for rings 1 and 7, respectively (Fig. 6). The other rings showed normal growth.

Figure 5. Image of two adjacent sectors in a wood sample of Purshia tridentata. Values in white font are areas of sectors in rings. The violet font values indicate the eccentric percentages of each ring. Sectors 4 and 5 of wood sample #7 of stem #1 are shown.

Figure 6. Image of two adjacent sectors in a wood sample of Purshia tridentata. Values in white font are areas of sectors in rings. The violet font values indicate the eccentric percentages of each ring. Sectors 4 and 5 of wood sample #9 of stem #2 are shown.

The second hypothesis is that eccentric growth occurs among consecutive stem samples that were 8 mm apart. Two examples were shown. A comparison of two sector #4 from wood samples

Khasroo

The Manhattan Scientist, Series B, Volume 6 (2019)

#7 and #8 of stem #1 showed eccentricity values of 66, 90, 98, 170 and 105% for rings 2, 5, 6, 7 and 8, respectively (Fig. 7). A comparison of two sector #4 from wood samples #8 and #9 of stem #2 showed eccentricity values of 80, 31, 33, 41, 30 and 221% for rings 2, 3, 4, 5, 6, and 7, respectively (Fig. 8). The other rings showed normal growth.

Figure 7. Image of two sectors in two wood samples of Purshia tridentata. Values in white font are areas of sectors in rings. The green font values indicate the eccentric percentages of each ring. Sector 4 of wood samples #7 and #8 of stem #1 are shown. Wood samples are 8 mm apart.

Figure 8. Image of two sectors in two wood samples of Purshia tridentata. Values in white font are areas of sectors in rings. The green font values indicate the eccentric percentages of each ring. Sector 4 of wood samples #8 and #9 of stem #2 are shown. Wood samples are 8 mm apart.

Discussion The objective of this research was to determine the frequency of eccentric growth in stems of Purshia tridentata. From the results, it was successfully showed that (1) eccentricity is localized among two adjacent stem sectors within a stem sample, and (2) eccentric growth is localized within two consecutive 8 mm stem samples within a stem. For a comparison of two adjacent sectors of the same wood sample, eccentricity occurred for rings 1 and the outer rings. For a comparison of two sectors of the consecutive wood samples, eccentricity occurred for rings 1 and the outer rings. The patterns were similar for the two comparisons. Taken together for the limited data shown, it appears that eccentricity occurs in the first year of growth and also in the outer rings. Eccentricity in the first ring for P. tridentata is similar to that of A. tridentata. Eccentricity occurs in A. tridentata at the bases of determinate flowering branches (Evans et al., 2012). The wood samples from the current study did provide samples for first year growth to determine a similar phenomenon for P. tridentata. Eccentricity values were higher for comparisons for sectors of two consecutive wood samples

The Manhattan Scientist, Series B, Volume 6 (2019)

Khasroo

than for two adjacent sectors of the same wood sample. This result is expected since the wood rings from consecutive wood samples are farther away (8 mm) compared with adjacent sectors of the same wood sample. Analyses of additional samples will determine if these results are representative of stems of P. tridentata.

Acknowledgements This research was funded by the Catherine and Robert Fenton Endowed Chair to Dr. Lance S. Evans. The author is indebted to Kody Rominger for stem samples of P. tridentata.

References Bilbrough, C.J., and R.H. Richards (1991). Branch architecture of sagebrush and bitterbrush: use of a branch complex to describe and compare patterns of growth. Canadian Journal of Botany 69:1288–1295. Diettert, R.A. 1938. The morphology of Artemisia tridentata Nutt. Lloydia 1:3-74 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. Evans, L. S., Kharran, T. A., Pena, I., and Kitchen, S. G. (2019). Quantification of eccentricity in stems of Artemisia tridentata Nutt. https://scholarsarchive.byu.edu/wnan/vol79/iss3/12/. Hopkins, W.G..Huner, N.P.A. (2004). Introduction to Plant Physiology 3rd Edition. John Wiley and Sons, inc. New York. Love, J.S., J. Borklund, M. Vahala, J. Hertzberg, J. Kangasjarvi, and B. Sunberg (2009). Ethylene is an endogenous stimulator of cell division in Populus. Proceedings of the National Academy of Sciences of the United States of America 106:5984–5989 MacMahon, J. (1985). The Audubon Society nature guides: Deserts. New York: Knopf. Mauseth, J.D.( 2014). Botany An Introduction to Plant Biology 5th Edition. Jones Bartlett Learning, LLC. Burlington, MA. Schweingruber, F. H., Bor̈ner A., and Schulze, E.-D. (2008). Atlas of woody plant stems: evolution, structure, and environmental modifications. Berlin: Springer. Scarinci, M., Encarnacion, K., Pineda, A. R., and Evans, L. S. (2017). Visualization of Xylary Rings of Stems of Artemisia tridentata spp. Wyomingensis. Universal Journal of Applied Mathematics, 5 (2), 28-33. doi: 10.13189/ujam.2017.050203 Snedecor, G. W., and Cochran, W. G. (1967). Statistical methods: 6th ed. Ames: Iowa State University Press. Taylor, T. (2009). Introduction to Paleobotany, How Fossil Plants are Formed. Biology and Evolution of Fossil Plants, 1-42. doi: 10.1016/b978-0-12-373972-8.00001-2

Molecular analysis and prevalence of Giardia lamblia in Geukensia demissa collected from two beaches in New York City Jailinne Lopezâ&#x2C6;&#x2014; Department of Biology, Manhattan College Abstract. Giardia lamblia is an intestinal parasite that causes giardiasis in humans. The goal of this research was to determine the prevalence of G. lamblia in the Geukensia demissa mussel species from both Orchard Beach and Clason Point in New York City. Specimens were collected in the fall of 2018 and each specimen was dissected to harvest the mantle, gills, foot, digestive gland, abductor muscle, and hemolymph. A total of 20 specimens were collected from Clason Point and a total of 42 specimens were collected from Orchard Beach. The prevalence of G. lamblia for specimens from Clason Point was 55% and 4.80% from Orchard Beach. The G. lamblia parasite was found in all of the tissues except for the foot and hemolymph. Further, the mantle tissue had the highest prevalence of G. lamblia for Clason Point (20%), while the digestive gland tissue had the highest prevalence for Orchard Beach (2.4%). Following Sanger sequencing and BLAST analysis of the positive samples, the following genotypes of G. lamblia were found: 14JBG, W3, B46, A44, 81f and VANC/96/UBC/126. Only the 14JBG genotype was found in Orchard Beach and Clason Point, all other genotypes were found in Clason Point only.

Introduction Bioindicators are species used to detect pollutants and pathogens in their surrounding environment (Adell et al., 2014). Specifically in marine environments, bioindicators are used to provide information on contaminants such as heavy metals (Hu et al., 2019). Some marine biondicators include fish and aquatic plants (Hu et al., 2019). Bionindicators can also be used for the detection of certain viruses and parasites (Miller et al., 2005). Bivalves are often used as bio-indicators of their surrounding environment due to their stationary nature and expansive dispersal across geographic regions. In prior years, bivalves have been used to detect aquatic contamination of pesticides and heavy metals (Miller et al., 2005; Oâ&#x20AC;&#x2122;Connor et al., 2002). However, in more recent years, bivalves have been used to detect bacteria, parasites and viruses (Miller et al., 2005). Since these bivalves are filter feeders, they are more susceptible to harboring waterborne contaminants (Miller et al., 2005). These bivalve mollusks are of particular interest because oocysts of G. lamblia can be detected in them before they can be detected in the surrounding marine environment (Miller et al., 2005). Previous studies have shown that bivalves such as Geukensia demissa, Mya areneria, Crassostrea virginica, and Mytillus edulis can in fact be used as bioindicators for several intestinal parasites, specifically, G. lamblia (Tei et al., 2016). Giardia lamblia is a single celled flagellated protozoan parasite that exists in two forms, cyst and trophozoite (Wolfe, 1992). The cyst form of the parasite is ingested with contaminated food or water and can survive in water for up to three months. This specific protozoan uses fecal-oral contact as its main mode of transmission and infects humans through sewage and contamination of â&#x2C6;&#x2014;

Research mentored by Ghislaine Mayer, Ph.D.

The Manhattan Scientist, Series B, Volume 6 (2019)

Lopez

land and rivers by animal or human feces (Baldursson et al., 2011). When the G. lamblia cysts are ingested, they become trophozoites in the duodenum and trophozoites then become cysts as they transit the small intestine and are excreted in feces in order to infect a new host (Wolfe, 1992). G. lamblia is the most common intestinal parasite that affects the United States and is linked to many waterborne diarrhea outbreaks as well as giardiasis (Adam et al., 2001). It is estimated that there are 1.2 million annual cases of giardiasis (Painter et al., 2015). Further, more frequent diagnosis of giardiasis is seen in the northern part of the US and mostly affects small children (Painter et al., 2015). Transmission of G. lamblia is usually accomplished through direct fecal-oral contact (Fig. 1). There are several assemblages of G. lamblia, categorized A-H according to which organism they infect (Fantinatti et al., 2016). Assemblages A and B infect humans, dogs, and cats, assemblages C and D commonly infect dogs, while assemblage E is found in herd animals, assemblage F in cats, assemblage G in rats and mice and assemblage H in seals (Fantinatti et al., 2016).

Figure 1. Life cycle of G. lamblia. Adapted from Esch et al. (2013).

Orchard Beach is a public beach located in Bronx, New York and Clason Point is a peninsula that is located in the Soundview section of the Bronx, New York, which includes both the Bronx and East River. The overarching goal of this study is to determine the prevalence of G. lamblia in Geukensia demissa from both Orchard Beach and Clason Point using molecular analysis techniques.

Materials and Methods Collection sites Geukensia demissa is a ribbed mussel that is commonly found in Atlantic salt marshes at about mid-tide and filter feeds as its main mechanism of nutrition, thus making it susceptible to a wide range of pollutants and parasites. Sixty-two samples of Geukensia demissa were collected from both Orchard Beach and Clason Point on September 24, 2018. Clason Point is a peninsula located

Lopez

The Manhattan Scientist, Series B, Volume 6 (2019)

in the Soundview section of the Bronx in New York while Orchard Beach is a public beach located in the Bronx as well that empties into the Long Island Sound. Samples were collected at low tide on September 24, 2018, and after collection were stored in a freezer at -80◦ C. Samples collected included Geukensia demissa, Mytillus edulis, Crassostrea virginica (oysters), and Mya areneria.

Figure 2. Map of collection sites (Clason Point and Orchard Beach). https://www.google.com/maps/search/orchard+beach+ and+clason+point/@40.83199,73.8625798,13.18z

DNA extraction and molecular analysis Each specimen was thawed and separated into the following tissues: mantle, gills, abductor muscle, hemolymph, digestive gland, and foot. DNA was extracted from each of these tissues using the Qiagen DNeasy blood and tissue extraction kit following the manufacturer’s protocol (Germantown, MD, USA). A total of 20 samples from Clason Point and a total of 42 samples from Orchard Beach were analyzed. Following DNA extraction, Polymerase Chain Reaction (PCR) was performed targeting the β-giardin gene, unique to G. lamblia DNA. For the first step of the PCR reaction, the forward primer, (Gia7F) was 50 -AAGCCCGACGACCTCACCCGCAGTGC-30 and the reverse primer Gia7R was 50 -GAGGCCGCCCTGGATCTTCGAGACGAC-30 (Piekarska et al., 2016). For the second step, amplification was performed with the Gia7 Nested Forward primer (50 -GAACGAACGAGATCGAGGTCCG-30 ) and with the Gia759 Nested Reverse Primer (50 CTCGACGAGCTTCGTGTT-30 (Piekarska et al., 2016). Purified G. lamblia DNA was used as the positive control and water was used as the negative control. The PCR products were visualized using agarose gel electrophoresis stained with ethidium bromide and a UV light source. Positive samples were sent out for Sanger sequencing.

Results The prevalence of G. lamblia in G. demissa collected in 2018 from Clason Point and Orchard Beach was determined for each organism and each tissue. A total of 11/20 (55%) samples from Clason Point (Fig. 3) and a total of 2/42 (4.8%) from Orchard Beach (Fig. 4) tested positive for G. lamblia DNA. Samples from Clason Point had a higher prevalence, which suggests more contamination in this site.

100

The Manhattan Scientist, Series B, Volume 6 (2019)

Lopez

Figure 4. Representative agarose gel electrophoresis showing positive sample for G. lamblia DNA from Orchard Beach. Top Lane 1: 100 bp marker, 3: + control, 4: 11D, 20: - control.

Figure 3. Representative agarose gel electrophoresis showing positive samples for G. lamblia DNA from Clason Point. Top Lane 1: 100bp marker, 3: + control, 6: 13A; 7: 14A; 12: 17M, 13: 16D; 16: 9M; 18: 6G; 20: - control. Bottom Lane 3: 7D; 5: 2G; 6: 8D; 11:4M

The prevalence of G. lamblia DNA was also found for each individual tissue in each of the samples collected. For Clason Point, a prevalence of 10.5% was observed in the gills, 11.1% in the abductor muscle, followed by 15% in the digestive gland, and 20% in the mantle. The mantle of G. demissa was the tissue with the highest prevalence in this site. On the other hand, Giardia was not detected in the foot and hemolymph of G. demissa for Clason Point. In contrast to Clason Point, the prevalence of Giardia in the gills and digestive glands was 2.4%. Giardia was not detected in the abductor muscle, foot, mantle, and hemolymph. In summary, difference in the tissue distribution of Giardia in G. demissa was observed between the two collection sites. Table 1. Prevalence of G. lamblia in Geukensia demissa from Clason Point, and Orchard Beach in 2018 Clason Point

Orchard Beach

11/20 (55%)

2/42 (4.8%)

Table 2. Prevalence of G. lamblia in tissue from G. demissa collected from Clason Point and Orchard Beach New York in 2018

Clason Point Orchard Beach

Abductor Muscle

Digestive Gland

Foot

Mantle

Gills

2/18 11.1% 0/41 0%

3/20 15% 1/41 2.43%

0/1 0% 0 /13 0%

4/20 20% 0/42 0%

2/19 10.5% 1/42 2.38%

Hemolymph 0/5 0% 0/4 0%

Lopez

The Manhattan Scientist, Series B, Volume 6 (2019)

101

Sanger sequencing of the positive samples confirmed G. lamblia in all positives. Further, the only genotype found in both Clason Point and Orchard Beach was 14JBG. Genotypes W3, B46, A44, 81f, and VANC/96/UBC/126 were found in Clason Point (Table 3). Table 3. G. lamblia genotype distribution for Geukensia demissa collected from Clason Point and Orchard Beach. Sample ID

Genotype

2G Clason 4M Clason 6G Clason 8D Clason 9M Clason 11D Orchard 12M Clason 13A Clason 14A Clason 16D Clason 17M Clason 30G

14 JBG 14JBG W3 B46 A44 14JBG 81f 81f 81f VANC/96/UBC/126 VANC/96/UBC/126

Assemblage Assemblage Assemblage Assemblage Assemblage Assemblage Assemblage Assemblage Assemblage Assemblage Assemblage Unknown

A (AI) A (AI) A A A (A4) A (AI) A (AII) A (AII) A (AII) A A

Table 4. G. lamblia tissue distribution for Geukensia demissa collected from Clason Point and Orchard Beach in 2018 Isolates & Assemblages Clason Point Orchard Beach

14JBG A (AI) 2/11 18.2% 1/2 50%

W3 A 1/11 9.1%

B46 A 1/11 9.1%

A44 A (A4) 1/11 9.1%

81f A (AII) 3/11 27.3%

VANC/96/UBC/126 A 2/11 18.2%

According to the data collected, all positive samples belonged to assemblage A, which affects humans, cats and dogs (Table 4). In addition to the eight assemblages, there are sub-assemblages AI, AII, and AIII, which infect various hosts. Sub-assemblage AI affects various mammals from humans to cats to dolphins, while sub-assemblage AII mainly affects humans and sub-assemblage AIII affects deer (Feng et al., 2011). It was found that for the Clason Point site, 18.2% of positive Geukensia demissa samples belonged to the AI sub-assemblage, 27.3% belonged to the AII subassemblage, and 9.1% belonged to the A4 assemblage (Table 4). This indicates that the probable source for this site was due to sewage overflow. On the other hand, for Orchard Beach positive Geukensia demissa samples belonged to the AI sub-assemblage, which mainly affects dogs and cats. This indicated that the probable source of contamination for this site was the improper handling of waste, especially coming from people walking their dogs. Results after performing a BLAST (Basic Local Alignment Search Tool) analysis showed that several genotypes of positive samples were found across the world. The 14JBG isolate was found and isolated in Sao Paolo, Brazil by University of Campinas and the 81f isolate was found in sheep

102

The Manhattan Scientist, Series B, Volume 6 (2019)

Lopez

feces in Greece (Ligda et al., 2017). Pursuing this further, other genotypes such W3, it was found in genomic DNA at the Medical Parasitology department in Cairo University. Genotype B46 was found in sheep feces from China at the College of veterinary medicine in Sichuan and the genotype for isolate A44 was found in calf DNA from Rome (Lalle et al., 2005).

Discussion The prevalence of G. lamblia in G. demissa was 55% for Clason Point and 4.8% for Orchard Beach. For Clason Point, a prevalence of 10.5% was seen in the gills, 11.1% in the abductor muscle, 20% in the mantle, and 15% in the digestive gland. Giardia was not detected in the foot and hemolymph for G. demissa collected from Clason Point. On the other hand, for Orchard Beach, a prevalence of 2.4% was seen in the gills and digestive gland tissue. However, no Giardia was detected in the abductor muscle, foot, mantle, and hemolymph. These differences in prevalence between tissues of different sites could be attributed to the fact that the source of contamination is different for each location. Genotypes also varied by location, however genotype 14JBG was found at both Clason Point and Orchard Beach. All other genotypes which were W3, B46, A44, 81f, and VANC/96/UBC/126 were found at the Clason Point location. There was a higher diversity of G. lamblia at Clason Point than at Orchard Beach. From 2016 to 2018, the prevalence of G. lamblia in G. demissa collected from Orchard Beach increased by 1.5% from 2016 to 2017 from 28% to 42% and then had a drastic reduction from 42% to 4.8% from 2017 to 2018. This reduction might be due to the close monitoring and more frequent testing performed in Orchard Beach. Further, for G. demissa collected from Clason Point, there was a steady increase year to year. The prevalence was 19% in 2016, 36.1% in 2017, and 55% in 2018. This steady increase is due to the fact that Clason Point is not as frequently tested for pollutants since it is not known for swimming or fishing. The presence of G. lamblia in G. demissa collected from these public New York waterways is an indication that there are G. lamblia oocysts which might be due to from sewage contaminating the water as well as improper management of waste.

Conclusions The above research has shown the prevalence of Giardia lamblia in Geukensia demissa collected from Clason Point and Orchard Beach in New York City in fall 2018. The stated results have an impact both on public health as well as water quality in the surrounding environment. The observed similar genotypes in the different parts of the world can be attributed to the fact that both collection sites experience a vast majority of sewage overflow and are surrounded by many wastewater plants. Combined Sewage Overflows (CSOs) occur when excess wastewater accumulates and flows into New York City waterways (Levine, 2019). This happens quite often in NYC since both rainwater and wastewater accumulate in the same pipes. According to the New York State Department

Lopez

The Manhattan Scientist, Series B, Volume 6 (2019)

103

of Environmental Conservation, 2018 saw CSO events once every three days, which was a 44% increase from 2016. Pursuing this further, 2018 was a year with record high precipitation for New York City, with the annual rainfall accumulating to 65.55 inches, thus accounting for the increase in Combined Sewage Overflow events. This might account for the increased prevalence of G. lamblia found in the samples collected from both Clason Point. Some next steps to consider with this research include advocating for better water treatment plans for New York City. Further, to advocate for more frequent water testing for all waterways, not just the ones the public frequents swimming in. Change is especially imminent since under the Clean Water Act, United States waters should be fishable and swimmable and free of any pollutant that may cause illness. The fact that the data presented indicate that both Clason Point and Orchard Beach are polluted, is an indication that these CSO events are an issue and the wastewater treatments in New York should be addressed, perhaps by creating an excess sewage storage tunnel, to hold the excess wastewater until it can be redirected into the wastewater treatment plan (Chaisson, 2019).

Acknowledgment This work was financially supported by the Linda and Dennis Fenton ’73 endowed biology research fund. The author would like to thank Dr. Ghislaine Mayer for giving her the opportunity to work on this research.

References Adam, R. D. “Biology of Giardia lamblia.” Clinical Microbiology Reviews, vol. 14, no. 3, 2001, pp. 447–475., doi:10.1128/cmr.14.3.447-475.2001. Adell, A. D., Smith, W. A., Shapiro, K., Melli, A., Conrad, P. A. “Molecular Epidemiology of Cryptosporidium Spp. and Giardia Spp. in Mussels (Mytilus Californianus) and California Sea Lions (Zalophus Californianus) from Central California.” Applied and Environmental Microbiology, vol. 80, no. 24, 2014, pp. 7732-740., doi:10.1128/aem.02922-14. Baldursson, S., and Karanis, P. “Waterborne Transmission of Protozoan Parasites: Review of Worldwide Outbreaks – An Update 2004–2010.” Water Research, vol. 45, no. 20, 2011, pp. 6603-6614., doi:10.1016/j.watres.2011.10.013. Chaisson, C. “When It Rains, It Pours Raw Sewage into New York City’s Waterways.” NRDC, 2019, www.nrdc.org/stories/when-it-rains-it-pours-raw-sewage-new-york-citys-waterways. Fantinatti, M., Bello, A.R. Fernandes, O., Da-Cruz, A. M. “Identification of Giardia lamblia Assemblage E in Humans Points to a New Anthropozoonotic Cycle.” The Journal of Infectious Diseases, Volume 214, Issue 8, 2016, Pages 1256-1259. doi: 10.1093/infdis/jiw361 Esch, K. J., and Petersen, C. A . “Transmission and Epidemiology of Zoonotic Protozoal Diseases of Companion Animals.” Clinical Microbiology Reviews, vol. 26, no. 1, 2013, pp. 58–85. doi:10.1128/cmr.00067-12.

104

The Manhattan Scientist, Series B, Volume 6 (2019)

Lopez

Feng, Y., and Xiao, L. “Zoonotic Potential and Molecular Epidemiology of Giardia Species and Giardiasis.” Clinical Microbiology Reviews, vol. 24, no. 1, Jan. 2011, pp. 110-140., doi:10.1128/cmr.00033-10. Hu, C., Yang, X., Gao, L., Zhang, P., Li, W., Dong, J., Li, C., Zhang, X. “Comparative Analysis of Heavy Metal Accumulation and Bioindication in Three Seagrasses: Which Species Is More Suitable as a Bioindicator?” Science of the total environment, vol. 669, 2019, pp. 41-48. doi:10.1016/j.scitotenv.2019.02.425. Lalle M., Pozio E., Capelli G., Bruschi F., Crotti D., Cacciò S.M. “Genetic Heterogeneity at the β-Giardin Locus among Human and Animal Isolates of Giardia Duodenalis and Identification of Potentially Zoonotic Subgenotypes.” International Journal for Parasitology, vol. 35, no. 2, 2005, pp. 207-213. doi: 10.1016/j.ijpara.2004.10.022. Levine, L. “A Wet 2018 Saw Sharp Rise in NYC Sewage Dumping: 1 in 3 Days.” NRDC, 15 Apr. 2019, www.nrdc.org/experts/larry-levine/wet-2018-saw-sharp-rise-nyc-sewage-dumping1-3-days Ligda, P., Claerebout, E., Zdragas, A., Casaert, S. and Sotiraki, S. “Cryptosporidium and Giardia in different water catchments within a high dense farming area in Greece.” 2017. https://academic.oup.com/jid/article/214/8/1256/2388198 Miller, W. A., Atwill., Edward R., Gardner, I. A., Miller, M. A., Fritz, H. M., Hedrick, R. P., Melli, A. C., Barnes, N. M., Conrad, P. A. “Clams (Corbicula Fluminea) as Bioindicators of Fecal Contamination with Cryptosporidium and Giardia Spp. in Freshwater Ecosystems in California.” International Journal for Parasitology, vol. 35, no. 6, 2005, pp. 673-684. doi:10.1016/j.ijpara.2005.01.002. O’Connor, T. P. “National Distribution of Chemical Concentrations in Mussels and Oysters in the USA.” Marine Environmental Research, vol. 53, no. 2, 2002, pp. 117–143., doi: 10.1016/s01411136(01)00116-7. Painter, J., Gargano, J. W., Collier, S. A., Yoder, J. S. “Giardiasis Surveillance - United States, 2011-2012. “Centers for Disease Control and Prevention, 2015, www.cdc.gov/mmwr/preview /mmwrhtml/ss6403a2.htm. Piekarska J., Bajzert J., Gorczykowski M., Kantyka M., Podkowik M. “Molecular Identification of Giardia Duodenalis Isolates from Domestic Dogs and Cats in Wroclaw, Poland.” Annals of Agricultural and Environmental Medicine, vol. 23, no. 3, 2016, pp. 410-415. doi:10.5604/12321966.1219178. Tei, F. F., Kowalyk, S., Reid, J. A., Presta, M. A., Yesudas, R. and Mayer, D.C.G. “Assessment and Molecular Characterization of Human Intestinal Parasites in Bivalves from Orchard Beach, NY, USA.” International Journal of Environmental Research and Public Health, vol. 13, no. 4, 2016, p. 381. doi:10.3390/ijerph13040381. Wolfe, M. S. “Giardiasis.” Clinical Microbiology Reviews, vol. 5, no. 1, 1992, pp. 93–100., doi:10.1128/cmr.5.1.93.

Mexican columnar cacti spines with regard to bark coverage Catherine Anne McDonough∗ Laboratory of Plant Morphogenesis, Department of Biology, Manhattan College

Abstract. Spines are a key feature found exclusively on cactus plants. They serve various functions such as defense from predators, temperature regulation and transpiration. Additionally, bark coverage is a phenomenon found on over twenty columnar cactus species in the Americas. This study observes the spines of five columnar cactus species of central Mexico (Cephalocereus columna-trajani, Neobuxbaumia macrocephala, N. mezcalaensis, N. tetetzo, and Pachycereus hollianus). Mature cacti (>4 m in height) showed fewer original spines than younger cactus plants (<2 m in height) When compared with increasing bark coverage, the number of mature original spines decreased. In three species, C. columna-trajani, N. macrocephala, and N. mezcalaensis, 75% of spines were absent when bark coverage was 80% or more. Each species, however, showed a unique relationship between bark coverage and numbers of spines. Additionally, these three species also produced hair-like spines as well as emergent spines. It was established that the occurrence of hair-like spines almost always coincided with emergent spines. However, there was no relationship between bark coverages and hair-like spines or emergent spines.

Introduction Spines are a feature nearly exclusive to cactus plants. Spines are modified leaves that extend from the areole, an axillary bud on the cactus surface (Gibson and Nobel 1986). As cells elongate and stretch, they are filled with lignin, giving rise to waterproof, stiff spines (Gibson and Nobel 1986). This rigidity gives spines their many functions. Some examples of these functions are protection from predators, shading cactus tissue from sunlight, temperature regulation, and transpiration. Unfortunately, these spines are often lost when the cactus surface decays (Evans and L’Abbate, 2018; Gibson and Nobel, 1986). Research has previously shown that over twenty species of columnar cacti throughout the Americas experience a phenomenon known as “bark” (Evans et al., 1994a; Evans et al., 1994b; Evans, 2005; Evans and Macri, 2008). Being that the bark had begun at the equatorial facing surfaces in these studies, it is thought that the bark is caused by increased sunlight exposure. The bark coverage then extends to the other surfaces of the cactus until it is completely covered in bark. Once bark coverage appears on a cactus surface, gas exchange is no longer possible in that tissue, and the tissue subsequently dies. A previous study negatively correlated the number of spines with bark coverage in Saguaro cactus plants (Evans and L’Abbate, 2018). This purpose of the current study was to (1) quantify the different types of spines, (2) quantify the loss of different types of spines in relation to bark coverage, and (3) investigate emergent spines. ∗

Research mentored by Lance Evans, Ph.D.

106

The Manhattan Scientist, Series B, Volume 6 (2019)

McDonough

Materials and Methods Field Conditions In May of 2018, data was collected for the following five species of columnar cacti: Cephalocereus columna-trajani (Karw. ex Pfeiff.): 36 plants; Neobuxbaumia macrocephala [(F.A.C.Weber ex K.Schum.) E.Y.Dawson]: 82 plants; N. mezcalaensis [(Bravo) Backeb.]: 101 plants; N. tetetzo [(F.A.C.Weber ex K.Schum.) Backeb.]: 74 plants; Pachycereus hollianus [(F.A.C.Weber) Buxb.]: 76 plants. Four species were located at the Tehuacán-Cuicatlán Biosphere Reserve, San Juan Raya (18.10◦ N, 97.21◦ W), Puebla, Mexico. This area is a protected reserve maintained by locals, minimizing human interference with the study. C. columna-trajani was located on a steep hill alongside Highway 125 (18.2◦ N, 97.3◦ W) where human access was difficult. Collection of data was not random in order to properly analyze the effect of sunlight exposure. Each cactus species was verified with The International Plant Name Index (www.ipni.org). All cacti sampled were at least 4 m in height and were free from surrounding vegetation. This ensured all cacti were adult plants and did not experience shading from sunlight. A few young plants (under 2 m) were observed to establish a complete arrangement of spines. Cacti have two main anatomical features: crests and troughs (Fig. 1). Crests are the protrusions met by two troughs along the cactus surface. Data and photographs were taken in an 8 cm section 1.7 m above ground for the crests nearest the cardinal directions: south, east, north and west, resulting in four data points. Bark coverage percentages were estimated visually to the nearest 5%, and the data replicated previous approximations from digital images (Evans and De Bonis, 2015, Evans et al., 1995; 2005; 2013). Figure 1. Images of the surfaces of columnar cactus species from relatively young and older stems. A-E. Young crests showing a full complement of spines. A: Cephalocereus columna-trajani. B. Neobuxbaumia macrocephala. C. Neobuxbaumia mezcalaensis. D. Neobuxbaumia tetetzo E. Pachycereus hollianus. F-J. Older crests showing fewer spines and various percentages of bark coverage. F: Cephalocereus columna-trajani. G. Neobuxbaumia. macrocephala. H. Neobuxbaumia mezcalaensis. I. Neobuxbaumia tetetzo. J. Pachycereus hollianus.

Evaluation of spines An optimum number of spines was established by counting the spines found on young cacti (< 2m). The spines counted on the young cacti were compared to those found in Benson (1982) and Anderson (2001), and they were found to be comparable. The optimum was compared to the number of spines found on older (>4 m) cacti to determine how well spines were retained with age and bark coverage.

McDonough

The Manhattan Scientist, Series B, Volume 6 (2019)

107

The number of original spines and emergent spines were compared to the crest bark coverage. Original spines were determined as the spines initially grown with the cactus. Emergent spines were spines that were regrown after original spines were broken or fell off. The two types of spines were also compared to each other in Table 1. Bark coverage was also compared with “fuzzy hair-like” spines. Hair-like spines were thin and small, and they appeared in clusters creating a “fuzzy” area (Evans and L’Abbate, 2018). Bark coverage was broken into four main categories: Class A (0-24% bark coverage), Class B (25-49% bark coverage), Class C (50-74% bark coverage), and Class D (75%-100% bark coverage). Table 1. Comparison of original and emergent spines for cactus species from central Mexico. Characteristic

Original

Emergent

Location

Not associated with thin hair-like spines

Associated with thin hair-like spines

Direction

Relatively constant

Random

Lengths

Fully elongated

Variety of lengths

Color

Mostly all the same Spines all the same age

Varies Some younger than others

Appearance

Many jagged and broke Dried out Dull

Not jagged or broken Not dried out Shiny

Curvatures

None

Occasionally

Darkened

Can be darkened but usually covered by hair-like spines

Areole

Statistical analyses Numbers of original spines within each class were compared to each other using pairwise T-tests (Snedecor and Cochran, 1967). Then, pairwise T-tests were also used to determine a relationship between hair-like spines and emergent spines. These tests were done for all species.

Results As previously stated, the optimum number of original spines was determined by examining young cacti <2 m in height. These numbers are comparable to those found by Anderson in 2001 (Table 2, Fig. 1). Table 3 provides detailed descriptions on the spines found on each species. On average, the number of original radial spines decreased by half. From Class A to Class D, the number of C. columna-trajani original radial spines decreased 50% (Table 4). For the same amount of bark, central spines reduced 75%. In the case of N. macrocephala, radial spines were reduced 90% from Class A to Class D. Furthermore, there were no central spines present on crests with 75% bark. For N. mezcalaensis, original radial spines decreased by half from Class A to Class D. Additionally for N. mezcalaensis, the number of central spines decreased by 75%. There was little to no effect on the spine counts of N. tetetzo and P. hollianus in regard to bark coverage.

108

The Manhattan Scientist, Series B, Volume 6 (2019)

McDonough

Table 2. Numbers of spines on columnar cactus species from central Mexico compared with those in published sources. Species

Source

Type of spine Radial

Central

Cephalocereus columna-trajani

Current Anderson

20-30 14-18

8-11 5-8

Neobuxbaumia macrocephala

Current Anderson

20-28 8-12

3 1-3

Neobuxbaumia mezcalaensis

Current Anderson

9 5-9

1 1-4

Neobuxbaumia tetetzo

Current Anderson

10-13 8-13

1-3 1

Pachycereus hollianus

Current Anderson

14 12-14

3-5 3-5

It was found that on four of the five species (C. columna-trajani, N. macrocephala, N. mezcalaensis, and N. tetetzo,) produce hair-like spines. For three of these species (C. columna-trajani, N. macrocephala, N. mezcalaensis), emergent spines were documented in data and photographs (Fig. 2). Almost all crests of C. columna-trajani, N. macrocephala, and N . mezcalaensis had at Table 3. Description of original spines on young, older cacti with no bark, older cacti with bark and emergent spines

Species / Young cacti characteristic Cephalocereus columna-trajani Reddish black centrals Color White radials Yes Swollen base Thick centrals Thickness Thin radials None Curvature Neobuxbaumia macrocephala Reddish black centrals Color Brown/Gray radials Slightly Swollen base Medium centrals Thickness Thin radials None Curvature Neobuxbaumia mezcalaensis Color Swollen base Thickness Curvature

Original spines Older no bark

Older bark present

Yes Thick centrals Thin radials None

Black with white base Patchy Yes Thick centrals Thin radials None

Gray/Black Patchy Slightly Medium centrals Thin radials None

White/Gray Patchy Slightly Medium centrals Thin radials None

Gray/Black

White with brown tips

White/Light gray

No Thin None

Gray Patchy No Thin None

Emergent spines

Black No Thin to medium None Reddish brown No Thin to medium Some curved White with brown tips Pale yellow w/ brown tips No Thin None

(continued on next page)

McDonough

The Manhattan Scientist, Series B, Volume 6 (2019)

109

Table 3. (continued from previous page) Original spines

Species/ characteristic Neobuxbaumia tetetzo

Young cacti

Thickness

Emergent spines

Black

Yes Thick centrals Medium radials None

Yes

Black w/ some white near base Patchy Yes

Medium Some slightly curved

N/A

Gray Yes Very thick centrals Thin radials None

N/A N/A

Color Swollen base

Older bark present

Older no bark

Curvature Pachycereus hollianus Color White with black base Yes Swollen base Very thick centrals Thickness Thin radials None Curvature

N/A N/A N/A

N/A N/A

Table 4. Numbers of original spines relative to bark coverage percentages on crests of stem surfaces of columnar cacti. (Within each species and within each category, values followed by a different letter are significantly different at p < 0.05 while values followed by the same letter are not significantly different.) 1 Number of cactus surfaces for sample. Species Cephalocereus columna-trajani

Neobuxbaumia macrocephala

Neobuxbaumia mezcalaensis

Neobuxbaumia tetetzo

Pachycereus hollianus

Class (Bark Coverage)

Number of crest samples1

Central

Radial

Young stems A (0 to 24%) B (25 to 49%) C (50 to 74%) D (75 to 100%) Young stems A (0 to 24%) B (25 to 49%) C (50 to 74%) D (75 to 100%) Young stems A (0 to 24%) B (25 to 49%) C (50 to 74%) D (75 to 100%) Young stems A (0 to 24%) B (25 to 49%) C (50 to 74%) D (75 to 100%) Young stems A (0 to 24%) B (25 to 49%) C (50 to 74%) D (75 to 100%)

– 136 43 27 30 – 174 38 32 44 – 182 22 48 156 – 136 44 27 29 – 158 20 28 98

8 to 11 2.22a 1.50a 1.67a 0.54b 3 0.80a 0.34b 0.10b 0.00b 1 0.82a 0.50b 0.22ab 0.18b 1 1.03 1.00 1.00 1.00 3 to 5 2.92 2.91 2.80 2.90

20 to 30 10.8a 5.90a 11.3a 4.38b 20 4.28a 1.21b 1.58b 0.41b 7 to 10 6.52a 3.47b 3.43b 3.11b 8 to 13 7.44a 6.09a 4.46b 5.25ab 14 11.10 11.09 9.74 9.86

Total 13.0a 7.4a 13.0a 4.88b 5.10a 1.55b 1.83b 0.41b 7.31a 4.97b 3.65b 3.28b 7.47a 7.09a 5.46b 6.25ab 15.9 15.8 14.31 14.58

110

The Manhattan Scientist, Series B, Volume 6 (2019)

McDonough

least one emergent spine when hair-like spines were present (Table 5). In contrast, when hair-like spines were not present, there was almost never emergent spines. It was also found that neither hair-like spines or emergent spines had any relationship with bark coverage in any species. Table 5. Numbers of emergent spines on crests of stems of columnar cactus species. Values for mean number of spines followed by a different letter are significantly different at p < 0.05. Hair-like spines Species

Present

Absent

Cephalocereus columna-trajani Mean number of spines1 Number of samples with no emergent spines Number of samples with one or more emergent spines Samples with no emergent spines (%) Samples with one or more emergent spines (%)

1.64a 29 61 32 68

0.40b 108 10 92 8

Neobuxbaumia macrocephala Mean number of spines Number of samples with no emergent spines Number of samples with one or more emergent spines Samples with no emergent spines (%) Samples with one or more emergent spines (%)

3.24a 6 31 16 84

0.31b 202 47 81 19

Neobuxbaumia mezcalaensis Mean number of spines Number of samples with no emergent spines Number of samples with one or more emergent spines Samples with no emergent spines (%) Samples with one or more emergent spines (%)

1.23a 86 100 46 54

0.24b 194 26 88 12

Figure 2. Images of surfaces of columnar cactus species with original and emergent spines present. A: Cephalocereus columna-trajani. Aa, Ae: Both of these areoles show old spines with one emergent spine on the right side of each areole. Ab, Ac, Ad: These areoles only contain old brown/gray spines. B. Neobuxbaumia macrocephala. All areoles show old spines and multiple emergent spines localized in the apical region of the areole. C. Neobuxbaumia mezcalaensis. Cc, Ce: These areoles showold spines and multiple emergent spines. Ca, Cd: These areoles show only old spines.

McDonough

The Manhattan Scientist, Series B, Volume 6 (2019)

111

Discussion The goal of this report was to compare the number of original and emergent spines in relation to cactus surface bark. The younger cacti (<2 m) always had more original spines than older cacti (>4 m). It has been previously published that spines are lost with age (Evans and Lâ&#x20AC;&#x2122;Abbate, 2018), but this is the first documentation giving reason for that. Data showed that bark coverage negatively correlates with the number of original spines in mature cacti. Three of the species have thinner bases at the spines (C. columna-trajani, N. macrocephala, and N. mezcalaensis), and in those species, they lose roughly 75% of original spines once bark coverage exceeds 80%. The other two species had swollen spine bases (N. tetetzo and P. hollianus), and only lose 10% of original spines at 80% or more bark coverage. This shows that the swollen bases of spines may aid in spine retention due to increased contact with the cactus tissue. On four of the five species (C. columna-trajani, N. macrocephala, N. mezcalaensis, N. tetetzo), hair-like spines were found on the older barked cacti. The hair-like regions were found primarily on the apical region of the areole. Therefore, the emergent spines were found also on the apical region of the areole. Hair-like spines have been noted previously (Gibson and Nobel, 1968; Evans and Lâ&#x20AC;&#x2122;Abbate, 2018), but not a single paper has documented their possible function. It is important to note that hair-like spines of this study were not the glaucids that are found on Cylindropuntia and Opuntia species because the hair-like spines are soft and will not puncture the skin (Gibson and Nobel, 1968; Benson, 1982; Anderson, 2001). Emergent spines were found on three species (C. columna-trajani, N. macrocephala, and N. mezcalaensis). On areoles with hair-like spines, the average/maximum numbers of emergent spines were 1.6/8, 3.2/12, and 1.2/17 for C. columna-trajani, N. macrocephala, N. mezcalaensis, respectively. These data were juxtaposed to those of areoles without the hair-like spines. Here, the average/maximum numbers of emergent spines were 0.4/5, 0.31/4 and 0.24/6 for C. columna-trajani, N. macrocephala, N. mezcalaensis respectively. Emergent spines were obviously distinct from original spines as they appeared different in color and texture (Table 1). Original spines grow in a set pattern repeated on all areoles. The emergent spines are randomly placed but within region of the hair-like spines. Emergent spines appeared the same color of the young original spines which differentiated them from the older and darker original spines. Additionally, emergent spines even appeared curved while original spines were typically straight. Lastly, emergent spines were shorter in length, suggesting they grew more recently than the original spines. This documentation is the first account of spines re-emerging on mature cactus plants.

Acknowledgments This work was supported by the Linda and Dennis Fenton â&#x20AC;&#x2122;73 endowed biology research fund. Additional support was provided by the Catherine and Robert Fenton Endowed Chair to Lance S. Evans. The author thanks Claudia Ramirez and Phillip Dombrovskiy for photography

112

The Manhattan Scientist, Series B, Volume 6 (2019)

McDonough

and data collection in the field. The author appreciates the personnel, particularly Felix Reyes, at the Tehuacán-Cuicatlán Biosphere Reserve, San Juan Raya, Puebla, Mexico for daily guidance, cooperation, assistance and encouragement for the procurement of data within the Reserve.

References Anderson E. F. 2001. The Cactus Family. Timber Press. Portland, OR. Benson L. 1982. The Cacti of the United States and Canada. Stanford Univ. Press. Stanford, CA. Evans L. S. 2005. Stem surface injuries to Neobuxbaumia tetetzo and Neobuxbaumia mezcalaensis of the Tehuacan Valley of Central Mexico. J. Torrey Bot. Soc. 132: 33–37. Evans L.S., De Bonis M. 2015. Predicting morbidity and mortality of saguaro cacti(Carnegiea gigantea). J. Torrey Bot. Soc. 142: 231–239. Evans L.S., L’Abbate 2018. Areole changes during bark formation on saguaro cacti. Haseltonia 24: 55-62. Evans L.S., Macri A. 2008. Stem surface injuries of several species of columnarcacti of Ecuador. J. Torrey Bot. Soc. 135: 475–482. Evans L.S., Cantarella V.A., Stolte K.W., Thompson K.H. 1994a. Phenologicalchanges associated with epidermal browning of saguaro cacti at Saguaro National Monument. Environ. Exp. Bot. 34: 9–17. Evans L.S., McKenna C., Ginochio R., Montenegro R., Keisling R. 1994b. Surfaceinjuries of several cacti of South America. Environ. Exp. Bot. 34: 285–292. Evans L.S., Sahi V., Ghersini S. 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., Boothe P., Baez A. 2013. Predicting morbidity and mortality for a saguaro cactus (Carnegiea gigantea) population. J. Torrey Bot. Soc. 140: 247–255. Gibson A.C., Nobel P.S. 1986. The Cactus Primer. Harvard Univ. Press. Cambridge, MA Snedecor, G.W. and Cochran, W.G. 1967. Statistical Methods, Sixth Edition. The Iowa State University Press. Ames www.ipni.org

Predicting bark coverage on saguaro cacti (Carnegiea gigantea) Olivia Printyâ&#x2C6;&#x2014; Laboratory of Plant Morphogenesis, Department of Biology, Manhattan College Abstract. In the Americas there are more than 20 species of tall, columnar cactus species that show bark coverages on their surfaces. From 1980 to present day, saguaro cacti (Carnegiea gigantea) have shown mortality rates of approximately 2.3% per year. This occurs as a result of bark accumulation on the surfaces of the cacti. A specific group of saguaro cacti located in Tucson, Arizona have been studied for 23 years, beginning in 1994 and ending in 2017. It has been determined with 95% probability that a cactus will be dead within 8 years once the bark coverage on the north-right trough exceeds 80% coverage. The purpose of this study was to predict bark coverage on north-right trough surfaces for saguaro cacti (Carnegiea gigantea) using bark coverages on several predictor surfaces. Over 55,000 data points were generated using data from the twelve surfaces of the 1,149 cacti used in this study. Machine learning programs DEC Trees and Validate Model were used to determine the accuracy of using predictive surfaces to determine the amount of bark coverage on the north-right troughs. Additionally, the machine learning programs Random Forest and kNN were also used to determine the best predictive surfaces. Results show that the most important indicators of bark formation on the north-right trough are the west-left trough and the north-left trough.

Introduction Saguaro cacti are a species of cacti native to southern Arizona and northern Mexico (Anderson, 2001). The cacti are tall, reaching 10 m in height, and are long-lived, with lifespans reaching up to 300 years (Steenbergh and Lowe, 1977). Due to high levels of sunlight exposure, the cacti experience an increase in epicuticular waxes in the stomata of the cacti, which then results in reduced levels of gas exchange (Evans et al., 2001). As a result, the cacti have a limited ability to perform photosynthesis and other metabolic processes, leading to limited cactus growth. Subsequently, the surfaces of the cactus saguaro cacti experience epidermal browning, or barking (Fig. 1). The barking and reduced gas exchange lead to cactus death (Evans et al., 2005; 2013; Evans and De Bonis, 2015; De Bonis et al., 2017). This epidermal browning begins on the southern-facing surfaces of the cacti. The cacti used in this study are located along the 32â&#x2014;Ś latitude. At this location, sunlight hits the southern facing surfaces four times more than the northern facing surfaces. The barking then progresses around both the eastern and western facing surfaces before finally reaching the northern facing surfaces. The northern facing surfaces are the last to brown (Fig. 2). A cactus will be dead within eight years once the north-right trough of the cactus has more than 80% bark (Evans et al., 2013). The purpose of this study was to better understand the importance of the north-right trough in predicting cactus death. To do this, machine learning programs were used to gain insight into the relationships between the rates of barking on the north-right trough and the other eleven cactus surfaces. The data were be separated into fast, slow, and normal (average) rates of bark formation. â&#x2C6;&#x2014;

Research mentored by Lance Evans, Ph.D.

114

The Manhattan Scientist, Series B, Volume 6 (2019)

Printy

Figure 1. Visual representation of the barking process on saguaro cacti. Barking will occur on the southern surfaces, specifically the southern crest, first and then travel around the eastern and western surfaces. The barking will then extend to the northern surfaces, which will be the last surfaces to experience bark coverage. The yellow star designates the North-right trough. The red stars designate Pair A, the North-left trough and the West-left trough. The purple stars designate Pair B, the North-left trough and the East-left trough. The green stars designate Pair C, the West-right trough and South-left trough.

Figure 2. The stages of bark formation on saguaro cacti. A cactus with no barking is shown with the trough and crest labeled. Crests protrude from the cactus and are generally exposed to more sunlight than troughs, which are angled. A cactus with more extensive bark coverage is also shown, as is a cactus that is completely covered in bark. Note that on this cactus, there are no spines located on the crest of the cactus.

Materials and Methods Field conditions In this study, a population of 1149 saguaro cacti were evaluated over a 23-year time period. These cacti were a part of 50 permanent plots established in 1994 in Tucson Mountain Park (Fig. 3). Evaluations occurred every 8 years; they took place in 1994, 2002, 2010 and 2017. Each cactus evaluated is listed using plot and cactus number. This designation is used in the database resulting from the evaluations. Data sets generated Samples of cacti ribs were evaluated for bark coverage for each of the four evaluation periods (1994, 2002, 2010, 2017). Each cactus has twelve ribs. Each rib is composed of two troughs, or indentations, and a crest, or a protrusion. The troughs flank the crest on either side. The ribs

Printy

The Manhattan Scientist, Series B, Volume 6 (2019)

115

Figure 3. A population of saguaro cacti found in Tucson Mountain Park in Tucson, Arizona.

evaluated were those most closely facing the cardinal directions (north, south, east and west). For each of these ribs, the crests and the troughs were evaluated. All samples were taken from a height of 1.75 cm above ground and had a length of 8 cm. From these evaluations, percentage of green coverage was determined. These percentages were subsequently converted into percent bark coverage. This information, along with the corresponding plot and cactus numbers, was placed into a Microsoft Excel File. This file is titled Master File. This information was uploaded for each of the four sampling periods, resulting in over 55,000 data points Structure of analysis The Master File data was uploaded into the MATLAB program Validate Model. Validate Model determined the qualifications of the data as having fast, slow, or normal rates of bark formation using predictor surfaces. For this program, three sets of predictor surfaces were used. The first used the north-left trough and west-left trough, which will be referred to as Pair A. The next set used the north-left trough and east-left trough, which will be referred to as Pair B. The final set used the west-right trough and south-left trough, which will be referred to as Pair C. Any cacti that had values greater or lesser than two standard deviations from normal were qualified as either fast or slow. Validate Model identified the fast, slow, and normal populations, resulting in three populations for each pair. This was done for each of the four sampling periods. Therefore, there were 36 unique populations. Each population was assigned a set of predictive surfaces (A, B, or C) and a rate (fast, slow, or normal), as well as an evaluation period. Once the populations of fast, slow and normal cacti were established, decision trees were produced and other evaluations were run to further distinguish between the populations of cacti.

116

The Manhattan Scientist, Series B, Volume 6 (2019)

Printy

Computer programs Predicting bark coverage 1. The data points used were provided by the Master File 2. The data from the Master File was uploaded into Validate Model. The data supplied by the Master File is arranged in a distribution by the program. Selective surfaces are provided to the algorithm to use to predict the rates of barking on those selected surfaces. A histogram (Fig. 4) displaying the range of cacti within the selection and the standard deviation of the selection is provided by the program. The algorithm shows both “fast” outliers, which are cacti whose rates of bark coverage are two standard deviations above the mean, and “slow” outliers, which are cacti whose rates of bark coverage are two standard deviations below the mean. 3. The information provided by Validate Model was then uploaded into Leave One Out. the information produced by Validate Model is analyzed by the algorithm to compare the predicted barking percentage and the observed barking percentage. 4. The data is then uploaded into REM High Low. The information produced by Leave One Out is analyzed by the algorithm and the data provided by Master File. Multiple sets of data were produced by the algorithm, one containing the information for fast outliers, one for slow outliers, and one for cacti considered to have a normal, or average, rate of barking. Each cactus listed as fast, slow or normal accompanied by the data for all four sampling periods (1994, 2002, 2010, and 2017). Predicting accuracy 1. The REM High Low program generates data sets to aid in the process of predicting barking accuracy. 2. the DEC Trees program compared the data for fast and slow outliers to that of the normal cacti. The program produces a confusion matrix and decision tree for each comparison after analyzing the fast, slow and normal cacti for each sampling period (1994, 2002, 2010, 2017). The accuracy of barking on the north-right trough is determined by these products. Random Forest The data in the Master File was uploaded into Random Forest, a statistical algorithm program that uses cross-validation to produce a list of the most important surfaces in predicting cactus death. Random Forest program produces and analyzes hundreds of decision trees in order to do this. The program must be trained. To do this, the Master File data for the 2017 evaluation period was used. Then, only the data for a select number of surfaces was uploaded in order to retrain the program to select for the most important surfaces. kNN kNN is an algorithm program used to compare cacti that are barking at similar rates in order to determine whether or not the qualification of a cactus as fast or slow by Validate Model is accurate. The Master File data is uploaded into kNN in order to do this. kNN looks only at the rates of bark

Printy

The Manhattan Scientist, Series B, Volume 6 (2019)

117

formation for the 2017 data. The program does not require training, unlike Random Forest. The Master File data was uploaded into the program twice. The first analysis used all twelve surfaces. The second analysis excluded data for the north-right trough. Table 1. Comparison of the accuracies produced by DEC for Pair A (north-left trough and west-left trough), Pair B (north-left trough and east-right trough) and Pair C (west-right trough and south-right trough) for all four sampling Outlier type

Range of accuracies

Total average accuracy

Pair A (north-left trough and west-left trough)

Fast Slow

96.9-98.5% 97.1-97.9%

97.6%

Pair B (north-left trough and east-right trough)

Fast Slow

91.3-97.5% 97.2-98.3%

96.5%

Pair C (west-right trough and south-right trough)

Fast Slow

88.4-90.1% 88.0-89.2%

88.9%

Results The purpose of this study was to predict bark coverage on the north-right trough using several predictor surfaces. Machine learning programs were used to determine the accuracy of the predictor surfaces in predicting barking rates on the north-right troughs. This was done by partitioning the rates of barking into fast, slow, and normal (average) barking rates. In order to do this, the populations of fast, slow and normal cacti must be determined. This was done using Validate Model. Predictive surfaces were used to create these populations within Validate Model. Fig. 4 shows the Validate Model output. Figure 4. The histogram produced by the MATLAB program Validate Model. It shows the distribution of samples collected over all four time periods. The slow outliers, or any data point two significant differences less than the mean, are located on the left portion of the histogram, while the fast outliers, or any data point two significant differences greater than the mean, can be found on the right portion of the histogram.

Table 2 shows the data for the prediction of bark coverage using Pair A and Pair B data from Validate Model for each of the four time periods. The accuracy of these predictive surfaces was then examined by DEC Trees using decision trees (Fig. 5). The accuracy for fast outliers in 1994 was 97.1%. The accuracy was 96.9% in 2002. The accuracy was 98.1% in 2010. The accuracy for slow outliers was 97.9% in 1994. The accuracy was 97.3% in 2002 and 97.1% in 2010. The accuracy for slow outliers was 97.3% in 2017. Using the north-right trough and west left trough

118

The Manhattan Scientist, Series B, Volume 6 (2019)

Printy

as predictive surfaces, DEC Trees determined an accuracy of 97% or greater for all four evaluation periods.

Figure 5. The decision trees produced from the analysis of fast outliers in Pairs A, B, and C by MATLAB.

For all fast surfaces, the determining surface was a north facing trough. For all slow surfaces, a western facing surface was the determining surface. In addition to producing decision trees, DEC Trees also produced confusion matrices. Each decision tree had its own confusion matrix, which analyzed the accuracy of the sorting of the cacti as fast or slow. The results of the confusion matrices determined the percent accuracy of the decision tree. The results of the confusion matrices for fast outliers of all four time periods were listed in Table 2.

Printy

The Manhattan Scientist, Series B, Volume 6 (2019)

119

Table 2. Confusion matrices of predicted and actual numbers of cacti with fast vs. normal rates of bark coverage on the north-right trough based off of data for fast north left troughs and fast west left troughs for each sampling period. A. 1994 Normal-Fast (NL, WL)

B. 2002 Normal-Fast (NL, WL)

Predicted Fast Normal Actual

Fast Normal Accuracy

3 5

| |

Predicted Fast Normal

5 465 97.9%

Actual

C. 2010 Normal-Fast

Fast Normal Accuracy

| |

3 4

7 462 96.9%

D. 2017 Normal-Fast

Predicted Fast Normal Actual

1 8

| |

Predicted Fast Normal

5 466 98.1%

Actual

Fast Normal Accuracy

5 3

| |

3 467 98.5%

Using all 12 surfaces, Random Forest was trained by all 4168 samples. The program listed the most important surfaces as the north-right trough, north-left trough, and west-left trough. Then, by Table 3. Results of the training of all 12 surfaces vs. the testing of only 2017 data by Random Forest for all four sampling periods.

0 1b NR NL WL ER SL SR EL WR SC WC EC NC

All 12 surfaces training

2017 data removed training

test on 2017 data test

4168 samples

3128 samples

567 samples

3126 0

3 1039

568 0

0 474

0 568

1 0

importance

mean decrease accuracy

20.98 14.29 11.02 7.76 5.18 4.44 4.19 4.11 3.74 3.58 2.83 2.47

10.59 9.57 6.69 4.48 4.44 4.09 4.09 3.45 3.56 2.88 2.29 2.25

113.87 78.0 100.29 42.17 14.73 13.34 11.67 10.28 8.76 8.3 3.02 4.68

Accuracy: 99.9%

120

The Manhattan Scientist, Series B, Volume 6 (2019)

Printy

removing the 2017 data, the program was retrained. The results showed the same rankings of importance amongst the surfaces, as shown in Table 3. Random Forest was also trained using only 11 of the 12 surfaces, excluding the north-right trough. 3128 samples were used to train the program. Random Forest named the north-left trough and west-left trough as the most important surfaces in predicting bark. Subsequently, Random Forest was retrained using only the 2017 data. Again, the north-left trough and west-left troughs were named as the most important surfaces. Table 4 displays the results. Table 4. Results of the training of 11 surfaces, excluding the north-right trough, vs. the testing of only 2017 data by Random Forest for all four sampling periods. 11 surfaces training

2017 data removed training

test on 2017 data test

3126 samples

3128 samples

1140 samples

0 b

0 1b NL WL ER SL EL NC SR WR WC SC EC

2550 0

11 565

561 7

0 474

682 0

1 358

importance

mean decrease accuracy

15.027 12.661 5.747 3.708 3.368 3.259 3.189 3.005 2.962 2.172 1.780

75.4 42.2 38.0 22.0 7.1 3.2 3.6 3.6 2.0 1.1 0.9

305.9 376.9 41.2 14.4 11.7 22.9 11.4 11.0 4.1 9.5 0.0

Accuracy: 99.9%

Once the Random Forest analyses were complete, the new Master File was uploaded into the kNN program. There is no training required for this program, unlike Random Forest. kNN produced two confusion matrices. The confusion matrix produced using all surfaces for the 2017 time period produced an accuracy of 100%. The second confusion matrix excluded the north-right trough and had an accuracy of 99.33%. These results are listed in Table 5.

Discussion The purpose of this study was to predict bark coverage on the north-right trough using several predictor surfaces. Machine learning programs were essential in this process. Specifically, Validate Model and DEC Trees were necessary for their analyses of bark coverage. Validate Model was used

Printy

The Manhattan Scientist, Series B, Volume 6 (2019)

121

Table 5. Confusion matrices produced by kNN through the training of various surfaces vs. the testing of only 2017 data A. 2017, all surfaces included

B. 2017, north-right trough excluded

Predicted Alive Dead Actual

Alive Dead Accuracy

| |

568 0

0 474

100%

Predicted Alive Dead Actual

Alive Dead Accuracy

| |

561 7

0 474

99.33%

to determine rates of bark formation. DEC Trees determined the accuracy of Validate Modelâ&#x20AC;&#x2122;s selections. Bark coverage has been documented on more than 20 tall, long-lived cacti in the Americas. This epidermal browning begins on the southern-facing surfaces of the cacti before progressing around both the eastern and western facing surfaces before reaching the northern facing surfaces. These cacti are located along the 32â&#x2014;Ś latitude, where sunlight hits the southern facing surfaces four times more than the northern facing surfaces. The bark formation will begin on the crests of the cacti ribs before progressing inward towards the troughs. Due to these patterns of bark formation, the last surface to experience barking will be the north-right trough. A cactus will be dead within 8 years once the north-right trough reaches more than 80% bark coverage on the north-right trough. The predictability of the north-right trough was determined through the use of three sets of paired surfaces. Pair A (north-left trough and west-left trough) was determined to have the best predictability by the DEC Trees program. The proximity to the north-right trough makes the northleft trough a good predictor surface. The west-left trough is one of the first troughs to experience barking and was chosen as a predictor surface for this reason, despite there being a wide range of bark coverages for this surface in the populations studied by this research. Together, the west-left trough and north-right trough indicate what may be happening on the north-right trough. High accuracies were produced using these two predictor surfaces, as seen in Table 1. The fast cacti have an average accuracy of 97.9% for all four time periods, while the slow cacti have an average accuracy of 97.4%. Table 1 also shows the average accuracies using Pair B (north-left trough and east-right trough) and Pair C (west-right trough and south-right trough). Pair B had similar accuracies, although they were lower than those for Pair A. Pair C had accuracies that were much less similar and not as accurate as those produced using Pair A as the predictive surfaces. The surfaces used in Pair C are located the farthest away from the north-right troughs. For the analyses using Random Forest, the same surfaces were determined to be the most important: the north-right trough, the north-left trough and the west-left trough. When the program was trained using all twelve surfaces, a value of importance of 17.79 was given to the northright trough (Table 3). The north-left trough and west left trough had values of 10.59 and 9.57,

122

The Manhattan Scientist, Series B, Volume 6 (2019)

Printy

respectively (Table 3). When only 11 surfaces were used to train Random Forest and the north-right trough was excluded, the north-left trough was listed as having the highest value of importance. This value was 75.4. The west left trough was listed as the second most important surface, with a value of importance of 42.2 (Table 4). It is indicated by these results that the north-right trough, north-left trough and west-left trough are the most important surfaces in predicting cactus death. The data produced by the kNN program supports the Random Forest results. kNN also suggests that the three surfaces given the most importance by Random Forest are the best indicators of cactus death. The two confusion matrices produced by kNN both have accuracies above 99% (Table 5). The notion that the north-facing surfaces are the best predictor surfaces is supported by these results.

Acknowledgement This work was supported by the Linda and Dennis Fenton â&#x20AC;&#x2122;73 endowed biology research fund. The author is grateful to the Catherine and Robert Fenton Endowed Chair Dr. Lance Evans for additional support for this research. The author is also indebted to George Kennedy for his work on the DEC Trees program (see Appendix 1).

References Anderson, E. F. 2001. The Cactus Family. Timber Press. Portland OR, 776 p. De Bonis, M. L. Barton and L.S. Evans. 2017. Rates of bark formation on surfaces of saguaro cacti (Carnegiea gigantea). Journal of the Torrey Botanical Society. 144: 1-8. 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: 181â&#x20AC;&#x201C;187. Evans, L. S., A. J. Young, and Sr. J. Harnett. 2005. Changes in the scale and bark stem surfaces injuries and mortality rates of a saguaro (Carnegiea gigantea) cacti population in Tucson Mountain Park. Can. J. Bot. 83: 311-319. 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., and M. DeBonis. 2015. Predicting Morbidity and Mortality of Saguaro Cacti (Carnegiea gigantea) J. Torrey Bot. Soc. 142: 231-239. 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.

The interaction of Amphotericin B: An ab initio study Alon Brownâ&#x2C6;&#x2014; Department of Chemistry, Manhattan College

Abstract. In the determination of Amphotericin B and its interaction with Ergosterol ab initio calculations were performed with the PM3 basis set. Values for Ergosterol, Amphotericin B, Amphotericin B with Ergosterol, Amphotericin B with Ergosterol and Chloride, and Amphotericin B with Ergosterol and Carbonate were calculated using Spartan 18.

Introduction Amphotericin B (Am B) is an antibiotic in the class of macrolides, it has been shown to have both antibiotic and antifungal activity. The structure of Am B has two hydrophilic and one lipophilic section as shown in Fig. 1.

Figure 1. Amphotericin B

These sections allow it to have a unique interaction with the cell membrane. Previous research [1, 2] has suggested that the unique hydrophilic regions of Am B combine with Ergosterol; a steroid component found within the membrane of fungi. Ergosterol shown in Fig. 2, is mostly hydrophobic with a single hydroxyl group. â&#x2C6;&#x2014;

Research mentored by Joseph F. Capitani, Ph.D.

124

The Manhattan Scientist, Series B, Volume 6 (2019)

Brown

Figure 2. Two Ergosterol compounds with the hydroxyl group circled

This hydroxyl group is proposed by Gray et al. [3] to create a strong hydrogen bond with a hydroxyl group of Am B. This interaction along with the large lipophilic region of Am B leads to a hydrophobic pore within the membrane as shown in Fig. 3. The large pore formed by Am B increases the cells permeability and is proposed to allow ions to leak out resulting in fungal cell death [4].

Figure 3. Ergosterol with Amphotericin B

Recent research by Katrina et al. [5] has proposed that Am B will interact within the cystic fibrosis transmembrane conductance regulator (CFTR) which is defective in individuals with cystic â&#x2C6;&#x2019; fibrosis. This interaction is proposed to restore HCOâ&#x2C6;&#x2019; 3 and Cl secretion and increased airway surface liquid pH in cultured airway epithelia from people with cystic fibrosis. In the present study, ab initio calculations were performed on Figs. 1, 2, and 3 along with the addition of chloride ion, and a carbonate ion within Fig. 3.

Brown

The Manhattan Scientist, Series B, Volume 6 (2019)

125

Results and Calculations The calculations were performed using Spartan-18 [6] on a Dell OptiPlex 9020 with an Intel core i5-4570 at the Kakos Center for Scientific Computing. All structures were optimized using Parameterized Model number 3 (PM3) semi empirical method [7]. In evaluation of Ergosterol there is little space between two compounds as seen in Fig. 4.

Figure 4. Two Ergosterol compounds shown sideways with a space filling model calculated with PM3

The calculated minimum distance between any two carbon atoms on opposing compounds is 3.282 AĚ&#x160;. The electrostatic potential map shows an evenly distributed neutral charge distributed between the two compounds with a small area of low potential on the hydroxyl groups Fig. 5.

Figure 5. Electrostatic potential map of 2 ergosterol compounds calculated with PM3

126

The Manhattan Scientist, Series B, Volume 6 (2019)

Brown

With the addition of Am B there is a visible pore as shown in Figs. 3 and 6.

Figure 6. Ergosterol with Amphotericin B side view space filling model calculated with PM3

The calculated minimum distance between any two carbon atoms on opposing compounds is 7.774 Å, and the calculated minimum distance between any two oxygens is 5.864 Å. This increase in at minimum 2.582 Å in the space leads to a visible pore. This channel is evident in the electrostatic potential map in Fig. 7.

Figure 7. Electrostatic potential map of Ergosterol and Amphotericin B calculated with PM3

There is a separation of charge potential between the two compounds, allowing for potentially charged ions to travel within the gap. Further analysis was implemented on the effect of a Cl− ion within the pore (Fig. 8). Cl− was chosen because of its proposed effect in CF patients.

Brown

The Manhattan Scientist, Series B, Volume 6 (2019)

127

Figure 8. Ergosterol with Amphotericin B and Cl− space filling model calculated with PM3

The chloride ion appears trapped between the two Am B complexes. The electrostatic potential map (Fig. 9), shows one fully decentralized compound with low electrostatic potential.

Figure 9. Ergosterol with Amphotericin B and two Cl- electrostatic potential map front and side view calculated with PM3

The chloride ions are trapped within the Am B complexes creating a bridge between the two. The distance between the chloride ion between the complexes and the three nearest oxygen atoms are; 2.771 Å, 2.761 Å, and 2.799 Å. The two closest carbon atoms on opposing compounds are 6.062 Å these same two carbons were 9.802 Å away in the analysis without the chloride ion. The two carbons that were 7.774 Å and the two oxygen that were 5.864 Å apart in the analysis without the chloride ion, are now calculated to be 6.111 Å and 5.994 Å apart respectively. Am B complexes with carbonate ion between were similarly chosen because of carbonates similarly proposed effect in CF patients results correlate to those seen with the chloride ion Fig. 10. The circled carbonate ion in Fig. 10 has hydrogen bonding with the Am B complex. The electrostatic potential map Fig. 11, shows a similar result to the chloride ion in that it is trapped within the Am B complexes creating a bridge between the two. Here there is a clear low electrostatic

128

analysis without the chloride ion, are now calculated to be 6.111 Å and 5.994 Å apart respectively. Am B complexes with carbonate ion between were similarly chosen because of carbonates similarly proposed effect in CF patients results correlate to those seen with the chloride ion figure 10.

The Manhattan Scientist, Series B, Volume 6 (2019)

Brown

Figure 10: Ergosterol with Amphotericin B and carbonate ball and spoke model calculated with PM3

Figure 10. ErgosterolThe with Amphotericin B and carbonate ball and spoke circled carbonate ion in figure 10 has hydrogen bonding with the Am Bmodel calculated with PM3 complex. The electrostatic potential map figure 11, shows a similar result to the chloride ion in that it is trapped within the Am B complexes creating a bridge between the two. Here there is a clear low electrostatic potential in the middle and a neutral charge among the rest of the two complexes. The distance between the carbonate ion and the three nearest oxygen atoms on the Am B complexes are; 2.735 Å, 2.726 Å, and 3.373 Å. The distance between the two oxygen atom on the bicarbonate ions and the hydrogens on the Am B complex forming a bond are 1.767 Å, and 1.773 Å. The two closest carbon atoms on opposing Am B compounds are 5.836 Å these same two carbons were 8.075 Å away in the analysis without the carbonate ion. Towards the bottom of the complex there is now on oxygen and carbon that are 3.787 Å apart these were 9.280 Å away in the analysis without the carbonate ion due to the shift created by the carbonate interaction.

Figure 11. Ergosterol with Amphotericin B and carbonate ion electrostatic potential map front and side view calculated with PM3

potential in the middle and a neutral charge among the rest of the two complexes. The distance between the carbonate ion and the three nearest oxygen atoms on the Am B complexes are; 2.735 Å, 2.726 Å, and 3.373 Å. The distance between the two oxygen atom on the bicarbonate ions and the hydrogens on the Am B complex forming a bond are 1.767 Å, and 1.773 Å. The two closest carbon atoms on opposing Am B compounds are 5.836 Å these same two carbons were 8.075 Å away in the analysis without the carbonate ion. Towards the bottom of the complex there is now on oxygen and carbon that are 3.787 Å apart these were 9.280 Å away in the analysis without the carbonate ion due to the shift created by the carbonate interaction.

Discussion The effect of Am B on the membrane of a cell has been suggested to create a large pore. The resulting calculations done with PM3 agree with this assumption and calculate the pore size to be larger than 5.864 Å at any given point. With the addition of the chloride ion the two compounds appear to join as shown in Figs. 8 and 9. While the distance between the two Am B is now 5.994 Å

Brown

The Manhattan Scientist, Series B, Volume 6 (2019)

129

apart at any point for the chloride calculation, there is now this connection through the chloride ion and its interaction with three neighboring oxygens. When analyzing the calculated natural charge there is a small difference on the three nearest oxygens as shown in Fig. 12.

Figure 12. Comparison of the natural charge on three oxygen of the Am B complex with the chloride and without, using PM3

Where the charges on the oxygen interacting with the chloride ion are -0.424, -0.414, and -0.423; the charges are -0.325, -0.337, and -0.306 respectively on the measurement of the same oxygen without chloride. This difference does not account for the strong interaction with the chloride ion but could explain the low electrostatic potential along with the -0.795-natural charge on the chloride ion that must be dispersed. Further research should be conducted on the interaction between the chloride ion and Am B. While the analysis suggests that the chloride ion would remain fixed within the complexes with the addition of a concentration gradient the chloride ion can be pulled through the pore, and the interaction with Am B should be enough to keep grabbing additional chloride ions over time. In measurements of two Am B complexes with the addition of the carbonate ion the two compounds appear similarly to join, with a minimal distance of 3.787 Å apart at any point between the two. When analyzing the calculated natural charge there is a small difference on the three nearest oxygen to carbonate as shown in Fig. 13. Where the charges on the oxygen interacting with the carbonate ion are -0.394, -0.390, and -0.355; the charges are -0.320,

Figure 13. Comparison of the natural charge on three oxygen of the Am B complex with the carbonate ion and without, using PM3

-0.305, and -0.336 respectively on the measurement of the same oxygen without carbonate. This is a minimal difference, and in this case the strong interaction is evident in the clear hydrogen bonding between the carbonate ion and Am B with a bond distance of 1.767 Å, and 1.773 Å. Further research should be conducted on the interaction between the carbonate ion and Am B. While the analysis suggests that the carbonate ion would remain fixed, the hydrogen bonding can happen at

130

The Manhattan Scientist, Series B, Volume 6 (2019)

Brown

many different places along the neighboring complexes. Again with the addition of a concentration gradient the carbonate ion should be able to be pulled along the pore creating different hydrogen bonds along the path. While the calculations presented in this paper provide a good introduction on the understanding of Am B, it is clear that density functional calculations with a large basis set need to be conducted to further analyze the results. The calculations should also checked with CFTR for validation that it works similar to Ergosterol. In addition, other ions need to be analyzed to extend the scope. To further simulate the in vivo affects calculations should be reworked in the presence of water and with a concentration gradient.

Acknowledgments The author wishes to thank Dr. Joseph Capitani, for his mentorship and continuous support throughout the project. Mr. Michael J. Kakos â&#x20AC;&#x2122;58, and Mrs. Aimee Rusinko Kakos for their donation of the Kakos Center for Scientific Computing where calculations were performed.

References [1] Brajtburg J., Powderly W.G., Kobayashi G.S., Medoff G. (1990). Amphotericin B: current understanding of mechanisms of action. Antimicrobe Agents Chemother. 34:183-8 [2] Yano, T., Itoh, Y., Kawamura, E., Maeda, A., Egashira, N., Nishida, M., Kurose, H., Oishi, R. (2009). Amphotericin B-induced renal tubular cell injury is mediated by Na+ Influx through ion-permeable pores and subsequent activation of mitogen-activated protein kinases and elevation of intracellular Ca2+ concentration. Antimicrobe Agents Chemother. 53:1420-1426. doi: 10.1128/AAC.01137-08 [3] Gray K.C., Palacios D.S., Dailey I., Endo M.M., Uno B.E., Wilcock B.C., and Burke M.D. (2012). Amphotericin primarily kills yeast by simply binding ergosterol. Proc. Natl. Acad. Sci. 109:2234-2239 [4] Kinsky, S.C. (1970). Antibiotic interaction with model membranes. Annu. Rev. Pharmacol. 10, 119-142. [5] Muraglia K.A., Chorghade R.S., Kim B.R., Tang X.X., Shah V.S., Grillo A.S., Daniels P.N., Cioffi A.G., Karp P.H., Zhu L., Welsh M.J., Burke M.D. (2019) Small-molecule ion channels increase host defenses in cystic fibrosis airway epithelia. Nature 567 (7748), 405-408. doi: 10.1038/s41586-019-1018-5 [6] Spartan 18 Wavefunction, Inc. 18401 Von Karman Ave., Suite 370 Irvine, CA 92612 [7] Mesa-Arango A.C., Scorzoni L., Zaragoza O. (2012). It only takes one to do many jobs: Amphotericin B as antifungal and immunomodulatory drug. Front Microbiol. 3: 286. doi: 10.3389/fmicb.2012.00286

Synthesizing aromatics using iridium catalyzed ketone alkylations and classical methods Sharron Fernandez∗ and Jennifer LaPoff∗ Department of Chemistry and Biochemistry, Manhattan College Abstract. This research focuses on the synthesis of fragrance compounds that contain phenyl rings with a polar functional group (alcohol or ketone) and a 2-carbon spacer. This was accomplished two ways, one of which is using classical reactions learned in second year organic lectures and labs. These include transfer hydrogenation, aldol condensations, and Grignard reactions. A more modern method uses a transition metal catalyst to formulate multistep reactions in one reaction setup. The new school method used was a ketone alkylation reaction, which is used in industry to join small alcohols together to make larger branched alcohols. The products were analyzed via IR and NMR spectroscopy and compared to commercial samples.

Introduction The goal of this experiment was to conduct two categories of organic chemical reactions to synthesize aromatic fragrance compounds. The products of both classifications of reactions were analyzed to conclude which reaction type is most convenient in synthesizing aromatics. The first set of reactions used traditional reactions similar to the ones practiced by first and second semester organic chemistry students. These consist of an aldol condensation, transfer hydrogenation, sodium borohydride-based reduction, and ethyl and methyl Grignard reactions. The second category, known as a ketone alkylation reaction, performs the same series of reactions as the first category, in just one step and in the same reaction vessel using an Ir catalyst. The expectation was that the Ir catalyzed ketone alkylators are optimized to give the highest yield of fragrance compounds. Because the product synthesized using classical methods consists of multiple reactions, there is more room error, and more opportunities for the product to gain impurities along the way. The ketone alkylation is a single reaction, therefore there is less room for error, and as a result the product should be cleaner.

Background The conventional methods that were used, or aldol condensation, transfer hydrogenation, sodium borohydride reduction and Grignard reactions are generally well-known reactions that have been around for quite some time. The aldol condensation, which was first introduced by Charles Wurtz in 1872, is the reaction that takes place between an enolate ion and carbonyl compound in the presence of an acid/base catalyst. The resulting product is either a beta-hydroxy aldehyde or ketone, which is further dehydrated to form a conjugated enone (Sigma Aldrich 2019). This is famous method in which aldehydes and ketones are prepared in industry. Transfer hydrogenation ∗

Research mentored by James McCullagh, Ph.D.

132

The Manhattan Scientist, Series B, Volume 6 (2019)

Fernandez + LaPoff

reactions are much older in the world of organic chemistry. They are generally used in the addition of a hydrogen to a compound, in which hydrogen is donated from one compound to the other. Sodium borohydride-based reduction is the process in which a carbonyl is reduced to an aldehyde in the absence of hydrogen, as the borohydride ion acts as a reducing agent. Sodium borohydride reacts with carbonyl compounds much like Grignard reagents, however they function as hydride donors rather than carbanion sources (Wiley, 1968). Grignard reagents are formed by reacting metal with alkyl, aryl or vinyl halide compounds in solvents such as diethyl ether or tetrahydrofuran. Grignard reagents are very strong bases that react well with compounds that contain hydroxide groups. Grignard reactions are also strong nucleophiles known to react with carbonyl compounds to form carbon-carbon bonds. The compounds being synthesized are benzyladine acetone, benzylacetone, lilac pentanol, muget carbinol, 4-phenyl-2-butabol, and 1-phenyl-3-pentanol (Figs. 1-6):

Figure 1. Benzyladiene acetone ; floral scent

Figure 2. Benzylacetone; sweet floral / strawberry scent

Figure 3. Lilac pentanol; sweet lilac scent

Figure 4. Muget carbinol; floral lily / violet scent

Figure 5. 4-phenyl-2-butnaol; floral mimosa scent (Good Scents Company)

Figure 6. 1-phenyl-3-pentanol; sweet lilac scent (Surburg and Panten, 2006).

One of the most common ways that these compounds are made commercially is with an aldol condensation followed by hydrogenation and possible reduction of the ketone as shown below (Fig. 7) (Volkov et al., 2015).:

Fernandez + LaPoff

The Manhattan Scientist, Series B, Volume 6 (2019)

133

Figure 7. The overall schematic for the old school reactions, using an aldol condensation, hydrogen catalyzed hydrogenation and reduction reaction to yield desired product.

In a more modern method, the [Ir (COD) Cl]2 catalyst is capable of doing the same multistep process in one reaction flask by way of the ketone alkylationreaction (Koda et al., 2009). The reaction looks unrealistic to most chemists; it only appears this way due to the several reactions that take place at the same time in the same reaction flask as shown below (Fig. 8)

Figure 8. The overall schematic for the iridium catalyzed ketone alkylation reaction to yield desired product. Only one reaction is required to yield desired product, as several reactions take place in one reaction flask.

Procedural overview Illustrated below (Fig. 9) is the overall classical synthesis method:

Figure 9. The overall reaction including appropriate starting material and products for the â&#x20AC;&#x153;old schoolâ&#x20AC;? reactions

The reaction begins with the aldol condensation of benzaldehyde and acetone to form benzylidene acetone. Transfer hydrogenation of benzylidene acetone results in the formation of benzylacetone, which is then used in three separate reactions. The first one being a sodium borohydride

134

The Manhattan Scientist, Series B, Volume 6 (2019)

Fernandez + LaPoff

reduction to form 4-phenyl-2-butanol. The second reaction was a methyl Grignard reaction which resulted in the formation of muget carbinol. The last reaction was an ethyl Grignard reaction in which lilac pentanol was synthesized. The first step in the procedure, the aldol condensation was performed by my partner. The transfer hydrogenation of benzylidene acetone to benzylacetone was completed by me. This reaction was a large-scale reaction and required a 1:100 starting material to catalyst. Keep in mind that because the catalyst is 5% Pd/ alumina, a 1:5 starting material to catalyst by weight is actually 1:100. Benzylidene acetone, sodium formate, potassium carbonate, and 1,2-propanediol were stirred at 160◦ C along with 5% Pd/alumina catalyst and the reaction was refluxed for 90 minutes. After the reflux the reaction was vacuum filtered and extracted with MTBE and saturated sodium chloride solution. The solvent was boiled off on a sand bath set to 120◦ C. This reaction was performed a total of nine times, with the starting material, reflux time and temperature changed in each reaction to assess which combination would give us the cleanest product and most yield (Table 1). Table 1. Differences and percent yield in transfer hydrogenation reactions run. Standard reaction conditions: 0.35 g benzylidene acetone, 7 mL 1,2 propanediol, 50 mg of potassium carbonate, 500 mg sodium formate, and 120 mg of 5% Pd / Alumina. Reaction number

Modifications to standard reaction conditions

Reflux time & temperature

% yield

Notes

Standard conditions

60 minutes @ 160◦ C

72%

Product only

2.521 g γ- terpene was used in place of sodium formate (H donor)

60 minutes @ 160◦ C

73%

16% conversion

Standard conditions

40 minutes @ 160◦ C

66%

82% conversion

90 minutes @ 180◦ C

61%

76% conversion

90 minutes @ 170◦ C

86%

75% conversion

90 minutes @ 230◦ C

58%

70% conversion

90 minutes @ 230◦ C

85%

63% conversion

90 minutes @ 230◦ C

81%

Product only

90 minutes @ 160◦ C

87%

Product only

4 5

7 8 9

Large scale reaction 1.0 g benzylidene acetone Reaction ran under acidic conditions (7 mL 0.25 M HCl in 1,2-propanediol used in place of potassium carbonate) Reaction ran under acidic conditions (7 mL 0.25 M HCl in 1,2-propanediol used in place of potassium carbonate) Reaction ran under acidic conditions (7 mL 0.25 M HCl in1,2-propanediol used in place of potassium carbonate) Large scale reaction1.0 g benzylidene acetone Large scale reaction 1.0 g benzylidene acetone

The methyl and ethyl Grignard reactions of benzylacetone to form lilac pentanol and muget carbinol were completed by me. Premade 3M Grignard reagent was used, as many complications could occur during the preparation of the reagent that would inhibit it to react properly. 3M

Fernandez + LaPoff

The Manhattan Scientist, Series B, Volume 6 (2019)

135

methyl/ethyl Grignard in diethyl ether was reacted with benzylacetone in tetrahydrofuran and refluxed for 30 minutes on a sand bath set to 120◦ C. The reaction apparatus was placed on an ice bath, quenched with 2M hydrochloric acid, and the tetrahydrofuran was boiled off. The product was extracted with MTBE, and saturated sodium chloride solution. The solvent (MTBE) was then boiled off on a sand bath set to 120◦ C. The ketone alkylation reactions were run. The overall schematic for the reaction is shown in Fig. 10. The goal of these reactions was to synthesize 1-phenyl-3-pentanone using the ideal ligand

Figure 10. The overall schematic for the “new school” ketone alkylation reaction. The desired product, as well as three by products were synthesized.

(Taguchi et al., 2004). The iridium catalyst, [Ir (COD)Cl]2, KOH, and phosphine were measured in a glove bag pumped with argon gas. The only variant between each ketone alkylation reaction run was the type, and amount of phosphine ligand used. The reactions were air sensitive and run in pressurized tubes filled with argon gas. Benzyl alcohol and methyl ethyl ketone were then added to the pressurized tube with the starting material and the reaction was refluxed for four hours at 130◦ C. After reflux the contents of the reaction tube were vacuum filtered through silica gel. The reaction tube was then rinsed with MTBE and acetone and the rinsings were also vacuum filtered. The filtrate was then rotavapped to remove the solventfrom the final product. These reactions were run a total of four times, and only variant between each reaction being the amount and type of phosphine used.

Results Our results are summarized below in Tables 1-3. Table 1 includes the starting material and reflux time/ temperature of all the transfer hydrogenation reactions, with a standard reaction as well as detailed changes in every reaction that deviated from the standard reaction. Percent yield gives an idea of how well each reaction ran.

136

The Manhattan Scientist, Series B, Volume 6 (2019)

Fernandez + LaPoff

Table 2 lists the ethyl and methyl Grignard reactions run as well as the starting material used and reflux time/ temperature. Percent yield is included to give an idea of how well the reaction ran. Table 2. Starting materials and percent yields for methyl and ethyl Grignard reactions Type of reaction Methyl Grignard

Ethyl Grignard

Starting material

Reflux time & temperature

% yield

30 minutes @ 120◦ C

76%

30 minutes @ 120◦ C

77%

0.520 g benzylacetone 11.3 mL tetrahydrofuran 4.5 mL Grignard reagent 0.504 g benzylacetone 11.3 mL tetrahydrofuran 4.5 mL Grignard reagent

Table 3 includes the type and amount of phosphine ligand used and reflux time/ temperature in the ketone alkylation reactions. Percent conversion was included to give an idea of the types and amounts of byproducts were formed in comparison to the desired product. Table 3. Guerbet reactions. Standard reaction conditions: 1 mole % of [Ir(COD)Cl]2, 0.5-2 mole % of phosphine, 20 mole % of KOH, 0.2 mL of benzyl alcohol, and 0.77 mL of methyl ethyl ketone were placed in a sealed tube under argon atmosphere and heated to 130◦ C. Percent conversion Reaction number 1 2 3 4

Amount and type of phosphine P(otol)3 phosphine 2:1 equivalence relevant to catalyst P(otol)3 phosphine 1:1 equivalence relevant to catalyst tricyclohexyl phosphine 2:1 equivalence relevant to catalyst DIPHOS phosphine 1:1 equivalence relevant to catalyst

Reflux time & temperature 240 minutes @ 130◦ C 480 minutes @ 130◦ C 240 minutes @ 130◦ C 240 minutes @ 130◦ C

Benzyl alcohol

Desired product

Other product

Aldehyde

52%

151%

8.7%

4.8%

60%

173%

7.5%

35%

114%

57%

5.7%

75%

97.3%

< 10%

1.33%

IR analysis was used to get an idea of the major functional groups that might have been present in the product compound. NMR analysis was used as a major means of identifying the structure and compounds present in the products. NMR spectra were taken on commercially synthesized versions of each expected product and used as a comparison to NMR spectra of products synthesized in lab. The reference spectra helped determine if the reactions ran to completion and if the product was synthesized as expected. The expected product for the transfer hydrogenation reaction was benzylacetone. In Fig. 11, the NMR spectra for the product of the transfer hydrogenation reaction had peaks at 7.24 ppm (bs, 5H), 2.809 ppm (m, 4H), and 2.373 ppm (s, 3H) that matched the peaks present in the reference spectra of benzylacetone. The expected product for the methyl Grignard was 2-

Fernandez + LaPoff

The Manhattan Scientist, Series B, Volume 6 (2019)

137

Figure 11. H1 NMR of transfer hydrogenation reaction #9

methyl-4-phenyl-2-pentanol (Muget carbinol). In Fig. 12, the NMR spectra for the product of the methyl Grignard reaction, peaks were present at 7.23 ppm (bs, 5H), 2.71 ppm (t, 2H), 1.76 ppm (t, 2H), and 1.28 ppm (s, 6H). These peaks were consistent with the peaks present in the Reference Muget carbinol spectra. Fig. 13 represents spectra for product of the ketone alkylation reaction.

Figure 12. H1 NMR of methyl Grignard reaction

Peaks present from 8 - 7.2 ppm are assumed to be overlapping of the hydrogens present in

138

The Manhattan Scientist, Series B, Volume 6 (2019)

Fernandez + LaPoff

the phenyl ring while peaks 3.70 ppm (m, 4H), 1.49 ppm (q, 2H), and 1.14 ppm (t, 3H) were consistent with some of peaks present from the reference spectra. Evidence was present in Fig. 13 that suggested the ketone alkylation reaction did not run to completion.

Figure 13. H1 NMR of ketone alkylation #4

Discussion Grignard reactions The purchase of premade Grignard reagent greatly aided in a smoother reaction. Since the Grignard was purchased, there was far less room for error than if it were made in lab, given its sensitivity to air. Procedural data as well as percent yield was noted in Table 2. The methyl Grignard and ethyl Grignard were run with the same ratios of starting material, under the same conditions. It was observed that after the initial reflux, it was vital to place reaction apparatus in ice bath before quenching, to prevent violent spluttering upon addition of 2 M HCl. This is due to the fact that premade Grignard reagent reacts violently with acid at room temperature. Transfer hydrogenation reactions Procedural data, as well as percent yield conversions for transfer hydrogenation reactions 19 were noted in Table 1. Reaction #1, the first transfer hydrogenation reaction was the standard reaction run to test the parameters of this class of reaction. Potassium carbonate was initially added to reaction #1 as a partial catalyst, as it is known to speed up these types of reactions. In Reaction #2 the base, sodium formate was replaced with Îł-terpene to determine if it would act as a better

Fernandez + LaPoff

The Manhattan Scientist, Series B, Volume 6 (2019)

139

hydride source. Based on NMR analysis and percent yield, it was determined that sodium formate acted as a better hydride source, which is why it was reused in all other reactions. The reflux time in reaction #3 was cut down by 20 minutes to see how a lower reflux time would impact the final product. NMR analysis proved that cutting the reflux time down to 40 minutes inhibited the reaction from running to completion, so the reflux time was increased to 90 minutes in Reaction #4 and the temperature was raised to 180◦ C. Additionally, reaction #4 was run on a large scale and the initial starting material was increased by 2.89 times the original. The impact of the potassium carbonate on the completion of the reaction was tested in reaction #5, when potassium carbonate was replaced by HCl in 1,2-propanediol. The small scale of starting material was also used once again. The reflux time and temperature also remained the same as reaction #4. The acidic reaction ran well, but not as well as under basic conditions. The reaction was close to completion, but did not reach the end. In reaction #6, the acid was used again instead of the base. The reflux time also remained the same, and the temperature was increased to 260◦ C, to determine if raising the temperature would improve reaction conditions. Reaction #7 reacted with the same conditions as reaction #6. In reaction #8, the starting material was increased once again by 2.89 times the original small-scale amount (0.350 g), and potassium carbonate was used in place of HCl in 1,2propanediol. The reflux conditions remained the same as reaction #7. Reaction #9 was a repeat of the conditions observed in reaction #8, however, the reflux temperature was lowered to 160◦ C. Ketone alkylation reactions Because this class of reaction is relatively new to the world of chemistry, there were less sources to reference when problems in reactivity occurred. Procedural data, as well as percent conversions for ketone alkylation reactions 1-4 were noted in Table 3. The initial trial, reaction #1, was the standard reaction and was tweaked based on trial and error. NMR analysis proved the desired product was not as we had hoped so a different amount and type of ligand [P(otol)3 ] was used in reaction #2. The reaction time was also raised from 4 hours to 8 hours to determine if the reaction needed a longer reflux time to run to completion. Once again, NMR analysis of reaction #2 proved the reaction to be unsuccessful, so another amount and type of phosphine (tricyclohexyl phosphine) was used in reaction #3. NMR analysis concluded the reaction was close to completion but did not quite provide us with the product we were looking for. In reaction #4, a different amount and type of ligand (DIPHOS) was used once again. NMR determined reaction #4 to be closest to completion.

Conclusion In the transfer hydrogenation reactions, through manipulating the starting material as well as temperature, and reflux time, we were able to piece together the most ideal reaction. Percent yield, IR, and NMR analysis helped determine that reaction 9 (as shown in Table 1) yielded the best product. The transfer hydrogenation reaction favored higher temperature, as well as longer reflux time compared to the original reaction. The reaction was also more compatible with potassium carbonate (base) as opposed to HCl in 1,2-propanediol (acid).

140

The Manhattan Scientist, Series B, Volume 6 (2019)

Fernandez + LaPoff

Although the traditional reactions were more time consuming, they proved to work better and more efficiently than the Guerbet reactions. The Guerbet reactions were more complicated and required a lot of adjusting in order to figure out the optimal way to run the reaction so that the desired product was synthesized. The starting material remained consistent in all the reactions, however the type and amount of ligand was changed. NMR analysis helped calculate percent conversion of each reaction. Based on percent conversion the most accurate reaction was reaction number 4 as shown in Table 2. The methyl and ethyl Grignard reactions obtained high percent yields of 76% and 77% respectively. IR and NMR analysis also proved appropriate product formation in each reaction. All the reactions mentioned above require more tweaking, as this project is still ongoing.

Acknowledgments This study was supported by a donation from Kenneth G. Mann ’63. The authors would like to thank Dr. James McCullagh for allowing them to be a part of his research and for guiding them through new methods and concepts.

References Goodscents Flavor Company: Flavor, Fragrance, Food and Cosmetics Ingredients information. http://the-goodscentscompany.com Koda, K., Matsu-Ura, T., Obora, Y., and Ishii, Y. (2009). Guerbet Reaction of Ethanol ton-Butanol Catalyzed by Iridium Complexes. Chemistry Letters, 38(8), 838-839. doi: 10.1246/cl.2009.838 Sigma Aldrich: Aldol Condensation Reaction. (n.d.). https://www.sigma aldrich.com/technicaldocuments/articles/chemistry/aldol-condensation-reaction.htm Surburg, H., and Panten, J. (2006). Common fragrance and flavor materials: preparation, properties and uses. Weinheim: Wiley-VCH. Taguchi, K., Nakagawa, H., Hirabayashi, T., Sakaguchi, S., and Ishii, Y. (2004). An Efficient Direct α-Alkylation of Ketones with Primary Alcohols Catalyzed by [Ir(cod)Cl]2/PPh3/KOH System without Solvent. Journal of the American Chemical Society, 126(1), 72-73. doi: 10.1021/ja037552c Volkov, A., Gustafson, K. P. J., Tai, C.-W., Verho, O., Bäckvall, J.-E., and Adolfsson, H. (2015). Mild Deoxygenation of Aromatic Ketones and Aldehydes over Pd/C Using Polymethylhydrosiloxane as the Reducing Agent. Angewandte Chemie, 127(17), 5211-5215. doi: 10.1002/ange.201411059

Synthesis of commercial fragrance compounds possessing an aromatic ring linked to a polar functional group by a 2 C spacer Jennifer LaPoffâ&#x2C6;&#x2014; and Sharron Fernandezâ&#x2C6;&#x2014; Department of Chemistry and Biochemistry, Manhattan College Abstract. Fragrances come from different compounds that can be produced naturally or created synthetically. In this work, we synthesized commercial fragrance compounds that possess a phenyl ring along with a polar alcohol or ketone connected by a two-carbon spacer. There were two ways of synthesizing the compounds: a traditional method and a new modern method. The traditional method uses reactions learned in second-semester organic chemistry class to be used in a capstone four-week experiment for second-semester organic chemistry laboratory. I optimized the aldol condensations and the hydride reductions along with two transfer hydrosilylations. The new modern method uses a transition metal catalyst that can-do multiple steps in one reaction setup. There were three products seen the desired, its isomer, and the reduced form of the desired product. Products that are produced will be characterized from obtaining and analyzing infrared (IR) spectra and proton nuclear magnetic resonance (H1 NMR) spectra along with comparisons to commercial samples.

Introduction In the fragrance industry, there are a number of naturally occurring and synthetic compounds that have an aromatic ring attached to a polar functional group via a two-carbon spacer. Specifically, in this piece of work, the polar functional groups that are connected to the two-carbon spacer are either an alcohol group or a ketone group [1]. Fig. 1 gives the general structures that can be seen from the previous description. Examples of commercial fragrance compounds containing those specific characterizations can be seen in Fig. 2 below. Each one of the compounds in Fig. 2 possesses a different aroma.

Figure 1. General features of desired odorants

A common way to commercially make the fragrance compounds is using an aldol condensation followed by a transfer hydrogenation and reduction using a hydride source or a Grignard reaction[1]. This general reaction flow would be considered the traditional method of producing the desired product which can be seen below in Fig. 3. â&#x2C6;&#x2014;

Research mentored by James McCullagh, Ph.D.

142

The Manhattan Scientist, Series B, Volume 6 (2019)

LaPoff + Fernandez

Figure 2. General features of desired odorants

Figure 3. General reaction flow for the traditional method

Fig. 3 gives a brief synopsis of the compounds being produced. As it can be seen, in the general reaction flow, a base is used in the aldol condensations, a hydrogen donor is used in the hydrogenations, and a reducing agent to reduce the compounds. An aldol condensation is an addition reaction between aldehydes and/ or ketones producing a Î˛-hydroxyketone or a Î˛-hydroxyaldehyde then a dehydration of the products occurs. A transfer hydrogenation is reducing bonds of a compound using a hydrogen donor and a catalyst. A variation of a transfer hydrogenations is a transfer hydrosilylation, which uses a monomer or a polymer hydrosilane as the hydrogen donor [2, 3]. A hydride reduction uses reducing agents to reduce a compound. A Grignard also reduces a compound, but also adds more carbons to the compound. Beside the traditional method, it has been seen that an iridium metal catalyst has shown the capability to condense the multiple-step reaction into one reaction setup referred to as the new modern method [4, 5, 6]. It can be accomplished in one glassware since multiple reactions are happening at once in pressurized vessels. There are two main reactions are a ketone alkylation with alcohols and the Guerbet reaction. Ketone alkylation reactions start with a ketone and an

LaPoff + Fernandez

The Manhattan Scientist, Series B, Volume 6 (2019)

143

Figure 4. The general ketone alkylation reaction flow.

alcohol to produce a bigger ketone. The standard ketone alkylation reaction conditions are 1mole % [Ir(COD)2Cl]2, phosphine ligand (0.5 - 2 mole %) , 20 mole % KOH, 0.203 mL benzyl alcohol, and 0.766 mL of 2-butanone run for 4 hours at 130â&#x2014;Ś C. The reaction flow of ketone alkylation with alcohols in Fig. 4 [5]. These parameters are changed respectively depending on results obtained and data that was analyzed. The reaction flow of Guerbet reaction can be seen in Fig. 5 [4]. The Guerbet reaction starts with two small alcohols and uses the same iridium catalyst as well along with a base and an alkene to join them together to make bigger alcohols. The spectra collected from the products produced were IR spectra and H1 NMR spectra. IR spectra allows the analysis of key functional groups in the compound produced. H1 NMR spectra takes it one step further and allows you to determine the structure of compounds.

Figure 5. Guerbet reaction flow.

Method and Materials As mentioned, there is a traditional and new modern method in synthesizing fragrance compounds. The traditional method uses versions of reactions that can be used with beginners in the organic chemistry lab. The specific reaction flow for the traditional method is represented in Fig. 6. The initial targets of the reaction flow are lilac pentanol and 4-phenyl-2-butanol. For the traditional method, I performed the aldol condensations and the hydride reductions along with two transfer hydrosilylations. Each compound produced from beginning to end has a different aroma. For the aldol condensations performed, 2 mL of benzaldehyde and 4.5 mL (exp. 1) / 8.4 mL (exps. 2, 3) acetone were added to the flask. The base used was either 10% sodium hydroxide or barium hydroxide. If barium hydroxide was used it had to be finely ground 0.3 g and then it was added to the flask. The reaction refluxed for either an hour or an hour and a half. After reflux, the residual barium hydroxide was removed using vacuum filtration with silica gel. The solvent was evaporated off on a hot plate. If sodium hydroxide was used, 20 mL was added and refluxed for an

144

The Manhattan Scientist, Series B, Volume 6 (2019)

LaPoff + Fernandez

Figure 6. Specific reaction flow for traditional method

hour. Then the reaction was separated using 20 mL of tert methyl butyl ether (MTBE) along with 10 mL of water. The organic layer was extracted with 10mL of saturated sodium chloride. The organic layer was dried and the solvent was evaporated off using a hot plate. To purify the product, using barium hydroxide, it was separated using MTBE and water, then using saturated sodium chloride. The organic layer was dried and the solvent was evaporated off. Then the product from evaporation was filtered using vacuum filtration with silica gel and a mixture of acetone and hexane as the solvent. The solvent was boiled off using a hot plate. For the hydride reduction reaction, 0.26 g of sodium borohydride, 6.8 mL of methanol, and 0.5 g of benzylacetone were added to a flask. The reaction stirred at room temperature for 40 minutes. For reaction 5, the methanol was boiled off and then 10 mL of water was added. Next, the hydride was quenched using 5 mL of hydrochloric acid. The reaction was extracted using 10 mL of MTBE twice. The organic layer was extracted once with 10 mL of saturated sodium chloride. The organic layer was dried and the solvent was evaporated off using a hot plate. For reaction 4, the sodium borohydride was quenched with acid and then the methanol was boiled off then water was added. For the transfer hydrosilylation reaction, 0.5 g of benzylidene acetone, 0.06 g of 5 palladium/alumina catalyst, 9 mL of methanol, and 0.1 mL of concentrated hydrochloric acid were added to a flask. Next, 0.6 mL of the polymer, poly(methylhydrosiloxane) or 2.2 mL of the monomer, triethylsilane, was added and then refluxed for 30 minutes. Steam distillation stripped off the methanol then water was added to distill off the product. The product was extracted twice with 10 mL of pentane, which was dried and evaporated off using a hot plate. The new modern method targets the same compounds as the traditional method along with benzylacetone. The specific reaction flow of the ketone alkylation reactions performed can be observed from Fig. 7. The standard reaction conditions for the ketone alkylationâ&#x20AC;&#x2122;s were 1 mole % [Ir(COD)2 Cl]2 , phosphine ligand (0.5 - 2 mole %), 20 mole % KOH, 0.203 mL benzyl alcohol, and 0.766 mL 2-butanone ran 4 hours at 130â&#x2014;Ś C. The catalyst and phosphine ligands were weighed out in a glove bag filled two times with argon gas. The pressurized tube was taken out of the bag plugged with a rubber septa which was purged once with argon gas. While having the argon gas

LaPoff + Fernandez

The Manhattan Scientist, Series B, Volume 6 (2019)

145

Figure 7. Specific reaction flow for traditional method

running pumped in, the liquid reactants (benzyl alcohol and 2-butanone) were added. Argon gas was then run through the system. The reaction refluxed for 4 hours. After 4 hours, the catalyst was vacuum filtered off using silica gel. Then a rotary evaporator stripped off any solvents. For all of the reactions done, traditional and new modern, IR spectra and H1 NMR spectra were run on the products.

Results and Discussion Since we know what the product is supposed to look like from the traditional method reactions, we can calculate percent yields along with determining the hydrogens of the product using the H1 NMR spectra. Experiments one, two, and three were aldol condensations. The desired compound for the aldol condensations was benzylidene acetone, which has a balsam aroma. The limiting reactant for all three was benzaldehyde. The base used in experiment one was sodium hydroxide and the base used in experiments two and three were barium hydroxide. The percent yield for experiment one was 87.71%. For experiment two, the percent yield was 82.74%. For experiment three, the crude percent yield was 86.21% and the pure percent yield was 44.79%. In these set of experiments, the cleanest and most efficient one was experiment three. In spectrum 1, the pure product from experiment three can be observed including the hydrogen integrations. H1 NMR (CDCl3 ) δ (ppm): 7.46 (m,6H), 6.70 (d, 1H), 2.35 (s, 3H). IR spectrum, spectrum 2, (cm−1 ): 1667.52 (C=O), 1610.15 (C=C), 750.09 (monosubstituted). These spectra give a clean visual of the what the product is to look like from analyzing spectra. Experiments four and five were hydride reductions. The reducing agent used for both was sodium borohydride and the limiting reagent was benzylacetone. The desired compound for the reductions was 4-phenyl-2-butanol, which has a floral peony heliotrope aroma. The percent yields are 50.10% for experiment four and 76.17% for experiment five. In spectrum 3, the H1 NMR spectra of the product from experiment five can be observed including the hydrogens. H1 NMR (CDCl3 ) δ (ppm): 1.274 (d, 3H), 3.210 (s, 1H), 3.776 (quintet, 1H), 1.574 (ms,2H), 7.228 (s, 5H), 2.7 (m,2H). IR spectrum, spectrum 4, (cm−1 ): 3360.36 (O-H), 3026.54 - 3084.37 (sp2 C-H), 2861.35 - 2966.98 (sp3 C-H), 745.47, 699.02 (monosubstituted). Experiments six and seven were transfer hydrosilylation reactions. The desired compound for the hydrosilylations was benzylacetone, which has a floral balsam aroma. The limiting reactant was benzylidene acetone. For experiment 6, a polymer called [poly(methylhydrosiloxane)] was

146

The Manhattan Scientist, Series B, Volume 6 (2019)

LaPoff + Fernandez

Spectrum 1: Experiment 3 aldol condensation H1 NMR spectrum

Spectrum 2: IR spectrum of Experiment 3 - aldol condensation

Spectrum 3: H1 NMR spectrum of Experiment 5 – hydride reduction

Spectrum 4: IR spectrum of Experiment 5 – hydride reduction

Spectrum 6: IR spectrum of Experiment 6 – transfer hydrosilylation

Spectrum 5: H1 NMR spectrum of Experiment 6 – transfer hydrosilylation

LaPoff + Fernandez

The Manhattan Scientist, Series B, Volume 6 (2019)

147

used instead of sodium formate, which was used in the hydrogenations. The percent yield for this experiment was 38.10%. Even though the yield was low, but from analyzing the H1 NMR spectra it can be observed that it was a clean reaction with only the desired product produced. For experiment 7, a monomer was used called triethylsilane and the percent yield was 244.8%. This yield was extremely large because the product acquired contained over reduced product and other impurities thus making it not a pure sample. In Spectrum 5, the H1 NMR of the product from experiment six can be observed including the hydrogens. H1 NMR (CDCl3 ) δ (ppm): 2.120 (s, 1H), 2.748 (q, 2H), 2.872 (q, 2H), 7.210 (s, 5H). IR spectrum, Spectrum 6, (cm−1 ): 3027.32 3084.70 (sp2 C-H), 2863.32, 2926.17 (sp3 C-H), 1715.94 (C=O), 749.72 (monosubstituted). However, for the new modern methods, we calculated the percent yields of the three products formed from calculating the percent conversion of the starting material determined from the NMR integrations. Three compounds could’ve been produced from the ketone alkylations. One was the desired compound, 1-phenyl-3-pentanone, its isomer, 3-methyl-4-phenyl-2-butanone, and the reduced form, 1-phenyl-3-pentanol. The desired product was observed to have a strawberry jam aroma. Experiments 8, 9, 11, 12, 13, 14, and 15 were ketone alkylation reactions, Table 1, while Table 1. Ketone alkylation with alcohols data Experiment

conversion

yield

compound O

H1 NMR

Ligand

Reaction Notes

Yes

PPh3

Benzyl alcohol + acetone

Yes

PPh3

Benzaldehyde + 2-propanol

Yes

PPh3

Benzyl alcohol + 2-butanone

Yes

DPPP 1%

Benzyl alcohol + 2-butanone

CH3 CH3

37%

33.78% O O

40.54%

CH3 CH3

CHO 3 CH3

CH3

O CH3

O OH

CH3 CH3

O O

CH3

97%

75.26%

CH3 CH3

CH3

OH O CH3

98%

7.14%

CH3

CH3 O

CH3

CH3 CH3

483.67%

OH O O

14.97%

CH3 CH3

CH3 CH3 CH3

(continued on next page) OH OH CH3 CH3

148

The Manhattan Scientist, Series B, Volume 6 (2019)

LaPoff + Fernandez

Table 1. (continued from previous page) Experiment

conversion

yield

compound

CH3

H1 NMR

Ligand

Reaction Notes

Yes

DPPM 1%

Benzyl alcohol + 2-butanone

Yes

BINAP 1%

Benzyl alcohol + 2-butanone

Yes

DPPM 2%

Benzyl alcohol + 2-butanone

O CH3

97%

108.76%

O O

CH3

66.5%

91.73%

CH3 CH3

CH3

OH O CH3

81.5%

9.82%

CH3

CH3 O

CH3

CH3 CH3

78.53%

O O CH3 CH3

26.17%

CH3 CH3 CH3

OH experiment ten was a Guerbet reaction, Table 2. It can be seen in the tables below that each reaction CH CH varied with ligands and starting materials. The percent conversions were calculated from using the percent of benzyl alcohol in the mixture. Then the percent yields were found depending on what products were analyzed observing the H1 NMR spectrum. The greatest conversion percentage is 98% coming from experiment 12 meaning that only 2% of starting material was not converted. The best yield for the desired product that was using 1% DPPM ligand from experiment 13 giving less than 1% of impurities causing the yield to be slightly over 100%. The sample spectra are from experiment 15, which contains the three products that could’ve been produced. Spectrum 7 is the H1 NMR spectrum and Spectrum 8 is the IR spectrum. In the NMR spectrum of experiment 15, it can be observed that there are three products formed with slight starting material not converted. In the IR spectra, you can distinctly observe an O-H peak at 3464.24 cm−1 belonging to the reduced form of the desired product. The Guerbet Reaction did not work with our preliminary starting materials and solvents. There was no conversion; i.e. no yield; thus expressing the need to change its parameters including starting material, ligands, or adjusting reflux time.

3 3

Table 2. Guerbet reaction data Experiment

conversion

yield

compound

H1 NMR

Ligand

Starting material

–

Yes

1-octene

Benzyl alcohol + 2-butanol

LaPoff + Fernandez

The Manhattan Scientist, Series B, Volume 6 (2019)

Spectrum 7: H1 NMR spectrum of experiment 15 ketone alkylation reaction containing peaks for all three products analyzed.

149

Spectrum 8: IR spectrum of experiment 15 ketone alkylation reaction

Conclusion The traditional methods used in the beginning part of the research is more or less done and we hope to try it in an organic chemistry lab class in the spring semester. The percent yields calculated from the data obtained overall varied from 38.10% being the smallest to 87.71% being the largest. The desired products for each set of reactions were analyzed and contain few impurities if any. The new modern method seems to be working fairly well for our preliminary test substrates: benzyl alcohol and 2-butanone. Some reactions didnâ&#x20AC;&#x2122;t convert and there were reactions that did convert to products. The best conversion was experiment 12, and the best yield of the desired product only was experiment 13. In the future, we would like to develop a more accurate method for determining percent conversions and product ratios which will involve of HPLC, high-performance liquid chromatography. Along with the use of TLC plates and chromatography columns to help purify by separating the compounds produced. We would also like to test the reaction over a wide range of starting materials involving the use of pressurized vessels for lower boiling starting materials. This would hopefully help get the Guerbet Reaction working starting with two alcohols instead of a ketone and an alcohol.

Acknowledgments This study was supported by a donation from Kenneth G. Mann â&#x20AC;&#x2122;63. The authors would like to thank Dr. James McCullagh for allowing them to be a part of his research and for guiding them through new methods and concepts.

References [1] Bauer, K., Garbe, D., and Surburg, H. (2001). Common Fragrance and Flavor Materials: Preparations, Properties, and Uses. 4th ed. Wiley-VCH, NY.

150

The Manhattan Scientist, Series B, Volume 6 (2019)

LaPoff + Fernandez

[2] Volkov, A., Gustafson, K. P., Tai, C., Verho, O., Bckvall, J., and Adolfsson, H. (2015). Mild Deoxygenation of Aromatic Ketones and Aldehydes over Pd/C Using Polymethylhydrosiloxane as the Reducing Agent. Angew. Chem. Int. Ed. Engl.;54(17):5122-6. doi: 10.1002/anie.201411059. Epub 2015 Feb 26. [3] Tuokko, S., Honkala, K., and Pihko, P. M. (2017). Pd/C-Catalyzed Hydrosilylation of Enals and Enones with Triethylsilane: Conformer Populations Control the Stereoselectivity. ACS Catal. 71, 480-493 [4] Matsu-ura, T., Sakaguchi, S., Obora, Y., and Ishii, Y. (2006). Guerbet Reaction of Primary Alcohols Leading to β-Alkylated Dimer Alcohols Catalyzed by Iridium Complexes. J. Org. Chem. 2006;71(21):8306-8. [5] Taguchi, K., Nakagawa, H., Hirabayashi, T., Sakaguchi, S., and Ishii, Y. (2004). An Efficient Direct α-Alkylation of Ketones with Primary Alcohols Catalyzed by [Ir(COD)Cl]2/PPh3/KOH System without Solvent. J. Am. Chem. Soc. 2004;126(1):72-3. [6] Koda, K., Matsu-ura, T., Obora, Y., and Ishii, Y. (2009). Guerbet Reaction of Ethanol to nButanol Catalyzed by Iridium Complexes. Chemistry Letters 38(8):838-839

Reinforced learning approach to reordering sentences in extractive text summarization William Kulpâ&#x2C6;&#x2014; Department of Computer Science, Manhattan College Abstract. The amount of information is increasing at an unprecedented rate and methods must be created to efficiently analyze and utilize the data. Automatically generating summaries allows the most prominent information to be captured. Reinforcement learning has been proven to be beneficial to text summarization and could lead to more refined and readable summaries. By reordering the structure of processed sections of text, a DQN agent can learn a global policy to generate summaries. Cumulative and individual sentence scores generated by TextRank can allow an agent to improve upon generated summaries when compared to reference summaries. A global policy can be learned based upon the results obtained from training on the Newsroom dataset, but more research should be performed to explore further applications.

Introduction Automatic text summarization is the process of picking the most relevant pieces of text to make a cohesive and coherent summary [1]. Summarization is designed to process, categorize, and generate coherent paragraphs to make text data more digestible. Researchers, scholars, businesses, and others would benefit from better text summarization models. Choosing the sections to be included and excluded is not a trivial task and research continues to challenge the standard of human generated summaries. There are two main types of text summarization, extractive and abstractive [1]. The extractive method focuses on identifying key sentences and sub paragraphs capturing the meaning of the passage. Once these parts have been selected, they are put together and submitted as a summary. The abstractive model focuses on understanding the main points of the article and creating a summary based on natural linguistics. Both models rely on preprocessing of the text for the algorithms to create a summary. Preprocessing starts with common words being removed from the text to avoid unnecessary bias. After removal, an extractive summarizer would apply various algorithms to rank words and sentences. Quality of a generated summary depends on the various attributes of sentences. A simple Theoretical Model of Importance for Summarization, proposes a framework of importance to identify the most valuable information. Importance incorporates attributes of relevancy, position, and ngrams to determine an importance score [2]. Importance can be easily incorporated into text summarizers, like various n-gram summarizers like TextRank, and learned by supervised machine learning. â&#x2C6;&#x2014;

Research mentored by Arafat Abu Mallouh, Ph.D.

152

The Manhattan Scientist, Series B, Volume 6 (2019)

Kulp

An abstractive summarizer converts the preprocessed words into embeddings based on global word datasets through Word2Vec and Flair. Embeddings of words are utilized to make accurate vector representations. These vector representations are utilized in Sequence to Sequence summarizers as input for the encoder, which are then passed to the decoder. The decoder attempts to create a coherent summary from extracted information.

Extractive Extractive text summarization techniques are built upon the analysis of sentence attributes. Extractive methods are implemented with two main steps, the preprocessing and processing step. In the preprocessing step, sentence boundaries are determined, unimportant words are eliminated, and a stemming is performed. Stemming is the process of determining a word’s radix, or origin, to determine its semantics. Generated summaries by extractive techniques, suffer from a lack of coherency and presentation bias, therefore they fail to capture the message of longer articles [3]. Furthermore, it does not attempt to capture the ‘essence,’ since it only uses the sentences available. However, it is still an important technique to study as many of the mechanisms are fundamental to abstractive summarization. Preprocessing, importance scores, key words, and sentence order are pivotal for the function of abstraction. Further research in extractive summarization allows for advances in preprocessing/processing, which can be implemented later in abstractive methods. Additionally, the computation needed to run an extractive summarizer is much less compared to abstractive summarization [3]. There are various algorithms implemented in extractive summarization. A basic extractive summarization technique, referred to as Term Frequency Inverse Document Frequency (TFIDF), performs a keyword analysis and calculates the probability of relevance, a higher score indicates the chance it will be included in the summary [1]. This technique has mostly been incorporated into the cluster methods to better analyze large sets of documents.

Abstractive Abstractive text summarization attempts to generate linguistically original summaries, while retaining the main points. Abstractive summarizers are built upon the popular Sequence-to-Sequence (Encoder Decoder) model. Any sized input can be converted into a fixed length output. These are built with a linear chain of Long Short-term Memory (LSTMs), a modification of the Recurrent Neural Network (RNN). The power of the LSTM is to adjust the information retained and discarded in the node’s memory. By adjusting the weights of the LSTMs, various bits of information can be saved for the decoder to generate words. Another important element of the Encoder Decoder models are attention mechanisms or intradecoder mechanisms, which are a solution to repeating in summaries by functioning like shortterm information. Global and local attention mechanisms attempt to normalize the weights and

Kulp

The Manhattan Scientist, Series B, Volume 6 (2019)

153

information held by all the nodes in the network to generate more controlled summaries [4]. While these models have been shown to be effective, extractive summarization still generate high Rouge scores and do not suffer from repetition or control because of challenges associated with natural linguistics [3]. There are inherent difficulties trying to generate coherent summaries with only semantics. The greatest issue is accurately representing the complex theme in a non-repetitive summary.

Rouge metric The ROUGE metric is a popular metric utilized as a benchmark for the performance of text summarizers. There are various metrics designed to analyze how much of the original content is contained in the generated summary. ROUGE-N scores are made of ROUGE-1, ROUGE-2, and ROUGE-3, measuring the incremental increases in the n-bits between ideal and generated summary [5]. ROUGE-L, ROUGE-W, and ROUGE-S are other metrics designed to measure how well the main idea of a summary is captured. ROUGE-l calculates the Longest Common Subsequence (LCS) and ROUGE-W is the weighted LCS. The metric of ROUGE-S is normally used for abstractive summarization since it calculates the number of skipped bigrams, giving a good representation of how well the language captures meaning [5]. For the purpose of this study, ROUGE-N was primarily utilized to determine how similar the extracted sentences matched the information contained in reference summary. These scores can be further broken down into a precision and recall score to accurately judge the summary. Precision of the ROUGE-N score refers to how well the summary covers information contained in the reference summary. The recall score extends off this by giving a ratio of the reference summaries redundant verbiage to information covered, or the effectiveness of the words.

Reinforcement learning Reinforcement learning is a relatively new part of machine learning (ML), fundamentally different than other approaches in traditional ML. Reinforcement learning is learning how to maximize a numerical reward signal. The learner is not told which actions to take, but instead must discover which actions yield the most reward by trying them. The basic idea of the reinforcement learning is to capture the most important aspects of the real problem facing a learning agent interacting over time with its environment to achieve a goal (Fig. 1). A learning agent must be able to sense the state of its environment to some extent and must be able to take actions that affect the state. Reinforcement learning is different from supervised learning, such that it doesnâ&#x20AC;&#x2122;t have specified output to learn an output/input mapping. Supervised learning has an external knowledgeable supervisor providing a set of mapped examples. The supervisor is trying to implement a function

154

The Manhattan Scientist, Series B, Volume 6 (2019)

Kulp

Figure 1. Overview of the DQN working in the environment (sentences)

based on the data provided. Unsupervised learning attempts to find hidden structure in the provided input data and classify categories. Reinforcement learning attempts to maximize a reward signal and learn a policy based on the strength of the reward [6]. There are various models of reinforcement learning, varying in how information is analyzed, stored, and learned after manipulating the environment. Q-learning is an off-policy model of reinforcement learning, employing an epsilon greedy policy of learning for decision making. The agent will select an action based on previous saved states and select the best option, otherwise choose at random, learning exclusively through trial and error. Other models of reinforcement learning use policy learning in order to learn more efficiently and avoid under/overfitting. These models are known as Deep Q-Network (DQN), Double DQN (DDQN), and State-action-reward-state-action (SARSA). Unlike Q-learning (model free reinforcement learning), a DQN has an implemented deep neural network (DNN) to analyze the state-action pairs produced by the agent to better assist learning. The implementation of deep learning solves multiple problems in reinforcement learning. For example, datasets, especially video input, can contain a huge volume of frames, creating volatility in the learning process. Experience replay samples sections of buffers from the input of the dataset to act more like a supervised network. A DQN also utilizes a target network to correct for an unstable error calculation [7]. These action state-pairs can be utilized in text summarization to maximize the ROUGE reward metric.

Background to problem Abstractive summarizers can generate understandable and accurate summaries for short articles, even with a variety of content. Longer articles, scientific papers, are too expansive and structurally complicated for Sequence to Sequence models and need to be fine-tuned depending on the structure of data [3]. Reinforcement Learning can help improve the process of text summarization. Models imple-

Kulp

The Manhattan Scientist, Series B, Volume 6 (2019)

155

mented in reinforcement learning have proven to be highly efficient and produce state of the art results in video games like Artari [6]. Applications of deep learning have proven to be a great implementation for the analysis of an action space for reinforced learning agents. The problem of text summarization can be translated into a reinforcement learning problem, where the learning agent must explore the environment (text) to search for the best parts that would maximize preferred features such as the ROUGE metric. The agentâ&#x20AC;&#x2122;s goal is to maximize various features of the summaries through a modified framework of ASLR [4]. The goal of the research was to identify key strategies in text summarization and utilize a simple extractive reinforcement model for text summarization. Many reinforced approaches to extractive summarization focus on maximizing the ROUGE metric as a function compiling the summary.

Related work Extractive techniques The fuzzy logic system grants the chance to quantify unconventional mediums like text summarization. The normal preprocessing steps are accomplished and then the sentences are ranked based on the parameters for sentences. A value from 0 to 1 is obtained for each and the summary is constructed from these values. The fuzzy system consists of a fuzzier, inference engine, defuzzifier, and the rule set. In fuzzification, the sentences are converted into linguistic values, which then be passed to the inference engine to gather the linguistic values. Finally, the linguistic scores are converted back into crisp (0 or 1) based on membership rules [1]. In cluster-based extraction, documents are usually based on a logical progression of topics, usually broken explicitly or implicitly into sections. Text clustering groups the sentences of documents into related categories based on various learned/defined categories. After creation of the clusters, sentences are selected depending on how similar they are to the cluster, to avoid cluster bias. Key term frequency, similarity to the first sentence in the document, and position in the paragraph are considered when assigning an importance score to a sentence. The total score will determine the probability a sentence is included in the final summary [1]. A graph-based approach represents sentences as nodes in an undirected graph. Edges are formed between the nodes forming subgraphs form topics. When a user is querying for information, the various subgraphs can be chosen to best reflect the content of the query. The edges connecting each node also indicate the importance of the sentence, more edges indicate a more important sentence. The paths of subgraphs can also play an important part in understanding and optimizing sentence cohesion [1]. TextRank, the algorithm used for preprocessing in this study, utilizes a graph-based approach and selects sentences based on the PageRank score. An importance score is calculated by iterating through random circuits to determine the interconnectedness of nodes [8].

156

The Manhattan Scientist, Series B, Volume 6 (2019)

Kulp

Reinforcement learning has been applied to text summarization. In the literature, Ryang and Abewaka proposed a framework for automatic text summarization using reinforcement learning (ASLR).The ALSR framework proved a global policy can be learned off a reward in text summarization. Unfortunately, the agent normally converges sub optimally, and it is highly dependent on the scoring function and processing preformed [9]. Further research in the field was performed by Narayan [10] and Molla [11] both provided further evidence of the importance and usefulness of the global policy with reinforcement learning. The models and experiments are limited in scope but show promising results in the field traditional supervised approaches. Reinforcement Learning has also been effective in ordering sentences based on embedding scores. Gyoung’s work [12] proves simple features of “meaning and position” are enough to train a global policy in the learning process. The DQN also was trained effectively and obtained encouraging results. Abstractive techniques A popular abstractive summarization technique is the graph-based method. The graph functions in the same manner as extractive techniques, however, the graph is used for more than subgraphs and querying. The graphs are analyzed for links between each node and an optimal path is found. Where it starts to differentiate itself from the extractive method by quantifying and creating relationships between nodes. These nodes will be selected by the algorithm for the summary, but phrases and sentences will be interjected in order to create coherency. Multi module techniques is another technique derived from graph-based theory. The nodes represent concepts and the edges are links between the concepts. The major focus of this technique is Natural Language Generation (NLG), which has a focus on taking different types of input, not only text, converting it to machine language, and creating a report. The algorithm can shift through large amounts of data already collected from various sources and generate a summary. Google’s TextSum is the state of the art open sourced abstractive text summarizer, which utilizes deep learning in order to create titles from the first two sentences of a document. This Encoder Decoder model has been used for a variety of summarizer models and further improved upon. An example of the system is simpleNLG, which is an open sourced Java API, which generates simple summaries based on input. While there are more complete abstractive multimodal systems available. However due to the open source nature of simpleNLG, the algorithm creates very complex structures, which can be developed into more complex synaptic structures [3]. Reinforcement learning has been applied in the Encoder Decoder model for abstractive summarization. The work by Paulus, implements a Reinforced training loop to train the novel-attention encoder decoder [13]. The best results were obtained with a mixture of intra-attention and RL training, compared to combined reinforced learning and machine learning.

Kulp

The Manhattan Scientist, Series B, Volume 6 (2019)

157

Methodology In this work, deep reinforcement learning was utilized for text summarization. The following objectives were explored: • • •

•

Investigate state of the art text summarization strategies Analyze the benefits of a DQN agent in extractive text summarization. Learn if the DQN agent can learn a global policy for sentence selection sequence based on the structure of a summary. If so, analyze and optimize the deep reinforcement learning algorithm for text summarization

DQN Agent The DQN agent utilized in this work is the model provided by the deep learning package PyTorch built using python. The tutorial provided by Adam Paszke, is a simple DQN implementation in python. The learning is in deterministic and does not include any further stochastic policy for transitions in the reward signal. The Q function conforms to the Bellman equation and is minimized in error by the Huber loss correction [7]. A convolutional neural network (CNN) is implemented in the agent and attempts to predict the sequence of actions based on the input. The output will be learned based on the strength of the reward returned from the environment. A DQN utilizes replay memory to reduce stochastic training decisions. Batches are randomly sampled to decorrelate information further preventing early convergence while training. Replay memory has shown to greatly improve the performance of the agent (Fig. 2). Replay memory is an

Replay Memory Training Loop 1. Choose random / policy action 2. Sample Environment 3. Record memory update 4. Optimize 5. Occassional target_net update

Random Batch

Optimize (4) Choose action

Policy Net Update (5) Target Net

Figure 2. DQN loop of training progression based upon replay learning [7]

158

The Manhattan Scientist, Series B, Volume 6 (2019)

Kulp

TextRank Preprocessing An article from the Newsroom dataset is preprocessed and the sentences with scores are returned to the environment

Environment The environment passes the score vector to the agent.

action

DQN Agent Agent iteratively chooses whether to include the current sentence based on the score vector and score total.

Next Summary

Reward The rouge-1 F1 metric is calculated after compilation of summary. The score is returned to agent.

Store Transition Agent stores the actions it takes to compile the summary and the final score.

Figure 3. Overview of learning loop for the DQN agent

important element of a DQN agent, it solves the problem of convergence on irrelevant information for the agent. By randomly sampling the stored transitions of the action-state pairs, an agent can ensure a more stochastic data distribution. This will ensure the agent does not train with information with links within the batch [14]. The Automatic Summarization using Reinforcement Learning (ASLR) algorithm was modified for this study (Fig. 3). The algorithm was chosen as it turns extractive summarization into a solvable reinforcement problem. Without iterating over the same summary multiple times in the original ASLR framework, the agent will only have one chance to choose the sentences. TextRank preprocessing Preprocessing of text was performed with the TextRank, which included stop word removal, common words (similarity) calculations, and graph-based analysis. The selected sentences were appended with scores and used to reset the environment for the DQN agent.

Kulp

The Manhattan Scientist, Series B, Volume 6 (2019)

159

The preprocessing was stop word removal and PageRank analysis to return a list of importance scores. A PageRank is performed after the similarity between every sentence pairing is complete and a graph is constructed. Random iterations are then performed over the weighted graph to rank the sentences. When the environment is reset all sentences are returned with attached scores. Environment The environment for the DQN agent consists of available sentences from the preprocessing of the article. When the environment is reset, the selected sentences are available for the agent to choose to include into the current summary. Once the length of the summary exceeds the reference length, the agent is done making actions. A reward is calculated and stored as a transition in memory, along with all the sentence actions. A transition is recorded for every action the agent makes. When a sentence is added to the summary, the previous importance score and updated score are added to the agent’s transition. A reward is granted when the summary exceeds the length of the reference summary. Table 1 is an overview of pseudo code which demonstrates how the agent stores actions. Table 1. Code for the progression of actions in the environment (modified ASLR) [9] Input: Document D = [1 xn] while not done s ← (∅, ∅, 0) //initial state e=0 while s is not terminated do a˜p(a|s; θ, τ ) //selects action with current policy 0

s , u ← execute(s, a)

//observes next state and receives score update θ ← θ + αδe //learning with current policy p ← storetrans (scoreprim , scoresec , a) //store transition s → s0 //update environment

end while

0 s , r ← quality(s0 )

//reward calculated for compiled summary p ← storetrans (scoref inal , r) //reward stored

end while

The agent can either choose to include the sentence or pass over it. If the total length of the summary does not exceed the reference summary, it will pass through the sentences again. The agent will try to find a global policy within the structure of various summaries across the dataset

160

The Manhattan Scientist, Series B, Volume 6 (2019)

Kulp

used. A global policy can be learned off individual scores of sentences [9], however the agent will attempt to learn based off the structure of a summary and chose sentences to include.

Rewards The reward for the agent is the ROUGE-1 F1 metric. The reward will be passed to the agent after the summary is compiled. The full state will be stored as a transition for the agent, including the TextRank scores, cumulative score, and reward. As the agent increases the Rouge score a larger reward signal will be granted to the agent and will be likely to choose actions to get larger rewards. The structure of the importance score vector will be information for the agent to choose sentences to include in the summary.

Experiments The model was trained on the Newsroom training dataset hosted by Cornell University. The data set has the full articles and the corresponding reference summary. The rouge-1 score was utilized as the reward when the agent was done compiling a summary. The Newsroom dataset contains 1.3 million articles written by authors from various publications. These are used by summarization models to maximize various ROUGE metrics. The agent was trained on the large training dataset and tested on the test set [15]. The following parameters were passed to the agent for training, gamma = 0.99, epsilon = 1.0, alpha = 0.003, batch size = 64, number of actions = 2, and input dimensions = [1]. The parameters were chosen arbitrarily and could be analyzed further for possible performance increase and a decrease in training time. The gamma value specified how much a single training sample weighs on the learning of the agent. The batch size specifies how many samples (summaries) are trained on in one training iteration and kept in memory. The agent can either choose to include the next sentence or not, so the agent can take 2 actions. The input dimensions were specified at 1 because a vector of importance scores was 1-dimensional. The epsilon and alpha values relate to learning rate of the agent. The learning rate (alpha) specifies the effect one sample has on the agentâ&#x20AC;&#x2122;s learning. The epsilon value further allows for control over whether decisions are made on memory or at random.

Results and Discussion There is a recent upsurge in popularity of text summarization due to large amounts of data being accumulated. Extractive and Abstractive strategies have both shown promise in condensing the information for better analysis. Abstractive summarization is becoming the more dominant form of summarization, since natural linguistics are created from the interpretation of available data. However, extractive summarization is valuable to the abstractive technique. Abstractive frameworks can benefit from better words embeddings, keyword analysis, sentence ranking, and

Kulp

The Manhattan Scientist, Series B, Volume 6 (2019)

161

other attributes. Additionally, when reinforcement learning is added to text summarization, more coherent summaries can be produced. Reinforcement learning has also been proven to further increase summarization model performance. The proposed framework for extractive summarization in this study intended to find if a DQN agent could learn a global policy for extractive summarization based on a vector of importance scores available in the environment. A simple DQN model was used for the experiment and all preprocessing was performed by Textrank (Fig. 2). Transitions were stored as an increase in score based on sentence selected. The agent received a reward when the summary exceeded the length of the reference summary. The agent learns in a framework like the structure of ASLR, except no repetition of summary compilation and individual scores (Table 2). The Cornell Newsroom dataset was used to train and test the agent’s performance on extractive summarization. Table 2. Average Rouge-1 F1 score on Newsroom dataset subset model comparison [15]

TextRank Abs-N Pointer-C Pointer-S Pointer-N RL Approach

Extractive

Mixed

32.43 6.08 28.34 37.29 39.11 33.16

22.30 5.67 20.22 23.71 25.48 18.59

Based upon the results of training on the mixed Newsroom dataset, the agent averaged an 18.59 Rouge-1 F1 score and a 33.16 Rouge-1 F1 score on extractive summaries (Table 1). While the mixed performance does not match up to TextRank’s published score of 22.30. The performance of the agent is equivalent to the Pointer C model (20.22), falling short of the Pointer-N (25.48) and Pointer-S (23.71) models for the mixed dataset (Table 2). The Sequence to Sequence model with attention falls short with low scores for all reported datasets. Under the extractive portion, the agent does much better against other models reported in the newsroom results. TextRank achieves a score of 32.43 for the F1 rouge-1 score, which is close to the other Pointer C (28.34), Pointer S (37.29), and Pointer N (39.11). Deep Transfer Reinforcement Learning for Text Summarization performed well on the mixed Newsroom dataset (21.39) [16], outperforming the simple DQN framework on the mixed dataset. The training curve for the extractive subset (Fig. 5) and slope (Table 3) displays a slight increase in scores over the course of the first 50 summaries. The epsilon did converge quickly after training started, even when adjusted significantly. The dataset is large, and convergence will likely happen before all samples are processed. From Fig. 4, it is evident when the agent attempted to compile an abstractive summary, it would lead to a reward close to 0. The stored transitions with 0 probably biased the agent’s sentence policy selection, which lead to an increase in stochastic

162

The Manhattan Scientist, Series B, Volume 6 (2019)

Kulp

training and noisy memory batches to analyze (Fig. 4), resulting in a negative slope for the mixed dataset (Table 3).

Figure 4. Sample of training progression on mixed Newsroom dataset

Figure 5. Training progression on extractive subset of Newsroom dataset

Table 3. Slopes of learning on different subsets of the Cornell Newsroom dataset Slope (Rouge/Summary) Mixed Newsroom Dataset Extractive Subset Newsroom Dataset

-1.7386 E-6 1.96268 E-6

The training curve of the agent over the extractive subset versus the mixed newsroom dataset, showed a slight increase in stabilization. Training on the extractive subset of the newsroom dataset produced a slope of 1.96 E-6 (Table 2). Based on less biased samples, the agent was able to make better decisions and only needed a 1/3 the number of samples. The hyper parameters discussed in the methodology were chosen arbitrarily and could be investigated further for future improvements. The agentâ&#x20AC;&#x2122;s learning could probably be improved based on changes to learning rate and learning policy. The models and importance structure (TextRank vector) used are relatively simple implementations. These serve as a foundation for further research in more complex and robust models. While the DQN learning policy and network were relatively simple, the results could be further improved by using state of the reinforcement learning models. There are various implementations of sentence analysis in text summarization. The TextRank algorithm is a simple PageRank analysis of the sentences. There are proposed and implemented models using detailed sentence, keyword, and structure analysis. Implementing these would capture more information for the RL agent to learn and apply.

Kulp

The Manhattan Scientist, Series B, Volume 6 (2019)

163

The newsroom dataset is one of the only available free to use datasets available for researchers. Further consideration on how the dataset is constructed is another area for performance enhancement. A dataset should be conducive to training based on structure of an article. The agent’s stochastic learning curve is attributed to summaries which far too often many produce low scoring transitions. The proposed ASLR framework proved a global policy could be learned by a RL agent [9]. This work further proves a RL agent can learn a policy based on different information and in a different framework. Further research should be dedicated to increasing the performance of RL in text summarization.

Conclusion Reinforcement learning has practical applications for text summarization. In this research a DQN agent was utilized for extractive text summarization. The objective was to see if the agent could learn a policy for sentence selection based on a vector of sentence scores. The training curve has many stochastic jumps in training on extractive subset, but there is an increase in overall ROUGE-1 F1 scores over all samples. On the extractive subset, the agent scored a 33.16 rouge-1 F1 score, surpassing TextRank results. Though the model does not produce state of the art scores, the framework for text summarization provides further support for investigating reinforcement learning in text summarization.

Acknowledgments This work was supported by the School of Science Research Scholars Program. The author thanks Dr. Abu Mallouh for serving as his advisor.

References [1] C. Saranyamol and L. Sindhu, “A survey on automatic text summarization,” Int. J. Comput. Sci. Inf. Technol, vol. 5, no. 6, pp. 7889-7893, 2014. [2] M. Peyrard, “A Simple Theoretical Model of Importance for Summarization,” in Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 2019, pp. 10591073. [3] N. Moratanch and S. Chitrakala, “A survey on abstractive text summarization,” in 2016 Intern. Conference on Circuit, power and computing technologies (ICCPCT), 2016: IEEE, pp. 1-7. [4] A. M. Rush, S. Chopra, and J. Weston, “A neural attention model for abstractive sentence summarization,” arXiv:1509.00685, 2015. [5] C.-Y. Lin, “Rouge: A package for automatic evaluation of summaries,” in Text summarization branches out, 2004, pp. 74-81. [6] V. Mnih et al., “Playing atari with deep reinforcement learning,” arXiv:1312.5602, 2013. [7] A. Paszke. “Reinforcement Learning (DQN) Tutorial.” https://pytorch.org/tutorials /intermediate/reinforcement q learning.html (accessed.

164

The Manhattan Scientist, Series B, Volume 6 (2019)

Kulp

[8] R. Mihalcea and P. Tarau, “Textrank: Bringing order into text,” in Proceedings of the 2004 conference on empirical methods in natural language processing, 2004, pp. 404-411. [9] S. Ryang and T. Abekawa, “Framework of automatic text summarization using reinforcement learning,” in Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, 2012: Association for Computational Linguistics, pp. 256-265. [10] S. Narayan, S. B. Cohen, and M. Lapata, “Ranking sentences for extractive summarization with reinforcement learning,” arXiv preprint arXiv:1802.08636, 2018. [11] D. Mollá, “Towards the Use of Deep Reinforcement Learning with Global Policy For Querybased Extractive Summarisation,” arXiv preprint arXiv:1711.03859, 2017. [12] G. H. Lee and K. J. Lee, “Automatic Text Summarization Using Reinforcement Learning with Embedding Features,” in Proceedings of the Eighth International Joint Conference on Natural Language Processing (Volume 2: Short Papers), 2017, pp. 193-197. [13] R. Paulus, C. Xiong, and R. Socher, “A deep reinforced model for abstractive summarization,” arXiv preprint arXiv:1705.04304, 2017. [14] V. Mnih et al., “Asynchronous methods for deep reinforcement learning,” in International conference on machine learning, 2016, pp. 1928-1937. [15] M. Grusky, M. Naaman, and Y. Artzi, “Newsroom: A dataset of 1.3 million summaries with diverse extractive strategies,” arXiv preprint arXiv:1804.11283, 2018. [16] Y. Keneshloo, N. Ramakrishnan, and C. K. Reddy, “Deep Transfer Reinforcement Learning for Text Summarization,” in Proceedings of the 2019 SIAM International Conference on Data Mining, 2019: SIAM, pp. 675-683.

Computing directions of maximal skew in a dataset Michael Rozyckiâ&#x2C6;&#x2014; Department of Computer Science, Manhattan College Abstract. The direction in which the skew of a dataset is maximized was computed for the purpose of use with binary classification. In the computation of the directions of maximal skew in a dataset, computational tools including MATLAB, Python, Singular, Macaulay2, and Julia were used. The result of the computation was then demonstrated visually using components of a dataset exhibiting skew. The result was also used in combination with a logistic regression classifier.

Introduction Importance of Skew The skewness of a real-valued random variable is a measure of the asymmetry of its probability distribution about its mean. Examples of skew found within univariate data are visualized in Fig. 1. In Fig. 1(a), the data visually has two regions, the dense left section and the elongated section forming a tail to the right of the dense section. This is positively skewed. Fig. 1(b) contains a dense right section, and an elongated section forming a tail to the left. This is negatively skewed. The asymmetrical spread of the data is the characteristic that this research seeks to maximize for use in data classification. The data used in Fig. 1 are components of the Pima Indians diabetes dataset [1]. This dataset as a whole contains significant skew, and contains 8 attributes which are associated with whether the patient tested positive or negative for diabetes. These two attributes were selected for the demonstration of the skewness of data because they are the most positively and negatively skewed components, respectively. In a previous study of data classification, DeBonis compared the accuracies of a classification method when it considered skewness and when it did not, and found that incorporating the skewness of each component of the dataset for classification reduced the error rate [2]. It is possible that the projection of multivariate data in the direction of maximal skew will produce similar benefits. The present research explores the computational process of finding the direction of maximal skew in a dataset, and by doing so creates the means for future research to consider the direction of maximal skew for multivariate data of any number of attributes for classification. Data classifiers and binary classification A data classifier is a tool used for sorting data into various groups. One way of creating a data classifier is by using a dataset containing data points where the desired categorizations, called classes, are already known for each data point. Since the classes are known in the considered â&#x2C6;&#x2014;

Research mentored by Lawrence C. Udeigwe, Ph.D.

166

The Manhattan Scientist, Series B, Volume 6 (2019)

Rozycki

500

200

Frequency

400 300 200

100

100 0

150

225

300

375

450

525

600

675

750

Value Range

825

900 More

Value Range

Figure 1. Histograms of Pima Indian Diabetes dataset components 5 and 3, respectively. These are measurements of 2-hour serum insulin levels (µU/mL) (left); and diastolic blood pressure levels (mm Hg) (right).

example, the statistical properties of the data can be used to determine a pattern for each class. The class to which a new data point belongs can then be predicted by how well the new data fits to one of the patterns. A binary data classifier, such as the logistic regression classifier, relies on the separation of two classes, say C1 and C2 , within a given dataset D, such that D = C1 ∪ C2 . The better the separation of the classes, the more accurate the classification of new data will be. A poorly separated dataset may lead to poor classification performance. Binary classification algorithms are applicable to bioinformatics, where a set of patient diagnostics is used to aid in the detection of conditions such as diabetes [1]. The computation of the direction of maximal skew is important so that all avenues for increasing the accuracy of data classification may be explored and used. Organization of the paper In the next section, the method used to find the direction of maximal skew in a dataset is explained. Then the logistic regression classifier is reviewed, as well as how the direction of maximal skew is used with the classifier. In the section that follows, the results of this research are presented. This is then followed by a section in which the tools used to compute the direction of maximal skew are reviewed and their purpose is explained. In the final section are the observations made on the research and the possibilities stemming from the research.

Mathematical Methods and Algorithms Find the direction of maximal skew The direction of maximal skew for a dataset D is obtained through the following algorithm adapted from previous work by DeBonis [3]. 1. Shift the points of the dataset, D, such that the mean of each attribute becomes 0. 2. Sphere the dataset to obtain a sphered data matrix in the following way: (a) Find the eigenvalues of DT D = [λ1 , λ2 , ..., λd ] and let Λ be a matrix whose diagonal is made up of these eigenvalues.

Rozycki

The Manhattan Scientist, Series B, Volume 6 (2019)

167

(b) Find the eigenvectors of DT D, {v1 , v2 , ..., vd }, and let P be the matrix whose columns are these eigenvectors. 1 (c) Let S represent the sphered data matrix and compute S = DP Λ− 2 3. Use the sphered data matrix to create a system of polynomial equations in the following way:     x1 x1  x2   x2      (1) [A1 x1 , A2 x2 , · · · , Ad xd ]  ..  =  ..  . . xd xd Where Ak = S T Dk S and Dk = diag(S1k , S2k , . . . , Snk ) 4. Solve the system of equations to obtain the solution vectors of the form:       x11 x21 xn1 x12  x22  xn2       ~ ~1 =  ~ X  ..  , X 2 =  ..  , . . . , Xn =  ..   .   .   .  x1d x2d xnd

(2)

(3)

5. Find which of the real valued solutions maximizes the function:

F (X) = X T x1 A1 + x2 A2 + . . . + xd Ad X

(4)

and call it X ∗ . 6. Finally, the direction of maximal skew is computed as: V = P T X∗

(5)

Logistic regression classifier In this research the binary classification technique of logistic regression was used with the direction of maximal skew to attempt classification. Logistic regression is performed using a sigmoid function 1 g(a) = 1 + e−a which is capable of classifying nonlinear relationships. A graph of the logistic curve is shown in Fig. 2. It is known that g(−a) = 1 − g(a), and observed that 0 < g(a) < 1 for any real number a, and as such the sigmoid function can be used to express probability.

168

The Manhattan Scientist, Series B, Volume 6 (2019)

Rozycki

Figure 2. The Sigmoid function curve. This curve is used to separate the data belonging to the binary classes such that data on one side should belong to the first class and the data on the other side belong to the second class.

The logistic regression classifier uses the sigmoid function and its ability to express probability in the following way: • Provided a dataset D = {x̃1 , x̃2 , . . . , x̃N } with the target values T = {t1 , t2 , . . . , tN }, where ( 1 if xi ∈ C1 ti ∈ {0, 1} and ti = 0 if xi ∈ C2 a weight vector, w, can be trained. • The weight vector is trained using the method of gradient descent [4]: wτ +1 = wτ − rτ ∇E(wτ ) where ∇E(w) =

(6)

N X (yn − tn )xn . n=1

Here r is the learning rate and yn = g(w xn ) • The weight vector is the result of a series of calculations, each using the result of the calculation before it, with the initial calculation using a predetermined value. This iterative process is captured in equation 6 where τ denotes the current iteration and τ + 1 denotes the next iteration. This process considers points of the training dataset repeatedly until an optimal weight vector is reached. • The weight vector is used to weigh a new point, xj , and determine to which class that point belongs. This is determined by the following method: if yj = g(wT xj ) ≥ 0.5 then it is said that xj ∈ C1 , otherwise xj ∈ C2 . T

Incorporating direction of maximal skew for classification As previously reviewed, a binary classifier relies on separating data into two classes. The direction of maximal skew is used in the following way. The dataset is projected onto the plane

Rozycki

The Manhattan Scientist, Series B, Volume 6 (2019)

169

where the skew of the data is maximized. The projection of data onto a plane can be imagined as a cloud, representing the data, hovering over the ground, representing a plane. The water of the cloud is changed to snow, and it collapses onto the ground. Rather than trying to observe each drop of water in the air, each snowflake is considered on the ground. The logistic regression classifier is then applied to the projected data. Since the data was projected onto the plane of the direction of maximal skew, the asymmetry of the data was maximized. With the asymmetry maximized, it is predicted that this will aid the classifier in separating the data by positioning the data in a way that keeps the data from each class away from each other. This serves as an alternative to principal component analysis, which finds the directions of maximal variance for projections of a dataset. The direction of maximal variance is the direction of the plane where a dataset will exhibit maximal variance among the data. Maximal variance is when the distance among the data points is greatest. This provides a data distribution where there is minimal overlapping of points. The role of the direction of maximal skew or maximal variance is like that of a client at a clothing store. The client provides a description of their styles to a wardrobe stylist. The better the description provided to the stylist, the better the chance the client will receive a wardrobe which meets their desire. The stylist is the classification algorithm, while the client is the projection providing a description. If the stylist simply looked at all of the clientâ&#x20AC;&#x2122;s outfits, it may be difficult to find a pattern. But by providing a good description, the stylist can sort different outfits to each of the styles with more accuracy. So when the client next returns and asks the stylist for a certain style, the client will be shown exactly what they would like to see and not waste any time.

Results Direction of maximal skew for 2-dimensional data To obtain the maximal direction of skew, all solutions to the polynomial system of equations created in step 4 of finding the direction of maximal skew must be found. Once found, the maximal direction can be determined. For 2-dimensional data, the direction of maximal skew can also be found by testing all possible directions for skew values. This was done considering two components of the Pima Indians dataset for the purpose of testing the computed direction of maximal skew. These components showed the most skew, one being most positive skew and the other being most negative. The objective of this experiment was to produce a result which could be confirmed through attempting all directions, as well as through visual analysis. Fig. 3 shows the skew of the data when projected in every possible direction, as well as the point corresponding to the direction of maximal skew. The direction of maximal skew obtained in Fig. 3 is drawn in Fig. 4 alongside the data. In these figures, it is observed that the computed direction of maximal skew is a direction in which skew is maximized. The computed direction corresponds with a maximal point on the curve in Fig. 3. And in Fig. 4 a visual relationship can be seen where the projection of the data points onto the illustrated line on the direction of maximal skew would produce a skewed distribution.

170

The Manhattan Scientist, Series B, Volume 6 (2019)

Rozycki

Figure 3. Skewness of the data vs. direction of skew (theta), when projected in that direction. The direction of maximal skew as computed with the methods above is marked with a diamond.

Figure 4. The direction of maximal skew represented by a line and the data which it maximizes the skew of when the data is projected onto the direction.

Directions of maximal skew The skewness of the components of a multivariate dataset provided insight about the skew of the dataset as a whole. Used in this research were the UCI Machine Learning Database datasets labeled Appendicitis, Pima Indian Diabetes, and Indian Liver [5]. These are some of many bench-

Rozycki

The Manhattan Scientist, Series B, Volume 6 (2019)

171

mark datasets often used when testing new machine learning techniques. These specific datasets were selected because of the skew present in each of the attributes of the datasets. Each dataset has two classes to which data may belong. In Table 1 and Table 2 the skew of the individual components within the Pima Indian Diabetes and Appendicitis datasets, respectively, were computed for each class within each dataset, as well as for each dataset as a whole. Table 1. The skew of each component of the Pima Indian Diabetes dataset. Here, the skew of each component is considered individually. These are used as a preliminary assessment of the skew of a dataset. Class

Skew 1

Skew 2

Skew 3

Skew 4

Skew 5

Skew 6

Skew 7

Skew 8

1.1108

0.1726

-1.8044

0.0311

2.4912

-0.6639

2.0002

1.5669

0.5009

-0.4928

-1.9327

0.1153

1.8335

0.0006

1.7127

0.5784

All

0.8999

0.1734

-1.8400

0.1092

2.2678

-0.4281

1.9162

1.1274

Table 2. The skew of each component of the Appendicitis dataset. Here, the skew of each component is considered individually. These are used as a preliminary assessment of the skew of a dataset. Class

Skew 1

Skew 2

Skew 3

Skew 4

Skew 5

Skew 6

Skew 7

0.0735

-1.5392

-0.1052

1.2950

1.8570

-1.6110

-0.0126

2.3203

-0.0047

2.1656

1.8569

1.2974

-0.0115

2.3726

All

0.2750

-1.1425

0.0586

1.3850

2.0116

-1.1950

0.2028

A visualization of the skew and how it may differ between the classes within a dataset is seen in Fig. 5, where attribute 7 of the Appendicitis dataset is considered. Looking at this attribute, the distribution of the data points differ between the two classes. Those points which belong to the first class experience little skew, while those points which belong to the second class experience greater skew. This shows that the skew of this attribute is beneficial to the separation of the two classes.

Figure 5. Histograms of attribute 7 of the Appendicitis dataset. (a) Presents the distribution of the data points belonging to class 1 of the dataset, while (b) presents the distribution of the data points belonging to class 2.

Details on the datasets originating from the UCI machine learning database [5] which were

172

The Manhattan Scientist, Series B, Volume 6 (2019)

Rozycki

used for testing are found in Table 3. Here, the number of data entries for each dataset as well as the number of attributes contained within each dataset are presented. Table 3. Table of datasets from the UCI machine learning database. Dataset Appendicitis Pima Indian Diabetes Indian Liver

Instances

Attributes

106 768 579

7 8 9

For each dataset, the logistic regression classifier was applied. Tests were performed using the direction of maximal skew, as well as with the direction of maximal variance as a baseline for comparison. The classes of the Appendicitis dataset are whether the patient has appendicitis or does not have it. The classes of the Pima Indian Diabetes dataset are whether the patient has diabetes or does not have it. The classes of the Indian Liver dataset are whether a patient has a liver disorder or does not have one. The results of this are located in Table 4. Table 4. Average rate of incorrect classification after performing 10-fold cross validation 10 times with the Logistic Regression classifier. Dataset Appendicitis Pima Indian Diabetes Indian Liver

Direction max skew

Max variance

15.0 ± 8.7% 33.3 ± 7.9% 28.7 ± 2.5%

15.4 ± 7.3% 25.4 ± 4.5% 29.2 ± 3.2%

The classification algorithm was run a total of 100 times for each method (direction of maximal skew, direction of maximal variance) by using 10-fold cross validation. 10-fold cross validation works by taking a dataset and shuffling the data. Then the data is split into 10 groups. Testing is then performed by using only 1 of the 10 groups to train a classification model, and then testing its accuracy with the other 9 groups. This is repeated 10 times in total, each time using a different group for training, and the mean performance of the algorithm is recorded. Computational time The computation time required to solve multi-polynomial equations grows at an exponential rate with respect to the number of equations and variables. The run time for solving one set of multi-polynomial equations was measured to provide insight on the feasibility of repeatedly solving these systems for directions of maximal skew. The results of these measurements using two different Computer Algebra System environments are shown in Fig. 6.

Computational tools

The computation of the direction of maximal skew was performed using tools from MATLAB, Python, and a Computer Algebra System (CAS). Each was selected for its set of tools that the

Rozycki

The Manhattan Scientist, Series B, Volume 6 (2019)

173

20,000

Times (sec)

15,000

Julia Macaulay2

10,000

5,000

0 0

Number of Attributes Figure 6. Computation time comparison of Computer Algebra Systems Julia and Macaulay2 using numerical methods to solve the polynomial system of equations resulting from equation 3 in finding the direction of maximal skew.

others did not have. This process is illustrated in Fig. 7.

Figure 7. Flow diagram of the implementation for computing the direction of maximal skew in a dataset.

MATLAB and Python MATLAB provides many tools which simplify mathematical computation. For this reason, MATLAB is the primary environment in which this research was conducted. Steps 1, 2, 5, and 6 of finding the direction of maximal skew, as well as the logistic regression classifier, are run in MATLAB. In order to solve the nonlinear polynomial system of equations, found in step 4 of the algorithm, the assistance of a Computer Algebra System is needed. While step 3 of the algorithm could be performed within MATLAB, the tools found within Python reduce the complexity of this task considerably. As such, the data is transferred from MATLAB to Python, which then completes step 3 by generating the polynomial system. Python

174

The Manhattan Scientist, Series B, Volume 6 (2019)

Rozycki

accomplishes this through the use of the SymPy library. SymPy is a Python library containing some Computer Algebra System components, particularly supporting symbolic mathematical computation and symbolic equations which can be easily converted to strings. Python is used to generate a script containing the polynomial system which is provided to a Computer Algebra System to solve. After the solutions are found, Python filters these and returns to MATLAB the real-valued solutions. Computer Algebra Systems In this research, we chose among Singular [6], Macaulay2 [7], and Julia [8] as our Computer Algebra System to solve the polynomial system in step 4. Here, we will review the differences each CAS presents. Singular employs symbolic mathematical approaches to find the solutions of the polynomial system. This provides exact results, meaning that it is guaranteed to find all solutions to the polynomial system. This provides the direction of maximal skew as needed for the success of the algorithm. In implementing with Singular, we found that it was not a feasible approach. Singular’s engine is designed to compute the solutions serially, unable to make use of significant computing tools such as GPUs, multi-core CPUs, multiple CPUs, or threading. In addition to this the time required for Singular to solve one polynomial system exceeds an hour when the system is composed of 6 variables. One need only remember the computational complexity follows an exponential growth rate to recognize the impracticality of Singular for finding the direction of maximal skew in a dataset with more than 6 attributes. Investigation into the parallelization of the algorithms within Singular found that such a task would not be possible within the scope of this research. In contrast to Singular, Macaulay2 and Julia use the approach of Homotopy Continuation to solve the polynomial system. Homotopy Continuation is a numerical approximation method to find solutions of various problems [9]. Homotopy Continuation performs significantly better than the approach used by Singular (see Fig. 6). The number of attributes reaches 16 before homotopy continuation requires as much time to solve a polynomial system as Singular’s algorithm needed for 6. Homotopy Continuation approximates and regresses towards the solution of what it is solving. While the mathematics behind this method may not be straight forward, one may consider the example of a successive approximation register, used in the process of converting an analog signal to a digital signal, to understand numerical approximation. When an analog voltage signal is received, it may have any value because of the continuous nature of analog readings. This is problematic for a computing system requiring a digital value composed of 0’s and 1’s. A computer can only store a value to a finite precision, which varies by architecture. For this example let us consider a 4 bit register. Suppose that the most accurate approximation of the analog value is 0101. The successive approximation register (SAR) begins by filling the most significant bit with a 1. The SAR now contains the value 1000. This is compared back with the analog value, which we

Rozycki

The Manhattan Scientist, Series B, Volume 6 (2019)

175

may observe to be the target value. The approximated value is larger, so the SAR reverts the bit to 0. It now fills the second most significant bit with 1. The SAR contains 0100, while the target is at 0101, thus the SAR value is smaller. The 1 is retained and the SAR fills the next bit with 1. The SAR contains 0110, which is larger. Thus the SAR reverts the bit to 0 and is back at 0100. The SAR fills the final bit with 1. The values are compared, and it is found that they are approximately equivalent. Homotopy Continuation is vastly more complex in nature, but applies the principle of numerical approximation to find the solutions. The performance of Julia differs from Macaulay2 in two primary ways. The first of these is that in testing, Julia required approximately 30 seconds to initialize its environment before beginning to solve the system, while Macaulay2 was immediate. This difference made Macaulay2 more efficient for datasets containing fewer than 13 attributes. At the point of 14 attributes, the second of the differences is observed. Both Computer Algebra Systems apply algorithms which primarily compute serially. After the core of the computation process, which runs in serial, is complete, there is a shorter component to each algorithm which runs in parallel. The parallel component of Macaulay2â&#x20AC;&#x2122;s algorithm requires a period of time to complete, while Juliaâ&#x20AC;&#x2122;s completes in only a moment. This makes Julia more efficient than Macaulay2 for datasets with 14 or more attributes. Hardware and system environment This research was performed using the Manhattan College School of Science Dionysus supercomputer, containing 24 CPU cores with x86 64 Architecture and 64GB RAM running an Ubuntu environment. This environment was ideal for the computation of the direction of maximal skew as it provides the efficiency important to scientific computing. With many scripting environments in use, the process scheduling algorithms employed by the Linux kernel allowed the calculations to be performed seamlessly.

Observations and Discussion The computational method used is well demonstrated in Fig. 4, where the direction of maximal skew is observable. Additionally, the skew of the components of each class provide knowledge of the nature of the datasets. Tables 1 and 2, representing the components of the Pima and Appendicitis datasets, respectively, show that the datasets are indeed skewed. Another observation may be attained from these. Within the Pima dataset, the corresponding components of the classes exhibit similar skews, while this is not the case within the Appendicitis dataset. It is possible that the performance of classification is worse when the skews of the classes of a dataset are similar. This theory comes from the Appendicitis dataset being classified with higher accuracy than was the Pima dataset. This may be explored further in future research. We found that for our chosen datasets, using the direction of maximal skew performed comparably to principal component analysis. This, however, may be further improved by taking additional steps of axis rotation when applying the direction of maximal skew.

176

The Manhattan Scientist, Series B, Volume 6 (2019)

Rozycki

The computation of the direction of maximal skew in a dataset benefits from the abundance of computational tools available to perform the calculations. This is contrasted, however, by the limitations imposed by the computation time. The required time to compute the direction of maximal skew was influential on the direction of the research. One must consider the relation among run time, classification performance, and practicality. The time to calculate the directions of maximal skew limits the potential application of maximal skew directions for classification because the time to compute becomes unfeasible with respect to solving the polynomial system of equations. Numerical approximation methods to solve polynomial systems are effective at finding the solutions for up to 16 attributes in a dataset, at which point the complexity becomes impractical for application.

Acknowledgement This research was advised by Dr. Lawrence C. Udeigwe, and funded by the Jasper Scholars Summer Research Program at Manhattan College. Additional acknowledgement goes to Dr. Mark DeBonis for his assistance and previous work which made this research possible.

References [1] National Institute of Diabetes and Digestive and Kidney Diseases, Pima Indians Diabetes Database. https://www.kaggle.com/uciml/pima-indians-diabetes-database. Received May 1990, updated 2017. [2] M. J. DeBonis, Using skew for classification. Int. J. Patt. Recogn. Artif. Intell., 29(3), 1-17 (2015). [3] M. J. DeBonis, The Maximal Direction Of Any Nth Order Moment. International Journal of Mathematics and Statistics, 15(2), 1-9 (2014). [4] C. M. Bishop, Pattern Recognition and Machine Learning. Springer-Verlag Berlin, Heidelberg (2006). [5] D. Dua and C. Graff, UCI Machine Learning Repository [http://archive.ics.uci.edu/ml]. Irvine, CA: University of California, School of Information and Computer Science (2019). [6] W. Decker, G.-M. Greuel, G. Pfister, H. SchoĚ&#x2C6;nemann, S INGULAR 4-1-2 - A computer algebra system for polynomial computations. http://www.singular.uni-kl.de (2019). [7] D. R. Grayson and M. E. Stillman, Macaulay2, a software system for research in algebraic geometry. https://faculty.math.illinois.edu/ (2019). [8] J. Bezanson, A. Edelman, S. Karpinski, and V. B. Shah, Julia: A Fresh Approach to Numerical Computing. SIAM Rev. https://doi.org/10.1137/141000671, 59(1), 65-98 (2017). [9] Homotopy Continuation. https://www.juliahomotopycontinuation.org/ (2019).

Cloud-based automated sound synthesis using machine learning Kyle Watkins∗ Department of Computer Science, Manhattan College Abstract. The ability to generate sounds mimicking different instruments and moods through software, has provided sound designers and producers freedom to quickly and conveniently convert their musical visions to reality. The quality or type of sound generated by these software can be altered by modifying a number of parameters that the software synthesizes. However, this is a confusing and laborious process, especially for users with little to no background in digital signal processing. This summer research project aims to offer a cloud-based solution that automates the generation of different types of sounds as specified by the user. Using machine learning techniques, the implemented prototype has currently been trained from a large database of existing sounds. A user-facing component has been implemented that uses the trained knowledge to enable users to choose a combination of instruments, sound types, and moods and create a sound tailored to those specific requirements. Additionally, for all generated sounds, the tool also offers suggestions for parameters that can be tweaked and still match the desired sound, if desired by the user. The efficacy of the proposed approach in generating suitable sounds is evaluated through surveys and feedback from professional music producers.

Introduction Sound synthesis is the generation of sound through electronic hardware or software. It is usually composed of an oscillator which generates a wave-form such as sine, saw, triangular, etc., which runs through filters, and is ultimately controlled by gain (volume). Changing any of these settings allows sound designers to create virtually infinite types of sounds. To generate any particular type of sound, suitable values have to be given to these parameters. For a particular sound, some of these parameters need to be fixed at an exact value to maintain the quality of that sound. Other parameters may be tweaked within certain range, and still maintain the quality of the sound. A set of parameters and their values is known as a preset, and can be loaded in software to generate the sound. An example of a software which generates sound is an audio plugin. Audio plugins are used as instruments, and loaded into a Digital Audio Workstation (DAW) which can be installed on a computer, laptop, or smartphone to record music. Audio plugins come in four formats: Virtual Studio Technology (VST) which is used by most DAWs, AudioUnits (AU), Real Time Audio Suite (RTAS), or an Avid Audio eXtention (AAX). Sound design can be challenging for users who do not understand the working of synthesizers, and the intricate relations between the various parameter settings. This summer research project aims to close this gap using machine learning and the cloud, making sound design simple and automated for users with minimum background in sound design. The main objective is to create an easy to use VST plugin which can generate sound based on the tags that the user specifies. The user will be able to select tags such as “dark” “piano,” and the software will generate a dark piano preset. After the sound is generated, suggestions on what knobs (hence, the parameters) should ∗

Research mentored by Kashifuddin Qazi, Ph.D.

178

The Manhattan Scientist, Series B, Volume 6 (2019)

Watkins

and shouldnâ&#x20AC;&#x2122;t be manipulated in order to maintain the given tags are shown to the user, in this case the dark piano. Currently, software such as Massive (used for the prototype in this research) do not allow for the examination of all parameters therefore, it is unknown whether the parameters which could not be looked into, will change the results [1].This research project will aid sound designers in creating new unheard-of sounds, and also teach users which parameters they can and cannot manipulate in order to maintain the tagâ&#x20AC;&#x2122;s quality. Additionally, the proposed tool can help experienced sound designers reduce the time spent searching for presets online, which is anecdotally considered one of the most time-consuming aspects. Finally, the tool will provide producers, sound designers, and songwriters freedom from a fixed set of presets that theyâ&#x20AC;&#x2122;ve used repeatedly, and offer an unlimited number of generated sounds at the click of a button. The entire tool can exist and learn in a cloud-based environment, enabling users to utilize it without any local installation.

Methodology and Implementation Fig. 1 shows an overview of the proposed approach. The entire approach can be split into two phases: 1. Training a machine learning algorithm using a large dataset of pre-existing sound presets 2. Implementing a front-end to select various tags, and generate a sound preset matching the tags

Figure 1. Overview of the approach

As shown in the figure, the training phase can be further broken down into: 1. Creating a dataset of pre-existing sound presets 2. Extracting the values of the various parameters for each preset

Watkins

The Manhattan Scientist, Series B, Volume 6 (2019)

179

3. For all presets of the same tag, running each parameterâ&#x20AC;&#x2122;s values through the machine learning algorithm 4. Saving the model generated for each parameter of each tag The implementation of the front-end and sound generation requires the following steps: 1. 2. 3. 4.

Offer a graphical user interface (GUI) to choose the tags Use the trained model to calculate the values of the parameters Identify the fixed parameters, and provide ranges for the non-fixed parameters Generate a sound preset with the calculated parameters

In order to evaluate the proposed technique, and create a prototype, Massive the VST plugin (Fig. 2) was chosen to generate the sounds, and analyze the values of the individual parameters. The tool VST Plugin Analyzer (VPA) (Fig. 3) was chosen to extract the parameter values of a given preset in Massive [2]. While Massive has a few limitations, such as unavailability of the parameters that control the wave form, compared to other VST plugins, it allows access to all other required parameters.

Figure 2. Screenshot of Massive

Figure 3. Screenshot of VPA

To create the dataset for training of the algorithm, fifty-five thousand pre-existing presets with various tags were mined and loaded into Massive. The files that are loaded into Massive are native to Massive (nsvm file types). Due to this, the nsvm type had to go through multiple converting processes in order to be in the proper format. Firstly, the presets had to be renamed by number to keep everything ordered. Puloverâ&#x20AC;&#x2122;s Macro Creator was used to go down the list of presets and rename them by number. After this part was done, it was time to open Massive inside of VPA [3]. Doing this allowed the presets to be opened within Massive then saved as an fxp (default preset type) file through VPA. Puloverâ&#x20AC;&#x2122;s Marco Creator was used to allow automation to constantly convert nsvm types to fxp types. Preset2Excel was then used to convert all the fxp type files into separate excel sheets, so that each parameter in the presets was lined up with its value in an xml file. Again, a script was developed to use Preset2Excel and automatically convert each fxp to an

180

The Manhattan Scientist, Series B, Volume 6 (2019)

Watkins

xml file. Finally, all the Excel sheets belonging to presets with the same tag were merged into one Excel sheet using Puloverâ&#x20AC;&#x2122;s Macro Creator. At the end of the previous steps, the dataset ended up as one xml file per tag. The file included the parameter names and values of all the presets with that tag. Linear Regression was chosen to learn the parameter values for each tag. This was programmed using Python in PyCharm and the TensorFlow package [4]. For each parameter of a particular tag, all available values were given to the linear regression model, which then tried to find the best fit line between all the values (example: Fig. 4). In each case, 98% of the values were chosen as the training set. Once the model trained on the entire training set, it was used to generate predictions for each value in the testing set (remaining 2%) [5]. If the predicted value and the actual value were near equal, the algorithm marked the parameter as a constant value for that particular tag. However, if the predicted value deviated substantially from the actual value, the parameter was marked non-constant for that tag. If the constants are greater than non-constants then presumably, the constant value is important to keep the preset sounding as it should. In this case, the user should keep this parameter at the fixed value. However, if the value of a parameter varies substantially among the testing set, it implies that the parameter could be tweaked with different values, and still leave the preset sounding as it should. In this case, the algorithm generates a range of values that would work for that parameter. The user should stay within this range when tweaking the preset (Fig. 5). After the file is generated it is converted to be loaded into VST Plugin Generator (VPG) then is converted to a default preset type [6].

Figure 5. PyCharm output Figure 4. OSC1-PITCH plotted

For the frontend and GUI, Java was used for its easily accessible vanilla library for GUI programming. The way the software was programmed allowed for new tags to be added if need be. Currently the software has four tags that the user can pick from (bass, lead, synth, and pad). Finally, the Python environment was placed in the cloud using Amazon Web Services (AWS). An AWS ec2 instance with Linux Ubuntu 18.04 was used and setup with the required environment

Watkins

The Manhattan Scientist, Series B, Volume 6 (2019)

181

(python, pip, and EB CLI). The setup allows for Java to call the Python code from the cloud. As seen in Fig. 6 the code is running on GitHub Bash which allows access to the cloud from a local computer. An example run is â&#x20AC;&#x153;python test.py bassâ&#x20AC;?, which generates a bass preset and displays it.

Figure 6. Screenshot of code running on AWS

Experimental Evaluations and Discussions In order to evaluate the proposed approach, an experiment was setup to survey professional sound designers. The survey consisted of ranking three sounds from four tags (Bass, Lead, Synth, and Pad). For each tag, the surveyees were provided one randomly generated sound, one preexisting sound, and one sound generated by the proposed method. Each sound had to be blindly ranked from 1 to 5, 5 being a sound most appropriate to its corresponding tag. The survey was sent to about 30 sound designers, out of which six have responded at the time of preparing this report. Table 1 reports the results of the survey received so far, while Fig. 7 summarizes the results with median scores for each tag. As is intuitive, the presets which created randomly in Fl Studios have the lowest median score for each tag. Interestingly, for the bass, lead, and pad tags, the autogenerated sounds were ranked higher than the actual pre-existing presets. For the synth tag, the auto-generated sound was ranked higher than the random sound but was ranked lower than the pre-existing preset. This observation can be attributed to the fact, that the three highest rated tags had the largest training data sets, with about twenty-six hundred presets. The synth preset on the

182

The Manhattan Scientist, Series B, Volume 6 (2019)

Watkins

Table 1. Individual survey scores for different sounds Tag

Random

Generated

Pre-existing

Bass

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

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

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

Generated

Random

Lead

Synth

Pad

Regular

Bass

Lead

Synth

Pad

Figure 7. Median of preset scores

other hand, had about twelve hundred presets. It is likely that the synth tag should produce something more accurate with a larger dataset. Another observation was with the constant/non-constant parameters generated for each tag. The generated presets from the bass demonstrated more parameters with constant values, as opposed to synth with most parameters non-constant, as seen in Figs.

Watkins

The Manhattan Scientist, Series B, Volume 6 (2019)

183

8 and 9. It can therefore also be deduced, that the synth, pad, and lead tagsâ&#x20AC;&#x2122; wide variety of sounds lead to it having more non-constant parameters.

Figure 8. Screenshot of Bass parameter values

Figure 9. Screenshot of Synth parameter values

184

The Manhattan Scientist, Series B, Volume 6 (2019)

Watkins

Conclusion This project proposed applying machine learning techniques to aid in sound design. A large database of pre-existing presets was used to train a model to learn the various values of a sound’s parameters. A front-end prototype was implemented that allowed users to select various tags for the sound they desired, and autonomously generated a suitable sound matching those tags. Through a survey-based evaluation it was demonstrated that the sounds auto-generated by the prototype are equivalent in quality to manually made sounds. Further work is envisioned to test various other machine learning techniques, as well as using larger datasets to train the model for more tags. Currently, the prototype has been ported to Amazon’s cloud-based service, which enables users to utilize the tool without implementing or installing it locally.

Acknowledgments This work was supported by the School of Science Research Scholars Program.

References [1] Native Instruments, “Massive” [Software]. https://www.nativeinstruments.com/en/products /komplete/synths/massive/?gclid=EAIaIQobChMIgumtrZDY5AIVgYnICh3IsgzAEAQYASA BEgIR3vD BwE [2] Christian Budde, “VST Plugin Analyzer” [Software]. Available from http://www.pcjv.de /applications/measurement-programs/vst-plug-in-analyser-2-0/ [3] Rodolfo U. Batista, “Pulover’s Macro Creator” [Software]. Available from https://www. macrocreator.com/download/ [4] JetBrains, “PyCharm.” Available from https://www.jetbrains.com/pycharm/download/#section =windows [5] Tech With Tim, Python Machine Learning Tutorial 1-7 [Online Video]. Available from https://www.youtube.com/watch?v=ujTCoH21GlA [6] François Mazen, “VST Plugin Generator” [Software]. Available from http://vst-preset- generator.org/download

Air pollutants and childhood asthma in the Bronx Jovan Gonzalez∗ Department of Chemistry and Biochemistry, Manhattan College Abstract. Children in New York City are two times more likely to be hospitalized for asthma then a child in the United States as a whole. The Bronx, in New York City, has the highest rates of childhood asthma than the rest of the city. Most of the cases of childhood asthma come from the South Bronx, which is also one of the lowest income areas of the city. Mott Haven and Hunts Point are two neighborhoods in the South Bronx that have the highest rates of childhood asthma in New York City. This research seeks to examine the air quality of The Bronx to see if the air near schools and playgrounds are affected by heavy traffic and to compare the air quality of lower income neighborhoods in the South Bronx with higher income neighborhoods, such as, Fieldston and Pelham Bay. To gather data of the air quality in the Bronx, four sensors were used to provide data on their respective air pollutants. Volatile organic compounds, carbon dioxide, particulate matter, and ultra violet light measurements were taken at eight locations in The Bronx. After all the data was collected, an average for was taken for each pollution for the respective location. It was shown that all the South Bronx locations have higher levels of particle matter measurements than the more affluent neighborhoods: Fieldston and Pelham Bay. Mott Haven, neighborhood with high rates of asthma, has the highest amount of VOCs measured. The research shows that heavy traffic does have an impact on the air quality of the playgrounds in which children in the South Bronx play. If you use the rates of asthma in those same neighborhoods, it can be seen that it affects the health of the children as well.

Introduction Asthma is a respiratory condition that affects more than 25 million Americans nationwide [1]. In an asthma attack, one’s airways begin to narrow and contract and become filled with mucus. This makes it difficult for the individual to breath and can be fatal if not treated quickly. Asthma attacks often are triggered by different environmental factors in the air that get into one’s lungs. Asthma is the leading chronic disease in children [2]. In 2000, children in New York City were almost twice as likely to be hospitalized for asthma as children in the United States as a whole [3]. Asthma is widespread and a major issue in New York City, and has gained the attention of NYC Department of Health. There must be environmental factors in NYC that is causing such a large asthma prevalence. This not just a major problem in New York City. Many cities across the United States have higher rates of asthma compared to more rural or suburban areas of the country. Urban areas are hot spots for asthma and other respiratory conditions. The Bronx is one of the five boroughs that make up New York City. It currently has the highest rates of childhood asthma in all of New York City. The highest rates of childhood asthma are in the South Bronx, which are also some of the lowest income areas of the borough. The areas with the lowest income, also have the most asthma diagnoses and asthma hospitalization rates. The two neighborhoods which suffer the most rates of childhood asthma are Mott Haven and Hunts Point. As seen in Fig. 1, the neighborhoods of Mott Haven and Hunts Point ∗ Research mentored by Yelda Hangun-Balkir, Ph.D., and Veronique Lankar, Ph.D.

186

Manhattan Scientist, Series B, Volume 6 (2019)

Gonzalez

have the darkest color. Children from ages 0-12 in those neighborhoods have been hospitalized from 73-119 children per 100,000 in 2014. In the Bronx, 17.4% of children up to 12 years old have been diagnosed with asthma. Asthma related ER visits for groups aged from 5-14 in 2014 has increased by 39% at a rate of 483 per 10,000. Bronx asthma rates have stayed almost stable, while every other borough has seen a decrease in asthma hospitalizations as seen in Fig. 2. Minorities and people of lower income/below the poverty line seem to be the most susceptible to being hospitalized due to asthma compared to those of higher income. The Bronx has the lowest median household income according to the U.S. Figure 1. Map of childhood asthma rates in NYC census bureau. According to data from NYU Furman Center, the medium household income of Mott Haven in 2017 was $21,366, that is 66% less than the median income of New York City, which is $62,040. 44.2% of people living in Mott Haven live below the poverty line. Mott Haven has similar demographics as Mott Haven. More affluent neighborhoods, such as Fieldston, Riverdale, have less people in poverty and higher median incomes. The average medium income of Fieldston, based on the 2017 American Community Survey, is $80,932, and has significantly lower rates of asthma hospitalizations than the South Bronx.

Figure 2. ED visits of children up to age 4. (Bronx Community Health Dashboard - Montefiore Hospital)

Gonzalez

Manhattan Scientist, Series B, Volume 6 (2019)

187

The South Bronx is often nicknamed â&#x20AC;&#x153;island of pollutionâ&#x20AC;? by residents who live there [4]. The South Bronx is surrounded by three major highways: Major Deegan, Bruckner, and Cross Bronx expressways. Many public schools and playgrounds in the South Bronx are in very close proximity or across the street from these major highways. The times student leave school or go to the playgrounds after school, is the time traffic started to become the heaviest due to rush hour approaching. Children in the South Bronx after school are exposed to emissions from heavy traffic. Also, the two neighborhoods mentioned previously with the highest rates of asthma, Hunts Point and Mott Haven, are the more industrialized neighborhoods in the borough. Hunts Point has the Hunts Point Produce Market which bring 15,000 diesel semitrucks driving through the neighborhood to deliver food and contributes greatly to traffic congestion [5]. Mott Haven has the Port Morris and many other sewage plants for the department of sanitation. The new fresh direct building in Mott Haven also brings in about 1000 fresh direct diesel semi-trucks through Mott Haven. Many air pollutants are emitted from the exhaust of gas fueled vehicles, especially diesel trucks. Fossil fuel powered vehicles release CO2 , volatile organic compounds (VOCs), and nitrous oxides (NOx ) into the air. During summer when the most UV radiation is received; the VOCs and NOx start to react together under intense heat and sunlight. The product of that reaction is ground level ozone. Ground level ozone is toxic, and colorless gas that is harmful to humans in high concentrations. According to the EPA, breathing ozone can trigger a variety of health problems including chest pain, coughing, throat irritation, and airway inflammation. It also can reduce lung function and harm lung tissue. Sensitive groups such as, people with asthma, children, and the elderly are the most at risk from breathing air with ozone. Particulate matter (PM) is another pollutant that is emitted from vehicles. Particulate matter, according to the EPA, are microscopic solids/dust particles or liquid droplets that are so small they can be inhaled and cause serious health problems. PM smaller than 2.5 micrometers in diameter is the most dangerous. PM of that size can go the deepest into your lungs and even your bloodstream. It can cause many respiratory issue and even cancer with long term exposure. The South Bronx provides an ideal context in which to investigate the effects of traffic-related air pollution on children with asthma, because large numbers of children with asthma live and attend schools or play in playgrounds near highways that experience heavy traffic with trucks and cars. This research is intended to survey the air quality of different Bronx neighborhoods, and to compare the air quality of lower income neighborhoods and higher income neighborhoods. Lastly, to see if areas with poor air quality high rates of childhood asthma have, using data from New York City DOHMH Environment & Health Data Portal and if heavy traffic does have an impact on air quality.

Methods and Materials Eight sites, in different neighborhoods in the Bronx were chosen for air quality measurements. The lower income, and South Bronx locations chosen were, Millbrook Playground in Mott Haven,

188

Manhattan Scientist, Series B, Volume 6 (2019)

Gonzalez

134 Playground in Mott Haven, Lyons Square Playground in hunts Point, Prospect Playground and Tremont Park in Tremont, and Noble Playground in West Farms. The uptown Bronx locations, or more affluent areas chosen were Horace Mann School in Fieldston, Pelham Bay Park, and Van Cortlandt Park. All South Bronx locations were strategically chosen to be a playground or school in very close proximity to a highway, in order to assess the effects of heavy traffic on the air quality in areas that children and see a link with childhood asthma (Fig. 3). Air measurements were obtained using four different air pollutant sensors manufactured by the company Adafruit. The open source programming software, Arduino, was used to help program and calibrate the sensors acquired for measurements. The sensors were calibrated and assembled with help from the physics department. The sensors collected the measurements and the data is processed through the Arduino board and stored on a data logger with an SD card. A sensor to

Figure 3. Map of site locations in the Bronx

Figure 4. Particulate matter sensor

detect levels of ozone in ppb and ppm was used. measured volatile organic compounds (VOCs) in ppb and CO2 in ppm. A sensor that measured particulate matter (PM) in several different sizes: 0.3 µm, 0.5 µm, 1 µm, 2.5 µm, 5 µm, 50 µm. The particulate matter sensor works by counting the amount of particulate in each size respectively in 0.1 L of air. Lastly, a sensor for UV was used, which provided the amount of UVA, UVB, and UV index by measuring the µW/cm2 . Each location was visited twice a week from June 12 to August 12. The locations were visit at similar times from 3:00pm-4:00pm to measure the air when rush hour is about the begin and traffic starts becoming heavy. Fieldston, Van Cortlandt Park, Hunts Point, and Pelham Bay Park

Gonzalez

Manhattan Scientist, Series B, Volume 6 (2019)

189

were taken every Monday and Wednesday. Mott Haven Locations, Prospect Playground. Noble Playground, and Tremont park were taken Tuesday and Thursdays. Fridays the location sets would be alternated every week. Air measurements were taken by placing the sensors down in a secure place at the respected location, then switch them on. After the measurements were collected and data stored on the SD card, and all the data points were averaged on excel as shown in Fig. 2 for that respective day. After the data collection of the whole summer was complete, an average was taken of all the daily averages to provide a summer wide average from June-August (Table 1). Table 1. Measured amounts, per 0.1 L of air, of particulate matter (PM) of several different sizes; at Hunts Point on 07/17/2019. The table shows an example of the data collected for a single particulate matter measurement. The sensors provide the number of particles in 0.1 L in air. The instrument takes ten data points for the particulate matter of each respective size, as listed. Particulate matter size data point # 1 2 3 4 5 6 7 8 9 10 Average

0.3µm

0.5µm

1µm

2.5µm

5µm

50µm

4644 4545 3327 3408 3180 2901 3174 3174 3660 3660

1465 1445 1041 1063 981 895 1003 1003 1154 1154

304 304 245 243 187 182 199 199 221 221

14 16 14 14 16 16 6 6 4 4

2 2 4 4 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0 0

3567.3

1120.4

230.5

1.2

Results After the end of the research, the VOC and PM measurements provided data that was close to previous assumptions. South Bronx locations had high levels of PM and Mott Haven had the highest amount of VOCs out of all the neighborhoods. The results are bar charts based on the final average of all the data at the end of the research. Fig. 5 shows the measurements of VOCs in parts per billion (ppb) and CO2 in parts per million (ppm). The highest VOC measured was in the Mott Haven neighborhood with an average of 1540.745 ppb from June 12 till August 12, 2019. Mott Haven was one of the neighborhoods with the highest rates of childhood asthma hospitalizations. Fieldston, which is one of the affluent neighborhoods, has the second highest VOC measurement with an average of 921.529 ppb. The CO2 measured is approximately 400 ppm in all neighborhoods. Fig. 6 shows the average amount of ozone measured during the summer in ppb. The ozone measurements were generally low across all the neighborhoods. EPA standards for exposure to ozone for more than 8 hours is less than 70 ppb. Every neighborhood measured has ozone levels

190

Manhattan Scientist, Series B, Volume 6 (2019)

Gonzalez PPB

VOC

µg/m^3

CO2 50

1000

Noble Playground

Prospect Playground

Tremont Park

Hunts Point

Mott Haven

Figure 5. Measurements of VOC (in ppb) and CO2 (in ppm)

Pelham Bay Park

Van Cortlandt

Noble Playground

Prospect Playground

Tremont Park

Hunts Point

Mott Haven

Pelham Bay Park

Van Cortlandt

Fieldston

0 Fieldston

Figure 6. Ozone measurements

well below 70 ppb. Though Prospect Playground and Nobile Playground have the highest average ozone measured, they do not exceed the EPA limit. Fig. 7 shows the number of particles measured in three different sizes/diameter. The sensor measured PM in 0.3, 0.5, and 1 micrometer. The sensor counts the number of particles for the respective size in 0.1 liters of air. The data shows that all the South Bronx (lower income areas) have high amounts of PM in the air than Fieldston and Pelham Bay park (Wealthier areas). Van Cortlandt Park also has high levels of PM in the air as well, which was unexpected. 0.3µm

0.5µm

1µm

UVA

2000

UVB

5000 1000

Noble Playground

Prospect Playground

Tremont Park

Hunts Point

Mott Haven

Pelham Bay Park

Van Cortlandt

Noble Playground

Prospect Playground

Tremont Park

Hunts Point

Mott Haven

Pelham Bay Park

Van Cortlandt

Figure 7. Particulate matter measured per 0.1 L of air

Fieldston

0 Fieldston

Figure 8. UV measurements (in µW/cm2 )

Fig. 8 shows the amount UVA and UVB rays averaged over the summer. UV rays are expected to be high since it is summer. Neighborhoods with lower UV levels are due to the measurement location being in shade due to tree coverage. High UV radiation in heavy traffic conditions can lead to the formation of ground level ozone.

Conclusion The South Bronx in New York City has neighborhoods with many people below the poverty line and on public assistance. These same neighborhoods also have some of the highest rates of

Gonzalez

Manhattan Scientist, Series B, Volume 6 (2019)

191

asthma in the city. The New York City health department considers indoor air quality as one of the main reasons of asthma in children. Though indoor air quality does have an effect on respiratory health, outdoor air quality should be a factor that gets more attention and research as well. Since this epidemic is affecting mostly individuals of lower income, there environmental injustice evident in New York City. The data in the results section does support the initial assumption that the South Bronx has lower air quality than their wealthier counterparts. This is evident by the PM measurements in figure 8. PM, as previously mentioned, can be very dangerous if inhaled for long periods of time. All the South Bronx locations had high PM levels close to 2000 particles in 0.1 L of air. During the research, the PM ratings would be consistently high, unlike Fieldston or Pelham Bay whose measurements were never constantly high. In the future, a different type of statistical analysis of the data will be used to not have one day of random spiked data skew the averaged too heavily. Van Cortlandt Park measurements were surprising, because the PM levels were not expected to be so high. Initial possible reason for the high PM levels was that it is a park, so particles from the soil or vegetation may be the cause. Pelham Bay Park, is a large park system as well like Van Cortlandt, but the PM measurements are much lower for reasons that are unclear and more research is required. This research is being continued in Van Cortlandt Park during theFall 2019 semester to have a better sense and understanding of the air quality in the park system. The data in the research shows that heavy traffic is affecting the air quality in many neighborhoods. This can be seen in the high levels of PM over the Bronx and the high VOC levels in Mott Haven. It is evident that there is an issue with these pollutants, and it is affecting the health of children and the people living in these neighborhoods. The asthma data shows that there the health of the individuals living there are affected. Possible solutions to this problem can be simply, planting more trees to provide vegetative Figure 9. Heavy traffic at Prospect Playground barriers between children in the parks. Comparing Hunts Point and Mott Haven, neighborhoods with highest asthma rates, VOC levels provide an example of this. The playground at Hunts Point, is surrounded by a barrier of bushes and trees. Hunts Point had low VOC levels, despite heavy truck traffic driving on the Bruckner daily. The playground at Mott Haven, had little to know vegetative barrier between the highway and the park, and the VOC levels were the highest of all the neighborhoods. More research will be required test of this assumption is true, but the planting of trees is an initiative being developed by the city to improve air quality. Another possible aid to combat air pollution is the adafruit sensors and Arduino systems. These systems are cheap, affordable and easy to use. This can provide residents

192

Manhattan Scientist, Series B, Volume 6 (2019)

Gonzalez

with sensors that give them information of the air quality at their exactly location. This type of information can help provide people in the neighborhoods enough information to take required actions for their health, and provide people with an awareness of the air they are breathing.

Acknowledgment This work was supported by the School of Science Summer Scholars Program

References [1] CDC.gov. (2018). CDC - Asthma - Data and Surveillance - Asthma Surveillance Data. http://www.cdc.gov/asthma/asthmadata.htm [2] CDC.gov. (2018). Asthma | Healthy Schools | CDC. https://www.cdc.gov/healthyschools /asthma [3] Garg R, Karpati A, Leighton J, Perrin M, Shah M. Asthma Facts, Second Edition. New York City Department of Health and Mental Hygiene, May 2003. [4] Butini, C. (2018, January 30). Asthma By The Numbers. Retrieved October 4, 2019, from https://medium.com/asthma-in-the-south-bronx/asthma-by-the-numbers-73553b2c9621 [5] Hunt Point Truck Study, URS/Goodkind & Oâ&#x20AC;&#x2122;Dea, Inc., 2004.

Determination of heavy metals in Tibbetts Brook Tatianna Peraltaâ&#x2C6;&#x2014; Department of Chemistry and Biochemistry, Manhattan College Abstract. Tibbetts Brook is a stream flowing north to south from the city of Yonkers into the Bronx, where it concludes in Van Cortlandt Park. Heavy metal contamination and accumulation is a serious problem found in bodies of water, especially those in urban areas, due to abundant sources of toxins, non-biodegradable properties, and accumulative behavior of heavy metals. This study determines the contamination levels of heavy metal concentrations in the Tibbetts Brook by using an Atomic Absorbance Spectrometry machine. In partnering with The Friends of Van Cortlandt Park Organization, nine different sites were chosen for this study. The sites were selected to include pipe outfalls entering the lake, the weir where the lakeâ&#x20AC;&#x2122;s outflow enters the sewer system, and a pipe that drains a salt pile north in the park. The research project highlights the importance of river and stream pollution in urban setting, in the context of determining the various ways that the human population affects our environment. This project also displays the importance of environmental science, by showcasing how humans affect our environment.

Introduction Heavy metals are a major concern in aquatic environments due to their toxicity, persistency and tendency to accumulate in organisms. The most abundant source for these metals is industrial and domestic waste as well as agricultural runoff. The fate of heavy metals in aquatic environments is affected by processes such as precipitation and dissolution. These processes are also affected by factors such as pH, temperature and dissolved oxygen. Tibbetts Brook runs from Westchester county to the Bronx, where it serves as an important watershed. Originally Tibbetts Brook naturally flowed into the Harlem River [1] In the early 1900s, it was decided that the land was too unstable for development, due to its high moisture retention, so the brook was rerouted into the sewer system [2]. Inadvertently, as a result of this rerouting the combined sewer and storm water now flow into the Harlem River. This is because the water treatment plant pipes are unable to withstand high quantities of water. Daylighting is a technique used in urban areas to help reduce polluted runoff, lower flash flood possibilities as well as improve the livability of overall environments. [3]. In this case daylighting will help reroute Tibbetts Brook back into the Harlem River. The nine sites were purposely selected to include pipe outflows, drainage areas, sewer system entry and exit points as well as areas popular for human interaction. Mining, working with metals and steel production, as well as burning coal and certain wastes can release zinc into the environment [4]. A common use for zinc is to coat steel and iron to prevent rust and corrosion; this process is called galvanization. If high levels of zinc are present in soils, the metal can seep into groundwater. Industries can release dust containing higher levels of zinc into the air we breathe. Eventually, the zinc dust will settle out onto the soil and surface waters. â&#x2C6;&#x2014;

Research mentored by Yelda Hangun-Balkir, Ph.D.

194

The Manhattan Scientist, Series B, Volume 6 (2019)

Peralta

Rain and snow also can remove zinc dust from the air. Most of the zinc in lakes, rivers and streams does not dissolve, but settles to the bottom. Some fish in these waters may contain high levels of zinc. Manganese exists naturally in rivers, lakes, and underground water. It may also be released to water from natural deposits, industrial wastes (iron, steel, acid mine drainage), and the use of pesticides. Iron and steel plants also release manganese into the atmosphere. Small amounts of manganese are needed for growth and good health in children, but excess manganese can result in nervous system problems. Too much manganese, however, can injure the cerebrum which controls body movements, an extreme exposure can also cause respiratory problems and sexual dysfunction. Iron exists naturally in rivers, lakes, and underground water. It may also be released to water from industrial wastes, refining of iron ores, and corrosion of iron containing metals. The combination of naturally occurring organic material and iron can be found in shallow wells and surface water. Electroplating and textile industries release relatively large amounts of chromium in surface waters. Leaching from topsoil and rocks is the most important natural source of chromium entry into bodies of water [5]. A common health issue caused by over exposure to chromium is irritation of the lining of the nose, runny nose, and breathing problems (asthma, cough, shortness of breath, wheezing)

Environmental Science Environmental science could be defined as the study of the effects of natural and unnatural processes, and of interactions of the physical components of the planet on the environment [6]. So essentially how humans affect the environment and how the environment affects humans both positively and negatively. Environmental science combines multiple diciplines into its field of study. It is not only backed by biology and chemistry but also geography, humanities, physiology and history. Environmental scientists research the ways humans have affected natural habitats, in order to do so we must lookback into the history of the geography of areas that we deem affected by human development. In this case we had to research the area surrounding Tibbetts Brook (Fig. 1) and how the physical area has changed over time to fit the needs of humans and our population growth. Tibbetts Brook flowed naturally into the Harlem river and was redirected into the sewer system so that buildings used for commercial and residencies could be built. This negatively affected the quality of both water sources and has caused issues regarding water leakage as well as the seepage of hazardous waste into areas exposed to humans. The nine sites were purposely selected to include pipe outflows, drainage areas, sewer system entry and exit points as well as areas popular for human interaction. They were also spread out so that we could collect in both Bronx county as well as Westchester county. This ensured that we were testing the brook in areas that are highly affected by human pollutants. Although we

Peralta

The Manhattan Scientist, Series B, Volume 6 (2019)

195

Figure 1. Map showcasing the nine sites where the water samples were collected

wanted to prove the fact that the brook does not contain high amounts of metals so that it could be redirected into the Harlem river, we wanted to do so fairly. Thus, we tested sites that were predicted to hold higher levels of metals due to human activities. Along the brook are different industrial sites that deliver pollutants into the brook. The Major Deegan Expressway runs along the southern end of the brook, the Van Cortlandt golf course and other recreation areas are major contributors to pollution seepage into the brook.

Goals and Hypothesis The goal of this study was to determine the heavy metal pollution levels in Tibbetts Brook, the study examined the concentrations of four heavy metals: iron (Fe), manganese (Mn), chromium (Cr) and zinc (Zn) in various locations throughout the brook. One of the main goals of this project

196

The Manhattan Scientist, Series B, Volume 6 (2019)

Peralta

was to measure and ensure that the brooksâ&#x20AC;&#x2122; heavy metal levels were in range so that the water could be allowed to flow back into the Harlem River, without altering the water quality. As a collective we also wanted to ascertain whether or not the different water collecting sites displayed significant differences in heavy metal concentrations. Although we are testing the water quality, it is impossible to restore the historic path of Tibbetts Brook, the goal is to help alleviate some of the flooding in the area. A well-designed water management system here could help reroute two billion gallons of freshwater annually from the cityâ&#x20AC;&#x2122;s overtaxed sewer system, while also channeling away hundreds of millions of gallons of rainwater. The NYC Parks department along with the Friends of Van Cortlandt Park Alliance have a plan to redirect the water from Tibbetts brook, starting in Kingsbridge, underground so that it is rerouted directly into the Harlem River rather than the sewage system [2]. It was expected that each collection site would yield different heavy metal concentrations, based on the different amounts of exposure to human pollutants. It was also expected to determine that the contaminant levels of Tibbetts brook are low enough to be directed back into the Harlem river without changing the water quality of either water source.

Methods and Materials Water samples were collected every two weeks from nine different locations from January to August. Atomic absorption spectroscopy was used to determine the concentrations of the heavy metals in the samples. Atomic absorption is an analytical method used for elemental analysis of a compound or a mixture, this is the best method for determining several elements in the same sample [7]. Even when trace amounts are present, without chemical separation of the elements. The collected water samples were digested with nitric acid. Stock solutions of each element were used as standards for each element found in the collected water samples.

Results Using the Atomic Absorption spectroscopy machine, we were able to measure trace amounts of metal levels without degrading the samples. The results that we gathered were in range with what was expected and was in line with what the Environmental Protection Agency (EPA) [8] says is safe for urban water sources. The results also supported the hypothesis that we came up with, it was expected that the metal levels would be in loin with what was set forth by the EPA. This makes it easier to sell the case that Tibbetts Brook could be redirected back into the Harlem River without altering water levels. By proving that the metal levels from the brook are not hazardous, we are supporting efforts and providing actual data to support the idea of redirecting the brook to its natural path. ater quality of either water source.

Peralta

The Manhattan Scientist, Series B, Volume 6 (2019)

Figure 2. Atomic absorption results for Iron, Manganese, Zinc, and Chromium

197

198

The Manhattan Scientist, Series B, Volume 6 (2019)

Peralta

Conclusion The essence of this research project was to highlight the importance of environmental science as a field. Environmental science is defined by many as an interdisciplinary field that studies how humans affect the environment and how the environment affects humans. This field studies the natural and unnatural processes of the Earth, as well as interactions between physical components of the planet and the environment in relation to humans and human activity. What is often found is the fact that human interactions with the environment are majority negative. For example, the main goal of this research project was to determine whether or not Tibbetts Brooks’ water was in range, based on what the EPA has set forth as acceptable regarding metal levels inside of natural water sources. By determining the ranges of metal levels in the brook we could relay this information to government officials so that the brook could be rerouted back to the Harlem river where it naturally flowed before human interaction.

Acknowledgments This work was supported by the Manhattan College Jasper Summer Scholars Program. This project would not have been possible without the efforts and knowledge of Dr. Yelda Balkir from Manhattan College and John Butler and Alex Bryne of the Friends of Van Cortlandt Park association.

References [1] “Van Cortlandt Park.” Van Cortlandt Park Highlights - Tibbetts Brook : NYC Parks. [2] Kensinger, Nathan. “One of NYC’s Underground Rivers May Soon Be Brought Back to Life.” Curbed NY. Curbed NY, December 20, 2018. [3] “Daylighting Streams: Breathing Life Into Urban Streams And Communities.” American Rivers. [4] Tchounwou, Paul B, Clement G Yedjou, Anita K Patlolla, and Dwayne J Sutton. “Heavy Metal Toxicity and the Environment.” Experientia supplementum (2012). U.S. National Library of Medicine, 2012. [5] “Toxic Substances Portal - Chromium.” Centers for Disease Control and Prevention. Centers for Disease Control and Prevention. [6] “What Is Environmental Science?” EnvironmentalScience.org. [7] “Table 1 EPA Sample Processing Method for Metallic Element Analysis.” Atomic absorption Spectroscopy. [8] “Metals.” EPA. Environmental Protection Agency, June 6, 2019.

Cluster detection in a dataset using the BCM and Oja synaptic learning rules Michael Campiglia∗ Department of Computer Engineering, Manhattan College Abstract. Understanding the process in which neurons learn has been a continuously growing area of research throughout recent years. This study seeks to explore two learning models for sensory neurons and apply them to the machine learning task of clustering. More specifically, an algorithm was created to detect clusters in datasets using model BCM and Oja neurons. This algorithm was designed to work well with linearly seperable datasets.

Introduction One key topic in the study of machine learning and artificial neural networks is determining how neurons learn [1]. If you take a look at the anatomy of the neuron, as shown in Fig. 1, you will find a few key components that make communication and learning possible. The first major component is the axons, which allow the neurons to send outgoing signals. Then there are the dendrites and soma, which receive and store incoming signals, respectively. Finally, there is the synapse, where the axon from one neuron meets the dendrites of another, and communication takes place. As this communication occurs, presynaptic signals sent from the axons impact the postsynaptic action potential in the soma, and if the potential rises high enough, the target soma generates its own action potential and the neuron is said to ‘fire.’ In the past few decades, extensive research has been conducted on this field of interest, and specific learning rules have been formulated to model the changing connections, or synaptic weights, between neurons.

Figure 1. Anatomy of a neuron. (Credit: Bear et al. [2])

Knowledge on how neurons learn in biology plays an essential role in machine learning and artificial intelligence, since models developed in neuroscience have fueled these fields. In fact, ∗

Research mentored by Lawrence Udeigwe, Ph.D.

200

The Manhattan Scientist, Series B, Volume 6 (2019)

Campiglia

the basis of the growing field of unsupervised learning is the Hebbian learning theory, which proposes that the strength of connection between neurons is proportional to the degree of correlation between pre and post synaptic activity. In other words, if neuron A continuously takes a part in firing neuron B, a change will occur in the cells such that efficiency of A in firing B is increased [3]. Hebbian theory provides the foundational learning algorithm for neural and artificial network models, providing a method to determine how to alter the snyaptic weights between model neurons. Hebbâ&#x20AC;&#x2122;s rule has also resulted in the development of new sensory learning models, such as the two explored in this study, the BCM and Oja learning rules. These mathematical efforts and learning rules have played a huge part in allowing researchers to continue to model and study both artificial and biological neural networks. One prime example is exhibited through the use of BCM theory in an experiment to model monocular deprivation and the concept of ocular dominance plasticity, the idea that brief deprivation of vision in one eye directly reduces neuronal responses in the closed eye and increases responses in the open one, in mice [4]. In this particular experiment, the BCM learning rule modeled the ocular dominance shift during reverse suture, which is the process of allowing vision in only one eye for a period of time and then reversing the orientation. This process resulted in the portrayal of two key effects, including noise replacing the original retinal activity in the newly closed eye, and a sharp reduction of cortical activity in the neuron, demonstrating the concept of synaptic plasticity. Another example is how the Oja ruleâ&#x20AC;&#x2122;s property of being a principal component analyzer (PCA) has led to defining the Oja rule for a layer of parallel neurons and modifiying the Oja rule to perform independent component analysis (ICA), which can find statistically independent components in a dataset [5, 6].

Figure 2. Example of data clustering. [7]

In the current study, we work on adapting these models to detect clusters in a dataset. As defined in [8], clusters consist of objects or data points that are in some way similar amongst themselves, and clustering is the process of dividing data into these specific groups of similar objects.

Campiglia

The Manhattan Scientist, Series B, Volume 6 (2019)

201

Clustering is a very active research field involved in statistics, data mining, and machine learning, and implementing efficient clustering algorithms is essential in handling the huge amounts of data that today’s technology has allowed us to explore. An example of data point clustering is shown in Fig. 2 . This paper will first explore how to model a neuron that can ‘learn’, and then focus on using this artificial neuron to detect clusters in datasets.

Methods How to model a neuron One of the simplest ways to model a neuron comes from a generalization of Hebb’s Rule and results in a linear neuron, y = w · x, where     w1 x1  ..   ..  w =  .  and x =  .  . (1) wm xm

In this equation, x is a vector set of input stimuli, w is a vector of all the synaptic weights associated to each input, and y is the sum of all the synaptic responses. This version, and actually all neuron models using Hebb’s rule are unstable, which is why neural networks usually rely on other stabilized learning models, such as BCM theory and Oja’s rule.

Figure 3. Artificial lnear model neuron [9].

BCM theory The BCM theory originates from the 1982 proposal on the theory of neuron selectivity by Bienenstock, Cooper, and Munro [10]. In this formulation, Hebb’s rule is stabilized by incorporating a nonlinear function of the postsynaptic activity φ(y), that changes its sign at a specific value θ, called the modification threshold. If the response y is above θ, the weight vector w is driven in the direction of the corresponding input x, and driven opposite the direction of the input if the output

202

The Manhattan Scientist, Series B, Volume 6 (2019)

Campiglia

is below θ. This behavior results in a temporal competition between incoming input patterns [11]. However, the key to BCM theory that solves the instability problems of Hebb’s rule is making the modification threshhold θ a super-linear function of the neuron’s output, meaning that it ‘slides’ with the activity of the neuron. With this addition, the threshhold is able to act as a negative feedback system, allowing θ to stay ahead of the growth caused by synaptic activity. An example of the super-linear θ is shown in Fig. 4.

Figure 4. Super-linear function θ. [8]

While there are many forms of the φ function, the one used in this study follows the ordinary differential equation form of the BCM modification equation, resulting in: y =w · x ẇ =ηy(y − θ)x

θ̇ =1/τ (y 2 − θ).

(2) (3) (4)

When looking at the one dimensional model, we get a stable fixed point of y = θ = 1 Evaluating this, y = 1 = wx, w = 1/x, meaning that the weight compensates depending on the size of the input. In order to simulate the BCM learning rule, the equations can be discretized to y (n) =w(n) · x(n)

θ(n+1) =θ(n) + 1/τ (y 2 − θ(n) )

w(n+1) =w(n) + ηy(y − θ(n+1) )x(n) ,

(5) (6) (7)

where n represents the current iteration, η is the learning rate and τ is the threshold averaging constant (sometimes referred to as the memory constant). Oja’s Rule Oja’s rule was also developed as a variant of the Hebb’s rule to avoid the problems that may arise from the instability of Hebb’s equation. To fix the instability problem, a decay term is included on the weights, resulting in the form ẇ = yx − y 2 w. With Oja’s Rule, increases in

Campiglia

The Manhattan Scientist, Series B, Volume 6 (2019)

203

connection strength for certain synapses are accompanied by decreases in connection strength for other synapses of the same neuron, which results in spatial competition between inputs [9]. The equations for Oja’s rule are described below: y =w · x

(8) (9)

ẇ =yx − y w. In order to perform simulations with Oja’s rule, the equations are also discretized to

y (n) =w(n) · x(n)

(n+1)

(n)

+ η(yx

(10) (n)

−y w

(n)

(11)

where η is the learning rate. Detecting clusters using sensory neurons To have a better understanding for the idea of detecting clusters using an artificial neuron, it is important to take a look at two related concepts for neurons. One is synaptic plasticity, which states that the strength of connection between two neurons can change over time [13]. The other is neuronal selectivity, which states that over a period of training, a neuron learns to react and fire strongly to certain stimuli while reacting weakly to others [14]. The idea of detecting clusters using the BCM and Oja learning rules on an artificial neuron is based on the premise that if a neuron becomes selective to a stimulus input, then that neuron should also be selective to every other stimulus near it. If we consider a data point to be treated as a stimulus input, we can then translate this concept to a dataset and group nearby data points into clusters. With this idea in mind, we can form the basis of the algorithm by focusing on grouping the responses of the artificial neuron. For both BCM and Oja’s rule, the neuron is trained by running many iterations of the discretized learning rule equations on a dataset of inputs, X. The dataset consists of a m x n matrix, where m represents the number of rows of inputs and n represents the dimensionality of the input. Thus, each row of X represents a data point x. For each iteration, a row of X is randomly chosen from the full set of inputs, and the response becomes y = w · x. Over time, the synaptic weight vector w will adjust itself from the learning rule and will ultimately become selective to an input, or in this case, a cluster of inputs.

Algorithm The cluster detection algorithm was written and tested in MATLAB. The function takes in two parameters, the matrix of inputs X, described earlier, and the number of clusters expected to be in the data set, denoted Clusters . The algorithm then begins a loop, running for a set number of iterations, numItr. Each iteration, a row of X, which represents a stimulus input, is chosen at

204

The Manhattan Scientist, Series B, Volume 6 (2019)

Campiglia

random to train the neuron using either the BCM or OJA learning rule, depending on the user’s choice. As each iteration takes place, the resulting synaptic response and weight vectors are stored into the matrices Y and W respectively, where Y has the same number of rows as inputs, W has m rows and numItr columns, and W has n rows and numItr columns, referring back to the parameters mentioned earlier while describing the m × n matrix.

Next, the algorithm takes the last column of Y, which represents the last iteration of the synaptic responses for all of the inputs, and performs a sort on this data, storing it in a new array, YSort. After the responses are sorted, a for loop iterates through each index of the array and the difference between each value and its predecessor is recorded and stored in the array YDiff . The maximum value in YDiff is recorded as maxDiff , which represents the largest jump between two synaptic responses and is thus a demarcation between clusters. After this, the index of maxDiff within the YDiff array is found. The value of YSort at this index is stored in the variable BreakPoint, as this represents the value of the last synaptic response before a cluster changes. After this is complete, a new for loop iterates throughout the last column of Y, this time splitting up the data depending on whether the index is greater or less than the BreakPoint. The smaller group is considered to be the first cluster of the data, and the inputs in the cluster are removed from X. After this, the remaining inputs are kept within the data set, and the algorithm is repeated until it reaches the predefined number of clusters. Algorithm 1 Train BCM Neuron with dataset X Initialize θ = 0, w = a non-zero vector for k = 1 to numItr do x ← random row of X y ←x·w w ← w + ηy(y − θ)xT θ ← θ + τ1 (y 2 − θ) end for Y ←X·w Algorithm 2 Train Oja Neuron with dataset X Initialize w = a non-zero vector for k = 1 to numItr do x ← random row of X y ←x·w w ← w + η(yxT − y 2 w) end for Y ←X·w

Campiglia

The Manhattan Scientist, Series B, Volume 6 (2019)

205

Algorithm 3 Cluster Detection on data set X for k = 1 to Clusters -1 do Train neuron using algorithm 1 or 2 to return Y YSort ← sort(Y) for i = 1 to Size(Y) -1 do YDiff ← Take difference of consecutive Y indeces end for maxDiff ← max(YDiff) maxIndex ← Find index where YDiff = maxDiff BreakPoint ← value of YSort at maxIndex for i = 1 to Size(Y) do if Y(i) ≤ BreakPoint then Store into LowerGroup else Store into UpperGroup end if if LowerGroup > UpperGroup then Remove UpperGroup inputs from X and store into new Cluster else Remove LowerGroup inputs from X and store into new Cluster end if end for end for

Results

For all of my results, I ran the datasets through 500,000 iterations, with parameters η = 0.01 and τ = 10. I began testing the cluster detection algorithm by using an artificially created 2D data set as my set of inputs. I created three clusters, each consisting of 40-70 data points, distinctly seperated from one another at different locations of the x-y plane. With the BCM learning rule, the clustering algorithm worked extremely well with this test set, detecting the clusters to be what one expect them to be by looking at them. This is shown in Fig. 5. Original Data set

-1

Algorithm Detected Clusters

1.5

0.5

-0.5

0.5 -0.5

1.5

-1

-0.5

0.5

1.5

-0.5

Figure 5. Original 2D data set vs. Algorithm defined clusters using BCM. -1 -1

206

The Manhattan Scientist, Series B, Volume 6 (2019)

Campiglia

I also tested a similar dataset with an Oja neuron, with results exhibited in Fig. 6. Original Data set

-1

Algorithm Detected Clusters

1.5

0.5

-0.5

0.5

1.5

-1

-0.5

0.5

1.5

-0.5

Figure 6. Original 2D data set vs. Algorithm defined -1clusters using OJA. -1

Next, an artificial 3D data set was created and I performed the cluster detection algorithm using BCM and Ojaâ&#x20AC;&#x2122;s rule, as shown in Figs. 7 and 8. Original Data set

Algorithm Detected Clusters

1.5

0.5

1 1

0.5

0.5 0.5 0

0.5 0

Figure 7. Original 3D data set vs. Algorithm defined clusters using BCM. Original Data set

Algorithm Detected Clusters

0.5

0 1

1 1 0.5

0.5 0

1 0.5

0.5 0

Figure 8. Original 3D data set vs. Algorithm defined clusters using OJA.

Finally, I tested the algorithm on a real dataset from the UCI Machine Learning Repository

Campiglia

The Manhattan Scientist, Series B, Volume 6 (2019)

207

[15]. The first real dataset that I tested was the Iris dataset. This dataset included five variables, with the first four representing the lengths and widths of the petals and sepals of Iris flowers, and the fifth variable representing their class type, in this case being either Iris Setosa, Iris Versicolor, or Iris Virginica. The dataset ran through the algorithm with all four variables, but since you cannot visualize a four dimensional dataset, only the first three variables are plotted in Figs. 9 and 10. Original Clusters

Algorithm Detected Clusters Setosa Versicolor Virginica

Setosa Versicolor Virginica

0.8

0.6

0.4

0.2

0 5

0 0.5 8

4 6

0.8

0.4

0.7 0.6

0.3

5 2

0.5 0.2

0.4

Figure 9. Original Iris data set classes vs. Algorithm defined clusters using BCM.

Original Clusters

Algorithm Detected Clusters Setosa Versicolor Virginica

Setosa Versicolor Virginica

0.8

0.6

0.4

0.2

0 5

0 0.5 8

7 6

0.8

0.4

0.7 0.6

0.3

5 2

0.5 0.2

0.4

Figure 10. Original Iris data set classes vs. Algorithm defined clusters using OJA.

The other real dataset that I tested was the Seeds dataset. This dataset included eight variables, with the first seven representing the various parameters of the seeds, including the area, perimeter, compactness, asymmetry, etc., and the eighth variable representing their class type, in this case being either type 1, 2, or 3. The dataset ran through the algorithm with all seven variables, but to visualize the clusters, only three of the variables are plotted in Figs. 11 and 12.

208

The Manhattan Scientist, Series B, Volume 6 (2019) Original Classes

Campiglia

Algorithm Detected Clusters Seed 1 Seed 2 Seed 3

Seed 1 Seed 2 Seed 3

-2

-2 2

2 2 0

2 0

0 -2

-2

Figure 11. Original seeds data set classes vs. Algorithm defined clusters using BCM.

Original Classes

Algorithm Detected Clusters Seed 1 Seed 2 Seed 3

Seed 1 Seed 2 Seed 3

-2

-2 2

2 2 0

2 0

0 -2

-2

Figure 12. Original seeds data set classes vs. Algorithm defined clusters using OJA.

Discussion For the most part, the results of this study supported the idea that data points near each other would trigger similar levels of synaptic responses, and therefore would be grouped into clusters after being ran through the algorithm. The clearest example of this is displayed in Fig. 5, where the distinctly separated groups of data were grouped into visually expected clusters using the BCM neuron. One reason for these results on this artificial dataset is because the clusters are linearly separable, meaning that a plane exists which can be used to separate each cluster from the rest of the dataset [12]. We found that the best results for all datasets occurred when Îˇ = 0.01 and Ď&#x201E; = 10. In both artificial datasets for the Oja neuron, Fig. 6 and Fig. 8, the clusters formed by the algorithm did not match with how one would group the data visually. To look further into why this was happening, I plotted the synaptic responses of all the inputs, organized in a way so that the visual clusters followed one another, during each iteration of the algorithm. After completing

Campiglia

The Manhattan Scientist, Series B, Volume 6 (2019)

209

Figure 13. Original synaptic responses vs. Responses after a cluster was removed

this, it was found that the synaptic responses of the groups were originally similar to one another and distinctly unique from the other clusters, supporting the idea of neuronal selectivity described earlier. However, it was discovered that after the first cluster was removed from the dataset and the neuron was retrained, the remaining synaptic responses did not show any selectivity, which can explain why the algorithm did not produce expected results for the Oja neuron. These findings are displayed in Fig. 13 above. It was also found that in both neurons for the Iris dataset, the cluster that was distinctly seperate from the rest of the data was detected with the algorithm, while the rest of the data formed slightly different clusters each test. The work completed in this study can serve as a basic feature detection tool and can help compress larger datasets by grouping certain columns into clusters. There are many situations where having an extra column which groups multiple attributes in a dataset can be very useful, such as in analyzing patterns or presenting statistical results to clients in a way that is easier to visualize and understand. One example of this applciation is seen in [16, 17], where Intrator et al. combined well-known statistical feature extraction models with the BCM learning rule to develop an algorithm for recognizing 3D objects from a 2D input image. One path of future work that could be taken with this study is exploring the implications and details of implementing a lateral inhibition network of BCM neurons and observing the similarities and differences with the new system. Another potential area of improvement would involve modifying the algorithm to handle other data types other than purely numerical data, as well being able to handle data that does not have the same number of attributes across the entire data set. We also plan to find ways to improve the Oja neuron results, with tactics such as weight or input normalization.

Acknowledgments This research was supported by the School of Science Research Scholars Program. The author would like to thank Dr. Lawrence Udeigwe for his support and guidance throughout the research.

210

The Manhattan Scientist, Series B, Volume 6 (2019)

Campiglia

References [1] Oja, Erkki.“Oja Learning Rule.” Scholarpedia, vol. 3, no. 3, 2008, p. 3612, 10.4249/scholarpedia.3612. [2] Bear, Mark F., Barry W. Connors, and Michael A. Paradiso, eds. Neuroscience. Vol. 2. Lippincott Williams and Wilkins, 2007. [3] Klein, Raymond. “Donald Olding Hebb.” Scholarpedia, vol. 6, no. 4, 2011, p. 3719, 10.4249/scholarpedia.3719. [4] Frenkel, Mikhail Y., and Mark F. Bear. “How monocular deprivation shifts ocular dominance in visual cortex of young mice.” Neuron 44.6 (2004): 917-923. [5] Oja, Erkki. “Principal components, minor components, and linear neural networks.” Neural networks 5.6 (1992): 927-935. [6] Oja, Erkki. “PCA, ICA, and nonlinear Hebbian learning.” Proc. Int. Conf. on Artificial Neural Networks (ICANN’95). 1995. [7] R. Ng and J. Han. “Efficient and effective clustering method for spatial data mining”. In: Proceedings of the 20th VLDB Conference, pages 144–155, Santiago, Chile, 1994. [8] Berkhin, Pavel. “A survey of clustering data mining techniques.” Grouping multidimensional data. Springer, Berlin, Heidelberg, 2006. 25-71. [9] Jayesh Bapu Ahire. Ostojic, Srdjan, and Nicolas Brunel. “From spiking neuron models to linear-nonlinear models.” PLoS computational biology 7.1 (2011): e1001056. [10] Bienenstock, Elie L., Leon N. Cooper, and Paul W. Munro. “Theory for the development of neuron selectivity: orientation specificity and binocular interaction in visual cortex.” Journal of Neuroscience 2.1 (1982): 32-48. [11] Blais, Brian S., et al. The role of the environment in synaptic plasticity: towards an understanding of learning and memory. Diss. Brown University, 1998. [12] Udeigwe, Lawrence C., Paul W. Munro, and G. Bard Ermentrout. “Emergent dynamical properties of the BCM learning rule.” The Journal of Mathematical Neuroscience 7.1 (2017) [13] Abbott, Larry F., and Sacha B. Nelson. “Synaptic plasticity: taming the beast.” Nature neuroscience 3.11s (2000): 1178. [14] Frégnac, Yves, and Michel Imbert. “Development of neuronal selectivity in primary visual cortex of cat.” Physiological Reviews 64.1 (1984): 325-434. [15] Dua, D. and Graff, C. (2019). UCI Machine Learning Repository. Irvine, CA: University of California. [http://archive.ics.uci.edu/ml] [16] Intrator, Nathan, et al. Three Dimensional Object Recognition Using an Unsupervised Neural Network: Understanding the Distinguishing Features. No. TR-63. Brown Univ Providence RI Inst for Brain and Neural Systems, 1992. [17] Intrator, Nathan. “Feature extraction using an unsupervised neural network.” Connectionist Models. Morgan Kaufmann, 1991. 310-318.

Using laterally inhibiting neurons to detect clusters Sebastian Peña∗ Department of Mathematics, Manhattan College Abstract. It is known that, over time, a sensory neuron ‘learns’ to be selective and respond strongly to a particular set of stimuli and weakly to other stimuli. The BCM theory models the learning process of a sensory neuron. Treating points in a data set as stimuli, this learning model can ‘teach’ a neuron to respond strongly to a subset of the data set and weakly to the remainder of the data set. In the present work, the BCM theory is modified to incorporate multiple neurons that inhibit one another, and result in a competitive environment where each neuron ends up ‘choosing’ a cluster in the data set.

Introduction While the physical organization of the central nervous system is established during embryonic development, the connections between neurons can be modified through outside influence. This capacity is termed neural plasticity [1]. The modification occurs at the synapses. There are several mathematical models that can simulate the underlying mechanism of neural plasticity. The basis for many of these is Hebb’s rule (named after Donald Hebb) where it is proposed that when neuron A repeatedly participates in firing neuron B, the strength of the action of A onto B increases [2]. This implies a system where changes in synaptic strengths in a neural network is a function of the pre-synaptic and post-synaptic neural activities [3]. The standard relationship between a stimulus and the response it generates is y = w · x, where: • x is a pattern vector representing a pre-synaptic stimulus. • w is a vector of synaptic efficacies or weights [4]. This vector is what ultimately decides what type of effect the stimulus, in the form of a pattern vector, will have on a neuron. • y is a scalar response of a neuron to the stimulus. A relatively high numerical value for y implies a strong response. What differentiates the several learning rules that model neural plasticity is the mechanism that guides the development of the weight vector. Hebb’s rule allows the weight vector to grow indefinitely. Over sufficient time, all the responses from a neuron become saturated. In this scenario it becomes impossible to distinguish the responses and identify to which stimulus each response corresponds. The BCM learning rule named after Ellie Bienenstock, Leon Cooper, and Paul Munro includes a dynamic threshold which prevents the weight vector, w, and in turn the response, y, from ‘blowing up.’ A response above this threshold strengthens the active synapse ∗

Research mentored by Lawrence Udeigwe, Ph.D.

212

The Manhattan Scientist, Series B, Volume 6 (2019)

Peña

while a response below the threshold leads to a weakening of the active synapse [3]. The threshold is continuosly adapted to control the changing weight vector and as a result, the response. Sensory neurons which include visual, auditory, tactile, and olfactory neurons directly receive stimuli from the environment. The learning mechanism of these neurons can be modeled using the BCM learning rule and over many iterations they will be significantly selective with the stimuli they receive. Selectivity refers to the ability of a neuron to ‘choose’ a group of stimuli for which it will produce a strong response and a weak response for all other stimuli. The higher the selectivity, the easier it becomes to identify the stimuli corresponding to the strong responses. The selectivity can be further improved through a modification of the BCM rule incorporating another biological observation: lateral inhibition [5]. Lateral inhibition is a process by which a sensory neuron that gets excited by a stimulus, inhibits the response of its neighboring neuron to the same stimulus. The ultimate consequence of the modification of the model is that each neuron participating in the setup, will eventually become selective to a different set of stimuli. A slight deviation needs to be taken to relate the properties described above to a particular application; namely clustering. Clustering can be considered the most important unsupervised learning problem [6]. Its applications are in many fields such as machine learning, pattern recognition, image analysis, bioinformatics and data mining [6]. It is a procedure by which structure is found in a collection of data where class types are not available. The objects belonging to one cluster are similar to each other while significantly different from the objects of another cluster. Data points having a small euclidean distance can be thought of as similar objects. Consider treating the data points of a data set as stimuli being presented to multiple neurons that laterally inhibit each other. It is thus reasonable to assume that when the neurons become selective (after being trained using the BCM learning rule), each neuron will choose a distinct cluster in the data set. The goal of this project is to develop a clustering algorithm that makes use of this underlying mechanism.

Methods General BCM learning rule The BCM learning rule comprises two equations. Shown in their differential form below, they deal with the rate of change of the weights, and the rate of change of the dynamic threshold with respect to time. dw = y(y − θ)x (1) τw dt dθ τθ = (y 2 − θ) (2) dt where: • •

x is a pattern vector representing a pre-synaptic stimulus. w is a vector of synaptic efficacies or weights.

Peña • • • •

The Manhattan Scientist, Series B, Volume 6 (2019)

213

y is a scalar response of a neuron to the stimulus. θ is the dynamic threshold. τw is the time scale factor which adjusts how fast the weights change with respect to time. τθ is the time scale factor which adjusts how fast the dynamic threshold changes with respect to time. Through investigation by Udeigwe et al. [3] it has been determined that τw is significantly larger than τθ so as to reach stability.

Adaptation of BCM learning rule for incorporating lateral inhibition This section implements the mathematical changes to the BCM learning rule necessary to incorporate the phenomenom of lateral inhibition. The modification is exclusive to the computation of the net response of a neuron to a stimulus. Under the new formulation, the inhibition caused by neighboring neurons will be taken into account. For the sake of simplicity, assume there are three neurons a, b, and c, laterally inhibiting each other. Looking at neuron a individually, the following is assumed [3]: sa = w a · x (3)

where: • x is a pattern vector representing a pre-synaptic stimulus. • w a is the synaptic weight vector for neuron a. • s a is the partial activity (scalar) induced by a stimulus x to neuron a. Similar equations can be formed for the other two neurons. However, to simplify the process, the partial activities for all neurons can be calculated in one step using matrices:     sa wa s =  sb  =  wb  x. sc wc s is a vector containing the partial activities of all the participating neurons as scalars. wa , wb , wc are the synaptic weight vectors, for each neuron, which are now laid out as rows. Since each neuron laterally inhibits each neuron next to it, there must be a way of quantifying this inhibition. The inhibition parameter which serves that purpose, is represented by the letter γ [3]. Still assuming the presence of only 3 neurons, the net activities of the neurons can now be computed using matrices in one step as well: y = G−1 s (4) where

and

  ya  y = yb  yc 

 1 γ γ G = γ 1 γ  γ γ 1

214

The Manhattan Scientist, Series B, Volume 6 (2019)

Peña

Assuming n neurons are in the system, the dimension of G must always be n by n. A diagram of the laterally inhibiting neurons is aptly described in Fig. 1: The negative symbol adjacent to the γ

Figure 1: Three laterally inhibiting neurons

symbol implies the inhibitory effect of two of the neurons to the third neuron. Discretization of BCM learning rule In order to observe the development of the responses as well as that of the weights, equations 1 and 2 need to be discretized. The following are the discrete modes of both: θi+1 = θi +

1 2 (y − θ) τθ

wi+1 = wi + ηy(y − θ)x

(5) (6)

η, the learning rate of a neuron is the equivalent of 1/τw . However, the procedure of lateral inhibition requires the participation of multiple neurons. Therefore to avoid building a tedious algorithm with countless loops, computing θ and w using equations 5 and 6 respectively for every participating neuron, a matrix form of the same procedure can be adopted. Continuing with the assumption that there are 3 neurons: a, b, and c, the following are obtained:  2   !  i+1  i ya θa θa θa  θb  =  θb  + 1  yb  −  θb  (7) τw yc θc θc θc  i+1  i   wa wa ya (ya − θa )xi  wb  =  wb  + dt · η ·  yb (yb − θb )xi  (8) wc wc yc (yc − θc )xi

The xi in equation 8 represents a particular stimulus selected for which to consider a corresponding response by 3 neurons. The source for many xi is a data set, X. In the iterative learning

Peña

The Manhattan Scientist, Series B, Volume 6 (2019)

215

process of the neurons, an xi can be selected randomly from the data set under analysis. xi is the ith row of the data set. Given enough iterations, the weight vectors of every single neuron converge to constant values. These stable weight vectors can be used to yield the stable responses of the neurons to all data points (stimuli). In general: R = XW T

(9)

where: • W is the matrix containing the weight vectors, for each neuron, as rows. • X is a data set. • R is a matrix with the following characteristics: – The elements of the ith row represent the responses of the neurons to the ith stimulus. – Assuming the number of rows in the data set is m, and n is the number of neurons in the system, the dimension of R is m × n.

Outline of Algorithms Applying the modified BCM learning rule to compute neural responses The discretized equations are implemented in an algorithm where the responses of each neuron are calculated for every single stimulus. This process is repeated enough times through the iterations, so that at the end all the weight vectors (and in turn neuron responses), will have become stable. The input parameters for the procedure are the following: • Number of neurons, n. • Initial weight matrix, W . Each row of W 0 0 is the weight vector associated with one neuron. Assuming n = 3, W0 must have 3 rows. All the initial row vectors of W0 have random elements very close to 0 in value. • Input data set X. This matrix contains the row vectors x which will play the role of input i stimuli. • Time step, dt. • Inhibition parameter, γ. • Total number of iterations. • Time scale factors for the weights and sliding threshold, τ w and τθ respectively. Algorithm 1 determines neural responses, R. Using lateral inhibiting neurons for clustering data points Algorithm 1 has as its output the variable R which contains the responses of every neuron for every data point in X. A new R is computed for every iteration. To cluster the data points from data set X, only the R computed in the last iteration is required. Every row of R corresponds to the responses of the neurons to one stimulus. Every column corresponds to a particular neuron. In any given row, the neuron with highest numerical response keeps the corresponding data point for its cluster. Algorithm 2 performs the clustering procedure:

216

The Manhattan Scientist, Series B, Volume 6 (2019)

Peña

Algorithm 1 Algorithm of BCM Learning Rule with Lateral Inhibition procedure BCM NETWORK(n, W0 , X, dt, γ, Iterations, τw , τθ ) for i ← 1 to Iterations do Select random row vector xi from X. Multiply the row vector xi with W0 resulting in s vector. Construct G using γ. Find the inverse of G and multiply it by s to obtain net responses. Calculate θ. Calculate new W . Multiply X by the tranpose of W to obtain R. Algorithm 2 Using lateral inhibtion to cluster data points Assign values to the parameters. Load the data set. Apply Algorithm 1 and obtain output R. Create an array that is able to store n data clusters of different sizes, side by side. for i ← 1 to length of data set do for j ← 1 to n do if the jth element in the ith row of R = the maximum element in the ith row of R then ith row of X is appended to jth cluster. else Append row of zeros to the jth cluster. Turn array into numeric matrix for plotting.

Results Certain required parameters for the procedure in Algorithm 1 are constant and cannot be modified in the real biological process. The approximate values these variables can take, are based on previous investigation by Udeigwe et al. [3]. These values are maintained constant for the analysis of all the different data sets used in the results that follow. τw = 100 τθ = 10 γ = .25 dt = .05 The tasks achieved by the algorithms include: •

•

Observe closely the evolution of a few responses by a few neurons over the increasing number of iterations. This task provides a way into understanding why clustering can be performed. The stimuli causing the responses are artificially created to be sufficiently distinct from each other. Cluster the data points of different data sets by using the neuron responses corresponding to different stimuli represented by the data points.

Peña

The Manhattan Scientist, Series B, Volume 6 (2019)

217

Development of responses generated by neighboring neurons Algorithm 1 can be used to observe the evolution of each neuron’s response for every single data point in a data set. This is done by observing how R changes with every iteration. To clearly demonstrate the development in a graph, three very distinct points can be synthesized and appended together as a matrix, X. (Graphing the evolution of responses for an entire data set would create visual disorder). The three points are shown in Fig. 2.

Figure 2: Three input patterns

For the sake of consistency three neurons are chosen for the implementation of Algorithm 1. Correspondingly, three graphs are shown in Fig. 3. Each graph corresponds to the responses of one individual neuron to the three different data points, with the number of iterations as the independent variable. Over time, each neuron develops a strong selectivity and hence has a strong response to one data point, and develops a weak response for the other two data points. Table 1 shows the final iteration net responses of each neuron to the three data points. All three neurons have a data point for which a higher net response was generated compared to the other two points. Table 1: Final responses of neurons to three different stimuli. Stimulus

Neuron A

Neuron B

Neuron C

(1, 1, .05) (1, .05, 1) (.05, 1, 1)

.6361 2.4815 .6425

.5954 .5892 2.4714

2.5976 .6606 .6740

One can assume that similar data points, having a small euclidean distance, will correspond to similar responses by each neuron. For example, a data point very close to (1, 1, 0.05) will lead neurons a and b to produce low responses, while leading neuron c to produce a high response. Therefore, in a backwards fashion, the responses corresponding to many data points from a data set can be used to determine different clusters. A neuron will ‘choose’ a cluster for which it produces a high response. Other neighboring neurons will in turn ‘choose’ other different clusters for which they will produce a high response. The data points in each cluster should ideally be similar and be

218

The Manhattan Scientist, Series B, Volume 6 (2019)

(a) First neutron: second point is favored

PenĚ&#x192;a

(b) Second neutron: first point is favored

Figure 3: Evolution of responses of 3 neurons for 3 stimuli, versus the number of iterations

very different from data points of another cluster. Clustering the Seeds data set The seeds data set contains seven attributes for three distinct wheat varieties (three data classes). These attributes correspond to different physical parameters of wheat kernels [7]. There are a total of 210 data points. For the purpose of plotting the data points, only three of its attributes are selected as a matrix and assigned to the variable X. For the sake of visual comparison, X is plotted with a color assigned to each of its data points according to its class, which can either be 1, 2, or 3. This plot is then compared to the plot of the clusters determined by Algorithm 2. Given

PenĚ&#x192;a

The Manhattan Scientist, Series B, Volume 6 (2019)

219

that there are 3 classes in the data set, it is most appropriate to set n equal to 3, so that 3 clusters form in using Algorithm 2. The results are shown in Fig. 4. The plot with points grouped by class

(a) Distinction by class type

(b) Lateral inhibition clusters

Figure 4: Comparison of clustering by class and lateral inhibition

type, shows several points belonging to different classes being mixed together. Hence there are no definite delineations that separate the groups from each other. The clusters obtained by lateral inhibition show virtually no mixing of data points assigned to different clusters. It is also visibly clear that many points in the two plots do not coincide in the way they were assigned to a specific color. In other words, several points of distinct classes were clustered together. Clustering the Iris dataset A second data set chosen to test Algorithm 2 was the popular iris data set which contains four attributes corresponding to four physical parameters of the genus Iris. There are 150 data points for three Iris species; Iris Versicolor, Iris Virginica and Iris Setosa which correspond to 3 class types: 1, 2 and 3. To be able to plot the data, only three of its attributes are selected. Due to the similarity in structure with the seeds data set, the same procedure was followed. The results are shown in Fig. 5. Independent of the color assigned to the points, the two plots seem to suggest the presence of two distinct group of points. Algorithm 2 can identify these two distinct clusters rather easily if only two neurons are allowed to participate (two neurons will choose two clusters). If three neurons are instead allowed to participate (corresponding to there being three class types in the data set) there is significant loss of clarity in the distinctiveness of the clusters. Application in two-dimensional data set Algorithm 2 can be applied to two dimensional inputs. For this purpose, two attributes of the seeds data set are selected as X. All other parameter values are kept the same except for the total

220

The Manhattan Scientist, Series B, Volume 6 (2019)

(a) Distinction by class type

PenĚ&#x192;a

(b) Lateral Inhibition clustering.

Figure 5: Comparison of grouping by class and lateral inhibition clustering

number of iterations. The results are shown in Fig. 7. The plots show the great relative success of

(a) Distinction by class type

(b) Lateral Inhibition clustering.

Figure 6: Comparison of clustering by class and lateral inhibition

lateral inhibition in mimicking the classification of points by class. Like Fig. 6, Fig. 7 demonstrates the difference between the plot of points distinguished by class and the plot of clusters determined by lateral inhibition, with respect to different groups mixing. The plot with points grouped by class has several overlapping points whereas the plot of clusters determined by lateral inhibition does not.

PenĚ&#x192;a

The Manhattan Scientist, Series B, Volume 6 (2019)

221

Discussion The modification of BCM learning rule by incorporating lateral inhibition greatly increases selectivity. Over many iterations each neuron yields very different responses for very different stimuli. This result allows the model to cluster similar stimuli together. However, the nature of the data set, as well as chance, plays a role in affecting the effectiveness of clustering. The clustering algorithm is more effective if the data points in a given data set already have groups that are compact and separated from each other. The number of iterations has a significant effect on the effectiviness of clustering. This parameter was chosen arbitrarily for all of the sections of the project. At the same time, it was observed that there was not a strong consistency in the way the data points were clustered. Every time the algorithm would be implemented, keeping every parameter constant, significantly different visual results were obtained. This was especially pronounced with the iris data set. The structure of the iris data set allows its points to be grouped in a variety of valid ways.

Conclusion Algorithm 2 seems to work most appropriately when the clusters defined by many data points are significantly, and visibly in the case of plot, distinct. Its relative success demonstrates that the model simulates effectively the concept of stimulus selectivity and neural plasticity. The model could potentially be useful in predicting or guiding how a network of neurons would respond to a continuous external stimuli. There needs to be some form of determining the appropriate number of iterations to run the algorithms, depending on the type of data set employed. It would be also important to understand why certain number of iterations are appropriate for specific types of data sets and how can this be related to the real biological process of neural learning. Other parameters which were chosen based on previous investigation, perhaps can be modified to optimize selectivity, and in turn clustering efficiency. Data sets containing more than three class types would be useful for comparing the clusters one could obtain by implementing Algorithm 2, with the â&#x20AC;&#x2DC;clustersâ&#x20AC;&#x2122; of points with the same class.

Acknowledgements This work was supported by the School of Science Research Scholars Program. The author wishes to thank Dr. Udeigwe for his mentorship and patience as well as the school of science for its support in motivating research.

References [1] Campbell N., Reece Berkeley J. B., Urry L. A., Cain M. L., Wasserman S. A., Minorsky P. V., and Jackson R. B. (2014). Neuron, Synapses, and signaling. Campbell Biology 10th Edition. Chapter 48, pp. 1061-1094 [2] Hebb D. (1949). The organization of behavior. Wiley, New York. [3] Udeigwe L C, Munro P. W., and Ermentrout G. B. (2017). Emergent Dynamical Properties of the BCM Learning Rule. Journal of Mathematical Neuroscience. pp. 2-20

222

The Manhattan Scientist, Series B, Volume 6 (2019)

PenĚ&#x192;a

[4] Blais B. S. (1998). The Role of the Environment in Synaptic Plasticity: Towards an Understanding of Learning and Memory. pp. 1-13 [5] Castellani G. C., Intrator N., Shouval H., Cooper L. (1999). Solutions of the BCM learning rule in a network of lateral interacting nonlinear neurons. New Comput Neural Syst. pp. 111â&#x20AC;&#x201C;21. [6] Madhulatha T. S. (2012). An Overview On Clustering Methods. IOSR Journal of Engineering, Vol. 2(4) pp: 719-725. arXiv:1205.1117 [cs.DS] [7] Charytanowicz M., Niewczas J., Kulczycki P., Kowalski P. A., Lukasik S., and Zak S. (2012). Seeds Data Set. UCI Machine Learning Repository

Monte Carlo computer investigations of higher generation ideal dendrimers Brandon Thrope∗ Department of Mathematics, Manhattan College Abstract. A polymer is a molecule consisting of many small identical units bonded together and a dendrimer is a type of symmetrical tree structured polymer. This study uses Monte Carlo simulations to study properties of tri-functional dendrimers having forty-five, ninety-three or one hundred and eighty-nine branches. Excellent agreement is found with theoretical predictions.

Introduction Polymers are macromolecules formed by the bonding of a large number of small identical units. The architecture of a polymer is determined by how these units are linked together. The simplest polymer is a linear chain. Three or more linear chains can be grown from a single junction to form a star polymer. If the ends of each chain are then used in turn repeatedly as junctions for further growth one obtains a dendrimer. These polymers have the shape of a branching tree and can be classified by their functionality (number of linear branches growing from a junction), the spacer size (the number of units in a linear branch), and the generation. A generation is defined by the number of rings of branches starting from a single central junction. Fig. 1 illustrates a typical dendrimer. The diagram displays a 93-branch dendrimer. This is the fourth generation in its family. The circles represent the junctions. The first three branches attached to the central junction represent a star polymer. The next ring, containing six more branches, represents the first dendrimer generation, a molecule with nine branches. The second generation has twenty-one branches, the third, forty-five and the fourth, ninety-three. The fifth generation will have one hundred and eighty-nine branches so that its structure would include another ring of branches. Two more branches would extend from each circle, adding 96 branches to the structure for a total of 189. The reason for the 9 − 21 − 45 − 93 − 189 branch numbers is that the number being added doubles; 12 branches are added to 9 to get 21, 24 to 21 to get 45, etc. The number of units in each branch is uniform throughout the entire structure and can be any value. The total number of units, N , in each dendrimer is N = NB ·(n−1)+1 where NB is the number of branches defined by the dendrimer generation and n is the number of units in each branch. In our study N ranged from 901 to 4915. Dendrimers have received a lot of attention [1] because they have applications in medical research, such as in pharmaceutical drug development. Their branched structure allows these molecules to be folded up into cages which can carry drugs into the body. The symmetry, geometry, and properties of the dendrimers examined will be of interest to pharmaceutical researchers. ∗

Research mentored by Marvin Bishop, Ph.D.

224

The Manhattan Scientist, Series B, Volume 6 (2019)

Thrope

Figure 1. 93-branch dendrimer (fourth generation)

In earlier work de Regt et al. [2] studied the properties of low generation tri-functional ideal structures containing nine or twenty-one branches. In this paper we extend their investigations to three higher generations, containing forty-five, ninety-three, and one hundred eighty-nine branches, respectively.

Method Monte Carlo (MC) simulations generate random configurations and can be used to research polymeric materials. The dendrimers are grown on a three dimensional simple cubic lattice. The central unit is placed at the origin of the coordinate system. Then a random number, from 0 to 5, is generated. This number selects one of the six possible directions for placing the second unit a distance one apart from the central unit: a random number of 0 indicates plus 1 in the X-direction for the next unit in that branch, 1 indicates -1 for X, 2 is +1 for Y , 3 is -1 for Y , 4 is +1 for Z, and 5 is -1 for Z. Each successive placement of units is performed by this procedure. Since ideal dendrimers are the focus of this research the units are allowed to overlap. This aspect makes the model less realistic but greatly decreases the amount of computer time needed. Even so, the bigger systems of the ideal model required a month of continuous running time on the School of Scienceâ&#x20AC;&#x2122;s supercomputer system. The C code was compiled and run in a Linux environment which made controlling unattended batch runs easy. The original MC code of de Regt et al. [2] was generalized to enhance both its clarity and efficiency. The new code was tested by reproducing their results for nine and twenty-one branch structures. Each completed dendimer is an independent configuration and properties were computed by averaging over each of these random samples. In all cases, 100,000 random samples were employed. The mean-square radius of gyration [3], hS 2 i, is used to measure the total size of a dendrimer. The mean-square radius of gyration of an object composed of N identical units, around its center of mass, is

Thrope

The Manhattan Scientist, Series B, Volume 6 (2019) N 1 X hS i = h(Xi − XCM )2 + (Yi − Y CM )2 + (Zi − ZCM )2 i N i 2

225

(1)

where Xi , Yi and Zi are the components of the position of the i-th unit and the center of mass is defined by N N N 1 X 1 X 1 X XCM = Xi ; Y CM = Yi ; ZCM = Zi . (2) N i N i N i

The braces hi indicate an average over multiple dendrimer constructions.

Sheng et al. [4] determined an exact equation for the radius of gyration of ideal tri-functional dendrimers: hS 2 iN 2 =3m3 (−1 + 6 · 2G1 + (3G1 − 5)22G1 ) hl2 i − 0.5m(m − 1)(2G1 − 1)(9 · 2G1 m − 7m + 2),

(3)

where N = 3m(2G1 − 1) + 1, hl2 i is the mean-square branch length (1 for the cubic lattice used in the MC simulations), and G1 is one more than the generation number. For linear chains, hS 2 i, follows the well-known [5] scaling law, hS 2 i = C(N − 1)2ν .

(4)

In this equation, the value of C depends on what polymer model is used, but the exponent 2ν is universal and is equal to 1 for ideal linear chains. Sheng et al. [4] found that for a given generation, the size of an ideal dendrimer has the same exponent value. In order to measure the compactness of the dendrimer’s structure, the g-ratio was calculated. This calculation involves the ratio of the radius of gyration of a dendrimer and the corresponding linear polymer chain. They must both contain the same number of units: hS 2 i hS 2 il

(5)

6(W1 − W2 ) . (n/m)3

(6)

F1 (F 2G (F1 F2 G − F1 G − F1 − F2 ) (F2 − 1)3 2 + F2G (2F1 + F1 G − F1 F2 G − G + GF22 ) − F1 + F2 )

(7)

There is an exact solution for the g-ratio determined by Wawrzyriska et al. [6] : g= Here, n is the number of bonds (N − 1), W1 =

226

The Manhattan Scientist, Series B, Volume 6 (2019)

and W2 =

F12 (F2G − 1)2 F1 (F2G − 1) . + 2(F2 − 1)2 3(F2 − 1)

Thrope

(8)

The overall shape of the dendrimer can be found using the radius of gyration tensor. This tensor is composed of different eigenvalues. 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 [7]. hS 2 i is equal to the average trace of the radius of gyration tensor, λ1 +λ2 +λ3 . The average asphericity hAi of a dendrimer in three dimensions is defined by Rudnick and Gaspari [8, 9] as + * P3 2 i>j (λi − λj ) . (9) hAi = P 2( 3i=1 λi )2

hAi is a number between 0 and 1, with 0 representing a perfectly spherical object, while an asphericity of 1 indicates that the units form a rod shape. The prolateness, hP i, is also used to identify the overall shape. It is defined as * + 27(λ1 − λ)(λ2 − λ)(λ3 − λ) hP i = (10) P ( 3i=1 λi )3 where λ is

λ=

λ1 + λ2 + λ3 3

(11)

Another important polymer property is the scattering function [10], S(k). Formally, this is defined as N N 1 X X ik·(Rm −Rl ) S(k) = 2 e (12) N l m Here, N is the number of units in the polymer, k is the scattering vector and Rl and Rm are the respective positions of the l-th and m-th units. After averaging over the angles in three dimensions, the scattering function becomes * N N + 1 X X sin x S(k) = 2 (13) N x 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 [5], S(k) =

2(x − 1 + e−x ) . x2

(14)

Thrope

The Manhattan Scientist, Series B, Volume 6 (2019)

227

Dendrimers are not linear and each structure will have its own equation for S(k). The linear equation is included in one of the figures for comparison with the behavior of other dendrimers. The exact [2] 9-branch dendrimer S(k) follows the equation: S(k)9 =

2(3 + x − 3e−x/9 − 12e−x/3 + 12e−4x/9 ) x2

(15)

whereas, the form factor for 21-branch dendrimers is: S(k)21 =

2 (9 + x − 3e−x/21 + 6e−2x/21 − 12e−x/7 − 48e−5x/21 + 48e−2x/7 ) x2

(16)

Results Tables 1-3 show the property values found for the different branch dendrimers as a function of the number of units N . hλi values for each of the eigenvalues are included, as well as, hAi and hP i. The radius of gyration is presented for both the dendrimer, and 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 26.79(3) would mean 26.79 ± 0.03. Table 1. Effect of the number of units, N , for 45-branch dendrimers. Property/N

901

1531

1756

2206

hλ1 i hλ2 i hλ3 i hAi hS 2 i hP i hS 2 i`

26.79(3) 12.78(1) 6.57(1) 0.160(1) 46.14(4) 0.095(1) 149.51(24)

45.45(5) 21.60(2) 11.09(1) 0.161(1) 78.13(7) 0.096(1) 254.05(42)

52.11(6) 24.81(3) 12.70(1) 0.161(1) 89.62(8) 0.096(1) 291.51(47)

65.54(8) 31.09(3) 15.93(2) 0.162(1) 112.56(10) 0.097(1) 366.89(60)

Table 2. Effect of the number of units, N , for 93-branch dendrimers. Property/N

931

1861

2791

3721

hλ1 i hλ2 i hλ3 i hAi hS 2 i hP i hS 2 i`

17.23(2) 9.15(1) 5.23(1) 0.121(1) 31.61(2) 0.059(1) 154.54(25)

34.28(4) 18.12(2) 10.32(1) 0.123(1) 62.73(5) 0.060 (1) 309.18(50)

51.36(5) 27.11(2) 15.44(1) 0.123(1) 93.91(7) 0.060(1) 465.57(76)

68.39(7) 36.12(3) 20.54(2) 0.123(1) 125.05(10) 0.060(1) 619.62(101)

228

The Manhattan Scientist, Series B, Volume 6 (2019)

Thrope

Table 3. Effect of the number of units, N , for 189-branch dendrimers. Property/N

946

2080

3592

4915

hλ1 i hλ2 i hλ3 i hAi hS 2 i hP i hS 2 i`

10.53(1) 6.12(5) 3.83(1) 0.091(1) 20.48(1) 0.036(1) 157.12(26)

22.94(2) 13.26(1) 8.26(1) 0.092(1) 44.47(3) 0.037(1) 345.21(56)

39.55(4) 22.76(2) 14.14(1) 0.093(1) 76.45(5) 0.038(1) 599.47(98)

54.10(5) 31.12(2) 19.32(1) 0.094(1) 104.53(7) 0.038(1) 819.05(134)

The tables contain some interesting results. For any dendrimer generation, hAi and hP i level off as the number of total units in the dendrimer increases. Another observation is that hS 2 i is always higher for the linear chains than it is for the dendrimers. This is to be expected because the dendrimer is less spread out than a linear chain and would therefore have an average shorter distance from its center of mass to an outer unit. Table 4 demonstrates that the radius of gyration data reported in Tables 1-3 are in excellent agreement with the Sheng et al. [4] exact equation. To determine the scaling exponent for the radii of gyration, Eq. 4, a C program was used to fit the hS 2 i data to a power function. Fig. 2 presents a log-log plot of the data for each distinct generation. The exponents found are: 0.995(1), 0.992(1), and 0.988(1) for 45, 93 and 189 branches, respectively. These are close to the expected theoretical exponent of 1.0. Table 4. Comparison of hS 2 i MC to the exact predictions of Sheng et al. [4]. N 901 1531 1756 2206

45 Branches MC Exact 46.14(4) 78.13(7) 89.62(8) 112.56(10)

46.161 78.195 89.636 112.517

N 931 1861 2791 3721

93 Branches MC Exact 31.61(2) 62.73(5) 93.91(7) 125.05(10)

31.578 62.723 93.868 125.014

N 946 2080 3592 4915

189 Branches MC Exact 20.48(1) 44.47(3) 76.45(5) 104.53(7)

20.469 44.484 76.503 104.520

The data in the tables are for finite N whereas most of the theories are for infinite N . Thus, the data for hAi and hP i have been extrapolated using a linear function in 1/N . The g-ratios were extrapolated after first determining their error by relating the error in a ratio to the error in the numerator and the error in the denominator. This extrapolation procedure is illustrated in Fig. 3 for hAi. As the number of units increases, 1/N gets smaller, so that when 1/N → 0, N becomes infinite. Hence, extrapolation should remove all finite N effects.

Thrope

The Manhattan Scientist, Series B, Volume 6 (2019)

229

Figure 2. Log-log plot of the radius of gyration vs.N for each dendrimer generation: circles 45-branch, squares 93-branch, triangles 189-branch, lines best linear fit.

Figure 3. Extrapolation of A data: circles 45-branch, squares 93-branch, triangles 189-branch, lines best fit.

Table 5 shows the extrapolation results for the various dendrimer properties examined in this research. Each of the extrapolations are compared to the exact calculations, determined using the Wei [11, 12] method. Excellent agreement is shown between all the extrapolated MC results and the theoretical predictions. Table 5. Property results for 45, 93 and 189-branch dendrimers. Property

45 Extrapolated

Wei

93 Extrapolated

Wei

189 Extrapolated

Wei

g â&#x2C6;&#x2019; ratio hAi hP i

0.306(2) 0.163(2) 0.098(1)

0.305088 0.1623100 0.0967390a

0.201(1) 0.124(1 ) 0.061(1)

0.200938 0.1237990 0.0609874a

0.128(1) 0.094(1) 0.039(1)

0.127062 0.0944898 0.0389064a

Note that the Wei method defines hP i with an additional factor of 1/2.

230

The Manhattan Scientist, Series B, Volume 6 (2019)

Thrope

The values of the asphericity reflect how spherical and symmetric a dendrimer is; an hAi value of 0.0 represents a perfect sphere. The data show that, as the number of branches increases, the values for hAi and hP i decrease. De Regt et al [2] found that hAi changed from 0.266(1) to 0.211(1) when going from the first generation 9-branch dendrimer to the second generation 21branch dendrimer; similarly hP i changed from 0.233(3) to 0.153(2). These values indicate that higher generation dendrimers have a more symmetric and spherical shape. In Table 6 the hAi and g-ratio results for ideal dendrimers are compared to Wawrzyriska et al. [13] values for a MC self-avoiding walk (SAW) model. One notes that the excluded volume effects in the SAW become more and more important as the generation number increases. Indeed, the excluded volume model is forced to be more symmetrical because the units cannot overlap. Table 6. Comparsion of ideal and SAW dendimers. NB 9 21 45 93 189 a

Ideal hAi

g-ratio

0.266(1) 0.211(1)a 0.163(2) 0.124(1) 0.094(1)

0.606(1) 0.443(1)a 0.306(2) 0.201(1) 0.128(1)

SAW hAi 0.2682 0.1968 0.1390 0.0972 0.0685

g-ratio 0.6068 0.4485 0.3180 0.2199 0.1509

See reference [2]

Fig. 4 shows the variation of a Kratky plot for the MC 93-branch dendrimer S(k) calculations as the total number of units in all the branches is increased from 931 to 3727. It is clear that number dependent effects are quite small if one uses a sufficiently large number of units in the simulations.

Figure 4. Number dependence for the 93-branch dendrimer scattering function.

Thrope

The Manhattan Scientist, Series B, Volume 6 (2019)

231

Fig. 5 shows results for the largest N , S(k) calculations for linear chains and 9, 21, 45, 93 and 189-branch dendrimers compared to available exact equations. This plot confirms the observation that higher generation dendrimers are more symmetrical.

Figure 5. Variation of the scattering function with dendrimer generation and number of units: squares, linear 2080; diamonds, 9-branch 1342; up triangles, 21-branch 3130; down triangles, 45-branch 2206; side triangles, 93-branch 3727; circles, 189-branch 4915; lines, exact equations.

Conclusions Three different dendrimer generations were studied 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. The use of these kinds of simulations gives researchers the ability to accurately model the properties of ideal dendrimers. In terms of future research, one could study other families of dendrimers, by having more branches per junction. The study was performed under ideal conditions, so the units could overlap with each other. Extending the simulations to the case in which the units interact with each other could lead to interesting conclusions about the behavior of dendrimers in the real world.

Acknowledgments This work was supported by the School of Science Research Scholars Program. The author would like to thank his advisor Dr. Marvin Bishop for helping him throughout this project. In addition, he is grateful to the mathematics department at Manhattan College for providing a place for him to work and the Manhattan College Computer Center and the Kakos Center for Scientific

232

The Manhattan Scientist, Series B, Volume 6 (2019)

Thrope

Computing for grants of computer time. He would also like to thank Dr. Robin de Regt for performing the analytical calculations of dendrimer properties. Lastly, he acknowledges the Jaspers Summer Research Program and its leaders for allowing him to participate in such a great program.

References A.W. Bosman, H.M. Hanssen, and E.W. Meijer, Chem. Rev,, 99, 1665 (1999). R. de Regt, C. von Ferber, M. Bishop, and T. Hamling, Physica A, 516, 50 (2019). P.J. Flory, Principles of Polymer Chemistry, (Cornell University Press, Ithaca, 1953). Y.J. Sheng, S. Jiang, and H.K. Tsao, Macromolecules, 35, (2002) 7865. P.G. de Gennes, Scaling Concepts in Polymer Physics, (Cornell University Press, Ithaca, 1979). E. Wawrzyriska, S. Eisenhaber, P. Parzuchowski, A. Sikorski, and G. Zifferer, Macromol. Theory Simul., 23, (2014) 288. [7] K. Solc and H. Stockmeyer, J. Chem. Phys., 54, 2756 (1971). [8] J. Rudnick and G. Gaspari, Science, 237, 384 (1987) and references therein [9] J. Rudnick and G. Gaspari, J. Phys. A, 19, L191 (1986). [10] P.J. Flory, Statistical Mechanics of Chain Molecules,(Hanser, Munich, 1989). [11] G. Wei, Physica A, 222, 152 (1995). [12] G. Wei, Physica A, 222, 155 (1995). [13] E. Wawrzyriska, A. Sikorski, and G. Zifferer, Macromol. Theory Simul., 24, (2015) 477. [1] [2] [3] [4] [5] [6]

Optimizing data acquisition for deep learning in magnetic resonance imaging Quinn Q. Torres∗ and Marcus L. Wong∗ Department of Mathematics, Manhattan College Abstract. Magnetic Resonance Imaging (MRI) uses large magnets, radio waves, and a computer system to show detailed images inside a person’s body. A downside of MRI is that it takes a long time to gather the data. To address the time required to collect the data, we explored three different acquisitions using half and a fourth of the data causing blur, aliasing, and a combination of both. The goal was to produce the best reconstructed image using a Convolutional Neural Network (CNN). Our CNN reduced blurring from the blurred image and aliasing from the alias + blur image but we found the CNN is unable to remove aliasing from only aliased images. This shows the importance of data acquisition in reconstruction with neural networks.

Introduction MRI is a useful tool in the medical field to look at tissues inside the body. However, a downside of MRI is that it is slow and therefore costly. Research is being conducted in order to optimize the speed of MRI. Our research explores how to optimize the speed by choosing less data to collect. Specifically, we divided the research into two parts, in one part we used 2× accelerated data, and in the other part we used 4× accelerated data. We used a convolutional neural network to reconstruct our accelerated images [1]. Our research will illustrate how the reconstructed images perform compared to each different type of data acquisition for 2× and 4× data accelerations in a similar way to previous work in compressed sensing [2].

Background Fourier transform The Fourier transform is a mathematical function that converts images into a representation using sines and cosines. Fourier transforms place an image in its frequency domain and display both low and high frequencies. We created our Fourier transforms in such a way where the low frequencies with the larger structural information of the image are closer to the center, and the higher frequencies that contain the smaller details of the image are towards the edge [3]. Masking, Aliasing and Blurring Masking is an image processing tool to take specific data from the image by setting some pixel values to zero and other pixel values to 1. The pixel values of zero are chosen to not be taken for further processing while the pixel values of one are. Different masks end up in images being processed in different ways (Fig. 1). For example, applying a mask that only processes the ∗

Research mentored by Angel Pineda, Ph.D.

234

The Manhattan Scientist, Series B, Volume 6 (2019)

Torres + Wong

frequencies in the middle of the Fourier transform results in a blurred image, while a mask that takes every other column of the image, results in an aliased image. Aliasing is an effect that causes different signals to become indistinguishable from one another when sampled. In simpler terms, aliasing displays an image overlapping with itself.

Figure 1. The masks we used for 2× accelerated data acquisition. Image A, every other line is collected; image B, the middle 32 columns is collected, and image C has every other line collected and the middle eight columns

Neural networks A neural network can be thought of as a black box that takes in one or multiple inputs and then processes them into one or multiple outputs. The black box contains layers that consist of neurons. The goal of the neural network is to find the appropriate weighted connections from one layer to another [4, 5, 6]. One example of a neural network would be having the inputs of a neural network be aspects of a house like number of bedrooms and square footage, and the output be the price of the house based on those inputs. A diagram of a neural network structure can be seen in Fig. 2.

Figure 2. A neural network with three nodes as the input layer, four nodes as the hidden layer, and one node as the output layer. Each node passes along information to the nest layer that have corresponding weights associated with them.

Methods We used three different data acquisitions, alias, blur, and alias + blur. The purpose of these different data acquisitions is to simulate how the image would look like if we only processed specific data from the Fourier transform of a 64×64 image. We used the data acquisitions for 2× and 4× accelerated data, which means that for two times accelerated data we acquired 32 columns,

Torres + Wong

The Manhattan Scientist, Series B, Volume 6 (2019)

235

Figure 3. The masks we used for four times accelerated data acquisition. Image A is collecting every other two columns of data, image B is collecting the middle 16 columns of data, and image C is collecting every other 2 columns and the middle 3 columns. The areas that are black represent data not being collected, while the area that is white represents the data being collected.

while for four times accelerated we acquired 16 columns of the Fourier transform. Fig. 3 shows images on how the three different data acquisition masks looked like for 4Ă&#x2014; accelerated data. Our data set consisted of 1000 64Ă&#x2014;64 sub images from an MRI volume. The order of the images was randomized; 800 images were used for training and 200 images were used for testing. We then took the inverse Fourier transform of the MRI data after the specific mask was applied and ran it through our convolutional neural network. A convolutional neural network is a type of neural network used in image recognition and processing. In addition, CNNs consist of convolutional layers that learn specific aspects of the image [4]. Fig. 4 gives the structure and number of parameters of the convolutional neural network that was used in our work. For the purposes of our research we decided to keep our network simple.

Figure 4. The table shows the progression of each layer in the convolutional neural network, the shape of the resulting output after each layer, and the number of parameters for each layer. The diagram shows the order of each layer in the convolutional neural network we used. It starts with the input layer which takes in the MRI images, then the convolutional layer which extracts features from our images, a batch normalization layer normalizes specified quantities of data to speed up the learning process and then repeats until it reaches the final layer, a convolutional layer, a batch normalization layer, and a convolutional layer for the final layer.

236

The Manhattan Scientist, Series B, Volume 6 (2019)

Torres + Wong

Results 2× Data Acceleration Fig. 5 shows the results for the 2× aliased acquisition. The image reconstructed using the neural network is a blurred version of the aliased image. The 2× blurred acquisition results are shown in Fig. 6. The reconstructed image removes some of the blurring. In the aliased + blur image (Fig. 7), the reconstructed image removes some of the blurring and aliasing.

Figure 5. Original image, image from the 2× aliased acquisition and the reconstructed image using the neural network. The RMSE between original and aliased image was 0.101 and the RMSE between original and reconstructed image was 0.102. The reconstructed image blurred the aliased image but had similar RMSE for this specific image.

Figure 6. Original image, image from the 2× blurred acquisition and the reconstructed image using the neural network. The RMSE between original and blurred image was 0.025 and the RMSE between original and reconstructed image was 0.029. The reconstructed image removed some of the blur from the blurred image but had a worse RMSE for this specific image.

4× Data Acceleration The results for a four times (4×) accelerated acquisition magnify the challenges of acquiring less data as seen in Figs. 8, 9, and 10.

Torres + Wong

The Manhattan Scientist, Series B, Volume 6 (2019)

237

Figure 7. Original image, image from the 2Ă&#x2014; aliased and blurred acquisition and the reconstructed image using the neural network. The RMSE between original and alias + blur image was 0.041 and the RMSE between original and reconstructed image was 0.037. The reconstructed image removed some of the blur and aliasing from the blurred and aliased image and had a better RMSE for this specific image.

Figure 8. Original image, image from the 4Ă&#x2014; aliased acquisition, and the reconstructed image using the neural network. The RMSE between original and aliased image was 0.112 and the RMSE between original and reconstructed image was 0.134. The reconstructed image blurred the aliased image and had a larger RMSE for this specific image.

Figure 9. Original image, image from the 4Ă&#x2014; blurred acquisition, and the reconstructed image using the neural network. The RMSE between original and blurred image was 0.040 and the RMSE between original and reconstructed image was 0.045. The reconstructed image removed some of the blur from the blurred image but had a worse RMSE for this specific image.

238

The Manhattan Scientist, Series B, Volume 6 (2019)

Torres + Wong

Figure 10. Original image, image from the 4× aliased, and blurred acquisition and the reconstructed image using the neural network. The RMSE between original and alias + blur image was 0.066 and the RMSE between original and reconstructed image was 0.063. The reconstructed image removed some of the blur and aliasing from the blurred and aliased image and had a slightly better RMSE for this specific image. There were significantly more artifacts in this image than in the 2× acceleration.

Average mean squared error To better assess the performance of our methods, we computed the average mean squared error (MSE) using our testing set for each of the acquistions (Table 1). We see that the 2× error is consistently less than the 4× error. The acquisition with the smallest MSE was the 2× blur and the one with the largest MSE was the 4× alias. Table 1. The average mean squared error results for the three different data acquisitions for 2× and 4× accelerated data. 2× Alias Blur Alias + Blur

4× −3

8.5 × 10 5.0 × 10−4 1.7 × 10−3

1.0 × 10−2 1.2 × 10−3 4.6 × 10−3

Training performance of CNN Fig. 11 shows the root mean square error (RMSE) for the training of the CNN for the 2× alias and blur acquisition as a funtion of the number of passes through the data (epochs). This plot is representative of the training of the neural network for other acquisitions. We used 50 epochs for all of our results.

Discussion Based on Figs. 5 and 8, we can see that our reconstructed image does not significantly reduce the aliasing from a purely aliased acquisition. Based on the images from Figs. 6 and 9, our CNN reduces some blurring from the purely blurred acquisitions. Figs. 7 and 10 show that our CNN reduces aliasing and blur from the aliased and blurred acquisition. These images suggest that the

Torres + Wong

The Manhattan Scientist, Series B, Volume 6 (2019)

239

Figure 11. A graph of the training performance of the convolutional neural network with the number of epochs on the x-axis and the average root mean squared error on the y-axis. As the number of epochs increases, the average root mean squared error decreases and plateaus around 45-50 epochs.

neural network works better when low frequencies (large scale features) are included in the data set. Table 1 shows that for 2Ă&#x2014; and 4Ă&#x2014; data acceleration, the lowest mean square error was for the blurred images. However, MSE may not be the best metric for image comparison and for future work we would like to look into other metrics like structural similarity (SSIM) [7]. Ultimately, we would like to explore the performance of these methods on clinical tasks [8]. In addition, we would also like to test different data acquisitions like spokes, spirals, and more complex acquisitions [2].

Conclusion We found that the reconstructed image decreases blurring from the blurred image and aliasing from the alias + blur image. In addition, we found the reconstructed image is unable to remove aliasing from the purely aliased data. In this work we showed the importance of data acquisition in the performance of neural network reconstructions.

Acknowledgements This project was supported by the School of Science Summer Research Scholar Program. We thank our advisor Dr. Pineda for his mentoring and Brother Daniel Gardner for coordinating the summer scholars program activities.

References [1] Wong ML, Torres QQ, Optimizing Network Architecture for Deep Learning in Magnetic Resonance Imaging (MRI), The Manhattan Scientist, 2019, 6, 241-248. [2] Ortega E, Saenz RV, Importance of Sampling Pattern and Regularization in Under-Sampled Magnetic Resonance Imaging (MRI), Dimensions, 2015, 101-112.

240

The Manhattan Scientist, Series B, Volume 6 (2019)

Torres + Wong

[3] Gonzalez RC, Woods RE, Digital Image Processing 4ed, 2017, Pearson, Hoboken, NJ. [4] Hyun CM, Kim HP, Lee SM, Lee S, Seo JK, Deep Learning for Undersampled MRI Reconstruction, Physics in Medicine and Biology, 2018, 63 (13), 135007. [5] Ronneberger O, Fischer P, Brox T, U-Net: Convolutional Networks for Biomedical Image Segmentation, Medical Image Computing and Computer-Assisted Intervention (MICCAI), Springer, LNCS, 2015, 9351, 234â&#x20AC;&#x201C;241. [6] Knoll F, Hammernik K, Kobler E, Pock T, Recht MP, Sodickson DK, Assessment of the Generalization of Learned Image Reconstruction and the Potential for Transfer Learning, Magnetic Resonance in Medicine, 2018, 81, 116â&#x20AC;&#x201C;128. [7] Wang Z, Bovik AC, Sheikh HR, Simoncelli EP, Image Quality Assessment: From Error Visibility to Structural Similarity, IEEE Transactions on Image Processing , 2004, 13 (4), 600-612. [8] Miedema H, Classification of Magnetic Resonance Imaging (MRI) Data Using Small Sample Sizes, The Manhattan Scientist, 2017, 4, 233-240.

Optimizing network architecture for deep learning in magnetic resonance imaging Marcus L. Wongâ&#x2C6;&#x2014; and Quinn Q. Torresâ&#x2C6;&#x2014; Department of Mathematics, Manhattan College Abstract. Magnetic Resonance Imaging (MRI) is crucial in the medical field because it gives excellent visualization of soft tissue. A common application of MRI is to detect tumors in the brain. Because medical images from MRI are slow to generate due to data acquisition, there has been recent research delving into accelerated methods incorporating a deep learning approach. This research explores different architectures of a convolutional neural network (CNN) to reconstruct the highest possible image quality of a brain using half the data. Our results reveal that a simple CNN is able to decrease aliasing and blurring in MRI reconstruction using half the time of a conventional acquisition.

Introduction Magnetic Resonance Imaging is used on a daily basis to help doctors care for their patients. Ongoing research has been trying to discover ways to accelerate how medical images are generated and find deep learning algorithms to improve the reconstruction of the MRI images. This research mainly pertains to the optimal methods for scientists to accelerate data acquisition for MRI images and to reconstruct medical images with the minimal amount of data possible. Achieving these goals will not only help doctors obtain high quality images, but will also be beneficial for patients who experience claustrophobia when positioned inside the scanner. From a deep learning perspective, this research work presents a variety of convolutional neural network architectures and compares and contrasts the image reconstruction performance from different models. The work presented in this paper combines multiple concepts of deep learning and relates partially to a recent architecture known as the U-Net [1, 2].

Data acquisition and Fourier transform Image data acquisition is an important process to specify how much data to collect from an image. This is simulated using a method known as masking. Masking determines what data is collected from a fully sampled image [3]. There are various mask patterns to acquire image data such as column stripes or spokes. The larger structures of an image lie in the low frequencies [4]. Meanwhile, the high frequencies contain the smaller details. The Fourier transform allows up to compute and visualize the low and high frequencies of an image. When the specific mask is applied to the Fourier transform of the data and an image is then generated using the inverse Fourier transform. This process can blur or distort the original image by overlapping several details over each other. This occurs because only specific parts of the image information is being captured and observed. The masking from Fig. 1 would cause aliasing and blurring to an image [5]. â&#x2C6;&#x2014;

Research mentored by Angel Pineda, Ph.D.

242

The Manhattan Scientist, Series B, Volume 6 (2019)

Wong + Torres

Figure 1. From left to right: An original image of a part of the brain, Fourier transform of the image, aliasing and blurring mask, distorted image of the original image using only the data collected using the mask. Note that there are artifacts in the aliased and blurred image which are repeated parts of the original image. This type of artifact is called aliasing. The aliased and blurred image is also blurred when compared to the original image.

Deep learning architectures Neural network A neural network is a network that is inspired by how humans learn. In a neural network there are multiple hidden layers that contain nodes. Each layer is responsible for extracting valuable features when analyzing an image or piece of information [1]. In between each hidden layer are weights being mapped from one layer to the next which can be analogous to how our brain sends signals from a neuron to another neuron. This is the main body of a neural network that enables a machine to learn and extract information to carry out a specific task. Fig. 2 shows an example of a neural network.

Figure 2. A neural network with an input layer, followed by two hidden layers and ending with an output layer. An example in applying this neural network would be a binary classification problem deciding whether there is a cat or no cat in a picture. An image would be inserted in the input layer. Next the two hidden layers would be tasked in analyzing the information and features of the image. Lastly, the output layer produces a yes or no result for whether there is a cat in the image.

Wong + Torres

The Manhattan Scientist, Series B, Volume 6 (2019)

243

Convolutional neural network A version of a neural network applicable to image data is a convolutional neural network (CNN). The convolutional neural network has been a popular network in a variety of tasks such as image reconstruction and classification problems. It is more complex than a simple neural network because it has a number of different kinds of layers including layers where the output is obtained by convolving a kernel with the image. This network also has some special components such as a max pooling layer which is responsible for extracting the largest feature value in a subsection of an image [1]. In this research, we took the structure of a CNN and added more layers to evaluate its performance for an image reconstruction task. U-Net Architecture An extension of a CNN is known as the U-Net architecture. The first half of the U-Net is similar to a standard CNN which is known as the contracting path. However, what makes the UNet unique is that it contains an expansion path which extends a CNN architecture in an upward route [2]. The whole network is shaped similar to a parabola that is concave up, hence the name UNet. This architecture is much larger due to the expansion path. Also, the contracting path reduces the complexity of the image and is designed to capture a lower dimensional representation of the image. Our U-Net architecture is inspired by [2, 6] but it is a significantly simplified version.

Methods In our research, we created a total of six different architectures. We initially only planned to have a U-Net implementation for our image reconstruction task. However after reviewing the results from the U-Net implementation, we decided to establish a simpler approach. The alternative was to implement three versions of a same convolutional neural network. A same CNN is a convolutional neural network but all of its convolutional layers contain the same number of feature maps and preserves the dimension of the input image to the output image. All together, three of the architectures followed the general U-Net structure, while the other three were same convolutional neural networks. We first titled our three U-Net architectures U-Net A, U-Net B, and U-Net C where the architecture got progressively smaller in terms of its feature mapping and number of layers. The same convention was used for the same convolutional neural network implementation. We titled the three same CNNs as CNN A, CNN B, and CNN C where each CNN got smaller by removing two of the same convolutional layers respectively. An output summary of our â&#x20AC;&#x153;Same CNN Câ&#x20AC;? is displayed in Fig. 3 which presents the specifics such as the number of trainable parameters and details of each layer in the networks. The same CNN C was our simplest CNN. The number of parameters for all six architectures studied are shown in Table 1.

244

The Manhattan Scientist, Series B, Volume 6 (2019)

Wong + Torres

Figure 3. Summary of our same CNN C version which has a total of 6 layers. Its convolution blocks consist of 64 features for its feature mapping. Throughout each layer, the CNN analyzes 64 features and ends its last layer with an output shape of (64, 64, 1) to preserve the dimension of the 64 Ă&#x2014; 64 Ă&#x2014; 1 image. The diagram of our same convolutional neural network version C architecture. The batch normalization layer normalizes the output in between each convolutional block and contributes to regularization for minimizing the chances of over-fitting [7].

Table 1. Number of parameters for all six networks architectures studied. Parameters

U-Net A

U-Net B

U-Net C

CNN A

CNN B

CNN C

Trainable parameters

740,353

149,377

260,737

186,625

38,401

We trained our different U-Net and same convolutional neural networks with our training and testing data. Our data set consisted of 1000 subimages obtained from a 3D MRI volume of a brain. The subimages were divided into 800 images for the training set and 200 images for the testing set. All neural networks were trained using 50 epochs (passes through the training set) with batch size of 16.

Results Results of the reconstructed testing images from same CNN A, CNN B, and CNN C are displayed in Fig. 4. All three networks produced similar results. Results of the reconstructed testing images from U-Net Type A, U-Net Type B, and U-Net Type C are displayed in Fig. 5. All models ran on 50 epochs with a batch size of 16.

Wong + Torres

The Manhattan Scientist, Series B, Volume 6 (2019)

Figure 4. The first row of images display the original and the aliased and blurred image of a corner of a brain. The second row are the three reconstructed images from the three different CNNs. All three neural networks reproduce an image with some aliasing but the amount of aliasing is less than in the image produced without the neural network (aliased + blur image). Despite the aliasing, the three CNNs produce reasonable versions of the original image.

Figure 5. “U-Net Type A” reconstructs a blurred image that captures a corner of a brain. It undoes the aliasing compared to the U-Net B and C results but has more blurring. “U-Net Type B” reconstructs the image with both blurring and aliasing. The aliasing is noticeable especially in the top left corner of the image. “U-Net Type C” reconstructs a poor image where the blurring is severe. Many of the fine details are lost and has the most aliasing.

245

246

The Manhattan Scientist, Series B, Volume 6 (2019)

Wong + Torres

Fig. 6 is a graph of the training performance for our same CNN C. The x-axis is labeled as the number of epochs that go up to 50. Meanwhile, the y-axis is set to a log scale to measure the root mean squared error loss over the epoch interval.

Figure 6. The training performance graph for the same CNN C architecture. After its significant drop near epoch 5, the training performance oscillates up and down and then eventually plateaus by the 50th epoch.

Tables 2 and 3 display the average mean squared error for the 800 training and 200 testing images for the same CNN and U-Net architectures. Table 2. Average mean squared error table of the three same CNN architectures. All three CNNs have relatively low variance with similar testing MSE values than the training MSE values. Data type Training MSE Testing MSE

Same CNN A −3

1.50 × 10 1.47 × 10−3

Same CNN B

Same CNN C

−3

1.66 × 10−3 1.59 × 10−3

1.89 × 10 1.87 × 10−3

Table 3. Average mean square error table of the three U-Net Type architectures. From Fig. 5, U-Net Type C produced a poor image. This relates to the higher testing MSE value compared to the U-Net Type A and B versions. Data type Training MSE Testing MSE

U-Net type A

U-Net type B

U-Net type C

−3

3.07 × 10−3 2.90 × 10−3

1.57 × 10 1.59 × 10−3

1.41 × 10 1.40 × 10−3

Discussion

An appropriate expansion for this research would be to develop an improved version of our smallest U-Net type and same convolutional neural network. We did discover an extremely small

Wong + Torres

The Manhattan Scientist, Series B, Volume 6 (2019)

247

U-Net type C architecture that only had three convolutional blocks which reconstructed a low quality image with noticeable blurring and aliasing. We would begin with this network and slowly expand it slowly but keeping it as simple and minimal as possible. Then, we can apply a larger data set and evaluate how well this small network can perform for image reconstruction. It would also be interesting to explore the variability of the results on the training and testing images. Lastly, we noticed that applying the mean squared error as our loss function was not an accurate presentation for our results. Although CNN C and U-Net Type A had extremely similar testing MSE values, their reconstructed images looked different. The CNN C reconstruction had aliasing but U-Net Type A only had blurring. This metric is unable to tell the difference between aliasing and blurring. In the future, it would be interesting to apply structural similarity (SSIM) [8] as our measurement to see how it relates to the differences in image quality. The final step would be to see how the reconstructed images perform in clinical tasks [9].

Conclusion Our research reveals that a simple convolutional neural network has the ability to perform well in this image reconstruction task. More specifically, our same convolution neural network version C architecture worked well compared to the other architectures that had more layers. Our research also suggests that a U-Net architecture may not be necessary to perform this image reconstruction task because the overall results show that a same CNN has similar performance than a U-Net architecture. This research only covered two types of neural network architectures. The work can be expanded to incorporate other networks that have been applied to other tasks and evaluate which networks are suitable for optimizing the reconstruction and quality of MRI images.

Acknowledgements This work was supported by the School of Science Research Scholars Program. We would like to thank our research advisor Dr. Pineda for guiding our research and Brother Daniel Gardner for coordinating the activities of summer research program.

References [1] Hyun CM, Kim HP, Lee SM, Lee S, Seo JK, Deep Learning for Undersampled MRI Reconstruction, Physics in Medicine and Biology, 2018, 63 (13), 135007. [2] Ronneberger O, Fischer P, Brox T, U-Net: Convolutional Networks for Biomedical Image Segmentation, Medical Image Computing and Computer-Assisted Intervention (MICCAI), Springer, LNCS, 2015, 9351, 234â&#x20AC;&#x201C;241. [3] Ortega E, Saenz RV, Importance of Sampling Pattern and Regularization in Under-Sampled Magnetic Resonance Imaging (MRI), Dimensions, 2015, 101-112. [4] Gonzalez RC, Woods RE, Digital Image Processing 4th ed, 2017, Pearson, Hoboken, NJ. [5] Torres QQ, Wong ML, Optimizing Data Acquisition for Deep Learning in Magnetic Resonance Imaging (MRI), The Manhattan Scientist, 2019, 6, 233-240.

248

The Manhattan Scientist, Series B, Volume 6 (2019)

Wong + Torres

[6] Knoll F, Hammernik K, Kobler E, Pock T, Recht MP, Sodickson DK, Assessment of the Generalization of Learned Image Reconstruction and the Potential for Transfer Learning, Magnetic Resonance in Medicine, 2018, 81, 116â&#x20AC;&#x201C;128. [7] Ioffe S, Szegedy C, Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift, Proceedings of the 32nd International Conference on International Conference on Machine Learning, 2015, 37, 448-456. [8] Wang Z, Bovik AC, Sheikh HR, Simoncelli EP, Image Quality Assessment: From Error Visibility to Structural Similarity, IEEE Transactions on Image Processing , 2004, 13 (4), 600-612. [9] Miedema H, Classification of Magnetic Resonance Imaging (MRI) Data Using Small Sample Sizes, The Manhattan Scientist, 2017, 6, 235-240.

CP studies on VH produced Higgs boson decay processes and background estimates Sarah Reese∗ Department of Physics, Manhattan College Abstract. Continuing checks on Higgs properties can function as a check on the validity of the Standard Model. Precision measurements on the Higgs Boson have now become one of the key tasks of the LHC experiment. Physical phenomena such as deviations beyond the Standard Model can be studied via Effective Field Theories. CP violation is a promising candidate in the search for new physics. Associated (VH) production in hadron-hadron collider experiments are a useful means to study possible CP Violation in the Higgs sector. In this paper, CP observables are constructed around the four lepton final state. We see that angular distributions are strongly sensitive to CP violation at current experimental limits of Effective Field Theory parameters.

Introduction Researchers at the ATLAS and CMS experiments discovered a new particle with a mass of 125 GeV, believed to be the Higgs boson [1, 2]. While the discovery was a massive leap forward in terms of our understanding, there are still many properties of the Higgs which have yet to be understood. Studying these properties have the double effect of working as a check on the Standard Model’s validity and as a search for new physics. Standard Model (SM) deviations and many models of physics beyond the Standard Model (BSM) are described through the context of Effective Field Theories (EFTs). At energies below a scaling factor, Λ, higher order observables expand the SM Lagrangian from the 4 dimensional operator basis to higher order terms. Operators remain consistent with the SU (3)C × SU (2)L × (D) U (1)Y symmetry with dimension D > 4 and are parameterized via Wilson coefficients, ci . Below is the generalized EFT Lagrangian with operators up to 8 degrees of freedom, LEF T = LSM +

X 1 (5) (5) X 1 (6) (6) X 1 (7) (7) X 1 (8) (8) (ci )Oi + 2 (ci )Oi + 3 (ci )Oi + 4 (ci )Oi . (1) Λ Λ Λ Λ i i i i

It is clear how the Lagrangian above can easily be generalized. The only limit is one’s ability to create nth dimensional operators. Ultimately that is quite difficult due to the specific Lorentz and gauge symmetries of the Standard Model. For Higgs physics only dimension-6 are the most relevant, so we can drop dimension-5 and dimension-7 operators while using dimension-8 operators can act as a bound on dimension-6 Wilson coefficients. The violation of CP symmetry in the Higgs sector would be evidence of a major inconsistency in the SM. CP symmetry is the product of the charge conjugate and parity symmetries. It means ∗

Research mentored by Rostislav Konoplich, Ph.D.

250

The Manhattan Scientist, Series B, Volume 6 (2019)

Reese

that if you swap a particle for its antiparticle then create a mirror image of the system the laws of physics will remain the same. The effect of this would be the first evidence of new physics towards an explanation to the baryon asymmetry of the universe. Detection of the CP violation can be detected through kinematic observables [3]. For this reason, EFTs are advantageous because they allow for the inclusion of kinematic distributions in the model. The Higgs Basis is an EFT model convenient for leading order LHC analysis which spans dimension-6 space [4]. Operators are separated from those already known to be strongly constrained from previous experiments and are expressed in a mass eigenstate. This relates the directly to physical, LHC-observable quantities. The HZZ interaction is written as follows LD=6 HZZ =

H g 2 + g 02 [(1 + Î´czz )v 2 ZÂľ Z Âľ + czz ZÂľÎ˝ Z ÂľÎ˝ + cĚ&#x192;zz ZÂľÎ˝ ZĚ&#x192; ÂľÎ˝ ]. v 4

Where Î´czz , czz , cĚ&#x192;zz are the couplings corresponding to new physics, g and g 0 are the SU (2)L and U (1)Y couplings respectively, v is the vacuum expectation value, and the field strength tensors are defined as ZÂľÎ˝ = â&#x2C6;&#x201A; Âľ ZÎ˝ â&#x2C6;&#x2019; â&#x2C6;&#x201A; Î˝ ZÂľ and ZĚ&#x192;ÂľÎ˝ = 21 ÂľÎ˝Ď Ď&#x192; ZĎ Ď&#x192; . The case Î´czz = czz = cĚ&#x192;zz = 0.0 corresponds to the Standard Model. The presence of cĚ&#x192;zz introduces CP-violation.

Motivation For the time being,â&#x2C6;&#x161;searches for new physics and particles at the LHC are limited by center of mass collision energy, s = 13TeV. The HL-LHC is a project to increase the Luminosity of LHC from 146.9 inverse femtobarn to 3000 inverse femtobarn. This should increase the event rate for the qq â&#x2020;&#x2019; Z â&#x2020;&#x2019; ZH â&#x2020;&#x2019; e+ eâ&#x2C6;&#x2019; Âľ+ Âľâ&#x2C6;&#x2019; signal by k-factor 20 (Ď&#x192; â&#x2030;&#x2C6; 6.45 attobarn). The associated production process involves the same HZZ vertex and four-lepton final state as the intensively studied gluon-gluon fusion process, but their features are very different. In the case of the ggF process, there are two vector propagators and the angular correlations between the final state leptons are expected. In case of ZH process, the situation is much more implicit. Despite the fact that the muon pair is connected to the Higgs, some special angular correlations take place due to their sensitivity to the transverse momentum carried by the scalar particle. In addition the signal process with the Higgs boson, there are so-called background processes which have the same four-lepton final state, mimicking the signal.

The H â&#x2020;&#x2019; Âľ+ Âľâ&#x2C6;&#x2019; was chosen for its easy detectability as compared to the recently detected and much more probable H â&#x2020;&#x2019; bb decay mode. Since their signal to background ratios are comparable (â&#x2C6;ź 10â&#x2C6;&#x2019;4 ), H â&#x2020;&#x2019; Âľ+ Âľâ&#x2C6;&#x2019; is the better choice. The background was estimated as qq â&#x2020;&#x2019; Z(Îł)Z(Îł) â&#x2020;&#x2019; l+ lâ&#x2C6;&#x2019; l+ lâ&#x2C6;&#x2019; at the SM level with Ď&#x192; â&#x2030;&#x2C6; 10.1 femtobarn.

Studies Simulations of the CP violating effects were made using two samples generated with the Monte Carlo generator MadGraph5 [5]. A version of the Higgs Basis has not been made that

Reese

The Manhattan Scientist, Series B, Volume 6 (2019)

251

is compatible with MadGraph5. The Higgs Characterisation model has been widely validated and a set of mapping formulas exists to translate the HB Lagrangian to HC [4]. Previous work in the project was done validating these formulas and the Higgs Basis translation. For the HZZ interactions the translations are: 1 + δ cz Λ (g 2 + g 02 ) Λ g2 kSM = kAZZ = − c̃zz kAW W = − c̃zz . cosα v sinα v sinα The first sample was generated at the Standard Model level only. The second was done as by including terms equivalent to a 25% BSM contribution. The values of the Wilson coefficients were chosen such that σSM +BSM = 1.25σSM which is the current experimental limit on the cross section [6]. Samples were made with α = 45◦ and Λ = 1 TeV for 100,000 events. Table 1 displays the mapped coupling coefficients. Table 1. Mapped coupling coefficients Higgs basis coupling Sample 1

δcz = 0.0; c̃zz = 0.0

Sample 2

δcz = 0.0; c̃zz = 1.3

Higgs characterisation couplings √ kSM = 2; kAZZ = 0.000; kAW W = 0.000 √ kSM = 2; kAZZ = −4.097; kAW W = −3.139

A version of the Feynman diagram used for the signal sample is displayed Fig. 1. As explained the new physics corrections lie in the HZZ vertex.

Figure 1. Example Feynman diagram of signal process

The Higgs decay and production topology, as taken in the rest frame of the virtual Z is shown in Fig. 2. Of the angles shown, Φ, Φ1 , the cosine of θ∗ are particularly sensitive to CP effects. Additionally, mass of the virtual Z boson, MZ1 , is responsive to CP violation. This follows naturally from the diagram since all of those variables are derived from the new physics vertex. The background was assumed as Standard Model only. This is consistent with other recent Higgs studies [7]. It was generated with the same parameters as sample 1 for 100,000 events. Fig. 3 shows an example of one the Feynman diagrams used for this generation.

252

The Manhattan Scientist, Series B, Volume 6 (2019)

μ+

Reese

θ1 Z1

θ*

μq-

Φ1

θ2 Z2 Φ

Figure 2. Higgs production and decay topology

Figure 3. Example Feynman diagram of Background process

Twin cuts were made on the signal and background samples. Cuts were done inside of MadGraph5 in order to preserve statistics. Cuts on four lepton, e+ e− , and µ+ µ− masses were done in order to further increase the signal to background ratio. Cuts on the transverse momentum, pseudorapidity, and ∆R distance of the leptons was also applied to simulate detector effects. Table 2 displays the values of all cuts. Table 2. mll > 50.0 GeV

∆Rll > 0.05

m4l > 150.0 GeV

ηl < 2.5

pTl > 5.0 GeV

Figs. 4-6 display the angular distributions for pp̄ → Z → x0Z → 4` , 13 TeV. Fig. 7 shows the virtual Z boson mass.

Reese

The Manhattan Scientist, Series B, Volume 6 (2019)

Figure 4. Distributions of the angle Φ for signals and the background simulations

253

Figure 5. Distributions of the angle Φ1 for signals and the background simulations

Figure 6. Distributions of cos θ∗ for signals and the background simulations

Figure 7. MZ1 distributions for signals and background simulatons

254

The Manhattan Scientist, Series B, Volume 6 (2019)

Reese

Conclusions This paper demonstrates a efficient method to search for CP violation in the Higgs sector at HL-LHC. For samples made within the current experimental limit on cross section, the variables Φ, Φ1 , cos θ∗ , and MZ1 from the Background, SM, and SM+BSM levels are distinguishable. We can conclude that the SM only background hypothesis was sufficient. Associated Higgs boson production proves to have viable angular correlations. It should also be noted that our experimental cuts greatly affect the signal distributions as well as the background reduction. Future work on this project will be done on generalizing the signal process to qq → Z → ZH → l+ l− l+ l− and adjusting cuts for more efficient suppression of the background processes. We hope to find new observables in the interference variables created by adding more decay modes.

Acknowledgements This project was supported by the School of Science Research Scholars Program. The work of Rostislav Konoplich and Sarah Reese was also partially supported by the US National Science Foundation under Grant No.PHY-1402964 and a fellowship from the Michael J. ’58 and Aimee Rusinko Kakos endowed chair in science.

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.Lett. B716 (2012) 1-29. doi: 10.1016/j.physletb.2012.08.020. [2] The CMS Collaboration, “Observation of a new boson at a mass of 125 GeV with the CMS experiment at the LHC.” Phys. Lett. B 716 (2012) 30. doi: 10.1016/j.physletb.2012.08.021. [3] J. Brehmer, F. Kling, T. Plehn, and T. M. P. Tait, “Better Higgs-CP Tests Through Information Geometry.” Phys. Rev. D 97, 095017 (2018). doi: 10.1103/PhysRevD.97.095017. [4] A. Falkowski, “Higgs Basis: Proposal for an EFT basis choice for LHC HXSWG.” (2017). https://cds.cern.ch/record/2001958/files/LHCHXSWG-INT-2015-001 2.pdf [5] R. Frederix, S. Frixione, V. Hirschi, D. Pagani, H.-S. Shao, and M. Zaro, “The automation of next-to-leading order electroweak calculations.” doi: 10.1007/JHEP07(2018)185. [6] G. Durieux, C. Grojean, J. Gu, and K. Wang. “The leptonic future of the Higgs.” J. High Energ. Phys. (2017) 2017:14. doi: 10.1007/JHEP09(2017)014 [7] The CMS Collaboration, “Measurements of the Higgs boson width and anomalous HVV couplings from on-shell and off-shell production in the four-lepton final state.” Phys. Rev. D 99, 112003 (2019). doi: 10.1103/PhysRevD.99.112003.

Studies of Higgs boson properties and search for new physics with ATLAS Joseph Thomas∗ Department of Physics, Manhattan College Abstract. The Standard Model is a brilliant model for the interactions of particles with the Electroweak force, Strong force, and Electromagnetism. It is not complete however and there are several observed physical events it cannot fully explain. To explain these observations, we need new physics. Current particle physicists are limited by the energy level of the Large Hadron Collider (13 TeV). Due to the steep price of building the LHC it is not feasible to just keep building larger and larger accelerators in the hopes of discovering new physics. One solution is to use an Effective Field Theory to model higher energy physical interactions at energy levels that can be achieved with the LHC. The goal of my research was to use the Standard Model Effective Field Theory to observe the effect an additional term added to the Standard Model has on the Higgs boson cross section. How will this new term put limits on parameters of new physics? Is it possible to simplify the model further by disregarding the higher order terms? As we would come to learn, the problem with trying to oversimplify a model is that error starts to creep in and becomes appreciable. The cross-sectional limits could be found and compared using a derived equation and a simulated plot. These results are now being used for analysis of experimental data at the LHC.

Introduction The Standard Model (SM) is the unified theory of matter comprised of quarks, leptons, and bosons that explains three of the fundamental forces. The SM can accurately predict almost every interaction in particle physics even at high energies. Yet we know it’s not complete because it fails to explain dark matter, matter/antimatter asymmetry, cosmic inflation, and neutrino masses. To explain these discrepancies, we must come up with new physics, which we believe may be hiding at higher energies than the LHC can produce. SM Effective Field Theory (EFT) allows us to explore all possible effects of heavy new physics on low energy observables. An EFT allows us to make an approximation for a physical event in order to simplify the model. EFT limits the degrees of freedom to only the relevant ones at the given energy level. This simplifies the calculations and hopefully scales up to larger, more complete models. The objectives of my research were: (1) to observe the effects of an additional Beyond Standard Model (BSM) term(s) on the Higgs boson cross section; (2) to understand the relationship that all the terms have vs just the quadratic terms on our cross-sectional calculations (see equations under Methods below); and (3) to find the limits and boundary conditions for the CP odd and the CP even cases. CP here stands for charge-parity symmetry, which is the combination of charge conjugation and parity. Put plainly if you take a particle and its antiparticle and take its mirror image then flip it and swap the charge, physics should act the same on both particles. That is charge-parity symmetry. The process with the Higgs boson could be separated into 3 different ∗

Research mentored by Rostislav Konoplich, Ph.D.

256

The Manhattan Scientist, Series B, Volume 6 (2019)

Thomas

cases, CP odd, CP even, and the mixed case. In general, if a system is identical to the original system after a parity transformation, the system is said to have even parity. If the end result is the negative of the original, its parity is odd.

Methods The search for new physics begins by adding an additional term to the SM. That additional term is represented here as c (see below Eq. 1). C is comprised of 3 terms cHW, cHB, and cHWB for the CP Even case and tcHW, tcHB, and tcHWB for the CP Odd case. These terms are known as Wilson coefficients. Wilson coefficients are coupling constants that affect the local (electroweak) interaction operator in the Lagrangian for the SM. The “t” in front is used to differentiate between the CP Odd and CP Even cases. σ here represents the Higgs boson cross section and is proportional to the integral of the amplitude squared (Eqs. 2 and 3). dF is equal to the product from i = 1 to 4 while d3 li is the phase space. There are 4 leptons produced in the final state (as depicted in the diagram of Fig. 1) which makes this a 12th dimensional integral (4×3). For my research I focused on Vector Boson Fusion (VBF) as the method for Higgs boson production.

Figure 1. A Feynman diagram showing VBF production and decay of a Higgs boson. This picture provides an overview of the entire process. Here two W bosons collide together to produce the Higgs which subsequently decays into two Z bosons of which each then decays into an electron/positron pair.

We could calculate sigma by performing an expansion in Eq. 4. However, in order to calculate sigma, we first needed to find out the value of the coefficients in front of all 21 terms. To do this we ran a simulation using Madgraph and used the model: SMEFTsim A U35 MwScheme UFO v2 1. This simulation would take in our three additional coupling values (tcHW, tcHB, and tcHWB), calculate the 12-dimensional integral mentioned above and provide us with a value for the cross section: σ that aligns with the SM and experimental evidence. If we repeat this process 20 more times, we are then able to use a 21×21 matrix to solve for all the coefficients. At the end of this process we have an equation (see below Eq. 9) that can be used to predict the Higgs boson cross section given any value for tcHW, tcHB, and tcHWB. With this equation, we can find the upper and lower limits for these coupling values as well as observe the effect of

Thomas

The Manhattan Scientist, Series B, Volume 6 (2019)

257

using just the quadratic terms vs all the terms. The procedure mentioned above only works for the case of CP odd case where the cubic and linear terms can be dropped. To get a similar equation for CP even case, we had to include two more coupling values for a total of five: cHW, cHB, cHWB, cHD, and cHBox (Eq. 10). A rough estimate for the expansion of a 5-term polynomial yielded an equation with over 1000 terms. The computational time would increase dramatically as would the range of the limits. Due to this we set cHD and cHBox equal to 0 and performed the expansion once again, although this time the cubic and linear terms could not be dropped. This led to an additional 13 terms being added to Eq. 4. After running 34 simulations and creating a 34×34 matrix we were able to create an equation to predict the Higgs boson cross section for the CP even case, so long as cHD and cHBox were fixed at 0. With equations for the CP odd and CP even cases, we were able to calculate the limits for each coupling value. To do this we created a Monte Carlo Program in python which would randomly select 3 values for cHW, cHB, and cHWB and plug them into our equations giving us a value for σ. We then compared this σ value to the SM value (which is determined when all coupling values = 0) to ensure a ratio between 1 and 1.25 for the CP odd case and between 0.75 and 1.25 for the CP even case. This was to ensure that our cross section matched with experimental data found by ATLAS. The program would iterate a number of times (∼10, 000) and compare the new coupling values to the old value to check whichever value was highest or lowest depending upon whether we were searching for an upper or lower limit. At the end the program would display the 3 highest or lowest values for cHW, cHB, and cHWB giving us limits on the Wilson coefficients obtained from a 25% limit on the cross section. To observe the effect that only the quadratic terms had vs all the terms, we would run simulations using our model and 3 different coupling values and compare the σ value it gave us with the value that our equations yielded. To check only the quadratic terms, we removed all the other terms from our equations and plugged in the same 3 coupling values that we ran our simulation with. We compared the final sigma values to determine whether it was within acceptable error. Equations f = (1 + c)

(1)

f 2 = (1 + c) (1 + c) Z σ ∼ |f |2 d3 li Z Z 2 σp ∼ (1 + c) dF = 1 + c + c2 dF = (1 + c2 ) CP Odd

(2)

σd ∼ (1 + c)2 2

σ = σp σd = (1 + c2 ) = 1 + c2 + c4

(3) (4) (5) (6)

258

The Manhattan Scientist, Series B, Volume 6 (2019)

σ ∼ (1 + c)2 (1 + c)2 ∼ 1 + c + c2 + c3 + c4 CP Even

Thomas

(7)

2 P σ =(qSM σSM + p1 tcHW 2 + p2 tcHB 2 + p3 tcHW B 2 + p12 tcHW tcHB + p13 tcHW tcHW B + p23 tcHBtcHW B) 2 2 2 2 ∗ (qSM dD SM + d1 tcHW + d2 tcHB + d3 tcHW B + d12 tcHW tcHB + d13 tcHW tcHW B + d23 tcHBtcHW B)

(8)

The fully expanded equation for the Higgs boson cross section in the CP odd case has a total of 21 terms with 21 different coefficients:

(9) An additional 13 terms are necessary for the CP even case:

(10)

Results We found that as we increased the number of couplings the range of the limit increased significantly. To find a limit that was consistent every time we fixed one of the coupling values at 0 and used the Monte Carlo to find the other two. This effect can be seen in Fig. 2. These graphs are showing the limits only in the CP odd case. The linear and cubic terms are discarded leaving only the quadratic and quartic terms. As a result, the upper and lower limits appear symmetric. The graphs in Fig. 3 compare the limits calculated when all terms are used vs. just the quadratic terms. Only when the coupling values were quite small and one coupling was fixed at 0 did we receive a good approximation. As the coupling values increased above a certain point it quickly became apparent that all the terms would be needed to be within acceptable error. When all three couplings were varied all terms became necessary no matter how small we made the coupling values.

Thomas

The Manhattan Scientist, Series B, Volume 6 (2019)

259

Figure 2. Cross section versus the Wilson coefficients. The contours correspond to the SM cross section, SM + 25% BSM, SM + 50% BSM, and so on. The limits found are displayed below the plots. The last plot demonstrates that increasing the number of couplings from 2 to 3 significantly increases the range. The x-axis corresponds to the tcHW Wilson coefficient, the y-axis to the tcHB coefficient, and the z-axis to the tcHWB coefficient.

Figure 3. These plots overlay the limits calculated when all the terms were used with the limits calculated using only the quadratic terms (CP Odd). There is an appreciable difference that gets larger as the range of the limits gets further away from the SM. The x-, y-, and z-axes all correspond to the same Wilson coefficients as in Fig. 2.

260

The Manhattan Scientist, Series B, Volume 6 (2019)

Thomas

The CP even case differs substantially compared to the CP odd case. The linear and cubic terms cannot be discarded and that has a significant effect on the upper and lower limits. They are no longer symmetric as they were for the CP odd case. cHD, cHBox, and a third coupling were all fixed at 0 in order to get reliable data for the upper and lower limits.

Figure 4. These plots were made in the same manner as Fig. 2 except that we used the equation for the CP even case. The linear and cubic terms could not be ignored in these calculations. As a result, the limits are no longer symmetric. The x-, y-, and z-axes all correspond to the same Wilson coefficients as in Fig. 2.

Both linear and quadratic terms were used to approximate the limit when 3 of the coupling values were fixed at 0. However, contrary to what we found in the CP odd case, there were no coupling values that yielded a sufficient approximation when using solely the linear and quadratic terms. All terms were needed to find the limits. If the linear and quadratic terms were enough to make a sufficient approximation then we could estimate the Higgs boson cross sectional limit when all 5 coupling values are varied by using just the linear and quadratic terms. Instead we only reinforced the notion that if we wanted to know how the cross-section changes as all 5 couplings change we would need to do the full expansion and run over 1000 simulations just to find all the terms.

Figure 5. These plots overlay the limits calculated when all the terms were used with the limits calculated using only the quadratic and linear terms (CP even). The difference is more pronounced here for the CP even case than it is for the CP odd case due to the additional linear/cubic terms. The x-, y-, and z-axes all correspond to the same Wilson coefficients as in Fig. 2.

Thomas

The Manhattan Scientist, Series B, Volume 6 (2019)

261

pp̄ → hjj → 4`jj CP even

Figure 6. Visual deviations from the Standard Model observed in the Φ distribution in the presence of new physics.

Conclusion

After running close to 200 simulations each taking 10 – 30 minutes and testing a wide range of coupling values we came to the following conclusions. We found that for small coupling values a good approximation could be calculated using just the quadratic terms. However, once the coupling values became greater than about 0.5, we needed all the terms in order to get an accurate value for σ. We also found that as we increased the number of couplings, we significantly increased the range of the limits. By setting some of the couplings equal to zero we can repeatedly find an upper and lower limit for all 6 of the Wilson coefficients (cHW, cHB, cHWB, tcHW, tcHB, tcHWB) that fit within experimental data. Currently these results are being used in analysis of experimental data at the Large Hadron Collider. Finally, we were able to confirm that, for the CP even case, the limits should no longer be symmetrical. Due to the limited time and computational power available to us, we were unable to generate an equation for the CP even case where all 5 of the couplings could be varied. The additional two couplings expand our equation from a manageable 34 terms to one with over 1000 terms. To fully understand the effect of all 5 terms on the Higgs boson cross section and its limits we would need to setup a computer to constantly run all 1000 simulations without pause. If each simulation finished in 20 minutes this process would take ∼14 continuous days to complete.

Acknowledgments

This work was supported by the Michael J. ’58 and Aimee Rusinko Kakos endowed chair in science. Additional support was received from the US National Science Foundation under Grant No.PHY-1402964

262

The Manhattan Scientist, Series B, Volume 6 (2019)

Thomas

References [1] A. Falkowski, M. González-Alonso, A. Greljo, and D. Marzocca, “Global Constraints on Anomalous Triple Gauge Couplings in the Effective Field Theory Approach.” Phys. Rev. Lett. 116, 011801 (2016) [2] M. Carena, C. Grojean, M. Kado, and V. Sharma, “Status of Higgs boson physics.” In Review of Particle Physics, Chin. Phys. C40 100001 (2016) p. 172. doi: 10.1088/16741137/40/10/100001 [3] S. Weinberg, “A Model of Leptons.” Phys. Rev. Lett. 19, 1264 (1967)

4513 Manhattan College Parkway, Riverdale, NY 10471 manhattan.edu