Chemicals Management Software Guide 29
the OECD would not be involved in that development. “This is work that would happen before the draft Defined Approach reaches the OECD.” He also says that he can see a role for AI in identifying promising test methods that can be developed into test guidelines.
Validation
“This use of the in vitro and in chemico assays and the artificial neural networkbased DA are the first use of such information in regulatory risk assessment,” it added. Skin sensitisation is the endpoint for which the adoption by the regulatory community of this kind of application of machine learning has progressed furthest. However, the EPA’s partner in the isothiazolinones project – the National Toxicology Program’s Interagency Center for the Evaluation of Alternative Toxicological Methods (Niceatm) – is working on the application of machine learning for other endpoints, primarily endocrine disruption and acute oral systemic toxicity.
Computational crowdsourcing Dr Kleinstreuer describes the Niceatm approach to this work as “computational crowdsourcing” via projects with collaborators around the world. “We have a great group of chem-informaticians and computational scientists in the Niceatm group,” she says. “But, you know, we recognise that we’re certainly not ‘the only show in town’. There are many different groups around the world that are really excellent in the area of applying machine learning to things like predicting chemical toxicity, and they have their own particular preferences in the different types of algorithms they use and the different types of molecular descriptors or chemical features they focus on.”
Niceatm does the ‘grunt work’ of collecting and curating data, which is then split into training and evaluation sets. The training set is shared with collaborators in academia, industry and regulatory agencies, who feed it into their models. Once those models have adapted to the training sets, they are sent to Niceatm, which uses the evaluation set to assess performance qualitatively and quantitatively.
Meanwhile back in the EU, the Commission’s JRC is investigating the use of AI for identifying chemical properties of concern, the substances with those properties and their modes of action. Applied to the scientific literature, the technology might, for example, identify chemicals with thyroid disrupting properties and extract the evidence for reference. “This has an immediate practical application in our work on the validation of in vitro methods,” says Clemens Wittwehr at the JRC. “But it could eventually also be used by researchers and decision makers for their own purposes,” he adds.
In the final stage, the project builds one or more ‘consensus’ models, incorporating all of the individual submissions.It is a system that compensates for the weaknesses of any one model, Dr Kleinstreuer says.
The organisation is also evaluating how suitable the validation principles of the European Centre for the Validation of Alternative Methods (Ecvam) are for AI technology.
Whether regulators in other countries will follow the lead of the US EPA and adopt this application of machine learning will depend on the extent to which they trust the algorithms.
Mr Wittwehr says that the JRC will aim to bridge the regulatory and scientific domains in relation to AI.
Bob Diderich, head of the OECD’s environmental health and safety division, which runs the organisation’s test guidelines programme, has said previously that AI presents a validation challenge. When AI is involved in computational methods, “there is no longer a transparent algorithm”, because it changes as it learns, he told an event in Helsinki last year. He says now that he can “see a role” for AI in the development of DAs, helping to combine several test methods to increase their predictive power. But he is clear that
“While AI might produce great knowledge, this knowledge must be augmented by a ‘credibility’ factor to turn it into actionable evidence.” That will only be achieved if AI helps to answer the typical questions regulators ask and the underlying algorithms can be assessed for properties that can improve their trustworthiness, such as transparency, fairness, explainability and robustness, he says.