Animal testing has traditionally been used to assess the safety of new chemicals before exposure to humans. More sustainable approaches to safety testing are desirable, not only due to the huge ethical concerns, but also because of the significant costs and time scales associated with in vivo methods that limit scalability.
Machine learning models such as deep neural networks represent an attractive alternative to animal testing. The suitability of these models for safety critical applications, however, is highly controversial due to their susceptibility to adversarial attacks, which can make their predictions unreliable. Consequently, there has been a significant interest in the deep learning community in analysing the stability of these networks and developing techniques for achieving robustness. At a more fundamental level, however, the current machine learning models that are used to predict the toxicity of chemicals remain black box approaches that do not offer any insights into their predictions, and their widespread adoption in chemical safety assessment remains challenging in the absence of such transparency.
For such models to be accepted by the diverse range of stakeholders in chemical risk assessment (chemical industry, regulators and toxicologists), they must provide clear insight into how and why predictions are made from a biological and chemical perspective. The overall assessment must be broken down into a series of intermediate results and auxiliary predictions that are open for expert scrutiny and independent verification. Such information supports accountable and responsible decision making in chemical safety assessment which has a direct impact on human health.
This project will combine the speed and flexibility of machine learning with chemical properties derived from molecular simulations in a synergistic, transparent machine learning approach to accountable and responsible chemical safety testing. Molecular simulations can be used to observe the interactions that occur between the body and toxic compounds, understand why compounds are toxic and quantify properties that lead to chemical toxicity, making such an approach highly interpretable. Thus, predictions made using these new machine learning models will be supported by clear insight into the biological and chemical origins of toxicity that can be scrutinised by experts at multiple levels. Auxiliary quantities that are not directly required in the assessment of toxicity, but help experts assess the validity and consistency of the prediction, will also be computed. Our focus will be on building transparent models for toxicological outcomes of particular importance including cancer, respiratory and skin sensitisation and aquatic toxicity.
This project is associated with the UKRI Centre for Doctoral Training (CDT) in Accountable, Responsible and Transparent AI (ART-AI). Further details of the ART-AI CDT can be found at: https://cdt-art-ai.ac.uk.
Applicants should hold, or expect to receive, a first or upper-second class honours degree in a related subject. Applicants should have taken a mathematics unit or a quantitative methods course at university or have at least grade B in A level maths or international equivalent.
Enquiries about the application process (detailed at https://cdt-art-ai.ac.uk/apply-now/) should be sent to [Email Address Removed].
Formal applications should be made via the University of Bath’s online application form: https://samis.bath.ac.uk/urd/sits.urd/run/siw_ipp_lgn.login?process=siw_ipp_app&code1=RDUCM-FP02&code2=0003
Start date: 4 October 2021.
Based on your current searches we recommend the following search filters.
Based on your current search criteria we thought you might be interested in these.