Explainable AI for Molecules - AiChemist

Optimising biological activity and physico-chemical properties, while minimising their toxicity, are objectives when developing new compounds in chemical industries. Advanced machine learning (AI) methods are indispensable to this process. They are also increasingly used in environmental chemistry to identify compounds damaging to the environment and humans. Traditional machine learning (ML) methods provide reliable predictions though only for compounds similar to the training set, thus defining their applicability domain (AD). Emerging representation learning approaches can efficiently approximate the physical interactions of molecules with an accuracy comparable to physics-based methods in only fractions of time. Models based on these representations should have much larger AD due to pre-training on large chemical sets of theoretical values. Here we will develop and benchmark representation learning approaches, addressing their accuracy and ADs, using public and in-house data for endpoints ranging from chemical reactions to toxicity. While explainable AI (XAI) methods are actively developing in the ML community, there is a gap with their use in chemistry, i.e. there is a need to translate their results to the end users, chemists and regulatory bodies. Since the research program is tightly coupled with the target users - large companies, regulatory agencies and SMEs - it provides a clear path for technology transfer from academia to industry. AiChemist will provide structured training to its fellows through a combination of online courses and schools, strengthening European innovation capacity in the education of specialists in AI methods. The fellows will receive comprehensive training in transferable skills. The complementary expertise and strong commitment of the partners make this ambitious innovative research program realistic via the proper allocation of individual tasks and resources, as described below.