Ahead of this session at Pharma IQ Live: Transforming drug discovery through artificial intelligence, Jan Wenzel, a principal scientist in digital toxicology at Sanofi, tells us how artificial intelligence (AI) is helping to make drugs safer, earlier in the drug discovery phase.
Pharma IQ: What key things are driving the use of AI in computational toxicology?
Safety-related attrition is mostly related to, for example, pharmacokinetic or general toxic properties so it is important to control these in the early stages of drug development where the selection and optimization of molecules happens.
AI-driven technologies have the potential to supplement or even replace experimentation, at lower cost and higher throughput. At Sanofi we want to increase the proportion of non-animal methods to significantly reduce the necessity of vivo animal testing as much as possible. This is where the new predictive AI models come into play, because with these new AI or machine learning technologies, we can predict biological and chemical properties from compounds in every stage of drug development.
This could help in early-stage drug development to profile compounds and of course, later on to avoid safety issues. In terms of marketing products, all these AI methods could help to identify and classify mutagenic impurities, for example.
Pharma IQ: In which areas of Sanofi’s work has AI had the most impact?
JW: I can speak only for preclinical safety research here. We use AI very early on in drug development. One example amongst many others is where we identify a secondary unwanted protein target and want to minimize the interaction with molecules. In our case, AI supports all of our research projects for all different indications, it is important for all phases of drug development.
Pharma IQ: Are there any particular projects you have worked on using AI that you would like to highlight?
JW: At the end of 2019 we realized that multitasking neural networks or AI methods could make a huge improvement to the predictive performance of computer models. We also published ‘response maps’ as an easy method to visualize and interpret the impact of models to guide our teams.
In 2020, we published a method on predicting phototoxicity. Some types of drug molecules interact with UV light and patients taking this compound might get sunburn and skin eruption. With our computer model, we are able to identify how risky a compound is regarding photosensitization.
Pharma IQ: How accurate are these predictive models?
JW: Usually the predictive quality of this model is very high, and it really helps us to identify the phototoxicity of a compound in the early stages. Then we can easily react and provide guidance on which part of the molecule needs to be transformed to reduce the risk of phototoxicity.
Pharma IQ: What do you think the next breakthrough will be for AI in drug development?
JW: I think there are two areas that would be quite helpful if they were improved. One is a kind of integrated and automated AI system. Exemplarily, if you want to design compounds having certain interactions with a protein, then the computer model would be able to suggest a possible compound with the desired chemical and biological properties.
Interpretable AI is a very important topic. Usually if you work with a classic neural network model which is the basic technology behind artificial intelligence methods, you cannot identify directly which property of a molecule or which input parameters drive the specific endpoint you are predicting. For successful drug development it would be very helpful to know why a molecule is predicted to have a certain liability. This would provide direct AI guidance to design safe and efficient drugs.
Pharma IQ Live: Transforming drug discovery through artificial intelligence takes place on July 5-6. Catch Jan Wenzel’s session where he will be presenting a case study on the application of AI and quantum chemical methods in a pharma environment on July 6 at 11am EDT (2pm BST). Register for free here.