Building physics into AI to discover new medicines

Despite their significant hype, machine learning (ML) models have struggled to solve one of the biggest challenges in modern biophysical chemistry: structure-based drug design. The task of designing a molecule that binds to a particular biomolecular target is extremely challenging for ML for two fundamental reasons: (1) the amount of experimental data is low and (2) the data type, 3D coordinates of molecules, is not native to most current ML models. Both of these challenges are problems of physics, and in this talk I will propose that the key to solving them is to build physics directly into our ML models. I will present our work building equivariant Euclidean Neural Networks (e3nns), ML models that directly incorporate the symmetries of 3D space, to solve two important problems for the chemistry of drug design. First, I will show how e3nns can learn the electronic structure of molecules with dramatically less training data than non-equivariant models. Second, I will highlight how coupling e3nns with generative ML methods can achieve state-of-the-art models for de novo molecular generation. Both of these applications highlight the critical role inductive bias plays in machine learning and how getting this right for molecular systems will be key to unlocking the future of AI-driven chemistry.

Speaker: Josh Rackers, Genentech