Predicting protein structure, mutations and dynamics with deep generative models

Proteins are biomolecules that play a key role in biology, biotechnology and medicine. They adopt specific yet dynamic three-dimensional (3D) shapes that are important for their functions. Successfully predicting these 3D shapes using a computer has been a major open problem in science for more than 50 years. Near the end of 2020, the problem was solved by a method based on artificial intelligence (AI) called AlphaFold, developed by the UK-based company DeepMind, a subsidiary of Google. However, in the course of 2021, the considerable limitations of AlphaFold became apparent. Notably, AlphaFold cannot predict the effect of disruptive mutations (so-called missense mutations) nor the dynamics of proteins. We think we can address these issues by developing better computational and mathematical models of protein structure (based on so-called deep generative models, an AI approach that excels at capturing noise and uncertainty) and better ways to train these models (based on a mathematical theory known as Stein’s method). At the heart of the project lies the following fundamental question: how can mathematics, statistics, physics and computer science be effectively combined to model a system as complex as a dynamic protein?