Towards a foundation model for protein biophysics

Deep learning and specifically generative models have had a profound impact on science and technology. For protein science - after the sequence and structure revolution - the key outstanding challenge is the lack of a scalable computational or experimental method to accurate determine the biomolecular basis of function - protein conformations, binding states and their equilibrium probabilities. Here we introduce deep learning methods to extract information about the equilibrium distribution of protein molecules from various data sources and to efficiently sample structures and compute equilibrium properties from it. We describe the advantages and limitations of such a method, discuss benchmarks and scaling laws with respect to simulation and experimental training data.