Diffusion Generative Models for Protein Structure Prediction and Beyond

Deep learning has fueled a recent revolution in structural biology, with AlphaFold recognized as a solution to the longstanding problem of protein structure prediction. However, related problems of interest, such as protein-ligand docking and predicting conformational ensembles, remain difficult. We explore and demonstrate the ability of diffusion models — a recent and influential paradigm in generative AI — to address such problems. First, we develop EigenFold, a structure prediction method that samples multiple structures for a given sequence and represents a step towards understanding conformational ensembles. Second, we present DiffDock, a diffusion model which docks a small molecule to a protein at significantly higher success rates than existing physics-based methods.