Protein Diffusion Models as Statistical Potentials
Roney, J., Ou, C., Ovchinnikov, S.
bioRxiv·2025
Machine learning has driven rapid progress in protein structure prediction and design, but key challenges remain such as predicting protein structures when evolutionary information is unavailable, modeling full conformational landscapes, and capturing the thermodynamics of mutations and conformational changes. To address these problems we developed ProteinEBM, an Energy-Based Model of protein conformational space. ProteinEBMs energies can be used to rank protein structure correctness, predict protein structures, sample from protein conformational landscapes, and predict the energetic effects of mutations. Across all of these tasks, ProteinEBM shows performance competitive with or exceeding previous machine learning and physics-based methods. We see ProteinEBM as an important step in developing physically-grounded machine learning models for protein science.