{"paper_id":"1d6676b5-23a3-4b49-b4de-006baece31fd","body_text":"Abstract\nMachine 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. ProteinEBM’s energies can be used to rank protein structure correctness, predict the energetic effects of mutations, sample from protein conformational landscapes, predict protein structures, and simulate protein folding pathways. Across all of these tasks, ProteinEBM shows performance competitive with or exceeding previous machine learning and physics-based methods, including state-of-the-art performance at predicting the effects of mutations on protein stability.\nCompeting Interest Statement\nThe authors have declared no competing interest.\nFootnotes\njamesron{at}mit.edu\nchenxiou{at}mit.edu\nFigured 3D corrected to show results from latest model.","source_license":"CC-BY-4.0","license_restricted":false}