BioEmu

BioEmu is a diffusion-based generative model from Microsoft Research that predicts protein structural ensembles — the full landscape of conformational states a protein occupies — rather than a single static fold. It generates thousands of statistically independent conformations per hour on a single GPU and is trained on more than 200 milliseconds of molecular-dynamics simulation data alongside experimental structures and stability measurements. The model captures functional motions including cryptic-pocket formation, domain rearrangements, and local unfolding, and predicts relative free energies with roughly kcal/mol accuracy.

BioEmu shifts protein modeling from a static-structure picture to a dynamic, ensemble picture, addressing the long-standing limitation that real proteins flex and breathe in ways relevant to function and drug binding. By amortizing the cost of microsecond-scale molecular dynamics and ensemble-resolving experimental techniques, it enables researchers to design drugs that target specific conformational states inaccessible to single-structure methods. Within Microsoft’s AI for Science and Health Futures programs, BioEmu is a central piece of the push to make protein dynamics — not just structure — a computable, scalable input to therapeutic discovery.