Microsoft open-sources BioEmu-1 to map moving proteins

BioEmu-1 is Microsoft's open-source deep learning model designed to simulate the structural ensembles of moving proteins. While breakthroughs like AlphaFold provide accurate static snapshots—much like a single frame of a film—BioEmu-1 generates the structural variations a protein adopts over time, bringing scientists closer to watching the full movie.

What it does

The tool's biggest draw is throughput: running on a single GPU, BioEmu-1 can sample thousands of physical protein structures per hour.

In drug discovery and molecular biology, researchers normally turn to molecular dynamics (MD) simulations to track how molecules flex and bind. However, classic MD often demands months of cluster time or custom computer architectures to simulate meaningful transitions.

BioEmu-1 slashes those compute demands. In benchmark comparisons against D. E. Shaw Research simulations of Protein G, it accurately matched the equilibrium structural distribution while requiring 10,000x to 100,000x fewer GPU hours.

It also handles unseen targets remarkably well. When tested on LapD, a protein from the cholera bacterium Vibrio cholerae, the model correctly predicted both the bound and unbound conformations with c-di-GMP molecules, despite never seeing them in training. BioEmu-1 even generated intermediate states that have never been experimentally observed.

Beyond visual conformations, counting its folded versus unfolded samples yields folding free energies that align closely with wet-lab experimental benchmarks.

How it works and availability

Microsoft achieved this balance by training BioEmu-1 across three distinct data layers.

First, it used clustered AlphaFold Database (AFDB) structures to outline distinct structural islands. Next, extensive MD simulation trajectories taught it to map physically realistic movements around those islands. Finally, fine-tuning on experimental folding stability datasets calibrated the sampling probabilities between folded and unfolded states.

Martin Steinegger, a professor at Seoul National University, noted that protein dynamics is the next frontier following static prediction, calling BioEmu-1 a significant milestone for rapid generative sampling of free-energy landscapes.

The project is now open-source, following a preprint released last December, and is also featured in Azure AI Foundry Labs.

Why it matters

Biotech teams and structural biologists should evaluate BioEmu-1 right away. Most therapeutics work by stabilizing or blocking dynamic intermediate shapes rather than static ground states, and turning months of simulation into an afternoon run is a massive unlock.

That said, we see this as an early step. While orders-of-magnitude faster, generative samplers still carry the risk of plausible-looking hallucinations on deeply novel targets. The smartest play today is treating BioEmu-1 as an agile hypothesis generator to guide downstream wet-lab assays, rather than a full replacement for rigorous physics-based solvers.