# Microsoft BioEmu2: Open Weights for Biomolecular Structure and Dynamics

BioEmu2 moves beyond predicting one molecular structure, with a public preview that explores how molecules change shape and interact.

By Paras Madan | 2026-10-20


Microsoft Research AI for Science released a public preview of BioEmu2 on October 20, 2026. The model predicts biomolecular structures, samples alternative configurations and propagates structures through time. Its code and weights are available under MIT terms, giving researchers a new tool for studying molecular motion rather than only static shapes.

[Frank Noé's announcement](https://x.com/FrankNoeBerlin/status/2112472173336240499) frames the problem clearly: knowing a molecule's structure does not fully explain how it works. Shape changes and interactions also matter. The interesting question is whether an AI model can make those behaviours easier to investigate without creating false confidence in its predictions.

## What BioEmu2 actually predicts

The [model card](https://github.com/microsoft/bioemu2/blob/main/MODEL_CARD.md) describes an approximately 500-million-parameter diffusion model that generates all-atom 3D structures. It supports multi-chain complexes, ligands and nucleic acids, with optional multiple sequence alignments and contact information to condition predictions.

An ensemble is a collection of possible configurations, rather than one supposedly definitive structure. That makes the output useful for asking how a system might vary, not just what a single representation looks like.

There are important boundaries. BioEmu2 takes supplied sequences; it does not generate new sequences or predict binding affinities. Microsoft describes the outputs as computational hypotheses requiring validation, not experimental measurements. It is not validated for patient decisions.

The model card also warns that sampling frequencies are not calibrated probabilities or quantitative free energies. More appearances in a set of generated structures do not automatically establish how often that configuration occurs in nature.

## Three workflows instead of one structure file

The [official repository](https://github.com/microsoft/bioemu2) exposes structure prediction, equilibrium sampling and time propagation through a command-line interface and Python API. It also supports conditioning on contacts and steering sampling toward desired properties.

That combination makes the release more interesting than another model that returns a single folded structure. A researcher can investigate different plausible states and ask what might happen after a particular starting configuration, while keeping the underlying system consistent.

The distinction between these modes should stay visible in any product built around them. A structure prediction, a collection of equilibrium samples and a sequence of time-ordered states answer different questions. Presenting all three as a smooth molecular animation would make the interface simpler but the scientific meaning less clear.

## How time propagation works

The [propagation documentation](https://github.com/microsoft/bioemu2/blob/main/docs/propagation.md) makes an important distinction: a propagation call samples an endpoint at a future time. Multiple returned samples are not necessarily consecutive frames in a trajectory.

For time-ordered output, the package provides a separate trajectory workflow that repeatedly propagates the system. The guide recommends time lags from 1 to 1,000 nanoseconds, the range the model mostly encountered during training. Starting structures must contain all required atoms; propagation cannot fill in missing ones.

This matters when interpreting demos. A convincing movie does not, by itself, tell you whether its frames represent a simulated sequence or independent samples. Developers should preserve the sampling method, time lag and validation information alongside exported structures.

## What researchers need to run it

The documented setup requires Linux and an NVIDIA GPU. macOS and Windows are not supported. Python 3.12, 3.13 and 3.14 are supported, with benchmark metrics evaluated on 3.12. Larger systems can use context-parallel sampling across at least four GPUs.

There is no sensible universal hardware promise here. Molecular system size, sample count and execution settings need to be considered together. Before reserving substantial compute, researchers should check the current installation guide and reproduce a small documented example on their intended hardware.

The repository lists several tested GPU families, while the model card describes a narrower tested configuration. That documentation difference is another reason not to treat a particular GPU as a guaranteed minimum for every workload.

## Why this release is worth watching

The useful change is the availability of a shared model and implementation that researchers can inspect and evaluate across structure and dynamics workflows. Open access makes comparison easier; it does not make the model's conclusions experimentally established.

For scientific AI products, the next layer of useful engineering is around the model: clear provenance, reproducible settings and honest displays of uncertainty. A fast prediction becomes valuable when a researcher can understand what it represents and decide what evidence should come next.

I’m Paras. I work at the intersection of AI, engineering and distribution.

I help technical founders turn complex research and product launches into clear stories that developers and founders want to read and share.

[Book a 30-minute call](https://cal.com/paras-madan/30min).