All-atom diffusion model that samples equilibrium ensembles and time-lagged dynamics for protein, nucleic acid and ligand complexes.
No providers recorded yet. Browse all providers
A protein that opens and closes around its substrate, or a peptide that docks into a groove and lets go again, is not described by any single structure. Molecular dynamics can recover those motions, but a millisecond-scale event can cost GPU-years of simulation, and static cofolding models return one pose per prediction. BioEmu2 (styled BioEmu-2 in its model card) is a generative model that sidesteps both limits: given the chemical composition of a biomolecular system, it draws statistically independent all-atom structures from an approximation of the equilibrium ensemble, and it can also sample where a given structure is likely to be after a chosen lag time.
BioEmu2 was developed by Microsoft Research AI for Science and released as a public preview on 20 October 2026, with the checkpoint named bioemu2-preview and the package at version 0.1.0rc0. It succeeds BioEmu-1, which emulated equilibrium ensembles of single protein chains. The successor widens the scope to multi-chain complexes, nucleic acids, ligands and other chemical components, and adds time-lagged dynamics through implicit transfer operator models, a capability its predecessor did not have.
No preprint or peer-reviewed paper accompanies the preview. The model card, the repository documentation and the release announcement are the primary sources, and all performance figures are reported by the developers.
EmuSession API.BioEmu2 is a diffusion-based all-atom score model of approximately 500 million parameters, consistent with the roughly 2 GB bioemu2-preview checkpoint. Inputs are biomolecular sequences with an optional multiple sequence alignment, retrieved from ColabFold or supplied as A3M files, plus optional contact and method conditioning. Outputs are all-atom Cartesian coordinates generated by reverse diffusion and written as XTC trajectories with an mmCIF topology. Inference runs on Linux with an NVIDIA GPU of compute capability 7.0 or newer; the README reports testing on V100, T4, A100, RTX A6000 and B200 GPUs, although the model card names an 80 GB A100 as the smallest device tested.
Training ran from August to September 2026 on public data and Microsoft-generated simulations: 189,745 experimental PDB structures in the Boltz-1 preprocessing, about 7.2 million AlphaFold Database monomers, 753,271 AFDB homodimer and 22,157 heterodimer multiconformation sets, 50,651 Teddymer structures derived from AlphaFold 2 predictions, and seven in-house molecular dynamics collections covering CATH domains, dimers, heterodimers and PDBbind 2020 protein-ligand systems. The developers evaluate structure prediction against cofolding models on the Boltz-1 test set using RMSD, TM-score, lDDT and DockQ. In the release announcement they report folding accuracy competitive with leading models on FoldBench and Runs N' Poses, benchmark free-energy surfaces predicted with about 1 kcal/mol error, and a 5 ms peptide-binding process sampled in about 40 minutes rather than the more than ten GPU-years molecular dynamics would need.
BioEmu2 targets structural biologists and computational chemists who need conformational heterogeneity rather than a single model: contrasting the states a protein adopts with and without a ligand or partner chain, and studying how a bound nucleic acid or second protein reshapes its ensemble. Propagation lets researchers ask how a structure evolves over nanoseconds to microseconds without running explicit dynamics, and contact conditioning brings experimentally derived distance restraints into sampling. The model card positions outputs as computational hypotheses for research use; it excludes quantitative binding-affinity and absolute free-energy prediction, warns that sampling frequencies are not calibrated Boltzmann probabilities, and notes weaker results on orphan proteins with shallow alignments and on complexes above 10,000 tokens.
BioEmu2 extends the emulator idea from single proteins to the heterogeneous complexes that drug discovery and mechanistic biology work with, and it pairs equilibrium sampling with a learned propagator in one open model. The code and weights are released under the MIT license with a detailed model card and a training-data inventory, so the community can test the developers' claims directly. Because the release is a preview with no accompanying paper, its benchmarks have not been independently reproduced, and the checkpoint may change before a versioned release.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.