Overview
| Name | BioEmu-2 · aliases: BioEmu-2, bioemu2, bioemu2-preview |
|---|---|
| Full name | A product name plus generation, with no expanded form in the repository |
| Ambiguous names (not auto-linked) | BioEmu |
| Family | This record covers the second generation only; the first handles single proteins and its results may not be carried over
|
| Organization | Microsoft Research AI for Science |
| Organization type | Big-tech lab |
| Country | United States, Germany |
| Model class | Structural ensembles (all-atom diffusion) |
| Domain | Structure & co-folding |
| First release | † partly verified · new weights Basis: Release month on the model card, with training in August and September that year Note: No primary source gives the exact day source |
| Other milestones | |
| Latest update | · code release Basis: Pre-release tag published Pre-release tag v0.1.0rc0 on 20 October 2026, with the model hub last modified the day before; repository activity through 16 November 2026 is automated dependency bumps. Note: The site previously recorded no releases or tags, which the releases page contradicts source The most recent sourced event, not a claim that nothing changed since. |
| Status | Active · Public preview and a pre-release tag within 12 months |
| Openness | Code: Open · MIT Weights: Open · MIT Weights and code under MIT, with the developer stating research use only and advising against commercial or real-world use without further testing |
| Benchmarks reported | 0 counted · none of the 44 |
| Wet-lab validation | None recorded |
| Unverified fields | † first_release, release_milestones.first_public_evidence, release_milestones.weights_release |
| As of | 2026-11-16 |
What it does
Samples equilibrium structural ensembles of biomolecules and can predict folded structures, extending the first generation from single proteins to multi-chain complexes, ligands and nucleic acids, taking structure files, sequences or multiple sequence alignments as input. The model card states it does not generate sequences or predict affinity, that sampling frequencies are not calibrated probabilities and that outputs are hypotheses rather than measurements. It shares the ensemble route with OpenComplex2 and differs from co-folding models such as AlphaFold3 in returning a distribution rather than one structure.
Benchmarks reported
None of the 44 benchmarks. This model is on the zero-pairing list.
How it evaluates itself: Model card evaluation areas
Other evaluations (outside the 44)
- Evaluation areas listed on the model card (Own task set) Self-reported: Lists folding, multi-conformation sampling, a dynamics-emulation holdout and large systems, with metric names but no reported values
Developers report: a foundation model for structures and ensembles of complexes, ligands and nucleic acids; the model card gives no quantitative results and states the limitations are not quantitatively characterised.
Independent evidence: No independent evidence found. It has no result among the 44 benchmarks here and no third-party evaluation was found
Training data
Group in Figure 7: Public data only.
| Class | Value | Detail |
|---|---|---|
| a. Public databases | Yes | About 189,700 systems from a public structure archive, plus monomer, homodimer and heterodimer entries from a public predicted-structure archive and one multimer set source |
| b. Private or licensed data | No | The model card states only public data and the developer's own data are used source |
| c. Own wet-lab data used for training | No | The in-house data are molecular dynamics trajectories, computationally generated rather than wet-lab source |
| d. Wet lab used only for validation | No |
Wet-lab validation
None. The developers report no wet-lab validation.
Background
- Organization: The developer is a large technology company's science-AI research division, formed in 2022 from existing teams across Europe, North America and Asia with two new European labs, funded from the company research budget.
- Funding: No separate funding; it is internal corporate research.
- Products and partnerships: The first generation was published in a general science journal in 2025, released under MIT and offered through the company's model platform, with a July 2026 customer story describing a pharmaceutical company using it there; whether the second generation is on the platform is not officially stated.
- Related models: OpenComplex2, AlphaFold 3, Boltz-2
Reading notes
- For now there is a model but no numbers: with no reported values it cannot be called better or worse than its predecessor or its peers.
- The new complex and nucleic-acid capability has no public evaluation, and the pose benchmarks here score single structures, which does not match its goal.
- The dynamics trajectories are produced in-house, so their coverage is the ceiling, and the model card itself says the training set leans towards well-studied protein families.
- Sampling frequencies are not calibrated probabilities, so they cannot be read as free energies and downstream use needs its own validation.
Caveats
- Neither model card nor repository reports any evaluation value.
- First release is month-level, with no primary day-level source.
- Weights are open but the developer limits them to research use.
- The in-house training data are computed trajectories, not wet-lab data.
- Results from the first generation may not be carried over to this one.
Sources
- Model card (repository)
- Code repository and releases page (repository)
- First-generation journal paper 2025 (journal)
Corrections
1 correction
| Field | Was | Now | Basis | Found |
|---|---|---|---|---|
| latest_update | 2026-11-16 (automated dependency-upgrade commit) | 2026-10-20 (pre-release tag v0.1.0rc0) | source | 2026-11-16 |
Part of the AI for Bio benchmarks survey · all 112 models · open in the explorer. Record as of 2026-11-16.