BioEmu2

BioEmu2 webserver overview

A crystal structure of adenylate kinase shows one conformation, but the enzyme spends its time opening and closing two lid domains around ATP. BioEmu2 is built for that gap. Given a molecular system, it draws many independent all-atom structures from the ensemble the model learned from PDB entries, AlphaFold Database models and Microsoft's molecular dynamics data. Sixteen samples of the 214-residue adenylate kinase take about three minutes on ProteinIQ.

BioEmu2 0.1.0rc0 from Microsoft Research AI for Science runs here with the bioemu2-preview checkpoint. It handles proteins, RNA, DNA, ligands, modified residues and covalent bonds in one system.

The webserver offers three tasks: equilibrium sampling, structure prediction and propagation of starting structures by a time lag. Each run returns BioEmu2's own XTC frames, mmCIF topology, per-frame physicality checks and the resolved system file.

BioEmu, the first version, samples single protein monomers only. Use BioEmu2 for complexes, ligands, nucleic acids, structure prediction or anything time-dependent.

How BioEmu2 works

BioEmu2 is a diffusion model with about 500 million parameters. A trunk reads the system as tokens (one per standard residue, one per heavy atom of a ligand or modified residue), together with the alignment and any contacts, and computes embeddings once per job. A denoiser then starts each sample from random atom coordinates and removes noise over 200 steps until an all-atom structure remains.

The same weights serve every task. Equilibrium sampling tells the model to produce structures like its molecular dynamics training data, while structure prediction asks for static structures like those in the PDB and AlphaFold Database and uses a step scale of 1.5 to sharpen toward one fold.

Propagation adds a starting structure and a lag time as conditions, so the denoiser draws where that structure is likely to be after the lag. Steering samplers apply potentials during denoising that penalize clashes and broken bonds or favour a property you define.

Pricing

BioEmu2 is billed by measured runtime at 20 credits per minute, rounded up to a whole credit for each job. Submission reserves at least 20 credits and returns whatever the finished run does not use. The quote shown before submission is an estimate: about one minute to start, load the model and embed the system, plus 9.4 seconds per frame, or 21 seconds per frame with a steering sampler, at 200 denoising steps. A ColabFold alignment adds an estimated minute; for Trp-cage the whole job took 31 seconds of BioEmu2 time.

The estimate does not yet account for system size, so a large complex runs longer than its quote and a small peptide shorter; the charge follows the actual runtime. On ProteinIQ, 16 equilibrium samples of adenylate kinase ran for 186 seconds, and 4 steered structures of myoglobin with heme took 135 seconds. Alignment depth has little effect: adenylate kinase with 16,866 aligned sequences and with its query alone took about the same time.

Inputs

Build the system from molecule cards, as for Boltz-2: one card per protein, DNA, RNA, SMILES ligand or CCD ligand, each with a copy count. Chains are lettered in card order with copies expanded, and the label on each card shows its chain IDs. Contacts and covalent bonds refer to those IDs. For propagation, the cards are optional when the starting structure's mmCIF already describes the system.

Alignments come from the public ColabFold server by default, through BioEmu2's own retrieval: one alignment per unique protein sequence, plus paired alignments for complexes of different proteins. To use your own instead, attach an A3M to every protein card; BioEmu2 takes all alignments from ColabFold or all from files, so attach one to every protein or to none. Turn off Generate MSA to keep sequences on ProteinIQ; proteins without an A3M then run in single-sequence mode, which the job reports as a warning. The cards become a BioEmu2 system YAML, kept as records/system-from-cards.yaml.

Settings

Task and system

Sampling

Precision (bf16-mixed), matmul precision (high) and batch size (all samples together) stay at BioEmu2's defaults. Multi-GPU context-parallel sampling for systems larger than one GPU is not available.

Steering potentials

The three steering samplers start from BioEmu2's ten default physicality potentials, which penalize bond length and angle errors, van der Waals overlaps, chain clashes, ligand distance violations, wrong chirality and non-planar ligand bonds. An uploaded potentials file replaces that set. Each entry names a class with _target_ and passes its constructor arguments:

potentials:
  - _target_: bioemu2.denoising.steering.potentials.BondLengthPotential
    sigma_tolerance: 10.0
    parameters: {guidance_weight: 0.05, resampling_weight: 0.1}

Targets must be BioEmu2's potential classes (PoseBustersPotential, InterChainVDWOverlapPotential, IntraChainVDWOverlapPotential, PolymerLocalClashPotential, SymmetricChainCOMPotential, StereoBondPotential, ChiralAtomPotential, PlanarBondPotential, ContactPotential, TemplateReferencePotential, BondLengthPotential, BondAnglePotential) or its schedules (ExponentialInterpolation, PiecewiseStepFunction, LinearInterpolation). Other Hydra controls and ${...} interpolation are not accepted.

Outputs

Structures loads the topology with every frame for playback. Files lists each download.

Load samples.xtc with topology.cif in MDTraj, VMD, PyMOL or ChimeraX. Coordinates are in nanometres, as XTC stores them.

Understanding results

The frames in samples.xtc from equilibrium sampling and structure prediction are a bag of independent structures, not a time series. Frame 7 is not later than frame 6, and neighbouring frames are unrelated. Propagation endpoints are also independent: endpoint i is a structure one lag time after starting frame i, superposed on it. No task returns a time-ordered trajectory.

Structure prediction and equilibrium sampling answer different questions. Prediction uses a step scale of 1.5, which sharpens the distribution toward the most likely fold, so its frames should agree with each other. Equilibrium sampling uses 1.0 and is meant to spread across the conformations the model considers accessible. A tight equilibrium ensemble is a result, not a failure, and a spread-out prediction is a warning about the input.

Physicality checks

Each record in physicality.json has a status of passed, rejected or not_checked, and for rejected frames the first failed check in reason. Checks cover protein chains only: ligands, RNA, DNA and single-residue chains are never checked, and a system with no protein chain reports not_checked. Rejected frames stay in the XTC, so filter by frame index when you need clean structures.

no_violations allows none of these, and high_quality_per_residue divides the counts by the number of protein residues (for example, more than 0.05 backbone bond errors per residue). A rejected frame most often means a clash or a broken bond, which a steering sampler such as edm_gd is designed to reduce. The run log and the job's warnings report how many frames failed.

On ProteinIQ, all 16 unsteered adenylate kinase samples, and all 4 edm_gd myoglobin structures with heme passed high_quality. Passing means plausible geometry, not a correct conformation: compare against known structures with RMSD or TM-score, and check geometry in detail with MolProbity.

What the ensemble can and cannot tell you

How often a conformation appears in the samples is not a calibrated Boltzmann probability or a free energy, according to the BioEmu2 model card. Use the ensemble to see which states exist and how they differ, not to read off populations. Quality falls with shallow alignments, so orphan proteins, designed sequences and some viral proteins are less reliable, and a query-only A3M removes evolutionary information entirely. BioEmu2 does not predict binding affinity.

For propagation, BioEmu2 recommends lags from 1 to 1000 ns, the range the model mostly saw in training. Endpoints at longer lags look more and more like equilibrium samples, and beyond a few milliseconds the starting structure no longer matters. For physics-based dynamics of the same system, OpenMM and GROMACS run molecular dynamics simulations.

A seed fixes the random numbers sampling starts from, but GPU arithmetic is not bit-for-bit repeatable between runs. On ProteinIQ, repeated seeded runs with the steering samplers gave slightly different coordinates; edm_fkc varies even within one machine, because its guidance takes gradients through the denoiser.

Sources

• BioEmu2 source code, documentation and model card: github.com/microsoft/bioemu2, release v0.1.0rc0.
• BioEmu2 model checkpoints: huggingface.co/microsoft/bioemu2.