Multi-Modal Foundation · microsoft
BioEmu
Generative model predicting protein equilibrium structural ensembles and thermodynamic stabilities from sequence.
What BioEmu is
BioEmu (Biomolecular Emulator) is a deep learning system developed by researchers at Microsoft Research AI for Science. It predicts equilibrium structural ensembles and thermodynamic properties of proteins directly from their primary amino acid sequence. Rather than outputting single static configurations, it emulates the Boltzmann distribution of conformations that proteins adopt in solution, generating thousands of statistically independent structures per hour on a single graphics processing unit.
A major challenge in structural biology and drug discovery is capturing the dynamic nature of proteins, including functional conformational shifts, local unfolding, and transient or cryptic binding pockets. Conventional molecular dynamics (MD) simulations require significant computational effort to reach the millisecond timescales needed to observe these events. BioEmu bridges this gap by directly generating representative thermodynamic ensembles and providing relative free energy estimates with errors around 1 kcal/mol relative to long-timescale MD and experimental data.
The model provides mechanistic insights into how mutations impact protein fold stability and produces candidate conformations for experimental validation and functional analysis.
How it works
BioEmu is based on the DiG deep generative architecture.
Key facts
- Architecture
- DiG
- Parameter count
- 31 million
- Fine-tuning variant
- Fine-tuned with AMBER simulation data (BioEmu 1.0)
- Software requirement
- Python >= 3.10
What you can use it for
- Protein equilibrium ensemble generation
- Rapidly sample thousands of statistically independent backbone conformations to emulate Boltzmann-distributed structural ensembles from amino acid sequences.
- Conformational change and cryptic pocket discovery
- Model functional movements such as large-scale domain rearrangements, local unfolding transitions, and the opening of cryptic or transient ligand-binding sites.
- Folding free energy and stability prediction
- Quantify thermodynamic stability differences (folding free energy changes) to evaluate the impact of amino acid mutations on protein fold stability.
- Hypothesis generation for molecular mechanisms
- Formulate experimentally testable hypotheses on the structural basis of mutation-induced destabilisation or allosteric regulation.
Run it here
We list this model but cannot run it yet: no pipeline_tag — needs a human.
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Achieves a relative free energy mean absolute error of 0.76 kcal/mol compared to experimental folding measurements, with a Pearson correlation of 0.66.
- Emulates molecular dynamics equilibrium free energy landscapes with a mean absolute error of 0.91 kcal/mol on in-house benchmark sets.
- Demonstrates high coverage of functional structural states, including 85% for domain motions and 85% for bound cryptic pocket conformations.
- Samples thousands of statistically independent conformations per hour using a single GPU.
Watch out for
- Outputs structures strictly in backbone frame representation, requiring separate external pipelines for side-chain reconstruction and full equilibration.
- Shows low prediction quality for protein-protein interactions and cannot model multi-chain protein complexes.
- Cannot represent explicit interactions with external chemical entities such as small-molecule ligands.
- Does not design or generate novel protein sequences, operating solely on provided sequences.
- Inherits biases and approximations present in its synthetic training data, including AlphaFold2 predictions and MD simulation force fields.
Inputs and outputs
You provide
Sequence or text
A protein or nucleotide sequence, or free text — whichever this model was trained on.
You get back
Mean-pooled representation, one vector per input.
Licence
Check the licence before commercial use · mit
The mit licence carries terms worth reading before you depend on it.
Questions
Can I use BioEmu commercially?
Check the licence before commercial use. The mit licence carries terms worth reading before you depend on it.
What structural format does BioEmu generate?
The model directly outputs protein structures in a backbone frame representation. Interfaces are provided to connect these backbones to side-chain reconstruction and molecular dynamics equilibration workflows.
Can BioEmu model protein-protein complexes or multi-chain systems?
No. The model is not designed for multi-chain assemblies and demonstrates low prediction quality for protein-protein interactions.
Does BioEmu model binding interactions with small molecules or ligands?
No. The model does not explicitly incorporate interactions with small molecules or non-protein chemical entities, though it can sample holo-like apo pocket conformations.
How accurate are the free energy estimates produced by the model?
BioEmu predicts folding free energies with a mean absolute error of 0.76 kcal/mol against experimental stability benchmarks, and matches molecular dynamics state free energies within 0.91 kcal/mol.
How to cite it
Lewis, Sarah, Tim Hempel, José Jiménez Luna, Michael Gastegger, Yu Xie, Andrew Y. K. Foong, Victor García Satorras, et al. (2024). Scalable Emulation of Protein Equilibrium Ensembles with Generative Deep Learning. bioRxiv. doi:10.1101/2024.12.05.626885.
Written from BioEmu's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.