Protein Structure · MIT × Recursion
Boltz-2
Predicts 3D complex structures and binding affinities for protein-ligand pairs in a single forward pass.
What Boltz-2 is
Boltz-2 is a biomolecular foundation model developed by MIT and Recursion that predicts the three-dimensional structures of protein-ligand complexes alongside their binding affinities. Designed for early-stage drug discovery, it addresses the computational bottleneck of traditional molecular modelling by evaluating complex structures and binding energetics simultaneously.
Physics-based free-energy perturbation methods yield accurate affinity calculations but require extensive computation. Boltz-2 approaches the accuracy of these physics-based approaches while operating roughly 1,000 times faster, enabling rapid virtual screening and candidate triage.
How it works
The model takes a protein sequence and a small-molecule ligand SMILES representation, predicting all-atom coordinates for the co-folded complex together with binding affinity metrics in a single forward pass. It supports execution with NVIDIA cuEquivariance acceleration kernels on compatible hardware and can query multiple sequence alignments via an external server.
Affinity evaluation is handled via two distinct outputs trained with different objectives and datasets. A binary probability score assesses whether a candidate molecule is a true binder rather than a decoy, while a continuous value estimates affinity as log10(IC50) in micromolar units to guide lead optimisation.
Key facts
- Version
- 2.1.0
- Tasks
- Structure prediction, binding affinity
- Input modalities
- Protein sequence, small-molecule SMILES
- Output modalities
- PDB coordinates, binding affinity scores, per-residue pLDDT
- Hardware requirement
- 40 GB GPU memory
What you can use it for
- Hit discovery
- Screen compound collections to separate genuine binders from non-binding decoys using the binary affinity probability output.
- Hit-to-lead optimisation
- Rank analogues and quantify small chemical modifications using the continuous log10(IC50) affinity prediction.
- Protein-ligand complex modelling
- Generate all-atom coordinates for protein-ligand complexes alongside per-residue pLDDT confidence scores in a single step.
Run it here
No installation and no GPU of your own: fill in the inputs and it runs on our servers. Results stay in your browser unless you sign in.
Runs on Managed GPU endpoint
Runs on managed compute. Availability and startup time depend on current capacity.Run this model
Boltz-2
Strengths and limits
Good at
- Matches the accuracy of physics-based free-energy perturbation methods at roughly 1,000 times the speed.
- Predicts full all-atom coordinates and binding affinity metrics concurrently in one forward pass.
- Provides separate specialised outputs for binary binder discrimination and continuous affinity quantification.
Watch out for
- Requires 40 GB of GPU memory for standard operation.
- Inference on CPU hardware is significantly slower than on GPU accelerators.
Inputs and outputs
You provide
Target protein sequence
Single-letter amino acid code. Whitespace and FASTA headers are stripped automatically.
Ligand SMILESoptional
Optional. Omit to fold the protein alone.
You get back
All-atom coordinates in PDB format, viewable and downloadable.
Binding affinity estimate for the protein–ligand pair.
pLDDT score per residue (0–100). Above 70 is generally reliable.
Licence
Commercial use permitted · MIT
The MIT licence allows use in commercial research and products.
Questions
Can I use Boltz-2 commercially?
Commercial use permitted. The MIT licence allows use in commercial research and products.
How do I run Boltz-2 without installing anything?
Use the form under “Run it here” on this page: fill in the inputs and press Run model. It runs on our servers on 1× A100 40GB, typically in 45s, and a run costs about 252 credits. Visitors without an account get 1,000 credits a month.
What inputs does the model take?
It accepts a target protein sequence and a ligand SMILES string, defined via input YAML configuration files.
What outputs does Boltz-2 produce?
It generates an all-atom PDB structure, per-residue pLDDT confidence scores from 0 to 100, and two binding affinity predictions: a binary binder probability and a continuous log10(IC50) estimate.
Which affinity output should be used for analogue design?
The continuous prediction value, log10(IC50) in micromolar units, is designed for lead optimisation to capture subtle changes in affinity between closely related binders.
How to cite it
Passaro, S., Corso, G., Wohlwend, J., Reveiz, M., Thaler, S., Somnath, V. R., Getz, N., Portnoi, T., Roy, J., Stark, H., Kwabi-Addo, D., Beaini, D., Jaakkola, T., and Barzilay, R. (2025). Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction. bioRxiv. doi: 10.1101/2025.06.14.659707.
Written from Boltz-2's own documentation (code). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.