Use Boltz-2 Online
Commercially Available Boltz-2 No-Code Web Server
Boltz-2: Biomolecular Structure Prediction and Small-Molecule Binding Affinity Prediction
Boltz-2 is a structural biology foundation model that predicts biomolecular complex structures and small-molecule binding affinity in a single framework. It was developed by researchers at MIT CSAIL, the MIT Jameel Clinic, Valence Labs, Recursion and ETH Zurich, and it succeeds Boltz-1 while building on ideas from AlphaFold3. According to the Boltz repository, the code and model weights are released under the MIT license, which permits academic and commercial use.
Predicting how a small molecule binds a protein, and how tightly, is central to hit discovery and lead optimization. Physics-based free-energy perturbation (FEP) methods can be accurate but are computationally expensive, while docking and earlier machine learning scores are fast but often unreliable at ranking close analogs. Structure predictors such as AlphaFold3 and Boltz-1 model complexes well but do not directly output affinity.
Boltz-2 addresses this gap by jointly modeling the 3D complex and an affinity estimate. The authors report that it approaches the accuracy of FEP-style methods on benchmarks while being at least 1,000 times more computationally efficient.
Attribute | Detail |
|---|---|
Model type | Structural biology foundation model for complex structure and binding affinity |
Modalities | Proteins, RNA, DNA, ligands, antibodies, TCR/pMHC |
Trunk | 64 PairFormer layers (48 in Boltz-1) |
Affinity outputs | Binding likelihood (binder vs. decoy) and a continuous value on a log10 micromolar scale |
Optional controls | Experimental-method conditioning, contact and pocket constraints, templates, steering potentials |
License | MIT (per the GitHub repository) |
How Boltz-2 Predicts Complex Structure and Binding Affinity
Boltz-2 pairs a structure-prediction trunk with a dedicated affinity module, so a single run can return both the predicted pose and an estimate of binding strength.
Structure trunk: A deeper PairFormer trunk (64 layers) with a denoising module inherited from Boltz-1 and a confidence module that reports PAE, PDE and ipTM-style scores.
Controllability: Users can condition predictions on the experimental method (such as X-ray, NMR or molecular dynamics), supply contact and pocket distance constraints, and integrate multi-chain templates. A Boltz-2x variant adds physics-based steering.
Affinity module: A PairFormer over protein-ligand and intra-ligand pairs feeds two heads: a binary binding likelihood meant for separating binders from decoys, and a continuous affinity value meant for comparing related ligands during hit-to-lead work.
Training data: Structures from the PDB, molecular dynamics ensembles, distillation data, and affinity measurements curated from ChEMBL, BindingDB, PubChem and other sources.
Boltz-2 Benchmark Results for Affinity Prediction and Virtual Screening
The paper evaluates affinity prediction on physics-style benchmarks, a blind challenge and screening tasks.
Benchmark | Reported result |
|---|---|
FEP+ subset (CDK2, TYK2, JNK1, P38) | Average Pearson R of 0.66 |
Full OpenFE set (876 hit-to-lead measurements) | Approaches OpenFE performance |
CASP16 affinity track (140 pairs, 2 targets) | Outperforms the top-ranked participants out of the box |
MF-PCBA hit discovery | Enrichment factor of 18.4 at the 0.5% threshold |
TYK2 prospective screen | All top 10 predicted binders from the 64,960-compound Enamine kinase library were confirmed by ABFE calculations |
Structure prediction
On unseen PDB complexes Boltz-2 matches or moderately improves on Boltz-1, edges out Chai-1 and ProteinX, and trails AlphaFold3. It also shows stronger correlation with molecular dynamics fluctuations (RMSF) than Boltz-1, BioEmu and AlphaFlow on the mdCATH and ATLAS sets.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with Boltz-2 on Tamarind Bio
Virtual screening: Rank large compound libraries by predicted binding likelihood to prioritize hits, as the authors demonstrate on the TYK2 prospective screen.
Hit-to-lead optimization: Compare closely related ligands using the continuous affinity value to guide analog selection.
Generative design: The paper pairs Boltz-2 scores with a SynFlowNet generator over Enamine REAL building blocks to propose candidates predicted to bind.
Complex modeling: Predict structures of protein, nucleic acid, antibody and ligand assemblies, optionally constrained by pockets, contacts or templates.
How to Use Boltz-2 on Tamarind Bio
Open the tool: Log in to tamarind.bio and select Boltz-2.
Define the complex: Enter the protein, DNA or RNA sequences and the ligand as a SMILES string, as Boltz-2 expects.
Add optional context: If exposed, provide MSAs, templates, contact or pocket constraints, or an experimental-method condition.
Request affinity: If exposed, enable affinity prediction and specify the ligand to score.
Run the job: Submit and wait for the prediction to complete.
Download results: Retrieve the predicted structure and confidence scores, plus the binding likelihood and affinity value when requested.
Interpret: Use the binary likelihood to separate binders from decoys and the continuous value to compare related ligands, keeping each output to its intended use.
Continue downstream: Feed top candidates into further docking, simulation or experimental follow-up.
Things to Keep in Mind
Affinity predictions depend on accurate predicted structures and pockets, and the module does not explicitly handle cofactors, ions, water or multimeric partners.
Performance varies strongly across assays; on eight internal Recursion assays, Pearson correlation exceeded 0.55 on only three.
Structure prediction is similar to predecessors, with difficulty on large complexes and large conformational changes, and gains in molecular dynamics modeling are modest.
Affinity labels are noisy, and the two affinity outputs were trained with different supervision, so each should be used only in its intended context.
Source: Passaro S., Corso G., Wohlwend J., et al., "Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction", bioRxiv preprint (2025), DOI 10.1101/2025.06.14.659707. Code and weights: https://github.com/jwohlwend/boltz