Boltz-2 on the platform

Boltz-2 co-folds a complex: you give it the chains — protein, DNA, RNA and small-molecule ligands as SMILES — and it predicts the assembled 3D structure in one pass, with per-chain and interface confidence scores. It needs no receptor structure, which is the difference from DiffDock, where you supply a receptor and dock a ligand into it.

Running boltz2_predict needs a signed-in account; it is not part of the guest tool set.

What Boltz-2 does here

• Predicts one complex from sequence. 1 to 20 chain entries per job, each typed protein, dna, rna or ligand.
• Scores the interfaces, not just the fold. Per-chain pLDDT and pTM, per-chain interface pTM, and one complex_iptm for the whole assembly.
• Optionally predicts binding affinity for a complex that contains a ligand chain. The number is reported as log10(IC50) with IC50 in µM — lower is stronger. See results for what that value is and is not.
• Runs on GPU as an async job, typically 5–30 minutes depending on chain count and length.

When to use it

• You have sequences and no trustworthy receptor structure — an assembly with no deposited structure, or a target whose close homologs are well represented in the PDB. Accuracy degrades on targets with no close PDB relative, and the model is documented not to capture ligand-induced conformational change, so a mutant or induced-fit pocket is the weakest case rather than the strongest. See translatability.
• You want the whole complex geometry: protein–protein, antibody–antigen, protein–DNA/RNA, or protein–ligand.
• You want an orthogonal check on a docked pose: co-fold the same receptor sequence plus ligand and compare the pocket geometry with the DiffDock pose.
• You want a model-based affinity estimate alongside the structure, knowing it is a prediction and not a measurement.

When to use something else

Limits to know before you run

• Chain count: at most 20 chain entries per job. chain_id is a single character.
• MSAs are off by default. Without use_msa_server, protein chains run in single-sequence mode (msa: empty), and the result flags this as single_sequence_mode. Upstream describes predictions in this mode as suboptimal. Turning the server on sends your sequences to the public ColabFold MMseqs2 service and can add 30 seconds to 10 minutes.
• Affinity needs a ligand chain and more than one chain. With a single chain, or with no ligand entry, no affinity is produced.
• Affinity has a ligand-size ceiling of 128 heavy atoms. Above it, the platform submits the job with affinity turned off and tells you so, so the structure still comes back instead of the whole run failing.
• Affinity is a predicted log10(IC50 [µM]), not an assay readout. Compare it with measured IC50/Kd on your own series before you lean on it. Published testing finds it useful for ranking within one congeneric series and unreliable across series — see translatability.
• One structure per job, with no steering potentials. The platform requests a single diffusion sample and does not enable Boltz's physics-based steering, templates, pocket constraints or experimental-method conditioning. Published accuracy figures were generally obtained with MSAs on and many samples per system, so they describe a richer configuration than this one.

Citation

Passaro S, Corso G, Wohlwend J, et al. Boltz-2: Towards Accurate and Efficient Binding Affinity Prediction. bioRxiv (2025). doi: 10.1101/2025.06.14.659707

Source code: jwohlwend/boltz.

Read next