Boltz-2

You already have a binder sequence and the target it should hit, and you want to know whether they actually stick together before you order DNA. Returns the predicted complex and a 0-to-1 interface confidence score. Trained on antibody-antigen complexes, so it gives a genuine second opinion next to AlphaFold2.

Folds a binder and its target together and tells you how confident it is that they touch. It was trained on antibody-antigen complexes and works from sequence alone, which makes it a genuinely independent second opinion next to AlphaFold2 multimer: when the two agree the complex is probably real, and when they disagree that is worth knowing before you order DNA. Returns the folded complex plus ipTM (confidence in the interface), pTM (confidence in the whole complex) and per-residue confidence for every design. Boltz-2, Passaro et al., bioRxiv 2025.

When it fits:

• You designed binders with another tool here and need them scored against the antigen you actually care about.
• You have natural or near-natural antibody fragments (scFv or Fab), nanobodies or peptides and want a fast independent fold of each.
• AlphaFold2 interface confidence has topped out across your candidates and you want a second, differently trained opinion before committing to synthesis.

You will need:

• Antigen PDB or mmCIF (single chain; the binder is added separately).
• One or more binder sequences (scFv, nanobody, peptide, anything that folds as a single protein chain), 20 to 400 aa each.
• Optional: a list of antigen hotspot residue numbers to count contacts against (1-indexed on the chosen antigen chain).

Each run uses a preset that sets the scale and scope:

Single-sequence (fast)

YAML ``msa: empty`` per chain. The right choice for designed sequences (MPNN, RFantibody, BindCraft, BoltzGen, RFdiffusion, PXDesign outputs) where no informative MSA exists. Runs on an A100-40GB.

With MSA (slower, natural sequences)

Boltz fetches MSAs from the public ColabFold MMseqs2 endpoint at runtime. Intended for natural / near-native sequences; slower than the single-sequence default, since it also fetches MSAs. On the one head-to-head we have run it separated a designed binder from decoys WORSE than the single-sequence default, so prefer that unless you know your sequence has natural relatives.

Per-design folded complex PDB + ipTM, pTM, complex_pLDDT, complex_iplddt, and hotspot contact count. Strict-pass classification (complex_pLDDT > 85, ipTM > 0.7, n_hotspot_contacts > 4) surfaces which designs are worth ordering.

Where a tool reports them, the scores mean:

ipTM

Predicted confidence in the contact between two chains, on a 0 to 1 scale. Higher is better: 0.65 to 0.75 depending on the tool. Individual tools set their own pass bar a little either side of that — this guide's own results summary above states this tool's. On a multi-chain target the number may cover the target's own chain–chain interface as well as the binder's, so read the per-tool note on the results table before comparing designs on it.

pLDDT

Per-residue confidence in the predicted fold. Higher means the model is more sure of that part of the structure.

i_pAE and pAE

Predicted alignment error, at the interface (i_pAE) or across the whole structure (pAE). Lower is better.

Passaro et al., bioRxiv 2025