Boltz-2

Boltz-2 is the open-weights diffusion co-folder closest in surface to AlphaFold3: a YAML describing protein, DNA, RNA, and ligand chains in, mmCIF plus pTM/ipTM/pLDDT confidences out, with an optional small-molecule affinity head. Among our four co-fold skills it is the default for binder-validation campaigns — fully open MIT weights and the fastest sampler; pick chai1 when you want a second independent model for consensus, openfold3 when AF3-faithful settings matter, and esmfold2 when you can live without an MSA. Code and weights are MIT (PyPI boltz, github.com/jwohlwend/boltz).

Running it

# complex.yaml
version: 1
sequences:
  - protein:
      id: A
      sequence: MVTPEGNVSLVDESLLVGVTDEDRAVRS...   # target
  - protein:
      id: B
      sequence: AIQRTPKIQVYSRHPAENG...            # binder
  - ligand:
      id: L
      smiles: 'N[C@@H](Cc1ccc(O)cc1)C(=O)O'      # or  ccd: SAH

boltz predict complex.yaml \
    --use_msa_server --out_dir out/ --recycling_steps 3 --diffusion_samples 5

Each protein chain needs an MSA; without one the run exits before the model loads. --usemsaserver queries api.colabfold.com (expect a 30–90 s pause per chain) and is the right default unless you already have an .a3m to name under msa: in the YAML. Setting msa: empty forces single-sequence mode — that is an accuracy sacrifice, not a speed or memory optimization, because the MSA search runs on CPU before the GPU stage starts.

Per input the output lands at out/boltzresultscomplex/predictions/complex/. Read confidencecomplexmodel0.json first: iptm > 0.5 is the community pass line for an interface, complexplddt > 0.7 for the fold itself, and confidencescore is the weighted aggregate the structures are ranked by. Structures themselves are complexmodel{0..N-1}.cif (or .pdb with --outputformat pdb).

Affinity head

Add a properties: block naming one ligand chain as the binder and Boltz-2 predicts protein–small-molecule binding affinity alongside the structure:

properties:
  - affinity:
      binder: L            # the ligand chain id, not the protein

Output gains affinitycomplex.json next to the confidence file: affinitypredvalue is log10(IC50 in μM) — lower is tighter (≈0 → 1 μM, −3 → 1 nM); affinityprobability_binary is the 0–1 binder-vs-non-binder score and is what to rank hits by. One affinity ligand per input; the binder must be a ligand chain (no protein–protein affinity), and Boltz v2.2.x caps affinity ligands at 128 atoms. FASTA inputs cannot request affinity at all.

msa: empty is an accuracy hit, not a memory save

Single-sequence mode has been suggested elsewhere as a way to fit smaller GPUs. It does not help: the MSA search is CPU-side, so --usemsaserver versus msa: empty changes nothing about peak VRAM. If you OOM, lower --diffusionsamples or --maxparallel_samples, or move to an 80 GB tier; do not trade away the MSA for it.

Missing fast kernels are slow, not fatal

ImportError for cuequivarianceopstorch or its libcueops.so means the compiled triangle-kernel package is not on the loader path. --nokernels falls back to the reference PyTorch path — roughly 2× slower, numerically identical, so it is the right unblock for a one-off and the wrong choice for a campaign.

Errors worth recognizing

Next: compute clash and interface metrics on passing complexes, or feed them back to proteinmpnn for another design round.