DiffDock — batch and sequence-only workflows

The main SKILL.md covers the single-complex path. Two further entry points share the same inference module and the same YAML-overwrites-CLI caveat.

Batch / small-library docking

Prepare one CSV row per complex and hand it to --protein_ligand_csv:

complex_name,protein_path,ligand_description,protein_sequence
6lu7_frag001,6lu7.pdb,CC(=O)Nc1ccc(O)cc1,
6lu7_frag002,6lu7.pdb,COc1ccc(C#N)cc1,

cd $DIFFDOCK_REPO
python3 -m inference \
  --config default_inference_args.yaml \
  --protein_ligand_csv batch.csv \
  --out_dir out/

Each row gets its own subdirectory under --out_dir. complex_name may be left empty (an index is used). ligand_description accepts a SMILES string or a path to an .sdf/.mol2; one ligand per row, so a multi-ligand SDF needs to be split first.

For libraries beyond a few hundred ligands, raise samples_per_complex only in a copied YAML (see the CLI-overwrite gotcha) and shard the CSV across jobs — the embedding step holds ESM-2 650M in GPU memory for the whole run.

Sequence-only receptor (ESMFold first)

If a receptor structure is not available, leave protein_path empty and fill protein_sequence; DiffDock folds the chain with ESMFold before docking. That path needs fair-esm installed; the esm2_t33_650M_UR50D checkpoint is fetched on first use into ~/.cache/torch/hub/checkpoints. Expect the fold to dominate wall-clock for receptors over ~400 residues, and treat poses against an ESMFold model with more scepticism than against a crystal structure: pocket geometry inherits the fold's pLDDT.

Reading confidence across a library

Confidence logits are calibrated within a complex, not across ligands. To compare ligands, rescore each rank1.sdf with an external scorer; using the raw DiffDock confidence as a virtual-screening ranker conflates "well-placed pose" with "tight binder" and the upstream FAQ explicitly warns against it.