diffdock

DiffDock-L

DiffDock-L is a blind pose predictor: given a protein structure and a ligand, it samples ligand placements over the whole surface with a diffusion model and ranks them with a separately trained confidence head. The confidence score correlates with pose correctness, not with binding free energy — DiffDock does not predict whether or how tightly the ligand binds, so for hit triage you still pair it with a scorer (GNINA, MM-GBSA) or with boltz's affinity head. For protein–protein and nucleic-acid co-folding, route to boltz or chai1. Code and weights are MIT (github.com/gcorso/DiffDock).

Running it

cd $DIFFDOCK_REPO   # a clone of github.com/gcorso/DiffDock
python3 -m inference \
  --config default_inference_args.yaml \
  --protein_path target.pdb \
  --ligand_description "COc1ccc(C#N)cc1" \
  --out_dir out

For more than one complex, give --protein_ligand_csv batch.csv instead of the two single-complex flags; the CSV has four columns — complex_name, protein_path, ligand_description (SMILES or an .sdf/.mol2 path), and protein_sequence. Leave protein_path empty and fill protein_sequence to have DiffDock fold the receptor with ESMFold first; that path and a larger-library screening recipe are in references/workflows.md.

Under --out_dir/<complex_name>/ each sample is written as rank{N}_confidence{score}.sdf, plus a copy of rank1.sdf for convenience. The confidence value in the filename is a logit, so it is unbounded and can be negative; among samples for the same complex higher is better, but values are not comparable across different complexes or ligands.

The YAML config overwrites your CLI flags

inference.py loads --config default_inference_args.yaml after argparse and replaces every key it finds, so passing --samples_per_complex 40 or --model_dir ... on the command line is silently ignored if the same key sits in the YAML. To change sampling depth or any other key the YAML defines, copy the YAML, edit the copy, and point --config at it.

The first run is silent for ~11 minutes and needs ≥32 GB host RAM

Before the first complex, DiffDock precomputes SO(3) and torus lookup tables. That step is silent on stderr, takes ~11 minutes, and may exhaust a small machine. Use a probed SSH context with at least 64 GiB RAM and precompute the tables while building the environment; do not assume a quiet Run has crashed.

The README's --ligand works on the CLI by accident — use --ligand_description

The upstream README shows --ligand, which only works because argparse prefix-matches it to the real flag --ligand_description. That shortcut is CLI-only: as a CSV column header or YAML key, ligand matches nothing and the row is silently treated as having no ligand. Spell the flag and the column header out in full.

Wisp execution

Use python only for bounded interactive checks. For a long or GPU-backed workload, require a selected and probed ssh:<alias> context and load remote-compute-ssh. Put the documented invocation in a self-contained project script, activate the remote environment explicitly, stage only small files with input_paths, and make the command write to a known absolute remote result path. Submit it with run_in_context and register that exact ssh:// path in output_specs. Call monitor_run once when waiting is needed, get_run once for a snapshot, or cancel_run to stop. Do not send a scheduler submission through the SSH-direct runner.

Errors worth recognizing

Next: rescore the rank1.sdf poses before ranking ligands against each other — boltz's affinity head is the in-tree option — since the DiffDock confidence head alone is not an affinity predictor.