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
bashcd $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 on the default Modal
tier runs out of host memory and SIGKILLs mid-precompute. Set
provider_params.modal.memory: 65536 (or your provider's equivalent), and do
not assume a hang means a crash.
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.
Errors worth recognizing
| You see | It means / do this |
|---|---|
ValueError: not allowed to raise maximum limit at startup | setrlimit(NOFILE, 64000) exceeds the sandbox hard limit — sed the constant in inference.py to min(64000, rlimit[1]). |
| Silent SIGKILL a few minutes into the SO(3) precompute | Host RAM exhausted — see the gotcha above. |
python3: not found | You are on the upstream rbgcsail/diffdock image — that one runs from /home/appuser/DiffDock under micromamba. |
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.
