diffdock

1# DiffDock-L
2 
3DiffDock-L is a blind pose predictor: given a protein structure and a ligand,
4it samples ligand placements over the whole surface with a diffusion model and
5ranks them with a separately trained confidence head. The confidence score
6correlates with pose correctness, not with binding free energy — DiffDock does
7not predict whether or how tightly the ligand binds, so for hit triage you
8still pair it with a scorer (GNINA, MM-GBSA) or with `boltz`'s affinity head.
9For protein–protein and nucleic-acid co-folding, route to `boltz` or `chai1`.
10Code and weights are MIT (github.com/gcorso/DiffDock).
11 
12## Running it
13 
14```bash
15cd $DIFFDOCK_REPO   # a clone of github.com/gcorso/DiffDock
16python3 -m inference \
17  --config default_inference_args.yaml \
18  --protein_path target.pdb \
19  --ligand_description "COc1ccc(C#N)cc1" \
20  --out_dir out
21```
22 
23For more than one complex, give `--protein_ligand_csv batch.csv` instead of the
24two single-complex flags; the CSV has four columns — `complex_name`,
25`protein_path`, `ligand_description` (SMILES or an `.sdf`/`.mol2` path), and
26`protein_sequence`. Leave `protein_path` empty and fill `protein_sequence` to
27have DiffDock fold the receptor with ESMFold first; that path and a
28larger-library screening recipe are in `references/workflows.md`.
29 
30Under `--out_dir/<complex_name>/` each sample is written as
31`rank{N}_confidence{score}.sdf`, plus a copy of `rank1.sdf` for convenience.
32…