Model Book/DiffDock

DiffDock

B

MIT

Molecular Dockingopen
Updated 3 months ago
NextIn take

DiffDock was the first to show that diffusion models could do molecular docking — a genuinely important contribution — but the field has moved past it. Boltz-2 does the same thing better and in a unified model.

Specifications

ArchitectureDiffusion generative model over ligand poses in the SE(3) space
Parameters~50M
Training DataPDBBind refined set, PDB protein-ligand complexes
LicenseMIT license. Fully open weights and code.
Hardware1x GPU with 16GB+ VRAM
Inference CostSelf-hosted — ~$0.005/docking on cloud GPU
API AvailableNo
Weights AvailableYes

Benchmark Performance

BenchmarkScore
PoseBusters58% valid
PDBBind38% top-1 < 2Å

When to Use This

  • +Blind docking where the binding site is unknown
  • +Research pipelines needing a fully open docking method

When NOT to Use This

  • Production virtual screening where accuracy is paramount — Boltz-2 is better
  • When you know the binding site — classical methods may suffice

Production Readiness

research

Known Users

  • Academic drug discovery labs

Grade Rationale

B

Pioneered the diffusion approach to molecular docking and remains a solid baseline. But Boltz-2 and AF3 have since surpassed it on accuracy, and the PoseBusters pass rate shows room for improvement on physical validity.

Sources

Update History

2026-01-10Initial entry