DiffDock-L is an open-source diffusion model for blind small-molecule docking — predicting how a ligand binds a protein without a known binding pocket — released under the MIT license. It returns ranked 3D poses with a confidence score per pose, not a binding-affinity estimate. MoleculeDesk manages DiffDock-L's pinned CUDA-specific Python environment and model/language-model checkpoints so you don't have to hand-configure PyTorch Geometric's CUDA extension wheels yourself.

Status: coming soon. The adapter, manifest, and pinned provenance are implemented and unit-tested. An install on a real NVIDIA RTX 3090 now gets through dependencies, model weights and the CUDA check, but DiffDock-L's one-time setup (about 30 minutes of precomputed lookup tables) has not yet been verified end to end, and there is no run benchmark yet. See Verification status below.

What DiffDock-L is used for

DiffDock-L is designed for:

Blind protein-ligand docking — predicting where and how a small molecule binds a protein, without requiring a known binding site.
Pose ranking — generating multiple candidate 3D poses per complex, ranked by a learned confidence score.

DiffDock-L does not predict binding affinity — for that, see Boltz-2.

Requirements


Model version	v1.1
License	MIT
Python	3.10, via `uv`
Platforms	`linux-x64` + NVIDIA CUDA only (required, no CPU path implemented for this model)
Input format	`.json` job file referencing a protein `.pdb` and a ligand `.sdf`/`.mol2` file, or a ligand SMILES string
Output	top pose (`.sdf`), all ranked poses (`.sdf`), a structured `poses_summary.json` with rank/confidence/atom-count

DiffDock-L's pinned dependency stack (torch==1.13.1+cu117 plus matching CUDA-built PyTorch Geometric extension wheels) is CUDA/Linux-specific with no Apple Silicon build available, and upstream exposes no CPU/device-selection flag at all — device selection is DiffDock-L's own automatic CUDA-or-CPU logic. MoleculeDesk requires a working CUDA install (checked at install-verification time) so a run is never silently downgraded to CPU.

Job format

{
  "jobName": "example-complex",
  "proteinPath": "protein.pdb",
  "ligand": { "path": "ligand.sdf" }
}

ligand accepts exactly one of path (an .sdf/.mol2 file, resolved relative to the job file) or smiles (a SMILES string). jobName is optional and, if present, must be a safe identifier (letters, digits, ./_/-; no path separators or ..) — it becomes a directory name under DiffDock-L's own output path.

Source

MoleculeDesk pins DiffDock-L to a specific upstream commit on main — which already ships the current DiffDock-L architecture by default (upstream's own legacy v1 model is reachable only via an old git tag) — plus the exact model-weights release asset and the ESM2 language-model checkpoint DiffDock-L's own pipeline always downloads to compute protein embeddings. See the upstream DiffDock repository for full details on the model itself.

DiffDock-L benchmarks

DiffDock-L is coming soon. Pace and cost numbers will be published here when it ships.

Verification status

pnpm typecheck, pnpm test, pnpm build, and pnpm pack:verify all pass with DiffDock-L registered. moldesk install diffdock correctly fails closed on machines without a matching platform (for example Darwin/arm64 with no NVIDIA GPU). On a real Linux x64 RTX 3090 the install now passes the dependency, model-weight and CUDA stages. DiffDock-L also runs a long one-time precomputation (its diffusion lookup tables, about 30 minutes on one CPU core) that MoleculeDesk caches at install time, and that step has not yet completed in a verified install. The real live gate — a clean install plus a docking run against the checked-in example job — is still outstanding:

moldesk install diffdock
moldesk run diffdock examples/diffdock/job.json

Until that run completes and its results are recorded, DiffDock-L stays status: planned in the registry — MoleculeDesk never marks a model beta/available from unit tests alone.

FAQ

Does moldesk install diffdock work today? Not yet — it requires Linux with an NVIDIA GPU, which was unavailable when this integration was built. The adapter, manifest, and tests are complete; only the real hardware run remains.

What license is DiffDock-L under? MIT, including the model weights. MoleculeDesk doesn't redistribute DiffDock-L's weights or the third-party ESM2 checkpoint it depends on; installing pins and downloads them from their original sources.

Does DiffDock-L predict binding affinity? No — it predicts ranked docking poses with a confidence score. For binding-affinity prediction, see Boltz-2.

Run DiffDock-L for blind small-molecule docking