DiffDock (MolComp)

Quick links

• GitHub: https://github.com/gcorso/DiffDock
• Paper: https://arxiv.org/abs/2210.01776

What MolComp runs

MolComp runs DiffDock inference in AWS Batch and uploads the full output folder to your Workspace:
user/<your_id>/jobs/<job_id>/

Upstream DiffDock supports both single-complex docking and batching via CSV. MolComp’s Playground is optimized for single-complex docking (protein PDB + ligand).

Inputs (MolComp Playground)

Required:

• Protein structure (PDB upload)

Ligand (choose one):

• SMILES string
• Ligand file upload (SDF/MOL2/MOL). If provided, the ligand file takes precedence over SMILES.

Parameters

MolComp passes parameters to DiffDock as CLI flags:

• params.<key> = value becomes --<key> <value>
• params.<key> = true becomes --<key>
• empty / false values are omitted

If you need an upstream flag that is not present as a first-class UI field, use “Advanced (JSON merge)” and add it there.

Common parameters in the Playground

• samples_per_complex (poses per target; higher = slower)
• inference_steps (diffusion steps; higher = slower)
• batch_size (throughput; too high can OOM)
• num_workers
• no_random, no_final_step_noise, ode
• sigma_schedule, actual_steps
• save_visualisation, keep_local_structures
• model_dir, ckpt, confidence_model_dir, confidence_ckpt (advanced; only if you know what you’re doing)

Optional: config
Upstream commonly uses --config default_inference_args.yaml. If your runner image includes that file, you can set config accordingly.

Advanced JSON merge examples

Example: fewer poses, deterministic

{
  "samples_per_complex": 5,
  "no_random": true,
  "inference_steps": 20
}

Example: enable a config file (if present in the container)

{
  "config": "default_inference_args.yaml"
}

Outputs

MolComp uploads the runner output folder to your Workspace under:
user/<your_id>/jobs/<job_id>/

Typical DiffDock outputs are docked poses (often SDF) plus logs/diagnostics. Exact filenames vary by version/checkpoint; MolComp uploads the entire output directory to avoid losing artifacts.

Notes / gotchas

• Provide a reasonable protonation/tautomer for the ligand (DiffDock will dock what you give it).
• If you upload a ligand file and also provide SMILES, the ligand file wins.
• If you use “Advanced (JSON merge)”, keys must match upstream flag names (without the leading --).