---

name: docking-inference description: "Plan, validate, command-build, and troubleshoot DiffDock command-line docking prediction runs for single complexes or batch CSV inputs." disable-model-invocation: true metadata: disco-role: operating license: MIT

Docking Inference

Use this sub-skill when a user wants to run, adapt, or debug DiffDock command-line docking prediction with python -m inference. It covers single-complex and batch CSV inputs, model/config choices, output layout, confidence-ranked SDF files, optional reverse-process visualization, and safe validation before expensive runs.

Start Here

1. Check runtime prerequisites in the root install/runtime reference when that file is present, especially the Torch, PyG, RDKit, ProDy, ESM/OpenFold, and CUDA notes.
2. Pick the input mode in input-output-formats.md: batch CSV or single complex.
3. Build a dry command with scripts/build_inference_command.py; it prints a command and never runs inference.
4. Validate CSV or single-complex inputs with scripts/validate_inference_inputs.py; it uses only Python standard-library modules.
5. Review cli-reference.md, configuration.md, and troubleshooting.md before launching a long model run.

Routes

• Batch docking: use a CSV with complex_name, protein_path, ligand_description, and protein_sequence; see input-output-formats.md.
• Single-complex docking: provide either --protein_path or --protein_sequence, plus --ligand_description; see cli-reference.md.
• Model/checkpoint overrides: use --model_dir, --confidence_model_dir, --ckpt, and --confidence_ckpt; see configuration.md.
• Output inspection: expect one output directory per complex and confidence-ranked SDF names; see input-output-formats.md.
• Failure recovery: diagnose imports, CPU/GPU behavior, model downloads, invalid inputs, RDKit parsing, ESMFold sequence paths, and GNINA caveats in troubleshooting.md.

Boundaries

• For the graphical Gradio wrapper and browser UI, route to web-ui.
• For benchmark metrics, aggregate RMSD reporting, and evaluation datasets, route to evaluation-benchmarks.
• For training, dataset preparation, and ESM embedding preparation for evaluation/training, route to training-data.
• This sub-skill does not run DiffDock, download models, import heavy DiffDock modules, or require the original repository examples to remain available.