Drug Design Pipeline

Overview

This skill provides a deterministic pipeline orchestrator (pipeline.py) that auto-chains multiple drug discovery skills into reproducible workflows. Instead of manually invoking 10+ scripts in the correct order, the agent runs a single command that handles file wiring, schema validation, and manifest logging at every stage.

Why a script instead of manual chaining:

• Guarantees correct execution order — the agent cannot skip or reorder stages
• Validates I/O contracts between stages — catches schema mismatches early
• Logs every invocation to _script_manifest.jsonl — critique agent can verify the full trace
• Stops on first failure with clear diagnostics — no silent errors

Pipeline Mode Selection

Choose the mode based on what the user wants:

Trigger phrases: "drug discovery pipeline", "find drugs", "design drugs", "screen compounds", "druggability assessment", "de novo design", "lead optimization"

Quick Start

Full Pipeline (target → drug candidates)

python scripts/pipeline.py \
    --mode full \
    --protein target.pdb \
    --output-dir results/ \
    --top-n 10

Lead Optimization (improve an existing hit)

python scripts/pipeline.py \
    --mode lead-opt \
    --protein target.pdb \
    --ligand hit_compound.sdf \
    --output-dir lead_opt_results/ \
    --top-n 20

Virtual Screening (screen a compound library)

python scripts/pipeline.py \
    --mode screen \
    --protein target.pdb \
    --library compound_library.sdf \
    --output-dir screening_results/ \
    --top-n 50

Target Assessment (is this druggable?)

python scripts/pipeline.py \
    --mode assess \
    --protein target.pdb \
    --output-dir assessment/

De Novo Design (generate novel molecules)

python scripts/pipeline.py \
    --mode denovo \
    --protein target.pdb \
    --output-dir denovo_results/ \
    --top-n 15

From Sequence (no PDB available)

python scripts/pipeline.py \
    --mode full \
    --sequence "MKTLLLTLLLGLLVSSALA..." \
    --output-dir results/

Pipeline Stages Reference

Full Pipeline (--mode full)

Lead Optimization (--mode lead-opt)

Virtual Screening (--mode screen)

Target Assessment (--mode assess)

De Novo Design (--mode denovo)

I/O Contract Reference

These are the JSON schemas each stage expects from its upstream stage:

pockets.json (pocket-detection → docking, druggability, denovo)

{
  "pockets": [
    {
      "center": [10.5, 22.3, 15.0],
      "volume_A3": 542.8,
      "residues": ["ASP189", "SER195"]
    }
  ]
}

Critical field: pockets[0].center must be [float, float, float] — dock.py reads this directly.

affinity.json (binding-affinity → consensus)

{
  "predictions": [
    {
      "pose_id": 1,
      "predicted_pKd": 7.2,
      "predicted_dG_kcal": -9.8,
      "confidence": "moderate"
    }
  ]
}

consensus.json (final output)

{
  "rankings": [
    {
      "pose_id": 1,
      "consensus_score": 0.85,
      "consensus_rank": 1,
      "individual_ranks": {"predict": 1, "rescore": 2}
    }
  ]
}

Output Directory Structure

After --mode full:

pipeline_results/
├── pockets.json              # Detected binding pockets
├── druggability.json         # Pocket druggability scores
├── candidates.sdf            # Generated molecules (pre-filter)
├── filtered.sdf              # Drug-like molecules (post-filter)
├── docking/
│   ├── poses.sdf             # Docked poses
│   └── scores.csv            # Docking scores
├── interactions.json          # Protein-ligand interactions
├── affinity.json              # Binding affinity predictions
├── mmgbsa.json                # MM/GBSA rescoring
├── consensus.json             # Final consensus ranking
├── complex_3d.html            # Interactive 3D viewer
├── pipeline_report.json       # Stage timings and status
└── _script_manifest.jsonl     # Full invocation log (for critique)

Script Reference

Error Recovery

Related Skills

• pocket-detection: Standalone pocket detection with visualization
• binding-affinity: Standalone affinity prediction and consensus scoring
• denovo-design: Standalone molecule generation with multiple strategies
• molecular-docking: Standalone docking with Vina/DiffDock
• structure-prediction: Standalone protein structure prediction from sequence
• molecule-visualization: Standalone 2D/3D molecular visualization

References

• Eberhardt, J. et al. "AutoDock Vina 1.2.0." J. Chem. Inf. Model. 61, 3891-3898 (2021).
• Corso, G. et al. "DiffDock: Diffusion Steps, Twists, and Turns for Molecular Docking." ICLR (2023).
• Le Guilloux, V. et al. "Fpocket." BMC Bioinformatics 10, 168 (2009).
• Wang, R. et al. "The PDBbind database." J. Med. Chem. 47, 2977-2980 (2004).