docking - SKILL.md Agent Skill

name: docking description: Use when performing protein-ligand docking, virtual screening, or structure-based drug design. Covers receptor preparation (protonation, pocket definition), AutoDock Vina/Gnina docking engines, high-throughput virtual screening pipelines, pose analysis with interaction fingerprints, and ensemble docking for protein flexibility.

Docking — Protein-Ligand Docking & Virtual Screening

AutoDock Vina 1.2 · Gnina · pdbfixer · ProLIF · fpocket. For structure-based drug design: binding mode prediction, virtual screening, and lead optimization by docking.

When to Use This Skill

Predicting how a small molecule binds to a protein (binding mode / pose)
Virtual screening: ranking a library of compounds by predicted binding affinity
Validating a pharmacophore hypothesis in 3D structural context
Ensemble docking to account for protein flexibility
Re-scoring docking poses with physics-based (MM-GB/SA) or CNN-based scoring
Fragment-based screening (→ see fbdd skill for growing/linking)

Decision Tree — Docking vs. Other Methods

Input: protein structure 3D?
  NO  → ligand-based methods (pharmacophore, QSAR, similarity search)
  YES → docking

Compound set size?
  > 10 000   → VS pipeline (references/virtual-screening.md)
  10–10 000  → standard docking batch (references/vina-gnina.md)
  < 10       → manual docking + careful pose analysis

Goal: binding mode accuracy vs. ranking accuracy?
  Binding mode → high exhaustiveness, Gnina CNN rescoring
  Ranking      → standard Vina + clustering + MM-GB/SA rescore

Protein structure source?
  X-ray / CryoEM  → direct prep (references/protein-prep.md)
  Homology model  → validate first (→ homology-modeling skill)
  AlphaFold       → check pLDDT > 80 in pocket region before docking

Quick Start

import subprocess
from pathlib import Path

# 1. Prepare receptor (pdbfixer + obabel → PDBQT)
# See references/protein-prep.md for full workflow

# 2. Prepare ligand
import subprocess
subprocess.run([
    "obabel", "ligand.sdf", "-O", "ligand.pdbqt",
    "--gen3d", "-h"
], check=True)

# 3. Run Vina
result = subprocess.run([
    "vina",
    "--receptor", "receptor.pdbqt",
    "--ligand",   "ligand.pdbqt",
    "--center_x", "10.5",
    "--center_y", "-2.3",
    "--center_z", "14.1",
    "--size_x",   "20",
    "--size_y",   "20",
    "--size_z",   "20",
    "--exhaustiveness", "16",
    "--num_modes", "9",
    "--out", "docked.pdbqt"
], capture_output=True, text=True, check=True)

# 4. Parse best score
for line in result.stdout.splitlines():
    if line.strip().startswith("1 "):
        print("Best score:", line.split()[1], "kcal/mol")
        break

Router — What to Read

Task	Reference
Clean PDB, add H, assign protonation, define docking box	`references/protein-prep.md`
AutoDock Vina / Gnina docking, parameters, scoring	`references/vina-gnina.md`
High-throughput VS pipeline, metrics (BEDROC, EF), filtering	`references/virtual-screening.md`
Extract interactions (H-bonds, hydrophobic, π), ProLIF, clustering	`references/pose-analysis.md`
Ensemble docking, protein flexibility, MD snapshots	`references/ensemble-docking.md`

Key Tools

Tool	Install	Role
`vina`	`conda install -c conda-forge autodock-vina`	Docking engine (empirical scoring)
`gnina`	prebuilt binary or Docker	CNN scoring function
`pdbfixer`	`conda install -c conda-forge pdbfixer`	PDB cleaning, H addition, missing residues
`openbabel`	`conda install -c conda-forge openbabel`	Format conversion → PDBQT
`prolif`	`pip install prolif`	Protein-Ligand Interaction Fingerprints
`fpocket`	`conda install -c conda-forge fpocket`	Pocket detection / box definition
`propka`	`pip install propka`	pKa prediction for protonation
`meeko`	`pip install meeko`	Ligand PDBQT prep (better than obabel for Vina)

Scoring Function Reference

Method	Score type	Precision	Speed	Use case
Vina empirical	ΔG (kcal/mol)	★★★	★★★★★	VS, binding mode
Gnina CNN	unitless affinity	★★★★	★★★★	Rescoring, pose selection
MM-GB/SA	ΔG (kcal/mol)	★★★★	★★	Lead opt rescoring
FEP/TI	ΔΔG (kcal/mol)	★★★★★	★	Precise relative ranking

Installation

# Vina + OpenBabel (required)
conda install -c conda-forge autodock-vina openbabel

# pdbfixer + propka (receptor prep)
conda install -c conda-forge pdbfixer
pip install propka

# ProLIF (pose analysis)
pip install prolif

# meeko (better ligand PDBQT prep)
pip install meeko

# fpocket (pocket detection)
conda install -c conda-forge fpocket

# Gnina (CNN rescoring) — prebuilt binary
wget https://github.com/gnina/gnina/releases/latest/download/gnina
chmod +x gnina && mv gnina ~/.local/bin/

# Verify Vina
vina --version   # AutoDock Vina 1.2.x

Related Skills

force-fields → MM-GB/SA rescoring after docking
mdanalysis → generate conformational ensemble for ensemble docking
homology-modeling → build receptor when no crystal structure available
pharmacophore → pharmacophore-constrained docking, pose validation
free-energy → FEP/TI for accurate ΔΔG after docking hit identification
py3Dmol → 3D visualization of poses inline
Scripts: chem_filter.py --lipinski → pre-filter library before VS
Scripts: chem_3d.py → generate 3D conformers for ligand prep