Drug Docking Analysis

drug-docking-analysis

Goal

To analyze virtual screening docking results by computing score distributions, ligand efficiency metrics, and (when labeled actives/inactives are available) enrichment statistics. This skill sits between drug-docking-vina and downstream refinement stages, providing quantitative assessment of docking campaign quality.

Important: this skill does not assess pose quality. A compound can receive an excellent Vina score with a physically implausible pose (internal clashes, strained torsions, mis-assigned bond orders). Always pair this analysis with drug-pose-validation (PoseBusters) before acting on the top-ranked compounds.

Outputs:

• Score distribution KDE plot
• Score vs. molecular weight scatter (visualizes Vina's size/lipophilicity bias)
• Ligand efficiency (LE, BEI, SEI) distributions
• ROC curve with AUC (when labels available)
• Enrichment factor bar chart at 1%, 2%, 5%, 10%, 20% (when labels available)
• Enriched results CSV with per-compound efficiency metrics

Instructions

0. Prerequisites: produce a docking_ranked.csv

This skill expects a ranked CSV produced by drug-docking-vina's collect_results.py. If you are starting from raw drug-docking-vina JSON output, run the collect step first:

${CLAUDE_SKILL_DIR}/../../venv/run cpu python ${CLAUDE_SKILL_DIR}/../drug-docking-vina/scripts/collect_results.py \
  --results docking/results/docking_results.json \
  --library_csv library/library_master.csv \
  --output_dir docking/analysis/

The collect step joins the docking scores with the library CSV to pull SMILES, labels, and (when present) parent_compound_id / microstate_id columns. See drug-docking-vina SKILL.md step 5 for details.

1. Basic analysis (no labels)

When you have docking results but no active/inactive labels:

${CLAUDE_SKILL_DIR}/../../venv/run cpu python ${CLAUDE_SKILL_DIR}/scripts/analyze_docking.py \
  --docking_csv docking/docking_ranked.csv \
  --output_dir docking/analysis/

This produces score KDE, score vs. MW, and ligand efficiency plots.

2. With enrichment analysis (labeled library)

When your library has known actives and inactives:

${CLAUDE_SKILL_DIR}/../../venv/run cpu python ${CLAUDE_SKILL_DIR}/scripts/analyze_docking.py \
  --docking_csv docking/docking_ranked.csv \
  --library_csv library/library_master.csv \
  --active_label active \
  --inactive_label inactive \
  --output_dir docking/analysis/

The --library_csv must have compound_id and label columns. If labels are already in the docking CSV, the library CSV is not needed.

3. Aggregate microstates before ranking

If the library contained enumerated protomers/tautomers (each parent compound appearing multiple times under different compound_id values), you must collapse microstates back to best-score-per-parent before computing any ranking metric. Without aggregation, compounds with more enumerated forms get extra chances to rank high and inflate the apparent library size, biasing enrichment.

Pass --parent_id_col parent_compound_id to opt in:

${CLAUDE_SKILL_DIR}/../../venv/run cpu python ${CLAUDE_SKILL_DIR}/scripts/analyze_docking.py \
  --docking_csv docking/docking_ranked.csv \
  --parent_id_col parent_compound_id \
  --output_dir docking/analysis/

collect_results.py propagates parent_compound_id and microstate_id from the library CSV automatically when those columns exist, so if you prepared ligands with drug-ligand-prep in its enumeration mode you should use this flag.

See examples/README.md for a side-by-side comparison of aggregated vs. unaggregated analysis on a small synthetic library.

4. Interpret results

Score distribution: A healthy screen typically shows a unimodal distribution with the bulk of scores in the -6 to -8 kcal/mol range and a tail extending toward -9 or beyond; the tail is where you look for hits. A bimodal distribution often means the library contains structurally distinct subsets (e.g. actives + fillers with very different MW profiles) and should be investigated before ranking.

Score vs. MW: Vina scores are biased toward larger, more lipophilic molecules. This plot reveals the bias. If actives cluster in a different MW range than inactives, the enrichment may be driven by size rather than binding complementarity, and you should re-rank by a size-normalized metric (LE or BEI) or apply an MW cutoff.

Ligand efficiency (LE): LE = -score / heavy_atom_count. Normalizes for molecular size. LE > 0.3 kcal/mol/HA is a common rule-of-thumb threshold for drug-like efficiency, but note that this threshold was derived from calibrated experimental binding free energies (Hopkins et al., 2004), not from Vina scores. Vina scores correlate with binding affinity but are not on the same kcal/mol scale as true free energies, so applying the 0.3 threshold directly to Vina-derived LE is common practice but not rigorously justified. Treat the dashed line on the plot as a visual reference, not a hard cutoff. BEI (-score*1000/MW) and SEI (-score*1000/TPSA) provide alternative normalizations; TPSA here is RDKit's Ertl 2D method via Descriptors.TPSA.

Enrichment: ROC AUC > 0.7 indicates decent discrimination. EF1% > 5x indicates meaningful early enrichment. Interpret in context: AUC and EF depend heavily on the decoy set composition, so report the decoy source alongside the metrics and prefer relative comparisons between protocol variants over absolute enrichment claims.

Pose quality: This skill reports score-based metrics only. None of these plots tell you whether the docked poses are physically reasonable. Always run drug-pose-validation (PoseBusters) on the top-ranked compounds before committing them to MD or experimental validation; a great score with a clashing or strained pose is a false positive waiting to happen.