PharmaClaw Toxicology Agent

Overview

Predictive toxicology and drug safety profiling agent for the PharmaClaw pipeline. Screens drug candidates via RDKit descriptors, rule-based filters (Lipinski Ro5, Veber), QED scoring, and PAINS alerts to flag safety risks early in the discovery process.

Quick Start

# Analyze a compound
python scripts/tox_agent.py "CC(=O)Nc1ccc(O)cc1"

# Default (ethanol)
python scripts/tox_agent.py

Capabilities

Check	Method	Threshold
Lipinski Ro5	MW, LogP, HBD, HBA	MW>500, LogP>5, HBD>5, HBA>10
Veber Rules	TPSA, Rotatable Bonds	TPSA>140, RotB>10
QED Score	RDKit QED module	0-1 (higher = more drug-like)
PAINS Alerts	Substructure matching	Known assay interference patterns
Ring Analysis	Aromatic/total ring count	Complexity indicator

Decision Tree

• SMILES input → RDKit descriptor calculation → Rule-based screening
• Lipinski violations = 0 AND PAINS = 0 → Risk: Low
• Any violations or PAINS hits → Risk: Medium/High
• High risk flagged → Recommend analogs via Chemistry Query / IP Expansion

Output Format

{
  "lipinski_viol": 0,
  "veber_viol": 0,
  "qed": 0.737,
  "pains": 0,
  "risk": "Low",
  "props": {
    "mw": 151.2,
    "logp": 1.02,
    "tpsa": 49.3,
    "hbd": 2,
    "hba": 2,
    "rotb": 1,
    "rings": 1,
    "arom": 1
  }
}

Risk Classification

• Low: No Lipinski violations, no PAINS alerts, QED > 0.5
• Medium: 1-2 Lipinski violations OR low QED
• High: 3+ Lipinski violations, PAINS hits, or multiple Veber violations

Chain Integration

• Receives from: Chemistry Query (SMILES), Pharmacology (ADME flags)
• Feeds into: IP Expansion (safer derivative suggestions), Synthesis (avoid toxic intermediates)
• Cross-references: Market Intel (FAERS adverse events for similar structures)

Dependencies

• rdkit-pypi — Molecular descriptors, QED, substructure matching

Scripts

• scripts/tox_agent.py — Main agent: ToxAgent class with analyze(smiles) method

Limitations

• PAINS screening uses simplified substructure set (production should use full PAINS catalog)
• No Ames mutagenicity or hERG channel prediction (descriptor-based proxies planned)
• LD50 estimation not yet implemented (QSAR model planned for future version)