brycewang-stanford/madd-drug-discovery-guide

MADD: Multi-Agent Drug Discovery Guide

Overview

MADD (Multi-Agent Drug Discovery) is a multi-agent system that automates key stages of the drug discovery pipeline — target identification, molecule generation, property prediction (ADMET), docking simulation, and lead optimization. Specialized agents collaborate to propose, evaluate, and refine drug candidates, reducing the manual effort in early-stage drug discovery research.

Agent Pipeline

Target Protein
      ↓
  Target Analysis Agent (binding site, druggability)
      ↓
  Molecule Generation Agent (de novo design)
      ↓
  Property Prediction Agent (ADMET screening)
      ↓
  Docking Agent (binding affinity estimation)
      ↓
  Optimization Agent (lead optimization cycle)
      ↓
  Report Agent (candidate ranking + rationale)

Usage

from madd import DrugDiscoveryPipeline

pipeline = DrugDiscoveryPipeline(
    llm_provider="anthropic",
    tools=["rdkit", "autodock_vina", "admet_predictor"],
)

# Run discovery pipeline
results = pipeline.discover(
    target_protein="6LU7",  # PDB ID (SARS-CoV-2 Mpro)
    target_site="active_site",
    constraints={
        "molecular_weight": (200, 500),    # Lipinski
        "logP": (-0.4, 5.6),
        "hbd": (0, 5),
        "hba": (0, 10),
        "tpsa": (0, 140),
    },
    num_candidates=100,
    optimization_rounds=3,
)

# Top candidates
for i, mol in enumerate(results.top_candidates[:5]):
    print(f"\nCandidate {i+1}: {mol.smiles}")
    print(f"  Docking score: {mol.docking_score:.2f} kcal/mol")
    print(f"  QED: {mol.qed:.3f}")
    print(f"  Synthetic accessibility: {mol.sa_score:.2f}")
    print(f"  ADMET: {mol.admet_summary}")

ADMET Prediction

from madd.agents import ADMETAgent

admet = ADMETAgent()

# Predict ADMET properties for a molecule
props = admet.predict("CC(=O)Oc1ccccc1C(=O)O")  # Aspirin

print(f"Absorption: {props.absorption}")
print(f"Distribution: {props.distribution}")
print(f"Metabolism: {props.metabolism}")
print(f"Excretion: {props.excretion}")
print(f"Toxicity: {props.toxicity}")
print(f"BBB penetration: {props.bbb_penetration}")
print(f"CYP inhibition: {props.cyp_inhibition}")
print(f"hERG liability: {props.herg_risk}")

Molecule Generation

from madd.agents import MolGenAgent

gen = MolGenAgent(method="reinforcement_learning")

# Generate molecules targeting a binding site
molecules = gen.generate(
    target_pdb="6LU7",
    binding_site="active_site",
    num_molecules=500,
    diversity_threshold=0.5,  # Tanimoto diversity
    constraints={
        "drug_likeness": True,  # Lipinski + Veber
        "novelty": True,        # Not in ChEMBL
    },
)

print(f"Generated: {len(molecules)}")
print(f"Drug-like: {sum(1 for m in molecules if m.is_drug_like)}")
print(f"Novel: {sum(1 for m in molecules if m.is_novel)}")

Lead Optimization

from madd.agents import OptimizationAgent

optimizer = OptimizationAgent()

# Optimize a lead compound
optimized = optimizer.optimize(
    lead_smiles="c1ccc(-c2ncc(F)c(N)n2)cc1",
    objectives=[
        ("docking_score", "minimize"),
        ("qed", "maximize"),
        ("sa_score", "minimize"),
        ("solubility", "maximize"),
    ],
    num_iterations=50,
    keep_scaffold=True,  # Maintain core structure
)

for mol in optimized.pareto_front[:5]:
    print(f"SMILES: {mol.smiles}")
    print(f"  Docking: {mol.docking_score:.2f}")
    print(f"  QED: {mol.qed:.3f}")

Use Cases

1. Hit discovery: Generate novel drug candidates for targets
2. Lead optimization: Improve properties of promising compounds
3. ADMET screening: Predict pharmacokinetic properties
4. Virtual screening: Score large molecule libraries
5. Drug repurposing: Evaluate known drugs for new targets

References

• MADD GitHub
• RDKit — Chemistry toolkit
• AutoDock Vina — Molecular docking