fadrienne/apps/molecular-optimization

Molecular Optimization

Overview

Lead optimization is the bottleneck of drug discovery — modifying a hit compound to improve potency, selectivity, and ADMET properties without breaking what already works. LLMs frequently generate invalid SMILES or propose modifications that don't appear in the actual structure.

This skill implements an iterative optimization protocol based on three peer-reviewed approaches:

• MT-Mol (Kim et al., 2025): Multi-agent tool-based reasoning with verification — SOTA on 17/23 PMO benchmark tasks
• DrugR (Liu et al., 2026): Explicit liability reasoning before generation — 18× improvement over blind generation
• MultiMol (Yu et al., 2025): Generate-then-rank with scaffold preservation — 82.3% multi-objective success rate

The core loop: Analyze → Identify Liabilities → Generate Candidates → Verify → Evaluate & Rank → Iterate.

When to Use This Skill

• Lead optimization: Improve ADMET properties of a hit while preserving potency
• Scaffold hopping: Find new scaffolds that maintain key pharmacophoric features
• Property-driven design: Generate analogs targeting specific property improvements (lower LogP, reduce hERG, improve solubility)
• Multi-objective optimization: Balance multiple properties simultaneously

Do NOT use this skill for:

• De novo design from scratch (use denovo-design instead)
• Simple property prediction without optimization (use admet-prediction)
• Docking or binding affinity estimation (use molecular-docking, binding-affinity)

Related Skills

• admet-prediction: Compute ADMET properties (this skill uses it internally)
• admet-reasoning: Interpretable ADMET analysis with mechanistic explanations
• smiles-validation: Strict SMILES parsing and structural verification
• rdkit: Core molecular operations
• medchem: Medicinal chemistry filters and transformations

Installation

Required dependencies

pip install rdkit-pypi numpy pandas

Optional dependencies

pip install PyTDC datamol

• PyTDC: Access to TDC ADMET predictors for enhanced property scoring
• datamol: Convenient molecular manipulation utilities

Core Workflows

1. Single-Molecule Optimization

Optimize one molecule for improved properties:

python scripts/optimize.py \
    --smiles "c1ccc(NC(=O)c2ccccc2Cl)cc1" \
    --targets "LogP<3,hERG<0.3,QED>0.5" \
    --max-iterations 3 \
    --candidates 8 \
    --output results.json

2. Batch Optimization

Optimize a CSV of molecules:

python scripts/optimize.py \
    --input leads.csv \
    --smiles-col SMILES \
    --targets "LogP<3,hERG<0.3" \
    --output optimized.csv

3. Verification Only

Verify a proposed SMILES matches a claimed modification:

python scripts/verify_smiles.py \
    --original "c1ccccc1" \
    --proposed "c1ccc(O)cc1" \
    --claimed-modification "Added hydroxyl group at para position"

4. Candidate Comparison

Compare multiple candidates against a reference:

python scripts/compare_candidates.py \
    --reference "c1ccc(NC(=O)c2ccccc2Cl)cc1" \
    --candidates candidates.csv \
    --output comparison.html

Script Reference

Optimization Protocol Detail

Step 1: ANALYZE

Compute molecular descriptors: MW, LogP, TPSA, HBA, HBD, RotBonds, QED, aromatic rings, Murcko scaffold. Flag properties outside ADMET target thresholds.

Step 2: IDENTIFY LIABILITIES

Rank flagged properties by severity (hERG > DILI > CYP > solubility). For each, identify the structural feature causing the liability and propose a specific modification.

Step 3: GENERATE CANDIDATES

Apply proposed modifications via:

• Bioisosteric replacement (e.g., phenyl → pyridine, amide → sulfonamide)
• Functional group addition/removal
• Ring system modification
• Chain length adjustment

Generate 4-8 candidates per iteration.

Step 4: VERIFY

For each candidate:

1. Chem.MolFromSmiles() — discard if None
2. Scaffold preservation: Murcko scaffold match
3. Tanimoto similarity (ECFP4): flag if < 0.4
4. Structural verification: confirm claimed modification exists

Step 5: EVALUATE & RANK

Recompute descriptors, build comparison table, score by net liability improvement (+1 per fix, -0.5 per new liability).

Step 6: ITERATE

If no improvement after 3 iterations, return best found with honest assessment.

ADMET Target Thresholds