synthetic-sciences/openscience/multi-objective-optimization

Multi-Objective Molecular Optimization

Overview

Real drug design is never single-objective. A useful molecule must simultaneously satisfy potency, selectivity, solubility, metabolic stability, and safety constraints. This skill implements Pareto-aware optimization that balances multiple properties without collapsing to a single weighted score.

Based on:

• MultiMol (Yu et al., 2025): 82.3% multi-objective success rate with generate-then-rank
• MOLLM (Ran et al., 2025): LLMs as genetic operators for multi-objective molecular design
• DrugR (Liu et al., 2026): Multi-granular reward balancing across property groups

When to Use This Skill

• "Improve potency while keeping hERG safe" — classic multi-objective lead optimization
• Balancing ADMET tradeoffs — LogP vs solubility, BBB penetration vs peripheral safety
• Pareto analysis — identify which candidates best balance competing objectives
• Property-constrained generation — generate molecules within a defined property box

Do NOT use this skill for:

• Single-property optimization (use molecular-optimization)
• Property prediction without optimization (use admet-prediction)

Related Skills

• molecular-optimization: Single-objective iterative optimization
• admet-prediction: Compute properties used as objectives
• admet-reasoning: Understand why properties need improvement

Installation

pip install rdkit-pypi numpy pandas

Optional

pip install matplotlib  # For Pareto front visualization

Core Workflows

1. Multi-Objective Optimization

python scripts/pareto_optimize.py \
    --smiles "c1ccc(NC(=O)c2ccccc2Cl)cc1" \
    --objectives "LogP:minimize:3.0,QED:maximize:0.5,TPSA:range:20:130" \
    --candidates 16 \
    --output pareto_results.json

2. Pareto Analysis of Existing Candidates

python scripts/pareto_optimize.py \
    --input candidates.csv \
    --objectives "LogP:minimize:3.0,QED:maximize:0.5" \
    --mode analyze \
    --output pareto_front.json

3. Property Radar Plot

python scripts/property_radar.py \
    --reference "original_smiles" \
    --candidates optimized.csv \
    --output radar.png

Script Reference