ToolMol: AI drug discovery framework improves binding affinity by 35%
An LLM agent with evolutionary algorithms designs better drug molecules, boosting binding by 35%.
A new AI framework called ToolMol, developed by Andrew Zhou and colleagues at UC San Diego, merges large language models with evolutionary algorithms to accelerate drug discovery. The system treats molecular generation as a multi-objective optimization problem, using an agentic LLM that calls RDKit-backed functions (e.g., bond addition, substitution) to make targeted modifications to ligand candidates. These modifications are guided by a genetic algorithm that evolves the population toward desired properties like binding affinity and synthesizability.
On benchmarks against three protein targets, ToolMol outperformed existing methods by over 10% in predicted binding affinity and over 35% in gold-standard absolute binding free energy calculations. Analysis of chain-of-thought traces revealed that tool-calling allowed the LLM to faithfully execute planned modifications, effectively leveraging its chemical knowledge. The work, published on arXiv, demonstrates how combining LLM agents with evolutionary search can produce more valid and potent drug-like molecules.
- ToolMol combines a multi-objective genetic algorithm with an LLM agent using RDKit-based tools for precise molecular edits.
- Achieves >10% improvement in predicted binding affinity and >35% better absolute binding free energy scores over prior methods.
- Evaluated on three protein targets, producing drug-like and synthesizable ligands with high validity.
Why It Matters
This framework could drastically shorten drug candidate design cycles by generating more potent, synthesizable molecules faster.