Research map / Large language models / Agents, code and math
Efficient Evolutionary Search Over Chemical Space with Large Language Models
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes MolLEO, which uses chemistry-aware LLMs (GPT-4, BioT5, MoleculeSTM) as crossover and mutation operators inside a Graph-GA evolutionary search to optimize molecules with fewer oracle calls.
Problem
Black-box molecular optimization with evolutionary algorithms is sample-inefficient because random crossover and mutation generate many poor candidates that require expensive oracle evaluations.
Approach
Replace random Graph-GA crossover and mutation with LLM-proposed edits conditioned on text descriptions of the objective, combined with selection by fitness and Tanimoto similarity to the top molecule, so that each oracle call is more likely to yield an improved candidate.
Key results
- MolLEO (GPT-4) achieves total top-10 AUC of 17.862 over 23 single-objective PMO tasks, ranking first and outperforming all baselines (Graph GA total 13.823).
- MolLEO (GPT-4) outperforms Graph GA on all three multi-objective tasks in hypervolume and summation metrics.
- MolLEO (BioT5) ranks second overall in single-objective PMO tasks with a total of 15.424, with open-source models showing gains over Graph GA.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Replace random Graph-GA crossover with LLM-based crossover (ablation) | MolLEO (GPT-4) · JNK3 | Top-10 AUC: 0.553 ± 0.136 -> 0.790 ± 0.027 | Helped |
| Use LLM mutation instead of random mutation with LLM crossover (ablation) | MolLEO (GPT-4) · JNK3 | Top-10 AUC: 0.790 ± 0.027 -> 0.794 ± 0.026 | Helped |
| Add selection of top molecules and distance-based offspring pruning to LLM mutation (ablation) | MolLEO (MolSTM) · JNK3 | Top-10 AUC: 0.631 ± 0.327 -> 0.730 ± 0.188 | Helped |
| Increase the selected top-scoring mutation candidates from 20 to 30 (ablation) | MolLEO (MolSTM) · JNK3 | Top-10 AUC: 0.680±0.213 -> 0.730±0.188 | Helped |
| Increase the selected top-scoring mutation candidates from 30 to 50 (ablation) | MolLEO (MolSTM) · JNK3 | Top-10 AUC: 0.730±0.188 -> 0.627±0.250 | Hurt |
| Add retrieval-augmented search using ZINC 250K molecules (ablation) | MolLEO (GPT-4) · JNK3 | Top-10 AUC: 0.790±0.027 -> 0.830±0.047 | Helped |
| Increase offspring count from 20 to 70 per generation (ablation) | MolLEO (GPT-4) · JNK3 | Top-10 AUC: 0.731±0.012 -> 0.790±0.027 | Helped |
| Increase offspring count from 70 to 200 per generation (ablation) | MolLEO (GPT-4) · JNK3 | Top-10 AUC: 0.790±0.027 -> 0.785±0.022 | Hurt |
| Use GPT-4 instead of GPT-3.5 (ablation) | MolLEO · JNK3 | Top-10 AUC: 0.669±0.104 -> 0.790±0.027 | Helped |
| Use GPT-4 instead of GPT-3.5 (ablation) | MolLEO · isomer_c9h10n2o2pf2cl | Top-10 AUC: 0.902±0.021 -> 0.874±0.053 | Hurt |
| Use Graph-GA rules instead of no rules (ablation) | MolLEO (GPT-4) · JNK3 | Top-10 AUC: 0.765±0.047 -> 0.790±0.027 | Helped |
| Replace Graph-GA rules with SMILES-GA rules (ablation) | MolLEO (GPT-4) · perindopril_mpo | Top-10 AUC: 0.600±0.031 -> 0.583±0.031 | Hurt |
Surprising findings
- Open-source LLMs (BioT5, MoleculeSTM) degrade when multiple objectives are included in the prompt.
- MolLEO (GPT-4) performs worse than Graph GA on the adenosine A2A receptor docking task.
Limitations
- All benchmarks are proxies for real chemical properties and may not reflect real-world performance.
- Open-source LLM performance degrades in multi-objective prompts.
- GPT-4 underperforms Graph GA on some docking tasks.
Experimental setup
Datasets: ZINC 250K PMO benchmark tasks TDC oracles ZINC20 PubChem
Benchmarks: PMO TDC AutoDock Vina docking tasks (DRD3, EGFR, A2A)
Metrics: AUC top-10 hypervolume docking score rank
Baselines: REINVENT Augmented Memory Graph GA GP BO DST
Builds on: Graph-GA PMO benchmark MoleculeSTM BioT5
Closest papers by meaning
- LLM-Augmented Chemical Synthesis and Design Decision Programs (ICML 2025 · similarity 0.89)
- Procedural Synthesis of Synthesizable Molecules (ICLR 2025 · similarity 0.88)
- From Artificially Real to Real: Leveraging Pseudo Data from Large Language Models for Low-Resource Molecule Discovery (AAAI 2024 · similarity 0.88)
- Searching for High-Value Molecules Using Reinforcement Learning and Transformers (ICLR 2024 · similarity 0.88)
- LICO: Large Language Models for In-Context Molecular Optimization (ICLR 2025 · similarity 0.88)
- Beyond Chemical QA: Evaluating LLM's Chemical Reasoning with Modular Chemical Operations (NeurIPS 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.