MolEvolve: LLM-Guided Evolutionary Search for Interpretable Molecular Optimization

2026-03-25

AI Summary

This work addresses the poor interpretability of deep learning in molecular optimization and its difficulty in handling activity cliffs—abrupt changes in molecular properties induced by minor structural modifications—by proposing a large language model (LLM)-guided autonomous forward-planning framework. The approach formulates molecular discovery as a symbolic evolutionary search problem, leveraging the LLM to automatically generate an executable library of chemical operations and integrating Monte Carlo Tree Search (MCTS) with cheminformatics tools like RDKit for on-the-fly planning. For the first time, it enables the generation of transparent, human-readable chemical reasoning chains without manual feature engineering, effectively capturing the discontinuities in structure–activity relationships. Experimental results demonstrate that the method outperforms existing baselines in molecular property optimization while providing interpretable optimization pathways and mechanistic insights.

Technology Category

• Search and Optimization: Learning to Search
• Planning, Routing, and Scheduling: Planning with Language Models
• Machine Learning: Optimization

Application Category

• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions

Abstract

Despite deep learning's success in chemistry, its impact is hindered by a lack of interpretability and an inability to resolve activity cliffs, where minor structural nuances trigger drastic property shifts. Current representation learning, bound by the similarity principle, often fails to capture these structural-activity discontinuities. To address this, we introduce MolEvolve, an evolutionary framework that reformulates molecular discovery as an autonomous, look-ahead planning problem. Unlike traditional methods that depend on human-engineered features or rigid prior knowledge, MolEvolve leverages a Large Language Model (LLM) to actively explore and evolve a library of executable chemical symbolic operations. By utilizing the LLM to cold start and an Monte Carlo Tree Search (MCTS) engine for test-time planning with external tools (e.g. RDKit), the system self-discovers optimal trajectories autonomously. This process evolves transparent reasoning chains that translate complex structural transformations into actionable, human-readable chemical insights. Experimental results demonstrate that MolEvolve's autonomous search not only evolves transparent, human-readable chemical insights, but also outperforms baselines in both property prediction and molecule optimization tasks.

Problem

• interpretability
• activity cliffs
• molecular optimization
• representation learning
• structure-activity relationship

Innovation

• LLM-guided evolution
• interpretable molecular optimization
• Monte Carlo Tree Search
• activity cliffs
• symbolic chemical operations