MolReAct: LLM-Guided Framework for Synthesizable Drug Lead Optimization
MolReAct represents an innovative framework aimed at enhancing lead optimization during drug discovery by merging Large Language Models (LLMs) with synthesis-constrained action spaces. This method treats lead optimization as a Markov Decision Process utilizing validated reaction templates, ensuring that molecular alterations align with practical synthetic pathways. An LLM agent, equipped with tools, functions as a responsive reaction environment, employing specialized chemical analysis instruments to pinpoint reactive sites and functional groups, subsequently suggesting concise sets of chemically valid transformations. A specific policy model is developed through Group Relative Policy Optimization. This framework overcomes the shortcomings of current techniques that either focus on property scores without ensuring synthesizability or depend on costly enumeration within extensive reaction networks, while steering clear of chemically invalid structures typical in direct LLM-based molecular generation.
Key facts
- MolReAct is a framework for lead optimization in drug discovery.
- It uses a Markov Decision Process over synthesis-constrained action spaces.
- Action spaces are defined by validated reaction templates.
- A tool-augmented LLM agent serves as a dynamic reaction environment.
- The agent invokes specialized chemical analysis tools to identify reactive sites and functional groups.
- The policy model is trained via Group Relative Policy Optimization.
- Existing approaches either prioritize property scores without enforcing synthesizability or rely on expensive enumeration over large reaction networks.
- Direct application of LLMs to molecular generation frequently produces chemically invalid structures.
Entities
—