Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LLM-Augmented Chemical Synthesis and Design Decision Programs

Haorui Wang, Jeff Guo, Lingkai Kong, Rampi Ramprasad, Philippe Schwaller, Yuanqi Du, Chao Zhang

TL;DR

The paper investigates leveraging large language models to perform constrained retrosynthesis planning and synthesizable molecular design. It introduces a route-centric encoding and two complementary paradigms: (i) LLMs as single-step predictors coupled with search, and (ii) LLMs as autonomous synthesis-route samplers governed by an evolutionary framework (LLM-Syn-Planner), with retrieval-augmented initialization, three-level evaluation, and partial rewards. It further extends to synthesizable molecular design through LLM-Syn-Designer, which uses MolLEO to optimize molecules under synthesizability constraints. Across USPTO and Pistachio datasets, the route-generation approach matches or surpasses traditional baselines in retrosynthesis planning and demonstrates effective balancing of fitness with synthesis feasibility, illustrating the potential of LLMs as decision-making engines in chemistry. The work highlights the importance of route representations, partial reward signals, and in-context exemplars, suggesting a shift toward full-route planning and evolutionary optimization for practical chemical discovery.

Abstract

Retrosynthesis, the process of breaking down a target molecule into simpler precursors through a series of valid reactions, stands at the core of organic chemistry and drug development. Although recent machine learning (ML) research has advanced single-step retrosynthetic modeling and subsequent route searches, these solutions remain restricted by the extensive combinatorial space of possible pathways. Concurrently, large language models (LLMs) have exhibited remarkable chemical knowledge, hinting at their potential to tackle complex decision-making tasks in chemistry. In this work, we explore whether LLMs can successfully navigate the highly constrained, multi-step retrosynthesis planning problem. We introduce an efficient scheme for encoding reaction pathways and present a new route-level search strategy, moving beyond the conventional step-by-step reactant prediction. Through comprehensive evaluations, we show that our LLM-augmented approach excels at retrosynthesis planning and extends naturally to the broader challenge of synthesizable molecular design.

LLM-Augmented Chemical Synthesis and Design Decision Programs

TL;DR

The paper investigates leveraging large language models to perform constrained retrosynthesis planning and synthesizable molecular design. It introduces a route-centric encoding and two complementary paradigms: (i) LLMs as single-step predictors coupled with search, and (ii) LLMs as autonomous synthesis-route samplers governed by an evolutionary framework (LLM-Syn-Planner), with retrieval-augmented initialization, three-level evaluation, and partial rewards. It further extends to synthesizable molecular design through LLM-Syn-Designer, which uses MolLEO to optimize molecules under synthesizability constraints. Across USPTO and Pistachio datasets, the route-generation approach matches or surpasses traditional baselines in retrosynthesis planning and demonstrates effective balancing of fitness with synthesis feasibility, illustrating the potential of LLMs as decision-making engines in chemistry. The work highlights the importance of route representations, partial reward signals, and in-context exemplars, suggesting a shift toward full-route planning and evolutionary optimization for practical chemical discovery.

Abstract

Retrosynthesis, the process of breaking down a target molecule into simpler precursors through a series of valid reactions, stands at the core of organic chemistry and drug development. Although recent machine learning (ML) research has advanced single-step retrosynthetic modeling and subsequent route searches, these solutions remain restricted by the extensive combinatorial space of possible pathways. Concurrently, large language models (LLMs) have exhibited remarkable chemical knowledge, hinting at their potential to tackle complex decision-making tasks in chemistry. In this work, we explore whether LLMs can successfully navigate the highly constrained, multi-step retrosynthesis planning problem. We introduce an efficient scheme for encoding reaction pathways and present a new route-level search strategy, moving beyond the conventional step-by-step reactant prediction. Through comprehensive evaluations, we show that our LLM-augmented approach excels at retrosynthesis planning and extends naturally to the broader challenge of synthesizable molecular design.
Paper Structure (40 sections, 3 equations, 4 figures, 6 tables, 1 algorithm)

This paper contains 40 sections, 3 equations, 4 figures, 6 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Problem Formulation
  3. Retrosynthesis Planning
  4. Synthesizable Molecular Design
  5. Methodology
  6. Route formatting
  7. LLM as a single-step prediction model
  8. LLM as a synthesis pathway sampler
  9. Optimization for synthesizable molecular design
  10. Experiments
  11. Experimental Setup
  12. Dataset.
  13. Baseline.
  14. Metrics.
  15. Configuration .
  16. ...and 25 more sections

Figures (4)

  • Figure 1: Overview of the LLM-Syn-Planner. 1. INITIALIZATION: Based on the target molecule, reaction routes of similar molecules are retrieved and scored by the SC score sc-score. 2. EVALUATION: The LLM generates new routes which are evaluated. 3. SELECTION: Starting from invalid steps in the reaction routes, the SC score of the molecules at this step are computed and the top $n_c$ routes are selected. 4. MUTATION: Starting from these invalid steps, the LLM proposes mutations to modify the molecules and/or reactions at this step. Repeat until a solution is found or the budget is reached.
  • Figure 2: Different route formats for retrosynthesis planning
  • Figure 3: Fitness score of the best molecule found by each molecule optimization method. Only LLM-Syn-Designer (GPT) here ensures the synthesizability of the found molecule.
  • Figure 4: Top 1 molecule of jnk3 found by LLM-Syn-Designer.