Enhancing Chemical Reaction and Retrosynthesis Prediction with Large Language Model and Dual-task Learning

TL;DR

This work tackles two essential challenges in LLM-based chemical synthesis: data scarcity for instruction tuning and the bidirectional nature of reaction and retrosynthesis. It introduces ChemDual, an enhanced LLaMA-based framework that builds a $4.4$M BRICS-derived Molecule-Fragments dataset, employs a multi-scale tokenizer, and utilizes dual-task learning to jointly optimize forward and backward synthesis tasks. Empirical results on Mol-Instruction and USPTO-50K show state-of-the-art performance in both reaction and retrosynthesis prediction, with additional support from docking analyses that yield drug-like, protein-binding compounds. The study demonstrates the value of large-scale, domain-specific instruction data and bidirectional training in advancing chemically informed LLMs and driving design-oriented capabilities in drug discovery.

Abstract

Chemical reaction and retrosynthesis prediction are fundamental tasks in drug discovery. Recently, large language models (LLMs) have shown potential in many domains. However, directly applying LLMs to these tasks faces two major challenges: (i) lacking a large-scale chemical synthesis-related instruction dataset; (ii) ignoring the close correlation between reaction and retrosynthesis prediction for the existing fine-tuning strategies. To address these challenges, we propose ChemDual, a novel LLM framework for accurate chemical synthesis. Specifically, considering the high cost of data acquisition for reaction and retrosynthesis, ChemDual regards the reaction-and-retrosynthesis of molecules as a related recombination-and-fragmentation process and constructs a large-scale of 4.4 million instruction dataset. Furthermore, ChemDual introduces an enhanced LLaMA, equipped with a multi-scale tokenizer and dual-task learning strategy, to jointly optimize the process of recombination and fragmentation as well as the tasks between reaction and retrosynthesis prediction. Extensive experiments on Mol-Instruction and USPTO-50K datasets demonstrate that ChemDual achieves state-of-the-art performance in both predictions of reaction and retrosynthesis, outperforming the existing conventional single-task approaches and the general open-source LLMs. Through molecular docking analysis, ChemDual generates compounds with diverse and strong protein binding affinity, further highlighting its strong potential in drug design.

Figures (8)

• Figure 1: (a) The upper subgraph is an example of the similarity between fragments and reactants for a molecule "OCCCN1CCOCC1". The lower subgraph is the similarity distribution between fragments and reactants for all molecules. (b) Single-task learning paradigm (left sub-figure) and dual-task learning paradigm of ChemDual (right sub-figure). (c) The Exact Match score (%) on the chemical reaction prediction task by using or not using the dual-task learning.
• Figure 2: The overall of proposed ChemDual.
• Figure 3: Examples of instruction of fine-tuning dataset.
• Figure 4: Comparison of ablation experiments for reaction prediction and retrosynthesis prediction on Mol-Instruction.
• Figure 5: Visualizations of LLaMa and ChemDual using the t-SNE.