Large Language Model Assistant for Molecular Design

The design of organic molecules is pivotal in advancing research and innovation within chemistry. Traditionally, the development of molecules with targeted properties has depended heavily on the expertise of chemists engaging in prolonged trial-and-error processes. Although computational molecular design methods have emerged, their full potential remains underutilized, largely due to difficulties in integrating experimental feedback and empirical expertise.

The ideal approach is a collaborative integration of human expertise with computational strategies. Large Language Models (LLMs), powerful artificial intelligence systems trained on extensive datasets, have recently emerged as promising facilitators for this human-machine collaboration. These models excel at producing human-like interactions across a variety of domains, enabling intuitive and dynamic exchanges between chemists and computational tools.

In this work, the research team demonstrates the potential of general-purpose LLMs in facilitating targeted molecular design through a case study involving the prediction of organic structure-directing agents (OSDAs) for zeolites. Zeolites are crystalline microporous materials whose synthesis heavily relies on selecting suitable OSDAs to fit their internal cavities. By incorporating physicochemical affinity scores and empirical chemical insights into interactions via natural language prompts, the team developed a novel molecular design workflow powered by LLMs.

This innovative computational workflow exemplifies effective human-AI collaboration, showcasing significant predictive accuracy and reliability. Remarkably, the approach identified experimentally validated OSDAs, closely related analogs, and entirely novel yet promising candidate molecules, underscoring its potential to revolutionize molecular design strategies.

Interactive Molecular Design Using Large Language Models

Papers
Journal: Chemistry of Materials
Title: Knowledge-Informed Molecular Design for Zeolite Synthesis Using General-Purpose Pretrained Large Language Models Toward Human-Machine Collaboration
Authors: Shusuke Ito, Koki Muraoka*, Akira Nakayama*
DOI: 10.1021/acs.chemmater.4c02726