A Multi-Modal Large Language Model for Free-Form, Open-Ended, and Interactive Prediction of Properties and Mechanisms of Candidate Drug Molecules

ICML 2025 Workshop FM4LS Submission74 Authors

Keywords: Drug discovery, molecule property prediction, QSAR, multi-modal large language model

TL;DR: DrugChat generates free-form predictions on various molecular properties with strong zero-shot generalization

Abstract: Accurately predicting the mechanisms and properties of candidate drug molecules is critical for advancing drug discovery. However, existing models are often limited to structured outputs, fixed task sets, and static, one-shot predictions. We present DrugChat, a multi-modal large language model that addresses these limitations through three key capabilities: (i) free-form text generation for predicting complex drug attributes such as indications, pharmacodynamics, and mechanisms of action; (ii) generalization to an open-ended set of tasks via prompt-based multi-task learning; and (iii) interactive, multi-turn dialogue for dynamic exploration for molecules. DrugChat integrates a molecular graph encoder, a molecular image encoder, and an instruction-tuned large language model. Pretrained on 248 million molecule-bioactivity records, DrugChat outperforms existing baselines across both unstructured and structured tasks, demonstrating strong zero-shot generalization.

Submission Number: 74