Transforming Molecular Science With Large Language Models: From Molecule Understanding to Autonomous Scientific Discovery

Large language models (LLMs) are driving a paradigm shift in molecular science, transitioning from molecule understanding to autonomous scientific discovery. By learning from multimodal molecular representations through advanced training techniques, LLMs address traditional limitations in molecular research, including expert dependency and limited experimental scalability. We analyze strategies for cross-modal alignment and domain adaptation that enable LLMs to interpret complex molecular semantics. This deep semantic understanding translates into broad versatility across key downstream applications, encompassing retrieval tasks like molecule-text matching, prediction tasks like reaction outcome inference, and generative tasks including zero-shot molecule design. Crucially, we highlight how LLMs integrate with robotic platforms to establish closed-loop autonomous discovery systems, where AI agents automate hypothesis generation, experimental planning, and iterative validation. While these advances accelerate exploration of molecular space, persistent challenges remain in scaling experimental validation and bridging symbolic reasoning with physical experimentation. This review provides a comprehensive roadmap for leveraging LLMs to redefine scientific discovery in molecular science.