MoleCode: A graph-explicit molecular language for LLMs

Researchers have introduced MoleCode, a training-free, graph-explicit molecular language designed for large language models (LLMs). Unlike SMILES, which compresses molecular topology into linear strings, MoleCode represents atoms, bonds, branches, and rings as typed entities with persistent identifiers and explicit relations. This makes molecular structure directly readable, editable, and auditable within the language context, allowing LLMs to operate on structure rather than reconstructing it from syntax. The approach improves frontier LLMs across molecular reasoning, editing, generation, and analysis tasks, particularly when structural access is limiting, such as with unfamiliar molecules. The paper is available on arXiv.