Natural language-guided generation of molecular structures for single-molecule electronics.
Abstract
As molecular science advances toward autonomous scientific discovery, enabling AI models to directly generate valid molecular structures from natural language descriptions has emerged as a critical requirement. While large language models (LLMs) have demonstrated the ability to map discrete chemical space into continuous semantic representation for general molecular generation, their application to single-molecule electronics (SME) remains largely unexplored. This is primarily due to the specific topological requirements of the field, such as the necessity of an anchor-backbone-anchor motif for electrode connection. In this work, we propose SingMolT5, a fine-tuned molecular LLM for natural language-guided generation of single molecules. We systematically evaluated the structural generation performance using both in-context learning (ICL) and fine-tuning strategies based on a carefully curated domain-specific instruction-tuning dataset. Results show that while ICL can capture basic molecular syntax, only the fine-tuning model can reliably enforce the anchor-backbone-anchor motif required for single-molecule measurements. More importantly, SingMolT5 achieved high structural prediction accuracy, along with strong generalization to unseen structural variations not included in the training dataset. This work provides a proof-of-concept demonstration for a text-to-structure workflow in SME, serving as a foundational step toward integrating natural language models into future autonomous workflows for single-molecule research.