A New Language for AI-Driven Chemistry

A team led by Jiacheng Xiong introduces ReactSeq, a novel reaction description language designed to enhance the way artificial intelligence interprets and models chemical reactions. Unlike traditional representations that merely list reactants and products, ReactSeq encodes detailed atomic and bond-level transformations, providing a step-by-step view of chemical processes. This richer encoding significantly boosts the performance of language models on retrosynthesis benchmarks and enables more interpretable, human-in-the-loop AI applications in chemistry. Beyond synthesis prediction, ReactSeq facilitates exploration of chemical reaction space and improves tasks like yield prediction and experimental planning. The authors position ReactSeq as a crucial bridge, aligning the symbolic rigor of chemistry with the data-hungry capabilities of modern AI.