Bridging chemistry and artificial intelligence by a reaction description language

With the fast-paced development of artificial intelligence, large language models are increasingly used to tackle various scientific challenges.

With the rapid development of artificial intelligence (AI), large language models (LLMs) are increasingly being used to address various scientific challenges. A crucial step in this process is converting domain-specific data into a format suitable for LLMs, typically a sequence of tokens. In chemistry, molecules are commonly represented by molecular linear notations, and chemical reactions are depicted as pairs of reactants and products. However, this approach does not capture the atomic and bond changes that occur during reactions, which are essential for chemical understanding and manipulation. To bridge this gap and facilitate seamless integration between chemistry and LLMs, we introduce ReactSeq, a reaction description language that decomposes chemical reactions into a series of molecular editing operations.

Each ReactSeq token corresponds to a specific atomic or bond modification, enabling a step-by-step unfolding of the chemical transformation. We trained a language model for retrosynthesis prediction using ReactSeq and observed that it consistently outperformed existing methods in all benchmark tests. Furthermore, the model demonstrated promising emergent abilities, such as performing multistep synthesis planning in response to user requests and providing explanations for its predictions. To delve deeper into the capabilities of LLMs in navigating chemical space, we trained a model to predict reaction yield based on ReactSeq representations and achieved high performance in this task.

Our analysis indicates that the model learned to evaluate the feasibility of reactions based on chemical principles, highlighting the potential of LLMs to go beyond empirical patterns and develop a chemical understanding of the data. Finally, we used ReactSeq to generate universal and reliable representations of chemical reactions, facilitating efficient retrieval of relevant experimental procedures from literature databases. This capability paves the way for seamless integration between theoretical predictions and experimental observations, ultimately advancing chemical discovery and invention.