Adapting Language Models for Retrosynthesis Prediction

Retrosynthesis, a commonly used technique used in organic synthesis, involves deconstructing complex molecules into simpler precursors, thus enabling the efficient synthesis of target compounds. Recently, advancements in deep learning and large language models (LLMs) have offered a transformative approach to this domain. In this study, we presented a fine-tuning language model tailored explicitly for retrosynthesis prediction. By fine-tuning this model on extensive chemical reaction data (USPTO-50K), we successfully proposed appropriate retrosynthetic pathways for approximately 92% of cases in the test set. Through detailed case studies and step-by-step retrosynthesis predictions, our model demonstrated proficiency in identifying reaction types, centers, and reagents, albeit with occasional deviations that reflect its creative approach to reaction prediction. These findings highlight the model's potential as a valuable tool in organic synthesis planning.