YieldFCP: Enhancing Reaction Yield Prediction via Fine-grained Cross-modal Pre-training

Predicting chemical reaction yields is a critical yet challenging task in organic chemistry. While integrating multi-modal information has shown promise, existing methods typically encode the entire reaction in different modalities and then align these embeddings for the same reactions. Such a coarse-grained modal fusion strategy may neglect atomic-level interactions crucial for accurate predictions. Recognizing the crucial role of modal fusion in multi-modal learning and the limitations of current methods in real-world scenarios, we propose YieldFCP, a reaction $\underline{\text{Yield}}$ prediction model based on $\underline{\text{F}}$ine-grained $\underline{\text{C}}$ross-modal $\underline{\text{P}}$re-training. Its cross-modal projector links the molecular SMILES sequence with 3D geometric data, focusing on the atomic-level interactions to achieve fine-grained modal fusion and enhance yield prediction. YieldFCP is pre-trained on a large-scale dataset leveraging cross-modal self-supervised learning techniques. Experimental results on the high-throughput experiments, real-world electronic laboratory notebook, and real-world organic reaction publication datasets demonstrate the effectiveness of our approach. Particularly, YieldFCP outperforms the state-of-the-art methods in real-world scenarios and successfully recognizes key components that determine reaction yields with valuable interpretability.