Abstract
Accurately predicting chemical reaction yields can accelerate reaction optimization by prioritizing promising conditions; however, the complexity of multicomponent reactions and the limited availability of high-quality datasets remain significant challenges. While machine learning has achieved substantial progress in molecular property prediction, reaction-level modeling requires representations that capture three-dimensional structures, intercomponent interactions, and mechanistically critical features. In this study, we investigate the effective extension of molecular pretraining to reaction yield prediction from two complementary perspectives. First, we apply a reaction-aware architectural design based on molecular representation models pretrained via partial denoising to multicomponent reactions. Each reaction component is encoded as a three-dimensional stereoisomer embedding and concatenated into a reaction-level representation, with multi-head attention modeling intercomponent dependencies. We show that this approach achieves robust performance and generalizes well to reactions involving previously unseen components. Second, we design an architecture that extracts intermolecular relational embeddings through mechanistically meaningful pre-supervised learning using electron source-sink annotations derived from reaction mechanisms, followed by an active fine-tuning process. This mechanism-aware extension leverages large-scale reaction data to provide informative inductive bias and improves performance in data-constrained settings. We further confirm that this approach outperforms models that simply combine individual molecular embeddings, as it directly captures chemically meaningful information relevant to reaction outcomes. These results suggest promising directions for developing machine-learning models for reaction yield prediction that are more closely aligned with underlying chemical mechanisms.