Reaction Center and Class Prediction via Cross-Attentive Multi-Task Learning

Understanding chemical reactions requires integrating fine-grained molecular transformations with broader semantic context. The underlying reaction mechanism is often defined by both the location of structural changes and the global reaction type. Modeling atom mappings between reactants and products is a key enabler of predicting several reaction properties; however, most existing models depend on external mapping tools, which introduce noise and limit end-to-end learnability. We propose MaRCC (Mapping-Assisted Reaction Center and Classification), a multi-task graph neural network that jointly performs reaction center identification and reaction classification, while incorporating atom mapping as a differentiable auxiliary task. MaRCC features a dual-level architecture: a soft cross-attention mechanism aligns product atoms to reactants for local reactivity prediction, and a global classification head infers reaction types from pooled graph embeddings. A learned atom mapping module provides alignment priors that guide both attention and representation learning.Evaluated on the USPTO-50K benchmark, MaRCC achieves state-of-the-art results across all core tasks, including an F1 score of 98.3 for atom reactivity, 98.0\% Top-1 edit localization accuracy, and 98.7\% reaction classification accuracy. Ablation studies demonstrate that mapping-guided attention and multi-task supervision yield consistent improvements in accuracy, while facilitating interpretable alignment between reactants and products.By unifying atom mapping, local reactivity, and global transformation prediction within a chemically grounded framework, MaRCC advances a structured, interpretable, and high-fidelity understanding of reactions. The architecture offers practical utility for synthesis planning, reaction annotation, and automated molecular design.