Structured Chemical Reaction Modeling with Multitask Graph Neural Networks

This work introduces MARCC (Mapping-Assisted Reaction Center and Classification), a multitask graph neural network for modeling chemical reactions in a structured manner.MARCC jointly learns atom mapping (graph matching), reaction center prediction (binary classification), and reaction classification (global classification) within a unified architecture.By coupling these interdependent tasks, the model captures the structural relationships between local reactivity and global transformation semantics.Experiments on the USPTO-50K benchmark demonstrate that multitask supervision improves both accuracy and interpretability compared to single-task and product-only baselines.The framework provides a general approach for graph-based reaction understanding and structured chemical prediction.