A cross-attentive multi-task graph learning framework for chemical reaction modeling.

Area of Science:

• Computational chemistry
• Machine learning in chemistry

Background:

• Chemical reaction understanding requires integrating local molecular edits with global context.
• Existing methods often treat atom mapping and reaction classification separately, limiting end-to-end learning.
• A unified approach is needed to bridge fine-grained transformations and broader reaction semantics.

Purpose of the Study:

• To develop a unified multi-task graph neural network architecture for joint prediction of atom mappings, reaction centers, and reaction classes.
• To improve the end-to-end learnability of chemical reaction understanding by integrating local and global information.
• To provide a more structured and interpretable framework for analyzing chemical reactions.

Main Methods:

• Developed MARCC (Mapping-Assisted Reaction Center and Classification), a multi-task graph neural network.
• Integrated a mapping-guided cross-attention mechanism for aligning reactants and products.
• Employed a dual-graph design for reasoning about bond-level transformations.
• Utilized pooled product embeddings for global reaction classification.

Main Results:

• Achieved state-of-the-art performance on the USPTO-50K benchmark with 98.2% atom mapping accuracy and 97.2% reaction classification accuracy.
• Demonstrated high Top-1 edit localization accuracy (99.1%) using a mapping-guided approach.
• Showcased competitive performance even in a products-only setting, comparable to specialized methods.
• Ablation studies confirmed the benefits of mapping-guided attention and multi-task supervision for accuracy and interpretability.

Conclusions:

• MARCC successfully unifies atom-level alignment, local reactivity, and global classification for enhanced chemical reaction understanding.
• The MARCC framework offers improved interpretability and accuracy in predicting reaction properties.
• MARCC has potential applications in reaction annotation, template discovery, and mechanism inference, with possible extensions to biochemical domains.