Relation-Aware Pretraining and Reaction Center Modeling for Chemical Reaction Graph Representation Learning

Area of Science:

• Computational chemistry
• Machine learning for chemistry
• Chemical informatics

Background:

• Learning informative representations of chemical reactions is vital for downstream tasks like predicting reaction conditions and yields.
• Current models face challenges in simultaneously capturing global reactant-product relationships and localized structural changes at the reaction center.
• A unified framework is needed to address these limitations in chemical reaction representation learning.

Purpose of the Study:

• To propose the Knowledge-Aware Graph Transformer (KAGT), a novel framework for chemical reaction graph representation learning.
• To develop a model that can effectively capture both global and local chemical reaction information within a unified structure.
• To enhance the performance of AI-driven chemical synthesis tasks through improved reaction representations.

Main Methods:

• KAGT utilizes a shared graph Transformer encoder integrated with a condition-attributed reaction knowledge graph.
• Pretraining involves relation-aware learning using molecular graphs as entities and reaction descriptors (solvent, catalyst, temperature) as relation attributes.
• Key techniques include masked atom modeling, 3D geometric denoising, and explicit reaction-center modeling guided by atom mapping.

Main Results:

• KAGT demonstrates strong performance in reaction classification, reaction condition prediction, and yield prediction across multiple downstream tasks.
• Evaluations show KAGT outperforms existing baseline models in these chemical synthesis applications.
• Qualitative analysis confirms that learned reaction center scores align with chemically plausible transformation sites.

Conclusions:

• KAGT provides a transferable representation framework for AI-driven chemical synthesis, effectively bridging the gap in current models.
• The model's ability to capture both global and local reaction features leads to improved performance in key synthesis prediction tasks.
• KAGT represents a significant advancement in learning informative representations of chemical reactions for computational chemistry and materials science.