Bond-Aware Molecular Graph Learning With Multi-Graph Interleaved Message Passing.

Area of Science:

• * Cheminformatics
• * Machine Learning
• * Computational Chemistry

Background:

• * Graph neural networks (GNNs) excel at molecular property prediction by treating molecules as homogeneous graphs.
• * The inherent heterogeneity of chemical bonds is often overlooked in current GNN models.
• * This limits the ability of GNNs to fully capture complex molecular structures and their properties.

Purpose of the Study:

• * To address the limitations of homogeneous graph representations in GNNs for molecular property prediction.
• * To develop a novel multi-graph learning model that accounts for bond heterogeneity.
• * To improve the accuracy and performance of GNNs in predicting molecular properties.

Main Methods:

• * Construction of bond-centric graphs to explicitly represent bond heterogeneity.
• * Development of a multi-graph learning model incorporating augmented bond graph views and bond coding for atom features.
• * Introduction of the interleaved message passing graph neural network (IMPGNN) for integrated cross-graph information during node representation learning.
• * Implementation of a structure-aware pooling mechanism for enhanced graph representation.

Main Results:

• * The proposed structure-aware pooling mechanism achieved up to a 45.7% improvement compared to simple sum pooling.
• * The IMPGNN model demonstrated superior performance on molecular property prediction tasks.
• * The method surpassed existing approaches, including multimodal models, on 75% of evaluated benchmark datasets.

Conclusions:

• * Accounting for bond heterogeneity in molecular graphs significantly enhances GNN performance.
• * The proposed bond-centric multi-graph learning approach offers a powerful new direction for molecular representation learning.
• * This work provides a foundation for more accurate and robust molecular property prediction using graph neural networks.