AS-MPNN: An Adaptive Substructure-Aware Message Passing Neural Network for Enhanced Polymer Property Prediction

Polymer property prediction is essential in the discovery and design of polymers. Graph Neural Networks (GNNs) are widely used for this task due to their ability to process molecular graphs. However, existing GNN-based models often represent graphs at the atomic level, overlooking critical substructural or functional group information, which limits their ability to capture the multiscale characteristics of polymers. Recent studies suggest integrating substructural information to capture more advanced chemical features, but current substructure fragmentation methods rely on predefined rules, lacking flexibility. To address this, we propose the Adaptive Substructure-aware Message Passing Neural Network (AS-MPNN), which enhances polymer property prediction by combining atom-level and substructure-level representations. By integrating Morgan fingerprints with a Random Forest (RF) model, AS-MPNN flexibly captures key substructures, ensuring optimal fragmentation. Benchmarking on multiple datasets demonstrates the model's exceptional performance and strong generalization across tasks.