FraHNN: A Fragment-Based Hypergraph Neural Network Model for Molecular Property Prediction

Molecular property prediction is an essential part of AI-driven drug design and discovery. Graph neural networks (GNNs) have become powerful tools for molecular property prediction. However, most of the existing GNNs represent molecules at the atomic level, ignoring the information about molecular fragments. It is these fragments containing pharmacophores that often determine molecular properties. In this work, we present FraHNN: Fragment-based Hypergraph Neural Network, a hypergraph learning framework that introduces molecular fragments to obtain more complete molecular information. Specifi-cally, FraHNN represents molecules as nodes of the hypergraph and fragments as hyperedges, using the hypergraph convolution algorithm to learn and update features of molecular nodes. Our proposed model has been tested on nine bench-mark datasets. The experimental results show that FraHNN achieves excellent performance on both classification and regression tasks for molecular property prediction. Additionally, visualization studies also indicate the model’s strong capability in molecular representation learning.