Topology-aware directional graph representation learning for capturing structural and topological patterns in molecular toxicity graphs

For many years, understanding how to learn expressive molecular graph representations for molecular toxicity prediction remains a fundamental challenge in graph representation learning and computational toxicology. Molecular graphs represent atoms and chemical bonds as nodes and edges. Most graph neural networks (GNNs) learn primarily through local neighbourhood aggregation. This local processing may limit the representation of global topology and long-range structural dependencies. Consequently, important structural characteristics associated with molecular toxicity may be insufficiently encoded. To address these limitations, we propose a novel TDGL model, which is a topology-aware directional graph learning framework. Our model combines the directional message propagation learning mechanism with persistent-homology-based representations. The directional encoder captures heterogeneous local structural information among neighbouring atoms, while the topological branch extracts multi-scale graph characteristics and represents them through persistence images. By integrating these complementary representations, our TDGL model incorporates both local molecular structure and broader topological organisation within a unified learning framework. Furthermore, an attention-based fusion mechanism is introduced to adaptively integrate structural and topological representations. We conducted extensive experiments on different real-world molecular datasets that demonstrate the effectiveness of our proposed TDGL model over state-of-the-art (SOTA) graph-learning methods.