Topology-Guided Multi-Level Graph Learning: A Unified Framework for Molecular Property Prediction

Accurate molecular property prediction underpins progress in drug discovery, medicinal chemistry, and molecular design and depends critically on the quality of the learned molecular representations. While Graph Neural Networks (GNNs) and Graph Transformers have advanced this field considerably, most existing approaches emphasize a single structural perspective and struggle to jointly capture local connectivity, higher-order substructures, explicit topology, and long-range dependencies in a single learning process. This work introduces Topology-Guided Multi-Level Graph Learning (TMGL), a unified framework that integrates topology-enhanced node encoding, motif-guided representation learning, a Multi-Scale Graph Isomorphism Network, topology-aware graph transformer attention, adaptive multi-scale feature fusion, and cross-level contrastive learning into a single end-to-end architecture. By coupling topology preservation and motif consistency objectives with hierarchical contrastive alignment across node-, motif-, and graph-level embeddings, TMGL enables these complementary structural views to interact directly rather than being combined post hoc. The framework was evaluated on seven MoleculeNet benchmark datasets spanning binary classification, multi-label classification, and regression tasks. TMGL achieved Receiver Operating Characteristic Area Under the Curve (ROC-AUC) scores of 0.937, 0.960, 0.962, and 0.723 on BACE, BBBP, ClinTox, and SIDER, respectively, and RMSE values of 0.611, 1.293, and 0.436 on ESOL, FreeSolv, and Lipophilicity, outperforming a broad set of state-of-the-art baselines on most classification benchmarks and on Lipophilicity, while remaining competitive on ESOL and FreeSolv. Ablation experiments confirmed that each architectural component contributes meaningfully to overall performance, and embedding visualizations further illustrated the discriminative, class-consistent structure of the learned representation space. These results demonstrate the value of jointly modeling complementary structural perspectives for molecular graph learning and highlight the potential of the TMGL framework for applications in computational chemistry and drug discovery.