HTF-DTI: Hypergraph-Based high-low order topological fusion for drug-target interaction prediction.
Abstract
Objective
Accurate drug-target interaction (DTI) prediction can accelerate candidate prioritization in drug discovery and repurposing. However, existing graph-based methods may inadequately integrate high-order neighborhood dependencies with the low-order drug-target topology and often rely on shallow feature fusion for pairwise interaction modeling.
Methods
We developed HTF-DTI, a hypergraph-based framework comprising a high-low order Structure-Level Fusion Module (HLSFM) and a Graph-Enhanced Dual-Path Interaction Module (GEDIM). Drug and target hypergraphs were constructed from K-nearest-neighbor relationships derived from interaction profiles and integrated with the original bipartite DTI topology. Graph attention networks learned topology-aware representations, whereas GEDIM generated joint drug-target pair representations through bidirectional information exchange. HTF-DTI was evaluated using five-fold stratified cross-validation on the Luo and Zheng datasets and the four Yamanishi subsets, with AUROC and AUPR as the primary evaluation metrics.
Results
HTF-DTI achieved AUROC/AUPR values of 0.9795/0.9769 on the Luo dataset and 0.9659/0.9593 on the Zheng dataset. Across the Yamanishi subsets, it achieved the highest AUROC on GPCRs, Es, ICs, and NRs subsets and the highest AUPR on the latter three subsets. Ablation experiments showed performance degradation after removing either high-low order topology fusion or dual-path interaction learning. Parameter sensitivity, robustness, and case study analyses further supported the stability and biological plausibility of the predictions.
Conclusion
Joint modeling of high-low order topology with bidirectional pair interaction provides an effective framework for prioritizing candidate DTIs. The resulting predictions should be interpreted as computational hypotheses requiring independent database and experimental validation.