HGSAF: Heterogeneous graph structural alignment and distillation model with attentive fusion for microbe-drug association prediction.
Abstract
The rapid emergence of antimicrobial resistance necessitates AI-driven computational strategies to navigate the molecular diversity landscape of microbe-drug interactions. While graph neural networks have shown promise in predicting such associations, existing methods struggle with the intertwined challenges of processing topological and semantic features in isolation and handling severe class imbalance inherent in biomedical datasets. In this work, we devised a heterogeneous graph structural alignment and distillation with an attentive fusion model to predict microbe-drug associations. Our model constructs heterogeneous biological graphs alongside meta-path-induced subgraphs to capture complementary views of microbial and drug entities. A layer-adaptive graph transformer is devised to simultaneously encode topological and positional information while mitigating noise and over-smoothing. Critically, we introduce a reciprocal distillation module for dual-path heterogeneous graph learning that mutually transfers structural and semantic knowledge between pathways, eliminating the need for explicit negative sampling during representation alignment. For the final association prediction, unobserved pairs are sampled as negative examples for supervised training, following standard practice. A cross‑graph alignment loss, jointly optimized with focal loss, improves robustness against class imbalance under controlled sampling ratios (1:1 and 1:5). Extensive evaluations, including comparisons against both domain-specialized predictors and general-purpose graph neural network backbones, demonstrate that HGSAF consistently achieves superior performance and highlight the methodological robustness and potential for broader application in biomedical network analysis.