Dual-Channel Multiscale Graph Transformer with Adversarial Contrastive Learning and Low-Rank Disentangled Stratified Negative Sampling for Drug Repositioning

Drug repositioning accelerates therapeutic discovery, but existing computational methods are hampered by representation collapse, noisy supervision, and suboptimal negative sampling. To address these limitations, we introduce MGTAL-DR, a novel graph learning framework that integrates a dual-channel transformer architecture with adversarial contrastive learning and a purely negative sampling strategy. Its parallel graph encoders capture both multiscale similarity patterns and heterogeneous biological semantics. In one channel, structural neighborhoods are expanded via diffusion-based propagation to encode varying granularities of similarity. In the other, cross-entity relationships are contextualized through meta-path-guided attention over a unified drug-disease-protein graph. Adversarial perturbations enhance latent space robustness to prevent noise-induced collapse, while a low-rank decomposition strategy isolates informative hard negatives by disentangling global association trends from local residual signals. Together, these components sharpen the decision boundary under sparse and noisy conditions. Extensive experiments demonstrate that MGTAL-DR achieves state-of-the-art performance across three benchmark data sets. Furthermore, a case study on Alzheimer's disease highlights its practical utility by successfully identifying promising therapeutic candidates, thereby validating its real-world potential.