MVGCL: Noise-Robust Multi-Source Similarity Fusion and Type-Aware Dual-Pathway Learning for Drug Repositioning

Drug repositioning can accelerate therapeutic discovery, but computational drug-disease association prediction remains constrained by noisy multisource similarities, heterogeneous biomedical semantics, and sparse supervision. We present MVGCL, a type-aware dual-pathway framework that combines noise-robust multisource similarity fusion with heterogeneous biological network representation learning and aligns the two pathways through a distribution-aware hierarchical contrastive strategy to improve representation consistency under sparse and noisy supervision. Under 10-fold cross-validation protocols consistent with prior work, MVGCL achieves the highest AUC and AUPR on all three benchmark data sets, reaching 0.9538/0.9511 on B-data set, 0.9871/0.9890 on C-data set, and 0.9818/0.9836 on F-data set, with consistent performance across varying negative sampling ratios and entity-wise split settings. Under a leakage-controlled protocol in which GIP kernels were recomputed exclusively from the training DDAs within each fold, MVGCL retained the highest AUC and AUPR among the compared GIP-based models, with AUC values of 0.9029, 0.9422, and 0.9145 on B-data set, C-data set, and F-data set, respectively. Case analyses on Alzheimer's disease and Parkinson's disease further show that MVGCL ranks literature-supported drugs highly and prioritizes biologically plausible candidates for follow-up investigation, supporting its utility for evidence-aware drug repositioning.