DRCMDA: A Dual-View Drug Repositioning Framework with Cluster-Aware Structured Masked Reconstruction and Diffusion-Based Metapath-Graph Augmentation.

Drug repositioning aims to identify new therapeutic indications for existing drugs, yet current deep learning approaches on multi-source biological data face limitations in both homogeneous and heterogeneous network modeling. In homogeneous similarity networks, random masking-based self-supervised learning neglects intrinsic clustering structures of biological entities and fails to capture high-order semantics, while in heterogeneous networks, sparse associations limit the effectiveness of metapath-based reasoning. To address these challenges, we propose DRCMDA, a dual-view drug repositioning framework that combines cluster-aware structured masked reconstruction with diffusion-based metapath-graph augmentation. The homogeneous module employs cluster-guided column permutation perturbations and a Teacher-Student distillation mechanism to learn robust, high-level representations, while the heterogeneous module leverages a diffusion model to generate diverse synthetic metapath graphs from the learned graph distribution. Furthermore, DRCMDA employs dual-view contrastive learning and node-level feature fusion to align and integrate complementary information across homogeneous and heterogeneous views. Experiments on three benchmark datasets demonstrate that DRCMDA consistently outperforms state-of-the-art methods across multiple evaluation metrics, with case studies and molecular docking analyses confirming its translational potential in identifying promising therapeutic candidates.