Dual-view global and mechanism-aware propagation for disease-gene association prediction

Understanding the associations between diseases and genes is essential for uncovering disease mechanisms and guiding therapeutic development. Given the high cost and time-consuming nature of experimental validation, computational prediction methods have become an effective approach for exploring potential disease–gene associations. Capturing both global network structure and latent mechanistic signals remains challenging in heterogeneous disease–gene networks. This paper proposes DGMAP, a dual-view global and mechanism-aware propagation framework for disease-gene association prediction. DGMAP first constructs a neighborhood view based on inter-node similarity. It then introduces a cross-view collaborative propagation mechanism to capture global structural information. A mechanism-aware disentanglement strategy is further designed to uncover latent interaction patterns. In addition, intra- and inter-view contrastive learning objectives are incorporated to improve node representations by aligning structural and mechanistic embeddings. Extensive experiments show that DGMAP achieves consistently better overall performance than existing methods in disease-gene association prediction. The Alzheimer’s disease case study further demonstrates its practical value: DGMAP identifies biologically relevant candidate genes, with 15 of the top 20 predictions supported by literature. Mechanism visualization and network bias analysis also show that these predictions reflect coherent and diverse disease-related mechanisms.