CSMSA: Cross-Space Multiscale Adaptive Link Prediction for ceRNA-Mediated Multimolecular Disease Regulatory Networks

Abstract

Regulatory interactions associated with diseases are pivotal for elucidating the molecular mechanisms that drive disease progression and promoting precision medicine. Nevertheless, existing research algorithms often overlook the potential dynamic synergistic-competitive mechanisms between different ceRNA regulatory networks and lack cross-space learning capabilities across multiple heterogeneous graph structures, making it difficult to comprehensively capture the multidimensional molecular regulatory biological mechanisms in disease data with different structural densities. Therefore, we propose the cross-space multiscale adaptive learning framework (CSMSA) that integrates a heterogeneous five-layer ceRNA regulatory network and introduces an adaptive cross-space learning mechanism to dynamically capture complementary and specific interactions and effectively learn the intrinsic biological regulatory mechanisms. Moreover, the CSMSA framework employs a multi-scale feature fusion strategy that hierarchically learns node embeddings by integrating local structural information and global topological features from heterogeneous graphs to enhance predictive performance and robustness across complex datasets of varying sizes. Comprehensive evaluations on three independent datasets show that CSMSA surpasses existing methods in the multimolecular disease prediction task (Max AUC = 0.9880, Max AUPR = 0.9829), thereby providing a reliable new paradigm for probing disease regulatory links.