Topology-aware relational learning on heterogeneous networks for lncRNA-disease association prediction

Long non-coding RNAs (lncRNAs) play an irreplaceable role in critical physiological processes such as cell cycle regulation, chromatin remodeling, and tumor suppression, yet traditional wet-lab molecular experiments are expensive and time-consuming. Consequently, there is an urgent need to develop efficient computational prediction models to prioritize disease-related lncRNAs, aiming to discover potential associations and reveal their underlying biological mechanisms.

A novel computational architecture named TARLHN is proposed. First, a hierarchical path attention encoder models multi-hop biological information transfer pathways over a heterogeneous network, employing a self-organizing dynamic routing mechanism to adaptively adjust semantic structural weights. Secondly, to counteract stochastic network noise and high sparsity, a multi-task equilibrium-driven representation learning strategy utilizing a generator-discriminator game is introduced to achieve structural homeostasis. Finally, an auxiliary edge prediction task captures evolutionary topological patterns to generate robust association scores.

Evaluated through a rigorous 10-fold cross-validation protocol with fixed negative sampling pools, TARLHN demonstrated exceptional predictive stability, achieving outstanding performance with an average AUC of 0.9638 and AUPR of 0.9682. Furthermore, extensive case studies confirmed its effectiveness in screening potential lncRNAs associated with breast neoplasms and squamous cell carcinoma, with top-ranked candidates successfully validated by independent literature.

This approach efficiently decodes the complex interactive topologies between lncRNAs and human diseases through topology-aware network representation and equilibrium-driven distribution alignment. It provides more effective theoretical guidance for interpreting lncRNA-mediated cellular control systems and discovering potential therapeutic targets.