Deciphering circRNA-drug sensitivity associations via global-local heterogeneous matrix factorization and hypergraph contrastive learning

Abstract

Growing evidence highlights the critical role of circular RNAs (circRNAs) as regulators of cellular drug sensitivity, significantly influencing drug efficacy. While matrix factorization has proven feasible in uncovering circRNA-drug sensitivity associations, existing methods rely solely on decomposing the association matrix and fail to efficiently incorporate richer biological information. Moreover, current predictive models are limited in representing multi-perspective relationships and higher-order relationships in circRNA-drug sensitivity associations. To address these limitations, we propose a novel model based on global-local heterogeneous matrix factorization and hypergraph contrastive learning (HMFHCL). HMFHCL first calculates the global and local similarities of circRNAs and drugs, then constructs global and local circRNA-drug heterogeneous networks and performs matrix factorization of the adjacency matrices of these networks to extract information-rich feature representations. By utilizing multi-source information, HMFHCL effectively captures richer structural features and reveals potential connections in heterogeneous networks. Next, we constructed multiple circRNA and drug hypergraphs using global and local association matrices to capture higher-order interactions between nodes via hypergraph convolution. To enhance feature learning, comparative learning is applied to both global and local views of circRNA/drugs, effectively mining the similarities and differences between global structures and local details, improving the model’s ability to perceive underlying patterns and the consistency of representation learning. Finally, HMFHCL integrates circRNA and drug features from different perspectives to predict circRNA-drug associations effectively. Comprehensive experiments on three benchmark datasets demonstrate that HMFHCL outperforms state-of-the-art models, highlighting its superior ability in uncovering complex circRNA-drug sensitivity associations.