GMIMDA: Interpretable miRNA-Disease Association Prediction via Game Optimization and Multi-view Representation Learning

miRNAs influence cellular functions by regulating gene expression and interacting with diverse biomolecules within the cell. Accurate prediction of miRNAdisease associations (MDA) plays a crucial role in disease diagnosis, treatment, and drug development. However, existing computational methods focus on network structure and ignore multi-view information such as linear and non-linear when extracting miRNA and disease features. In addition, these models are generally “blackbox” in nature, which limits the understanding of their prediction mechanisms. Therefore, we propose an interpretable computational method based on game optimization and multi-view representation learning (GMIMDA) for predicting MDA. Specifically, GMIMDA first integrates multisource similarity views using similarity network fusion (SNF) technique and constructs multi-view complementary representations by extracting linear, nonlinear, and graph structure representations of miRNA and disease via singular value decomposition with weighted scaling (SVDWS), nonnegative matrix factorization (NMF), and graph convolutional networks (GCNs), respectively. Second, to enhance the interactivity between nodes, GMIMDA uses graph features as initial strategies of participants in game theory modeling and selects optimal strategies based on Nash equilibrium. Finally, GMIMDA weightedly fuses the results of multi-view representation learning and inputs them into the interpretable Kolmogorov-Arnold Networks (KANs) for prediction, using symbolic activation functions to reveal the mapping relationship between input and output. Experimental findings demonstrate that GMIMDA outperforms existing approaches in two different databases. Case studies provide additional evidence supporting the effectiveness of GMIMDA in identifying unknown miRNAs of disease-associated in practical applications.