PMLocMSCAM: Predicting miRNA Subcellular Localisations by miRNA Similarities and Cross-Attention Mechanism

Abstract

Many studies have shown that microRNAs (miRNAs) play key roles in some important processes and human complicated diseases. In addition, they also have specific physiological roles at different cellular sites. Therefore, identifying their subcellular localisation is very urgent to systemically understand their physiological functions. In this study, we propose a computational method, called PMLocMSCAM, to predict miRNA subcellular localisation based on miRNA similarities and cross-attention mechanism. PMLocMSCAM implements a multimodal integration framework that systematically processes miRNA sequence data, miRNA-mRNA association networks with mRNA subcellular localisation annotations, miRNA-disease associations, and miRNA-drug association networks. The architecture initiates with intrinsic feature extraction through Smith-Waterman alignment for sequence similarity computation and disease ontology-based functional similarity derivation. Subsequent heterogeneous network embedding employs Node2vec for topological feature learning across three interaction modalities (miRNA-disease, miRNA-drug, and miRNA-mRNA networks), enhanced by hypergraph convolution to capture higher-order relationships through incidence matrix decomposition. Localisation-specific patterns are propagated via miRNA-mRNA interaction weights, culminating in a multi-head attention mechanism that dynamically fuses five feature matrices—miRNA sequence features, miRNA-disease association features, miRNA-drug association features, miRNA-mRNA association features, and miRNA-mRNA localisation features. These integrated representations are processed through residual-connected multilayer perceptrons to generate probabilistic predictions across seven subcellular compartments, establishing an end-to-end computational paradigm for multimodal miRNA localisation analysis. In order to assess the prediction performance of our method and compare it with other miRNA subcellular localisation computational methods, we conduct 10-fold cross validation (10-CV) and independent test dataset. The AUC (area of receiver operating characteristic curve) and AUPR (area of precision-recall curve) are used as metrics. The experiment results show that the average AUC and AUPR values exceed 0.9182 and 0.8487 on 10-CV, respectively. The AUC and AUPR values also reach 0.9157 and 0.8469 on independent test dataset, respectively. It is superior with compared methods. The ablation experiment results also further that PMLocMSCAM can effective predict miRNA subcellular localisations and provide help to understand their physiological functions.