Prediction of miRNA-disease associations using a dual-channel contrastive model based on multi-source information fusion

In the early prevention and treatment of diseases, it is crucial to identify potential associations between microRNAs (miRNAs) and diseases accurately. In recent years, deep learning methods have provided a new computational strategy for exploring these associations. However, integration of multi-source datasets introduces computational complexity, resulting in time-consuming and unstable performance when applying deep learning to accurately predict miRNA-disease associations (MDAs). Here, we propose a computational approach for predicting MDAs based on a dual-channel contrastive model with multi-source information fusion (DCCM-MSIF). Firstly, the heterogeneous graph transformation (HGT) method is utilized to capture potential relationships in multi-source biological data; subsequently, the multi-view features are integrated to learn sequence and graph structure information separately in a dual-channel manner, and residual linkage and multi-module comparative learning are utilized to enhance the model performance. Finally, potential MDAs are predicted by a multilayer perceptron (MLP). The experimental results show that the AUC of DCCM-MSIF reaches 96.89% on HMDD v3.2 and 95.01% on HMDD v2.0, which is better than most of the methods. The importance of the modules and the generalization performance of DCCM-MSIF are further validated by ablation experiments and three cancer case studies. This work aims to provide new ideas and references for further exploring potential MDA mechanisms and promoting clinical diagnosis.