DCTGM: A Novel Dual-channel Transformer Graph Model for miRNA-disease Association Prediction

Studies have shown that as non-coding RNAs, miRNAs regulate all levels of life activities and most pathological processes. Therefore, identifying disease-related miRNAs is essential for disease diagnosis and treatment. However, traditional biological experiments are highly uncertain and time-consuming. Hence, advanced intelligent computational models are needed to address this problem. We propose a dual-channel transformer graph model, named DCTGM, to learn multi-scale representations for miRNA-disease association prediction. Specifically, DCTGM includes a transformer encoder (TE) and GraphSAGE encoder (GE). The TE intensely captures the important interaction information between miRNA-disease pairs, and the GE aggregates multi-hop neighbor information of miRNA-disease association heterograph to enrich node features. Then, an attention module is proposed to aggregate the dual-channel interactive representations, and we adopt a multi-layer perceptron (MLP) to predict the miRNA-disease association scores. The fivefold cross-validation experimental results demonstrate that our proposed DCTGM achieves the AP of 92.735%, F1 of 84.430%, accuracy of 85.255%, and ROC of 93.012%. In addition, we conduct case studies on brain neoplasms, kidney neoplasms, and breast neoplasms. The extensive experiments show that the dbDEMC database validates 100% of the top 20 predicted miRNAs associated with these diseases. This model can effectively predict the potential mirNA-disease association. Experiments have shown that miRNA associated with a new disease can also be predicted.

Automated summary

This study proposes a dual-channel transformer graph model (DCTGM) for predicting miRNA-disease associations by learning multi-scale representations. Experimental results show that the model is effective in predicting potential miRNA-disease associations.