Multi-granularity transformer contrastive learning and feature reconstruction for prediction of disease-related miRNAs

The increasing research confirms the function of microRNAs (miRNAs) play a vital role in the diagnosis, treatment, and prognosis of diseases. Screening the potential candidate miRNAs related to the diseases by the computational methods may be helpful for reducing the experimental cost for discovering the actual disease-miRNA associations. Most of the recent methods exploited the data from multiple sources to infer the association tendencies between miRNAs and diseases. However, these methods did not completely learn the features of the multi-source data from multiple granularities. This paper present an association prediction model based on multi-granularity transformer contrastive learning and feature reconstruction convolution (APMF). First, the strongly correlated node sequences are constructed by random walks, and then the features of the positive and negative node sequences for the target node are learned. Second, the multi-granularity transformer strategy is designed to refine the features of the miRNA, disease, and lncRNA nodes from fine-granularity and coarse-granularity, respectively. Third, the contrastive learning strategy is presented to maximize the consistency among the features of positive samples and the difference among the features of positive samples and negative ones. Finally, the feature reconstruction convolution is constructed to capture the diverse semantics of features and suppress the noises of the data. The experimental results indicate that APMF outperforms several state-of-the-art methods for prediction of disease-related miRNAs. The ablation experiments demonstrate the effectiveness of the multi-granularity transformer contrastive learning and the feature reconstruction convolution. The case analysis over three diseases also shows APMF’s ability in discovering the potential miRNA candidates for the diseases.