Refining Protein-Level MicroRNA Target Interactions in Disease from Prediction Databases Using Sentence-BERT

Abstract

MicroRNAs (miRNAs) are non-coding RNAs approximately 22 nucleotides in length. miRNAs regulate gene expression by binding to target mRNAs, inhibiting translation, or promoting mRNA degradation. miRNAs are of great importance in the development of disease. So far, a variety of miRNA prediction tools have been available, which are able to identify and predict miRNA sequences and their target genes, providing key support for studying miRNA regulatory networks, resolving disease mechanisms and developing potential therapeutic strategies. In this paper, we propose a unique prediction approach that predicts miRNA-gene relationships based on the disease semantic strings of miRNAs and their potential target genes, named MicroRNA-Gene Connection Building Method. The method was trained on more than 1,900 pairs of experimental and predictive data consisting of 10 cancer-related genes and evaluated using stratified cross-validation (StratifiedKFold) with a model accuracy of more than 90%. Subsequently, we used this model to classify 10 genes related to immunity and inflammation with a classification accuracy of over 85%.