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

MicroRNAs (miRNAs) regulate gene expression by binding to mRNAs, inhibiting translation, or promoting mRNA degradation. miRNAs are of great importance in the development of various diseases. Currently, numerous sequence-based miRNA target prediction tools are available, however, only 1% of their predictions have been experimentally validated. In this study, we propose a novel approach that leverages disease similarity degree between miRNAs and genes as a key feature to further refine human sequence-based predicted miRNA target interactions (MTIs). To quantify the similarity degree of diseases, we fine-tuned the Sentence-BERT model. Our method achieved an F1 score of 0.88 in accurately distinguishing human protein-level experimentally validated MTIs (functional MTIs, validated through western blot or reporter assay) and predicted MTIs. Moreover, this method exhibits exceptional generalizability across different databases. We applied the proposed method to analyze 1,220,904 human MTIs sourced from miRTarbase, miRDB, and miRWalk, encompassing 6,085 genes and 1,261 pre-miRNAs. Our model was trained in miRTarBase 2022. However, we accurately identified 90% (518/574) of the updated functional MTIs in miRTarbase 2025. This study has the potential to provide valuable insights into the understanding of miRNA-gene regulatory networks and to promote advancements in disease diagnosis, treatment, and drug development.