Increasing numbers of studies have highlighted the significance of miRNAs in human gene regulation, with their dysregulation linked to a variety of diseases, including cancers and neurological disorders1. miRNAs play a crucial role through post-transcriptional gene suppression, targeting the majority of human genes, with potential to act as highly selective therapeutics2-3. In order to further our understanding of their biological roles, including in disease and use as therapeutics, a comprehensive map of miRNA functions, including their targets and repression levels is required.
Due to experimental constraints, a number of computational models4-6 for prioritising potential miRNA-target mRNA repression activities have been proposed to assist functional characterisation, but with limited predictive performance and restricted usability. To address these limitations, we developed PRIMITI - a PRedictive model for the identification of new MIRNA-Target mRNA Interactions. PRIMITI utilises machine learning to model the molecular interactions between miRNAs and target sites in 3’-untranslated regions (3’-UTR) and to predict miRNA-mRNA suppression activities. To counteract natural publication bias, negative sample selection was used to produce reliable negative samples, resulting in high-quality data. PRIMITI then used information on single nucleotide polymorphisms (SNPs) and general physicochemical descriptors to improve the characterization of functional target sites.
On cross-validation and independent blind test sets, PRIMITI achieved robust predictive performances, achieving Matthew’s correlation coefficients of up to 0.81 for identification of miRNA-target sites and up to 0.77 for prediction of the resulting mRNA repression activity. The model was further validated with an external blind microarray dataset7 and a list of experimentally validated miRNA-mRNA interactions collected from miRTarbase and Tarbase8,9. Across both validations, PRIMITI successfully predicted miRNA-mediated repression, outperforming current state-of-the-art methods4-6, confirming the model’s utility for preliminary screening. We have made PRIMITI publicly available through a user-friendly web server at https://biosig.lab.uq.edu.au/primiti.