Benchmarking microRNA Target Prediction Algorithms Using Single-Cell Co-Sequencing Data

(1) Background: MicroRNAs (miRNAs) are small non-coding RNAs that play pivotal roles in the post-transcriptional regulation of gene expression, influencing a wide range of physiological and pathological processes. Accurately identifying miRNA targets is crucial for understanding miRNA modes of action. To this aim, a plethora of algorithms have been developed to predict miRNA targets, each employing distinct methodologies and relying on different features. The limited overlap among target predictions generated by various algorithms underscores the necessity for comprehensive and independent benchmarks to evaluate their performance. (2) Methods: We selected seven algorithms among the most popular ones to perform a benchmark with an original approach using recently published datasets of miRNA-mRNA co-sequencing at the single-cell level. We used Gene Set Enrichment Analysis to assess algorithm's capabilities to predict sets of targets statistically anti-correlated in expression with miRNAs. We worked with both co-sequencing datasets of human and mouse single-cells. (3) Results: Our benchmark shows high performances for mirDIP, which corresponds to the consensus result of 24 different algorithms, for human miRNAs. In human cell lines, Diana microT, TargetScan, miRmap, and miRDB also provide excellent results, while RNA22 and miRWalk exhibited poorer results. In mouse primary cells, Diana microT leads, closely followed by miRmap. Intriguingly, RNA22 performs better in mouse primary cells than in human cell lines. Moreover, our benchmark highlights the benefit of reducing targets to the experimentally validated ones. Finally, we demonstrated that performance varies depending on the number of targets used, with TargetScan performing better than mirDIP when considering only a few dozen targets.