miRBench: A Comprehensive microRNA Binding Site Prediction Training and Benchmarking Dataset

Abstract

Benchmarking drives innovation by providing a standardised framework that challenges existing methods and encourages the development of more advanced and effective solutions. Having well-curated benchmarking datasets has allowed machine learning experts to enter and revolutionise various fields of bioinformatics, such as protein structure prediction. Here, we present a set of benchmarking datasets for the yet unsolved task of accurate microRNA (miRNA) binding site prediction. miRNAs are small, non-coding RNAs that regulate gene expression post-transcriptionally. miRNAs are loaded onto Argonaute family proteins (AGO) and serve as guides to target specific binding sites on messenger RNA through partial sequence complementarity, the exact rules of which remain to be fully understood. The precise identification of miRNA binding sites is crucial for the further understanding of these master regulators of gene expression. Despite decades of attempts to accurately predict miRNA binding sites, the field has been unable to conclusively determine the exact rules of miRNA binding. An important issue is the relative paucity of unbiased experimental datasets that could be used to train more accurate machine learning models. In recent years, novel experimental approaches have revolutionised the field, producing thousands of unbiased miRNA-binding site sequence pairs. Recently, over one million such pairs were produced and made publicly available, posing the potential to revolutionise the field. In this manuscript, we process and produce several comprehensive, well-annotated, and easy-to-use benchmark datasets from these experimental approaches. We use these datasets to benchmark state-of-the-art methods. We make the benchmarking datasets easy to access programmatically with a few lines of Python code, aiming to stimulate further machine learning research in the field by lowering the barrier to entry for machine learning experts into the field of miRNA binding site prediction. All code, datasets, and prediction tools are accessible at https://github.com/katarinagresova/miRBench/releases/tag/v1.0.0.