From Binding Sites to Target Genes: Reframing Human microRNA Target Prediction

MicroRNA (miRNA) target prediction is often treated as a single problem, yet it encompasses two distinct computational tasks: identifying whether a miRNA can bind a particular target RNA binding site, and predicting the change in gene expression that follows from miRNA perturbation. Here, we trace the evolution of the field of miRNA target prediction in humans from early seed- and conservation-based algorithms, to deep learning models trained on large-scale interaction datasets. We note that early frameworks obscured binding site recognition and gene-level regulation by aggregating binding site-level scores into gene-level predictions, whereas newer experimental and computational advances have enabled a clearer separation of the tasks. Deep learning has substantially advanced miRNA binding site prediction, but progress in miRNA target gene prediction remains constrained by limited perturbation data, inconsistent data processing and annotations, heterogeneous definitions of regulatory response, uneven predictor coverage, and the absence of comprehensive benchmarks. We also highlight the importance of negative-example construction, which can introduce dataset-specific biases and inflate apparent model performance. We propose two complementary benchmarked community challenges: the prediction of miRNA-binding site interactions, and the genome-wide prediction of miRNA-mediated changes in target gene expression in defined biological contexts. Treating these as distinct computational tasks, while recognising their biological interdependence, would enable more meaningful evaluation of methodological advances and provide a stronger foundation for modelling miRNA-mediated regulatory networks.