DeepExoMir: A Reproducible RNA Language Model Framework for CLIP-Seq-Supported MicroRNA Target-Site Prioritization

MicroRNAs regulate gene expression post-transcriptionally, yet target prediction faces a credibility gap: published methods drop sharply against CLIP-seq-validated negatives. We present DeepExoMir, a deep learning framework integrating frozen RiNALMo RNA language model embeddings with biologically informed features. Under a dual-probe ablation protocol on three miRBench test sets, DeepExoMir reaches mean AU-PRC 0.855, surpassing eight retrained baselines (paired bootstrap p<0.001). Evolutionary conservation and duplex structure prove largely redundant with language-model priors, motivating a structure-free Lite variant (0.863). On nine exosomal miRNAs from a companion melanogenesis study, DeepExoMir recovers literature-validated targets and ranks canonical pigmentation regulators (KITLG, MITF, TYRP1) in the top 5%.