miRstring: An RNA language model enables mature miRNA decoding and artificial small RNA design across species

Abstract

Introduction MicroRNAs (miRNAs) are processed from structured precursors and subsequently loaded into Argonaute proteins to repress target mRNAs. Accurate decoding of mature miRNAs from precursor sequences is fundamental to miRNA annotation and the rational engineering of artificial miRNAs. However, existing computational approaches are largely developed for humans, with limited ability to generalize across diverse plants, animals, and viruses. Moreover, a unified data-driven framework that links mature-miRNA decoding with the predictive design of artificial miRNA precursors is still lacking, restricting the scalable application of miRNA biotechnology across species. Objectives We develop and evaluate miRstring, a biogenesis-aware RNA language framework for decoding the four boundaries of mature miRNAs from precursor context across species, and further investigate its utility for cross-species miRNA annotation and rational artificial miRNA design. Methods miRstring, a biogenesis-aware RNA language framework, was developed using 77,708 miRNA precursors spanning 414 species. The model was evaluated against existing methods under both family-held-out and species-held-out settings to assess its ability to generalize to unseen miRNA families and species. Attention-based analyses were further performed to examine whether miRstring captured biologically relevant features associated with miRNA processing, and the framework was subsequently applied to predictive design of artificial miRNA precursors. Results miRstring substantially outperformed existing methods across species-held-out evaluations and, notably, retained strong performance under the more stringent family-held-out setting. It accurately identified mature-miRNA start sites, while attention was enriched at endonuclease-cleaved miRNA/miRNA* boundaries, revealing biologically meaningful processing features. Importantly, miRstring-guided precursor design markedly enhanced artificial miRNA-mediated target repression, demonstrating its utility for rational miRNA engineering. Conclusion miRstring provides a scalable framework linking cross-species mature-miRNA annotation with predictive artificial miRNA design, extending computational miRNA analysis toward artificial intelligence-driven small-RNA biotechnology.