UniSplicer: A deep-learning framework for accurate splice-site prediction and splice-altering mutation detection across diverse taxa

RNA splicing removes non-coding introns from pre-mRNA to generate mature mRNA in eukaryotes, and accurate identification of splice sites is essential for understanding gene structure and regulation. Traditional gene annotation and splice-site prediction rely heavily on high-quality genome assemblies, extensive functional characterization, and substantial computational and experimental resources, limiting their applicability to many non-model species. Here, we present UniSplicer, a deep-learning-based training framework capable of generating accurate intron splice-site prediction models for diverse species using relatively limited transcriptomic data. UniSplicer-based models (http://www.unisplicer.com) consistently outperform existing prediction tools across a wide range of taxa, including plants, fungi, and metazoans. UniSplicer prediction scores serve as reliable indicators of mutational effects across multiple types of splice mutants. Moreover, application of the UniSplicer Arabidopsis thaliana model identified genes in natural ecotypes that exhibit aberrant splicing due to sequence variation near splice sites, suggesting potential roles in environmental adaptation. Collectively, UniSplicer-based models achieve high predictive accuracy and provide insight into how sequence variation drives splicing alterations across large genomic datasets.