Advances and challenges of splicing prediction with AI

RNA alternative splicing is a fundamental post-transcriptional mechanism whose dysregulation drives various human diseases. Predicting splicing outcomes is therefore a central challenge in precision medicine. This Review traces the evolution of computational approaches from early statistical heuristics to modern artificial intelligence frameworks. We dissect the methodologies that shape predictive performance, including training data scale, output resolution, splicing event quantification and model complexity. Among these factors, quantitative assessment of splicing events and the increasing complexity of model architectures represent particularly critical axes that define both biological interpretability and computational feasibility. We further describe how these models empower translational applications, from annotating variant effects to guiding antisense oligonucleotide development. Nonetheless, persistent challenges remain, including the interpretation of deep-intronic mutations, isoform-level reconstruction and integration of multimodal data. Together, these perspectives define both the progress achieved and the opportunities ahead for splicing prediction in genomics and medicine.