Interpretable machine learning enables de novo mapping of cell type-specific RNA splicing regulation from scRNA-seq data

Context-dependent regulation of alternative splicing (AS), largely mediated by RNA-binding proteins (RBPs), is a key post-transcriptional mechanism shaping diverse biological processes. Experimental approaches for probing splicing regulation, such as CLIP-based assays and RBP perturbations, suffer from low throughput, poor physiological relevance, and bulk resolution that overlooks cellular heterogeneity. Here, we present CASREL (cell-specific alternative splicing regulation inference via explainable learning), a machine learning framework that reconstructs RBP-AS regulatory circuitry directly from single-cell RNA sequencing data without reliance on prior RBP-RNA binding annotations. CASREL integrates ensemble learning with SHAP-based interpretation to uncover RBP-AS associations, leveraging unique features of single-cell splicing profiles, including polarized isoform usage, minimal averaging of regulatory programs, and resilience to expression noise. Applications across diverse tissues and cell types demonstrate its robustness, accuracy, and biological relevance, establishing CASREL as a powerful method for de novo mapping of cell-specific RNA splicing regulation in physiological contexts.