Recurrent Splicing Programs Define a Conserved Axis of Tumor Heterogeneity and Prognosis

Alternative splicing is a widespread but largely undercharacterized contributor to tumor heterogeneity, lineage plasticity, and clinical outcomes. To systematically define the pan-cancer landscape of alternative splicing, we leveraged a sensitive unsupervised strategy to iteratively discover and functionally describe the universe of coordinated and uncoordinated splicing events across 25 malignancies and >10,000 patients. This analysis uncovered more than 1,000 distinct splicing-based tumor subtypes, including ~200 with prognostic significance and captured three times the number of poor survival splicing events than conventional analyses. Among these tumor subtypes, we describe a recurrent, poor-prognosis lung adenocarcinoma subtype enriched for but not genetically dependent on RBM10 mutations, and a neuroendocrine-like breast cancer subtype exhibiting brain-specific microexon splicing. These subtypes recur across multiple tumor types, revealing conserved splicing programs that transcend tissue of origin. Comparison of splicing across malignancies revealed nine core pan-cancer modules linked to regulation of distinct RNA binding proteins and reciprocal regulation of lineage differentiation programs. These findings establish an orthogonal classification system based on splicing dysregulation, independent of canonical splicing factor mutations, with implications for prognostic biomarker discovery and therapeutic targeting.