Introduction: Loss of nuclear TDP43 function and disruption of RNA splicing are central molecular features of amyotrophic lateral sclerosis. However, a unified tissue level metric that captures TDP43 related splicing abnormalities across heterogeneous human spinal cord samples remains limited. In this study, we developed a machine learning based framework to quantify alternative splicing dysregulation associated with ALS and to enable individual level disease assessment.

Methods: Alternative splicing events were first defined in a low heterogeneity induced pluripotent stem cell model with TDP43 knockdown. These splicing features were then integrated using nonlinear machine learning and projected onto postmortem human spinal cord transcriptomes from sporadic ALS patients and controls. Each sample was summarized as a single splicing-based score that reflects TDP43 loss of function related RNA processing defects.

Results: The splicing score was robust across independent cohorts and spinal cord regions and reliably distinguished sporadic ALS from healthy controls. In inherited ALS, the score recapitulated expected differences across genetic subtypes. Importantly, the score was associated with patient survival and neuropathological measures and showed improved performance compared to single marker splicing metrics.

Conclusions: Samples with high splicing scores exhibited coordinated gene expression changes involving inflammatory pathways as well as cytoskeletal and cilia related processes, supporting a biologically meaningful axis of disease heterogeneity. Together, this work demonstrates that alternative splicing informed machine learning models provide an interpretable and scalable approach to quantify ALS molecular pathology and link core RNA processing dysfunction to clinically relevant disease burden.