Decoding splicing variants in high-throughput sequencing: a functional validation approach integrating deep learning tools.

High-throughput genome and exome sequencing have uncovered numerous intronic variants in disease genes, yet predicting their impact on pre-mRNA splicing is still challenging. While deep learning tools have improved variant prioritization, their ability to detect atypical regulatory mechanisms remains unknown. We combined predictions from multiple complementary splicing prediction tools with functional RNA studies to evaluate the splicing consequences of intronic variants identified in patients with neurodegenerative diseases. Nine intronic variants with uncertain impact on splicing were selected, including six deep-intronic variants, two non-canonical splice-site variants, and one canonical splice-site variant associated with an atypical phenotype. Three variants were predicted to disrupt a donor splice site, five to create a cryptic donor splice site, and one was located close to a putative acceptor site. Transcript analysis in blood revealed splicing abnormalities in eight of the nine cases. All algorithms failed at least once: deep learning tools missed specific donor site losses or cryptic splice-site activation events, while motif-based approaches failed to predict some donor site activation or generated false-positive predictions. The only variant not directly located in a splice site was predicted to activate a deep intronic splicing enhancer. Motif-based tools suggested the creation of an SRSF2 protein-dependent enhancer. Functional studies based on minigene assays support a role for SRSF2 in promoting pseudoexon inclusion. These findings highlight the complementarity of splicing prediction tools and the need to integrate them into diagnostic pipelines, with transcript-level confirmation remaining critical for accurate pathogenicity assessment.