EPBAM: annotation-robust detection of alternative splicing events and a systematic benchmarking of current tools

Alternative splicing (AS) is a major driver of transcriptomic diversity, yet its computational analysis remains challenged by inconsistent event definitions across tools and by the sensitivity of quantification methods to annotation quality. Here, we present a comprehensive benchmark of seven AS detection tools-replicate Multivariate Analysis of Transcript Splicing (rMATS), Super-fast Pipeline for alternative splicing analysis, version 2 (SUPPA2), Modeling Alternative Junction Inclusion Quantification + Visualization Of Inferred Local splicing Alternatives (MAJIQ+VOILA), EventPointerST (EPST), Shiba, LeafCutter, and the newly introduced EventPointer BAM (EPBAM)-evaluated across three complementary datasets: reverse transcription PCR (RT-PCR) validated experiments, simulated data, and Lexogen Spike-In RNA Variant (SIRV). The inclusion of SIRV, which provide experimentally derived sequencing data with fully controlled ground truth, is central to our evaluation framework, enabling rigorous assessment of quantification accuracy under real technical conditions. We assess tool performance under three annotation scenarios-complete, incomplete, and overloaded-reflecting the annotation uncertainty commonly encountered in practice. EPBAM extends the EventPointer framework by enabling de novo event detection coupled with a coverage-corrected Ψ quantification strategy. Our results demonstrate that annotation quality substantially drives tool performance: annotation-dependent tools achieve peak accuracy under complete annotations, whereas EPBAM exhibits the most robust and consistent behaviour across annotation conditions, recovering unannotated events with low false discovery rates. No single tool dominates across all evaluation metrics, highlighting the importance of dataset diversity and standardized benchmarking practices for meaningful tool comparison in the AS field.