Abstract Accurate characterization of alternative splicing and gene fusion events from RNA-seq is crucial for understanding cancer biology, however, this task remains inherently challenging due to factors such as complex splice junction architecture, ambiguous read mapping caused by multi-mapped reads, and the presence of chimeric transcripts that confound standard analysis pipelines. We present DeepSAP, a splice-aware RNA-seq aligner that integrates Transcriptome-Guided Genomic Alignment (TGGA) as implemented in GSNAP with advanced Transformer-based Splice Junction Scoring (TSJS) to overcome these challenges and achieve greater sensitivity and specificity in detecting true splicing events. DeepSAP leverages a fine-tuned DNABERT model, trained on curated splice donor and acceptor site sequences drawn from multiple species, utilizing sequence windows of varying lengths (90, 150, 200, and 400 bp) around the splice junctions. Among the different fine-tuned models evaluated, DNABERT MS150 demonstrated superior performance in distinguishing true splice sites, and is directly integrated into the DeepSAP workflow to re-score and prioritize candidate splice junctions identified by GSNAP TGGA, effectively combining both sequence-driven plausibility in local DNA context and strongly supported by RNA-seq reads. To evaluate DeepSAP in a clinically relevant setting, we applied it to a prostate cancer RNA-seq cohort (PRJNA579899, University Hospital Zurich). In this cohort, DeepSAP consistently identifies numerous high-confidence, novel splice junctions absent from current gene annotations and frequently missed or only weakly detected by current state of the art RNA-seq aligners. Notably, in FOXA1, DeepSAP resolves a novel donor site with a double-adenine substitution that creates a previously unannotated exon-intron boundary. This junction shows coherent read coverage and high transformer derived donor and acceptor probabilities, whereas alternative aligners fail to produce a consistent splice junction at the same locus. A similar pattern is observed in the ERG oncogene, where DeepSAP recovers an additional complex, unannotated splice junction that is not captured by other aligners. The results highlight DeepSAP’s ability to recover complex, previously undescribed splice junctions in tumor RNA-seq data and demonstrate that coupling GSNAP TGGA with TSJS substantially improves alignment sensitivity and specificity, enabling improved detection of candidate oncogenic splice events in cancer RNA-seq samples. Citation Format: Pankaj Vats, Fadel Berakdar, Thomas D. Wu, Tong Zhu, Mehrzad Samadi. Uncovering novel splice junctions in prostate cancer using DeepSAP abstract. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 1418.
Vats et al. (Fri,) studied this question.