RNA sequencing (RNA-seq) has opened up new frontiers in our understanding of gene expression and the complexities of how genes are regulated. One of the most significant breakthroughs in this field is the ability to examine how RNA is processed and spliced in various ways, providing insights into diseases, gene function, and even new therapeutic targets. However, the analysis of RNA-seq data comes with its own set of challenges. In particular, dealing with complex splice junctions and accurately identifying variations, like insertions and deletions (indels), can be very difficult.

Researchers have been working on improving the accuracy of RNA-seq data alignment, and one of the most promising approaches comes from a new method called DeepSAP. Developed by researchers at NVIDIA Corporation, DeepSAP uses an innovative combination of transcriptome-guided genomic alignment and advanced deep learning models to achieve highly accurate splice junction detection.

Overview of the DeepSAP workflow

DeepSAP utilizes the TGGA GSNAP aligner initially to align short-read RNA-seq data. The TGGA aligner tries three methods for transcriptome alignment, exact search, prevalence search, and extension search, to find a satisfactory solution, and otherwise resorts to genomic alignment. In the next phase, it incorporates the fine-tuned DNABERT MS150 transformer model to enhance splice junction detection. This second stage analyzes all alignments generated by TGGA GSNAP for each query read, recalibrates the MAPQ (mapping quality) scores for multi-mapped reads and applies soft clipping for splice junctions that exhibit either low transformer scores or flanking bases with suboptimal Phred quality scores.

The beauty of DeepSAP lies in how it uses a transformer-based deep learning model to score splice junctions more accurately. This model can capture intricate sequence patterns surrounding splice donor and acceptor sites, improving the overall interpretation of RNA-seq data. The method demonstrated a remarkable performance in benchmark tests, achieving the highest mean F1 score of 0.971 for splice junction detection, far outperforming other RNA-seq tools like DRAGEN, novoSplice, STAR, HISAT2, and Subjunc.

What makes this approach even more revolutionary is how it blends traditional transcriptome-guided alignment with cutting-edge machine learning, enabling researchers to better understand the complex splicing events that contribute to gene function and disease. This tool opens up a lot of possibilities for more precise analysis in both basic research and clinical applications, helping scientists tackle some of the most complex RNA-seq data challenges.

With DeepSAP, we are one step closer to making RNA sequencing an even more powerful tool for researchers and clinicians alike. The results could not only help advance our understanding of complex diseases but also improve how we develop new therapeutic strategies.

Availability – https://github.com/clara-parabricks-workflows/DeepSA