DeepSAP

Aligning RNA-seq reads is deceptively hard. Reads that span exon-exon boundaries must be split across introns that can be tens of kilobases long, and an aligner has to decide, from a handful of bases of overhang, whether a candidate gap is a real splice junction or an artifact. Conventional spliced aligners resolve this with hand-tuned penalty schemes over canonical GT-AG dinucleotides, which works well for abundant, annotated junctions and poorly for novel, rare, or non-canonical ones.

DeepSAP, from NVIDIA's Clara Parabricks team, attacks the problem from two directions at once. It first aligns reads with Transcriptome-Guided Genomic Alignment (TGGA) in GSNAP, which tries to place a read against a known transcriptome before falling back to the genome, and then rescores every candidate junction with a transformer that reads the raw sequence context around the donor and acceptor sites. The two components are complementary: TGGA is highly sensitive to introns near read ends, while the transformer supplies a learned, annotation-free judgement about whether a given sequence context looks like a genuine splice site.

The scoring model is not trained from scratch. DeepSAP fine-tunes DNABERT, the 6-mer genomic language model, into a task-specific checkpoint called DNABERT_MS150, and ships it inside a Docker container so that users score their own FASTQ or BAM inputs without any training step. The preprint appeared in April 2025 and the peer-reviewed version was published in Genome Biology in 2026.

#Key Features

• Two-stage alignment and rescoring: Reads are aligned with transcriptome-guided GSNAP, then every reported junction is rescored by the transformer and the MAPQ values are recalibrated accordingly.
• Pretrained genomic language model backbone: The junction classifier is a fine-tuned DNABERT with 6-mer tokenization, reading a 150-base window centred on the candidate donor or acceptor motif.
• Multi-species fine-tuning: The shipped DNABERT_MS150 checkpoint was fine-tuned on splice junctions drawn from thirteen species, giving it broader transferability than the human-only variants.
• Zero-retrain deployment: The workflow is distributed as an NGC container that runs end-to-end alignment plus scoring, or scores pre-existing BAM files from any aligner emitting the XA tag.
• GPU-accelerated throughput: Junctions are scored in batches of 2,048 with fp16 inference, validated on A100, H100, and RTX A6000 hardware at human-genome scale.

#Technical Details

The authors benchmarked three pretrained genomic transformers as scoring backbones — DNABERT 6-mer, DNABERT-2, and the 2.5B-parameter Nucleotide Transformer — across window sizes of 90, 150, 200, and 400 bases and across annotation sources including GRCh38 RefSeq, GENCODE release 44, a Plasmodium set, and a thirteen-species multi-species set. Training examples pair labelled donor and acceptor motifs with negative loci sampled to match the splicing-signal composition and count of the positives; the RefSeq90 set contains 630,424 sequences and RefSeq400 contains 58,760. DNABERT consistently beat DNABERT-2: Matthews correlation coefficients on held-out data were 0.965 for DNABERT RefSeq150, 0.958 for DNABERT MS150, 0.964 for the Nucleotide Transformer, and 0.934 for DNABERT-2 RefSeq150. MS150 was selected for release because its multi-species training clustered true junctions more cleanly, particularly on the malaria data.

On the Baruzzo simulated human benchmark, the full DeepSAP workflow reached the highest mean F1 score for splice junction detection at 0.971, ahead of DRAGEN (0.933), novoSplice (0.914), STAR (0.821), Subjunc (0.770), and HISAT2 (0.662), with additional gains in indel identification.

#Applications

DeepSAP targets production RNA-seq pipelines where junction-level accuracy matters: transcript quantification and isoform assembly, novel splice-variant discovery, fusion detection, and clinical transcriptome analyses where a missed or spurious junction propagates into a variant call. Because the container also accepts pre-aligned BAM files, groups already committed to another aligner can apply the transformer rescoring as a post-processing step rather than replacing their alignment stage.

#Impact

DeepSAP is a practical demonstration that genomic language models can be dropped into the interior of a mature bioinformatics workflow — not as a replacement for the aligner, but as a learned scoring function that supplies exactly the signal the heuristic was approximating. Its ablation across DNABERT, DNABERT-2, and the Nucleotide Transformer is also a useful public data point: the largest or newest backbone was not the best one for this task. Openness is mixed. The GitHub repository holds the inference scripts and manuscript data under Apache-2.0, but the DNABERT_MS150 checkpoint is distributed only inside the NGC container under the NVIDIA Models Community License, so the weights cannot be inspected or fine-tuned independently, and the fine-tuning notebook ships with its code cells stripped.