Learning the rules of RNA splicing

Splicing removes introns at single-nucleotide precision, yet the GT and AG that mark most intron ends occur millions of times in a genome; they are necessary, not sufficient. Branch points, splicing enhancers and silencers, and broader sequence context decide which sites are used and how often. Alternative choices create isoforms; wrong choices disrupt proteins and cause disease. SpliceAI (Jaganathan et al., 2019) showed that a deep network reading 10,000 nucleotides of flanking context can recognize splice sites from sequence alone, and Pangolin (Zeng and Li, 2022) extended prediction to tissue-specific usage.

I treat a splice-site model as a component of genomics workflows. Its output should improve a concrete decision: which RNA-seq junction to keep, which gene model to trust, which variant to follow up.

Questions I’m pursuing

What should the model recognize? Scanning a chromosome for splice sites and judging a donor–acceptor pair proposed by an aligner are different tasks. The input, the training data, and the evaluation should match the decision being made.

How do we separate new biology from noise? An unannotated junction in RNA-seq may be a real isoform or an alignment artifact, and read count alone does not settle it. Sequence gives evidence independent of the alignment, and the two are stronger together.

How far does a model travel? Species differ in intron length, sequence composition, and annotation quality. A human model carries some signal elsewhere; a retrainable one can learn the target species directly.

What I have built

Splam: scoring junctions in RNA-seq alignments

Splam (Chao et al., 2024) evaluates a candidate intron as a unit. It reads 400 nt centered on the donor and 400 nt centered on the acceptor, scores both sites in one pass, and reports the weaker of the two. Because it was trained on junctions from multiple annotated isoforms, alternative splicing counts as signal rather than error. In 20 human RNA-seq samples, removing low-scoring junctions raised intron precision by 14 to 37 percentage points at a cost of about 1% recall or less, and improved the transcripts assembled from the cleaned alignments. Without retraining, it recalled 91%, 87%, and 80% of junctions in chimpanzee, mouse, and Arabidopsis, against 57%, 47%, and 20% for SpliceAI at the same threshold.

OpenSpliceAI: a retrainable SpliceAI

OpenSpliceAI (Chao et al., 2025) reimplements SpliceAI in PyTorch and exposes the full cycle: data preparation, training, transfer learning, calibration, genome-scale prediction, and variant scoring. In mouse, zebrafish, honeybee, and Arabidopsis, models trained on the target species outperformed the human SpliceAI model, and transfer from a human checkpoint converged faster and more stably than training from scratch in the settings we tested. The point is a model anyone can inspect, reproduce, and retrain.

Future directions

Recognizing a splice site is the first step. What I want to model is splicing as a quantity: how often each site and junction is used, how those choices combine into transcripts, and how a variant shifts them in the tissue where it matters. A splice score is evidence, not a verdict. It earns its place when it is calibrated, combined with RNA-seq rather than substituted for it, and judged by whether the transcripts and variant calls built on it are better.