1977 was when the trail went cold: researchers caught RNA being cut and reassembled in ways the old gene manuals had not warned them about, and in the nearly 50 years since, dozens of motif scanners, statistical gumshoes, neural nets, and transformer models have tried to predict the splice job before biology pulls it off. Most got a piece of the story. None closed the case.

The victim, if you want to call it that, is pre-mRNA. It starts life as a messy transcript, all useful exons and disposable introns, like a ransom note assembled by a committee. The spliceosome comes in with molecular scissors, removes the introns, and stitches the exons together. Alternative splicing makes the plot thicker: one gene can produce different RNA messages depending on tissue, timing, disease state, and cellular mood. Biology loves options. Clinicians love certainty. That is where the trouble starts.

In their 2026 Nature Genetics Review, Ning Shen, Ningyuan You, and Chang Liu survey the whole precinct: early splicing heuristics, deep-learning models, training-data choices, output resolution, event quantification, and the stubborn mess of translating predictions into medicine (DOI: 10.1038/s41588-026-02629-4). It reads less like a victory lap and more like a detective’s wall: red string, coffee rings, and one suspicious intronic variant standing under a streetlamp.

The Old Clues Were Too Short

Early tools watched for obvious splice signals, especially the donor and acceptor sites near exon-intron borders. Useful, sure. But biology does not commit crimes only near the front door. Variants can lurk deep in introns, nudge regulatory motifs, awaken cryptic splice sites, or cause exon skipping. The genome is not a tidy spreadsheet. It is a warehouse after a power outage.

Deep learning changed the surveillance game. SpliceAI showed that a model could look across long DNA context and predict donor and acceptor sites from primary sequence (Jaganathan et al., 2019). Pangolin pushed toward tissue-specific splice strength. These systems do not “understand” RNA in the human sense. They are very expensive pattern hounds. Still, give them enough sequence and RNA-seq evidence, and they start finding footprints older tools missed.

The Suspects Wear Transformers Now

The newer models have gotten flashier coats. Transformer-based systems borrow the attention trick from language models: decide which parts of a long sequence deserve scrutiny. If a neural network were a detective squad, attention would be the one person who actually read every witness statement before accusing the butler.

Recent work shows why this matters. One 2024 transformer model scanned raw 45,000-nucleotide sequences and beat SpliceAI on splice-site detection benchmarks (Jónsson et al., 2024). SpliceTransformer tackled tissue-specific splicing and disease-linked variants, reporting tissue-aware signals across ClinVar-scale variant sets (You et al., 2024). Splam took another angle, pairing donor and acceptor sites like the spliceosome actually does, then using deep learning to clean up spliced alignments (Chao et al., 2024).

That last detail matters. Models that predict one nucleotide at a time can miss the chemistry of the room. Splicing is a relationship problem. Donors and acceptors need each other, like crooked partners in a rain-soaked alley.

The Medical Motive

Why chase this case? Because splicing errors show up in rare disease, cancer, neurological disorders, and variant interpretation. A DNA change that looks harmless on paper can wreck an RNA transcript in practice. That is the kind of quiet villain medicine hates.

The Review highlights two big translational leads. First: variant annotation. Tools such as SpliceVault use large RNA-seq resources to predict what kind of mis-splicing a variant may cause, not just whether something smells off (Dawes et al., 2023). Second: therapy design. Antisense oligonucleotides can bind RNA and redirect splicing, basically slipping the spliceosome a new set of instructions. Recent AI/ML work has even designed splice-switching oligonucleotides and validated candidates in triple-negative breast cancer models (Fronk et al., 2024).

If these results keep holding up outside benchmark alleys, the payoff is practical: fewer mystery variants, faster prioritization for lab testing, better ASO target selection, and a cleaner path from genome sequence to clinical hypothesis. Not magic. More like a better flashlight.

The Case Is Still Open

The hard parts remain hard. Deep-intronic mutations still wear disguises. Isoform-level reconstruction is still a foggy dock at midnight. Multimodal integration - DNA sequence, RNA-seq, long reads, tissue context, chromatin, protein binding - asks models to juggle knives while paying rent. And as Shen and colleagues stress, bigger architectures can improve accuracy while making interpretation and compute costs uglier.

The field is moving from “Where is the splice site?” to “What transcript appears, in which tissue, under which condition, and can we safely change it?” That is a much better question. Also a much meaner one.

The loss curve told a story. It was not a happy one. But for once, the detectives have better shoes.

References

• Shen, N., You, N. & Liu, C. “Advances and challenges of splicing prediction with AI.” Nature Genetics (2026). DOI: 10.1038/s41588-026-02629-4. PMID: 42350808
• Jaganathan, K. et al. “Predicting splicing from primary sequence with deep learning.” Cell 176, 535-548 (2019). DOI: 10.1016/j.cell.2018.12.015
• Dawes, R. et al. “SpliceVault predicts the precise nature of variant-associated mis-splicing.” Nature Genetics 55, 324-332 (2023). DOI: 10.1038/s41588-022-01293-8
• Jónsson, B. A. et al. “Transformers significantly improve splice site prediction.” Communications Biology 7, 1616 (2024). DOI: 10.1038/s42003-024-07298-9
• You, N. et al. “SpliceTransformer predicts tissue-specific splicing linked to human diseases.” Nature Communications 15, 9129 (2024). DOI: 10.1038/s41467-024-53088-6
• Chao, K. H. et al. “Splam: a deep-learning-based splice site predictor that improves spliced alignments.” Genome Biology 25, 243 (2024). DOI: 10.1186/s13059-024-03379-4
• Fronk, A. D. et al. “Development and validation of AI/ML derived splice-switching oligonucleotides.” Molecular Systems Biology 20, 676-701 (2024). DOI: 10.1038/s44320-024-00034-9

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.