Skip to content

Cryptic Splice Variants

Summary

Cryptic splice variants are non-canonical sequence modifications that introduce or destroy splice donor or acceptor sites at locations distant from the normal intron-exon boundaries. Although historically under-annotated, modern deep learning tools like SpliceAI have enabled systematic genome-wide predictions of these cryptic events, demonstrating that their phenotypic consequences are often indistinguishable from canonical predicted loss-of-function (pLoF) mutations.

Overview

Canonical splicing mutations typically affect the highly conserved $+1, +2$ donor or $-1, -2$ acceptor dinucleotides at exon-intron boundaries. In contrast, cryptic splice variants can occur deep within introns or exons. When activated, they lead to aberrant splicing, intron retention, or exon skipping, which can introduce premature stop codons or disrupt critical protein domains.

Characterization in Large-Scale ExWAS

In the exome-wide association study of blood lipids in 1,158,017 individuals (Koyama et al. 2026), the authors highlighted: - Identification: Cryptic splice variants were identified using SpliceAI with a Delta Score (DS) threshold $>0.8$. - Effect Size equivalence: The phenotypic effect size (median $\beta = 1.092$) of identified cryptic splice variants was equivalent to that of canonical pLoF variants ($P = 0.71$). - Splicing donor 5th base: A strong enrichment of cryptic splice variants was observed at the splice donor 5th base (e.g. APOA5 rs200831171, associated with significantly elevated triglycerides, $\beta = 1.10$).

Beyond Splice-Site Scoring: Junction-Level Modelling

Delta-score approaches such as SpliceAI answer one question — how much does this variant change the probability that a given base acts as a donor or acceptor? Splicing outcomes can in fact be modelled at three levels: site probability, competitive selection among candidate sites (splice site usage), and the identity and strength of specific introns (splice junction prediction).[3] Site-level predictors capture only the first, which is why exon-skipping and exon-extension events — where the disrupted base is not itself the whole story — are systematically harder to call.

AlphaGenome predicts all three levels plus RNA-seq coverage, and reports that its splice junction scorer alone outperformed prior methods on every splicing benchmark except deep-intronic/synonymous ClinVar and MFASS.[3] On ClinVar classification of deep intronic (>6 bp from a splice site) and synonymous (>3 bp) variants — the cryptic category proper — it reached auPRC 0.66 versus 0.64 for Pangolin.[3]

Interpretation: the improvement attributable to the junction scorer specifically indicates that a share of cryptic-splice pathogenicity is only visible when the competition between introns is modelled, not just the strength of individual sites. This has a practical consequence for pipelines that reclassify variants as pLoF on a delta-score threshold alone: such thresholds are calibrated on site-level evidence and will miss variants whose effect is on junction selection. Neither cited study evaluates the reclassification pipelines directly.[2][3]

See Also

Citations

  • [1] Jaganathan, K. et al. (2019). Predicting splicing from primary sequence with deep learning. Cell, 176(3), 535–548. DOI: 10.1016/j.cell.2018.12.015
  • [2] Koyama, S. et al. (2026). Exome-wide association study of blood lipids in 1,158,017 individuals from diverse populations. Nature Genetics, 58(6), 1268-1279. DOI: 10.1038/s41588-026-02613-y. Source paper: s41588-026-02613-y.pdf
  • [3] Avsec, Ž. et al. (2026). Advancing regulatory variant effect prediction with AlphaGenome. Nature, 649, 1206–1217. DOI: 10.1038/s41586-025-10014-0. Source paper: s41586-025-10014-0.pdf Supports: the three levels of splicing modelling; the junction-scorer ablation result; the ClinVar deep-intronic/synonymous auPRC comparison. Location: "Improved splicing variant predictions"; Figs. 3a, 3h.