SpliceAI’s Precomputed Scores Miss Splice-Altering Variants in Thousands of Rare Disease Genomes

SpliceAI, a widely used deep learning algorithm that predicts whether genetic variants are likely to disrupt splicing, has a hidden weakness in the precomputed scores that many diagnostic pipelines rely on, according to a study published in Genome Medicine. Researchers led by Alexandra C. Martin-Geary and Ruebena Dawes found that outdated gene annotations and genome version inconsistencies affect 8.34% of the theoretical single nucleotide variants covered by the precomputed scores, and compromise prediction accuracy in nearly 35% of disease-associated genes on the PanelApp green list.

SpliceAI’s precomputed predictions, released in 2019, cover all theoretical SNVs and small indels and are embedded in variant annotation tools such as VEP and dbNSFP, allowing laboratories to batch-annotate large numbers of variants quickly and cheaply. The team compared the original gene annotation file used to generate these scores against MANE v1.0 Select transcripts and identified transcript annotation and genome build inconsistencies that had gone largely unexamined. The impact of these factors on the prioritisation of likely splice-altering variants in cohort studies had not previously been established.

To quantify the consequences for diagnostics, the researchers re-analysed variants in 660 genes linked to neurodevelopmental disorders identified in 7,221 participants with these conditions from the Genomics England 100,000 Genomes Project. When they re-ran SpliceAI with updated annotations and parameters, the number of predicted splice-altering variants rose by 18.2% compared with variants identified using the precomputed scores alone. Assessed retrospectively against confirmed diagnoses in the 100,000 Genomes Project, these additional predictions corresponded to an 11.7% increase in diagnostic yield relative to what the precomputed scores would have flagged.

The findings suggest that laboratories relying solely on the 2019 precomputed scores may be missing clinically relevant splice-altering variants in patients with rare genetic disorders. Because the precomputed resource remains a useful and efficient tool for annotating variants in large cohorts, the authors do not call for abandoning it; instead, they document specific, correctable limitations that can cause clinically important predictions to be overlooked in variant prioritisation pipelines.

As a practical remedy, the team released SpliceAI-splint, a command-line tool that lets users scan a VCF file for variants whose potential splice-altering effects would be missed by the precomputed scores. By flagging these variants for full re-scoring with current annotations, the tool aims to close the gap between the convenience of batch annotation and the accuracy of up-to-date, transcript-aware SpliceAI predictions, offering a mitigation path for diagnostic and research pipelines built around the 2019 resource.

Subject of Research: Limitations of precomputed SpliceAI scores in prioritising splice-altering variants in rare disease genomes

Article Title: Using SpliceAI to triage splice-altering variants in 7,220 individuals with rare conditions highlights limitations of the precomputed scores

Article References: Martin-Geary, A. C., Lecoquierre, F., Walker, S., Whiffin, N., & Dawes, R. (2026). Using SpliceAI to triage splice-altering variants in 7,220 individuals with rare conditions highlights limitations of the precomputed scores. Genome Medicine. https://doi.org/10.1186/s13073-026-01797-6

Image Credits: AI Generated

DOI: 10.1186/s13073-026-01797-6

Keywords: SpliceAI, splice-altering variants, precomputed scores, Genomics England, 100,000 Genomes Project, neurodevelopmental disorders, variant annotation, MANE transcripts, diagnostic yield, SpliceAI-splint, Genome Medicine, rare genetic disorders