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

Background SpliceAI is a deep learning algorithm that predicts whether genetic variants are likely to affect splicing. Precomputed SpliceAI predictions for all theoretical SNVs and small indels were released in 2019, and are used by variant annotation tools such as VEP and dbNSFP to batch annotate large numbers of variants quickly. While the use of precomputed scores vastly reduces the computational cost of retrieving SpliceAI predictions, several limitations to their utility have emerged, for which the impact on the prioritisation of likely splice-altering variants in cohort studies has not been established. Methods We identify transcript annotation and genome version inconsistencies impacting the precomputed SpliceAI scores by comparing the original gene annotation file with MANE v1.0 Select transcripts. To quantify the impact that these factors may have in variant prioritisation pipelines for diagnostic purposes, we re-analysed variants in 660 genes linked to neurodevelopmental disorders (NDD), identified in 7,221 participants with NDD from the Genomics England 100,000 Genomes Project (100kGP). Results We find outdated gene annotations and genome version inconsistencies impact 8.34% of theoretical SNVs included in the SpliceAI precomputed scores, and affect the accuracy of SpliceAI predictions in 34.98% of disease-associated panelApp green genes. In 7,221 Genomics England participants, re-running SpliceAI with updated annotations and parameters results in an increase of 18.2% predicted splice-altering variants compared with variants identified using precomputed scores alone. These variants constitute a diagnostic increase of 11.7% versus the precomputed scores, assessed retrospectively using confirmed diagnoses in 100kGP. To mitigate issues with the precomputed scores we provide ‘SpliceAI-splint’- a command-line tool to allow users to identify variants within a VCF file whose potential splice-altering effects would be missed by the precomputed scores. Conclusions: Although the precomputed scores continue to represent a useful resource in the annotation of variants in cohorts with genetic disorders, we find key limitations that can lead to their missing clinically relevant splice-altering variants, and provide a command-line tool to mitigate these issues.