Background: It is estimated that up to 50% of all disease causing variants disrupt splicing. Due to its complexity, our ability to predict which variants disrupt splicing is limited, meaning missed diagnoses for patients. The emergence of machine learning for targeted medicine holds great potential to improve prediction of splice disrupting variants. The recently published SpliceAI algorithm utilises deep neural networks and has been reported to have a greater accuracy than other commonly used methods.
Methods and findings: The original SpliceAI was trained on splice sites included in primary isoforms combined with novel junctions observed in GTEx data, which might introduce noise and de-correlate the machine learning input with its output. Limiting the data to only validated and manual annotated primary and alternatively spliced GENCODE sites in training may improve predictive abilities. All of these gene isoforms were collapsed (aggregated into one pseudo-isoform) and the SpliceAI architecture was retrained (CI-SpliceAI). Predictive performance on a newly curated dataset of 1,316 functionally validated variants from the literature was compared with the original SpliceAI, alongside MMSplice, MaxEntScan, and SQUIRLS. Both SpliceAI algorithms outperformed the other methods, with the original SpliceAI achieving an accuracy of ∼91%, and CI-SpliceAI showing an improvement at ∼92% overall. Predictive accuracy increased in the majority of curated variants.
Conclusions: We show that including only manually annotated alternatively spliced sites in training data improves prediction of clinically relevant variants, and highlight avenues for further performance improvements.