CI-SpliceAI-Improving machine learning predictions of disease causing splicing variants using curated alternative...

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Up to 50% of disease-causing variants can disrupt RNA splicing.
• Current methods for predicting splice-disrupting variants are limited, leading to potential missed diagnoses.
• Machine learning offers a promising approach to enhance the prediction of splice-disrupting variants.

Purpose of the Study:

• To investigate if training SpliceAI on a curated dataset of validated splicing sites improves its predictive accuracy.
• To compare the performance of a retrained SpliceAI model (CI-SpliceAI) against the original SpliceAI and other established methods.

Main Methods:

• The SpliceAI algorithm was retrained (CI-SpliceAI) using only validated, manually annotated primary and alternatively spliced GENCODE sites.
• Gene isoforms were collapsed into a single pseudo-isoform for training.
• Predictive performance was evaluated on a new dataset of 1,316 functionally validated variants.

Main Results:

• Both SpliceAI algorithms outperformed MMSplice, MaxEntScan, and SQUIRLS.
• The original SpliceAI achieved an accuracy of approximately 91%.
• The retrained CI-SpliceAI model demonstrated an improved overall accuracy of approximately 92%.

Conclusions:

• Training machine learning models with manually annotated alternatively spliced sites enhances the prediction of clinically relevant variants.
• The findings suggest avenues for further improvements in splice-disrupting variant prediction.
• This refined approach holds potential for more accurate diagnoses in genetic diseases.