Deep learning-based prediction of tissue-specific splice sites in the human neural retina

Abstract

Splice prediction tools can be used to identify splice-altering variants in patients with inherited diseases. Since splicing is tissue-specific, a predicted splice defect may vary depending on the tissue of interest. Current splice prediction tools often neglect splice junctions from the human neural retina, which can lead to the misinterpretation of benign and pathogenic variants in patients with inherited retinal diseases (IRDs). To address this issue, we developed a retina-specific splice prediction tool based on the model architecture of SpliceAI. In addition to retraining the existing model using splice junctions derived from retina RNA-sequencing data, we also improved the models generalization through hyperparameter optimizations such as dropout. The resulting retina-specific model was validated on retina-enriched exons and variants known to cause retina-specific splicing defects. Although the retina model correctly identified five more retina-enriched splice sites than the GTEx model, it did not enhance predictions for variants with tissue-specific splicing defects nor did it identify novel pathogenic variants in a set of genetically unexplained IRD cases. Despite these limitations, the retina-specific SpliceAI model shows promising results and should be applied to a larger patient cohort to uncover novel splice-altering variants in IRD patients.