Learning alternative RNA splicing with applications to mutagenesis

Thesis

Alternative RNA splicing is the process by which a single gene generates a large diversity of transcripts, but how the DNA sequence results in particular transcript abundances remains incompletely known. We present a two-stage machine learning algorithm that models this process by predicting the frequency of splice junction usage, with a particular focus on transcripts with low relative expression. We find that nonsense-mediated decay plays an important role in shaping transcript abundance, resulting in splice-altering mutations leaving no evidence in RNA-seq libraries. Using the model, we build a pipeline for annotating putative disease-causing mutations, and show state-of-art performance on HGMD mutations known to affect splicing. We also apply the model to common variation associated with disease phenotypes in GWAS, and show that variants annotated to affect splicing are significantly enriched, indicating that splicing plays a key mechanistic role in human disease aetiology arising from common genetic variation. These variants are found both near annotated splice-sites and in deep-intronic regions. Our model predicts that the deep-intronic variants often modify gene expression through activation of cryptic alternative splice sites followed by nonsense-mediated decay (AS-NMD), providing further evidence of AS-NMD as a common mechanism of gene regulation.

Authors: Sanders, E

• DOI: 10.5287/ora-dppeqe4xa
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: splicing; machine learning; NMD