Alternative Splice Site Prediction with Deep Learning

Alternative splicing of mRNA is tightly regulated in different tissues and developmental stages and its disruption is one of the leading mechanisms that cause genetic disease in humans. Gene splicing depends on the precise interaction of hundreds of cis-motifs that control the action of trans-acting splicing factors. These regulatory factors have been investigated in great detail, but it remains challenging to predict the splicing pattern of a transcript from its sequence, which led to the development of computational “splicing codes” using machine learning methods. I present multiple models that learn regulatory sequence features ab initio using convolutional neural networks instead of requiring a hand-engineered feature library. The first model is trained to distinguish true splice sites from decoys and improves on the performance of a previous support vector machine model. Another model is trained to predict the splicing pattern of pairs of alternative 3′ splice sites, achieving 95% auROC on this task. When this model was applied to the DBASS3 database of genomic variants that activate cryptic splice sites, it was able to identify 63% of cryptic splice sites correctly. Next, I developed the competitive splice site model (COSSMO), which, for the first time, can predict complex, non-binary splicing patterns. COSSMO can analyze a variable number of alternative acceptor or donor splice sites, uses convolutional and recurrent neural networks to learn from directly from sequence, and was trained on over 15M sites. COSSMO achieves 77% accuracy on the task of identifying the dominant splice sites from, on average, 90 candidate sites and matches the true splicing pattern with an R² of 71.3%. On a dataset of deep intronic variants, COSSMO can identify cryptic splice sites that are located hundreds of kilobases within the intron. An ensemble model of MaxEntScan and COSSMO achieves an auROC of 92.6% at discriminating between cryptic splice site-activating variants and common variants nearby. Applying COSSMO to the HGMD dataset of disease-causing variants showed that thousands of missense or nonsense variants also disrupt splicing and that intronic variants up to a distance of 1000 nt affect splicing.Ph.D