Predicting exonization in the human genome with a deep learning model.

exonization, or the recruitment of intronic elements into gene sequences, has contributed to functional diversification; however, its extent and the ways in which it influences gene regulation are not fully understood. We developed an unbiased approach to predict exonization events from genomic sequences implemented in a deep learning model, eXAlu, that overcomes the limitations of tissue or condition specificity and the computational burden of RNA-seq analysis. The model captures previously reported characteristics of exonized sequences and can predict sequence elements important for exonization. Using eXAlu, we estimate the number of elements in the human genome undergoing exonization to be between 55-110K, 11-21 fold more than represented in the GENCODE gene database. Using RT-PCR we were able to validate selected predicted exonization events, supporting the accuracy of our method. Lastly, we highlight a potential application of our method to identify polymorphic insertion exonizations in individuals and in the population from whole genome sequencing data.