Deep learning models based on distributed feature representations for alternative splicing predictions

Alternative splicing (AS) is a fundamental step in mRNA maturation and gene expression. The advancement in RNA sequencing technologies has shed light on the role of AS in increasing protein isoform diversity. AS is recognized to be involved in the regulation of both physiological and pathological functions, hence it is an essential part of the study of gene regulation development and diseases. With the recent advances in machine learning, there is an interest in developing accurate deep learning based computational models for AS prediction. In this paper, we propose a convolutional neural network and multilayer perceptron models to tackle the AS prediction task as classification and regression. These models use feature representations learned from genomic data and cellular context. Unlike previous works which use hand-crafted feature extraction, we propose an automatic feature learning approach to avoid explicit and predefined feature extraction. The proposed approach is based on the adaptation of two extensively used natural language processing techniques, namely word2vec and doc2vec. In order to understand the effects of different representation learning techniques, many experiments have been conducted to predict AS based on the cassette exons and cell type. Overall, experimental results on five tissues data set prove that learning features from genome sequence add a significant improvement to AS outcome prediction in both classification and regression tasks.