A DEEP LEARNING-BASED APPROACH FOR ALTERNATIVE SPLICING EVENTS PREDICTION IN THE HUMAN GENOME USING GENOMIC SEQUENCE DATA

Abstract

The generation of messenger RNA (mRNA) transcripts and protein isoforms from a single gene represents an example of basic post-transcriptional regulation mechanism, which greatly contributes to transcriptomic and proteomic diversity in eukaryotic species. Accurate prediction of alternative splicing events is essential for understanding regulation of the gene, cellular differentiation and the molecular causes of many complex diseases and genetic disorders. However, conventional experimental approaches used to detect splicing events (e.g. RNA sequencing and manual validation) are often costly, labor intensive and time consuming. Therefore, computer techniques are successful alternatives to extensive splicing analysis. In the past few years, deep learning algorithms have made tremendous strides in prediction accuracy and pattern recognition in biological sequence data. In this work, an improved deep learning framework for predicting alternative splicing events in the human genome is presented. Doing this without any special feature engineering by hand, the proposed model is based on convolutional neural networks (CNNs), which extract discriminative sequence characteristics from genomic data automatically. The experimental data consists of sequences of high and low inclusions, which are preprocessed and encoded prior to using the data for model training and evaluation. Widely used assessment criteria, such as the area under the receiver operating characteristic curve (AUC) and classification accuracy, are used to evaluate model performance. Experimental results show that the proposed approach achieves an AUC of 0.975, which is superior to other baseline methods reported in the literature. The results show that deep learning-based methods are useful tools for genomic research since they can produce extremely reliable and accurate predictions of alternative splicing events. Moreover, the proposed framework allows for precise and efficient prediction of alternative splicing mechanism, which can be used in future studies in computational genomics, precision medicine, disease gene analysis and bioinformatics.