Prediction of m6A Reader Substrate Sites Using Deep Convolutional and Recurrent Neural Network

Abstract

N6-methyladenosine (m6A), one of the most common post-transcriptional mRNA modifications, has been proved to correlate with multiple biological functions through the process of binding to specific m6A reader proteins. Various m6A readers exist among the genome of human beings, however, owing to the scarce wet experiments related to this topic, the binding specificity of proteins was not elucidated. Therefore, a deep learning approach combined with CNN and RNN frameworks was generated to predict the epitranscriptome-wide targets of six m6A reader proteins (YTHDF1-3, YTHDC1-2, EIF3A). Additionally, layer-wise relevance calculation was conducted to obtain each input feature contribution and tried to explain the model training process. Finally, we achieved superior performance in the classification, with an average AUROC of 0.942 in EIF3A full transcript, higher than the typical conventional machine learning algorithms (SVM) under the same condition. Moreover, we quantified the most optimal sequence length (1001bp) during the m6A reader substrate prediction. This research paves the way for further RNA methylation target prediction and functional characterization of m6A readers.