Abstract: DNA-N6 methyladenine (6-mA) methylation modification is one of the most important epigenetic modification markers. The aberrant 6-mA modification can affect gene expression and lead to serious diseases. Therefore, the work of predicting the 6-mA site is of great significance for the understanding of the pathogenesis and treatment of diseases. In this paper, a long short-term memory (LSTM) neural network based on K-mer encoding method and one hot encoding method is proposed to predict methylation sites.Firstly, the information content of gene sequence is increased through K-mer coding method. Secondly, the information content after one hot encoding is converted into a composite encoding matrix. The LSTM model can extract more feature dimensions and types from the encoding matrix, to improve the prediction performance of the LSTM model for gene sequence. The cross validation experiment show that the proposed method can achieve accuracy of 93.7% on benchmark datasets. The sensitivity, specificity and matthews correlation coefficient of the trained model were 93.0%, 94.5% and 0.875, which outperformed existing 6-mA prediction methods. On the other six different species datasets, the proposed method can achieve the area under the curve (AUC) values from 0.9055 to 0.9262,which shows the applicability of the proposed method on animals, plants and microorganisms methylation tasks. The proposed method was applied on rice gene NC_ 029258.1, and the predictions were verified by the recently published online prediction tools. The results show that 86% to 96% of the prediction results are supported by these tools, indicating that the proposed method can be applied to large-scale site prediction and analysis of different species.

Key words: methylation site prediction, deep learning, long short-term memory network, compound features