DeepTESite: The Prediction of Protein Arginine Methylation Sites Using Amino Acids Sequence Symmetric Position Encodings Based on Transformer Encoder

Methylation is a type of Post-translational modification, which is closely associated with various diseases. The methylation sites prediction is very important for revealing methylation molecular mechanism. The deep learning model Transformer and the Multi-Head Attention Mechanism have been applied and have demonstrated promising performance in sites prediction. However, the position encoding of Transformer model don't consider the structure particularity of protein sequences, and the Multi-Head Attention Mechanism insufficiently focuses on unidirectional flow of information within extensive genomic sequences, resulting in the challenging of accuracy and position encoding computation. To solve the above problems, we propose DeepTESite model based on Transformer Encoder. In our DeepTESite model, Amio Acids Sequence Symmetric Position Encodings is first proposed based on the methylation symmetry hypothesis that arginine methylation of some histones show a certain spatial or functional symmetry in specific sequences for halving the position encoding computation. Then the Bidirectional Multi-Head Attention Mechanism is applied to extract features in sequential and spatial information. The comprehensive experiments demonstrate that our DeepTESite outperforms existing state-of-the-art methylation sites prediction methods, which achieves an accuracy of 87.88% and enhances computational complexity. Experimental results based on our proposed DeepTESite in phosphorylation site prediction further show significant improvement compared to other models. These findings suggest that first proposed Amio Acids Sequence Symmetric Position Encodings is reasonable and provides a promising solution for protein arginine methylation sites prediction.