Study Of Prediction Algorithm Of Protein Post Translation Modification Sites Based On Machine Learning

Protein post-translational modification is involved in almost all cellular activities,which is a key mechanism to increase protein diversity and plays a very important role in regulating the structure and function of proteins.In addition,the abnormality of posttranslational modification may cause protein dysfunction and major diseases,and a variety of regulatory enzymes involved in post-translational modifications have become potential drug targets.Therefore,the in-depth study and analysis of posttranslational modification has important value for revealing the regulation of cell activity and finding drug targets.In the past decades,a large number of experimental methods for identifying posttranslational modification sites have emerged,which has promoted the development of post-translational modification site prediction research.However,experimental identification methods are time-consuming and labor-intensive.In the face of rapidly increasing protein data,it is difficult to meet the needs of scientific research only by relying on experimental methods to identify protein post-translational modification sites.Therefore,it is necessary to develop a computational method that can quickly predict a large number of post-translational sites.However,most of the current computational methods are limited to protein sequence features,resulting in limited performance in predicting post-translational modification sites.Recent studies have shown that the cross-talk information between different post-translational modification sites is also important to the prediction of post-translational modification sites.Therefore,we construct a site-specific modification profile that can effectively encode post-translational modification cross-talk information,and combine protein sequence to predict post-translational modification sites.In addition,we propose a deep transfer learning method based on domain-adaptive technique to solve the problem of small sample size in species-specific ubiquitination site prediction.The main contributions of this dissertation are as follows:(1)We construct a novel site-specific modification profile by using various posttranslational modification sites collected from multiple databases.Furthermore,we propose a post-translational modification site predition method called PTM-ssMP by combining protein sequence information and site-specific modification profile.In this dissertation,corss-validation and the independent test set are used to evaluate the performance of PTM-ssMP,which is also used for comparing with existing posttranslational modification site prediction methods.The results show that site-specific modification profile can effectively improve the accuracy of post-translational site prediction and PTM-ssMP achieves significant imporvemt compare with other existing methods.In addition,we also develop an online tool based on the PTM-ssMP method for the prediction of multiple post-translational modification sites.(2)Furthermore,we propose a deep learning based method Deep-ssMP that combines protein sequence and site-modification profile to predict post-translational sites.Deep-ssMP is consist of feature extraction sub-network and feature fusion subnetwork,which are used to extract the features of protein sequences and sitemodification profiles and effectively fuse the above features respectively.Experimental results show that Deep-ssMP is superior to other common deep neural networks and existing post-translational modification site prediction methods.(3)We propose a deep transfer learning method,DeepTL-Ubi,to solve the problem of small sample size in species-specific ubiquitination sit prediction.In order to achieve effective knowledge transfer,DeepTL-Ubi adopts a novel siamese network to learn the common feature space,and at the same time,the species transfer loss function is designed to reduce the distribution difference of ubiquitination samples of different species.Experimental results show that DeepTL-Ubi is superior to other existing ubiquitination site prediction methods.