Deep Learning-based Prediction Of Protein Acylation Sites In Rice

Post-translational modification of proteins is involved in almost all cellular regulatory processes in organisms,and has a very important regulatory role.The posttranslational modification of rice protein is related to the process of rice metabolism and starch synthesis,which in turn affects the yield of rice,and the yield of rice is closely related to people’s lives.Therefore,it is of great value to determine whether the posttranslational modification of lysine occurs in rice proteins.The traditional way of judgment is through biological experiments,which requires high time and economic cost.The emergence of high-throughput mass spectrometry technology has brought more abundant experimental data to the problem of protein post-translational modification,which also provides a data basis for researchers to predict protein post-translational modification sites by computational methods.However,the existing prediction models have a strong dependence on feature engineering,and cannot make good use of the multisource features obtained by various feature extraction methods.To solve this problem,this thesis proposes some new prediction frameworks through deep learning methods.The main contributions of this thesis are as follows:（1）Propose a method for predicting post-translational modifications of lysines in rice proteins based on multi-source feature selection by applying table networks.The method is based on multi-source feature selection and screening of redundant features,constructs an optimal sub-feature set,and realizes the prediction of rice protein posttranslational modification sites on the optimal sub-feature set through a table network.In order to solve the feature interference between different feature extraction methods in multi-source feature sets,this thesis improves the framework and proposes a method for predicting post-translational modifications of lysine in rice proteins combined with multichannel attention mechanism by applying table networks.The improved method extracts information from feature subsets from different sources in the multi-source features respectively,and adjusts the features of the multi-channel table network with the weights learned by the channel attention module,and achieves better prediction performance.（2）Propose a multi-scale amino acid word vector method.The method takes into account the multiple scale information of amino acids in protein sequences at the same time,and can obtain more abundant amino acid vector expressions from protein sequences.The experimental results show that the multi-scale amino acid word vectors can improve the prediction performance of the model for six types of post-translational modification sites on rice proteins,compared with the vector representations learned from a single amino acid.（3）Propose a rice protein post-translational modification prediction method using multi-head attention.The method framework uses multi-scale amino acid word vectors as input to construct a multi-head attention-based prediction model,which can learn attention weights among channels of amino acid word vectors of different scales and adjust the characteristics of each channel.The experimental results show that the rice protein PTM prediction framework based on multi-head attention can achieve better prediction performance than the existing protein site prediction models on the problem of PTM site prediction in rice proteins.The AUC index of this model is 3.0%-10.3% higher than the existing model,and the accuracy is 2.0%-4.0% higher than the existing model.