Protein-ATP Binding Site Prediction Based On 1D-convolutional Neural Network

In the process of life activities,the interaction between proteins and ligands is universal and important.It is meaningful to predict the interaction between protein and ligand for understanding the function of proteins and drug design.Among them,Adenosine-5’-triphosphate(ATP)is an important ligand,it can provide energy for living organisms by hydrolyzing and breaking high-energy phosphate bonds.It is direct energy source of most of living organisms.The method of determining protein ligand interaction and binding site by wet experiment has the problems of high cost and time-consuming.Therefore,there is an urgent need to develop a simple and efficient intelligent computing method to predict the binding sites of proteins and ligands and predicting binding sites of protein ligands based on machine learning has become a hot issue.In this paper,we focus on the prediction of protein-ATP binding sites for ATP,main work is as follows:(1)Study the binding relationship between protein and ATP,and proposes to use neural network with convolution layer to take advantage of the local relationship characteristics of ATP binding sites.According to the formation principle of binding pocket,protein secondary structure with structural information and solvent accessibility information are selected to participate in the prediction.The problem of data imbalance is found and solved by using the method of random sampling.(2)A method based on one-dimensional convolutional neural network is proposed to predict protein-ATP binding sites:firstly,extract features from multiple perspectivesthe,such as position specific score matrix information,secondary structure information and solvent accessibility information,and then express the features of each residue;secondly,in order to cope with the serious imbalance between positive samples(binding sites)and negative samples(non-binding sites),use random under-sampling technique to build a balanced training dataset.Finally,on the balanced dataset,one-dimensional convolutional neural network model is designed to predict the ATP binding sites of protein.Strict cross-validation and independent test are carried out on the standard data set,and the results are compared with the mainstream protein-ATP binding sites prediction model in the field.The experimental results verify the effectiveness of the proposed method.(3)A prediction method of predicting protein ATP binding sites based on onedimensional deep network is proposed.Based on the previous experimental method,the length uncertainty of training data is solved by using residual network,and the restriction of sliding window is solved by using the whole protein sequence to participate in training each time.Experimental results show that the deep network is effective for protein ATP binding.