| Protein is a complex biopolymer with amino acids as its basic unit,it usually interacts with other proteins to achieve protein-related functions.Protein-protein interaction sites are essentially the amino acid residues interacting on the protein surface.It is beneficial to understand the principle and mechanism of protein-protein interaction by studying which residues are involved in the interaction process.The identification of protein-protein interaction sites can be applied to studies such as deciphering new drug targets and drug design,and it is one of the important topics in molecular biology and bioinformatics.The research based on traditional biological experimental methods is very timeconsuming and labor-consuming,and the prediction ability is limited.With the rapid rise of machine learning and the continuous accumulation of protein sequence data,computational methods have gradually become a hotspot in the field of bioinformatics.However,there are far fewer interacting residues in human body than non-interacting residues.In other words,serious class-imbalance exists in protein sequence data,which limits the prediction ability of the classification model.Based on the above background,this paper proposes a protein-protein interaction site prediction method based on deep learning.The main contents are summarized as follows:1.A down sampling technique based on Near Miss algorithm is proposed to eliminate the impact of class imbalance.Firstly,the feature representation is generated for the acquired protein sequence data,and each residue is encoded into a vector with a specific length by using the physical and chemical properties of protein.After the extracted features are fused,the label of each residue is combined to form the input dataset of the model.In order to solve the problem of class imbalance,the Near Miss algorithm is adopted to downsample the input dataset.By measuring the distance between positive and negative samples,the representative majority samples are selected to keep the number of samples consistent with the minority samples,so as to achieve the balance of data distribution.2.A deep learning method combining improved deep autoencoder and multilayer perceptron(NM-DAEMLP)is proposed to realize the classification of target residues.The balanced dataset is fed into the constructed deep autoencoder,and sparse constraints are added to eliminate redundant features,so as to better extract the representative information in the data.Finally,the high-level abstract features of the hidden layer in the middle of the network are used for the downstream classification task.Considering that the features extracted by the deep autoencoder already contain enough information,a simple multilayer perceptron is added to identify the category of amino acid residues.The trained model is tested on the testset,and the evaluation indexes such as accuracy,recall and F1-measure are obtained.At the same time,two public datasets are used as independent verification sets to verify the robustness of the model,so as to achieve the best prediction effect.In order to verify the performance of the proposed model,NM-DAEMLP is compared with several commonly used machine learning classification algorithms and some existing models.The accuracy of NM-DAEMLP in two independent validation sets was 88.1% and 85.2%,the recall rate was 80.5% and 76.1%,the precision was98.2% and 93.0%,the F1-measure was 87.3% and 83.7%,and the Matthews correlation coefficient was 80.7% and 75.2%,respectively.Experimental results show that NMDAEMLP is significantly superior to other classification algorithms,and the proposed model has better prediction performance and generalization ability than the existing prediction models.Therefore,the NM-DAEMLP model proposed in this paper can provide some reference value for the prediction of protein-protein interaction sites. |
