Research On Prediction Of Protein-protein Interactions Based On Deep Neural Network And Ensemble Learning

Protein is the material basis of many lives,and the interaction between proteins is a prerequisite for most biological functions.At the same time,the study of protein interaction is conducive to drug and diseases discovery and promote drug research.Therefore,the study of protein interaction has become one of the contemporary research hotspots.The existing prediction of protein interaction is mainly including two methods: traditional highthroughput experimental methods and computational biology methods.The high cost and time-consuming become the main factors hinder the development of traditional experimental method.Thus,it is increasingly urgent to develop efficient computational methods for predicting protein-protein interactions.This paper proposes two computational methods for predicting protein interaction based on deep neural network and ensemble learning.The main research contents are as follows:(1)This paper proposes a computational method based on structural deep network embedding and random forest for predicting protein interactions.Firstly,the stack autoencoder is used to obtain the attribute features of macro-molecules.Secondly,the structural deep network embedding method is used to obtain another feature vector which called behavior features on the molecular heterogeneous network.After combining the attribute features and the behavior features,the random forest classifier is used to predict protein interactions.When predicting protein interactions,the results of average accuracy,precision,sensitivity,specificity and MCC come to be 83.12%,84.58%,81.02%,85.22%,and66.30%,respectively.We also compare the proposed method with multiple single classifiers.The comprehensive experimental results illustrate that the method proposed can effectively predict protein-protein interactions.(2)Predicting protein interactions based on protein sequences.This paper proposes the computational method based on local optimal orientation pattern and rotation forest for predicting protein-protein interaction.Firstly,each protein sequence is converted into a two-dimensional matrix through using the position-specific scoring matrix.Secondly,the local optimal orientation pattern is used to extract feature vectors from scoring matrix to reduce the noisy influence of the final result.Finally,the rotation forest is used to predict protein interactions.When predicting the data sets of Yeast,Human,and H.pylori,the results of average accuracy achieved 90.48%,94.14% and 94.89%,respectively.In order to better evaluate the effectiveness of the proposed method,we compared the results of proposed method with various classifiers.And we also compared the proposed method with previously works.A great variety of experiments demonstrate that the proposed methodbased on local optimal orientation pattern and rotation forest for predicting protein interactions can effectively predict protein interaction.