Protein And RNA Flexibility Prediction Based On Point Cloud Convolutional Neural Network

As everyone knows that proteins are the carriers of life.Protein flexibility plays important roles in various biochemical processes in the living organisms,prediction of the intrinsic flexible motions based on the tertiary structure of proteins is helpful for our understanding of the mechanism of protein functions.Protein flexibility is an important scientific problem in the research field of protein structure-function relationship.Compared with proteins,RNAs flexibility prediction is relatively late,and their structure is also complex to analyze.The development of mathematical models and tools to effectively predict proteins and RNAs flexibility can help us to understand the spatial structure and biological functions of proteins and RNAs.Also,it can provide important information for design of drug and nanomolecular shapes as well as the study of cancer and other related diseases,has important practical application value.Convolutional neural network(CNN)is one of the most popular models in deep learning,it has strong performance in extracting feature information of different levels.CNN has been successfully used to a number of research topics and is a very popular machine learning algorithm.In recent years,CNN method has been successfully applied to the study of proteins flexibility,and has attracted more and more attention from the majority of biological researchers.In the present work,based on the idea of Point Net method developed in the computer vision research,a CNN model based on point cloud was proposed to predict the protein and RNA flexibility.In this model,the point clouds introduced into the network was expressed with the atomic coordinates of proteins and RNAs,and in order to keep the global rotation invariance and permutation invariance of the point cloud,we used a spatial transformation network and the pooling operations,respectively.In addition,considering the varied sizes of different proteins and RNAs,a new mini-batch optimization strategy was proposed,the strategy aims to introduce the mini-batches of protein and RNA molecules containing different number of residues into the network.The Pearson correlation coefficient was used as the evaluation function for the training of the model.Besides that,in order to further enhance the performance of the network,in the part of symmetric pooling operation,the max-pooling and the average-pooling were concatenated to better extract the global features of protein and RNA structures.In order to verify the validity of the model,it was iteratively trained by proteins and RNAs data.The results show that the mean Pearson correlation coefficients between the predicted and experimental temperature factors of proteins and RNAs data sets by the proposed model are better than the widely used Gaussian network model.