Research On Protein-Protein Interactions Prediction Algorithm With Deep Learning Methods

Most organisms carry out the cell activities by protein-protein interactions(PPI).The deep research about PPI can help humans understand the mechanism of life activities and apply the protein functions better.Nowadays,PPI prediction algorithms based on deep learning methods can overcome the shortcomings of traditional biological experimental methods and achieve great effects.As for the massive protein data,they extract protein features and train the models to predict PPI.However,the existing methods based on deep learning mainly have these defects:feature extraction only considers the sequence features of proteins and ignores others;The training process of sequence models is time-consuming;There is no effective method to predict PPI based on protein tertiary structure.Based on the above defects,in this paper,we propose three algorithms to improve the current research on PPI.The algorithms based on fusion features and the ResNet predict PPI based on protein sequences,they can improve the accuracy and speed of PPI prediction respectively.The algorithm based on 3D grids is the first in the research field to use deep learning methods to predict PPI based on protein tertiary structure and it can link protein structures and functions.The main contents of this paper are as follows:The PPI prediction algorithm based on fusion features considers the sequence features,local features,physical features and structural features of proteins,and combine them together to achieve PPI prediction.This algorithm can improve the accuracy of PPI prediction based on protein sequences.The PPI prediction algorithm based on the ResNet combines the advantages of powerful feature extraction capabilities of the ResNet with deep layers and convolutional neural network that can take full advantage of the GPU performance for efficient computing to achieve PPI prediction.This algorithm can ensure high accuracy and shorten the training time greatly.The PPI prediction algorithm based on 3D grids adopts a processing method to transform the representation of the protein tertiary structure into a 3D grid,because the representation of the atoms in the protein tertiary structure is similar to that of the pixels in the 3D image.Then we use the 3D ResNet to extract the deep features of the 3D grid and achieve PPI prediction.