Method and system for identifying protein domain based on protein three-dimensional structure image

A system for identifying a protein domain based on the protein three-dimensional structure image, comprising:

• a data acquisition and preprocessing module configured to obtain protein information, protein domain annotation information, a protein three-dimensional experimental structure, a protein three-dimensional computational structure and protein secondary structure annotation information, so as to generate a training set and a target set,
• wherein generating the training set comprises:
• extracting three-dimensional coordinates of carbon, nitrogen and oxygen atoms in a main chain from the protein three-dimensional experimental structure to construct three-dimensional atomic point cloud, giving, after standardized preprocessing, each atomic point a semantic label of a protein domain to which each atomic point belongs, and forming a first protein three-dimensional structure image as the training set;
• and
• wherein generating the target set comprises:
• extracting three-dimensional coordinates of carbon, nitrogen and oxygen atoms in a main chain from the protein three-dimensional computational structure to construct three-dimensional atomic point cloud, and forming, after standardized preprocessing, a second protein three-dimensional structure image as the target set;
• a model constructing module, configured to construct a point cloud segmentation model, and comprising:
• constructing, by the atomic points in the first and second protein three-dimensional structure images, a local directed graph according to a K-nearest neighbor classification algorithm;
• and
• constructing the point cloud segmentation model based on a dynamic graph convolutional neural network according to the local directed graph;
• wherein the point cloud segmentation model comprises a local feature extraction layer, a global feature extraction layer and a segmentation layer;
• the first and second protein three-dimensional structure images are inputted into the point cloud segmentation model after feature extraction, local features are extracted through the local feature extraction layer, global features are extracted through the global feature extraction layer, and the local features and the global features are integrated through the segmentation layer to output a score of a protein domain category label of each atomic point;
• wherein the local feature extraction layer comprises a plurality of edge-connected convolutional layers which are sequentially connected, a local directed graph feature of each atomic point is input into each of the plurality of edge-connected convolutional layer, and each edge-connected convolutional layer outputs the local feature of each atomic point after an edge-connected convolutional operation;
• an output of the previous edge-connected convolutional layer is taken as an input of the next edge-connected convolutional layer;
• wherein the global feature extraction layer comprises a multi-layer perceptron and a pooling layer, the local features of the atomic points outputted by all edge-connected convolutional layers in the local feature extraction layer are integrated as an input of the multi-layer perceptron, a local feature set of the atomic point cloud is outputted, and then the global features are outputted after a global pooling operation of the pooling layer;
• and
• wherein the segmentation layer comprises a plurality of multi-layer perceptron and a softmax regression layer, the local features of the atomic points outputted by all edge-connected convolutional layers in the local feature extraction layer and the global features outputted by the global feature extraction layer are integrated as an input, and a score of the protein domain category label of each atomic point is generated after operations of the plurality of multi-layer perceptron and the softmax regression layer;
• a model training module configured to extract features of the first protein three-dimensional structure image in the training set and train the point cloud segmentation model;
• and
• a model prediction module configured to identify a protein domain of the target set after feature extraction by using the trained point cloud segmentation model.