A method for finding K most similar neighbors of proteins based on graph neural network

Through a graph neural network-based method, combining features such as amino acid sequence, physical and chemical properties and geometric properties, a protein interaction network was constructed, and the graph convolution and graph attention model was used to solve the problem of the feature information used in the search of K most similar neighbors of proteins. The full problem improves the accuracy of prediction.

CN117594127BActive Publication Date: 2024-11-15BEIJING INST OF TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202311558758.2
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2023-11-21
Publication Date
2024-11-15
Estimated Expiration
2043-11-21

AI Technical Summary

Technical Problem

The prior art cannot fully utilize the interaction relationship and network topological characteristics between proteins in the search for K most similar neighbors of proteins, resulting in insufficient utilization of feature information and insufficient prediction accuracy.

Method used

Using a graph neural network-based method, combining the amino acid sequence, physical and chemical properties and geometric properties of proteins, a protein interaction relationship network is constructed, and a graph convolution and graph attention model is used to perform node feature learning to calculate the K best proteins. Similar neighbors.

Benefits of technology

The characteristic information of proteins and the topological characteristics of interaction networks are fully utilized to improve the prediction accuracy of the most similar neighbors of proteins.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117594127B_ABST
    Figure CN117594127B_ABST
Patent Text Reader

Abstract

The present invention proposes a method for searching K most similar neighbors of a protein based on a graph neural network. The purpose of this method is to find the K neighbors that are most similar to the query point feature in the feature space of the sample, so as to mine the potential similarity relationship of the protein. The proposed method is a graph neural network model with node feature embedding, which can make full use of the network topological characteristics of the interaction relationship between proteins, the amino acid sequence, physicochemical properties and geometric properties of the protein, and generate the feature vector of the protein node. The K most similar neighbors of the protein are solved according to the node feature vector. The graph neural network consists of a graph convolutional neural network, a graph attention network, an objective function and a nearest neighbor solving module. The K most similar neighbors are the K nodes with the shortest Euclidean distance to the target node or the most similar to the target node. Experimental results show that the method proposed by the present invention can obtain a higher accuracy rate in mining the K most similar neighbors of proteins.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of mining K most similar neighbors of a protein, and in particular to a method for searching K most similar neighbors of a protein based on a graph neural network. Background Art

[0002] Finding the K most similar neighbors of a protein is one of the important tasks in the fields of bioinformatics and structural biology. It aims to find the K proteins that are most similar to the structure of a given target protein from a large-scale protein structure database. With the rise of graph neural networks, it can be used to learn and represent the graph structure of protein structure in order to search for protein structure similarity. Existing deep learning methods are studied from four aspects: methods based on sequence alignment, methods based on structure alignment, methods based on feature vectors, and methods based on traditional machine learning. Existing methods have certain limitations. They cannot fully utilize the interaction relationship between proteins and the topological characteristics of protein networks. A multi-feature fusion graph neural network model is proposed. It combines the characteristics of protein amino acid sequence, physicochemical properties, and geometric properties, and considers the interaction between proteins and the network topological characteristics of proteins. The graph neural network model is used to mine the potential similarity relationship of proteins. Summary of the invention

[0003] The purpose of the present invention is to overcome the above-mentioned technical shortcomings and deficiencies and provide a method for searching the K most similar neighbors of a protein based on a graph neural network.

[0004] The technical solution adopted by the present invention is as follows:

[0005] The protein K most similar neighbors search method based on graph neural network includes the following steps:

[0006] Step 1: Obtain the characteristics of the protein, such as amino acid sequence, physicochemical properties, and geometric properties;

[0007] Step 2: Feature fusion and construction of protein interaction network (PPI);

[0008] Step 3: Build a graph neural network model;

[0009] Step 4: Calculate the K most similar neighbors of the protein and input the vector representation of the protein node output by the model into the K nearest neighbor calculation module;

[0010] Step 1: Obtain the characteristics of the protein such as amino acid sequence, physicochemical properties and geometric properties. The specific steps are as follows:

[0011] Step 1.1, obtain the encoded representation of the amino acid sequence of the protein. The traditional approach is to use One-hot encoding. The One-hot encoding represents each amino acid as a vector, in which only one position is 1 and the other positions are 0. When there are many types of amino acids, the dimension of the representation will be very high, which will increase the complexity of calculation and storage. The present invention combines the Word2Vec word embedding model in the NLP task. After the amino acid sequence is numerically represented, it is input into the Word2Vec model and trained using the Skip-gram and CBOW models. When Word2Vec is applied to protein sequences, each amino acid can be regarded as a unit similar to a word, and the Word2Vec model can be trained on the protein sequence.

[0012]

[0013] Where w is the characteristic representation of amino acids, which is a K-dimensional vector, and v w and are two ways of expressing amino acid w, and p is the goal to be optimized so that when input w I When outputting w j The probability of the amino acid sequence being the highest is obtained by sampling the amino acid sequence and inputting it into the Word2Vec model for training. The word vector representation for each amino acid can be obtained, and the similarity and semantic relationship between amino acids can be captured. Finally, the protein feature F1 is obtained;

[0014] Step 1.2, obtaining the coding representation of the physicochemical properties of the protein, the physicochemical properties of the protein mainly include isoelectric point (PI), solubility, hydrophobicity index, ultraviolet absorption, etc. The present invention uses numerical coding to encode the above quantitative properties into a K-dimensional feature vector. Obtain the physicochemical feature F2;

[0015] Step 1.3: Obtain the coding representation of the geometric properties of the protein. For the geometric properties of the protein, the present invention adopts a feature representation method based on coordinate coding. In this method, the atomic coordinates of each amino acid residue are represented as a vector or matrix in three-dimensional space to obtain the geometric property F3.

[0016] Step 2: Feature fusion and construction of protein interaction network (PPI). The specific steps are as follows:

[0017] Step 2.1: Feature fusion: The three types of feature vectors in step 1 are horizontally spliced ​​to construct a high-dimensional feature F = {F1, F2, F3};

[0018] Step 2.2: Construct a protein interaction network (PPI). Obtain protein-related interaction data from String or other protein interaction databases. The data structure is represented by e=(u,v), where e represents the interaction between two proteins, and u and v represent two different protein structures. The final protein interaction network G=(V,E) is constructed using the above data, where G represents the network itself, V represents the node, and E represents the edge, ultimately forming a protein interaction network.

[0019] Step 3: Build a graph neural network model. The specific steps are as follows:

[0020] Step 3.1, data preprocessing. The present invention is trained based on a self-supervised learning method and does not require manual annotation for training sets or test sets. Training can be completed by sampling the edges of the protein interaction network. However, since the number of positive samples (i.e., the edges existing in the protein network) is relatively small and there is a class imbalance problem (i.e., the number of negative samples is far greater than the positive samples or the number of non-existent edges is far greater than the number of existing edges), using only the original data for training can easily lead to a biased model prediction that all protein pairs are negative samples. Therefore, the negative sampling and positive-negative sample balance methods are combined to construct a data set;

[0021] Step 3.2, MLP layer. After the data set is constructed, the training set is input into the first MLP layer of the model for nonlinear transformation.

[0022] Y=Wx+b

[0023] Y=Relu(Y)

[0024] Where x is the input feature vector, w is the weight matrix, b is the bias term, Y is the output feature vector, and ReLU is a common activation function. The MLP layer learns appropriate weights and biases to perform linear transformation and nonlinear activation on the input features, thereby extracting higher-order and more abstract feature representations. By stacking multiple MLP layers, the network can gradually extract more complex features;

[0025] Step 3.3, graph convolution layer, inputs the output of the MLP layer to the transmission and aggregation of node features of the graph convolution layer. This enables each node to obtain information about surrounding nodes. Feature propagation enables the graph convolution layer to have a certain degree of local perception and context perception. The calculation formula of the graph convolution layer can be expressed as follows:

[0026]

[0027] where Z∈R N×F represents the updated node features, and F is the output feature dimension. It is the result of normalizing the adjacency matrix. D×F is a learnable weight matrix. f(·) is an activation function, which is usually ReLU or other nonlinear activation functions;

[0028] Step 3.4: Graph Attention Layer: Input the output of the graph convolution layer to the graph attention layer to enhance the aggregation and propagation of node features by adaptively learning the importance weights between nodes. The graph attention layer converts the relationship between each node and its neighboring nodes into an attention coefficient, and then uses the attention coefficient as a weight to perform weighted aggregation on the features of neighboring nodes. The calculation formula of the graph attention layer can be expressed as follows:

[0029]

[0030] Among them, H i represents the feature vector of node i, and N(i) represents the set of neighboring nodes of node i. ij represents the attention coefficient of node i and neighbor node j, which is used to measure the importance of node i to neighbor node j. W is a learnable weight matrix, and f(·) is an activation function, which is usually applicable to ReLU or other nonlinear activation functions.

[0031] Step 3.5: Inference prediction layer: Input the output of the graph attention layer to the inference prediction layer. Calculate the deviation between the predicted value and the true value as the loss. Then, optimize the model parameters through the back propagation algorithm.

[0032]

[0033] Where N1 represents the number of positive examples, N2 represents the number of negative examples, and p uv represents the joint probability distribution of the existence of an edge for the node pair (u, v), Representative node pair There is no joint probability distribution of edges;

[0034] Step 3.6: Generate protein embedding features. After the model training is completed, the feature representation of each protein node will be output as an M-dimensional feature vector. This feature vector is the final vector representation of the protein node.

[0035] Step 4: Calculate the K most similar neighbors of the protein. Input the vector representation of the protein node output by the model into the K nearest neighbor calculation module. The calculation steps are as follows:

[0036] Step 4.1: Select a distance measurement method. First, you need to select an appropriate distance measurement method to measure the similarity between nodes. Common distance measurement methods include Euclidean distance, Manhattan distance, cosine similarity, etc.

[0037] Step 4.2: Calculate the distance. According to the selected distance metric, calculate the distance or similarity between the node and all other nodes. This can be achieved by calculating the distance between node features.

[0038] Step 4.3: Select K nearest neighbors. For each node, select K nearest neighbors from the calculated distance or similarity. You can select the nearest K nodes in order of distance from small to large, or select the most similar K nodes in order of similarity from large to small, and output the results.

[0039] Advantages and beneficial effects of the present invention

[0040] The present invention proposes a method for calculating the K most similar neighbors of a protein based on a graph neural network. The method is a graph neural network model with multiple features fusion. It combines the characteristics of the protein's amino acid sequence, physicochemical properties, and geometric properties, and considers the interactions between proteins and the network topology characteristics of proteins. The graph convolution and graph attention models are used to learn the protein node features. The feature information of proteins and the network topology characteristics of the interaction between proteins are fully utilized. It provides a new method for the field of protein information mining, solves the problem of insufficient utilization of feature information in the process of mining the most similar neighbors of proteins, and improves the accuracy of prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a workflow diagram of the present invention for solving the problem of finding K similar neighbors of a protein;

[0042] Figure 2 It is a network structure diagram of a protein network;

[0043] Figure 3 It is a structural diagram of the graph neural network and KNN model. DETAILED DESCRIPTION

[0044] The present invention is described in detail below with reference to the accompanying drawings and examples.

[0045] A method for finding the K most similar neighbors of a protein based on graph neural network. The overall implementation process of finding the K most similar neighbors of a protein is as follows: Figure 1 As shown; solve the K most similar neighbors for the protein network of four nodes, the network structure is as follows Figure 2 As shown, the structure of the graph neural network and KNN model Figure 3 The specific implementation plan includes the following contents and steps:

[0046] Step 1: Obtain the characteristics of the protein such as amino acid sequence, physicochemical properties and geometric properties. The specific steps are as follows:

[0047] Step 1.1: Obtain the encoded representation of the amino acid sequence of the protein. Obtain the amino acid sequence of protein A and input its amino acid sequence into the Word2Vec model to obtain its protein feature F1 = {x1, x2, x3, …, x k};

[0048] Step 1.2: Obtain the coded representation of the physicochemical properties of the protein, obtain the numerical representation of the physicochemical properties of protein A, and construct the feature vector F2 = {y1, y2, y3, ..., y k};

[0049] Step 1.3: Obtain the coded representation of the geometric properties of the protein, obtain the numerical representation of the geometric properties of protein A, and represent it as a vector F3 = {z1, z2, z3} in a three-dimensional space.

[0050] Step 2: Feature fusion and construction of protein interaction network (PPI). The specific steps are as follows:

[0051] Step 2.1: Feature fusion: Construct a high-dimensional feature F = {F1, F2, F3} by horizontally splicing the three types of feature vectors in step 1. Repeat steps 1.1 to 1.3 to construct feature vector representations of proteins B, C, and D;

[0052] Step 2.2: Construct a protein interaction network (PPI). Obtain protein interaction data from the String database and construct a protein interaction network. The final protein network is as follows: Figure 2 As shown, A, B, C and D represent nodes, and edges represent the interaction relationships between proteins, ultimately forming a protein interaction relationship network.

[0053] Step 3: Build a graph neural network model. The specific steps are as follows:

[0054] Step 3.1: Data preprocessing: negative sampling of the protein network to obtain non-existent edges (A, C) and (B, D), and sampling of positive edges (A, B) and (B, C);

[0055] Step 3.2, MLP layer, input the network structure data and protein feature data into the MLP layer;

[0056] Step 3.3: Graph convolution layer: Input the output of the MLP layer into the graph convolution layer;

[0057] Step 3.4: Graph Attention Layer: Input the output of the graph convolutional layer into the graph attention layer.

[0058] Step 3.5: Inference prediction layer: Input the output of the graph attention layer to the inference prediction layer to perform loss calculation and back propagation;

[0059] Step 3.6: Generate protein embedding features. After the model training is completed, feature representations of protein nodes A, B, C, and D will be generated, which are M-dimensional feature vectors. The final features are shown in Table 1.

[0060] Table 1 Feature representation of each node in the protein network

[0061] node Feature Representation A <![CDATA[N={a1,a2,a3,…,a M }]]> B <![CDATA[N={b1,b2,b3,…,b M }]]> C <![CDATA[N={c1,c2,c3,…,c M }]]> D <![CDATA[N={d1,d2,d3,…,d M }]]>

[0062] Step 4: Calculate the K most similar neighbors of the protein. The specific steps are as follows:

[0063] Step 4.1, select the distance measurement method, select cosine similarity as the distance measurement method, and its formula is as follows;

[0064]

[0065] Step 4.2, calculate the distance, and calculate the similarity between protein A and all other proteins according to the cosine similarity calculation formula. When the cosine similarity is equal to 1, it means that the directions of the two vectors are exactly the same; when the cosine similarity is equal to -1, it means that the directions of the two vectors are completely opposite; when the cosine similarity is equal to 0, it means that the directions of the two vectors are orthogonal. Therefore, the larger the cosine similarity, the higher the similarity of the two vectors; conversely, the smaller the cosine similarity, the lower the similarity of the two vectors. The calculation results are shown in Table 2;

[0066] Table 2 Similarity between A and other proteins

[0067] node Similarity value B 0.3 C 0.1 D 0.5

[0068] Step 4.3, select K nearest neighbors, select the two neighbors most similar to protein A, and according to the similarity values ​​in Table 2, the two neighbors most similar to protein A are finally obtained as protein B and protein D. The above description is only a preferred example of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various changes and variations. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention shall be included in the scope of the claims of the present invention.

Claims

1. A method for finding the K most similar neighbors of a protein based on graph neural network, characterized in that: The following steps are involved: Step 1, obtaining the amino acid sequence, physicochemical properties and geometric properties of the protein, specifically including obtaining the coding representation feature F1 of the amino acid sequence of the protein, obtaining the coding representation feature F2 of the physicochemical properties of the protein, and obtaining the coding representation feature F3 of the geometric properties of the protein; Step 2: feature fusion and construction of protein interaction network (PPI), specifically including feature fusion to obtain high-dimensional feature F and construction of protein interaction network (PPI); Step 3: Build a graph neural network model, which includes combining negative sampling and positive and negative sample balance methods to build a data set, extracting more complex features through the MLP layer, extracting structural information through the graph convolution layer, strengthening the aggregation and propagation of node features through the graph attention layer, performing inference prediction, and generating the final embedded features of the protein; Step 4: Calculate the K most similar neighbors of the protein, which specifically includes selecting a distance measurement method, calculating the distance, and selecting K nearest neighbors.

2. The method for searching K most similar neighbors of a protein based on a graph neural network according to claim 1, characterized in that: The step 1 of obtaining the amino acid sequence, physicochemical properties and geometric properties of the protein specifically includes obtaining the coding representation feature F1 of the amino acid sequence of the protein, obtaining the coding representation feature F2 of the physicochemical properties of the protein, and obtaining the coding representation feature F3 of the geometric properties of the protein. The specific steps are as follows: Step 1.1, obtain the encoding representation feature F1 of the amino acid sequence of the protein: the traditional approach is to use One-hot encoding, where One-hot encoding represents each amino acid as a vector, in which only one position is 1 and the other positions are 0. When there are many types of amino acids, the dimension of the representation will be very high, which will increase the complexity of calculation and storage. The present invention combines the Word2Vec word embedding model in the NLP task, and the amino acid sequence is numerically represented and then input into the Word2Vec model for training using the Skip-gram and CBOW models. When Word2Vec is applied to protein sequences, each amino acid can be regarded as a unit similar to a word, and the Word2Vec model is trained on the protein sequence. The specific formula can be expressed as: Where w is the characteristic representation of amino acids, which is a K-dimensional vector, and v w and are two ways of expressing amino acid w, and p is the goal to be optimized so that when input w I When outputting w j The probability of having the highest value is 1. By sampling the amino acid sequence and inputting it into the Word2Vec model for training, we can obtain the word vector representation for each amino acid and capture the similarity and semantic relationship between amino acids, and finally obtain the protein feature F1. Step 1.2, obtaining the coding representation feature F2 of the physicochemical properties of the protein: the physicochemical properties of the protein mainly include isoelectric point (PI), solubility, hydrophobicity index, and ultraviolet absorption properties. The present invention uses numerical coding to encode the above quantitative properties into a K-dimensional feature vector to obtain the physicochemical feature F2; Step 1.3, obtaining the coding representation feature F3 of the geometric properties of the protein: For the geometric properties of the protein, the present invention adopts a feature representation method based on coordinate coding. In this method, the atomic coordinates of each amino acid residue are represented as a vector or matrix in three-dimensional space to obtain the geometric property feature F3.

3. The method for searching protein K most similar neighbors based on graph neural network according to claim 1, characterized in that: The feature fusion and construction of the protein interaction relationship network (PPI) described in step 2 specifically include feature fusion to obtain a high-dimensional feature F and construction of a protein interaction relationship network (PPI), and the specific steps are as follows: Step 2.1: Feature fusion to obtain high-dimensional feature F: The three types of feature vectors in step 1 are horizontally spliced ​​to construct a high-dimensional feature F = {F1, F2, F3}; Step 2.2, construct a protein interaction relationship network (PPI): obtain protein-related interaction relationship data from String or other protein interaction databases. The data structure is expressed as e=(u,v), where e represents the interaction relationship between two proteins, and u and v represent two different protein structures respectively. The final protein interaction network G=(V,E) is constructed through the above data, where G represents the network itself, V represents the node, and E represents the edge, finally forming a protein interaction relationship network.

4. The method for searching protein K most similar neighbors based on graph neural network according to claim 1, characterized in that: The construction of the graph neural network model described in step 3 specifically includes combining negative sampling and positive and negative sample balance methods to construct a data set, extracting more complex features through the MLP layer, extracting structural information through the graph convolution layer, strengthening the aggregation and propagation of node features through the graph attention layer, performing inference prediction, and generating the final embedded features of the protein. The specific steps are as follows: Step 3.1, combining negative sampling and positive-negative sample balance methods to construct a data set: The present invention is trained based on a self-supervised learning method, which does not require manual annotation for training sets or test sets. It is necessary to sample the edges of the protein interaction network to complete the training. However, due to the relatively small number of positive samples and the existence of class imbalance problems, using only the original data for training can easily lead to the model's biased prediction that all protein pairs are negative samples. Therefore, the negative sampling and positive-negative sample balance methods are combined to construct a data set; Step 3.2: Extract more complex features through the MLP layer: After the dataset is constructed, the training set is input into the first MLP layer of the model for nonlinear transformation. Y=Wx+b Y=Relu(Y) Where x is the input feature vector, w is the weight matrix, b is the bias term, Y is the output feature vector, ReLU is a common activation function, and the MLP layer learns appropriate weights and biases to perform linear transformation and nonlinear activation on the input features, thereby extracting higher-order and more abstract feature representations. By stacking multiple MLP layers, the network can gradually extract more complex features. Step 3.3: Extract structural information through the graph convolution layer: The output of the MLP layer is input into the transmission and aggregation of node features of the graph convolution layer, so that each node can obtain information about surrounding nodes. Feature propagation enables the graph convolution layer to have a certain degree of local perception and context perception. The calculation formula of the graph convolution layer can be expressed as follows: where Z∈R N×F represents the updated node features, F is the output feature dimension, is the result of normalizing the adjacency matrix, W∈R D×F is a learnable weight matrix, f(·) is an activation function, usually ReLU or other nonlinear activation functions; Step 3.4: Strengthen the aggregation and propagation of node features through the graph attention layer: Input the output of the graph convolution layer to the graph attention layer, and strengthen the aggregation and propagation of node features by adaptively learning the importance weights between nodes. The graph attention layer converts the relationship between each node and its neighboring nodes into an attention coefficient, and then uses the attention coefficient as a weight to perform weighted aggregation on the features of neighboring nodes. The calculation formula of the graph attention layer can be expressed as follows Among them, H i represents the feature vector of node i, N(i) represents the set of neighbor nodes of node i, α ij represents the attention coefficient of node i and neighbor node j, which is used to measure the importance of node i to neighbor node j. W is a learnable weight matrix, and f(·) is an activation function, which is usually applicable to ReLU or other nonlinear activation functions. Step 3.5: Perform inference prediction: Input the output of the graph attention layer to the inference prediction layer, calculate the deviation between the predicted value and the true value as the loss, and then optimize the model parameters through the back propagation algorithm. Where N1 represents the number of positive examples, N2 represents the number of negative examples, and p uv represents the joint probability distribution of the existence of an edge for the node pair (u, v), Representative node pair There is no joint probability distribution of edges; Step 3.6: Generate the final embedded features of the protein: After the model training is completed, the feature representation of each protein node will be output as an M-dimensional feature vector, which is the final vector representation of the protein node.

5. The method for searching protein K most similar neighbors based on graph neural network according to claim 1, characterized in that: The calculation of the K most similar neighbors of the protein described in step 4 specifically includes selecting a distance measurement method, calculating the distance, and selecting K nearest neighbors. The specific steps are as follows: Step 4.1: Select a distance measurement method: First, you need to select an appropriate distance measurement method to measure the similarity between nodes. Common distance measurement methods include Euclidean distance, Manhattan distance, and cosine similarity. Step 4.2, calculate the distance: According to the selected distance measurement method, calculate the distance or similarity between the node and all other nodes, which can be achieved by calculating the distance between node features; Step 4.3: Select K nearest neighbors: For each node, select K nearest neighbors from the calculated distance or similarity. You can select the nearest K nodes in order of distance from small to large, or select the most similar K nodes in order of similarity from large to small, and output the results.