A compound-protein interaction prediction method integrating network topology information
By incorporating network topology information into a compound-protein interaction prediction method, node centrality and relevance metrics are calculated, and the cross-attention mechanism of the transformer model is utilized to improve the accuracy of compound-protein interaction prediction.
Patent Information
- Application Number
- CN202210491027.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-07
- Publication Date
- 2024-11-08
- Estimated Expiration
- 2042-05-07
AI Technical Summary
Existing deep learning models have failed to effectively integrate network topology information, resulting in insufficient accuracy in predicting compound-protein interactions.
By constructing a compound-protein interaction prediction method that integrates network topology information, calculating the centrality and relevance measures of nodes, and utilizing the cross-attention mechanism of the transformer model to process protein and compound features, multimodal information is integrated.
It improves the accuracy of compound-protein interaction prediction, demonstrating superior precision, recall, accuracy, and AUC compared to other methods.
Smart Images

Figure CN114678081B_ABST
Abstract
Description
Technical Field
[0001] The invention belongs to the field of bioinformatics, and in particular relates to a compound-protein interaction prediction method integrating network topology information. Background Art
[0002] Proteins are the basis for organisms to realize life activities and play a wide range of important roles in organisms. Drugs are usually compounds with special properties. They affect the function of proteins by binding to specific proteins in organisms, thereby producing drug efficacy. The study of compound-protein interactions is an important part of drug design and is of great significance to drug development. In order to improve the efficiency of drug research and development, many prediction models based on deep learning have been developed, but existing models fail to explicitly integrate network topology information into the model. Summary of the invention
[0003] The purpose of the present invention is to provide a compound-protein interaction prediction method integrating network topology information, which can effectively improve the accuracy of compound-protein interaction prediction.
[0004] To achieve the above object, the present invention adopts the following technical solution:
[0005] A compound-protein interaction prediction method integrating network topology information comprises the following steps:
[0006] Step 1: Preprocess the data;
[0007] Step 2: Construct a compound-protein interaction network based on the data set and calculate the degree of each node in the interaction network as the centrality measure of the node;
[0008] Step 3: For each pair of compounds and proteins in the data set, calculate the number of common neighbors between each adjacent node of the compound and the protein in the interaction network as a measure of the correlation between the compound and the protein; similarly, obtain the correlation measure between the protein and the compound;
[0009] Step 4: Build a transformer-based binary classification model, assign a real-valued embedding vector to each node based on the obtained node centrality measure, and add it to the node features.
[0010] Step 5: Assign a learnable scalar to each possible value of the correlation between each pair of nodes, and use it as the bias term of the cross-attention module in the model described in step 4.
[0011] Step 6: Finally, use the fully connected layer to output the predicted probability.
[0012] Furthermore, the step 1 is specifically as follows:
[0013] Step 1.1: Preprocess the compound-protein interaction data, protein sequence information, and compound SMILES data by removing outliers, randomly generating negative examples, and randomly dividing the data set;
[0014] Step 1.2: Encode protein sequence using seqvec model;
[0015] Step 1.3: Use rdkit to extract compound features and adjacency matrix of compound graph.
[0016] Furthermore, the step 2 is specifically as follows:
[0017] Step 2.1: Take each compound and each protein in the original data set as a node, and the positive interactions between pairs of compounds and proteins as edges to construct a compound-protein interaction network.
[0018] Step 2.2: Calculate the number of neighbor nodes of each node in the network as the degree centrality of the node.
[0019] Furthermore, the step 3 is specifically as follows:
[0020] Step 3.1: Calculate and save the number of common neighbors between all proteins and between all compounds in the interaction network.
[0021] Step 3.2: For each pair of compounds and proteins in the data set, find the number of common neighbors between each adjacent node of the compound and the protein in the interaction network based on the results saved in step 3.1, and record the maximum value as the correlation measure of the compound to the protein.
[0022] Step 3.3: For each pair of compound and protein in the data set, find the number of common neighbors between each adjacent node of the protein and the compound in the interaction network according to the results saved in step 3.1, and record the maximum value as the correlation measure of the protein to the compound.
[0023] Furthermore, the step 4 is specifically as follows:
[0024] Step 4.1: Build a traditional transformer model, remove the decoder's position encoding and transform the mask from a lower triangular matrix to an adjacency matrix so that the decoder can only see adjacent nodes.
[0025] Step 4.2: Assign a real-valued embedding vector to each node based on its centrality measure and add it to the node features as follows:
[0026] F=X+Z deg (1)
[0027] Where F represents the final new feature vector. X represents the initial feature vector of amino acids or atoms. Z is a learnable embedding vector specified by the degree of the protein or compound node.
[0028] Furthermore, the step 5 is specifically as follows:
[0029]
[0030] in, It is the traditional method of calculating attention weights. The function φ is defined by the correlation between nodes.
[0031] B φ is a learnable scalar indexed by the output value of the function φ and is shared across all layers.
[0032] Compared with the prior art, the present invention has the following beneficial effects:
[0033] 1. The present invention takes into account the network topology information and integrates the node centrality coding and correlation coding in the model, so that the model can contain more effective information;
[0034] 2. A transformer-based binary classification model was constructed, and the cross-attention mechanism was used to process the relationship between protein features and compound features to integrate multimodal information and improve the accuracy of interaction prediction. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1 It is a schematic diagram of the node correlation calculation method;
[0036] Figure 2 It is a schematic diagram of the system principle;
[0037] Figure 3a is a line graph comparing the performance of the embodiment of the present invention with other methods on a human data set;
[0038] Figure 3b It is a line graph comparing the performance of the embodiments of the present invention with other methods on the C. elegans dataset. DETAILED DESCRIPTION
[0039] The present invention will be further described below with reference to the accompanying drawings and examples.
[0040] like Figure 1As shown in the figure, the present invention proposes a compound-protein interaction prediction method that integrates network topology information. First, based on the network topology information of the compound and protein, the centrality of each node and the correlation of paired data are calculated, and then a model is established to predict the potential interaction relationship between the compound and the protein. Finally, the performance of the model is evaluated using corresponding indicators. In order to make the correlation calculation method more intuitive, the present invention establishes a corresponding schematic diagram, as shown in the figure. Figure 1 As shown. A compound-protein interaction network is constructed through the data set, and then a projection operation is performed on it. By calculating the number of common neighbors of each adjacent node of node 1 in the interaction network and node 2 in the paired data, the correlation measure of node 1 to node 2 is obtained. The upper part of the schematic diagram represents the original interaction network. It can be seen that it is essentially a bipartite graph. The lower part represents the projection of the original network to the compound set and the projection to the protein set. If two protein nodes are connected to the same compound, the two proteins are connected, and the weight of the edge is the number of common compounds. The same treatment is performed on the compound nodes. The correlation is calculated based on the obtained projection, as shown in Figure 1 As shown, if we want to find the correlation between x1 and y1, we first need to find the node with the most common neighbors with y1 among the other neighbors of x1 except y1, and take the number of common neighbors. Here, x1 has only two neighbors, and only y2 meets the conditions. There are two common neighbors between y2 and y1, namely x1 and x2, so the correlation between x1 and y1 is 2. Then find the node with the most common neighbors with x1 among the other neighbors of y1, which is Figure 1 , the number of common neighbors of x2 and x1 is 2, which is taken as the correlation of y1 to x1. Finally, a learnable scalar is assigned to each possible value obtained, which will be used as a bias term in the cross-attention module, that is, correlation coding, adding more effective information to the model. The purpose of the present invention is to predict potential interactions using a compound-protein interaction prediction model that combines centrality coding and correlation coding.
[0041] like Figure 2 As shown, the specific embodiment process of the present invention is:
[0042] Step 1: First, obtain compound-protein interaction pairs from the human and Caenorhabditis elegans datasets. The human dataset contains 3369 positive interactions between 1052 compounds and 852 proteins; the Caenorhabditis elegans dataset contains 4000 positive interactions between 1434 compounds and 2504 proteins. After removing outliers, randomly generate negative samples with the same number of positive samples and randomly divide the training, validation, and test sets. Obtain amino acid embeddings through the seqvec model, which is pre-trained on elmo using a large protein database. The SMILES descriptors of the molecules are used to obtain atomic embeddings through rdkit, an open source toolkit for cheminformatics.
[0043] Step 2: Take each compound and each protein in the two data sets as a node, and the positive interactions between paired compounds and proteins as edges to construct a compound-protein interaction network. Calculate the number of neighbor nodes of each node in the network as the degree centrality of the node.
[0044] Step 3: For each pair of compounds and proteins in the data set, calculate the number of common neighbors between each adjacent node of the compound in the interaction network and the protein as a measure of the correlation between the compound and the protein, which represents the number of compounds that the proteins that can interact with the compound have in common with the target protein. Intuitively, two protein nodes that already have many common compounds have a stronger tendency to have more common compounds in the future; similarly, the correlation measure of the protein to the compound is obtained.
[0045] Step 4: Build a transformer-based binary classification model, remove the decoder's positional encoding and transform the mask from a lower triangular matrix to an adjacency matrix so that the decoder can only see adjacent nodes. Assign a real-valued embedding vector to each node based on the obtained node centrality measure, add it to the original feature matrix of the node, as shown in formula (1), and input it into the model as a new feature matrix.
[0046] F=X+Z deg (1)
[0047] Step 5: Assign a learnable scalar to each possible value of the correlation between each pair of nodes, and use it as the bias term of the cross-attention module in the model described in step 4, as shown in formula (2).
[0048]
[0049] in, It is the traditional method of calculating attention weights. The function φ is defined by the correlation between nodes.
[0050] B φ is a learnable scalar indexed by the output value of the function φ and is shared across all layers.
[0051] Step 6: Finally, use the fully connected layer to output the predicted probability.
[0052] Verification of the effectiveness of the present invention:
[0053] Through comparative experiments, the performance of the present invention is evaluated on five indicators respectively. The results of the comparison between the present invention and other methods are as follows: Figure 3a and Figure 3b As shown, the best indicators of the present invention on the test set reached precision: 0.997, recall: 1, accuracy: 0.999, F1score: 0.998, and AUC: 1. The verification results show that the performance of the present invention is better than other methods.
Claims
1. A compound-protein interaction prediction method integrating network topology information, characterized in that: The following steps are involved: Step 1: Preprocess the data; Step 2: Construct a compound-protein interaction network based on the data set and calculate the degree of each node in the interaction network as the centrality measure of the node; Step 3: For each pair of compounds and proteins in the data set, the number of common neighbors between each adjacent node of the compound in the interaction network and the protein is calculated as a measure of the relevance of the compound to the protein; Similarly, the correlation measure of protein to compound is obtained; Step 4: Build a transformer-based binary classification model, assign a real-valued embedding vector to each node based on the obtained node centrality measure, and add it to the node features; Step 4.1: Build a traditional transformer model, remove the decoder's position encoding and transform the mask from a lower triangular matrix to an adjacency matrix so that the decoder can only see adjacent nodes; Step 4.2: Assign a real-valued embedding vector to each node based on its centrality measure and add it to the node features as follows: F=X+Z deg (1) Among them, F represents the final new feature vector; X represents the initial feature vector of the amino acid or atom; Z is a learnable embedding vector specified by the degree of the protein or compound node; Step 5: Assign a learnable scalar to each possible value of the correlation between each pair of nodes, and use it as the bias term of the cross-attention module in the model described in step 4; Step 6: Finally, use the fully connected layer to output the predicted probability.
2. The compound-protein interaction prediction method integrating network topology information according to claim 1, characterized in that The step 1 is specifically as follows: Step 1.1: Preprocess the compound-protein interaction data, protein sequence information, and compound SMILES data by removing outliers, randomly generating negative examples, and randomly dividing the data set; Step 1.2: Encode protein sequence using seqvec model; Step 1.3: Use rdkit to extract compound features and adjacency matrix of compound graph.
3. The compound-protein interaction prediction method integrating network topology information according to claim 1, characterized in that The step 2 is specifically as follows: Step 2.1: Take each compound and each protein in the original data set as a node, and the positive interactions between pairs of compounds and proteins as edges to construct a compound-protein interaction network; Step 2.2: Calculate the number of neighbor nodes of each node in the network as the degree centrality of the node.
4. The compound-protein interaction prediction method integrating network topology information according to claim 1, characterized in that The step 3 is specifically as follows: Step 3.1: Calculate and save the number of common neighbors between all proteins and the number of common neighbors between all compounds in the interaction network; Step 3.2: For each pair of compounds and proteins in the data set, find the number of common neighbors between each adjacent node of the compound and the protein in the interaction network according to the results saved in step 3.1, and record the maximum value as the correlation measure of the compound to the protein; Step 3.3: For each pair of compound and protein in the data set, find the number of common neighbors between each adjacent node of the protein and the compound in the interaction network according to the results saved in step 3.1, and record the maximum value as the correlation measure of the protein to the compound.
5. The compound-protein interaction prediction method integrating network topology information according to claim 1, characterized in that The step 5 is specifically as follows: in, It is the traditional method of calculating attention weights. The function φ is defined by the correlation between nodes. B φ is a learnable scalar indexed by the output value of the function φ and is shared across all layers.
