Traditional Chinese medicine target relation prediction method based on multi-attention graph neural network learning

Abstract

The invention belongs to the field of bioinformatics, and particularly provides a traditional Chinese medicine target relation prediction method based on multi-attention graph neural network learning. The method comprises the following steps: firstly, constructing a heterogeneous network comprising a plurality of entities such as traditional Chinese medicines, targets, diseases and prescriptions and a plurality of edge types so as to better describe the relationship between the traditional Chinese medicines and the targets; then, a method based on a multi-attention graph neural network is provided to learn fusion feature representation of traditional Chinese medicine and targets based on a heterogeneous network; and finally, realizing the prediction of the traditional Chinese medicine-target relationship based on a deep neural network DNN. Based on the multi-source heterogeneous data and the heterogeneous relationship, the traditional Chinese medicine and the target are subjected to fusion feature representation by using the multi-attention graph neural network, more comprehensive information is considered, and the prediction accuracy is effectively improved while the traditional Chinese medicine target relationship prediction is realized.

Description

Technical Field

[0001] The present invention belongs to the field of bioinformatics, and specifically provides a method for predicting the relationship between traditional Chinese medicine targets based on multi-attention graph neural network learning. Background Art

[0002] Accurately identifying potential TCM-target interactions is a key step in TCM discovery. In the past few decades, many scientific researchers have conducted many related biological experiments to identify TCM-target interactions, but the identification of TCM-target interactions based on biological experiments is time-consuming and expensive; therefore, it is of great significance to develop effective TCM-target interaction identification and prediction methods.

[0003] Existing computational methods for identifying drug-target interactions are mainly divided into three categories: text mining-based methods, biometric-based methods, and network-based methods. Among them, text mining-based methods extract information from literature and use descriptions of drugs and targets as features to identify drug-target interactions. For example, the prior art proposes a semantic similarity-based model using random forest and support vector machine methods to identify drug-target interactions. Biometric-based methods extract biological features of drugs and targets and use machine learning models to identify drug-target interactions. For example, the prior art uses drug-drug and target-target similarities based on chemical and genomic data, and uses support vector machine methods to convert the drug-target interaction identification problem into a binary classification problem. Network-based methods can describe complex interactions between different types of biological entities (such as drugs and proteins). For example, the prior art integrates a variety of drug-related information and protein-related information, and calculates the similarity matrix of drugs and proteins through multiple networks. Then, the similarity matrix is ​​restarted by random walk and singular value decomposition in turn to obtain low-dimensional feature representations of drugs and proteins to identify drug-target interactions. It can be seen that the methods based on text mining are affected by the differences in semantic expression, and the conflicts between different documents limit their performance; moreover, the existing methods do not consider more comprehensive information, such as TCM-disease associations, TCM interactions, prescription-TCM relationships, and disease-protein associations, etc., and usually extract features based on each interaction type separately, and connect the features based on each interaction type together. These methods cannot model heterogeneous information in a single heterogeneous network; at the same time, the existing methods do not take into account the inconsistency of feature weights of the same nodes in different relationships; in addition, the existing methods are mainly divided into two independent steps of feature extraction and TCM-target interaction prediction. In the TCM-target interaction prediction task, the parameters involved in the feature extraction step cannot be optimized, which limits the prediction performance. Summary of the invention

[0004] The purpose of the present invention is to propose a method for predicting the relationship between traditional Chinese medicine and targets based on multi-attention graph neural networks. A heterogeneous network containing multiple biological entities (including traditional Chinese medicine, targets, diseases, and prescriptions) is creatively designed. A model based on multi-attention graph neural networks is used to learn the low-dimensional feature representation of drugs and targets, and the traditional Chinese medicine-target interaction is predicted based on the learned features.

[0005] To achieve the above object, the technical solution adopted by the present invention is:

[0006] A method for predicting the relationship between traditional Chinese medicine targets based on multi-attention graph neural network learning, characterized in that it comprises the following steps:

[0007] Step 1. Collect data and build a heterogeneous network;

[0008] Obtain prescription data, Chinese medicine data, target data and disease data from the ETCM database to construct a Chinese medicine-target association network A mt , Chinese medicine-disease association network A md , Chinese medicine-prescription association network A mf , Chinese medicine interaction network A mm , target interaction network A tt , Chinese medicine-Chinese medicine similarity network S mm , Target-Disease Association Network A td , target-target similarity network S tt There are eight networks in total, which are combined to form a heterogeneous network;

[0009] Step 2. Initialize the feature vectors of the TCM nodes in the heterogeneous network;

[0010] Step 3. Perform similarity calculations on the TCM nodes and target nodes of the heterogeneous network to obtain the adjacency matrix of the heterogeneous network;

[0011] Step 4. Build a graph convolutional neural network model based on the graph attention network GAT, and extract the node feature representation of each node in each graph convolution layer in the heterogeneous network according to the adjacency matrix;

[0012] Step 5. Use the self-attention mechanism to perform multi-layer graph convolution fusion on the node feature representation of the heterogeneous network to obtain the fusion feature representation of each traditional Chinese medicine node and target node in the heterogeneous network;

[0013] Step 6. Concatenate the fusion feature representations of the TCM node to be predicted and the target node to form a TCM-target feature vector and send it to the TCM-target relationship prediction model, which outputs the TCM-target relationship prediction result.

[0014] Furthermore, in step 2, the characteristic data of the TCM node include nature, flavor, meridians and toxicity, specifically "cold, hot, warm, cool, neutral; sour, bitter, sweet, pungent, salty; lung, pericardium, heart, large intestine, triple burner, small intestine, stomach, gallbladder, bladder, spleen, liver, kidney; toxicity";

[0015] For the modifier description features, 2 -1 , 2 0 and 2 1 For the attribution characteristics, binary is used for quantization, thereby obtaining the initialized feature vector of each traditional Chinese medicine node.

[0016] Furthermore, in step 3, the similarity calculation process of the traditional Chinese medicine nodes is:

[0017] For any two TCM nodes, the cosine similarity between them is calculated based on the feature vectors, which is specifically expressed as:

[0018]

[0019] in, and They represent the feature vectors of TCM node a and TCM node b respectively, and k represents the feature number in the feature vector;

[0020] Set the TCM node similarity threshold θ1. If sim(a,b)>θ1, it is determined that TCM node a and TCM node b are similar. Otherwise, they are not similar.

[0021] Furthermore, in step 3, the similarity calculation process of the target node is:

[0022] For any two targets, BLAST was used to calculate the target similarity;

[0023] A target similarity threshold θ2 is set. If the target similarity is greater than the threshold θ2, the two targets are judged to be similar. Otherwise, they are not similar.

[0024] Furthermore, in step 3, the adjacency matrix is ​​specifically expressed as:

[0025] Corresponding to the TCM-target association network A mt , Chinese medicine-disease association network A md , Chinese medicine-prescription association network A mf , Chinese medicine interaction network A mm , target interaction network A tt Target-disease association network A td , the original association relationship is represented as a [0,1] matrix, 0 means there is no association relationship, 1 means there is an association relationship; corresponding to the Chinese medicine-Chinese medicine similarity network S mmSimilarity network S with target-target tt , the similarity relationship is represented as a [0,1] matrix, where 0 indicates no similarity and 1 indicates similarity.

[0026] Furthermore, in step 4, the graph convolutional neural network model adopts a three-layer stacked structure, wherein the feature representation of each graph convolution layer is:

[0027]

[0028] Among them, i represents node information, l represents the number of layers in the graph convolutional neural network, r represents the edge relationship type information, and α ij represents the attention coefficient, represents the normalization constant, represents the symmetric square root normalization constant, φ(·) represents the ReLU activation function; represents the neighborhood set of node i with respect to edge relation r, represents the neighborhood set of node j with respect to edge relationship r; represents the training parameters of the lth graph convolutional layer with edge relationship r, represents the feature representation of node i in the lth graph convolutional layer, Represents the neighborhood feature representation of node j in the lth graph convolutional layer, and the original feature representation of each node is initialized as a one-hot vector; represents the feature representation of node i in the l+1th graph convolutional layer; a (l) represents the training parameters of the l-th graph convolutional layer, and σ(·) represents the sigmoid activation function.

[0029] Furthermore, in step 5, the query vector, key vector, and value vector are defined in the self-attention mechanism as follows:

[0030]

[0031] Among them, q l , k l 、v l Denote the query vector, key vector, and value vector of the attention mechanism respectively, W Q , W K Represents the training parameters of the self-attention mechanism;

[0032] For any two graph convolutional layers, the attention coefficient is expressed as:

[0033]

[0034] Among them, e mn represents the attention coefficient of the m-th graph convolution layer and the n-th graph convolution layer, m,n∈(1,2,3);

[0035] Then, for node i, the feature representation after multi-layer graph convolution fusion is:

[0036]

[0037] in, Represents the feature representation of node i in the lth graph convolutional layer of the graph convolutional neural network, l∈(1,2,3);

[0038] Thus, the fusion feature representation of each TCM node and each target node is obtained.

[0039] Furthermore, in step 6, the TCM-target relationship prediction model adopts a three-layer DNN network.

[0040] Based on the above technical solution, the beneficial effects of the present invention are:

[0041] The present invention provides a method for predicting the relationship between traditional Chinese medicine and targets based on a multi-attention graph neural network. Firstly, a heterogeneous network including multiple entities such as traditional Chinese medicine, targets, diseases, prescriptions and multiple edge types is constructed to better describe the relationship between traditional Chinese medicine and targets; then, a method based on a multi-attention graph neural network is proposed to learn the fusion feature representation of traditional Chinese medicine and targets based on the heterogeneous network; finally, the relationship between traditional Chinese medicine and targets is predicted based on a deep neural network DNN.

[0042] Compared with the prior art, the present invention takes into account more comprehensive information, such as TCM-disease associations, TCM interactions, target interactions, prescription-TCM relationships, TCM-target associations, and disease-target associations, by using multi-source heterogeneous data and heterogeneous relationships and calculating related similarity relationships; uses graph neural networks to learn node feature representations of heterogeneous graphs, models heterogeneous information in a single heterogeneous network and considers node weights in different relationships; finally, performs final node feature learning through a self-attention mechanism to predict TCM targets, and improves prediction accuracy by training deep neural networks; in addition, the final TCM-target interaction prediction task can also optimize the parameters involved in the feature extraction step and improve model performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a flow chart of the method for predicting the relationship between traditional Chinese medicine targets based on multi-attention graph neural network learning in the present invention.

[0044] Figure 2 This is a model structure diagram of the traditional Chinese medicine target relationship prediction method based on multi-attention graph neural network learning in the present invention. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical solutions and beneficial effects of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0046] This embodiment provides a method for predicting the relationship between traditional Chinese medicine targets based on multi-attention graph neural network learning, and its process is as follows: Figure 1 As shown, the model structure is as follows Figure 2 As shown, the specific steps include:

[0047] Step 1. Collect data and build a heterogeneous network;

[0048] Obtain prescription data, Chinese medicine data, target data and disease data from the ETCM database to construct a Chinese medicine-target association network A mt , Chinese medicine-disease association network A md , Chinese medicine-prescription association network A mf , Chinese medicine interaction network A mm , target interaction network A tt , Chinese medicine-Chinese medicine similarity network S mm , Target-Disease Association Network A td , target-target similarity network S tt There are eight networks in total, which are combined to form a heterogeneous network;

[0049] Step 2. Initialize the feature vectors of the TCM nodes in the heterogeneous network;

[0050] The characteristics of traditional Chinese medicine were counted and initialized. The traditional characteristics of traditional Chinese medicine include nature, flavor, meridians and toxicity, which are "cold, hot, warm, cool, neutral; sour, bitter, sweet, spicy, salty; lung, pericardium, heart, large intestine, triple burner, small intestine, stomach, gallbladder, bladder, spleen, liver, kidney; toxicity", a total of 23 characteristics. All characteristics were quantified, and for the characteristics described by modifiers, such as "mild cold", "cold" and "severe cold", 2 were used respectively. -1 , 2 0 and 2 1 Quantify the representation; for the attribution characteristics, binary quantization is used, that is, 1 represents the existence of the characteristic, and 0 represents the non-existence of the characteristic; thus, the initialization feature vector of each traditional Chinese medicine node is obtained;

[0051] Step 3. Perform similarity calculations on the TCM nodes and targets of the heterogeneous network to obtain the adjacency matrix of the heterogeneous network;

[0052] For any two TCM nodes, the cosine similarity between them is calculated based on the feature vectors, which is specifically expressed as:

[0053]

[0054] in, and They represent the feature vectors of TCM node a and TCM node b respectively, and k represents the feature number in the feature vector;

[0055] Set the TCM node similarity threshold θ1. If sim(a,b)>θ1, it is determined that TCM node a and TCM node b have similarity, otherwise, they do not have similarity.

[0056] For any two targets, BLAST is used to calculate the target similarity. The similarity between two protein sequences is identified by finding local sequence similarity. The higher the similarity score output by BLAST, the higher the similarity between the two targets.

[0057] Set the target similarity threshold θ2. If the target similarity is greater than the threshold θ2, the two targets are judged to be similar. Otherwise, they are not similar.

[0058] According to the similarity judgment of the above-mentioned TCM nodes and targets and the original association relationship, the adjacency matrix of the heterogeneous network is constructed; among them, corresponding to the TCM-target association network A mt , Chinese medicine-disease association network A md , Chinese medicine-prescription association network A mf , Chinese medicine interaction network A mm , target interaction network A tt Target-disease association network A td , the original association relationship is represented as a [0,1] matrix, 0 means there is no association relationship, 1 means there is an association relationship; corresponding to the Chinese medicine-Chinese medicine similarity network S mm Similarity network S with target-target tt , the similarity relationship is represented as a [0,1] matrix, where 0 means no similarity and 1 means similarity;

[0059] Step 4. Build a graph convolutional neural network model based on the graph attention network GAT, and extract the node feature representation of each node in each graph convolution layer in the heterogeneous network according to the adjacency matrix;

[0060] The key step in identifying the relationship between Chinese medicine and target is feature extraction. The present invention generates feature representations of Chinese medicine and targets in each layer through a heterogeneous graph convolutional neural network. The heterogeneous network constructed by the present invention has four types of nodes: Chinese medicine, target, disease, and prescription, and contains eight types of edge relationships: Chinese medicine-Chinese medicine interaction, target-target interaction, Chinese medicine-target association, Chinese medicine-disease association, Chinese medicine-prescription association, target-disease association, Chinese medicine-Chinese medicine similarity and target-target similarity. A heterogeneous graph convolutional neural network based on a graph attention network GAT is used to perform message transmission on different types of edges to achieve feature aggregation between related nodes and extract features of Chinese medicine and targets in the heterogeneous network. When the graph convolutional neural network has only one layer, the node feature representation obtained by the model will only aggregate the first-order information of its neighbors, while a graph convolutional neural network with multiple layers stacked can enable the feature representation to effectively convolve its high-order neighbor information. Therefore, the present invention adopts a three-layer stacked graph convolutional neural network. In each layer, the feature representation of each node is generated by aggregating the features of adjacent nodes connected by different types of edges.

[0061] Specifically, in each graph convolution layer of the graph convolutional neural network, the type information of the edge is modeled as:

[0062]

[0063] Among them, i represents node information, l represents the number of layers in the graph convolutional neural network, and r represents the edge relationship type information. represents the normalization constant, represents the symmetric square root normalization constant, φ(·) represents the ReLU activation function; represents the neighborhood set of node i with respect to edge relationship r. Similarly, represents the neighborhood set of node j with respect to edge relationship r; represents the training parameters of the lth graph convolutional layer with edge relationship r, represents the feature representation of node i in the lth graph convolutional layer, Represents the neighborhood feature representation of node j in the lth graph convolutional layer, and the original feature representation of each node is initialized as a one-hot vector; Represents the feature representation of node i in the l+1th graph convolutional layer, that is, the feature representation after feature aggregation;

[0064] GAT assigns different weights to adjacent nodes during the node information aggregation process, that is, different weights are assigned to different nodes in a heterogeneous network. GAT performs linear transformation on neighboring nodes respectively, and then assigns values ​​to the nodes in the graph to calculate the attention coefficient. Therefore, GAT is used to enhance the feature learning of nodes; the attention coefficient is specifically expressed as:

[0065]

[0066] Among them, α ij represents the attention coefficient, a (l) represents the training parameters of the l-th graph convolution layer, and σ(·) represents the sigmoid activation function; therefore, the graph convolution layer based on GAT can be finally expressed as:

[0067]

[0068] After the above steps, the feature representation of each layer of each node in the heterogeneous network can be obtained;

[0069] Step 5. Use the self-attention mechanism to perform multi-layer graph convolution fusion on the node feature representation of the heterogeneous network to obtain the fusion feature representation of each traditional Chinese medicine node and target node in the heterogeneous network;

[0070] In conventional methods, the output of the last graph convolution layer is usually taken as the final feature representation of the Chinese medicine node or target node, or multiple layers of feature representations are concatenated as the final feature representation of the Chinese medicine node or target. Different from conventional methods, the present invention uses a self-attention mechanism to prevent information loss.

[0071] Specifically, in the self-attention mechanism, the query vector, key vector and value vector are defined as:

[0072]

[0073] Among them, q l , k l 、v l Denote the query vector, key vector, and value vector of the attention mechanism respectively, W Q , W K Represents the training parameters of the self-attention mechanism;

[0074] For the output of any two layers of graph convolutional layers, the attention coefficient is calculated as:

[0075]

[0076] Among them, e mn represents the attention coefficient of the m-th graph convolution layer and the n-th graph convolution layer, m,n∈(1,2,3);

[0077] Then, for node i, the feature representation after multi-layer graph convolution fusion is:

[0078]

[0079] in, Represents the feature representation of node i in the lth graph convolutional layer of the graph convolutional neural network, l∈(1,2,3);

[0080] Thus, the fusion feature representation of each TCM node and each target node is obtained;

[0081] Step 6. Concatenate the fusion feature representations of the TCM node to be predicted and the target node to form a TCM-target feature vector and send it to the TCM-target relationship prediction model, and the model outputs the TCM-target relationship prediction result;

[0082] After obtaining the fusion feature representation of TCM and target, a deep neural network DNN is used to predict the TCM-target relationship; specifically, the fusion feature representation h of a given TCM node i The fusion feature representation h of the target node j , concatenate them to form the Chinese medicine-target feature vector X = [h i ,h j ] and used as the input vector of the three-layer DNN model, which outputs the prediction results of the relationship between traditional Chinese medicine and target points;

[0083] In the three-layer DNN model, the first layer uses the ReLU function to construct 256 nodes, the second layer selects ReLU as the activation function, and the last layer uses the sigmoid activation function to achieve binary classification; binary cross entropy is used as the loss function, and RMSProp is used as the optimizer to complete the training of the three-layer DNN model; in the loss function, when the predicted value and the true value are equal, the loss is 0, otherwise, the loss is a positive number, and the greater the difference between the predicted value and the true value, the greater the loss.

[0084] The beneficial effects of the present invention are described in detail below in conjunction with experiments.

[0085] In this embodiment, the known Chinese medicine-target associations are taken as positive samples, and the same number of unknown Chinese medicine-target associations are randomly selected as negative samples, and the 10-fold cross-validation method is used for evaluation; for the positive samples, 10% of the entire labeled Chinese medicine-target associations are randomly selected as the test set, and the remaining 90% are used as the training set; for the negative samples, the method of generating the training set and the test set is the same as the method of generating the positive sample; the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPR) are used as the evaluation criteria of the model, and the performance results are shown in Tables 1 and 2. It can be seen from the table that in the Chinese medicine-target relationship prediction task, the AUROC and AUPR values ​​of the present invention reach 0.8569 and 0.8655, respectively, which proves the effectiveness of the present invention.

[0086] Table 1

[0087]

[0088] Table 2

[0089]

[0090]

[0091] The above description is only a specific implementation mode of the present invention. Any feature disclosed in this specification, unless otherwise stated, can be replaced by other alternative features that are equivalent or have similar purposes; all the disclosed features, or all the steps in the methods or processes, except for mutually exclusive features and / or steps, can be combined in any way.

Claims

1. A method for predicting the relationship between traditional Chinese medicine targets based on multi-attention graph neural network learning, characterized in that: The following steps are involved: Step 1. Collect data and build a heterogeneous network; Obtain prescription data, Chinese medicine data, target data and disease data from the ETCM database to construct a Chinese medicine-target association network A mt , Chinese medicine-disease association network A md , Chinese medicine-prescription association network A mf , Chinese medicine interaction network A mm , target interaction network A tt , Chinese medicine-Chinese medicine similarity network S mm , Target-Disease Association Network A td , target-target similarity network S tt There are eight networks in total, which are combined to form a heterogeneous network; Step 2. Initialize the feature vectors of the TCM nodes in the heterogeneous network; Step 3. Perform similarity calculations on the TCM nodes and target nodes of the heterogeneous network to obtain the adjacency matrix of the heterogeneous network; Step 4. Build a graph convolutional neural network model based on the graph attention network GAT, and extract the node feature representation of each node in each graph convolution layer in the heterogeneous network according to the adjacency matrix; Step 5. Use the self-attention mechanism to perform multi-layer graph convolution fusion on the node feature representation of the heterogeneous network to obtain the fusion feature representation of each traditional Chinese medicine node and target node in the heterogeneous network; Step 6. Concatenate the fusion feature representations of the TCM node to be predicted and the target node to form a TCM-target feature vector and send it to the TCM-target relationship prediction model, which outputs the TCM-target relationship prediction result.

2. The method for predicting the relationship between traditional Chinese medicine targets based on multi-attention graph neural network learning according to claim 1, characterized in that: In step 2, the characteristic data of the TCM node include nature, flavor, meridian and toxicity, specifically "cold, hot, warm, cool, neutral; sour, bitter, sweet, pungent, salty; lung, pericardium, heart, large intestine, triple energizer, small intestine, stomach, gallbladder, bladder, spleen, liver, kidney; toxicity"; For the modifier description features, 2 -1 , 2 0 and 2 1 For the attribution characteristics, binary is used for quantization, thereby obtaining the initialized feature vector of each traditional Chinese medicine node.

3. The method for predicting the relationship between traditional Chinese medicine targets based on multi-attention graph neural network learning according to claim 1, characterized in that: In step 3, the similarity calculation process of traditional Chinese medicine nodes is as follows: For any two TCM nodes, the cosine similarity between them is calculated based on the feature vectors, which is specifically expressed as: in, and They represent the feature vectors of TCM node a and TCM node b respectively, and k represents the feature number in the feature vector; Set the TCM node similarity threshold θ1. If sim(a,b)>θ1, it is determined that TCM node a and TCM node b are similar. Otherwise, they are not similar.

4. The method for predicting the relationship between traditional Chinese medicine targets based on multi-attention graph neural network learning according to claim 1, characterized in that: In step 3, the similarity calculation process of the target node is: For any two targets, BLAST was used to calculate the target similarity; A target similarity threshold θ2 is set. If the target similarity is greater than the threshold θ2, the two targets are judged to be similar. Otherwise, they are not similar.

5. The method for predicting the relationship between traditional Chinese medicine targets based on multi-attention graph neural network learning according to claim 1, characterized in that: In step 3, the adjacency matrix is ​​specifically expressed as: Corresponding to the Chinese medicine-target association network A mt , Chinese medicine-disease association network A md , Chinese medicine-prescription association network A mf , Chinese medicine interaction network A mm , target interaction network A tt Target-disease association network A td , the original association relationship is represented as a [0,1] matrix, 0 means there is no association relationship, 1 means there is an association relationship; corresponding to the Chinese medicine-Chinese medicine similarity network S mm Similarity network S with target-target tt , the similarity relationship is represented as a [0,1] matrix, where 0 indicates no similarity and 1 indicates similarity.

6. The method for predicting the relationship between traditional Chinese medicine targets based on multi-attention graph neural network learning according to claim 1, characterized in that: In step 4, the graph convolutional neural network model adopts a three-layer stacked structure, wherein the feature representation of each graph convolution layer is: Among them, i represents node information, l represents the number of layers in the graph convolutional neural network, r represents the edge relationship type information, and α ij represents the attention coefficient, represents the normalization constant, represents the symmetric square root normalization constant, φ(·) represents the ReLU activation function; represents the neighborhood set of node i with respect to edge relation r, represents the neighborhood set of node j with respect to edge relationship r; represents the training parameters of the lth graph convolutional layer with edge relationship r, represents the feature representation of node i in the lth graph convolutional layer, Represents the neighborhood feature representation of node j in the lth graph convolutional layer, and the original feature representation of each node is initialized as a one-hot vector; represents the feature representation of node i in the l+1th graph convolutional layer; a (l) represents the training parameters of the l-th graph convolutional layer, and σ(·) represents the sigmoid activation function.

7. The method for predicting the relationship between traditional Chinese medicine targets based on multi-attention graph neural network learning according to claim 1, characterized in that: In step 5, the query vector, key vector, and value vector are defined in the self-attention mechanism as follows: Among them, q l , k l 、v l Denote the query vector, key vector, and value vector of the attention mechanism respectively, W Q , W K Represents the training parameters of the self-attention mechanism; For any two graph convolutional layers, the attention coefficient is expressed as: Among them, e mn represents the attention coefficient of the m-th graph convolution layer and the n-th graph convolution layer, m,n∈(1,2,3); Then, for node i, the feature representation after multi-layer graph convolution fusion is: in, Represents the feature representation of node i in the lth graph convolutional layer of the graph convolutional neural network, l∈(1,2,3); Thus, the fusion feature representation of each TCM node and each target node is obtained.

8. The method for predicting the relationship between traditional Chinese medicine targets based on multi-attention graph neural network learning according to claim 1, characterized in that: In step 6, the TCM-target relationship prediction model uses a three-layer DNN network.

Images