Drug repositioning method based on graph generative adversarial network and variational autoencoder

Abstract

The application relates to the technical field of bioinformatics, in particular to a drug repositioning method based on a graph generative adversarial network and a variational autoencoder, which comprises the following steps: a drug similarity matrix, a target similarity matrix and a drug-target interaction matrix are utilized to construct a drug-target heterogeneous network, an adjacency matrix and an initial feature matrix; the variational autoencoder is adopted to utilize the feature extraction and fusion functions of a convolutional neural network to encode initial features in the adjacency matrix and the initial feature matrix of the drug-target into hidden variables; based on the drug-target heterogeneous network, an adversarial model is utilized to strengthen the feature representation of drugs and targets in the hidden variables; after multiple iterations of the adversarial model, the hidden variables in the generator are extracted to construct a prediction matrix; and the unknown drug-target interaction is predicted. The application has good performance in drug-target interaction prediction and good effect in drug repositioning.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of bioinformatics, and particularly relates to a drug repositioning method based on a graph generative adversarial network and a variational autoencoder. BACKGROUND

[0002] In the past few decades, targeted and personalized therapy has become the most promising strategy for treating complex human diseases. Therefore, the important task of predicting drug-target interaction (DTI) has been actively studied. The target protein family with medicinal value, such as enzyme, ion channel, G protein-coupled receptor (GPCR) and nuclear receptor, can regulate its function through interaction with drug compounds. In most cases, the traditional method of predicting DTI in vitro and in vivo is desirable, but the experimental and testing process is both time-consuming and expensive. At present, the computational methods for DTI prediction are increasingly popular in drug repositioning, as they effectively avoid the tedious and lengthy drug experimental process of traditional methods. Researchers can take advantage of the efficiency and scalability of these methods to quickly provide supporting evidence for experimental results.

[0003] In recent years, machine learning-based techniques have been widely used for DTI prediction. For example, a binary local model was constructed and a support vector machine (SVM) was used to predict DTI. A random forest-based computational method was proposed to predict unknown DTI by integrating compound chemical properties, biological properties and indications of DTI. It is proposed to use regularized least squares and Gaussian interaction profile (GIP) kernel to predict the interaction of new drugs and targets. A regularized linear classification model based on the least absolute shrinkage and selection operator (LASSO) was proposed to predict DTI. An algorithm was proposed to consider the global network topology by a restarted random walk algorithm to weight the features of drugs and targets, and then train a K-nearest neighbor model to predict the affinity score of drugs and targets. Although the prediction method based on traditional machine learning models has achieved good performance, they are still limited to small-scale data sets.

[0004] With the explosive growth of large-scale biomedical data, deep learning gradually replaces traditional machine learning and becomes the mainstream technology in the DTI prediction task, because it is more suitable for large-scale data sets than traditional machine learning. A model based on convolutional neural network (CNN) is proposed, which uses the one-dimensional representation of the protein sequence of the target and the compound of the drug to predict the binding affinity of the drug-target pair. A deep learning model is developed, which extracts features through a multi-modal deep autoencoder to predict potential DTIs. A deep learning model based on long short-term memory neural network is developed. A DTI prediction framework is designed by combining the denoising autoencoder and CNN, which can effectively deal with noise and incomplete information from heterogeneous data sources. A molecular interaction conversion model based on transformer and CNN is developed to predict unknown DTIs. The deep learning-based prediction method usually directly extracts high-level features from data to perform the DTI prediction task. In order to improve the evaluation accuracy, these methods greatly increase the number of neural network layers, which often leads to the homogenization of node features of drugs and targets. Intuitively, it is like that all the top-ranked predictions are drugs related to a certain class of targets, and such results are not diverse enough, rely on specific data sets, and have weak generalization ability.

[0005] Recently, researchers have introduced graph neural networks (GNN) into the DTI prediction task, because GNN can capture complex hidden features behind heterogeneous data. A simple but high-performance prediction model based on residual GNN and DTI attention network is developed. A prediction model for the DTI prediction task is designed based on the multi-graph convolutional network (GCN) and the deep residual network (ResNet). A multi-channel GNN based on graph attention network is designed to predict DTIs. The GCN is improved using the transformer network, which overcomes the problem that the GCN does not learn enough edge features. Existing methods do not consider the consistency of the intrinsic relationship between drugs and targets in different feature spaces, which often leads to poor results for drug repositioning. The invention defines the correlation between drug or target features to measure this relationship, and maintains the consistency of the correlation between drug and target features in different feature spaces.

[0006] In summary, there are still some problems that can be considered and solved for the existing DTI prediction methods. For example: the prediction method based on the traditional machine learning model is still limited to small-scale data sets; the current method does not fully consider the consistency of the inherent internal relationship between drug and target characteristics in different characteristic spaces; and the current graph neural network model increases the number of network layers to improve prediction accuracy, but too many layers will cause the node characteristics of drugs and targets to be too homogeneous, often leading to poor drug repositioning results. SUMMARY

[0007] The summary part of the application is used to introduce the concepts in a brief form, which will be described in detail in the specific embodiment part. The summary part of the application is not intended to identify the key features or essential features of the claimed technical solution, nor is it intended to limit the scope of the claimed technical solution.

[0008] In order to solve the technical problem of poor drug repositioning effect, the application provides a drug repositioning method based on a graph generative adversarial network and a variational autoencoder.

[0009] The application provides a drug repositioning method based on a graph generative adversarial network and a variational autoencoder, which comprises:

[0010] A drug-target heterogeneous network, an adjacency matrix and an initial feature matrix are constructed using known drug similarity matrices, target similarity matrices and drug-target interaction matrices;

[0011] The variational autoencoder uses the feature extraction and fusion function of the convolutional neural network to encode the initial features in the drug-target adjacency matrix and the initial feature matrix into latent variables;

[0012] Based on the drug-target heterogeneous network, the feature representation of drugs and targets in the latent variables is strengthened using an adversarial model, wherein the adversarial model is a generative adversarial network with an attention mechanism;

[0013] After the adversarial model is iterated for a preset number of times, the latent variables in the generator are extracted to construct a prediction matrix;

[0014] The unknown drug-target interaction is predicted according to the prediction matrix.

[0015] Further, the construction of the drug-target heterogeneous network, the adjacency matrix and the initial feature matrix using the known drug similarity matrix, the target similarity matrix and the drug-target interaction matrix comprises:

[0016] The value in the drug-drug interaction matrix corresponding to the formula is determined using the known drug similarity matrix:

[0017]

[0018] wherein d (i,j) is the value of the ith row jth column in the drug-drug interaction matrix, S i,j denotes the similarity score of the ith drug and the jth drug in the drug similarity matrix, and μ and σ represent the mean and variance of all samples in the drug similarity matrix, respectively;

[0019] determining a target-target interaction matrix using the target similarity matrix;

[0020] completing information in the drug-drug interaction matrix, the target-target interaction matrix and the drug-target interaction matrix in a matrix completion manner to construct an adjacency matrix of drug-target;

[0021] completing information in the drug similarity matrix, the target similarity matrix and the drug-target interaction matrix in a matrix completion manner to construct an initial feature matrix of drug-target;

[0022] constructing a heterogeneous network of drug-target based on the interaction relationship in the adjacency matrix of drug-target, wherein when the value of the ith row jth column in the adjacency matrix of drug-target is 1, it indicates that the ith and the jth nodes have an edge in the heterogeneous network, and the value of 0 means no edge.

[0023] Further, the training process of the adversarial model comprises:

[0024] the generator included in the adversarial model is used to construct a traversal tree using all drug and target nodes in a breadth-first traversal manner, to calculate the probability of selecting the next hop node in the search process using the feature representation of drugs and targets in the latent variable, and to generate drug and target pairs with interaction;

[0025] the discriminator included in the adversarial model is used to receive node pairs sampled from the heterogeneous network of drug-target and node pairs generated by the generator, and to give the possibility of the existence of a connection relationship by the classifier included in the discriminator according to the features of the node pairs;

[0026] In the training process of the adversarial model, the correlation consistency of drug or target features is maintained by minimizing the correlation error between drug or target features in the latent variable and the initial feature matrix, so as to achieve the purpose of maintaining the inherent chemical properties and attributes of drugs and targets unchanged;

[0027] The adversarial model updates the information between drugs and targets in the latent variable after each iteration to strengthen the feature representation of drugs and targets in the latent variable.

[0028] Further, the predicting the unknown drug-target interaction according to the prediction matrix comprises:

[0029] The predicting the unknown drug-target interaction is completed according to ranking of the filling values in the prediction matrix, wherein the filling values in the prediction matrix correspond to the possibility of existence of the interaction of a pair of drug-target.

[0030] Further, the steps of implementing the generator and the discriminator comprise:

[0031] The latent variable Z is represented by Z G and Z D in the generator and the discriminator respectively, Z i and Z j represent the feature vectors of the i-th row and the j-th row thereof;

[0032] The generator is used to implement the following steps:

[0033] The attention score of the drug d i and the target t j is calculated in the following manner:

[0034] e ij =tanh((W G Z I +b G ) T (W G Z j ))

[0035]

[0036] wherein W G and b G are the weight and bias parameters of the attention layer in the generator respectively, is the set of neighbors of the drug d i .

[0037] The new representation of the node i is:

[0038]

[0039] wherein the node i is a drug or a target;

[0040] The sample provided by the generator to the discriminator is obtained by sampling from a breadth-first search tree constructed in advance, V cur represents the currently searched node, and V next represents the next node to be searched, wherein the probability of the next hop in the sampling path corresponds to the following formula:

[0041]

[0042] wherein Z' i and Z' j are the representations of V G and V cur in Z' next respectively, is the set of all neighbors of the ith node;

[0043] The discriminator is used to implement the following steps:

[0044] The formula corresponding to the new feature of the ith node is:

[0045]

[0046] wherein e ij = tanh((W D Z i +b D ) T (W D Z j )) and W D and b D are the trainable weights and bias parameters of the attention layer in the discriminator respectively;

[0047] The discriminator simultaneously receives the drug-target node pairs sampled from the heterogeneous network and the node pairs from the generator, and uses the classifier to measure the input node pairs, the formula corresponding to the probability of interaction between the node pairs is:

[0048]

[0049] D out is fed back to the generator to participate in the gradient update of the generator, so that the generator updates the latent variable;

[0050] The correlation between the features of the drugs and the targets is predefined before the training of the adversarial model, and the consistency of the correlation between the features of the drugs and the targets is maintained in a regularized manner;

[0051] The formula of the correlation before training is:

[0052]

[0053] wherein S(i,j) is the correlation before training;

[0054] The formula of the correlation after training is:

[0055]

[0056] wherein S'(i,j) is the correlation after training, Z i and Z jZ G ;

[0057] Minimize To achieve the optimization goal, the corresponding formula is:

[0058]

[0059] Wherein, lambda is a parameter for adjusting the contribution of consistency to overall loss, W is a trainable weight parameter in attention mechanism, gamma is a weight factor for controlling the influence of attention weight on the model, and k is the total number of samples;

[0060] The generator and the discriminator are iterated in an antagonistic game manner, the feature representation learned by the generator is extracted, and a prediction matrix is constructed to achieve the purpose of predicting drug-target interaction.

[0061] The present application has the following beneficial effects:

[0062] The drug repositioning method based on the graph generation adversarial network and the variational autoencoder of the application is based on the similarity of drugs, the similarity of targets and the drug-target interaction information, and uses the encoder model and the generation adversarial network model for joint training to obtain reliable feature representation for drug-target interaction prediction. Firstly, the drug-target heterogeneous network, the adjacency matrix and the initial feature matrix are constructed in the form of matrix completion using the data. Secondly, the variational autoencoder is used to encode the drug and target adjacency matrix and the initial feature into the hidden variable using the feature extraction and fusion function of the convolutional neural network. Then the attention mechanism and the generation adversarial network are used to strengthen the feature representation of the drug and target in the hidden variable. The generator uses all the drug and target nodes to construct a traversal tree in a breadth-first traversal manner, and uses the representation of the drug and target in the hidden variable to calculate the probability of selecting the next hop node in the search process, so as to generate the drug and target pair considered to have interaction by the generator. The discriminator receives the node pair sampled from the drug-target heterogeneous network and the node pair generated by the generator, and gives the possibility of the existence of the connection relationship of the node pair by its own classifier. The model updates the information between the drug and target in the hidden variable after each iteration to achieve the purpose of strengthening. In the training process, the correlation error between the drug or target features in the hidden variable and the initial feature matrix is minimized to maintain the correlation consistency of the drug or target features, so as to realize the purpose of maintaining the inherent chemical properties and attributes of the drug and target unchanged. Finally, the prediction matrix is constructed by the hidden variable of the drug and target learned by the generator of the adversarial model, so as to realize the purpose of predicting the drug-target interaction. Tests on multiple data sets show that the application has good performance in drug-target interaction prediction. Since the chemical properties and functions of drugs and targets are inherent and not easy to change. Drugs with similar chemical structures usually have similar binding targets or treat similar diseases. It can be inferred that there is a similar intrinsic relationship between the features of different drugs or targets. In the training of the graph neural network, the features of the nodes are mapped from one space to another. The existing method does not consider the consistency of this intrinsic relationship in different feature spaces. The application defines the correlation between the drug or target features to measure this relationship, and maintains the consistency of the correlation between the drug and target features in different feature spaces. BRIEF DESCRIPTION OF DRAWINGS

[0063] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, and the advantages thereof, below, a brief introduction will be given to the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are only some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor based on these drawings.

[0064] Figure 1 Flow chart of the drug repositioning method based on graph generative adversarial network and variational autoencoder according to the present application;

[0065] Figure 2 Schematic diagram of the construction of drug-target heterogeneous network according to the present application;

[0066] Figure 3 Workflow diagram of GGANDTI according to the present application;

[0067] Figure 4 Schematic diagram of the construction and sampling process of BFS tree according to the present application;

[0068] Figure 5 Schematic diagram of the comparison of node correlation before and after training according to the present application;

[0069] Figure 6 Schematic diagram of the performance of average AUROC and AUPR at different λ according to the present application;

[0070] Figure 7 Schematic diagram of the performance comparison on different balanced datasets according to the present application. DETAILED DESCRIPTION

[0071] In order to further illustrate the technical means and effects taken by the present application to achieve the predetermined object, the specific embodiments, structures, features and effects of the technical solutions proposed by the present application are described in detail below in combination with the drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. In addition, the specific features, structures or characteristics in one or more embodiments can be combined in any suitable form.

[0072] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which the present application belongs.

[0073] The present application provides a drug repositioning method based on graph generative adversarial network and variational autoencoder, which comprises the following steps:

[0074] Using known drug similarity matrix, target similarity matrix and drug-target interaction matrix, a drug-target heterogeneous network, an adjacency matrix and an initial feature matrix are constructed;

[0075] Using the feature extraction and fusion function of convolutional neural network, the variational autoencoder encodes the initial features in the adjacency matrix and the initial feature matrix of drug-target into latent variables;

[0076] Based on the drug-target heterogeneous network, the feature representation of the hidden variables of the drug and the target is strengthened by using an adversarial model, wherein the adversarial model is a generative adversarial network with an attention mechanism.

[0077] After multiple iterations of the adversarial model, the hidden variables in the generator are extracted to construct a prediction matrix.

[0078] According to the prediction matrix, the unknown drug-target interaction is predicted.

[0079] The above steps are described in detail as follows:

[0080] Reference Figure 1 , shows the flow of some embodiments of the drug repositioning method based on the graph generative adversarial network and the variational autoencoder according to the present application. The drug repositioning method based on the graph generative adversarial network and the variational autoencoder comprises the following steps:

[0081] Step S1, using the known drug similarity matrix, target similarity matrix and drug-target interaction matrix, constructing the drug-target heterogeneous network, adjacency matrix and initial feature matrix.

[0082] In some embodiments, the known drug similarity matrix, target similarity matrix and drug-target interaction matrix can be used to construct the drug-target heterogeneous network, adjacency matrix and initial feature matrix.

[0083] It should be noted that the Yamanishi dataset is a gold standard dataset. Yamanishi et al. named and divided it into four benchmark sub-datasets according to protein families, including: G protein-coupled receptors (GPCR), ion channels (IC), nuclear receptors (NR) and (4) enzymes (E). Each sub-dataset includes drug similarity, target similarity and drug-target interaction. The Luo dataset includes drug-target interaction, drug-drug interaction, target-target interaction, and two similarity information: drug-drug similarity, target-target similarity. Table 1 lists the statistical data of the final used dataset.

[0084] Table 1

[0085]

[0086]

[0087] The data marked with "*" in Table 1 is counted after converting the similarity matrix into a binary interaction matrix.

[0088] As an example, this step may include the following steps:

[0089] The first step is to use the known drug similarity matrix and statistical inference to set a cutoff value to transform the drug similarity matrix into a drug-drug interaction matrix. The transformation strategy formula corresponding to the values ​​in the drug-drug interaction matrix is ​​as follows:

[0090]

[0091] Where, d (i,j) S is the value in the i-th row and j-th column of the drug-drug interaction matrix. i,j This represents the similarity score between drug i and drug j in the drug similarity matrix. μ and σ represent the mean and variance of all samples in the drug similarity matrix, respectively. Referring to the standard normal distribution table, when the significance level α ≤ 0.05, the corresponding Z-test statistic (Zi = 0.05) is... α It should be greater than or equal to 1.64.

[0092] It should be noted that the construction of drug-target heterogeneous networks depends on information about the interaction between the drug and the target. For example... Figure 2 As shown, let D = {d1, d2, ..., d} m} and T={t1,t2,…,t n Let} represent m drug nodes and n target nodes, respectively. Drug d i and target t j The interaction is caused by dt (i,j) Indicates. Drug d i and drug d j The interaction is represented by d (i,j) Indicates. Target t i and target t j The interaction is caused by t (i,j) This indicates that for the Luo dataset, drug-target interactions, drug-drug interactions, and target-target interactions are known. However, for the Yamanishi dataset, only the DTI (Drug-Target Interaction) is known. Therefore, this invention obtains drug-drug interactions and target-target interactions by transforming the similarity matrix.

[0093] Hypothesis testing is a statistical inference method used to determine differences between samples and between a sample and the population. This invention uses the critical value of a one-sample Z-test based on hypothesis testing principles as a threshold to truncate similarity scores, thus transforming the similarity matrix into a binary interaction matrix. This transformation relationship can be expressed using a transformation strategy formula. express.

[0094] The second step is to use the target similarity matrix to determine the target-target interaction matrix.

[0095] For example, the first step included in step S1 can be referred to as an example, the target similarity matrix is converted into a target-target interaction matrix.

[0096] The third step is to complete the information of the drug-drug interaction matrix, the target-target interaction matrix and the drug-target interaction matrix in the form of matrix completion, and construct the adjacency matrix of drug-target.

[0097] The fourth step is to complete the information of the drug similarity matrix, the target similarity matrix and the drug-target interaction matrix in the form of matrix completion, and construct the initial feature matrix of drug-target.

[0098] The fifth step is to construct the heterogeneous network of drug-target based on the interaction relationship in the adjacency matrix of drug-target.

[0099] Wherein, when the value of the i-th row and j-th column in the adjacency matrix of drug-target is 1, it means that the i-th and j-th nodes have edges in the heterogeneous network, and the value is 0, which is considered to have no edge.

[0100] It should be noted that the present application calculates the Z score (Z α =(S ij -μ) / σ) between drugs and targets. If Z α ≥1.64, it means that the sample to be tested is significantly different from the population, that is, it is considered that the drug d i and d j (target t i and t j ) exist interaction. If dt (i,j) =1, it is considered that there is an edge between d i and t j in the drug-target network. The processing method of drug-drug network and target-target network is the same. By detecting the existence of edges in the interaction matrix, drug-drug network, target-target network and drug-target network can be constructed. The drug-target heterogeneous network is constructed by integrating the three networks. The statistical data of the converted binary network is shown in Table 1.

[0101] Based on the topological information of the drug-target heterogeneous network, the present application constructs its corresponding adjacency matrix and initial feature representation. From the perspective of network construction, the adjacency matrix (A) can be described as follows:

[0102]

[0103] Wherein, A DD , A TT and A DT are the interaction matrices of drug-drug network, target-target network and drug-target network, respectively. is A DT transpose.

[0104] Next, the drug similarity matrix and the target similarity matrix are represented by S D and S T respectively. Similarly, the initial feature matrix (X) of drug and target nodes can be described by the following equation:

[0105]

[0106] where X ∈ R (m+n)×(m+n) Each row of X represents the feature vector of a node. The first m and last n vectors represent the features of drugs and targets, respectively. Take drugs as an example, the feature vector X i of drug d i is concatenated by and . is a vector representation composed of the similarity between drug d i and all other drugs, is a vector representation composed of the interaction between drug d i and all other targets. Similarly, the feature vector X j of target t j is concatenated by and .

[0107] In step S2, the variational autoencoder is used to encode the initial features in the drug-target adjacency matrix and the initial feature matrix into latent variables using the feature extraction and fusion function of the convolutional neural network.

[0108] In some embodiments, the variational autoencoder can be used to encode the initial features in the drug-target adjacency matrix and the initial feature matrix into latent variables using the feature extraction and fusion function of the convolutional neural network.

[0109] As an example, the variational autoencoder can be used to encode the initial features in the drug-target adjacency matrix and the initial feature matrix into latent variables using the feature extraction and fusion function of the convolutional neural network.

[0110] It is worth mentioning that the work of the generator is to fit the real characteristics of drugs and targets as much as possible to improve the final prediction performance. Therefore, before conducting adversarial training, the features of drugs and targets need to be encoded for initializing the generator and discriminator. In this part, the variational graph autoencoder (VGAE) is used to encode the features of drugs and targets. VGAE is a framework for learning graph data based on variational autoencoder, which can learn the latent variables of undirected graph according to the data distribution. VGAE includes an encoding part and a decoding part. The encoding part takes the adjacency matrix (A) and the initial feature matrix (X) as input and obtains the latent variable Z as output. The corresponding formula is as follows:

[0111]

[0112]

[0113] where, is the degree matrix. I is the identity matrix. μ and σ are the mean and variance of the data distribution, respectively. W0, W μ and W σ are trainable weight parameters. Φ μ and φ σ are the operations for calculating the mean and variance, respectively. Then, the latent variable is obtained using the reparameterization trick as follows:

[0114] Z = μ + σ * ε

[0115] Here ε ~ N(0, 1).

[0116] In the decoding part, VGAE decodes from Z and reconstructs a new adjacency matrix The corresponding formula of the decoder is as follows:

[0117]

[0118] The loss function includes two parts. The first part is the binary cross-entropy between A and , while the second part is the KL divergence between the latent variable distribution q(Z|X, A) and the standard normal distribution p(Z). p(A|Z) represents the conditional probability of having an edge in . The corresponding formula is as follows:

[0119]

[0120] The process of encoding features is performed independently throughout the workflow. This part is only used to obtain the latent variable Z of drugs and targets, as it is the basis for subsequent work.

[0121] Step S3, based on the drug-target heterogeneous network, the feature representation of drugs and targets in the latent variable is strengthened by using an adversarial model.

[0122] In some embodiments, based on the drug-target heterogeneous network, the feature representation of drugs and targets in the latent variable can be strengthened by using an adversarial model.

[0123] Wherein, the adversarial model is a generative adversarial network with attention mechanism.

[0124] It should be noted that compared with traditional graph representation learning methods such as DeepWalk, LINE and Node2Vec, graph generative adversarial network (GraphGAN) has achieved better performance on graph structure data. Nevertheless, GraphGAN still does not pay enough attention to neighbor nodes. This part introduces attention mechanism to improve GraphGAN. In addition, a correlation constraint is defined to ensure that the distribution of drugs and targets is consistent in different feature spaces. This helps to obtain more reliable feature representation consistent with the chemical properties and functions of drugs and targets.

[0125] As shown in Figure 3 , this part regards drug-target interaction as connectivity in the graph, aiming to learn two models: (1) Generator. It assumes that for each node (V c ) in the graph, there is a latent true connectivity distribution related to it, which contains the connectivity distribution of V c relative to all other nodes in the graph. The edges in the graph can be regarded as observed samples generated from these connectivity distributions, and then the generator can learn the node representation by maximizing the likelihood of the edges in the graph. (2) Discriminator. It does not regard the edge as the edge generated from the connectivity distribution, but learns a classifier to directly predict the existence of the edge. In other words, the discriminant model takes two nodes V i and V j as features, and predicts the probability of the existence of the edge between the two nodes according to the classifier learned by itself.

[0126] As shown in Figure 3 , first, the constructed initial feature matrix X and adjacency matrix A are encoded into latent variable Z. Then, the generator and discriminator are initialized using Z, and then the information between drugs and targets in Z G is strengthened by using the generator sampling, attention mechanism and adversarial training. Finally, the final Z G is used to perform prediction.

[0127] The generator attempts to reduce the distance between the distribution of the generator and the real distribution, so the generator aims to minimize the log probability that the discriminator correctly assigns a negative label to a sample generated by the generator. The features of drugs and targets in the generator are initialized by the latent variables Z from the encoder. In order to flexibly capture the relationship between global nodes and local nodes, an attention mechanism is introduced and a new representation containing the importance of node neighbors is output.

[0128] As an example, the training process of the adversarial model can include the following steps:

[0129] Firstly, the adversarial model needs to train two modules: the generator and the discriminator, and the training of the two modules can include the following sub-steps:

[0130] The first sub-step, the generator included in the adversarial model is used to build a traversal tree using all drug and target nodes in a breadth-first traversal manner, and the features of drugs and targets in the latent variables are used to calculate the probability of selecting the next hop node in the search process, so as to generate drug and target pairs with interactions.

[0131] For example, the generator included in the adversarial model is used to build a traversal tree using all drug and target nodes in a breadth-first traversal manner, and the features of drugs and targets in the latent variables are used to calculate the probability of selecting the next hop node in the search process, so as to generate drug and target pairs that the generator itself considers to have interactions.

[0132] The second sub-step, the discriminator included in the adversarial model is used to receive node pairs sampled from the heterogeneous network of drugs and targets and the node pairs generated by the generator, and give the possibility of the existence of connection relationship of the node pairs as features by the classifier included in the discriminator.

[0133] For example, the discriminator included in the adversarial model is used to receive node pairs sampled from the heterogeneous network of drugs and targets and the node pairs generated by the generator, and give the possibility of the existence of connection relationship of the node pairs as features by its own classifier.

[0134] The second step, in the training process of the adversarial model, the correlation consistency of drug or target features is maintained by minimizing the correlation error between the drug or target features in the latent variables and the initial feature matrix, so as to achieve the purpose of maintaining the inherent chemical properties and attributes of drugs and targets unchanged.

[0135] The third step, the adversarial model updates the information between drugs and targets in the latent variables after each iteration, in order to strengthen the feature representation of drugs and targets in the latent variables.

[0136] As another example, the generator and the discriminator can implement the following steps:

[0137] The first step is to use Z in both the generator and discriminator. G and Z D Z indicates i and Z j This represents the eigenvectors of its i-th and j-th rows.

[0138] The second step involves using a generator to implement the following sub-steps:

[0139] First sub-step, drug d i and target t j The features are represented in the generator by Z. i and Z j It indicates that drug d i and target t j Attention score is calculated as follows:

[0140] e ij =tanh((W G Z i +b G ) T (W G Z j ))

[0141] Among them, W G and b G These are the weights and bias parameters of the attention layer in the generator, respectively. i and Z j Shared weight parameter W G To eliminate the negative impact of outlier data, attention scores need to be normalized:

[0142]

[0143] in, It is drug d i The collection of neighbors.

[0144] In the second sub-step, the new representation of node i (drug or target) is:

[0145]

[0146] In this context, node i represents a drug or target.

[0147] It should be noted that the output Z′ of the attention layer i It contains the neighborhood information of node i. The generator can be based on Z′. G Drug-target pairs are selected. Each node (drug or target) is selected as the root node, and then a tree is constructed using breadth-first search (BFS) based on the connections to the heterogeneous drug-target network. The construction process is as follows: Figure 4 As shown in (1), the sampling strategy is as follows:Figure 4 (2) is shown. Sampling starts from the root node (V)(drug d) i or target t j Start by traversing the branches in sequence. After passing through the multi-branch node (V)... cur When selecting the next node V, choose the node V. next The probability can be expressed as G. out express.

[0148] The third sub-step involves the generator providing samples to the discriminator, obtained from a pre-constructed breadth-first search tree, V. cur V represents the node currently found in the tree. next Let represent the next node to be searched. The probability of the next hop in the sampling path is calculated using the following formula:

[0149]

[0150] Among them, Z′ i and Z′ j They are Z′ G China V cur and V next The expression . It is the set of all neighbors of the i-th node. Subsequent nodes are sampled in the same way until the selected node (V) is reached. next ) is the current node (V) cur The parent node (V) f Next, V cur These will be recorded as sampling points. At this point, an edge sample (V, V) can be obtained. cur If both the sampling node and the root node are drug nodes or target nodes, the current sampling node is discarded, and sampling is started from the root node.

[0151] Third, the discriminator is used to implement the following sub-steps:

[0152] It should be noted that the discriminator measures drug d i and target t j There are probabilities of margins existing, so the discriminator aims to maximize the log probability of assigning the correct label to both positive and negative samples. Similar to the generator's initialization, the discriminator's feature representation is initialized with the latent variable Z from the encoder. Because the generator's capability depends on the discriminator, improving the discriminator's capability allows the generator to better fit the true distribution. The same attention mechanism is introduced here to enhance the discriminator's recognition ability. (drug d) i and target t j The features are represented in the discriminator by Z respectively. i and Z j express.

[0153] The first sub-step, the formula corresponding to the new feature of the i-th node is:

[0154]

[0155] where e ij = tanh((W D Z i +b D ) T (W D Z j )) is the activation function of the attention layer in the discriminator, and W D and b D are the trainable weights and bias parameters of the attention layer in the discriminator, respectively.

[0156] The second sub-step, the discriminator receives the drug-target node pairs sampled from the heterogeneous network and the node pairs from the generator at the same time, and uses the classifier to measure the input node pairs, the formula corresponding to the probability of interaction between the node pairs is:

[0157]

[0158] For example, the discriminator receives some edge samples randomly sampled from the drug-target heterogeneous network and other edge samples sampled from the BFS tree of the generator. The discriminator measures the connectivity between the drug d D and the target t i according to Z′ j . The feature representations of the drug d i and the target t j are represented by Z′ D , Z′ i and Z′ j in Z′ out , respectively. The formula corresponding to the probability of interaction between them is:

[0159] The third sub-step, D out is fed back to the generator to participate in the gradient update of the generator, so that the generator updates the latent variable.

[0160] For example, D out is fed back to the generator to participate in the gradient update of the generator, helping the generator to update the latent variable.

[0161] The fourth sub-step, the correlation between the features of the drug and the target is predefined before the training of the adversarial model, and the consistency of the correlation between the features of the drug and the target is maintained in a regularized manner.

[0162] It should be noted that in the adversarial model, the type of node is completely ignored. This means that the inherent relationship between different drugs or targets is not considered during the training process. Since the chemical properties and functions of drugs and targets are inherent and not easily changed, it can be inferred that this inherent relationship is also stable and unchanging during the learning process of graph features. Therefore, maintaining this inherent relationship is necessary for feature learning of biological networks. The present invention defines the correlation between drug or target features to measure this relationship and maintains the consistency of the correlation between drug and target features in different feature spaces. The features of drug d i and target t j are represented as Z i and Z j , respectively.

[0163] The fifth sub-step, the correlation formula before training is:

[0164]

[0165] where S(i,j) is the correlation before training.

[0166] The sixth sub-step, the correlation formula after training is:

[0167]

[0168] where S'(i,j) is the correlation after training, and the representation of Z i and Z j comes from Z G in the generator.

[0169] The seventh sub-step minimizes to achieve the optimization goal, and the corresponding formula is:

[0170]

[0171] where λ is a parameter that adjusts the contribution of consistency to the overall loss, W is a trainable weight parameter in the attention mechanism, γ is a weight factor that controls the influence of attention weight on the model, and k is the total number of samples.

[0172] It should be noted that in order to maintain the correlation consistency of nodes during prediction, the loss function of the adversarial model is set to The present invention uses the Adam optimizer to optimize the loss function. Finally, the features from the generator are used to predict unknown DTI:

[0173]

[0174] where the features of drug and target are represented as Z d and Zt .

[0175] The GAN aims to fit more reliable drug and target features by adversarial learning and constraining relevant consistency.

[0176] In the fourth step, the generator and the discriminator are iterated in an adversarial game, and the feature representation learned by the generator is extracted to construct a prediction matrix to achieve the purpose of predicting drug-target interactions.

[0177] For example, the generator and the discriminator are iterated in an adversarial game, and finally the ability of the generator is improved to a stable level. Finally, the feature representation learned by the generator is extracted to construct a prediction score matrix to achieve the purpose of predicting drug-target interactions.

[0178] In step S4, after the adversarial model is iterated for a preset number of times, the hidden variable in the generator is extracted to construct a prediction matrix.

[0179] In some embodiments, after the adversarial model is iterated for a plurality of times, the hidden variable in the generator is extracted to construct a prediction matrix.

[0180] As an example, after the adversarial model is iterated for a preset number of times, the hidden variable in the generator is extracted to construct a prediction matrix to achieve the purpose of predicting drug-target interactions.

[0181] In step S5, according to the prediction matrix, the interaction of unknown drug-target is predicted.

[0182] In some embodiments, according to the prediction matrix, the interaction of unknown drug-target can be predicted.

[0183] As an example, according to the ranking of the fill values in the prediction matrix to determine the candidate drug-target pairs, the prediction of the interaction of unknown drug-target is completed, wherein the fill value in the prediction matrix corresponds to the possibility of the existence of the interaction of a pair of drug-target. The larger the fill value is, the more likely there is a potential interaction between the drug and the target. According to this, the ranking is performed to determine the candidate drug-target pairs, and the prediction of the interaction of unknown drug-target can be completed.

[0184] Optionally, the experimental verification is as follows:

[0185] The area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AUPR) are standard evaluation metrics often used in DTI prediction tasks. The receiver operating characteristic (ROC) curve is composed of a set of key points generated by continuously moving the "threshold" of the classifier. These key points can be calculated from four kinds of prediction results: the number of edges that actually exist and are predicted to exist is denoted as TP; the number of edges that actually exist but are predicted not to exist is denoted as FN; the number of edges that actually do not exist but are predicted to exist is denoted as FP; the number of edges that actually do not exist and are predicted not to exist is denoted as TN. The coordinates of these key points correspond to the true positive rates (TPRs) and false positive rates (FPRs) at different thresholds, respectively. Their corresponding formulas are:

[0186]

[0187] TPRs focus on all positive samples, and FPRs focus on all negative samples. Neither TPRs nor FPRs depend on the specific distribution of classes, nor do they change with changes in class distribution. This is both the advantage and disadvantage of AUROC. The dataset used by the present invention is imbalanced, and if there are too many negative samples, FPRs will not increase significantly, which can lead to overly optimistic estimates of the ROC curve. In this case, AUPR is considered to provide more information and provide a better evaluation reference value, as both axes of the precision-recall (P-R) curve focus on positive samples. Similarly, the P-R curve is composed of another set of key points generated by continuously moving the "threshold" of the classifier. The coordinates of these key points correspond to the Precision and Recall at different thresholds, respectively:

[0188]

[0189] The present invention uses ten-fold cross-validation (CV) to evaluate the performance of GGANDTI and compare it with the most advanced frameworks. The present invention divides the randomly shuffled edge pairs (drug-target) into ten subsets with the same percentage and ensures that the number of positive edge pairs (negative edge pairs are similar) in all subsets is equal. Because the data is imbalanced, using standard CV can result in some subsets having no positive edge pairs, which can lead to errors and meaningless training. To balance the data, the present invention randomly under-samples the negative edge samples until the number of positive and negative edge pairs in each subset is equal. In the training phase, 9 subsets are used to fit the model, and the remaining 1 subset is used to evaluate the model. In ten-fold cross-validation, each subset is selected in turn as the test set. The present invention calculates the average of the evaluation results of the 10 test subsets to report the performance of the model. To obtain more reliable evaluation results, the present invention repeats the cross-validation ten times and calculates their average.

[0190] GGANDTI was compared with five state-of-the-art drug target prediction frameworks, including DTINet, deepDR, deepDTNet, MolTrans, and EEG-DTI. To ensure sufficient fairness, the present invention trained and evaluated all frameworks using the same dataset. All results were based on ten-fold cross-validation and the reliable values of repeating this process 10 times. Moreover, the original DTINet, deepDR, deepDTNet, EEG-DTI also introduced drug-side effect and drug-disease topology information. However, this part of information was not used in the dataset of the present invention. Because the topology information related to drugs or targets alone is not an advantage feature, but introduces new noise. The main features should be those that are closely related to both drugs and targets (drug-target interactions, drug similarity, and target similarity). Therefore, in the benchmark test of the present invention, this part of information was deleted to maintain consistency with other models, while ensuring that it does not affect the training and evaluation of the original model.

[0191] Based on the evaluation of the performance of multiple methods, Table 2 reports the average AUROC and AUPR values of multiple methods on the Yamanishi and Luo datasets.

[0192] Table 2

[0193]

[0194]

[0195] The present application compares GGANDTI with several other methods. As shown in Table 2, GGANDTI performs best in the comparison of average AUROC and AUPR. This indicates that the method of the present application is effective in the task of identifying unknown DTIs. On the Yamanishi dataset, GGANDTI achieves consistent improvement over other baseline methods, possibly benefiting from the superior performance of the graph generative adversarial model and the constraint on feature consistency. Unlike the Yamanishi dataset, the Luo dataset has higher sparsity, but the average AUPR performance of GGANDTI is still very good compared to other state-of-the-art methods. In the later experiments, the present application specifically evaluates the adaptability of all methods to imbalanced data. In addition, in the Nuclear receptor dataset, there are fewer drug nodes and protein nodes in the heterogeneous network, so the performance of most methods is not ideal. Nevertheless, the performance of GGANDTI is also better than other advanced methods. The possible reason is that the attention mechanism enhances the importance of neighbors during training, so that the graph generative adversarial model can learn more latent feature information. In the following experiments, the present application verifies this conjecture. In general, from the experimental results, it can be concluded that GGANDTI performs better than some advanced methods in the task of identifying unknown DTIs.

[0196] The ablation experiments and analysis are as follows:

[0197] GGANDTI is composed of multiple parts such as an attention module and a consistency constraint module. In this section, the present application conducts ablation experiments to evaluate the effectiveness of different parts in the framework.

[0198] The impact of the generative adversarial module on the prediction results. To prove that the generative adversarial module helps to improve the performance of the framework, the present application separates GGANDTI and then compares it with the part without the generative adversarial module. The results are reported in Table 3. "Without GAN module" means that the results are obtained only by the features of the variational autoencoder without improving GraphGAN, and "GGANDTI full model" means that the results are obtained based on the complete prediction framework.

[0199] Table 3

[0200]

[0201] The AUROC and AUPR are used by the present application to evaluate the performance of GGANDTI on the Yamanishi and Luo datasets. In the Yamanishi dataset, the method with the generative adversarial module tends to perform better than the method without the generative adversarial module. In the Luo dataset, the generative adversarial module also shows strong distribution fitting ability. This is sufficient to show the important contribution of the generative adversarial module to the DTI prediction task. In addition, even in the Nuclear receptor dataset with a small number of drug-target pairs, the performance of the adversarial model is often better. It can be seen that the generative adversarial module can effectively capture the potential features of the nodes in the heterogeneous network, which is of great help to the mining of unknown DTIs. Overall, the results fully demonstrate the adaptability of the graph generative adversarial module to different datasets in the DTI prediction task.

[0202] The attention mechanism emphasizes the importance of neighboring nodes by assigning different weights to the neighbors of a node to enhance the representation of the node. In the ablation experiment, the present application removes the attention mechanism and only uses the remaining part to verify whether the attention module captures the deep neighbor information of the node and evaluates the contribution of the attention mechanism to the overall model performance. The results are shown in Table 4. "GGANDTI full model" represents the complete prediction framework, and "Without attention" represents the framework without the attention layer.

[0203] Table 4

[0204]

[0205] After removing the attention module, the average AUROC and AUPR values both decreased, and the AUPR further decreased. By focusing on and weighting the neighbor information of the drug or target node, different importance is implicitly assigned to different nodes in the neighborhood. It can be seen that the attention mechanism can improve the framework performance to some extent.

[0206] Because the chemical properties and functions of drugs or targets should remain unchanged regardless of changes in the feature space, there is a hypothesis that drug or target pairs exhibit consistent correlation in different feature spaces. Without introducing consistency constraints, for example, in the Enzyme dataset, 30 feature vectors of drugs or targets are randomly selected in the feature space. Then, the pairwise correlation of them before and after training is calculated, and a heat map is drawn. Here, the premise is that the spatial distribution of different nodes is different. As shown in Figure 5As shown, the pair-wise correlations before (left) and after training (right) are somewhat inconsistent. The darker the color, the stronger the correlation. It can be observed that the correlations between nodes are different. The correlation of some drugs or targets changes relatively small, while others change relatively large. For drugs, their chemical structures and fingerprints are inherent and not easy to change. Since drugs have unobserved targets and unknown indications, the information of proteins and indications interacting with these drugs is likely to be incomplete. The lack of this information can lead to changes in the correlation between drugs in different spaces. Therefore, it is meaningful to maintain the pair-wise correlation of drugs or targets in different spaces.

[0207] To verify the impact of correlation consistency on the performance of the framework, the present application evaluates the performance of the model by setting different contribution parameters. Taking the dataset Enzyme as an example, the present application uses the average AUROC and AUPR to evaluate the performance of the framework. The results are shown in Figure 6 The contribution parameter λ of the constraint term is selected from {0, 10 -1 ,10 -2 ,10 -3 ,10 -4 ,10 -5}. When λ is 0, it means that the result is obtained by the framework without consistency constraints. As can be seen from the bar chart, when λ is selected from {10 -1 ,10 -2}, the performance gradually improves. When λ is selected from {10 -4 ,10 -5}, the performance gradually decreases. Considering that DTI is added as a subpart of the feature matrix when constructing the initial feature matrix, it is speculated that the incomplete feature information in the DTI matrix leads to this problem. The introduction of correlation consistency can achieve a balance on the selection of λ, which helps to reduce the impact of incomplete information on performance. In general, this idea of introducing consistency constraints is reasonable.

[0208] To verify the applicability of GGANDTI to data with different imbalances, the present application removes part of the known associations in the dataset to change the imbalance. In this way, the imbalance level is set to several levels {0, 10%, 30%, 50%, 70%}, where the percentage represents the proportion of removed positive samples to the total positive samples. As shown in Figure 7As shown, most methods have varying degrees of performance loss as imbalance increases. deepDR performs the worst and is the least adaptable to imbalanced data, with the fastest performance drop. DTINet performs averagely on datasets with fewer nodes, possibly because it uses random walks, which do not learn node features well when there are few positive samples. deepDTNet integrates autoencoder embeddings and Positive-Unlabeled (PU) matrix completion, both of which work together to outperform DTINet and deepDR. When the matrix is too sparse, EEGDTI cannot capture enough neighborhood information to infer a high-quality DTI matrix. Moreover, since only the same number of negative samples as positive samples can be used during training, a large number of negative samples that can contain valuable information are discarded, which limits the performance of MolTrans on extremely imbalanced data. On the smallest dataset, Nuclear receptor, all baseline methods do not adapt well to this data. GGANDTI's performance is still noteworthy. Compared to Yamanishi's dataset, Luo's dataset has higher sparsity from the beginning, which can affect the performance of most methods. Overall, GGANDTI has the smallest rate of performance loss. By comparing GGANDTI with several other methods, the present application can observe that GGANDTI performs well on multiple datasets with different balances.

[0209] To demonstrate the ability of GGANDTI in discovering potential DTIs, the present application uses all known DTIs in four datasets to train the model, taking the Yamanishi dataset as an example, and uses it to predict the target proteins of all drugs. After excluding known interactions from the prediction results, several DTI candidates are obtained, and they are sorted in descending order according to their interaction prediction scores. The top 10 DTI candidates with the highest interaction scores are finally selected. Finally, the KEGG public database is queried to obtain evidence to verify this prediction. The top 10 candidates of drug-target pairs are shown in Table 5.

[0210] Table 5

[0211]

[0212]

[0213]

[0214] The data marked with "*" in Table 5 means that the supporting evidence has not been found in the given database.

[0215] Generally, based on the assumption that similar drugs can usually interact with the same protein, relevant evidence can be searched to support the prediction of the present application. For example, in Table 5, the first candidate in the Enzyme dataset is taken; D00448 and D00567 are both anti-inflammatory drugs and are used to treat rheumatoid arthritis. Moreover, there is a known interaction between D00567 and has5742. Therefore, it is inferred that there may be some degree of interaction between DB00448 and hsa5742. In addition, proteins in the same metabolic pathway interact, co-express specific biological functions. Therefore, if a drug can interact with a protein, it can also affect the expression of other proteins in the same pathway as the protein. For example, in the first candidate in the ion channel dataset, D00553 and D00738 have the same protein metabolic pathway (hsa04728, dopaminergic synapse), so there may be potential interaction between them. The present application retrieves relevant evidence from the database to support this prediction. Overall, KEGG can infer 34 of all the candidate DTIs listed in Table 5.

[0216] The present application integrates drug similarity, target similarity and drug-target interaction information, and uses an encoder model, an attention mechanism and a generative adversarial network model for joint training, and defines a regularization to maintain the inherent correlation between drugs and targets. By making full use of known biological information, the interaction between drugs and targets is predicted. Experimental results show that this method can effectively predict the interaction between drugs and targets.

[0217] The above examples are only used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacement for part of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application, and should be included in the protection scope of the present application.

Claims

1. A drug relocation method based on graph generative adversarial networks and variational autoencoders, characterized in that, Includes the following steps: Using known drug similarity matrices, target similarity matrices, and drug-target interaction matrices, a heterogeneous network, adjacency matrix, and initial feature matrix of drugs and targets are constructed. A variational autoencoder is used to encode the initial features in the adjacency matrix and initial feature matrix of the drug-target into latent variables by utilizing the feature extraction and fusion functions of a convolutional neural network. Based on a drug-target heterogeneous network, an adversarial model is used to enhance the feature representation of drugs and targets in latent variables. The adversarial model is a generative adversarial network with an attention mechanism, which includes a generator and a discriminator. After a preset number of iterations of the adversarial model, the latent variables in the generator are extracted to construct the prediction matrix; Based on the prediction matrix, predict unknown drug-target interactions; The steps involved in implementing the generator and discriminator include: The latent variable Z is used in the generator and discriminator respectively. and express, and This represents the eigenvectors of its i-th and j-th rows; The generator is used to implement the following steps: drug and target Attention score is calculated as follows: in, and These are the weights and bias parameters of the attention layer in the generator. It is a drug The set of neighbors; node The new representation is: Among them, nodes It is a drug or target; The samples provided by the generator to the discriminator are obtained by sampling from a pre-constructed breadth-first search tree. This represents the node currently found in the tree. Let represent the next node to be searched. The probability of the next hop in the sampling path is calculated using the following formula: in, and They are middle and The expression, It is the first The set of all neighbors of a node; The discriminator is used to implement the following steps: No. The formula corresponding to the new features of each node is: in, , and These are the trainable weights and bias parameters of the attention layer in the discriminator; The discriminator simultaneously receives drug-target node pairs sampled from a heterogeneous network and node pairs from the generator, and uses a classifier to measure the probability of interaction between the input node pairs. The formula corresponding to this probability is: Will Feedback is given to the generator, participating in the generator's gradient update, so that the generator updates its latent variables; Before training the adversarial model, the correlation between the features of the drug and the target was predefined, and the consistency of the correlation between the features of the drug and the target was maintained by regularization. The correlation formula before training is: in, It is the correlation before training; The correlation formula after training is: in, It is the correlation after training. and The representations of are respectively from the generator ; minimize To achieve the optimization goal, the corresponding formula is: in, It is a parameter that adjusts the contribution of consistency to the overall loss. These are the trainable weight parameters in the attention mechanism. These are weighting factors that control the impact of attention weights on the model. It is the total number of samples; The generator and discriminator iterate in an adversarial game manner, extracting the feature representations learned by the generator to construct a prediction matrix in order to predict drug-target interactions.

2. The drug relocation method based on graph generative adversarial networks and variational autoencoders according to claim 1, characterized in that, The process of constructing a heterogeneous drug-target network, adjacency matrix, and initial feature matrix using known drug similarity matrices, target similarity matrices, and drug-target interaction matrices includes: The formula for determining the values ​​in the drug-drug interaction matrix using a known drug similarity matrix is ​​as follows: in, It is the value in the i-th row and j-th column of the drug-drug interaction matrix. This represents the similarity score between the i-th drug and the j-th drug in the drug similarity matrix. and Let represent the mean and variance of all samples in the drug similarity matrix, respectively. The target-target interaction matrix is ​​determined using the target similarity matrix; The drug-drug interaction matrix, target-target interaction matrix, and drug-target interaction matrix are supplemented with matrix completion information to construct the drug-target adjacency matrix; The drug similarity matrix, target similarity matrix, and drug-target interaction matrix are supplemented with matrix completion information to construct the initial feature matrix of drug-target interaction. Based on the interaction relationships in the adjacency matrix of drug-target, a heterogeneous network of drug-target is constructed. When the value of the i-th row and j-th column in the adjacency matrix of drug-target is 1, it indicates that there is an edge between the i-th and j-th nodes in the heterogeneous network. When the value is 0, it is considered that there is no edge.

3. The drug relocation method based on graph generative adversarial networks and variational autoencoders according to claim 1, characterized in that, The training process of the adversarial model includes: The adversarial model includes a generator that constructs a traversal tree using all drug and target nodes in a breadth-first traversal manner, and calculates the probability of selecting the next hop node during the search process using the feature representations of drugs and targets in the latent variables, generating drug and target pairs with interactions. The adversarial model includes a discriminator that receives node pairs sampled from the heterogeneous drug-target network and node pairs generated by the generator, and uses the classifier included in the discriminator to give the probability that there is a connection relationship between the node pairs as features. During the training of the adversarial model, the correlation consistency of drug or target features is maintained by minimizing the correlation error between latent variables and drug or target features in the initial feature matrix, so as to maintain the inherent chemical properties and attributes of drugs and targets unchanged. The adversarial model updates the information between the drug and the target in the latent variables after each iteration in order to enhance the feature representation of the drug and the target in the latent variables.

4. The drug relocation method based on graph generative adversarial networks and variational autoencoders according to claim 1, characterized in that, The prediction of unknown drug-target interactions based on the prediction matrix includes: The candidate drug-target pairs are determined by sorting the filling values ​​in the prediction matrix, thus completing the prediction of unknown drug-target interactions. The filling values ​​in the prediction matrix correspond to the probability of the existence of a drug-target interaction pair.

Images