A method for predicting microorganism and drug relationships based on multi-relational graphs

Abstract

The application discloses a kind of prediction method of microorganism and drug relationship based on multi-association graph, through microorganism-drug association database, the association network of microorganism-drug is constructed, and further obtains interaction network Net1, Net2 and Net3;Establish the graph neural network model of introduction regularization, input graph neural network model with Net1, Net2, Net3 combine the multi-modal attribute graph of microorganism-drug to obtain embedding representation Z1, Z2 and Z3, and embedding representation Z1, Z2 and Z3 are input into graph neural network and trained, to obtain trained graph neural network;Finally, the microorganism-drug association effect in microorganism-drug data set is predicted by trained graph neural network model.The application constructs the node feature of biological and drug that can be explained, and considers the sparsity problem brought by existing microorganism-drug association data set, greatly improves the prediction accuracy of microorganism-drug association effect.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of biological information, and in particular to a method for predicting the relationship between microorganisms and drugs based on a multi-association graph. BACKGROUND

[0002] In recent years, the focus of research in the medical field is to explore the relationship between microbial community imbalance and drug efficacy and toxicity, however, the complex mechanism of the interaction between microbial community and drugs in the human body is still lacking comprehensive understanding. Electronic methods for identifying potential microorganism-drug associations are an effective complement to traditional experimental methods, which not only help to screen candidate compounds for drug development, but also help to discover new knowledge to understand the mechanism of microorganism-drug interaction. The recent increase in the availability of cumulative biomedical data on microorganisms and drugs provides an excellent opportunity for machine learning methods to predict microorganism-drug correlations.

[0003] Currently, there are two types of computational methods for predicting the relationship between microorganisms and drugs.

[0004] The first type of method mainly focuses on similarity measures, such as the HMDAKATZ method using KATZ measures, but such measures are too simple to fully reflect similarity, which can lead to inaccurate association identification.

[0005] The second type of method uses graph-based learning methods, which use rich semantic information in graph data representation, and have better prediction ability compared to previous methods based on similarity measures. There are two common graph representation learning methods: meta-path and graph convolutional network.

[0006] The meta-path algorithm mainly uses edge information of microorganism-drug associations for prediction. The meta-path algorithm combines metapath2vec with neural network recommendations to learn low-dimensional embedding representations of microorganisms and drugs. The meta-path algorithm does help to improve the prediction ability of the model, but it relies too much on edge information, and when new drugs or new microorganisms are introduced, there is naturally no existing edge information to consider, which can lead to prediction failure.

[0007] Compared with meta-path, the GCN method can not only capture edge information, but also capture node information. Therefore, in the current method, the use of the GCN method to predict microorganism-drug correlations has attracted widespread interest. Long et al. first applied a GCN encoder to the method GCNMDA for microorganism-drug associations, and introduced a conditional random field into the GCN hidden layer. Deng et al. used a variational graph autoencoder (VGAE) to construct a multi-modal graph, as the VGAE can generate multiple interpretable node features to predict MDAs, but did not consider using the microorganism-drug-disease association network to construct more connected graphs.

[0008] In summary, the existing method for predicting the correlation between microorganisms and drugs has the problem of being unable to construct rich and interpretable node features of biological and drugs, therefore, how to comprehensively utilize the correlation among microorganisms, drugs and diseases, construct more connected graph networks, and improve the prediction accuracy of the correlation between microorganisms and drugs is a technical problem that needs to be solved in the field. SUMMARY

[0009] The present application aims to overcome the deficiencies of the prior art and provide a prediction method for the correlation between microorganisms and drugs based on multiple association graphs, which can take into account the sparsity problem caused by the existing microorganism-drug correlation dataset.

[0010] To achieve the above-mentioned purpose, the technical solution provided by the present application is:

[0011] A prediction method for the correlation between microorganisms and drugs based on multiple association graphs, comprising:

[0012] Constructing a microorganism-drug association network from a microorganism-drug association database, which is referred to as Net1;

[0013] Retrieving the relevant interactions between microorganisms from a microorganism database in the microorganism-drug association database, and retrieving the relevant interactions between drugs from a drug database in the microorganism-drug association database; constructing an interaction network according to the relevant interactions between microorganisms and the relevant interactions between drugs, which is referred to as Net2;

[0014] Deriving a corresponding microorganism-drug network Net3 based on the meta-path "microorganism-disease-drug";

[0015] Constructing a microorganism-drug multi-modal attribute graph;

[0016] Establishing a graph neural network model with regularization based on Net1, Net2, Net3 and the microorganism-drug multi-modal attribute graph, and training the graph neural network model;

[0017] Predicting the correlation between microorganisms and drugs in the microorganism-drug dataset through the trained graph neural network model.

[0018] Further, the construction of the microorganism-drug multi-modal attribute graph comprises:

[0019] Constructing a similarity feature matrix of drugs and a second attribute feature matrix of drugs based on drug data in the drug database, Net1 and Net2;

[0020] Constructing a similarity feature matrix of microorganisms and a second attribute feature matrix of microorganisms based on microorganism data in the microorganism database, Net1;

[0021] According to the similarity feature matrix of the drug and the similarity feature matrix of the microorganism, a microorganism-drug similarity feature network is constructed;

[0022] According to the second attribute feature matrix of the drug and the second attribute feature matrix of the microorganism, a microorganism-drug second attribute feature network is constructed;

[0023] The microorganism-drug similarity feature network and the microorganism-drug second attribute feature network are combined to obtain a multi-modal attribute graph of the microorganism-drug.

[0024] Further, according to the drug data in the drug database, Net1 and Net2, the similarity feature matrix of the drug and the second attribute feature matrix of the drug are constructed, comprising:

[0025] The molecular structure similarity of the drugs in the drug database is calculated using the SIMCOMP2 tool to obtain a molecular structure similarity matrix DS of the drugs struct (d i ,d j ), which represents the molecular structure similarity between drug d i and drug d j ;

[0026] The drug-drug interaction spectrum in Net2 is represented by matrix DIP to obtain a standardized kernel bandwidth:

[0027]

[0028] Where τ represents the standardized kernel bandwidth, τ' is the original bandwidth, and DIP(d i ) represents the interaction of drug d i with other drugs, and nd represents the number of drugs in Net1;

[0029] The Gaussian kernel similarity DS gauss (d i ,d j ) between drug d i and drug d j is obtained:

[0030] DS gauss (d i , d j ) = exp (-τ||DIP(d i )-DIP(d j )|| 2 )

[0031] DIP(d j ) represents the interaction of drug d j with other drugs;

[0032] The similarity feature matrix of the drug is represented as S d (d i ,d j ):

[0033]

[0034] The drug network of the topological properties of the drug data in the drug database is constructed by the random walk method with restart, random drift and restart are performed on the drug network until the drug network converges, the probability distribution vector of each drug is obtained, and the second attribute feature matrix F of the drug is constructed d ∈R nd×nd .

[0035] Further, the formula of random drift and restart is:

[0036]

[0037] Wherein, represents the probability of the i th node of the drug network moving to other nodes at time t+1, θ is the restart probability, T is the transition probability matrix, represents the initial probability vector of the i th node of the drug network, represents the probability of the i th node of the drug network moving to other nodes at time t.

[0038] Further, according to the microorganism data in the microorganism database, Net1, the similarity feature matrix of the microorganism and the second attribute feature matrix of the microorganism are constructed, including:

[0039] The functional similarity of the microorganism data in the biological database is calculated by using Kamneva tool, and the similarity feature matrix S of the microorganism is obtained m ∈R nm×nm , wherein nm represents the number of microorganisms in Net1; the similarity between microorganism m i and microorganism m j is represented as S m (m i ,m j );

[0040] The original gene sequence of the microorganism data is coded;

[0041] All the coded sequences are padded with zeros, so that all the padded sequences have the same length;

[0042] All the padded sequences are analyzed by using principal component analysis method, and a k-dimensional matrix is obtained, and the second attribute feature matrix of the microorganism is represented as F m ∈Rnm×k .

[0043] Furthermore, based on the drug similarity feature matrix and the microorganism similarity feature matrix, a microorganism-drug similarity feature network X is constructed. simility :

[0044]

[0045] S d This is the similarity feature matrix of drugs;

[0046] A microbial-drug second attribute feature network X is constructed based on the second attribute feature matrices of drugs and microorganisms. secondary :

[0047]

[0048] F d This is the feature matrix of the drug's second attribute;

[0049] Combining the microbial-drug similarity feature network and the microbial-drug second attribute feature network yields the multimodal attribute map X of microbial-drug pairs.

[0050] X = [X] simility ,X secondary ].

[0051] Furthermore, based on Net1, Net2, Net3, and the multimodal attribute graphs of microorganisms and drugs, a regularized graph neural network model is established and trained, including:

[0052] A feature matrix for microorganisms and drugs is established based on the multimodal attribute graph of microorganisms and drugs; Net1, Net2, and Net3 are converted into corresponding adjacency matrices.

[0053] The microbial-drug feature matrix and the adjacency matrices corresponding to Net1, Net2, and Net3 are input into the encoder to obtain the latent variable Z as the encoder output;

[0054] The encoder comprises a two-layer graph convolutional neural network and a two-layer regularized graph convolutional network; the expression for the graph convolutional neural network is:

[0055]

[0056] Where GCN represents Graph Convolutional Neural Network, X is the microbial-drug feature matrix, A is the microbial-drug adjacency matrix, and W... 0 W represents the parameters of the first graph convolutional neural network layer used for training. 1 W represents the parameters of the second graph convolutional neural network layer used for training.0 ∈R m×f W 1 ∈R m×f The learnable matrix is ​​initialized with random numbers and continuously updated during training; ReLU = max(0, ·) is a linear activation function. It is a symmetric normalized adjacency matrix. The expression for a symmetric normalized adjacency matrix is: Where D is the degree matrix of the microbial-drug adjacency matrix A;

[0057] The expression for a regularized graph convolutional network is:

[0058]

[0059] Wherein, GNCN represents Regularized Graph Convolutional Network. I N It is an N-order identity matrix. for The degree matrix is ​​denoted by s, where s is the scaling factor, and g(XW) is used to introduce L2 regularization, with the formula:

[0060]

[0061] in, Set the weight matrix to W∈R m×f The elements of the weight matrix are initialized using random numbers, f is the dimension of the node embedding representation (set by hyperparameters), and n = nd + nm is used as the number of nodes, m = 2 × (nd + nm) is used as the matrix dimension of the nodes, nd is the number of drugs, and nm is the number of microorganisms. And W generate a vector after feature transformation.

[0062] The mean of the normal distribution is calculated using the constructed regularized graph convolutional network, and the variance of the normal distribution is calculated using the graph convolutional neural network. The corresponding normal distribution is constructed using the specific mean and variance of each real sample. Then, the latent variable Z is obtained by sampling the normal distribution using the reparameterization method.

[0063] The formula for calculating the mean of the normal distribution is:

[0064] μ = [μ1, μ2, ..., μ n ] T =GNCN(X,A,s)

[0065] Where μ is the mean of the normal distribution;

[0066] The variance of the normal distribution is obtained by squared standard deviation, and the formula for calculating standard deviation is:

[0067] log sigma = [log sigma1,..., log sigma f ] T = GCN(X, A)

[0068] where sigma is the standard deviation of the normal distribution;

[0069] The reparameterization method is used to sample the hidden variable Z from the normal distribution Q(Z|X, A) i and Z j The equation for calculating the hidden variable Z i is: Z i = mu + sigma * epsilon i The equation for calculating the hidden variable Z j is: Z j = mu + sigma * epsilon j ;

[0070] where epsilon i ~ N(0, 1), epsilon j ~ N(0, 1), the hidden variable Z i is the hidden variable of node i in all microorganism nodes and drug nodes, and the hidden variable Z j is the hidden variable of node j in all microorganism nodes and drug nodes;

[0071] The hidden variable Z is input into the decoder, and the decoder decodes the encoded multi-modal attribute graph in the form of inner product decoding to generate a reconstructed graph for calculating the association probability;

[0072] The decoding process is as follows:

[0073] The decoder takes the inner product between the hidden variables Z i and Z j , reconstructs the original adjacency matrix A using the inner product to obtain the reconstructed graph, and based on the reconstructed graph, represents the association probability P of the given hidden variables p(Aij|Z i , Z j ) that the microorganism and the drug exist association, and completes the association prediction through the association probability;

[0074] The calculation method of the association probability P is as follows:

[0075]

[0076] where phi is a Sigmoid function;

[0077] The loss function is defined, and the loss function is made to tend to converge by the stochastic gradient descent method, to obtain the hidden variables Z1, Z2 and Z3

[0078] The Z1, Z2 and Z3 are input into a DNN classifier of a graph neural network model, the number of training times epoch is set to k2, the training process adopts a stochastic gradient descent, and the training is stopped when a loss function converges, and a trained graph neural network model is obtained.

[0079] Further, the calculation formula of the microbe-drug feature matrix is as follows:

[0080]

[0081] Wherein, X is a microbe-drug feature matrix, is a content feature vector of node i in the multi-modal attribute graph of microbe-drug;

[0082] The specific method of converting the microbe-drug association network into a microbe-drug adjacency matrix is as follows:

[0083] The adjacency matrix A of the microbe-drug association network is obtained. nd×nm If there is a known association between nodes i and j in the microbe-drug association network, the element Aij in the adjacency matrix A is set to 1, and if there is no known association between nodes i and j in the microbe-drug association network, the element Aij in the adjacency matrix A is set to 0; thereby obtaining the microbe-drug adjacency matrix A.

[0084] Further, after obtaining the trained graph neural network model, the accuracy of the graph neural network model is verified, and the specific steps of the verification are as follows:

[0085] The k-fold cross-validation framework is introduced, and under the k-fold cross-validation framework, all known microbe-drug association data on the existing microbe-drug association database is randomly divided into k1 groups, and each group in the k1 groups is randomly sampled to obtain a subset of unknown association pairs with the same size batch as the test set, and the remaining known association pairs are selected as the training set.

[0086] The test set is input into the trained graph neural network model to obtain a classification result.

[0087] If the classification result is positive, it is predicted that the microbe and the drug have an association, and if the classification result is negative, it is predicted that the microbe and the drug have no association.

[0088] According to the classification result, the AUC value of the trained graph neural network model is obtained; according to the AUC value, the accuracy of the graph neural network model is verified.

[0089] Further, the AUC value of the trained graph neural network model is obtained according to the classification result, including:

[0090] The training set is input into the model to obtain a reconstruction graph of the training set by the current model, and the scores of the edges between the nodes in the reconstruction graph of the training set by the current model are taken as the association probabilities, which are valued between 0 and 1;

[0091] The association probability is used as a classification threshold, and when other association probabilities are greater than the classification threshold, it is considered to be predicted as a positive sample, and when other association probabilities are less than the classification threshold, it is considered to be predicted as a negative sample.

[0092] According to the association relationship between microorganisms and drugs in the training set, the label true value of the edge in the training set is obtained, and the label true value is valued as 0 or 1, wherein 0 represents that there is no edge, that is, there is no association relationship, and actually it is a negative sample, and 1 represents that there is an edge, that is, there is an association relationship, and actually it is a positive sample;

[0093] The true positive rate and the false positive rate under each classification threshold are counted:

[0094]

[0095] Wherein, TPRate is the true positive rate, FPRate is the false positive rate, TP is the true positive rate, which represents the number of samples that are actually negative samples and are predicted as positive samples, FN is the false negative rate, which represents the number of samples that are actually positive samples and are predicted as negative samples, FP is the false positive rate, which represents the number of samples that are actually negative samples and are predicted as positive samples, and TN is the true negative rate, which represents the number of samples that are actually negative samples and are predicted as negative samples;

[0096] The ROC curve is drawn with FPRate as the horizontal axis and TPRate as the vertical axis, and the area of the ROC curve is calculated using the microelement method, that is, the AUC value.

[0097] Compared with the prior art, the principles and advantages of the present scheme are as follows:

[0098] The present scheme constructs the association network of microorganisms and drugs through the microorganism-drug association database, and further obtains the interaction networks Net1, Net2 and Net3; a graph neural network model with regularization is established, the Net1, Net2, Net3 and the multi-modal attribute graph of microorganisms and drugs are input into the graph neural network model to obtain the embedding representations Z1, Z2 and Z3, and the embedding representations Z1, Z2 and Z3 are input into the graph neural network for training to obtain a trained graph neural network; finally, the trained graph neural network model is used to predict the microorganism-drug association in the microorganism-drug data set. The present scheme constructs the interpretable node features of biological and drug, and considers the sparsity problem brought by the existing microorganism-drug association data set, which greatly improves the prediction accuracy of the microorganism-drug association. BRIEF DESCRIPTION OF DRAWINGS

[0099] In order to more clearly illustrate the technical solutions of the embodiments of the present application or the prior art, the services required in the embodiments or the prior art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0100] Figure 1 A principle flow chart of a microorganism and drug relationship prediction method based on a multi-association graph according to the present application;

[0101] Figure 2 A flow chart of constructing a multi-modal attribute graph in a microorganism and drug relationship prediction method based on a multi-association graph according to the present application;

[0102] Figure 3 A flow chart of obtaining an association probability in a microorganism and drug relationship prediction method based on a multi-association graph according to the present application. DETAILED DESCRIPTION

[0103] The present application will be further described below in combination with specific embodiments:

[0104] As shown in the drawings, the microorganism and drug relationship prediction method based on a multi-association graph according to the present embodiment comprises the following steps: Figure 1 S1, constructing a microorganism-drug association network by a microorganism-drug association database (MDAD), and the association network is referred to as Net1;

[0105] S2, retrieving microorganism-microorganism related interactions from a microorganism database in the microorganism-drug association database, and retrieving drug-drug related interactions from a drug database in the microorganism-drug association database; constructing an interaction network according to the microorganism-microorganism related interactions and the drug-drug related interactions, and the interaction network is referred to as Net2;

[0106]

[0107] ​S3, Microbial disease associations were collected from HMDAD database (http: / / www.cuilab.cn / hmdad / ) and Disbiome (https: / / disbiome.ugent.be / home). Here, the example derived associations involving microbes contained in MDAD (MDAD is a sub-database of HMDAD) (Sun et al., 2018). As a result, 59 microbe-disease associations between 29 microbes and 20 diseases were downloaded. Drug-disease associations were downloaded from CTD database (http: / / ctdbase.org / ). The example selected entries containing drugs in MDAD. Finally, 168 drug-disease associations between 82 drugs and 21 diseases were obtained. The corresponding microbe-drug network Net3 was inferred based on the meta-path “microbe-disease-drug”.

[0108] S4, As shown in Figure 2 , a microbe-drug multimodal attributed graph was constructed, and the specific process was as follows:

[0109] S4-1, according to the drug data in the drug database, Net1 and Net2, the similarity feature matrix of the drug and the second attribute feature matrix of the drug were constructed:

[0110] The SIMCOMP2 tool was used to calculate the molecular structure similarity of the drugs in the drug database, and the molecular structure similarity matrix DS of the drugs was obtained struct (d i ,d j ), which represents the molecular structure similarity between drug d i and drug d j ;

[0111] The matrix DIP represents the drug-drug interaction spectrum in Net2, and the standardized kernel bandwidth is obtained:

[0112]

[0113] Where τ represents the standardized kernel bandwidth, τ' is the original bandwidth, and DIP(d i ) represents the interaction of drug d i with other drugs, and nd represents the number of drugs in Net1;

[0114] The Gaussian kernel similarity DS i (d j ,d gauss ) between drug d i and drug d j was obtained:

[0115] DS gauss (di , d j )=exp(-τ||DIP(d i )-DIP(d j )|| 2 )

[0116] DIP(d j ) represents the interaction of drug d j with other drugs;

[0117] The similarity feature matrix of the drug is represented as S d (d i ,d j ):

[0118]

[0119] The drug network of the topological properties of the drug data in the drug database is constructed by the random walk method with restart, random drift and restart are performed on the drug network until the drug network converges, the probability distribution vector of each drug is obtained, and the second attribute feature matrix F d ∈R nd×nd of the drug is constructed.

[0120] Specifically, the formula of random drift and restart is:

[0121]

[0122] wherein, represents the probability of the i th node of the drug network moving to other nodes at time t+1, θ is the restart probability, T is the transition probability matrix, represents the initial probability vector of the i th node of the drug network, represents the probability of the i th node of the drug network moving to other nodes at time t.

[0123] S4-2, according to the microbial data in the microbial database, Net1, the similarity feature matrix of the microorganism and the second attribute feature matrix of the microorganism are constructed, and the specific process is as follows:

[0124] S4-2-1, the functional similarity of the microbial data in the biological database is calculated by using Kamneva tool, and the similarity feature matrix S m ∈R nm×nm of the microorganism is obtained, wherein nm represents the number of microorganisms in Net1; the similarity between microorganism m i and microorganism m j is represented as S m (m i ,m j );

[0125] S4-2-2, encode the original gene sequence of the microorganism data;

[0126] S4-2-3, pad all the encoded sequences with zeros to make all the padded sequences have the same length;

[0127] S4-2-4, analyze all the padded sequences using principal component analysis to obtain a k-dimensional matrix, and represent the second attribute feature matrix of the microorganism as F m ∈R nm×k .

[0128] S4-3, construct a microorganism-drug similarity feature network X simility :

[0129]

[0130] S d is the similarity feature matrix of the drug;

[0131] S4-4, construct a microorganism-drug second attribute feature network X secondary :

[0132]

[0133] F d is the second attribute feature matrix of the drug;

[0134] S4-5, combine the microorganism-drug similarity feature network and the microorganism-drug second attribute feature network to obtain a multi-modal attribute graph X of the microorganism-drug:

[0135] X=[X simility ,X secondary ]。

[0136] S5, according to Net1, Net2, Net3 and the multi-modal attribute graph of the microorganism-drug, establish a graph neural network model with regularization, and train the graph neural network model;

[0137] This step specifically includes:

[0138] According to the multi-modal attribute graph of the microorganism-drug, establish a feature matrix of the microorganism-drug; convert Net1, Net2, Net3 into corresponding adjacency matrices;

[0139] Input the microorganism-drug feature matrix and the adjacency matrices corresponding to Net1, Net2, Net3 into the encoder to obtain hidden variables Z as the encoder output;

[0140] The encoder comprises two layers of graph convolutional neural networks and two layers of regularized graph convolutional networks; the expression of the graph convolutional neural network is:

[0141]

[0142] wherein GCN represents the graph convolutional neural network, X is a microbe-drug feature matrix, A is a microbe-drug adjacency matrix, W 0 represents parameters of the first graph convolutional neural network layer for training, W 1 represents parameters of the second graph convolutional neural network layer for training, W 0 ∈R m×f , W 1 ∈R m×f and a random number is used to assign an initial value to the element of the learnable matrix, which is constantly updated in the process of training; ReLU = max(0, ·) is a linear activation function, is a symmetric normalized adjacency matrix, and the expression of the symmetric normalized adjacency matrix is: wherein D is a degree matrix of the microbe-drug adjacency matrix A;

[0143] The expression of the regularized graph convolutional network is:

[0144]

[0145] wherein GNCN represents the regularized graph convolutional network, I N is an N-order identity matrix, is a degree matrix of , s is a scaling coefficient, g(XW) is used to introduce L2 regularization, and the formula is:

[0146]

[0147] wherein the weight matrix is set as W∈R m×f and a random number is used to assign an initial value to the element of the weight matrix, f is the dimension of the node embedding representation, which is set by a hyperparameter, then n = nd+nm is taken as the number of nodes, m = 2 × (nd+nm) is taken as the matrix dimension of the nodes, nd is the number of drugs, nm is the number of microbes, and according to and W, a feature-transformed vector

[0148] The mean of the normal distribution is calculated using the constructed regularized graph convolutional network, and the variance of the normal distribution is calculated using the graph convolutional neural network. The corresponding normal distribution is constructed using the specific mean and variance of each real sample. Then, the latent variable Z is obtained by sampling the normal distribution using the reparameterization method.

[0149] The formula for calculating the mean of the normal distribution is:

[0150] μ = [μ1, μ2, ..., μ n ] T =GNCN(X,A,s)

[0151] Where μ is the mean of the normal distribution;

[0152] The variance of the normal distribution is obtained by squared standard deviation, and the formula for calculating standard deviation is:

[0153] logσ = [log σ1, ..., logσ] f ] T =GCN(X,A)

[0154] Where σ is the standard deviation of the normal distribution;

[0155] The latent variable Z is sampled from the normal distribution Q(Z|X, A) using a reparameterization method. i and Z j Calculate the latent variable Z i The equation is: Z i =μ+σ*ε i ; Calculate the latent variable Z j The equation is: Z j =μ+σ*ε j ;

[0156] Where ε i ~N(0,1), ε j ~N(0,1), latent variable Z i Let Z be the latent variable of node i in all microbial and drug nodes. j For all microbial nodes and drug nodes, there is a hidden variable for node j;

[0157] The latent variable Z is input into the decoder, which uses inner product decoding to decode the encoded multimodal attribute map and generate a reconstructed map for calculating the association probability.

[0158] The decoding process is as follows:

[0159] Decoder retrieves hidden variable Z i and Z jinner product between them, the original adjacency matrix A is reconstructed by using the inner product to obtain a reconstructed graph, and the given hidden variable p(Aij|Z i ,Z j The association probability P between the microorganism and the drug is calculated based on the reconstructed graph, and the association prediction is completed by the association probability.

[0160] The calculation method of the association probability P is:

[0161]

[0162] Wherein, φ is a Sigmoid function.

[0163] The loss function is defined, and the loss function is made to converge by the stochastic gradient descent method, and the hidden variables Z1, Z2 and Z3

[0164] Z1, Z2, Z3 are input into the DNN classifier of the graph neural network model, the training times epoch are set to k2, the training process adopts the stochastic gradient descent, and the training is stopped when the loss function converges, and the trained graph neural network model is obtained.

[0165] S6, the microorganism-drug association in the microorganism-drug data set is predicted by the trained graph neural network model. Before prediction, the trained graph neural network model needs to be verified, and the specific verification steps are as follows:

[0166] The k-fold cross validation framework is introduced, and under the k-fold cross validation framework, all known microorganism-drug association data on the existing microorganism-drug association database is randomly divided into k1 groups, and each group is randomly sampled with the same size batch of unknown association pairs as a test set, and the remaining known association pairs are selected as a training set.

[0167] The test set is input into the trained graph neural network model to obtain the classification result.

[0168] If the classification result is positive, it is predicted that the microorganism and the drug have association, and if the classification result is negative, it is predicted that the microorganism and the drug have no association.

[0169] According to the classification result, the AUC value of the trained graph neural network model is obtained; according to the AUC value, the accuracy of the graph neural network model is verified.

[0170] According to the classification result, the AUC value of the trained graph neural network model is obtained, including:

[0171] The training set is input into the model to obtain the reconstructed graph of the current model on the training set, and the score of the edge between each node in the reconstructed graph of the current model on the training set is taken as the association probability, and the association probability is valued between 0 and 1.

[0172] The association probability is taken as a classification threshold, when other association probabilities are greater than the classification threshold, it is considered to be predicted as a positive sample, and when other association probabilities are less than the classification threshold, it is considered to be predicted as a negative sample.

[0173] According to the association relationship between microorganisms and drugs in the training set, the label true value of the edge in the training set is obtained, and the label true value is 0 or 1, wherein 0 represents that there is no edge, that is, there is no association relationship, and actually it is a negative sample, and 1 represents that there is an edge, that is, there is an association relationship, and actually it is a positive sample;

[0174] The true positive rate and the false positive rate under each classification threshold are counted:

[0175]

[0176] Wherein, TPRate is the true positive rate, FPRate is the false positive rate, TP is the true positive rate, which represents the number of samples that are actually negative samples and are predicted to be positive samples, FN is the false negative rate, which represents the number of samples that are actually positive samples and are predicted to be negative samples, FP is the false positive rate, which represents the number of samples that are actually negative samples and are predicted to be positive samples, and TN is the true negative rate, which represents the number of samples that are actually negative samples and are predicted to be negative samples.

[0177] The ROC curve is drawn with FPRate as the horizontal axis and TPRate as the vertical axis, and the area of the ROC curve is calculated using the microelement method, that is, the AUC value.

[0178] In order to verify the effectiveness of the method for predicting the relationship between microorganisms and drugs based on the multi-association graph, the method and five existing methods are run on the MDAD data set using the default parameter settings, and the AUC value is used as the performance evaluation index. The larger the AUC value, the higher the accuracy of the method.

[0179] In this embodiment, 5-fold cross-validation and 10-fold cross-validation are performed on all methods including the present application. The experimentally verified drug combination is randomly divided into 5 or 10 subsets of the same size, each of which is used as a test set in turn, and the remaining is used for training the model. In order to eliminate the bias of random sampling, the process is repeated 10 times, and the final AUC score is calculated according to the average of the AUC values in the 10 repeated verifications, and the final AUC score is used as the performance index to evaluate the accuracy of each method.

[0180] The verification results are shown in the following table. In 5-fold cross-validation, the method adopted by the present application is represented by G3GNAEMDA, and the final AUC score of the present application is the highest among all methods. In 10-fold cross-validation, the final AUC score of the present application is also the highest among all methods. Therefore, the accuracy of the present application is better than that of the existing methods.

[0181] Validation fold method GAE NTSHMDA LAGCN EGATMDA G3GNAEMDA 5-fold cross-validation 0.852 0.870 0.940 0.951 0.963 10-fold cross-validation 0.866 0.885 0.942 0.963 0.971

[0182] The above-described embodiments are only preferred embodiments of the present application, and are not intended to limit the scope of the present application. Any changes made in the shape, principle, and the like of the present application should be covered by the scope of the present application.

Claims

1. A method for predicting the relationship between microorganisms and drugs based on multi-association graphs, characterized in that, include: A microbial-drug association network was constructed using a microbial-drug association database, and this network is referred to as Net1. Microbial-microbial interactions were retrieved from the microbial database within the microbial-drug association database, and drug-drug interactions were retrieved from the drug database within the same database. An interaction network, termed Net2, was constructed based on these microbial-microbial and drug-drug interactions. Based on the meta-path "microbe-disease-drug", the corresponding microbe-drug network Net3 is derived; Constructing a multimodal attribute graph of microorganisms and drugs; Based on Net1, Net2, Net3 and the multimodal attribute graph of microorganism-drug, a graph neural network model with regularization is established and trained. Predict microbial-drug associations in a microbial-drug dataset using a trained graph neural network model; Constructing a multimodal attribute graph of microorganisms and drugs, including: Based on drug data in the drug database, Net1, and Net2, construct a drug similarity feature matrix and a drug second attribute feature matrix; Based on microbial data from the microbial database and Net1, construct a similarity feature matrix and a second attribute feature matrix for microorganisms. Based on the similarity feature matrix of drugs and the similarity feature matrix of microorganisms, a microorganism-drug similarity feature network is constructed. Construct a microbial-drug second attribute feature network based on the second attribute feature matrix of drugs and the second attribute feature matrix of microorganisms; The microbial-drug similarity feature network and the microbial-drug second attribute feature network are combined to obtain a multimodal attribute map of microbial-drug; Based on drug data from the drug database, Net1, and Net2, construct a drug similarity feature matrix and a drug second attribute feature matrix, including: Using the SIMCOMP2 tool, the molecular structure similarity of drugs in a drug database is calculated, resulting in the molecular structure similarity matrix DS. struct (d i ,d j ), which represents drug d i and drug d j Molecular structural similarity between them; Using matrix DIP to represent the drug-drug interaction spectrum in Net2, the normalized kernel bandwidth is obtained: Where τ represents the normalized kernel bandwidth, τ' is the raw bandwidth, and DIP(d i ) indicates drug d i Interactions with other drugs, where nd represents the amount of drug in Net1; Find the drug d i and drug d j Gaussian kernel similarity DS between gauss (d i ,d j ): DS gauss (d i ,d j )=exp(-τ||DIP(d i )-DIP(d j )|| 2 ) DIP(d j ) indicates drug d j Interactions with other drugs; The drug similarity feature matrix is ​​represented as S d (d i ,d j ): A drug network representing the topological attributes of drug data in a drug database is constructed using a random walk method with restarts. Random drifts and restarts are performed on the drug network until it converges, yielding the probability distribution vector for each drug. The second attribute feature matrix F of the drugs is then constructed. d ∈R nd×nd ; Based on microbial data from the microbial database and Net1, a similarity feature matrix and a second attribute feature matrix of microorganisms are constructed, including: The Kamneva tool was used to calculate the functional similarity of microbial data in a biological database, resulting in the microbial similarity feature matrix S. m ∈R nm×nm Where nm represents the number of microorganisms in Net1; and m represents the number of microorganisms. i and microorganisms m j The similarity between them is represented by S. m (m i ,m j ); Encode the raw gene sequences of microbial data; All encoded sequences are padded with zeros to make all padded sequences the same length; Principal component analysis was used to analyze all the filled sequences to obtain a k-dimensional matrix. The secondary attribute feature matrix of the microorganisms was then represented by the k-dimensional matrix F. m ∈R nm×k .

2. The method for predicting the relationship between microorganisms and drugs based on multiple association graphs according to claim 1, characterized in that, The formulas for random drift and restart are: in, Let θ represent the probability that the i-th node in the drug network moves to another node at time t+1, where θ is the restart probability and T is the transition probability matrix. Let represent the initial probability vector of the i-th node in the drug network. Let represent the probability that the i-th node in the drug network moves to another node at time t.

3. The method for predicting the relationship between microorganisms and drugs based on multiple association graphs according to claim 1, characterized in that, Based on the drug similarity feature matrix and the microorganism similarity feature matrix, a microorganism-drug similarity feature network X is constructed. simility : S d This is the similarity feature matrix of drugs; A microbial-drug second attribute feature network X is constructed based on the second attribute feature matrices of drugs and microorganisms. secondary : F d This is the feature matrix of the drug's second attribute; Combining the microbial-drug similarity feature network and the microbial-drug second attribute feature network yields the multimodal attribute map X of microbial-drug pairs. X=[X simility ,X secondary ]。 4. The method for predicting the relationship between microorganisms and drugs based on multiple association graphs according to claim 3, characterized in that, Based on Net1, Net2, Net3, and the multimodal attribute graphs of microorganisms and drugs, a regularized graph neural network model is established and trained, including: A feature matrix for microorganisms and drugs is established based on the multimodal attribute graph of microorganisms and drugs; Net1, Net2, and Net3 are converted into corresponding adjacency matrices. The microbial-drug feature matrix and the adjacency matrices corresponding to Net1, Net2, and Net3 are input into the encoder to obtain the latent variable Z as the encoder output; The encoder comprises a two-layer graph convolutional neural network and a two-layer regularized graph convolutional network; the expression for the graph convolutional neural network is: Where GCN represents Graph Convolutional Neural Network, X is the microbial-drug feature matrix, A is the microbial-drug adjacency matrix, and W... 0 W represents the parameters of the first graph convolutional neural network layer used for training. 1 W represents the parameters of the second graph convolutional neural network layer used for training. 0 ∈R m×f W 1 ∈R m×f The learnable matrix is ​​initialized with random numbers and continuously updated during training; ReLU = max(0, ·) is a linear activation function. It is a symmetric normalized adjacency matrix. The expression for a symmetric normalized adjacency matrix is: Where D is the degree matrix of the microbial-drug adjacency matrix A; The expression for a regularized graph convolutional network is: Wherein, GNCN represents Regularized Graph Convolutional Network. I N It is an N-order identity matrix. for The degree matrix is ​​denoted by s, where s is the scaling factor, and g(XW) is used to introduce L2 regularization, with the formula: in, Set the weight matrix to W∈R m×f The elements of the weight matrix are initialized using random numbers, f is the dimension of the node embedding representation (set by hyperparameters), and n = nd + nm is used as the number of nodes, m = 2 × (nd + nm) is used as the matrix dimension of the nodes, nd is the number of drugs, and nm is the number of microorganisms. And W generate a vector after feature transformation. The mean of the normal distribution is calculated using the constructed regularized graph convolutional network, and the variance of the normal distribution is calculated using the graph convolutional neural network. The corresponding normal distribution is constructed using the specific mean and variance of each real sample. Then, the latent variable Z is obtained by sampling the normal distribution using the reparameterization method. The formula for calculating the mean of the normal distribution is: μ=[μ1,μ2,…,μ n ] T =GNCN(X,A,s) Where μ is the mean of the normal distribution; The variance of the normal distribution is obtained by squared standard deviation, and the formula for calculating standard deviation is: logσ=[logσ1,…,logσ f ] T =GCN(X,A) Where σ is the standard deviation of the normal distribution; The latent variable Z is sampled from the normal distribution Q(Z|X, A) using a reparameterization method. i and Z j Calculate the latent variable Z i The equation is: Z i =μ+σ*ε i ; Calculate the latent variable Z j The equation is: Z j =μ+σ*ε j ; Where ε i ~N(0,1), ε j ~N(0,1), latent variable Z i Let Z be the latent variable of node i in all microbial and drug nodes. j For all microbial nodes and drug nodes, there is a hidden variable for node j; The latent variable Z is input into the decoder, which uses inner product decoding to decode the encoded multimodal attribute map and generate a reconstructed map for calculating the association probability. The decoding process is as follows: Decoder retrieves hidden variable Z i and Z j The inner product between the two is used to reconstruct the original adjacency matrix A, thus obtaining the reconstructed graph. Based on the reconstructed graph, a given latent variable p(Aij|Z) is represented. i Z j The probability P of association between microorganisms and drugs is used to predict the association. The correlation probability P is calculated as follows: Where φ is a sigmoid function; Define a loss function and use stochastic gradient descent to make the loss function converge, obtaining the latent variables Z1, Z2, and Z3. Input Z1, Z2, and Z3 into the DNN classifier of the graph neural network model, set the number of training epochs to k2, and use stochastic gradient descent during training. Stop training when the loss function converges to obtain the trained graph neural network model.

5. The method for predicting the relationship between microorganisms and drugs based on multiple association graphs according to claim 4, characterized in that, The formula for calculating the characteristic matrix of microorganisms-drugs is as follows: Where X is the microbial-drug characteristic matrix, It is the content feature vector of node i in the multimodal attribute graph of microorganism-drug; The specific method for transforming the microbial-drug association network into a microbial-drug adjacency matrix is ​​as follows: Obtain the adjacency matrix A∈R of the microorganism-drug association network. nd×nm If there is a known association between nodes i and j in the microbial-drug association network, then element Aij in adjacency matrix A is set to 1; if there is no known association between nodes i and j in the microbial-drug association network, then element Aij in adjacency matrix A is set to 0; thus, the microbial-drug adjacency matrix A is obtained.

6. The method for predicting the relationship between microorganisms and drugs based on multiple association graphs according to claim 5, characterized in that, After obtaining the trained graph neural network model, the accuracy of the graph neural network model is verified. The specific steps for verification are as follows: A k-fold cross-validation framework is introduced. Under the k-fold cross-validation framework, all known microbial-drug association data in the existing microbial-drug association database are randomly divided into k1 groups. A subset of randomly sampled unknown association pairs with the same batch size in each of the k1 groups is selected as the test set, and the remaining known association pairs are selected as the training set. The test set is input into the trained graph neural network model to obtain the classification results; If the classification result is positive, it is predicted that the microorganism and the drug are associated; if the classification result is negative, it is predicted that the microorganism and the drug are not associated. The AUC value of the trained graph neural network model is obtained based on the classification results; the accuracy of the graph neural network model is verified based on the AUC value.

7. The method for predicting the relationship between microorganisms and drugs based on multiple association graphs according to claim 6, characterized in that, The AUC value of the trained graph neural network model is obtained based on the classification results, including: Input the training set into the model to obtain the reconstruction graph of the training set by the current model. Take the score of the edge between each node in the reconstruction graph of the training set by the current model, and record it as the association probability. The association probability takes a value between 0 and 1. The association probability is used as a classification threshold. When other association probabilities are greater than the classification threshold, the sample is considered to be predicted as a positive sample. When other association probabilities are less than the classification threshold, the sample is considered to be predicted as a negative sample. Based on the association between microorganisms and drugs in the training set, the true values ​​of the edge labels in the training set are obtained. The true values ​​of the labels are 0 or 1, where 0 indicates that there is no edge and no association, which means that the sample is actually negative, and 1 indicates that there is an edge and an association, which means that the sample is actually positive. Calculate the true positive rate and false positive rate at each classification threshold: Wherein, TPRate is the true positive rate, FPRate is the false positive rate, TP is the true positive rate, which represents the number of samples that are actually negative but are predicted as positive, FN is the false negative rate, which represents the number of samples that are actually positive but are predicted as negative, FP is the false positive rate, which represents the number of samples that are actually negative but are predicted as positive, and TN is the true negative rate, which represents the number of samples that are actually negative but are predicted as negative. Plot the ROC curve with FPRate on the horizontal axis and TPRate on the vertical axis, and calculate the area under the ROC curve, i.e., the AUC value, using the method of infinitesimal elements.

Images