A microbiome-drug association prediction method based on normalized graph neural network
By using normalized graph neural network and multimodal attribute map in microbial-drug association prediction, combined with L2 regularization, the problem of insufficient prediction accuracy of isolated nodes is solved, and the accuracy of microbial-drug association prediction is significantly improved.
Patent Information
- Application Number
- CN202210690052.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-06-17
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2042-06-17
AI Technical Summary
The prior art is difficult to effectively predict complex interactions between microorganisms and drugs, especially in isolated nodes, resulting in insufficient prediction accuracy.
The microbial-drug association prediction method based on normalized graph neural network is adopted to construct multimodal attribute graphs of microbial-drugs and introduce L2 regularized graph neural network model to prevent model overfitting and improve prediction accuracy.
It effectively avoids the zero trend of isolated nodes, improves the accuracy and model performance of microbial-drug association prediction, and is better than the existing technology.
Smart Images

Figure CN115148303B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of bioinformatics, and more specifically, to a microorganism-drug association prediction method based on a normalized graph neural network. Background Art
[0002] In recent years, research in the medical field has focused on exploring the relationship between microbial community imbalance and drug efficacy and toxicity. However, there is still a lack of comprehensive understanding of the complex mechanisms of interaction between microbial communities and drugs in the human body. At present, new drug development faces two major challenges. On the one hand, there are few types of antibiotics to be discovered, and most work focuses on optimizing or combining known compounds. It is difficult to culture target bacteria under laboratory conditions, and most drugs fail during experiments. On the other hand, the number of drug-resistant bacteria is also growing at an alarming rate. More and more studies have shown that microorganisms and drugs have close interactions, and some connections between microorganisms and drugs have been confirmed by culture experiments, but they are not enough to elucidate the complex interaction mechanisms between human microorganisms and drugs. Therefore, there is an urgent need to develop an effective method to systematically explore possible connections between microorganisms and drugs.
[0003] There are currently two main types of computational methods used to predict the relationship between microorganisms and drugs.
[0004] The first type of methods mainly focuses on similarity measurement. For example, in [1] L. Zhu, G. Duan, C. Yan and J. Wang, "Prediction of Microbe-Drug Associations Based on KATZ Measure," 2019 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), 2019, pp. 183-187, doi: 10.1109 / BIBM47256.2019.8983209., Zhu et al. proposed the HMDAKATZ method using the KATZ metric. However, this type of metric is too simple and cannot fully reflect the similarity, which will lead to inaccurate association identification.
[0005] The second category of methods uses graph representation learning methods, which use rich semantic information in graph data representation and have better predictive ability than previous similarity metric-based methods. There are currently two common graph representation learning methods: meta-path and graph convolutional network (GCN).
[0006] The metapath algorithm mainly uses the edge information of microbial drug association for prediction. For example, in [2] Y. Long and J. Luo, "Association Mining to Identify Microbe Drug Interactions Based on Heterogeneous Network Embedding Representation," in IEEE Journal of Biomedical and Health Informatics, vol. 25, no. 1, pp. 266-275, Jan. 2021, doi: 10.1109 / JBHI.2020.2998906. Long et al. proposed a new heterogeneous network embedding representation framework HNERMDA in microbial drug association prediction, which combines metapath2vec with neural network proposals to learn low-dimensional embedding representations of microorganisms and drugs. The metapath algorithm does help improve the predictive ability of the model, but it relies too much on edge information. When introducing new drugs or new microorganisms, considering that there is no existing edge information, it will naturally lead to prediction failure.
[0007] Compared with meta-paths, GCN methods can capture not only edge information but also node information. Therefore, among the current methods, the use of GCN methods to predict microbe-drug associations has attracted widespread interest. In [3] Yahui Long, MinWu, Yong Liu, Chee Keong Kwoh, Jiawei Luo, Xiaoli Li, Ensembling graph attention networks for human microbe–drug association prediction, Bioinformatics, Volume 36, Issue Supplement_2, December 2020, Pages i779–i786,. Long et al. first applied the GCN encoder to the microbe-drug association method GCNMDA and introduced the conditional random field (CRF) into the GCN hidden layer. Subsequently, Long et al. proposed a node-level attention GCN method EGATMDA to learn the embedding of nodes (i.e., microbes and drugs), which can effectively preserve the target neighbors of the graph and only retain relevant information. However, the work of Long et al. failed to construct node features containing biological information (e.g., drug structure similarity and microbial sequence similarity). Therefore, [4] Deng L, Huang Y, Liu X, Liu H. Graph2MDA: a multi-modal variational graph embedding model for predicting microbe-drug associations. Bioinformatics. 2021 Nov 23: btab792. doi: 10.1093 / bioinformatics / btab792. Epub ahead of print. PMID: 34864873. LEI et al. constructed a multimodal node attribute graph and used VGAE to extract node representations, which were then used for Graph2MDA of the DNN classifier. However, this method requires the complete extraction of the connected graph and cannot be used for isolated nodes. Similarly, GAE based on GCN encoder has not been directly used for microbial drug association prediction, and it is difficult for the graph embedding model based on GCN encoder to effectively utilize the information of isolated nodes without connection information observed in our experiments. Summary of the invention
[0008] In order to overcome the problem that the above-mentioned prior art cannot be used for isolated nodes, the present invention provides a microorganism-drug association prediction method based on a normalized graph neural network, introduces regularization into the neural network model to prevent the model from overfitting, thereby solving the isolated node problem and improving the accuracy of microorganism-drug association prediction.
[0009] In order to solve the above technical problems, the technical solution of the present invention is as follows:
[0010] A method for predicting microbial-drug association based on a normalized graph neural network comprises the following steps:
[0011] S1: Use the existing microbiome-drug association database to construct a microbiome-drug association network;
[0012] S2: Constructing a multimodal attribute graph of microbes and drugs;
[0013] S3: Based on the microbial-drug association network and the microbial-drug multimodal attribute graph, a graph neural network model with regularization is established;
[0014] S4: Use the existing microorganism-drug association database to construct training sets and test sets, and use the training set to train the regularized graph neural network model; input the test set into the trained graph neural network model to predict the microorganism-drug association in the test set.
[0015] In the above scheme, regularization is introduced into the neural network model to prevent the model from overfitting, thereby solving the problem of isolated nodes and improving the accuracy of microbial-drug association prediction.
[0016] Preferably, the step S1 specifically comprises:
[0017] A microorganism-drug association network is established using an existing microorganism-drug association database, wherein the microorganism-drug association data in the existing microorganism-drug association database is derived from clinical reports or experimental verifications, and the nodes of the microorganism-drug association network include microorganisms and drugs with known associations in the existing microorganism-drug association database.
[0018] Preferably, the step S2 specifically includes the following steps:
[0019] A1: Construct the similarity feature matrix of drugs and the second attribute feature matrix of drugs;
[0020] A2: Construct the similarity feature matrix of microorganisms and the second attribute feature matrix of microorganisms;
[0021] A3: Based on the similarity feature matrix of drugs and the similarity feature matrix of microorganisms, a microorganism-drug similarity feature network is constructed; based on the drug second attribute feature matrix and the microorganism second attribute feature matrix, a microorganism-drug second attribute feature network is constructed, and the microorganism-drug similarity feature network and the microorganism-drug second attribute feature network are combined to obtain a microorganism-drug multimodal attribute graph.
[0022] Preferably, the specific process of step A1 is:
[0023] Use the SIMCOMP2 tool to calculate the molecular structure similarity of drugs and obtain the molecular structure similarity matrix DS of drugs struct (di, dj);
[0024] The drug-drug interaction spectrum is represented by the matrix DIP, that is, DIP(d i ) indicates drug d i For interactions with other drugs, μ is used to represent the normalized kernel bandwidth, then:
[0025]
[0026] Among them, μ′ is the original bandwidth, which is set to 1;
[0027] The drug d i and drugs j The Gaussian kernel similarity between gauss (d i ,d j ), then:
[0028] DS gauss (d i ,d j )=exp(-μ||DIP(d i )-DIP(d j )|| 2 );
[0029] The similarity feature matrix of drugs is represented as S d (d i ,d j ), then:
[0030]
[0031] The topological property network of drugs is constructed by random walk method with restart, and random drift and restart are performed on the drug network, so that the formal definition of restart is:
[0032]
[0033] in, represents the probability that the i-th node moves to other nodes at time t+1, θ is the restart probability, T is the transition probability matrix, and p i (0) ∈R n×1 represents the starting probability vector of the i-th node, p i (t) ∈R n×1 represents the probability of the ith node moving to other nodes at time t. After the above formula converges, the probability distribution vector of each drug is obtained, and the second attribute feature matrix F of the drug is constructed based on it. d ∈R nd×nd , nd represents the number of microorganisms in the association network.
[0034] Preferably, the specific process of step A2 is:
[0035] Kamneva tool was used to calculate the microbial functional similarity and obtain the microbial similarity feature matrix S m ∈R nm×nm , where nm represents the number of microorganisms in the association network, S m (m i , m j ) represents microorganism m i and microorganisms j similarities between;
[0036] The original gene sequence of the microorganism is encoded, all sequences are filled with zeros to make all sequences of the same length, and finally the principal component analysis method is used to obtain the microbial sequence genus. The second attribute feature matrix of the microorganism is represented by a k-dimensional matrix F m ∈R nm×k .
[0037] Preferably, the specific process of step A3 is:
[0038] According to the drug similarity feature matrix and the microbial similarity feature matrix, a microbial-drug similarity feature network X is constructed. similitN , calculated by the following formula:
[0039]
[0040] Construct the microbial-drug second attribute characteristic network X based on the drug network topology attribute matrix and microbial sequence genus secondary , calculated by the following formula:
[0041]
[0042] The microbe-drug similarity feature network and the microbe-drug second attribute feature network are combined to obtain the microbe-drug multimodal attribute graph X, which is calculated by the following formula:
[0043] X=[X simility ,X secondary ].
[0044] Preferably, in step S3, the specific process of the graph neural network model is:
[0045] B1: Construct a GNCN encoder with L2 regularization, use the GNCN encoder to encode the multimodal attribute graph, and use the GNCN encoder to obtain the embedded representation of each node, that is, the latent variable Z;
[0046] B2: According to the latent variable Z, the encoded multimodal attribute graph is decoded using inner product decoding to generate a reconstructed graph.
[0047] Preferably, the specific process of step B1 is:
[0048] According to the multimodal attribute graph of microorganisms and drugs, a feature matrix of microorganisms and drugs is established, and a heterogeneous matrix of microorganisms and drugs is constructed. In the heterogeneous matrix, vi represents the microorganism or drug of any node. The heterogeneous matrix is expressed by the following formula:
[0049]
[0050] Among them, X is the characteristic matrix of microorganism-drug, is the content feature vector of node vi;
[0051] Set the learnable matrix to be W∈R m×f , and use random numbers to initialize the elements of the learnable matrix, where f is the dimension of the node embedding representation, which is set by the hyperparameters, and then n = nd + nm is used as the number of nodes, m = 2 × (nd + nm) is used as the feature dimension of the node, according to and W generate the vector after feature transformation The specific formula is as follows:
[0052]
[0053] Set the scaling constant s∈R, which represents the norm of the propagated hidden features and generates the normalized feature transformation vector by the GNCN network It is calculated by the following formula:
[0054]
[0055] The specific process of using the GNCN encoder to encode the microbial-drug association network and the microbial-drug multimodal attribute graph is expressed as:
[0056]
[0057] Among them, A∈R nd×nm is the adjacency matrix of the association network described in step S1. If there is a known association between nodes i and j in the association network, then the element A in A i j Set to 1, otherwise 0;
[0058] g(XW) is used to introduce L2 regularization, and its formula is:
[0059]
[0060] Among them, I N is the N-order unit matrix;
[0061] for The degree matrix of
[0062] GNCN network propagates normalized vectors to generate node embedding vectors It is calculated by the following formula:
[0063]
[0064] According to the node embedding vector Generate the node embedding matrix, that is, generate the hidden variable Z∈R of the GNCN encoder n ×f , and its calculation formula is:
[0065] Z = GNCN(X,A,s).
[0066] Preferably, the specific process of step B2 is:
[0067] The specific process of the decoder using the inner product to reconstruct the original image is:
[0068]
[0069] in, is the reconstructed image, and σ is the sigmoid activation function.
[0070] Preferably, the specific process of step S4 is:
[0071] Define the loss function: The loss function is the binary cross entropy between the multimodal attribute graph and the reconstruction graph, as follows:
[0072]
[0073] Among them, L is the loss function, N is the total number of all nodes, y represents the value of an element in the adjacency matrix A, which can be 0 or 1. Represents the reconstructed adjacency matrix The value of the corresponding element in is between 0 and 1;
[0074] Under the framework of k-fold cross-validation, all known microorganism-drug association data on the existing microorganism-drug association database are randomly divided into k groups, and each subset of randomly sampled unknown association pairs with the same batch size is selected as the test set, and the remaining known association pairs are selected as the training set. The regularized graph neural network model is trained with the training set, and the number of early stopping times is preset. During the training process, when the loss function converges, the AUC value of the trained graph neural network model is judged in the subsequent training within the number of early stopping times. If the AUC value of each time is not greater than the previous AUC value, the training is stopped, otherwise, the training continues;
[0075] The AUC value is calculated by the following method:
[0076] Input the training set into the model to obtain the reconstruction graph of the current model for the training set, take the score of the edge between each node in the reconstruction graph of the current model for the training set, and record it as the association probability, which takes a value between 0 and 1;
[0077] According to the association between microorganisms and drugs in the training set, the true value of the label of the edge in the training set is obtained. The true value of the label takes a value of 0 or 1, where 0 means that there is no edge, that is, there is no association, that is, it is actually a negative sample, and 1 means that there is an edge, that is, there is an association, that is, it is actually a positive sample;
[0078] Take each association probability as the classification threshold. When other association probabilities are greater than the classification threshold, they are considered to be predicted as positive samples. When other association probabilities are less than the classification threshold, they are considered to be predicted as negative samples.
[0079] The true positive rate and false positive rate under each classification threshold are calculated. The specific calculation formula is as follows:
[0080]
[0081]
[0082] Among them, TPRate is the true positive rate, FPRate is the false positive rate, TP is the true positive rate, which indicates the number of samples that are actually negative samples predicted as positive samples, FN is the false negative rate, which indicates the number of samples that are actually positive samples predicted as negative samples, FP is the false positive rate, which indicates the number of samples that are actually negative samples predicted as positive samples, and TN is the true negative rate, which indicates the number of samples that are actually negative samples predicted as negative samples;
[0083] The ROC curve is drawn with FPRate as the horizontal axis and TPRate as the vertical axis, and the ROC curve area, i.e., AUC value, is calculated using the differential element method;
[0084] After stopping training, the test set is input into the trained graph neural network model, and the association between microorganisms and drugs in the test set is predicted based on the generated reconstructed graph.
[0085] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:
[0086] The present invention avoids the zero trend of the isolated node paradigm by introducing L2 regularization, thereby improving the prediction performance of the isolated node model. BRIEF DESCRIPTION OF THE DRAWINGS
[0087] Figure 1 A schematic diagram of a process of predicting a microorganism-drug association based on a normalized graph neural network in the present invention;
[0088] Figure 2 A schematic diagram of the process of constructing a multimodal attribute graph of microorganisms and drugs in the present invention;
[0089] Figure 3 This is a schematic diagram of the process of introducing the regularized graph neural network model in the present invention. DETAILED DESCRIPTION
[0090] The drawings are for illustrative purposes only and should not be construed as limiting the present patent;
[0091] In order to better illustrate the present embodiment, some parts in the drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product;
[0092] It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.
[0093] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.
[0094] Example 1
[0095] A microbiome-drug association prediction method based on normalized graph neural network, such as Figure 1 As shown, the steps include:
[0096] S1: Use the existing microbiome-drug association database to construct a microbiome-drug association network;
[0097] S2: Constructing a multimodal attribute graph of microbes and drugs;
[0098] S3: Based on the microbial-drug association network and the microbial-drug multimodal attribute graph, a graph neural network model with regularization is established;
[0099] S4: Use the existing microorganism-drug association database to construct training sets and test sets, and use the training set to train the regularized graph neural network model; input the test set into the trained graph neural network model to predict the microorganism-drug association in the test set.
[0100] In the specific implementation process, regularization is introduced into the neural network model to prevent the model from overfitting, thereby solving the problem of isolated nodes and improving the accuracy of microbiome-drug association prediction.
[0101] Example 2
[0102] More specifically, the step S1 is as follows:
[0103] A microorganism-drug association network is established using an existing microorganism-drug association database, wherein the microorganism-drug association data in the existing microorganism-drug association database is derived from clinical reports or experimental verifications, and the nodes of the microorganism-drug association network include microorganisms and drugs with known associations in the existing microorganism-drug association database.
[0104] More specifically, Figure 2 As shown, step S2 specifically includes the following steps:
[0105] A1: Construct the similarity feature matrix of drugs and the second attribute feature matrix of drugs;
[0106] A2: Construct the similarity feature matrix of microorganisms and the second attribute feature matrix of microorganisms;
[0107] A3: Based on the similarity feature matrix of drugs and the similarity feature matrix of microorganisms, a microorganism-drug similarity feature network is constructed; based on the drug second attribute feature matrix and the microorganism second attribute feature matrix, a microorganism-drug second attribute feature network is constructed, and the microorganism-drug similarity feature network and the microorganism-drug second attribute feature network are combined to obtain a microorganism-drug multimodal attribute graph.
[0108] More specifically, the specific process of step A1 is:
[0109] Use the SIMCOMP2 tool to calculate the molecular structure similarity of drugs and obtain the molecular structure similarity matrix DS of drugs struct (di, dj);
[0110] The drug-drug interaction spectrum is represented by the matrix DIP, that is, DIP(d i ) indicates drug d i For interactions with other drugs, μ is used to represent the normalized kernel bandwidth, then:
[0111]
[0112] Among them, μ′ is the original bandwidth, which is set to 1;
[0113] The drug d i and drugs j The Gaussian kernel similarity between gauss (d i ,d j ), then:
[0114] DS gauss (d i ,d j )=exp(-μ||DIP(d i )-DIP(d j )|| 2 );
[0115] The similarity feature matrix of drugs is represented as S d (d i ,d j ), then:
[0116]
[0117] The topological property network of drugs is constructed by random walk method with restart, and random drift and restart are performed on the drug network, so that the formal definition of restart is:
[0118]
[0119] in, represents the probability that the i-th node moves to other nodes at time t+1, θ is the restart probability, T is the transition probability matrix, and p i (0) ∈R n×1 represents the starting probability vector of the i-th node, p i (t) ∈R n×1 represents the probability of the ith node moving to other nodes at time t. After the above formula converges, the probability distribution vector of each drug is obtained, and the second attribute feature matrix F of the drug is constructed based on it. d ∈R nd×nd , nd represents the number of microorganisms in the association network.
[0120] More specifically, the specific process of step A2 is:
[0121] Kamneva tool was used to calculate the microbial functional similarity and obtain the microbial similarity feature matrix S m ∈R nm×nm , where nm represents the number of microorganisms in the association network, S m(m i , m j ) represents microorganism m i and microorganisms j similarities between;
[0122] The original gene sequence of the microorganism is encoded, all sequences are filled with zeros to make all sequences of the same length, and finally the principal component analysis method is used to obtain the microbial sequence genus. The second attribute feature matrix of the microorganism is represented by a k-dimensional matrix F m ∈R nm×k .
[0123] More specifically, the specific process of step A3 is as follows:
[0124] According to the drug similarity feature matrix and the microbial similarity feature matrix, a microbial-drug similarity feature network X is constructed. simility , calculated by the following formula:
[0125]
[0126] Construct the microbial-drug second attribute characteristic network X based on the drug network topology attribute matrix and microbial sequence genus secondary , calculated by the following formula:
[0127]
[0128] The microbe-drug similarity feature network and the microbe-drug second attribute feature network are combined to obtain the microbe-drug multimodal attribute graph X, which is calculated by the following formula:
[0129] X=[X simility ,X secondary ].
[0130] More specifically, in step S3, if Figure 3 As shown, the specific process of the graph neural network model is as follows:
[0131] B1: Construct a GNCN encoder with L2 regularization, use the GNCN encoder to encode the multimodal attribute graph, and use the GNCN encoder to obtain the embedded representation of each node, that is, the latent variable Z;
[0132] B2: According to the latent variable Z, the encoded multimodal attribute graph is decoded using inner product decoding to generate a reconstructed graph.
[0133] More specifically, the specific process of step B1 is:
[0134] According to the multimodal attribute graph of microorganisms and drugs, a feature matrix of microorganisms and drugs is established, and a heterogeneous matrix of microorganisms and drugs is constructed. In the heterogeneous matrix, vi represents the microorganism or drug of any node. The heterogeneous matrix is expressed by the following formula:
[0135]
[0136] Among them, X is the characteristic matrix of microorganism-drug, is the content feature vector of node vi;
[0137] Set the learnable matrix to be W∈R m×f , and use random numbers to initialize the elements of the learnable matrix, where f is the dimension of the node embedding representation, which is set by the hyperparameters, and then n = nd + nm is used as the number of nodes, m = 2 × (nd + nm) is used as the feature dimension of the node, according to and W generate the vector after feature transformation The specific formula is as follows:
[0138]
[0139] Set the scaling constant s∈R, which represents the norm of the propagated hidden features and generates the normalized feature transformation vector by the GNCN network It is calculated by the following formula:
[0140]
[0141] The specific process of using the GNCN encoder to encode the microbial-drug association network and the microbial-drug multimodal attribute graph is expressed as:
[0142]
[0143] Among them, A∈R nd×nm is the adjacency matrix of the association network described in step S1. If there is a known association between nodes i and j in the association network, then the element A in A i j Set to 1, otherwise 0;
[0144] g(XW) is used to introduce L2 regularization, and its formula is:
[0145]
[0146] Among them, I N is the N-order unit matrix;
[0147] for The degree matrix of
[0148] GNCN network propagates normalized vectors to generate node embedding vectors It is calculated by the following formula:
[0149]
[0150] According to the node embedding vector Generate the node embedding matrix, that is, generate the hidden variable Z∈R of the GNCN encoder n ×f , and its calculation formula is:
[0151] Z = GNCN(X,A,s).
[0152] More specifically, the specific process of step B2 is:
[0153] The specific process of the decoder using the inner product to reconstruct the original image is:
[0154]
[0155] in, is the reconstructed image, and σ is the sigmoid activation function.
[0156] More specifically, the specific process of step S4 is:
[0157] Define the loss function: The loss function is the binary cross entropy between the multimodal attribute graph and the reconstruction graph, as follows:
[0158]
[0159] Among them, L is the loss function, N is the total number of all nodes, y represents the value of an element in the adjacency matrix A, which can be 0 or 1. Represents the reconstructed adjacency matrix The value of the corresponding element in is between 0 and 1;
[0160] Under the framework of k-fold cross-validation, all known microorganism-drug association data on the existing microorganism-drug association database are randomly divided into k groups, and each subset of randomly sampled unknown association pairs with the same batch size is selected as the test set, and the remaining known association pairs are selected as the training set. The regularized graph neural network model is trained with the training set, and the number of early stopping times is preset. During the training process, when the loss function converges, the AUC value of the trained graph neural network model is judged in the subsequent training within the number of early stopping times. If the AUC value of each time is not greater than the previous AUC value, the training is stopped, otherwise, the training continues;
[0161] The AUC value is calculated by the following method:
[0162] Input the training set into the model to obtain the reconstruction graph of the current model for the training set, take the score of the edge between each node in the reconstruction graph of the current model for the training set, and record it as the association probability, which takes a value between 0 and 1;
[0163] According to the association between microorganisms and drugs in the training set, the true value of the label of the edge in the training set is obtained. The true value of the label takes a value of 0 or 1, where 0 means that there is no edge, that is, there is no association, that is, it is actually a negative sample, and 1 means that there is an edge, that is, there is an association, that is, it is actually a positive sample;
[0164] Take each association probability as the classification threshold. When other association probabilities are greater than the classification threshold, they are considered to be predicted as positive samples. When other association probabilities are less than the classification threshold, they are considered to be predicted as negative samples.
[0165] The true positive rate and false positive rate under each classification threshold are calculated. The specific calculation formula is as follows:
[0166]
[0167]
[0168] Among them, TPRate is the true positive rate, FPRate is the false positive rate, TP is the true positive rate, which indicates the number of samples that are actually negative samples predicted as positive samples, FN is the false negative rate, which indicates the number of samples that are actually positive samples predicted as negative samples, FP is the false positive rate, which indicates the number of samples that are actually negative samples predicted as positive samples, and TN is the true negative rate, which indicates the number of samples that are actually negative samples predicted as negative samples;
[0169] The ROC curve is drawn with FPRate as the horizontal axis and TPRate as the vertical axis, and the ROC curve area, i.e., AUC value, is calculated using the differential element method;
[0170] After stopping training, the test set is input into the trained graph neural network model, and the association between microorganisms and drugs in the test set is predicted based on the generated reconstructed graph.
[0171] Example 3
[0172] In order to verify that the accuracy of the microorganism-drug association prediction method based on normalized graph neural network provided by the present invention is better than the existing methods, this embodiment uses the default parameter settings, runs this method and 5 existing methods on the MDAD data set, and uses the AUC value as the performance evaluation indicator. The larger the AUC value, the higher the accuracy of the method.
[0173] The specific process is as follows:
[0174] All methods were subjected to 5-fold cross-validation and 10-fold cross-validation, that is, the experimentally verified drug combinations were randomly divided into 5 or 10 subsets of equal size, each of which was used as a test set in turn, and the remaining were used to train the model. In order to eliminate random sampling bias, the process was repeated 10 times, and the final AUC score was calculated based on the average of the AUC values in the 10 repeated validations. The final AUC score was used as a performance indicator to evaluate the accuracy of each method. The larger the final AUC score, the higher the accuracy of the method.
[0175] The validation results are shown in the following table. In the 5-fold cross validation, the final AUC score of the microorganism-drug association prediction method based on the normalized graph neural network provided by the present invention is the highest among all methods, which is 09676. In the 10-fold cross validation, the final AUC score of the microorganism-drug association prediction method based on the normalized graph neural network provided by the present invention is also the highest among all methods, which is 0.9709. Therefore, the accuracy of the microorganism-drug association prediction method based on the normalized graph neural network provided by the present invention is better than the existing methods.
[0176] Table 1
[0177]
[0178] Example 4
[0179] In order to further verify the effectiveness of the microorganism-drug association prediction method based on normalized graph neural network provided by the present invention, this example applies the present invention to two drugs and two microorganisms, predicts the top 20 microorganisms or drugs associated with them, and then verifies the accuracy of the prediction results by searching literature reports. During the prediction process, the drug or microorganism in the case study is regarded as an isolated node.
[0180] The specific process is as follows:
[0181] Using the microorganism-drug association prediction method based on normalized graph neural network provided by the present invention, the drug curcumin was predicted, and the top 20 microorganisms associated with the drug curcumin were obtained, of which 16 have been confirmed to be associated in relevant publications. The top 20 predicted microorganisms associated with the drug curcumin and related publications are shown in Table 2.
[0182] Table 2
[0183]
[0184] Using the microorganism-drug association prediction method based on normalized graph neural network provided by the present invention, the drug gallic acid was predicted, and the top 20 microorganisms associated with it were obtained, of which 16 have been confirmed to be associated in relevant publications. The top 20 predicted microorganisms associated with the drug gallic acid and related publications are shown in Table 3.
[0185] Table 3
[0186]
[0187] Using the microorganism-drug association prediction method based on normalized graph neural network provided by the present invention, the microorganism Staphylococcus aureus was predicted, and the top 20 drugs associated with the microorganism Staphylococcus aureus were obtained, of which 15 have been confirmed to be associated in relevant publications. The top 20 predicted drugs associated with the microorganism Staphylococcus aureus and related publications are shown in Table 4.
[0188] Table 4
[0189]
[0190] Using the microorganism-drug association prediction method based on normalized graph neural network provided by the present invention, the microorganism Pseudomonas aeruginosa was predicted, and the top 20 drugs associated with the microorganism Pseudomonas aeruginosa were obtained, of which 15 have been confirmed to be associated in relevant publications. The top 20 predicted drugs associated with the microorganism Pseudomonas aeruginosa and related publications are shown in Table 5.
[0191] Table 5
[0192]
[0193] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.
Claims
1. A method for predicting microbial-drug association based on normalized graph neural network, characterized in that: Includes steps: S1: Use the existing microbiome-drug association database to construct a microbiome-drug association network; S2: Constructing a multimodal attribute graph of microbes and drugs; S3: Based on the microbial-drug association network and the microbial-drug multimodal attribute graph, a graph neural network model with regularization is established; S4: Use the existing microbe-drug association database to build training sets and test sets, and use the training set to train the regularized graph neural network model; input the test set into the trained graph neural network model to predict the microbe-drug association in the test set; The step S1 is specifically as follows: Establishing a microorganism-drug association network using an existing microorganism-drug association database, wherein the microorganism-drug association data in the existing microorganism-drug association database is derived from clinical reports or experimental verifications, and the nodes of the microorganism-drug association network include microorganisms and drugs with known associations in the existing microorganism-drug association database; The step S2 specifically includes the following steps: A1: Construct the similarity feature matrix of drugs and the second attribute feature matrix of drugs; A2: Construct the similarity feature matrix of microorganisms and the second attribute feature matrix of microorganisms; A3: Based on the similarity feature matrix of drugs and the similarity feature matrix of microorganisms, a microorganism-drug similarity feature network is constructed; based on the second attribute feature matrix of drugs and the second attribute feature matrix of microorganisms, a microorganism-drug second attribute feature network is constructed, and the microorganism-drug similarity feature network and the microorganism-drug second attribute feature network are combined to obtain a multimodal attribute graph of microorganisms-drugs; The specific process of step A1 is as follows: Use the SIMCOMP2 tool to calculate the molecular structure similarity of drugs and obtain the molecular structure similarity matrix DS of drugs struct (di, dj); The drug-drug interaction spectrum is represented by the matrix DIP, that is, DIP(d i ) indicates drug d i For interactions with other drugs, μ is used to represent the normalized kernel bandwidth, then: Among them, μ′ is the original bandwidth, which is set to 1; The drug d i and drugs j The Gaussian kernel similarity between gauss (d i ,d j ), then: DS gauss (d i ,d j )=exp(-μ||DIP(d i )-DIP(d j )|| 2 ); The similarity feature matrix of drugs is represented as S d (d i ,d j ), then: The topological property network of drugs is constructed by random walk method with restart, and random drift and restart are performed on the drug network, so that the formal definition of restart is: in, represents the probability that the i-th node moves to other nodes at time t+1, θ is the restart probability, T is the transition probability matrix, and p i (0) ∈R n×1 represents the starting probability vector of the i-th node, p i (t) ∈R n×1 represents the probability of the ith node moving to other nodes at time t. After the above formula converges, the probability distribution vector of each drug is obtained, and the second attribute feature matrix F of the drug is constructed based on it. d ∈R nd×nd , nd represents the number of microorganisms in the association network; The specific process of step A2 is as follows: Kamneva tool was used to calculate the microbial functional similarity and obtain the microbial similarity feature matrix S m ∈R nm×nm , where nm represents the number of microorganisms in the association network, S m (m i , m j ) represents microorganism m i and microorganisms j similarities between; The original gene sequence of the microorganism is encoded, all sequences are filled with zeros to make all sequences of the same length, and finally the principal component analysis method is used to obtain the microbial sequence genus. The second attribute feature matrix of the microorganism is represented by a k-dimensional matrix F m ∈R nm×k ; The specific process of step A3 is as follows: According to the drug similarity feature matrix and the microbial similarity feature matrix, a microbial-drug similarity feature network X is constructed. simility , calculated by the following formula: Construct the microbial-drug second attribute characteristic network X based on the drug network topology attribute matrix and microbial sequence genus secondary , calculated by the following formula: The microbe-drug similarity feature network and the microbe-drug second attribute feature network are combined to obtain the microbe-drug multimodal attribute graph X, which is calculated by the following formula: X=[X simility ,X secondary ]; In step S3, the specific process of the graph neural network model is as follows: B1: Construct a GNCN encoder with L2 regularization, use the GNCN encoder to encode the multimodal attribute graph, and use the GNCN encoder to obtain the embedded representation of each node, that is, the latent variable Z; B2: According to the latent variable Z, the encoded multimodal attribute graph is decoded using inner product decoding to generate a reconstructed graph.
2. According to claim 1, a method for predicting microorganism-drug association based on normalized graph neural network is characterized in that: The specific process of step B1 is as follows: According to the multimodal attribute graph of microorganisms and drugs, a feature matrix of microorganisms and drugs is established, and a heterogeneous matrix of microorganisms and drugs is constructed. In the heterogeneous matrix, vi represents the microorganism or drug of any node. The heterogeneous matrix is expressed by the following formula: Among them, X is the characteristic matrix of microorganism-drug, is the content feature vector of node vi; Set the learnable matrix to be W∈R m×f , and use random numbers to initialize the elements of the learnable matrix, where f is the dimension of the node embedding representation, which is set by the hyperparameters, and then n = nd + nm is used as the number of nodes, m = 2 × (nd + nm) is used as the feature dimension of the node, according to and W generate the vector after feature transformation The specific formula is as follows: Set the scaling constant s∈R, which represents the norm of the propagated hidden features and generates the normalized feature transformation vector by the GNCN network It is calculated by the following formula: The specific process of using the GNCN encoder to encode the microbial-drug association network and the microbial-drug multimodal attribute graph is expressed as: Among them, A∈R nd×nm is the adjacency matrix of the association network described in step S1. If there is a known association between nodes i and j in the association network, then the element A in A ij Set to 1, otherwise 0; g(XW) is used to introduce L2 regularization, and its formula is: Among them, I N is the N-order identity matrix; for The degree matrix of GNCN network propagates normalized vectors to generate node embedding vectors It is calculated by the following formula: According to the node embedding vector Generate the node embedding matrix, that is, generate the hidden variable Z∈R of the GNCN encoder n×f , and its calculation formula is: Z = GNCN(X,A,s).
3. The microorganism-drug association prediction method based on normalized graph neural network according to claim 2 is characterized in that: The specific process of step B2 is: The specific process of the decoder using the inner product to reconstruct the original image is: in, is the reconstructed image, and σ is the sigmoid activation function.
4. The method for predicting microorganism-drug association based on normalized graph neural network according to claim 3, characterized in that: The specific process of step S4 is as follows: Define the loss function: The loss function is the binary cross entropy between the multimodal attribute graph and the reconstruction graph, as follows: Among them, L is the loss function, N is the total number of all nodes, y represents the value of an element in the adjacency matrix A, which can be 0 or 1. Represents the reconstructed adjacency matrix The value of the corresponding element in is between 0 and 1; Under the framework of k-fold cross-validation, all known microorganism-drug association data on the existing microorganism-drug association database are randomly divided into k groups, and each subset of randomly sampled unknown association pairs with the same batch size is selected as the test set, and the remaining known association pairs are selected as the training set. The regularized graph neural network model is trained with the training set, and the number of early stopping times is preset. During the training process, when the loss function converges, the AUC value of the trained graph neural network model is judged in the subsequent training within the number of early stopping times. If the AUC value of each time is not greater than the previous AUC value, the training is stopped, otherwise, the training continues; The AUC value is calculated by the following method: Input the training set into the model to obtain the reconstruction graph of the current model for the training set, take the score of the edge between each node in the reconstruction graph of the current model for the training set, and record it as the association probability, which takes a value between 0 and 1; According to the association between microorganisms and drugs in the training set, the true value of the label of the edge in the training set is obtained. The true value of the label takes a value of 0 or 1, where 0 means that there is no edge, that is, there is no association, that is, it is actually a negative sample, and 1 means that there is an edge, that is, there is an association, that is, it is actually a positive sample; Take each association probability as the classification threshold. When other association probabilities are greater than the classification threshold, they are considered to be predicted as positive samples. When other association probabilities are less than the classification threshold, they are considered to be predicted as negative samples. The true positive rate and false positive rate under each classification threshold are calculated. The specific calculation formula is as follows: Among them, TPRate is the true positive rate, FPRate is the false positive rate, TP is the true positive rate, which indicates the number of samples predicted as positive samples when they are actually positive samples, FN is the false negative rate, which indicates the number of samples predicted as negative samples when they are actually positive samples, FP is the false positive rate, which indicates the number of samples predicted as positive samples when they are actually negative samples, and TN is the true negative rate, which indicates the number of samples predicted as negative samples when they are actually negative samples; The ROC curve is drawn with FPRate as the horizontal axis and TPRate as the vertical axis, and the ROC curve area, i.e., AUC value, is calculated using the differential element method; After stopping training, the test set is input into the trained graph neural network model, and the association between microorganisms and drugs in the test set is predicted based on the generated reconstructed graph.
