A method and system for predicting the association between microorganisms and diseases based on adversarial regularized graph neural networks

Abstract

The present invention discloses a method and system for predicting microbial-disease associations based on an adversarial regularized graph neural network. The method comprises the following steps: S1. Acquiring data on the association of microorganisms and diseases and constructing a microorganism-disease association matrix; S2. Obtaining comprehensive similarity matrices for microorganisms and diseases based on the microorganism-disease association matrix; S3. Constructing a heterogeneous microorganism-disease network using the microorganism-disease association matrix, the comprehensive microorganism similarity matrix, and the comprehensive disease similarity matrix; S4. Using a graph neural network to predict associations based on the heterogeneous microorganism-disease network; and training the model using the loss function of the adversarial model as a regularizer added to the loss function of the graph neural network. This solution transforms the prediction problem into a graph-structured data processing problem and uses a graph neural network to capture the topological structure and attribute information of nodes in the heterogeneous network. This method predicts associations using the graph neural network, achieving excellent prediction results.

Description

Technical Field

[0001] The present invention belongs to the field of computer bioinformatics, and in particular relates to a method and system for predicting the association between microorganisms and diseases based on an adversarial regularized graph neural network. Background Art

[0002] The microbiome encompasses a diverse array of tiny organisms, including bacteria, viruses, fungi, and parasites, that inhabit the bodies of animals and plants, as well as the surrounding environment. The intricate relationship between microbes and disease is well established, with microbial communities closely interacting with their human hosts, exerting positive, negative, or neutral effects. These interactions may underlie the pathogenesis of numerous diseases, including infections with bacteria, viruses, fungi, and parasites.

[0003] Discovering associations between microbes and diseases has important scientific and clinical implications. Traditional methods for predicting microbial-disease associations rely primarily on statistical analysis and simple machine learning models. These methods have limitations when processing large-scale, complex microbial-disease data. For example, they cannot effectively capture topological information between nodes and multi-source similarity information, resulting in insufficient prediction accuracy and generalization capabilities. Summary of the Invention

[0004] The purpose of the present invention is to address the problems existing in the prior art and propose a method and system for predicting the association between microorganisms and diseases based on adversarial regularized graph neural networks.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for predicting the association between microorganisms and diseases based on an adversarial regularized graph neural network, the method comprising:

[0007] S1. Obtain data on the association between microorganisms and diseases and construct a microorganism-disease association matrix;

[0008] S2. Obtaining a microbial comprehensive similarity matrix and a disease comprehensive similarity matrix based on the microbial-disease association matrix;

[0009] S3. Construct a microbe-disease heterogeneous network using the microbe-disease association matrix, the microbe comprehensive similarity matrix, and the disease comprehensive similarity matrix;

[0010] S4. Using a graph neural network to predict associations based on the microbe-disease heterogeneous network;

[0011] The loss function of the adversarial model including the generator and the discriminator is added as a regular term to the loss function of the graph neural network for model training.

[0012] In the above-mentioned method for predicting the association between microorganisms and diseases based on adversarial regularized graph neural networks, in step S1, each row / column in the microorganism-disease association matrix represents a microorganism, each column / row represents a disease, and an element in the matrix represents whether there is an association between the microorganism and the disease;

[0013] In the above-mentioned method for predicting the association between microorganisms and diseases based on adversarial regularized graph neural networks, step S2 specifically includes:

[0014] According to the microorganism-disease association matrix, the topological structure similarity and functional similarity between microorganisms are obtained respectively, and the topological structure similarity and functional similarity between microorganisms are weightedly combined to obtain the microorganism comprehensive similarity matrix;

[0015] According to the microorganism-disease association matrix, the topological structure similarity and functional similarity between diseases are obtained respectively, and the topological structure similarity and functional similarity between diseases are weightedly combined to obtain the disease comprehensive similarity matrix.

[0016] In the above-mentioned method for predicting the association between microorganisms and diseases based on adversarial regularized graph neural networks, the topological similarity includes GIP core similarity and HIP similarity. GIP core similarity and HIP similarity measure the similarity between microorganisms and diseases based on the network topology, while functional similarity measures the similarity based on biological characteristics.

[0017] The microbial comprehensive similarity matrix and disease comprehensive similarity matrix are expressed as 、 :

[0018]

[0019]

[0020] 、 、 They represent the GIP core similarity, HIP similarity and functional similarity of microorganisms respectively;

[0021] 、 、 Represent the GIP nuclear similarity, HIP similarity and functional similarity of the disease, respectively.

[0022] In the above-mentioned microorganism-disease association prediction method based on adversarial regularized graph neural network, in step S3, the constructed microorganism-disease heterogeneous network is represented by the following matrix:

[0023]

[0024] in, represents the comprehensive similarity matrix of microorganisms;

[0025] represents the disease comprehensive similarity matrix;

[0026] represents the microbe-disease association matrix;

[0027] yes The transposed matrix of .

[0028] Microorganisms and diseases with similarity scores greater than a specific value, such as greater than 0.5, are considered to be associated. The above-mentioned microorganism-disease heterogeneous network graph is constructed based on the association relationships between microorganisms, between diseases, and between microorganisms and diseases. That is, this microorganism-disease heterogeneous network is a heterogeneous graph used to indicate whether there is an association between microorganisms and microorganisms, diseases and diseases, and microorganisms and diseases.

[0029] In the above-mentioned method for predicting the association between microorganisms and diseases based on adversarial regularized graph neural network, in step S4, the graph neural network includes an encoder and a decoder;

[0030] The encoder selects a subset of neighboring nodes for each node in the microorganism-disease heterogeneous network, aggregates the feature vectors of the selected neighboring nodes, and updates the feature vector of the current node by combining the features of the current node and the aggregated neighboring nodes through a nonlinear transformation. The aggregation and updating process is repeated, and the topological structure information and attribute information of the node are gradually captured through a multi-layer neural network, ultimately generating a low-dimensional feature vector for the node.

[0031] The inner product of the low-dimensional feature vector is used as a decoder, association prediction is performed through the low-dimensional node representation, and the cross entropy loss function is used to calculate the difference between the prediction result and the true label.

[0032] The heterogeneous network graph is the input of the graph neural network (representing the nodes and the associations between nodes), and the comprehensive similarity matrix of microorganisms and diseases is the initial feature input of the nodes.

[0033] In the aforementioned microbe-disease association prediction method based on adversarial regularized graph neural network, the encoder construction method in the graph neural network includes:

[0034] The initial feature vectors of microbial nodes and disease nodes are defined as follows:

[0035]

[0036] represents the microorganism / disease node;

[0037] Representation node Feature vector at the initial layer;

[0038] is the initial feature vector of the node, i.e., the comprehensive microbial similarity matrix / comprehensive disease similarity matrix;

[0039] , It is the set of all nodes in the biological-disease heterogeneous network graph, which means traversing all nodes in the graph ;

[0040] For each node , and then sample its neighbor nodes to get the node representation , aggregate it:

[0041]

[0042] in, Representation node In the Neighborhood aggregate feature vector of the layer;

[0043] For nodes Neighbor nodes of is the network layer number, indicating the number of hops to access neighbors, Indicates that the average aggregation function MEAN is used to calculate the average value of each dimension;

[0044] node Through nonlinear transformation, it updates its own representation by using its own representation and the aggregated representation of its neighbors. , ;

[0045] Get the final encoder output ,

[0046] Get the feature representation Z of each node based on the encoder,

[0047] in, Representation node In the The feature vector of the layer, Indicates traversing all nodes in the graph ;

[0048] represents a nonlinear activation function; Indicates the The weight matrix of the layer;

[0049] Represents a concatenation operation, which concatenates two vectors together;

[0050] Representation node exist The feature vector of the layer;

[0051] Represents the set of all nodes in the graph;

[0052] The feature matrix representing the disease node; The feature matrix representing the microbial nodes;

[0053] represents the number of disease nodes; represents the number of microbial nodes; Represents the dimension of the feature vector.

[0054] First, by aggregating the information of neighboring nodes and performing nonlinear transformations, the feature vector of each node is gradually updated. Finally, the feature vectors of all nodes are concatenated into a feature matrix Z for subsequent prediction tasks.

[0055] In the aforementioned microbe-disease association prediction method based on adversarial regularized graph neural network, the decoder in the graph neural network is constructed in the following ways:

[0056] Matrix-based , using its inner product as a decoder, performing association prediction through low-dimensional node representation, and using the cross entropy loss function to calculate the distance between the prediction result and the true label:

[0057]

[0058]

[0059] Represents the association prediction result;

[0060] represents the activation function, which is used to map the inner product result to the interval [0,1];

[0061] represents the cross entropy loss function;

[0062] Indicates the The association vector of each node in the original association matrix;

[0063] Indicates the The association prediction results of the nodes, the association vector in the prediction association matrix;

[0064] Represents the cross entropy distance function, which is used to calculate the distance between two elements;

[0065] The decoder passes the feature matrix The inner product and activation function get the predicted correlation matrix Then, the original correlation matrix is ​​calculated using the cross entropy distance function and the predicted correlation matrix The difference between them and summing over all nodes gives the loss function of the graph neural network This loss function is used to guide the training of the model, enabling the model to generate more accurate node feature vectors.

[0066] In the above-mentioned microorganism-disease association prediction method based on adversarial regularized graph neural network, in step S4, the generator of the adversarial model outputs a pseudo microorganism-disease association feature matrix based on random noise sampled from a predefined distribution;

[0067] The discriminator of the adversarial model distinguishes the real microorganism and disease association feature matrix generated by the graph neural network The pseudo-microorganism and disease association feature matrix generated by the generator in the adversarial model ;

[0068] The adversarial model’s loss function is added as a regularization term to the graph neural network’s loss function to feed back the difference between positive and negative samples to the graph neural network during the adversarial model training process:

[0069]

[0070] represents the graph neural network loss function, represents the loss function of the adversarial model, Represents the weight of the adversarial model loss term.

[0071] A microbe-disease association prediction system based on an adversarial regularized graph neural network is configured to receive an association matrix input containing microbe-disease pairs to be predicted and predict whether the microbe-disease pairs to be predicted are associated;

[0072] The association matrix input includes known associations between the microorganism to be predicted and the disease in the microorganism-disease pair to be predicted, as well as known associations between the disease to be predicted and the microorganism;

[0073] The system includes a similarity acquisition module, a graph network construction module, and a prediction module;

[0074] A similarity acquisition module is used to calculate the GIP core similarity, HIP similarity, and functional similarity of the microorganism to be predicted based on the correlation matrix input, thereby obtaining a comprehensive similarity matrix of the microorganism to be predicted;

[0075] Calculate the GIP core similarity, HIP similarity, and functional similarity of the disease to be predicted, and then obtain a comprehensive similarity matrix of the disease to be predicted;

[0076] A graph network construction module is used to construct a heterogeneous network of the microorganisms to be predicted and the diseases to be predicted based on the association matrix input, the comprehensive similarity matrix of the microorganisms to be predicted, and the comprehensive similarity matrix of the diseases to be predicted;

[0077] The prediction module is used to implement the graph neural network constructed by the above method, taking the heterogeneous network of the microorganism to be predicted and the disease to be predicted as input, and outputting the association prediction result of the microorganism to be predicted-disease pair.

[0078] The advantages of the present invention are:

[0079] 1) This approach uses biomedical data, such as disease-gene associations, gene-gene interactions, and protein-protein interactions, to derive functional similarities between microbes and diseases. Furthermore, based on the topological relationships between microbe and disease networks, structural similarities between microbes and diseases are calculated using GIP core similarity and HIP similarity. Finally, a heterogeneous microbe-disease network is constructed based on the microbe similarity network, disease similarity network, and known microbe-disease associations.

[0080] 2) This solution builds a microbe-disease heterogeneous network based on the above method, transforming the prediction problem into a graph-structured data processing problem. It also uses graph neural networks to capture the topological structure and attribute information of nodes in the heterogeneous network, enabling association prediction through graph neural networks. This approach fully utilizes the advantages of graph neural networks in processing graph-structured data to achieve excellent prediction results.

[0081] 3) The graph neural network constructed in this scheme includes an encoder and a decoder. After using the encoder to extract the node representation, the association representation between microorganisms and diseases and the random noise of the generator are input into the discriminator of the adversarial model for true or false judgment. Finally, the loss function of the adversarial model is added as a regular term to the loss function of the graph neural network. During the training process of the adversarial model, the difference between positive and negative samples is fed back to the graph neural network, significantly enhancing the robustness and predictive ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 Shown is a flow chart of the model construction of the microorganism and disease association prediction method based on the adversarial regularized graph neural network of the present invention;

[0083] Figure 2 FIG2 is a schematic diagram showing the principle of the method for predicting the association between microorganisms and diseases based on adversarial regularized graph neural networks of the present invention;

[0084] Figure 3 Shown is a prediction flow chart of the microorganism-disease association prediction method based on adversarial regularized graph neural network of the present invention;

[0085] Figure 4 Shown is the performance comparison of different models on the HMDAD dataset;

[0086] Figure 5 Performance comparison of different models on the Disbiome dataset. DETAILED DESCRIPTION

[0087] like Figure 1 and Figure 2 As shown in the figure, this scheme provides a method for predicting the association between microorganisms and diseases based on adversarial regularized graph neural networks. The specific steps include:

[0088] S1. Obtain data on the association between microorganisms and diseases and construct a microorganism-disease association matrix;

[0089] S2. Based on the microbe-disease association matrix, obtain the microbe comprehensive similarity matrix and the disease comprehensive similarity matrix respectively;

[0090] S3. Construct a heterogeneous microbiome-disease network using the microbiome-disease association matrix, the microbiome comprehensive similarity matrix, and the disease comprehensive similarity matrix.

[0091] S4. Using Graph Neural Networks to Predict Associations Based on Microbe-Disease Heterogeneous Networks;

[0092] The loss function of the adversarial model containing the generator and the discriminator is added as a regular term to the loss function of the graph neural network for model training.

[0093] Specifically, the association data in step S1 of this example used known microbe-disease associations from the HMDAD database, including 450 experimentally confirmed associations involving 39 diseases and 292 microbes, and 4,351 associations between 218 diseases and 1,052 microbes from Disbiome. Statistics for these two microbe-disease association datasets are shown in Table 1.

[0094] Table 1 Disease and microbial association data

[0095]

[0096] use represents the microbe-disease association matrix, where and Represents the number of microorganisms and diseases respectively. For any given microorganism and diseases , if they have a known association, then ,otherwise , as shown in Table 2:

[0097] Table 2 Microorganism-disease association matrix

[0098]

[0099] In this embodiment S2, in order to obtain the comprehensive microbial similarity matrix and the comprehensive disease similarity matrix, the following steps are specifically included:

[0100] S2-1. Calculate the GIP (Gaussian Interaction Profile) core similarity and HIP (Heterogeneous Interaction Profile) similarity of microorganisms respectively;

[0101] The formula for calculating the similarity of microbial GIP cores is as follows:

[0102] (1)

[0103] (2)

[0104] 、 Microorganisms 、 interaction profiles in the microbe-disease association matrix;

[0105] is the norm of the interaction profile, and the norm is calculated as follows:

[0106] (3)

[0107] Representation matrix The elements in the rows and columns of Representation matrix the number of rows, Representation matrix D The number of columns.

[0108] The formula for calculating disease GIP nuclear similarity is as follows:

[0109] (4)

[0110] (5)

[0111] 、 Diseases 、 Interaction profiles in the microbe-disease association matrix.

[0112] The calculation formula of microbial HIP similarity is as follows:

[0113] (6)

[0114] 、 Represents microorganisms 、 The interaction profile in the microbe-disease association matrix, i.e., the association vector with all diseases or other microbes.

[0115] in express and The number of elements in unequal, express The number of elements in .

[0116] The disease HIP similarity calculation formula is as follows:

[0117] (7)

[0118] express and The number of elements in unequal, express The number of elements in .

[0119] S2-2. Calculate disease functional similarity;

[0120] Regarding the functional similarity of diseases, similar diseases often interact with similar genes. This scheme calculates the functional similarity of diseases based on the functional associations between disease-related genes. First, gene interactions are downloaded from the HumanNet database, where each gene-gene interaction is associated with a log-likelihood score (LLS). Then, for each disease, and ,set up and , and Respectively represent and Then, define the gene set and Functional similarities between diseases , as shown in the formula:

[0121] (8)

[0122] (9)

[0123] in, 、 Gene sets and the number of genes in It's the gene In gene sets The feature similarity score in .

[0124] Indicates gene In gene sets The feature similarity score in ;

[0125] For genes and The functional similarity score between them is the log-likelihood score of gene interaction downloaded from the HumanNet database, and is calculated as follows:

[0126] (10)

[0127] in and are the maximum and minimum values ​​of LLS in HumanNet, respectively.

[0128] S2-3. Calculate microbial functional similarity;

[0129] Microbial functional similarity The calculation is based on the protein function information provided in the String v11 database. First, the protein set of each microorganism and its protein interaction information (including confidence score) are extracted from the STRING v11 database. Then, the protein sets of two microorganisms are paired and the confidence score between each pair of proteins is calculated. Finally, the average of all confidence scores is used as the microbial functional similarity. :

[0130] (11)

[0131] in, and Represents microorganisms and A collection of proteins, and Represents microorganisms and The number of protein sets, Indicates protein pair and STRING confidence score.

[0132] 、 and The values ​​of the three matrices range from 0 to 1. 、 and The values ​​of the three matrices range from 0 to 1. Combining the three matrices creates a comprehensive microbial similarity matrix and a comprehensive disease similarity matrix. GIP kernel similarity measures the similarity between microorganisms and diseases based on network topology, while functional similarity measures similarity based on biological characteristics. The resulting comprehensive microbial similarity matrix and comprehensive disease similarity matrix are shown in the following formulas:

[0133] (12)

[0134] (13)

[0135] 、 、 They represent the GIP core similarity, HIP similarity and functional similarity of microorganisms respectively;

[0136] 、 、 Represent the GIP nuclear similarity, HIP similarity and functional similarity of the disease, respectively.

[0137] In step S3, the heterogeneous network pass 、 and Build:

[0138] (14)

[0139] represents the comprehensive similarity matrix of microorganisms;

[0140] represents the disease comprehensive similarity matrix;

[0141] represents the microbe-disease association matrix;

[0142] yes The transposed matrix of .

[0143] This embodiment S4 specifically includes:

[0144] S4-1. Encoder construction;

[0145] The calculated comprehensive similarity matrix between microbial nodes and disease nodes is used as the initial node feature vector for each node.

[0146] (15)

[0147] in for the microorganism / disease node, is the initialization feature vector of the node, that is, the similarity matrix of microorganisms or diseases.

[0148] For each node , and then sample its neighbor nodes to get the node representation , aggregate it:

[0149] (16)

[0150] in, Representation node In the Neighborhood aggregate feature vector of the layer;

[0151] For nodes Neighbor nodes of is the network layer number, indicating the number of hops to access the neighbor. Indicates that the average aggregation function MEAN is used to calculate the average value of each dimension.

[0152] node Through nonlinear transformation, it updates its own representation by using its own representation and the aggregated representation of its neighbors. After calculating the maximum number of hops K to visit neighbors, the final graph autoencoder output is obtained , that is, the feature representation of the node , and finally obtain the characteristic matrix of microorganisms and diseases:

[0153] (17)

[0154] (18)

[0155] (19)

[0156] in, Representation node In the The feature vector of the layer;

[0157] represents a nonlinear activation function; Indicates the The weight matrix of the layer;

[0158] Represents a concatenation operation, which concatenates two vectors together;

[0159] Representation node exist The feature vector of the layer; Representation node The final eigenvector of

[0160] Represents the set of all nodes in the graph; Represents the feature matrix of all nodes;

[0161] and Representing low-dimensional topological representations of disease nodes and microbial nodes, respectively.

[0162] In this embodiment, the encoder uses GraphSAGE (Graph Sample and Aggregate), the output dimension is set to 16, and a 3-layer graph neural network architecture is adopted.

[0163] S4-2. Decoder construction;

[0164] Matrix-based , using its inner product as a decoder, the decoding function infers the original input through the low-dimensional node representation. Decoding results It should be close to the label matrix associated with the N original inputs (i.e., microorganisms and diseases). The cross entropy loss function is used to calculate The distance between the original input label and the distance function is defined by cross entropy:

[0165] (20)

[0166] (twenty one)

[0167] (twenty two)

[0168] This approach constructs a global representation of microbe-disease associations to enhance local information. Random noise sampled from a predefined distribution is input into the generator. The goal of the generator is to maximize the probability that the discriminator will classify the generated samples as real samples, while the goal of the discriminator is to maximize its ability to accurately distinguish between real data and generated data. The calculation is as follows:

[0169] (twenty three) The distribution of real association data between microorganisms and diseases randomly sampled for each training, Represents the discriminator's distribution from the real data Samples sampled in The expected value of the output;

[0170] The distribution of microorganism and disease association data generated by the generator, Represents the distribution of data generated by the discriminator from the generator Samples sampled in The expected value of the output;

[0171] Represents the weight of the gradient penalty term;

[0172] is the interpolation between the real sample and the generated sample,

[0173] (twenty four)

[0174] is the discriminator's interpolation sample The second norm of the gradient of .

[0175] The loss function Added as a regularization term to the graph neural network, during the training process of WGAN-GP, the difference between positive and negative samples is fed back to the encoder, enabling the encoder to extract more effective node information to distinguish positive and negative samples:

[0176] (25)

[0177] By learning to distinguish the correlation features produced by GraphSAGE from those generated by the generator, the discriminator forces GraphSAGE to more effectively capture the structural information of the entire graph, thereby enhancing the overall expressiveness of the model.

[0178] This method acquires data on the association of microorganisms and diseases, integrates their topological and functional similarities to construct a heterogeneous network, and utilizes a GraphSAGE-based graph neural network to learn the topological and attribute representations of nodes in the heterogeneous microorganism-disease network. This method also incorporates an adversarial regularized autoencoder graph neural network embedding model to address the inherent limitations of traditional GraphSAGE autoencoders in capturing global information. By utilizing WGAN-GP as the regularizer in the graph neural network, this method mitigates the limitations typically encountered when processing graph data with long-range dependencies or global structural features, significantly improving the model's accuracy in predicting microorganism-disease associations.

[0179] This embodiment also discloses a microorganism-disease association prediction system based on an adversarial regularized graph neural network, which is used to receive an association matrix input containing microorganism-disease pairs to be predicted to predict whether the microorganism-disease pairs to be predicted are associated; the association matrix input includes known associations between the microorganism to be predicted and the disease in the microorganism-disease pairs to be predicted, as well as known associations between the disease to be predicted and the microorganism.

[0180] Specifically, the system includes a similarity acquisition module, a graph network construction module, and a prediction module;

[0181] The similarity acquisition module is used to calculate the GIP core similarity, HIP similarity, and functional similarity of the microorganism to be predicted based on the association matrix input, thereby obtaining a comprehensive similarity matrix of the microorganism to be predicted; and calculate the GIP core similarity, HIP similarity, and functional similarity of the disease to be predicted, thereby obtaining a comprehensive similarity matrix of the disease to be predicted;

[0182] A graph network construction module is used to construct a heterogeneous network of the microorganisms to be predicted and the diseases to be predicted based on the association matrix input, the comprehensive similarity matrix of the microorganisms to be predicted, and the comprehensive similarity matrix of the diseases to be predicted;

[0183] The prediction module is used to output the association prediction result of the microorganism to be predicted and the disease to be predicted using the heterogeneous network of the microorganism to be predicted and the disease to be predicted as input.

[0184] like Figure 3 As shown, the prediction process of the above system for predicting the microorganism-disease pair is as follows:

[0185] receiving an association matrix input containing microorganism-disease pairs to be predicted;

[0186] The association matrix input includes the known associations between the microorganism to be predicted and the disease in the microorganism-disease pair to be predicted, as well as the known associations between the disease to be predicted and the microorganism. As shown in the following table, disease 3 is known to be associated with microorganisms 1 and 3, and microorganism 2 is associated with diseases 1 and 2. The association relationship between disease 3 and microorganism 2 can be predicted by inputting the following association matrix input:

[0187] Table 3 Correlation matrix input

[0188]

[0189] The similarity acquisition module calculates the GIP core similarity, HIP similarity, and functional similarity of the microorganism to be predicted based on the correlation matrix input, and then obtains the comprehensive similarity matrix of the microorganism to be predicted. Taking Table 3 above as an example, the comprehensive similarity matrix of microorganism 2 is obtained;

[0190] Calculate the GIP core similarity, HIP similarity, and functional similarity of the disease to be predicted, and then obtain the comprehensive similarity matrix of the disease. Taking Table 3 above as an example, the comprehensive similarity matrix of disease 3 is obtained;

[0191] The graph network construction module constructs a heterogeneous network of the microorganisms to be predicted and the disease to be predicted based on the association matrix input, the comprehensive similarity matrix of the microorganisms to be predicted, and the comprehensive similarity matrix of the disease to be predicted, that is, constructs a heterogeneous network of disease 3 and microorganism 2. The heterogeneous network graph contains information such as the association between disease 3 and other diseases 3, the association between microorganism 2 and other microorganisms, the association between disease 3 and other microorganisms, and the association between microorganism 2 and other diseases;

[0192] The prediction module outputs the association prediction results of the predicted microorganism-disease pair based on the heterogeneous network of the predicted microorganism and the predicted disease.

[0193] If the prediction result is related, the system output is as shown in Table 4 below. Otherwise, the system output is as shown in Table 5 below:

[0194] Table 4 System output matrix

[0195]

[0196] Table 5 System output matrix

[0197]

[0198] Furthermore, in this scheme, the model uses 5-fold cross-validation to evaluate model performance. The known microbe-disease associations are evenly divided into 5 subsets, one of which is used for testing, and the remaining four are used for training in each iteration. Positive samples consist of all known microbe-disease pairs, while negative samples consist of unknown microbe-disease pairs. During training, negative samples equal to the number of positive samples are randomly selected in each epoch, ensuring that different negative samples are used in different epochs. Five metrics are used to evaluate model performance: area under the ROC curve (AUC), area under the precision-recall curve (AUPR), accuracy, MCC (Matthews correlation coefficient), and F1 score (harmonic mean of precision and recall). Higher values ​​of these metrics indicate better performance.

[0199] In the model, the training process includes 200 epochs, a learning rate of 0.001, the Ada optimizer, and a batch size of 32. Finally, the performance is compared with eight state-of-the-art microbiome-disease association prediction models, including:

[0200] (1) KATAHMDA, a calculation method based on KATZ.

[0201] (2) NTSHMDA, a restart-based random walk model.

[0202] (3) BiRWHMDA, a model based on double random walks.

[0203] (4) NGRHDMA, a recommendation-based approach that combines neighbor-based collaborative filtering with a graph-based scoring method.

[0204] (5) BRWMDA, a random walk-based method.

[0205] (6) WMGHDMA, a metagraph-based computational method.

[0206] (7) GRNMFHMDA, a model based on matrix factorization.

[0207] (8) GATMDA, a deep learning computational method for graph attention networks with inductive matrix completion.

[0208] For fair comparison, the baseline models were run on the HMDAD and Disbiome datasets using their respective default settings.

[0209] like Figure 4 As shown in the figure, on the HMDAD dataset, the AUC of the proposed method is 0.9891±0.0057, AUPR is 0.9902±0.0128, accuracy is 0.8778±0.0021, MCC is 0.7649±0.0022, and F1 score is 0.8675±0.0014, which are the highest indicators among all models.

[0210] like Figure 5 As shown in the figure, on the Disbiome dataset, our proposed method achieved an AUC of 0.9328±0.0072, an AUPR of 0.9233±0.0089, an accuracy of 0.8736±0.0029, an MCC of 0.7471±0.0045, and an F1 score of 0.8736±0.0023, also achieving the highest performance. Therefore, we can see that our proposed method significantly improves the model's predictive performance for microbe-disease associations by first obtaining a comprehensive similarity matrix, then constructing a heterogeneous microbe-disease network, and then combining it with a graph neural network. Furthermore, adversarial regularization is introduced into the graph neural network using an adversarial network.

[0211] The specific embodiments described herein are merely illustrative of the spirit of the present invention. Persons skilled in the art may make various modifications, additions, or substitutions to the described specific embodiments without departing from the spirit of the present invention or exceeding the scope of the appended claims.

Claims

1. A method for predicting the association between microorganisms and diseases based on adversarial regularized graph neural networks, characterized in that: The method includes: S1. Obtain data on the association between microorganisms and diseases and construct a microorganism-disease association matrix; S2. Obtaining a microbial comprehensive similarity matrix and a disease comprehensive similarity matrix based on the microbial-disease association matrix; S3. Construct a microbe-disease heterogeneous network using the microbe-disease association matrix, the microbe comprehensive similarity matrix, and the disease comprehensive similarity matrix; S4. Using a graph neural network to predict associations based on the microbe-disease heterogeneous network; The loss function of the adversarial model including the generator and the discriminator is added as a regular term to the loss function of the graph neural network to perform model training; The graph neural network includes an encoder and a decoder; The generator of the adversarial model outputs a fake feature matrix based on random noise sampled from a predefined distribution; The discriminator of the adversarial model distinguishes the true feature matrix produced by the graph neural network And the fake feature matrix generated by the generator in the adversarial model ; The adversarial model’s loss function is added as a regularization term to the graph neural network’s loss function to feed back the difference between positive and negative samples to the graph neural network during the adversarial model training process: represents the graph neural network loss function, represents the loss function of the adversarial model, Represents the weight of the adversarial model loss term; The loss function of the adversarial model is The calculation is as follows: The distribution of real association data between microorganisms and diseases randomly sampled for each training, Represents the discriminator's distribution from the real data Samples sampled from The expected value of the output; The distribution of microorganism and disease association data generated by the generator, Represents the distribution of data generated by the discriminator from the generator Samples sampled from The expected value of the output; Represents the weight of the gradient penalty term; is the interpolation between the real sample and the generated sample, is the discriminator's interpolation sample The second norm of the gradient of .

2. The method for predicting the association between microorganisms and diseases based on adversarial regularized graph neural networks according to claim 1, characterized in that: In step S1, in the microorganism-disease association matrix, each row / column represents a microorganism, each column / row represents a disease, and an element in the matrix indicates whether there is an association between the microorganism and the disease.

3. The method for predicting the association between microorganisms and diseases based on adversarial regularized graph neural networks according to claim 2, characterized in that: Step S2 specifically includes: According to the microorganism-disease association matrix, the topological structure similarity and functional similarity between microorganisms are obtained respectively, and the topological structure similarity and functional similarity between microorganisms are weightedly combined to obtain the microorganism comprehensive similarity matrix; According to the microorganism-disease association matrix, the topological structure similarity and functional similarity between diseases are obtained respectively, and the topological structure similarity and functional similarity between diseases are weightedly combined to obtain the disease comprehensive similarity matrix.

4. The method for predicting the association between microorganisms and diseases based on adversarial regularized graph neural networks according to claim 3, characterized in that: The topological similarity includes GIP core similarity and HIP similarity. GIP core similarity and HIP similarity measure the similarity between microorganisms and diseases based on the network topology, while functional similarity measures the similarity based on biological characteristics. The microbial comprehensive similarity matrix is ​​the average of the microbial GIP core similarity, HIP similarity, and functional similarity; The disease comprehensive similarity matrix is ​​the average of the disease's GIP core similarity, HIP similarity, and functional similarity.

5. The method for predicting the association between microorganisms and diseases based on adversarial regularized graph neural networks according to claim 4, characterized in that: In step S3, the constructed microorganism-disease heterogeneous network includes a microorganism comprehensive similarity matrix, a disease comprehensive similarity matrix, a microorganism-disease association matrix, and a transposed matrix of the microorganism-disease association matrix.

6. The method for predicting the association between microorganisms and diseases based on adversarial regularized graph neural networks according to claim 5, characterized in that: In step S4, The encoder selects a subset of neighboring nodes for each node in the microorganism-disease heterogeneous network, aggregates the feature vectors of the selected neighboring nodes, and updates the feature vector of the current node by combining the features of the current node and the aggregated neighboring nodes through a nonlinear transformation. The aggregation and updating are repeated, and the topological structure information and attribute information of the node are gradually captured through a multi-layer neural network, ultimately generating a low-dimensional feature vector for the node. The inner product of the low-dimensional feature vector is used as a decoder, association prediction is performed through the low-dimensional node representation, and the cross entropy loss function is used to calculate the difference between the prediction result and the true label.

7. The method for predicting the association between microorganisms and diseases based on adversarial regularized graph neural networks according to claim 6, characterized in that: The encoder construction methods in graph neural networks include: The initial feature vectors of microbial nodes and disease nodes are defined as follows: represents the microorganism / disease node; Representation node Feature vector at the initial layer; is the initial eigenvector of the node, Indicates traversing all nodes in the graph ; For each node , to its neighbor nodes u After sampling, the node representation is obtained , aggregate it: in, Representation node In the Neighborhood aggregate feature vector of the layer; h u k-1 Representation node u exist k- Feature vector of layer 1; node Through nonlinear transformation, it updates its own representation by using its own representation and the aggregated representation of its neighbors. , and get the final encoder output ; And obtain the feature representation of each node based on the encoder ; in, Representation node In the The feature vector of the layer; The feature matrix representing the disease node; Represents the feature matrix of microbial nodes.

8. The method for predicting the association between microorganisms and diseases based on adversarial regularized graph neural networks according to claim 7, characterized in that: The decoder construction methods in graph neural networks include: Based on the matrix Z, its inner product is used as a decoder, and the association prediction is performed through the low-dimensional node representation. The cross entropy loss function is used to calculate the distance between the prediction result and the true label: Represents the association prediction result; represents the activation function; represents the cross entropy loss function; Indicates the The association vector of each node in the original association matrix; Indicates the The association prediction results of nodes; represents the cross entropy distance function.

9. A microorganism-disease association prediction system based on adversarial regularized graph neural network, characterized in that: Used to receive an association matrix input containing the microorganism-disease pairs to be predicted to predict whether the microorganism-disease pairs to be predicted are associated; The association matrix input includes known associations between the microorganism to be predicted and the disease in the microorganism-disease pair to be predicted, as well as known associations between the disease to be predicted and the microorganism; The system includes a similarity acquisition module, a graph network construction module, and a prediction module; A similarity acquisition module is used to calculate the GIP core similarity, HIP similarity, and functional similarity of the microorganism to be predicted based on the correlation matrix input, thereby obtaining a comprehensive similarity matrix of the microorganism to be predicted; Calculate the GIP core similarity, HIP similarity, and functional similarity of the disease to be predicted, and then obtain a comprehensive similarity matrix of the disease to be predicted; A graph network construction module is used to construct a heterogeneous network of the microorganisms to be predicted and the diseases to be predicted based on the association matrix input, the comprehensive similarity matrix of the microorganisms to be predicted, and the comprehensive similarity matrix of the diseases to be predicted; A prediction module, configured to output an association prediction result of a microorganism-disease pair using the heterogeneous network of the microorganism to be predicted and the disease to be predicted as input; The prediction module is constructed by the method described in any one of claims 1 to 8.

Images