A method for predicting microbe-drug associations

Abstract

The application relates to a microorganism-drug correlation prediction method and belongs to the technical field of bioinformatics. The application is aimed at solving the problems of low efficiency and high cost in the existing method for obtaining the microorganism-drug correlation. The application comprises the following steps: obtaining a microorganism-drug correlation matrix A, obtaining a microorganism-drug spectrum kernel similarity matrix M g , g , obtaining a drug-microorganism structure similarity matrix D s , M s , obtaining a microorganism-drug fusion similarity matrix M F , D F , obtaining a matrix X by splicing A, M F and D F , training and fine-tuning a VAE-GAN model by using X to obtain a microorganism-drug nonlinear feature matrix M v , D v , constructing a matrix A' by taking the first k elements of A, decomposing A' to obtain a microorganism-drug linear feature matrix V k , U k , connecting A, D v , M v , V k , U k in series to obtain a final embedding feature T, training an MLP by using T, inputting the final embedding features of the drug and the microorganism to be predicted into the trained MLP model, and obtaining the correlation between the drug and the microorganism. The application is used for predicting new correlations between microorganisms and drugs.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics technology, and in particular to a method for predicting microbial-drug associations. Background Technology

[0002] Microbes play a vital role in the human body, often referred to as the "forgotten" organ. They protect our health, including fighting pathogens, improving metabolism, and boosting immunity. However, when beneficial functions are lost or harmful functions are present in the microbiome, they can have significant impacts on health and disease. The loss of beneficial functions and the activity of harmful microbes can lead to certain diseases. Furthermore, the use of antibiotics during treatment can collaterally damage the microbiome. For example, Clostridium difficile infection is associated with the loss of beneficial microbial function and decreased diversity; some studies have shown that obesity is associated with a significant reduction in Bacteroidetes and an increase in Firmicutes; Firmicutes are also reduced in patients with type 2 diabetes. To treat these diseases, antibiotics are generally used to restore lost beneficial functions and eliminate harmful microbial activity. However, the overuse of antibiotics can lead to antibiotic resistance, not only reducing the effectiveness of antibiotics but also threatening human health. Studies have shown that over 70% of bacteria are resistant to at least one common antibiotic. Therefore, there is an urgent need to identify the association between microbes and drugs in order to minimize harm to human health. At the same time, the association between microbes and drugs can guide the search for new drug combinations and drug repositioning for drug-resistant microbes.

[0003] Beyond their role in regulating disease and health, microorganisms may also be involved in modulating drug activity and toxicity. Some studies have shown that the microbial community plays a crucial role in drug activity and toxicity. For example, certain microorganisms increase digoxin toxicity, decrease morphine clearance, increase the morphine AUC of certain Pseudomonas aeruginosa, increase the area under the curve (AUC) of simvastatin by 221%, alter warfarin activity, and increase irinotecan toxicity. However, our current understanding of the mechanisms linking microorganisms and drugs is insufficient to explain why some drugs exhibit good responses in certain populations while others may cause side effects. Therefore, further research into the interactions between microorganisms and drugs is needed to better understand their roles in the human body and to guide the development of more personalized and effective therapeutic agents.

[0004] Traditional methods involve synthesizing and evaluating the activity of hundreds to thousands of compounds to demonstrate the bioactivity, selectivity, bioavailability, and low toxicity of microbe-drug relationships. However, this approach typically requires several years of experimentation, which is time-consuming and labor-intensive. Therefore, current methods for obtaining microbe-drug relationships are characterized by low efficiency and high cost. Summary of the Invention

[0005] The purpose of this invention is to address the problems of low efficiency and high cost in existing methods for obtaining microbial-drug associations, and to propose a method for predicting microbial-drug associations.

[0006] The specific process of a microbiome-drug association prediction method is as follows:

[0007] Step 1: Obtain a dataset of microbial-drug associations with known relationships and construct a microbial-drug association matrix A;

[0008] Step 2: Obtain the microbial Gaussian interaction spectrum kernel similarity matrix M using the microbial-drug correlation matrix A. g Drug Gaussian Interaction Spectrum Kernel Similarity Matrix D g ;

[0009] Step 3: Obtain the drugs and microorganisms contained in the microbial-drug association dataset, and obtain the structural similarity matrix D between drugs. s Sequence similarity matrix M between microorganisms s ;

[0010] Step 4: Use the matrix M obtained in Step 2 g And the matrix M obtained in step three s Obtain the microbial fusion similarity matrix M F Using the matrix D obtained in step two g And the matrix D obtained in step three s Obtain the drug fusion similarity matrix D F ;

[0011] Step 5: Using the microbial-drug correlation matrix A obtained in Step 1, matrix M... F and D F Concatenate the components to obtain the associated feature matrix X;

[0012] Step 6: Train the VAE-GAN model using matrix X, and then fine-tune the VAE-GAN model to obtain the microbial latent nonlinear feature matrix M. v and the drug's potential nonlinear characteristic matrix D v. ;

[0013] Step 7: Take the first k largest diagonal elements of the microbial-drug correlation matrix A to construct a diagonal matrix A′. Use the singular value decomposition (SVD) method to decompose matrix A′ to obtain the microbial linear characteristic matrix V. k And the linear characteristic matrix U of the drug k ;

[0014] Step 8: Combine the microbial-drug association matrix A obtained in Step 1 with the matrix D obtained in Step 6. v and M v. Step 7 Obtained V k and U k The final embedding features are obtained by concatenation, and the final embedding features are used as the second training set. The MLP model is trained using the second training set to obtain the trained MLP model.

[0015] Step 9: Obtain the drugs and microorganisms with the predicted association. Obtain the potential nonlinear feature matrix and linear feature matrix of the drugs with the predicted association. Obtain the potential nonlinear feature matrix and linear feature matrix of the microorganisms with the predicted association. Associate the microorganism-drug association matrix A, the potential nonlinear feature matrix and linear feature matrix of the drugs with the predicted association, and the potential nonlinear feature matrix and linear feature matrix of the microorganisms with the predicted association to obtain the final embedding features of the drugs and microorganisms with the predicted association. Input the final embedding features of the drugs and microorganisms with the predicted association into the trained MLP model to obtain the association between the drugs and microorganisms with the predicted association.

[0016] Further, the step one of obtaining a microbial-drug association dataset with known associations and constructing a microbial-drug association matrix A specifically involves:

[0017]

[0018] in, A(i, j) is the element in the i-th row and j-th column of matrix A, N d N represents the number of drug types. m It refers to the number of microbial species, d i It is the i-th drug, m j It is the j-th microorganism.

[0019] Furthermore, in step two, the microbial Gaussian interaction spectrum kernel similarity matrix M is obtained using the microbial-drug correlation matrix A. g Drug Gaussian Interaction Spectrum Kernel Similarity Matrix D g Specifically:

[0020]

[0021]

[0022]

[0023]

[0024] Among them, M g (m j m j′ ) represents the Gaussian interaction spectrum kernel similarity between the j-th microorganism and the j′-th microorganism, D g (d i d i′ ) represents the Gaussian interaction spectrum kernel similarity between the i-th drug and the i′-th drug, η m It is the normalized microbial nuclear broadband, η d It is the normalized drug nuclear broadband, η′ m It is the nuclear broadband of primitive microorganisms, η′ d It is the original drug nuclear broadband. It is the vector of the j-th column of matrix A. It is the j′ column vector of matrix A. It is the vector in the i-th row of matrix A. It is the vector of the i′ row of matrix A.

[0025] Furthermore, in step three, the structural similarity matrix D between drugs is obtained. s Sequence similarity matrix M between microorganisms s Specifically:

[0026] Step 3: 1. Obtain the drug chemical formula and input it into the SMILES system to obtain the structural similarity matrix D between drugs. s :

[0027]

[0028] Among them, D s (d i d i′ ) is drug d i With drug d i′ Structural similarity between them;

[0029] Step 3.2: Retrieve the similarity M between microbial protein sequences from the STRING v11 database. s (m j m j′ This allows us to obtain the sequence similarity matrix M among microorganisms. s .

[0030] Furthermore, in step four, the matrix M obtained in step two is used... g And the matrix M obtained in step three sObtain the microbial fusion similarity matrix M F Using the matrix D obtained in step two g And the matrix D obtained in step three s Obtain the drug fusion similarity matrix D F Specifically:

[0031]

[0032]

[0033] Among them, D F (d i d i′ ) is drug d i With drug d i′ Fusion similarity, M F (m j m j′ ) is a microorganism m j With microorganisms m j′ The similarity of fusion.

[0034] Furthermore, in step five, the microbial-drug correlation matrix A obtained in step one is used to transform matrix M. F and D F Concatenate the components to obtain the associated feature matrix X, specifically:

[0035]

[0036] in, It is D F The i-th row of drug characteristics, It is M F The microbial characteristics of the jth row, for and By combining the correlation matrix A, X i′ It is the i′-th row of matrix X.

[0037] Furthermore, in step six, the VAE-GAN model is trained using matrix X, and then fine-tuned to obtain the microbial latent nonlinear feature matrix M. v and the drug's potential nonlinear characteristic matrix D v. Specifically:

[0038] Step 61: Establish the VAE-GAN model;

[0039] The VAE-GAN model includes: a generator and a discriminator;

[0040] The generator is a variational autoencoder (VAE), comprising: an encoding layer and a decoding layer;

[0041] The encoder layer uses matrix X to perform variational operations to obtain the intermediate layer hidden vector z, and then sends the intermediate layer hidden vector z to the decoding layer.

[0042] The process of obtaining the intermediate layer latent vector z using variational operations on matrix X is as follows:

[0043] z = μ + e*exp(ε)

[0044] ε = logσ 2

[0045] μ = Encoder(X) = f1(X)

[0046] σ 2 =Encoder(X)=f2(X)

[0047] Where μ is the mean of the data in matrix X, ε is an intermediate variable, and σ 2 f1() is the variance of the data in matrix X, f2() is the forward propagation function of the mean in the coding layer, f2() is the forward propagation function of the variance in the coding layer, e is the logarithm of the variance, and e follows a standard Gaussian distribution N(0,1);

[0048] The spatial distribution of the intermediate layer latent vector z is as follows:

[0049]

[0050] Where p(z|X) refers to the conditional probability distribution of the latent vector z, P(X) is the distribution of data in matrix X, and p(z) is the spatial distribution of the latent vectors in the intermediate layer.

[0051] The decoding layer reconstructs matrix X using the intermediate layer latent vector z to obtain the reconstructed matrix X′, and sends X′ to the discriminator;

[0052] The process of reconstructing matrix X using the intermediate layer latent vector z to obtain the reconstructed matrix X′ is as follows:

[0053] X′=Decoder(z)=f3(z)

[0054] Where f3() is the forward propagation function from the encoding layer to the decoding layer;

[0055] The discriminator is a WGAN-GP discriminator; the discriminator is used to determine the probability estimate of the real data for X′, specifically:

[0056] V P =Dis({X,X′},{C,C′})

[0057] Where Dis() represents the discriminator's prediction of the input data, V PIt is a probability estimate that the sample is real data, C is the real label of X, C′ is the label of the data generated by the generator, and X′ is the data generated by the generator;

[0058] Step 62: Use matrix X to train the VAE-GAN model and obtain the trained VAE-GAN model;

[0059] The loss function for training the encoding layer is as follows:

[0060]

[0061] Among them, Loss KL It is the loss function of the coding layer;

[0062] The loss function for training the decoding layer is as follows:

[0063]

[0064] Among them, Loss recon It is the loss function of the decoding layer;

[0065] The loss function for training the discriminator is as follows:

[0066]

[0067] Where λ is the penalty coefficient. This represents data points obtained by uniformly sampling between the real data distribution and the generated data distribution. Indicates the discriminator pair The gradient;

[0068] Step 63: Use the data in matrix X as positive samples in the first training set, randomly obtain the same number of negative samples as the positive samples, and concatenate the positive and negative samples in pairs to form the first training set. Use the first training set to retrain the generator in the trained VAE-GAN model, and use the retrained VAE-GAN model as the fine-tuned generator.

[0069] Step 64: Input a random noise matrix of the same size as matrix X into the fine-tuned VAE-GAN model to obtain the microbial latent nonlinear feature matrix M. v and the drug's potential nonlinear characteristic matrix D v. ;

[0070] The random noise matrix consists of random numbers uniformly distributed from -1 to 1.

[0071] Furthermore, the generator in the trained VAE-GAN model is retrained using the first training set with the following loss function:

[0072]

[0073] X f =Connecting({X known X unknown},{C1,C0})

[0074] Among them, G(X) known G(X) represents the generator's regeneration of known associated data; unknown ) represents the generator's regeneration of unknown related data, X f This is the first training set, X known X is a positive sample in the first training set. unknown C1 is the label of the positive sample in the first training set, and C0 is the label of the negative sample in the first training set.

[0075] Furthermore, in step seven, the first k largest diagonal elements of the microbial-drug correlation matrix A are used to construct a diagonal matrix A′, and the singular value decomposition (SVD) method is used to decompose matrix A′ to obtain the microbial linear feature matrix V. k And the linear characteristic matrix U of the drug k Specifically:

[0076]

[0077] Among them, U k It is the linear characteristic matrix of the drug, V k It is the linear characteristic matrix of microorganisms, ∑ k Let A be the singular value matrix of A′, and k be the matrix factorization dimension.

[0078] Furthermore, in step eight, the microbial-drug correlation matrix A obtained in step one and the matrix D obtained in step six are combined. v and M v. Step 7 Obtained V k and U k The final embedding features are obtained by concatenation, and these features are used as the second training set to train the MLP model, resulting in a trained MLP model.

[0079] Step 81: Combine matrices A and U k With D v The final drug embedding feature matrix Z is obtained by concatenation. d , A, V k With M v. The final embedding feature matrix Z of microorganisms is obtained by concatenation. m Using Z d and Z m Obtain the final embedding feature matrix T:

[0080]

[0081] Z d =connecting(U k A, D v )

[0082] Z m =connecting(V k A T M v )

[0083] Among them, Z d It is the final embedding feature matrix of the drug, Z m It is the final embedded feature matrix of microorganisms. It is Z d The i-th row in It is Z m In the j-th row, T ij It is drug d i and microorganisms m j The final embedded features after association;

[0084] Step 82: Use the final embedded feature matrix T as the second training set, train the MLP model using the second training set, and use the trained MLP as the microbiome-drug association prediction model.

[0085] The MLP model is trained using the second training set, specifically as follows:

[0086] T l =ReLU(T) l-1 w l +b l )

[0087]

[0088] Where l is the number of hidden layers; T l It is the output of the l-th hidden layer, w l b is the learnable parameter matrix of the l-th hidden layer. l It is the bias of the l-th hidden layer. It is the microbiome-drug association prediction score;

[0089] The following loss function is used to train the MLP model:

[0090]

[0091] Among them, y ij These are the labels of the positive samples in the second training set. 1-y represents the probability of predicting a positive sample. ijThese are the labels of the negative samples in the second training set. y represents the probability of a predicted negative sample, N is the total number of microbe-drug association pairs in the second training set, and y + Let y represent the set of positive sample labels in the second training set. - It is the set of negative sample labels in the second training set.

[0092] The beneficial effects of this invention are as follows:

[0093] This invention employs a method that fuses microbial Gaussian interaction spectrum kernel similarity with microbial sequence similarity to obtain a microbial fusion similarity matrix; and a method that fuses drug structural similarity with drug Gaussian interaction spectrum kernel similarity to obtain a drug fusion similarity matrix. The fusion of these two similarities fills in potential association information. The VAE of this invention achieves transformation mapping by constraining the latent variable z to follow a standard normal distribution (KL divergence) and reconstructing the data, thereby extracting latent features from the data's latent distribution. The proposed VAE-GAN model combines the advantages of VAE and GAN, enabling better control over the latent distribution of data while stabilizing model training, making latent feature extraction more efficient and thus improving the efficiency of obtaining microbial-drug associations. This invention uses VAE-GAN to extract nonlinear features and singular value decomposition to obtain linear features. The linear and nonlinear features are concatenated with the association matrix and embedded into a multilayer perceptron with two hidden layers. This concatenates latent linear and nonlinear features without losing original information, further revealing unknown associations hidden beneath the data. This invention eliminates the need for extensive experiments to obtain associations between microorganisms and drugs, reducing experimental costs. Attached Figure Description

[0094] Figure 1 A general flowchart for constructing microbe-drug associations;

[0095] Figure 2 A detailed process diagram for constructing microbe-drug associations;

[0096] Figure 3 A diagram showing the detailed process of building the VAE-GAN method model;

[0097] Figure 4 This is a diagram illustrating the latent space sampling process of the data in the VAE model.

[0098] Figure 5 Detailed diagram of VAE model reparameterization;

[0099] Figure 6 This is a schematic diagram of SVD matrix decomposition.

[0100] Figure 7This is a schematic diagram of data stitching before training the VGMDA model.

[0101] Figure 8 The ROC plots for the VGMDA model at 2-fold, 5-fold, and 10-fold folds are shown. Detailed Implementation

[0102] Specific implementation method one: as follows Figure 1-2 As shown, the specific process of a microbial-drug association prediction method in this embodiment is as follows:

[0103] Step 1: Obtain known microbial-drug association datasets and construct a microbial-drug association matrix A, specifically as follows:

[0104]

[0105] in, A(i, j) is the element in the i-th row and j-th column of matrix A, N d N represents the number of drug types. m It refers to the number of microbial species, d i It is the i-th drug, m j It is the j-th microorganism;

[0106] Step 2: Obtain the microbial Gaussian interaction spectrum kernel similarity matrix M using the microbial-drug correlation matrix A. g Drug Gaussian Interaction Spectrum Kernel Similarity Matrix D g Specifically:

[0107]

[0108]

[0109]

[0110]

[0111] Among them, M g (m j m j′ ) represents the Gaussian interaction spectrum kernel similarity between the j-th microorganism and the j′-th microorganism, D g (dx,d i′ ) represents the Gaussian interaction spectrum kernel similarity between the i-th drug and the i′-th drug, η m It is the normalized microbial nuclear broadband, η d It is the normalized drug nuclear broadband, η′ m It is the nuclear broadband of primitive microorganisms, η′ d It is the original drug nuclear broadband. It is the j-th column vector of matrix A (microorganism m) jGaussian interaction spectrum), It is the j′ column vector of matrix A. It is the vector in the i-th row of matrix A (drug m) i Gaussian interaction spectrum), It is the vector of the i'th row of matrix A;

[0112] In this step, the Gaussian kernel function is a commonly used similarity metric. It determines the degree of similarity between two vectors by calculating the Euclidean distance between them, η. m The Gaussian kernel controls the smoothness of similarity calculations. The larger the value, the greater the local influence range of the Gaussian kernel function; if the value is too small, it is prone to overfitting in classification tasks.

[0113] Step 3: Use SMLES to obtain the structural similarity matrix D between drugs. s Microbial protein sequences were retrieved using the STRING v11 database to obtain the microbial sequence similarity matrix M. s Specifically:

[0114] Step 3: 1. Obtain the drug chemical formula and input it into SMILES (Simplified Molecular Input Linear System) to obtain the structural similarity matrix D between drugs. s :

[0115]

[0116] Among them, D s (d i d i′ ) is drug d i With drug d i′ Structural similarity between them;

[0117] In this step, the SMILES system obtains the structural similarity matrix between drugs as follows: First, it collects the standard SMILES (Simplified Molecular Language) for all drugs from the DrugBank database and calculates the chemical fingerprints of all drugs using the Chemical Development Kit tool. Finally, it measures the structural similarity between drugs based on the Tanimoto scores of the drug chemical fingerprints. The SMILES system is a string representation for chemical molecular structures. It can accurately describe the chemical structure of all drug compounds using a single line of symbols.

[0118] Step 3.2: Retrieve the similarity M between microbial protein sequences from the STRING v11 database (protein-protein interaction network https: / / string-db.org). s (m j mj′ This allows us to obtain the sequence similarity matrix M of the microorganisms. s ;

[0119] Among them, M s (m j m j′ ) is a microorganism m j With microorganisms m j′ Similarity between them.

[0120] Step 4: Use the matrix M obtained in Step 2 g And the matrix M obtained in step three s Obtain the microbial fusion similarity matrix M F Using the matrix D obtained in step two g And the matrix D obtained in step three s Obtain the drug fusion similarity matrix D F Specifically:

[0121]

[0122]

[0123] Among them, D F (d i d i′ ) is drug d i With drug d i′ Fusion similarity, M F (m j m j′ ) is a microorganism m j With microorganisms m j′ The similarity of fusion.

[0124] Step 5: Using the microbial-drug correlation matrix A obtained in Step 1, matrix M... F and D F Matrix concatenation yields the associated feature matrix X, specifically:

[0125]

[0126] in, It is D F The i-th row of drug characteristics, It is M F The microbial characteristics of the jth row, for and By combining the correlation matrix A, X i′ It is the i′-th row of matrix X;

[0127] Step 6: Train the VAE-GAN model using matrix X, and then fine-tune the VAE-GAN model to obtain the microbial latent nonlinear feature matrix M. v and the drug's potential nonlinear characteristic matrix D v Specifically:

[0128] Step 61: Establish the VAE-GAN model;

[0129] The VAE-GAN model includes: a generator and a discriminator;

[0130] The generator is a variational autoencoder (VAE), comprising an encoding layer and a decoding layer; the overall process of the variational autoencoder is as follows: Figure 4 As shown;

[0131] The encoder layer uses matrix X to perform variational operations to obtain the intermediate layer hidden vector z, and then sends the intermediate layer hidden vector z to the decoding layer.

[0132] The method of obtaining the intermediate layer hidden vector z using matrix X is as follows:

[0133] z = μ + e*exp(ε)

[0134] ε = logσ 2

[0135] μ = Encoder(X) = f1(X)

[0136] σ 2 =Encoder(X)=f2(X)

[0137] Where μ is the mean of the data in matrix X, ε is an intermediate variable, and σ 2 f1() is the variance of the data in matrix X, f2() is the forward propagation function of the mean in the coding layer, f2() is the forward propagation function of the variance in the coding layer, e is the logarithm of the variance, and e follows a standard Gaussian distribution N(0,1);

[0138] The spatial distribution of the intermediate layer latent vector z is as follows:

[0139]

[0140] Where p(z|X) refers to the conditional probability distribution (posterior probability) of the latent vector z; P(X) is the distribution of data in matrix X, and p(z) is the spatial distribution of the latent vectors in the intermediate layer;

[0141] The decoding layer reconstructs matrix X using the intermediate layer latent vector z to obtain the reconstructed matrix X′, and sends the reconstructed matrix X′ to the discriminator;

[0142] The process of reconstructing matrix X′ using the intermediate layer latent vector z to obtain the reconstructed matrix X′ is as follows:

[0143] X′=Decoder(z)=f3(z)

[0144] Where f3() is the complete forward propagation function from the encoding layer to the decoding layer;

[0145] The discriminator is a WGAN-GP discriminator; the discriminator is used to determine the probability estimate of the real data for X′, specifically:

[0146] V P =Dis({X,X′},{C,C′})

[0147] Where Dis() represents the discriminator's prediction of the input data; V P C is the predicted value, ranging from 0 to 1, representing the probability estimate that the input sample is the real data; C is the true label of X, C′ is the label of the data generated by the generator, and X′ is the data generated by the generator.

[0148] Step 62: Use matrix X to train the VAE-GAN model and obtain the trained VAE-GAN model;

[0149] The loss function for the encoding layer is the KL divergence, which provides a metric for how well two probability distributions match. The smaller the difference between the two distributions, the smaller the KL divergence. The loss function for training the encoding layer is as follows:

[0150]

[0151] Among them, Loss KL It is the loss function of the coding layer;

[0152] To compare the differences between the generated samples and the real samples, the mean squared error loss function between the real data X and the reconstructed data X′ is used to train the loss function of the decoding layer, specifically:

[0153]

[0154] Among them, Loss recon It is the loss function of the decoding layer;

[0155] The loss function for training the discriminator is obtained as follows:

[0156] The discriminator uses binary cross-entropy as its loss function, first enabling it to distinguish between real and fake samples:

[0157] Loss D1=-(log Dis(X, C)+log(1-Dis(X′, C′)))

[0158] Then, make the generated samples as confusing as possible to the discriminator:

[0159] Loss D2 = -log Dis(X′, C)

[0160] Therefore, the total loss of the discriminator is:

[0161] Loss D =Loss D1 +Loss D2

[0162] Among them, Loss D It is the loss function of the discriminator, Loss D1 Loss D2 It is an intermediate variable;

[0163] Because generative adversarial networks (GANs) are difficult to train and have stringent convergence conditions, they are prone to problems such as vanishing gradients and mode collapse. Furthermore, the inclusion of variational autoencoders makes training even more challenging. Therefore, we used an improved GAN network, WGAN-GP, which introduces a gradient penalty to avoid these problems. The improved discriminator's loss is no longer logarithmic, and a penalty term is introduced. Thus, the discriminator's loss function is modified as follows:

[0164]

[0165] Where λ is the penalty coefficient, λ = 10. It is an X distribution X ~ P r Distribution X′ ~ P g Distribution data obtained by uniform sampling along a straight line, i.e. This represents data points uniformly sampled from the real data distribution and the generated data distribution; where This indicates that the discriminator is a pair. The output, Indicates the discriminator pair The gradient of Pr represents the true data distribution, and Pg represents the generated data distribution. This is because the optimal annotator follows a straight line with unit gradient norm between samples coupled from Pr and Pg;

[0166] Step 63: Use the data in matrix X as positive samples in the first training set. Randomly obtain the same number of negative samples as the positive samples. Concatenate the positive and negative samples pairwise to form the first training set. Use the first training set to fine-tune the generator, specifically as follows:

[0167] X f=Connecting({X known X unknown},{C1,C0})

[0168] Among them, X f The input data for model fine-tuning is the first training set, X. known X is the known correlated feature data, i.e., the positive samples in the first training set. unknown C1 is the label of the known associated feature data, and C0 is the label of the unknown associated feature data.

[0169] The fine-tuned generator loss function is:

[0170]

[0171] Among them, G(X) known G(X) represents the generator's regeneration of known associated data; unknown ) indicates that the generator regenerates unknown related data, further optimizing the generator.

[0172] Step 64: Input a random noise matrix of the same size as matrix X into the fine-tuned VAE-GAN model to obtain the microbial latent nonlinear feature matrix M. v and the drug's potential nonlinear characteristic matrix D v. ;

[0173] The random noise matrix consists of random numbers uniformly distributed from -1 to 1.

[0174] In this step, a traditional autoencoder (AE) consists of two parts: an encoder and a decoder. The encoder process compresses the original data into a low-dimensional vector; the decoder then restores the original data from the low-dimensional vector. However, this approach is more like rote memorization. When we want the decoder to generate some data that has never appeared in the dataset before, the AE model is powerless because the latent variables of a traditional autoencoder are fixed and invisible.

[0175] Variational autoencoders (VAEs) address the problem of fixed and invisible latent variables in traditional autoencoders. During encoder training, a constraint is added to ensure that the generated latent vector *z* largely follows a unit Gaussian distribution (obtaining the mean and variance of *z*). *z* is then sampled from this Gaussian distribution and used by the decoder to generate new data. VAEs are unsupervised generative network structures based on Variational Bayes (VB) inference and incorporating features of deep learning and statistics.

[0176] Generative Adversarial Networks (GANs) are generative models that learn by having two neural networks compete against each other. GANs can learn generative tasks without using labeled data. A GAN consists of a generator and a discriminator. The generator randomly samples from the latent space as input, and its output should mimic real samples in the training set as closely as possible. The discriminator takes real samples or the generator's output as input, and its goal is to distinguish the generator's output from real samples as much as possible. The generator and discriminator compete against each other and continuously learn, ultimately making it impossible for the discriminator to determine whether the generator's output is real. Combining Variational Autoencoders (VAEs) with GANs can better obtain the latent distribution of data and extract latent features. We use a VAE as the generator in the GAN network, and capture the nonlinear features of microorganisms and drugs by obtaining the latent feature Gaussian distribution z of the association feature matrix X. The general structure of the VAE-GAN model is as follows: Figure 3 As shown. In this step, directly sampling from a Gaussian distribution would prevent backpropagation, thus making network training impossible, as... Figure 5 As shown. Therefore, we need to reparameterize the mean and variance by first sampling from a standard Gaussian distribution, and then multiplying by the standard deviation plus the mean to obtain the latent vector z.

[0177] Step 7: Take the first k largest diagonal elements of the microbial-drug correlation matrix A to construct a diagonal matrix A′. Use the singular value decomposition (SVD) method to decompose matrix A′ to obtain the microbial linear characteristic matrix V. k And the linear characteristic matrix U of the drug k Specifically:

[0178]

[0179] Among them, U k It is the linear characteristic matrix of the drug, V k It is the linear characteristic matrix of microorganisms, ∑ k Let A' be the singular value matrix, ∑ k All elements except those on the main diagonal are 0, and k is the matrix factorization dimension.

[0180] This step is as follows: Figure 6 As shown, we only focus on the first k singular values ​​of A. That is, by retaining the first k largest diagonal elements, we construct a k*k diagonal matrix, resulting in a new diagonal matrix U. k Accordingly, the first k columns are taken to form N. d *k-order matrix, for V kSimplify in the same way to obtain N. m If we have a k-order matrix, then the reconstructed matrix A′ approximates the original matrix A, thus achieving data dimensionality reduction.

[0181] Step 8: Combine the microbial-drug association matrix A obtained in Step 1 with the matrix D obtained in Step 6. v and M v. Step 7 Obtained V k and U k By concatenating the associated pairs, a training set is obtained. The MLP model is then trained using this training set to obtain the trained MLP model. Specifically:

[0182] Step 81, as follows Figure 7 As shown, the microbial-drug association matrix A obtained in step one and the matrix D obtained in step six are used to... v and M v Step 7 Obtained V k and U k By concatenating the associated pairs, the final embedded feature matrix T is obtained, specifically:

[0183]

[0184] Z d =connecting(U k A, D v )

[0185] Z m =connecting(V k A T M v )

[0186] Among them, Z d It is the final embedding feature matrix of the drug, Z m It is the final embedded feature matrix of microorganisms. It is Z d The i-th row in It is Z m In the j-th row, T ij It is drug d i and microorganisms m j The final embedded features after association, T∈R (1152 *2)×(691+206) ; 1152*2 indicates that there are 1152 pairs of positive and negative samples; 691 indicates the microbial embedding dimension; 206 indicates the drug embedding dimension;

[0187] Step 82: Use the final embedded feature matrix T as the second training set, train the MLP model using the second training set, and use the trained MLP as the microbial-drug association prediction model.

[0188] The second training set is fed into the MLP for learning and training to predict unknown microorganisms-drugs, as follows:

[0189] T l =ReLU(T) l-1 w l +b l )

[0190]

[0191] Where l = 2 represents the number of hidden layers; T l This represents the output of the l-th hidden layer; w l b represents the learnable parameter matrix of the l-th hidden layer; l This indicates the bias of the l-th hidden layer; Microbe-drug association prediction score;

[0192] The MLP model is trained using the Adam optimizer and by minimizing the error of the binary cross-entropy loss function.

[0193]

[0194] Among them, y ij These are the labels of the positive samples in the second training set. 1-y represents the probability of predicting a positive sample. ij These are the labels of the negative samples in the second training set. y represents the probability of a predicted negative sample, N is the total number of microbe-drug association pairs in the second training set, and y + Let y represent the set of positive sample labels in the second training set. - It is the set of negative example sample labels in the second training set;

[0195] Positive examples represent known relationships, while negative examples represent unknown relationships.

[0196] Step 9: Obtain the drugs and microorganisms with the predicted association. Obtain the potential nonlinear feature matrix and linear feature matrix of the drugs with the predicted association. Obtain the potential nonlinear feature matrix and linear feature matrix of the microorganisms with the predicted association. Associate the microorganism-drug association matrix A, the potential nonlinear feature matrix and linear feature matrix of the drugs with the predicted association, and the potential nonlinear feature matrix and linear feature matrix of the microorganisms with the predicted association to obtain the final embedding features of the drugs and microorganisms with the predicted association. Input the final embedding features of the drugs and microorganisms with the predicted association into the trained MLP model to obtain the association between the drugs and microorganisms with the predicted association.

[0197] Example: To verify the beneficial effects of the present invention, the following experiments were conducted:

[0198] This embodiment uses 2-, 5-, and 10-fold cross-validation to verify the performance of the proposed evaluation model VGMDA. The ROC graphs obtained based on 2-CV, 5-CV, and 10-CV are shown below. Figure 8 As shown in Table 1, the AUC values ​​of 2-CV, 5-CV, 10-CV, and other models are compared.

[0199] Table 1. Comparison of AUC values ​​of the VGMDA model and other models under 2, 5, and 10-fold cross-validation on the same dataset.

[0200]

[0201] In the prediction results, this invention proved the top 15 drugs associated with Escherichia coli and the top 15 microorganisms associated with aspirin through relevant literature. The verification results are shown in Tables 2 and 3.

[0202] Table 2. Top 15 drugs associated with E. coli

[0203]

[0204] Table 3. Top 15 microorganisms associated with aspirin

[0205]

[0206]

[0207] This invention, based on known microorganism-drug relationships, first derives their correlation matrix A, and then calculates the Gaussian interaction spectrum kernel matrix D between the drug and the microorganism. g M g The drug structure similarity matrix D was obtained using the Chemical Development Kit and STRING v11. s Microbial sequence similarity matrix M s This approach enriches known association information by integrating Gaussian similarity matrices and structural similarity (sequence similarity) matrices.

[0208] Traditional VAEs and GANs are used to generate data, and combining them can more effectively obtain the latent distribution of data. In this invention, however, they are not used to generate data, but rather to learn the latent distribution of data, thereby obtaining the latent nonlinear features of the data. On the other hand, this invention incorporates SVD technology to obtain the latent linear features of microorganisms and drugs. The linear and nonlinear data of microorganisms and drugs are concatenated with the correlation matrix A to obtain the final data for training our MLP.

[0209] Through performance evaluation, the AUC values ​​of the VGMDA model under 2-, 5-, and 10-fold cross-validation all reached above 0.92, indicating that our model has certain application value.

Claims

1. A method for predicting microbial-drug associations, characterized in that... The specific process of the method is as follows: Step 1: Obtain a dataset of microbial-drug associations with known relationships and construct a microbial-drug association matrix A; Step 2: Obtain the microbial Gaussian interaction spectrum kernel similarity matrix M using the microbial-drug correlation matrix A. g Drug Gaussian Interaction Spectrum Kernel Similarity Matrix D g ; Step 3: Obtain the drugs and microorganisms contained in the microbial-drug association dataset, and obtain the structural similarity matrix D between drugs. s Sequence similarity matrix M between microorganisms s ; Step 4: Use the matrix M obtained in Step 2 g And the matrix M obtained in step three s Obtain the microbial fusion similarity matrix M F Using the matrix D obtained in step two g And the matrix D obtained in step three s Obtain the drug fusion similarity matrix D F ; Step 5: Using the microbial-drug correlation matrix a obtained in Step 1, matrix M... F Sum matrix D F Concatenate the components to obtain the associated feature matrix X; Step 6: Train the VAE-GAN model using matrix X, and then fine-tune the VAE-GAN model to obtain the microbial latent nonlinear feature matrix M. v and the drug's potential nonlinear characteristic matrix D v. ; Step 7: Take the first k largest diagonal elements of the microbial-drug correlation matrix a to construct a diagonal matrix a′. Use the singular value decomposition (SVD) method to decompose matrix a′ to obtain the microbial linear characteristic matrix V. k And the linear characteristic matrix U of the drug k ; Step 8: Combine the microbial-drug association matrix A obtained in Step 1 with the matrix D obtained in Step 6. v and M v. Step 7 Obtained V k and U k The final embedding features are obtained by concatenation, and the final embedding features are used as the second training set. The MLP model is trained using the second training set to obtain the trained MLP model. Step 9: Obtain the drugs and microorganisms with the predicted association. Obtain the potential nonlinear feature matrix and linear feature matrix of the drugs with the predicted association. Obtain the potential nonlinear feature matrix and linear feature matrix of the microorganisms with the predicted association. Associate the microorganism-drug association matrix A, the potential nonlinear feature matrix and linear feature matrix of the drugs with the predicted association, and the potential nonlinear feature matrix and linear feature matrix of the microorganisms with the predicted association to obtain the final embedding features of the drugs and microorganisms with the predicted association. Input the final embedding features of the drugs and microorganisms with the predicted association into the trained MLP model to obtain the association between the drugs and microorganisms with the predicted association.

2. The method for predicting microbial-drug associations according to claim 1, characterized in that: The step one, obtaining a microbial-drug association dataset with known associations and constructing a microbial-drug association matrix A, specifically involves: in, A(i,j) is the element in the i-th row and j-th column of matrix A, N d It refers to the number of drug types, N. m It refers to the number of microbial species, d i It is the i-th drug, m j It is the j-th microorganism.

3. The method for predicting microbial-drug associations according to claim 2, characterized in that: In step two, the microbial Gaussian interaction spectrum kernel similarity matrix M is obtained using the microbial-drug correlation matrix A. g Drug Gaussian Interaction Spectrum Kernel Similarity Matrix D g Specifically: Among them, M g (m j m j′ ) represents the Gaussian interaction spectrum kernel similarity between the j-th microorganism and the j′-th microorganism, D g (d i d i′ ) represents the Gaussian interaction spectrum kernel similarity between the i-th drug and the i′-th drug, η m It is the normalized microbial nuclear broadband, η d It is the normalized drug nuclear broadband, η′ m It is the nuclear broadband of primitive microorganisms, η′ d It is the original drug nuclear broadband. It is the vector of the j-th column of matrix a. It is the j′ column vector of matrix A. It is the vector in the i-th row of matrix A. It is the vector of the i′ row of matrix A.

4. The method for predicting microbial-drug associations according to claim 3, characterized in that: The step three involves obtaining the inter-drug structural similarity matrix D. s Sequence similarity matrix M between microorganisms s Specifically: Step 3:

1. Obtain the drug chemical formula and input it into the SMILES system to obtain the structural similarity matrix D between drugs. s : Among them, D s (d i d i' ) is drug d i With drug d i' Structural similarity between them; Step 3.2: Retrieve the similarity M between microbial protein sequences from the STRING v11 database. s (m j m j′ This allows us to obtain the sequence similarity matrix M among microorganisms. s .

5. The method for predicting microbial-drug associations according to claim 4, characterized in that: The matrix M obtained in step two is used in step four. g And the matrix M obtained in step three s Obtain the microbial fusion similarity matrix M F Using the matrix D obtained in step two g And the matrix D obtained in step three s Obtain the drug fusion similarity matrix D F Specifically: Among them, D F (d i d i′ ) is drug d i With drug d i′ Fusion similarity, M F (m j m j′ ) is a microorganism m j With microorganisms m j′ The similarity of fusion.

6. The method for predicting microbial-drug associations according to claim 5, characterized in that: In step five, the microbial-drug correlation matrix A obtained in step one is used to transform matrix M. F and D F Concatenate the components to obtain the associated feature matrix X, specifically: in, It is D F The i-th row of drug characteristics, It is M F The microbial characteristics of the jth row, for and By combining the correlation matrix A, X i′ It is the i′-th row of matrix X.

7. The method for predicting microbial-drug associations according to claim 6, characterized in that: In step six, the VAE-GAN model is trained using matrix X, and then fine-tuned to obtain the microbial latent nonlinear feature matrix M. v and the drug's potential nonlinear characteristic matrix D v Specifically: Step 61: Establish the VAE-GAN model; The VAE-GAN model includes: a generator and a discriminator; The generator is a variational autoencoder (VAE), comprising: an encoding layer and a decoding layer; The encoder layer uses matrix X to perform variational operations to obtain the intermediate layer hidden vector z, and then sends the intermediate layer hidden vector z to the decoding layer. The process of obtaining the intermediate layer latent vector z using variational operations on matrix X is as follows: z = μ + e*exp(ε) ε=logσ 2 μ = Encoder(X) = f1(X) σ 2 =Encoder(X)=f2(X) Where μ is the mean of the data in matrix X, ε is an intermediate variable, and σ 2 f1() is the variance of the data in matrix X, f2() is the forward propagation function of the mean in the coding layer, f2() is the forward propagation function of the variance in the coding layer, e is the logarithm of the variance, and e follows a standard Gaussian distribution N(0,1); The spatial distribution of the intermediate layer latent vector z is as follows: Where p(z|X) refers to the conditional probability distribution of the latent vector z, P(X) is the distribution of data in matrix X, and p(z) is the spatial distribution of the latent vectors in the intermediate layer. The decoding layer reconstructs matrix X using the intermediate layer hidden vector z to obtain the reconstructed matrix X′, and sends X′ to the discriminator; The process of reconstructing matrix X using the intermediate layer latent vector z to obtain the reconstructed matrix X′ is as follows: X′=Decoder(z)=f3(z) Where f3() is the forward propagation function from the encoding layer to the decoding layer; The discriminator is a WGAN-GP discriminator; the discriminator is used to determine the probability estimate of the real data for X′, specifically: V P =Dis({X,X′},{C,C′}) Where Dis() represents the discriminator's prediction of the input data, V P It is a probability estimate that the sample is real data, C is the real label of X, C′ is the label of the data generated by the generator, and X′ is the data generated by the generator; Step 62: Use matrix X to train the VAE-GAN model and obtain the trained VAE-GAN model; The loss function for training the encoding layer is as follows: Among them, Loss KL It is the loss function of the coding layer; The loss function for training the decoding layer is as follows: Among them, Loss recon It is the loss function of the decoding layer; The loss function for training the discriminator is as follows: Where λ is the penalty coefficient. This represents data points obtained by uniformly sampling between the real data distribution and the generated data distribution. Indicates the discriminator pair The gradient; Step 63: Use the data in matrix X as positive samples in the first training set, randomly obtain the same number of negative samples as the positive samples, and concatenate the positive and negative samples in pairs to form the first training set. Use the first training set to retrain the generator in the trained VAE-GAN model, and use the retrained VAE-GAN model as the fine-tuned generator. Step 64: Input a random noise matrix of the same size as matrix X into the fine-tuned VAE-GAN model to obtain the microbial latent nonlinear feature matrix M. v and the drug's potential nonlinear characteristic matrix D v. ; The random noise matrix consists of random numbers uniformly distributed from -1 to 1.

8. The method for predicting microbial-drug associations according to claim 7, characterized in that: The generator in the pre-trained VAE-GAN model is retrained using the first training set with the following loss function: X f =Connecting({X known ,X unknown },{C1,C0}) Among them, G(X) known G(X) represents the generator's regeneration of known associated data; unknown ) represents the generator's regeneration of unknown related data, X f This is the first training set, X known X is a positive sample in the first training set. unknown C1 is the label of the positive sample in the first training set, and C0 is the label of the negative sample in the first training set.

9. The method for predicting microbial-drug associations according to claim 8, characterized in that: In step seven, the first k largest diagonal elements of the microbial-drug correlation matrix a are used to construct a diagonal matrix a′, and the singular value decomposition (SVD) method is used to decompose matrix a′ to obtain the microbial linear feature matrix V. k And the linear characteristic matrix U of the drug k Specifically: Among them, U k It is the linear characteristic matrix of the drug, V k It is the linear characteristic matrix of microorganisms, ∑ k Let be the singular value matrix of a′, and k be the matrix factorization dimension.

10. The method for predicting microbial-drug associations according to claim 9, characterized in that: In step eight, the microbial-drug correlation matrix a obtained in step one and the D obtained in step six are combined. v and M v Step 7 Obtained V k and U k The final embedding features are obtained by concatenation, and these features are used as the second training set to train the MLP model, resulting in a trained MLP model. Step 81: Combine matrices a and U k With D v The final drug embedding feature matrix Z is obtained by concatenation. d , will a, V k With M v. The final embedding feature matrix z of microorganisms is obtained by concatenation. m Using Z d and z m Obtain the final embedding feature matrix T: Z d =connecting(U k ,a,D v ) Z m =connecting(V k ,a T ,M v ) Among them, Z d It is the final embedding feature matrix of the drug, z m It is the final embedding feature matrix of microorganisms, Z d It is Z d The i-th row in It is Z m In the j-th row, T ij It is drug d i and microorganisms m j The final embedded features after association; Step 82: Use the final embedded feature matrix T as the second training set, train the MLP model using the second training set, and use the trained MLP as the microbiome-drug association prediction model. The MLP model is trained using the second training set, specifically as follows: T l =ReLU(T l-1 In l +b l ) Where l is the number of hidden layers; T l It is the output of the l-th hidden layer, w l b is the learnable parameter matrix of the l-th hidden layer. l It is the bias of the l-th hidden layer. It is the microbiome-drug association prediction score; The following loss function is used to train the MLP model: Among them, y ij These are the labels of the positive samples in the second training set. 1-y represents the probability of predicting a positive sample. ij These are the labels of the negative samples in the second training set. y represents the probability of a predicted negative sample, N is the total number of microbe-drug association pairs in the second training set, and y represents the probability of a predicted negative sample. + Let y- represent the set of positive sample labels in the second training set, and let y- represent the set of negative sample labels in the second training set.

Images