Intestinal microorganism-disease association prediction method based on deep learning
By constructing metabolite-mediated association networks and training a deep neural network, the method addresses the limitations of existing methods, enabling accurate prediction of microbe-disease associations, including new microbial markers.
Patent Information
- Application Number
- CN202411430539.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-14
- Publication Date
- 2025-01-03
AI Technical Summary
Existing methods for predicting intestinal microorganism-disease associations lack consideration of biological significance and cannot leverage deep learning to explore new microbial species related to diseases, limiting their applicability in diagnosis and treatment.
Construct intestinal metabolite-metabolite, microbe-metabolite, and disease-metabolite association networks using point-binary correlation, merge these networks to form a potential species-level association network, and use a deep neural network model trained with known labels to predict microbe-disease associations.
Accurately predicts relationships between microorganisms and diseases, especially for new microbial markers, enhancing diagnostic and treatment capabilities.
Smart Images
Figure CN119252346A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of artificial intelligence technology, and specifically relates to a method for predicting intestinal microorganism-disease association based on deep learning. Background Art
[0002] The intestinal microbiome is a complex microbial ecosystem whose composition and host disease progression are mutually influential. This means that changes in the composition of the intestinal microbiome can affect the progression of the disease, and the progression of the disease can also change the composition of the intestinal microbes. Predicting the association between intestinal microbes and diseases is very helpful in providing a theoretical basis for the development of new diagnostic methods, treatments, and preventive measures.
[0003] At present, some methods for predicting the association between microorganisms and diseases have emerged. Some of these methods assume that there are similar interaction patterns between microorganisms with similar functions and diseases, abstract the association between microorganisms and diseases into a similarity matrix, and use social network methods to calculate the number of walks between microbial nodes and disease nodes, converting the prediction of microorganism-disease association into a counting problem, and considering the influence of walk length to complete the prediction. Another part of the method also first constructs a microorganism-disease association matrix. The difference is that it applies a double random walk algorithm on the constructed heterogeneous network to obtain the degree of association between microorganisms and diseases. Another part of the method abstracts the microorganism-disease association into two cost functions in the microbial space and the disease space, thereby converting the problem into a minimum optimization problem, unifying the two classifiers through the weighted average method and finally outputting the microorganism-disease association. Although these methods can complete the prediction of microorganism-disease association, these methods still have the following problems: 1. These methods directly calculate the association between microorganisms and diseases through algorithms, and do not take into account the biological significance of these associations, that is, metabolites are the bridge between microorganisms and diseases. At present, a variety of microbial metabolites are considered to be intermediates that affect host cells and regulate host physiology, and are key to human health and disease. Imbalance or disruption of intestinal microorganisms has been shown to lead to the production of harmful metabolites, such as the metabolites lipopolysaccharide and trimethylamine oxide, which can induce inflammation, oxidative stress and insulin resistance. In contrast, probiotics have the ability to produce beneficial metabolites, including short-chain fatty acids and bile acids, which can initiate multiple signaling pathways in host cells, such as G protein-coupled receptors and nuclear receptors.
[0004] 2. These prediction methods are all based on non-deep learning methods, which means that they can only predict the association between known microorganisms and diseases. They cannot use the powerful learning ability of deep learning to explore new microbial species related to diseases, resulting in the inability to apply the association between new species and diseases to specific diagnosis and treatment. Summary of the invention
[0005] In order to solve the technical problems existing in the above background technology, the present invention provides a method for predicting intestinal microorganism-disease association based on deep learning, comprising: The intestinal metabolite-metabolite interaction network, microbe-metabolite association network and disease-metabolite association network were constructed respectively; The metabolite-metabolite interaction network, the microorganism-metabolite association network and the disease-metabolite association network are constructed by point-binary correlation formula to obtain a microorganism-metabolite association network and a disease-metabolite association network, and the microorganism-metabolite association network and the disease-metabolite association network are merged to obtain a potential species-level intestinal microorganism-disease association network with metabolites as a bridge; The potential species-level intestinal microorganism-disease association network with metabolites as a bridge is used as the input of the deep neural network model, and the known microorganism-disease association labels are used and binarized as the target of network model learning, and the deep neural network model is fully trained to achieve model convergence, and the model parameters of the round with the highest accuracy are taken as the final trained model weights; Leverage fully trained model weights to predict model-agnostic species-level gut microbe-disease associations.
[0006] As a further illustration of the present invention, the construction of the intestinal metabolite-metabolite interaction network, the microorganism-metabolite association network and the metabolite-disease association network respectively specifically includes: The compound search tool STITCH was used to construct an intestinal metabolite-metabolite interaction network; The association between species-level intestinal microorganisms and intestinal metabolites retrieved using the STITCH was extracted from the databases getMGene and NJS16 to form an intestinal microorganism-metabolite association network; The association between diseases and intestinal metabolites retrieved using STITCH was searched through the HMBD database to obtain a disease-metabolite association network.
[0007] As a further illustration of the present invention, the microorganism-metabolite association network and the disease-metabolite association network are constructed by the point-by-point correlation formula, and the microorganism-metabolite association network and the disease-metabolite association network are merged to obtain a potential species-level intestinal microorganism-disease association network with metabolites as a bridge, specifically including: The calculation process of the point-two column correlation formula is as follows: ;
[0008] in, is the point-biserial correlation coefficient, is the mean of the continuous variable when the binary variable is 1, represents the mean of the continuous variable when the binary variable is 0; and Binary samples are 1 or 0 in length. n represents the total sample size, represents the standard deviation of continuous variables; In order to reconstruct the microbe-metabolite interaction network, the point biserial correlation coefficients of the microbe-metabolite and metabolite-metabolite interaction networks were first calculated. The reconstructed microbe-metabolite interaction network was defined as follows: ; is the reconstructed microorganism-metabolite association matrix, is the point biserial correlation coefficient between different microorganisms and metabolites; The disease-metabolite association network is represented as the following matrix: ;
[0009] is the reconstructed disease-metabolite association matrix, is the point biserial correlation coefficient between different diseases and metabolites; The microbe-metabolite association matrix and the disease-metabolite association matrix are combined to obtain a potential species-level intestinal microbe-disease association network with metabolites as a bridge. The potential species-level intestinal microbe-disease association network with metabolites as a bridge is expressed as the following matrix:
[0010] A potential species-level microbe-disease association matrix constructed using metabolites as a bridge.
[0011] As a further explanation of the present invention, the deep neural network model includes a three-layer neural network, three fully connected layers and an activation function layer, wherein each layer of the neural network includes a convolutional layer, a pooling layer and a nonlinear layer.
[0012] As a further illustration of the present invention, the training process of the deep neural network model includes: The input data is processed by the convolution layer to capture local features and extract useful information. Then, it is processed by the pooling layer for dimensionality reduction while retaining important feature information. Finally, the ReLU nonlinear activation function is used to introduce rich nonlinear features, and the input data is finally mapped into a high-level semantic feature vector of length 128. The feature vector is processed through three fully connected layers, the dimension of the feature vector is gradually mapped to 1, and finally a single value is obtained as the prediction result of the microorganism-disease association through the operation of the Sigmoid activation function.
[0013] As a further illustration of the present invention, in model training, the mean square error is used as the loss function in training, the adam optimizer is used to optimize the update of the gradient in the network, and 20% of the neurons in the network are randomly frozen. 20 rounds of training are performed from scratch based on all the training data. Random Gaussian noise is added during the training process, and finally the convergence of the model loss function is achieved.
[0014] Compared with the prior art, the present invention has the following beneficial technical effects: This method uses metabolites as a bridge to fully learn the potential characteristics between microorganisms-metabolites and metabolites-diseases. It can accurately predict the relationship between microorganisms and diseases, especially in predicting the relationship between new microbial markers and diseases.
[0015] Other features and advantages of the technical solution will be described in the subsequent description, and partly become apparent from the description, or understood by implementing the technical solution. The purpose and other advantages of the technical solution can be achieved and obtained through the structures specifically pointed out in the written description and the drawings.
[0016] The technical solution of the present technical solution is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] The accompanying drawings are used to provide further understanding of the present technical solution and constitute a part of the specification. Together with the embodiments of the present technical solution, they are used to explain the present technical solution and do not constitute a limitation to the present technical solution.
[0018] FIG1 is a schematic diagram of the intestinal microorganism-disease association prediction process based on deep learning provided by the present invention; FIG2 is a schematic diagram of the reasoning process of the trained deep neural network model provided by the present invention.
[0019] FIG3 shows the prediction performance results of the three methods in the present invention. DETAILED DESCRIPTION
[0020] The preferred embodiments of the present technical solution are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present technical solution, and are not used to limit the present technical solution.
[0021] like Figure 1As shown, the present invention provides a method for predicting intestinal microorganism-disease association based on deep learning, comprising the following steps: S1: Construct intestinal metabolite-metabolite interaction network, microorganism-metabolite association network and disease-metabolite association network respectively.
[0022] In order to achieve the above construction process, it is necessary to ensure that the association network composed of the three data of intestinal microorganisms, metabolites and diseases contains the same elements. Therefore, in some specific embodiments, the above network can be obtained by the following construction process: first, the compound search tool STITCH is used to construct the intestinal metabolite-metabolite interaction network; then the association between intestinal microorganisms and metabolites previously retrieved using STITCH is extracted from the database getMGene and NJS16, which is the intestinal microorganism-metabolite association network; then the association between diseases and intestinal metabolites previously retrieved using STITCH is searched through the HMBD database, which is the disease-metabolite association network. A total of 10,187 species-level intestinal microorganism-metabolite association data, 5,398 intestinal metabolite-disease data, and 428,346 intestinal metabolite-metabolite association data were retrieved.
[0023] S2: The metabolite-metabolite interaction network, microbe-metabolite association network and disease-metabolite association network were constructed through the point-binary correlation formula to obtain the microbe-metabolite association network and the disease-metabolite association network, and the microbe-metabolite association network and the disease-metabolite association network were merged to obtain the potential species-level intestinal microbe-disease association network with metabolites as the bridge.
[0024] The present invention fully considers metabolites as a key bridge to predict the association between microorganisms and diseases. Therefore, the present method uses a microorganism-metabolite association network to fully capture the potential characteristics of microorganisms and metabolites, a metabolite-metabolite association network to fully learn the interaction characteristics between metabolites, and a metabolite-disease association network to fully learn the characteristics between metabolites and diseases. Through these three association networks, the purpose of fully utilizing metabolites as a key bridge to predict the association between microorganisms and diseases is achieved.
[0025] Specifically, the microorganism-metabolite association network and the disease-metabolite association network are constructed by the point-binary correlation formula, and the microorganism-metabolite association network and the disease-metabolite association network are merged to obtain a potential species-level intestinal microorganism-disease association network with metabolites as a bridge, including: Since the microbe-metabolite and metabolite-disease association networks are expressed as binary values, that is, 0 or 1 represents irrelevant or relevant, and the microbe-microbe association network is expressed as a continuous value, indicating the degree of association between different microorganisms, in order to fully characterize the relationship between the three and adapt to the input of the deep learning model, the point-by-point correlation formula is used to quantify the potential association between the three. The specific calculation process is as follows: ; in, is the point-biserial correlation coefficient, is the mean of the continuous variable when the binary variable is 1, represents the mean of the continuous variable when the binary variable is 0; and Binary samples are 1 or 0 in length. n represents the total sample size, Represents the standard deviation of continuous variables.
[0026] In order to reconstruct the microbe-metabolite interaction network, the point biserial correlation coefficients of the microbe-metabolite and metabolite-metabolite interaction networks were first calculated. The reconstructed microbe-metabolite interaction network was defined as follows: ; is the reconstructed microorganism-metabolite association matrix, is the point biserial correlation coefficient between different microorganisms and metabolites; The disease-metabolite association network is represented as the following matrix: ; is the reconstructed disease-metabolite association matrix, is the point biserial correlation coefficient between different diseases and metabolites; The microbe-metabolite association matrix and the disease-metabolite association matrix are combined to obtain the potential species-level intestinal microbe-disease association network with metabolites as the bridge. The potential species-level intestinal microbe-disease association network with metabolites as the bridge is expressed as the following matrix: ; A potential species-level microbe-disease association matrix constructed using metabolites as a bridge.
[0027] S3: Use the potential species-level gut microbe-disease association network with metabolites as the bridge as the input of the deep neural network model, use the known microbe-disease association labels and binarize them as the learning target of the network model, fully train the deep neural network model to achieve model convergence, and take the model parameters of the round with the highest accuracy as the final trained model weights.
[0028] In some specific embodiments, corresponding to the construction process of the above three association networks, microorganism-disease associations can be retrieved in gutMDisorder and GMrepo as labels required for the training process, and a total of 1,609 microorganism-disease association data are retrieved as training targets. In addition, an unrelated microorganism-disease association network is constructed to solve the problem of uneven positive and negative samples during training.
[0029] In some specific embodiments, the above-mentioned deep neural network model includes a three-layer neural network, three fully connected layers and an activation function layer, wherein each layer of the neural network includes a convolution layer, a pooling layer nonlinear layer, and fully extracts the effective features in the three networks. First, the input data is processed by the convolution layer to capture local features and extract useful information. Then, after the dimensionality reduction processing of the pooling layer, the spatial dimension of the feature is reduced, while retaining the most important feature information. This process helps to improve the computational efficiency of the network and reduce the risk of overfitting. Finally, rich nonlinear features are introduced through the ReLU nonlinear activation function, so that the model can learn more complex feature relationships. After these three operations, the input data is mapped to a high-level semantic feature vector with a length of 128. Subsequently, the feature vector is further processed by three fully connected layers, and its dimension is gradually mapped to 1. The last step of this process is to use the single value obtained as the prediction result of the microorganism-disease association through the operation of the Sigmoid activation function.
[0030] In model training, known microorganism-disease association labels were used and binarized as the target of network learning, mean square error was used as the loss function in training, the adam optimizer was used to optimize the update of gradients in the network, and 20% of the neurons in the network were randomly frozen to prevent overfitting of the model. All training data were used to train the model from scratch for 20 rounds. Random Gaussian noise was added during the training process to improve the robustness of the model, and the convergence of the model loss function was achieved.
[0031] S4: Use fully trained model weights to predict model-unknown microbe-disease associations.
[0032] By fully training under the labels of known microorganism-disease associations to achieve model convergence, a weight file recording the model parameters will be generated. By reading the weight file into the network model, the following can be accomplished: Figure 2 The reasoning process shown is used to predict the association between unknown microorganisms and diseases.
[0033] In the specific implementation process, the method provided by the present invention can be developed based on Python language, using the scipy library to construct microorganism-metabolite and metabolite-disease association networks, and using the TensorFlow framework according to Figure 1 The structure shown in the figure is used to construct a deep neural network model. The network model first consists of three consecutive convolutions, pooling, and nonlinear operations. The size of the convolution kernel is , and the number of convolution kernels is 32, 64, and 128, respectively. The pooling adopts the maximum pooling operation, that is, taking the maximum value in the receptive field. The nonlinear operation is completed using the Relu activation function. Then the features pass through the fully connected layers with dimensions of 640 and 320. Each fully connected operation is followed by a Relu nonlinear operation. Finally, the features are sent to a fully connected layer with a dimension of 1. After a Sigmoid nonlinear operation on one side, the final predicted probability of microorganism-disease association is obtained. During the training process, Gaussian noise is randomly added to enhance the robustness of the model. The mean square error loss function and Adam optimizer are used to update the parameters in the model. After 20 rounds of training, the model converges and the model parameters of the round with the highest accuracy are taken as the final model weights. Finally, the fully trained weights can be used to predict the association between microorganisms and diseases that are unknown to the model.
[0034] Method performance verification: The 5-fold and 10-fold cross-validation methods were used to verify the performance of this method. All known associations between intestinal microorganisms and diseases were randomly and evenly divided into 5 parts or 10 parts, 4 parts or 9 parts were used for training, and the other parts were used to evaluate the performance of the method. In order to reduce the potential bias caused by random sample division, this process was repeated 20 times. The performance of the model of the present invention is evaluated based on AUC, and the results show that the AUC of the 5-fold and 10-fold cross-validation of this method are 0.9754 and 0.9780, respectively. At the same time, three mainstream methods, KATZHMDA, BiRWHMDA and LRLSHMDA, were compared. To ensure consistency, we used the same positive examples for the above three methods to evaluate the corresponding prediction performance. In the end, the auc of the 5-fold cross-validation of the three methods were 0.8014, 0.8131 and 0.5189, respectively, and the auc of the 10-fold cross-validation were 0.7981, 0.8086 and 0.5233, respectively, which are significantly lower than our method. Figure 3 shown.
[0035] Subsequently, we performed 16S rRNA gene sequencing and WGS raw data analysis, and performed literature validation on the intestinal microbes predicted to be associated with the top three diseases in our prediction results, including inflammatory bowel disease (IBD), rheumatoid arthritis (RA), and obesity. Five datasets (PRJEB7949, PRJNA450340, PRJNA401981, PRJEB12123, and PRJNA637782) were analyzed, and new microbial biomarkers were identified that were not present in the positive group but were predicted to be associated with IBD, RA, and obesity by our method. In the literature validation, we found that subsequent studies confirmed our predictions, as shown in Table 1 below.
[0036] Table 1
[0037] Obviously, those skilled in the art can make various changes and modifications to the technical solution without departing from the spirit and scope of the technical solution. Thus, if these modifications and variations of the technical solution fall within the scope of the claims of the technical solution and their equivalents, the technical solution is also intended to include these modifications and variations.
Claims
1. A method for predicting intestinal microorganism-disease association based on deep learning, characterized in that: include: The intestinal metabolite-metabolite interaction network, microbe-metabolite association network and disease-metabolite association network were constructed respectively; The metabolite-metabolite interaction network, the microorganism-metabolite association network and the disease-metabolite association network are constructed by point-binary correlation formula to obtain a microorganism-metabolite association network and a disease-metabolite association network, and the microorganism-metabolite association network and the disease-metabolite association network are merged to obtain a potential species-level intestinal microorganism-disease association network with metabolites as a bridge; The potential species-level intestinal microorganism-disease association network with metabolites as a bridge is used as the input of the deep neural network model, and the known microorganism-disease association labels are used and binarized as the target of network model learning, and the deep neural network model is fully trained to achieve model convergence, and the model parameters of the round with the highest accuracy are taken as the final trained model weights; Leverage fully trained model weights to predict model-agnostic species-level gut microbe-disease associations.
2. The intestinal microorganism-disease association prediction method based on deep learning according to claim 1, characterized in that: The construction of the intestinal metabolite-metabolite interaction network, the microorganism-metabolite association network and the metabolite-disease association network respectively specifically includes: The compound search tool STITCH was used to construct an intestinal metabolite-metabolite interaction network; The association between species-level intestinal microorganisms and intestinal metabolites retrieved using the STITCH was extracted from the databases getMGene and NJS16 to form an intestinal microorganism-metabolite association network; The association between diseases and intestinal metabolites retrieved using STITCH was searched through the HMBD database to obtain a disease-metabolite association network.
3. The intestinal microorganism-disease association prediction method based on deep learning according to claim 1, characterized in that: The microorganism-metabolite association network and the disease-metabolite association network are constructed by the point-binary correlation formula, and the microorganism-metabolite association network and the disease-metabolite association network are merged to obtain a potential species-level intestinal microorganism-disease association network with metabolites as a bridge, specifically including: The calculation process of the point-two column correlation formula is as follows: ； in, is the point-biserial correlation coefficient, is the mean of the continuous variable when the binary variable is 1, represents the mean of the continuous variable when the binary variable is 0; and Binary samples are 1 or 0 in length. n represents the total sample size, represents the standard deviation of continuous variables; In order to reconstruct the microbe-metabolite interaction network, the point biserial correlation coefficients of the microbe-metabolite and metabolite-metabolite interaction networks were first calculated. The reconstructed microbe-metabolite interaction network was defined as follows: ； is the reconstructed microorganism-metabolite association matrix, is the point biserial correlation coefficient between different microorganisms and metabolites; The disease-metabolite association network is represented as the following matrix: ； is the reconstructed disease-metabolite association matrix, is the point biserial correlation coefficient between different diseases and metabolites; The microbe-metabolite association matrix and the disease-metabolite association matrix are combined to obtain a potential species-level intestinal microbe-disease association network with metabolites as a bridge. The potential species-level intestinal microbe-disease association network with metabolites as a bridge is expressed as the following matrix: ； A potential species-level microbe-disease association matrix constructed using metabolites as a bridge.
4. The intestinal microorganism-disease association prediction method based on deep learning according to claim 1, characterized in that: The deep neural network model includes a three-layer neural network, three fully connected layers and an activation function layer, wherein each layer of the neural network includes a convolutional layer, a pooling layer and a nonlinear layer.
5. The intestinal microorganism-disease association prediction method based on deep learning according to claim 4, characterized in that: The training process of the deep neural network model includes: The input data is processed by the convolution layer to capture local features and extract useful information. Then, it is processed by the pooling layer for dimensionality reduction while retaining important feature information. Finally, the ReLU nonlinear activation function is used to introduce rich nonlinear features, and the input data is finally mapped into a high-level semantic feature vector of length 128. The feature vector is processed through three fully connected layers, the dimension of the feature vector is gradually mapped to 1, and finally a single value is obtained as the prediction result of the microorganism-disease association through the operation of the Sigmoid activation function.
6. The intestinal microorganism-disease association prediction method based on deep learning according to claim 1, characterized in that: In model training, the mean square error is used as the loss function in training, the adam optimizer is used to optimize the update of the gradient in the network, and 20% of the neurons in the network are randomly frozen. 20 rounds of training are performed from scratch based on all the training data. Random Gaussian noise is added during the training process to finally achieve convergence of the model loss function.