A disease prediction method and system based on a variational neural network

Abstract

The application discloses a disease prediction method and system based on a variational neural network, and the method comprises the following steps: step 1: extracting DNA, RNA and metabolites of samples, amplifying the DNA and RNA information into a library suitable for high-throughput sequencing, obtaining sequencing original data by using a high-throughput sequencing technology, performing species annotation and function annotation on the original data after processing, obtaining the abundance data of a metagenome and the pathway data of a metatranscriptome, and obtaining the metabolite abundance data of a metabolome by using a mass spectrometry analysis method; step 2: pre-processing the multi-omics data, including data conversion and normalization processing; step 3: inputting the processed abundance data of the metagenome, the pathway data and the metabolite abundance data into a trained algorithm framework, obtaining a probability value of getting sick, and the output result of the algorithm framework is getting sick or not getting sick. The application can be used for integrating incomplete microbiome multi-omics data, predicting diseases and finding disease-related biomarkers.

Description

Technical Field

[0001] This invention relates to a disease prediction method and system based on variational neural networks, belonging to the field of disease prediction technology. Background Technology

[0002] The human gut microbiota is a complex microbial ecosystem with a genome of approximately 3 million cells, 150 times larger than the human host genome. In fact, the gut microbiota plays a crucial role in human metabolism by synthesizing enzymes not encoded by the human genome. These enzymes promote the breakdown of polysaccharides and polyphenols, facilitate nutrient absorption, and provide protection against pathogens. A growing body of research indicates that dysbiosis of the gut microbiota may be closely associated with various diseases, particularly those affecting the gastrointestinal system. To elucidate the relationship between the microbiota and human health, various omics technologies have emerged, such as metagenomics, metatranscriptomics, and metabolomics. Each provides information on a molecular mechanism or biological process at a specific omics level. In recent years, increasing research has shown that combining omics data often provides more complete information and a better understanding of microbial ecology, which can increase the accuracy of human disease prediction, improve the robustness of analyses, and discover important biomarkers. Notably, multi-omics microbiome data encompasses various types of data and is known for its heterogeneity, sparsity, and high dimensionality. Given these characteristics, data processing requires specialized analytical methods to facilitate deeper understanding and knowledge discovery. High-performance machine learning methods have received considerable attention in the biological field, and a large number of models have been developed to fully utilize the potential of multi-omics information.

[0003] Incomplete omics data are common in publicly available databases, attributable to various factors such as limited funding, ethical considerations, and privacy concerns, which can affect sample availability. This poses a significant challenge to ensemble analysis. In such cases, sample discarding or mean imputation can be considered. However, the former will significantly reduce the number of usable samples, while the latter may severely distort the true distribution of the data. Existing machine learning algorithms for disease prediction using incomplete multi-omics data suffer from two main problems: first, they cannot effectively extract relevant features from high-dimensional omics data and filter out irrelevant features; second, they struggle to achieve flexible integration of incomplete multi-omics data while fully utilizing its information for efficient prediction. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides a disease prediction method and system based on variational neural networks. This method utilizes incomplete intestinal multi-omics data for disease prediction, collects human intestinal fecal samples, and obtains bacterial flora abundance data, pathway data, and metabolite abundance data of the samples through sequencing and analysis techniques. The multi-omics data is preprocessed, and the multi-omics data of the samples are substituted into a trained algorithm framework to obtain the probability value of disease. The output of the algorithm framework is divided into two categories: diseased and not diseased.

[0005] On the one hand, the present invention provides a disease prediction method based on variational neural networks, comprising the following steps:

[0006] Step 1: Extract DNA, RNA and metabolites from the sample, amplify the DNA and RNA information into a library suitable for high-throughput sequencing, obtain raw sequencing data using high-throughput sequencing technology, process the raw data and perform species annotation and functional annotation to obtain metagenomic microbial abundance data and metatranscriptome pathway data, and obtain metabolome metabolite abundance data using mass spectrometry analysis.

[0007] Step 2: Preprocess the multi-omics data, including data transformation and normalization.

[0008] Step 3: Substitute the processed microbial abundance data, pathway data, and metabolite abundance data into the trained algorithm framework to obtain the probability value of disease. The output of the algorithm framework is diseased or not diseased.

[0009] In one embodiment of the present invention, the data conversion and normalization process in step 2 includes the following steps:

[0010] Step 2.1: Perform the following transformations on the data to enable reasonable analysis using neural networks;

[0011] x = log2(2x + 0.00001)

[0012] Where x represents the bacterial community abundance data;

[0013] Step 2.2: If the bacterial community abundance data has been transformed, then the following normalization process is performed on each type of omics data:

[0014]

[0015] Where x mean It is the average value of the omics data, x max It is the maximum value in this omics dataset, x min It is the minimum value in this omics data.

[0016] In one embodiment of the present invention, step 3, in which the algorithm framework uses multi-omics data to derive the disease probability value, includes the following steps:

[0017] Step 3.1: The processed microbial abundance data, pathway data, and metabolite abundance data are compiled into three matrices. Let's assume the matrix representing the v-th omics data is... d from n samples v The matrix is ​​composed of several features. First, each matrix undergoes feature selection through a trained feature selection layer. The calculation process is as follows:

[0018] u v =x v ·s v

[0019] in It is a linear transformation matrix that approximates the one-hot form after training, and u is obtained by feature selection for each omics dataset. v ∈R n×F ;

[0020] Step 3.2: After feature selection, the omics data are passed through a trained encoder consisting of fully connected layers and activation functions to obtain the latent representation of each omics. The calculation process is as follows:

[0021]

[0022] μ v +∈·σ v =z v

[0023] in This represents the nonlinear transformation process of a multilayer neural network. During network training, ∈ is randomly sampled from a standard normal distribution; after model training, ∈ is fixed at 0. First, u... v The mean μ of the latent representation is obtained. v and variance σ v Then, the reparameterization technique is used to obtain the latent representation z of each omics data. v ;

[0024] Step 3.3: Simply integrate incomplete multi-omics data with arbitrary missing values ​​using a joint omics encoder to obtain the joint latent representation z. The calculation process is as follows:

[0025]

[0026]

[0027] μ+∈·σ=z,∈~N(0,1)

[0028] Where V represents the number of omics data in the sample, μ0 and σ0 represent the mean and variance of the prior distribution, ∈ is randomly sampled from the standard normal distribution during network training, and ∈ is fixed to 0 after model training; μ is the mean of the existing omics data. v and variance σ v The mean μ and variance σ of the joint omics data are obtained by integration, and the latent representation z of the joint omics data is obtained by using the reparameterization technique.

[0029] Step 3.4: The latent representation z of the joint omics data is used by a trained joint predictor consisting of fully connected layers and activation functions to derive the disease probability value. The calculation process is as follows:

[0030]

[0031] Where f ψ The nonlinear transformation process of a multilayer neural network is represented by the latent representation z of joint omics data, which yields a predictive label for whether a sample is sick.

[0032] In one embodiment of the present invention, the training process of the algorithm framework includes the following steps:

[0033] Step S1: Collect intestinal fecal samples from healthy individuals and individuals with the target disease predicted by the diagnostic model. Manually label the population, marking fecal samples from sick individuals as 1 and fecal samples from healthy individuals as 0. Obtain multi-omics data corresponding to the samples through sequencing and analysis technologies, or collect publicly available data to construct a multi-omics database and obtain multi-omics data of labeled fecal samples.

[0034] Step S2: Perform data transformation and normalization on the multi-omics data;

[0035] Step S3: Divide the labeled dataset into a training set and a test set. Use the training set data to perform supervised training of the algorithm framework and test it on the test set.

[0036] In one embodiment of the present invention, the supervised training process of the algorithm framework using training set data in step S3 includes the following steps:

[0037] Step S3.1: The microbial abundance data, pathway data, and metabolite abundance data of the training set are compiled into three matrices. Let's assume the matrix representing the v-th omics data is... d from n samples v Composed of several features, each matrix first undergoes a linear transformation through a feature selection layer:

[0038] T e =T0·(T E / T0) e / E

[0039]

[0040] u v =x v ·ε v

[0041] Where E represents the total number of training iterations, e represents the iteration number, and T0 and T... E These are the hyperparameters of the algorithm model, set to 10 and 0.1 respectively, γ v Here are the parameters of the fully connected layer, ε is randomly sampled from a uniform distribution (0,1), and softmax() represents an activation function. After the above transformations, x... v Get u v ∈R n×F ;

[0042] Step S3.2: The omics data, modified through the feature selection layer, is passed through an encoder consisting of fully connected layers and activation functions to obtain the latent representation of each omics. The calculation process is as follows:

[0043]

[0044] μ v +∈·σ v =z v

[0045] in θ represents the nonlinear transformation process of a multilayer neural network. v The parameters of the neural network are ∈, randomly sampled from a standard normal distribution; first by u v The mean μ of the latent representation is obtained. v and variance σ v Then, the reparameterization technique is used to obtain the latent representation z of each omics data. v ;

[0046] Step S3.3: Simply integrate incomplete multi-omics data with arbitrary missing values ​​using a joint omics encoder to obtain the joint latent representation z. The calculation process is as follows:

[0047]

[0048]

[0049] μ+∈·σ=z,∈~N(0,1)

[0050] Where V represents the number of omics samples, μ0 and σ0 represent the mean and variance of the prior distribution, and ∈ is randomly sampled from a standard normal distribution; the mean μ of the existing omics data is used to determine the omics data. v and variance σ vThe mean μ and variance σ of the joint omics data are obtained by integration, and the latent representation z of the joint omics data is obtained by using the reparameterization technique.

[0051] Step S3.4: Latent representation z of each omics data v The joint latent representation z is passed through a predictor consisting of fully connected layers and activation functions, and a joint predictor, respectively, to obtain the specific omics prediction probability value y. v and the final predicted probability value The calculation process is as follows:

[0052]

[0053]

[0054] in and f ψ This represents the nonlinear transformation process of a multilayer neural network. Both and ψ are parameters of the neural network;

[0055] Step S3.5: Calculate the loss based on the model's loss function, perform gradient backpropagation to update the parameters of the model's neural network, and the calculation process is as follows:

[0056]

[0057]

[0058] L T =L J +α∑ v∈V L v

[0059]

[0060] Where α and β are the balance coefficients combining different losses, and N(μ0,σ0) represents the prior distribution. v ,σ v Let ) denote the distribution of the v-th omics latent representation, N(μ,σ) denote the distribution of the joint latent representation, KL() denotes the KL divergence, i.e., the relative entropy, between the two distributions, and y be the one-hot true label. v Predict probability values ​​for specific omics. Here, λ represents the final predicted probability value, n is the number of samples, and λ is the learning rate during algorithm training. These are the parameters of the neural network in the algorithm framework; thus, the neural network parameters are updated after one training cycle of the model.

[0061] On the other hand, the present invention also provides a disease prediction system based on variational neural networks, which applies the aforementioned disease prediction method based on variational neural networks. The system includes:

[0062] The feature selection layer module is used to select important features for linear transformation of omics data;

[0063] The encoder module is used to randomly encode omics data into latent representations;

[0064] The joint encoder module is used to integrate the latent representations of various omics; and

[0065] The predictor and joint predictor modules are used to provide label inference, i.e. disease prediction results, from the latent representation encoded by the encoder module.

[0066] In one embodiment of the present invention, the feature selection layer module is a Boolean matrix.

[0067] In one embodiment of the present invention, the encoder module is composed of a fully connected neural network.

[0068] In one embodiment of the present invention, the joint encoder module is a computing module.

[0069] In one embodiment of the present invention, the predictor and joint predictor module are composed of a fully connected neural network.

[0070] This invention provides a disease prediction method and system based on variational neural networks, which has the following advantages: First, it proposes a novel framework based on variational neural networks, capable of integrating incomplete multi-omics microbiome data to predict diseases and identify disease-related biomarkers. Second, it introduces specific distributions, allowing the selection of features most relevant to the target disease within each microbiome, thereby improving model interpretability. Third, the algorithm utilizes the information bottleneck principle to construct the loss function for model training, promoting the learning of single-omics and joint-omics latent representations, resulting in high prediction accuracy and robustness. Fourth, it does not have high requirements for the completeness of multi-omics data, allowing for flexible utilization of the available omics data in samples for disease prediction. Attached Figure Description

[0071] Figure 1 This is a flowchart of the disease prediction method based on variational neural networks of the present invention. Detailed Implementation

[0072] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0073] like Figure 1 As shown, the present invention provides a disease prediction method based on variational neural networks, which includes the following steps in some embodiments:

[0074] Step 1: Collect intestinal fecal samples from the target population. First, extract DNA, RNA, and metabolites from the samples. Then, amplify the DNA and RNA information into libraries suitable for high-throughput sequencing. Next, use high-throughput sequencing technology to obtain raw sequencing data. After processing the raw data, perform species annotation and functional annotation to obtain metagenomic microbial abundance data and metagenomic pathway data. In addition, obtain metabolite abundance data of the metabolome using mass spectrometry analysis.

[0075] Considering the cost of sequencing, if there are too many samples, it will take a lot of time and resources to obtain complete multi-omics data. Therefore, it is easy for the multi-omics data of the samples to be incomplete. By substituting the existing omics data of the samples into the trained model, the missing omics data will not participate in the integration calculation, and the prediction accuracy can still be high.

[0076] Step 2: Preprocess the multi-omics data, including data transformation and normalization.

[0077] Optionally, in some embodiments, the data transformation and normalization process in step 2 includes the following steps:

[0078] Step 2.1: To perform reasonable analysis using neural networks, the data first needs to be transformed. The relative abundance of the bacterial community is considered in this invention. These values ​​range from 0 to relatively large actual values, with most ranges from 10. -1 Up to 10 -4 Considering that low abundance characteristics may play an important role in health status, the following logarithmic transformation was designed and applied:

[0079] x = log2(2x + 0.00001) (1)

[0080] To avoid issues with the origin value, 0.00001 was added. Here, x represents the bacterial community abundance data.

[0081] Step 2.2: If the bacterial community abundance data has been transformed, then the following normalization process is performed on each type of omics data:

[0082]

[0083] Where x mean It is the average value of the omics data, x max It is the maximum value in this omics dataset, x min It is the minimum value in this omics data.

[0084] Step 3: Substitute the processed microbial abundance data, pathway data, and metabolite abundance data into the trained algorithm framework to obtain the probability value of disease in the population to be tested. The output of the algorithm framework is divided into two categories: diseased and not diseased.

[0085] Optionally, in some implementations, step 3, in which the algorithm framework uses multi-omics data to derive the disease probability value, includes the following steps:

[0086] Step 3.1: The processed microbial abundance data, pathway data, and metabolite abundance data are compiled into three matrices. Let's assume the matrix representing the v-th omics data is... d from n samples v The matrix is ​​composed of several features. First, each matrix undergoes feature selection through a trained feature selection layer. The calculation process is as follows:

[0087] u v =x v ·s v (3)

[0088] in It is a linear transformation matrix that approximates the one-hot encoding form after training, and u is obtained by feature selection for each omics dataset. v ∈R n×F .

[0089] Step 3.2: After feature selection, the omics data are passed through a trained encoder consisting of fully connected layers and activation functions to obtain the latent representation of each omics. The calculation process is as follows:

[0090]

[0091] μ v +∈·σ v =z v (5)

[0092] in This represents the nonlinear transformation process of a multilayer neural network. During network training, ∈ is randomly sampled from a standard normal distribution; after model training, ∈ is fixed at 0. First, from μ... v The mean μ of the latent representation is obtained. v and variance σ vThen, the reparameterization technique is used to obtain the latent representation z of each omics data. v .

[0093] Step 3.3: Simply integrate incomplete multi-omics data with arbitrary missing values ​​using a joint omics encoder to obtain the joint latent representation z. The calculation process is as follows:

[0094]

[0095]

[0096] μ+∈·σ=z,∈~N(0,1) (8)

[0097] Where V represents the number of omics samples, set to 3 in this embodiment; μ0 and σ0 represent the mean and variance of the prior distribution, set to 0 and 1 respectively in this embodiment; ∈ is randomly sampled from a standard normal distribution during network training, and ∈ is fixed to 0 after model training. The mean μ of the existing omics data... v and variance σ v The mean μ and variance σ of the joint omics data are obtained by integration, and the potential representation z of the joint omics data is obtained by using the reparameterization technique.

[0098] Step 3.4: The latent representation z of the joint omics data is used by a trained joint predictor consisting of fully connected layers and activation functions to derive the disease probability value. The calculation process is as follows:

[0099]

[0100] Where f ψ The nonlinear transformation process of the multilayer neural network is represented by the latent representation z of the joint omics data, which yields the predictive label of whether the population to be tested is sick according to the embodiments of the present invention.

[0101] Optionally, in some embodiments, the training process of the algorithm framework of the disease prediction method based on variational neural networks includes the following steps:

[0102] Step S1: Collect intestinal fecal samples from healthy individuals and individuals with the target disease predicted by the diagnostic model. Manually label the population, marking fecal samples from sick individuals as 1 and fecal samples from healthy individuals as 0. Obtain multi-omics data corresponding to the samples through sequencing and analysis technologies, or collect publicly available data online to construct a multi-omics database and obtain multi-omics data of labeled fecal samples.

[0103] Step S2: Perform data transformation and normalization on the multi-omics data.

[0104] Step S3: Divide the labeled dataset into a training set and a test set. Use the training set data to perform supervised training of the algorithm framework and test it on the test set.

[0105] To further illustrate, the steps of a single training process in step S3, where the algorithm framework utilizes the training set data for supervised training, are described below:

[0106] Step S3.1: The microbial abundance data, pathway data, and metabolite abundance data of the training set are compiled into three matrices. Let's assume the matrix representing the v-th omics data is... d from n samples v Composed of several features, each matrix first undergoes a linear transformation through a feature selection layer:

[0107] T e =T0·(T E / T0) e / E (10)

[0108]

[0109] u v =x v ·s v (12)

[0110] Where E represents the total number of training iterations, which is set to 2000 in this embodiment of the invention, and e represents the iteration number, T0 and T E These are the hyperparameters of the algorithm model, set to 10 and 0.1 respectively, γ v Here are the parameters of the fully connected layer, ε is randomly sampled from a uniform distribution (0,1), and softmax() represents an activation function. After the above transformations, x... v Get u v ∈R n×F .

[0111] Step S3.2: The omics data, modified through the feature selection layer, is passed through an encoder consisting of fully connected layers and activation functions to obtain the latent representation of each omics. The calculation process is as follows:

[0112]

[0113] μ v +∈·σ v =z v (14)

[0114] in θ represents the nonlinear transformation process of a multilayer neural network. v The parameters of the neural network are ∈, randomly sampled from a standard normal distribution. First, u... v The mean μ of the latent representation is obtained.v and variance σ v Then, the reparameterization technique is used to obtain the latent representation z of each omics data. v .

[0115] Step S3.3: Simply integrate incomplete multi-omics data with arbitrary missing values ​​using a joint omics encoder to obtain the joint latent representation z. The calculation process is as follows:

[0116]

[0117]

[0118] μ+∈·σ=z,∈~N(0,1) (17)

[0119] Where V represents the number of omics data in the sample, set to 3 in this embodiment; μ0 and σ0 represent the mean and variance of the prior distribution, set to 0 and 1 respectively in this embodiment; and ∈ represents random sampling from a standard normal distribution. The mean μ of the existing omics data is used as the basis for the calculation. v and variance σ v The mean μ and variance σ of the joint omics data are obtained by integration, and the potential representation z of the joint omics data is obtained by using the reparameterization technique.

[0120] Step S3.4: Latent representation z of each omics data v The joint latent representation z is passed through a predictor consisting of fully connected layers and activation functions, and a joint predictor, respectively, to obtain the specific omics prediction probability value y. v and the final predicted probability value The calculation process is as follows:

[0121]

[0122]

[0123] in and f ψ This represents the nonlinear transformation process of a multilayer neural network. Both and ψ are parameters of the neural network.

[0124] Step S3.5: Calculate the loss based on the model's loss function, perform gradient backpropagation to update the parameters of the model's neural network, and the calculation process is as follows:

[0125]

[0126]

[0127] L T =L J +α∑ v∈VL v (twenty two)

[0128]

[0129] Wherein, α and β are balance coefficients that combine different losses, and in this embodiment of the invention, they are set to 1 and 0.001, respectively.

[0130] N(μ0,σ0) represents the prior distribution, N(μ v ,σ v Let ) denote the distribution of the v-th omics latent representation, N(μ,σ) denote the distribution of the joint latent representation, KL() denotes the KL divergence, i.e., the relative entropy, between the two distributions, and y be the one-hot true label. v Predict probability values ​​for specific omics. The final predicted probability value is given by λ, where n is the number of samples and λ is the learning rate during algorithm training, which is set to 0.01 in this embodiment. These are the parameters of the neural network within the algorithm framework. This process is repeated to complete one training iteration of the model and update the neural network parameters.

[0131] This invention also provides a disease prediction system based on variational neural networks, which applies the aforementioned disease prediction method based on variational neural networks. In some embodiments, the system includes:

[0132] The feature selection layer module is used to select important features for linear transformation of omics data;

[0133] The encoder module is used to randomly encode omics data into latent representations;

[0134] The joint encoder module is used to integrate the latent representations of various omics; and

[0135] The predictor and joint predictor modules are used to provide label inference, i.e. disease prediction results, from the latent representation encoded by the encoder module.

[0136] Optionally, in some implementations, the feature selection layer module is a Boolean matrix.

[0137] Optionally, in some embodiments, the encoder module is composed of a fully connected neural network.

[0138] Optionally, in some embodiments, the joint encoder module is a computing module.

[0139] Optionally, in some implementations, the predictor and joint predictor module are composed of a fully connected neural network.

[0140] This invention processes and integrates multi-omics data separately, making full use of the information within each omics while also mining the interaction information between omics. It is applicable to the prediction of samples with incomplete multi-omics data, reducing the requirements for the completeness of multi-omics data of samples, and providing a technical means for predicting diseases using intestinal multi-omics data.

[0141] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A disease prediction system based on variational neural networks, characterized in that, The system, applied to a disease prediction method based on variational neural networks, comprises: The feature selection layer module is used to select important features for linear transformation of omics data; The encoder module is used to randomly encode omics data into latent representations; The joint encoder module is used to integrate the latent representations of various omics; and The predictor and joint predictor modules are used to provide label inference, i.e., disease prediction results, from the latent representation encoded by the encoder module; The disease prediction method based on variational neural networks includes the following steps: Step 1: Extract DNA, RNA and metabolites from the sample, amplify the DNA and RNA information into a library suitable for high-throughput sequencing, obtain raw sequencing data using high-throughput sequencing technology, process the raw data and perform species annotation and functional annotation to obtain metagenomic microbial abundance data and metatranscriptome pathway data, and obtain metabolome metabolite abundance data using mass spectrometry analysis. Step 2: Preprocess the multi-omics data, including data transformation and normalization. Step 3: Substitute the processed microbial abundance data, pathway data, and metabolite abundance data into the trained algorithm framework to obtain the probability value of disease. The output of the algorithm framework is diseased or not diseased. The data transformation and normalization process in step 2 includes the following steps: Step 2.1: Perform the following transformations on the data to enable reasonable analysis using neural networks; in, This represents the abundance data of the bacterial community; Step 2.2: If the bacterial community abundance data has been transformed, then the following normalization process is performed on each type of omics data: in It is the average value of the omics data. It is the maximum value in this omics dataset. It is the minimum value in this omics dataset; Step 3 of the algorithm framework utilizes multi-omics data to derive the disease probability value, including the following steps: Step 3.1: The processed microbial abundance data, pathway data, and metabolite abundance data are compiled into three matrices, assuming they represent the first... Matrix of omics data Depend on one sample The matrix is ​​composed of several features. First, each matrix undergoes feature selection through a trained feature selection layer. The calculation process is as follows: in It is a linear transformation matrix that approximates the one-hot form after training, obtained by feature selection for each omics dataset. ; Step 3.2: After feature selection, the omics data are passed through a trained encoder consisting of fully connected layers and activation functions to obtain the latent representation of each omics. The calculation process is as follows: in This represents the nonlinear transformation process of a multilayer neural network. When training the network, random samples are taken from a standard normal distribution. After the model training is completed... Fixed at 0; first by Obtain the mean of the latent representation and variance Then, the reparameterization technique is used to obtain the latent representation of each omics data. ; Step 3.3: Simply integrate incomplete multi-omics data with arbitrary missing values ​​using a joint omics encoder to obtain a joint latent representation. The calculation process is as follows: , in This represents the number of omics systems represented by the sample. and These represent the mean and variance of the prior distribution. When training the network, random samples are taken from a standard normal distribution. After the model training is completed... Fixed at 0; based on the mean of existing omics data and variance The ensemble yields the joint omics mean. and variance Using reparameterization techniques to obtain potential representations of joint omics data ; Step 3.4: Potential Representation of Joint Omics Data The probability of illness is obtained by a trained joint predictor consisting of fully connected layers and activation functions. The calculation process is as follows: in Representing the nonlinear transformation process of multilayer neural networks, and the latent representation based on joint omics data. Obtain a predicted label indicating whether a sample is sick.

2. The disease prediction system based on variational neural networks according to claim 1, characterized in that, The training process of the algorithm framework includes the following steps: Step S1: Collect intestinal fecal samples from healthy individuals and individuals with the target disease predicted by the diagnostic model. Manually label the population, marking fecal samples from sick individuals as 1 and fecal samples from healthy individuals as 0. Obtain multi-omics data corresponding to the samples through sequencing and analysis technologies, or collect publicly available data to construct a multi-omics database and obtain multi-omics data of labeled fecal samples. Step S2: Perform data transformation and normalization on the multi-omics data; Step S3: Divide the labeled dataset into a training set and a test set. Use the training set data to perform supervised training of the algorithm framework and test it on the test set.

3. The disease prediction system based on variational neural networks according to claim 2, characterized in that, The supervised training process of the algorithm framework in step S3 includes the following steps: Step S3.1: The bacterial community abundance data, pathway data, and metabolite abundance data of the training set are compiled into three matrices, assuming they represent the first... Matrix of omics data Depend on one sample Composed of several features, each matrix first undergoes a linear transformation through a feature selection layer: , Where E represents the total number of training iterations, and e represents the iteration number. and These are the hyperparameters of the algorithm model, set to 10 and 0.1 respectively. These are the parameters of the fully connected layer. From uniform distribution Random sampling in the middle, This represents an activation function that, after the above linear transformation, becomes... get ; Step S3.2: The omics data, modified through the feature selection layer, is passed through an encoder consisting of fully connected layers and activation functions to obtain the latent representation of each omics. The calculation process is as follows: in This represents the nonlinear transformation process of a multilayer neural network. For the parameters of the neural network, Random sampling from a standard normal distribution; first by Obtain the mean of the latent representation and variance Then, the reparameterization technique is used to obtain the latent representation of each omics data. ; Step S3.3: Simply integrate incomplete multi-omics data with arbitrary missing values ​​using a joint omics encoder to obtain a joint latent representation. The calculation process is as follows: , in This represents the number of omics systems represented by the sample. and These represent the mean and variance of the prior distribution. Random sampling from a standard normal distribution; based on the mean of existing omics data. and variance The ensemble yields the joint omics mean. and variance Using reparameterization techniques to obtain potential representations of joint omics data ; Step S3.4: Potential representation of each omics data and joint potential representation The specific omics prediction probability values ​​are obtained by passing the predictor, which consists of a fully connected layer and an activation function, and the joint predictor, respectively. and the final predicted probability value The calculation process is as follows: in and This represents the nonlinear transformation process of a multilayer neural network. and These are all parameters of the neural network; Step S3.5: Calculate the loss based on the model's loss function, perform gradient backpropagation to update the parameters of the model's neural network, and the calculation process is as follows: in, and It is a balance coefficient that combines different losses. Describe the prior distribution, Indicates the first The distribution of potential representations of omics, Describe the distribution of the joint latent representation. This represents the calculation of the KL divergence, or relative entropy, between two distributions. True labels in one-hot format. Predict probability values ​​for specific omics. This is the final predicted probability value. For the sample size, The learning rate during algorithm training. These are the parameters of the neural network in the algorithm framework; thus, the neural network parameters are updated after one training cycle of the model.

4. The disease prediction system based on variational neural networks according to claim 1, characterized in that, The feature selection layer module is a Boolean matrix.

5. A disease prediction system based on a variational neural network according to claim 1, characterized in that, The encoder module is composed of a fully connected neural network.

6. A disease prediction system based on a variational neural network according to claim 1, characterized in that, The joint encoder module is a computing module.

7. A disease prediction system based on a variational neural network according to claim 1, characterized in that, The predictor and joint predictor modules are composed of fully connected neural networks.

Images