Metabolite-disease association prediction method and system
Through non-negative matrix decomposition and ZINB graph autoencoder extraction of metabolites and disease characteristics, the problem of existing methods ignoring the characteristics of input characteristics during feature extraction is solved, and the accuracy and reliability of metabolite-disease association prediction is significantly improved. .
Patent Information
- Application Number
- CN202411325535.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-23
- Publication Date
- 2024-12-27
AI Technical Summary
The existing metabolite-disease association prediction methods ignore the characteristics of input features during feature extraction, such as zero value and discrete distribution, resulting in the model being unable to fully capture the true structure and pattern of the data, affecting the accuracy and reliability of the prediction.
Using metabolite-disease association prediction method based on non-negative matrix decomposition and ZINB graph autoencoder, the non-negative matrix decomposition module and ZINB graph autoencoder are constructed by obtaining the comprehensive similarity between disease and metabolites, and combining with the graph volume Integrate neural networks and bilinear decoders, extract low-dimensional features of metabolites and diseases, and perform iterative optimization to improve prediction accuracy.
Effectively dealing with the discrete distribution and zero-value situation of features improves the accuracy and reliability of metabolite-disease association prediction, improves the model's performance in learning feature interaction, and enhances the ability to capture data patterns.
Smart Images
Figure CN119207775A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of bioinformatics, and in particular to a metabolite-disease association prediction method and system based on non-negative matrix decomposition and ZINB graph autoencoder. Background Art
[0002] The occurrence and development of diseases are usually related to abnormal levels of metabolites in the body. These metabolites include various small molecule compounds, such as amino acids, sugars, lipids, etc., which play a key role in the metabolic pathways in organisms. Traditional biological experimental methods require a lot of manpower to conduct repeated experiments to study the relationship between metabolites and diseases. This method is not only time-consuming and labor-intensive, but also inefficient. Therefore, developing efficient computational methods to predict the association between metabolites and diseases has become a research hotspot.
[0003] The latest solution is to use a deep learning-based method called MDA-AENMF (Predicting Metabolite–Disease Associations Based on Auto-Encoder and Non-Negative Matrix Factorization). This method combines three techniques, autoencoder (AutoEncoder), graph autoencoder (GAE), and non-negative matrix factorization (NMF), to extract features of metabolites and diseases. The autoencoder is used to learn a low-dimensional representation of the data, the graph autoencoder uses graph structure information to further extract features, and the non-negative matrix factorization helps maintain non-negative constraints in the data. Finally, a multi-layer perceptron (MLP) classifier is used to predict the association between metabolites and diseases.
[0004] Although existing autoencoder methods can extract metabolite and disease features to a certain extent, they usually only focus on the direct reconstruction of the input information, while ignoring the characteristics of the input features themselves. For example, the features input to the encoder may contain a large number of zero values and be discretely distributed, and the variance of these features may be smaller than the mean. These characteristics have an important impact on the final extracted features, but they are often not given enough attention in the current autoencoder framework. This neglect may cause the model to fail to fully capture the true structure and pattern of the data, thereby affecting the accuracy and reliability of the prediction. Summary of the invention
[0005] In view of this, the present invention provides a metabolite-disease association prediction method and system, which performs metabolite-disease association prediction based on non-negative matrix decomposition and ZINB graph autoencoder to fully capture the real structure and pattern of the data and improve the accuracy and reliability of the prediction.
[0006] To this end, the present invention provides the following technical solutions:
[0007] The present invention provides a metabolite-disease association prediction method based on non-negative matrix decomposition and ZINB graph autoencoder, comprising:
[0008] Obtain comprehensive disease similarity and comprehensive metabolite similarity;
[0009] A metabolite-disease association prediction model based on non-negative matrix factorization and zero-inflated negative binomial distribution ZINB graph autoencoder is constructed; the model includes: a non-negative matrix factorization NMF module, a ZINB-based graph autoencoder and a bilinear decoder; the non-negative matrix factorization NMF module is used to decompose the metabolite-disease association matrix into two low-rank matrices to obtain low-dimensional features of diseases and metabolites; the ZINB-based graph autoencoder includes: an encoder composed of a graph convolutional neural network GCN and a ZINB decoder, GCN encodes the metabolite comprehensive similarity matrix, the disease comprehensive similarity matrix and the metabolite-disease association matrix, the features encoded by GCN are input into the ZINB decoder to reconstruct the network output parameters, and the network output parameters are continuously iterated and optimized. After the optimization is completed, the final output of GCN is used as the feature of the disease and metabolite containing graph structure and data characteristics; the bilinear decoder is predicted based on the low-dimensional features of the disease and metabolites and the features of the disease and metabolites containing graph structure and data characteristics;
[0010] The metabolite-disease association matrix, the disease comprehensive similarity and the metabolite comprehensive similarity are input into the metabolite-disease association prediction model to obtain a prediction result of whether the metabolite and the disease are associated or not.
[0011] Furthermore, the comprehensive disease similarity is formed by integrating disease semantic similarity, disease Gaussian kernel similarity and disease similarity based on information entropy;
[0012] The comprehensive metabolite similarity is formed by integrating metabolite structural similarity, metabolite Gaussian kernel similarity and metabolite similarity based on information entropy.
[0013] Further, calculating the semantic similarity of the diseases includes: constructing a directed acyclic graph using medical subject terms as descriptors of the diseases, and calculating the semantic similarity of the diseases therefrom;
[0014] Calculating the disease Gaussian kernel similarity, including: inputting the disease related vector into the Gaussian kernel function to obtain the disease Gaussian kernel similarity;
[0015] Calculating disease similarity based on information entropy, including: using information entropy and common information between metabolites and diseases to obtain disease similarity based on information entropy;
[0016] Calculating the metabolite structural similarity, including: obtaining a vector describing the chemical properties of each metabolite, and then calculating the metabolite structural similarity based on the vector;
[0017] Calculating the metabolite Gaussian kernel similarity, including: inputting the metabolite related vector into the Gaussian kernel function to obtain the metabolite Gaussian kernel similarity;
[0018] Calculating metabolite similarity based on information entropy includes: using information entropy and common information between metabolites and diseases to obtain metabolite similarity based on information entropy.
[0019] Furthermore, the comprehensive disease similarity is formed by integrating disease semantic similarity, disease Gaussian kernel similarity and disease similarity based on information entropy, including:
[0020] Normalize the disease semantic similarity, disease Gaussian kernel similarity, and disease similarity based on information entropy;
[0021] The K nearest neighbor algorithm was used to calculate the local affinities between the diseases in the three disease similarity matrices, retaining the previously preset percentage of local affinities and setting the rest to 0;
[0022] The similarities of the three diseases are continuously updated iteratively until the termination condition is reached. After the iteration is completed, the similarities of the three diseases are added and averaged, and the matrix after the addition and average is added to its transpose and then averaged as the comprehensive similarity of the disease;
[0023] The comprehensive metabolite similarity is integrated by metabolite structural similarity, metabolite Gaussian kernel similarity and metabolite similarity based on information entropy, including:
[0024] The metabolite structural similarity, metabolite Gaussian kernel similarity, and information entropy-based metabolite similarity were normalized;
[0025] The K nearest neighbor algorithm was used to calculate the local affinities between metabolites in the three metabolite similarity matrices, retaining the previously preset percentage of local affinities and setting the rest to 0;
[0026] The similarities of the three metabolites are continuously updated iteratively until the termination condition is reached. After the iteration is completed, the similarities of the three metabolites are added and averaged, and the matrix after the addition and average is added to its transpose and then averaged as the comprehensive similarity of the metabolites.
[0027] Furthermore, the method also includes: testing the model.
[0028] Further, testing the model includes:
[0029] Model variants are constructed by removing the non-negative matrix factorization module, replacing the non-negative matrix factorization, removing the ZINB decoder, and replacing the ZINB decoder with a regular graph autoencoder to test the importance of different modules to the overall results of the model.
[0030] Furthermore, before the prediction, it also includes: using a 1:1 sampling method to randomly select negative samples with the same number as the positive samples.
[0031] The present invention also provides a metabolite-disease association prediction system based on non-negative matrix decomposition and ZINB graph autoencoder, comprising:
[0032] A similarity acquisition module is used to obtain comprehensive disease similarity and comprehensive metabolite similarity;
[0033] A model building module is used to build a metabolite-disease association prediction model based on non-negative matrix factorization and zero-inflated negative binomial distribution ZINB graph autoencoder; the model includes: a non-negative matrix factorization NMF module, a ZINB-based graph autoencoder and a bilinear decoder; the non-negative matrix factorization NMF module is used to decompose the metabolite-disease association matrix into two low-rank matrices to obtain low-dimensional features of diseases and metabolites; the ZINB-based graph autoencoder includes: an encoder composed of a graph convolutional neural network GCN and a ZINB decoder, GCN encodes the metabolite comprehensive similarity matrix, the disease comprehensive similarity matrix and the metabolite-disease association matrix, the features encoded by GCN are input into the ZINB decoder to reconstruct the network output parameters, and the network output parameters are continuously iterated and optimized. After the optimization is completed, the final output of GCN is used as the feature of the disease and metabolite containing graph structure and data characteristics; the bilinear decoder is predicted based on the low-dimensional features of the disease and metabolites and the features of the disease and metabolites containing graph structure and data characteristics;
[0034] The prediction module is used to input the metabolite-disease association matrix, the disease comprehensive similarity and the metabolite comprehensive similarity into the metabolite-disease association prediction model to obtain a prediction result of whether the metabolite and the disease are associated or not.
[0035] Advantages and positive effects of the present invention:
[0036] In the present invention, first, by integrating multiple similarities of metabolites and diseases, metabolite comprehensive similarity and disease comprehensive similarity networks are constructed respectively, laying a solid foundation for subsequent feature extraction. Secondly, non-negative matrix decomposition and graph autoencoder based on ZINB (zero-inflated negative binomial distribution) are used for feature extraction, which enables the model to effectively handle the discrete distribution of features, a large number of zero values, and the situation where the variance is less than the mean. In addition, the extracted metabolites and disease features are spliced and input into the bilinear decoder, further improving the performance of the model in learning feature interaction, thereby improving the accuracy of metabolite-disease association prediction. Compared with common deep learning methods, the introduction of the ZINB decoder of the present invention takes full account of data characteristics and provides strong support for improving prediction accuracy.
[0037] In summary, the present invention has significant advantages in improving the accuracy of metabolite-disease association prediction, and is of great value for the discovery of actual diseases, judgment of disease development stages, and subsequent treatment, and is expected to promote research and application development in related fields. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.
[0039] Figure 1 It is a flowchart of a metabolite-disease association prediction method based on non-negative matrix factorization and ZINB graph autoencoder in an embodiment of the present invention;
[0040] Figure 2 Schematic diagram of the architecture of the ZINB-based graph autoencoder in an embodiment of the present invention. DETAILED DESCRIPTION
[0041] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.
[0042] It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.
[0043] like Figure 1 As shown, in an embodiment of the present invention, a metabolite-disease association prediction method based on non-negative matrix decomposition and ZINB graph autoencoder is proposed. The method includes three parts: data preparation, model building and model testing.
[0044] (I) Data preparation:
[0045] Step 11: Obtain the initial similarity data of diseases and metabolites. Use medical subject terms to construct a directed acyclic graph as a descriptor of the disease, from which the semantic similarity of the disease is calculated; input the SMELLS of the metabolites into the PaDEL-Descriptor software to obtain vectors describing the chemical properties of each metabolite, and then calculate the structural similarity of the metabolites based on these vectors; according to the Gaussian kernel function, input the disease and metabolite related vectors respectively, and obtain the disease Gaussian kernel similarity and metabolite Gaussian kernel similarity respectively; using information entropy and the common information between diseases and metabolites, we can obtain information-based disease similarity and information-based metabolite similarity.
[0046] Step 12: Integrate the three similarities of disease and metabolites into a comprehensive similarity. Taking disease as an example, the three similarities obtained in step 1 are normalized; then the K nearest neighbor algorithm (KNN) is used to calculate the local affinities between the diseases in the three similarity matrices respectively, retaining the top 10% of the local affinities and setting the rest to 0; the three similarities are continuously iterated and updated until the termination condition is reached. After the iteration is completed, they are added and averaged, and the matrix after the addition and average is added and averaged with its transpose and then averaged as the comprehensive similarity of the disease. The comprehensive similarity of metabolites is also obtained using the same method.
[0047] (II) Model building:
[0048] Step 21: Negative sample extraction.
[0049] Due to the imbalance of positive and negative samples, a 1:1 sampling method is adopted to randomly select negative samples with the same number of positive samples.
[0050] Step 22: Non-negative matrix factorization.
[0051] In order to obtain low-dimensional features of diseases and metabolites, the metabolite-disease association matrix is decomposed into two low-rank matrices through non-negative matrix factorization (NMF), and the product of these two low-rank matrices should be as close as possible to the original undecomposed matrix. Continuously iterate and optimize, when the relative difference standard meets the set conditions, the two low-rank matrices obtained at this time are the low-dimensional features of diseases and metabolites.
[0052] Step 23: Graph autoencoder based on zero-inflated negative binomial distribution (ZINB).
[0053] In order to obtain the features of diseases and metabolites containing graph structure and data characteristics, the metabolite comprehensive similarity matrix, disease comprehensive similarity matrix and metabolite-disease association matrix were input into the ZINB-based graph autoencoder. Figure 2 As shown in the figure, the ZINB-based graph autoencoder consists of an encoder composed of GCN and a ZINB decoder. The features encoded by GCN are input into the ZINB decoder to reconstruct the dropout probability, mean and dispersion of the network output, and the parameters of these outputs are continuously optimized iteratively. After the optimization is completed, the final output of GCN is used as the features of diseases and metabolites containing graph structures and data characteristics.
[0054] Step 24: The extracted features are concatenated and sent to the bilinear decoder for prediction.
[0055] (III) Model testing:
[0056] Step 31: Use various similarity calculation methods to obtain metabolite and disease similarities, and integrate them to obtain comprehensive similarity. Use the non-negative matrix decomposition module to decompose the metabolite-disease association matrix to obtain low-dimensional features of metabolites and diseases. Input the metabolite comprehensive similarity matrix, disease comprehensive similarity matrix and metabolite-disease association matrix into the ZINB-based graph autoencoder to obtain features containing graph structure and data characteristics. Finally, the features extracted by the two modules are concatenated and the scores are calculated through a bilinear decoder.
[0057] Step 32: Test various parameters and important modules that affect model performance. Observe the impact of the non-negative matrix factorization module, ZINB decoder, etc. in the model on the model test results. Develop model variants such as removing the non-negative matrix factorization module, replacing the non-negative matrix factorization, removing the ZINB decoder, and replacing the ZINB decoder with a regular graph autoencoder to test the importance of different modules to the overall model results.
[0058] Step 33: Test the model’s ability to identify potential metabolite-disease associations. For several common diseases in real life, predict and observe the model’s ability to identify relevant metabolites.
[0059] In the above embodiment, multiple similarities between metabolites and diseases are first obtained and integrated into comprehensive metabolite similarities and comprehensive disease similarities, respectively. Then, the features of metabolites and diseases are extracted by non-negative matrix decomposition and a graph autoencoder based on ZINB, and finally the extracted features are spliced and input into a bilinear decoder for metabolite-disease association prediction. Compared with the common deep learning method, the present invention introduces a ZINB decoder, which can fully consider the discrete distribution of features and the presence of a large number of zero values, as well as the characteristics that the variance of the features is less than the mean when extracting features. The method of the present invention helps to improve the accuracy of metabolite-disease association prediction, and has certain value for actual disease discovery and judgment of the stage of disease development and subsequent treatment of the disease.
[0060] Corresponding to the metabolite-disease association prediction method in the above embodiment, this embodiment further provides a metabolite-disease association prediction system based on non-negative matrix decomposition and ZINB graph autoencoder, comprising:
[0061] A similarity acquisition module is used to obtain comprehensive disease similarity and comprehensive metabolite similarity;
[0062] A model building module is used to build a metabolite-disease association prediction model based on non-negative matrix factorization and zero-inflated negative binomial distribution ZINB graph autoencoder; the model includes: a non-negative matrix factorization NMF module, a ZINB-based graph autoencoder and a bilinear decoder; the non-negative matrix factorization NMF module is used to decompose the metabolite-disease association matrix into two low-rank matrices to obtain low-dimensional features of diseases and metabolites; the ZINB-based graph autoencoder includes: an encoder composed of a graph convolutional neural network GCN and a ZINB decoder, GCN encodes the metabolite comprehensive similarity matrix, the disease comprehensive similarity matrix and the metabolite-disease association matrix, the features encoded by GCN are input into the ZINB decoder to reconstruct the network output parameters, and the network output parameters are continuously iterated and optimized. After the optimization is completed, the final output of GCN is used as the feature of the disease and metabolite containing graph structure and data characteristics; the bilinear decoder is predicted based on the low-dimensional features of the disease and metabolites and the features of the disease and metabolites containing graph structure and data characteristics;
[0063] The prediction module is used to input the metabolite-disease association matrix, the disease comprehensive similarity and the metabolite comprehensive similarity into the metabolite-disease association prediction model to obtain a prediction result of whether the metabolite and the disease are associated or not.
[0064] For the metabolite-disease association prediction system of the embodiment of the present invention, since it corresponds to the metabolite-disease association prediction method in the above embodiment, the description is relatively simple. For relevant similarities, please refer to the description of the metabolite-disease association prediction method in the above embodiment, which will not be described in detail here.
[0065] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A metabolite-disease association prediction method based on non-negative matrix factorization and ZINB graph autoencoder, characterized in that: include: Obtain comprehensive disease similarity and comprehensive metabolite similarity; A metabolite-disease association prediction model based on non-negative matrix factorization and zero-inflated negative binomial distribution ZINB graph autoencoder is constructed; the model includes: a non-negative matrix factorization NMF module, a ZINB-based graph autoencoder and a bilinear decoder; the non-negative matrix factorization NMF module is used to decompose the metabolite-disease association matrix into two low-rank matrices to obtain low-dimensional features of diseases and metabolites; the ZINB-based graph autoencoder includes: an encoder composed of a graph convolutional neural network GCN and a ZINB decoder, GCN encodes the metabolite comprehensive similarity matrix, the disease comprehensive similarity matrix and the metabolite-disease association matrix, the features encoded by GCN are input into the ZINB decoder to reconstruct the network output parameters, and the network output parameters are continuously iterated and optimized. After the optimization is completed, the final output of GCN is used as the feature of the disease and metabolite containing graph structure and data characteristics; the bilinear decoder is predicted based on the low-dimensional features of the disease and metabolites and the features of the disease and metabolites containing graph structure and data characteristics; The metabolite-disease association matrix, the disease comprehensive similarity and the metabolite comprehensive similarity are input into the metabolite-disease association prediction model to obtain a prediction result of whether the metabolite and the disease are associated or not.
2. A metabolite-disease association prediction method based on non-negative matrix factorization and ZINB graph autoencoder according to claim 1, characterized in that: The comprehensive disease similarity is formed by integrating disease semantic similarity, disease Gaussian kernel similarity and disease similarity based on information entropy; The comprehensive metabolite similarity is formed by integrating metabolite structural similarity, metabolite Gaussian kernel similarity and metabolite similarity based on information entropy.
3. A metabolite-disease association prediction method based on non-negative matrix factorization and ZINB graph autoencoder according to claim 2, characterized in that: Calculating the semantic similarity of diseases, including: using medical subject terms to construct a directed acyclic graph as a descriptor of the disease, and calculating the semantic similarity of the disease; Calculating the disease Gaussian kernel similarity, including: inputting the disease related vector into the Gaussian kernel function to obtain the disease Gaussian kernel similarity; Calculating disease similarity based on information entropy, including: using information entropy and common information between metabolites and diseases to obtain disease similarity based on information entropy; Calculating the metabolite structural similarity, including: obtaining a vector describing the chemical properties of each metabolite, and then calculating the metabolite structural similarity based on the vector; Calculating the metabolite Gaussian kernel similarity, including: inputting the metabolite related vector into the Gaussian kernel function to obtain the metabolite Gaussian kernel similarity; Calculating metabolite similarity based on information entropy includes: using information entropy and common information between metabolites and diseases to obtain metabolite similarity based on information entropy.
4. A metabolite-disease association prediction method based on non-negative matrix factorization and ZINB graph autoencoder according to claim 2 or 3, characterized in that: The comprehensive disease similarity is formed by integrating disease semantic similarity, disease Gaussian kernel similarity and disease similarity based on information entropy, including: Normalize the disease semantic similarity, disease Gaussian kernel similarity, and disease similarity based on information entropy; The K nearest neighbor algorithm was used to calculate the local affinities between the diseases in the three disease similarity matrices, retaining the previously preset percentage of local affinities and setting the rest to 0; The similarities of the three diseases are continuously updated iteratively until the termination condition is reached. After the iteration is completed, the similarities of the three diseases are added and averaged, and the matrix after the addition and average is added to its transpose and then averaged as the comprehensive similarity of the disease; The comprehensive metabolite similarity is integrated by metabolite structural similarity, metabolite Gaussian kernel similarity and metabolite similarity based on information entropy, including: The metabolite structural similarity, metabolite Gaussian kernel similarity, and information entropy-based metabolite similarity were normalized; The K nearest neighbor algorithm was used to calculate the local affinities between metabolites in the three metabolite similarity matrices, retaining the previously preset percentage of local affinities and setting the rest to 0; The similarities of the three metabolites are continuously updated iteratively until the termination condition is reached. After the iteration is completed, the similarities of the three metabolites are added and averaged, and the matrix after the addition and average is added to its transpose and then averaged as the comprehensive similarity of the metabolites.
5. The method for predicting metabolite-disease association based on non-negative matrix factorization and ZINB graph autoencoder according to claim 1, characterized in that: Also includes: The model was tested.
6. The method for predicting metabolite-disease association based on non-negative matrix factorization and ZINB graph autoencoder according to claim 5, characterized in that: Testing the model includes: Model variants are constructed by removing the non-negative matrix factorization module, replacing the non-negative matrix factorization, removing the ZINB decoder, and replacing the ZINB decoder with a regular graph autoencoder to test the importance of different modules to the overall results of the model.
7. The method for predicting metabolite-disease association based on non-negative matrix factorization and ZINB graph autoencoder according to claim 1, characterized in that: Before prediction, it also includes: using a 1:1 sampling method to randomly select negative samples with the same number of positive samples.
8. A metabolite-disease association prediction system based on non-negative matrix factorization and ZINB graph autoencoder, characterized in that: include: A similarity acquisition module is used to obtain comprehensive disease similarity and comprehensive metabolite similarity; A model building module is used to build a metabolite-disease association prediction model based on non-negative matrix factorization and zero-inflated negative binomial distribution ZINB graph autoencoder; the model includes: a non-negative matrix factorization NMF module, a ZINB-based graph autoencoder and a bilinear decoder; the non-negative matrix factorization NMF module is used to decompose the metabolite-disease association matrix into two low-rank matrices to obtain low-dimensional features of diseases and metabolites; the ZINB-based graph autoencoder includes: an encoder composed of a graph convolutional neural network GCN and a ZINB decoder, GCN encodes the metabolite comprehensive similarity matrix, the disease comprehensive similarity matrix and the metabolite-disease association matrix, the features encoded by GCN are input into the ZINB decoder to reconstruct the network output parameters, and the network output parameters are continuously iterated and optimized. After the optimization is completed, the final output of GCN is used as the feature of the disease and metabolite containing graph structure and data characteristics; the bilinear decoder is predicted based on the low-dimensional features of the disease and metabolites and the features of the disease and metabolites containing graph structure and data characteristics; The prediction module is used to input the metabolite-disease association matrix, the disease comprehensive similarity and the metabolite comprehensive similarity into the metabolite-disease association prediction model to obtain a prediction result of whether the metabolite and the disease are associated or not.