Method for predicting multiple degenerative diseases and screening markers based on DNA methylation

Abstract

The invention provides a method for predicting multiple degenerative diseases and screening markers based on DNA methylation, and belongs to the technical field of biological information. The method comprises the following steps: firstly, acquiring DNA methylation data from a GEO database, and screening out CpG loci with non-zero feature importance through a random forest algorithm; performing multi-scale feature extraction on the screened CpG sites by using a multi-scale one-dimensional convolutional neural network, and capturing local and global features; then, 110 key CpG loci related to various degenerative diseases are screened out through interaction network analysis of GO, KEGG and protein; and finally, constructing a multi-classification disease prediction model based on the residual network, and performing classification prediction on neurodegenerative diseases, bone degenerative diseases, breast cancers and health states. According to the invention, the screening accuracy of degenerative diseases is improved, a scientific biomarker screening basis is provided, and support of precise medical treatment and early intervention is facilitated.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics, and in particular to a method for predicting multiple degenerative diseases and screening markers based on DNA methylation. Background Art

[0002] With the aging of the population, degenerative diseases such as cancer, Parkinson's disease, Alzheimer's disease, rheumatoid arthritis, etc. have become important health problems worldwide. These diseases are closely related to aging, and their pathogenesis involves a variety of physiological changes and gene regulatory mechanisms. DNA methylation, as an important epigenetic modification, plays a key role in human development, gene expression regulation, and disease-related processes. Studies have shown that the state of DNA methylation changes significantly with age. It can not only serve as a marker of biological age, but is also closely related to the occurrence and development of a variety of age-related diseases such as autoimmune diseases, neurological diseases, and cancer. Therefore, early diagnosis and screening of degenerative diseases based on DNA methylation data is of great significance in medical research.

[0003] In the existing technology, linear regression and machine learning methods are widely used in the diagnosis of cancer and degenerative diseases. For example, methods such as Lasso regression, semi-supervised learning framework, and deep neural network have been used to predict multiple single diseases and have achieved good prediction results. However, existing studies usually focus on a single disease and ignore the possible common mechanisms between different degenerative diseases. In addition, the research of traditional biomarkers is also mostly focused on a single type of disease, which limits its application in multiple diseases.

[0004] At present, there are few studies on multi-classification prediction of degenerative diseases, and most studies use small-scale DNA methylation data, which limits the generalizability and accuracy of the models. At the same time, many methods in the prior art have performance bottlenecks when processing high-dimensional data, especially in the process of screening representative CpG sites for multi-disease prediction, which often face problems such as low computational efficiency and feature redundancy. In addition, most existing biomarker identification methods are only applicable to a few diseases and cannot effectively cover a wide range of degenerative diseases. Therefore, the prior art is insufficient in exploring the common mechanisms of multiple degenerative diseases, screening of widely applicable biomarkers, and efficient feature selection. Summary of the invention

[0005] In view of the problems existing in the above-mentioned prior art, the present invention provides a method for predicting multiple degenerative diseases and screening markers based on DNA methylation.

[0006] The present invention proposes a method for predicting multiple degenerative diseases and screening markers based on DNA methylation, comprising the following steps:

[0007] S1: DNA methylation datasets collected from the GEO database;

[0008] S2: Random forest algorithm was used to screen CpG sites and retain CpG sites with non-zero feature importance;

[0009] S3: Based on a multi-scale one-dimensional convolutional neural network, the feature combination of the optimal CpG site is determined by gradually increasing the CpG site input;

[0010] S4: CpG sites were analyzed by bio-omics analysis through Gene Ontology (GO) functional enrichment analysis, Kyoto Encyclopedia of Genes and Genomes (KEGG) signaling pathway enrichment analysis, and STRING protein interaction network analysis to screen out key CpG sites;

[0011] S5: Construct a multi-classification degenerative disease prediction model based on residual network, and the model outputs the classification probabilities of neurodegenerative diseases, bone degenerative diseases, breast cancer and health status;

[0012] S6: Louvain algorithm cluster analysis was used to verify the classification effect and disease association of CpG sites;

[0013] S7: Apply SHAP values ​​to interpret model output, quantify the contribution of CpG sites to disease prediction, and improve model interpretability.

[0014] Furthermore, the step S1 includes:

[0015] S11: 142 DNA methylation datasets were obtained from the GEO database, blood samples were retained, and 39 were selected for analysis;

[0016] S12: 12,578 common CpG sites were extracted from 39 data sets and divided into training and test sets in a ratio of 8:2 for feature screening.

[0017] Furthermore, the step S2 includes:

[0018] S21: 12,578 CpG sites were used as input features of the random forest model to evaluate the feature importance of each CpG site and screen out CpG sites that contribute significantly to disease prediction;

[0019] S22: sites with zero feature importance were removed, and 6438 CpG sites with non-zero feature importance were retained;

[0020] S23: The screening effect is evaluated by accuracy, precision, recall and MCC evaluation indicators.

[0021] Furthermore, the step S3 includes:

[0022] S31: Use a multi-scale one-dimensional convolutional neural network to extract features from 6438 CpG sites and capture local and global features;

[0023] S32: Through five-fold cross validation, the number of CpG sites was gradually increased to evaluate the impact of different feature combinations on model performance;

[0024] S33: The results showed that the model performed best when 600 CpG sites were input; 600 optimal CpG sites were eventually determined, providing a basis for the construction of subsequent disease prediction models.

[0025] Further, the step S4 includes:

[0026] S41: Functionally annotate 600 CpG sites through GO functional enrichment analysis to reveal their biological functions and identify key biological processes related to degenerative diseases;

[0027] S42: KEGG signaling pathway analysis revealed the enrichment of CpG sites in multiple disease-related pathways;

[0028] S43: Use the STRING tool to construct a protein interaction network and identify key protein interaction modules that are highly related to degenerative diseases;

[0029] S44: Through GO, KEGG and STRING analysis, 110 key CpG sites related to degenerative diseases were finally screened out, revealing their potential biological mechanisms.

[0030] Furthermore, the step S5 comprises:

[0031] S51: 15 DNA methylation datasets were selected from the GEO database, covering four categories: neurodegenerative diseases, bone degenerative diseases, breast cancer and healthy samples. The datasets were balanced and adjusted to construct a multi-classification disease prediction model.

[0032] S52: Based on 110 key CpG sites, a multi-classification disease prediction model was constructed, and local and global features were extracted through residual networks;

[0033] S53: The features are reduced in dimension through the global average pooling layer, and finally the classification probabilities of the four types of diseases are output through the Softmax activation layer.

[0034] Further, the step S6 includes:

[0035] S61: Unsupervised cluster analysis of the multi-classification prediction results of 110 CpG sites was performed using the Louvain algorithm to evaluate the clustering of degenerative disease samples and the distribution of disease types;

[0036] S62: Clustering results verify the effectiveness of CpG sites in distinguishing different diseases and reveal the similarities and potential correlations of diseases in molecular characteristics.

[0037] Further, the step S7 includes:

[0038] S71: After cluster analysis, the SHAP value was used to interpret the model output, quantify the contribution of each CpG site to disease classification, and identify the key sites that contributed most to the prediction;

[0039] S72: Through SHAP analysis, we further reveal the role of key CpG sites in disease classification and their biological mechanisms, improve model interpretability and provide theoretical support for biomarkers.

[0040] The present invention has the following beneficial effects:

[0041] (1) The present invention combines DNA methylation data with machine learning and deep learning methods to achieve accurate prediction of multiple degenerative diseases. By screening out 110 key CpG sites, the present invention significantly improves the prediction ability of the model, with a classification accuracy of 0.984, which has the advantages of high precision and high efficiency, and provides a reliable tool for the early diagnosis of multiple degenerative diseases.

[0042] (2) The present invention effectively screens and optimizes CpG sites that contribute significantly to disease prediction by combining random forests and multi-scale one-dimensional convolutional neural networks, reduces feature redundancy, improves the computational efficiency of the model, and reduces the complexity of data processing.

[0043] (3) The present invention uses the Louvain algorithm for cluster analysis and interprets the model output through SHAP value to quantify the contribution of each CpG site to disease prediction, thereby improving the interpretability of the model, revealing the potential biological mechanism of CpG sites, and providing a theoretical basis for the accurate classification of diseases and the effectiveness of biomarkers.

[0044] (4) The present invention adopts a biomarker screening method that is widely applicable to a variety of degenerative diseases. It can not only predict a single disease, but also effectively identify multiple diseases with a common mechanism of action, thereby expanding the application scope of existing technologies and having significant clinical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1The overall technical flow chart of the method for predicting various degenerative diseases and screening markers of the present invention;

[0046] Figure 2 Schematic diagram of the architecture of the degenerative disease diagnosis model based on a multi-scale one-dimensional convolutional neural network of the present invention;

[0047] Figure 3 Schematic diagram of the structure of the multi-classification degenerative disease prediction model of the present invention. DETAILED DESCRIPTION

[0048] In order to more clearly understand the above-mentioned objects, features and advantages of the present invention, the present invention is described in detail below with reference to the accompanying drawings and specific embodiments.

[0049] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. The embodiments described are only a part of the embodiments of the present invention, rather than 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 are within the scope of protection of the present invention.

[0050] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the present invention herein are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0051] One embodiment of the present invention, as Figure 1 As shown, a method for predicting multiple degenerative diseases and screening markers based on DNA methylation, the method comprising:

[0052] S1: DNA methylation datasets collected from the GEO database;

[0053] S2: Random forest algorithm was used to screen CpG sites and retain CpG sites with non-zero feature importance;

[0054] S3: Based on a multi-scale one-dimensional convolutional neural network, the feature combination of the optimal CpG site is determined by gradually increasing the CpG site input;

[0055] S4: CpG sites were analyzed by bio-omics analysis through Gene Ontology (GO) functional enrichment analysis, Kyoto Encyclopedia of Genes and Genomes (KEGG) signaling pathway enrichment analysis, and STRING protein interaction network analysis to screen out key CpG sites;

[0056] S5: Construct a multi-classification degenerative disease prediction model based on residual network, and the model outputs the classification probabilities of neurodegenerative diseases, bone degenerative diseases, breast cancer and health status;

[0057] S6: Louvain algorithm cluster analysis was used to verify the classification effect and disease association of CpG sites;

[0058] S7: Apply SHAP values ​​to interpret model output, quantify the contribution of CpG sites to disease prediction, and improve model interpretability.

[0059] The working principle of the above technical solution is as follows: First, 142 publicly available DNA methylation datasets were collected from the GEO database, covering the Illumina Infinium HumanMethylation27K, 450K and EPIC platforms. In order to ensure the representativeness of the samples, we retained samples derived from blood tissues, and finally screened out 39 datasets for further analysis. In subsequent steps, 12,578 common CpG sites were extracted from these 39 datasets. These sites will be used in the feature screening process of the random forest algorithm and the multi-scale one-dimensional convolutional neural network (MSDCNN) to evaluate their contribution to disease prediction, and help identify key features that have an important impact on disease prediction.

[0060] Next, the random forest algorithm was used to perform feature screening on the extracted 12,578 CpG sites. The importance of each site to disease prediction was evaluated through the random forest model, and invalid sites were eliminated based on the feature importance score. Finally, 6,438 CpG sites that made significant contributions to disease prediction were retained. The screening successfully reduced redundant features, improved the efficiency of model calculation, and maintained the accuracy of model prediction.

[0061] Subsequently, a multi-scale one-dimensional convolutional neural network (MSDCNN) was used to further extract features from the 6438 retained CpG sites. MSDCNN extracts local and global features of CpG sites through convolution kernels of different sizes, thereby capturing the complex nonlinear relationships between sites. During the model training process, the number of input CpG sites was gradually increased through five-fold cross-validation, and the impact of different feature combinations on model performance was evaluated, and finally the 600 best-performing CpG sites were determined.

[0062] Next, we conducted a bio-omics analysis on the 600 CpG sites that were screened. Through GO functional enrichment analysis, KEGG signaling pathway analysis, and STRING protein interaction network analysis, we deeply revealed the biological processes, cellular components, and molecular functions involved in these CpG sites. Through these analyses, we screened out 110 key CpG sites that are closely related to a variety of degenerative diseases. These sites have important biological significance in the occurrence and development of a variety of diseases, revealing the potential molecular mechanisms of the diseases.

[0063] After completing feature screening and biological analysis, a multi-classification degenerative disease prediction model ResDegNet based on residual network (ResNet) was constructed. The model takes the 110 selected key CpG sites as input, first extracts features from DNA methylation data through a one-dimensional convolutional layer, and then further extracts local and global features through three residual modules to ensure the effective transmission of information in the deep network. The residual connection mechanism solves the gradient vanishing problem that may occur in the deep learning process, and improves the training efficiency and prediction accuracy of the model. Finally, through the fully connected layer and Softmax activation function, the model outputs the classification probabilities of neurodegenerative diseases, bone degenerative diseases, breast cancer and health status. The model uses the category with the highest probability value among the four types of diseases as the prediction result, that is, determines the type of disease to which the sample belongs.

[0064] In order to verify the classification effect of the model, the Louvain algorithm was used to perform unsupervised cluster analysis on the model output results. Cluster analysis was used to observe the cluster distribution of different degenerative disease samples in the feature space, further verifying the effectiveness of these CpG sites in disease classification. At the same time, the distribution relationship between samples was analyzed to reveal the potential correlation between different diseases.

[0065] Finally, to improve the interpretability of the model, SHAP values ​​were used to interpret the model output results. The contribution of each CpG site to disease classification was quantified by SHAP values, and CpG sites that made important contributions to different disease classifications were identified. This analysis process not only improved the interpretability of the model, but also provided theoretical support for the application of CpG sites as potential biomarkers in degenerative diseases.

[0066] The effect of the above technical solution is: by collecting DNA methylation data from the GEO database and screening out key CpG sites, the present invention effectively improves the accuracy and efficiency of the degenerative disease prediction model. By screening features through random forests and multi-scale one-dimensional convolutional neural networks, the computational efficiency of the model is optimized while maintaining high-precision prediction capabilities. The application of residual networks solves the gradient vanishing problem in deep learning and further enhances the generalization ability of the model. Biomics analysis reveals the potential mechanism of CpG sites in disease occurrence and provides an important basis for early diagnosis and personalized treatment of diseases. The application of SHAP values ​​provides interpretability for the model, helps identify the key sites that contribute most to disease prediction, and enhances the potential for clinical application of this method.

[0067] One embodiment of the present invention, as Figure 2 As shown, the multi-scale one-dimensional convolutional neural network determines the feature combination of the optimal CpG site by gradually increasing the CpG site input; including:

[0068] A multi-scale one-dimensional convolutional neural network was used to extract features from 6438 CpG sites to capture local and global features;

[0069] Through five-fold cross validation, the number of CpG sites was gradually increased to evaluate the impact of different feature combinations on model performance;

[0070] The results showed that the model performed best when 600 CpG sites were input; 600 optimal CpG sites were eventually determined, providing a basis for the construction of subsequent disease prediction models.

[0071] The working principle of the above technical solution is: 6438 CpG sites that were initially screened are processed through a multi-scale one-dimensional convolutional neural network (MSDCNN). The core of the design of MSDCNN is the use of convolution kernels of different sizes (3, 5, 7), which enables the model to capture both the local detail features and the global pattern of CpG sites, and then extract the complex nonlinear relationship in the DNA methylation pattern. Smaller convolution kernels can capture local fine features, while larger convolution kernels help learn features in a larger range. This multi-scale structure improves the model's ability to identify complex disease features.

[0072] In the specific operation, we gradually optimized the model performance through five-fold cross validation, starting from 150 CpG sites, and increasing 150 sites each time. The model made predictions and evaluated the effects of different input feature combinations until 6438 sites were used to complete the test. This process helps to discover the impact of different feature combinations on the model prediction effect, and gradually optimizes the selection of CpG sites by evaluating the accuracy, precision, recall and MCC indicators.

[0073] The results of cross-validation show that when 600 CpG sites are input, the performance indicators of the model reach the best level. Compared with using all 6438 sites, the combination of 600 CpG sites not only ensures the prediction accuracy of the model, but also performs better in terms of precision and MCC indicators. Although the recall rate has dropped slightly, the overall performance remains at a high level, indicating that the synergistic effect of these 600 sites in disease prediction is more significant.

[0074] In addition, by gradually increasing the number of CpG sites for testing, MSDCNN can flexibly capture the interactions between sites, thereby effectively optimizing the feature selection process of the model. In the end, the 600 key CpG sites screened by MSDCNN not only improved the computational efficiency of the model, but also reduced the interference of redundant features, significantly improving the accuracy of predicting disease types.

[0075] These 600 CpG sites will be further screened in subsequent GO functional enrichment analysis, KEGG signaling pathway analysis, and protein interaction network analysis. This will ensure that the biological functions of these sites in different degenerative diseases have important disease relevance, providing a strong basis for the early diagnosis and treatment of subsequent diseases.

[0076] The effect of the above technical solution is: through the multi-scale one-dimensional convolutional neural network (MSDCNN), the local and global characteristics of CpG sites are effectively captured, and the model's ability to recognize complex disease characteristics is enhanced. By gradually increasing the number of input CpG sites through five-fold cross-validation, the model achieved the best prediction performance at 600 CpG sites, significantly improving the accuracy, precision and MCC indicators of the model. Compared with using all 6438 sites, the screening of 600 key CpG sites not only reduces redundant features and optimizes computational efficiency, but also maintains a high recall rate and overall performance. This process ensures the synergistic effect of the screened CpG sites in disease prediction, providing a more accurate basis for subsequent biological analysis and disease diagnosis.

[0077] One embodiment of the present invention, as Figure 3 As shown, the multi-classification degenerative disease prediction model based on the residual network is constructed, and the model outputs the classification probabilities of neurodegenerative diseases, bone degenerative diseases, breast cancer and health status; including:

[0078] We selected 15 DNA methylation datasets from the GEO database, covering four categories: neurodegenerative diseases, bone degenerative diseases, breast cancer, and healthy samples. We then balanced and adjusted the datasets to build a multi-classification disease prediction model.

[0079] Based on 110 key CpG sites, a multi-classification disease prediction model was constructed, and local and global features were extracted through residual networks;

[0080] The features are reduced in dimension through the global average pooling layer, and finally the classification probabilities of the four types of diseases are output through the Softmax activation layer.

[0081] The working principle of the above technical solution is as follows: First, the model is based on 15 DNA methylation data sets selected from the GEO database. After processing, these data sets only retain the 110 key CpG sites previously screened as input. The DNA methylation expression matrix is ​​feature extracted through a one-dimensional convolutional layer. The function of the one-dimensional convolutional layer is to extract local features of different dimensions from the DNA methylation data, and perform preliminary dimensionality reduction and feature analysis on the input data through convolution operations. In order to further reduce the computational complexity, the convolutional layer is connected to the MaxPooling layer, which aggregates information by reducing the size of the feature map, thereby improving computational efficiency.

[0082] Next, the model further extracts features through three residual modules. In each residual module, the convolution layer not only extracts local features, but also maintains the information flow between input features and output features through residual connections. The role of residual connection is to add the input signal directly to the output, avoiding the problem of gradient vanishing and improving the stability of the network. Through the residual module, the model can gradually extract deep-level features and global features between CpG sites, and then accurately capture the DNA methylation patterns under different disease states.

[0083] After feature extraction, the model uses a global average pooling layer to reduce the dimensionality of multi-dimensional features. This layer compresses high-dimensional data into a low-dimensional feature vector by taking a global average of the feature maps output by each convolution kernel. This operation not only reduces the risk of overfitting, but also improves the computational efficiency of the model. Finally, the model outputs probability values ​​for four categories through a fully connected layer and uses a Softmax activation function, corresponding to neurodegenerative diseases, bone degenerative diseases, breast cancer, and health status. The Softmax function converts the output of the model into a probability distribution, and the probability value of each category is between 0 and 1, thereby ensuring the interpretability of the output results. Finally, the model uses the category with the highest probability value among the four categories of diseases as the prediction result, that is, determines the type of disease to which the sample belongs.

[0084] In order to verify the performance of the model, the present invention evaluates the model through five-fold cross validation and independent test sets. In cross validation, the data set is divided into five parts, four of which are selected for training each time, and one is used for validation, ensuring the generalization ability and stability of the model on different data sets. Through cross validation, the performance of the model has been comprehensively evaluated, including classification accuracy, precision, recall rate and F1 value indicators, all of which have reached a high level. In particular, the classification accuracy rate reached 0.984, which proved the predictive ability of the model. At the same time, the model was also evaluated on an independent test set to further verify its generalization ability on unknown data.

[0085] The effect of the above technical solution is: by constructing a multi-classification degenerative disease prediction model based on a residual network, the model can effectively extract local and global features in DNA methylation data and enhance the ability to capture complex nonlinear relationships. The use of residual modules improves the stability of the deep network, avoids the gradient vanishing problem, and improves the training efficiency and prediction accuracy of the model. Through five-fold cross-validation and independent test set evaluation, the superior performance of the model in different disease classifications has been demonstrated, especially the classification accuracy of 0.984, showing the significant role of 110 key CpG sites in disease prediction. The model can not only accurately distinguish between neurodegenerative diseases, bone degenerative diseases, breast cancer and healthy states, but also has good generalization ability, which is suitable for disease prediction of unknown data.

Claims

1. A method for predicting multiple degenerative diseases and screening markers based on DNA methylation, characterized in that: The following steps are involved: S1: DNA methylation datasets collected from the GEO database; S2: Random forest algorithm was used to screen CpG sites and retain CpG sites with non-zero feature importance; S3: Based on a multi-scale one-dimensional convolutional neural network, the feature combination of the optimal CpG site is determined by gradually increasing the CpG site input; S4: CpG sites were analyzed by bio-omics analysis through Gene Ontology (GO) functional enrichment analysis, Kyoto Encyclopedia of Genes and Genomes (KEGG) signaling pathway enrichment analysis, and STRING protein interaction network analysis to screen out key CpG sites; S5: Construct a multi-classification degenerative disease prediction model based on residual network, and the model outputs the classification probabilities of neurodegenerative diseases, bone degenerative diseases, breast cancer and health status; S6: Louvain algorithm cluster analysis was used to verify the classification effect and disease association of CpG sites; S7: Apply SHAP values ​​to interpret model output, quantify the contribution of CpG sites to disease prediction, and improve model interpretability.

2. The method for predicting multiple degenerative diseases and screening markers based on DNA methylation according to claim 1, characterized in that: The step S1 comprises: S11: 142 DNA methylation datasets were obtained from the GEO database, blood samples were retained, and 39 were selected for analysis; S12: 12,578 common CpG sites were extracted from 39 data sets and divided into training and test sets in a ratio of 8:2 for feature screening.

3. The method for predicting multiple degenerative diseases and screening markers based on DNA methylation according to claim 1, characterized in that: The step S2 comprises: S21: 12,578 CpG sites were used as input features of the random forest model to evaluate the feature importance of each CpG site and screen out CpG sites that contribute significantly to disease prediction; S22: sites with zero feature importance were removed, and 6438 CpG sites with non-zero feature importance were retained; S23: The screening effect is evaluated by accuracy, precision, recall and MCC evaluation indicators.

4. The method for predicting multiple degenerative diseases and screening markers based on DNA methylation according to claim 1, characterized in that: The step S3 comprises: S31: Use a multi-scale one-dimensional convolutional neural network to extract features from 6438 CpG sites and capture local and global features; S32: Through five-fold cross validation, the number of CpG sites was gradually increased to evaluate the impact of different feature combinations on model performance; S33: The results showed that the model performed best when 600 CpG sites were input; 600 optimal CpG sites were eventually determined, providing a basis for the construction of subsequent disease prediction models.

5. The method for predicting multiple degenerative diseases and screening markers based on DNA methylation according to claim 1, characterized in that: The step S4 comprises: S41: Functionally annotate 600 CpG sites through GO functional enrichment analysis to reveal their biological functions and identify key biological processes related to degenerative diseases; S42: KEGG signaling pathway analysis revealed the enrichment of CpG sites in multiple disease-related pathways; S43: Use the STRING tool to construct a protein interaction network and identify key protein interaction modules that are highly related to degenerative diseases; S44: Through GO, KEGG and STRING analysis, 110 key CpG sites related to degenerative diseases were finally screened out, revealing their potential biological mechanisms.

6. The method for predicting multiple degenerative diseases and screening markers based on DNA methylation according to claim 1, characterized in that: The step S5 comprises: S51: 15 DNA methylation datasets were selected from the GEO database, covering four categories: neurodegenerative diseases, bone degenerative diseases, breast cancer and healthy samples. The datasets were balanced and adjusted to construct a multi-classification disease prediction model. S52: Based on 110 key CpG sites, a multi-classification disease prediction model was constructed, and local and global features were extracted through residual networks; S53: The features are reduced in dimension through the global average pooling layer, and finally the classification probabilities of the four types of diseases are output through the Softmax activation layer.

7. The method for predicting multiple degenerative diseases and screening markers based on DNA methylation according to claim 1, characterized in that: The step S6 comprises: S61: Unsupervised cluster analysis of the multi-classification prediction results of 110 CpG sites was performed using the Louvain algorithm to evaluate the clustering of degenerative disease samples and the distribution of disease types; S62: Clustering results verify the effectiveness of CpG sites in distinguishing different diseases and reveal the similarities and potential correlations of diseases in molecular characteristics.

8. The method for predicting multiple degenerative diseases and screening markers based on DNA methylation according to claim 1, characterized in that: The step S7 comprises: S71: After cluster analysis, the SHAP value was used to interpret the model output, quantify the contribution of each CpG site to disease classification, and identify the key sites that contributed most to the prediction; S72: Through SHAP analysis, we further reveal the role of key CpG sites in disease classification and their biological mechanisms, improve model interpretability and provide theoretical support for biomarkers.

Images