Osteoarthritis biomarker screening method and system based on trans-cnn feature fusion

Abstract

This invention relates to a method and system for screening osteoarthritis biomarkers based on Trans-CNN feature fusion, and relates to the field of medical technology. It addresses the problem that in existing osteoarthritis biomarker screening research, traditional Transformer models excel at capturing global relationships of features in input text sequences, while traditional CNN models focus more on local feature relationships. The fusion of these two approaches can reduce the accuracy of osteoarthritis diagnosis. The proposed Trans-CNN feature fusion model cleverly combines the global context awareness capability of Transformer with the local feature extraction advantage of CNN through a dynamic gating mechanism. This enables a more comprehensive capture of the complex regulatory relationships in DNA methylation data, thus facilitating the screening of osteoarthritis biomarkers.

Description

Technical Field

[0001] This invention relates to the field of medical technology, specifically to a method and system for screening biomarkers for osteoarthritis based on Trans-CNN feature fusion. Background Technology

[0002] Osteoarthritis (OA) is a complex degenerative disease that is most prevalent in the elderly and has become one of the fastest-growing health conditions globally. Traditional CT and MRI scans often diagnose OA in its later stages based on the degree of intersegmental osteophytes, which typically occurs in the middle to late stages and requires long-term treatment.

[0003] Epigenetics refers to the mechanisms by which genes are controlled and traits are thus controlled through reversible and heritable chemical modifications without altering the DNA sequence. Epigenetic modifications involve multiple regulatory processes related to gene expression, including DNA methylation, histone modification, chromatin remodeling, and non-coding RNA regulation. DNA methylation, primarily occurring on the cytosine (C) sites of CPG islands, is considered a major epigenetic regulatory mechanism. Depending on the specific region of the CPG island, methylation at CPG sites may promote or inhibit gene expression. In recent years, research has increasingly focused on disease analysis through DNA methylation. Recent studies have shown that DNA methylation is one of the important factors contributing to the pathogenesis and pro-inflammatory properties of osteoarthritis (OA).

[0004] In the screening of biomarkers for osteoarthritis, a method of cross-selection using multiple traditional machine learning models is often employed. With the continuous development of deep learning models in text sequence classification, their application in disease diagnosis is becoming more widespread. This invention, based on the fact that DNA methylation regulates gene expression and thus affects disease traits, proposes a deep learning model that better captures the relationship between methylation and disease for disease diagnosis. Traditional Transformer models excel at capturing global relationships of features in input text sequences, while traditional CNN models focus more on local feature relationships.

[0005] In summary, in the traditional research on screening biomarkers for osteoarthritis, the traditional Transformer model is good at capturing global relationships of features in the input text sequence, while the traditional CNN model focuses more on local feature relationships. The fusion of the two will lead to problems such as reduced accuracy, efficiency and generalization of osteoarthritis diagnosis. Summary of the Invention

[0006] To overcome the shortcomings of traditional machine learning models in capturing the relationship between different methylation sites and the disease in the screening research of osteoarthritis biomarkers, this invention fully considers the possible influence relationships between long-distance methylation sites and local methylation sites. Combining the Transformer's ability to capture global relationships of features with CNN's focus on local feature relationships, this invention proposes a screening method and system for osteoarthritis biomarkers based on Trans-CNN feature fusion.

[0007] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution:

[0008] Option 1: This invention proposes a method for screening osteoarthritis biomarkers based on Trans-CNN feature fusion, the method comprising the following steps:

[0009] Step 1: Obtain genome-wide methylation sites from the GEO database to form the original feature dataset;

[0010] Step 2: Based on the original feature dataset obtained in Step 1, high-potential candidate sites with statistical significance and biological prior support are initially screened out, forming differential methylation feature subsets and gene region feature subsets, respectively.

[0011] Step 3: The differential methylation feature subset is processed by SVM-RFE to select the site combination with the highest average accuracy. The differential methylation feature subset is then processed by the Random Forest RF model to select the site combination with the highest average accuracy. The common site of the two combinations is taken as the stable differential site set T-sites selected from the differential methylation feature subset. The gene region feature subset is processed by SVM-RFE to select the site combination with the highest average accuracy. The random forest RF model is then used to select the site combination with the highest average accuracy. The common site of the two combinations is taken as the stable gene region site set G-sites selected from the gene region feature subset. The combination of T-sites and G-sites, U-sites, is used as the feature subset that simultaneously possesses differential methylation sites and gene region methylation sites as the input set of the diagnostic model.

[0012] Step 4: Construct a Trans-CNN feature fusion diagnostic model. Calculate the feature importance of each whole-genome methylation site using the trained Trans-CNN feature fusion diagnostic model. Sort the whole-genome methylation sites according to the feature importance score, select the sites with the highest number of ACC to form the final combination of methylation biomarkers for osteoarthritis diagnosis, and retrain the final Trans-CNN diagnostic model to complete the screening of biomarkers for osteoarthritis.

[0013] Furthermore, in a preferred embodiment, step 1 further includes a step of screening common sites for whole-genome methylation sites.

[0014] Furthermore, a preferred embodiment is provided, wherein the method for initially screening high-potential candidate sites with statistical significance and biological prior support in step 2, and constructing differential methylation feature subsets and gene region feature subsets respectively, is as follows:

[0015] Step 2.1: Perform a two-sample t-test on each methylation site and set a joint significance threshold, which includes a statistical significance threshold and a biological significance threshold;

[0016] Step 2.2: Combine the statistical significance and biological significance threshold after FDR correction to sort all sites comprehensively, and select the top 5000 differentially methylated sites to form a subset of differentially methylated features;

[0017] Step 2.3: Obtain the list of effector genes;

[0018] Step 2.4: For each list of effector genes, define the core regulatory region as 2000bp upstream to 2000bp downstream of the transcription start site, and extract all methylation sites falling within the defined region to form a subset of gene region features.

[0019] Furthermore, a preferred embodiment is provided, wherein the method for feature selection using the SVM-RFE model and the random forest model in step 3 is as follows:

[0020] The differential feature subsets are processed by the SVM-RFE model and the random forest model, respectively, to select the combination of sites with the highest average accuracy for disease classification. The intersection of the site combinations selected by the SVM-RFE model and the random forest model is then taken. At the same time, the above operation is repeated for the gene region feature subsets, and the combination of the special diagnosis combination with the highest accuracy for each of the SVM-RFE model and the random forest model is taken.

[0021] Furthermore, a preferred embodiment is provided, wherein the method for taking the union of the gene region feature subsets selected by the SVM-RFE model and the random forest model in step 3 is as follows:

[0022] The intersection of the optimal feature subsets selected from the differential methylation feature subsets by the SVM-RFE model and the random forest model constitutes the robust differential site set;

[0023] The intersection of the optimal feature subsets selected from the feature subsets of gene regions by the SVM-RFE model and the random forest model constitutes the robust gene region locus set;

[0024] The robust differential loci set and the robust gene region loci set are combined to form a comprehensive candidate loci set, which is used as input to the deep learning model.

[0025] Furthermore, a preferred embodiment is provided, wherein the Trans-CNN feature fusion diagnostic model constructed in step 4 includes:

[0026] The input layer and embedding layer are used to receive the methylation β value vector of the integrated candidate site set;

[0027] The parallel feature extraction branch includes a Transformer encoder branch, which receives the embedded vector sequence. The Transformer encoder branch contains a multi-layer Transformer encoder, each layer equipped with a multi-head self-attention mechanism and a feedforward neural network. Through the self-attention mechanism, the association weight between any two sites in the sequence is calculated, thereby capturing long-range dependencies and global contextual information across the entire genome.

[0028] The CNN branch is used to receive the embedded vector sequence. The CNN branch includes two one-dimensional convolutional layers that capture the spatial patterns and short-range co-regulatory relationships between local neighboring sites, and is followed by activation functions and pooling layers.

[0029] The feature mapping and dynamic gating fusion module is used to project the outputs of the Transformer branch and the CNN branch onto the same fusion dimension through independent linear layers to obtain feature vectors Pa and Pb.

[0030] The gated vector generation module concatenates Pa and Pb along the feature dimension to obtain Z = [Pa; Pb]. Z is then input into a gated network, which consists of two fully connected layers; the final output is a gated vector g.

[0031] The adaptive fusion module is used to perform a weighted summation of the differential methylation feature subset and the gene region feature subset using the gating vector g, i.e., Fused_feature = g ⊙ Pa + (1 - g) ⊙ Pb, where ⊙ represents element-wise multiplication.

[0032] Furthermore, a preferred embodiment is provided, wherein the method for calculating the output gating vector g is as follows:

[0033]

[0034] Where W1 and W2 are weight matrices, and b1 and b2 are bias terms.

[0035] Option 2: A biomarker screening system for osteoarthritis based on Trans-CNN feature fusion, the system comprising:

[0036] The original feature dataset construction module is used to obtain whole-genome methylation sites from the GEO database to form a gene region feature dataset;

[0037] The differential methylation feature subset and gene region feature subset construction module are used to initially screen high-potential candidate sites with statistical significance and biological prior support based on the gene region feature dataset obtained by the gene region feature dataset construction module, which respectively constitute the differential methylation feature subset and gene region feature subset.

[0038] The input module is used to process the differentially methylated feature subsets through SVM-RFE to obtain the site combination with the highest average accuracy, and through the random forest RF model to obtain the site combination with the highest average accuracy. The common site of the two combinations is taken as the stable differentially methylated site set T-sites selected from the differentially methylated feature subsets. Similarly, the gene region feature subsets are processed through SVM-RFE to obtain the site combination with the highest average accuracy, and through the random forest RF model to obtain the site combination with the highest average accuracy. The common site of the two combinations is taken as the stable gene region site set G-sites selected from the gene region feature subsets. The combination of T-sites and G-sites, U-sites, serves as the feature subset that simultaneously possesses both differentially methylated sites and gene region methylated sites, and is used as the input set for the diagnostic model.

[0039] The screening module is used to construct a Trans-CNN feature fusion diagnostic model. It calculates the feature importance of each methylation site in the U-sites using the trained Trans-CNN feature fusion diagnostic model. The whole genome methylation sites are sorted according to the feature importance score. The sites with the highest number of ACC are selected to form the final combination of methylation biomarkers for osteoarthritis diagnosis. The final Trans-CNN diagnostic model is then retrained to complete the screening of biomarkers for osteoarthritis.

[0040] Option 3: A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described in Option 1.

[0041] Option 4: A computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor executes the method described in Option 1.

[0042] The advantages of this invention are:

[0043] This invention discloses a method and system for screening osteoarthritis biomarkers based on Trans-CNN feature fusion. It combines a dual-path initial screening approach of "differential analysis + effector gene region screening," robust selection through integrated machine learning of "SVM-RFE & RF," and a multi-level screening process involving fine-grained feature fusion and importance assessment using a Trans-CNN deep learning model. This progressive process maximizes the final biomarker combination's high discriminative power, strong robustness, and clear biological significance.

[0044] The Trans-CNN feature fusion model described in this invention cleverly combines the global context awareness capability of Transformer with the local feature extraction advantage of CNN through a dynamic gating mechanism. This enables it to capture the complex regulatory relationships in DNA methylation data more comprehensively, thereby achieving classification performance significantly better than a single model in OA diagnosis tasks.

[0045] The biomarker combination of 220 DNA methylation sites identified in this invention achieved excellent performance with an ACC of 1.00 and an AUC of 1.00 on the independent test set. On the external independent validation set, the ACC also reached 0.9516, demonstrating extremely high diagnostic accuracy and good generalization, providing a new approach for developing early detection of osteoarthritis (OA).

[0046] This invention, by integrating enrichment analysis and PPI network analysis, discovers potential core pathway interactions that jointly participate in the pathogenesis of OA, providing novel and highly promising candidate targets and research directions for subsequent experimental verification and targeted drug development.

[0047] This invention is also applicable to fields such as deep learning. Attached Figure Description

[0048] Figure 1 This is a volcano map of 5000 sites screened by the t-test as described in Implementation Method 1.

[0049] Figure 2 The diagrams shown in Implementation Method 1 illustrate the SVM-RFE model and RF model screening. (a) is a diagram of the common loci after SVM-RFE and RF screening of 5000 loci selected by t test. (b) is a diagram of the common loci after SVM-RFE and RF screening of 4291 loci near the effect gene. (c) is a diagram of the combination of the two common loci.

[0050] Figure 3 This is a schematic diagram of the Trans-CNN feature fusion model architecture described in Implementation Method 1.

[0051] Figure 4The following are schematic diagrams of ROC curves for diagnosis using the Trans-CNN model as described in Implementation Method 1: (a) is a schematic diagram of ROC curves for diagnosis using the Trans-CNN model with the T-sites test set input; (b) is a schematic diagram of ROC curves for diagnosis using the Trans-CNN model with the G-sites test set input; and (c) is a schematic diagram of ROC curves for diagnosis using the Trans-CNN model with the U-sites test set input.

[0052] Figure 5 The diagram shows the variation curves based on the selected sites as described in Implementation Method 1. (a) is a diagram showing the ACC comparison of the first K sites in U-sites for model diagnosis. (b) is a diagram showing the ROC curve of the test set for constructing the final diagnostic model using 220 sites. (c) is a diagram showing the structure of the confusion matrix using the external independent test set GSE63106.

[0053] Figure 6 This is a schematic diagram of PPI analysis as described in Implementation Method 1.

[0054] Figure 7 The following are schematic diagrams of the analysis described in Implementation Method 1: (a) is a schematic diagram of GO enrichment analysis, and (b) is a schematic diagram of WikiPathways enrichment analysis. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0056] Implementation Method 1, see Figures 1 to 7 This embodiment describes a method and system for screening osteoarthritis biomarkers based on Trans-CNN feature fusion. The primary objective of this embodiment is to overcome the shortcomings of existing technologies and provide a combination of osteoarthritis DNA methylation biomarkers with excellent diagnostic performance and clear biological significance.

[0057] This invention provides a systematic method for efficiently screening the above-mentioned combinations of biomarkers. This method integrates differential analysis, gene region screening, machine learning robust selection, deep learning feature fusion model, and a precision diagnostic model for osteoarthritis based on the above-mentioned combinations of biomarkers.

[0058] Step 1: Multi-source data integration and standardized preprocessing

[0059] Data Acquisition: A systematic search was conducted from the Gene Expression Omnibus (GEO) database using the keywords "Osteoarthritis" and "DNA methylation". The preferred training set included GSE63695 and GSE73626, totaling 125 samples (94 OA patients and 31 normal controls). The preferred independent external test set was GSE63106 (62 OA patients), used to evaluate the model's generalization ability.

[0060] Data cleaning and standardization:

[0061] Common site screening: Only methylation sites that are present in all training datasets are retained to ensure consistency in model application.

[0062] Batch effect correction: The ComBat algorithm (based on the empirical Bayesian framework) is used to effectively eliminate technical variations introduced from different detection batches.

[0063] Data standardization and sorting: The methylation β values ​​are standardized to a uniform precision, forming a standardized input matrix.

[0064] Step 2: Preliminary screening of candidate features based on a dual-path strategy

[0065] This step aims to initially screen high-potential candidate sites with statistical significance and biological prior support from hundreds of thousands of methylation sites across the entire genome.

[0066] 1. Pathway A: Differential methylation analysis (DMP screening)

[0067] Significance filtering: A two-sample t-test was performed on each methylation site (patient group vs. control group), and a joint significance threshold was set: statistical significance p-value < 0.05 and biological significance (difference in methylation level) |Δβ| > 0.2.

[0068] Comprehensive Ranking and Screening: All sites are comprehensively ranked by combining the FDR-corrected p-value and Fold Change (FC) value. The top 5000 differentially methylated sites are selected to form a differentially methylated feature subset (T-pool), such as... Figure 1 .

[0069] 2. Pathway B: Screening of regulatory regions of effector genes

[0070] Gene list acquisition: Obtain an experimentally validated list of genes with OA effects from authoritative OA published literature.

[0071] Region definition and site extraction: For each effector gene, its core regulatory region is defined as 2000 bp upstream to 2000 bp downstream of the transcription start site (TSS) (TSS±2000 bp). From the preprocessed data, all methylation sites falling within this region are extracted to form a gene region feature subset (G-pool).

[0072] Step 3: Robust Feature Selection Based on Ensemble Machine Learning

[0073] To reduce feature dimensionality, remove redundancy and noise, and improve the robustness and generalization of the selected features, this step uses two different machine learning models for feature selection and takes their intersection.

[0074] Model training and feature ranking:

[0075] SVM-RFE (Support Vector Machine Recursive Feature Elimination): Using T-pool and G-pool as inputs, SVM-RFE recursively removes the features with the smallest weights and reconstructs the model based on the remaining features. The importance of the features is ultimately ranked according to the order in which they are removed. Five-fold cross-validation is used to evaluate model performance in each recursive round.

[0076] Random Forest (RF): Similarly, it uses T-pool and G-pool as inputs. Random Forest evaluates feature importance by calculating the average decrease in node impurity that each feature brings across numerous decision trees.

[0077] Determining the optimal feature subset:

[0078] The intersection of the optimal feature subsets selected by SVM-RFE and RF from the T-pool constitutes the robust set of dissimilar sites (T-sites), such as... Figure 2 As shown in (a).

[0079] The intersection of the optimal feature subsets selected by SVM-RFE and RF from the G-pool constitutes a robust set of gene region sites (G-sites), such as... Figure 2 As shown in (b).

[0080] Feature set union: The union of the T-sites and G-sites sets forms a comprehensive candidate site set (U-sites), which serves as the input to subsequent deep learning models, such as... Figure 2 As shown in (c).

[0081] Step 4: Constructing and training a Trans-CNN feature fusion diagnostic model

[0082] To address the challenge of capturing the complex global and local relationships between methylation sites, this invention constructs a novel deep learning model that combines Transformer and CNN in parallel processing and dynamic fusion.

[0083] Model architecture:

[0084] Input layer and embedding layer: Receives the methylation β value vectors of U-sites. A fully connected embedding layer maps the numerical values ​​of all sites into vector representations, enhancing the model's expressive power.

[0085] Parallel feature extraction branch:

[0086] The Transformer encoder branch receives the embedded vector sequence. This branch contains a multi-layer Transformer encoder, each layer equipped with a multi-head self-attention mechanism and a feed-forward network. Through the self-attention mechanism, the model can calculate the association weight between any two sites in the sequence, thereby effectively capturing long-range dependencies and global contextual information across the entire genome.

[0087] CNN branch: Also receives the embedded vector sequence. This branch contains two one-dimensional convolutional layers (Conv1D) that capture spatial patterns and short-range co-regulatory relationships between local neighboring sites, followed by activation functions and pooling layers.

[0088] Feature mapping and dynamic gating fusion module:

[0089] Feature mapping: The outputs of the Transformer branch (dimension: B, d_model) and the CNN branch (dimension: B, cnn_out) are projected onto the same fusion dimension (fusion_dim) through independent linear layers to obtain feature vectors Pa and Pb (both with dimensions: B, fusion_dim), where B is the batch size.

[0090] Gated vector generation: Pa and Pb are concatenated along the feature dimension to obtain Z = [Pa; Pb] (dimension: B, 2×fusion_dim). Z is then input into a gating network consisting of two fully connected layers. The final output is a gated vector g (dimension: B, fusion_dim), with each element having a value range of [0, 1].

[0091] The calculation formula is:

[0092] Where W1 and W2 are weight matrices, and b1 and b2 are bias terms.

[0093] Adaptive fusion: Weighted summation of features from two sources using a gating vector g.

[0094] Fused_feature = g ⊙ Pa + (1 - g) ⊙ Pb

[0095] Here, ⊙ represents element-wise multiplication. This mechanism allows the model to automatically learn and dynamically allocate the contribution ratios of the Transformer and CNN across different feature dimensions, achieving adaptive complementarity and enhancement of global and local information.

[0096] Classification output layer: The fused features are first reduced in dimensionality by global average pooling, then input into a multilayer perceptron (MLP), and finally output the probability that the sample is OA.

[0097] Model architecture diagram as follows Figure 3 As shown.

[0098] Model training and evaluation:

[0099] The dataset (T-sites, G-sites, U-sites) was divided into a training set and an internal test set in a ratio of 80:20. During training, five-fold cross-validation was used to fine-tune the hyperparameters and prevent overfitting.

[0100] The performance metrics of the model on the validation and test sets for each dataset were recorded, including accuracy (ACC), F1 score (F1), and AUC. The dataset with the best performance (U-sites in this study) was determined through comparison. The classification performance results for T-sites, G-sites, and U-sites are shown in Table 1, and the ROC curves for the test set are shown in... Figure 4 As shown.

[0101] Table 1. Performance metrics of models on the validation and test sets for each dataset.

[0102]

[0103] Final biomarker determination: On the optimal dataset (U-sites), the feature importance of each locus was calculated using a trained Trans-CNN model. Loci were ranked according to their importance scores, and the model was trained for diagnosis using combinations of these loci. The loci with the highest number of ACC (220) were selected to form the final methylation biomarker combination for osteoarthritis diagnosis, such as... Figure 5 As shown in (a). Using this combination, the final Trans-CNN diagnostic model was retrained, and the ROC curve on the test set is shown in Figure (a). Figure 5 As shown in (b).

[0104] Step 5: Multi-level biological function analysis of biomarkers

[0105] The final selected biomarker combinations will be interpreted in depth.

[0106] Gene mapping and protein-protein interaction (PPI) network construction and analysis:

[0107] Gene annotation: Selected methylation sites are mapped to their nearest related genes using the Illumina HM450K annotation file.

[0108] Network construction: The mapped genes are input into the STRING database to obtain protein-protein interaction information. A minimum interaction confidence score is set.

[0109] Key node selection: Calculate the topological characteristics of each node in the network, including degree centrality and betweenness centrality. Use these indicators to screen for key hub features in the network, such as… Figure 6 .

[0110] Functional enrichment analysis:

[0111] Gene ontology is applied to the mapped gene set. Figure 7 Analysis revealed signaling pathways and biological processes closely related to the pathological process of OA.

[0112] In summary, the Trans-CNN model constructed in this invention combines the advantages of both approaches. By incorporating a dynamic weight gating mechanism, it considers both global relationships and local mechanisms across all feature dimensions, enhancing the capture of complex relationships between features. This method not only improves the accuracy and efficiency of osteoarthritis diagnosis but also provides new insights for biomarker screening.

[0113] Those skilled in the art will understand that the above description is merely a preferred embodiment of the present invention, and the features described in the various embodiments and / or technical solutions of this disclosure can be combined or combined in various ways, even if such combinations or combinations are not explicitly described in this disclosure. This is not intended to limit the present invention. 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.

[0114] Although preferred embodiments of the present invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended technical solutions are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the present invention. Clearly, those skilled in the art can make various modifications and variations to the present invention without departing from its spirit and scope. Thus, if these modifications and variations of the present invention fall within the scope of the present invention and its equivalents, the present invention also intends to include these modifications and variations.

Claims

1. A method for screening osteoarthritis biomarkers based on Trans-CNN feature fusion, characterized in that, The method includes the following steps: Step 1: Obtain genome-wide methylation sites from the GEO database to form the original feature dataset; Step 2: Based on the original feature dataset obtained in Step 1, high-potential candidate sites with statistical significance and biological prior support are initially screened out, forming differential methylation feature subsets and gene region feature subsets, respectively. Step 3: The differential methylation feature subset is processed by SVM-RFE to select the site combination with the highest average accuracy. The differential methylation feature subset is then processed by the Random Forest RF model to select the site combination with the highest average accuracy. The common site of the two combinations is taken as the stable differential site set T-sites selected from the differential methylation feature subset. The gene region feature subset is processed by SVM-RFE to select the site combination with the highest average accuracy. The random forest RF model is then used to select the site combination with the highest average accuracy. The common site of the two combinations is taken as the stable gene region site set G-sites selected from the gene region feature subset. The combination of T-sites and G-sites, U-sites, is used as the feature subset that simultaneously possesses differential methylation sites and gene region methylation sites as the input set of the diagnostic model. Step 4: Construct a Trans-CNN feature fusion diagnostic model. Calculate the feature importance of each whole-genome methylation site using the trained Trans-CNN feature fusion diagnostic model. Sort the whole-genome methylation sites according to the feature importance score, select the sites with the highest number of ACC to form the final combination of methylation biomarkers for osteoarthritis diagnosis, and retrain the final Trans-CNN diagnostic model to complete the screening of biomarkers for osteoarthritis.

2. The method for screening osteoarthritis biomarkers based on Trans-CNN feature fusion according to claim 1, characterized in that, Step 1 also includes a step of screening common sites for whole-genome methylation sites.

3. The method for screening osteoarthritis biomarkers based on Trans-CNN feature fusion according to claim 1, characterized in that, The method for initially screening high-potential candidate sites with statistical significance and biological prior support in step 2, and constructing differential methylation feature subsets and gene region feature subsets respectively, is as follows: Step 2.1: Perform a two-sample t-test on each methylation site and set a joint significance threshold, which includes a statistical significance threshold and a biological significance threshold; Step 2.2: Combine the statistical significance and biological significance threshold after FDR correction to sort all sites comprehensively, and select the top 5000 differentially methylated sites to form a subset of differentially methylated features; Step 2.3: Obtain the list of effector genes; Step 2.4: For each list of effector genes, define the core regulatory region as 2000bp upstream to 2000bp downstream of the transcription start site, and extract all methylation sites falling within the defined region to form a subset of gene region features.

4. The method for screening osteoarthritis biomarkers based on Trans-CNN feature fusion according to claim 1, characterized in that, The method for feature selection using the SVM-RFE model and the random forest model in step 3 is as follows: Using subsets of differentially methylated features and subsets of gene region features as inputs, the SVM-RFE model recursively removes the features with the smallest weights and reconstructs the model based on the remaining features. Finally, the importance ranking is determined according to the order in which the subsets of differentially methylated features and subsets of gene region features are removed.

5. The method for screening osteoarthritis biomarkers based on Trans-CNN feature fusion according to claim 1, characterized in that, The method for taking the union of the gene region feature subsets selected by the SVM-RFE model and the random forest model in step 3 is as follows: The intersection of the optimal feature subsets selected from the differential methylation feature subsets by the SVM-RFE model and the random forest model constitutes the robust differential site set; The intersection of the optimal feature subsets selected from the feature subsets of gene regions by the SVM-RFE model and the random forest model constitutes the robust gene region locus set; The robust differential loci set and the robust gene region loci set are combined to form a comprehensive candidate loci set, which is used as input to the deep learning model.

6. The method for screening osteoarthritis biomarkers based on Trans-CNN feature fusion according to claim 5, characterized in that, The Trans-CNN feature fusion diagnostic model constructed in step 4 includes: The input layer and embedding layer are used to receive the methylation β value vector of the integrated candidate site set; The parallel feature extraction branch includes a Transformer encoder branch, which receives the embedded vector sequence. The Transformer encoder branch contains a multi-layer Transformer encoder, each layer equipped with a multi-head self-attention mechanism and a feedforward neural network. Through the self-attention mechanism, the association weight between any two sites in the sequence is calculated, thereby capturing long-range dependencies and global contextual information across the entire genome. The CNN branch is used to receive the embedded vector sequence. The CNN branch includes two one-dimensional convolutional layers that capture the spatial patterns and short-range co-regulatory relationships between local neighboring sites, and is followed by activation functions and pooling layers. The feature mapping and dynamic gating fusion module is used to project the outputs of the Transformer branch and the CNN branch onto the same fusion dimension through independent linear layers to obtain feature vectors Pa and Pb. The gated vector generation module concatenates Pa and Pb along the feature dimension to obtain Z = [Pa; Pb]. Z is then input into a gated network, which consists of two fully connected layers; the final output is a gated vector g. The adaptive fusion module is used to perform a weighted summation of the differential methylation feature subset and the gene region feature subset using the gating vector g, i.e., Fused_feature = g ⊙ Pa + (1 - g) ⊙ Pb, where ⊙ represents element-wise multiplication.

7. The method for screening osteoarthritis biomarkers based on Trans-CNN feature fusion according to claim 1, characterized in that, The output gating vector g is calculated as follows: Where W1 and W2 are weight matrices, and b1 and b2 are bias terms.

8. A biomarker screening system for osteoarthritis based on Trans-CNN feature fusion, characterized in that, The system includes: The original feature dataset construction module is used to obtain whole-genome methylation sites from the GEO database to form a gene region feature dataset; The differential methylation feature subset and gene region feature subset construction module are used to initially screen high-potential candidate sites with statistical significance and biological prior support based on the gene region feature dataset obtained by the gene region feature dataset construction module, which respectively constitute the differential methylation feature subset and gene region feature subset. The input module is used to process the differentially methylated feature subsets through SVM-RFE to obtain the site combination with the highest average accuracy, and through the random forest RF model to obtain the site combination with the highest average accuracy. The common site of the two combinations is taken as the stable differentially methylated site set T-sites selected from the differentially methylated feature subsets. Similarly, the gene region feature subsets are processed through SVM-RFE to obtain the site combination with the highest average accuracy, and through the random forest RF model to obtain the site combination with the highest average accuracy. The common site of the two combinations is taken as the stable gene region site set G-sites selected from the gene region feature subsets. The combination of T-sites and G-sites, U-sites, serves as the feature subset that simultaneously possesses both differentially methylated sites and gene region methylated sites, and is used as the input set for the diagnostic model. The screening module is used to construct a Trans-CNN feature fusion diagnostic model. It calculates the feature importance of each methylation site in the U-sites using the trained Trans-CNN feature fusion diagnostic model. The whole genome methylation sites are sorted according to the feature importance score. The sites with the highest number of ACC are selected to form the final combination of methylation biomarkers for osteoarthritis diagnosis. The final Trans-CNN diagnostic model is then retrained to complete the screening of biomarkers for osteoarthritis.

9. A computer storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method described in any one of claims 1-7.

10. A computer device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the program to implement the method of any one of claims 1-7.

Images