A schizophrenia DNA methylation classification method and system based on GNN-Transformer fusion

Abstract

The application discloses a schizophrenia DNA methylation classification method and system based on GNN-Transformer fusion, relates to the field of bioinformatics, and solves the problems that traditional machine learning methods in existing methylation classification methods depend on manual feature engineering and cannot automatically capture high-order nonlinear relationships between CpG sites and the like, and the method comprises the following steps: acquiring two groups of schizophrenia DNA methylation data in a GEO database, selecting common CpG site intersections from the two groups of schizophrenia DNA methylation data, and pre-processing the selected common CpG site intersections to jointly stratify sampling according to disease states and data set sources; taking the selected common CpG site intersections as nodes of a graph, and constructing a heterogeneous weighted graph through a three-layer edge weight strategy; constructing a GNN-Transformer fusion model to realize double-path feature extraction and cross-modal Cross-Attention interaction; and adopting a multi-seed training strategy of five random seeds to optimize the constructed GNN-Transformer fusion model by taking FocalLoss and LabelSmoothing as target functions.

Description

Technical Field

[0001] This invention relates to the field of bioinformatics technology, specifically to a method and system for classifying DNA methylation in schizophrenia based on GNN-Transformer fusion. Background Technology

[0002] Schizophrenia is a severe mental disorder with a global prevalence of approximately 1%, characterized by high disability rates and a significant disease burden. Current clinical diagnosis relies primarily on symptom descriptions and behavioral observations, lacking objective molecular biomarkers, leading to diagnostic delays and a high risk of misdiagnosis.

[0003] DNA methylation is an important epigenetic modification, and studies have shown that schizophrenic patients exhibit abnormal methylation patterns at specific CpG sites. The Illumina HumanMethylation450K chip can simultaneously detect the methylation levels (beta values) of approximately 450,000 CpG sites, providing a high-throughput data foundation for genome-wide epigenetic research. However, this type of data is characterized by high dimensionality (the number of features far exceeds the number of samples), strong noise, and complex nonlinear genomic associations, making it difficult for traditional machine learning methods to fully model its inherent spatial topology and long-range dependencies.

[0004] Existing methylation classification methods have the following limitations: First, traditional machine learning methods (such as SVM and random forest) rely on manual feature engineering and cannot automatically capture high-order nonlinear relationships between CpG sites; second, existing deep learning methods (such as MLP and 1D-CNN) treat CpG sites as independent features and ignore the genomic spatial associations between sites; third, although graph neural network methods can model topological structures, they lack modeling of long-range dependencies at the sequence level; fourth, existing methods mostly use a single random seed for training, making it difficult to guarantee the stability of the results. Summary of the Invention

[0005] This invention addresses several shortcomings in existing methylation classification methods. Traditional machine learning methods rely on manual feature engineering, failing to automatically capture high-order nonlinear relationships between CpG sites. Treating CpG sites as independent features neglects genomic spatial associations. While graph neural networks can model topological structures, they lack modeling of long-range sequence-level dependencies. Furthermore, many existing methods employ single random seeds for training, leading to inconsistent results. Therefore, this invention proposes a GNN-Transformer fusion-based DNA methylation classification method and system for schizophrenia. The invention achieves these technical problems through the following technical solutions: Option 1: This invention proposes a DNA methylation classification method for schizophrenia based on GNN-Transformer fusion, the method comprising the following steps: Step 1: Obtain two sets of schizophrenia DNA methylation data from the GEO database, select the intersection of common CpG sites in the two sets of schizophrenia DNA methylation data, preprocess the selected common CpG site intersection, and perform joint stratified sampling based on disease status and dataset source. Step 2: Take the intersection of the common CpG sites selected in Step 1 as nodes of the graph, and construct a heterogeneous weighted graph using a three-layer edge weighting strategy; Step 3: Construct a GNN-Transformer fusion model to achieve dual-path feature extraction and cross-modal Cross-Attention interaction, and input it into a 3-layer MLP classifier to output two classes of logits for classification prediction; Step 4: Based on the two types of logits output in Step 3, the GNN-Transformer fusion model constructed in Step 3 is optimized using multiple sub-training strategies with 5 random seeds, with Focal Loss and Label Smoothing as the objective functions.

[0006] Furthermore, a preferred embodiment is provided, wherein the method for joint stratified sampling based on disease state and dataset source in step 1 is as follows: Two sets of schizophrenia DNA methylation data were divided into a training set, a validation set, and a test set, with a ratio of 8:1:1.

[0007] Furthermore, a preferred embodiment is provided in which, for the training set, the proportion of missing values ​​for each CpG site is calculated, and sites with a missing rate exceeding 5% are removed; for the validation set and test set, the mean of the training set is used to fill the gaps, and the methylation level values ​​of all CpG sites are cropped to [0.001, 0.999].

[0008] Furthermore, a preferred embodiment is provided in which step 1 further includes the step of fitting the training set using a normalized scaler method and then performing Z-score normalization on all sets.

[0009] Furthermore, a preferred embodiment is provided, wherein the method for constructing a heterogeneous weighted graph in step 2 using a three-layer edge weighting strategy is as follows: Step 2.1: Construct Layer 1 strong biological prior connections, establishing connections for CpG site pairs with the same gene name and a genomic distance of less than 10kb; establishing connections for site pairs marked as CpG Islands within a 10kb window on the same chromosome; Step 2.2: Construct Layer 2 intermediate biological prior connections, establish connections for CpG site pairs on the same chromosome with a genomic distance of less than 50kb; give site pairs of the same type of gene region a weighting of 1.2 times; Step 2.3: Construct Layer 3 statistical correlation connections. Calculate the absolute value matrix of Pearson correlation coefficients for all CpG site pairs on the training set, take the top 5% as the threshold, and establish correlation edges for site pairs exceeding the threshold; limit the total Layer 3 degree of each node to no more than 40; the edge weight is 0.5×|r|, where r is the Pearson correlation coefficient; Step 2.4: Based on steps 2.1 to 2.3, complete the construction of the heterogeneous weighted graph, calculate the connected components of the heterogeneous weighted graph, establish repair edges for isolated nodes using the KNN strategy to ensure that the graph is a connected graph; finally, normalize the weights of all edges to the interval [0.01, 1.0].

[0010] Furthermore, a preferred implementation method is provided, wherein the method for constructing the GNN-Transformer fusion model in step 3 is as follows: Step 3.1: Construct the input projection layer and GNN branch, and map the methylation level value of each CpG site to a 64-dimensional feature vector through linear transformation; Step 3.2: Construct a Transformer branch. After mapping the methylation level values ​​of the original CpG sites through a value embedding layer, a learnable position embedding is superimposed and input into a 2-layer TransformerEncoder. Step 3.3: Construct a Cross-Attention bidirectional interaction module. The Cross-Attention bidirectional interaction module includes attention interactions in two directions: GNN to Transformer and Transformer to GNN. Each direction is equipped with an independent Q / K / V projection matrix, Dropout, and residual LayerNorm, so that the feature representations of the two branches are mutually aware and cross-modal feature complementarity is achieved. Step 3.4: Extract AttentionPooling, Max Pooling, and Global Mean Pooling from the GNN and Transformer features after the Cross-Attention bidirectional interaction module. Concatenate the six feature vectors into a 384-dimensional comprehensive representation and input it into a 3-layer MLP classifier. The output logits of the two classes are used for classification prediction.

[0011] Furthermore, a preferred embodiment is provided, wherein the method for optimizing the GNN-Transformer fusion model constructed in step 3 in step 4 using Focal Loss and Label Smoothing as objective functions is as follows: The loss function and optimization strategy are configured. The loss function adopts a combination of Focal Loss and Label Smoothing: Focal Loss reduces the loss weight of easily distinguishable samples through modulation factors; Label Smoothing softens hard labels to alleviate model overfitting. The optimizer is implemented using AdamW, and the learning rate scheduling is implemented using CosineAnnealingLR.

[0012] Option 2: A DNA methylation classification method for schizophrenia based on GNN-Transformer fusion, the method comprising the following steps: The data acquisition module is used to acquire two sets of schizophrenia DNA methylation data from the GEO database, select the intersection of common CpG sites from the two sets of schizophrenia DNA methylation data, preprocess the selected common CpG site intersection, and perform joint stratified sampling based on disease status and dataset source. The heterogeneous weighted graph construction module is used to construct a heterogeneous weighted graph by taking the intersection of common CpG sites selected in the data acquisition module as nodes of the graph and using a three-layer edge weighting strategy. The classification module is used to build a GNN-Transformer fusion model, realize dual-path feature extraction and cross-modal Cross-Attention interaction, and input it into a 3-layer MLP classifier, outputting two classes of logits for classification prediction; The optimization module is used to optimize the GNN-Transformer fusion model constructed in step 3 based on the logits of the two classes output by the classification module, using multiple sub-training strategies with 5 random seeds and Focal Loss and Label Smoothing as objective functions.

[0013] 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.

[0014] 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.

[0015] The advantages of this invention are: The present invention describes a schizophrenia DNA methylation classification method and system based on GNN-Transformer fusion. Through a three-layer hierarchical graph construction strategy, namely Layer 1 biological prior connections, Layer 2 genomic distance connections, and Layer 3 statistical correlation connections, the method effectively encodes genomic prior knowledge into a weighted graph structure, which significantly improves the biological rationality of the graph compared with purely data-driven graph construction methods.

[0016] The present invention describes a schizophrenia DNA methylation classification method and system based on GNN-Transformer fusion, which combines a dual-path fusion architecture of GNN and Transformer with a Cross-Attention bidirectional interaction mechanism to simultaneously capture local graph topological information and global sequence dependencies, thus overcoming the limitations of a single model and exhibiting strong feature complementarity.

[0017] The present invention describes a schizophrenia DNA methylation classification method and system based on GNN-Transformer fusion, which employs a 6-way hybrid pooling strategy, namely two paths each of attention pooling, max pooling, and mean pooling, to fully aggregate multi-granularity feature representations and improve classification robustness.

[0018] The present invention describes a method and system for classifying DNA methylation in schizophrenia based on GNN-Transformer fusion. This method employs repeated training with five random seeds and mean ± standard deviation (MSD) statistical reporting to ensure the reproducibility and statistical reliability of the experimental results.

[0019] This invention is also applicable to fields such as the fusion of graph neural networks (GNN) with Transformer models and methods for classifying DNA methylation data in schizophrenia (SCZ). Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall model architecture of NN-Transformer as described in Implementation Method 1.

[0021] Figure 2 This is a schematic diagram of the three-layered map construction strategy for CpG sites described in Implementation Method 1.

[0022] Figure 3 This is a schematic diagram of the 5-seed ROC curve and confusion matrix of the internal test set described in Implementation Method 1.

[0023] Among them, (a) is a schematic diagram of the ROC curve of the receiver operation features, (b) is a schematic diagram of the confusion matrix, and (c) is a schematic diagram of the performance comparison of individual model seeds.

[0024] Figure 4 This is a schematic diagram of the ROC curve and confusion matrix of the external verification queue described in Implementation Method 1. Detailed Implementation

[0025] 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.

[0026] Implementation Method 1, see [link] Figures 1 to 4 This embodiment describes a method and system for classifying DNA methylation in schizophrenia based on GNN-Transformer fusion. It employs a hierarchical graph construction strategy to map CpG sites, a dual-path feature extraction combining graph neural networks and Transformers, cross-modal cross-attention interaction, and multi-path pooling fusion. The method specifically includes the following steps: Step 1: Data Loading and Preprocessing This implementation uses Illumina HumanMethylation450K microarray data from two independent cohorts in the GEO database: GSE147221 (n=679, Schizophrenia=348, Control=331) and GSE84727 (n=847, Schizophrenia=414, Control=433), resulting in a total of 1526 samples (Schizophrenia=762, Control=764), with a case-control ratio close to 1:1. The two datasets have 484057 and 420374 CpG loci, respectively; common CpG loci were retained after intersection analysis for subsequent analysis. The datasets were pre-divided into a training set (n≈1221), a validation set (n≈153), and a test set (n≈153) at a fixed ratio of 8:1:1. Stratified sampling (disease state × dataset source) was used during partitioning to ensure consistent case / control ratios and dataset source ratios across the three subsets, guaranteeing experimental reproducibility.

[0027] Furthermore, step 1 also includes the following steps: Step 1.1: Find the intersection of common CpG sites in the two datasets (GSE147221 and GSE84727), filter out cross-reactive probes and multi-site mapping probes, remove low-quality probes and retain high-confidence CpG sites; calculate the missing value ratio of each CpG site in the training set, remove sites with a missing value ratio of more than 5%, fill the remaining missing values ​​of the retained sites with the mean of the training set, and fill the validation set and test set with the mean of the training set to prevent data leakage; then prune all beta values ​​to [0.001, 0.999] to eliminate the influence of extreme values.

[0028] Step 1.2: In the R environment, the ComBat method of the sva package is used to perform batch effect correction on the training set, using age, gender, blood cell ratio, and methylated smoking score as covariates to remove batch effects caused by the source of the dataset; the validation set and test set are frozen using the affine transformation parameters fitted to the training set to strictly avoid data leakage.

[0029] Step 1.3: After fitting the training set with StandardScaler, perform Z-score standardization on all sets; the validation and test sets strictly use the mean and standard deviation of the training set to prevent data leakage.

[0030] Step 1.4: Using a joint stratified sampling strategy of disease status and dataset source at a fixed ratio of 8:1:1, the data is pre-divided into training set, validation set and test set to ensure that the case / control ratio and dataset source ratio are consistent in the three subsets, thus ensuring the reproducibility of the experiment.

[0031] Step 1.5: Load the Illumina HumanMethylation450K annotation file and extract annotation information such as chromosome number, genome coordinates, gene name, gene region type, and CpG island relationship for each CpG site, for use in subsequent graph construction and interpretability analysis.

[0032] Step 2: Construction of a three-layer weighted map of CpG sites Application of hierarchical graph construction strategy: Biological prior knowledge and statistical correlation are organically integrated through a three-layer heterogeneous graph structure. Layer 1 to Layer 3 correspond to biological confidence from strong to weak, respectively. The hierarchical weight design enables the model to prioritize the use of high-confidence genomic functional association information, avoiding the problem of traditional pure data-driven graph construction methods ignoring biological semantics, and providing a semantically rich graph structure foundation for subsequent GNN feature aggregation.

[0033] This invention uses the 1000 CpG sites selected in step 1 as nodes of a graph, and constructs a heterogeneous weighted graph using a three-layer edge weighting strategy, as follows: Figure 2 As shown.

[0034] Furthermore, step 2 includes the following steps: Step 2.1: Construct Layer 1 strong biological prior connections (weight boost=1.0). Connections are established for CpG site pairs with the same gene name and a genomic distance of less than 10kb, based on the biological basis that methylation sites of the same gene exhibit synergistic regulatory relationships. Connections are also established for site pairs marked as CpG Islands within a 10kb window on the same chromosome, based on the biological basis that the methylation status of sites within CpG islands is highly correlated. A maximum of 50 sites per gene and a maximum of 30 sites per CpG island are limited to avoid graph degradation due to hyperconnections.

[0035] Step 2.2: Construct Layer 2 moderate biological prior connections (weight boost=0.7, distance decay). Establish connections for CpG site pairs on the same chromosome with a genomic distance of less than 50kb. The edge weights are calculated using the exponential decay function w=0.7×exp(-d / 16667), with higher weights for closer distances. Site pairs in the same type of gene region (e.g., both located in the TSS or Body region) are given a 1.2x weight boost. Weak edges with weights below 0.05 are filtered out.

[0036] Step 2.3: Construct Layer 3 statistical correlation connections (weight boost=0.5). Calculate the absolute value matrix of Pearson correlation coefficients for all CpG site pairs on the training set, take the top 5% as the threshold, and establish correlation edges for site pairs exceeding the threshold; limit the total Layer 3 degree of each node to no more than 40; the edge weight is 0.5×|r|, where r is the Pearson correlation coefficient.

[0037] Step 2.4: After the graph is constructed, perform connectivity repair. Calculate the connected components of the graph, and for isolated nodes, establish repair edges using the KNN strategy (the 3 nodes with the highest correlation to the principal components) to ensure the graph is connected; finally, normalize the weights of all edges to the interval [0.01, 1.0].

[0038] Step 3: Constructing the GNN-Transformer fusion model This invention constructs a GNNTransformerModel fusion model with the following hyperparameter settings: hidden_dim=64, number of attention heads n_heads=4, and Dropout=0.4. The model structure is as follows: Figure 1 As shown, it includes the following core components: (1) Input projection layer: The single-dimensional beta value of each CpG site is mapped to a 64-dimensional feature vector through linear transformation, providing a unified input representation for subsequent dual-path feature extraction.

[0039] (2) GNN branch: A two-layer GCNConv graph convolutional network is used, with residual connections and LayerNorm layers for feature normalization. Dropout=0.4 is used to prevent overfitting. The graph convolution uses the layered graph structure and edge weights constructed in step two to aggregate neighbor node information and capture the local genomic topological relationships of CpG sites.

[0040] (3) Transformer branch: After mapping the original beta value through the value embedding layer, the learnable position embedding is superimposed and input into a 2-layer TransformerEncoder (Pre-LayerNorm structure, nhead=4, FFN dimension 256, Dropout=0.4) to capture the global long-range dependencies between CpG sites.

[0041] (4) Cross-Attention bidirectional interaction module: It includes attention interaction in two directions: from GNN to Transformer (GNN output is Query, Transformer output is Key / Value) and from Transformer to GNN (reverse). Each direction is equipped with an independent Q / K / V projection matrix, Dropout and residual LayerNorm, so that the feature representations of the two branches are mutually aware and achieve cross-modal feature complementarity.

[0042] (5) 6-way hybrid pooling: AttentionPooling (learnable attention-weighted mean), max-over-nodes and global mean pooling are extracted from the GNN features and Transformer features after Cross-Attention respectively. The 6 feature vectors are concatenated into a 384-dimensional (64×6) comprehensive representation.

[0043] (6) 3-layer MLP classifier: The network structure is 384→128→64→2. Each layer is equipped with LayerNorm, ReLU activation and Dropout, and outputs two classes (Control / SCZ) logits for classification prediction.

[0044] Furthermore, step 3 specifically includes the following steps: Step 3.1: Construct the input projection layer and GNN branch. A linear transformation maps the single-dimensional beta value of each CpG site to a 64-dimensional feature vector, providing a unified input representation for subsequent dual-path feature extraction. The GNN branch employs a two-layer GCNConv graph convolutional network, combined with residual connections and LayerNorm layers for feature normalization, with Dropout=0.4 to prevent overfitting. The graph convolution utilizes the hierarchical graph structure and edge weights constructed in Step 2 to aggregate neighbor node information and capture the local genomic topological relationships of CpG sites.

[0045] Step 3.2: Construct the Transformer branch. After mapping the original beta values ​​through a value embedding layer, learnable positional embeddings are superimposed and input into a 2-layer TransformerEncoder (Pre-LayerNorm structure, attention heads nhead=4, feedforward network dimension 256, Dropout=0.4) to capture global long-range dependencies between CpG sites.

[0046] Step 3.3: Construct the Cross-Attention bidirectional interaction module. This module includes attention interactions in two directions: from GNN to Transformer (with GNN output as Query and Transformer output as Key / Value) and from Transformer to GNN (in the opposite direction). Each direction is equipped with an independent Q / K / V projection matrix, Dropout, and residual LayerNorm, enabling the feature representations of the two branches to be mutually aware and achieve cross-modal feature complementarity.

[0047] Step 3.4: Construct a 6-way hybrid pooling and MLP classifier. Attention pooling (learnable attention-weighted mean), max pooling, and global mean pooling are extracted from the GNN and Transformer features after Cross-Attention, resulting in a total of 6 feature vectors concatenated into a 384-dimensional (64×6) comprehensive representation. This is then input into a 3-layer MLP classifier (network structure: 384→128→64→2, each layer equipped with LayerNorm, ReLU activation, and Dropout), outputting logits for two classes (healthy control / schizophrenia) for classification prediction. Application of the GNN-Transformer fusion model: The GNN branch mines local collaborative regulatory relationships between CpG sites from a graph topology perspective, while the Transformer branch models global long-range dependencies from a sequence perspective. The two achieve bidirectional perception and complementarity at the feature level through Cross-Attention, effectively overcoming the problem of insufficient ability of a single model to model complex associations in methylated data. The 6-way hybrid pooling strategy fully aggregates multi-granularity feature representations, further improving the robustness of classification.

[0048] Step 4: Multi-random Seed Training This invention uses 5 preset random seeds for repeated training, with each training session being completely independent.

[0049] Furthermore, step four includes the following steps: Step 4.1: Configure the loss function and optimization strategy. The loss function adopts a combination of Focal Loss (modulation factor gamma=2.0) and Label Smoothing (smoothing coefficient epsilon=0.1): Focal Loss reduces the loss weight of easily distinguishable samples through the modulation factor, focusing on difficult samples; Label Smoothing softens hard labels and alleviates model overfitting. The optimizer uses AdamW (learning rate lr=5e-4, weight decay=5e-4), and the learning rate scheduling adopts CosineAnnealingLR (T_max=250, minimum learning rate eta_min=1e-6).

[0050] Step 4.2: Perform repeated training with multiple random seeds. Five preset random seeds are used for completely independent repeated training. The training configuration is as follows: batch size = 32, maximum training epochs = 250, early stopping patience = 30 (with validation set AUC as the monitoring metric), and gradient clipping max_norm = 1.0 to prevent gradient explosion. Finally, the performance metrics are reported in the form of mean ± standard deviation (mean ± std) of the five random seeds to ensure the repeatability and statistical reliability of the experimental results.

[0051] Step 5: Internal Testing and External Independent Validation Furthermore, step five includes the following steps: Step 5.1, Internal Test Set Evaluation. Using an internal test set that is from the same source as the training set but strictly separate, evaluate the AUC, ACC, F1, Precision, and Recall metrics for models trained with five random seeds. Plot various sub-ROC curves (including the mean ± standard deviation confidence band) and confusion matrices, such as... Figure 3 As shown in Table 2, the performance of the model is compared with that of the mainstream deep learning and machine learning methods (1D-CNN, Random Forest, GCN, GAT) to verify the effectiveness of the model of the present invention.

[0052] Step 5.2, External Independent Validation. External generalization validation was performed on two independent public GEO datasets: GSE80417 (675 cases, 293 with schizophrenia / 322 healthy controls) and GSE152027 (493 cases, 290 with schizophrenia / 203 healthy controls). Figure 4As shown. External data processing flow: Automatically detect the row and column directions of the beta matrix, fill in the column mean for missing CpG sites, and standardize using the StandardScaler parameters of the training set (without refitting). The model can directly infer without retraining to verify the cross-dataset generalization ability of the method of this invention. Ablation experiments simultaneously analyze the contributions of the Cross-Attention module, edge weight design, GNN branch, and Transformer branch to the overall performance, as shown in Table 1.

[0053] like Figure 3 As shown, the method of the present invention achieved an average AUC of 0.890±0.030 for 5 seeds on the internal test set (153 cases in total, SCZ: 61 cases / Control: 76 cases), an AUC of 0.917 for the best seed (Seed 121), an average ACC of 0.799±0.022, and an average F1 of 0.787±0.032.

[0054] As shown in Table 1, the ablation experiments show that removing the Cross-Attention module resulted in a 6.1% performance decrease (AUC: 0.8915→0.8305), removing the Edge Weight resulted in a 6.1% performance decrease (AUC: 0.8915→0.8304), removing the GNN branch (Transformer only) resulted in a 6.3% performance decrease (AUC: 0.8915→0.8286), and removing the Transformer branch (GNN only) resulted in a 14.3% performance decrease (AUC: 0.8915→0.7487). This verifies the effectiveness of each component, with the Transformer branch contributing the most to the model performance.

[0055] Table 1 Ablation Experiment Results

[0056] As shown in Table 2, comparative experiments show that the method of the present invention outperforms baseline methods such as 1D-CNN (0.8822±0.0066), RandomForest (0.8668±0.0082), GCN (0.7463±0.0061), and GAT (0.7397±0.0030) in TestAUC (0.8915±0.030), verifying the effectiveness of the GNN and Transformer fusion architecture.

[0057] like Figure 4As shown, the method of this invention achieved average AUCs of 0.781±0.051 and 0.671±0.055 on the external validation sets GSE80417 (675 cases, SCZ:353 / Ctrl:322) and GSE152027 (493 cases, SCZ:290 / Ctrl:203), respectively. The optimal seed achieved an AUC of 0.831 on GSE80417 and an AUC of 0.722 on GSE152027, verifying the model's cross-dataset generalization ability.

[0058] 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.

[0059] 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 DNA methylation classification method for schizophrenia based on GNN-Transformer fusion, characterized in that, The method includes the following steps: Step 1: Obtain two sets of schizophrenia DNA methylation data from the GEO database, select the intersection of common CpG sites in the two sets of schizophrenia DNA methylation data, preprocess the selected common CpG site intersection, and perform joint stratified sampling based on disease status and dataset source. Step 2: Take the intersection of the common CpG sites selected in Step 1 as nodes of the graph, and construct a heterogeneous weighted graph using a three-layer edge weighting strategy; Step 3: Construct a GNN-Transformer fusion model to achieve dual-path feature extraction and cross-modal Cross-Attention interaction, and input it into a 3-layer MLP classifier to output two classes of logits for classification prediction; Step 4: Based on the two types of logits output in Step 3, the GNN-Transformer fusion model constructed in Step 3 is optimized using multiple sub-training strategies with 5 random seeds, with Focal Loss and Label Smoothing as the objective functions.

2. The schizophrenia DNA methylation classification method based on GNN-Transformer fusion according to claim 1, characterized in that, The method for joint stratified sampling based on disease status and dataset source in step 1 is as follows: Two sets of schizophrenia DNA methylation data were divided into a training set, a validation set, and a test set, with a ratio of 8:1:

1.

3. The schizophrenia DNA methylation classification method based on GNN-Transformer fusion according to claim 2, characterized in that, For the training set, the proportion of missing values ​​for each CpG site is calculated, and sites with a missing rate exceeding 5% are removed. For the validation and test sets, the mean of the training set is used for padding, and the methylation level values ​​of all CpG sites are cropped to [0.001, 0.999].

4. The schizophrenia DNA methylation classification method based on GNN-Transformer fusion according to claim 1, characterized in that, Step 1 also includes the step of fitting the training set using the normalized scaler method and then performing Z-score normalization on all sets.

5. The schizophrenia DNA methylation classification method based on GNN-Transformer fusion according to claim 1, characterized in that, The method for constructing a heterogeneous weighted graph using a three-layer edge weighting strategy in step 2 is as follows: Step 2.1: Construct Layer 1 strong biological prior connections, establishing connections for CpG site pairs with the same gene name and a genomic distance of less than 10kb; establishing connections for site pairs marked as CpG Islands within a 10kb window on the same chromosome; Step 2.2: Construct Layer 2 intermediate biological prior connections, establish connections for CpG site pairs on the same chromosome with a genomic distance of less than 50kb; give site pairs of the same type of gene region a weighting of 1.2 times; Step 2.3: Construct Layer 3 statistical correlation connections. Calculate the absolute value matrix of Pearson correlation coefficients for all CpG site pairs on the training set, take the top 5% as the threshold, and establish correlation edges for site pairs exceeding the threshold; limit the total Layer 3 degree of each node to no more than 40; the edge weight is 0.5×|r|, where r is the Pearson correlation coefficient; Step 2.4: Based on steps 2.1 to 2.3, complete the construction of the heterogeneous weighted graph, calculate the connected components of the heterogeneous weighted graph, establish repair edges for isolated nodes using the KNN strategy to ensure that the graph is a connected graph; finally, normalize the weights of all edges to the interval [0.01, 1.0].

6. The schizophrenia DNA methylation classification method based on GNN-Transformer fusion according to claim 1, characterized in that, The method for constructing the GNN-Transformer fusion model in step 3 is as follows: Step 3.1: Construct the input projection layer and GNN branch, and map the methylation level value of each CpG site to a 64-dimensional feature vector through linear transformation; Step 3.2: Construct a Transformer branch. After mapping the methylation level values ​​of the original CpG sites through a value embedding layer, a learnable position embedding is superimposed and input into a 2-layer TransformerEncoder. Step 3.3: Construct a Cross-Attention bidirectional interaction module. The Cross-Attention bidirectional interaction module includes attention interactions in two directions: GNN to Transformer and Transformer to GNN. Each direction is equipped with an independent Q / K / V projection matrix, Dropout, and residual LayerNorm, so that the feature representations of the two branches are mutually aware and cross-modal feature complementarity is achieved. Step 3.4: Extract AttentionPooling, Max Pooling, and Global Mean Pooling from the GNN and Transformer features after the Cross-Attention bidirectional interaction module. Concatenate the six feature vectors into a 384-dimensional comprehensive representation and input it into a 3-layer MLP classifier. The output logits of the two classes are used for classification prediction.

7. The schizophrenia DNA methylation classification method based on GNN-Transformer fusion according to claim 1, characterized in that, The method for optimizing the GNN-Transformer fusion model constructed in step 3 in step 4, using Focal Loss and Label Smoothing as objective functions, is as follows: The loss function and optimization strategy are configured. The loss function adopts a combination of Focal Loss and Label Smoothing: Focal Loss reduces the loss weight of easily distinguishable samples through modulation factors; Label Smoothing softens hard labels to alleviate model overfitting. The optimizer is implemented using AdamW, and the learning rate scheduling is implemented using CosineAnnealingLR.

8. A DNA methylation classification method for schizophrenia based on GNN-Transformer fusion, characterized in that, The method includes the following steps: The data acquisition module is used to acquire two sets of schizophrenia DNA methylation data from the GEO database, select the intersection of common CpG sites from the two sets of schizophrenia DNA methylation data, preprocess the selected common CpG site intersection, and perform joint stratified sampling based on disease status and dataset source. The heterogeneous weighted graph construction module is used to construct a heterogeneous weighted graph by taking the intersection of common CpG sites selected in the data acquisition module as nodes of the graph and using a three-layer edge weighting strategy. The classification module is used to build a GNN-Transformer fusion model, realize dual-path feature extraction and cross-modal Cross-Attention interaction, and input it into a 3-layer MLP classifier, outputting two classes of logits for classification prediction; The optimization module is used to optimize the GNN-Transformer fusion model constructed in step 3 based on the logits of the two classes output by the classification module, using multiple sub-training strategies with 5 random seeds and Focal Loss and Label Smoothing as objective functions.

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