Peripheral blood dna methylation markers for predicting frontal cortex thickness

Abstract

The present application relates to a peripheral blood DNA methylation marker for predicting frontal lobe cortex thickness, the marker is one or several of middle frontal gyrus thickness, superior frontal gyrus thickness, inferior frontal sulcus thickness, middle frontal sulcus thickness. The present application not only constructs a high-performance prediction model, but also screens out high-reproducible methylation markers through strict Bootstrap stability analysis, and the reproducibility of these markers in other independent data sets reaches more than 85%, which provides a reliable feature screening evidence base for related research teams.

Description

Technical Field

[0001] This invention belongs to the field of biomedical technology, and specifically relates to a peripheral blood DNA methylation marker for predicting the thickness of the frontal cortex. Background Technology

[0002] Frontal cortex thickness, as a key indicator of brain structure, has profound and multidimensional biological significance. The most crucial biological function of the frontal lobe is executive control, responsible for coordinating working memory, suppressing irrelevant interference, and guiding goal-oriented complex behaviors. The efficiency of this higher cognitive function is directly related to the thickness of the dorsolateral prefrontal cortex. Greater cortical thickness generally indicates a denser concentration of neurons and richer synaptic connections in this region, thus providing a stronger neural basis for executive control. Studies have shown a significant positive correlation between cortical thickness in this region and an individual's working memory capacity and cognitive flexibility test scores; conversely, pathological or age-related thinning often leads to a decline in executive functions such as planning, task switching, and problem-solving. (Fjell, AM, et al. (2015). Development and aging of cortical thickness correspond to genetic organization patterns. Proceedings of the National Academy of Sciences.)

[0003] Changes in the thickness of the frontal cortex are closely linked to the risk of various neuropsychiatric diseases. Studies have shown that thinning of the frontal cortex can be observed in the early stages of Alzheimer's disease; meanwhile, in patients with schizophrenia, the thickness of the frontal cortex is significantly reduced, and its neuronal density is abnormal. (Cullen TJ, Walker MA, Eastwood SL, Esiri MM, Harrison PJ, Crow TJ. Anomalies of asymmetry of pyramidal cell density and structure in dorsolateral prefrontal cortex in schizophrenia. British Journal of Psychiatry. 2006;188(1):26-31. doi:10.1192 / bjp.bp.104.008169)

[0004] In summary, the thickness of the frontal cortex is not only a core element in understanding individual differences in cognitive abilities and disease susceptibility, but it also constitutes the neural basis for higher cognitive functions such as executive control. Furthermore, it is an important structural marker for the early identification of neuropsychiatric disorders such as Alzheimer's disease and schizophrenia. Therefore, in-depth research on this indicator provides indispensable biological evidence for assessing brain health, predicting the risk of cognitive decline, and exploring targeted intervention strategies.

[0005] The mechanisms underlying changes in frontal cortex thickness are not fully understood, but current research suggests that DNA and other factors are potential biomarkers. Studies have reported a significant correlation between SKA2 gene methylation and the prefrontal cortex in individuals with post-traumatic stress disorder (PTSD). (Sadeh N, Spielberg JM, Logue MW, Wolf EJ, Smith AK, Lusk J, Hayes JP, Sperbeck E, Milberg WP, McGlinchey RE, Salat DH, Carter WC, Stone A, Schichman SA, Humphries DE, Miller MW. SKA2 methylation is associated with decreased prefrontal cortical thickness and greater PTSD severity among trauma-exposed veterans. Mol Psychiatry. 2016 Mar;21(3):357-63. doi: 10.1038 / mp.2015.134. Epub 2015 Sep 1. PMID: 26324104; PMCID: PMC4760869.)

[0006] DNA methylation is an important component of systems biology and a key technology in precision medicine. It can reveal the physiological and pathological state of an organism in real time, reflecting changes related to its phenotype. Currently, numerous DNA methylation biomarker prediction models based on blood, urine, feces, and tissue samples are used to predict brain structure-related phenotypes.

[0007] However, no studies have yet used DNA methylation to predict individual differences in frontal cortex thickness. Therefore, developing a highly sensitive and specific integrated DNA methylation biomarker prediction model to accurately assess frontal cortex thickness is of great significance for improving the early identification of executive dysfunction and the ability to warn of neurological disease risks while ensuring non-invasive and convenient detection. Summary of the Invention

[0008] The technical problem to be solved by this invention is to provide a peripheral blood DNA methylation biomarker for predicting the thickness of the frontal cortex. Through rigorous Bootstrap stability analysis, highly reproducible methylation biomarkers were screened. These biomarkers achieved a reproducibility rate of over 85% in other independent datasets, providing a reliable evidence basis for feature screening for research teams in related fields.

[0009] This invention provides a peripheral blood DNA methylation biomarker for predicting the thickness of the frontal cortex, wherein the biomarker is one or more of the following: G_front_middle_thickness, G_S_frontomargin_thickness, S_front_inf_thickness, and S_front_middle_thickness.

[0010] The present invention also provides a diagnostic product based on the aforementioned peripheral blood DNA methylation marker.

[0011] Furthermore, the diagnostic product is a reagent kit, a methylation panel, or a chip. The chip is preferably an EPIC chip.

[0012] Furthermore, the peripheral blood is obtained by venous blood collection using EDTA anticoagulant tubes or sodium citrate anticoagulant tubes.

[0013] This invention also provides a model for predicting the thickness of the frontal cortex, the method for constructing the model comprising the following steps:

[0014] S1. Obtain a sample set with complete phenotypic and methylation data;

[0015] S2. Preprocess the above sample set to obtain a dataset that can be used for subsequent analysis;

[0016] S3. In the above dataset, linear correlation analysis is performed for each phenotype to screen CpG sites that are significantly related to the phenotype as candidate features. Then, a dual-track analysis system is established, and the samples are randomly divided into training set and validation set in a 5:5 ratio to construct continuous variable prediction model and binary classification prediction model respectively.

[0017] S4. Summarize and output all model results in a unified manner.

[0018] Furthermore, during the construction process, Bootstrap sampling technique is used to evaluate feature stability, and feature importance is comprehensively evaluated by synthesized scores (0.7 × frequency + 0.3 × performance). Dynamic convergence conditions are set (Spearman correlation > 0.95 and ranking change < 2) to automatically stop iteration.

[0019] Furthermore, based on the feature combinations selected for stability, multiple linear regression, LASSO, random forest, Elastic Net, and Stacking ensemble models are constructed respectively, and prediction results and performance evaluations are output.

[0020] The present invention also provides a system for predicting the thickness of the frontal cortex, the system comprising the following modules: a data acquisition module for acquiring methylation level data of the marker in the sample to be tested; a data analysis module for predicting the data in the data acquisition module using the model; and an output module for outputting the prediction results.

[0021] The beneficial effects are as follows:

[0022] (1) The blood sample collection used in this invention is simple, and the DNA methylation detection technology is mature and stable. In addition, the methylation data can be stored for a long time and analyzed in batches, which is suitable for large-scale cohort studies. The acceptance of the tested population is high, and it is also convenient for clinical and scientific researchers to promote its application.

[0023] (2) Compared with traditional single phenotype or single model analysis methods, this invention first corrects the interference of potential confounding factors such as sample heterogeneity and technical batch effect through rigorous data preprocessing and feature stability screening. It can accurately reflect the association between multiple phenotypes and methylation markers, and significantly reduce the false positive rate and false negative rate.

[0024] (3) This invention is the first to construct a dual-track prediction system for continuous variables and binary classification tasks. By integrating various machine learning algorithms (including LASSO, random forest, Elastic Net, etc.) and performing Bootstrap stability assessment, it screens out combinations of methylation biomarkers with high reproducibility, avoiding the limitations of single algorithm or single phenotype analysis. At the same time, it combines Stacking ensemble learning to construct a robust multi-phenotype prediction model, and performs cross-validation and independent testing. In practical applications, multiple phenotypes (such as EAA-related indicators) can be accurately predicted using only blood DNA methylation data. This helps researchers and clinicians identify high-risk individuals or explore molecular mechanisms based on the prediction results, providing support for early intervention and personalized medicine, and has broad scientific research and clinical translational value.

[0025] (4) The multi-phenotypic prediction model provided by this invention performs well in performance evaluation and is validated using multiple metrics (applicable to continuous and binary classification tasks). In continuous prediction tasks, the average R-value is [missing information]. 2 The ROC curves achieved a mean squared error (MCI) of over 0.68, with the area under the ROC curve exceeding 0.82 in both the training and validation sets for binary classification tasks. Accuracy, sensitivity, and specificity all exceeded 0.79. More importantly, this invention not only constructed a high-performance prediction model but also screened highly reproducible methylation biomarkers through rigorous bootstrap stability analysis. These biomarkers achieved a reproducibility rate of over 85% in other independent datasets, providing a reliable evidence base for feature selection for research teams in related fields. This feature selection method based on stability assessment can effectively reduce the trial-and-error costs for other research teams in the biomarker validation stage, shorten the research cycle, and provide methodological reference and practical basis for the discovery and validation of disease biomarkers. The prediction model combines excellent predictive performance and reliability, and the combination of methylation biomarkers it screened also provides important clues for subsequent mechanism research, offering a powerful tool for the application of DNA methylation in complex phenotypic studies. Attached Figure Description

[0026] Figure 1 This is a flowchart of the prediction model construction process of the present invention.

[0027] Figures 2-6 This is an image showing the effect of the peripheral blood DNA methylation marker of the present invention. Detailed Implementation

[0028] The present invention will be further illustrated below with reference to specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. Furthermore, it should be understood that after reading the teachings of this invention, those skilled in the art can make various alterations or modifications to the invention, and these equivalent forms also fall within the scope defined by the appended claims.

[0029] Example 1

[0030] This embodiment provides a peripheral blood DNA methylation marker for predicting the thickness of the frontal cortex. The marker is one or more of the following: G_front_middle_thickness, G_S_frontomargin_thickness, S_front_inf_thickness, and S_front_middle_thickness.

[0031] like Figure 1As shown, this embodiment also provides a model for predicting the thickness of the frontal cortex. The method for constructing the model includes the following steps:

[0032] Step 1: Create a dataset

[0033] 1.1 Sample preparation:

[0034] To ensure the accuracy and reliability of subsequent DNA methylation detection results, the collection of fresh blood samples must follow strict standard procedures. EDTA anticoagulant tubes (purple caps) or sodium citrate anticoagulant tubes (blue caps) are preferred for venous blood collection, as these two anticoagulants cause minimal interference with subsequent DNA extraction and methylation analysis. Heparin anticoagulant tubes should be avoided, as heparin severely inhibits subsequent enzymatic reactions. Collected samples should be temporarily stored and transported at 4°C and must be processed within 24 hours of collection. The core processing step is to separate the target cell population, such as peripheral blood mononuclear cells (PBMCs), from whole blood using density gradient centrifugation (commonly using Ficoll reagent) to eliminate the dilution effect of highly heterogeneous blood cells on the methylation signal. The separated cell pellet should be immediately placed in an ultra-low temperature freezer at -80°C or liquid nitrogen for long-term storage until the DNA extraction stage. Aseptic techniques must be used throughout the process, and sample information must be recorded to maximize the integrity of genomic DNA and prevent degradation and unnatural changes in DNA methylation patterns.

[0035] 1.2 Data Acquisition:

[0036] The EPIC chip-based data acquisition process begins with the acquisition of high-quality genomes. First, DNA is extracted from frozen PBMC cell pellets, and its purity, concentration, and integrity are rigorously controlled using instruments such as Nanodrop or Qubit. After passing quality control, approximately 500 ng of genomic DNA undergoes bisulfite treatment. This step is crucial for the success of the experiment; it converts unmethylated cytosine to uracil, while methylated cytosine remains unchanged, thus creating differences in methylation status at the sequence level. The treated DNA is then hybridized on an EPIC chip, which covers over 850,000 CpG sites in the human genome, encompassing multiple important functional regions such as enhancers, gene promoters, and coding regions. The chip reads the hybridization signals via fluorescence scanning, and the resulting raw data file (IDAT file) contains the fluorescence intensity values ​​for each probe. This data will lay a solid foundation for subsequent bioinformatics analyses, including background correction, normalization, and screening of differentially methylated sites / regions.

[0037] Step 2: Data Preprocessing

[0038] 2.1 First, a linear correlation analysis was performed for each phenotype, and the results were obtained.

[0039] 2.2 Read the pre-flight phenotype table and the overall methylation matrix, align them according to the fixed sample ID (Sample_Name), and filter out samples that are common to both; then automatically traverse all numerical phenotype columns, remove samples with missing values ​​for each phenotype, and find the intersection with the methylation data to ensure that only samples that "have the phenotype value and have methylation data" participate in the analysis.

[0040] 2.3 For each phenotype, a round of CpG feature screening is performed using the linear correlation results. Then, these CpGs and their corresponding common samples are extracted from the global methylation matrix to form a "CpG×sample" submatrix.

[0041] Step 3: Model Building

[0042] 3.1 Construction of Continuous Prediction Model

[0043] 3.1.1 Model building data preprocessing: Prioritize the use of samples with the ID prefix ST; remove low variance features and reduce noisy / redundant features.

[0044] 3.1.2 10-fold cross-validation: Perform at least 100 bootstrap sampling analyses for each phenotype. In each round, 50% of the sample is randomly selected as the training set, and the remaining 50% is used to validate the model.

[0045] 3.1.3 Bootstrap sampling is used to perform single CpG linear regression screening on each round of subsamples. Using 50% of the samples, the top 20 significant CpGs are selected using single CpG linear regression. The stability of the feature ranking is checked every 10 iterations. Then, R² is calculated using multiple models including LASSO, linear regression, random forest, Elastic Net, and Kernel Ridge on the screened features. 2 The frequency of selection and corresponding performance of each CpG were statistically analyzed. Every 10 times, a comprehensive score was calculated for each feature using the following formula: Composite_Score = 0.7 × selection frequency + 0.3 × average R 2 .

[0046] 3.1.4 Sort by comprehensive score to obtain the current ranking; calculate Spearman correlation with the previous batch ranking and add the average change in rank of the top 20; if the correlation > 0.95 and the average rank change < 2, it is considered that "feature ranking is stable", and Bootstrap is stopped early to obtain the original result, and the result is processed according to R. 2 The results obtained by filtering for AUC >0.5 and AUC >0.75 are shown in Table 1.

[0047] Table 1. Overview of the predictive performance of frontal cortex thickness-related features

[0048]

[0049] 3.1.5 Finally, construct multiple final models:

[0050] Multiple linear regression (using only stable features), LASSO (continuous) (all features, lambda.min), random forest (stable features), Elastic Net (all features), Kernel Ridge (stable features, and number of features < number of samples), and Stacking ensemble use LM / RF / KR predictions as machine learning output, then use linear regression as a two-layer model. For linear regression and stacking, explicit prediction equation strings are constructed, such as:

[0051] Target phenotype (pheno) = intercept + β1·CpG1 + β2·CpG2 + …

[0052] 3.2 Construction of Binary Classification Prediction Model

[0053] 3.2.1 Model building data preprocessing: Prioritize the use of samples with the ID prefix ST; remove low variance features and reduce noisy / redundant features.

[0054] 3.2.2 Ten-fold cross-validation: Perform at least 100 bootstrap sampling analyses for each phenotype. In each round, randomly select 50% of the sample as the training set and use the remaining 50% to validate the model.

[0055] 3.2.3 Constructing Binary Classification Labels and Automatic Backtracking Mechanism

[0056] For continuous phenotypic values, two categories of labels, "high" and "low", are constructed according to a pre-configured binary classification strategy (binary_threshold_type), including:

[0057] (a) Median segmentation: Using the median of all samples as the threshold, samples with phenotypic values ​​greater than the median are marked as high, and the rest are marked as low;

[0058] (b) Quantile extreme group segmentation (q30_70): Calculate the 30th percentile (Q30) and 70th percentile (Q70) of the phenotype in all samples, and only retain the "extreme samples" outside these two quantiles, that is: samples with phenotype values ​​≤ Q30 are retained and marked as low; samples with phenotype values ​​≥ Q70 are retained and marked as high; samples between Q30 and Q70 (the middle 40% interval) are all discarded and do not participate in binary classification modeling.

[0059] This creates a "low" and a "high" extreme group to train a clearer binary classification model. The preferred split is q30_70. If this split results in only one class or extreme imbalance (e.g., too few samples in one class), it automatically reverts to a robust split based on the median of the entire sample to ensure sufficient samples for modeling in both classes.

[0060] 3.2.4 For each round of bootstrapping samples, a univariate logistic regression (Ventricle_Class ~ CpG) is fitted to each CpG feature. The CpGs are ranked according to their p-values, and the top 20 most significant CpGs are selected as candidate features for that round. The number of times these features are selected in the current round is recorded. In the same round, using all candidate CpG features as independent variables, a LASSO logistic regression model is trained on the bootstrapping samples, and the area under the ROC curve (AUC) is calculated on the validation set. The AUC contribution of this round is allocated to the CpGs pre-screened in this round and accumulated in the average AUC statistics of the corresponding features. Every 10 iterations, based on the statistical results so far, the reconstruction comprehensive score for each CpG is calculated.

[0061] ; .

[0062] Sort the composite features from highest to lowest to obtain the current feature ranking, and compare it with the ranking of the previous batch (first 10 rounds) using Spearman correlation and the average change of the top 20 rankings: if the ranking correlation coefficient is > 0.95 and the average change of the top 20 rankings is < 2, then the feature ranking is considered to have stabilized and the Bootstrap is stopped early; otherwise, continue iterating until the maximum number of iterations is reached.

[0063] Step 4: Comprehensive evaluation and result output of the multi-phenotypic prediction model

[0064] After completing the continuous and binary classification task modeling for each phenotype, this study summarizes and outputs all model results in a unified manner:

[0065] 4.1 Single-Table Result Management:

[0066] For each phenotype, save it in a separate directory named after the phenotype:

[0067] (a) Continuous task related files: including feature frequency and performance analysis, Bootstrap iteration results, final model performance evaluation, stable feature list, feature ranking history trajectory, and RDS files for LASSO, linear regression, random forest, ElasticNet, kernel ridge regression and Stacking models;

[0068] (b) Documents related to the binary classification task: including the comprehensive score of binary classification features, Bootstrap performance records, test set prediction results, final LASSO-logistic model and feature summary, etc.

[0069] 4.2 Performance Summary of Cross-Phenotype Models:

[0070] The script iterates through the result list of all phenotypes and compares the key performance indicators (such as R) of continuous and binary classification tasks. 2 The results of multi-phenotype modeling (such as RMSE, MAE, AUC, etc.) are combined into a summary table to provide an overview and comparison of the results, which facilitates the rapid identification of phenotypes with better predictive performance and their corresponding combinations of methylation features.

[0071] Experimental results:

[0072] Characteristics of the study population:

[0073] Data from 134 adult crew members on long-distance voyages were included. Participants were recruited through community recruitment and health check-up centers, and none had a history of clinical diagnosis of neurological diseases. Exclusion criteria included: a history of stroke, traumatic brain injury, or neurodegenerative diseases (such as Alzheimer's disease or Parkinson's disease); severe mental illness (such as schizophrenia or a major depressive episode); contraindications to MRI scans (such as implanted metal devices or claustrophobia); recent (within 6 months) use of medications that may significantly affect DNA methylation patterns (such as chemotherapy drugs or immunosuppressants); and a known systemic disease that may affect epigenetics (such as autoimmune diseases or advanced liver or kidney disease).

[0074] All participants underwent high-resolution brain MRI scans (for precise measurement of frontal cortex thickness) and peripheral venous blood collection (for whole-genome DNA methylation microarray analysis). Plasma separation was performed within 2 hours of blood sample collection, and the leukocyte layer was cryopreserved at -80°C until DNA extraction. All participants signed written informed consent forms prior to enrollment, agreeing to the use of their imaging data, blood samples, and molecular data in this study to construct and validate a predictive model of "peripheral blood characteristics-brain structural phenotype."

[0075] The central tendency and interquartile distribution of the thickness of the four frontal lobe brain regions in the population are as follows (units are based on the original brain image processing output): Middle frontal gyrus thickness (G_front_middle_thickness): mean 2.8196, median 2.8395, Q1=2.4585, Q3=3.0385, minimum 1.8780, maximum 3.3950; Superior frontal margin gyrus thickness (G_S_frontomargin_thickness): mean 2.6035, median 2.6035, Q1=2.2200, Q3=3.0310, minimum 1.3140, maximum 3.1920;

[0076] Frontal sulcus thickness (S_front_inf_thickness): mean 2.5345, median 2.5440, Q1=2.2010, Q3=2.8605, minimum 1.4120, maximum 3.6450; Frontal middle sulcus thickness (S_front_middle_thickness): mean 2.5761, median 2.5530, Q1=2.4270, Q3=2.7660, minimum 1.9450, maximum 3.0260.

[0077] Furthermore, the differences in thickness among individuals within the same brain region were significant and could be quantified by specific numerical values: the range of thickness in each brain region was 1.5170 (3.3950-1.8780) for the middle frontal gyrus, 1.8780 (3.1920-1.3140) for the superior frontal gyrus, 2.2330 (3.6450-1.4120) for the inferior frontal sulcus, and 1.0810 (3.0260-1.9450) for the middle frontal sulcus; the interquartile range (IQR) was 0.5800 (3.0385-2.4585) for the middle frontal gyrus, 0.8110 (3.0310-2.2200) for the superior frontal gyrus, 0.6595 (2.8605-2.2010) for the inferior frontal sulcus, and 0.3390 (2.7660-2.4270) for the middle frontal sulcus. The above results indicate that not only is there a significant gap between extreme individuals (minimum and maximum values), but the middle 50% of the population (Q1 to Q3) also exhibits stable dispersion.

[0078] Therefore, the frontal cortical thickness phenotype predicted by this invention has sufficient variability and distinguishability within the population (Range approximately 1.0810–2.2330, IQR approximately 0.3390–0.8110), providing a clear and quantifiable basis for the distribution of brain structure in the research population for subsequent screening of biomarkers based on peripheral blood DNA methylation characteristics and construction of a frontal cortical thickness prediction model.

[0079] Using four cortical thickness phenotypes as prediction targets (Pheno), we compared models such as LASSO regression, ElasticNet, random forest, and Stacking ensemble learning in continuous tasks; in binary classification tasks, we used the final LASSO-logistic model for validation, and uniformly adopted the q30_70 extreme grouping strategy (segmentation at 30% and 70% quantiles, with Cut_Info recording the threshold of each phenotype).

[0080] Stacking ensemble learning showed the best overall performance across the four phenotypes, with its R... 2 The mean is 0.9743 (range 0.9663–0.9814); Random Forest (stable feature) R 2 The mean is 0.9577 (range 0.9412–0.9726). In comparison, Elastic Net and LASSO have lower R... 2 The means were 0.8922 and 0.8954 respectively, showing significant fluctuations; multiple linear regression (stability feature) performed the weakest, with R... 2 The mean is 0.4679 (range 0.3608–0.5631). Further, based on the statistics of the "optimal model for each phenotype," the optimal R-values ​​for the four phenotypes are... 2 The mean was 0.9798 (range 0.9663-0.9937), corresponding to RMSE of 0.0170-0.0590 and MAE of 0.0128-0.0459. The optimal results for each phenotype are as follows:

[0081] (1) Frontal gyrus thickness (G_front_middle_thickness): Elastic Net, R 2 =0.9838;

[0082] (2) Frontal margin thickness (G_S_frontomargin_thickness): Stacking, R 2 =0.9753;

[0083] (3) Frontal sulcus thickness (S_front_inf_thickness): Stacking, R 2 =0.9663;

[0084] (4) Frontal sulcus thickness (S_front_middle_thickness): Elastic Net, R 2 =0.9937.

[0085] Overall results for the binary classification task show that after using the q30_70 extreme grouping (retaining only low and high quantile samples), the number of training set samples N_train=28 and the number of test set samples N_test=10-11 for each phenotype; the overall low / high sample count is approximately 19-20 / 18-20. The final LASSO-logistic model achieved a mean AUC of 0.8100 (range 0.60-1.00), a mean Accuracy of 0.6500 (range 0.50-0.80), a mean Sensitivity of 0.55 (range 0.40-0.60), and a mean Specificity of 0.75 (range 0.60-1.00) across the four phenotypes. The superior frontomargin thickness (G_S_frontomargin_thickness) showed the best discriminative performance (AUC=1.00, Accuracy=0.80, Specificity=1.00).

[0086] Meanwhile, to ensure the reproducibility of methylation marker screening, a Bootstrap stability assessment was conducted for each target phenotype: the frequency_Ratio of repeated inclusion of CpG sites was calculated during repeated sampling, and a comprehensive score was constructed by combining the model performance indicators.

[0087] In continuous prediction tasks, the composite score is set to 0.7 × Frequency Ratio + 0.3 × Avg R. 2 This refers to a model with a high frequency of repeated inclusion (Frequency_Ratio) and a goodness of fit (Avg_R) corresponding to the inclusion round. 2 (A better CpG site)

[0088] In the binary classification task, Composite = 0.7 × Frequency_Ratio + 0.3 × Avg_AUC is used, which refers to CpG sites with a high repeated inclusion rate (Frequency_Ratio) and good model discrimination ability (Avg_AUC) in the corresponding selection round.

[0089] Each phenotype was selected using the 70th percentile of the comprehensive score as a threshold. The top 30% of the CpGs in terms of comprehensive score were used to obtain a stable feature set, resulting in stable feature tables: continuous model markers.xlsx and binary classification model markers.xlsx.

[0090] In the continuous model feature selection results, the stable feature sets corresponding to the four frontal lobe phenotypes were of consistent size. Based on a frequency ratio (FR) ≥ 0.3, the number of superior features corresponding to each phenotype is as follows:

[0091] (1) Frontal middle gyrus thickness (G_front_middle_thickness): 19;

[0092] (2) Frontal margin thickness (G_S_frontomargin_thickness): 19;

[0093] (3) Frontal sulcus thickness (S_front_inf_thickness): 18;

[0094] (4) Frontal sulcus thickness (S_front_middle_thickness): 18.

[0095] In the feature selection results of the binary classification model, the stable feature sets corresponding to the four frontal lobe phenotypes were of consistent size. Based on FR ≥ 0.3, the number of superior features corresponding to each phenotype were as follows:

[0096] (1) Frontal middle gyrus thickness (G_front_middle_thickness): 16;

[0097] (2) Frontomargin thickness (G_S_frontomargin_thickness): 16;

[0098] (3) Frontal sulcus thickness (S_front_inf_thickness): 15;

[0099] (4) Frontal sulcus thickness (S_front_middle_thickness): 17.

[0100] Meanwhile, the overall distribution of stable features shows a very consistent "dominant distribution pattern" across the four phenotypes. In gene structural regions (features): Body and IGR dominate, accounting for approximately 72.5% of all phenotypes in the continuous model and approximately 64.5% in the binary model. This indicates that candidate CpGs are not concentrated solely in promoters, but rather fall more heavily within gene bodies / intergenic regions, consistent with the common characteristic in genome-wide methylation screening that "stable signals originate from broader genomic regions."

[0101] On the CpG island related annotations (CGI), opensea is dominant, with opensea accounting for the highest proportion (approximately 59%-64%) across all phenotypes. This indicates that candidate sites are more biased towards regions far from the CpG island (which may contain enhancers, distal regulatory elements, etc.), rather than the traditional single pathway of the "CpG island promoter".

[0102] The superiority of the "preferred feature" described in this invention is reflected in at least the following two aspects: (1) High repeatability: It is stably and repeatedly selected in multiple rounds of Bootstrap resampling iterations, reducing the sensitivity of feature selection to random sample fluctuations, thereby improving the reproducibility of marker screening results; (2) More robust prediction contribution: Avg_R 2 A high level indicates that it is not only "easily selected repeatedly" but also makes a more stable contribution to the fitting / interpretation ability of continuous prediction models.

[0103] This embodiment focuses on four cortical thickness phenotypes, establishing continuous prediction models and binary classification discriminant models respectively, and performing system evaluation. Multiple algorithms are compared in continuous tasks. As shown in Table 2, the ensemble learning model, represented by Stacking, exhibits the best overall performance, with the optimal model for the four phenotypes achieving the highest R-value. 2 The mean value reached 0.9743 (range 0.9663–0.9814), and both RMSE and MAE were at low levels, demonstrating strong prediction accuracy and stability. In the binary classification task, under extreme grouping conditions of q30_70 and small sample size, LASSO-logistic still achieved good discrimination results (mean AUC 0.8100, highest AUC 1.00). Figures 2-6 As shown in Table 3, the accuracy, sensitivity, and specificity metrics are also presented, verifying the effectiveness of the model in classification scenarios.

[0104] Table 2. Comparison of the effects of various algorithms for the continuous model of frontal cortex thickness.

[0105]

[0106] Table 3 Performance Indicators of the Binary Classification Model for Frontal Cortex Thickness

[0107]

[0108] More importantly, this embodiment not only reports model performance but also introduces Bootstrap stability analysis to screen CpG features for reproducibility: reproducibility is represented by the inclusion rate (FR), and FR is correlated with model performance (Avg_R for continuous tasks). 2 The binary classification task (Avg_AUC) is used to weight a composite score, which is then used to form a stable feature set with a 70th percentile threshold. This achieves "high-performance prediction / discrimination" and "reproducible biomarker screening based on stability assessment," providing a reliable basis and feasible process for subsequent biomarker validation, panel construction, and mechanism research.

Claims

1. A peripheral blood DNA methylation marker for predicting the thickness of the frontal cortex, characterized in that, The markers are one or more of the following: the thickness of the middle frontal gyrus, the thickness of the superior frontal gyrus, the thickness of the inferior frontal sulcus, and the thickness of the middle frontal sulcus.

2. A diagnostic product based on the peripheral blood DNA methylation marker as described in claim 1.

3. The diagnostic product according to claim 2, characterized in that, The diagnostic products are reagent kits, methylation panels, or chips.

4. The diagnostic product according to claim 2, characterized in that, The peripheral blood was obtained by venous blood collection using EDTA anticoagulant tubes or sodium citrate anticoagulant tubes.

5. A model for predicting the thickness of the frontal cortex, characterized in that, The method for constructing the model includes the following steps: S1. Obtain a sample set with complete phenotypic and methylation data; S2. Preprocess the above sample set to obtain a dataset that can be used for subsequent analysis; S3. In the above dataset, perform linear correlation analysis for each phenotype and select CpG sites that are significantly related to the phenotype as candidate features. Subsequently, a dual-track analysis system was established, and the samples were randomly stratified into training and validation sets in a 5:5 ratio to construct continuous variable prediction models and binary classification prediction models respectively. S4. Summarize and output all model results in a unified manner.

6. The model according to claim 5, characterized in that, During the construction process, Bootstrap sampling technique is used to evaluate feature stability, synthetic scores are used to comprehensively evaluate feature importance, and dynamic convergence conditions are set to automatically stop iteration.

7. The model according to claim 5, characterized in that, Based on the feature combinations selected for stability, we construct ensemble models for multiple linear regression, LASSO, random forest, Elastic Net, Kernel Ridge, and Stacking, respectively, and output prediction results and performance evaluations.

8. A system for predicting the thickness of the frontal cortex, characterized in that, The system includes the following modules: A data acquisition module is used to acquire methylation level data of the biomarker of claim 1 in the sample to be tested; The data analysis module is used to predict the data in the data acquisition module using the model described in claim 5; The output module is used to output the prediction results.

Images