A method for constructing a DNA methylation microhaplotype-based age prediction model

Abstract

This invention discloses a method for constructing an age prediction model based on DNA methylation microhaplotypes, belonging to the field of bioinformatics. Through bioinformatics analysis of methylation data from multiple whole blood samples, DNA methylation microhaplotype sites for accurate age prediction are selected using the sparse group Lasso algorithm. Further, a convolutional neural network is used for age regression modeling, and the stability and accuracy of the model are verified in multiple independent cohorts. This model provides a novel and effective technical means for age prediction in epigenetics, with high scalability and application prospects.

Description

Technical Field

[0001] This invention relates to a method for constructing an age prediction model based on DNA methylation microhaplotypes, belonging to the field of bioinformatics technology. This model is mainly applied in forensic medicine for individual age estimation, assessment of healthy aging status, and risk assessment of age-related diseases. By detecting the methylation level of age-related CpG sites, aggregating them into groups, and combining feature screening and modeling analysis, a rapid, efficient, and stable method for predicting human age is developed. Background Technology

[0002] Aging is a biological process that occurs over time, characterized by a decline in the body's ability to maintain homeostasis and repair itself, functional decline, and increased vulnerability. Among numerous biomarkers, DNA methylation, as an epigenetic modification, exhibits regular changes with age across different tissues and cell types, and these changes are highly reproducible and predictable. Based on these characteristics, current technologies typically detect the methylation levels of multiple CpG sites and integrate information from multiple sites using regression analysis, elastic networks, or machine learning methods to predict human age. These methods have led to various age prediction models based on DNA methylation, collectively known as epigenetic clocks.

[0003] Most existing epigenetic clock models rely on the methylation levels of multiple CpG loci for construction and employ statistical or machine learning methods to integrate information from multiple CpG loci to predict an individual's age. For example, Horvath et al. developed an epigenetic clock based on 353 CpG loci in 2013, stably predicting the age of 51 tissue / cell types; in the same year, Hannum et al. developed an age prediction model for whole blood samples based on 71 CpG loci. However, most existing epigenetic clocks rely on the methylation level of a single CpG locus as input features, often failing to fully capture the co-methylation signals between adjacent CpG loci, thus limiting the model's accuracy and cross-cohort applicability.

[0004] Previous studies have shown that DNA methylation deposition and demethylation are regulated by specific enzyme systems, and the methylation levels of adjacent CpG sites often exhibit a distance-dependent co-methylation relationship. Short-distance CpG sites show stronger correlations; therefore, analyzing the co-methylation patterns of adjacent CpG sites provides richer information than analyzing methylation changes at individual sites. However, most existing methods fail to fully utilize this synergistic change, resulting in the ineffective extraction of local methylation signals. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides a method for constructing an age prediction model based on DNA methylation microhaplotypes (DMHs). Bioinformatics analysis of methylation data from multiple whole blood samples is performed, and DMH loci for accurate age prediction are selected using the Sparse Group Lasso (SGL) algorithm. Further, a Convolutional Neural Network (CNN) is employed for age regression modeling, and the model's stability and accuracy are validated in multiple independent cohorts. This model provides a novel and effective technique for age prediction in epigenetics, exhibiting high scalability and application potential.

[0006] This invention provides a method for constructing an age prediction model based on DMH, employing the following technical solution: This invention provides a method for constructing a DMH-based age prediction model. Through bioinformatics analysis of methylation data from multiple whole blood samples, DMH loci for accurate age prediction are selected using the SGL algorithm. Further, a convolutional neural network is used for age regression modeling, and the stability and accuracy of the model are verified in multiple independent cohorts. Specifically, the method includes: The methylation standard matrix files (Series Matrix Files) of the human whole blood cohort GSE40279 were downloaded from the Gene Expression Omnibus (GEO) database, sourced from the Illumina Infinium HumanMethylation450 platform. Simultaneously, clinical data on various phenotypic characteristics of the samples in the cohort, including age, blood source, and microarray platform, were obtained from the GEO data. The 656 healthy samples in GSE40279 were divided into a training set (524 samples) and a test set (132 samples) at an 8:2 ratio. Data preprocessing: Due to missing data and inconsistent formatting in the human whole blood cohort GSE40279, the R package CHAMP was used for quality control to remove SNP colocation probes, non-CpG region probes, multiple matching probes, and sex chromosome-related sites to eliminate age-independent interference. Unhealthy or age-deficient samples were deleted from the cohort. Samples with more than 5% missing probe values ​​were removed to minimize potential bias caused by low-quality samples. Furthermore, probes with more than 10% missing values ​​were excluded. The remaining missing data was imputed using the KNN algorithm (K=5) from Python's scikit-learn library. The KNN method identifies k spatially similar or adjacent samples in the dataset and uses these k samples to estimate the value of missing data points. Missing values ​​are imputed using the neighborhood average of the k samples found in the dataset, resulting in preprocessed data. Assembly and extraction of DMH: Based on the official annotations of the Illumina BeadChip 450K methylation microarray, a custom Python script was used to extract the physical coordinates of all CpG sites in the hg19 genome from the preprocessed data. The coordinates were then converted to hg38 using the `liftOver` function of the R package `rtracklayer` in conjunction with the `hg19ToHg38.over.chain` file provided by UCSC. Following the principle that adjacent CpG sites should not exceed 50 bp in the DMH definition, these sites were sequentially grouped and aggregated to assemble all possible DMHs on the chip. The first CpG site in each group is the starting CpG of the DMH. Then, compare the coordinate distance between the second CpG site and the starting CpG to see if it is less than 50. If it is less than 50, continue to compare the coordinate distance between the third CpG site and the starting CpG. If the requirements of DMH are met, continue the above process. If not, combine the first two CpGs into one DMH and output the DMH. At the same time, move the starting CpG backward and repeat the above process. Write a custom Python script to extract the corresponding β value of the site contained in each real queue of the DMH from the reference file of the assembled DMH, and concatenate the site β values ​​of the DMH into a feature column using a specified delimiter. Finally, a total of 44,878 DMHs are assembled. SGL Screening for Age-Related Features: The SGL method was implemented in Python to screen for age-related features in the 44,878 DMHs. The L2 regularization strength was 0.1, and the L1 regularization strength was 0.05. First, an SGL was fitted to all training samples to derive the regression coefficients for each locus. The absolute coefficients of all loci within each DMH group were summed and used as the importance score for that group. Then, the bootstrap method (bootstrap=50) in sklearn was used to repeatedly sample the training set, and the screening frequency of the selected DMHs (i.e., non-zero groups) in the resampling was counted. Finally, the sum of the absolute coefficients within each group obtained from the SGL fitting and the screening frequency obtained from the bootstrap method were combined to evaluate the importance of each DMH group to age. 29 age-related DMHs were selected, assembled from 64 CpG loci. Convolutional Neural Network Modeling: The 29 age-related DMHs consist of varying numbers of CpG sites. To ensure consistent dimensionality of the model input features, the DMH containing the most CpG sites was used as the standard (L_max = 3 CpG), and other DMHs were padded to the right with a padding value of 0. To better distinguish between true and padded values ​​in the DMHs during forward propagation, a missing mask was used to preserve the true structure of the input features as much as possible, with a true value of 1 and a padded value of 0. The model input is represented by a shape tensor (N, M, L), where N represents the number of samples, M represents the number of DMHs, and L represents the number of CpG sites in the longest DMH. In the network, each DMH is encoded as a fixed-length one-dimensional sequence with a β value and a corresponding mask. Two channels are used, employing one-dimensional convolutional layers and residual blocks to capture local co-methylation patterns between CpG sites. The hidden layer dimension is 32, and Dropout = 0.3 is used to prevent overfitting. Training uses the AdamW optimizer combined with Cosine Annealing learning rate scheduling and early stopping strategies; ensuring consistent age distribution between training and test sets, with stratified sampling of age spans over 30 years, resulting in the DMH-AgeNet model after training; recording and outputting the loss, mean absolute error (MAE), root mean square error (RMSE), median absolute error (MedAE), Pearson correlation coefficient (R), and coefficient of determination (R²) for each round of training and validation sets during training. 2 This is for use in subsequent model performance evaluation.

[0007] The DMH-AgeNet model was compared with three classic epigenetic clock models (Hannum, Horvath, and PhenoAge) in terms of feature overlap at the locus level. The 29 age-related DMHs used were split into individual CpG loci, and the intersection relationship of the CpG locus sets was analyzed.

[0008] Among the 29 age-related DMHs, the CpG with the largest and most stable correlation coefficient was selected to represent the DHM for visualization. The vertical axis represents the DMH and the horizontal axis represents the sample. The samples are arranged in hierarchical clusters, and the position on the chromosome is marked on the right. The gray bars related to age represent the age of each sample (18-101 years).

[0009] Gene annotation of age-related DMHs: The 29 age-related DMHs used in the DMH-AgeNet model were annotated.

[0010] Validation and Evaluation of the DMH-AgeNet Model on External Validation Sets: To further verify the universality of the constructed DMH-AgeNet model, it was externally validated on six independent blood cohorts: GSE42861, GSE55763, GSE147221, GSE84727, GSE72775, and GSE72680. All six independent cohorts were downloaded from the GEO database and underwent the same preprocessing steps as GSE40279. Healthy samples were selected to ensure data consistency and comparability. In each independent cohort, DMH data was used as input features to predict the age of each sample group. The model performance metrics of the DMH-AgeNet model on different samples were obtained by comparing the predicted age with the actual age.

[0011] Performance comparison of DMH-AgeNet model with three classic epigenetic clocks: To comprehensively evaluate the model performance of DMH-AgeNet, the six external validation sets were input into three classic epigenetic clock models (Hannum, Horvath, and PhenoAge) and the predicted ages were calculated using the R package methylclock (v1.0.1). The performance metrics were compared with those of DMH-AgeNet.

[0012] Preferably, the 29 age-related DMHs are: chr1:3164415-3164431, chr1:25257624-25257629, chr1:68512777-68512807, chr1:151298905-151298954, chr1:207925081-207925112, chr2:106015739-106015771, chr2:182545555-18 2545572, chr2:197458521-197458548, chr3:10206254-10206258, chr3:160167977-160167990, chr5:63 461216-63461232, chr6:11044877-11044894, chr6:30140113-30140145, chr7:75931606-75931608, chr7 :130418315-130418324、chr7:130419116-130419159、chr7:130419485-130419514、chr8:16860088-168 60121, chr8:91094811-91094847, chr8:131455249-131455276, chr9:86755499-86755504, chr10:323458 64-32345866, chr12:7023346-7023352, chr12:21926447-21926476, chr13:95952902-95952937, chr15:83332065-83332087, chr17:56832991-56833000, chr18:77139072-77139085 and chr20:62168831-62168878.

[0013] Preferably, the 64 CpG sites are: cg14361627, cg08097417, cg07955995, cg16867657, cg24724428, cg21572722, cg06639320, cg22454769, cg24079702, cg17621438, cg07850154, cg09692396, cg04604946, cg26685941, and cg05412028. , cg16567172, cg07571951, cg23729020, cg10779981, cg21868031, cg25584930, cg20592391, cg04528819, c g20426994, cg08327269, cg19711579, cg03767531, cg04027548, cg03352106, cg21674704, cg15707833, cg2 6161329, cg16477091, cg04733838, cg10106965, cg01541867, cg13868084, cg17474524, cg14010720, cg124 02251, cg26290632, cg00510111, cg26537076, cg07553761, cg15618978, cg13612317, cg14776578, cg14149 172, cg01974375, cg22901840, cg20149168, cg24884543, cg15526954, cg02281167, cg03771840, cg18240463, cg09990962, cg20695936, cg13106389, cg18087266, cg04451175, cg27110277, cg08719712 and cg02778245.

[0014] Compared with the prior art, the present invention has the following beneficial effects: This invention constructs an age prediction model based on DMH, successfully screening a set of age-related DMHs. SGL can perform feature selection between groups and achieve sparsity within groups, making it very suitable for screening DHMs. This invention uses a convolutional neural network to characterize and regress the joint methylation pattern of adjacent CpG sites within the DMH. Compared with the scheme using the methylation level of a single CpG site, it makes fuller use of local co-methylation information, thereby improving the expression efficiency of age information. Furthermore, it was validated in six independent cohorts and obtained stable prediction performance, indicating that the age prediction model has good cross-cohort applicability and robustness. Attached Figure Description

[0015] Figure 1 This is a schematic diagram illustrating the method for constructing an age prediction model based on DNA methylation microhaplotypes according to the present invention. Figure 2 This is a heatmap of the normalized eigenvalues ​​(z-scores) of age-related methylation microhaplotypes in the DMH-AgeNet model of this invention. Figure 3 For the performance evaluation of the DMH-AgeNet model built based on convolutional neural networks in this invention: A is the training set; B is the test set; Figure 4 For the annotation and gene localization of age-related DMH in this invention; Figure 5 This is the queue dataset used in this invention; Figure 6 This is the external validation result of the DMH-AgeNet model constructed based on the convolutional neural network in this invention. Detailed Implementation

[0016] To make the present invention more apparent and understandable, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0017] This invention provides a method for constructing an age prediction model based on DMH. By performing bioinformatics analysis on methylation data from multiple whole blood samples, DMH sites for accurate age prediction are selected using the SGL algorithm. Furthermore, age regression modeling is performed using a convolutional neural network, and the stability and accuracy of the model are verified in multiple independent cohorts.

[0018] Example 1 This invention provides a method for constructing an age prediction model based on DMH, such as... Figure 1 As shown, it specifically includes: The methylation standard matrix file of the human whole blood cohort GSE40279 was downloaded from the GEO database, sourced from the Illumina Infinium HumanMethylation450 platform. Simultaneously, clinical data on various phenotypes corresponding to the samples in the cohort, including age, blood source, and microarray platform, were obtained from the GEO data. The 656 healthy samples in GSE40279 were divided into a training set (524 samples) and a test set (132 samples) at an 8:2 ratio. Data preprocessing: Data from the human whole blood cohort GSE40279 contained missing data and inconsistent formats. The R package CHAMP was used for quality control to remove SNP colocation probes, non-CpG region probes, multiple matching probes, and sex chromosome-related loci to eliminate age-independent interference. Unhealthy or age-related missing samples were deleted. Samples with more than 5% missing values ​​among probes were removed to minimize potential bias caused by low-quality samples. Furthermore, probes with more than 10% missing values ​​were excluded. The remaining missing data was imputed using the KNN algorithm (K=5) from Python's scikit-learn library. The KNN method identifies k spatially similar or adjacent samples in the dataset and uses these k samples to estimate the value of missing data points. Missing values ​​are imputed using the neighborhood average of the k samples found in the dataset, resulting in preprocessed data. Assembly and extraction of DMH: Based on the official annotations of the Illumina BeadChip 450K methylation microarray, a custom Python script was used to extract the physical coordinates of all CpG sites in the hg19 genome from the preprocessed data. The coordinates were then converted to hg38 using the `liftOver` function of the R package `rtracklayer` in conjunction with the `hg19ToHg38.over.chain` file provided by UCSC. Following the principle that adjacent CpG sites should not exceed 50 bp in the DMH definition, these sites were sequentially grouped and aggregated to assemble all possible DMHs on the chip. The first CpG site in each group was the starting CpG of the DMH. Next, the coordinate distance between the second CpG site and the starting CpG is compared to see if it is less than 50. If it is less than 50, the coordinate distance between the third CpG site and the starting CpG is compared. If the requirements of DMH are met, the above process continues. If not, the first two CpGs are combined into one DMH and output. At the same time, the starting CpG is moved backward, and the above process is repeated. A custom Python script is written to extract the corresponding β values ​​of the sites contained in each real queue of the DMH from the reference file of the assembled DMH, and concatenate the site β values ​​of the DMH into a feature column using a specified delimiter. Finally, a total of 44,878 DMHs are assembled. SGL Screening for Age-Related Features: The SGL method was implemented in Python to screen 44,878 DMHs for age-related features. The L2 regularization strength was 0.1 and the L1 regularization strength was 0.05. First, SGL was fitted once on all training set samples to derive the regression coefficients of each locus. The absolute coefficients of all loci within each DMH group were summed and used as the importance score of the group. Then, the bootstrap method in sklearn (bootstrap=50) was used to repeatedly sample the training set, and the screening frequency of the selected DMHs (i.e., non-zero groups) in the resampling was counted. Finally, the sum of the absolute coefficients within the group obtained from fitting SGL once and the screening frequency obtained from the bootstrap method were combined to evaluate the importance of each DMH group to age. 29 age-related DMHs were screened, which were assembled from 64 CpG loci, as shown in Table 1. Table 1. Age-related DMH and its constituent CpG sites identified by SGL screening.

[0019] The results of SGL screening for age-related DMHs are shown in Table 2. The chromosome coordinate interval represents each DMH. By comparing the correlation coefficient and screening frequency of the DMHs processed by SGL, the following 29 DMHs that are highly age-related and have good stability in repeated evaluations were finally identified. Table 2 Regression coefficients for SGL screening of age-related DMH

[0020] Convolutional Neural Network Modeling: The 29 age-related DMHs consist of varying numbers of CpG sites. To ensure consistent dimensionality of the model input features, the DMH containing the most CpG sites was used as the standard (L_max = 3 CpG), and other DMHs were padded to the right with a padding value of 0. To better distinguish between true and padded values ​​in the DMHs during forward propagation, a missing mask was used to preserve the true structure of the input features as much as possible; the true value was 1, and the padded value was 0. The model input is represented by a shape tensor (N, M, L), where N represents the number of samples, M represents the number of DMHs, and L represents the number of CpG sites in the longest DMH. In the network, each DMH is encoded as a fixed-length one-dimensional sequence with a β value and a corresponding mask. One-dimensional convolutional layers and residual blocks are used to capture local co-methylation patterns between CpG sites. The hidden layer dimension is 32, and Dropout = 0.3 is used to prevent overfitting. Training uses the AdamW optimizer combined with Cosine Annealing learning rate scheduling and early stopping strategies; ensuring consistent age distribution between training and test sets, with stratified sampling across 30-year age ranges to obtain the DMH-AgeNet model after training; recording and outputting the loss, mean absolute error (MAE), root mean square error (RMSE), median absolute error (MedAE), Pearson correlation coefficient (R), and coefficient of determination (R²) for each round of training and validation sets during training. 2 This is for use in subsequent model performance evaluation.

[0021] The 29 DMHs used were broken down into individual CpG sites. Analysis of the intersection relationships among the CpG site sets revealed that the 64 CpG sites used by DMH-AgeNet had minimal overlap with the three classic epigenetic clock models: sharing 9, 2, and 2 CpG sites with the Hannum clock (71 CpGs), Horvath clock (353 CpGs), and PhenoAge clock (513 CpGs), respectively. Furthermore, the number of CpG sites shared by all four models was zero. The predictive features extracted by DMH-AgeNet showed significant differences from the three classic epigenetic clocks. The model predicts age by capturing local co-methylation patterns. In the model built based on the 29 DMHs, 52 CpG sites were not present in these classic epigenetic clocks. These sites may be related to the co-methylation effect between DMH sites, thus introducing a new dimension of information for age prediction.

[0022] Among 29 age-related DMHs, the CpG with the highest and most stable correlation coefficient was selected to represent the DHM for visualization; for example Figure 2As shown, the vertical axis represents DMH, and the horizontal axis represents the samples. The samples are arranged in hierarchical clusters, and the positions on the chromosomes are marked on the right. The age-related gray bars represent the age of each sample (18-101 years old). Overall, most DMHs show monotonic changes related to age in the DMH standardized feature value heatmap. The z-score of some DMHs increases with age, while that of others decreases with age.

[0023] like Figure 3 As shown in Table 3, DMH-AgeNet performed as follows on the training set: the Pearson correlation between predicted age and actual age was R≈0.96, and the coefficient of determination was R0.96. 2 R² ≈ 0.92; errors were MAE ≈ 3.29 years, MedAE ≈ 2.64 years, and RMSE ≈ 4.34 years; the model maintained similar performance in the test set: R² ≈ 0.95, R² ≈ 0.92; 2 The training and testing timelines were approximately 0.90, MAE approximately 3.40 years, MedAE approximately 2.95 years, and RMSE approximately 4.35 years. The overall difference between training and testing was small, indicating good generalization and no obvious overfitting. Table 3 Evaluation results of DMH-AgeNet modeling

[0024] Gene annotation of age-related DMH: Annotate the 29 age-related DMHs used in DMH-AgeNet, such as... Figure 4 As shown, 24 DMHs were located in the promoter-adjacent region, with 2 in the distal spacer region, 1 in an intron, 1 in an exon, and 1 in the 3'UTR. Subsequently, each DMH was mapped to the nearest gene, revealing genes associated with age or epigenetic regulation, such as ELOVL2, RUNX3, KLF14, TRIM15, and NFATC1. Statistical analysis showed that approximately 83% were located in the promoter-adjacent region. This distribution suggests that these DMHs are prevalent in gene regulatory regions and may influence gene transcription by altering the methylation levels of promoters or other regulatory elements.

[0025] Validation and evaluation of the DMH-AgeNet model on an external validation set: To further verify the universality of the constructed DMH-AgeNet model, it was tested on six independent blood cohorts: GSE42861, GSE55763, GSE147221, GSE84727, GSE72775, and GSE72680 (e.g., Figure 5External validation was performed on (as shown); all six independent cohorts were downloaded from the GEO database and underwent the same preprocessing steps as GSE40279, selecting healthy samples to ensure data consistency and comparability; in each independent cohort, its DMH data was used as input features to predict the age of each group of samples, and the model performance index of the DMH-AgeNet model on different samples was obtained by comparing the predicted age with the actual age. Figure 6 As shown: The MAE of the DMH-AgeNet model is 3.38 years and the RMSE is 4.40 years in the GSE42861 queue; 3.89 years and 5.04 years in the GSE147221 queue; 3.31 years and 4.33 years in the GSE55763 queue; 3.84 years and 4.73 years in the GSE84727 queue; 4.61 years and 5.94 years in the GSE72680 queue; and 4.75 years and 6.14 years in the GSE72775 queue.

[0026] Performance comparison of the DMH-AgeNet model with three classic epigenetic clock models: To comprehensively evaluate the model performance of DMH-AgeNet, the six external validation sets were input into three classic epigenetic clock models (Hannum, Horvath, and PhenoAge). The predicted ages were calculated using the R package methylclock (v1.0.1), and the performance metrics were compared with DMH-AgeNet, as shown in Table 4. In five of the six independent validation cohorts (GSE42861, GSE147221, GSE55763, GSE84727, and GSE72680), the DMH-AgeNet model outperformed the other three classic epigenetic clock models in terms of MAE and RMSE. In the GSE72775 cohort, only Hannum clock slightly outperformed DMH-AgeNet, indicating that DMH features can capture age-related information and have good generalization ability across different cohorts.

[0027] Table 4 Performance comparison of the DMH-AgeNet model with three classic epigenetic clock models

[0028] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any form or substance. It should be noted that those skilled in the art can make various improvements and additions without departing from the present invention, and these improvements and additions should also be considered within the scope of protection of the present invention. Any modifications, alterations, and equivalent changes made by those skilled in the art based on the above-disclosed technical content without departing from the spirit and scope of the present invention are equivalent embodiments of the present invention. Furthermore, any modifications, alterations, and evolutions made to the above embodiments based on the essential technology of the present invention still fall within the scope of the technical solution of the present invention.

Claims

1. A method for constructing an age prediction model based on DNA methylation microhaplotypes, characterized in that, Bioinformatics analysis was performed on methylation data from multiple whole blood samples. A sparse-group Lasso algorithm was used to screen for DNA methylation microhaplotypes (DMH sites) suitable for accurate age prediction. Further age regression modeling was performed using a convolutional neural network, and the model's stability and accuracy were validated in multiple independent cohorts. Specifically, this included: The methylation standard matrix file of the human whole blood cohort GSE40279 was downloaded from the GEO database, sourced from the Illumina Infinium HumanMethylation450 platform. Simultaneously, clinical data on various phenotypes corresponding to the samples in the cohort were obtained from the GEO data, including information on age, blood source, and microarray platform. The 656 healthy samples in GSE40279 were divided into a training set of 524 samples and a test set of 132 samples at an 8:2 ratio. Data preprocessing: Due to missing data and inconsistent formatting in the human whole blood cohort GSE40279, quality control was performed using the R package CHAMP to remove SNP co-location probes, non-CpG region probes, multiple matching probes, and sex chromosome-related loci to eliminate age-independent interference. Samples from unhealthy or age-deficient groups were deleted. Samples with more than 5% missing probe values ​​were removed to minimize potential bias caused by low-quality samples. Furthermore, probes with more than 10% missing values ​​were excluded. The remaining missing data was imputed using the KNNImputer algorithm (K=5) from Python's scikit-learn library to obtain the preprocessed data. Assembly and extraction of DMHs: Based on the official annotations of the Illumina BeadChip 450K methylation microarray, a custom Python script was used to extract the physical coordinates of all CpG sites in the hg19 genome from the preprocessed data. The coordinates were then converted to hg38 using the liftOver function of the rtracklayer package in conjunction with the hg19ToHg38.over.chain file provided by UCSC. Then, following the principle that adjacent CpG sites should not exceed 50 bp in the DMH definition, these sites were sequentially grouped and aggregated to assemble all possible DMHs on the chip. A custom Python script was written to extract the corresponding β values ​​of the sites contained in each real cohort of the DMH from the reference file of the assembled DMHs, and the site β values ​​of the DMH were concatenated into a feature column using a specified delimiter. Finally, a total of 44,878 DMHs were assembled. Sparse group Lasso screening for age-related features: The sparse group Lasso method was implemented in Python to screen the 44,878 DMHs for age-related features. The L2 regularization strength was 0.1 and the L1 regularization strength was 0.

05. First, a sparse group Lasso was fitted to all training set samples to derive the regression coefficients of each site. The absolute coefficients of all sites within each DMH group were summed as the importance score of the group. Then, the bootstrap method in sklearn (bootstrap=50) was used to repeatedly sample the training set, and the screening frequency of the selected DMHs in the resampling was counted. Finally, the sum of the absolute coefficients within each group obtained from the sparse group Lasso and the screening frequency obtained from the bootstrap method were combined to evaluate the importance of each DMH group to age. 29 age-related DMHs were screened, which were assembled from 64 CpG sites. Convolutional Neural Network Modeling: The 29 age-related DMHs consist of varying numbers of CpG sites. To ensure consistent dimensionality of the model input features, the DMH containing the most CpG sites was used as the standard, and other DMHs were padded to the right with a padded value of 0. To better distinguish between true and padded values ​​in the DMHs during forward propagation, a missing mask was used to preserve the true structure of the input features as much as possible, with a true value of 1 and a padded value of 0. The model input was represented by shape tensors N, M, and L, where N represents the number of samples, M represents the number of DMHs, and L represents the number of CpG sites in the longest DMH. In the network, each DMH was encoded as a fixed-length one-dimensional sequence with two channels: a β value and a corresponding mask. One-dimensional convolutional layers and residual blocks were used to capture local co-methylation patterns between CpG sites. The hidden layer dimension was 32, and Dropout = 0.3 was used to prevent overfitting. Training was performed using the AdamW optimizer combined with Cosine. Annealing learning rate scheduling and early stopping strategy; ensuring consistent age distribution between training and test sets, with stratified sampling of age spans over 30 years, resulting in the DMH-AgeNet model after training; recording and outputting the loss, mean absolute error, root mean square error, median absolute error, Pearson correlation coefficient, and coefficient of determination for each round of training and validation sets during training for subsequent model performance evaluation. Validation and evaluation of the DMH-AgeNet model on an external validation set: To further verify the universality of the constructed DMH-AgeNet model, it was externally validated on six independent blood cohorts: GSE42861, GSE55763, GSE147221, GSE84727, GSE72775, and GSE72680.

2. The construction method as described in claim 1, characterized in that, The 29 age-related DMHs are: chr1:3164415-3164431, chr1:25257624-25257629, chr1:68512777-68512807, chr1:151298905-151298954, chr1:207925081-207925112, chr2:106015739-106015771, chr2:182545555-182545572, chr2: 197458521-197458548, chr3:10206254-10206258, chr3:160167977-160167990, chr5:63461216-6346 1232, chr6:11044877-11044894, chr6:30140113-30140145, chr7:75931606-75931608, chr7:13041831 5-130418324, chr7:130419116-130419159, chr7:130419485-130419514, chr8:16860088-16860121, c hr8:91094811-91094847, chr8:131455249-131455276, chr9:86755499-86755504, chr10:32345864-32 345866, chr12:7023346-7023352, chr12:21926447-21926476, chr13:95952902-95952937, chr15:83332065-83332087, chr17:56832991-56833000, chr18:77139072-77139085 and chr20:62168831-62168878.

3. The construction method as described in claim 1, characterized in that, The 64 CpG sites are: cg14361627, cg08097417, cg07955995, cg16867657, cg24724428, cg21572722, cg06639320, cg22454769, cg24079702, cg17621438, cg07850154, cg09692396, cg04604946, cg26685941, cg05412028, cg 16567172, cg07571951, cg23729020, cg10779981, cg21868031, cg25584930, cg20592391, cg04528819, cg2 0426994, cg08327269, cg19711579, cg03767531, cg04027548, cg03352106, cg21674704, cg15707833, cg261 61329, cg16477091, cg04733838, cg10106965, cg01541867, cg13868084, cg17474524, cg14010720, cg1240 2251, cg26290632, cg00510111, cg26537076, cg07553761, cg15618978, cg13612317, cg14776578, cg141491 72, cg01974375, cg22901840, cg20149168, cg24884543, cg15526954, cg02281167, cg03771840, cg18240463, cg09990962, cg20695936, cg13106389, cg18087266, cg04451175, cg27110277, cg08719712 and cg02778245.

Images