Pathway-level epigenetic modeling illuminates the methylation architecture to asthma risk across tissues Authors: Zhongting Huang, Yuxin Liu, Fuqiang Cai, Jijun Zhu, Shenghan Wang, Pan Li, Shunjie Zhang, Zhijian Song, Weixin Liu, Huayong Liu, Junfang Chen Published: 2026 ABSTRACT BackgroundAsthma is a clinically heterogeneous airway disorder characterized by complex interactions between environmental exposures, immune activation, and molecular regulatory programs, whose underlying mechanisms are not fully elucidated by known genetic loci. DNA methylation serves as a mechanistic interface bridging genetic predisposition and environmental influences; however, most epigenetic studies remain confined to isolated CpG sites, lacking robust biological interpretability.MethodsWe developed a cross-tissue, multi-cohort, and mechanistically interpretable epigenetic framework to delineate pathway-level methylation mechanisms underlying asthma. Leveraging data from 908 participants across one combined training cohort and three independent validation cohorts, we constructed a linear support vector classifier based on pathway-derived methylation scores. Additionally, SHapley Additive exPlanations (SHAP) were applied to quantify the contributions of individual pathways. To assess the statistical significance of pathway contributions, one-sample t-tests were performed for each pathway’s SHAP values against zero, followed by Benjamini–Hochberg false discovery rate (FDR) correction to obtain adjusted p values.ResultsThe model exhibited reproducible and cross-tissue performance, achieving area under the curve (AUC) values of 0.792 (95% CI: 0.782–0.799; GSE65163) and 0.980 (95% CI: 0.974–0.990; GSE201872) in two airway epithelial cohorts, and 0.736 (95% CI: 0.690–0.771; GSE104471) in peripheral blood samples. Pathway interpretability analyses identified dominant roles of amino acid metabolism, epithelial and mesodermal developmental programs, metabolic-immune transport pathways, and neuroimmune signalling in shaping asthma-associated methylation patterns. Mediation analyses further revealed that these pathways influence asthma both directly and indirectly via eosinophil activity, epithelial proliferative dynamics, and nitric oxide-linked airway inflammation. Notably, pathways annotated by GO:0061205 and GO:0098727 exerted significant direct effects independent of immune intermediates.ConclusionsThis study describes a pathway-level methylation model designed for biological interpretability that shows associations with both clinical severity and latent molecular heterogeneity. It provides statistical evidence contributing to the understanding of epigenetic, immune, and metabolic signatures in asthma, offering a potential framework warranting further validation for precision respiratory medicine.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s13072-026-00685-y. Introduction Asthma is a highly prevalent and heterogeneous chronic inflammatory disease of the airways. It affects over 260 million people worldwide and contributes substantially to global morbidity and healthcare burden [1, 2]. Despite extensive research, the molecular mechanisms underlying asthma onset, persistence, and heterogeneity remain incompletely understood. Increasing evidence suggests that asthma arises from complex interactions among genetic susceptibility, environmental exposures, and immune dysregulation [3, 4]. However, the molecular intermediates that bridge environmental influences and phenotypic outcomes remain a central challenge in respiratory and immunological research. Understanding these mechanisms at a systems level is essential for improving disease prediction, prevention, and precision treatment [5, 6]. Genome-wide association studies (GWAS) have identified numerous asthma risk loci, including variants in IL33, IL1RL1, and ORMDL3. These genes implicate pathways in Th2 inflammation and epithelial barrier function [7, 8]. Yet, these genetic variants explain only a modest proportion of disease heritability and fail to account for the dynamic impact of environmental exposures such as allergens and pollutants [4, 9]. In recent years, epigenome-wide association studies (EWAS) have provided valuable insights into how DNA methylation mediates environmental effects on immune regulation and airway remodelling. Multiple studies report differential methylation in immune-related genes and cytokine pathways in asthma patients. This suggests epigenetic reprogramming plays a key role in disease susceptibility and persistence [10–13]. Nevertheless, most existing epigenetic studies remain constrained by several methodological and conceptual limitations. First, they often focus on individual CpG sites rather than functionally coherent pathways, limiting biological interpretability. Second, single-cohort or single-tissue designs hinder generalizability, as methylation patterns vary across tissues and populations. Third, current machine learning models for asthma prediction tend to maximize classification accuracy without addressing interpretability, thereby restricting their translational potential [14, 15]. These limitations collectively underscore the need for integrative, interpretable, and biologically grounded analytical frameworks. A critical gap persists in understanding how epigenetic alterations at the pathway level translate into measurable respiratory and immune phenotypes. Previous studies have established associations between methylation markers and asthma risk. However, few have elucidated causal mechanisms linking molecular variation to clinical outcomes [13, 16, 17]. Moreover, the extent to which pathway-level methylation signatures can generalize across tissues or populations remains unclear [18, 19]. Addressing these questions is essential for transforming asthma epigenetics from descriptive correlation toward predictive and mechanistic understanding. The present study aimed to construct a cross-tissue, multi-cohort, and biologically interpretable methylation-based model for asthma prediction. We hypothesized that pathway-level methylation signatures, aggregated from functionally related CpGs, capture essential biological processes that mediate disease risk and phenotypic variation. To test these hypotheses, we integrated data from four independent cohorts. These included airway epithelial and peripheral blood samples for cross-tissue validation. Using GO pathway-level score, we condensed high-dimensional methylation data into interpretable biological axes. A linear support vector classifier was implemented for prediction, selected for its balance of performance and interpretability. SHapley Additive exPlanations (SHAP) identified the most influential pathways. Mediation modelling was employed to quantify the direct and indirect effects of methylation on respiratory phenotypes. This combination of interpretable machine learning and causal modeling provides a unified framework. It bridges statistical prediction with biological inference [20–22]. By integrating pathway-level methylation profiling, interpretable artificial intelligence (AI), and causal inference, this study establishes a novel paradigm for mechanistically grounded epigenetic modeling of asthma. The framework improves predictive accuracy and biological interpretability. It also clarifies quantitative relationships between molecular pathways, immune processes, and clinical phenotypes. These findings may refine our understanding of asthma heterogeneity and support molecularly informed endotyping. This could advance the field of precision respiratory medicine. Methods Study design and datasets This study integrated multiple DNA methylation cohorts to develop and validate a biologically interpretable asthma classification model. Six publicly available datasets were included: GSE172365, GSE250513, GSE85566, GSE65163, GSE201872, and GSE104471 [23–28]. As summarized in Table 1, samples without disease labels were removed, and all datasets were processed under a unified analytical framework to ensure cross-cohort compatibility. Table 1Baseline demographic and clinical characteristics of participants across datasetsGEO IDTissueStatusSex (M/F)Age (mean ± SD)Sample sizeGroupGSE172365Cultured airway epithelial cellsControl76/4644.2 ± 13.9122 (59.2%)TrainCase50/3445.8 ± 14.484(40.8%)GSE250513nasal airway epithelial cellsControl65/11726.9 ± 15.3182(55.0%)Case60/8928.0 ± 17.8149(45.0%)GSE85566airway epithelial cellsControl14/2738.2 ± 11.341(35.7%)Case22/5239.3 ± 12.874(64.3%)GSE65163nasal-epitheliumControl17/1910.9 ± 0.936(50.0%)Test 1Case19/1711.1 ± 0.836(50.0%)GSE201872Primary bronchial epithelial cellsControl14/2738.2 ± 11.341(36.6%)Test 2Case21/5039.1 ± 12.771(63.4%)GSE104471PBMCCase12/2442.5 ± 17.336(50.0%)Test 3Control18/1830.2 ± 5.836(50.0%)Summary of publicly available datasets and baseline participant characteristics analyzed in this study. Datasets were obtained from the Gene Expression Omnibus (GEO). For participant characteristics, data are shown at the group level (Train, Test1, Test2, Test3) under the corresponding dataset column for each group. Platform and tissue are shown for each dataset individually. Continuous variables are presented as mean ± standard deviation (SD), and categorical variables as counts and percentages Table 1Baseline demographic and clinical characteristics of participants across datasetsGEO IDTissueStatusSex (M/F)Age (mean ± SD)Sample sizeGroupGSE172365Cultured airway epithelial cellsControl76/4644.2 ± 13.9122 (59.2%)TrainCase50/3445.8 ± 14.484(40.8%)GSE250513nasal airway epithelial cellsControl65/11726.9 ± 15.3182(55.0%)Case60/8928.0 ± 17.8149(45.0%)GSE85566airway epithelial cellsControl14/2738.2 ± 11.341(35.7%)Case22/5239.3 ± 12.874(64.3%)GSE65163nasal-epitheliumControl17/1910.9 ± 0.936(50.0%)Test 1Case19/1711.1 ± 0.836(50.0%)GSE201872Primary bronchial epithelial cellsControl14/2738.2 ± 11.341(36.6%)Test 2Case21/5039.1 ± 12.771(63.4%)GSE104471PBMCCase12/2442.5 ± 17.336(50.0%)Test 3Control18/1830.2 ± 5.836(50.0%)Summary of publicly available datasets and baseline participant characteristics analyzed in this study. Datasets were obtained from the Gene Expression Omnibus (GEO). For participant characteristics, data are shown at the group level (Train, Test1, Test2, Test3) under the corresponding dataset column for each group. Platform and tissue are shown for each dataset individually. Continuous variables are presented as mean ± standard deviation (SD), and categorical variables as counts and percentages Baseline demographic and clinical characteristics of participants across datasets 84 (40.8%) 182 (55.0%) 149 (45.0%) 41 (35.7%) 74 (64.3%) 36 (50.0%) 36 (50.0%) 41 (36.6%) 71 (63.4%) 36 (50.0%) 36 (50.0%) Summary of publicly available datasets and baseline participant characteristics analyzed in this study. Datasets were obtained from the Gene Expression Omnibus (GEO). For participant characteristics, data are shown at the group level (Train, Test1, Test2, Test3) under the corresponding dataset column for each group. Platform and tissue are shown for each dataset individually. Continuous variables are presented as mean ± standard deviation (SD), and categorical variables as counts and percentages The training dataset consisted exclusively of airway epithelial samples, whereas the remaining datasets were used for independent validation and cross-tissue evaluation. Specifically, GSE65163 and GSE201872 served as primary validation cohorts due to their comprehensive phenotypic annotations. GSE104471, derived from peripheral blood, was used to assess tissue generalizability of the classifier. All datasets were aligned by genomic coordinates. CpG probes with > 20% missingness were excluded, and missing methylation values were imputed using a K-nearest neighbor algorithm (K = 5), resulting in a harmonized dataset containing 340,047 CpG sites. These preprocessing steps ensured high data integrity and consistent feature representation across cohorts. All datasets included in this study were originally generated in accordance with the principles of the Declaration of Helsinki and relevant international ethical guidelines for human subject research. Detailed statements of ethical approval, consent to participate, and consent to publish for each dataset are provided in the Supplementary Materials. As this study involved secondary analyses of fully de-identified, publicly available data, no additional institutional review board approval or consent procedures were required for the present work. DNA methylation preprocessing and pathway feature construction DNA methylation β-values were first adjusted for age and sex effects, and then normalized using functional normalization implemented in the minfi R package. The adjusted methylation values were subsequently converted to M-values for downstream analysis [29]. To integrate DNA methylation data from multiple cohorts and tissue types, we adopted a data-driven normalization and modelling strategy consistent with standard machine learning practice. All methylation β-values were transformed to M-values and subsequently standardized using z-score normalization, with mean and standard deviation parameters estimated exclusively from the training set within each cross-validation fold and applied to the corresponding test set without re-estimation to prevent information leakage. This approach ensures that normalization does not introduce bias from test data while maintaining comparability across samples. Rather than applying explicit batch-correction algorithms, we relied on the supervised learning framework to implicitly learn and adapt to cohort-specific and tissue-specific variation through the pathway-level feature construction process. By aggregating individual CpG signals into biologically informed pathway scores, the model captures systemic epigenetic patterns that are more robust to tissue-specific noise and cohort-level technical variation than probe-level measurements. Tissue type and cohort identifiers were available as metadata but were not explicitly adjusted for in the primary analysis, allowing the model to learn relevant variation directly from the methylation data. We acknowledge that this approach differs from specialized harmonization methods such as ComBat, but the observed cross-tissue predictive performance in our validation analyses provides empirical support that the pathway-level features capture biological signals generalizable across tissue types and study cohorts. To construct robust and biologically interpretable pathway-level features, we employed the BioM2 (Biologically Informed Multi-Stage Machine Learning) framework [30]. This open-source R package integrates prior biological knowledge to stratify high-dimensional omics data. Our implementation with python coding followed a strict two-stage architecture as described in the published paper. For Stage-1, to enhance biological interpretability, CpG sites were mapped to Gene Ontology (GO) biological processes using R packages IlluminaHumanMethylation450k.db and org.Hs.eg.db annotations. We retained pathways containing between 5 and 300 genes with at least 10 mapped CpG sites to ensure biological specificity and statistical power. For each pathway, methylation data were standardized, and a predictive model was trained to capture the major coordinated methylation pattern. To generate robust pathway-level scores for Stage-2 input, we evaluated multiple candidate base learners. For each pathway, the optimal learner was selected based on performance metrics, while Pearson’s correlation coefficient was employed to identify and retain the most relevant methylation features within that pathway, thereby ensuring the highest fidelity of the biological signal. To address concerns regarding the uniformity of cell-type correction and the interaction between pathway aggregation and tissue heterogeneity, we conducted parallel analyses using two distinct input strategies: one utilizing raw beta values and another employing data regressed against key confounders, specifically age, sex, and the top ten principal components of the methylation matrix which serve as robust proxies for tissue and cell type composition. By comparing the predictive performance of these uncorrected versus adjusted inputs, we explicitly evaluated the extent to which cell type composition influences asthma label prediction within our pathway framework. Within each resampling fold of the cross-validation for both scenarios, CpGs inside a pathway were filtered based on their absolute Pearson correlation with the target phenotype, thereby ensuring that our conclusions regarding pathway significance remain robust across varying degrees of tissue heterogeneity. The Stage-1 correlation cutoffs explored include 0, 0.1, 0.3, 0.5 and 0.8. With current trainset, no pathways met the selection criteria in the current training set when the threshold was set above 0.5. In Stage-2, to minimize multicollinearity, we removed pathway scores with a pairwise correlation coefficient exceeding 0.8. Furthermore, only pathway scores showing a positive association with the asthma phenotype were included in the final modelling, consistent with the BioM2 protocol to reduce noise from inversely associated random features. The final predictive model was constructed on the Stage-2 dataset, which is a matrix with size N samples ⋅ M selected pathway scores. Model performance was evaluated using 10-repeated 5-fold cross-validation with each combination of hyper parameters. The predicted probability from each pathway-specific model was extracted as a composite epigenetic feature. This pathway-level aggregation captured coordinated methylation regulation among functionally related genes and provided a biologically meaningful dimensionality reduction for model construction. Model selection and construction To comprehensively assess the performance of different machine learning algorithms for asthma classification, we systematically benchmarked ten representative models, spanning linear models, support vector machines, ensemble methods, and neural networks. All models were implemented using scikit-learn (v1.2.2) under default hyper parameters provided by the officially released code, identical preprocessing and data partitioning schemes. For Linear models, there are Logistic Regression (Logreg), Stochastic Gradient Descent Classifier (SGD). With Support Vector Machine, Linear kernel Support Vector Classifier (SVC_linear) is chosen. Then, Ensemble methods include Random Forest (RF), Extra Trees (Etrees), Bagging Classifiers (Bagging). At last, Neural network was employed using Multilayer Perceptron Classifier (MLP) with two hidden layers (64 and 32 neurons), ReLU activation, and Adam optimizer. Each model was trained and evaluated using 5-fold stratified cross-validation with fixed random seeds to ensure reproducibility. Performance metrics included Accuracy, F1-score, Cohen’s \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\upkappa\:}$$\end{document} coefficient, and Area Under the Receiver Operating Characteristic Curve (AUC-ROC). For GO-specific modeling, each pathway was trained independently using the SVC_linear, which demonstrated the highest balance of accuracy, generalization, and interpretability. A grid search with 5-fold cross-validation was performed to optimize key hyperparameters, including penalty coefficient C, class weighting, convergence tolerance, and maximum iteration count. The optimal model was selected based on the mean cross-validated AUC. To ensure a fair comparison, all baseline models underwent hyperparameter optimization via grid search. Linear models including Logistic Regression, SGD, and Ridge classifiers were tuned across regularization strengths ranging from 0.01 to 10 or alpha values spanning three orders of magnitude. Support Vector Machines utilized linear kernels with C values of 0.1, 1, and 10. Ensemble methods such as Random Forests and Extra Trees explored estimator counts up to 300 with maximum depths varying from 5 to unlimited, while Bagging and AdaBoost classifiers tested estimator counts between 50 and 200 alongside learning rates from 0.1 to 1.0. MLP evaluated hidden layer architectures ranging from 32 to 128 units. All experiments were repeated across three distinct random seeds (42, 6, and 123), and we reported the mean performance of the optimal configuration for each algorithm (see Supplementary Table S1, S2). After training, the final optimized model was refitted on the complete training dataset, and all fitted parameters, standardization scalers, and pathway models were archived. Out-of-fold (OOF) and test-set prediction probabilities were recorded for subsequent SHAP interpretation and phenotypic correlation analysis. Model interpretability analysis To elucidate the contribution of each pathway to asthma prediction, we employed SHAP, a cooperative game theory-based approach that quantifies each feature’s marginal contribution to the model output. Using package shap.LinearExplainer, we computed SHAP values for each pathway-level methylation score in the final SVC_linear model. For each sample, SHAP values indicated the direction (positive or negative) and magnitude of each pathway’s contribution to the prediction. Global feature importance was quantified by the mean absolute SHAP value across all samples. The top 20 pathways ranked by SHAP importance were selected for biological interpretation, encompassing processes related to metabolic regulation, cellular homeostasis, and epitranscriptomic modification, including cell number maintenance, myo-inositol transport, and RNA guanine-N7-methylation, among others. Functional annotation was cross-referenced using the GO database, and results were visualized through summary plots and bar plots. Besides, we added the single sample t test to draw the significance level of each pathway-level scores with FDR correction. Clinical phenotypic correlation and association analysis To explore how model-derived features relate to biological and clinical phenotypes, we conducted correlation analyses using the GSE201872 dataset, which contains detailed clinical metadata. Missing values were imputed via KNN (K = 5) to maintain completeness. Clinical variables were categorized into three major domains: (1) Airway inflammation: bronchoalveolar lavage eosinophil (bal_eo) and neutrophil (bal_neut) proportions, and peripheral blood eosinophil absolute count (blood_eo_abs). (2) Pulmonary function: absolute and percent predicted forced expiratory volume in one second (fev1_abs, fev1_pct) and forced vital capacity (fvc_abs). (3) Inflammation and environmental exposure: fractional exhaled nitric oxide (feno), serum total IgE (ige), and current smoking status (current_smok). We then modeled associations between each phenotype and both true clinical labels (Label) and model-derived prediction probabilities (pred_prob). Continuous outcomes were analyzed via ordinary least squares (OLS) regression, and binary outcomes via logistic regression, adjusting for age, sex, and the first ten methylation PCs to mitigate confounding. Standardized Z-scores and 95% confidence intervals were computed, and significance was corrected for multiple testing using the Benjamini-Hochberg false discovery rate (FDR) method (Q < 0.05). Results were visualized as heatmaps indicating effect direction and significance levels (Q < 0.05, < 0.01, < 0.001). This analysis revealed that model predictions captured both observed disease phenotypes and latent immunometabolism variations, reflecting deeper biological processes beyond clinical observation. Mediation modeling with selection of phenotypic mediators Mediation analysis was performed to quantify causal pathways linking pathway-level methylation to respiratory phenotypes, using the semopy package in Python (v2.3.11). Based on the top 20 predictive GO pathways, the predicted probability from each pathway-specific model was defined as the exogenous variable. To identify robust phenotypic mediators linking pathway activities to asthma outcomes while avoiding multicollinearity, we employed a two-step correlation analysis strategy on all available clinical variables in the GSE201872 dataset. First, we calculated Pearson’s correlation coefficients across the entire set of candidate clinical factors to identify broad linear dependencies and exclude redundant features exhibiting high inter-correlation (absolute r over 0.7). Following this filtering, we performed a more rigorous validation on the shortlisted candidates using Spearman’s rank correlation coefficients, which are less sensitive to non-normal distributions and outliers common in clinical data. The final three mediators were selected only if they demonstrated low pairwise rank correlations with absolute coefficients below 0.5, thereby confirming that each variable captures a distinct biological dimension. For representation of top-ranked pathways, distinct from the predictive modelling stage, we utilized the first principal component (PC1) of each candidate pathway as the exposure variable, rather than the Stage-1 model’s output risk score. PC1 captures the dominant mode of intrinsic biological variation within the pathway’s methylation profile. Furthermore, to control for false positives arising from multiple testing, all p-values for direct and indirect effects in the mediation models were adjusted using the Benjamini-Hochberg False Discovery Rate procedure. Also, a predefined set of demographics, clinical, and technical covariates was included as exogenous variables regressed onto all endogenous nodes (i.e., methylation risk scores, phenotypic mediators, and the asthma outcome) in the model. These covariates comprised: age, sex, body mass index, current smoking status, self-reported race/ethnicity, recruitment source, experimental batch, ChIP batch, and maternal asthma history. To identify robust mediators linking epigenetic pathways to asthma, we implemented a systematic two-stage selection process using the GSE201872 validation cohort. First, we generated a candidate pool comprising all respiratory-related clinical phenotypes associated with the top 20 epigenetic pathways identified by SHAP value analysis in the discovery phase. Second, we constructed full mediation models for each candidate, and a phenotype was retained as a final mediator only if both the exposure-to-mediator and mediator-to-outcome path coefficients remained statistically significant after False Discovery Rate (FDR) correction using the Benjamini–Hochberg procedure. This rigorous filtering excluded alternative phenotypes such as spirometry indices and other inflammatory markers that failed to demonstrate significant associations in either path, thereby ensuring that the final model includes only variables with validated dual-linkage. To verify the independence of these mediators, we calculated pairwise Spearman correlation coefficients for all candidate phenotypes, and the analysis confirmed low inter-correlation among the selected trio with absolute correlation coefficients below 0.5, indicating distinct biological contributions involving local versus systemic inflammation and metabolic activity without significant multicollinearity. Three representative mediation models were constructed: First model incorporated paramesonephric duct development (GO:0061205), L-alanine catabolic process (GO:0042853), and immature T cell proliferation in thymus (GO:0033080) as exposures, with bronchoalveolar lavage eosinophil count (BAL-Eos) as the mediator and asthma label as the outcome. Second model employed paramesonephric duct development (GO:0061205), maintenance of cell number (GO:0098727), somatotropin secreting cell differentiation (GO:0060126), and positive regulation of cellular pH reduction (GO:0032849) as exposures, with peripheral blood eosinophil count as the mediator and asthma label as the outcome. Third model utilized paramesonephric duct development (GO:0061205), maintenance of cell number (GO:0098727), L-alanine catabolic process (GO:0042853), somatotropin secreting cell differentiation (GO:0060126), and positive regulation of cellular pH reduction (GO:0032849) as exposures, with fractional exhaled nitric oxide (FeNO) as the mediator and asthma label as the outcome. Results Study population A total of 908 participants from four independent datasets were analyzed, including one training cohort (n = 652) and three external validation cohorts (Test1, n = 72; Test2, n = 112; Test3, n = 72). The inclusion of multiple external cohorts enabled a stringent assessment of model generalizability across populations and tissues. As summarized in Table 1, the cohorts displayed distinct demographic and clinical profiles, reflecting differences in recruitment strategies. The Training and Test3 cohorts were similar in age, sex, and case-control ratio, whereas Test2 contained a higher proportion of females (68.8%) and cases (63.4%). Test1 represented a unique adolescent population (mean age ≈ 11 years), contrasting with the predominantly adult composition of the other datasets. Across all samples, the mean age was 33.63 ± 16.97 years, with 520 females (57.3%) and 388 males (42.7%). Case and control groups were balanced overall (49.6% vs. 50.4%). To mitigate technical variation, principal component analysis (PCA) of DNA methylation data was conducted, and the first three components were normalized across cohorts. The near-zero means of these principal components confirmed effective batch correction and ensured robust cross-cohort integration for downstream analyses. Model benchmarking and classifier selection We systematically compared the predictive performance of multiple machine learning algorithms using cross-validation within the training cohort. As detailed in Supplementary Table S2, SVC_linear achieved the highest overall performance, with accuracy = 0.767, F1-score = 0.74, and Cohen’s κ = 0.53. It also attained the highest AUC-ROC (0.843), significantly outperforming other classifiers. Logreg and SGD achieved stable yet slightly lower accuracies (0.760–0.765). Among ensemble methods, Bagging with grid search-based hyperparameters achieved the best overall balance (accuracy = 0.706, AUC = 0.75), whereas RF and Etrees exhibited overfitting, with inflated training but reduced validation performance. MLP reached comparable accuracy (0.733, AUC = 0.804) but required substantially longer computation time. Considering predictive performance, interpretability, and computational efficiency, SVC_linear was selected as the benchmark model for all downstream analyses. Its linear kernel structure allows direct interpretation of feature coefficients, aligning with the study’s goal of uncovering mechanistic pathways underlying asthma risk. Cross-tissue predictive performance The benchmark model demonstrated robust and generalizable performance across both internal and external validation cohorts. In the airway epithelial-based validation sets (Test 1 and Test 2), the model achieved AUC values of 0.79 and 0.98, respectively. The model showed high reproducibility within airway tissue. In the peripheral blood-based cohort (Test3), the AUC remained 0.74, confirming strong cross-tissue transferability. These findings suggest that the derived methylation features capture core biological processes relevant to asthma pathophysiology, rather than tissue-specific epigenetic noise. These results establish the model’s stability and generalization capacity across independent populations and biological compartments. Corresponding ROC curves from independent test predictions are shown in Fig. 1. Figure corresponding AUC results are summarized in Supplementary Table S3 and detailed experiments records are listed in Supplementary Table S4. Fig. 1ROC curves for the final model of BioM2 framework with SVC_linear as the base learner. Receiver operating characteristic (ROC) curves show the predictive performance of the final model using SVC_linear as the base learner, with raw CpG data as input. Gene range parameter is from 20 to 300. Correlation coefficient for Stage-1 selection is 0.1. Shaded area around each ROC curve represents the corresponding 95% confidence interval (CI) across independently repeated experiments for both cross validation and independent tests. The blue curve represents the mean cross-validation (CV) performance in the training cohort (AUC = 0.86, 95% CI = [0.820, 0.896]). The green, orange, and red curves correspond to the three independent test cohorts, with AUC values of 0.79 (95% CI = [0.782, 0.799]), 0.98 (95% CI = [0.974, 0.990]), and 0.74 (95% CI = [0.690, 0.771]), respectively. The dashed diagonal line indicates random prediction (AUC = 0.5) ROC curves for the final model of BioM2 framework with SVC_linear as the base learner. Receiver operating characteristic (ROC) curves show the predictive performance of the final model using SVC_linear as the base learner, with raw CpG data as input. Gene range parameter is from 20 to 300. Correlation coefficient for Stage-1 selection is 0.1. Shaded area around each ROC curve represents the corresponding 95% confidence interval (CI) across independently repeated experiments for both cross validation and independent tests. The blue curve represents the mean cross-validation (CV) performance in the training cohort (AUC = 0.86, 95% CI = [0.820, 0.896]). The green, orange, and red curves correspond to the three independent test cohorts, with AUC values of 0.79 (95% CI = [0.782, 0.799]), 0.98 (95% CI = [0.974, 0.990]), and 0.74 (95% CI = [0.690, 0.771]), respectively. The dashed diagonal line indicates random prediction (AUC = 0.5) Pathway-level methylation signatures associated with asthma status To identify pathway-level features associated with asthma methylation signatures, SHAP analysis was performed. The valine biosynthetic process (GO:0009099, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{adj}$$\end{document} = 0.152) exhibited the highest mean SHAP value (0.044), reflecting its highest predictive importance to model prediction. Other top-ranked pathways included paramesonephric duct development (GO:0061205, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{adj}$$\end{document} = 0.016), myo-inositol transport (GO:0015798, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{adj}$$\end{document} = 0.055), behavioral response to formalin-induced pain (GO:0061368, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{adj}$$\end{document} = 0.000), and atrial cardiac muscle cell development (GO:0055014, \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{p}_{adj}$$\end{document} = 0.061). Together, these findings suggest that amino acid metabolism, developmental regulation, and cellular signaling processes are prominently featured in the epigenetic landscape associated with asthma, as shown in Fig. 2. Corresponding statistics results were detailed in Supplementary Table S7. Fig. 2SHAP-based feature importance analysis. a Per-sample SHAP value distributions for the top 10 features. Points are colored by feature value, with a color gradient from blue (low values) to orange (high values). SHAP values range from − 0.20 to 0.15, reflecting the magnitude and directional effect of each feature on model predictions across individual samples. b The mean |SHAP| ranking of those features, showing each GO’s average contribution to model predictions. Colors gradient from blue to orange, corresponding to the range of −log10(Qvalues, where deeper orange indicates higher −log10(Q) (stronger statistical significance) and lighter blue indicates lower −log10(Q) (weaker significance). Statistical significance is indicated as Q < 0.05 (*), Q < 0.01 (**), and Q < 0.001 (***) SHAP-based feature importance analysis. a Per-sample SHAP value distributions for the top 10 features. Points are colored by feature value, with a color gradient from blue (low values) to orange (high values). SHAP values range from − 0.20 to 0.15, reflecting the magnitude and directional effect of each feature on model predictions across individual samples. b The mean |SHAP| ranking of those features, showing each GO’s average contribution to model predictions. Colors gradient from blue to orange, corresponding to the range of −log10(Qvalues, where deeper orange indicates higher −log10(Q) (stronger statistical significance) and lighter blue indicates lower −log10(Q) (weaker significance). Statistical significance is indicated as Q < 0.05 (*), Q < 0.01 (**), and Q < 0.001 (***) From a biological perspective, the prominence of valine biosynthesis might reflect a potential association between branched-chain amino acid metabolism and airway inflammation. The enrichment of paramesonephric duct and cardiac muscle developmental pathways is consistent with shared regulatory processes involved in epithelial remodelling and tissue differentiation. The involvement of myo-inositol transport may indicate metabolic-signaling interactions relevant to immune regulation, while the behavioral response to pain pathway could reflect neuroimmune patterns associated with airway hypersensitivity. Collectively, these results generate hypotheses regarding how metabolic and developmental pathways may relate to asthma-related methylation dynamics. Clinical correlates and biological relevance To evaluate the clinical and statistical associations of the model, we examined whether its prediction scores showed significant relationships with disease-associated gradients. Association analyses in the GSE201872 cohort demonstrated statistically significant correlations between the model’s predicted probabilities and diverse clinical phenotypes. Interestingly, while the true asthma labels were more strongly associated with airway cytology and lung function, the model predictions showed closer correlations with metabolic and methylation-derived features, such as the L-alanine catabolic process (GO:0042853) and tRNA stabilization (GO:0036416). These findings suggest that the model reflects molecular patterns that may extend beyond overt clinical manifestations. Overall, the model output show associations with both phenotypic severity and molecular heterogeneity, linking clinical features with molecular signatures. Group differences in BAL_EO levels across asthma status are visualized in Fig. 3 using violin and box plots. Association analysis results between phenotypes (e.g., BAL_EO, FeNO, IgE) and GO pathway scores are presented in Fig. 4. Fig. 3Violin and box plots of BAL_EO levels by asthma status. Violin plots with embedded box plots and scatter points show the distribution of BAL_EO levels in the control group (blue, N = 41) and asthma group (brown, N = 71). The violin shape represents data density. Box plots show the median, interquartile range, and whiskers. Individual samples are shown as scatter points. Statistical significance is marked with asterisks as Q ≤ 0.05 (*), Q ≤ 0.01 (**), and Q < 0.001 (***) Violin and box plots of BAL_EO levels by asthma status. Violin plots with embedded box plots and scatter points show the distribution of BAL_EO levels in the control group (blue, N = 41) and asthma group (brown, N = 71). The violin shape represents data density. Box plots show the median, interquartile range, and whiskers. Individual samples are shown as scatter points. Statistical significance is marked with asterisks as Q ≤ 0.05 (*), Q ≤ 0.01 (**), and Q < 0.001 (***) Fig. 4Associations between top GO pathways, model prediction probability, asthma status, and clinical phenotypes. Heatmap showing standardized regression effect sizes between the top ten GO pathways (ranked by mean absolute SHAP values), model prediction probability, asthma status, and key clinical phenotypes in the GSE201872 cohort. Corresponding SHAP values indicate predictor values are positively associated with phenotype levels. Statistical significance is indicated as Q ≤ 0.05 (*), Q ≤ 0.01 (**), and Q < 0.001 (***). GO:0061205 (paramesonephric duct development) and GO:0042853 (L-alanine catabolic process) were significantly correlated with bronchoalveolar eosinophil counts (BAL_EO, Q < 0.05) and peripheral blood eosinophil levels (Blood_EO, Q < 0.05). GO:0036416 (tRNA stabilization) was strongly associated with fractional exhaled nitric oxide (FENO, Q < 0.001), supporting the presence of this strong association captured by the model Associations between top GO pathways, model prediction probability, asthma status, and clinical phenotypes. Heatmap showing standardized regression effect sizes between the top ten GO pathways (ranked by mean absolute SHAP values), model prediction probability, asthma status, and key clinical phenotypes in the GSE201872 cohort. Corresponding SHAP values indicate predictor values are positively associated with phenotype levels. Statistical significance is indicated as Q ≤ 0.05 (*), Q ≤ 0.01 (**), and Q < 0.001 (***). GO:0061205 (paramesonephric duct development) and GO:0042853 (L-alanine catabolic process) were significantly correlated with bronchoalveolar eosinophil counts (BAL_EO, Q < 0.05) and peripheral blood eosinophil levels (Blood_EO, Q < 0.05). GO:0036416 (tRNA stabilization) was strongly associated with fractional exhaled nitric oxide (FENO, Q < 0.001), supporting the presence of this strong association captured by the model Pathway-level mediation modelling of clinical phenotypes Mediation modeling was performed to quantify how pathway-level methylation exerts direct and indirect influences on respiratory phenotypes. Based on the top GO pathways identified by the predictive models, the predicted probability from each pathway-specific methylation model was extracted to represent its dominant methylation regulation feature. Three representative mediation models were constructed to evaluate the causal architecture linking molecular, functional, and clinical layers. Only paths yielding an FDR-adjusted p-value less than 0.05 were deemed statistically significant. Also, with Spearman’s correlation coefficients < 0.5, final mediators included in the structural diagrams presented in Fig. 5. Complete results for correlation analyses were provided in Supplementary Fig. S1 and Supplementary Table S8. Fig. 5Mediation analysis reveals pathway-driven causal mechanisms linking immune activation to asthma severity. Mediation models were applied to evaluate potential associations between pathway model prediction probabilities (derived from top GO pathways) and real asthma status. Representative models show that pathway-level signals, including GO:0061205 (paramesonephric duct development), GO:0042853 (L-alanine catabolic process), and GO:0033080 (regulation of neutrophil homeostasis), relate to asthma severity via immune mediators such as bronchoalveolar eosinophil counts (BAL_EO), peripheral eosinophil abundance (Blood_EO_abs), and exhaled nitric oxide (FENO). Coefficient above each arrow is the corresponding standardized effect estimates. Positive effects are colored in red. Arrow widths are proportional to standardized mediation coefficients, highlighting both direct and indirect regulatory pathways between immune activation and clinical asthma severity Mediation analysis reveals pathway-driven causal mechanisms linking immune activation to asthma severity. Mediation models were applied to evaluate potential associations between pathway model prediction probabilities (derived from top GO pathways) and real asthma status. Representative models show that pathway-level signals, including GO:0061205 (paramesonephric duct development), GO:0042853 (L-alanine catabolic process), and GO:0033080 (regulation of neutrophil homeostasis), relate to asthma severity via immune mediators such as bronchoalveolar eosinophil counts (BAL_EO), peripheral eosinophil abundance (Blood_EO_abs), and exhaled nitric oxide (FENO). Coefficient above each arrow is the corresponding standardized effect estimates. Positive effects are colored in red. Arrow widths are proportional to standardized mediation coefficients, highlighting both direct and indirect regulatory pathways between immune activation and clinical asthma severity In the first model, the predicted probabilities of pathways GO_0061205 (adaptive immune response), GO_0042853 (neutrophil degranulation), and GO_0033080 (eosinophil activation) served as mediators linking bronchoalveolar eosinophil count (bal_eo) to asthma diagnosis (Label), showing an excellent model fit (CFI = 1.0147, RMSEA = 0, AIC = 18.00, BIC = 42.47). The second model incorporated GO_0061205, GO_0098727 (maintenance of cell number), GO_0060126 (epithelial cell proliferation), and GO_0032849 (positive regulation of leukocyte proliferation) as mediators between blood eosinophil count (blood_eo_abs) and asthma, also achieving strong fit indices (CFI = 1.0217, RMSEA = 0, AIC = 22.00, BIC = 51.90). The third model considered the predicted probabilities of GO_0061205, GO_0098727, GO_0042853, GO_0060126, and GO_0032849 as mediators linking fractional exhaled nitric oxide (FeNO) to asthma, showing the best overall fit (CFI = 1.0222, RMSEA = 0, AIC = 26.00, BIC = 61.34). Collectively, these mediation analyses indicate that pathway-specific predicted probabilities exert both direct and indirect effects on asthma outcomes via inflammatory cell activity, providing quantitative evidence for epigenetic regulation of respiratory phenotypes. Importantly, only GO:0060126 and GO:0098727 demonstrated significant direct effects on asthma, highlighting pathway-specific molecular contributions beyond mediated immune or functional influences. The identified pathways, particularly those involving epithelial proliferation (GO:0060126) and cell number maintenance (GO:0098727), are hallmarks of airway remodeling. Alongside eosinophil activation (GO:0033080), these mechanisms drive the inflammatory and structural changes that ultimately manifest as impaired lung function and asthma diagnosis (see Fig. 5). Modelling results were presented in Supplementary Table S5 and detailed in S9. Discussion This study presents a cross-tissue, multi-cohort, and interpretable epigenetic prediction model for asthma. By integrating pathway-level methylation components, interpretable machine learning, and mediation analysis, we provide quantitative observational evidence that methylation states within specific biological pathways influence respiratory and immune phenotypes. The findings collectively suggest an interconnected pattern of metabolic regulation, cellular homeostasis, and epi-transcriptomic modification that associates pathway-specific molecular alterations to clinical phenotypes. This integrative framework provides mechanistic insight into how epigenetic, metabolic, and immune processes converge to influence asthma pathogenesis. This integrative framework advances the understanding of epigenetic-immune-metabolic interactions associated with asthma pathogenesis. Most previous studies on asthma-related DNA methylation have focused on individual CpG sites or single-tissue analyses [10, 31, 32]. These efforts identified candidate loci such as IL4R, ORMDL3, and GSDMB [33–35], but they were limited in interpretability and cross-cohort reproducibility. The present study differs by modelling pathway-level methylation components that capture coherent biological processes rather than isolated epigenetic markers. By aggregating functionally related CpGs into pathway principal components, the analysis enhances robustness against noise and improves biological interpretability. Compared with genome-based prediction models, such as those built on polygenic or transcriptome-wide scores [36, 37], the pathway-level methylation framework developed here primarily reflects epigenetic patterns associated with environmental exposures and disease-related biological processes, and may indirectly reflect genetic regulation of epigenetic states, providing a more dynamic reflection of immune activation states [38]. Unlike most prior studies confined to one tissue type or via standard machine learning algorithm, our model demonstrated strong cross-tissue generalization, performing consistently across airway epithelial and peripheral blood samples [28, 39, 40]. This feature is essential for translational application, as it implies that the methylation signatures reflect systemic biological processes rather than localized changes. Importantly, the combination of explainable artificial intelligence and statistical modelling provides a methodological advance beyond prior black-box approaches. The integration of SHAP value with mediation analysis creates a reproducible bridge between statistical prediction and biological patterns, thereby redefining how epigenetic data can be used to generate hypotheses about regulatory pathways. Our results indicate that asthma susceptibility is closely reflected in coordinated pathway-level methylation rather than isolated CpG variation. Consistent with prior work showing that metabolic imbalance, epithelial dysregulation, and immune activation contribute to asthma pathogenesis [41, 42]. Our findings extend these observations. Aberrant methylation was identified within the top predictive pathways. These include pathways involved in metabolic regulation, cellular homeostasis, tissue remodelling, and immune signalling. These pathways show associations with immune balance and epithelial integrity markers characteristic of asthma. This provides a potential epigenetic mechanism explaining why these molecular dysregulations correlate with chronic airway inflammation and remodelling. The enrichment of pathways related to cell number maintenance suggests that the balance between apoptosis and proliferation may be important for preserving epithelial barrier integrity. Disruption of this balance may result in impaired repair and prolonged inflammation, which are hallmarks of airway remodelling. In addition, the identification of myo-inositol transport pathways suggests a possible contribution of metabolic signalling to immune activation and epithelial stability. Emerging evidence supports the role of inositol metabolism in regulating inflammatory responses and oxidative stress in airway diseases [43–45]. The detection of RNA guanine-N7-methylation further implies that epi-transcriptomic regulation may modulate stress adaptation and inflammatory gene expression, an area of growing interest in respiratory epigenetics [46–48]. Taken together, these observations support a systems-level model in which metabolic regulation, cellular homeostasis and RNA modification interact to set airway immune tone and may influence tissue remodelling. The results suggest that asthma pathogenesis cannot be fully understood through single-pathway or single-cell perspectives, but rather through integrated, multi-layered networks that couple metabolic and epigenetic regulation. The present study introduces several methodological advances that strengthen both the reliability and interpretability of epigenetic prediction. The use of pathway-level methylation principal components transforms complex CpG data into biologically meaningful variables, providing a structured view of molecular regulation. Multi-cohort integration with rigorous batch correction ensures that the findings are not artifacts of cohort-specific variation. The incorporation of SHAP-based model interpretation allows transparent quantification of each feature’s contribution to prediction, while mediation analysis quantifies the direct and indirect effects of methylation on lung function and immune cell indicators. These combined strategies move beyond association analysis toward a more interpretable understanding of disease mechanisms [49, 50]. Collectively, these analytical innovations establish a reproducible framework for translating high-dimensional epigenetic data into testable biological hypotheses. Such integration of statistical learning and associative inference represents an important step toward achieving greater mechanistic insight in molecular epidemiology though experimental validation remains necessary. At the broader disciplinary level, this work contributes to a conceptual shift from site-specific association studies to pathway-oriented, system-level modelling. The proposed framework quantitatively assesses coordinated methylation within biological pathways. It reveals associations with disease phenotypes. This paradigm expands the role of epigenetics from descriptive biomarker discovery to mechanistic understanding and clinical application [51–53]. The study shows that explainable artificial intelligence can be integrated into biomedical research. This enhances interpretability while maintaining rigor. In the context of respiratory and immune disorders, the approach provides a scalable foundation for molecular endotyping, individualized risk stratification, and therapeutic optimization. Over time, the incorporation of pathway-level methylation features into precision medicine pipelines could facilitate more effective monitoring of treatment response and disease progression [54, 55]. Several limitations should be acknowledged. The cross-sectional design of the datasets limits the ability to infer temporal or causal directionality. Longitudinal studies are needed. These should assess whether the observed methylation patterns precede disease onset or reflect downstream consequences. Experimental validation, including perturbation of candidate pathways in cellular or organoid models, would further clarify causal mechanisms. Although the current analysis includes both airway and peripheral blood tissues, other relevant compartments such as lymphoid or bone marrow tissues were not assessed [56–58]. Future studies could integrate additional omics layers. These include transcriptomic, proteomic, and metabolomic data. Such integration would strengthen the mechanistic framework. Furthermore, direct measurements of amino acid, polyol, and inositol metabolites would provide empirical support for the proposed metabolic-epigenetic links. Future research should explore the development of longitudinal biomarkers, such as a pathway-level epigenetic clock that tracks airway inflammation and therapeutic responsiveness. Extending this framework to single-cell or spatial epigenomic data may uncover cell-type-specific regulation that drives the observed systemic signals [59–61]. Conclusion In summary, this study provides an interpretable and quantitatively validated model that shows statistical associations among DNA methylation, immune regulation, and lung function in asthma. By combining multi-cohort epigenetic data with interpretable machine learning and mediation analysis, the research provides a hypothesis-generating framework that complements correlative observations with prioritized candidates for mechanistic investigation. The integration of pathway-level methylation with functional phenotypes establishes a new framework for precision respiratory medicine, offering both conceptual and methodological contributions that will guide future efforts in multi-omics integration and disease mechanism discovery. Electronic Supplementary Material Below is the link to the electronic supplementary material. Supplementary Material 1. Supplementary Material 1. Supplementary Material 2. Supplementary Material 2. Supplementary Material 3. Supplementary Material 3. Supplementary Material 4. Supplementary Material 4. Supplementary Material 5. Supplementary Material 5. Supplementary Material 6. Supplementary Material 6.