A computational framework that aggregates DNA methylation into biologically interpretable pathway scores has successfully differentiated individuals with asthma from healthy controls across both respiratory epithelium and peripheral blood, according to findings published in Epigenetics & Chromatin. Rather than examining isolated cytosine-phosphate-guanine dinucleotide sites, the researchers constructed a linear support vector classifier centered on coordinated biological processes mapped through the Gene Ontology database. Across independent multi-cohort evaluations encompassing 908 human participants, the pathway-level model attained an area under the receiver operating characteristic curve of 0.792 and 0.980 in two independent airway epithelial cohorts, while maintaining an area under the curve of 0.736 in peripheral blood samples. These reproducible findings indicate that systemic epigenetic signals mirror asthma pathophysiology across distinct anatomical compartments, capturing both overt clinical manifestations and latent variations in cellular metabolism.
Asthma is a clinically heterogeneous and chronic inflammatory disorder of the human airways that affects more than 260 million individuals worldwide, contributing substantially to global healthcare burdens. While previous genome-wide association studies identified specific risk loci including variants within IL33, IL1RL1, and ORMDL3 linked to epithelial barrier integrity and T-helper 2 inflammation, these genetic markers explain only a modest fraction of disease heritability and cannot capture dynamic environmental influences such as airborne pollutants and allergen exposures. Subsequent epigenome-wide association investigations demonstrated differential DNA methylation within inflammatory pathways and cytokine genes, highlighting epigenetic reprogramming as a regulatory interface bridging inherited susceptibility and environmental exposures. However, traditional epigenetic investigations have largely been constrained by single-cohort designs, reliance on single tissue types, and isolated probe-level evaluations that lack functional interpretability, while conventional machine learning classifiers frequently optimize numeric accuracy at the expense of biological transparency.
To overcome these methodological challenges, the investigators integrated six publicly accessible DNA methylation datasets from the Gene Expression Omnibus repository, specifically GSE172365, GSE250513, GSE85566, GSE65163, GSE201872, and GSE104471. All included patient data were originally collected under the ethical principles of the Declaration of Helsinki with appropriate institutional consents, allowing secondary analytical evaluation without additional institutional review board approvals. Following genomic coordinate alignment, probes with missingness rates exceeding 20 percent were filtered out, and remaining missing values were imputed using a K-nearest neighbor algorithm with five neighbors, yielding a harmonized matrix of 340,047 cytosine-phosphate-guanine sites. The final study population comprised 908 individuals divided into an airway epithelial training cohort of 652 individuals and three independent external validation cohorts, including an adolescent airway cohort of 72 subjects designated Test 1, an adult airway cohort of 112 subjects designated Test 2, and a peripheral blood cohort of 72 subjects designated Test 3.
Across the complete study population of 908 individuals, the mean age was 33.63 years with a standard deviation of 16.97 years, comprising 520 females representing 57.3 percent and 388 males representing 42.7 percent of the participants. Overall case and control distributions remained balanced at 49.6 percent asthmatic patients and 50.4 percent non-asthmatic controls, although demographic profiles varied across individual cohorts to enable rigorous testing of model generalizability. The training cohort and Test 3 exhibited balanced demographic distributions, whereas Test 2 comprised an older group with 68.8 percent females and 63.4 percent asthma cases, and Test 1 comprised an adolescent cohort with an average age of approximately 11 years. To mitigate technical variations across these heterogeneous recruitment sources without introducing synthetic distortions, principal component analysis was conducted on the methylation profiles, and normalization of the first three components achieved near-zero means that confirmed effective batch harmony across all cohorts.
The feature engineering pipeline employed the Biologically Informed Multi-Stage Machine Learning framework, known as BioM2, implemented through custom Python code following a two-stage analytical architecture. Raw methylation beta values were adjusted for age and sex, functionally normalized via the minfi package in R, converted into M-values, and standardized using z-scores where parameters were strictly calculated within the training fold to prevent information leakage into testing partitions. In Stage 1, cytosine-phosphate-guanine sites were annotated to Gene Ontology biological processes using IlluminaHumanMethylation450k and human organism databases, retaining candidate pathways containing between 5 and 300 genes with at least 10 mapped probes. Candidate features were tested across multiple absolute Pearson correlation thresholds against phenotypic asthma status, demonstrating that higher correlation cutoffs above 0.5 eliminated all viable pathways in the training data, while an optimal correlation threshold of 0.1 preserved robust feature representations across both unadjusted beta values and principal-component-adjusted inputs.
In Stage 2 of the framework, multicollinearity was controlled by filtering out any pathway scores exhibiting a pairwise correlation coefficient greater than 0.8, and only features displaying positive associations with asthma diagnosis were retained. The researchers systematically benchmarked ten supervised classification algorithms using scikit-learn version 1.2.2 under stratified five-fold cross-validation with three fixed random seeds to ensure reproducibility. Tested models included logistic regression, stochastic gradient descent, linear support vector classification, random forests, extra trees, bagging classifiers, and a multilayer perceptron neural network containing two hidden layers of 64 and 32 neurons with rectified linear unit activation. The linear support vector classifier achieved the most balanced performance across all evaluated metrics, recording an accuracy of 0.767, an F1-score of 0.74, a Cohen’s kappa coefficient of 0.53, and a cross-validated receiver operating characteristic area under the curve of 0.843, significantly outperforming ensemble algorithms that suffered from validation overfitting.
When applied to external validation datasets, the optimized linear support vector classifier confirmed high reproducibility and cross-tissue robustness. In the independent airway epithelial validation cohorts, the classifier achieved an area under the receiver operating characteristic curve of 0.792 with a 95 percent confidence interval from 0.782 to 0.799 in Test 1, and an area under the curve of 0.980 with a 95 percent confidence interval from 0.974 to 0.990 in Test 2. In the peripheral blood cohort of Test 3, the model achieved an area under the curve of 0.736 with a 95 percent confidence interval from 0.690 to 0.771. Because the supervised framework learned systemic epigenetic patterns directly from pathway-aggregated methylation values without applying explicit batch-correction algorithms like ComBat, this sustained performance across distinct biological compartments demonstrates that pathway-level features preserve core disease mechanisms while filtering out tissue-specific and technical noise.
To interpret the contributions of specific pathways to the predictive output, the authors implemented SHapley Additive exPlanations through the LinearExplainer algorithm, computing individual feature importance across all samples alongside one-sample t-tests against zero with Benjamini-Hochberg false discovery rate corrections. Among the top 20 ranked features, the valine biosynthetic process (GO:0009099) exhibited the highest mean absolute SHAP value of 0.044 with an adjusted p-value of 0.152, suggesting potential connections between branched-chain amino acid metabolism and airway inflammatory dynamics. Other top-ranked contributors exhibiting significant associations included paramesonephric duct development (GO:0061205) with an adjusted p-value of 0.016, behavioral response to formalin-induced pain (GO:0061368) with an adjusted p-value of 0.000, myo-inositol transport (GO:0015798) with an adjusted p-value of 0.055, and atrial cardiac muscle cell development (GO:0055014) with an adjusted p-value of 0.061, collectively underscoring developmental, metabolic, and neuroimmune involvement in asthma.
Correlation analyses performed within the deeply characterized GSE201872 cohort further illustrated the distinct clinical relevance of model-derived probabilities across three primary clinical domains: airway inflammation, pulmonary spirometry, and systemic inflammatory indicators. While true clinical diagnosis labels correlated most strongly with airway cytology and lung function measurements such as forced expiratory volume in one second and forced vital capacity, model-predicted probabilities aligned more closely with metabolic and epitranscriptomic pathways. Specifically, paramesonephric duct development and the L-alanine catabolic process (GO:0042853) demonstrated statistically significant correlations with bronchoalveolar lavage eosinophil proportions and absolute peripheral blood eosinophil counts under false discovery rate thresholds below 0.05. Furthermore, tRNA stabilization (GO:0036416) exhibited a highly significant correlation with fractional exhaled nitric oxide at an adjusted significance level below 0.001, indicating that the epigenetic model captures latent molecular heterogeneity that extends beyond routine physiological spirometry.
To examine whether pathway-level epigenetic alterations exert direct or indirect influences on clinical asthma manifestations, the authors performed structural equation mediation modeling using the Python semopy package. Phenotypic mediators were selected using a two-stage strategy requiring pairwise Spearman rank correlation coefficients below 0.5 to prevent multicollinearity, alongside dual-path significance after Benjamini-Hochberg adjustment, which retained bronchoalveolar lavage eosinophils, peripheral blood eosinophils, and fractional exhaled nitric oxide while excluding spirometry indices. To capture intrinsic biological variance within the mediation models, candidate pathways were represented by their first principal components rather than stage-one classification scores, with full adjustment for confounding covariates including age, sex, body mass index, smoking status, self-reported race, recruitment source, experimental batch, microarray chip batch, and maternal asthma history.
All three structural mediation models demonstrated excellent statistical fit indices, with root mean square errors of approximation of zero and comparative fit indices exceeding 1.01. The first model, incorporating paramesonephric duct development, L-alanine catabolism, and immature T cell proliferation in thymus (GO:0033080) mediated by bronchoalveolar eosinophils, achieved an Akaike information criterion of 18.00 and a Bayesian information criterion of 42.47. The second model linked blood eosinophil abundance to four developmental and homeostatic pathways, achieving an Akaike information criterion of 22.00 and a Bayesian information criterion of 51.90. The third model evaluated fractional exhaled nitric oxide across five pathways, yielding an Akaike information criterion of 26.00 and a Bayesian information criterion of 61.34. Crucially, while most pathways operated through inflammatory mediators, maintenance of cell number (GO:0098727) and epithelial cell proliferation (GO:0060126) maintained significant direct effects on asthma status independent of intermediate inflammatory cell counts.
The study authors highlighted several methodological limitations, noting that the cross-sectional design of existing public cohorts precludes definitive inferences regarding temporal sequence or causality, leaving open the question of whether observed methylation signatures precede disease onset or reflect secondary chronic inflammation. Furthermore, while the models successfully generalized across airway epithelium and peripheral blood, other pertinent tissues such as bone marrow and lymphatic organs were not examined, and direct biochemical quantification of amino acid and inositol metabolites was unavailable to experimentally confirm the proposed metabolic linkages. The researchers concluded that prospective longitudinal cohorts, functional perturbations in cellular or airway organoid models, single-cell spatial epigenomic profiling, and the integration of transcriptomic and metabolomic data will be essential to establish formal causal mechanisms and translate pathway-level epigenetic modeling into precision respiratory medicine.
Subject of Research: Pathway-level epigenetic modeling of DNA methylation in cross-tissue asthma risk
Article Title: Pathway-level epigenetic modeling illuminates the methylation architecture to asthma risk across tissues
Article References: Huang, Z., Liu, Y., Cai, F., Zhu, J., Wang, S., Li, P., Zhang, S., Song, Z., Liu, W., Liu, H., & Chen, J. (2026). Pathway-level epigenetic modeling illuminates the methylation architecture to asthma risk across tissues. Epigenetics & Chromatin, 19(1), Article 37. https://doi.org/10.1186/s13072-026-00685-y
Image Credits: AI Generated
DOI: 10.1186/s13072-026-00685-y
Keywords: DNA methylation, epigenetics, asthma, machine learning, support vector classifier, pathway modeling, Gene Ontology, SHAP analysis, cross-tissue genomics, mediation analysis, Pathway-level, epigenetic
News Source: Juliet Wilcox. (November 7, 2026). Pathway Epigenetic Model Uncovers Cross-Tissue DNA Methylation Patterns Linked to Asthma. Scienmag.



