This exploratory study prioritised immune-related methylation features (IL6, CXCR4, CDK11, CREBBP) across paediatric blood and infant buccal cells within an interpretable neural-network framework for age-related methylation and outcome-oriented molecular investigation, with Mendelian randomisation suggesting genetically supported associations of IL6 methylation with preterm birth and lower mean birth weight and of CXCR4 methylation with preterm birth, without establishing that methylation itself causes the outcomes.
Key Findings
Results
IL6, CXCR4, CDK11 and CREBBP showed the strongest coherent methylation signals and stable predictive contributions, whereas BCLAF1 contributed less consistently.
Identified through empirical-Bayes differential methylation analysis, feature attribution and gene-ablation analysis.
The primary modelling cohort was paediatric peripheral-blood leukocytes (GSE36054; 192 samples retained after quality control).
BCLAF1 was described as contributing less consistently than the other four genes.
Results
The prioritised loci mapped to immune and cell-regulatory pathways.
Pathways included cytokine-receptor interaction, chemokine signalling, G2/M checkpoint regulation, histone acetylation, and apoptosis-related pathways.
Pathway enrichment was performed as part of the integrated analysis pipeline.
Results
The biologically informed neural network predicted age with a cross-validated mean absolute error of 1.87 years and R2 of 0.781.
Performance was assessed using repeated 10-fold cross-validation.
The model was trained on the paediatric peripheral-blood dataset (192 samples after quality control).
Results
Removing the polybrominated-biphenyl (PBB)-associated feature filter yielded comparable age-prediction performance.
Without the filter, MAE was 1.91 years and R2 was 0.772.
With the filter, MAE was 1.87 years and R2 was 0.781.
The PBB exposure cohort comprised peripheral blood from 659 individuals.
Results
Prioritised features showed cross-tissue consistency in infant buccal cells.
The infant buccal cell dataset (n = 542) served as an independent early-life cross-tissue reference.
A neuronal nuclei dataset (n = 58) was also integrated in the study.
Results
Mendelian randomisation identified genetically supported associations of IL6 methylation with preterm birth and lower mean birth weight, and of CXCR4 methylation with preterm birth, after multiplicity correction.
Two-sample Mendelian randomisation was used following disease mapping.
Associations survived multiplicity correction.
The authors state these associations do not establish that methylation itself causes the outcomes.
Methods
The study integrated four Illumina 450K and EPIC methylation datasets using a multi-step exploratory pipeline.
Processing included cohort-specific QC and normalisation, batch adjustment, empirical-Bayes differential methylation, pathway enrichment, repeated 10-fold CV, feature attribution, cross-tissue assessment, gene ablation, disease mapping and MR.
The study is described as exploratory.
What This Means
This study looked at chemical tags on DNA, called methylation, that change as children grow and that may reflect immune system activity. Using data from blood samples of children, cheek swabs from infants, brain cell nuclei, and blood from people exposed to an industrial chemical (polybrominated biphenyl), the researchers built a neural network that was designed to reflect known biology. It highlighted several immune-related genes, mainly IL6, CXCR4, CDK11 and CREBBP, whose methylation patterns were most informative.
The model estimated children's ages with an average error of about 1.9 years, and performed about as well when a filter related to chemical exposure was removed. The key genes also showed similar patterns in infant cheek cells, suggesting the signals are not limited to blood. A genetics-based approach (Mendelian randomisation) suggested links between IL6 methylation and preterm birth and lower birth weight, and between CXCR4 methylation and preterm birth.
The authors stress that this is an exploratory study, and the genetic associations do not prove that methylation itself causes these outcomes. This research suggests these immune-related methylation features may be useful candidates for further investigation of ageing and early-life health, but confirmation in other studies would be needed.
Cheng X, Wu C, Yang X, Liu R. (2026). Immune-related DNA methylation signatures, age prediction, and cross-tissue assessment via biologically informed neural networks.. Human molecular genetics. https://doi.org/10.1093/hmg/ddag110