Epigenetic age predictions based on buccal swabs are more precise in combination with cell type-specific DNA methylation signatures

Abstract

Aging is reflected by highly reproducible DNA methylation (DNAm) changes that open new perspectives for estimation of chronological age in legal medicine. DNA can be harvested non-invasively from cells at the inside of a person's cheek using buccal swabs - but these specimens resemble heterogeneous mixtures of buccal epithelial cells and leukocytes with different epigenetic makeup. In this study, we have trained an age predictor based on three age-associated CpG sites (associated with the genesPDE4C, ASPA, and ITGA2B) for swab samples to reach a mean absolute deviation (MAD) between predicted and chronological age of 4.3 years in a training set and of 7.03 years in a validation set. Subsequently, the composition of buccal epithelial cells versus leukocytes was estimated by two additional CpGs (associated with the genes CD6 and SERPINB5). Results of this "Buccal-Cell-Signature" correlated with cell counts in cytological stains (R2 = 0.94). Combination of cell type-specific and age-associated CpGs into one multivariate model enabled age predictions with MADs of 5.09 years and 5.12 years in two independent validation sets. Our results demonstrate that the cellular composition in buccal swab samples can be determined by DNAm at two cell type-specific CpGs to improve epigenetic age predictions.

Keywords: aging; cell composition; epigenetic; epithelial cells; methylation; predictor; swab.

Figure 1. Epigenetic aging model for blood needs to be adjusted for buccal swabs

(A) Illustration of sample collection with a buccal swab. (B) Epigenetic age predictions of 55 mouth swab samples using an age predictor that was trained on blood samples as described before [4]. (C) For comparison, we demonstrate the predictions for 151 whole blood samples of our previous work [4]. (D) The multivariate model for age predictions was then retrained on pyrosequencing results of 55 mouth swab samples and validated on 55 independent additional samples that were analyzed in a different lab. (E-G) Correlation of β-values of age-associated CpG sites with chronological age. To this end, we used publically available datasets of blood (GSE41037), saliva (GSE28746), and mouth swabs (GSE50586). The CpG site cg17861230 corresponds to the neighboring CpG site in PDE4C that was used in the pyrosequencing models (because, the latter is not represented by Illumina Bead Chips). (H) β-values of the CpG site in the PDE4C gene in swab samples were determined by pyrosequencing and correlated with chronological age. (I) Age predictions based on DNAm levels at the CpG site in PDE4C. The linear regression model is depicted in (H). MAD = mean absolute deviation.

Figure 2. Prediction of the cellular composition in mouth swab samples

(A) Representative mouth swab smears with different proportions of leukocytes and epithelial cells. Smears of freshly harvested cells were stained with haematoxylin and eosin. (B) Mean β-values of CpGs on Illumina 27k Bead Chip in datasets of buccal swabs (GSE50586) and blood (GSE39981). Red arrows indicate CpG sites selected for the “Buccal-Cell-Signature”. (C) As additional criterion for suitable cell type-specific CpGs we used the sum of variances in both datasets. (D) Mean β-values at cg07380416 (CD6) and cg20837735 (SERPINB5) were compared in whole blood (GSE41037, GSE39981), hematopoietic subsets (GSE39981), saliva (GSE28746, GSE34035, GSE39560), and buccal swabs (GSE25892, GSE50586). Error bars represent standard deviation. (E, F) The percentage of buccal epithelial cells versus leukocytes was determined by cell counting in 11 stained mouth swab smears. DNAm levels at the two cell type-specific CpGs were determined by pyrosequencing and correlated with cell counts. (G) Linear regressions of both CpGs were combined into the Buccal-Cell-Signature. Predicted percentages of buccal epithelial cells correlated with cell counts. (H) Percentages of epithelial cells were subsequently estimated using the Buccal-Cell-Signature for 55 samples of the training set and 26 samples of the validation set. Error bars represent standard deviation.

Figure 3. Smoking, ethnicity and gender do not impact on DNAm at selected CpG sites

(A) DNAm levels at the three age-associated CpG sites (ITGA2B, ASPA, and PDE4C), and the two cell type associated CpGs (CD6 and SERPINB5) did not differ in blood samples of current smokers (red) and never-smokers (blue; GSE50660). In contrast, such differences were validated in three CpG sites, which have previously been described as smoking-associated. (B) DNAm profiles of pure nasal epithelial cells of smokers (red) and non-smokers (blue) did not demonstrate differences in the two cell type associated CpGs (GSE28368). (C) Pyrosequencing analysis of the Buccal-Cell-Signature in 36 samples with known smoking status did not reveal differences in the cellular composition of buccal swabs. (D) DNAm profiles of children (1 to 17 years) did not reveal significant differences between different ethnic groups (GSE36054; blue: black donor; red: white donor; black lines: Asian donor). (E) None of the five CpGs revealed gender-associated differences (GSE40279; blood samples of 40 to 50 year old donors; blue: female; red: male). * P < 0.05; *** P < 0.0005; Whiskers indicate 10% and 90% percentiles, respectively.

Figure 4. Buccal-Cell-Signature improves epigenetic age prediction

(A) The differences of chronological age and predictions by the 3-CpG-blood-model were compared to the predicted percentage of buccal epithelial cells (according to the Buccal-Cell-Signature). Deviations were higher in samples with more buccal epithelial cells. (B) In analogy, we compared age predictions by the 3-CpG-swab-model to the estimated percentage of buccal epithelial cells and here the impact of the cellular composition was less clear. (C) Combination of age-associated CpGs and Buccal-Cell-Signature in a multivariate regression model of five CpGs (5-CpG-model) facilitated age predictions in the training and validation set. (D) Mean absolute deviations of predicted and chronological age were significantly smaller in the validation set when using the 5-CpG-model as compared to the 3-CpG-swab-model. (E) The models for age-prediction were subsequently validated in a second, independent dataset of 37 samples (18 to 35 years). (F) Samples of the validation group were stratified by an age of 35 years. Comparison of the 3-CpG-swab-model and the 5-CpG-model revealed that the additional analysis of the Buccal-Cell-Signature was particularly relevant for samples of older donors (*** P < 0.0005).