Blood-based profiles of DNA methylation predict the underlying distribution of cell types: a validation analysis

Abstract

The potential influence of underlying differences in relative leukocyte distributions in studies involving blood-based profiling of DNA methylation is well recognized and has prompted development of a set of statistical methods for inferring changes in the distribution of white blood cells using DNA methylation signatures. However, the extent to which this methodology can accurately predict cell-type proportions based on blood-derived DNA methylation data in a large-scale epigenome-wide association study (EWAS) has yet to be examined. We used publicly available data deposited in the Gene Expression Omnibus (GEO) database (accession number GSE37008), which consisted of both blood-derived epigenome-wide DNA methylation data assayed using the Illumina Infinium HumanMethylation27 BeadArray and complete blood cell (CBC) counts among a community cohort of 94 non-diseased individuals. Constrained projection (CP) was used to obtain predictions of the proportions of lymphocytes, monocytes and granulocytes for each of the study samples based on their DNA methylation signatures. Our findings demonstrated high consistency between the average CBC-derived and predicted percentage of monocytes and lymphocytes (17.9% and 17.6% for monocytes and 82.1% and 81.4% for lymphocytes), with root mean squared error (rMSE) of 5% and 6%, for monocytes and lymphocytes, respectively. Similarly, there was moderate-high correlation between the CP-predicted and CBC-derived percentages of monocytes and lymphocytes (0.60 and 0.61, respectively), and these results were robust to the number of leukocyte differentially methylated regions (L-DMRs) used for CP prediction. These results serve as further validation of the CP approach and highlight the promise of this technique for EWAS where DNA methylation is profiled using whole-blood genomic DNA.

Keywords: DNA methylation; cell mixture analysis; leukocytes; mixture deconvolution; whole-blood.

Figure 1. Illustration of the blood cell mixture deconvolution approach. This approach involves, (A) constrained projection of DNA methylation profiles from a target methylation data set (S₁) onto a reference data set (S₀ ), which is comprised of the DNA methylation signatures for isolated white blood cell types (shapes reflect different white blood cell types). The result is an estimate of the underlying distribution of cell proportions (circle, triangle and hexagon) for each sample within S₁. (B) This approach assumes that the methylation signature for samples within S₁ are the weighted sum of the methylation signatures from individual white blood cell types, where the weights are proportional to the cell-type frequencies.

Figure 2. Complete blood cell (CBC) and predicted proportions of white blood cell types in the target methylation data set. CBC derived proportions [i.e., Ω^(CBC)] of white blood cell types in (A) whole-blood and (B) peripheral blood mononuclear cell (PBMCs) (i.e., devoid of granulocytes) for the samples in the target methylation data set. (C) Predicted proportions (i.e., ) of CD8+ T-lymphocytes (CD8T), CD4+ T-lymphocytes (CD4T), Natural killer cells (NK), B cells (Bcell), Monocytes (Mono) and Granulocytes (Gran) for the target samples using constrained projection (CP). Black bars denote the median and the red dashed bars denote the 75th and 25th percentiles for the predicted cell-type proportions. Colored points indicate subjects with replicate samples, where two points of the same color denote replicate samples for the same subject.

Figure 3. Comparison of the predicted and CBC derived proportions of monocytes and lymphocytes among the target samples. Scatter-plot of the predicted and CBC-derived proportions of (A) monocytes and (B) lymphocytes. Solid red lines represent the unity lines (i.e., y = x). Bland–Altman plots for (C) monocyte and (D) lymphocyte proportions. Y-axes represent the difference in the predicted and CBC-derived cell-type proportions, and X-axes represent the mean cell-type proportions based on CP prediction and CBC-based proportions. Red-dotted lines indicate the global bootstrap-based 95% prediction intervals for the difference in predicted and CBC-derived cell-type proportions.

Figure 4. Prediction performance as a function of the number of L-DMRs used in CP. (A) Pearson correlation between the predicted and CBC-derived proportions of monocytes (blue line) and lymphocytes (red line) as a function of the numbers of L-DMRs used in CP. (B) root mean squared error (rMSE) for monocytes and lymphocytes and (C) median (%) granulocytes as a function of the numbers of L-DMRs used in CP. (D) Pearson correlation between the predicted and CBC-derived proportions of monocytes and lymphocytes as a function of the numbers of non L-DMRs (negative controls) used in CP.