Swartz et al. [1] report that an epigenetic signature links socioeconomic status (SES) to amygdala function and depression, built on an impressive prospective design. Here, we discuss the methods for quantifying blood DNA methylation (DNAm) employed in this study, focusing on their use of principal component analysis (PCA) to collapse 20 target CpG sites and extract one PC without prior correction of blood cell-type heterogeneity in DNAm profiles. Since the primary driver of variability in DNAm within whole blood is cell type [2,3,4,5], we argue that the index used in this analysis is likely reflective of interindividual differences in cell-type proportions rather than the DNAm mechanism inferred by the authors. We demonstrate this position analytically using publicly available DNAm data.

CpG sites located in the proximal promoter of the serotonin transporter gene (SLC6A4) were targeted by Swartz et al. [1] using bisulfite pyrosequencing. CpG sites are densely located at many gene promoters, referred to as “CpG islands”, which are typically unmethylated [6]. Consequently, there tends to be a low dynamic range of DNAm across individuals for promoter sites containing CpG islands. These invariable measurements are partly attributed to the way that DNAm is quantified: a single CpG site is present two times per cell, one on each of a chromosome pair, and thus DNA can be 0, 50 (rarely) or 100% methylated. An average across thousands or millions of cells present in a typical biological sample results in DNAm values ranging between 0 and 100% methylated, with CpG island sites typically demonstrating values very close to zero. This is indeed the case with the 20 CpGs used by the authors (from a prior publication), which demonstrate a mean percent methylation across participants ranging between 0.88 and 3.57% methylated [7] (Supplementary Table 1). Considering that a 5% difference in DNAm is a commonly used threshold for a biologically meaningful effect [8] and that 5% is the error for pyrosequencing [9], 0.88–3.57% would be considered a low range.

Although the authors address the limitations of making inferences about DNAm in the brain using blood, the important influence of cell-type differences in DNAm variability is overlooked. DNAm plays an essential role in the differentiation of tissues and cell types, resulting in highly cell-type-specific DNAm patterns [10, 11]. Cell-type heterogeneity within blood is generally the largest contributor to DNAm variation [3,4,5, 12], which includes seven cell types (and subcategories within these types) with distinct DNAm profiles. Variation attributable to cell type in blood exceeds interindividual variability explained by age, ethnicity, or exposures, and thus emerges in top PCs [3,4,5, 12]. For instance, one group reported that the top two DNAm PCs from blood are highly significantly correlated with cell-type proportions in five publicly available DNAm data sets [5]. This large contribution of cell type is so robust that the top PC is statistically leveraged by some methods for cell-type correction [13], which has been argued to be valid for blood samples but not necessarily for other more homogeneous or disease-related tissues [2].

Figure 1: Below, we use two publicly available DNAm data sets [12, 14] to argue what has been shown before: that the primary driver of variability in DNAm data in blood, including at the SLC6A4 promoter, is due to cell-type proportions. First, for descriptive purposes, we show DNAm levels in isolated white blood cell types for the four most variable SLC6A4 CpGs (one of which belongs to a CpG island, showing very low DNAm levels). DNAm at each of these sites is highly associated with blood cell type. Because these cell-type- specific patterns were measured on an older array technology, there is limited representation and no overlap with the region investigated by Swartz et al. However, we also show data drawn from a second more recent data set using the Illumina EPIC array [15]. With greater coverage, this array captures one CpG included in the PC used by Swartz et al. Although this study ran the array on whole blood, reference-based methods can be applied to array data to bioinformatically estimate cell-type proportions [8]. We correlated the estimates of cell-type variability derived for each individual with (1) DNAm at the Swartz CpG site, (2) within 1000 bp of this region, (3) across SLC6A4, and (4) across all CpG sites assayed by the EPIC array, and show that in all cases, cell-type proportions and DNAm are extremely, highly correlated. Although not conclusive, this analysis strongly suggests that the variable probes contributing to the first PC used in the authors’ analysis do indeed reflect cell-type proportions rather than interindividual variability in DNAm.

Fig. 1
figure 1

a, b Associations between cell types and DNAm levels. a SLC6A4 variable CpG β values by blood cell types. b Associations between the first PC of cell count predictions and DNA methylation. Note: a B cell = CD19+ B cells; CD4T = CD4+ T cells; CD8T = CD8+ T cells; Eos = eosinophils; Gran = granulocytes; Mono = CD14+ monocytes; Neu = neutrophils; NK = CD56+ natural killer cells; PBMC = peripheral blood mononuclear cells; WB = whole blood. DNAm levels in isolated white blood cell types for the four most variable SLC6A4 CpGs drawn from Reinius et al., 2012 Infinium HumanMethylation450k array data [12]. Four CpGs from the SLC6A4 region with the highest standard deviations are shown. The second CpG is classified as located within a CpG island by the UCSC Annotation. For all four examples, β values are significantly associated with cell-type category (significant cell types with p < 0.05 are indicated by a “*” symbol). Note: b Data on DNA methylation from whole-blood samples and relations to cell-type proportions drawn from Guastafierro et al., 2017 Infinium HumanMethylationEPIC array data [14]. The first PC of cell counts was based on estimated counts of blood cell types produced by the “estimateCellCounts” function in the minfi package in R. Top left: β values of one CpG overlapping with the region investigated by Swartz et al. (Genome Build 37 location is 17:28562813). Top right: first PC from SLC6A4 promoter region (17:28561783:28563929, which is 1000 bp around the area included in the PC used by Swartz et al., and includes 12 CpGs). Bottom left: the first PC from all 31 SLC6A4 CpGs. Bottom right: the second PC from genome-wide PCA. The second rather than the first PC is plotted because the first/zeroth PC of DNA methylation quantified by array technology represents the variation in mean DNA methylation values of probes across the genome [17]. This is opposed to quantitative bisulfite pyrosequencing used by Swartz et al., or targeted gene regions, which capture a small number of highly correlated CpGs, opposed to epigenome-wide variability patterns

Full size image

This observation warrants an adjustment of the interpretation that DNAm status of the SLC6A4 promoter is predicted by SES, and in turn predicts amygdala reactivity, at least until the finding is replicated in the context of cell-type correction or cell-type confounding is explicitly tested in the author’s data. Indeed, blood cell-type proportions can themselves be related to environmental exposures [16]. We previously documented that after correcting for cell type in blood samples, a connection between DNAm and current SES was no longer present [17]. Moreover, it was recently reported that the ratio of inflammatory to antiviral white blood cell types, calculated bioinformatically using DNAm profiles, mediated the association between SES and chronic illness. Specifically, a higher ratio of monocytes and natural killer cells (i.e., markers of chronic inflammation) to T and B cells (adaptive immune system and antiviral cells) accounted for the relationship between low SES and chronic disease [18]. These reported interindividual cell-type differences in immune cells are consistent with a broad literature linking circulating immune cells to stress levels as an adaptive response of the body to threat [19].

It is thus essential to determine whether findings from the target article reflect DNAm variability, cell-type proportion differences related to SES, or a combination of the two. Regardless, these findings provide an intriguing example of the complexity of the biological processes potentially affected by SES.