The effect of 5-fluorouracil/leucovorin chemotherapy on CpG methylation, or the confounding role of leukocyte heterogeneity: an illustration

Abstract

Blood-based epigenome-wide association studies that aim at comparing CpG methylation between colorectal cancer (CRC) patients and controls can lead to the discovery of diagnostic or prognostic biomarkers. Numerous confounders can lead to spurious associations. We aimed to see if 5-fluorouracil (5-FU)/leucovorin chemotherapy administered to cases prior to the collection of their blood has an effect on methylation. 304 patients who received treatment and 273 who did not were profiled on the HumanMethylation450 array. Association tests were adjusted for confounders, including proxies for leukocyte cell counts. There were substantial methylation differences between these two groups that vanished once the leukocyte heterogeneity was accounted for. We observed a significant decrease of T cells in the treatment group (CD4+:p=10⁻⁶; CD8+:p=0.036) and significant increase of NK cells (p=0.05) and monocytes (p=0.0006). 5-FU/leucovorin has no effect on global and local blood-based methylation profiles, other than through differences in the leukocyte compositions that the treatment induced.

1. INTRODUCTION

In colorectal cancer (CRC), lymphocyte infiltrates are one of the strongest prognostic factors¹. Tumor infiltration of T cells is associated with longer survival². Changes in DNA methylation patterns of peripheral leukocytes could reflect tumor-mediated immune response or lifetime exposure to environmental risk factors³. Epigenome-wide association studies that aim at comparing blood-based methylation patterns between cases and controls can potentially lead to the discovery of diagnostic or prognostic biomarkers⁴.

In any association study, care must be placed in identifying sources of bias and confounding, and accounting for them⁵. Numerous sources of potential confounding have been described specifically in the context of DNA methylation studies. When comparing the methylation profile of lymphocytes in cases of CRC and controls, examples of potential confounders include alcohol consumption, tobacco use, genetic variants, sex, age and leukocyte heterogeneity. While moderate alcohol consumption confers protection against CRC, heavy alcohol intake is associated with increased risk^6–8. The absorption of folate, a key component of the one-carbon metabolism and an important methyl donor, is impaired by chronic alcohol use^9,10; this can potentially lead to global hypomethylation¹¹. With a two-fold increase in risk in current smokers compared to never smokers, tobacco use is strongly associated with colorectal adenomas¹², a precursor of colorectal tumors. Risk is also elevated for CRC, albeit not as high¹³. Tobacco smoking has a long-term effect on DNA methylation, enough so that derivation of blood biomarkers for past exposure is possible¹⁴. Single nucleotide polymorphisms (SNPs) can potentially be confounders, partly because associations in cis- and trans- between SNPs and methylation at CpG sites are widespread in the genome^15,16, and partly because of technical considerations: one or more SNP in the DNA sequence targeted by a probe designed to measure CpG methylation may cause efficiency differences in hybridization and may generate variation in signal intensity depending on the genotypes of the samples. This would create genotype-dependent measurement errors in methylation values¹⁷. When measuring the methylation profiles of DNA extracted from blood, differences in leukocyte compositions are known to create spurious associations^18,19. For leukocyte composition to be considered a confounder, different proportions in the types of leukocytes should be expected between the blood of CRC cases and controls. These differences could be caused indirectly, such as by age¹⁹ differences, or drugs administered to cancer patients such as adjuvant chemotherapy.

5-fluorouracil (5-FU), a uracil analog, is an adjuvant chemotherapy agent that improves disease-free and overall survival of resected stage III CRC patients²⁰. 5-FU alone was standard adjuvant chemotherapy for CRC for many years, and is an important component of modern CRC combination chemotherapy regimens. It exerts its effects by irreversibly inhibiting thymidylate synthase causing depletion of dTTP, as well as disruption of DNA synthesis and repair resulting in thymineless death and lethal DNA damage^20,21. The misincorporation of 5-FU metabolites into RNA and DNA confers profound effects on cellular metabolism and viability²⁰. 5-FU treatment inhibits synthesis of high molecular weight nuclear RNA and impairs methylation of low molecular weight nuclear 4S RNA²². In an in vitro DNA-based hybridization chain reaction assay, 5-FU inhibited the activity of DNA methyltransferase²³. In human lung cancer cells, expression of DNA methyltransferases was decreased with 5-FU treatment alone, but increased when combined with S-adenosyl methionine, a methyl donor of essential methyltransferase reactions, indicating that the 5-FU may modulate aberrant DNA methylation²⁴. In a human lung cancer cell line, higher concentrations of 5-FU induced significant DNA hypermethylation²⁵. In breast cancer locally advanced tumors, key cell cycle regulator genes were differentially methylated before and after combined treatment with 5-FU and mitomycin D²⁶. These data suggest that 5-FU treatment may affect DNA methylation of peripheral blood leukocytes, through direct and indirect mechanisms.

In what follows, we sought to evaluate if patients who received adjuvant 5-FU chemotherapy treatment prior to the collection of their blood for biobanking purposes display substantial CpG methylation profile differences compared to patients who did not receive such treatments.

2. MATERIALS AND METHODS

2.1 Sample description

CRC patients used in the present study are part of a bigger study of cases and controls that were enrolled in phase I of the Ontario Familial Colon Cancer Registry (OFCCR)^27,28. Briefly, probands were selected from incident CRC cases identified between July 1,1997 and June 30, 2000 from the population-based Ontario Cancer Registry. Patients were stratified based on whether or not they received treatment prior to the collection of their blood for biobanking purposes. The large majority of patients who received treatment received a combination of 5-fluorouracil and leucovorin. The few patients (<10) who received other drugs (Xeloda, Tomudex, CPT-11) prior to the collection of their blood were excluded from the analyses.

2.2 Methylation profiling

The set of patients used in the current study consist of a subset of a larger collection of 2,203 samples from 2,101 unique donors (1,103 cases of CRC and 998 controls) that were profiled on the HumanMethylation450 array from Illumina¹⁵. Lymphocyte pellets were extracted from whole blood using Ficoll-Paque PLUS (GE Healthcare). DNA was extracted from lymphocytes using phenol–chloroform or the Qiagen Mini-Amp DNA kit, except for 99 samples (90 cases, 9 controls), for which DNA was extracted from lymphoblastoid cell lines. We used 15 μl of DNA from all samples at concentrations averaging 90 ng/μl (20 ng/μl s.e.). DNA samples were bisulfite-converted using the EZ-96 DNA Methylation-Gold Kit (Zymo Research, Orange, CA); 4 μl of bisulfite-treated DNA was then analysed on the HumanMethylation450 BeadChip from Illumina according to the manufacturer’s protocol.

The OFCCR methylation data was deposited in dbGaP under the accession number phs000779.v1.p1.

2.3 Calculation of methylation ratios from intensities

We calculated the methylation ratios (proportion of molecules that are methylated at a given site, also known as β-value) as β=min(M,1)/(min(M,1)+min(U,1)) where M and U are, respectively, the methylated and unmethylated intensity signals (possibly background-corrected signals). This is a slight modification of GenomeStudio’s calculation: Illumina defines the ratio as β=min(M,0)/(min(M,0)+min(U,0)+100). In this expression, the offset (100) artificially moves the methylation value away from 1. Moreover, the possibility of the numerator being 0 makes the resulting distribution of β-values both discrete and continuous (with point mass at 0). We prefer our expression for β, which results in a continuous distribution that does not bias against a fully methylated state.

2.4 Normalization of methylation values

Based on method comparisons that we have made elsewhere¹⁵ (also see below), methylation values were derived from data that was background-corrected using NOOB (methylumi.bgcor from the R package methylumi v2.14.0)²⁹ followed by colour adjustments using Illumina’s normalization probes and algorithms; BMIQ³⁰ (from the R package wateRmelon v1.8.0) was then applied to the set of methylation values.

2.5 Sample exclusion

We excluded from association analyses: (1) samples for which DNA was extracted from lymphoblastoid cell lines; (2) samples that were outliers with respect to any one of the internal control probes (excluding probes designed to evaluate the background noise and probes designed to normalize the data), where an outlier is defined as any value more than three times the interquartile range away from the closest quartile; (3) samples that were outliers with respect to any one of the first 12 principal components calculated from the 131,045 most variable CpG sites¹⁵ (corresponding to the approximate location of the elbow of the eigenvalue scree plot), recomputed after exclusion of the samples in steps 1 and 2; (4) samples that were not of non-Hispanic white ancestry, either self-declared or by investigation of genetic ancestry using genome-wide SNP data; and (5) samples with intensities on the X or Y chromosome inconsistent with their gender. For duplicate samples, we only kept the one with the smallest number of sites with a detection p-value (see below) less than 0.01.

2.6 CpG site exclusion

The detection p-value is a quantitative assessment of the probability that the target sequence signal is distinguishable from negative controls³¹, and is included along with the intensity values. All β-values with a detection p-value less than 1% were treated as missing data. Sites with more than 1% missing values after sample exclusion were discarded.

We excluded a CpG site if: (1) its probe sequence aligned to multiple locations in the genome with >= 90% identity¹⁷; (2) a SNP resides at the cytosine or guanine base of the CpG site, or, in the case of Infinium I probes, at the position where single-base extension occurs, irrespective of its minor allele frequency (MAF); (3) a SNP resides elsewhere within the probe target sequence if its MAF is at least 5%^15,17. Minor allele frequencies were based on the samples of European ancestry from the 1000 Genomes Project³².

2.7 Statistical analysis

We sought to evaluate if 5-FU/leucovorin treatment prior to the collection of a patient’s blood for biobanking purposes has an effect on the methylation profile of DNA extracted from blood. Association between methylation levels and chemotherapy status was assessed using a linear regression model, where the methylation level at a CpG site was taken as the dependent variable after logit transformation, and chemotherapy status as the independent variable. In the linear regression model, covariates were included to remove unwanted variation. These included: age, sex, alcohol consumption status, smoking status, the methylation array identifier on which the sample was found (12 samples were included on each array), position of the sample on its array and the first two principal components (PCs) calculated from control probes.

To adjust for alcohol consumption, one categorical variable was used to indicate if the individual was a current drinker (at least one drink per week), if he or she has never drunk alcohol since they turned 50 (or 30, or 20, depending on their current age), or to indicate missing or ambiguous current drinker status. By construction of the questionnaire, the status can often not be derived unambiguously: the status is derived from questions such as, e.g., “since you turned 50, how many years in total did you consume at least one alcoholic beverage (of any type) a week?” An individual who is 65 year-old who answers 15 or 16 years (due to rounding) can reasonably be classified as a current drinker, but any other non zero answer is ambiguous (perhaps due to interruptions) and can lead to misclassification.

Smoking status³³ was assessed using a categorical variable to indicate current/never/former smoker status, or missing or ambiguous smoking-related data.

As proxies for unobserved (technical) covariates, PCs were derived from the set of control probes included on the arrays, whose function is to assess the performance of the assays: these probes assess bisulfite conversion, hybridization, extension, staining, specificity, stripping (“target removal” control probes), and overall performance (“non-polymorphic” control probes). In addition, dye-bias normalization probes and the out-of-band intensities from Infinium I probes²⁹ were used in the derivation of these PCs. All these controls probes were transformed into scaled summary variables from which PCs were derived, as described in Fortin et al ³⁴. There, the PCs were used to reshape the whole distribution of methylation values in each sample, upstream of association analyses (i.e., as a normalization method). Instead of applying this normalization procedure, we chose to use these PCs as covariates in a linear regression model, because they will further be used as described below, and also because we found this method to be outperformed by the NOOB+dye bias+BMIQ data normalization strategy used here. Adjusting for these PCs derived from control probes to remove unwanted variation is thus similar to the RUV-2 method³⁵ (Remove Unwanted Variation, 2-step) derived for gene expression microarrays, where the first step is to perform factor analysis (singular value decomposition or principal components) on a set of control (“housekeeping”) genes, and the second step is to use the first few of these factors as covariates in a linear regression model. Note that defining covariates using factors or principal components derived from methylation probes targeting actual CpG sites as opposed to control probes can potentially remove true biological differences if these differences are large³⁵. Of note, RUV-2-based normalization methods tailored specifically for methylation datasets have been published with available code during preparation of this manuscript³⁶.

To further account for leukocyte heterogeneity, we computed factors derived from 600 CpG sites chosen for their ability to discriminate the different blood cell types (CD4+ and CD8+ T cells, B cells, NK cells, monocytes and granulocytes)^19,37. We used the singular value decomposition (SVC) of the data matrix constructed from the 600 CpG sites, so that we could include the control probe principal components described above as covariates. This was done so that the factors (components) derived from the 600 leukocyte-discriminating sites do not capture the same unwanted variation than the principal components calculated from control probes. To derive these components while accounting for covariates, we used the RUV-2 code³⁵. We used the first 6 components, a choice driven by the 6 purified cell types used to derive the 600 leukocyte-discriminating sites. Percentages of each cell type in our lymphocyte samples were estimated using the estimateCellCounts function from the minfi R package v1.14.0^19,38.

The mean methylation values that we report were adjusted for all covariates using the adjust function from the R package epicalc v2.15.1.0. All other analyses used functions included with R v3.0.0.

3. RESULTS

3.1 Description of the data set

In the present study, genome-wide CpG methylation profiles detected using the Infinium HumanMethylation450 BeadChips were obtained from DNA extracted from the lymphocytes of 577 CRC patients with colon cancer enrolled in the Ontario Familial Colon Cancer Registry (OFCCR)^27,28, of whom 304 received chemotherapy treatments prior to the collection of their blood for biobanking purposes, and 273 did not. Table 1 further breaks down the sample composition based on covariates. The two groups did not differ significantly based on sex (p=0.93; Fisher exact test) or alcohol consumption (p=0.34). The two groups displayed differences based on smoking status (p=0.026), driven by a relatively larger proportion of former smokers among patients who received chemotherapy. Patients who received chemotherapy were moreover slightly older than patients who did not (p=0.0016; Wilcoxon rank-sum test).

Table 1.

Characteristics of CRC patients entering the study.

	No chemo received	5-FU received
N	273	304
% Females	49.1%	48.7%
Median age (range)	63 (29–77)	64 (35–76)
Smoking
Never	101	113
Former	120	159
Current	28	17
Ambiguous/missing	24	15
Alcohol use
Never	90	91
Current	80	80
Ambiguous/missing	103	133
Median time (days) since initiation of treatments (range)	NA	591 (51–1910)

In Lemire et al¹⁵, we used 129 pairs of duplicate samples to compare normalization strategies, which motivated the choice of NOOB+dye-bias+BMIQ (51 pairs for which both duplicates are found on the same 96-well DNA plates; 78 pairs for which both duplicates are found on different plates). For completeness, and because we were faced with a choice between using control probe-derived principal components from within a normalization method or instead as covariates in a linear regression framework, we compared the methylation values from NOOB+dye-bias+BMIQ normalization to values obtained from functional normalization³⁴ (from the R package minfi v1.14.0). Based on correlations in duplicates, Supplementary Figures 1–2 show a slightly greater reproducibility for the NOOB+dye-bias+BMIQ method.

We normalized the methylation data sets and identified and flagged sites with more than 1% missing data (13,032 sites), calculated after sample exclusions. We further identified CpG sites that were polymorphic irrespective of the minor allele frequency (MAF), sites for which at least one SNP with a MAF above 5% resided anywhere else underneath the probe sequence (69,974 sites derived from Chen et al.¹⁷) and sites that cross-hybridized with >90% identity to multiple regions (30,436 sites¹⁵). These overlapping counts amount to 93,675 unique sites that were discarded. We retained the methylation values at 380,189 autosomal CpG sites.

3.2 Effect of 5-FU on lymphocyte CpG methylation profiles

Figure 1 illustrates the confounding role of leukocyte heterogeneity in blood-based methylation profiles. For clarity, Figure 1 focuses on the 218,276 most variable CpG sites (sites for which the difference between the most and least methylated individuals is at least 5%, among the 95% of the individuals forming the central distribution of methylation values -- or 95% reference range). Ignoring the blood cell composition, we observed that a majority of sites (133,211) displayed lower methylation levels in those who received adjuvant chemotherapy treatment prior to blood sampling, albeit not significantly for the large majority. This asymmetry is nevertheless informative in terms of the global effect the drug has on methylation; the asymmetry affected mostly CpG sites with intermediate methylation levels (in the ~0.20–0.80 approximate range; Figure 1B). However, a dendrogram and heat map representation of the sites that showed the strongest association with chemotherapy status, in data sets of purified leukocyte subtypes³⁷, revealed that these sites also tended to show substantial differences between the different cell types (Figure 2). Of note, these sites particularly tended to discriminate T cells from other leukocyte types. Estimating the percentages of each leukocyte subtype from carefully chosen CpG sites¹⁹ revealed that chemotherapy patients have reduced CD8+ (p=0.036) and CD4+ T cells (p=9.8×10⁻⁷), and elevated NK cells (p=0.05) and monocytes (p=0.0006) (Figure 3). This is consistent with reports that 5-FU reshapes the composition of leukocyte subpopulations^39–43. Of the cell types that showed significant differences, only NK cells proportions were associated with time elapsed since the start of the treatment (p=0.00017; Figure 4); however NK cells increase in proportion with time, as opposed to reverting back to the level seen in those who did not receive the drug, as might be expected. Granulocytes (which are present in lymphocyte pellets as contaminants) also displayed an association with time (p=0.008). However, the association in granulocytes is mostly driven by 6 patients who started treatment 128 days or less before blood collection. The association was lost by excluding them (p=0.19) but did not affect significance in NK cells (p=0.0005). Note that it was necessary to control for the age of the patients, since blood cell proportions vary with age. Supplementary Figure 3 illustrates this; within the age range, cell type proportion estimates tend to increase or decrease with age, in directions consistent with published data¹⁹, except for granulocytes.

Figure 1. Association results.

(A) Volcano plot of the mean methylation differences between patients who underwent chemotherapy prior to collection of their blood and those who did not. Leukocyte heterogeneity is ignored. Each point from the underlying heat map is an autosomal CpG. Mean methylation differences are adjusted for covariates; the distribution of all mean methylation differences is illustrated by the boxplot. Significance is plotted on the negative log scale. (B) Heat map representation of the mean methylation level calculated across all samples and the direction of the difference in (A); the sign of the y-axis is the sign of the methylation difference displayed in (A). A point on the negative side correponds to a CpG site that is less methylated in those who received treatments. Colors range from white (high density of points) to green (low density of points).

Figure 2. CpG sites associated with treatment.

Top 2000 CpG sites associated with chemo treatment, prior to adjustment for leukocyte heterogeneity, evaluated in purified human leukocyte subtype methylation data sets. Colors range from dark blue (unmethylated CpG site) to red (fully methylated CpG site).

Figure 3. Leukocyte percentages.

Estimates of percentages of (A) CD8+ T cells, (B) CD4+ T cells, (C) NK cells, (D) B cells, (E) monocytes and (F) granulocytes, stratified by chemo treatment prior to the collection of blood. Significance calculated from a linear regression model, adjusting for age and sex.

Figure 4. Time since start of 5-FU treatments.

Estimates of percentages of (A) CD8+ T cells, (B) CD4+ T cells, (C) NK cells, (D) B cells, (E) monocytes and (F) granulocytes, as a function of time since start of 5-FU treatments. Significance calculated from a linear regression model, adjusting for age and sex.

Results from Figure 1 were thus likely to have been confounded by leukocyte heterogeneity. By including additional covariates designed to capture the effect of heterogeneity of leukocyte-derived DNA, the asymmetry apparently caused by chemotherapy treatment vanished (Figure 5). Moreover, no single CpG sites were associated with treatment at an FDR<5%. We conclude from these results that 5-FU/leucovorin has no effect on leukocyte methylation, both globally and locally.

Figure 5. Association results, adjusted.

See Figure 1 for details. Covariates to adjust for leukocyte heterogeneity are included in the model. Scaling is the same as in Figure 1.

4. DISCUSSION

In this study we illustrated the confounding role of leukocyte heterogeneity in methylation studies caused by drug treatment. Results have shown that 5-FU/leucovorin has no effect on global and local blood-based methylation profiles, other than through differences in the leukocyte compositions that the treatment induced. However, since only mixtures of white blood cells were profiled, it is still unclear if treatment could induce methylation differences in specific leukocyte subtypes. For human brain tissue samples that consist of mixtures of neuronal and glial cells, a framework for the deconvolution of the (mixed) methylation fractions into neuronal and glial components has been proposed and applied⁴⁴ to help identify differentially methylated regions specific to each cell type, as well as reduce confounding. The authors provided guidelines on how to apply their framework to mixtures consisting of more than two cell types, but acknowledge that robustness may be an issue considering the added noise from the larger number of cell fraction estimates⁴⁴. Such a decomposition of methylation fractions from whole blood would be specifically relevant when, for example, the focus of a study is on tumor T lymphocyte infiltrates; these are avenues to explore, but validation would be necessary.

5-FU treatment has been described to affect lymphocyte numbers in the peripheral blood of patients. In patients with recurrent or metastatic CRC, treatment with 5-FU/leucovorin caused significant decreases in the number of T cells⁴⁰. In the blood of patients with locally advanced rectal cancer, lymphocytes were the most affected as compared to monocytes, neutrophils, and WBCs at 4–12 weeks after chemoradiotherapy with 5-FU and leucovorin⁴³. At 6–8 weeks after 5-FU treatment, a sustained ~20% reduction in T-cells and an increase in monocyte percentage was documented in the blood of patients with multiple sclerosis³⁹. In mice, 5-FU monotherapy reduced the numbers of CD4 and CD8 positive T cells in tumor draining lymph nodes⁴¹. In a human Jurkat T lymphocyte cell line, administration of 5-FU induced death⁴².

The data and results presented here are consistent with these reports. This illustrates that powerful insights can be obtained from investigating a carefully selected set of leukocyte-specific CpG markers and that their use in methylation association studies can substantially reduce confounding.

Highlights.

• The CpG methylation profiles of leukocytes subpopulations display substantial heterogeneity

• 5-fluorouracil/leucovorin chemotherapy reshapes the composition of leukocyte subpopulations

• 5-fluorouracil/leucovorin chemotherapy has no effect on blood-based methylation profiles