Dietary Omega-3 Fatty Acid Intake Impacts Peripheral Blood DNA Methylation -Anti-inflammatory Effects and Individual Variability in a Pilot Study

Abstract

Omega-3 or n-3 polyunsaturated fatty acids (PUFAs) are widely studied for health benefits that may relate to anti-inflammatory activity. However, mechanisms mediating an anti-inflammatory response to n-3 PUFA intake are not fully understood. Of interest is the emerging role of fatty acids to impact DNA methylation (DNAm) and thereby modulate mediating inflammatory processes. In this pilot study, we investigated the impact of n-3 PUFA intake on DNAm in inflammation-related signaling pathways in peripheral blood mononuclear cells (PBMCs) of women at high risk of breast cancer.

PBMCs of women at high risk of breast cancer (n=10) were obtained at baseline and after 6 months of n-3 PUFA (5 g/day EPA+DHA dose arm) intake in a previously reported dose finding trial. DNA methylation of PBMCs was assayed by reduced representation bisulfite sequencing (RRBS) to obtain genome-wide methylation profiles at the single nucleotide level. We examined the impact of n-3 PUFA on genome-wide DNAm and focused upon a set of candidate genes associated with inflammation signaling pathways and breast cancer.

We identified 24,842 differentially methylated CpGs (DMCs) in gene promoters of 5507 genes showing significant enrichment for hypermethylation in both the candidate gene and genome-wide analyses. Pathway analysis identified significantly hypermethylated signaling networks after n-3 PUFA treatment, such as the Toll-like Receptor inflammatory pathway. The DNAm pattern in individuals and the response to n-3 PUFA intake are heterogeneous. PBMC DNAm profiling suggests a mechanism whereby n-3 PUFAs may impact inflammatory cascades associated with disease processes including carcinogenesis.

1. Introduction

Dietary n-3 polyunsaturated fatty acids (PUFAs) are associated with health benefits for a wide range of diseases in both preclinical and clinical studies, particularly for conditions with underlying chronic inflammation such as cardiovascular disease, rheumatoid arthritis, dementia, and cancer [1,2]. In humans, blood and tissue concentrations of eicosapentaenoic acid (EPA) and docosahexaenoic acid (DHA) are largely dependent on consumption of foods or supplements enriched in these fatty acids such as fatty fish or fish oils, given the inefficient conversion of plant-derived alpha linolenic acid to long chain n-3 PUFAs [3]. Intervention studies of dietary n-3 PUFAs/fish oil have shown increased EPA, DHA, and n-3:n-6 PUFA ratio in plasma, adipose tissue, and circulating cells such as erythrocytes and peripheral blood mononuclear cells (PBMCs) [4–6].

Clinical investigations with dietary EPA and DHA have demonstrated anti-inflammatory effects on PBMC responses. Fish oil supplementation reduces the production of inflammatory cytokines including interleukin 2 (IL-2), interleukin 1 (IL-1), and tumor necrosis factor alpha (TNFα) by stimulated mononuclear cells [7–9]. Healthy subjects taking fish oil (775 mg/d EPA) and borage oil (831 mg/d gamma linolenic acid) supplements for four weeks demonstrate decreased PBMC gene expression of PI3Kα and PI3Kγ, key mediators of pro-inflammatory signal transduction, and cytokines (IL-1, IL-10, IL-23) [10]. In older adults, treatment with 1.8 g/day EPA+DHA for six months also showed a decrease in PBMC pro-inflammatory gene expression involving the interleukin 6, MAP kinase, NF-κB, and Toll-like receptor signaling pathways [11]. In an Alzheimer’s disease cohort, supplementation with DHA-rich n-3 PUFA (1.7 g/d DHA, 0.6 g/d EPA) versus corn oil for 6 months led to decreased secretion of IL-6, IL-1β, and granulocyte colony stimulating factor by PBMCs stimulated ex vivo by lipopolysaccharide (LPS), without impacting TNFα, IL-10 [12]. In parallel, PBMC gene expression showed changes consistent with anti-inflammatory effects of omega-3 fatty acid intake [13].

Epigenetic regulation of gene expression may mediate, at least in part, the anti-inflammatory effects of dietary n-3 PUFAs. DNAm is one mechanism by which pro-inflammatory genes are silenced, as shown in prior reports of epigenetic regulation of inflammation in a wide range of cell types including PBMCs [14,15], CD4+ lymphocytes [16], and cancer cell lines [11,17–19]. Regarding the role of dietary n-3 PUFAs in modulating the epigenome, a cross-sectional study of DNAm in PBMCs of Yup’ik Alaskan Native Americans at the lowest and highest range of long chain marine n-3 PUFA intake showed potential impact of DNAm on anti-inflammatory pathway genes [20]. Modulating effects of n-3 PUFAs on DNAm that impact inflammatory and PUFA metabolism pathways are also evident in short term intervention trials of n-3 PUFAs [21–23]. However, to date, very few n-3 PUFA intervention studies conducted in humans investigate DNAm on a genome-wide scale [23].

Towards our goal of investigating n-3 PUFA in breast cancer prevention, we conducted this pilot study to determine whether DNAm changes could be detected in PBMCs. We utilized reduced representation bisulfite sequencing (RRBS) to analyze DNAm on a genome-wide level with single base-pair resolution as well as coverage of many more sites than array-based methods. Our work shows that DNAm changes after dietary n-3 PUFA treatment in women at high risk of breast cancer are detectable in PBMCs. Additionally, our findings implicate two inflammatory pathways and uncover variability in DNAm that may indicate variable treatment effects.

2. Materials and Methods

2.1. PBMCs

PBMCs were isolated via ficoll-hypaque separation from peripheral blood collected in acid citrate dextrose (ACD) blood tubes as part of a study of n-3 PUFA supplementation in women at high risk of breast cancer; the main results of the trial were previously reported [4]. In brief, 48 women at high risk of breast cancer were randomly assigned to one of four daily doses of n-3 PUFAs (0.84, 2.52, 5.04, or 7.56 g/d of EPA+DHA) for 6 months of treatment. The study was conducted with the approval of the Institutional Review Board of The Ohio State University and in accordance with ethical standards of the 1975 Helsinki Declaration and its later amendments. For full eligibility criteria, demographic, and anthropometric measures, see Table 1 and Yee et al. 2010 [4]. PBMCs were processed within 4 hours of collection and cryopreserved in liquid nitrogen until analysis. DNA quality and integrity were assessed with Thermo-Fisher Scientific™’s NanoDrop™ and Agilent™ BioAnalyzer™ instruments, respectively. In the present study, we utilized PBMCs obtained at 0 and 6 months of 5.04 g/d EPA+DHA from the 10 of 12 sample sets with highest increase in DHA/EPA at 6 months. This dose is utilized in an ongoing phase II trial by our team of investigators (clinicaltrials.gov ID NCT02295059).

Table 1.

Characteristics^* of Study Participants (n=10)

	Baseline	6 months
Age	50.8 ± 6.9
Weight (kg)	74.6 ± 11.5	74.6 ± 11.8
BMI	27.1 ± 4.9	27.1 ± 5.1
Waist (cm)	84.2 ± 8.9	85.3 ± 9.6
Waist to Hip Ratio	0.8 ± 0.07	0.8 ± 0.05
Premenopausal (n)	5
Postmenopausal (n)	5

^*mean values ± SD

For further details, see Yee et al. 2010 [4].

2.2. Fatty Acid and Serum Biomarker Experiments and Analysis

Fatty acid compositions for both the serum and the breast adipose tissues samples were reported with the main study findings in Yee et al., 2010, which also includes the methods utilized for these analyses [24]. As described in Yee et al., breast adipose samples were collected via fine needle aspiration of the breast. Fatty acids were reported a percentage of the total fatty acid content. Change in fatty acids from n-3 PUFA treatment were identified using a paired t-test to produce a p-value. The Benjamini-Hochberg false discovery rate (FDR) multiple test correction was applied to all p-values [25]. The serum biomarkers were reported with the main study findings; the methodology was previously described [4]. Briefly, serum samples were processed within 1 hour of venipuncture and stored at −80°C until analysis in duplicate for: Adiponectin and leptin were measured by radioimmunoassay (Millipore, St Charles, MO); glucose by enzymatic assay (Yellow Springs International, Yellow Springs, OH); insulin, SHBG, and hsCRP by chemiluminescence (Siemens Medical Solutions Diagnostics, Duluth, MN); IGF-I and IGFBP3 by enzyme immunoassay (Beckman Coulter, Webster, TX); C-peptide and estradiol by radioimmunoassay (Beckman Coulter, Webster, TX); and IL-6 and TNF-a by electrochemiluminescence (Meso Scale Discovery, Gaithersburg, MD). The assay was repeated for any duplicate values with a CV of .8%. Changes in these values as well as anthropometric values (see Table 1) were identified using a paired t-test and applying the Benjamini-Hochberg FDR multiple test correction. Similar to the original study, only triglyceride levels showed a significant change from n-3 PUFA treatment (FDR p-value = 0.046; Pretreatment mean = 81.4; Posttreatment mean = 56).

2.3. Power analysis and Sample Size Determination

To determine how many samples should be included in this study, we performed a power analysis based on detecting average DNAm differences at CpG loci of 5% in PBMCs. The variance for PBMC DNAm was determined using publicly available data. Control PBMC samples assayed for DNAm using the Illumina Infinium 450k array were used to determine the standard deviation (SD) in DNAm at CpGs as a function of DNAm (GEO series GSE57107). A fit of the SD values produced a maximum value at 50% DNAm (max of fit SD values = 0.30; data not shown). For the power analysis, a more conservative SD of 0.65 which corresponds to the 95th percentile of all CpGs was used. The power analysis was conducted using an FDR corrected alpha = 0.05 to account for multiple testing of the CpGs in the promoter region of our candidate genes. We found that 20 total samples (10 baseline and 10 after treatment) produced sufficient statistical power for detecting a 5% difference in average DNAm of 96%. Based on this analysis, we selected 10 paired samples (10 baseline and 10 after six months n-3 PUFA treatment) with highest level of adherence as determined by pill counting. Those samples also showed the most consistent increase in EPA and DHA of the 12 participants in the 6 capsule/day dosing arm of the n-3 PUFA phase I clinical trial [4]. The variation in our DNAm data, which was assayed using the RRBS method, was lower than in the public data used for the power analysis (Supplemental Figure 4).

2.4. DNAm Data Generation, Processing, and Quantification

DNAm data was generated for all samples using the reduced representation bisulfite sequencing (RRBS) method [26]. Prior to library generation, DNA was extracted using QIAamp DNA mini kit from QIAGEN. Sequencing libraries were created using the NEXTflex Bisulfite Library Prep Kit for Illumina Sequencing from BIOO following the manufacturer’s protocol. All samples were sequenced using 100bp paired-end technology. Sequenced reads were first trimmed for both adapter sequences and low quality using TrimGalore v0.4.0 (a wrapper for Cutadapt) which removed any detected Illumina adapters or bases with Phred quality scores below 20 from the 3’ end of each read [27]. Next, reads were aligned using Bismark v0.17.0 to the human genome (version GRCh38) [28]. Bismark alignment was performed using Bowtie 2 and the minimum alignment scoring function set to “L,−0.6,−0.6”. Sequencing quality was assessed using FastQC v0.10.1, trimming reports, and alignment reports to ensure low adapter, low quality trimming rates, and acceptable unique alignment rates for each sample. Additional reads were sequenced until each sample had 30 million reads that passed the quality filter. Success of the bisulfite conversion process was assessed by determining the incomplete conversion rate (i.e., the percent of unconverted C’s that occurred outside of CpG sequence context): each sample was required to have less than 2% incomplete conversion. After passing our quality assessment, DNAm was quantified as the percent of CpG sites that are methylated in the sample using Bismark’s “bismark_methylation_extractor” function. Additional processing of DNAm quantification (e.g., filtering CpGs based on read depth, extracting coverage, generating coverage plots) was performed by custom scripts written in R v3.3.2 and Python v2.7.8.

2.5. Summarizing DNAm Genome-wide

Global DNAm levels were determined using the total number of unconverted (methylated) cytosines at CpG sites divided by the total coverage at CpG sites genome wide. Only CpGs with at least 5 reads coverage in at least 3 samples in both treatment groups were included in the global DNAm average. The treatment group averages for all CpGs that achieved the coverage and sample threshold were plotted to create experiment wide DNAm scatter plots. All plotted CpGs were fitted using orthogonal regression (ODR function from the SciPy library in Python) with the standard deviation used as the error for each CpG (SciPy.org). Orthogonal regression, as opposed to the more traditional linear regression, was used as it provides the most accurate estimation of the error in the case of our null hypothesis of no DNAm change from n-3 PUFA treatment (illustrated by the y=x line). Linear regression assumes that only one variable produces the observed error for each measurement whereas orthogonal regression assumes that both variables contribute to the observed error.

2.6. Genomic Feature Summaries

DNAm for annotated genomic features were summarized differently. First, all CpGs within a region that had at least five reads coverage were averaged to produce an average DNAm for each region. Next, the averages for all regions that comprise a genomic feature were again averaged into one DNAm value that was used to represent the DNAm for that genomic feature. Feature summaries were performed on the genomic features where DNAm has been implicated as playing a role in regulating gene expression. Therefore, CpG islands, two definitions of the promoter region, CpG islands located in the distal promoter, gene bodies, and the 1^st coding exon of each gene were interrogated for DNAm changes. The two promoter regions were defined to create a long-range and a short-range promoter (10kb upstream, 1kb downstream and 1kb upstream, 1kb downstream, respectively, of the transcriptional start site (TSS)).

2.7. Candidate Gene List Development

All GDAC Firehose analysis results and data were downloaded to access their 8,586 gene DNAm table (Broad Institute TCGA Genome Data Analysis Center (GDAC) 2012). The DNAm table provided data in the format of 450k Infinium array beta-values for each of the 8,586 most varying probes which were associated with gene symbols. Pathway membership for each gene was obtained by querying against Gene Ontology, KEGG, and Reactome using each site’s API [29–31]. Genes involved in inflammation related pathways were kept resulting in 89 inflammation related genes that are highly variable in breast cancer. An additional 63 genes were identified by performing literature searches in June of 2015 and again in May of 2016 using combinations of the main keywords in both PubMed and Google Scholar: n-3 PUFA, fatty acids, n-3 PUFA, inflammation, DNA methylation, PBMC, cancer, breast cancer, prevention, biomarker. These 63 genes and their references are provided in Supplemental Table 2. The genes that resulted from the literature search were all reported to play a role in or have been affected (DNAm or expression levels) by fatty acid metabolism, fatty acid related inflammation, or breast cancer inflammation. The resulting 152 genes (Supplemental Table 2) are the candidate genes.

Unsupervised clustering was performed using the same non-negative matrix factorization (NMF) method used in the original GDAC analysis, but was implemented in R using the NMF package (v0.20.6) [32] for analysis convenience (Supplemental Figure 1A). Analysis in R was successful in reproducing sample membership of clusters reported by GDAC. To perform NMF on candidate genes, the full 450k Infinium Array data was downloaded from TCGA for the entire breast cancer cohort so that DNAm beta values could be obtained. Metadata (sample and experiment numbers as well as clinical classification) were downloaded so that sample IDs could be converted and matched. Using only the samples included in the original GDAC analysis and selecting the most varying probe for each candidate gene, the 152 candidate gene beta-value table was constructed. NMF was performed and quality measures calculated for 2-10 clusters. As in the GDAC analysis, the quality measures cophenetic similarity, dispersion, and silhouette coefficients all indicated that six clusters resulted in the best separation (Supplemental Figure 1C). Molecular markers ER, PR, and HER2, Triple Negative status, and PAM50 subtype, were all used to determine if the six resultant sample clusters corresponded to clinically relevant classifications. Both the original 8,586 GDAC clusters and the candidate gene clusters were annotated with clinical classification to observe associations. However, as with the original clustering, clustering did not stratify samples based on any single or combination of molecular markers included in the metadata. Although clustering could not be explained by molecular markers, we did observe that both the original GDAC cluster and our clustering showed a general separation between luminal and basal-like breast cancers. For both the GDAC gene list and our candidate gene list, the majority of the basal-like samples were present in a single cluster (Supplemental Figure 1 B&D). The stratification of these subtypes supports the observed differential expression in breast cancer marker genes that was also reported in the original GDAC analysis. Additionally, the fact that our candidate gene list performs similarly implies that our greatly reduced gene list is also able to discern differences in the genes responsible for resulting in different breast cancer subtypes.

2.8. DMC Analysis and Association with Gene Promoters

DMCs were determined between treatment groups using MethylKit v1.2.0 [33]. All CpGs with 10 reads coverage in at least 16 samples were input into MethylKit resulting in a total of 962,093 CpGs being tested (Figure 2A). A DMC had a SLIM multiple test corrected p-value less than 0.05 and at least a 5% average DNAm change between treatment groups. DMCs in candidate genes were found by intersecting all DMCs with the promoters and promoter associated CpG islands of the candidate genes. Promoter DMCs were identified as any DMC that was in a region 10k base-pair (bp) upstream and 1k bp downstream of transcriptional start sites of RefSeq genes (hg38). The rationale for selection of this longer distance promotor (instead of 1k up- and 1k down-stream) as the promoter region is: 1) the longer distance includes more distal regulatory features such as promoter associated CpG islands and distal transcription factors and 2) the restriction enzyme used for RRBS preferentially interrogates CG rich regions so expanding the promoter definition to include CpG islands allowed us to interrogate more genes [34,35]. To ensure that DMCs were associated with n-3 PUFA treatment effect and not explained by any confounders, we performed linear regression and correlation analysis for each DMC. Using linear regression, DNAm at DMCs of all samples were compared to both the fatty acid levels and estimated cell type composition as determined by methylCC [36]; DMCs where the resulting fit had an FDR p-value ≤ 0.1 with any confounding variable were removed. The maximum R² of the DMCs that passed this filtering process and were used in subsequent analyses was 0.58. Additionally, we used a paired t-test to test for changes in the estimated fraction after n-3 PUFA treatment for each cell type present (i.e. granulocyte, CD4+, CD8+, B cells, monocytes, natural killer cells); no significant changes were detected (data not shown).

Figure 2: Differential DNAm in a set of Candidate Genes after n-3 PUFA Treatment.

A) The flow chart depicts the processing involved in identifying differentially methylated CpGs (DMCs). All CpGs with at least 5 reads coverage in 80% of the samples (16/20 samples) were included in the analysis (962,093 CpGs). DMC analysis was performed using MethylKit. CpGs that had a multiple test corrected p-value (q-value) less than 0.05 and a change in DNAm (ΔDNAm) > 5% were considered DMCs. Candidate gene promoter DMCs were identified as those that overlap the genomic coordinates of the candidate gene promoters (defined as 10kb up-, 1kb downstream of the TSS). B) A flow chart for the development of the candidate gene list. The 152 candidate genes were based on either 1) genes with highly variable DNAm in the TCGA breast cancer cohort that involve inflammation or 2) literature review for inflammation related DNAm processes, DNAm and fatty acids or breast cancer development. C) The candidate genes (left bar) were enriched in DMCs when compared to all genes (right bar; hypergeometric test p-value = 0.026). D) Genome-wide promoters and candidate gene promoters were enriched in hypermethylation after n-3 PUFA treatment. The hypergeometric test was used to compare the ratio of hyper-to hypo-methylated genes in all tested gene promoters (left bar) to both the genome-wide promoter DMCs (** middle bar; p-value < 0.001) and the candidate gene promoters (* right bar; p-value = 0.018).

2.9. Enrichment of Pathways for Directional DNAm Change

Using Enrichr [37], candidate gene promoter DMCs were used to determine which inflammation-related WikiPathways were enriched in n-3 PUFA treatment DNAm changes [38,39]. Directional enrichment was determined using the genome-wide promoter associated DMCs instead of the candidate gene promoter DMCs that were used for general enrichment [40]. The DNAm change from n-3 PUFA treatment was determined for each gene in each pathway using all DMCs in the promoter region. We tested for an over-representation of either hyper-or hypomethylated DMCs in each identified pathway that included at least 8 DMCs. Only genes that had a DMC in the promoter region were used to determine directional DNAm enrichment. Because the genome-wide DMCs had a strong trend toward hypermethylation (7,129 hyper-vs 3,174 hypomethylated DMCs), the hypergeometric test was used to determine if the ratio of hyper- to hypomethylated DMCs in each pathway was significant given the genome-wide DMC background. Pathways were considered to be significantly enriched if the hypergeometric test produced a q-value < 0.05 (FDR corrected p-value).

2.10. Correlating DNAm with Fatty Acids, Serum Biomarkers, and Biometric Data

For each DNAm measure and each patient data type (i.e., total fatty acid, serum biomarkers, etc.), DNAm was correlated with the patient data using linear regression. The resulting linear model was assessed by producing an R² value and multiple test corrected (FDR) p-value to control for testing multiple DNAm measures and the multiple parameters included in the data type. Using the total fatty acids data type as an example, all total fatty acid measures (n-3 PUFA, n-6 PUFA, monounsaturated, saturated) were compared to DNAm of all six genomic features and the resulting p-values were corrected for 24 comparisons. These correlations were performed for the baseline values, the change after n-3 PUFA treatment and the baseline vs n-3 PUFA change. This analysis used DNAm measures of global DNAm, genomic feature summaries, and candidate gene promoter DMCs. Each DNAm measure was compared to the patient data for total fatty acids, individual fatty acids, serum inflammation markers (Adiponectin, Leptin, high-sensitivity C-reactive protein, IL-6, and TNFα), and biometric data (age, weight, BMI, waist, waist hip circumference). Additionally, we used the anthropometric and biomarker data as interaction terms in our regression models; none of these terms improved the relationship between DNAm and fatty acid measures. However, this pilot study was not powered or designed to explore the relationship between DNAm and fatty acid levels.

3. Results

3.1. n-3 PUFA treatment does not alter average genome-wide DNA methylation patterns

We previously reported the effects of four different doses of n-3 PUFAs on fatty acid profiles of women at high risk of breast cancer following six months of treatment [4]. This PBMC DNAm substudy focused on 10 participants from the 6 capsule/day arm with highest increases in DHA and EPA at 6 months. Analysis of serum and breast adipose fatty acid profiles of these samples yielded similar results to those previously reported [4]. The 10 women had an average age of 51 ± 6.8 years, weight of 74.4 ± 11.3 kg, BMI of 27.1 ± 5.0, waist of 84.7 ± 9.3 cm, and waist hip circumference ratio of 0.80 ± 0.06 (mean ± SD; see Table 1). Averaged pre-and post-treatment serum fatty acid profiles showed changes in the overall distribution of fatty acids, with significant increases in total n-3 PUFAs and decreases in total monounsaturated fatty acids and n-6 PUFAs after 6 months of study treatment (Figure 1A; Supplemental Table 1). Fatty acid profiles in adipose tissue also showed significantly higher n-3 PUFA content at six months compared to baseline (Figure 1A).

Figure 1: Effects of n-3 PUFA treatment on fatty acid profiles and averaged DNAm.

A) Average fatty acid content before (left; n=10) and after (right; n=10) 6 months of n-3 PUFA treatment. The fatty acid composition was measured in both serum (top) and adipose tissue (bottom) as previously detailed [4]. Asterisks indicate statistically significant changes from n-3 PUFA treatment (FDR p-value < 0.05). B) Global DNAm represented by all CpGs that had 5 reads in at least 3 samples. The solid gray line is the y=x line which indicates no change in average DNAm. The blue to yellow color scale indicates decreasing to increasing (respectively) number of CpGs represented at that point. Data were fitted using orthogonal regression (red dashed line) and standard regression (black dashed line). The fits showed no global effect from n-3 PUFA treatment. C) DNAm averaged at 0 and 6 months for six genomic features. Genes are RefSeq GRCh38 annotation, and genomic features are defined as follows: 1) CpG Islands (CGI) – UCSC genome browser CpG island track; 2) Promoter associated CGI–CGI that fall within 10kb up-/1kb downstream from gene transcriptional start sites (TSS); 3) Promoter – region defined as 10kb up-/1kb downstream from TSS; 4) Proximal promoter – promoter region defined as 1kb up-/1kb downstream of TSS; 5) first exon; 6) Gene body – entire gene coding region for all transcripts. There were no significant treatment effects (mean ± standard error) in these summary features by averaged DNAm.

To assess whether n-3 PUFA treatment which led to increased serum and adipose EPA and DHA could result in a global remodeling of DNAm patterns, we calculated the effect on average DNAm before and after treatment (n=10 matched pairs). For a genome-wide view of DNAm, the average DNAm values of all pass-filter CpGs (>=5 reads coverage in at least 80% of samples) were plotted and the line of best fit was used to determine whether n-3 PUFAs caused a global change in DNAm (see Methods). The resulting fit, with an R² of 0.999, indicates that DNAm averaged before and after treatment is unchanged at most CpG sites (Figure 1B). The average DNAm change genome-wide was 0.7% (50.5% pre-and 51.2% post-treatment, t-test p-value=not significant). We also determined the effect of dietary n-3 PUFAs on DNAm of specific regions of the genome by evaluating the change in DNAm in six annotated genomic features that are associated with gene regulation (e.g., gene promoters and CpG Islands). In these selected genomic features, average DNAm did not change after n-3 PUFA treatment (Figure 1C).

3.2. n-3 PUFAs resulted in promoter hypermethylation as assessed by genome-wide and candidate gene CpG-level analyses

To determine how n-3 PUFA effects on DNAm could potentially regulate gene expression, we investigated the average DNAm effects at individual CpG loci. A genome-wide differentially methylated CpG (DMC) analysis was conducted with MethylKit to compare the DNAm between 0 and 6 months [33]. All CpGs in the genome with sufficient read coverage were tested for changes in the average DNAm (see Methods). To limit the detected DNAm changes to only those caused by the n-3 PUFA treatment, DMCs with DNAm changes significantly correlated with estimated cell type were removed (180 DMCs removed; see Methods) [36]. This analysis at CpG loci identified a total of 30,793 DMCs following n-3 PUFA treatment (DNAm change >= 5% and multiple test corrected p-value<0.05; Figure 2A). To reduce the number of DMCs and focus on those more likely to regulate gene expression [41,42], the DMCs in the promoter region (10kb up- and 1kb down-stream from the TSS) were examined. However, this filtering still resulted in 10,261 DMCs located in 5,491 gene promoters.

Instead of performing a genome-wide analysis to investigate all DNAm-mediate effects from n-3 PUFA treatment, we focused the analysis based on our hypothesis that n-3 PUFA treatment impacts inflammatory and breast cancer pathways. We used the Cancer Genome Atlas (TCGA) breast cancer data set to identify regions where DNAm has the potential to change during breast tumorigenesis. Before conducting the analysis, we developed a set of candidate genes based on two criteria: 1) genes involved in inflammation that show high variation among the TCGA breast cancer and normal samples (89 genes; GDAC Analysis 2012; see Methods for more details) and 2) genes associated with breast cancer (36 genes) and fatty acid induced effects (33 genes) (Figure 2B; Supplemental Table 2). Using the resulting candidate genes (Supplemental Table 2), we demonstrated a similar stratification of the TCGA data when comparing the original 8,562 genes to our reduced set of 152 candidate genes (Supplemental Figure 1). From the genome-wide promoter DMCs, candidate gene DMCs were identified by selecting only those DMCs located in a candidate gene promoter region (Figure 2A; Supplemental Table 3). This approach both focused the analysis to our hypothesis and used a more stringent genome-wide multiple test correction to reduce false positives. The full genome-wide promoter DMCs were used later in our analysis to predict the effects of n-3 PUFA treatment.

From the 152 candidate genes, we identified 49 candidate genes that contained at least one DMC (compared to 5491 genome-wide gene promoters). Comparing the proportion of candidate gene DMCs to genome-wide DMCs, the candidate genes are significantly enriched in n-3 PUFA associated DMCs (hypergeometric test p-value = 0.026; Figure 2C-D). Both the genome-wide and candidate gene promoters are significantly enriched in hypermethylation after n-3 PUFA treatment, with 68% hypermethylated genome-wide promoters and 65% hypermethylated candidate gene promoters compared to 52% hypermethylation in all gene promoters (Figure 2D).

3.3. Pathway analysis identifies potential epigenetic mechanisms of n-3 PUFAs

To focus on those pathways of interest, we then used the candidate gene promoter DMCs to perform pathway analysis. We identified 33 pathways enriched for candidate gene promoter DMCs (adjusted p-value < 0.05; Supplemental Table 4). To determine which pathways were more likely to be up- or down-regulated by the n-3 PUFA-induced DNAm changes, we utilized the genome-wide promoter DMCs. The DNAm change in each gene promoter DMC was used to identify the pathways where n-3 PUFA treatment caused concerted treatment effects (all DMCs showing either hyper- or hypomethylation). Enrichment in directional DNAm was determined using the hypergeometric test in all pathways with at least 8 genome-wide promoter DMCs (minimum number of DMCs to achieve hypergeometric p-value < 0.05) [43]. This test yielded three pathways significantly enriched in n-3 PUFA associated DNA hypermethylation: the Focal Adhesion, Toll-like Receptor (TLR), and Leptin signaling pathways (Figure 3A). Of these pathways, the Focal Adhesion and TLR pathways had both enrichment for DNA hypermethylation and biologically relevant localization of DMCs (Supplemental Figures 2–3). By comparing the DMCs detected in these pathways (Figure 3B) and the genes that comprise each pathway (Figure 3C), we noted only two shared DMCs and two shared signal transduction pathways. The localization of DMCs within each pathway has potential biological relevance because of the canonical relationship between increased DNAm in gene promoters and gene silencing. Notably, the hypermethylated DMCs in the Focal Adhesion pathway are located downstream of cytokine, chemokine, and hormone signaling (Supplemental Figure 2). The TLR pathway has hypermethylated DMCs downstream of lipopolysaccharide (LPS) receptors and upstream of pro-inflammatory cytokines (Supplemental Figure 3). Taken together, these pathways represent two distinct mechanisms by which n-3 PUFAs may modulate inflammation through DNAm changes.

Figure 3: Directional Enrichment of DNAm Changes in Candidate Gene Associated Pathways.

Inflammation-related pathways were determined using DMCs located in candidate gene promoter regions. A) The hypergeometric test was used to compare the number of hyper-or hypomethylated DMCs observed in each pathway to the distribution of DMCs in promoter regions genome-wide (7,129 hyper-and 3,174 hypomethylated DMCs). Asterisks indicate hypergeometric FDR p-value <0.05. Daggers indicate that hypermethylated DMCs are localized in biologically relevant region(s) of each pathway. Of the pathways that were tested for directional enrichment, three had significant overrepresentation of hypermethylated DMCs (FDR p-value < 0.05; indicated by stars). Two pathways were both enriched in DMCs and biologically relevant localization of DMCs. B) Comparison of the promoter DMCs used to determine enrichment in the Focal Adhesion and TLR pathways shows the shared genes. C) Comparison of all genes included in the Focal Adhesion and TLR pathways shows an overlap of 21 genes. The shared genes represent two shared signaling cascades in both pathways.

3.4. Exploratory analyses of variability in PBMC DNAm after n-3 PUFA Treatment

Although average DNAm did not change following n-3 PUFA treatment in any of the six genomic regions analyzed (Figure 1C), we observed highly variable DNAm patterns when comparing DNAm change in the genomic features for each sample pair from 0 to 6 months (Figure 4A). Similarly, n-3 PUFAs did not affect average global DNAm (Figure 1B), but the change per individual was highly heterogeneous (Figure 4B; mean = 0.7%; minimum= −10.5%; maximum = 12.1%). The individual fatty acid profiles did not show such variability among the participants (Figure 4C).

Figure 4: n-3 PUFA Treatment Effects show High Variability Between Samples.

A) n-3 PUFA treatment produces highly variable effects on genome-wide DNAm between participants in gene associated features. Participants are labeled 1 through 10. The DNAm was averaged for each of the genomic features for each participant at 0 and 6 months of n-3 PUFA treatment. The change in DNAm from n-3 PUFA treatment (treated minus untreated) per individual is shown. B) The change in global DNAm also showed variable treatment effects between women. The average DNAm values from all CpGs with at least 5 reads coverage were used to determine global DNAm for each participant at 0 and 6 months. The change in DNAm from n-3 PUFA treatment (treated minus untreated) per individual is shown. C) Following n-3 PUFA treatment, the serum fatty acid profiles of all participants showed increases in n-3 PUFAs and decreases in monounsaturated fatty acids. Following n-3 PUFA treatment, participants differed in directional change of saturated fatty acids and n-6 PUFAs. For each fatty acid measure, the change after n-3 PUFA treatment is shown (treated minus untreated) for each sample pair. D-E) Variable n-3 PUFA treatment effects at the CpG level were assessed in the Focal Adhesion and TLR pathways. For each study participant, the number of hyper- and hypomethylated CpGs were counted for each of the DMCs detected in the two pathways. CpGs were counted only when the magnitude of the DNAm change was greater than 5% so that the variability was not overrepresented by small changes. Based on the average effects, more hypermethylated CpGs would be expected; however, several participants had PBMCs with a higher number of hypomethylated CpGs. In four participants, PBMCs showed relatively more CpG hypomethylation than hypermethylation in at least one pathway at 6 months of n-3 PUFAs; for participants 2 and 3, there was greater CpG hypomethylation in both pathways.

To investigate the basis for the observed individual vs average differences in DNAm in the 10 participants, we tested for correlations of DNAm with possible confounders such as BMI, age, waist, serum biomarkers (complete list in Methods section 2.2), fatty acid profiles (EPA, DHA, n-3 and n-6 PUFA, poly-and mono-unsaturated fatty acids), and estimated cell type composition (natural killer, monocytes, B-cells, CD8+ T-cells, CD4+ T-cells, granulocytes) using both baseline values and change from treatment. The analyses did not reveal significant correlations between potential confounders and DNAm (global or genomic features; see Methods for further details; data not shown).

Given the observed variability in global DNAm in PBMC between participants, we assessed whether the Focal Adhesion and TLR pathways, as identified by pathway analysis of statistically significant differences in average DNAm, also had variable DNAm change between participants at the CpG level. Some women showed PBMC DNAm changes in the opposite direction when compared to the average. To illustrate this variability, we counted the number of hypo-vs hyper-methylated CpGs for each study participant for all DMCs in each of the pathways; this showed that some participants had more hypomethylated CpGs (Figure 4D–E; CpGs with DNAm change > 5%).

4. Discussion

With this study, we demonstrate that n-3 PUFA supplementation in women at high risk of breast cancer elicits locus-specific DNAm changes in PBMCs. While average genome-wide CpG methylation is essentially unchanged in PBMCs after six months of n-3 PUFA supplementation, analysis of DNAm changes at CpGs (DMCs) revealed marked enrichment of hypermethylation in the promoter regions of candidate genes selected for potential for epigenetic regulation of inflammatory and breast carcinogenesis signaling cascades. Using the DNAm changes in gene promoters of the candidate genes, pathway analysis identified at least two pathways that might in part mediate n-3 PUFA anti-inflammatory effects.

Although many studies have investigated the effects of n-3 PUFA treatment on DNAm [18,19,44–49], most studies in humans have focused on a limited number of select genes or regions [21,22,50–52]. Few studies address genome-wide DNAm changes from n-3 PUFAs [23,53]. In comparison to Tremblay et al. [23]. notable differences with our study cohort and study design include the use of a lower EPA+DHA dose (3g/day vs 5g/day), shorter treatment duration (6 weeks vs 6 months), and a study population of overweight or obese men and women [4]. In another genome-wide approach to analysis of DNAm changes from n-3 PUFAs, Aslibeykan et al. conducted analyses of cross-sectional samples at the top and bottom deciles of n-3 PUFA intake to represent traditional Yup’ik versus modern U.S. dietary patterns, respectively [20]. As both of these prior investigations used the Infinium Human Methylation 450k array, our study is the first to use a genome-wide bisulfite sequencing assay of DNAm to identify DNAm changes from an n-3 PUFA intervention. Our study, when combined with these previous reports, helps build consensus that n-3 PUFA treatment modulates DNAm and works toward developing a better understanding of n-3 PUFA effects.

Our exploratory analyses of the effect of n-3 PUFA treatment on each sample pair led to observations of variable and even opposing DNAm effects from n-3 PUFA treatment at every scale in the genome: global DNAm (Figure 4B), gene associated genomic features (Figure 4A), and individual CpGs (Figure 4D-E). We attempted to explain this variability using linear regression methods; however, in this limited study, variability in DNAm patterns did not correlate with demographic or anthropometric features with known effects on DNAm such as age, menopausal status, and central obesity [54,55] Prior studies have noted individual variability in response to dietary n-3 PUFAs that relates to baseline parameters (e.g. fatty acid levels, BMI) [4] or genotype (e.g. variants in COX1, IL-6, FADS1, and FADS2) [50,56–58], with potential for differential treatment effects. With the lack of consensus regarding clinical use of n-3 PUFAs, as for depression, cardiovascular disease, and cancer [59–62], greater understanding of predictors of individual response to n-3 PUFA treatment is needed. Based on our observations of variable n-3 PUFA effects, PBMC DNAm profile might offer a method to measure an individual’s response to n-3 PUFAs. This observation of variability in individual DNAm response warrants further investigation in future n-3 PUFA intervention studies.

A number of studies have investigated n-3 PUFAs and have reported inflammation and fatty acid metabolism effects that support our findings in general [7–10,12,13,52]. Additionally, the changes in DNAm we observed in specific genes and pathways are also supported by prior studies. We observed both direct methylation of pro-inflammatory genes and hypermethylation of multiple mediators in pro-inflammatory pathways that have previously been reported to show changes in gene expression. The pro-inflammatory genes IRAK1, CXCL16, ALOX5, MAN2A1, HSD17B11 were hypermethylated in our study cohort, which corroborates decreases reported in previous n-3 PUFA studies [11,13]. We also observed hypermethylation upstream of the pro-inflammatory genes TNF, IL1B, IL6, IL12A which corroborate the findings of others [17,63,64]. Bowens et al. reported significant downregulation of the TLR and the MAP kinase signaling pathways, which we also identified as significantly enriched in hypermethylated DMCs [11]. In one of the most comprehensive reports of n-3 PUFA’s effects on DNAm, Tremblay et al. used the 450k Infinium array to demonstrate that 6 weeks of n-3 PUFA supplementation led to hypermethylated CpGs and changes in pathways that included the inflammation-related related TLR and PI3K pathways [23]. Thus, our analyses of DNAm changes following n-3 PUFA supplementation support the accumulating evidence that dietary fatty acid patterns impact both DNAm and gene expression. However, our pilot study was not meant to identify a mechanism of n-3 PUFA effects; larger confirmational studies that include gene expression studies are still needed.

In this pilot study, we developed a set of candidate genes based on previously reported n-3 PUFA effects (see Supplemental Table 2) so as to focus the investigation of DNAm changes at CpGs from a genome-wide assay. The RRBS experiment uses bisulfite conversion which is the gold standard for DNAm assays. Our focused analysis on a hypothesis-derived set of candidate genes was conducted with a genome-wide statistical cutoff which minimizes false positives while still focusing on inflammation. As the first opportunity to test the capability of our candidate genes to detect n-3 PUFA treatment effects, this study showed that the candidate genes were enriched in DNAm differences after n-3 PUFA treatment. Despite the relatively small sample size used in our pilot study, our six-month n-3 PUFA treatment of 5 g/day EPA+DHA detected a large number (30,793) of DNAm changes at CpGs [23]. In order to understand how the large number of observed n-3 PUFA specific DNAm changes involved inflammation signaling, we focused on pathways that involved our candidate genes and featured directionally concordant DNAm changes [40]. This approach identified at least two hypermethylated pathways involved in regulating inflammation with biologically relevant localization of DMCs with hypermethylation observed after n-3 PUFA treatment consistent with decreased inflammation or inflammatory potential: Focal Adhesion and TLR pathways. The enrichment for hypermethylation, and the lack of any enrichment for hypomethylation, was an intriguing result and contrary to our hypothesis (e.g., hypomethylation of anti-inflammatory genes/pathways). The TLR pathway is interesting both for effects on immune response and production of pro-inflammatory cytokines; additionally, n-3 PUFAs have been reported to modulate TLR signaling [63–66]. Likewise, previous research has implicated n-3 PUFA modulation of gene expression in the focal adhesion pathway [11] and of DNAm in the PI3K signaling component of the focal adhesion pathway [23]. Such concordance across studies of various designs and exposures lends support to the accumulating evidence that n-3 PUFAs impact these processes.

The limitations of this pilot study primarily relate to the small sample size. The sample size limited our ability to investigate the unexpected observation of inter-patient variability. However, we were able to detect DNAm changes at CpG loci which were supported by other n-3 PUFA intervention studies that have used similar sample sizes to detect DNAm changes from n-3 PUFA [21,23,53]. A larger study population is needed to explore the variability of DNAm changes with n-3 PUFA treatment as well as the interplay between DNAm and biological processes. Another limitation is the lack of a control or lower dose group which would have provided a useful comparison; nonetheless, this pilot effort with limited resources provides guidance for a larger study of the impact of dietary n-3 PUFA intake on PBMC DNAm. Additionally, we acknowledge the potential limitations of focusing on a predetermined set of candidate genes, combined with directional enrichment of DNAm in selection of pathways; however, this analytic approach enabled identification of two inflammation-related signaling networks that have also been reported by others as responsive to n-3 PUFAs. Finally, changes in DNAm may not indicate functional regulation of gene expression even when occurring in the promoter region [67]. We chose a broad definition of the promoter region and looked for concerted DNAm changes in pathways to increase the probability of identifying pathways under epigenetic regulation. Future studies will address the functional relationship between DNAm and gene expression and biologic response of n-3 PUFA treatment.

Taken together, our pilot study demonstrates changes in PBMC DNAm following a six-month, high dose regimen of n-3 PUFAs in women at high risk of breast cancer by assaying DNAm genome-wide in PBMCs. While changes in average global DNAm following n-3 PUFA treatment were not evident in this study of limited power, our analyses of DMCs identified hypermethylation of key pro-inflammatory signal transduction genes in the PI3K and TLR signaling pathways. Further research is warranted to evaluate the potential for individual responses in PBMC DNAm and the implications for monitoring and tailoring n-3 PUFA intake to promote optimal immune function for health outcomes.

Highlights.

• In women at high risk of breast cancer, dietary n-3 PUFA treatment for six months produces DNAm changes in human PBMCs.

• Genome wide methylation profiling of PBMCs revealed n-3 PUFA treatmentassociated DNAm changes in 24,842 CpG sites in 5,507 gene promoters.

• N-3 PUFA associated DNAm changes included hypermethylation of inflammation-related pathways including the Focal Adhesion and TLR pathways.

• Further research is warranted to evaluate the potential for individual responses in PBMC DNAm and the implications for monitoring and tailoring n-3 PUFA intake to optimize health outcomes.