Associations between dietary patterns and gene expression profiles of healthy men and women: a cross-sectional study

Two hundred and fifty four participants were recruited between September 2009 and December 2011 from the greater Quebec City metropolitan area through advertisements in local news as well as by electronic messages sent to university students/employees. Women who were pregnant or breastfeeding were excluded. To be eligible, participants had to be between 18 to 50 years of age, non-smokers and free of any thyroid or metabolic disorders requiring treatment, such as diabetes, hypertension, severe dyslipidemia and coronary heart disease requiring treatment. The body mass index (BMI) of the participants was between 25 and 40 kg/m². Subjects drinking regularly more than 2 drinks per day, taking omega-3 PUFA (n-3 PUFA) supplements 6 months prior to the study and other medication or supplement affecting lipid and lipoprotein metabolism were excluded. A total of 210 participants completed the protocol which is described elsewhere [22] and were included in this cross-sectional study. Subjects all provided written consent to participate into the study, which was approved by the ethics committees of Laval University Hospital Research Center and Laval University. This trial was registered at clinicaltrials.gov as NCT01343342.

Body weight, height, and waist circumference were measured according to the procedures recommended by the Airlie Conference [23]. BMI was calculated as weight per meter squared (kg/m²).

The morning after a 12-hour overnight fast and 48-h alcohol abstinence, blood samples were collected from an antecubital vein into vacutainer tubes containing EDTA. Blood samples were used to identify individuals with metabolic disorders, which were excluded. Plasma was separated by centrifugation (2500 × g for 10 minutes at 4°C), samples were aliquoted and frozen for subsequent analyses. Plasma total cholesterol (TC) and triglyceride concentrations were measured using enzymatic assays [24, 25]. Infranatant (d >1.006 g/ml) with heparin-manganese chloride was used to precipitate very low-density lipoprotein (VLDL) and low-density lipoprotein (LDL) and then determine high-density lipoprotein (HDL) cholesterol concentrations (HDL-C) [26]. The equation of Friedewald was used to estimate LDL-cholesterol (LDL-C) [27]. Non-HDL-C was calculated by subtracting HDL-C from TC. CRP was measured by nephelometry (Prospec equipment Behring) using a sensitive assay, as described previously [28]. Plasma apolipoprotein B-100 (apoB) concentrations were measured by the rocket immunoelectrophoretic method of Laurell, as previously described [29]. Glucose concentrations were determined enzymatically [30] and plasma insulin was measured by radioimmunoassay with polyethylene glycol separation [31].

Resting blood pressure measurements (three readings) were performed after a 10-min rest in a sitting position, phases I and V of Korotkoff sounds being respectively used for systolic (SBP) and diastolic (DBP) blood pressures [32].

Dietary intake of the past month was determined by a 91-items validated food frequency questionnaire (FFQ) [33] based on food habits of Quebecers, administered by a registered dietitian (RD). The RD asked participants how often they consumed each type of food: daily, weekly, monthly or none at all during the last month. To make sure each participant estimated correctly the portion eaten, examples of portion size were provided. Data obtained from FFQ were analysed using the Nutrition Data System for Research software version 2011, developed by the Nutrition Coordination Center (University of Minnesota, Minneapolis, MN). All the information was compiled and similar food items from the FFQ were grouped, as previously described [4]. Three criteria were used to form these groups: first, the similarity of nutrient profiles, second, the culinary usage of different types of food (similar to groups used in a previous study [11]) and third, the consideration of groups utilized in other studies to maintain consistency [34]. Some individual food items were classified separately when their composition differed considerably from other foods (for example, pizza or eggs) or when they represented a different dietary habit (for example, liquor, wine, beer and French fries). On this basis, thirty-seven food groups were formed as described by Paradis et al. [4]. Food items from only thirty-five food groups were consumed by the participants in the present study. From these thirty-five food groups, eight were not normally distributed even after logarithmic transformation and were excluded as well. Consequently, twenty-seven foods groups were used for factor analyses to generate dietary patterns.

Food patterns derivation methods have already been described in a previous study [4]. Briefly, variables with abnormal distribution where logarithmically transformed before further analyses. The FACTOR procedure from Statistical Analysis Software (SAS) was used to derive factors. To determine the number of factors to retain, components with eigenvalue > 1, values at Scree test and the interpretability were considered. Food groups with absolute factor loadings ≥ 0.30 were regarded as significant contributors to the pattern. The patterns (derived factors) were named according to the interpretation of the data and to previous literature [4]. Each participant was given a score for both dietary patterns. These scores were calculated from the sum of food groups multiplied by their respective factor loading with the SCORE procedure of SAS. These scores reflect the degree of each participant dietary intakes conforming to a dietary pattern. In order to form two groups for each dietary pattern, participants were divided according to their score. A score ≤ 0 was considered as «low» and a score > 0 was considered as «high».

For transcriptomics analyses, the first 30 individuals who completed the study were included (13 men and 17 pre-menopausal women). In human nutrition studies 5 to 10 individuals in each group appear to be sufficient to detect differently expressed genes [35]. The following methods have been described by Rudkowska et al. [22]. Briefly, peripheral blood mononuclear cells (PBMCs) were collected into an 8-ml Cell Preparation Tube (Becton Dickinson, Oakville, Ontario, Canada). Remarkable concordance (<80%) of gene expression profiles between PBMCs and different tissues including liver, kidney, stomach, spleen, prostate, lung, heart, colon and brain, has been reported [36]. Centrifugation at room temperature (1500g, 20 minutes) was executed to separate PBMCs. The RNeasy Plus Mini Kit (QIAGEN, Mississauga, Ontario, Canada) was used to extract total RNA according to the manufacturer’s protocol. Microarray analyses were performed after spectrophotometric quantification and verification of the total RNA quality on the Agilent 2100 Bioanalyser (Agilent Technologies, Palo Alto, CA, United States). None of the samples had RNA integrity number (RIN) values less than 8 [37]. Consequently, all samples were included in the microarray analysis. However, as described previously [22], one outlier was excluded due to abnormal hybridization results. Thus, further analyses were conducted with 29 participants. The Illumina TotalPrep RNA Amplification kit (Ambion, Austin, TX, United States) was used to amplify and label 200ng of total RNA. The quality of complimentary RNA (cRNA) was evaluated by capillary electrophoresis on Agilent 2100 Bioanalyzer. 37,804 genes were analysed via expression levels of 48,803 mRNA transcripts with the Human-6 v3 Expres-sion BeadChips^® (Illumina, San Diego, CA, United States). The McGill University and Génome Québec Innovation Center (Montreal, Quebec, Canada) performed hybridization according to the manufacturer’s instructions, as previously described [22]. Validation of the expression levels were assessed previously by polymerase chain reaction (PCR) [22].

Microarray data was analysed with Flexarray software [38]. The lumi Bioconductor package algorithm included in Flexarray software, was used to pre-process and normalize Illumina microarray data. Background correction was assessed using negative controls followed by log₂ to stabilize variance and quantile normalization. According to Shi et al. [39] the use of control probes during background correction minimize false discovery rate (FDR). Fold changes obtained with this method also relate more to fold changes observed with PCR validation tests [39]. Participants where then stratified according to sex and scores for Prudent and Western dietary patterns. To assess differences (separately for men and for women) in gene expression levels between high versus low scores for Prudent and Western dietary patterns, the Significant Analysis of Microarray (SAM) was performed. SAM is an adaptation of t-test for microarray data which assigns a score to each gene according to changes in gene expression relative to the standard deviation of repeated measurements [40]. SAM uses permutation of the repeated measurements to estimate FDR. Transcripts were declared differently expressed only when P < 0.05 and fold changes were either < 0.8 (down-regulated) or > 1.2 (up-regulated), as previously described [22].

Ingenuity Pathway Analysis (IPA) system (Ingenuity^® System, http://​www.​ingenuity.​com) was used to verify if differentially expressed genes were related more than expected by chance to networks, diseases and canonical pathways. IPA allows adding structure to the vast amount of data generated by microarrays. To begin, an input file containing fold changes and P values of all probe sets was uploaded into IPA system. From this file, dataset in Core Analysis was produced. General settings for IPA system as «Ingenuity^® Knowledge Base (genes)» and «considered only molecules and/or relationships for humans» were used. IPA calculates a P value based on the right-tailed Fisher’s exact test for each canonical pathway, which is a measure of the likelihood that the association of a data set with a pathway is due to random chance. A cutoff of 1.2 was set. Relevant pathways with P values smaller than 0.05, were taken into account. IPA suggests that canonical pathways with P values higher than 0.05 may also be biologically relevant. Furthermore, no direction is associated with pathways, in other words, they cannot be qualified as up- or down-regulated. Significantly relevant canonical pathways related to cancer, cardiovascular diseases, immune system and inflammation were considered. IPA also processes «Downstream Effects Analysis» which is based on expected causal effects between genes and functions. These causal effects are derived from literature and compiled in «Ingenuity^® Knowledge Base». «Downstream Effects Analysis» compares the direction of the differently expressed genes with expectations based on the literature and predicts for each function a direction change using the «regulation z-score algorithm». In other words, if the observed direction change is mostly consistent with a particular activation state (increase or decrease) then IPA make a directional prediction. The z-score algorithm is designed to reduce the chance that random data will generate significant predictions. Z-scores ≥ 2, indicate that the function is significantly increased and z-scores ≤ -2, indicate that the function is significantly decreased. IPA also calculates a P value with the Fisher’s Exact Test which represents the likelihood that the association between a set of genes from the experimental data set and a related function is due to random association.