Dietary patterns differently associate with inflammation and gut microbiota in overweight and obese subjects.

BACKGROUND: Associations between dietary patterns, metabolic and inflammatory markers and gut microbiota are yet to be elucidated.
OBJECTIVES: We aimed to characterize dietary patterns in overweight and obese subjects and evaluate the different dietary patterns in relation to metabolic and inflammatory variables as well as gut microbiota.
DESIGN: Dietary patterns, plasma and adipose tissue markers, and gut microbiota were evaluated in a group of 45 overweight and obese subjects (6 men and 39 women). A group of 14 lean subjects were also evaluated as a reference group.
RESULTS: Three clusters of dietary patterns were identified in overweight/obese subjects. Cluster 1 had the least healthy eating behavior (highest consumption of potatoes, confectionary and sugary drinks, and the lowest consumption of fruits that was associated also with low consumption of yogurt, and water). This dietary pattern was associated with the highest LDL cholesterol, plasma soluble CD14 (p = 0.01) a marker of systemic inflammation but the lowest accumulation of CD163+ macrophages with anti-inflammatory profile in adipose tissue (p = 0.05). Cluster 3 had the healthiest eating behavior (lower consumption of confectionary and sugary drinks, and highest consumption of fruits but also yogurts and soups). Subjects in this Cluster had the lowest inflammatory markers (sCD14) and the highest anti-inflammatory adipose tissue CD163+ macrophages. Dietary intakes, insulin sensitivity and some inflammatory markers (plasma IL6) in Cluster 3 were close to those of lean subjects. Cluster 2 was in-between clusters 1 and 3 in terms of healthfulness. The 7 gut microbiota groups measured by qPCR were similar across the clusters. However, the healthiest dietary cluster had the highest microbial gene richness, as evaluated by quantitative metagenomics.
CONCLUSION: A healthier dietary pattern was associated with lower inflammatory markers as well as greater gut microbiota richness in overweight and obese subjects. TRIAL REGISTRATION: ClinicalTrials.gov NCT01314690.

Introduction

Dietary pattern analysis is a useful way to consider the diet as a whole, rather than considering individual foods or nutrients. The analysis of dietary patterns provides an opportunity to investigate relationships between diet and health in nutritional epidemiology [1]–[3]. Some prospective studies have demonstrated a relationship between dietary patterns and weight changes [4]–[6], although not all studies are consistent [7].

Dietary patterns have also been associated with markers of systemic inflammation and risk of cardiovascular diseases [8]–[10]. A healthy dietary pattern (higher in fruits, vegetables, poultry, tea, fruit juices and whole grains) was found inversely related to systemic C-reactive protein (CRP), while a western dietary pattern was positively related to CRP in different populations [8]–[11]. A similar healthy eating pattern was also associated with reduced insulin resistance [12] and the risk of metabolic syndrome [10], [13]. The mechanisms behind these links still have to be elucidated. The gut microbiota is viewed as a pivotal actor linking “external” macro-environmental changes to the “internal” host biology particularly inflammation as well as metabolic and body weight homeostasis. Gut microbiota has been shown to be involved in the development of metabolic syndrome and low-grade inflammation associated with obesity via different mechanisms including lipopolysaccarides-Toll-like receptors/CD14 (LPS-TLRs/CD14) complex mostly in animal models [14]. To our knowledge, a limited number of studies have investigated the relationship between dietary patterns, gut microbiota, and host inflammatory levels in humans.

Recently, quantitative metagenomic approaches have provided the opportunity to study the gut microbiota in humans in more depth. Interestingly, a rapid adaptation of gut microbiota could be observed after 24 h of switching from a low-fat, plant polysaccharide-rich diet to a high fat high sugar “western” diet [15], whereas the diversity of species within the gut microbiota organized into identifiable clusters or enterotypes are correlated with long-term but not short term dietary patterns [16]. We found recently that when subjects were clustered in function of their bacterial gene count, low fecal bacterial gene richness was associated with impaired glucose homeostasis and higher low-grade inflammation during a weight loss dietary program in overweight/obese subjects [17]. Low gene richness might not only be linked with isolated foods, but rather with global dietary patterns, an aspect not yet explored in the later study. Therefore we aimed to explore in the same cohort of subjects at baseline before any dietary transition, the relationship between different dietary patterns, metabolic and inflammatory variables and gut microbiota evaluated by qPCR and next-generation sequencing methods. The results in the overweight/obese group were compared to a group of lean subjects.

Materials and Methods

The protocol for this trial and supporting CONSORT checklist are available as supporting information; see Checklist S1 and Protocol S1.

Subjects

Fifty overweight and obese (BMI: ≥25−<38 Kg/m2), but otherwise healthy subjects (8 males and 42 females), aged from 25 to 65 years, were recruited at the Human Nutrition Research Center (Pitié-Salpêtrière Hospital, Paris, France). Forty-five subjects (6 males and 39 females) completed the 7-day dietary records, bioclinical and fecal bacterial data and were included in the present analysis (Figure 1). Selection of subjects was based on the absence of inflammatory or infectious diseases, cancer or any history of gastrointestinal problems. Subjects with hepatic, renal or cardiac diseases were excluded. In addition, 17 normal weight healthy female (BMI: <25 >18 kg/m2) volunteers living in the same area as the obese subjects were recruited as a control group. Fourteen of these completed the dietary records and were included for reference purposes. No antibiotics were taken within 2-months prior to the start of the study. Subjects were enrolled between October 2008 and December 2009. The Ethical Committee of Hotel-Dieu Hospital, Paris, France, approved the protocol. All the subjects gave written informed consent. This clinical trial was registered before enrolment of participants in the EU Clinical trials Register under the identification number: 2008-001138-28, and in the French agency for the security of medications and health products - changed to ANSM) under the identification number: ID RCB 2008-A00406-49. It was also registered in the ClinicalTrials.gov under the identification number: NCT01314690, later due to some technical problems with the sponsor of the study. The authors confirm that all ongoing and related trials for this study are registered.

Figure 1

Consort Flowchart.

All selected participants underwent a series of tests following an overnight fast. Blood samples were taken to measure plasma glucose, insulin, lipids and some inflammatory markers. Body fat and fat-free mass distributions were measured using dual-energy X-ray absorptiometry (Hologic APEX, discovery W. (S/N 84030), version 3.0, Bedford, MA). Estimation of pancreatic β-cell function (HOMA B%) and insulin sensitivity (HOMA S%) were calculated using homeostasis model assessment: HOMA/CIGMA software.

Dietary assessment

Dietary intake was assessed using an unweighed dietary record for 7-days (unless specified otherwise). The records were self-completed and contained pre-defined spaces for the recording of each meal. Subjects were provided with instructions on how to complete the dietary records. On the visit day, the dietitian verified the information with the subject. All records were analyzed by using the computer software program PROFILE DOSSIER X029 (Audit Conseil en Informatique Médicale, Bourges, France), which has a food composition database initially made up of 400 food items representative of the French diet as described previously [18]. Nutrient intakes were generated for each subject based on the dietary data collected. Information on consumed foods could not be directly extrapolated from the program and were therefore coded manually into 26 food groups. Food groups were defined based on groups and sub-groups typically used in the reporting of dietary data in large surveys, such as the French Nutrition and Health Survey and program (Étude Nationale Nutrition Santé, ENNS) [19], [20]. Mean food and nutrient intakes in the overweight/obese subjects were calculated for each subject according to the 7-day diary (in 2 subjects only 6 days of data were available). The lean subjects completed a 3-day diary of the same format as those used for overweight and obese subjects. Habitual physical activity (Baecke questionnaire) was also evaluated.

Adipocyte morphology and immunohistochemical analysis

Subcutaneous abdominal adipose tissue samples were obtained by needle biopsy from the periumbilical area under local anesthesia (1% xylocaine) for measuring adipocyte diameter and immunohistochemical detection (HAM56, CD163). Sub-cutaneous adipose tissue samples were washed in physiological saline. A fresh aliquot was used to measure adipocyte diameter as described previously [21] by using Perfect Image software (Claravision. Orsay.France). Other aliquots were fixed overnight at 4°C in 4% paraformaldehyde, and then embedded in paraffin for immune histochemical detection. Antibody anti-HAM56 and anti-CD163 (DakoCytomation, Trappes, France) were used to target the macrophage cells in subcutaneous adipose tissue by the avidin-biotin-peroxidase method. The number of positively marked cells was counted by two blinded investigators as previously described [22]. We calculated the labeled cell number per field, and the adjusted number (percentage per 100 adipocytes).

Gut Microbiota

(qPCR and SOLiD methods).

Fecal samples were obtained on the morning of the study before breakfast after 12 hours of fasting. Whole fecal samples were self-collected in sterile boxes and stored at −20°C within 4 h. Samples were treated in the laboratory as 200-mg aliquots and stored at −80°C until further analysis. DNA were extracted using the method previously described [23].

Real time qPCR method

The procedure for qPCR was performed as previously described (23). Seven bacterial groups were detected: Clostridium leptum (C. leptum), Clostridium coccoides (C. coccoides), Bacteroides/Prevotella, Bifidobacterium, Lactobacillus/Leuconostoc/Pediococcus, Escherichia coli (E. coli) as well as Faecalibacterium prausnitzii (F. prausnitzii).

Metagenomic sequencing

Intestinal bacterial gene content was determined by high throughput ABI SOLiD sequencing technology of total fecal DNA as described previously by our group [17], an average of 76.5±36.5 (mean ± sd) million 35 base-long single reads were determined for each sample (a total of 393 Gb of sequence). By using corona_lite (v4.0r2.0), an average of 24.8±14.3 million reads per individuals were mapped on the reference catalog of 3.3 million genes [24]–[25] with a maximum of 3 mismatches. Reads mapping at multiple positions were discarded and an average of 14.2±8.1 million uniquely mapped reads per individuals were retained for estimating the abundance of each reference gene by using METEOR software [26]. Abundance of each gene in an individual was normalized with METEOR by dividing the number of reads that uniquely mapped to a gene by its nucleotide length. After that, normalized gene abundances were transformed in frequencies by dividing them with the total number of uniquely mapped reads for a given sample. The resulting set of gene frequencies, termed as microbial gene profile of an individual, was used for further analyses.

Identification of patients with low and high gene counts (LGC and HGC, respectively)

The two groups of patients were defined using the 480 000 gene threshold. Genes significantly different in groups of individuals were identified by Mann-Whitney tests using p-value threshold <0.0001. They were clustered by an abundance-based binning strategy, using the covariance of their gene frequency profiles among the individuals of the cohort, as described in the previous study [24]. Abundance of a given cluster in each individual was estimated as a mean abundance of 25 arbitrarily selected 'tracer' genes for each cluster; these values were close to those obtained by using all the genes of a cluster.

Biochemical measurements

Plasma glucose was measured by hexokinase method (ARCHITECT system, Abbott, Park, Illinois, USA). Plasma insulin was determined by chemiluminescence (ARCHITECT system, Abbott). Plasma triglyceride and free fatty acids (FFA) were measured with Biomérieux kits (Marcy l’Etoile. France), and total cholesterol, high density (HDL), and low density (LDL) lipoprotein cholesterol were measured with Labintest kits (Aix-en-Provence. France). High sensitive (hs) CRP was measured by immuno-nephelometry on an IMMAGE analyzer (Backman-Coulter, Villepinte, France)..Leptin and interleukin-6 (IL-6) were determined by using enzyme-linked immunosorbent assay (ELISA) kit (Quantikine, R&D Systems, Oxford, UK). The limit of detection for leptin was 7.8 pg/ml and the intra- and interassay variability was less than 3.5 and 5.5%, respectively. Adiponectin was determined by using ELISA (Bühlman, Basel, Switzerland). The limit of detection was 0.019 ng/ml and the intra- and interassay variability was less than 5.5%. Serum concentrations of 5 cytokines (eotaxin, vascular endothelial growth factor (VEGF), interferon gamma-induced protein 10 (IP10), monocyte chemotactic protein-1 (MCP-1) and macrophage inflammatory protein-1 beta (MIP-1b)) were simultaneously determined by multiplex biochip technology with the human cytokine multiplex panel kits (Bio-Rad, Hercules, CA. USA). A multiplex assay was performed according to the manufacturer’s instructions. Multi-analytic profiling was performed on the Luminex-200 system and the Xmap Platform (Luminex Corporation. Austin. TX. USA). Acquired fluorescence data were analyzed using Exponent software with standard curves obtained from serial dilutions of standard cytokine mixtures. Serum samples were diluted (1∶4). Plasma soluble CD14 (sCD14) was determined by using ELISA (Quantikine, R&D Systems, Oxford, UK). The limit of detection was 125 pg/ml and the intra- and interassay variability was less than 6.5 and 7.5%, respectively and plasma LPS was measured with the Limulus amebocyte lysate kinetic chromogenic methodology that measures the color intensity directly related to the endotoxin concentration in a sample, using Endosafe-MCS (Charles River laboratories, Lyon, France).

Statistical Analysis

Data were analyzed using SAS 9.2 and SAS Enterprise Guide 4.2 (SAS Institute Inc., Cary. NC. USA) and R version 2.13 Software (http://www.r-project.org). Homeostasis model assessment of insulin resistance (HOMA-IR), insulin sensitivity (HOMA-S%), and b-cell function (HOMA-B%) were calculated by HOMA Calculator from www.dtu.ox.ac.uk/homacalculator. All data are expressed as mean ± SEM. P values ≤0.05 were considered as significant unless specified otherwise. Bonferroni correction was used for the adjustment of the food and nutrient analysis.

Identification of dietary patterns

Cluster analysis was employed to derive dietary patterns from the data. Overweight and obese subjects were grouped based on similarities in eating behaviors. Prior to this analysis, all food categories were standardized to a mean of 0 and a standard deviation of 1 in order to ensure that quantities consumed were comparable across different categories. A comparison of two clustering methods for use with dietary data (Ward’s Agglomerative Hierarchical Clustering and k-means clustering method) was undertaken as described previously [27]. K-means was found to be the most appropriate method to use in the conditions of the present study. Group membership is determined by a series of steps. The algorithm starts by a set of selected seeds and an estimation of the centroid (the multidimensional equivalent of the mean). Each individual is assigned to the nearest centroid and temporary clusters are formed. The seeds are then replaced by the centroid for each group. The process is repeated until no further changes occur in the groups. The k-means function fits a user-specified number of centers (the parameter k), such as that the within-cluster sum of squares from these centers is minimized, based on the Euclidian distance or another chosen distance. Since the number of clusters must be established a prior, several solutions were compared with a varying number of clusters (from 1 to 4).

To visualize the food categories that significantly distinguish one pattern from another, a Canonical Discriminate Analysis was implemented through SAS (CANDISC procedure) in order to transform the food categories into two canonical axes that achieve maximum separation between the clusters. The canonical coefficients were used to assess the contributions of food categories to the separation by evaluating their signs and magnitude. Using canonical discriminate analysis for k-means, the 26 food categories were transformed into two canonical variables from the patterns. This approach permitted the visualization of the food categories that significantly distinguish one pattern from another in the overweight/obese subjects (Figure 2). Overweight and obese subjects were classified according to their general dietary pattern. Food or drink categories typically high in fat, sugar or salt and low in other nutrients and fibers were regarded as ‘less healthy’, whereas those typically low in fat, sugar or salt and higher in fiber, fruits and vegetables were regarded as ‘healthier’ [28], [29]. Higher or lower intakes (as appropriate) of such categories contributed to the assignment of a name to a particular group of subjects. The lean subjects were projected onto the clusters of the overweight and obese subjects to examine their distribution (Figure S1 in Supporting Information S1).

Figure 2

Canonical analysis: graphical representation of the food categories by cluster.

A graphical representation of the food categories that created the distinction between the clusters i.e. those which were strongly correlated with canonical axis (Can) or significantly different between clusters (KW test with Bonferroni correction). The can 1 axis separates Cluster1 from 2 or 3, the can 2 axis separates Cluster 2 from 1 or 3. If the food category is strongly correlated with the two canonical axes it separates the three clusters at the same time. Food categories shown in black characterise Cluster 1, in green characterise Cluster 2, and in red characterise Cluster 3. The projection of each food or drink category on each canonical axis represents the contribution of this category to the building of this canonical axis. Therefore, if a category has a high contribution to the first axis (e.g. fruit), it discriminates Cluster 1 from Cluster 2 or Cluster 3. The food categories with weak contribution (below 0.5 in the inner circle) are shown in blue. These categories do not contribute to the discrimination/characterization of the three clusters. The distance between the centre point of the figure and the food category represents the correlation to the canonical axis and therefore the contribution to the separation of clusters. Food categories close to the axis between Clusters 2 and 3 indicate that intakes are similar, as for yogurt. Food category names have been shorted in this figure for readability.

Links between dietary patterns and bioclinical variables including gut microbiota

Since the food and nutrient data were not normally distributed, Kruskal-Wallis tests were used to test for differences between the dietary clusters, while Wilcoxon rank-sum tests were used to test for individual differences between each set of 2 clusters in food and nutrient data. Considering the influence of age on clinical parameters, we performed stratified Kruskal-Wallis tests stratified by age groups (R package = coin) for analyzing the differences in clinical variables as well as gut microbiota between the dietary patterns. When Kruskal-Wallis tests were significant, differences between each set of 2 clusters were assessed with stratified post-hoc Nemenyi tests. The stratified tests for trend (R package = coin) were performed to confirm the trends of clinical markers across the dietary clusters (ordered in terms of healthiness starting with the least healthy).

Partial spearman tests were used to verify the link between gene counts/the 7 bacterial groups and dietary categories/bioclinical variables by adjusting for age. Links between dietary clusters and gene richness groups (low gene count/high gene count: LGC/HGC groups) were investigated using an age-stratified Cochran-Mantel-Haenszel test. Trends were assessed with Mantel-Extension test [30].

Canonical Correlation Analysis was used to maximize the correlation between the set of the food categories that significantly separate the clusters and a set of selected clinical parameters. Two pairs of canonical axes were determined; one pair was from the food categories and one pair from the clinical parameters. The canonical coefficients were used to assess the contributions of each food category and each clinical parameter to the correlation by evaluating their signs and magnitude. This approach permitted to visualize the association between the food categories that significantly distinguish one pattern from another and the selected clinical parameters.

Results

Overweight and obese subjects were divided into dietary clusters with the number of clusters based on the optimum results obtained using three statistical parameters and practical considerations given the interpretability of the nutritional data. The optimum number of clusters (k) was selected as three (k = 3). Food categories that significantly distinguish one pattern from another in the overweight and obese subjects are presented in Figure 2. The lean subjects were found to be unequally dispersed across the clusters (Figure S1 in Supporting Information S1). The three dietary clusters were then analyzed in relation to clinical and biological factors.

Subject general characteristics

Subject characteristics are shown in Table 1 and Table S1 in Supporting Information S1. In the clusters of overweight and obese subjects, no significant difference was detected in the proportion of women or men, however, mean age of subjects in Cluster 3 was higher (52.2±2.3 years), while mean age of subjects in Cluster 1 was lower (34.4±2.7 years) (p = 0.001 across the 3 clusters). Importantly, body weight and adiposity markers did not differ significantly between the 3 clusters (Table 1). Level of physical activity was also similar across the clusters. There were no differences in plasma glucose homeostasis including surrogate insulin resistance index (HOMA-IR) between the three clusters. A trend towards significance was seen in total- and LDL-cholesterol that were significantly different across the clusters after age adjustment, Cluster 1>Cluster 2> Cluster 3 (Figure 3). Results were significant after stratified tests by age for trend (p = 0.03 for total cholesterol and p = 0.01 for LDL-cholesterol) with the greatest difference seen between Clusters 1 and 3 (when comparing the medians Figure 3) but between cluster 1 and 2 (when comparing the mean values as in table 1).

Table 1

Subject Characteristics in the lean and the overweight/obese groups and in the 3 dietary clusters.

	Lean (n = 14 w)	All overweight/obese subjects (n = 45, 39w/6m)	Dietary pattern clusters in overweight/obese subjects			P value
			Cluster 1 n = 1412w/2m	Cluster 2 n = 1815w/3m	Cluster 3 n = 1312w/1m	Kruskal P*	Trend P**
Adiposity markers
Body weight (kg)	60.81±2.48	91.51±1.98	96.42±3.64	89.67±3.41	86.73±3.14	0.67	0.85
BMI (kg/m2)	22.62±0.58	33.2±0.55	33.71±0.98	33.55±0.87	32.91±1.30	0.83	0.54
Total Fat mass (%)	29.68±1.19	39.42±0.94	39.51±1.67	40.22±1.57	39.88±1.61	0.89	0.95
Waist circumference (cm)	75.20±1.80	106.11±1.41	108.68±2.50	104.28±2.45	106.00±2.89	0.30	0.42
Adipocyte diameter (µm)	53.06±1.2	108.8±1.09	108.39±1.83	108.56±2.18	108.71±1.87	0.91	0.81
Leptin (ng/ml)	20.68±3.24	50.68±3.16	57.50±6.14	49.28±4.86	47.22±5.46	0.97	0.84
Plasma glucose homeostasis and insulin sensitivity
Fasting glycaemia (mmol/l)	4.49±0.13	5.22±0.06	5.06±0.05	5.31±0.10	5.28±0.14	0.46	0.47
Fasting insulinemia (µU/ml)	6.27±0.65	8.93±0.62	9.69±1.53	8.69±0.70	7.29±0.72	0.63	0.77
HOMA-IR	0.79±0.09	1.17±0.08	1.25±0.20	1.14±0.09	0.96±0.09	0.64	0.74
Adiponectin (µg/ml)	9.36±1.09	14.13±0.9	15.28±1.69	14.30±1.25	14.40±1.86	0.61	0.37
Plasma lipid homeostasis
Total cholesterol (mmol/l)	4.39±0.2	5.32±0.12	5.52±0.26	5.29±0.13	5.29±0.28	0.09 †	0.03
HDL cholesterol (mmol/l)	1.57±0.1	1.41±0.05	1.37±0.10	1.48±0.08	1.35±0.09	0.87	0.87
LDL cholesterol (mmol/l)	2.51±0.18	3.3±0.11	3.62±0.22	3.22±0.10	3.26±0.26	0.05 ‡	0.01
Triglycerides (mmol/l)	0.68±0.07	1.31±0.12	1.19±0.18	1.28±0.17	1.45±0.29	0.61	0.46
Fasting FFA (mmol/l)	0.62±0.07	0.45±0.02	0.45±0.05	0.45±0.04	0.48±0.05	0.86	0.59
Inflammatory markers
hsCRP (mg/l)	1.45±0.41	3.79±0.44	4.07±0.99	3.64±0.70	3.31±0.67	0.94	0.93
IL-6 (pg/ml)	1.13±0.18	2.08±0.29	2.75±0.74	1.43±0.27	2.06±0.50	0.19	0.55
LPS (pg/ml)	1.57±0.23	2.19±0.21	2.13±0.24	2.00±0.35	2.79±0.55	0.42	0.54
sCD14 (pg/ml)	1358.73±78.92	1390.6±59.57	1607.9±99.1	1350.5±91.4	1263.1±135.5	0.05 †	0.01
Adipose tissue macrophages
HAM56	2.53±0.35	3.87±0.36	3.35±0.64	4.14±0.62	4.18±0.75	0.50	0.29
HAM56%	7.55±1.1	13.58±1.2	11.07±1.77	14.74±2.23	15.20±2.60	0.48	0.26
CD163	2.35±0.39	1.89±0.22	1.3±0.3	1.7±0.3	2.6±0.4	0.15	0.05
CD163%	6.66±1.18	6.67±0.72	4.6±0.9	6.3±1.12	8.7±1.2	0.17	0.07
Chemokines
MCP1 (pg/ml)	50.11±8.95	36±3.95	46.55±7.07	34.79±6.99	28.39±7.77	0.41	0.24
VEGF (pg/ml)	82.89±18.74	88.07±13.1	68.10±18.06	102.99±19.69	57.98±14.50	0.31	0.91
Eotaxin (pg/ml)	93.11±17.99	62.95±4.63	55.39±9.59	71.58±8.55	65.56±6.43	0.36	0.38
IP10 (µg/ml)	328.6±32.66	604.09±48.23	635.72±75.40	486.56±64.85	707.05±123.38	0.16	0.93
MIP-1b (pg/ml)	133.4±23.8	134.09±10.8	159.46±27.89	111.26±12.98	119.61±14.12	0.34	0.36

Data are presented as means ± SEM; n = 45 subjects. BMI: Body Mass Index; *Kruskal-Wallis rank sum test stratified by age groups. **tests for trend stratified by age. P value ≤0.05 is shown in bold italics. 0.05

Figure 3

Differences of metabolic and inflammatory markers after stratified Kruskal-Wallis tests in the 3 dietary pattern clusters.

Black, grey and white columns represent the median values of the parameters in Cluster 1, Cluster 2 and Cluster 3, respectively, after age adjustment (see Methods S1 in Supporting Information S1). *: significant differences (p≤0.05) between the 3 clusters after stratified Kruskal-Wallis tests, #: a tendency of differences (0.05

Data are presented as means ± SEM; n = 45 subjects. BMI: Body Mass Index; *Kruskal-Wallis rank sum test stratified by age groups. **tests for trend stratified by age. P value ≤0.05 is shown in bold italics. 0.05