Effects of changing from a diet with saturated fat to a diet with n-6 polyunsaturated fat on the serum metabolome in relation to cardiovascular disease risk factors

An 8-week, double-blind, randomized, controlled dietary intervention study was conducted at the Oslo and Akershus University College of Applied Sciences and the University of Oslo, Norway, from July 2012 until April 2014. Details of the study design and description of participants have been described previously [15]. In short, 99 healthy adults (58% females) aged 25–70 years with LDL-cholesterol ≥ 3.5 mmol/L, total cholesterol within the normal range for each age group, and triglyceride (TG) ≤ 2.6 mmol/L participated in this study. Included study subjects had BMI between 20 and 35 kg/m², and had a stable body weight the last three months prior to inclusion. The study subjects were excluded if they used lipid-lowering drugs.

Over a 2-week run-in period, the control food items were included in all subjects’ daily diets. Afterwards, subjects were randomly assigned 1:1 stratified by age and gender to a control diet (C-diet) group that included the control food items in their diet and an experimental diet (Ex-diet) group that included experimental food items.

The experimental food items were the same type of products as the control food items, but with an altered fatty acid composition (SFAs replaced with mostly n–6 PUFAs). The food items in the C-diet group or the Ex-diet group were butter-based spread or margarine-based spread, butter or liquid margarine, and olive oil or rapeseed and sunflower oil, respectively. Likewise, products, such as liver paté, cheese, bread, muesli cereals, cream, mayonnaise, and crème fraîche, were given to the study subjects in the C-diet group, and some of the SFAs were replaced with rapeseed and sunflower oils in the products for the Ex-diet group. More information on the study food items is described in detail earlier [15]. The participants were instructed to include these food items in their daily diet and otherwise continue to eat as normal. The participants included the food items in their daily diet and the products were mainly used for breakfast, lunch, and supper, and oils, butter/margarine, cream, and crème fraîche were used for cooking. Therefore, we do believe that the foods item mainly were used to freshly prepare their meals at home. Based on the minimum intake of the food items, the n–6 PUFA intakes (mainly linoleic acid) were 4.2 g/d and 12.9 g/d, the n-3 PUFA intakes (mainly α-linolenic acid) were 0.9 g/d and 1.5 g/d and the SFA intakes were 19.2 and 5.7 g/d for the C-diet and Ex-diet group, respectively, during the intervention. More information on the FAs composition of the food items is provided earlier [15]. Dietary intake was registered during the intervention, and all subjects completed a four-day pre-coded food diary which also included a validated photography booklet with four different portion sizes at the beginning and the end of the intervention period [15].

Clinical and blood laboratory assessments were performed at baseline and after 8 weeks of follow-up. The day before blood sampling, the subjects were told to refrain from alcohol consumption and vigorous physical activity. Venous blood samples were drawn after an overnight fast (≥ 12 h). Serum was obtained from silica gel tubes (Becton Dickinson Vacutainer Systems) and kept at room temperature for ≤ 30 min until centrifugation (1500 × g, 15 min). Fasting serum concentrations of total cholesterol, LDL cholesterol, HDL cholesterol, TG, glucose, and insulin were measured by standard methods at a routine laboratory (Fürst Medical Laboratory) as previously described [15].

As described in detail [22], serum protein precipitation was performed before serum samples were analysed by an ultra-performance liquid chromatography (UPLC) system coupled to a quadrupole time-of-flight mass spectrometer (Premier QTOF-MS, Waters Corporation, Manchester, UK). The sample analysis was based on previously described analytical methodologies [23]. Briefly, serum samples were placed in 96-well Sirocco™ plates and five μL aliquots of each sample were injected into an Acquity UHPLC (Waters) equipped with an HSS T3 C18 column (2.1 × 100 mm, 1.8 μm) coupled with an HSS T3 C18 pre-column (2.1 × 5 mm, 1.8 μm) (Waters), using a gradient operated for 10.0 min. Experimental conditions throughout the analysis are shown in Table 1SI. Blanks consisting of 0.1% formic acid in H₂0, external metabolomics standard mixtures, and a pooled sample (of each analytical batch) were injected after every 15 samples and at the end of each analytical batch.

The raw data from the LC–MS was converted to an intermediate.netCDF format with the DataBridgeTM utility provided with the MassLynx software (Waters). The data acquired in positive and negative ionization modes were preprocessed separately with MZmine (ver. 2.31) [24]. The optimized preprocessing parameters are given in the online supplementary information (Table 2SI). Subsequently, the pre-processed data were imported to the software program, Matlab R2017b (ver. 9.3.0; Mathworks Inc., MA, US). Feature filtering was applied: (1) the features present in blanks were removed, (2) features detected before 0.3 min and later than 9.42 min were excluded, (3) early (0.3–0.8 min) features having 4–9 as the first decimal place of the mass-to-charge ratio (m/z) were deemed implausible and removed, and (4) potential isotopes along with duplicates were excluded.

Intensities were corrected for inter-batch variation; the intensity of each feature detected in the samples in each batch was multiplied by the batch average and divided by the overall average across all batches [22]. Furthermore, intra-batch drift was corrected using the online software program Galaxy (ver. 2.1.2). Afterwards, features were grouped within a retention time (RT) range of 0.01 min if their correlation coefficients were higher than 0.7. Lastly, features from negative and positive modes were concatenated into a single final data set.

The sample size of this study was calculated based on the primary outcome as published previously [15]. Power calculations estimated that 180 subjects (including a 20% dropout rate) were required to obtain 80% power with a type I error of 5% to detect a difference of 8% in LDL cholesterol concentrations at the end of the study between the two groups. Post hoc analyses showed that the number of subjects recruited gave sufficient power with the observed 10% change in LDL cholesterol between the groups (n = 47 subjects in each group). Prespecified secondary outcomes were other lipids, inflammatory markers, metabolomics, and transcriptomics.

Data analysis was performed in Matlab® using PLS-toolbox (ver. 8.7, Eigenvector Research, MA, US) with baseline-corrected data (serum baseline samples subtracted from 8-week sample) to reduce the variations in the metabolic profile caused by individuality.

Firstly, principal component analysis (PCA) was applied to explore for grouping patterns, and secondly, partial least-squares discriminant analysis (PLS-DA) models was developed to identify the most differentiating features separating the Ex-diet and C-diet group. Double cross-validation was applied for PLS-DA model evaluation [25]. Briefly, 80% of the data were used as a training set and the remaining 20% was used as a test set. The samples in the training set were cross-validated to find the optimal complexity of the model (number of latent variables). The minimum numbers of latent variables providing the lowest cross-validation (CV) classification errors for the model were chosen and variables that had the lowest influence based on variable importance in projection (VIP) plots were removed. The model was then recalculated based on the reduced number of variables in the training set. This continued as long as the selected features improved the model performance by decreasing the CV classification error in the training set. The performance of the selected features in terms of differentiating the samples from the Ex-diet and the C-diet group was based on the rate of misclassification errors (ER) and area under the curve (AUC) of the test set samples. One hundred PLS-DA models with randomly selected training/test set pairs were generated and a mean of the test set ER and a mean of the AUC were calculated to evaluate the performance of the models and the final number of included features. Next, features present in > 70% of the models (from 100 training/test set pairs) were selected for the final feature set.

The associations between diet-specific serum metabolic profiles reflecting the diets and the clinical outcomes of interest, i.e. total cholesterol, TG, LDL cholesterol, HDL cholesterol, glucose, and insulin, were explored by linear models applied using R (ver. 3.6.1) [26]. The metabolic profiles were characterized by the selected features from the PLS-DA and translated into principal component scores for comparison with the clinical outcomes. Finally, causal mediation analysis between the metabolic profile scores and the clinical outcomes was performed using the mediation package [27] to explore if the altered metabolome as a mediator of the effect of the intervention on clinical outcomes, i.e. total cholesterol, TG, LDL-cholesterol, HDL-cholesterol, and glucose.

Unknown compounds with a signal area below 100 counts were excluded from further identification as MS/MS experiments are not feasible with such low intensities (Table 3SI). For unknown compounds with higher intensity than 100, MS/MS spectra (ionized with either 10, 20, and 30 eV or 20, 36, and 48 eV using argon/nitrogen) were recorded with a tandem mass spectrometer (VION from Waters) using the same prior UPLC separation as described above (Table 4SI). MS/MS spectra were compared with available databases, such as METLIN (https://​metlin.​scripps.​edu) [28], Human Metabolome Database (HMDB) (http://​www.​hmdb.​ca) [29], and produced by in silico fragmentation tools such as CFM-ID (http://​cfmid.​wishartlab.​com/​) [30], FingerID (https://​www.​csi-fingerid.​uni-jena.​de/​) [31] and LIPID MAPS^R (http://​www.​lipidmaps.​org) [32].

Authentic standards (Table 5SI) and the sample that contains the metabolites of interest most abundantly were run consecutively on the same platform for correct annotation. Furthermore, an algorithm was used to predict retention times of homologous series of phospholipids based on their physicochemical properties and their orthogonal data as described previously [33]. Prediction plots for lysophosphatidylcholines (lysoPCs) and PCs are shown in supplementary information Figs. 1SI and 2SI. With increasing numbers of carbon atoms in the fatty acyl chain in PCs and lysoPCs polarity decreases, resulting in a stepwise increase in retention time. Likewise, an increase of double bonds reduces the retention time in a predictable manner [33]. For several of the phospholipids, it was possible to detect both isomers with the different fatty acyl chains located in either sn1 or sn2 position. Authentic standards were used to validate the prediction plots.

The identified metabolites were classified with a level of identification I-V in accordance with the Metabolomics Standards Initiative classification scheme [34]. Examples of structure elucidations of level II identified compounds are provided in supplementary information Fig. 3SI.

The data pre-processing and further data filtering resulted in 740 and 2120 features for negative and positive modes, respectively. The final PLS-DA model performed well in terms of differentiating samples from the Ex-diet and C-diet groups with an average AUC and ER of 0.83 and 0.23, respectively. The result of the variable selection process provided 32 features that were present in at least 70% of the models (with 100 training/test pairs). A PCA based on these metabolites shows that the first two principal components explain 43.2% of the variation with good separation between the two groups (Fig. 1). The identified metabolites are presented in Table 2 together with their retention time, observed mass, annotation, and observed fragments from the MS/MS experiments. Separation and box plots of each of the metabolites are listed in supplementary information Table 7SI.

The serum metabolic pattern mainly reflects the different lipid compositions of the two diets. Among the selected metabolites, the Ex-diet group was dominated by unsaturated plasmalogens, e.g. PC(P-38:4) (Fig. 1). Furthermore, an unsaturated medium-chain carnitine, e.g. C8:1-carnitine, together with the bile acids lithocholic acid (LCA) and indolelactic acid (ILA) also characterized the Ex-diet group. In contrast, the C-diet group was mainly characterized by odd-numbered phospholipids, e.g. PC(31:0), PC(33:0), and PC(33:1). The C-diet group also had increased levels of the lysophospholipid, lysoPC(14:0), representing myristic acid present in butter and milk. Furthermore, a saturated medium-chain carnitine, C9:0-carnitine, and the short-term glycaemic marker, 1,5-anhydroglucitol (1,5-AG) (same as 1,5-anhydrosorbitol), also characterized the C-diet group.

From a hierarchical cluster analysis using Spearman correlations between the selected metabolites from the final PLS-DA model, associations between the compounds are shown in Fig. 2. The metabolites differentiating the two diet groups are forming two major clusters that show no or a very weak correlation. Furthermore, within the clusters, a mutual correlation is found in accordance with the PCA model (Fig. 1). Besides the two major clusters, there are several small sub-clusters; the phospholipids with lower levels in the Ex-diet group all showed a strong correlation with each other, especially two isomers of PC(33:1) (m/z 790.563, RT 8.14 and 8.25) and an unidentified phospholipid (m/z 776.5638, RT 8.13), but also two other unidentified metabolites, m/z 880.6077 and m/z 832.6107. Furthermore, C9:0-carnitine correlates with an unknown compound (m/z 155.0677, RT 6.28). Moreover, the two carnitines show a moderate inverse correlation, confirming their associations with the different diet groups.

The serum metabolites most strongly separating the two diet groups are listed in Table 2. The associations between PC1 and PC2 scores and the clinical outcomes are shown in Table 3. Total cholesterol, LDL-cholesterol, and TG levels are all significantly correlated with PC1 scores. PC1 is dominated by the odd-numbered PCs, e.g. PC(31:0), PC(33:0), and PC(33:1). Furthermore, total cholesterol and LDL-cholesterol levels are inversely correlated with PC2, represented by the two PC(P-38:4) plasmalogens and by C8:1-carnitine. Moreover, HDL-cholesterol levels correlated negatively with PC2. No associations were detected for glucose and insulin levels in relation to the diet-specific metabolic profiles.

Mediation analysis was performed between the clinical outcomes that were significantly associated with PC1 and PC2 scores and are presented in supplementary information Table 8SI. The metabolic profile scores revealed no causal mediation effect of the clinical outcomes related to serum triglycerides or cholesterol fractions.