Participant recruitment, inclusion and exclusion criteria, randomization and compliance have been described in detail previously . Briefly, a total of 142 healthy men and women, between 45 and 69 years were enrolled in the study. Of these, 125 participants were included in the per protocol analysis. Written informed consent for participation in the study was obtained from all participants and all procedures were conducted in accordance with the Declaration of Helsinki.
For the present analysis, we report on a predefined subgroup of 42 subjects selected from the main cohort using stratified sampling with respect to sex and treatment group (fourteen subjects per group). High-resolution lipid-analysis (HR-lipids), gas chromatography (GC)- and liquid chromatography (LC) mass spectrometry (MS) analysis was completed for the entire subset while gene expression was assessed in whole blood of 21 randomly selected subjects from the same subset (seven subjects per intervention group).
Study design and interventions
The study was a randomized, controlled, double-blind, three-arm, parallel-group intervention, designed to investigate the health effects of TFA intake from industrial and ruminant sources . The study was approved by the cantonal ethical committee of Bern, Switzerland and registered at www.clinicaltrials.gov (NCT00933322). Block randomization with stratification for gender and age was used to assign subjects to one of three diets: a diet with alpine butter rich in ruminant TFA (rTFA group), a diet with margarine rich in industrial TFA (iTFA group) and a diet with margarine without TFA (wTFA group, control).
The study protocol began with a run-in period of two weeks, during which all study participants followed the wTFA diet, and was followed by a 4-week intervention period with the assigned diets (rTFA, iTFA, wTFA) according to the randomization. All dietary fats for the study were provided to the subjects (33-36% of individual energy requirements) in the form of the designated study products plus 15-25 g/day rapeseed oil, as described in detail previously . The dietary fat supplements resulted in comparable levels of 2% energy intake from TFA for iTFA and rTFA diets. During all study phases, subjects adhered to a prescribed diet, defined by a dietitian to exclude all other dietary sources of TFA and meet individually calculated daily energy requirements. Details of the dietary restrictions and evaluation of dietary adherence are published elsewhere . At each visit, weight was assessed and BMI calculated.
Blood samples were collected from the antecubital vein at baseline (after the 2-week run-in period) and at the end of the intervention period (week 6) after an overnight fast. The preparation and processing of blood samples to assess clinical biochemistry and biomarkers of inflammation, coagulation and endothelium function have been described previously . For GC-MS and LC-MS analyses, serum was separated and stored in microtubes at -80°C until assayed. Whole blood was collected for microarray analysis in PAXgene blood RNA tubes and frozen at -20°C for 24 h before transfer to -80°C until further treatment.
High-resolution lipid analysis with GC-FID
Serum samples were prepared for analysis by addition of 15 µL of internal standard (C13:0, 7.5 µg/15 µL) to 100 µL of serum, followed by methylation of free FA with MeOH/HCl (25°C for 45 min). A post-reaction treatment for neutralization was applied with Na2CO3 and extraction was performed with 300 µL hexane. Samples were measured by gas chromatography with flame ionization detector (GC-FID) (6890, Agilent Technologies, Santa Clara, CA) with injection volumes of 0.5 µL as described previously . Butter and margarine samples were assayed previously by Radtke et al. , using the same GC-FID technique as for serum samples with sample preparation according to the method of Collomb and Bühler . The data was reprocessed and reintegrated here with some additional reference lipids included. Eighty-five features (66 single FA and 19 sum parameters) were quantitatively analysed in the serum and dietary fats.
Transcriptomic analysis and data preprocessing
Total RNA was extracted from whole blood samples and treated to deplete for globin mRNA according to the protocol described by Gille et al. . Whole genome transcript profiling was then performed with HG-U219 oligonucleotide expression probe arrays (Affymetrix, USA). Preparation of samples for profiling including reverse transcription, amplification, amplified RNA labelling, purification, fragmentation and hybridization steps were performed as defined previously . Arrays were measured in accordance with manufacturer’s recommendations using Affymetrix GeneAtlas™System. Raw array data was imported to the R environment (version 4.1.2)  and corrected for background noise, log2 transformed, normalized for inter-array variation (quantile normalization)  and summarized using the rma (Robust Multichip Average) function from affy (version 1.72.0) . In addition, probe sets were filtered to keep only: the most variable probe set per gene (based on standard deviation), probe sets assigned to gene symbols (hgu219.db, version 3.2.3)  and probe sets with average expression (log2) > 5. Of the 49’386 probe sets, 6’128 probe sets were retained for further analysis.
Untargeted GC-MS analysis and data pre-processing
Sample preparation for the untargeted analysis of the serum samples by GC-MS was based on the method published by Trimigno et al.  using a GC-MS 7890B/MS5977A (Agilent Technologies, Santa Clara, CA, USA) with a CombiPAL autosampler (CTC-Analytics AG, Zwingen, Switzerland). After deconvolution, features from subjects that only appeared in less than three samples were eliminated (n = 4828 remaining features). A second manual integration was conducted on 54 selected features that showed significant differences between treatment group responses to the interventions (see Statistical Analyses). The areas were normalized with the isotopically-labelled fructose. The features demonstrating a significant treatment effect were searched in EI-Mass Spectral Library (NIST 2017, Gaithersburg, MD 20899-6410, USA), masslib-library (www.masslib.com). The molecules demonstrating sufficient potential for identification were acquired from Sigma-Aldrich (Buchs, Switzerland) and analyzed by GC-MS.
Untargeted LC-MS analysis and data pre-processing
Serum samples were thawed, treated and analyzed according to the protocol described by Pimentel et al. , using a maXis 4G+ (Bruker, Bremen, Germany) coupled to a 3000 RS UHPLC (Thermo, Basel, Switzerland). Raw data (m/z values, retention times and intensities) were imported into Progenesis QI (V2.4) (Waters, Switzerland) for retention time alignment, peak detection, sample loading normalization, deconvolution of masses (including detection of adducts) and final export of peak volumes for further statistical treatment. The pre-processed dataset (~85’000 features) was imported into R (version 3.5.2)  and filtered in two steps. All features that showed a mean in QC samples (n = 21) that were < 4 times higher than the mean for the final three blanks were excluded. In addition, features with a relative standard deviation > 30% in the QC samples were excluded to leave a reduced dataset of 11’616 features. Identification was carried out by a mass search in the Human Metabolome Database (HMDB v4.0 2018 ) and NIST (National Institute of Standards and Technology Database, V14). Successive identification was then confirmed by LC-MS where exact mass and retention time of the metabolites were compared with standards. The identification was differentiated at four levels [38-40]. We report here in detail on metabolites identified at the highest level (level 1), i.e. compound identified with accurate retention time (+/- 10%) and mass (+/- 10 ppm) when confirmed with a chemical reference standard.
All statistical analyses were performed in the R environment (version 3.5.2) . The normality of baseline clinical characteristics and metabolomics datasets was explored through Shapiro-Wilk tests of normality and visually by quantile-quantile plots. Principal components analysis (PCA) was used with all datasets to visually inspect sources of variation, in particular to exclude the presence of batch effects, outliers or potential confounders (e.g. sex) (lipids and metabolomics datasets: ropls package, v1.4.4, 2016, ; transcriptomics dataset: stat function ‘prcomp’ visualized with ggplot2 (version 3.3.5) .
Non-parametric univariate statistics were used for the current analysis in view of the reduced study population size and multiple features that were not distributed normally. Baseline clinical characteristics, blood biochemistry, inflammation and endothelium biomarkers of the study population were compared by treatment group using the Kruskal-Wallis test (P < 0.05).
GC-FID Lipid analysis
The Kruskal Wallis test was used to compare the changes in lipid profiles in response to the three interventions. For significant results, a post hoc Wilcoxon signed-rank test was completed to compare each with adjustment of P-values for multiple testing using the Benjamini-Hochberg FDR method  (significance where PFDR-adjusted < 0.05). The sum parameters for the classes of fatty acids were integrated into this analysis. The response to the intervention for these analyses was defined for each participant as the delta change between measures at endpoint (week 6) and baseline (week 2). Features showing a significant difference between the three treatments were presented in heatmaps using the “Metabolomics” package (v0.1.4)  with distance calculated by “Canberra” method and clustering using the “ward.D2” method .
Gene and pathway analysis
The molecular effects of the dietary interventions were investigated by assessing the change in gene expression in whole blood (i.e. delta change between baseline and post-treatment), with delta values calculated using log2-transformed data, thus interpretable as log2-fold expression changes (log2FC). Differentially expressed genes for each group and between groups were identified by performing a paired moderated t-test on the responses to the intervention using R package Limma (Linear Models for Microarray Data) (version 3.50.0) (PFDR-adjusted < 0.05) .
Gene set enrichment analysis (GSEA)  was used to explore the results of the differential analyses. GSEA determines whether a predefined set of genes show statistically significant, consistent differences between two biological states by calculation of an enrichment score (ES) for each geneset. A ‘normalized’ ES (NES) is calculated to account for differences in number of genes in each geneset. GSEA was carried out for a selection of genesets (n = 20) from the Hallmark human reference geneset collection (mSigDB, Broad Institute, version 7.5.1) that describe pathways that are regulated by TFA, including lipid and cholesterol metabolism, inflammation, apoptosis, autophagy, coagulation, adipose tissue regulation and oxidative stress . Significance of enrichments were confirmed by comparing the ES to those obtained by random permutations (n = 1000 iterations) of the ranked gene with FDR correction for multiple testing (significance where PFDR-adjusted < 0.05) .
Untargeted metabolomics analysis
The same statistical workflow as described above for the GC-FID lipid analysis was applied for the untargeted LC-MS metabolomics dataset to identify molecular biomarkers that were differentially modulated by the interventions, (comparing delta change between baseline and post-treatment).For the untargeted GC-MS dataset, a two-step statistical analysis was applied to select putative compounds that could discriminate the different effects of the interventions (delta change between baseline and post-treatment) for the reintegration step of data processing. In the first step, univariate analyses were applied to identify metabolites showing a different response between the dietary treatments using the Kruskal-Wallis test (P < 0.05). This was complemented with the identification of further discriminating metabolites from orthogonal partial least squares - discriminant analysis (OPLS-DA) models using the Ropls package (v1.4.4, 2016)  to compare the effects of each pair of treatments (for valid models, the top 100 metabolites were selected for reintegration). After reintegration, the second step of analysis applied a Wilcoxon signed-rank test to compare the responses of the reintegrated metabolites to the treatment in pairwise assessments (PFDR-adjusted < 0.05).
Correlation analyses were conducted to assess the relationships between biomarkers measured previously , including blood biochemistry (glycemia, insulinemia, blood lipids) and biomarkers of inflammation and endothelium function and the FAs measured by targeted GC-FID analysis. Spearman’s correlation test was used to associate the biomarker and FA changes in response to the dietary interventions (delta change between baseline and post-treatment) with visualization by corrplot (v0.84)  with significance where PFDR-adjusted < 0.05 . This analysis was restricted to the fourteen clinical parameters that showed no significant difference between groups at baseline. The effect of the different dietary intervention on these associations was considered by repeating the analysis with Spearman’s partial correlation test to control for the effect of diet (ppcor Package, v 1.1) .