Metabolomics investigations and data collection and identification

110 μL of organic solvent (2/1 v/v chloroform:methanol) was added to 20 μL of serum and the sample was shaken at 30 Hz for 2 mins, then let to stand at room temperature for 0.5–1 hrs. The sample was then centrifuged at 14,000 rpm and 4°C for 3 mins, then 50 µL of the organic phase was transferred to a microvial and 70 µL of internal standards was added (phosphatidylserine (PS)(17:0/17:0), phosphatidylglycerol (PG)(17:0/17:0), phosphatidylethanolamine (PE)(17:0/17:0), monoacylglycerol (MG)(17:0/0:0/0:0), diacylglycerols (DGs)(17:0/17:0/0:0), TG(17:0/17:0/17:0), PC(19:0/19:0), PC(17:0/0:0), TG(16:0/16:0/16:0), ceramide (Cer)(16:0). Quality control samples were also prepared by pooling 10 μL from each extract. All samples were stored at −80°C until analysis.

The chromatographic separation was performed on an Agilent 1290 Infinity UHPLC-system connected to an Agilent 6550 Q-TOF mass spectrometer equipped with a jet stream electrospray ionization. 1 µL of the extracted serum was injected onto an Acquity UHPLC-Q-TOF-MS system. The samples were analyzed first in positive electrospray ionization polarity mode (ESI+) and then the instrument was switched to negative polarity mode (ESI−) and the samples were re-analyzed. In ESI+, the lipid species including TGs, DG, PC, PE, or sphingomyelins (SMs) are detected, whereas ESI– gives better sensitivity for phosphatidylinositols (PIs), PS, phosphatidic acids (PAs), ceramides, cardiolipins and nonesterified free fatty acids (FFAs).

MS data processing was performed using open-source software MZmine 2.18.18 Peaks were identified using a custom database search and normalized using lipid class-specific internal standards, and also utilizing MS/MS data. Unknown lipids were normalized with the closest eluting internal standard. The custom database used in this study was recently assessed as part of the NIST lipidomics ring study, which comprised 31 laboratories worldwide.19

We applied principal component analysis (PCA) on the data to have an overview of the CACS classes.20 We used orthogonal partial least squares discriminant analysis (OPLS-DA) to discriminate the 3 combinations of classes (NCC vs MCC; NCC vs SCC; MCC vs SCC) separately for the negative and positive modes. A model is considered statistically significant if it obtains relatively high Q2 values (according to cross-validation [CV]) with simultaneously low analysis of variance (ANOVA) and permutation test p-values (in this work α=0.05 for both). All variables were normalized to unit variance (z-scores).

Unpaired t-tests at the 95% confidence level were used to evaluate the statistical significance for lipid differences between the CACS classes. Benjamini–Hochberg correction with α=0.10 was used to control the false discovery rate (FDR).21 A logistic regression model (for each lipid) with a binary outcome was used to test models that were adjusted for age, sex, and use of statin medication. In order to be considered statistically significant, a variable had to present both a p-value of t-test <0.05 and an FDR p-value of <0.10. A trend of statistical significance was defined as either a p-value of t-test <0.05, or a p-value of <0.05 in the multivariate logistic regression models adjusted for age, sex, and statin treatment.

Fold changes were calculated using the median intensities of two CACS classes at a time. They were calculated as the highest median intensity among the two classes divided by the lowest of them. In case the class with lowest CACS has highest median intensity, the value is multiplied by −1.

For cluster analysis, ESI+ data were scaled to zero mean and all variables were normalized to unit variance. K-means clustering was applied, to the scaled data, with cluster numbers based on silhouette scores, to group lipids with similar profiles across all samples. The analysis was performed using R software, version 3 (R Foundation for Statistical Computing, Vienna, Austria) using function “kmeans” with the default algorithm.22 The number of clusters (cl) was determined to n=6, based on the silhouette score using package NbClust.23 ANOVA was applied to explain the average lipid profile in each lipid cluster by age and calcification group. The age-corrected average cluster profiles are visualized as barplot with grouped bars and 95% CIs, based on the Tukey adjustment.