2.2. Ethics Statement

The study adheres to the directives of the Declaration of Helsinki (1964).

2.3. Sample Preparation and ABO and Rh Determination

Plasma samples were obtained after overnight fasting and, after collection, were stored at −80 °C and handled according to standard operating procedures.²⁰ The ABO and Rh blood groups were determined by standard procedures using agglutination techniques.²¹

2.4. NMR Analysis

The mono-dimensional nuclear Overhauser effect spectroscopy (NOESY) ¹H spectra of plasma samples were acquired using a Bruker 600 MHz spectrometer (Bruker BioSpin s.r.l., Germany), operating at 600.13 MHz. For more details about NMR sample preparation, acquisition, and spectral processing, we refer the reader to the original publications.^17,22

Twenty-seven (27) metabolites were assigned and identified using in-house software developed based on standard line-shape analysis methods and with the help of matching routines of AMIX 7.3.2 (Bruker BioSpin) in combination with the BBIOREFCODE (Version 2–0–0; Bruker BioSpin) reference database and published literature (when available). In the Supporting Information (Figure S1), the assignment of a plasma spectrum is shown. The relative concentration of each metabolite was calculated by integrating the signals in the spectra; 114 lipoprotein fractions and subfractions were assigned and quantified using the AVANCE IVDr [Clinical Screening and In Vitro Diagnostics (IVD) research with B.I. Methods, Bruker BioSpin].²³ The direct integration of NMR signals was carried out. A list of metabolites and lipoprotein fractions and subfractions and lipids is given in Table 3.

2.5. Statistical Analysis

2.5.4. Permutation Test

The statistical significance of the RF classification models was determined with a permutation test using M = 1000 permutation. P-values were calculated by comparing the value model₀, obtained from the original, and non-permuted data with the values model₁, model₂, ..., model_M obtained from the M-times permutation-test. The P-value for a specific quality measure is calculated as follows³⁵

1

where |D_perm| is the number of permuted models for which a given quality measure is larger or equal to the quality measure from the original (non-permuted) model (model₀).

2.5.5. Robust Linear Regression

The association between plasma metabolites, lipids, and lipoproteins and the ABO and Rh groups was determined by linear regression models,^36,37 according to the following formula

2

where y_i is the abundance/concentration of the ith molecular feature (metabolite or lipid/lipoprotein), b_i is the estimated regression coefficient that quantifies the association between the ABO/Rh blood groups, adjusted for age and sex, and the molecular features, and ε_i is the residual part not accounted for by the other terms. To reduce the influence of outliers and high leverage points on the regression solutions, we used a robust version for the linear regression, where the fitting is done by iterated re-weighted least squares.^36,37 For each model, post-hoc pairwise comparisons among ABO blood group systems were also performed using the estimated marginal means.³⁸

The Benjamini–Hochberg method³⁹ was used to correct for multiple testing.

2.6. Software

All calculations were performed in R (version 4.0.3).⁴⁰ The function “missForest”, implemented in the missForest package, was used to impute the missing data.²⁴ RF models were built using the “randomForest” function, implemented in the R package RF,^31,41 growing a decision forest composed of 1000 trees, using default parameters. To estimate the significance of importance metrics for the RF models by the 1000-times permutation of the response variable, the “importance” function, implemented in the R package rfPermute, was used.⁴² The function “rlm”, implemented in the MASS R package, was used to perform the robust fitting of linear models.⁴³ Default parameters were used. The function “emmeans”, implemented in the emmeans R package,^38,44 was used to compute comparisons among specified factors and/or factor combinations in the linear models.

3.2. Exploration and Discrimination of Metabolites, Lipids, and Lipoproteins Associated with ABO and Rh Blood Groups

To explore comprehensively the metabolomic and lipoproteomic profiles (consisting of 27 metabolites and 114 lipoproteins) associated with the ABO and Rh blood group systems, we applied PCA on the n = 840 plasma samples. The PCA score plot in Figure ​Figure11a shows no clear separation among the A, AB, B, and O groups, suggesting that metabolic differences are too subtle to be resolved using an unsupervised multivariate approach. A similar lack of separation can be observed in the case of Rh⁺ and Rh^– profiles (Figure ​Figure11b).

Open in a separate window

Figure 1

(a) PCA model score plot [PC1 (11.4%) versus PC2 (9.2%) versus PC3 (5.9%)]. Each dot represents a single metabolic profile with colors denoting different groups of patients: n = 330, A blood group subjects (violet dots); n = 41, AB blood group subjects (dark orange dots); n = 78, B blood group subjects (green dots); and n = 391, O blood group subjects (dark blue dots). (b) PCA model score plot [PC1 (11.4%) versus PC2 (9.2%) versus PC3 (5.9%)]. Each dot represents a single metabolic profile colored by the different groups of patients: n = 710, Rh⁺ blood group subjects (red dots); and n = 130, Rh^– blood group subjects (light blue dots).

RF classification was applied to investigate whether subjects with different ABO and Rh blood group systems could be discriminated from the metabolic and lipoproteomic profiles. The results of the RF classification for ABO and Rh blood groups are given in Tables 5 and 6, respectively. Overall, we obtained extremely weak classification models to discriminate between the different ABO groups: A versus AB groups (AUC = 0.562, P-value = 0.04), A versus B groups (AUC = 0.514, P-value = 0.03), AB versus B groups (AUC = 0.535, P-value = 0.03), A versus O groups (AUC = 0.557, P-value = 0.01), and AB versus O groups (AUC = 0.568, P-value = 0.03). A non-significant predictive model (AUC = 0.503, P-value = 0.09) was obtained for the discrimination of B versus O groups (see Table 5). These results suggest that the metabolic and lipoproteomic profiles, as a whole, may be weakly associated with the ABO groups; why this is the case is not clear: a lack of association may be a consequence of the limited sample size, or relevant associations may be limited to a few metabolites and/or lipoprotein/lipid features.

In contrast, subjects with different Rh blood groups can be easily and accurately discriminated on the basis of their metabolite and lipoproteomic profiles (AUC = 0.808, P-value = 0.02) (see Table 6 and Figure ​Figure22). By evaluating the RF important variables, as reported in Figure ​Figure33, we observed that the subfraction LDL5 related to Apo B (Subfr. Apo B–LDL5), the subfraction LDL4 related to Apo B (Subfr. Apo B–LDL4), the subfraction HDL4 related to Apo A1 (Subfr. Apo A1–HDL4) and Apo A2 (Subfr. Apo A2–HDL4), the subfraction HDL2 related to cholesterol (Subfr. Chol–HDL2), and the lipid main fraction LDL related to Apo B (LMF Apo B–LDL) are the most relevant and significant variables in the model discriminating Rh⁺ with respect to Rh^– blood groups.

Open in a separate window

Figure 2

Balanced RF model score plot, taking into account the sexual dimorphism distribution. Each dot represents a single metabolic profile with colors denoting different groups of subjects. n = 51, randomly selected Rh⁺ blood group subjects stratified by sex (red dots); and n = 51, Rh^– blood group subjects stratified by sex (light blue dots).

Open in a separate window

Figure 3

Importance of metabolites and lipoproteins in the RF model for the classification of Rh⁺ and Rh^– subjects calculated using the mean decrease Gini index. Gray bars correspond to no significant variables. Red bars correspond to statistically significant (false discovery rate (FDR)-adjusted P-value < 0.05) important variables more representative in the Rh⁺ and Rh^– comparison obtained with the permutation test. Only variables with a mean decrease Gini score >0.7 are shown.

3.3. Association of ABO Blood Groups to Metabolites and Lipids

The association between the levels of circulating plasma metabolites, lipoproteins, and ABO and Rh groups was assessed by means of robust linear regression, correcting for sex and age. We observed significant association (P-value < 0.05) for 8 out of 114 lipoprotein fractions and subfractions with ABO groups; no significant association remains after correction for multiple testing (see Table 7). Since correction for multiple testing increases the risk of false-negative, especially in the case where (possibly) weak associations are tested on a large number of variables, we took an honest and pragmatic approach, presenting both corrected and uncorrected P-values and discussing the biological implications of the results for which P-values were significant before correction. Using robust linear modeling, we observed that the subfractions of HDL (in particular, HDL1 with a density of 1.063–1.100 kg/L and HDL2 with a density of 1.100–1.112 kg/L) and the subfraction of LDL (in particular, LDL2 with a density of 1.031–1.034 kg/L) turned out to be relevant in ABO lipidic and lipoproteomic differences. We observed non-statistically significant associations between the particle number of LDL2 (PN LDL2), the subfraction HDL1 related to Apo A1 (Subfr. Apo A1–HDL1), the subfraction HDL2 related to cholesterol (Subfr. Chol–HDL2), phospholipids (Subfr. PL–HDL2), the subfraction LDL2 related to free cholesterol (Subfr. Free Chol–LDL2), cholesterol (Subfr. Chol–LDL2), phospholipids (Subfr. PL–LDL2), and Apo B (Subfr. Apo B–LDL2) and the ABO groups. A post-hoc test on the 8 lipids and lipoproteins poorly associated with the ABO groups was also performed (see Supporting Information Table 1), highlighting that the non-statistically differences exist mainly between the A and AB, and the AB and O groups. No differences were observed between the B and O, and the A and O groups.

It has been shown in both experimental and clinical studies that higher plasma levels of LDL and cholesterol in non-O blood groups (A, AB, B) influence the susceptibility of these groups to cardiovascular diseases, while in the O blood group, higher levels of HDL tend to play a protective role in these systemic pathologies,⁴⁶⁻⁵⁰ but the molecular mechanisms by which these group-specific pathologies are influenced have not yet been elucidated.

Concerning the role played by apolipoproteins (Apo), especially Apo B, there is evidence that higher numbers of RBC-bound Apo B in the O blood group compared with the non-O blood groups are associated with an atheroprotective effect and the reduction of the risk of developing vascular diseases (i.e., venous thromboembolism, ischemic stroke, peripheral vascular thrombosis, and so forth).^46,51,52

The authors acknowledge the support and resources provided by Instruct-ERIC, a Landmark ESFRI project, and specifically the CERM/CIRMMP Italy Centre.