Sphingolipid composition of circulating extracellular vesicles after myocardial ischemia

Sphingolipids are structural components of cell membrane, displaying several functions in cell signalling. Extracellular vesicles (EV) are lipid bilayer membrane nanoparticle and their lipid composition may be different from parental cells, with a significant enrichment in sphingolipid species, especially in pathological conditions. We aimed at optimizing EV isolation from plasma and describing the differential lipid content of EV, as compared to whole plasma. As pilot study, we evaluated the diagnostic potential of lipidomic signature of circulating EV in patients with a diagnosis of ST-segment-elevation myocardial infarction (STEMI). STEMI patients were evaluated before reperfusion and 24-h after primary percutaneous coronary intervention. Twenty sphingolipid species were quantified by liquid-chromatography tandem-mass-spectrometry. EV-ceramides, -dihydroceramides, and -sphingomyelins increased in STEMI vs. matched controls and decreased after reperfusion. Their levels correlated to hs-troponin, leucocyte count, and ejection fraction. Plasma sphingolipids levels were 500-to-700-fold higher as compared to EV content; nevertheless, only sphingomyelins differed in STEMI vs. control patients. Different sphingolipid species were enriched in EV and their linear combination by machine learning algorithms accurately classified STEMI patients at pre-PCI evaluation. In conclusion, EV lipid signature discriminates STEMI patients. These findings may contribute to the identification of novel biomarkers and signaling mechanisms related to cardiac ischemia.


Results
Study cohort clinical profile. We analyzed peripheral venous blood samples of patients with a diagnosis of STEMI (n = 7) on presentation to the emergency department before primary percutaneous coronary intervention (PCI), as well as 24 h (hours) post-intervention, as compared to controls (n = 9), matched for cardiovascular risk and demographic parameters. The median time from chest pain onset to the presentation at the emergency department was 3 h. PCI was performed within 1 h from the first medical contact for all STEMI patients. Mean age was 62 ± 10.6 years, 68.8% were males, with a high prevalence of hypertension, and dyslipidemia. Study groups did not significantly differ from one another with respect to systolic and diastolic pressure, BMI, white blood cells count, creatinine, glomerular filtration rate, glycemia, total cholesterol, HDL, and triglycerides. As expected, patients with a diagnosis of STEMI displayed higher levels of high-sensitive (hs)-troponin and C-reactive protein (CRP), and lower left ventricular ejection fraction (LVEF%) at echocardiography 24 h post-PCI as compared to controls (p < 0.05 for all comparisons). Clinical and biochemical parameters are reported in Table 1. EV isolation and characterization. The classical EV isolation protocol using a serial centrifugation steps, here has been implemented to further remove cellular debris and blood contaminants. EV were isolated from 200 μL of platelet-free plasma by double series of sequential centrifugations as depicted in Fig. 1a. The final solution enriched in extracellular vesicles is referred hereby as EV.
Nanoparticle tracking analysis (NTA) was performed in parallel in pre-cleared plasma and enriched EV, the latter showed typical dimensional profile of small EV (90% of total particles were less or equal 227 ± 38.5 nm), with no contamination of larger particles (Fig. 1b). Western blot analysis confirmed enrichment of small EV also Scientific RepoRtS | (2020) 10:16182 | https://doi.org/10.1038/s41598-020-73411-7 www.nature.com/scientificreports/ referred elsewhere as exosomes 20 , based on the TSG101 expression (1.3-fold increase; p = 0.013), with negligible presence of antigens gauging the presence of cellular debris and/or lipoparticles as compared to plasma ( Fig. 1c and Figs. S1a,b, S2). As expected, whole plasma was positive for albumin, apolipoprotein APOA1, and for the endoplasmic reticulum chaperone marker GRP94 31 . Albumin and ApoA1 were also present in EV preparations, although their expression was markedly reduced (4.1-and 2.8-fold decrease, respectively; p < 0.01), whereas GRP94 was hardly expressed (22.6-fold decrease; p < 0.001; Figs. 1c, S1a,b, S2). Collectively, these data indicate that the plasma EV preparations obtained by our optimized protocol were enriched in exosomes with minimal presence of contamination in terms of albumin and lipoproteins, and negligible presence of cellular debris. Once the reliability of the isolation protocol had been assessed, we sought to evaluate differences between patient and matched controls in terms of EV-plasma concentration. NTA analysis clearly highlights a significant increase of plasma EV concentration in patients with STEMI as compared to controls (9.5E11 particles/mL [7.6E11-3.1E12] vs. 1.6E11 particles/mL [1.1E11-3.3E11]; p = 0.003). After reperfusion, the amount of circulating EV significantly decreased (4.9E11 n/mL [2.6E11-6.6E11]; p = 0.036; Fig. 2a). Moreover, EV size as measured by diameter was higher in patients with STEMI at pre-PCI evaluation compared to controls (182 nm  vs. 149 nm [129-175] p = 0.033; Fig. 2b and Table S1).
To further exclude the presence of cross-contaminating particle with similar size, plasma EV were then analyzed based on the expression of tetraspanins (CD9, CD63, CD81) by flow cytometry. To overcome the issue of size-related limit of detection, EV were bound to capture beads coated with specific antibodies. Mean fluorescence intensity values for CD9 and CD63 were significantly increased in patients with STEMI compared to controls ( [8.6-29.0]; p < 0.05 for both comparisons), and an increasing trend was observed also for CD81 (Fig. 2c). Levels of tetraspanin antigens decreased after reperfusion (Table S1). To validate our findings in accordance to International Society of Extracellular Vesicles (ISEV) recommendation 32 , we performed a semiquantitative analysis by WB for both surface and luminal EV specific proteins in a subset of subjects (4 controls vs. 4 patients with STEMI before and after PCI). Immunoblot assays showed an increase of canonical EV markers such as TSG101 and CD81 in pre-PCI samples as compared to control ones (1284.  5] a.u., respectively) (Figs. 2d and S1c,d). Finally, to explore the putative origin of circulating EV, we profiled their surface by using multiplex assay to simultaneously analyze 37 antigens (Table S2 and Fig. S3). Tetraspanins (CD63, CD9 and CD81), specific markers of EV 32 , were highly expressed on the surface of vesicles. Moreover, EV were found positive for markers of activated platelets (CD62p, CD41b, and CD42a), endothelium (CD31), and leukocytes (CD40). Markers of inflammatory cells (e.g. CD2, CD3, CD4, CD8, CD14, CD19, CD20, CD45; Table S2 and Fig. S3) were also found in a minor proportion of EV (less than 5%).

Sphingolipids composition of EV after myocardial infarction.
Sphingolipids composition was evaluated by LC-MS/MS analysis in whole plasma over isolated EV (Tables S3-S6). Representative chromatographic traces for EV samples (STEMI patient at pre-PCI evaluation vs. control patient) are shown in Fig. 3. www.nature.com/scientificreports/   www.nature.com/scientificreports/ p < 0.05 for all comparisons; Fig. 4 and Table S5). The sphingolipids concentration in EV was directly correlated to the peak level of high sensitive troponin (hs-troponin) reached by each patient (R ranging between 0.630 and 0.859; p < 0.01 for all comparisons) and inversely correlated to ejection fraction measurements via ultrasound echocardiography at 24 h post reperfusion (R ranging between − 0.545 and − 0.631; p < 0.05 for all comparisons; Fig. 4 and Table S7). Finally, EV sphingolipids concentration decrease 24 h after reperfusion (p < 0.05). As a result, the peak in EV sphingolipid concentration anticipated the peak of hs-troponin, which was reached at or after 24 h from PCI (Fig. 4). The presence of ceramides, dihydroceramides, and sphingomyelins in plasma was 500 to 700-fold higher as compared to EV. Similarly, to what found in EV, plasma sphingomyelin concentration significantly increased after STEMI before PCI and fell back to control level at 24 h after reperfusion (pre-PCI 420.6 pmol/mL [399.7-461.5; controls 169.3 pmol/mL [154.4-184.9]; post-PCI 149.9 pmol/mL [129.8-176.2]; p < 0.001). Moreover, sphingomyelin plasma levels significantly correlated with both hs-troponin and ejection fraction ( Fig. S4 and Tables S6, S7). On the contrary, plasma levels of ceramides and dihydroceramides did not differ between STEMI patients and controls; likewise, neither hs-troponin, nor ejection fraction significantly correlated with their concentrations in plasma (p > 0.05 for all comparisons; (Fig. S4 and Tables S6, S7). In addition, ceramide, dihydroceramide, and sphingomyelin total content within isolated EV was directly correlated to circulating white blood cells (R 0.591, 0.504, and 0.578; p < 0.05), contrary of plasma sphingolipids (Table S7).
We then investigated the contribution of 20 different species of sphingolipids to the total amount of differential expressed lipids in whole plasma versus isolated EV (Fig. 5). LC-MS/MS analysis showed that ceramide 18:1, and dihydroceramides 16:0 and 18:0 were not detectable in EV, whereas dihydroceramide 18:1 was not detectable in EV as well as in plasma. All the other ceramide, dihydroceramide, and sphingomyelin species except for ceramide 14:0 significantly increased after STEMI in EV (p < 0.05 for all comparisons; Table S5), with ceramide 16:0 being the most represented among ceramides, and sphingomyelin 16:0 among sphingomyelins. On the contrary, all sphingomyelin species and ceramide 24:0 increased after myocardial ischemia in plasma, whereas ceramides 14:0, 16:0, 18:1, 18:0, and 20:0 and dihydroceramides 18:0 and 24:0 significantly decreased in patients with a diagnosis of STEMI as compared to controls (Table S6).
To exclude confounding effects due to the presence of contaminants, we verified the presence of correlations between albumin or ApoA1 and sphingolipids levels in EV (Table S9 and Fig. S5). We found no significant correlations with sphingolipid levels. Furthermore, patients were stratified for fasting conditions (Table S10) so that levels of lipids were re-analyzed in STEMI patients at pre-PCI stage according to this stratification. We found no difference between patients with STEMI in fasting conditions as compared to those sampled at not fasting conditions (p > 0.05 for all comparisons).

Diagnostic modelling.
To verify the diagnostic relevance of the three main classes of lipids detected in EV over those enriched within whole plasma, ROC analysis was performed, and AUC was calculated. The gold standard biomarker for STEMI, hs-troponin was used as reference. Except for sphingomyelin, AUC of www.nature.com/scientificreports/ EV lipids was higher as compared to whole plasma. AUCs for EV-derived ceramides and dihydroceramides, as well as sphingomyelins from both EV and plasma, were not different from hs-troponin (p > 0.005; Fig. 7a-d).
Moreover, EV levels of ceramides, dihydroceramides, and sphingomyelins were combined in a specific signature using supervised machine learning algorithms via linear discriminant analysis. Canonical plot showed that this approach correctly discriminated all the patients with a diagnosis of STEMI (before and after primary PCI), from controls (Fig. 7e). Collectively, these data suggest the relevance of EV lipid signature as biomarker compared to hs-troponin for the diagnosis of STEMI.

Discussion
This study addresses sphingolipid profiling of circulating EV in patients with a diagnosis of STEMI, before and 24-h after primary PCI against control subjects as promising theranostic tool. Specifically, here we applied lipidomics with supervised machine learning algorithms to pinpoint EV sphingolipid signature in order to accurately discriminate patients with STEMI before and after PCI and controls. Our group recently described www.nature.com/scientificreports/ that in a preclinical rat model of MI, the large majority of sphingolipids species are increased in EV derived from plasma 27 . Here, we further confirmed modulation of lipid EV membrane in cardiovascular patients as well. Ceramide, dihydroceramide, and sphingomyelin EV content was significantly increased in patients with STEMI at pre-PCI evaluation, and correlated to the highest hs-troponin reached by each patient and to the ejection fraction at echocardiography 24 h post-PCI, as indices of myocardial injury severity. The levels of three lipid species in EV were not inferior to troponin as potential diagnostic biomarker for acute myocardial infarction. Regarding whole plasma composition, as for ceramides and dihydroceramides, the differences between patient and controls subjects revealed a trend in their decrease.
Sphingolipids are one of the major lipid constituents of eukaryotic cells and, beside their structural role, they also function as signaling molecules for maintenance of cell membrane integrity, apoptosis, cellular damage or oxidative stress response, cell growth, senescence and migration, endothelial cell function, and inflammatory signaling 13,[33][34][35] . Thus, profiling of circulating sphingolipids in patients may offer relevant information and allow identification of specific disorders and disrupted metabolic pathways 35 . Such diagnostic approach can be further implemented by recent technological advances in mass spectrometry (MS), which allows precise characterization of different lipid molecules 2 . Here, we provided additional sensitivity by quantifying sphingolipid species of circulating EV. In STEMI patients, systemic EV are mainly composed of vesicles that are released in the blood as result of acute thrombosis and inflammation 36 . Accordingly, we observed the expression of surface antigens from activated platelets (CD62p, CD41b, and CD42a), endothelium (CD31), and leukocytes (CD40), as previously described 36 . A minor proportion was positive for markers of inflammatory cells (e.g. CD2, CD3, CD4, CD8, CD14, CD19, CD20, CD45). A crucial aspect in characterizing circulating EV is represented by the cross-contamination with lipidic particles. To overcome this issue and to minimize the presence of contaminants without affecting the yield of EV enabling consistent MS analysis, we applied a double series of centrifugation steps interspaced by two washing steps. This allowed us to isolate plasma EV with reasonable specificity as demonstrated by western blotting analysis, in line with previous study showing that UC method still represent a suitable method for isolating EV from plasma 26 . Consistently, apolipoprotein A1 and albumin were not correlated to sphingolipid levels, and fasting conditions did not modify lipid profile in patients with STEMI.
Healthy individuals usually present a concentration of circulating plasma EV ranging between 10 7 and 10 9 EV/mL 37 , whereas lipoproteins are about 10 16 particles/mL 22 . Concomitantly, we found that sphingolipids were significantly more abundant in plasma than in EV (Tables S5 and S6). However, the significant increase in sphingolipids observed in EV of STEMI patients at pre-PCI stage, was mirrored in plasma by sphingomyelins (C16 to C24), and by ceramide C24:0 species. In STEMI patients at pre-PCI evaluation, EV sphingomyelin species (C16 to C24) and ceramide 24:0 were increased, as compared to controls, whereas they decreased after reperfusion. This is consistent with previous studies demonstrating an increase of circulating sphingomyelins 16:0 and 24:0 in patients with atherothrombotic diseases, such as MI or ischemic stroke 5,38 . Indeed, some sphingolipids species are enriched in atherosclerotic plaques 39 , thus possibly suggesting a role in their pathogenesis. Similarly, ceramides 16:0, and sphingomyelin have been previously associated with negative cardiovascular outcome and worsening of cardiac function in heart failure 1 . Moreover, the plasma signature of specific ceramides and dihydroceramide species has been proven as prognostic reference for major adverse cardiovascular events in patients with acute CAD 7,8,40 .
In our study we further confirmed such findings for plasma levels of sphingomyelin, but neither for ceramide (C14 to C22) nor dihydroceramide (C16 to C24). Although the concentration sphingolipid species in control  www.nature.com/scientificreports/ group is in line with the reference range of circulating sphingolipid in human plasma from healthy subjects 41 , our study suffers of limited numerosity of included patients, which can partially explain this discrepancy. Moreover study design and blood sampling modality (different anticoagulants) may influence LC-MS/MS results 41,42 . Notably, the increase of different ceramide species has been shown to concur to the upregulation of different acute phase proteins, and to leucocytes recruitment and activation [43][44][45] . Moreover, it is widely acknowledged that S1P possesses pleiotropic functions and orchestrates many cellular processes involved in cardiovascular pathophysiology 46 . This is consistent with our findings showing a direct correlation between sphingolipid EV content and leucocyte count. Inflammatory environment may also partially explain the increase in overall sphingolipids species carried by EV, as the secretion of inflammatory cytokines (interferon-γ, TNF-α, and IL1β) together with cellular stress, stimulate the de-novo synthesis of ceramides through the enzymatic activity of the dihydroceramide desaturase, which catalyze the conversion of dihydroceramides in ceramide 13,47,48 .
Sphingolipids and EV are not only measurable indicator of some biological state or condition in cardiovascular disease, but also causative agents in CAD progression. Growing evidence suggests that the blood levels of ceramides are associated with the exacerbation of myocardial ischemic disease and its complications 7,40,49 . Studies on murine models revealed that pharmacological inhibition of the synthesis of ceramides prevents heart failure after myocardial ischemia, decreasing ventricular remodeling, fibrosis, and inflammatory infiltrate 8,43,50 . Similarly, the administration of an inhibitor of neutral SMases which hydrolyzes sphingomyelin to ceramide, thus impairing EV biogenesis, namely GW4869, reduced the increase of circulating inflammatory EV after MI in rats, resulting in the preservation of left ventricular ejection fraction 27 .
To the best of our knowledge, this is the first study combining a protocol to isolate EV that meets current recommendation in terms of purity 32,51 , a robust targeted lipidomic MS-based platform for the accurate evaluation of less-abundant lipid species, and a machine learning algorithm to model a specific sphingolipid signature in www.nature.com/scientificreports/ STEMI patients. In such prospective, we demonstrated that EV sphingolipids can provide a theranostic platform to gain new information about disease status. This represents critical information to infer insights into developing novel biomarkers. This is further strengthened by considering that such approach is significantly different from traditional cardiovascular diseases lipid biomarkers such as cholesterol 6 . Nonetheless, limitations should be acknowledged as well. First of all, limited number of patients enrolled in the study, which will have to be replicated in larger cohorts. Plasma is a complex biofluid and we acknowledged residual albumin contamination in EV samples, which might affect EV count by NTA, and therefore should be taken in consideration; nevertheless, circulating EV increase in STEMI patients has been largely documented 52,53 . Moreover, antithrombotic treatment administered to patients with MI could have affected sphingolipid composition. However, all STEMI patients received standard treatment (i.e. dual anti-platelets therapy and unfractioned heparin); hence, we can assume that the comparison between STEMI patients at pre-PCI vs. 24 h post-PCI could not have been biased. Finally, the use of ultracentrifugation is not clinically feasible; therefore alternative, yet highly efficient, isolating method should be optimized to be translated into clinical practice.
In conclusion, this is the first description of a dynamic change in EV sphingolipid signature after STEMI in humans, thus attracting a considerable interest in the context of personalized medicine for patients with CAD. In such perspective, EV sphingolipid profiling before and after reperfusion could be used as prognostic tool for STEMI patient. In such scenario, circulating EV modulation may be used as therapeutic strategy to improve their prognosis.
Methods patient selection and blood sampling. Subjects were enrolled at the Fondazione Cardiocentro Ticino (Lugano, Switzerland). The study protocol complied with Helsinki Declaration and was approved by the "Ticino Cantonal Ethical Committee", Switzerland, and informed written consent was obtained from each participant. Patients were diagnosed with STEMI according to the European Society of Cardiology guideline 29 ; peripheral blood samples were collected on the presentation to the emergency department before primary PCI and after 24 h from reperfusion (named as pre-PCI and post-PCI samples). Patients were excluded in case of: (1) Age > 85 years; (2) Chest pain onset > 6 h; (3) Glomerular filtration rate < 30 mL/min; (4) Atrial fibrillation, ventricular tachycardia, ventricular fibrillation, cardiac arrest, or cardiogenic shock with indication to invasive device assistance; (5) Other non-ischemic cardiac disease (severe heart valve disease, chronic heart failure), or acute/chronic inflammatory disease (e.g. auto-immune disease, cancer, infections). Asymptomatic subjects underwent computed tomography or magnetic resonance imaging which documented the absence of significant coronary stenoses (> 30% diameter reduction) and were recruited as controls. All patients received a standard treatment with dual anti-thrombotic therapy (aspirin 150-300 mg plus either clopidogrel 300-600 mg, or prasugrel 60 mg, or ticagrelor 180 mg) and unfractioned heparin 70-100 IU/Kg. Control subjects, STEMI patients at post-PCI evaluation and 4 of 7 STEMI patients at pre-PCI evaluation were sampled in fasting conditions. EV isolation. Blood was collected in sodium citrate tubes and immediately centrifuged at 1600×g for 15 min at 4 °C to avoid platelet activation and separate plasma from cellular components. All hemolytic plasma samples were excluded from the study. Free-platelet plasma was differentially centrifuged at 3000×g for 20 min to remove cellular debris, and then at 10,000×g for 30 min, and 20,000×g for 15 min to remove apoptotic bodies and large particles 27 . EV were obtained from 300uL of free-platelet plasma by ultracentrifugation at 100,000×g (18 h) using a Beckman Optima Max-TL ultracentrifuge (Beckman Coulter). Plasma EV enriched pellet were resuspended in PBS 1X and underwent washing steps by further centrifuge steps at 10,000×g for 30 min, and 20,000×g for 15 min, with a final step of ultracentrifuge at 100,000×g (18 h) (Fig. 1a) as previously described 27 . Plasma-EV pellets were resuspended in 100 μL PBS, pH 7.4, and then stored at − 80 °C, prior to analysis. EV characterization. EV concentration and size were measured by nanoparticle tracking analysis (NTA), using NanoSight LM10 (Malvern Instruments, United Kingdom) equipped with a 405 nm laser and NTA 2.3 analytic software. Free-platelet plasma and EV samples were diluted 1:2000 in PBS 1X and exposed to a laser light source. Brownian movements of EV were recorded by a camera and nanoparticles size and number per mL were calculated by Stokes-Einstein equation. Three videos of 60 s were recorded for each sample to perform the analyses. Nanoparticles concentration was reported as number of particles per mL (n/mL).
After ultracentrifugation, purified EV were analysed by MACSPlex human Exosome Kit (Miltenyi, Bergisch Gladbach, Germany) with MACSQuant Analyzer-10 flow cytometer (Miltenyi, Bergisch Gladbach, Germany), as previously described 36,54 . EV samples were incubated overnight with capture-beads coated with specific antibodies for 37 EV surface antigens, and then with detection antibodies against CD9, CD63, CD81. The median fluorescence intensity (MFI) was analyzed after normalization for blank control.

Lipid extraction and sphingolipid content quantification. Sphingolipid extraction and targeted LC-
MS/MS analysis were performed as previously described 30,55 . Sphingolipids were assayed in 25 µL of plasma and in EV deriving from 300 µL of plasma. Both EV pellet and plasma were diluted to 100 µL water and after the addition of 850 µL methanol/chloroform mixture (2:1 v/v), samples were incubated overnight in an oscillator bath at 48 °C. Then, to enhance their recovery, alkaline methanolysis was performed by incubation at 37° for 2 h with 75 µL of potassium hydroxide 1 M in methanol. After neutralization with 75 µL of acetic acid 1 M in methanol, samples were evaporated. The residues were dissolved in 100 µL of methanol, centrifuged for 10 min at 13,400 RPM, and withdrawn in a glass-vial. Finally, samples were analyzed by LC Dionex 3000 Ulti-Mate (ThermoFisher Scientific) coupled to a tandem mass spectrometer AB Sciex 3200 QTRAP (AB Sciex). The separation was achieved by reversed-phase chromatography either using BEH C8 100 × 2.1 mm × 1.7 μm (for ceramides, dihydroceramides and sphingomyelins) or Restek Raptor C18 2.1 × 100 mm × 2.7 μm (for sphingoid bases) by mixing eluent A (0.2% formic acid 2 mM ammonium formate water-solution) and eluent B (methanol 0.2% formic acid 1 mM ammonium formate). Quantitative analysis was performed interpolating each peak area of analyte/area IS with calibration curve for each sphingolipid. Additional LC-MS/MS details can be found in the Supplementary materials (Supplementary Tables S3, S4). Sphingolipids concentration in EV were related to 1 mL plasma.
Statistics. IBM SPSS Statistics 22 (IBM Corp., Armonk, New York, USA), Python 3.5 (library, scikit-learn), and GraphPad PRISM 7.0a (La Jolla, California, USA) were used for analyses. The distribution of each parameter was assessed by Kolmogorov-Smirnov test. Normally distributed parameters were expressed as mean ± standard deviation (SD) and analyzed by T-student test or ANOVA test with post-hoc Bonferroni test for multiple comparisons. Non-normally distributed parameters were expressed as median [interquartile range] and analysed by Mann-Whitney's or Kruskal-Wallis's tests. Categorical variables were compared through chi square or Fisher tests. Correlations were assessed by Pearson's R test. P-values < 0.05 were considered significant. The analysis of ROC curves was used to assess the AUC and derive the diagnostic performance in STEMI patient discrimination.
Linear discriminant analysis (LDA) was used as strategy for features reduction to build the canonical plot and evaluate the discrimination power of EV sphingolipid composition as signature to classify patients with STEMI. LDA employs linear combinations of variables to maximize the separation between groups by increasing precision estimates with variance reduction 36 . The algorithm computes a set of coefficients for linear combination of each variable to estimate the diagnosis. The axes of the canonical plot are calculated by the LDA from weighted linear combinations of variables included in the model; each patient is indicated by a point. The crosses indicate the means of (canonical 1; canonical 2) for each group, the ellipses include patients with a linear combination coefficient that falls within the mean ± SD.