Uneven metabolic and lipidomic profiles in recovered COVID-19 patients as investigated by plasma NMR metabolomics.

COVID-19 is a systemic infectious disease that may affect many organs, accompanied by a measurable metabolic dysregulation. The disease is also associated with significant mortality, particularly among the elderly, patients with comorbidities, and solid organ transplant recipients. Yet, the largest segment of the patient population is asymptomatic, and most other patients develop mild to moderate symptoms after SARS-CoV-2 infection. Here, we have used NMR metabolomics to characterize plasma samples from a cohort of the abovementioned group of COVID-19 patients (n = 69), between 3 and 10 months after diagnosis, and compared them with a set of reference samples from individuals never infected by the virus (n = 71). Our results indicate that half of the patient population show abnormal metabolism including porphyrin levels and altered lipoprotein profiles six months after the infection, while the other half show little molecular record of the disease. Remarkably, most of these patients are asymptomatic or mild COVID-19 patients, and we hypothesize that this is due to a metabolic reflection of the immune response stress.

one dimensional

area under the curve

Carr‐Purcell‐Meiboom‐Gill

α‐1‐acid glycoprotein N‐acetyl‐glucosamino (N‐acetyl) signal A

α‐1‐acid glycoprotein N‐acetyl‐glucosamino (N‐acetyl) signal B

high‐density lipoprotein‐cholesterol

low‐density lipoprotein‐cholesterol

never infected cohort

orthogonal projections to latent structures discriminant analysis

principal component analysis

recovered cohort

undistinguishable from NI

statistically different from NI

receiver operating characteristic

supramolecular phospholipid composite

total cholesterol

trimethylsilyl propionate

uroporphyrin I

INTRODUCTION

COVID‐19 disease is caused by the infection of the coronavirus SARS‐CoV‐2 and currently represents a first‐order worldwide health crisis. Symptomatic patients first undergo an acute phase, which, in turn, can be divided into three well differentiated stages: early infection, characterized by mild symptomatology; a pulmonary phase, which may be accompanied by shortness of breath, and eventually a severe hyperinflammatory phase, which can lead to acute respiratory distress syndrome and/or heart failure. At the beginning of the pandemic these clinical features suggested the notion that COVID‐19 disease was essentially a severe form of pneumonia. Yet, it is nowadays widely accepted that SARS‐CoV‐2 exerts a systemic infection, that can affect almost all the vital organs of the body, including neurological inflammation, renal damage, liver dysfunction and increased atherosclerosis risk, among others.

The vast majority of COVID‐19 patients experience asymptomatic or mild infectious processes. Also, most patients who undergo a moderate or severe acute phase, pharmacologically treated or not, recover without showing further symptoms. Yet, a growing body of evidence indicates that a non‐negligible segment of people infected by SARS‐CoV‐2 (between 10% and 30%) do not return to a normal/healthy state, establishing a new deteriorated health baseline. This group includes recovered COVID‐19 patients who have a recurrence of PCR‐positive tests due to the possibility that SARS‐CoV‐2 RNAs can be reverse‐transcribed and integrated into their DNA, but also other patients who retain some symptoms and/or have developed novel long‐term effects after an asymptomatic infectious phase, grouped in the so‐called long‐term COVID‐19 or post‐acute COVID‐19 syndrome. , , A recent high‐dimensional characterization of post‐acute sequelae of COVID‐19 underlines the extent and the severity of these long‐term complications. Nevertheless, the molecular basis for the long‐term evolution of the COVID‐19 disease is much less understood.

Considering its systemic character, the underlying molecular mechanisms and the associated metabolic alterations of the SARS‐CoV‐2 infection have been extensively studied. , , , , , The group of Jeremy Nicholson and our group, among others, have shown that COVID‐19 disease dysregulates the metabolic and lipidomic profiles of serum: the lipoprotein distribution suggested an increase of the atherosclerotic risk of the infection, and the high levels of ketone bodies (acetoacetic acid, 3‐hydroxybutyric acid and acetone) and 2‐hydroxybutyric acid could induce liver damage associated with dyslipidaemia and oxidative stress. Moreover, porphyrin levels are abnormally elevated during the acute phase of COVID‐19 infection and they could be used as a potential biomarker for the disease. Furthermore, amino acids, biogenic amines and the tryptophan pathway are also largely misregulated, constituting a truly molecular marker of the disease.

Metabolic phenoreversion on a cohort of 27 non‐hospitalized patients has been recently analysed using a combination of NMR and LC/MS spectrometry to show that, remarkably, a subset of these patients still display a plethora of persistent biochemical pathophysiology, suggesting that the metabolic phenotyping approach may be suitable in the assessment of post‐infection COVID‐19 patients. Here, we have used NMR‐based metabolomics including porphyrin characterization on a large cohort of plasma samples to investigate the recovery of COVID‐19 patients (n = 69) and compared them with a similar large‐size group of reference samples (n = 71), from non‐infected subjects. The goal of the study was to check to what extent the phenotypical recovery associated with the lack of symptoms in the cohort under consideration also reflects metabolic phenoreversion.

METHODS

Recruitment of patients

The analysed plasma samples from recovered COVID‐19 patients (n = 69) come from a cohort of patients diagnosed by RT‐PCR assay targeting viral RNA from nasal swab samples and/or by IgG detection assay in blood. All patients recovered from the acute infection at least 3 months after the acute phase of the disease and 6 ± 2 months on average. Recovered patients were divided into disease severities: asymptomatic, mild (no hospitalization required), moderate (hospitalization required) and severe (ICU hospitalization required). The plasma samples of the healthy donors (n = 71) were collected after testing negative in the abovementioned SARS‐CoV‐2 assay. Both cohorts were collected within the same geographical region (Basque Country, Spain). The procedures, mainly temperature, clotting tube and clotting time, were the same.

All samples were collected through the Basque Biobank (http://www.biobancovasco.org) under a protocol approved by the Basque Committee of Ethics and Clinical Research (PI2020063). The methods were carried out in accordance with the approved guidelines. The Basque Biobank complies with the quality management, traceability and biosecurity, set out in Spanish Law 14/2007 of Biomedical Research and in Royal Decree 1716/2011. All study subjects provided written informed consent.

Sample preparation and 1H NMR acquisition

Blood samples were collected into commercially available EDTA‐treated tubes. Cells and platelets were removed from plasma by centrifugation for 15 min at 2000 g using a refrigerated centrifuge. The obtained plasma samples were stored at −80 °C until measured. Each aliquot of plasma sample was thawed at room temperature for 30 min. NMR samples were then prepared using a SamplePro Tube (Bruker BioSpin, Rheinstetten, Germany) robot system for liquid handling with integrated temperature control. Every sample was then (automatically) mixed with buffer (75 mM Na2HPO4, 2 mM NaN3, 4.6 mM sodium trimethylsilyl propionate‐[2,2,3,3‐2H4] (TSP) in 20% D2O, pH 7.4 ± 0.1) , at a 1:1 (v/v) ratio and a 5 mm NMR tube was filled with 600 μL of the mix. After manually shaking every sample for a few seconds, the NMR tubes were stored at 5 °C inside a tempered SampleJet automatic sample changer mounted on a 600 MHz in vitro diagnostic research (IVDr) spectrometer (Bruker BioSpin).

The samples of the control cohort (preCOVID; n = 280, serum) were collected in 2018/2019 in overnight fasting conditions, well before the start of the current COVID pandemic, by Osarten Kooperatiba Elkartea (Mondragon Cooperative) as extra aliquots during the routine annual medical tests of their employees in the Basque Country by OSARTEN.

The NMR spectrometer, equipped with a double resonance broadband probe (BBI) probehead with a z gradient coil and BOSS‐III shim system, was calibrated daily following strict standard operation procedures (SOPs) to ensure the highest spectral quality and reproducibility. All 1H‐NMR spectra were measured at 310 K and referenced to the TSP signal (0 ppm). Three different 1H NMR experiments were recorded per sample: a standard one‐dimensional (1D) 1H NOESY spectrum (noesygppr1d) with water presaturation, a 1D 1H Carr‐Purcell‐Meiboom‐Gill (CPMG) experiment (cpmgpr1d) implementing a T 2 filter to suppress the broad signals of proteins and other macromolecules, and a two‐dimensional J‐resolved experiment (jresgpprqf). The list of quantified metabolite or lipoprotein subclasses were acquired with Bruker's B.I. Quant‐PS and B.I. LISA methods (Bruker BioSpin), respectively. The α‐1‐acid glycoprotein N‐acetyl‐glucosamino (N‐acetyl) signal integrals were calculated as GlycA and GlycB in 2.06 ppm and 2.09 ppm, respectively, from the noesygppr1d experiment. Likewise, the signal at 3.23 ppm (SPCtotal) is a supramolecular phospholipid composite signal containing the N+‐(CH3)3 choline headgroup signals of lysophosphatidylcholines and other phospholipids. The identification of some metabolites in the 1D spectra was done using the Human Metabolome Database (HMDB) and Chenomx NMR Suite 8.6 software (Chenomx, Edmonton, Canada).

Porphyrin extraction and quantification

For porphyrin extraction, 100 μL of 6 M HCl was added to 100 μL of the healthy (n = 28) and COVID‐19 recovered (n = 40) plasma sample and incubated at 37 °C with shaking for 30 min. The samples were then centrifuged for 10 min at 10 000 g. The supernatant was transferred to a centrifuge tube with a cellulose acetate membrane filter of 0.22 μm pore size (Corning Costar Spin‐X, Corning, NY, US), and centrifuged for 10 min at 4000 g. All samples were immediately analysed using HPLC with SIL20AC HT multisampler, CTO‐10AS VP, LC‐20 AD, CBM‐20A and SPD‐M20A modules (Shimadzu, Kyoto, Japan).

Porphyrins were separated using a 60 min gradient elution with a two‐component mobile phase consisting of ammonium acetate (1 M, pH 5.16, Solvent A) and 100% acetonitrile (Solvent B) in a BDS Hypersil C18 (250 × 3 mm2; 5 μm particle size) column (Thermo Scientific, Waltham, MA, US). Gradient elution started upon injection at 0% B, which then increased to 65% B over 30 min, remained at 65% B for 5 min, returned to 0% B over 15 min, and remained at 0% B for 10 min in order to re‐equilibrate the column before the next sample injection. The flow rate was 1 mL/min. All analyses were performed at 20 °C with fluorescence detection at an excitation wavelength of 405 nm and emission wavelength of 610 nm. A standard commercial lyophilized porphyrin kit containing uroporphyrin I (URO I) dihydrochloride, HEPTA I heptamethyl ester, HEXA I hexamethyl ester, PENTA I pentamethyl ester and COPRO I tetramethyl ester (Frontier Scientific, Logan, UT, US) was used to calibrate the porphyrins in the chromatogram.

Statistical analysis

To facilitate a general overview over negative and recovered samples, principal component analysis (PCA) was used as a dimensionality reduction method to represent the first and second principal components in one score plot, taking plasma metabolites and lipoprotein subclasses as input variables after unit variance scaling.

A supervised classification model was built using orthogonal projections to latent structures discriminant analysis (OPLS‐DA) to distinguish between healthy donors and recovered patients, again using metabolites and lipoprotein subclasses as variables. A stratified sevenfold cross‐validation procedure was used to determine the number of orthogonal components using the area under the receiver operating characteristic curve (AUROC) as the performance metric. In all cases, the aim of the model was to detect which recovered people are clearly distinguishable from healthy donors, not to achieve the maximum classification performance. For each analysed metabolite or lipoprotein subclass, comparisons between pairs of groups (eg recovered versus negatives) were performed using two‐tailed independent Student t‐tests. Unadjusted p‐values were considered as statistically significant below 0.05. To account for the effect size, binary logarithms of fold changes (log2 FC) between group means were calculated. Standard errors for log2 FC were estimated through bootstrap resampling. Individual metabolite or lipoprotein subclass analyses were summarized in forest plots, where each variable is represented by its log2 FC value with its standard error and statistical significance.

To obtain a deeper understanding of the behaviour of individual samples in the comparison of recovered versus negatives, violin plots were used to represent data distribution and probability density. They included data points and internal boxplots representing median and interquartile range. A small random movement was applied to each point to minimize overplotting.

GlycA, GlycB and SPCtotal were estimated from integration of their specific bins (2.06, 2.09 and 3.23 respectively) after a bucketing process of NOESY spectra. Bucketed spectra (290 bins of 0.03 ppm width between 9.5 and 0.5 ppm excluding residual water, 4.7‐5.0 ppm) were normalized by total intensity in order to minimize the effect of different concentrations.

All analyses were performed using the R statistical software, Version 3.6.0 (http://cran.r-project.org/) and the following R packages: ade4 (Version 1.7‐16), factoextra (1.05), ggforestplot (0.1.0), metabom8 (0.4.4) and tidyverse (1.3.0).

RESULTS

Uneven metabolic phenoreversion in recovered COVID‐19 patients

To investigate the putative residual metabolic and lipidomic abnormal profiles after a long recovery period following the SARS‐CoV‐2 infection (6 ± 2 months), we used NMR metabolomics over a set of plasma samples from a cohort of 69 COVID‐19 patients (recovered, RE). As a control group we used plasma samples from a cohort of 71 healthy donors, collected at the same time and under the same conditions (never infected, NI). Table S1 reports the general characteristics of the cohorts, which are adequately matched in terms of age, but where males are dominant.

1D 1H NMR‐metabolomics spectra of plasma allowed the quantification of up to 41 metabolites and 112 lipoprotein subclasses with Bruker's B.I. Quant‐PS and B.I. LISA methods, as described in Section 2. An unsupervised PCA of the full list of metabolites and lipoprotein subclasses showed a significant overlap of the RE and NI groups (Figure 1A), suggesting metabolic and lipidomic phenoreversion after six months of SARS‐CoV‐2 infection. However, an orthogonal partial least‐squares discriminant analysis (OPLS‐DA) of plasma metabolites and lipoproteins depicted a much more complex scenario (Figure 1B). In fact, OPLS‐DA analysis aim can discriminate the NI group from the ensemble, since the model is set to correctly allocate 90% of NI donors to the predicted NI group (Figure 1C). Yet, the same model only identifies 51% of patients from RE group as showing an equivalent metabotype to NI, while the other 49% of RE patients have a distinctive metabotype. Thus, our sorting model revealed two different groups and prompted us to separate the RE cohort into two sub‐cohorts: RE_I (indistinguishable from NI) and RE_II (statistically different from NI). To demonstrate that RE_I is indistinguishable from NI we rebuilt the OPLS‐DA model without RE_I samples. This new model was able to achieve better classification performance (area under the curve, AUC = 0.92, Figure S1) and to correctly classify as NI the RE_I subgroup (Figure S2).

FIGURE 1

Multivariate unsupervised (PCA) and supervised (OPLS‐DA) analyses. A, Score plot of the first two principal components from PCA between NI (green) and RE (red) cohorts when considering all plasma metabolites and lipoprotein subclasses. Ellipses surround the area that includes 95% of confidence. B, Score plot from OPLS‐DA of metabolites and lipoprotein subclasses for NI (green) and RE (red) cohorts. The dashed vertical line corresponds to the 90% specificity to discriminate for the NI group, as indicated in the ROC curve from the OPLS‐DA model (C), where the cut‐point for t pred is indicated in red, with the corresponding specificity and sensitivity in parenthesis. Also, AUC and its 95% confidence interval are indicated

The GlycA signal, corresponding to a mixture of glycoproteins, is a well known inflammation marker. In Figure 2A it is shown that a previously reported cohort of COVID‐19 patients in acute phase presented elevated levels of the GlycA signal when compared with the healthy volunteers (preCOVID) but also when compared with other pneumonia patients. As compared with them, the NI group showed comparable basal levels of the inflammatory marker, while RE_I and RE_II groups are compatible with a model that still retained some inflammation (with the GlycA signal slightly more elevated, on average, for the RE_II group).

FIGURE 2

Inflammation markers: GlycA and uroporphyrins. A, Normalized GlycA values of every subject belonging to preCOVID, NI, RE_I and RE_II cohorts compared with equivalent analysis from people who suffered other pneumonias and COVID‐19 positive patients. , B, Normalized uroporphyrin levels for NI, RE_I and RE_II cohorts. Statistical differences between cohorts are represented as *** (p < 0.001) or **** (p < 0.0001)

Accumulation of porphyrins, despite their intrinsic variability, can be considered a subrogate biomarker for pneumonia and inflammation, and COVID‐19 patients have abnormally elevated porphyrin levels. URO I levels, as a representative example for porphyrin accumulation, showed similar (low) levels for NI and RE_I groups while they were elevated for the RE_II group (Figure 2B). The different results obtained in the URO I for the RE_I and RE_II sub‐cohorts underline their different phenoreversions, validating the independent analysis.

Metabolomics and lipidomics of recovered COVID‐19 patients

We first investigated the redistribution of plasma lipoproteins, by comparing the RE_I and the RE_II groups with the NI reference group. The RE_I cohort reflected a significant lipoprotein recovery, matching the healthy values in every lipoprotein subclass (Figure 3 and Table S3). Yet, for the VLDL subfractions, the triglyceride TG‐VLDL, free cholesterol VLDL and VLDL phospholipid subfractions had a tendency to remain elevated, and while this was not found to be statistically significant it is clearly reminiscent of the observations found for the COVID‐19 patients during the acute phase of the disease.

FIGURE 3

Plasma lipoprotein distribution in COVID‐19 RE_I (blue) and RE_II (orange) sub‐cohorts. Forest plot representing the fold change as compared with the NI cohort. Statistically significant differences are represented with filled circles (p < 0.05). Horizontal bars are the standard errors

The lipoprotein profile remained much more altered in the RE_II sub‐cohort: almost each subfraction of lipoprotein, triglycerides, cholesterol and phospholipids had an increment of their content as compared with the NI reference cohort (Figure 3 and Table S3). Nonetheless, although free cholesterol, total cholesterol (TC), low‐density lipoprotein‐cholesterol (LDL‐C) and high‐density lipoprotein‐cholesterol (HDL‐C) were higher for the RE_II sub‐cohort (Figures 3 and 4 and Table S3), the LDL‐C/HDL‐C ratio was no different from that of the healthy donors. Similarly, in case of apolipoproteins, elevated values of Apo‐B and Apo‐A1 did not significantly increase their ratio, a validated predictor of coronary heart disease risk. Figure S3 shows the comparison for the lipoprotein profiles for RE_II and COVID‐19 patients in the acute phase, the latter previously reported. Interestingly, the RE_II profile tends to compensate the altered profile found in COVID‐19 patients, even trespassing the values found in the normal population for the HDL and LDL subclasses.

FIGURE 4

Detailed distributions (violin plots) of some relevant lipoproteins. Each plot shows the distribution of individuals for the NI (green), RE_I (blue) and RE_II (orange) cohorts, and the median value for each cohort. Statistical differences between cohorts are represented as **** (p < 0.0001)

We then quantitatively analysed the low‐molecular‐mass metabolites in plasma samples of COVID‐19 RE_I and RE_II groups, as compared with the NI group (Figure 5 and Table S3). The RE_I group showed recovery in all metabolites, with no statistical deviations from the NI group observed for any metabolite. In turn, the RE_II group showed some metabolite alterations typical of COVID‐19 patients in acute phase: elevated levels of acetone, glycine, valine and alanine, and acidolysis due to elevated levels of glutamic, lactic, pyruvic and succinic acids (Figure S4). For this group, the glutamine/glutamate ratio was slightly reduced as compared with the NI cohort, while Fisher's ratio , was back to normal values.

FIGURE 5

Plasma metabolite distribution in COVID‐19 RE_I (blue) and RE_II (orange) sub‐cohorts. Forest plot representing the fold change as compared with the NI cohort. Statistically significant differences are represented with filled circles (p < 0.05). Horizontal bars are the standard errors

DISCUSSION

Metabolic alterations are consubstantial to COVID‐19: patients in the acute phase of SARS‐CoV‐2 infection had liver dysfunction, dysregulation of macrophage activity, alterations in platelet degranulation and the complement system, dysregulated levels of lipid metabolites and an exacerbated decrease in amino acid metabolism, all of them reflected in alterations in basal metabolism. Hence, it is important to investigate to what extent the recovery of the infective phase of the disease is accompanied by a metabolic phenoreversion.

SARS‐CoV‐2 cannot directly infect hepatocytes but the immune dysregulation could be associated with liver injury and pathology, and the liver could be impacted by hypoxia and elevated cytokines (as reviewed in Reference 22), and it is now fully accepted that SARS‐CoV‐2 infection causes an inflammatory storm that could aggravate the disease. , In this work, patients who had recovered from SARS‐CoV‐2 infection still had signs of inflammation, with the presence of α‐1‐antitrypsin and other glycated proteins (GlycA), one of the known inflammatory biomarkers of the disease. We have observed a similar trend as Holmes et al, where healthy patients with no symptoms showed a lower concentration of GlycA while the unhealthy patients presented higher intensity of the corresponding NMR signals for this metabolite. Moreover, discharged patients (RE cohort) had GlycA values in between those of healthy individuals and the COVID‐19 patients undergoing acute phase.

The most intriguing observation is that the metabolic landscape of recovered patients is not homogeneous, in line with previous analyses. This result prompted us to separate the RE group into two subsets, using a similarity test with the NI group. This was achieved using a combination of NMR‐based metabolomics and the porphyrin analysis, which disclosed two almost equally populated sub‐cohorts by OPS‐DA analysis. The two sub‐cohorts represent patients showing a high degree of metabolic phenoreversion (the RE_I sub‐cohort) versus others (the RE_II sub‐cohort) who showed metabolic alterations other than the ones found in COVID‐19 patients during the acute phase. Despite the differences between the sub‐cohorts, we observed that some cholesterol subfractions, the VLDLs, were still upregulated 6 months after the infection for both groups (RE_I and RE_II). Consistently, Wu et al reported in 2017 that patients had altered lipid metabolism 12 years after having SARS coronavirus. These differences were observed in TC, phospholipids and their respective particles and sub‐particles, and they were reminiscent of those observed during the acute phase of COVID‐19 disease, which were associated with liver dysfunction. , ,

The RE_II sub‐cohort showed specific features, including an upregulation of many lipoprotein subfractions that were found to be depleted in acute COVID‐19 patients. Indeed, the profile shown in Figure S3 suggested that the lipoprotein metabolism for the RE_II sub‐cohort rebooted from the disease, but still does not show full recovery. That said, LDL‐C/HDL‐C ratios had no differences from the healthy donors, suggesting liver balance, and consistent with the asymptomatic/mild natural history found in these patients. Moreover, the Apo‐B/Apo A1 ratio had a similar level to the NI cohort, again suggesting that the cardiovascular risk was not increased.

Regarding the metabolite profile of recovered COVID‐19 patients, we had seen that all metabolite values returned to healthy values in the RE_I sub‐cohort. In turn, for the RE_II group, there were some that did not, such as ketone bodies, which are a product of the β‐oxidation of the fatty acids in the matrix of liver mitochondria and are exported to other tissues, where they are oxidized to obtain energy. These metabolites were significantly increased during the acute phase of the disease. Interestingly, 2‐hydroxybutyric acid, a marker of oxidative stress, returned to average values in all recovered COVID‐19 patients investigated here. RE_II patients also experienced acidolysis, with increased levels of lactic and pyruvic acid, among others. Pyruvate in hypoxic conditions undergoes an anoxic respiration, producing lactate and increasing the concentration of NADH, which could inhibit allosterically the pyruvate dehydrogenase (PDH) complex, reducing the oxidation from pyruvate to acetyl‐CoA. 2‐oxoglutarate, also elevated in the RE_II sub‐cohort, is converted to succinyl‐CoA and produces NADH using the α‐ketoglutarate dehydrogenase complex, which is sensitive to reactive oxygen species (ROSs) and can be inhibited by this oxidative stress. Altogether, these alterations in the RE_II group point to a mild impairment in the mitochondrial function and the central carbon metabolism.

The glutamine and glutamate values were elevated in RE_II patients, but the glutamine/glutamate ratio, related to skeletal muscle energy, liver damage and septic shock, , was in equilibrium. Alanine, glycine, dimethylglycine and valine were more elevated in the same cohort of patients, but the Fisher ratio, a strong indicator of liver dysfunction, , showed normal values.

When compared with the available dataset of general characteristics, it is clear that most of the subjects belonging to the RE_II group are either asymptomatic or mild COVID‐19 patients. This counterintuitive notion has already been reported, and it has been attributed to a metabolic expression of the immunological stress that the asymptomatic (or mild) patients underwent.

Finally, it is important to describe the limitations of the study. The cohort is mostly populated by males, reasonably distributed among the sub‐cohorts, but giving a biased and incomplete view of the metabolic phenoreversion associated with COVID‐19. Unfortunately, attempts to expand the cohort to correct this situation would go beyond the scope of this work. On the other hand, the study used NMR spectroscopy only to metabolically characterize the plasma samples. Some pathways such as kynurenine and tryptophan degradation, intimately related to the COVID‐19 metabotype, can only be properly investigated using LC‐MS. Also, the GlycA, GlycB and SPC signals were determined by using the bin integrals in the 1D‐NOESY spectrum, where the signal is convoluted with other lipoprotein and/or EDTA signals. Signal deconvolution or using the DIRE experiment would improve the accuracy in GlycA and GlycB determination. Finally, we had access to a limited number of general characteristics from the donors, with no knowledge of some important factors such as BMI, smoking and other lifestyle habits that could potentially confound the binomial classification. That said, we believe that our study underlines the complexity and the systemic character of the disease at the molecular level, and it may be helpful for study of the disease progression towards full recovery and/or long‐term COVID syndrome.

Table S1. General characteristics of the analyzed cohorts Never Infected and Recovered, and of the sub‐cohorts RE_I and RE_II.

Table S2. General characteristics of the sub‐cohorts RE_I and RE_II.

Table S3. Univariate analysis for all variables and both recovered subgroups. FC: fold‐change; SE: standard error. Adjusted P‐values (Adj. P‐value) were calculated with Benjamini‐ Hochberg method.

Figure S1. ROC plot from OPLS‐DA model of RE_II vs. NI. The cut‐point for tpred, indicated in red, was determined following the Youden index criteria. It also shows its correspondent specificity and sensitivity in parenthesis. Area under the curve (AUC) and its 95% confidence interval is indicated are indicated in black.

Figure S2. Score plot from re‐built OPLS‐DA model of RE_II vs. NI. Never Infected (green) and RE_II (red) cohorts were used to build the model. Scores for RE_I cohort (open blue triangles) were predicted with this model, showing that they are undistinguishable from NI cohort.

Figure S3. Plasma lipoproteins subclasses distribution in COVID‐19 RE_II (black) and COVID‐19 positive (red) patients. Forest plot representing the fold‐change variation as compared to each NI cohort. Position of the circles define the mean concentration values and statistically significant differences with respect to NI are represented with filled circles (p value < 0.05). Horizontal bars are the 95% confidence interval.

Figure S4. Plasma metabolites distribution in COVID‐19 RE_II (black) and COVID‐19 positive (red) patients. Forest plot representing the fold‐change variation as compared to each NI cohort. Position of the circles define the mean concentration values and statistically significant differences with respect to NI are represented with filled circles (p value < 0.05). Horizontal bars are the 95% confidence interval.

Click here for additional data file.