Overview of study populationThe demographic data and clinical features of the COVID-19 patients admitted to the medical ward (MW) and intensive care unit (ICU), together with the demographic data of healthy controls (HC), are described in Table 1 and Table S1. No significant differences regarding age or sex were observed between both groups of patients and healthy controls. Most patients in both groups, MW and ICU, presented previous pathologies (80.43% and 88.24%, respectively) and also some cardiovascular risk factor (71.74% and 73.53%, respectively). Of these, overweight or obesity and dyslipidemia were more prevalent among patients in ICU (13.04% vs. 29.41% and 19.57% vs. 44.12%, respectively), although this increase was only significant for dyslipidemia (p = 0.018). As expected, certain clinical parameters were significantly more frequent in the ICU patients at the time of hospital admission, such as dyspnoea (71.11% vs. 91.18%, p = 0.028), pneumonia (64.44% vs. 88.24%, p = 0.016) and decreased oxygen saturation (p = 0.002). Laboratory findings also tended to reflect the severity of the disease between both patient groups, showing non-normal levels of the blood parameters measured in the ICU group. Particularly, we observed a statistically significant increase for lactate dehydrogenase (LDH, 48.89% vs. 76.47, p = 0.013) and fibrinogen (42.22% vs. 67.65%, p = 0.025), and a significant decrease in the lymphocyte count (37.78% vs. 79.41%, p < 0.001). The average length of hospitalization was also significantly higher in ICU patients (27.59 vs. 42.85 days, p < 0.001), as well as the mortality rate due to the disease (15.91% vs. 44.12%, p = 0.006).Table 1 Demographic and clinical characteristics of patients with COVID-19 admitted to MW and ICU, and demographic characteristics of HC.MW and ICU times associated to clinical variablesA bivariate analysis was performed in order to determine whether certain clinical characteristics (cardiovascular risk factors and laboratory results) were associated with length of hospital stay (Table S2). We observed that patients with high ferritin or C-reactive protein levels required more time in ICU than those with normal values, being the mean number of days 32.50 vs. 16.17 for ferritin measurements (p = 0.021) and 30.78 vs. 11 days for C-reactive protein levels (p = 0.048). Likewise, high ferritin concentration was significantly associated with longer hospitalization time in patients admitted to the medical ward (27.55 vs. 10.25 days, p = 0.023), those admitted to the ICU (47.64 vs. 25.83 days, p = 0.03), as well as in both groups together (37.08 vs. 15.44 days, p = 0.002). Furthermore, considering all COVID-19 patients, we observed that a lower number of lymphocytes in the blood was significantly associated with a longer hospital stay compared to those patients with a normal lymphocyte count (36.42 vs. 26.47 days, p = 0.015) and the same trend was observed in patients admitted to the ICU (47.67 vs. 28.86 days), although this difference did not reach statistical significance (p = 0.064). These observations are consistent with previous studies in which elevated ferritin and C-reactive protein, and decreased absolute lymphocyte counts were associated with critical disease, unfavorable evolution and increased mortality due to COVID-1922,23.Similarly, it has been demonstrated that the cardiometabolic status has an important impact on the clinical outcome of these patients24. Disorders as hypertension, dyslipidemia, obesity and diabetes mellitus have been shown to increase the risk of severe COVID-19 and mortality25. Accordingly, in our study, having one or more cardiovascular risk factors (including hypertension, obesity/overweight, dyslipidemia, and/or diabetes mellitus) was associated with longer hospital stay in the group of patients admitted to the ward (26.00 vs. 14.25 days, p = 0.022). In fact, patients with hypertension in this group spent a mean of 27.38 days in hospital compared to those with normal values, whose stay was 16.84 days (p = 0.015). In addition, when we analyzed all COVID-19 patients, we found that having dyslipidemia was also associated with a longer time in hospital (45.5 vs. 25.92 days, p = 0.009). These results reflect the importance of these variables in the clinical setting.Metabolic differences between COVID-19 patients and healthy controlsTo investigate the possible differences between COVID-19 patients and healthy controls, an NMR metabolomic approach was applied first and used over a set of serum samples and then over their corresponding urine aliquots, collected in both cases under the same conditions. All the metabolite assignments of both serum and urine samples are shown in representative 1H NMR spectra of each matrix on Figure S1A and Figure S1B and are described in Tables S3 and S4, respectively. It was possible the identification of 32 metabolites in serum samples, and 55 in the case of urine, falling into various classes of compounds, including amino acids, ketone bodies, organic acids, sugars, fatty acids, among others.First, an unbiased statistical analysis for the comparison of COVID-19 patients and healthy controls was performed by employing Principal Component Analysis (PCA) to 1H NMR data of each matrix. In Figs. 1A and 2A appear both PCA scores plots from serum and urine data, respectively. A slight clustering trend between the COVID-19 and control samples was observed. Therefore, to emphasize the differentiation between the groups and unveil potential biomarkers, supervised analysis using Orthogonal Partial Least Squares Discriminant Analysis (OPLS-DA) was applied to each 1H NMR data set. OPLS-DA model correlates the information of the samples, in this case, the COVID-19 or Control groups, with the spectral information obtained through NMR. The OPLS-DA scores plot (Figs. 1B and 2B for serum and urine, respectively) displayed a remarkable discrimination between samples from different groups in both matrices. Moreover, the Important Features Analysis, which reveals the discriminating molecules responsible for the separation of the spectroscopic profiles through Variable Importance in the Projection (VIP) score analysis (VIP > 1), identified specific metabolites such as trimethylamine-N-oxide (TMAO), phenylalanine, N-acetylglycoproteins (NAG), tyrosine, lysine, acetone, mannose, citrate, glycerol, and fatty acids, particularly unsaturated fatty acids (UFA), as the key molecules driving the separation between COVID-19 patients and healthy controls in the serum metabolome (Fig. 1C). Otherwise, Fig. 2C shows the Important Features Analysis with VIP > 1 for urine metabolites, that displays methionine, formate, fucose, arabinose, N-phenylacetylglycine, lysine, 3-methylhistidine, trigonelline, hippurate, 2-phenylpropionate, glutamine, creatinine, 3-indoxylsulphate and pseudouridine, as metabolites responsible for urine metabolome differentiation into COVID-19 patients and healthy controls. In addition, evaluation tools from univariate analysis were also taken into consideration for both matrices, including Fold Change (FC) analysis, p-values and Receiver Operating Characteristic (ROC) analysis (see below).Fig. 1(A) PCA scores plot, (B) OPLS-DA scores plot and (C) important features (VIP-plot from OPLS-DA) distinguishing healthy controls and COVID-19 serum samples. Scaling was done to Pareto. R2X = 0.79, R2Y = 0.90, Q2 (cum) = 0.83, CV-ANOVA = 2.12 x 10-32 for OPLS-DA, which was validated through Permutation Test (Figure S2). Important features (VIP > 1) plot shows those buckets that contain metabolites that increase (in red) or decrease (in blue) its content for each group. (D) Significantly enriched metabolic routes by COVID-19 built with serum metabolites with VIP > 1 from OPLS-DA model: 1) Phenylalanine, tyrosine and tryptophan biosynthesis, 2) Phenylalanine metabolism, 3) Glycerolipid metabolism and 4) Arginine and proline metabolism. The more affected metabolic pathways appear with a color gradient, from yellow (less significant) to red (most significant) and satisfy the following criteria: number of matching metabolites in the pathway > 1, false discovery rate (FDR) adjusted p < 0.05 and impact index > 0 (see Table S5).Fig. 2(A) PCA scores plot, (B) OPLS-DA scores plot, and (C) important features (VIP-plot from OPLS-DA) distinguishing healthy controls and COVID-19 urine samples. Scaling was done to Pareto. R2X = 0.72, R2Y = 0.85, Q2 (cum) = 0.73, CV-ANOVA = 1.01 × 10− 21 for OPLS-DA model which was validated through Permutation Test (Figure S3). Important features (VIP > 1) plot shows those buckets that contain metabolites that increase (in red) or decrease (in blue) its content for each group.To investigate more deeply the functional implications of these metabolites and their interactions in the context of COVID-19, a metabolic pathway analysis (MetPA) was performed. MetPA provides valuable information on the perturbations and alterations that occur in metabolic pathways, shedding light on their potential roles in the pathogenesis and clinical manifestations of the disease. The metabolic pathways associated with the discriminant metabolites in serum for the comparison between COVID-19 and healthy controls are shown in Fig. 1D. The most affected metabolic routes by COVID-19 disease with a higher impact index are “Phenylalanine, tyrosine and tryptophan biosynthesis”, “Phenylalanine metabolism”, “Glycerolipid metabolism” and “Arginine and proline metabolism”. In the cases of “Phenylalanine, tyrosine and tryptophan biosynthesis” (p, p-Holm and p-FDR < 0.05) and “Phenylalanine metabolism” (p and p-FDR < 0.05), these two routes involve two common metabolites: tyrosine and phenylalanine. With respect to “Glycerolipid metabolism” (p < 0.05), disturbed metabolites were glycerol and fatty acids (including acyl-, diacyl- and triacylglycerols, phosphatidic and lysophosphatidic acids) and, for “Arginine and proline metabolism” (p < 0.05), metabolites that hit this pathway were creatine and proline.Evaluation of potential biomarkers for COVID-19 infection and severityOne of the main questions in this past pandemic is why some people experienced severe symptoms of coronavirus while others remained asymptomatic. In this context, we conducted comparisons to deepen our understanding of the disease and its varying impacts. We analyzed metabolic differences between: (i) healthy individuals and COVID patients, (ii) patients admitted to the MW and those admitted to the ICU, and (iii) between individuals with good or poor prognoses. These comparisons aimed to shed light on the severity of symptoms (in case of comparison ii), and to investigate COVID-19 patients’ prognosis (in case of comparison iii), setting as poor prognosis those patients who required ICU admission or experienced fatal outcomes during their hospital stay. In this instance, multivariate data analysis did not yield valid models with satisfactory goodness-of-prediction (Q2) cumulative values. Consequently, we resorted to performing univariate data analysis instead. Under the criteria of fold change ≥ 1.1 or ≤ 0.9 (p < 0.05), several potential metabolite biomarkers in serum and urine in COVID-19 patients were observed, as shown in Table 2. Furthermore, in order to evaluate the diagnostic performance of metabolic biomarkers, ROC analysis was conducted and the area under curve (AUC) was calculated with AUC ≥ 0.6, ranging from poor (AUC = 0.6 to 0.7), fair (AUC = 0.7 to 0.8), good (AUC = 0.8 to 0.9) to excellent (AUC = 0.9 to 1) diagnostic ability26. This ROC analysis was performed for both matrices, serum and urine, and for the three aforementioned comparisons. Table 2 summarizes all the comparisons studied in serum and urine, detailing the FC, p-values, AUC, cutoff, sensitivity and specificity values.Table 2 Potential biomarkers in serum and urine for different comparisons: COVID-19 patients vs. HC, ICU vs. MW, and poor vs. good prognosis (PvsG).The biomarker ability, evaluated through AUC value, of those biomarkers obtained from the OPLS-DA models was, in most cases, higher than 0.7 in serum, and higher than 0.6 in urine. In general, highest values of AUC were obtained for the COVID-19 vs. controls comparison in serum samples. Furthermore, we evaluated the sensitivity and specificity for each metabolite, and both were found to be higher than 50% in each case.Regarding the metabolic changes between COVID-19 and control groups, higher levels of phenylalanine, tyrosine, lysine, creatine, proline, trimethylamine (TMA), mannose, acetone, NAG and fatty acids, especially UFA, were found in serum of COVID-19 group accompanied by decrease levels of TMAO and citrate. In urine, higher levels of methionine, fucose, lysine, hippurate, 2-phenylpropionate, glutamine, 3-indoxylsulphate and pseudouridine were observed for COVID-19 group in comparison to control group, that in turn revealed higher content of formate, 3-methylhistidine, trigonelline and creatinine.Concerning disease severity (ICU vs. MW), a decrease on TMAO, betaine and ethanol were found in serum samples from ICU patients, while metabolic changes on the contents of TMAO, hippurate, urea, dimethylamine (DMA), 3-hydroxybutyrate (3HB), fucose, 4-hydroxyphenyl acetate, 3-methyl-2-oxovalerate, 3-indoxylsulphate, tryptophan and formate were detected in urine.About prognosis, we found that a decrease on betaine and TMAO were associated to a negative outcome (poor prognosis) in serum, and a decrease in cis-aconitate, urea, formate, and creatinine, methanol, DMA, and TMAO in urine, together with an increase in creatine and hippurate also in urine.Relation between serum metabolites and clinical parametersBivariate analysis was performed to study the association between serum metabolite levels and fatal outcome in COVID-19 patients. Of total metabolites analyzed, TMAO levels were found to be significantly related with death due to the disease (p = 0.015). This is in accordance with our previous findings, in which TMAO was also found to be correlated with poor prognosis (ICU and death).We did not find metabolites with significant correlation with time in ICU, but in contrast, serum glucose levels were positively correlated with time spent in hospital in COVID-19 patients (p = 0.043, Fig. 3A). This reinforces what has been seen in other studies, in which high glucose levels have been associated with severe disease, probably derived from sustained inflammation resulting from infection27. However, this data should be taken with caution since the time spent in hospital is a variable with a multifactorial effect component and it would not be expected that a single metabolite would be a good indicator of this response variable on its own.In addition, numerous metabolites were significantly correlated to clinical features in COVID-19 patients, especially those related to cardiovascular risk factors and blood parameters of tissue/cell damage and clotting factors (Fig. 3B and Table S6). When we analyzed individually the group of patients admitted to medical ward, we observed some metabolites significantly correlated with having arterial hypertension or being overweight/obese (Fig. 3C and Table S7). Among them, ethanol, NAG, TMA, creatine and glycerol ranked among the top positive correlates to arterial hypertension, and in the same line, n-3 fatty acids, acetoacetate, proline, glycerol, tryptophan, tyrosine and triacylglycerol levels were the top to obesity/overweight. These metabolites are involved in numerous metabolic pathways, such us arginine and proline, galactose, and glycerolipid metabolisms, which were shared between these two conditions in the correlation studies. Similarly, ICU patients also showed metabolites positively correlated with cardiovascular risk factors (Fig. 3D and Table S8). In particular, the most significant metabolites were 3HB and choline for hypertension, creatine, tyrosine and acetate for overweight/obesity and TMAO for dyslipidemia, which are mainly involved in pathways of amino acid, carbohydrate and lipid metabolism. Some authors have suggested an association between the deregulation of these metabolic routes and high risk of severe COVID-195,28,29,30.Remarkably, we also detected that certain clinical variables were associated with the same metabolite in both groups of patients. Thus, the presence of one or more cardiovascular risk comorbidities was significantly associated with increased ethanol and glycerol. In particular, increased 3HB levels were related to hypertension, and elevated n-3 fatty acids, creatine, tyrosine, phenylalanine and acetate with obesity/overweight.Fig. 3Relations between clinical features and serum metabolite levels. (A) Scatterplot on correlation between serum glucose levels in COVID-19 patients and hospital time (R2 = 0.055, correlation coefficient (r) = 0.235, p-value < 0.05). Correlation analysis between clinical features and serum metabolite levels in COVID-19 patients (B) Considering MW and ICU groups together. (C) COVID-19 patients admitted to the MW. (D) COVID-19 patients admitted to the ICU. The color scale shows whether the relationship between two variables is positive (blue) or negative (violet). Spearman’s correlation coefficient was used for the analysis. Abbreviations: TAG: triacylglycerol, PC: phosphocholine, GPC: glycerophosphocholine, FA: fatty acids, PUFA: polyunsaturated FA, TMAO: trimethylamine, NAG: N-acetylglutamate, UFA: unsaturated FA, LDH: lactate dehydrogenase, IL-6: interleukin 6.Biomarker panels based on serum metabolites and clinical variablesThe use of biomarker panels has the advantage of better capturing inter-patient variability than individual metabolites, which is more feasible when they are transferred to large cohorts. Therefore, we designed biomarker panels capable of classifying COVID-19 patients by conducting logistic regression analyses.First of all, we focused on constructing panels based only on metabolites. The Least Absolute Shrinkage and Selection Operator (LASSO) method was used to select the metabolites that would constitute each panel. For the model comparing COVID-19 patients with HC, eight metabolites were initially identified by the LASSO method (isoleucine, ethanol, proline, TMAO, glucose, mannose, phenylalanine and betaine). Then, a Spearman’s correlation analysis was established to determine the correlations between pairs by setting a cut-off filter of 0.7, in such a way that the metabolites with a strong positive correlation were discarded to avoid fitting problems in the logistic regression model. As shown in Table S9, the metabolites with a high degree of correlation were proline, mannose and phenylalanine. Next, the Variance Inflation Factor (VIF) was calculated to check for multicollinearity among metabolites, and we removed those with VIF > 2.5 (ethanol and betaine). Thus, the final logistic regression model included the metabolites isoleucine, TMAO and glucose. The ROC curve for this panel showed outstanding discrimination (AUC = 0.91) reporting 86.08% of sensitivity and 83.87% of specificity (Figure S5A). In addition, we investigated a model to discriminate patients admitted to the MW from ICU patients. In this case, nine metabolites were identified by LASSO regression to build the model (ethanol, lactate, alanine, citrate, creatinine, TMAO, tyrosine, betaine and acetate). After subsequent correlation analyses (Table S10), some metabolites remained strongly correlated (lactate, alanine, citrate, creatinine and acetate) and were then discarded. In this case, no metabolite presented VIF > 2.5, therefore, the panel consisted of ethanol, TMAO, tyrosine and betaine. However, this panel did not provide sufficient discriminatory capacity (AUC = 0.697, sensitivity 45.45% and specificity 91.11%, Figure S5B). Interestingly, the biomarker panels included metabolites other than those shown to be more discriminating at the individual level.Finally, given the insufficient discriminatory capacity of the panel discriminating both types of COVID-19 patients, those admitted to the MW and ICU patients, we constructed a multivariate logistic regression model including both metabolite levels and clinical variables (analytical results or cardiovascular risks) as independent variables, with the aim of finding a predictor model to determine the risk of severe disease in COVID-19 patients. For this purpose, a previous selection of metabolites and variables to be included in the model was made by means of LASSO regression. Firstly, six metabolites and five clinical variables were identified (obesity/overweight, dyslipidemia, C-reactive protein, neutrophiles, lymphocytes, ethanol, lactate, TMAO, tyrosine, betaine, and acetate) and after subsequent correlation analysis, two metabolites (lactate and acetate) were removed (Table S11). As described in the previous panels, collinearity was then studied, and after eliminating those variables with VIF > 2.5, the final reduced model included: obesity/overweight, dyslipidemia, lymphocytes, ethanol, TMAO, tyrosine and betaine. According to the ROC curves performed to differentiate between ICU and MW patients, this classifier model showed a good discriminatory performance (AUC = 0.825), providing a sensitivity of 81.82% and a specificity of 71.11%, for a cut-off point of 0.41 (Figure S5C). These results suggest the contribution of conditions as low lymphocytes counts, and comorbidities as obesity/overweight and dyslipidemia to COVID-19 outcomes. Furthermore, the panel included only a few variables that would facilitate its clinical implementation in order to predict which of the hospital MW patients may require admission to the ICU due to potential severity.
