Multilayer metabolomic integration reveals bioenergetic disruption in Long COVID

Abstract

Background

Long COVID represents a significant health challenge, with 10–20% of patients with COVID-19 experiencing persistent multiorgan symptoms. The heterogeneity of clinical manifestations, combined with an incomplete understanding of the underlying molecular mechanisms, limits the improvement of patient management. Circulating metabolomic profiling constitutes a promising tool to address these limitations. In this context, we aim to investigate long-term metabolic disruptions in Long COVID through multilayer integration of plasma metabolites.

Methods

The study population included 42 survivors of critical COVID-19 who attended a comprehensive clinical evaluation conducted 12 months postdischarge. Plasma biochemicals, including lipoproteins, lipids, glycoproteins and metabolites were quantified using proton nuclear magnetic resonance spectroscopy (H-NMR). Circulating tricarboxylic acid (TCA) cycle intermediates and protein damage markers were detected by gas chromatography‒mass spectrometry (GC/MS). A machine learning-based feature selection approach was employed to identify the multilayered metabolic signature. Generalized additive models (GAMs) were used to explore associations between individual metabolites and specific dimensions of Long COVID.

Results

Univariate analysis revealed significantly elevated levels of alpha-ketoglutarate (aKG) and reduced levels of creatine in patients with Long COVID. A nine-metabolite and damage marker signature [aKG, carboxymethyl-cysteine (CMC), carboxymethyl-lysine (CML), creatine, fumarate, lactate, low density lipoprotein particle size (LDL-Z), 2-succinyl-cysteine (2SC) and tyrosine] was identified through the integration of Random Forest with Boruta and Sparse Partial Least Squares regression. This signature effectively classified patients with Long COVID (a cross-validated AUC of 0.91). In the GAM models, aKG, CMC, CML and creatine were associated with distinct Long COVID dimensions, including cognitive, functional and respiratory impairments.

Conclusions

Multilayer metabolomic integration reveals persistent bioenergetic disruption in patients with Long COVID. The identified metabolic profile offers promising biomarkers for medical decision-making. Modulating key metabolites could potentially mitigate specific symptoms of long COVID.

Graphical Abstract

Supplementary Information

The online version contains supplementary material available at 10.1186/s12967-026-07684-3.

Background

Long COVID, also referred to as post-acute sequelae of COVID-19 (PASC), is a debilitating condition that occurs in individuals with a history of probable or confirmed SARS-CoV-2 infection, generally three months from onset, with multiorgan clinical manifestations that last for at least two months and cannot be explained by another identifiable diagnosis [1]. The World Health Organization (WHO) calculates that roughly 10–20% of patients with COVID-19 experience mid- and long-term symptoms after resolution [2], which differ considerably according to the severity of the acute phase [3]. This results in increased use of health care resources, including the need for hospitalization and the use of nonconventional medication [4]. Altogether, the long-term aftermath of COVID-19 poses a great awareness in the global socioeconomic environment, with an estimation that approximately 30% of the pandemic health burden might arise from COVID-19-induced disability [5, 6].

Long COVID is defined as a complex syndrome affecting a great number of bodily organs and systems, including general or musculoskeletal, cardiovascular, neuropsychiatric, renal, endocrine, gastrointestinal and dermatologic dimensions [7, 8]. The interplay among multiple pathobiological pathways, such as the persistence of viral reservoirs, direct SARS-CoV-2-related tissue damage, dysbiosis, ongoing inflammation, autoimmunity and microthrombi phenomena, has been proposed as the main driver of Long COVID [9, 10]. Recent studies have highlighted the role of metabolic dysfunction in post-acute sequelae. For instance, cognitive and physical impairments after SARS-CoV-2 infection have been closely related to disturbances in amino acid metabolism, ultimately leading to attenuated mitochondrial function [11, 12]. In support of these results, Appelman et al. [13] described a systemic alteration of the tricarboxylic acid (TCA) cycle and glycolytic metabolites in Long COVID patients with reduced physical performance, indicating a shift from oxidative phosphorylation (OXPHOS) to glycolytic metabolism. Emerging evidence also points to mitochondrial oxidative stress as a determinant of Long COVID pathophysiology through the disruption of energy metabolism and the exacerbation of systemic inflammation [14]. SARS-CoV-2 can directly interact with mitochondrial metabolism, impairing OXPHOS and increasing reactive oxygen species (ROS) production [15]. According to independent investigations, these abnormalities in the redox state persist even months after resolution of the infection [16–18]. Additional metabolic analyses have revealed that oxidative stress induced by the immune response against SARS-CoV-2 infection might impair DNA repair pathways, permanently impacting the health status of these patients [19]. Overall, these mechanisms may constitute potential targets to identify biomarkers with diagnostic value and to guide therapeutic interventions, although the evidence is limited and inconclusive.

Metabolomics is an omics discipline that allows for the quantification of small molecules (i.e., fatty acids, amino acids, carbohydrates) and other chemically transformed intermediates [20]. Metabolites are the ultimate products of metabolism and subsequently have the most direct bond to phenotype, providing readable information into the biochemical state of an organism and into perturbations in response to pathological conditions [21]. Hence, comprehensive analysis of the human metabolome is emerging as a promising tool for understanding the molecular basis of a variety of pathologies [22, 23]. This particularly applies to complex diseases, in which clinical manifestations result from intricate crosstalk among different molecular effectors [24].

A major limitation of metabolomics is the great heterogeneity associated with the biological composition of distinct metabolites. The levels of these mediators usually exhibit modest correlations, thereby limiting the assignment of molecular meaning and the understanding of global relationships [25]. Machine learning (ML) algorithms constitute a great platform for the analysis of molecular information [26]. This methodology is able to detect relevant biological features across different molecular layers and how they interact with each other by retaining complex patterns from entire datasets [27]. Multilayered approaches that integrate distinct metabolic components and ML have the potential to determine the systemic consequences of whole subsets of metabolites and nontargeted axes of variation, providing further insight into disease development and progression [22].

Here, we investigated long-term metabolic disruption in Long COVID by analyzing circulating biochemical profiles, including lipoproteins, lipids, glycoproteins, low molecular weight metabolites (LMWMs), TCA cycle intermediates and nonenzymatic protein damage markers. The primary objective was to identify a multilayered metabolic signature associated with persistent symptoms in survivors of critical COVID-19 one year after hospital discharge. Additionally, we aimed to explore the novel associations between specific metabolites and different clinical dimensions of the syndrome. Our multilayer integration selected a subset of molecular components, suggesting mitochondrial and metabolic dysfunction are contributors to long-term symptoms following critical COVID-19. Several of these metabolites may be specifically related to different Long COVID dimensions.

Methods

Study design

This was a case‒control study including 42 patients admitted for COVID-19 at Hospital Universitari Arnau de Vilanova and Hospital Universitari Santa Maria in Lleida (Spain) between April 2020 and September 2021. Patients were recruited during their intensive care unit (ICU) stays and were scheduled for 12-month follow-up medical appointments after hospital discharge. The inclusion criteria included: i) age 18 years or older; ii) admission to the ICU due to confirmed SARS-CoV-2 infection; and iii) attendance at a “Post-COVID” evaluation 12 months after hospital discharge (median follow-up [P₂₅;P₇₅] = 367 [353;376] days). The exclusion criteria included: i) severe mental disability; ii) life expectancy of less than one year or palliative treatment; iii) follow-up in another department or institution; and iv) a previous diagnosis of diabetes, chronic kidney disease (glomerular filtration rate < 60 ml/min/1.73 m²), anemia, chronic respiratory disease or active malignant neoplasm.

Sociodemographic, clinical, pharmacological and laboratory data were recorded at ICU admission and at the follow-up visit and entered into the RedCap database. Incoherent or missing data were identified and reviewed by dedicated researchers.

Long COVID evaluation

A complete clinical evaluation was performed as previously described [28, 29]. General and respiratory symptoms, including dyspnea measured by the modified Medical Research Council (mMRC) Scale, ageusia, anosmia, dry and wet cough, asthenia and muscular fatigue, were evaluated during the consultation. To further assess the cognitive and physical states of patients, a set of standardized and validated questionnaires were employed: the British Columbia Cognitive Complaints Inventory (BC-CCI), the Hospital Anxiety and Depression Scale (HADS), mental and physical scores from the 12-Item Short Form Survey (SF-12) and the Functional Assessment of Chronic Illness Therapy (FACIT). Airway function was measured using a flow spirometer according to the guidelines of the American Thoracic Society (ATS) [30]. Respiratory variables were calculated as a percentage of the predicted value according to the European Community Lung Health Survey [31].

Long COVID was defined as dyspnea [mMRC scale > 0 [32]], cognitive impairment [BC-CCI > 4 [33]] and/or fatigue [FACIT < 34 [34]] persisting 12 months after hospital discharge, with no explanatory etiology other than severe SARS-CoV-2 infection. The controls were survivors of critical COVID-19 who fully recovered without dyspnea, cognitive impairment or fatigue at the same timepoint. Controls and patients with Long COVID were matched according to age and sex. Covariate balance was reached using the Nearest neighbor propensity score method, which paired every Long COVID patient with an available control according to the similarity of their propensity score values [35].

Sample collection

The samples were processed under standardized conditions with support from the IRBLleida Biobank (B0.000682) and the Biobank and Biomodels Platform ISCIII PT23/00032. Venous blood samples were collected in EDTA tubes (BD, NJ, USA) by venipuncture after an overnight fast and prior to the Long COVID evaluation (median follow-up [P25; P75] = 367 [353; 376] days after hospital discharge). Blood samples were fractionated by centrifugation at 1500 ×g at room temperature for 10 min. The plasma supernatant was immediately aliquoted and stored at −80 °C.

Biochemical analysis

Comprehensive circulating biochemical profiles, including lipoproteins, glycoproteins, LMWM, lipids, TCA or Krebs cycle intermediates and nonenzymatic protein oxidation markers, were assessed in the same patient-derived plasma samples using standardized protocols. Experiments were performed by trained staff blinded to clinical information. Details of metabolites’ identification and its levels of confidence are provided in Supplemental Tables S1 & S2.

Proton nuclear magnetic resonance spectroscopy analysis

Preprocessing included the dilution of 200 μl of each plasma sample in 50 μl of deuterated water and 300 μl of 50 nM phosphate buffer solution (PBS) at a pH of 7.4. Proton nuclear magnetic resonance (H-NMR) spectroscopy was performed using a Bruker Avance III 600 spectrometer operating at a proton frequency of 600.20 MHz.

• Lipoprotein analysis. The Liposcale® Test was employed to quantify the number and sizes of the main types of lipoproteins (very low-density lipoprotein [VLDL], low-density lipoprotein [LDL] and high-density lipoprotein [HDL]), the concentrations of their associated lipids and the concentrations of nine subclasses of the main lipoproteins, as previously described [36]. In brief, the lipid methyl group signal was deconvoluted using nine Lorentzian functions, where the area represented the concentration of the lipids associated with each lipoprotein subclass and the diffusion coefficient allowed for the calculus of the particle size. To extract the number of particles, the lipid concentration was divided by the volume of the specific particle. The variation coefficients for the particle number ranged from 2–4%, and those for the particle sizes were < 0.3%.

• Glycoprotein analysis. The region of the H-NMR spectrum related to the N-acetyl groups (2.15–1.90 ppm) was deconvoluted using three Lorentzian functions in order to calculate the specific concentrations of N-acetylglucosamine and N-acetyl galactosamine (GlycA) and N-acetylneuraminic acid (GlycB) [37]. The function total area was proportional to the concentration, whereas the signal shape, defined as the height-to-bandwidth (H/W) ratio, was related to the aggregation state of the protein–sugar bond in glycosylated proteins. The variation coefficients for the glycoproteins were lower than 3%.

• LMWM and lipid analysis. In-house softwares were developed to deconvolute the spectra associated with 15 different LMWMs (amino acids and carbohydrates, among others) [38] and 14 lipids, including cholesterol, glycerides, phospholipids and fatty acids [39]. The variation coefficients for LMWM and lipids varied between 6 and 18%.

Gas chromatography–mass spectrometry analysis

Gass chromatography–mass spectrometry (GC/MS) was carried out in an Intuvo 9000 GC system coupled to a 5977B mass selective detector (MSD) (both from Agilent Technologies, Barcelona, Spain). MassHunter Data Analysis and MassHunter Quantitative Analysis software programs (Agilent Technologies, Barcelona, Spain) were used to collect and quantify the data, respectively. Quantification was performed using standard curves. The signal was corrected using a Locally Weighted Scatter-plot Smoother (LOESS) approach [40]. National Institute of Standards and Technology (NIST) references were used for quality control and were injected every 10 samples.

• TCA cycle intermediates analysis. Pyruvate, succinate, fumarate, malate, alpha-ketoglutarate (aKG), aconitate, citrate and isocitrate were extracted from the plasma samples and derivatized as previously described [41, 42]. Briefly, 10 μL of plasma were extracted with cold methanol (1:3, v/v) containing dry low nitroxides (DLN) as an internal standard (IS) after a 1-h incubation at −20 °C. The mixture was subsequently centrifuged. The pellet was then dried in a SpeedVac concentrator, and the metabolites were silylated with MTBSTFA plus 1% TBDMCS. Finally, the silylated derivatives were diluted with acetonitrile and transferred to injection vials for selecting ion monitoring (SIM)-GC/MS analysis. Reference NIST and quality control samples were included. One microliter of each final solution was injected into an Intuvo 9000 GC equipped with an HP-5 MS UI capillary column coupled to a 5977- MSD (Agilent Technologies, Barcelona, Spain). The injection port was held at 250 °C in split mode (ratio of 1:10). The metabolites were separated by applying a temperature gradient: the initial temperature was 60 °C, increased to 300 °C at 15°C/min over 16 min and held for 4 min. For equilibration, the temperature was lowered to 70 °C for 2 min. Data acquisition was performed in SIM mode, with a mass range of 30–650 m/z. The MSD ion source temperature was set at 250 °C, and the transfer line temperature was set at 280 °C. Helium was used as the carrier gas. The detected ions were L-norleucine (NRL), m/z 200; pyruvate, m/z 259; fumarate, m/z 287; malate, m/z 287; succinate, m/z 289; aKG, m/z 431; aconitate, m/z 459; and (iso)citrate, m/z 459.

• Protein oxidation markers analysis. A total of seven biomarkers of protein nonenzymatic modifications indicating oxidative stress were detected, as described previously [43]. Markers of protein oxidation (protein carbonyl glutamic [GSA] and aminoadipic [AASA] semialdehydes), glycoxidation (carboxyethyl-lysine [CEL], carboxymethyl-lysine [CML] and carboxymethyl-cysteine [CMC]), lipoxidation (malondialdehyde-lysine [MDAL]) and mitochondrial stress (2-succinyl-cysteine [2SC]) were identified as trifluoroacetic acid methyl ester derivatives in acid hydrolyzed delipidated and reduced plasma protein samples by GC/MS. In summary, a sample volume containing 500 μg of protein was delipidated by adding methanol:chloroform (2:1, v/v) twice. After centrifugation, methanol-containing proteins were precipitated with cold trichloroacetic acid (TCA) and reduced with 500 mM sodium borohydride in 0.2 M borate buffer (pH 9.2) after overnight incubation at room temperature. The samples were subsequently reprecipitated with TCA, and a mixture of deuterated IS for every marker was added, including [2H8]Lys (d8-Lys), [2H5]5-hydroxy-2-aminovaleric acid (d5-HAVA), a stable derivative for GSA; [2H4]6-hydroxy-2-aminocaproic acid (d4-HACA), a stable derivative for AASA; [2H4]CEL (d4-CEL); [2H8]MDAL (d8-MDAL); [2H4]CML (d4-CML); U–13C315N-CMC (d-CMC); and, [2H2]2-SC (d2-2-SC). Then, the proteins were hydrolyzed at 155 °C with chloride acid (HCl) and dried in vacuo. Finally, N,O-trifluoroacetyl methyl ester (TFAME) derivatives were prepared by adding a methanolic-HCl solution followed by trifluoroacetic anhydride incubation. TFAME derivatives of the protein hydrolysate were resuspended in dichloromethane and transferred to injection vials for SIM-GC/MS analysis. Reference NIST samples and quality control samples were included. Four microliters of each stable derivative were injected in spitless mode into an Intuvo 9000 GC equipped with an HP-5 MS UI capillary column coupled to a 5977- MSD (Agilent Technologies, Barcelona, Spain). The injection port was held at 250 °C. The temperature was specifically programmed for chromatographic separation: 5 min at 110 °C, 2 min at 150 °C, 5 min at 240 °C, 15 min at 300 °C, and finally maintained at 300 °C for 5 min. The total runtime was 52 min, followed by 3 min post-run at 110 °C. Data acquisition was performed in SIM mode with a mass range of 30–650 m/z. The MSD temperature was set at 250 °C, and the temperature of the transfer line was set at 310 °C. Helium was used as the carrier gas. The detected ions were Lys and d8-Lys, m/z 180 and 187, respectively; HAVA and d5–HAVA, m/z 280 and 285, respectively; HACA and d4–HACA, m/z 294 and 298, respectively; CEL and d4-CEL, m/z 379 and 383, respectively; MDAL and d8-MDAL, m/z 474 and 482, respectively; CML and d4-CML, m/z 392 and 396, respectively; CMC and U–13C315N-CMC, m/z 271 and 275, respectively; and 2-SC and d2-2-SC, m/z 284 and 286, respectively.

Multilayer integration

Data preprocessing and standardization

Data from each metabolite were assessed for normality using the Shapiro‒Wilk test. Nonnormal metabolites were log-transformed for statistical purposes. Outlier analysis was conducted to identify incoherent data in the six independent datasets. Following a thorough examination, outliers were excluded before subsequent analysis.

Variable filtering

A consensus principal component analysis (CPCA) was used to reduce dimensionality and, subsequently, overfitting. The importance of the metabolites was determined based on their variability. In brief, a stepwise filtering approach was implemented in each subclass dataset [44]. First, PCA was calculated, including ten components, to ensure its exploratory nature. The data were scaled prior to unit variance. Second, components that explained at least 70% of the variability were selected. Eigenvalues matrix was considered to establish the cutoff. Third, the 70% of features that contributed the most to the definition of each component were included in the final multilayer dataset.

Machine learning multilayer feature selection agreement

The feature selection process was built by applying two different ML algorithms to the multilayer dataset, in order to further delimit the most informative metabolites for the discrimination of patients with Long COVID.

• Random forest by Boruta. Boruta is a relevant feature selection wrapper algorithm that is coupled with Random Forest (RF) for classification purposes [45]. Boruta created artificial shadow features by shuffling original features and compared their importance to the importance of the input metabolites. The algorithm underwent an iterative process (100 runs), during which it maintained only those attributes that exhibited higher importance values than their shadows. Finally, the metabolites selected in any of the runs were utilized to train the RF. The number of variables randomly sampled as candidates at each split (two) and number of trees (500) were optimized after consecutive cross-validated training (5-fold cross-validation, repeated ten times). The maximum accuracy and kappa were established as criteria. The classification performance of the method was evaluated using leave-one-out cross validation (LOOCV) on the RF model. The area under the receiver operating characteristic (ROC) curve (AUC) resulted from the average of the predicted probabilities extracted from the internal validations across the entire dataset.

• Sparse Partial Least Squares regression (sPLS). sPLS is a dimension reduction technique that decomposes data matrices into component scores by imposing an L₁ penalty [46]. The method selects metabolites according to the LASSO penalty on the loading vectors estimated by partial least squares regression. We initially built an exploratory sPLS model with the number of components set at ten. To determine the optimal values for the sparsity parameters and the number of components (two) and features (four and five, respectively) 5-fold cross-validation (repeated ten times) was employed. Maximum Euclidean distance was established as criterion. The classification performance of the tuned model was subsequently evaluated using the cross-validated AUC in the entire dataset.

Statistical analysis

The characteristics of the study population were summarized using descriptive statistics. The normality of the data was evaluated using the Kolmogorov–Smirnov test. Continuous variables were compared employing the Mann–Whitney U test or Student’s T test, whereas categorical variables were compared employing the Fisher’s exact test. Data are presented as medians [25th; 75th percentiles] or means (SDs) for continuous variables and as frequencies (percentages) for categorical variables. Differential detection of metabolites between study groups was performed using linear models with Empirical Bayes statistics [47]. The AUCs with their 95% confidence interval (CI) were used to assess the individual discriminative accuracy of the differentially abundant metabolites and their ratio. The relationships between the selected metabolites and the main clinical variables were estimated using generalized additive models (GAMs) with penalized cubic regression splines.

The p-value threshold defining statistical significance was fixed at < 0.05. Owing to the exploratory nature of the study, p-values were not adjusted for multiple comparisons. All the statistical analyses were performed using R software, version 4.3.2.

Results

Characteristics of the study population

The study included 42 survivors of critical COVID-19 whose samples were available for metabolomic profiling. Table 1 summarizes the demographic and clinical characteristics of these patients during the acute phase and at the 12-month follow-up. As anticipated, individuals experiencing sequelae reported dyspnea and cognitive complaints, with 50% of the cohort exhibiting abnormal fatigue levels. Other persistent symptoms encompassed cough and increased anxiety and depression scores. Long COVID was further associated with reduced pulmonary function and a marked decline in functional capacity. In terms of laboratory findings, leukocyte counts were abnormally elevated in survivors with Long COVID one year after hospital discharge.

Table 1.

Baseline and follow-up characteristics of the study population

	Control (N = 21)	Long COVID (N = 21)	p-value	N
Sociodemographic characteristics
Age (years)	63.0 [53.0;66.0]	59.0 [54.0;62.0]	0.363	42
Female	8 (38.1%)	9 (42.9%)	1.000	42
BMI (kg/m2)	29.6 (4.05)	30.2 (4.40)	0.657	42
Smoking history			0.710	39
Former	9 (47.4%)	7 (35.0%)
Nonsmoker	9 (47.4%)	11 (55.0%)
Current	1 (5.26%)	2 (10.0%)
Comorbidities
Hypertension	6 (28.6%)	7 (33.3%)	1.000	42
Type II diabetes mellitus	0 (0.00%)	0 (0.00%)	1.000	42
Obesity	12 (57.1%)	12 (57.1%)	1.000	42
Cardiovascular disease	1 (4.76%)	1 (4.76%)	1.000	42
Chronic lung disease	0 (0.00%)	0 (0.00%)	1.000	42
Asthma	3 (14.3%)	5 (23.8%)	0.697	42
Chronic kidney disease	0 (0.00%)	0 (0.00%)	1.000	42
Chronic liver disease	2 (9.52%)	1 (4.76%)	1.000	42
Autoimmune disease	0 (0.00%)	0 (0.00%)	1.000	42
Hematologic disorders	0 (0.00%)	0 (0.00%)	1.000	42
Rheumatic disease	0 (0.00%)	1 (4.76%)	1.000	42
Hospital stay
Hospital stay (days)	22.0 [16.0;34.0]	28.0 [13.0;35.0]	0.900	42
ICU stay (days)	13.0 [9.00;25.0]	13.0 [6.00;22.0]	0.332	42
High-flow cannula	20 (100%)	20 (100%)	1.000	40
Invasive mechanical ventilation	15 (71.4%)	11 (52.4%)	0.340	42
Noninvasive mechanical ventilation	18 (85.7%)	16 (76.2%)	0.697	42
Prone positioning	13 (61.9%)	8 (40.0%)	0.276	41
Antibiotic	17 (81.0%)	15 (71.4%)	0.717	42
Hydroxychloroquine	5 (23.8%)	4 (19.0%)	1.000	42
Corticoids	21 (100%)	20 (95.2%)	1.000	42
Remdesivir	0 (0.00%)	4 (19.0%)	0.107	42
Tocilizumab	16 (76.2%)	17 (81.0%)	1.000	42
12-months follow-up
Ageusia	0 (0.00%)	1 (4.76%)	1.000	42
Anosmia	0 (0.00%)	0 (0.00%)	1.000	42
Dry cough	0 (0.00%)	9 (42.9%)	0.001	42
Wet cough	1 (4.76%)	4 (19.0%)	0.343	42
Asthenia	0 (0.00%)	0 (0.00%)	1.000	42
Dyspnea (mMRC)	0.00 [0.00;0.00]	2.00 [1.00;2.00]	< 0.001	42
Abnormal ( > 0)	0 (0.00%)	21 (100%)	< 0.001	42
FACIT score	50.0 [48.0;51.0]	30.0 [17.0;38.2]	< 0.001	41
Abnormal ( < 34)	0 (0.00%)	10 (50.0%)	< 0.001	41
SF-12
Physical score	54.7 [50.1;55.9]	36.0 [25.0;42.4]	< 0.001	41
Mental score	53.8 [51.9;57.2]	45.0 [32.5;55.8]	0.019	41
BC-CCI score	0.00 [0.00;1.00]	9.00 [6.00;12.0]	< 0.001	42
Abnormal ( > 4)	0 (0.00%)	21 (100%)	< 0.001	42
BC-CCI category			< 0.001	42
None or minimal cognitive complaints	21 (100%)	0 (0.00%)
Mild cognitive complaints	0 (0.00%)	8 (38.1%)
Moderate cognitive complaints	0 (0.00%)	11 (52.4%)
Severe cognitive complaints	0 (0.00%)	2 (9.52%)
HADS
Depression score	0.00 [0.00;1.00]	4.00 [1.00;6.00]	< 0.001	42
Anxiety score	1.00 [0.00;2.00]	3.00 [1.00;9.00]	0.010	42
Pulmonary function
FEV1 (%)	98.0 [87.0;106]	81.8 [72.0;95.2]	0.012	41
FVC (%)	92.0 [80.0;97.0]	76.5 [66.8;88.2]	0.009	41
FEV1/FVC	82.7 [80.4;85.2]	83.1 [79.6;85.8]	0.845	41
TLC (%)	91.5 [81.3;101]	78.0 [71.5;88.0]	0.049	35
DLCO (% of predicted)	88.0 [78.5;95.0]	71.0 [63.2;82.5]	0.008	39
Laboratory parameters
Hemoglobin (g/dl)	14.4 [13.9;15.7]	14.7 [13.5;15.1]	0.392	42
Hematocrit (%)	43.3 [41.5;46.2]	43.1 [40.8;45.7]	0.521	42
Leukocyte count (x109/L)	6.12 [5.66;6.63]	7.53 [6.08;8.57]	0.021	42
Platelet count (x109/L)	223 [211;277]	269 [198;300]	0.458	42
Urea (mg/dL)	36.0 [33.0;42.0]	35.0 [30.0;40.0]	0.588	42
Creatinine (mg/dL)	0.82 [0.71;1.05]	0.85 [0.64;0.98]	0.633	42
Glomerular filtrate (mL/min/1.73 m2)	90.0 [78.1;90.0]	90.0 [78.4;90.0]	0.908	42
C-reactive protein (mg/L)	3.60 [1.60;5.20]	5.30 [2.50;11.9]	0.085	42
Ferritin (ng/L)	140 [95.3;205]	113 [71.7;188]	0.642	42
LDH (U/L)	377 (58.0)	374 (63.9)	0.866	42
D-dimer (mg/L)	150 [150;150]	150 [150;198]	0.268	41
Exertional capacity
6MWT Distance (m)	467 (86.7)	435 (93.4)	0.258	42
6MWT Oxygen saturation initial	96.0 [96.0;97.0]	96.0 [96.0;97.0]	0.808	41
6MWT Oxygen saturation final	95.5 [94.0;96.0]	96.0 [95.0;97.0]	0.453	41
6MWT Oxygen saturation minimal	95.0 [93.8;96.0]	95.0 [92.0;96.0]	0.905	41
6MWT Oxygen saturation average	96.0 [94.8;96.0]	96.0 [95.0;97.0]	0.522	41
Medication
Corticoids	6 (28.6%)	8 (40.0%)	0.659	41

Continuousvariables are expressed as the means (SD) or medians [P25; P75].Categorical variables are expressed as n (%). 6MWT: 6-minute walking test. BMI:body mass index. BC-CCI: British Columbia cognitive complaints inventory. DLCO:carbon monoxide diffusing capacity. FACIT: functional assessment of chronicillness therapy. FEV1: forced expiratory volume during the firstsecond. FVC: forced vital capacity. HADS: hospital anxiety and depressionscale. ICU: intensive care unit. LDH: lactate dehydrogenase. TLC: total lungcapacity. Sex refers to a set of biological attributes that are associated withphysical and physiological features (e.g., chromosomal genotype, hormonallevels, internal and external anatomy). The binary sex categorization(male/female) was designated at birth

Distinct plasma metabolic profiles in survivors with Long COVID

A total of 72 circulating analytes were quantified in the plasma samples by H-NMR spectroscopy and GC/MS. These analytes were categorized into six distinct subclasses: lipoproteins, glycoproteins, LMWM, lipids, TCA cycle intermediates and protein oxidation markers. To address the high biological and technical variability inherent to metabolomics, each subclass was analyzed independently. Following data normalization and quality control, four samples were excluded from further analyses (Supplemental Figure S1).

Univariate analysis revealed two metabolites with significant alterations in patients experiencing persistent Long COVID symptoms compared with controls. aKG levels were elevated (average difference = 0.306, p = 0.014), whereas creatine levels were reduced (average difference = −9.965, p = 0.017) in these patients compared with those without persistent symptoms (Fig. 1A&B). Other metabolites did not significantly differ between the study groups (Supplemental Table S3).

Fig. 1.

Differentially detected metabolites in Long COVID one year after hospital discharge. Violin plots of significant metabolites, alpha-ketoglutarate (aKG) (A) and creatine (B) and their ratio (C). The inner boxplots present the medians and P₂₅ and P₇₅ ranges of the concentration values of each metabolite and their ratio. P-values are displayed for aKG, creatine and their ratio. (D) receiver operating characteristic (ROC) curves for aKG, creatine and their ratio. The plots display the sensitivity versus the specificity of the metabolites and their ratio. The area under the curve (AUC) values with their 95% confidence intervals (CIs) are displayed for each curve

To enhance biological interpretability, we examined the ratio of aKG to creatine. Ratios have the potential to capture interconnected biochemical pathways and reduce internal variability, as supported by previous metabolomics studies [48]. Higher values were found in the Long COVID group (average difference = 0.049, p-value = 0.027) (Fig. 1C). However, this ratio demonstrated a discriminative ability comparable to those of the individual metabolites, with AUC (95% CI) values of 0.73 (0.57–0.89), 0.70 (0.54–0.86), and 0.75 (0.60–0.91) for aKG, creatine, and their ratio, respectively (Fig. 1D).

Multilayer plasma metabolic profiling of Long COVID

To assess metabolic disruptions comprehensively, relationships among metabolites from different subclasses were explored using a multilayer profiling approach coupled with ML. The set of features that contributed most significantly to the definition of each component was incorporated into the training multilayer dataset (Supplemental Figure S2). A total of 49 features across all metabolic layers were used for training the ML models, with lipid metabolism-related compounds accounting for 40% of the dataset (Fig. 2A).

Fig. 2.

Multilayer metabolite fingerprints associated with Long COVID one year after hospital discharge. (A) proportion of each level of metabolic information included in the study regarding the entire multilayer dataset used for training machine learning (ML) models (green) and the ultimate signature of metabolites (yellow). (B) supervised component analysis clustering through partial least square discriminant analysis (PLS-DA) classifying critical survivors who fully recovered and critical survivors with persistent symptoms. Each point represents a patient and is colored according to the study group. (C) best combination of metabolites selected by PLS-DA. The plots present the contribution of each metabolite to the two components that compose the model after cross-validation. (D) best subset of metabolites selected by Boruta. The bars show the importance of the contribution of each metabolite to the model. (E) receiver operating characteristic (ROC) curves for SPLS-DA and Random forest (RF) with Boruta models. Sensitivity versus specificity are plotted for both models. Area under the curve (AUC) values, along with their 95% confidence intervals (CIs) are included for each of the models. 2SC: 2-succinyl-cysteine. aKG: alpha-ketoglutarate. CMC: carboxymethyl-cysteine. CML: carboxymethyl-lysine. LDL - Z: low-density lipoprotein particle size. LMWM: low molecular weight metabolites. TCA: tricarboxylic acid

Feature selection was performed using two supervised ML algorithms: Boruta and sPLS-DA. sPLS-DA highlighted a subset of nine metabolites that accurately classified Long COVID patients, including aKG, CMC, CML, creatine, fumarate, lactate, LDL particle size (LDL-Z), 2SC and tyrosine (Fig. 2B& 2C). Meanwhile, Boruta identified aKG, CMC, CML and lactate as the most critical variables for defining this condition (Fig. 2D). Both ML models demonstrated high cross-validated performance, with AUC values of 0.91 (0.79–1.00) for sPLS-DA and 0.80 (0.63–0.97) for Boruta (Fig. 2E).

Given the superior discriminative ability of the sPLS-DA model, we focused on its nine-metabolite signature for subsequent analyses. This signature exhibited an overrepresentation of mediators of mitochondrial metabolism (Fig. 2A).

Associations between plasma metabolites and Long COVID dimensions

To further investigate the clinical relevance of the identified metabolites, GAMs were employed to evaluate associations between individual biomarkers and specific dimensions of Long COVID. These dimensions included pulmonary function (lung diffusing capacity [D_LCO] levels) and mental and physical health scores from the SF-12 questionnaire.

aKG exhibited a dose‒response relationship with the physical health score (effective degree of freedom [edf] = 2.397, p-value = 0.039), where higher levels were associated with greater impairment in functional capacity (Fig. 3A). The oxidative stress markers CMC (edf = 4.580, p-value = 0.002) and CML (edf = 4.456, p-value = 0.007) displayed nonlinear relationships with mental health scores, with cognitive impairments being particularly pronounced in patients with intermediate levels of these markers (Fig.s 3B &C). Creatine showed a direct linear relationship with D_LCO (edf = 1.000, p-value = 0.026), indicating that reduced creatine levels were linked to impaired lung diffusion capacity (Fig. 3D). Other metabolites in the sPLS-DA model did not exhibit significant associations with specific Long COVID dimensions (Supplemental Figures S2–S4).

Fig. 3.

Significant linear and nonlinear relationships between the metabolites in the signature and Long COVID dimensions. (A) generalized additive modeling (GAM) for physical score (Y axis) and the concentration of alpha-ketoglutarate (aKG) (X axis). GAM for mental score (Y axis) and the concentrations of carboxymethyl-cysteine (CMC) (B) and carboxymethyl lysine (CML) (C) (X axis). (D) GAM for lung diffusion capacity for carbon monoxide (D_LCO) (Y axis) and the concentration of creatine (X axis). The 95% confidence band for each GAM is colored in each plot, according to the Long covid variable presented. Effective degrees of freedom (edf) and p-values are indicated for each GAM

Given the consistent selection of aKG and creatine across all the analytical approaches, their associations with demographic and clinical variables were further explored (Supplemental Tables S4 & S5). Patients with higher aKG levels reported increased dyspnea, fatigue, cognitive complaints and cough. The upper stratum of the aKG also included survivors with a significant decrease in lung capacity. Conversely, lower creatine levels were strongly associated with reduced pulmonary function, as evidenced by lower D_LCO values. In addition, reduced creatine correlated with nonmodifiable factors, such as male sex and with altered renal function during the follow-up.

Discussion

The long-term persistence of multisystem symptoms has become clinically evident in a large percentage of survivors of critical COVID-19 [49]. Indeed, there is increasing concern about the chronification of post-acute sequelae and their potential impact on quality of life and health systems. A deeper understanding of their causes and pathophysiology, the development of therapeutic interventions, and, ultimately, the establishment of an effective model of care are imperative. In this study, we examined the relationship between Long COVID and a large array of circulating molecules from different metabolic levels. Our findings suggest that: i) patients with Long COVID present dysregulated levels of aKG and creatine one year after hospital discharge; ii) a specific multilayer metabolic fingerprint is associated with the persistence of symptoms in post-critically ill survivors; and iii) several metabolites within this signature correlate with distinct dimensions of Long COVID.

Analyzing single levels of molecular information often fails to provide a holistic perspective on complex, multifactorial conditions such as post-acute COVID-19 sequelae. To address this, integrative studies that explore multilayer interactions between different biological elements are highly encouraged [50, 51]. ML algorithms currently represent the gold standard for analyzing multidimensional biological data [52]. Various integration approaches, including orthogonal clustering and feature selection methods, have been proposed to refine the molecular classification of diseases [53, 54]. However, ML-based integration still faces hurdles before its widespread implementation in multi-omics research. A key limitation is the imbalance in datasets, which often contain numerous biological features but relatively few patients, increasing the risk of overfitting and reducing model generalizability [55]. Variable filtering processes help mitigate these issues by transforming high-dimensional data into lower-dimensional forms while preserving variability [44]. Here, we first combined different metabolic markers based on their degree of variance and then selected the most relevant metabolites explaining the persistence of symptoms using an ML-based feature selection consensus. Our results suggest that multilayer integration captures more critical features than single-dataset analyses, identifying nine key metabolites versus only two in univariate analyses.

Long COVID impacts multiple clinical domains, hindering its diagnosis [8]. Accumulating evidence has linked Long COVID syndrome with numerous markers in the immune and coagulation domains that exhibited limited diagnostic performance, including complement system, pro-inflammatory cytokines and platelet activators [56, 57]. To the best of our knowledge, no universal biomarker has been established to guide clinical decision-making in survivors of critical COVID-19. Metabolites constitute the ultimate biological response to environmental and pathological disturbances, offering a direct and timely link to disease phenotypes [58]. Thus, the identification of accessible and reliable metabolomic biomarkers could enhance the clinical management of post-acute sequelae. In our study, two metabolites identified critically ill survivors with persistent symptoms one year post-discharge (AUC: 0.70–0.75). Recent trends favor biomarker panels over single-molecule biomarkers, as detecting subtle changes in diverse metabolite types may enhance diagnostic accuracy [59]. Larger metabolite panels also improve the understanding of biological complexity, revealing novel associations and fluxes between biochemical pathways [60]. This is the case for the 47-plasma metabolite panel developed by An et al. [61] for the early diagnosis of breast cancer. Similarly, in respiratory diseases, combining plasma metabolomic profiling with ML has demonstrated high diagnostic performance for various tuberculosis types [62]. A panel of lipid metabolism intermediates retrieved by ML also showed good performance in discriminating between bacterial and COVID-19 pneumonia [63]. In this scenario, our multilayer ML integration markedly outperformed univariate classification approaches, achieving an accuracy improvement of up to 21% (AUC = 0.80–0.91) and revealing candidate biomarkers for Long COVID diagnosis.

In addition to their diagnostic potential, these novel metabolic alterations offer insights into Long COVID pathophysiology. The dysregulation of key catabolic substrates, such as αKG, creatine, fumarate and lactate, aligns with previous research indicating long-term disruptions in energy and mitochondrial metabolism in survivors of COVID-19 [13, 64, 65]. Within the mitochondria, the TCA cycle and OXPHOS are the primary pathways for generating adenosine triphosphate (ATP) and maintaining homeostasis. Fumarate and aKG are key intermediate metabolites of the TCA cycle and are supported by anaplerotic pathways. Elevated levels of both metabolites have been widely associated with impaired TCA cycle dynamics in preclinical models of various chronic diseases [66]. However, the role of aKG as a protective factor against systemic inflammation associated with long COVID cannot be disregarded [67]. Dietary supplementation with αKG has demonstrated anti-inflammatory effects in chronic inflammation and COVID-19 models [68, 69]. In muscle and brain cells, creatine kinase (CK) enzymes facilitate the conversion of creatine to phosphocreatine, a process that consumes ATP supplied by OXPHOS. When the cell requires a rapid source of ATP, CK can reverse this process, converting phosphocreatine back into creatine [70]. This dynamic interplay enables the cell to regulate the ATP supply from the mitochondria according to its energy demands [71]. The observation of significantly lower levels of creatine in patients with Long COVID implies the potential impairment of this essential homeostatic pathway. Such disruption could lead to reduced physical and cognitive performance [72–74], hallmarks of this syndrome. Alternatively, the direct link between creatine and the pulmonary dimension might be explained by the attenuation of prolonged inflammation and lung ischemic injury after acute disease [10], in parallel with what has been observed as a consequence of lung transplantation [75]. Taken together, supplementation emerges as a potential therapeutic strategy for post-acute sequelae [76] NCT06992414).

Energy metabolism relies on the continuous exchange of free electrons [77]. Disruptions in redox balance can have detrimental health effects [78]. Our findings indicate that survivors with persistent symptoms exhibit abnormal levels of αKG and lactate, key redox buffers that mitigate TCA cycle dysfunction [79, 80]. Furthermore, our multilayer model revealed an imbalance in the mitochondrial stress axis (fumarate/succinate). Excess fumarate, resulting from impaired TCA function, accumulates in the cytoplasm and reacts with thiol groups in crucial cellular proteins, leading to succination and inactivation [81, 82]. Elevated fumarate has been implicated in metabolic and respiratory diseases, such as airway obstruction and diabetes [83, 84]. Similarly, persistent cognitive complaints after critical SARS-CoV-2 infection exhibited a significant association with levels of other oxidative markers, CMC and CML. Early cognitive impairment during aging is mechanistically connected with elevated bioenergetic defects and a more oxidant state in the mitochondrial environment [85]. These molecular changes eventually trigger the formation of carbonyl groups and the subsequent irreversible oxidation and inactivation of several proteins that are essential for proper neuronal function [86]. These findings highlight the need to explore antioxidant therapies for Long COVID, which has also been identified as a factor in the acceleration of aging [87, 88].

The strengths of the current investigation include its real-world clinical setting with well-characterized medical evaluations and rigorous omics analyses validated by an established feature selection algorithm [89]. However, several limitations should be acknowledged. First, this was a single-center study including a modest number of patients. Model overfitting and the impact of type I error should not be discarded. Although the medical assessments were based on routine clinical practice, which suggests that our findings may be generalizable, a larger cohort from other institutions and other clinical settings is needed to confirm our results. Indeed, we are currently collecting data and biological samples from patients who have recovered from severe pneumonia and/or acute respiratory distress syndrome (ARDS) (ClinicalTrials.gov Identifier: NCT06083363). Second, potential confounding factors might have impacted our findings. Although the study population was selected after matching by age and sex, the identified metabolic signature could be influenced by other sociodemographic factors and clinical variables that are highly related to metabolism, including ethnicity, diet, physical activity and medication use during the 12-month follow-up. Third, we employed a targeted metabolomic approach; thus, the metabolic alterations identified may reflect biases inherent to previously studied metabolites. Additionaly, untargeted analyses could reveal other metabolites relevant to the pathogenesis of Long COVID. Fourth, the observational nature of our study impedes the establishment of causal relationships. We could not clearly state whether these metabolites were actively released in response to Long COVID-related disruptions. Further in vitro and in vivo studies, including experimental models, are needed to clarify the mechanistic roles of these molecular effectors and to validate their contribution to the Long COVID phenotype. Finally, it is important to acknowledge that the metabolic alterations described here represent only one aspect of the multifactorial mechanisms proposed to underlie Long COVID. Other biological processes, such as persistent viral reservoirs, chronic inflammation, autoimmunity, dysbiosis and microthrombotic phenomena, are also likely to contribute to the complex clinical manifestations of the syndrome.

Conclusions

In conclusion, our study provides valuable data for generating hypotheses and guiding future investigations. Multiple layers were integrated to select a subset of molecular components, which might suggest that mitochondrial and metabolic dysfunction are contributors to long-term symptoms following critical COVID-19. Several of these metabolites may be specifically related to different Long COVID dimensions. These findings offer potential biomarkers and therapeutic targets for clinical management.

Electronic supplementary material

Below is the link to the electronic supplementary material.

Abbreviation

2SC

2-succinyl-cysteine

aKG

Alpha-ketoglutarate

ARA+EPA

Arachidonic acid and eicosapentaenoic acid

ATP

Adenosine triphosphate

ATS

American Thoracic Society

AUC

Area under the receiver operating characteristic curve

BC-CCI

British Columbia Cognitive Complaints Inventory

bp

Base pair

C

Cholesterol

CK

Creatine kinase

CMC

Carboxymethyl-cysteine

CML

Carboxymethyl-lysine

CPCA

Consensus principal component analysis

DHA

Docosahexaenoic acid

D_LCO

Lung diffusing capacity for carbon monoxide

DLN

Dry low nitroxides

EC

Esterified cholesterol

edf

Effective degrees of freedom

FACIT

Functional Assessment of Chronic Illness Therapy

FC

Free cholesterol

GAM

Generalized additive modeling

Glyc

Glycoprotein

GC/MS

Gas chromatography‒mass spectrometry

H-NMR

Proton nuclear magnetic resonance spectroscopy

H/W

Height to bandwidth

HADS

Hospital Anxiety and Depression Scale

HACA

Aminocaproic acid

HAVA

Aminovaleric acid

ICU

Intensive care unit

IDL

Intermediate-density lipoprotein

IS

Internal standard

LA

Linoleic acid

LDL

Low-density lipoprotein

LMWM

Low molecular weight metabolites

LOESS

Locally Weighted Scatter-plot Smoother

LOOCV

Leave-one-out cross validation

LPC

Lysophosphatidylcholine

MDAL

Malondialdehyde-lysine

ML

Machine learning

mMRC

Modified Medical Research Council

MSD

Mass selective detector

NIST

National Institute of Standards and Technology

NRL

L-norleucine

OXPHOS

Oxidative phosphorylation

P

Particle number

PASC

Post-acute sequelae of COVID-19

PBS

Phosphate buffer solution

PCA

Principal component analysis

PL

Glycerophospholipids, except LPC

PLS-DA

Partial least square discriminant analysis

PUFA

Polyunsaturated fatty acids

RF

Random forest

ROC

Receiver operating characteristic

ROS

Reactive oxygen species

SF-12

Mental and physical scores from the 12-Item Short Form Survey

SIM

Selecting ion monitoring

TCA

Tricarboxylic acid

TFAME

N,O-trifluoroacetyl methyl ester

TG

Triglycerides

VLDL

Very low density lipoprotein

W3

Omega-3 fatty acids

w6+w7

Omega-6 and omega-7 fatty acids

W9

Omega-9 fatty acids

WHO

World Health Organization

Z

Diameter

Author contributions

JG, FB, RP, DdGC and GT were responsible for conceptualization, methodology, resources and funding acquisition. Data curation was done by MCGH, JG, IDB and DdGC. Investigation was developed by MCGH, NMM, ICM, MJ and NA. MCGH, NMM, IDB, NA and DdGC contributed to software development and formal analysis. Visualization was conducted by MCGH. DdGC supervised final results. MCGH, NMM, RP, DdGC and GT wrote the original draft. The review and editing of the manuscript was made by JG, IDB, FB and NA.

Funding

The current project was financially supported by SEPAR (1284/2022). This study was funded by the Instituto de Salud Carlos III (ISCIII) through the project “PI23/01205” and cofunded by the European Union. The COVID-Ponent study is funded by Institut Català de la Salut and Gestió de Serveis Sanitaris (GSS). We were further supported by: Program “estar preparados”; UNESPA (Madrid, Spain), Fundación Eugenio Rodríguez Pascual (FERP-2023-076), Programa de Becas Gilead a la Investigación Biomédica (GLD23_00063). Research by the authors was also supported by the Spanish Ministry of Science, Innovation, and Universities (Ministerio de Ciencia, Innovación y Universidades, PID2023-152233OB-I00) and the Generalitat of Catalonia: Agency for Management of University and Research Grants (2021SGR00990) to RP. This study was co-financed by FEDER funds from the European Union (“A way to build Europe”). MCGH held predoctoral fellowship Ayudas al Personal Investigador en Formación from IRBLleida/Diputación de Lleida. FB is supported by ICREA Academia program. DdGC has received financial support from Instituto de Salud Carlos III (Miguel Servet 2020: CP20/00041), co-funded by the European Union.

Data availability

The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.

Declarations

Ethical approval

The medical ethics committee of Hospital Universitari Arnau de Vilanova approved the protocol (CEIC/2273). The present study was performed in full compliance with the Declaration of Helsinki. The patients were provided with written information indicating the nature and goals of the study and signed an informed consent form. Patient and sample data were registered in a database with access restricted to authorized personnel. To ensure that sex, and other diversity dimensions (such as age) are properly accounted for, the study includes balanced participant selection. No exclusion criteria were set based on sex, age ethnic or cultural background.

Consent for publication

Not applicable.

Competing interests

NA is a stock owner of Biosfer Teslab and has a patent related to the lipoprotein profiling described in the present manuscript. The rest of authors declare not to have any conflicts of interest that may be considered to influence directly or indirectly the content of the manuscript.