Metabolite profile of COVID-19 revealed by UPLC-MS/MS-based widely targeted metabolomics.

The metabolic characteristics of COVID-19 disease are still largely unknown. Here, 44 patients with COVID-19 (31 mild COVID-19 patients and 13 severe COVID-19 patients), 42 healthy controls (HC), and 42 patients with community-acquired pneumonia (CAP), were involved in the study to assess their serum metabolomic profiles. We used widely targeted metabolomics based on an ultra-performance liquid chromatography-tandem mass spectrometry (UPLC-MS/MS). The differentially expressed metabolites in the plasma of mild and severe COVID-19 patients, CAP patients, and HC subjects were screened, and the main metabolic pathways involved were analyzed. Multiple mature machine learning algorithms confirmed that the metabolites performed excellently in discriminating COVID-19 groups from CAP and HC subjects, with an area under the curve (AUC) of 1. The specific dysregulation of AMP, dGMP, sn-glycero-3-phosphocholine, and carnitine was observed in the severe COVID-19 group. Moreover, random forest analysis suggested that these metabolites could discriminate between severe COVID-19 patients and mild COVID-19 patients, with an AUC of 0.921. This study may broaden our understanding of pathophysiological mechanisms of COVID-19 and may offer an experimental basis for developing novel treatment strategies against it.

Introduction

The COVID-19 pandemic remains an ongoing critical threat. As of January 20, 2022, 336,790,193 confirmed cases worldwide, including 5,560,718 deaths, have been reported by the World Health Organization (https://covid19.who.int/). Although COVID-19 has been effectively controlled in China, sporadic cases have emerged that counteracted the controlling efforts. In fact, in a 2021 Nature poll, many prominent experts expressed their opinion on the future of the SARS-CoV-2 virus, and nearly 90% believed that it would continue spreading globally for years to come (1).

Viral infections perturb the host metabolism, and the outcome depends on the stage of viral infection (2). Using mass spectrometry-based metabolomics, Shen et al. identified that the levels of over 100 lipid metabolites were abnormal during COVID-19 infection (3). The perturbed metabolites were mainly related to activation of the complement system, macrophage function, and platelet degranulation. Tiffany Thomas et al. reported that COVID-19 infection could trigger changes in host tryptophan metabolic pathways, and this metabolic reprogramming is also related to inflammation and immune regulation in vivo (4). However, our knowledge of the changes of host metabolic pathways after COVID-19 infection is still very limited. In addition, the mechanism driving metabolite change during the infection is poorly understood. Metabolomics is a relatively new method that quantifies metabolites in a biological sample and explores the occurrence and development of human diseases (5, 6). Unlike other omics approaches, it is highly conserved and endogenous, having a high potential to discover biomarkers. Most metabolomic studies of COVID-19 were nontargeted, but this method has certain limitations in identifying metabolites. Therefore, more advanced and reliable technology for identifying and quantifying metabolites is paramount to understand the metabolic changes in patients with COVID-19.

In this study, we first employed a widely targeting technology based on Metabolon ultra-performance liquid chromatography–tandem mass spectrometry (UPLC-MS/MS) to determine the global metabolic profile of patients with COVID-19. Next, we explored whether the COVID-19 infection can induce disturbances in metabolite abundance in order to explore its underlying pathophysiology.

Materials and methods

Blood sample collection and clinical data

The study was conducted according to the Declaration of Helsinki and was approved by the Ethics Committee of the First Affiliated Hospital of Zhejiang University School of Medicine, China. All patients and healthy volunteers signed the informed consent, and their fasting blood samples were collected in the morning hours. The diagnosis of CAP was based on the physical examination, clinical manifestations, and chest X-ray changes. All patients enrolled in present study were initial diagnosed, and blood sampling were performed before any intervention. A disposable EDTA-anticoagulated vacuum blood collection tube was used to store 5 ml of the freshly drawn peripheral blood. Plasma was separated and stored as aliquots at −80°C and clinical information was recorded ( ). The samples were further processed in a biosafety level 2 laboratory qualified for SARS-CoV-2 testing following the Laboratory Biosafety Guidelines for COVID-19 (2nd edition) issued by the National Health Commission of China. Plasma samples of 44 patients with COVID-19 (31 mild COVID-19 patients and 13 severe patients), 42 healthy volunteers, and 42 patients with community-acquired pneumonia (CAP) were collected. In addition, clinical information was recorded, and the sample information for all patients and healthy volunteers are presented in .

Table 1

The baseline clinical information of both groups.

Characteristic	Mild COVID-19	Severe COVID-19	Healthy controls	Community-acquired pneumonia	p-value (Mild vs Severe COVID-19)
n	31	13	42	42
Sex, n (%)					1.000
Female	15 (48.4%)	6 (46.2%)	20 (47.6%)	20 (47.6%)
Male	16 (51.6%)	7 (53.8%)	22 (52.4%)	22 (52.4%)
Age, median (IQR)	41 (35.5, 49)	52 (35, 55)	50 (35, 57)	46 (37.75, 55)	0.086

Full-spectrum metabolic identification and data acquisition

The plasma metabolites were determined using UPLC-MS/MS (ExionLC AD coupled to a QTRAP spectrometer) (https://sciex.com.cn/). The liquid-phase conditions were as follows: (1) chromatographic separation using Waters Acquity UPLC HSS T3 column (1.8 µm particle size, 100 mm × 2.1 mm; (2) analyte elution on a mobile phase with ultrapure water (0.1% formic acid) for phase A and acetonitrile (0.1% formic acid) for phase B; (3) elution gradients were water/acetonitrile (95:5 V/V) at 0 min, 10:90 V/V at 11.0 min, 10:90 V/V at 12.0 min, 95:5 V/V at 12.1 min, and 95:5 V/V at 14.0 min; (4) the flow rate was 0.4 ml/min, the column temperature was 40°C, and the injection volume was 2 μl. The MS/MS conditions were as follows: 500°C, electrospray ionization; 5500 V (positive) and −4500 V (negative); 55 psi, ion gas source I; 60 psi, gas source II; 25 psi, curtain gas; and high collision-activated dissociation parameters. The ions were scanned and detected according to an optimized decluttering potential and collision energy. They were subjected to qualitative determination based on the retention time, daughter–parent ion pair information, and secondary spectral data of the detected substances using the MetWare database (http://www.metware.cn/). The analytes were quantified using the QTRAP multiple-reaction monitoring mode. After collecting data for different samples, the area under the peak was scored separately for the chromatographic peaks of the extracted ions of all metabolites. Finally, score correction for the chromatographic peaks of the same metabolite in different specimens followed.

Quality control analysis

Quality control (QC) samples were prepared from the extract mixture of all tested samples and analyzed under the same conditions to monitor the reproducibility of the test. In the process of sample testing, the QC samples were analyzed once every 10 samples to ensure the stability of the analytical process. The repeatability of metabolite extraction and detection can be determined by overlapping analysis for the total ion current (TIC) of mass spectrometry detection of different QC samples. The results showed that the curve overlap of metabolites TIC is high, which means that the retention time and peak intensity were consistent, indicating that the signal stability of mass spectrometry was good when detecting the same sample at different times ( ). Pearson analysis was performed to evaluate the correlation between biological replicate samples. Pearson’s correlation coefficient (r2) was used as the evaluation index of correlation ( ). PCA is a standard unsupervised pattern recognition algorithm, which is mostly used to identify the dominant signals in multi-dimensional data. In this study, PCA was used to examine the overall differences in metabolic profiles among the groups and the variability within groups. The result indicated that the metabolic profiles in each group were clearly clustered and obviously separated, and strong clustering within QC samples (MIX) ( ).

Differential metabolite enrichment analysis

The metabolite data were analyzed with different univariate and multivariate analyses to identify differentially expressed metabolites. Principal component analysis (PCA) and orthogonal partial least squares-discriminant analysis (OPLS-DA) was performed in R (v. 4.1.1) and used to reduce data dimensionality and verify the separation trends between groups. OPLS-DA combines orthogonal signal correction (OSC) and partial least squares-discriminant analysis (PLS-DA) method. In this study, OPLS-DA was performed to decompose the X matrix information (metabolic profile) into Y (groups) correlation and irrelevance by OSC and PLS-DA, thereby screen differential metabolites by removing unrelated differences. The relative contents of the differentially expressed metabolites were normalized and centralized, and K-mean clustering (K-means) analysis was performed to investigate the changing trends of the relative metabolite contents in different samples. The OPLS-DA results yielded variable importance in projection (VIP) for each metabolite, and only those with VIP ≥ 1 were selected. The permutation test was used to evaluate the OPLS-DA model. In addition, metabolites with fold change ≥ 1.5 or ≤ 0.6 were identified as significantly differentially expressed. Their potential functions were investigated using the Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis. MetaboAnalyst database (https://www.metaboanalyst.ca/) was used to perform the metabolic set enrichment analysis (MSEA).

Supervised analysis of differential metabolites

The least absolute shrinkage and selection operator (LASSO) regression model and random forest algorithms were used to screen the candidate biomarkers for diagnosing COVID-19 (7, 8). The fivefold cross-validation was applied to assess the power of prognostic classifiers. Briefly, the differential metabolite profiles were first log2 transformed, and the whole cohort was randomly classified into five equal portions. Subsequently, the least absolute shrinkage was used to screen feature selection on four-fifths of the cohort, while the remaining fifth was applied for prediction. In addition, the receiver operating characteristic (ROC) curves were plotted to assess the predictive ability of all five iterations.

Statistical analysis

Experimental data are presented as mean ± standard deviation, and P < 0.05 was considered significant. Continuous parametric variables in the three groups were tested by the analysis of variance, while continuous nonparametric variables were tested using the Kruskal test. Qualitative data analysis was performed using the chi-square test. The above analyses were performed in R version 4.1.1 (R Foundation for Statistical Computing, Vienna, Austria).

Results

Identification of differentially expressed metabolites

More than 600 metabolites from 128 samples were identified using a widely targeted metabolomics approach based on the UPLC-MS/MS platform and the MetWare database. They belonged to various classes of organic molecules, including amino acids and their metabolites, bile acids, nucleotides and their metabolites, glycerophospholipids, fatty acyls, oxidized lipids, and organic acids and their derivatives. The unsupervised PCA was used to analyze the correlations between healthy controls and COVID-19 patients. The score of three-dimension PCA (3D-PCA) indicated that the metabolites could distinguish between COVID-19 patients and the healthy controls ( ), suggesting high data quality. In addition, the OPLS-DA analysis confirmed a clear tendency towards separating healthy controls from COVID-19 patients ( ). The OPLS-DA model was evaluated using 200 randomized permutations (Q2 = 0.894, p<0.005, R2Y=0.962, p<0.005) and the results showed that OPLS-DA exhibited the optimal discriminatory ability ( ). Unlike healthy controls, COVID-19 patients had 125 differentially expressed metabolites (56 upregulated and 69 downregulated) in the plasma ( ). Among the top 20 differentially expressed metabolites between the COVID-19 group and the control group ( ), N6-methyladenosine, 3-methylsalicylic acid, 2′-O-methyladenosine, 2-(α-D-mannosyl)-3-phosphate glyceride, dGMP, AMP, and 3′-Adenylic acid were most notable.

Figure 1

Identification of differentially expressed metabolites between COVID-19 patients and healthy controls. (A) Score plots of three-dimension principal component analysis (3D-PCA) discriminating between the metabolic profiles of healthy controls (HC) and COVID-19 patients (CP). (B) Orthogonal partial least-squares discriminant analysis (OPLS-DA) score plot of HC and CP. (C) Volcanic map of differentially expressed metabolites between HC and CP. Red dots represent upregulated metabolites, gray dots represent unchanged metabolites, and green dots represent downregulated metabolites. (D) The top 20 differentially expressed metabolites ranked by the value importance plot (VIP). (E) The top 20 differentially expressed metabolites ranked by the log2 fold change (Log2 FC).

In agreement, the scores of 3D-PCA and OPLS-DA also suggested a clear separation between COVID-19 patients and CAP patients ( ). The permutation test (Q2 = 0.945, p<0.005; R2Y=0.977, p<0.005) also indicated that the OPLS-DA model had the best discrimination effect ( ). The COVID-19 patients had 134 differentially expressed metabolites (99 downregulated and 35 upregulated) in the plasma ( ). Of these, N6-methyladenosine, 2′-deoxyinosine, and cis-11,14,17-eicosatrienoic acid (C20:3) were most significantly differentially expressed between the two groups ( ).

Figure 2

Identification of differentially expressed metabolites between COVID-19 patients and community-acquired pneumonia. (A) Score plots of three-dimension principal component analysis (3D-PCA) discriminating between the metabolic profiles of community-acquired pneumonia (CAP) and COVID-19 patients (CP). (B) Orthogonal partial least-squares discriminant analysis (OPLS-DA) score plot of CAP and CP. (C) Volcanic map of differentially expressed metabolites between CAP and CP. Red dots represent upregulated metabolites, gray dots represent unchanged metabolites, and green dots represent downregulated metabolites. (D) The top 20 differentially expressed metabolites ranked by the value importance plot (VIP). (E) The top 20 differentially expressed metabolites ranked by the log2 fold change (Log2 FC).

Metabolic pathway enrichment analysis of differential metabolites

Pathway enrichment analysis was performed using the KEGG database to elucidate the role of the differentially expressed metabolites in the progression of COVID-19. The KEGG annotation results for the COVID-19 group and the control groups were classified according to the pathway type and divided into 4 categories: organismal systems, metabolism, human diseases, and environmental information processing ( and ). The metabolites expressed only in COVID-19 patients were mainly involved in biliary secretion, ferroptosis, purine, glycerophospholipid, and arachidonic acid metabolism pathways ( and ). To further explore the pathway differences between the COVID-19 group and the control group, MSEA was carried out. It revealed several significantly different pathways such as, taurine and hypotaurine metabolism, primary bile acid biosynthesis, drug metabolism-other enzymes, and phenylalanine metabolism ( and ).

Figure 3

Pathway enrichment analysis of differentially expressed metabolites between COVID-19 patients and healthy controls. (A) Kyoto Encyclopedia of Genes and Genomes (KEGG) classification of differentially expressed metabolites between healthy controls (HC) and COVID-19 patients (CP). (B) KEGG pathway enrichment analysis of differentially expressed metabolites. (C) Metabolite Set Enrichment Analysis (MSEA) between HC and CP.

For the differential metabolites between the COVID-19 group and the CAP group, the KEGG pathways belonged to the same categories identified in the COVID-19 group and the control group ( and ). The enriched pathways were mainly associated with ferroptosis, purine, caffeine, and arachidonic acid metabolism ( and ). Moreover, MSEA uncovered differences in KEGG pathways between the two patient groups: primary bile acid biosynthesis, lysine degradation beta-alanine, histidine, taurine, and hypotaurine metabolism ( and ).

Figure 4

Pathway enrichment analysis of differentially expressed metabolites between COVID-19 patients and community-acquired pneumonia. (A) Kyoto Encyclopedia of Genes and Genomes (KEGG) classification of differentially expressed metabolites between community-acquired pneumonia (CAP) and COVID-19 patients (CP). (B) KEGG pathway enrichment analysis of differentially expressed metabolites. (C) Metabolite Set Enrichment Analysis (MSEA) between CAP and CP.

K-means analysis of differentially expressed metabolites

K-means analysis was performed to determine the variation trend of differential metabolites across the different groups. It revealed significant changing trends divided into six subclasses ( ). Among them, subclass 2, 5, and 6 were in line with expectations, which exhibited increasing or decreasing trends from the HC group to the CAP group to the COVID-19 group. Subsequently, KEGG analysis was performed and showed that the 24 metabolites from subclass 2 were significantly enriched in the phospholipase D and cAMP signaling pathways and porphyrin and chlorophyll metabolism ( and ). The 15 metabolites from subclass 5 were involved in metabolic pathways and fatty acid biosynthesis ( and ). Finally, the 65 metabolites from subclass 6 were associated with arachidonic acid and purine metabolism, taurine and hypotaurine metabolism, serotonergic synapse, and primary bile acid biosynthesis ( and ).

Figure 5

K-mean clustering analysis and pathway enrichment analysis. (A) K-means clustering analysis of differentially expressed metabolites across different groups. (B) KEGG enrichment analysis of metabolites in subcluster 2. (C) KEGG enrichment analysis of metabolites in subcluster 5. (D) KEGG enrichment analysis of metabolites in subcluster 6. HC, healthy control; CAP, community-acquired pneumonia; CP, COVID-19 patients.

Biomarker panel for identifying COVID-19 and its model evaluation

Venn diagrams of mild and severe COVID-19 groups were plotted to characterize the metabolic profile of COVID-19 patients. They showed that the two groups had 92 shared differentially expressed metabolites ( ). These metabolites were enriched with ferroptosis, arachidonic acid metabolism, bile secretion, purine metabolism, and platelet activation KEGG pathways ( and ). Furthermore, MESA indicated that the purine metabolism, primary bile acid biosynthesis, taurine and hypotaurine metabolism, and cysteine and methionine metabolism pathways were significantly differentially expressed between the two groups ( and ). Pearson correlation analysis also revealed a weak correlation between the metabolites ( ). Detailed metabolic information and classification are listed in .

Figure 6

Predictive value of differentially expressed metabolites in predicting COVID-19 patients. (A) Venn diagram showing differentially expressed metabolites between two groups: healthy controls (HC) vs. COVID-19 patients (CP) and community-acquired pneumonia (CAP) vs. CP. (B) KEGG enrichment analysis of common differentially expressed metabolites. (C) Metabolite Set Enrichment Analysis (MSEA) of common differentially expressed metabolites. (D) Pearson correlation of 92 differentially metabolites between the two groups. Only the index of metabolites in the database is shown, and detailed metabolic information and classification are listed in . (E) Receiver operating characteristic (ROC) curves of the two classifiers based on cross-validation in distinguishing COVID-19 patients (CP) from community-acquired pneumonia (CAP). (F) The ROC of the two classifiers based on the cross-validation in distinguishing CP from healthy controls (HC). LR, logistic regression; RF, random forest.

A rigorous machine-learning method was used to explore the predictive value of the differentially expressed metabolites in identifying COVID-19. First, we applied LASSO regression to screen for biomarker metabolites of COVID-19. Seven and fifteen metabolites were identified to discriminate the COVID-19 patients from CAP patients and HC, respectively ( ). Next, random forest and logistic regression algorithms were used to assess their performance via fivefold cross-validation. The ROC analysis showed that the diagnostic model combining seven metabolites performed outstandingly in distinguishing COVID-19 from CAP patients, with an area under the curve (AUC) equaling 1 ( ). In addition, they also performed well in discriminating the COVID-19 group from the control group, with an AUC close to 1 ( ).

Metabolic profile of severe COVID-19 patients

We also performed a differential metabolic analysis of the severe and mild COVID-19 groups to determine the metabolic profile of severe COVID-19. The plotted Venn diagrams of the three groups showed that 21 metabolites were expressed only in severe COVID-19 patients ( ). KEGG pathways were mainly classified into organismal systems, metabolism, human diseases, and environmental information processing ( and ). The enrichment analysis of the differential metabolites revealed they were enriched with purine metabolism, PI3K–Akt signaling, mTOR signaling, and renin secretion pathways ( and ). MSEA also indicated differentially expressed pathways between the groups ( and ): taurine and hypotaurine metabolism, primary bile acid biosynthesis, beta-alanine metabolism, histidine metabolism, and retinol metabolism.

Figure 7

Identification of metabolites specific to severe COVID-19 patients. (A) Venn diagram showing differentially expressed metabolites between three groups: healthy controls (HC) vs. COVID-19 patients (CP), community-acquired pneumonia (CAP) vs. CP, and mild COVID-19 (MC) patients vs. severe COVID-19 (SC). (B) KEGG classification of common differentially expressed metabolites. KEGG enrichment pathway analysis (C) and Metabolite Set Enrichment Analysis (MSEA) (D) of 21 common differentially expressed metabolites.

Subsequently, the differential analysis showed that AMP, dGMP, and sn-glycero-3-phosphocholine were remarkably upregulated in severe COVID-19 patients. By contrast, several carnitine family members were considerably downregulated versus HC subjects, CAP patients, and mild COVID-19 patients ( ). Pearson correlation analysis also uncovered a clear clustering trend among the differentially expressed metabolites. Since a strong correlation existed among the carnitine family members ( ), we applied a random forest algorithm to screen for severe COVID-19 biomarkers ( ). Next, the relative importance of each metabolite was ranked, and the top 10 metabolites (Carnitine C11: DC, Carnitine C6:0 Isomer 1, Carnitine C16:3, Carnitine C14:2, Carnitine C14:2-OH, Carnitine C8-OH, Carnitine C14:1, Carnitine C12-OH, Carnitine C12:1, and Carnitine C14-OH) were selected to construct a diagnostic model for distinguishing between severe COVID-19 patients and mild COVID-19 patients. ROC plots demonstrated that the AUC was 0.921 (0.841–1.000), indicating a relatively high specificity ( ).

Figure 8

Metabolite performance in predicting severe COVID-19. (A) Kruskal test revealing significant differences between all tested groups: healthy controls (HC), community-acquired pneumonia (CAP), mild COVID-19 (MC), and severe COVID-19 (SC) patients. All metabolite profiles were first log2 transformed and then normalized using Z-score standardization. (B) Pearson correlation analysis of differential metabolites. (C) The top ten ranking metabolites are indicated by random forest. (D) Predictive performance of biomarker metabolites evaluated by ROC analysis. *P < 0.05, **P < 0.01, ***P < 0.001, ****P < 0.0001.

Discussion

In 2020, the COVID-19 pandemic swept the globe with grave effects on human health and the social economy. Since then, governments worldwide have invested in a vast workforce and material resources to improve epidemic prevention measures and prevent the disease (9). Along with the extensive application of vaccines, its spread has been controlled to some extent (10). However, the daily number of new COVID-19 infections remains high, with more than 1 million new infections. Dr. Gregory Poland, head of the Vaccine Research Group at Mayo Clinic, predicted that the disease will continue spreading well into the next century. In addition, efficient, targeted treatments for COVID-19 infection are still lacking (11). Thus, identifying the changes of metabolic molecules and pathways during the viral infection is crucial for improving disease management and accelerating drug development.

Currently, untargeted and targeted metabolomics have been used to reveal the variation of acylcarnitines, fatty acid, triglycerides, and sphingomyelins in COVID-19 patients (12–15). In this study, we also found aberrant expression of amino acid, oxidized lipids, fatty acid, and carnitine in COVID-19 patients. Furthermore, arachidonic acid was identified to be significantly increased in COVID-19 patients, compared to HC. The imbalance of arachidonic acid has been confirmed to be associated with multiple inflammatory pathways (16–18). Multiple studies have indicated that arachidonic acid was related to the infection of the COVID-19 and inflammatory cytokine storm (19, 20). Thus, arachidonic acid deserves further investigation and may be a promising target for COVID-19 therapy. Metabolites exhibited great potential as biomarkers for diagnosing COVID-19. Yamilé López-Hernández et al. identified 3 metabolites combinations that can discriminate between COVID-19 patients and HC, with an AUC of 0.947 (21). Song JW et al. proposed ten metabolites that could distinguish between COVID-19 patients and HC (AUC=0.975) (22). Similarly, Barberis E et al. also revealed the diagnostic value of multiple metabolites with an AUC>0.900 for each model (12). In the present study, seven metabolites were identified based on multiple established machine learning, which can distinguish between COVID-19 patients and CAP patients with an AUC of 1.

In this study, 92 metabolites were specific to COVID-19 patients and included amino acids and their metabolites, fatty acyls, nucleotides and their metabolites, organic acids and their derivatives, and oxidized lipids. Atila, Alptug et al. first described the serum amino acid profile and reported dysregulation of amino acid metabolism in COVID-19 patients (23). Amino acids are precursors of many vital molecules and play a key role in immune cell function (24). Abnormal amino acid metabolism causes neurological symptoms and multiorgan failure. For instance, we know that even mild COVID-19 patients have a certain degree of neurological sequelae after recovering from the infection. Conversely, severe patients with COVID-19 may develop severe multiple organ failure during hospitalization and eventually die. Philips, Paez-Franco, and Rees et al. also studied the role of amino acid metabolism in the progression of COVID-19 infection (25–27). They discovered that the differentially expressed metabolites between the patients and HC were enriched with taurine and hypotaurine metabolic pathways. Taurine has a relatively high abundance in leucocytes and is associated with the inflammatory response (28). Therefore, we speculated that an overactive taurine pathway can drive the excessive immune response in COVID-19 patients. Because amino acids are required for viral replication and virulence, we also proposed amino acid pathways as promising targets for drug development.

The pathway enrichment analysis revealed significantly enriched ferroptosis and energy metabolism pathways in patients with COVID-19. Indeed, the serum of patients with COVID-19 showed an iron imbalance (29). Studies have shown that COVID-19 infection will attack hemoglobin, which is also an important pathogenesis step of COVID-19 (30). This will result in dissociation of the porphyrins from iron and releasing iron from stores into the circulation, causing iron overload and ferritin elevation (31). Bellmann-Weiler R et al. also reported that 88.2% of 259 hospitalized COVID-19 patients had an abnormal iron homeostasis, and 68.8% of anemic patients developed anemia of inflammation (32). Anemia and iron metabolism disorder were significantly associated with increased ICU admission and hospital mortality. Phua, Jason et al. found that serum ferritin levels were significantly elevated in patients with severe COVID-19 compared with mild COVID-19 (33). Furthermore, cellular assays indicated that COVID-19 can decrease the mRNA levels of glutathione peroxidase 4 (34). Most patients with severe COVID-19 disease present with multiorgan damage and failure, and lungs, liver, and kidneys are most affected (35). Ferroptosis is an iron-dependent type of programmed cell death involved in the pathogenesis of various body systems, including the heart, liver, kidney, lung, and intestine (36). Severe organ failure in COVID-19 patients typically starts approximately 14 days post-infection (32). Thus, it coincides with the onset of ferroptosis (37). Metabolomics exhibits robust performance in early diagnosing diseases and monitoring progression. In the present study, we employed multiple well-established machine learning to validate the performance of differential metabolites in diagnosing COVID-19 infection, with an AUC of 1. Therefore, our results suggested that research efforts should focus more on metabolic signatures of the disease.

Twenty-one metabolites were identified only in severe COVID-19 patients. Among them, nucleotides and their metabolites, including AMP and dGMP, and sn-glycero-3-phosphocholine were upregulated. In agreement, KEGG analysis demonstrated that purine metabolism pathways were enriched in the severe COVID-19 group. Purine metabolism disorders are responsible for several metabolic diseases. A central component of ATP is AMP, which further hydrolyzes to adenosine. Kondo, Yutaka et al. reported that AMP stimulation can rapidly slow down mitochondrial respiration in mouse and human neurons, thereby reducing dependence on oxygen under a hypo-metabolic state and protecting the brain and other organs from further damage (38). Furthermore, in patients with severe COVID-19, dyspnea and hypoxemia are common symptoms (39). These findings suggested that the AMP increase in severe COVID-19 patients is a self-protection mechanism against hypoxia. It is also worth noting significantly reduced carnitine levels in severe COVID-19 patients. Carnitine is a vitamin-like compound that plays an important role in fatty acid metabolism in granules (40). It is mainly biosynthesized in the kidney, liver, and brain and primarily stored in the skeletal muscle and heart (41). Maintaining carnitine balance is extremely important for regulating and sustaining normal physiological functions, including antioxidant, anti-apoptotic, anti-inflammatory, biomembrane-stabilizing, and anti-fibrosis (40). Low carnitine is associated with multiple diseases, such as advanced liver cirrhosis and cerebral hemorrhage (42). Since patients with severe COVID-19 usually exhibit metabolic disorders and multiple organ dysfunctions, the decrease of serum carnitine in the severe patients may be related to organ dysfunction (43). Furthermore, random forest analysis showed that AMP and carnitine have excellent performance in distinguishing between severe COVID-19 patients and mild COVID-19 patients. These results suggest that understanding the metabolic changes of AMP and carnitine during COVID-19 may advance monitoring disease progression.

However, a few limitations still remain to be noticed in this study. Due to the small number of patients with COVID-19 enrolled in the present study, the diagnostic efficacy of metabolites needs to be further evaluated. Therefore, the results still need further validation in larger prospective cohorts of COVID-19.

Taken together, the present study systematically delineated the serum metabolic profiles in HC subjects, patients with COVID-19, and CAP patients. We demonstrated that the serum metabolites may have a strong potential in identifying COVID-19 patients, especially those with severe disease. This study may provide new insights into COVID-19 diagnosis and treatment.

Data availability statement

The original contributions presented in the study are included in the article/ . Further inquiries can be directed to the corresponding author.

Ethics statement

The studies involving human participants were reviewed and approved by Yue Bei People’s Hospital. The patients/participants provided their written informed consent to participate in this study.

Author contributions

JL, Z-BL, Q-QL, YY, S-QZ, P-FK, and FZ collected blood samples and clinical information. JL, Z-BL and J-CL conceived the study and designed the experiment. JL, Z-BL, S-QZ, and J-CL participated in writing the manuscript. All authors read and approved the final manuscript.

Conflict of interest

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.