Targeted metabolomics identifies accurate CSF metabolite biomarkers for the differentiation between COVID-19 with neurological involvement and CNS infections with neurotropic viral pathogens

Abstract

Background: COVID-19 is primarily considered a respiratory tract infection, but it can also affect the central nervous system (CNS), which can result in long-term sequelae. In contrast to CNS infections by classic neurotropic viruses, SARS-CoV-2 is usually not detected in cerebrospinal fluid (CSF) from patients with COVID-19 with neurological involvement (neuro-COVID), suggesting fundamental differences in pathogenesis.

Methods: To assess differences in CNS metabolism in neuro-COVID compared to CNS infections with classic neurotropic viruses, we applied a targeted metabolomic analysis of 630 metabolites to CSF from patients with (i) COVID-19 with neurological involvement [n = 16, comprising acute (n = 13) and post-COVID-19 (n = 3)], (ii) viral meningitis, encephalitis, or myelitis (n = 10) due to herpes simplex virus (n = 2), varicella zoster virus (n = 6), enterovirus (n = 1) and tick-borne encephalitis virus (n = 1), and (iii) aseptic neuroinflammation (meningitis, encephalitis, or myelitis) of unknown etiology (n = 21) as additional disease controls.

Results: Standard CSF parameters indicated absent or low neuroinflammation in neuro-COVID. Indeed, CSF cell count was low in neuro-COVID (median 1 cell/µL, range 0-12) and discriminated it accurately from viral CNS infections (AUC = 0.99) and aseptic neuroinflammation (AUC = 0.98). 32 CSF metabolites passed quality assessment and were included in the analysis. Concentrations of differentially abundant (fold change ≥|1.5|, FDR ≤ 0.05) metabolites were both higher (9 and 5 metabolites) and lower (2 metabolites) in neuro-COVID than in the other two groups. Concentrations of citrulline, ceramide (d18:1/18:0), and methionine were most significantly elevated in neuro-COVID. Remarkably, triglyceride TG(20:1_32:3) was much lower (mean fold change = 0.09 and 0.11) in neuro-COVID than in all viral CNS infections and most aseptic neuroinflammation samples, identifying it as highly accurate biomarker with AUC = 1 and 0.93, respectively. Across all samples, TG(20:1_32:3) concentration correlated only moderately with CSF cell count (ρ = 0.65), protein concentration (ρ = 0.64), and Q-albumin (ρ = 0.48), suggesting that its low levels in neuro-COVID CSF are only partially explained by less pronounced neuroinflammation.

Conclusions: The results suggest that CNS metabolite responses in neuro-COVID differ fundamentally from viral CNS infections and aseptic neuroinflammation and may be used to discover accurate diagnostic biomarkers in CSF and to gain insights into differences in pathophysiology between neuro-COVID, viral CNS infections and aseptic neuroinflammation.

Keywords: Biomarker; COVID-19; Ceramides; Cerebrospinal fluid; Diagnosis; Encephalitis; Long COVID; Long-COVID; Meningitis; Metabolism; Neuroinflammation; SARS-CoV-2; Triglycerides.

Fig. 1

Cerebrospinal fluid (CSF) metabolite populations differ between COVID-19 with neurological involvement and non-COVID encephalitis/meningitis/myelitis. Principal component analysis (PCA) was performed based on 32 metabolites (detailed in Table S2) in the comparison between COVID-19 and viral central nervous system (CNS) infections [dCtrl (viral)] and clinical encephalitis/meningitis/myelitis without pathogen detection [dCtrl (unknown)], respectively. The y-axis label of A applies also to B and C. A COVID-19 vs. dCtrl (viral). B COVID-19 vs. dCtrl (unknown). C dCtrl (viral) vs. dCtrl (unknown). A PCA comprising all three groups is shown as Figure S1

Fig. 2

Classification of the cerebrospinal fluid (CSF) samples by unsupervised hierarchical clustering analysis. Analysis based on the same metabolite concentration data as used for the principal component analysis (PCA) in Fig. 1. The samples are organized along the x-axis, the diagnostic groups are indicated by a color code in the legend. The metabolites are clustered along the y-axis. Each colored cell on the map corresponds to the relative concentration of the analyte with respect to the mean-centered and divided by standard deviation of the analyte (z-score). A COVID-19 vs. dCtrl (viral). B COVID-19 vs. dCtrl (unknown). A clustering analysis based on mean metabolite concentrations per group is shown as Figure S3

Fig. 3

Identification of cerebrospinal fluid (CSF) metabolite biomarkers by receiver operating characteristic (ROC) curve analysis. Dispersion plots based on the same CSF metabolites as used for the principal component analysis (PCA) and clustering analyses shown in Figs. 1 and 2. The y-axis represents the ratio of mean concentrations in COVID-19/dCtrl (viral) or dCtrl (unknown). Area under the ROC curve (AUC) values are plotted along the x-axis. Each circle represents one metabolite, and the fill color indicates the asymptotic significance of the ROC curve. The y-axis label of A also applies to B. A COVID-19/dCtrl (viral). B COVID-19/dCtrl (unknown)

Fig. 4

Concentrations of the six best-validated metabolite biomarkers for the discrimination between COVID-19 with neurological involvement and dCtrl (viral). They also include the best five for the discrimination between COVID-19 and dCtrl (unknown). The red dotted line represents optimal cut-off values in the ROC curve, indicating the optimal trade-off between sensitivity and specificity, generated by the Youden index method. The y-axis labels of A and D also apply to B, C, E, F. A TG (20:1_32:2); B Cer(d18:1/18:0); C Cit; D Met; E Taurine; F SDMA. Significance of between-group differences is indicated as **p < 0.01, ***p < 0.001, ****p < 0.0001 (Mann–Whitney U-test)

Fig. 5

Correlation analysis between metabolite concentrations and standard blood and cerebrospinal fluid (CSF) parameters across all samples. Correlation is based on the same CSF metabolites as used for the analyses shown in Figs. 1 and 2. The y-axis represents the Spearman rank correlation coefficient. Each circle represents a specific metabolite, and the fill color indicates significance of correlation. The labelled metabolites are the best validated biomarkers with a Youden index ≥ 0.9 (Table 3). TG: triglyceride; Cer: ceramide; Met: methionine, Cit: citrulline

Fig. 6

Immunosuppressive therapy has only a minor effect on the cerebrospinal fluid (CSF) metabolome. A Principal component analysis (PCA) between the COVID-19 samples without immunosuppressive therapy and dCtrl (viral). The PCA was performed based on the same metabolites as in Fig. 1. B Differential abundance analysis based on the same data set as the PCA. The ratio of mean concentration (“fold change (FC)”, COVID-19/controls) is plotted log₂ transformed on the y-axis, adjusted p value log₁₀ transformed (corrected for multiple testing by Benjamini–Hochberg correction) on the x-axis. The threshold was set to FC ≥|1.5| and the adjusted p < 0.05. TG: triglyceride; Cer: ceramide; Met: methionine; C0 (carnitine) and C4 (butyrylcarnitine): acylcarnitines; Leu: leucine; Orn: ornithine; Met: methionine; Cit: citrulline; SDMA: symmetrical dimethylarginine