Targeted metabolomics for discrimination of systemic inflammatory disorders in critically ill patients.

The occurrence of systemic inflammatory response syndrome (SIRS) remains a major problem in intensive care units with high morbidity and mortality. The differentiation between noninfectious and infectious etiologies of this disorder is challenging in routine clinical practice. Many biomarkers have been suggested for this purpose; however, sensitivity and specificity even of high-ranking biomarkers remain insufficient. Recently, metabolic profiling has attracted interest for biomarker discovery. The objective of this study was to identify metabolic biomarkers for differentiation of SIRS/sepsis. A total of 186 meta-bolites comprising six analyte classes were determined in 143 patients (74 SIRS, 69 sepsis) by LC-MS/MS. Two markers (C10:1 and PCaaC32:0) revealed significantly higher concentrations in sepsis. A classification model comprising these markers resulted in 80% and 70% correct classifications in a training set and a test set, respectively.This study demonstrates that acylcarnitines and glycerophosphatidylcholines may be helpful for differentiation of infectious from noninfectious systemic inflammation due to their significantly higher concentration in sepsis patients. Considering the well known pathophysiological relevance of lipid induction by bacterial components, metabolites as identified in this study are promising biomarker candidates in the differential diagnosis of SIRS and sepsis.

critically ill patients ( 1,2 ). In contrast to the progress made in the symptomatic supportive therapy of organ dysfunction, diagnosis of the underlying infectious-induced etiology of SIRS is often delayed because clinical symptoms are broad and unspecifi c ( 3,4 ). Because delays in the administration of antimicrobial therapy increase mortality ( 5 ), biomarkers indicating infection in the critically ill are vital. Furthermore, avoiding antibiotics in patients with SIRS without infection reduces the risk for developing antibiotic resistance ( 6 ).
Many efforts have been made to fi nd biomarkers that allow discrimination of noninfectious SIRS and sepsis at an early stage of the disease. Most of those studies addressed the transcriptome or proteome for biomarker identifi cation (7)(8)(9). Even though 178 different biomarkers for sepsis have been proposed ( 7 ), most of them have only prognostic and not discriminative value. Only few have been suggested for differentiation of noninfectious induced SIRS from sepsis, and the majority have not been validated for clinical routine use ( 9 ). The only marker widely accepted in the clinical setting is procalcitonin. However, procalcitonin remains controversial for differentiation of SIRS/sepsis because it has been described to be elevated after major surgery, a common cause of SIRS without underlying infection (10)(11)(12)(13)(14).
Technical advances in mass spectrometry opened a new fi eld to this research. In a previous study, we identifi ed a peptide biomarker by proteome analysis that discriminates noninfectious induced SIRS from sepsis, which is being validated in other patient groups ( 15 ). Because SIRS and sepsis are accompanied by severe metabolic alterations (16)(17)(18)(19)(20), we hypothesize that a systematic analysis of the Abstract The occurrence of systemic infl ammatory response syndrome (SIRS) remains a major problem in intensive care units with high morbidity and mortality. The differentiation between noninfectious and infectious etiologies of this disorder is challenging in routine clinical practice. Many biomarkers have been suggested for this purpose; however, sensitivity and specifi city even of highranking biomarkers remain insuffi cient. Recently, metabolic profi ling has attracted interest for biomarker discovery. The objective of this study was to identify metabolic biomarkers for differentiation of SIRS/sepsis. A total of 186 metabolites comprising six analyte classes were determined in 143 patients (74 SIRS, 69 sepsis) by LC-MS/MS. Two markers (C10:1 and PCaaC32:0) revealed signifi cantly higher concentrations in sepsis. A classifi cation model comprising these markers resulted in 80% and 70% correct classifi cations in a training set and a test set, respectively. This study demonstrates that acylcarnitines and glycerophosphatidylcholines may be helpful for differentiation of infectious from noninfectious systemic infl ammation due to their signifi cantly higher concentration in sepsis patients. Considering the well known pathophysiological relevance of lipid induction by bacterial components, metabolites as identifi ed in this study are promising biomarker candidates in the differential diagnosis of SIRS and sepsis . The development of systemic infl ammatory response syndrome (SIRS) associated with multiple organ dysfunction is accompanied by high morbidity and mortality in a clinically suspected infection as evidenced by one or more of the following: white cells in a normally sterile body fl uid, perforated viscus, radiographic evidence of pneumonia in association with the production of purulent sputum, and a syndrome associated with a high risk of infection (e.g., ascending cholangitis). Patients admitted to the ICU after major cardiac or vascular surgery who did not fulfi ll SIRS criteria served as a control group (ICU controls). Patients less than 18 years old, pregnant, or lacking informed consent by their representatives were excluded. The local ethics committee approved this study.
According to a standardized protocol, 10 ml of blood was drawn into EDTA tubes (Sarstedt, Nuembrecht, Germany) immediately followed by the addition of 800 µl of a protease inhibitor cocktail (Complete Mini, Roche). Tubes were gently mixed and centrifuged at 3,750 rpm (2516 g at the middle of the tube) and 4°C for 10 min. Plasma aliquots were immediately frozen and stored at Ϫ 80°C.

Determination of metabolite concentration
PITC, ammonium acetate, and HPLC water were purchased from Merck; methanol (HPLC grade) and acetonitril (LC-MS grade) were purchased from Roth; and formic acid was purchased from Sigma Aldrich. Metabolite concentrations were obtained using the AbsoluteIDQ kit p180 (Biocrates Life Science AG, Austria) according to manufacturer's instructions on an API4000™ LC/MS/MS System (AB SCIEX, USA) equipped with an electrospray ionization source, an Agilent G1367B metabolome may lead to the identifi cation of new diseasespecifi c biomarkers. Although there are several studies addressing single metabolites or metabolite groups in sepsis, no study has investigated the differential metabolom of SIRS patients with and without infection. Because LC-MS/ MS allows simultaneous detection and quantifi cation of up to several hundred metabolites in one sample in a relatively short time, we chose to determine 186 metabolites comprising six analyte classes in these patients.

Patients and sample collection
Patients admitted to the intensive care unit (ICU) of the Department of Anesthesiology and Intensive Care Medicine, Jena University Hospital, from September 2002 until September 2003 were screened for eligibility. Patients were enrolled if they fulfi lled the sepsis criteria according to the ACCP/SCCM consensus conference or the criteria for SIRS of noninfectious origin with acute organ dysfunction, respectively (ICD-10-GM code) ( 3,4 ). In patients with SIRS of noninfectious origin, blood samples were obtained within 24 h after ICU admission or at the onset of symptoms. In patients with sepsis, samples were obtained within 24 h after onset of the fi rst sepsis-induced organ dysfunction. Sepsis patients had to have a microbiologically documented infection or analytes in a defi ned extracted ion count section to those of specifi c labeled internal standards or nonlabeled, nonphysiological standards (semiquantitative) provided by the kit plate. The nomenclature of phospholipids comprises a substance class of the same mass rather than a single substance. It includes the number of carbon atoms and the number of double bonds of both fatty acids linked to C 1 and C 2 of glycerophosphatidylcholine. The nature of fatty acid linkage is expressed as aa for diacyl or ae for acyl-alkyl.

Statistics
Statistical analysis was performed via SPSS Statistics 19 (IBM, USA). The dataset was divided into a training set and a test set. Normal distribution of the metabolite concentrations was tested. Depending on the outcome, Student's t -test (normal distribution) or the Mann-Whitney U test (no normal distribution) was chosen for determination of statistical signifi cance. The level of statistical signifi cance (0.05) was corrected with the Bonferroni method and then defi ned as 2.7 × 10 . ROC curves were obtained by plotting sensitivity against 1-specifi city. The obtained area under the curve (AUC) values are a measure for the predictive power of the analyte: the higher or lower the AUC (converges 1 or 0), the better is the classifi cation by the corresponding analyte. For the combination of analytes, a binary logistic regression analysis was performed for the training set and applied to the test set. The obtained predicted probabilities for correct classifi cation of each value can be plotted as ROC-curve. The proportion of correct classifi cation (sensitivity and specifi city) for the training set and the test set were obtained. autosampler, and the Analyst 1.51 software (AB SCIEX, USA). Samples (10 µl) were pipetted onto the spots of the kit plate. Spots were dried at room temperature in a nitrogen evaporator drying unit for 30 min. Twenty microliters of 5% PITC reagent was pipetted onto the spots and incubated for 20 min at RT. The plate was dried under the nitrogen evaporator for 60 min. Ammonium acetate (300 µl of 5 mM) in methanol was added to each well and incubated on a shaker (450 rpm) for 30 min. The plate was centrifuged at 100 g for 2 min, receiving about 250 µl sample in plate 1 (fl ow injection analysis [FIA] plate). The upper plate was removed, and 150 µl of each sample was transferred into a second plate (LC-MS plate). HPLC water (150 µl) was added to the LC-MS plate, and 500 µl of MS running solvent (Biocrates solvent diluted in methanol) was added to the FIA plate. The LC-MS plate was measured fi rst by scheduled multiple reaction monitoring, and the FIA plate was stored at 4°C. Sample (10 µl) was injected onto a Zorbax Eclipse XDB C18, 3 × 100 mm column (Agilent) coupled to a C18, 4 × 3 mm security guard precolumn (Phenomenex) and eluted with solvent A (HPLC water + 0.2% formic acid) and solvent B (acetonitrile + 0.2% formic acid). Peaks were integrated, and concentrations were obtained (in comparison to internal standards) with Analyst 1.51. Evaluation of calibration curves, blanks, quality controls, and samples was accomplished in the MetIQ software (an integral part of the kit). The FIA plate was measured by FIA, and detection of fragments was performed in multiple reaction monitoring mode. Sample (20 µl) was injected directly into the MS at a fl ow of 30 µl/min with MS running solvent. Concentrations were calculated and evaluated in the Analyst/MetIQ software by comparing measured patients in the training set were correctly classifi ed by the two markers (C10:1 + PCaaC32:0), and these results were reproducible in the test set (70% correct classifi ed samples). Moreover, the combination of more than two markers does not yield improvement in classifi cation, which can be traced back to a strong correlation of markers within one substance class (Fig. 3) .
The training set mainly consisted of gram-positive sepsis cases, whereas the test set contained more gram-negative cases ( Table 1 and Supplementary Table I ). The discrimination of SIRS and sepsis by selected analyte concentrations worked well in both sets ( Table 2 ), suggesting that the observed increase of analyte concentrations is a general host response to infection independent on the gram status of bacteria. However, only 11 out of 23 g negative sepsis cases compared with 13 out of 15 g positive sepsis cases (training set and test set) were classifi ed correctly (Supplementary  Table IIA ). This might suggest that C10:1 + PCaaC32:0 are more suitable for the detection of gram-positive sepsis, but this hypothesis needs to be confi rmed in a larger patient cohort .

DISCUSSION
Sepsis is accompanied by severe metabolic changes, such as an increase of plasma lipids, including triglycerides (TGs), phospholipids, and free fatty acids (FAs). In animals, plasma lipids can be induced by lipopolysaccharides (LPSs) and lipoteichoic acid, a component of the cell wall of gram-positive bacteria, due to increased lipolysis and impaired lipid catabolism ( 16 ). This seems to be RESULTS Samples of 143 patients were divided into a training set (30 sepsis, 33 noninfectious SIRS) for marker identifi cation and establishment of a model for discrimination of SIRS and sepsis and a test set (39 sepsis, 41 noninfectious SIRS) for validation of that model. Metabolite concentrations of 16 ICU control subjects were determined. Clinical characteristics of the patients are depicted in Table 1 .
Concentrations of 186 metabolites belonging to six analyte classes (acylcarnitines, amino acids, biogenic amines, glycerophospholipids, sphingolipids, and carbohydrates) were determined by LC-MS/MS. Analytes and their respective P values for these classes in the trainings set are shown in Fig. 1 . At a signifi cance level of 0.05, analytes from all classes but the carbohydrates were signifi cantly different. After Bonferroni correction, the level of significance was defi ned as 2.7 × 10 Ϫ 4 . At this level, only acylcarnitines and the glycerophospholipids were signifi cantly higher in sepsis samples. Boxplots of the respective markers are shown in Fig. 2 .
Metabolite concentrations for most of the 186 analytes in ICU controls were at the SIRS level (data not shown).
To analyze the diagnostic value of these markers, we included all signifi cant markers of the training set after Bonferroni correction into a model for classifi cation of samples in the SIRS and the sepsis group. For validation, we applied this model to a test set. The AUC values for the single markers and for the combination of markers as well as the percentages of correctly classifi ed patients for the test set and the training set are depicted in Table 2 . About 80% of Because PCs with two fatty acids of C16-C20 are major components of mammalian lipoproteins, we observed PCaaC32-PCaaC36 as the most signifi cant PCs in our samples. Furthermore Drobnik et al. (25) have shown decreased lysophosphatidylcholine/PC and increased ceramide/sphingomyelin ratios in septic patients compared with healthy control subjects. However, although comparison of sepsis patients with healthy control subjects might not be appropriate regarding our hypothesis, this is still in concordance with our results. Again, highly increased PCs seem to be sepsis specifi c because they are not detectable in SIRS samples without infection compared with ICU control subjects.
Recent data suggest that lipemia in sepsis is not only a reaction of the host to provide energy but is also an integral part of the innate immunity to neutralize bacterial toxins ( 26 ) by which the lipid-A moiety of LPS is embedded in the phospholipid layer of lipoproteins. This neutralizing effect has been shown for all lipoproteins in vitro and in vivo ( 27,28 ) and has been attributed to the PC content of lipoproteins.
Together, these data indicate that lipemia in systemic infl ammation, particularly the increase in PCs, is a specifi c host response to infection-induced infl ammation (bacteria and their toxins). Because it has been shown that this effect occurs within 2 h after LPS administration and is sustained for at least 24 h ( 29, 30 ), medium-chain acylcarnitines and phospholipids might be appropriate biomarkers for differentiation of sepsis and noninfectious SIRS in the early stage of the disease. the result of a coordinated, predominantly down-regulation of many proteins taking part in FA transport , FA oxidation (carnitine palmitoyl transferase, medium chain acyl CoA dehydrogenase [MCAD]), and the breakdown of triglycerides (LPL) (17)(18)(19)(20). A well characterized enzyme taking part in FA oxidation that is suppressed in sepsis is MCAD ( 17,19 ). This mitochondrial enzyme processes FAs bound to CoA (acyl-CoA) with a chain length of C5-C14. Inhibition of oxidation by MCAD leads to accumulation of medium-chain acyl-CoA in the mitochondria, which is quite toxic for the cell. As a consequence, medium-chain FA intermediates are transferred to carnitine and are capable of leaving the mitochondria and the cell as acylcarnitine. This is supported by our observation that most of the acylcarnitines increased in sepsis samples belong to medium-chain acylcarnitines ( Figs. 1 and 2 ). Suppression of MCAD has also been shown in zymosan-and turpentineinduced infl ammation and therefore has been suggested as a general response in infl ammation ( 21 ), but we could not confi rm this in our patients because the intense rise of MCAD intermediates seems to be sepsis specifi c and does not occur in noninfectious SIRS compared with ICU control subjects. However, these models might not properly refl ect the conditions in our SIRS patients.
In plasma, most lipids are bound to albumin, chylomicrons, and lipoproteins. The latter particles, specifi cally LDL and VLDL, have also been shown to be highly elevated in sepsis (22)(23)(24). We suppose that the elevated levels of phosphatidylcholines (PCs) observed in sepsis patients are attributed to high concentrations of lipoproteins. Shown are the AUCs for the single markers and their combination in a binary logistic regression analysis and the correct classifi cations in percent for the combination of two to six markers. The combination of more than two markers does not yield further improvement, probably due to strong correlation between markers within one analyte group.
we could exclude that differences in mechanical ventilation, severity of disease, age, or gender confound the fi ndings (data not shown), we cannot out rule that there are other factors that might infl uence the results. Thus, validation of our markers in independent, prospectively collected larger sample cohorts is mandatory. Considering that mass spectrometry, with analysis times of 5-10 min for determination of 186 metabolites in parallel, is fairly quick, metabolic markers or marker classes as identifi ed in this study may serve as very promising candidates in the differential diagnosis of sepsis and noninfectious SIRS in the clinical routine setting.
It is known that medication can alter metabolism. Therefore, we gathered information on blood products (antithrombin III, packed red blood cells, fresh frozen plasma, prothrombin complex concentrate, and platelet concentrate) and medications such as glucocorticoids (hydrocortisone, methylprednisolone, prednisolone), insulin, and propofol (a hypnotic agent often used in general anesthesia that has been described to infl uence metabolism), and tested for correlation with C10:1 and PCaaC32:0. None of the aforementioned parameters had a signifi cant infl uence on those markers (Supplementary Table III and Supplementary Figs. II and III ). Also, clinical data that might have confounded the results were tested for correlation with metabolic markers and differences in falsely and correctly classifi ed patients (Supplementary Table II  In conclusion, we were able to demonstrate that SIRS and sepsis patients can be distinguished with high sensitivity and specifi city by using two metabolic markers (C10:1 and PCaaC32:0). These markers belong to the MCAD intermediates and the most common glycerophospholipids, respectively. In animal models, they have been shown to be responsive to cell wall components of gram-negative and gram-positive bacteria, indicating that they are specifi c for infection, which we could confi rm in our patient cohort. However, the study has several limitations. Concentrations of phosphatidylcholines were determined semiquantitatively to nonphysiological standards. Another limitation is that the sample size is rather small. Although Fig. 3. ROC curve of C10:1, PCaaC32:0, and the combination of both markers. ROC curves were created for the best markers of the acylcarnitines (green) and glycerophosphatidylcholines (yellow) in the training set. Both markers were included in a binary logistic regression model and the obtained predicted probabilities plotted as ROC curve representing the combination of those markers (gray). Respective AUCs are shown for the single markers and the combination.