Potential Lipid Signatures for Diagnosis and Prognosis of Sepsis and Systemic Inflammatory Response Syndrome

Systemic inflammatory response syndrome (SIRS) and sepsis are two conditions which are difficult to differentiate clinically and which are strongly impacted for prompt intervention. This study identified potential lipid signatures that are able to differentiate SIRS from sepsis and to predict prognosis. Forty-two patients, including 21 patients with sepsis and 21 patients with SIRS, were involved in the study. Liquid chromatography coupled to mass spectrometry and multivariate statistical methods were used to determine lipids present in patient plasma. The obtained lipid signatures revealed 355 features for the negative ion mode and 297 for the positive ion mode, which were relevant for differential diagnosis of sepsis and SIRS. These lipids were also tested as prognosis predictors. Lastly, L-octanoylcarnitine was found to be the most promising lipid signature for both the diagnosis and prognosis of critically ill patients, with accuracies of 75% for both purposes. In short, we presented the determination of lipid signatures as a potential tool for differential diagnosis of sepsis and SIRS and prognosis of these patients.


Introduction
The definition of sepsis, as introduced in 2016, updated several concepts and brought some new ones. Now, sepsis is defined as a life-threatening organ dysfunction caused by a dysregulated host response to infection that includes immune as well as nonimmune responses [1]. All over the world, nearly 6 million people die of sepsis annually [2]. Systemic inflammatory response syndrome (SIRS) is a condition in which the patient presents two of the following signs: tachycardia, fever or hypothermia, leukocytosis or leukopenia and tachypnea. It may also occur in response to various forms of aggression such as infection, trauma or surgery. Almost all septic patients have SIRS, but not all SIRS patients are septic. As an exception to this theory, it has been suggested that there are subgroups of hospitalized elderly patients who do not meet criteria for SIRS on presentation but progress to severe infection and multiple organ dysfunction and death. For this reason, SIRS could be an element of confusion for the diagnosis, management plan or evolution assessment and, eventually, patient prognosis prediction [3]. Table 1 shows a statistical comparison of the two groups for the baseline characteristics of the participants involved in this study. No significant differences were found between the demographic characteristics of the groups. Other prognostic scores did not present a statistically significant difference. Almost no comorbidities were present in the SIRS group; this can be explained by the epidemiological characteristics of this group of patients as they were almost all victims of poly-trauma. A higher frequency of comorbidities is expected in patients in the sepsis group [23], leading to a statistically significant difference for systemic hypertension and for diabetes mellitus. The other comorbidities were less frequent in our patients with sepsis, so there were no statistically significant differences, even in the absence of the comorbidities in the SIRS group. No significant differences were found in the frequency of organ dysfunctions between the groups, since both diagnoses can lead to the occurrence of these dysfunctions. None of these variables had a statistically significant effect on the multiple linear regression model for diagnostic classification (Supplementary Table S3). All of the non-survivor patients in the study died during their intensive care unit (ICU) stay.

Analysis of Plasma Samples
In this study, 42 samples were assessed: 21 plasma samples from male patients diagnosed with sepsis and 21 plasma samples from male patients diagnosed with SIRS. After applying quality control (QC) and non-QC filters and making corrections, final numbers of 733 features for negative ion mode and 1703 features for positive ion mode were obtained. The obtained lipidome data were assessed using principal component analysis (PCA) for both negative ( Figure 1A) and positive ionization modes ( Figure 1B). In negative mode, both groups presented very close individual profiles which impeded complete separation of groups. In positive mode, there is a total overlap of the groups. Supplementary Figure S1 shows PCA for samples and QC, where high-quality data depict QC samples in clusters tighter than those observed for biological samples [24]. Other descriptive analyses, such as volcano plot and heatmap of clustered intensities, were performed for all the features. The results are represented in Figure 2 for both negative and positive modes. These descriptive results show that, despite the difficulty in differentiating groups by PCA, it is still possible to determine features with differential abundances.  Other descriptive analyses, such as volcano plot and heatmap of clustered intensities, were performed for all the features. The results are represented in Figure 2 for both negative and positive modes. These descriptive results show that, despite the difficulty in differentiating groups by PCA, it is still possible to determine features with differential abundances.

Analysis of Lipid Signatures for Diagnosis
In order to identify the most relevant features in the task of correctly classifying the samples by diagnosis (sepsis or SIRS), a selection of the lipid signatures was made with prediction models using the random forest (RF) method implemented in MetaboanalystR. Matching the obtained list of features from the RF model for negative and positive mode with Lipidmaps and Human Metabolome DataBase (HMDB) databases resulted in the annotation of 33 significant features as possible biomarkers for discriminant diagnosis between sepsis and SIRS ( Table 2). Annotated lipids such as L-palmitoylcarnitine, gamma-linolenyl carnitine, linoleyl carnitine and the omega 6 polyunsaturated fatty acid arachidonic acid were found in higher abundance in the sepsis patient's plasma and were significant contributors to differentiation among sepsis and SIRS. The predictive importance of these putatively identified lipids was evaluated in subsequent analyses.    Figure 3 shows the metabolic pathways most associated with the lipids found to be relevant. A large impact on the pathway is related to the importance of the compound within the metabolic network evaluated; a higher log (p) (or lower p-value) indicates the over-representation of the evaluated pathway in relation to the list of compounds consulted. Only 22 compounds were found in the HMDB database. Supplementary Table S2 shows information on the matched lipids and statistics of the enriched pathways.
large impact on the pathway is related to the importance of the compound within the metabolic network evaluated; a higher log (p) (or lower p-value) indicates the over-representation of the evaluated pathway in relation to the list of compounds consulted. Only 22 compounds were found in the HMDB database. Supplementary Table S2 shows information on the matched lipids and statistics of the enriched pathways.

Performance Evaluation of Diagnostic Lipid Signatures Used for Prognostic Prediction
With a more reduced but significant list of features, random forest for multivariate classification was used to assess features′ performances as possible signatures for prognostic classification ( Figure  4). This model had an average accuracy of 61.3% and an AUC = 0.676 (see Supplementary Figure S4). Supplementary Table S1 provides a complete list of ranked scores. Although this model presents

Performance Evaluation of Diagnostic Lipid Signatures Used for Prognostic Prediction
With a more reduced but significant list of features, random forest for multivariate classification was used to assess features performances as possible signatures for prognostic classification (Figure 4). This model had an average accuracy of 61.3% and an AUC = 0.676 (see Supplementary Figure S4). Supplementary Table S1 provides a complete list of ranked scores. Although this model presents low accuracy due to the small number of features selected for prognostic classification, its results enabled the identification of the most relevant features for further analysis.  The lipid L-Octanoylcarnitine was found to be the most relevant for the prognostic classification (samples from patients who survived and died), with a notable difference in importance in relation to the other lipids. To evaluate its individual importance in diagnostic and prognostic classification, classification prediction models were used based on random forest. To

Performance Evaluation of L-Octanoylcarnitine as Diagnostic and Prognostic Predictor
The lipid L-Octanoylcarnitine was found to be the most relevant for the prognostic classification (samples from patients who survived and died), with a notable difference in importance in relation to the other lipids. To evaluate its individual importance in diagnostic and prognostic classification, classification prediction models were used based on random forest. To build the classification models, eight samples were randomly selected and unlabeled (four from each group for each classification) to define a validation group. As a diagnostic signature ( Figure 5A Figure 6A shows a heatmap of clustered intensities of these lipids for samples grouped by diagnosis and subgrouped by prognosis. Figure 6B plots the differences in intensity by each group for the lipids found to be important for both categories.  Figure 6A shows a heatmap of clustered intensities of these lipids for samples grouped by diagnosis and subgrouped by prognosis. Figure 6B plots the differences in intensity by each group for the lipids found to be important for both categories.

Discussion
Sepsis, one of the major causes of death in the world, is a serious medical condition associated with high incidence and mortality rates [25]. The discovery of differentiators of patients with a high chance of poor outcome should optimize the selection of better treatment strategies. Similarly, early discrimination between sepsis and some other similar clinical condition, such as SIRS, would make better decision making possible by preventing the progression of the disease, even before organ dysfunction. Here, we have shown that the differences in the lipidomes of patients diagnosed with sepsis or SIRS are relevant for the patients' prognoses. Although differences in the prognosis of patients with sepsis or SIRS, detected either by different omic or clinical approaches [1,17,21,26], have been previously reported; in the present study, we report the co-occurrence of variations in lipid abundance with diagnostic and prognostic potential.

Discussion
Sepsis, one of the major causes of death in the world, is a serious medical condition associated with high incidence and mortality rates [25]. The discovery of differentiators of patients with a high chance of poor outcome should optimize the selection of better treatment strategies. Similarly, early discrimination between sepsis and some other similar clinical condition, such as SIRS, would make better decision making possible by preventing the progression of the disease, even before organ dysfunction. Here, we have shown that the differences in the lipidomes of patients diagnosed with sepsis or SIRS are relevant for the patients' prognoses. Although differences in the prognosis of patients with sepsis or SIRS, detected either by different omic or clinical approaches [1,17,21,26], have been previously reported; in the present study, we report the co-occurrence of variations in lipid abundance with diagnostic and prognostic potential.  :7)) and acylcarnitines (L-octanoylcarnitine, L-palmitoylcarnitine) were more abundant in the sepsis diagnosed group when compared with the SIRS group, while some species of sulfated steroid (dehydroepiandrosterone sulfate,), fatty acid esters of hydroxy fatty acids (FAHFA 36:4), were more abundant in cases with SIRS when compared with the sepsis cases. When comparing the groups of survivor and non-survivor patients, no univariate adjusted statistical difference (false discovery rate (FDR) p < 0.05) was found, but multivariate relevance was found, especially in L-octanoylcarnitine (more abundant in non-survivor patients) and FAHFA 36:4 (more abundant in survivor patients). In addition, the last two compounds present interaction between the two groups, with abundances codependent on prognosis and diagnosis.
Glycosphingolipids (GSLs) are a subclass of sphingolipids with glycans exposed to the extracellular space. These lipids are abundant components of the cell membrane [27]. GSLs are related to many biological processes including infections by specific pathogens as binding receptors at the surface of host cells [28]. GSLs play a role in immune cell function as a signal transducer (i.e., toxins or IgM antibodies) or in binding lipid rafts to trigger chemotaxis, phagocytosis and phagolysosome formation [28] and are involved in regulatory aspects of T cell biology [29]. Some clinical uses for GSLs, such as lipid-rafts for signaling the presence of pathogens, and pharmacological reduction of GSL are being actively studied [30]. However, so far, there have been no studies that describe or associate AS 1-5, a glycosylated N-acylsphingosine, with immune response, inflammation or infection so far.
Glycerophospholipids or phosphoglycerides are lipids with hydrophobic regions composed of two fatty acids linked to glycerol. Sphingolipids are lipids with a single fatty acid linked to a fatty amine, sphingosine. Both lipids are the main components of biological membranes. A wide variety of these compounds have been reported as differentials in assessing septic mortality [21] or in differentiating stages of sepsis and SIRS [20]. These compounds present an increase in abundance related to the severity of sepsis, being more abundant in septic shock and non-differential in non-infectious SIRS [31]. A confounding factor when analyzing these compounds is the variability of their abundance, sometimes decreased in sepsis, depending on the type and focus of infection (i.e., decrease in lysophosphatidylcholines in community-acquired pneumonia) [32]. This high variability has made its biological interpretation difficult. Interestingly, the compounds of this class identified in our study have a higher mean abundance in sepsis, although with weak univariate statistics but relevance in multivariate differentiation. These compounds are largely associated with lipid peroxidation, whose products may have pro-inflammatory and protective activity against infection [33].
Ceramides play essential roles in cell signaling and contrasting roles within cellular metabolism. Ceramide is involved in cellular responses related to stress, autophagy and apoptosis, whereas S1P, another bioactive lipid of the sphingolipid pathway, stimulates cell survival, proliferation and tissue regeneration [34]. However, it is necessary for further investigation to understand the effect of different lengths of acyl chains on this lipid class. Again, sphingolipids participate in the regulation of the phagosome/lysosome fusion, apoptosis or the inflammatory response [35], facilitating bacterial destruction.
Higher average importance for multivariate model and univariate significance of L-octanoylcarnitine and L-palmitoylcarnitine were found in the sepsis group and just low average importance for gamma-linolenyl carnitine and linoleyl carnitine for the same model. The quaternary ammonium compound carnitine and its acyl esters (acylcarnitines) are essential for the oxidative catabolism of fatty acids and thence for maintaining energy homeostasis in the human body. Downregulation of fatty acid oxidation is evidenced by an increased presence of acylcarnitines in plasma [36]. Their accumulation in the plasma is marked in sepsis non-survivors, indicating a possible mitochondrial dysfunction in energy production. Moreover, it was reported that non-survivor septic patients have mitochondrial dysfunction leading to deficient aerobic catabolism [37] and consequently elevated plasma concentrations of TCA cycle metabolites. Unused acylcarnitines are reversely transported to the cytoplasm and then into the plasma [38]. Levels of these lipids were found to be lower in SIRS and survivor patients, as reported by other studies [20]. In the present study, we looked for a particular abundance profile for prognostic and diagnostic classifications: L-octanoylcarnitine presented the highest abundance in non-survivor sepsis patients when compared to survivor SIRS patients (lowest abundance), non-survivor SIRS and survivor sepsis patients. Its importance was evaluated by univariate and multivariate prediction methods ( Figure 6), with good predictive performance for both diagnoses and prognoses ( Figure 5, which identifies it as a possible lipid signature). This compound is the physiologically active form of octanoylcarnitine, an intermediate fatty acid b-oxidation byproduct. In addition to indicating increased lipid oxidation, L-octanoylcarnitine may indicate increased lipid input [39]. A recent study identified low levels of L-octanoylcarnitine as a biomarker of breast cancer (100% positive predictive value) against samples from healthy individuals, in addition to presenting different levels depending on the size of the tumor, as well as high abundance in tumors with high expression of estrogen and progesterone receptors [40]. This may be related to the high metabolic demand of the tumor. Another study on prostate cancer showed a positive relationship between L-octanoylcarnitine levels and the risk of cancer progression in primary and metastatic samples [41]. There is currently no information that relates this acylcarnitine to sepsis, SIRS or the prognosis of these cases. However, a larger, stratified study covering a wider range of compounds (metabolites and proteins) is needed to infer the biological basis of their variable abundance in the cases presented here.
FAHFA 36:4 is a compound that presented a different pattern to those mentioned above. This fatty acid ester of hydroxy fatty acid was found to be more abundant in samples of surviving patients with SIRS when compared to non-survivors with sepsis (less abundant), survivors with sepsis and non-survivors with SIRS. These lipids are endogenous products present in food and mammalian tissues. To date, more than 16 FAHFA families have been determined. Structurally, each family has different fatty acid and hydroxy fatty acid compositions and multiple isomers by the ester bond position. These compounds have anti-inflammatory and anti-diabetic effects [42]. Although it is not known how they perform their biological activity, recent studies link FAHFA to erythroid nuclear factor 2-related factor 2 (Nrf2) [43]. Their presence is related to resolution or regulation of inflammation, including providing protection against potential infection [44]. Therefore, it is not clear whether the low abundance in patients with sepsis and in non-survivors is a depletion or a result of some altered pathway. No studies have been published that relate FAHFA to the progression and outcome of patients with sepsis or SIRS.
In conclusion, this lipidomics study carried out on plasma taken from male patients with sepsis or SIRS assessed relevant lipids for diagnosing. Then, identified lipids from the previous step were assessed as prognostic signatures. Finally, one relevant lipid, L-octanoylcarnitine, was found to be a promising signature for diagnosis and prognosis. Quantification studies of all relevant metabolites highlighted by this study and their physiological and altered levels in human plasma seem to be an interesting matter for further investigation.

Study Groups
The study samples came from the Universidade São Francisco (USF) Hospital, Bragança Paulista, São Paulo, Brazil. Male patients admitted to the ICU were evaluated. The project was approved by the Research Ethics Committee of the Universidade São Francisco (CAAE:51356315.5.0000.5514) and was developed at the Intensive Care Unit of Universidade São Francisco Hospital. The following inclusion criteria were adopted for the group of critically ill patients: individuals from 15 to 90 years of age admitted to the intensive care unit, either clinical or surgical, in the period. Female patients and patients receiving special diets were not included in the study to avoid gender-related and diet-related lipidomic profile bias [45]. Following SIRS definition criteria [46], 21 male patients with 2 or more signs of SIRS and no suspected or confirmed infection were selected for inclusion in the SIRS group. Patients with organ dysfunction and confirmed infection were selected for inclusion in the sepsis group. Clinical data were collected, including severity score (SAPS III and SOFA on the first day of hospitalization). Clinical and demographic data are provided in Table 1. Additionally, a logistic multiple linear regression model was implemented to evaluate the influence of non-lipidomic variables on the classification (diagnostic) variable.

Sample Collection, Preparation and LC-MS/MS Analysis
Blood samples were collected for daily laboratory monitoring of critically ill patients and aliquots of this material from the first 36 h of hospitalization were used to carry out the analyses of the present study. Labeled ethylenediamine tetraacetic acid (EDTA) blood samples were sent to the Multidisciplinary Research Laboratory of the USF, where the lipidomic analyses were performed. Centrifuged plasma samples were frozen at −80 • C. A mixture of samples from both groups was used as quality control (QC). This pooled sample was divided and extracted along with the remaining samples. CHCl 3 :MeOH solution (2:1, v/v) was used for extraction with 150 mL of plasma sample. Extracted samples were then vortexed for 30 s and centrifuged at 12,000× RPM for 5 min at 4 • C. The bottom organic layer (450 mL) was collected. Nitrogen-dried samples were stored at −20 • C to await analysis. A solution of isopropanol (IPA)/acetonitrile (ACN)/water (2:1:1, v/v/v) was used to reconstitute samples before analysis.

LC-MS Analysis
Following a method previously published by our group [47], untargeted LC-MS analysis was performed using an ACQUITY UPLC coupled to a XEVO-G2XS QTOF mass spectrometer (Waters, Manchester, UK). Liquid chromatography was performed using an Acquity UPLC CSHC18 column (2.1 × 100 mm, 1.7 mm, Waters). The volume of injection was 1 mL. MS E mode was used to separately record positive and negative ion modes in the range of 50-2000 m/z. The injection order was randomly defined and QC samples were analyzed after every ten injections.

Data Acquisition and Preprocessing
The peak alignment, deconvolution, selection of possible adducts and compound annotation based on MS E experiments were obtained using Progenesis QI 2.0 software (Nonlinear Dynamics, Newcastle, UK). Search parameters for putative annotation were precursor mass error 5 ppm and fragment tolerance 10 ppm. At this stage, putative identification using LIPID MAPS [48] database and the Human Metabolome Database (HMDB) [49] was defined by fragmentation score, mass accuracy and isotope similarity. Annotation of compounds was classified in accordance with the Metabolomics Standards Initiative (MSI) [50], where ions with some level of match with MS/MS database reached level 2 while compounds putatively identified by exact mass, using the mummichog algorithm, reached level 3. Progenesis QI generated a table of ion intensity by sample and ions. Ions were labeled according to their retention time and mass-to-charge (m/z) ratio. Preprocessed data are available as Supplementary Materials files: Spreadsheet_1 Sepsis SIRS negative mode, for negative mode; Spreadsheet_2 Sepsis SIRS positive mode for positive mode.

Statistical Analysis
MetaboAnalystR 3.0 [51], statTarget2 [52] and Bioconductor package manager using R programming language [53], were used to perform statistical analyses. Quality control based signal correction was performed using random forest implementation (QC-RFSC) [54]. According to the "80% rule" [55], peaks present in more than 80% of the samples of each group were kept for further analysis. The K-nearest neighbor algorithm was used to impute the remaining missing values. Further data filtering removed variables with low variance based on the interquartile range (IQR) [56]. Then, the corrected data were log-transformed and normalized using the Pareto scale [57].

Exploratory Analysis
For univariate descriptive analyses, a volcano plot was used to represent features with FDR-adjusted p-values < 0.05 using t-test and 2-fold intensity between groups for each m/z. Principal component analysis (PCA) was used to distinguish sample cluster distribution in the first two principal components. A heatmap and unsupervised hierarchical clustering of 50 features with the lowest adjusted p-value < 0.05 depicts differential peaks.

Analysis of Biomarkers for Diagnosis
The biomarker analysis module implemented in the MetaboanalystR package was used on the MS peak intensities table for all the samples for detecting relevant features for diagnostic classification. The random forest method, a classification ensemble algorithm, was used for classification and feature selection models. To construct ROC curves, balanced sub-sampling and Monte Carlo cross-validation (MCCV) with two thirds (2/3) of the samples for training were used to evaluate feature importance. The test subgroup (1/3 of samples) was used to build a classification model for top n (1 to 100) important features. The performance and confidence interval of each model were calculated, repeating the procedure multiple times. The RF model produces a reduced list of features ranked by value of importance. All the features obtained here were then used in the annotation stage.

Putative Identification of Lipids and Metabolomics Pathway Analysis
In addition to the putative identification using Progenesis QI described above, the mummichog V2 algorithm [58] was used for MS peaks, without prior annotation. This method identifies lipids based on mass-to-charge ratios (m/z), p-values, fold change, retention time and mixed analytical mode (positive and negative ions), which were used to interrogate the KEGG library. Molecular weight tolerance at 5 ppm and a customized adduct list were used. Only lipidic matched compounds with registered LipidMaps entries were kept. A final manually curated list of identified lipids was obtained using Progenesis QI putative identification and the mummichog-identified lipid list. Using the identified compound list, metabolomics pathway analysis (MetPA) was used to identify biological pathway impact associated with the differences between study groups.

Performance Evaluation of Diagnostic Biomarkers Used for Prognostic Prediction
To assess whether the lipids identified as diagnostic biomarkers could also be predictive for prognostic classification, these lipids were used to build a random forest predictive model for the prognosis. The most relevant lipid was further individually evaluated as a diagnostic and prognostic individual biomarker. For a more stringent evaluation as a possible biomarker, the predictive model tested a subgroup of unlabeled samples. A random forest model was then trained with the labeled subgroup of samples for a single compound, thus alleviating the training bias for which it was initially selected in the diagnostic classification. ANOVA two-way was used for final clustering and visualization of lipid relevant to both diagnosis and prognosis categories.

Acknowledgments:
We thank each participant involved in this research from the São Francisco University Hospital (HUSF).

Conflicts of Interest:
The authors declare no conflict of interest. The funders had no role in the design of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript, or in the decision to publish the results.

Abbreviations
The