S1PR1-associated Molecular Signature Predicts Survival in Patients with Sepsis

Abstract

Background:

Sepsis is a potentially life-threatening complication of an underlying infection that quickly triggers tissue damage in multiple organ systems. To date, there are no established useful prognostic biomarkers for sepsis survival prediction. Sphingosine-1-phosphate (S1P) and its receptor S1P receptor 1 (S1PR1) are potential therapeutic targets and biomarkers for sepsis, as both are active regulators of sepsis-relevant signaling events. However, the identification of an S1PR1–related gene signature for prediction of survival in sepsis patients has yet to be identified. This study aims to find S1PR1-associated biomarkers which could predict the survival of patients with sepsis using gene expression profiles of peripheral blood to be used as potential prognostic and diagnostic tools.

Methods:

Gene expression analysis from sepsis patients enrolled in published datasets from Gene Expression Omnibus were utilized to identify both S1PR1 related genes (co-expression genes or functional related genes) and sepsis survival related genes.

Results:

We identified 62-gene and 16-gene S1PR1-related molecular signatures (SMS) associated with survival of patients with sepsis in discovery cohort. Both SMS genes are significantly enriched in multiple key immunity-related pathways that are known to play critical roles in sepsis development. Meanwhile, the SMS performs well in a validation cohort containing sepsis patients. We further confirmed our SMSs, as newly developed gene signatures, perform significantly better than random gene signatures with the same gene size, in sepsis survival prognosis.

Conclusions:

Our results have confirmed the significant involvement of S1PR1 dependent genes in the development of sepsis and provided new gene signatures for predicting survival of sepsis patients.

Background

Although clinical medicine has advanced significantly in past decades, sepsis remains a significant cause of death in hospitalized patients around the world. Roughly one million cases result in an estimated 200,000 deaths every year in the United States ⁽¹⁾. Sepsis is a systemic syndrome triggered by infections, including bacterial infections. During infection, the host immune system interacts with pathogens and produces downstream inflammatory and anti-inflammatory responses leading to a global response which could result in multi-organ dysfunction and mortality (approximately 13% in sepsis-3 ⁽²⁾ (previous term is severe sepsis) patients ⁽³⁾ and 40% in septic shock patients) ⁽⁴⁾.

Fast and appropriate treatment is the foundation for therapy of sepsis. However, differentiating sepsis-3 patients from sepsis cases may be a problem. Currently, no biomarkers are available for sepsis which allow for a rapid and dependable prognosis for sepsis patients ⁽⁵⁾. Due to the complicity of the syndrome, a prognostic biomarker of mortality or disease development is not well established.

Sphingosine-1-phosphate (S1P) is a bioactive plasma-membrane sphingolipid which plays an essential role in multiple critical signaling pathways including immunomodulation ⁽⁶⁾. S1P production has been found to be well correlated with the susceptibility and severity of many inflammatory diseases such as asthma ⁽⁷⁾, autoimmunity ⁽⁸⁾, and sepsis ^{(9, 10)}. S1P signaling occurs via interactions with a family of G-protein-coupled S1P receptors (S1PR1–S1PR5). These receptors perform vital but distinct roles in a range of cellular processes variably characterized by their receptor specific downstream signaling pathways. S1PR1, the first identified S1P receptor and the most well characterized, is the receptor most predominantly expressed in the immune system. S1PR1 regulates multiple biological processes including cellular proliferation ⁽¹¹⁾, maintenance of human embryonic stem cells ⁽¹²⁾ migration of various types of immune cells including lymphocytes, regulatory T cells, and dendritic cells ^(13–15), and cytokine secretion ⁽¹⁶⁾. Interestingly, blood possesses the highest S1P concentration compared to all other body tissues ⁽¹⁷⁾, and the circulating plasma S1P maintains vascular integrity against endogenous and exogenous endothelial integrity disruptors including growth factors, pathogens, and toxins ⁽¹⁸⁾.

S1P-S1PR1 axis is considered to play a vital role in the development of sepsis. There are multiple reports demonstrating the molecular mechanism of S1PR1 in the pathogenesis of sepsis ⁽⁹⁾. One recent report suggested the S1PR1 agonist SEW2871 is effective in improving the microcirculation and protecting against kidney injury in a murine model of sepsis ⁽¹⁹⁾.

During infection and inflammation, the activation of S1PR1 by S1P is a negative regulator, reversing the increase in pulmonary vascular permeability induced by lipopolysaccharide (LPS) ⁽²⁰⁾. The ligation of S1P to S1PR1 maintains the homeostatic barrier within the vascular system. Therefore the S1P–S1PR1 axis restores the integrity of the vascular barrier and stabilizes blood vessels ⁽²¹⁾. Consequently, it is likely that activation of vascular S1PR1 reduces the severe complications in sepsis and is suggested to improve the survival rate of sepsis patients.

This study aimed to understand the involvement of S1PR1 in sepsis and identify S1PR1-related genes that also regulate sepsis survival. Firstly, we analyzed two Gene Expression Omnibus (GEO) datasets from whole blood samples of sepsis patients and built the first S1PR1-related gene signature, which we show can also predict the outcomes of sepsis patients. By analyzing gene expression data, we found a 62-gene and 16-gene S1PR1-associated molecular signature that can even predict survival of sepsis patients. These results suggest that peripheral blood gene expression data can be used to predict the survival of sepsis patients.

Methods

GEO datasets selection

We carried out a systematic search in the GEO database for clinical studies of sepsis. Search items included whole blood samples from sepsis patients and microarray datasets with survival data. Several studies were identified after exclusion of studies in animals, RNA-seq datasets or non whole blood samples. GEO: GSE54514 was designated as the discovery cohort and GEO: GSE33118 as the validation cohort. In GSE54514, daily whole blood samples were detected for up to 5 days for sepsis non-survivors (n=9) and sepsis survivors (n=26) ⁽²²⁾. The blood samples of GSE33118 included sepsis non-survivors (n=10) and sepsis survivors (n=10) (Table 1, Supplementary Table 1 and 2). For GSE54514, first day blood samples were collected within the first 24 hours of ICU admission. For GSE33118, blood samples were taken within 12 hours of diagnosis.

Table 1.

Demographics of Discovery and Validation Cohort.

Dataset accession	First Author	Cohort Description	No. Survivors	No. Non- survivors
GSE54514 (Discovery Cohort)	Parnell	Whole blood transcriptome of survivors and non-survivors of sepsis	26	9
GSE33118 (Validation Cohort)	Wolfgang	Whole blood transcriptome of the septic shock	10	10

Sepsis was diagnosis as documented bacterial infection in addition to the presence of at least two of the following clinical criteria: (a) fever (temperature > 100.4 °F (38 °C) or hypothermia (temperature < 96.8 °F (36 °C)); (b) heart rate > 90 beats/min; (c) tachypnea (> 20 breaths/min or PaCO₂ < 4.3 kPa); (d) white blood cell count > 12,000 cells/μL, or < 4,000 cells/μL, or with > 10% band forms ⁽²²⁾.

Sepsis survival-related genes and S1PR1-related genes

The limma package (https://bioconductor.org/packages/release/bioc/html/limma.html) was used to identify differentially expressed genes (DEGs) between non-survivors and survivors in the discovery cohort and was referred to as sepsis survival-related genes. We conducted Pearson correlation test to identify all the S1PR1-coexpressed genes in the discovery cohort as our S1PR1-related genes in method 1. For method 2, S1PR1-related genes were identified from the STRING database (https://string-db.org) which was based on protein interactions and signaling pathways.

Expression score and risk score

We allocated expression (for the whole genome) and risk score (for sepsis survival-related genes) for each patient using a linear combination of expression values of genes in the signature. The formula corresponding to expression and risk score are:

$$\mathrm{expression}\mathrm{score}=\sum _{i=1}^n\left(\frac{e_i-{\mu }_i}{s_i}\right)$$

$$\mathrm{risk}\mathrm{score}=\sum _{i=1}^n{W}_i\left(\frac{e_i-{\mu }_i}{s_i}\right)$$

Here, n is the count of genes included in gene signature in each dataset, W_i represents the weight value of ith gene (see in Table 2), e_i represents the expression level of the ith gene, and μ_i and s_i are the corresponding mean and standard deviation value for the ith gene among whole samples.

Table 2.

Clinical Characteristics of sepsis patients in Discovery and Validation Cohort.

	GSE54514		GSE33118
	Non-Survivors	Survivors	Non-Survivors	Survivors
Age, mean ± SD, Year	69.67±11.01	56.69±18.27	65.6±15.16	64.65±16.65
Female sex (%)	5 (56%)	16 (61.5%)	4 (40%)	4 (40%)
Apache II score	24.22±5.43	18.54±6.18	NA	NA

^*APACHE: acute physiology and chronic health evaluation.

Statistical analyses

All the statistical calculations were performed by the R 3_.5.2. Principal component analysis (scatterplot3d package) and ROC curves (pROC package) were used in this study to verify the differentiating power of this SMS on sepsis survival status. False discovery rate (FDR) < 0.05 was considered as statistically significant.

Results

S1PR1-correlated genes significantly overlap with sepsis survival-related genes

Two GEO datasets (GSE54514 and GSE33118) with 55 sepsis patients were included in our study, and Table 2 showed the clinical characteristics of study population. We utilized two methods to identify the S1PR1-related genes (Figure 1). The first method is to define the genes that co-expressed with S1PR1 as S1PR1-related genes. In total, we found 557 genes (Pearson’s r > 0.4 and FDR < 0.05) that were co-expressed with S1PR1 in the discovery dataset. Both positive and negative co-expression genes were included in our study. To determine the DEGs in the whole blood transcriptome that correlated with the survival outcome of sepsis patients, we analyzed the gene expression data from 9 non-survivors and 26 survivors in the discovery cohort. 1078 up-regulated and 1134 down-regulated genes in sepsis (fold change [FC] > 1.2 and FDR < 5%) were identified as survival-related genes.

Figure 1. Schematic of workflow of discovery of gene signature.

We used the DAVID Bioinformatics Resources 6.8 (https://david.ncifcrf.gov/) ^{(23, 24)} to identify enriched Kyoto Encyclopedia of Genes and Genomes (KEGG) pathways among the S1PR1-related and sepsis survival-related genes. Our analyses revealed several immune-pathways were significantly associated with S1PR1 co-expressed genes (adjusted-P value < 0.05) The top KEGG pathways were significantly associated with several pathways such as B cell receptor signaling pathways, cytotoxicity, and cell adhesion molecules (CAMs) (Figure 2A). Furthermore, we noticed that several innate immune or infection pathways are associated with the DEGs between sepsis survivors and non-survivors (Figure 2B).

Figure 2. 62-gene signature in sepsis patients.

Enriched pathways among the (A) sepsis survival-related genes and (B) S1PR1-related genes. (C) Venn plot represents sepsis survival and S1PR1-associated overlapping genes. (D) Enriched pathways among the 62 genes. (E) Heat map shows the 62 gene signature expression in the discovery cohort. Red represents increased gene expression, while blue means decreased gene expression; (F) Kaplan-Meier plot for the sepsis patients based on the expression level of the 62 genes.

Among these sepsis survival-related genes, 62 genes overlapped with the S1PR1-correlated genes (Table 3, Figure 2C, cumulative hypergeometric test: P < 0.05), confirming a pivotal role for S1PR1-signaling in sepsis pathogenesis. Here, several innate and adaptive immune-pathways were also found as enriched among the 62 genes (Figure 2D). However, these 62 proteins didn’t have strong protein interactions with each other (Supplementary Figure 1).

Table 3.

S1PR1-related gene signatures.

Gene Signature 1				Gene Signature 2
Gene Symbol	Weight	Gene Symbol	Weight	Gene Symbol	Weight
GPM6A	1	TBXAS1	−1	PIK3R1	1
CCDC69	1	NLRP3	−1	LAMP1	−1
SLAMF6	1	NAIP	−1	STAT3	−1
ADRB2	1	HERC1	−1	SELL	−1
IKZF1	1	PECAM1	−1	GNAI2	−1
STOM	1	CRTAP	−1	CXCR2	−1
UBL3	1	EIF4B	−1	PIK3CD	−1
SYT15	1	GAS7	−1	AKT1	−1
PLEKHA2	1	FAM53B	−1	FPR1	−1
AP1S2	1	PRKCB	−1	MAPK1	−1
CIAO1	1	LRRK2	−1	HCK	−1
PIP4K2A	1	SLC12A6	−1	LPAR2	−1
FYCO1	1	FNBP1	−1	LYN	−1
PIK3CG	1	SEC14L1	−1	C3AR1	−1
PIK3R1	1	ELMO1	−1	FPR2	−1
MIR22HG	1	FAM49A	−1	CXCL16	−1
PPP3CB	1	BIN2	−1
SESN1	1	DPYSL2	−1
ATP1B2	1	MYO5A	−1
GADD45B	1	IL17RA	−1
PREX1	1	KLF6	−1
DDX55	1	ARAP3	−1
TMCC3	1	AKNA	−1
NFATC3	1	CXCR2	−1
CCBE1	1	LCP2	−1
SH2D3C	−1	NAAA	−1
MAST3	−1	MYADM	−1
CD37	−1	ARHGAP15	−1
ITGAL	−1	KLF4	−1
MSRA	−1	ITM2B	−1
TOM1L2	−1	SLCO3A1	−1

In method 2, we utilized the STRING database (https://string-db.org/) to generate a list of all S1PR1-associated genes based on signaling pathways, experimental data, co-expression data, and public text collections ⁽²⁵⁾. 233 genes were found as having high confidence (> 0.7) interaction score with S1PR1 (Supplementary Table 3). Several immune system pathways (TNF signaling pathway, Toll-like receptor signaling pathway, and T, B cell receptor signaling pathway et al.) and vascular system pathways (EGFR tyrosine kinase inhibitor resistance, and VEGF signaling pathway) were enriched among the pathway-based S1PR1-related genes (Figure 5A). Most of the signaling pathways were consistent with the S1PR1-related pathways mentioned in the literature ⁽²⁶⁾.

Figure 5. 16-gene signature in sepsis patients.

(A) Enriched pathways among S1PR1-related genes. (B) Venn plot represents sepsis survival and S1PR1-associated overlapping genes. (C) Enriched pathways among the 16 genes. (D) PPI network for the 16-genes. (E) Heat map shows the 29 gene signature expression in the discovery cohort. Red represents increased gene expression, while blue means decreased gene expression. (F) Kaplan-Meier plot for the sepsis patients based on the expression level of the 16 genes.

16 genes overlapped with the sepsis survival-related genes and S1PR1-related genes in method 2 (Table 3, Figure 5B). KEGG analysis (Figure 5C) identifies several strongly enriched pathways, including Chemokine signaling pathway, B cell receptor signaling pathway, and VEGF signaling pathway et.al. Next, we wanted to visualize the interactions within the 16 genes with PPI (protein-protein interaction) network. To generate the PPI network, PPI data was acquired from the STRING database. The clustering analysis was also performed in the PPI network by MCODE module in Cytoscape 3.6.1 (Figure 5D). We obtained two clusters of proteins with more dense interactions than average. One group is related to Staphylococcus aureus infection, the other group is related to chemokine signaling pathway, T, B cell receptor signaling pathway, and VEGF signaling pathway et.al.

S1PR1 gene signature predicts clinical outcome in both discovery and validated cohorts

When arranged for hierarchical clustering, the expression heatmaps of the genes in both the 62-gene and 16-gene signatures were able to differentiate non-survivors from survivors (Figure 2E and 5E), which suggests the significant power of discriminating survival patients from non-survival patients. To compare the overall survival (OS) of patients with different gene signature expression, we grouped the patients based on the total gene expression of the 62 or 16 genes in the discovery cohort. Higher or lower than median value was used to differentiate high or low expression groups. The results (Figure 2F and 5F) showed that the low expression group for both two sets of genes had more inferior OS than the high expression group.

To make the prediction feasible with one unique “survival prognosis score”, a scoring system representing a linear combination of the 62-gene expression and 16-gene expression (Table 3) values with a weight value was constructed to allocate each patient with a score to measure the possibility of risk. A higher risk score represented a worse clinical outcome. Our results focused on both the discovery and validation cohort. The result is in line with our expectations; either the 62-gene or 16-gene risk scores from non-survivors were significantly higher than those of survivors in the discovery cohort, while more importantly, it was repeatedly observed in the independent validation cohort (Figure 3A and 6A). Therefore, both of our S1PR1 based gene signatures had the statistical power to predict clinical out come in sepsis. We named our two gene signatures as S1PR1-related molecular signatures (SMS).

Figure 3. The 62-gene signature-based sepsis risk score differentiates non-survivors from survivors in both the discovery and validation cohort.

(A) Violin plot of the risk scores in non-survivors and survivors. (B) ROC curves of the 62-gene signature in distinguishing non-survivors from survivors. (C) Predictive power of the 62-gene signature–based AUC values in the discovery and validation cohort compared with random gene signatures from the whole genome and sepsis survival-related genes.

Figure 6. The 16-gene signature-based sepsis risk score differentiates non-survivors from survivors in both the discovery and validation cohort.

(A) Violin plot of the risk scores in non-survivors and survivors. (B) ROC curves of the 16-gene signature in distinguishing non-survivors from survivors (C) Predictive power of the 16-gene signature–based AUC values in the discovery and validation cohort compared with random gene signature from the whole genome and sepsis survival-related genes.

Classification power of the gene signature

The receiver operating characteristic (ROC) curve was utilized to visualize how well the gene signatures can distinguish between non-survivors and survivors. For the 62-gene signature, the area under the curve (AUC) for the ROC curve were 1 and 0.86 for the discovery and validation cohort, respectively (Figure 3B). For the 16-gene signature, the AUC values were 0.98 and 0.83 for the discovery and validation cohort, respectively (Figure 6B). We investigated the classification performance of the SMS signature in the discovery and validation datasets. Principal component analyses (Figure 4 and 7) showed that the SMS thoroughly distinguished non-survival patients from survival patients in the discovery cohort, while exhibiting minimal overlaps in the validation cohort.

Figure 4.

3d PCA plot of the 62-gene signature for the discovery cohort (A) and validation cohort (B).

Figure 7.

3d PCA plot of the 16-gene signature for the discovery cohort (A) and validation cohort (B).

A bioinformatics study by David Venet et al. shows that most gene signatures randomly selected from the human genome with the same genes size sometimes were better than published gene signatures ⁽²⁷⁾. To confirm whether this issue existed in our study, we further applied resampling tests a total of 10,000 permeations to generate the expression score or risk score for each patient to generate the corresponding AUC value for each random gene signature. Each independent permeation randomly selected 62 or 16 genes of random sizes from the whole genome. It is of interest that the SMS gene signature had higher power on the classification of sepsis survival than randomly generated genes with the same gene count (better than 95% percent random gene signatures in the whole genome) (Figure 3C and 6C). This quality control data confirmed the significance of the prognostic power of SMS while demonstrating it was not population specific.

We next compared the performance of gene signature 1 and 2 (Table 4). The sepsis risk prediction accuracy is almost the same according to the P and AUC value. However, signature 2 has better internal protein interactions (PPI P value) than gene signature 1.

Table 4.

Statistical comparison between Gene Signature 1 and 2.

	P value		AUC Value
	Discovery	Validation	Discovery	Validation
Gene Signature 1	1.1.e-05	0.0073	1	0.86
Gene Signature 2	2.20e-05	0.014	0.98	0.83

Discussion

Sepsis is the primary cause of admission in the intensive care unit (ICU) ⁽³⁾. The pathology of sepsis and its associated systemic inflammatory response syndrome, which can lead to multi-organ dysfunction and even death, is still poorly understood. Lipoproteins, such as S1PR1, are increasingly identified as essential mediators influencing the progression and disease outcome of sepsis ^{(28, 29)}. Serum S1P levels are radically decreased and are inversely related to disease severity in sepsis ^{(29, 30)}. S1P binds to S1PR1 in immune or endothelial cells, leading to release of cytokines ⁽³¹⁾ and regulation of vascular integrity ⁽²⁰⁾, suggesting S1PR1 could be used as a therapeutic molecule in sepsis. However, S1PR1-related genes which also can be used for predicting survival in sepsis are still missing. Linking the gap between in vitro or in vivo model results and clinical diseases is necessary for medical researchers. Identifying gene expression signatures can disclose a variety of clinical and biological characteristics of patients’ samples. In this study, we analyze GEO datasets containing whole-genome gene expression data from both non-survival and survival patients with sepsis. Our results have several contributions as follow: 1. Confirmed the correlation of S1PR1 dependent signaling in sepsis survival with bioinformatics tools; 2. SMS is an “independent” prognostic marker of sepsis survival; 3. Potential to move forward to a second-generation biomarker for sepsis prognosis and therapy decisions, thus fulfilling the need for “precision medicine.”

Biomarkers for sepsis can be used for diagnosis, risk assessment, and survival prediction. 178 biomarkers have been verified for evaluating of sepsis ⁽³²⁾. However, no single biomarker such as C-reactive protein ⁽³³⁾ and Procalcitonin ⁽³⁴⁾ have adequate sensitivity or specificity for diagnosis and prognosis to be regularly used in clinical practice to date. Compared to the single-gene biomarker, multi-gene biomarkers and corresponding sepsis risk scoring methods showed higher AUC values in ROC curves ⁽³⁵⁾. More and more researches begin to focus on using Genome-wide expression analysis to better predict the clinical outcome of sepsis. Genome-wide expression analysis offers of examining the entire transcriptome of a tissue, and assessing gene expression changes without any bias. Our multi-gene biomarkers have been derived from SP1R1 signaling networks showed high prediction power, which was shown in both a discovery and validation cohort, has the potential to be extended to clinical trial in the future.

Compared to other sepsis biomarkers, our gene signatures have several advantages: (1) the two gene signatures, especially the 16 gene signature, have a strong relationship with signaling pathways and protein interactions. Our S1PR1 gene signatures both reflect the status of the immune and vascular systems during infections which was regulated by S1PR1, so this explains why our sets of genes are the pivotal factor for predicting the risk of sepsis patients; (2) the resampling tests which compared random gene signatures with our gene signatures has confirmed the significance of the prognostic power of our gene signatures; (3) we only included the sepsis samples from whole blood sample and microarray datasets, which made our results more consistent and comparable. Whole blood samples and microarray datasets meet our needs as a rapid and dependable prognosis for sepsis patients.

To generate validated and accurate bioinformatic information, we performed our studies with multiple layers of quality control. Firstly, one independent validation cohort was used. Secondly, the AUC values from ROC curves and PCA plots showed that these S1PR1-derived SMS are powerful tools to provide an essential prognostic method to distinguish sepsis patients with high risk from sepsis patients with low risk. Lastly, gene signatures selected may not have a better outcome predictor than random signatures from whole genome or sepsis survival-related genes ⁽²⁷⁾. So, we used resampling tests to reveal whether the 62-gene or 16-gene signature has more power prediction than random gene signatures. Our results show that our gene signatures have superior predictive power than that of most gene sets randomly chosen from whole genome or sepsis survival-related genes.

In this study, we showed that S1PR1-related gene signatures are capable of distinguishing patients with higher risk from other sepsis patients. However, our work in S1PR1-related gene signatures was only based on bioinformatics methods. We only included sepsis datasets with whole blood samples in this study, so the potential power of our gene signature needs to be verified in additional datasets which contained peripheral blood mononuclear cell (PBMC) samples or studies from larger multicenter. Additionally, we can test our gene signature in the devastating complication of sepsis such as acute respiratory distress syndrome (ARDS) which have similar underlying mechanisms with sepsis.

Conclusions

In conclusion, we obtained gene signatures containing 62 and 16 protein-coding genes, which we demonstrated to be reproducible predictors of clinical outcome in patients with sepsis. Thus, our results could have potential value in clinical evaluations and disease monitoring in patients with sepsis.

Supplementary Material

Supplemental Figure 1

Supplementary Figure 1. PPI network for the 62-genes.

Abbreviations:

AUC

area under the curve

DEGs

Differentially expressed genes

GEO

Gene Expression Omnibus

OS

Overall survival

PPI

Protein-protein interaction

ROC

receiver operating characteristic

S1PR1

Sphingosine-1-phosphate and its receptor S1P receptor 1

SMS

S1PR1-related molecular signatures