Identification of a 3-gene signature predicting 28-day mortality for sepsis

Abstract

Sepsis often leads to unpredictable consequences. The prognosis of sepsis has not been largely improved. We tried to construct a prognostic gene model related to the 28-day mortality of sepsis to identify the risk of mortality and improve the outcome early. We identified the modules associated with 28-day mortality by weighted gene co-expression network analysis from the microarray data of GSE65682. Protein–protein interaction network analysis and univariate Cox regression were conducted to identify hub genes for constructing a prognostic model. Finally, the model was evaluated for robustness. The correlation between the model and immune cells was investigated. The cyan module has a significant negative relationship with 28-day mortality. A risk model was developed to predict prognosis, utilizing macrophage expressed gene 1, CX3C chemokine receptor 1, and human leukocyte antigen-DRB1. The model’s expression was found to be higher in the group with lower risk, while the group with higher risk had a higher 28-day mortality rate. These findings were validated using both the test and whole sets. Three genes were positively associated with monocyte expression. We constructed a septic prognostic model with 3 genes, including macrophage expressed gene 1, CX3C chemokine receptor 1, and human leukocyte antigen-DRB1. The expression of them had a significant negative relationship with the 28-day mortality and may influenced monocyte function.

1. Introduction

Sepsis is a potentially life-threatening medical condition characterized by a dysregulated host response to an infection.^[1] Sepsis is a leading cause of mortality in intensive-care units (ICU).^[2] The specific molecular mechanisms of sepsis remain unclear. Early diagnosis and prompt initiation of appropriate treatment are crucial for managing sepsis.^[3] Sepsis patients are heterogeneous, making it difficult to identify those at high risk.^[4] Although several biomarkers have been proposed for early detection and prognostic prediction, their exact value remains unclear.^[4] A lack of effective biomarkers hampers early diagnosis and prognostic prediction in sepsis. Studying sepsis pathogenesis and finding new biomarkers is crucial for improving treatments and patient outcomes.

The technology of bioinformatics analysis has provided tools to study the molecular mechanism of diseases.^[5,6] This makes it possible to discover novel biomarkers and study the pathogenesis of sepsis at the gene level to improve the outcomes of sepsis patients.^[7,8] This study attempts to construct a prognosis-related gene model of sepsis related to the 28-day mortality. Herein, we obtained a large sample of gene expression data with survival information from the Gene Expression Omnibus (GEO) database. We utilized weighted gene co-expression network analysis (WGCNA) to identify multiple modules associated with 28-day mortality of sepsis. To identify the most important genes within these modules, we constructed a protein–protein interaction (PPI) network and identified the hub genes that are highly correlated with 28-day mortality. Univariate Cox regression and Lasso regression analyses were performed to construct and validate a prognostic model for sepsis and identify the hub genes associated with 28-day mortality. Figure 1 outlines the workflow of our study.

Figure 1.

The workflow of this study.

2. Materials and methods

2.1. Microarray data

We obtained microarray data and clinical information from the GEO dataset GSE65682 at http://www.ncbi.nlm.nih.gov/gds. GSE65682 is an expression profiling analysis based on the GPL13367 platform and includes 802 whole blood leukocyte transcriptome data from septic patients and healthy subjects. Sepsis was diagnosed according to the International Sepsis Forum Consensus Conference Definition. Sepsis patients older than 18 were included in the study. Blood samples for each patient were obtained within the first 24 hours of ICU admission. The present study is concerned with the 28-day mortality rate of sepsis. Consequently, the data of healthy subjects and patients lacking information on the 28-day mortality rate in this dataset were excluded. A total of 323 samples were excluded, including 281 samples with missing information about survival times and 42 samples of healthy subjects. The data arrays of 479 remaining septic patients were collected for this study, of which 107 died within 28 days of ICU admission.

2.2. Identifying key co-expressed modules using WGCNA

We used the WGCNA package (Department of Human Genetics and Department of Biostatistics, University of California, Los Angeles) in R to create gene co-expression networks from gene expression data profiles. To construct a scale-free network, we selected soft powers β using the pickSoftThreshold function. The topological overlap matrix was constructed based on the adjacency matrix. The topological overlap matrix underwent hierarchical clustering to identify gene modules represented by different colors. This categorization was based on the weighted correlation coefficient of gene expression patterns, dividing thousands of genes into multiple modules. The key modules were identified as the most significantly associated with 28-day mortality in co-expression.

2.3. Further screening of candidate genes by PPI construction

Potential hub genes with an interaction score ≥ 0.4 were chosen from the major co-expression modules using the STRING database (https://string-db.org/), allowing Cytoscape (Institute for Systems Biology, Seattle) (version 3.8.2) to see the network. Using the MCODE plugin of Cytoscape (degree cutoff = 2, max. depth = 100, k-core = 2, and node score cutoff = 0.2), we further chose putative hub genes to test them.

2.4. Development of the prognostic gene model for risk assessment

We used univariate Cox regression analysis to explore genes associated with the prognosis of sepsis. The putative prognostic genes were chosen based on the .05 P-value cutoff. The 479 samples were randomly split into 2 groups, the training set (n = 240) and the test set (n = 239), in a 5:5 ratio, to create a gene prognosis model. There was no discernible difference in the clinical characteristics of the 2 groups (Table 1). The Lasso regression analysis was applied to the training set. The 10-fold cross-validation technique was employed to evaluate the performance of the classification model. Using the regression coefficient of each gene, the risk score for each person was determined using the following formula:

Table 1.

The clinical features of the included patients of sepsis.

Characteristic	Whole set (n = 479)	Training set (n = 240)	Test set (n = 239)
Age (yr)	61.0 ± 14.8	60.8 ± 15.1	61.1 ± 14.5
Sex, n (%)
Female	206 (43)	102 (42.5)	104 (43.5)
Male	273 (57)	138 (57.5)	135 (56.5)
Diabetes, n (%)	84 (17.5)	42 (17.5)	42 (17.6)
The site of infection, n (%)
Pulmonary	188 (39.2)	93 (38.7)	95 (39.7)
Abdominal	49 (10.2)	21 (8.8)	28 (11.7)
Other	242 (50.6)	126 (52.5)	116 (48.6)
Follow up days (d)	23.2 ± 9.2	22.5 ± 9.9	23.9 ± 8.4
28-day mortality, n (%)	114 (23.8)	62 (25.8)	52 (21.8)

$${\displaystyle \begin{array}{l}{\displaystyle Riskscore={\displaystyle \sum _{k=1}^n}coefficient(genek)\ast Expk}\end{array}}$$

“n” denotes the number of selected prognostic genes in the formula above, “genek” denotes the kth selected gene, “coefficient” displays the estimated regression coefficient of the genes from the multivariate Cox regression analysis, and “Expk” indicates the kth selected gene’s expression value. Based on the median risk score, the sepsis training set from the chosen database was split into high-risk and low-risk groups. The correlation between the risk ratings and the potential genes is shown visually in the heatmap. The risk score model’s predictive power was assessed using receiver operating characteristic (ROC) curve analysis and Kaplan–Meier (K–M) survival analysis. To evaluate the LASSO regression model’s dependability and predictive power, internal validation was conducted. Subsequently, following the stratification of subjects according to the source of infection, the ROC curve analysis was employed once more to assess the predictive capability of the risk scoring model.

2.5. Immune infiltration using CIBERSORT analysis

CIBERSORT is a tool that analyzes gene expression data to calculate the proportion of 22 human immune cell subsets. It uses the LM22 gene file, which can be downloaded from the CIBERSORT web portal at http://CIBERSORT.stanford.edu/. We used LM22 to determine immune cell proportions in sepsis patients. Data with CIBERSORT P-value < .05 was visualized using ggplot2, corplot, and vioplot R packages.

2.6. Statistical analysis

R version 3.6.3 was used for statistical analyses. Categorical variables were compared using Chi-square or Fisher exact test. Survival statistics were analyzed using the K–M curve and the log-rank test. Cox regression analysis was performed to assess the association of genes with septic events. To evaluate the diagnostic value of the risk model, the ROC curve was used. The correlation between variables was calculated using Spearman rank correlation test. Statistical significance was determined by P-values < .05.

2.7. Compliance with ethical guidelines

Ethical approval was not necessary because this study was based on public databases.

3. Results

3.1. Identifying key co-expressed modules using WGCNA

Gene co-expression networks were constructed using 479 gene expression profiles. Hierarchical clustering analyses were performed, and obvious outlier samples were recognized and removed before WGCNA (Figure S1A, Supplemental Digital Content, https://links.lww.com/MD/P750). We chose soft powers β = 7 for the soft threshold parameter (Fig. 2A) to ensure a scale-free network. Based on the average linkage hierarchical clustering, 17 modules were identified (Figure S1B, Supplemental Digital Content, https://links.lww.com/MD/P750). Figure 2C shows the results of the module–trait relationships. Module eigengenes of the pink, salmon, cyan, black, lightcyan, magenta, and lightcyan modules were found to have the highest correlation with the subtypes (age, gender, 28-day mortality, survival time, abdominal, pulmonary, and diabetes, respectively). Given the exploratory nature of this study, we focused on the most robust module. The cyan module (r = -0.16, P = 3e-04) has a significant negative relationship with the 28-day mortality of sepsis patients. A total of 82 genes from the cyan module were used for PPI network analysis.

Figure 2.

Identification of modules associated with clinical information in the selected dataset from GSE131761. Dataset and the PPI network construction of the genes from the cyan module. (A) The scale-free fit index (left) and mean connectivity (right) for various soft-thresholding powers. (B) Heatmap of the 17 co-expression module–trait correlation. (C) PPI network of the 82 genes from cyan module visualization was performed with Cytoscape software. (D) The subnetworks of the PPI networks was identified using MCODE plugin in the Cytoscape. PPI = protein–protein interaction.

3.2. Further screening of candidate genes by PPI construction

An 82-gene PPI network was constructed using the STRING database. The resulting network had 81 nodes and 135 edges, with a PPI-enrichment P-value of 1 × 10^-16. It was visualized using Cytoscape software (Fig. 2B). The MCODE plugin in Cytoscape identified 14 hub genes and critical subnetworks from the PPI network nodes (Fig. 2D). Table 2 lists the functions of these hub genes.

Table 2.

Functional roles of the 14 hub genes.

No.	Gene symbol	Function
1	CX3CR1	Receptor for the CX3C chemokine fractalkine (CX3CL1); binds to CX3CL1 and mediates both its adhesive and migratory functions.
2	IL10RA	Receptor for IL-10; binds IL-10 with a high affinity.
3	RNASE6	Has a role in host defense; ribonuclease A family
4	HLA-DMA	Plays a key role in catalyzing the release of class II-associated invariant chain peptide (CLIP) from newly synthesized MHC class II molecules and releasing the peptide binding site to acquire antigenic peptides.
5	HLA-DMB	Plays an important role in catalyzing the release of CLIP from newly synthesized MHC class II molecules and releasing the peptide binding site to acquire antigenic peptides.
6	HLA-DPA1	Binds peptides derived from antigens that access the endocytic route of antigen presenting cells (APC) and presents them on the cell surface for recognition by the CD4 T cells.
7	FCGR3A	Receptor for the Fc region of IgG. Binds complexed or aggregated IgG and also monomeric IgG. Mediates antibody-dependent cellular cytotoxicity (ADCC) and other antibody-dependent responses.
8	CLEC10A	Regulate adaptive and innate immune responses. Binds in a calcium-dependent manner to terminal galactose and N-acetylgalactosamine units, linked to serine or threonine.
9	MPEG1	As a protein coding gene, it plays a pivotal role in the innate immune response after bacterial infection by inserting into the bacterial surface to form pores.
10	HLA-DRB1	Binds peptides derived from antigens that access the endocytic route of APC and presents them on the cell surface for recognition by the CD4 T cells.
11	CCR2	Receptor for the CCL2, CCL7 and CCL13 chemokines. Receptor for the beta-defensin DEFB106A/DEFB106B.
12	CD86	Receptor engaged in the costimulatory signal essential for T-lymphocyte proliferation and interleukin-2 production.
13	LY86	Cooperate with CD180 and TLR4 to mediate the innate immune response to bacterial lipopolysaccharide (LPS) and cytokine production.
14	CD68	Play a role in phagocytic activities of tissue macrophages, both in intracellular lysosomal metabolism and extracellular cell–cell and cell–pathogen interactions.

CX3CR1 = CX3C chemokine receptor 1, HLA-DR = human leukocyte antigen-DR.

3.3. Identification of a prognostic risk model in the training set

To assess the gene significance in sepsis, we analyzed the expression of 14 hub genes using univariate Cox regression in the training set. Our analysis revealed that 6 hub genes, namely macrophage expressed gene 1 (MPEG1), CX3C chemokine receptor 1 (CX3CR1), human leukocyte antigen-DR (HLA-DR)B1, HLA-DPA1, HLA-DMA, and HLA-DMB, were significantly related to survival time, as depicted in the forest plot (Fig. 3A). Furthermore, the Spearman correlation test showed a positive correlation between the expression of these genes (Fig. 3B). A LASSO regression was used to construct a 3-gene model, which included MPEG1, CX3CR1, and HLA-DRB1, from the candidate genes (Fig. 4A and B). The expression of the 3 genes was significantly lower in high risk individuals than in low-risk individuals, as shown in Figure 4C. The formulas used to calculate the prognostic risk score were as follows: k score = (-0.119 × MPEG1) + (-0.230 × CX3CR1) + (-0.046 × HLA-DRB1). In the training set, patients were divided into 2 groups based on their risk score. The median risk score (-2.530256) was used as the threshold to identify patients with high risk and low risk of mortality. Furthermore, our results showed that there is a strong positive correlation between the value of the risk score and the expression of 3 specific genes, as depicted in Figure 4D. On survival analysis, the 28-day survival of low-risk group patients was better than the high-risk group (P < .001, Fig. 5A). The 28-day survival area under the ROC curve (AUC) of the prognostic model was 0.66 (Fig. 5D). Subsequent stratification by the source of infection yielded 28-day survival AUC values of 0.64 and 0.8 for the pulmonary and abdominal infection subgroups, respectively, validating the model’s capacity for survival prediction (Figures S2A and S2D, Supplemental Digital Content, https://links.lww.com/MD/P750). Figure 6A shows the risk plot of high- and low-risk score groups, patient survival status, and risk gene expression data.

Figure 3.

Univariate regression analysis of 14 candidate genes to screen hub genes for prognosis. (A) Forest plot of the univariate regression analysis for 14 candidate genes, 6 genes were significantly related to the 28-days mortality, including MPEG1, CX3CR1, HLA-DRB1, HLA-DPA1, HLA-DMA, and HLA-DMB. (B) Spearman correlation analysis between the expression of the 6 genes. Positive correlation was marked with red and negative correlation with blue. (***P < .001). CX3CR1 = CX3C chemokine receptor 1, HLA-DR = human leukocyte antigen-DR, MPEG1 = macrophage expressed gene 1.

Figure 4.

LASSO regression construction the 3-genes signature prediction model. (A) LASSO coefficient profiles of the 6 genes. (B) Partial likelihood deviance was plotted versus log (λ), the dotted vertical lines were drawn at the optimal values using the minimum criteria and the 1-SE criteria. (C) The expressions of the 3 genes of the prediction model in the low and high risk groups. (***P < .001). (D) The correlation between risk score value and the expressions of the 3 genes. Red indicates positive correlation, green negative correlation.

Figure 5.

Survival analysis and prognostic performance of the 28-day mortality risk score model in sepsis. (A) Kaplan–Meier survival analysis of high risk and low-risk groups in the training set. (B) Time-dependent ROC curves for 28-days overall survival (OS) rate in the training set. (C, E) Kaplan–Meier survival analysis of high risk and low-risk groups in the test set and whole set, respectively. (D, F) Time-dependent ROC curves for 28-days OS in the test set and whole set, respectively. ROC = receiver operating characteristic.

Figure 6.

Prognosis and expression of risk genes of the high risk and low-risk sepsis patients. (A) Risk plot distribution, survival status, and expression of the risk genes in the training set. (B) Risk plot distribution, survival status (all patients were followed up for 28 days), and expression of the risk genes in the test set. (C) Risk plot distribution, survival status, and expression of the risk genes in the whole set.

3.4. Prognostic risk model verification

The prognostic risk model was tested using the test data to ensure its stability and reliability. The sepsis database test set was then classified into 2 categories: high risk (n = 119) and low risk (n = 120). The K–M survival curve demonstrates that patients with a high risk of mortality had poorer 28-day survival than those with a low risk (P < .001, Fig. 5B). The AUC of the test set was 0.65 (Fig. 5E). Specifically, the AUC values for the pulmonary and abdominal infection subgroups were 0.69 and 0.62, respectively, suggesting a robust capacity for survival prediction (Figures S2B and S2E, Supplemental Digital Content, https://links.lww.com/MD/P750). Figure 6B shows the risk distribution, patient survival status, and gene expression data of the risk genes in the test set. We then verified the model’s stability and reliability using all available data. The study used the same risk model for both groups, dividing all data into high-risk (n = 239) and low-risk (n = 240) categories. The 28-day survival rate of sepsis patients could be distinguished well between the low-risk and high-risk groups (P < .001, as shown in Fig. 5C). The AUC was 0.66 (as shown in Fig. 5F), and the AUC of the pulmonary infection and abdominal infection subgroups was 0.67 and 0.68, respectively (Figures S2C and S2F, Supplemental Digital Content, https://links.lww.com/MD/P750). Figure 6C displays the corresponding risk distribution, patient survival status, and gene expression data for all genes in GSE65682.

3.5. Establishment and evaluation of the multivariate Cox regression and nomogram

The GSE65682 dataset provides limited basic clinical information about patients’ immune status or inflammation, including age, gender, diabetes status, and the site of infection (abdominal, pulmonary, or other). These factors may affect the prognosis of septic patients. We integrated 3 gene-based risk scores with clinical features to construct a nomogram for predicting septic patient mortality. Of these, the risk score had the greatest impact on 28-day mortality in the training set, test set, and whole set [hazard ratio (HR) = 5.74, 95% confidence interval (CI): 2.54–13.0, P < .001; HR = 6.11, 95% CI: 2.59–14.4, P < .001; HR = 5.83, 95% CI: 3.26–10.39, P < .001, respectively]. The multivariate Cox regression analysis reveals that the risk score signature is a dependable and autonomous prognostic biomarker that can be used to assess the 28-day mortality of sepsis patients (see Fig. 7A–D).

Figure 7.

Multivariate Cox regression analysis in sepsis patient 28-days OS. (A)Forest plot of the multivariate Cox regression analysis in test set, the risk score signature showed significance in the multivariate Cox regression. (B) Forest plot of the multivariate Cox regression analysis in training set, the risk score signature showed significance. (C) Forest plot of the multivariate Cox regression analysis in whole set, the 28-days OS of sepsis was significantly correlated with the risk score, age and the infection in abdominal. (D) Nomogram predicting 28-days OS for sepsis patients in whole set. (*P < .05; ***P < .001).

3.6. Immune infiltration by CIBERSORT analysis

We analyzed the proportion of immune cells in whole blood from 239 low-risk and 240 high-risk samples using the selected GEO microarray data. The top 5 highest infiltrating proportions in whole blood were neutrophils, monocytes, T cells CD4 naive, T cells gamma delta, and macrophages M0 (Fig. 8C). The proportion of neutrophils, monocytes, activated NK cells, activated dendritic cells (DCs) and B-cell memory was higher in the low-risk group compared to the high-risk group (P-value < .05). The study conducted a correlation analysis of immune cells. The results showed that DCs activated and T cells follicular helper had the most synergistic effect, while neutrophils and monocytes had the most competitive effect (Fig. 8A). Monocytes, CD8 T cells, and activated NK cells were positively correlated with gene expression (Fig. 8B).

Figure 8.

Immune infiltration analyses for sepsis. (A) Correlation heatmap of 22 immune cell types. (B) The violin plot of immune cells, red for high risk group and blue for low-risk group. (B) The correlation between 6 hub genes and 22 immune cells. Red, positive; blue, negative. (*P < .05; **P < .01).

4. Discussion

In recent decades, the incidence and mortality of sepsis have remained high despite improvements in its diagnosis and treatment.^[9,10] To improve clinical outcomes, it is urgent to identify the underlying mechanism of sepsis and develop more sensitive and specific diagnostic and prognostic signatures.^[11] To identify effective prognostic indicators of sepsis, numerous studies have been conducted. These studies have included important clinical markers of sepsis such as Sequential Organ Failure Assessment (SOFA), procalcitonin, lactate, and systemic inflammatory response syndrome.^[12,13] However, the identification of an ideal marker may be a challenge. For example, the SOFA criteria were used as a cornerstone of the Sepsis-3 definition. Numerous studies have attempted to explore the relationship between SOFA and the prognosis of sepsis.^[14] However, the results have been conflicting. A systematic review and meta-regression analysis revealed no significant correlation between SOFA and mortality in sepsis.^[14] There is currently no single sensitive and specific biomarker for diagnosing sepsis. However, the use of multi-biomarkers has shown potential in improving the accuracy of sepsis prognosis prediction.^[15,16] The purpose of this study was to identify hub genes associated with sepsis-related 28-day mortality using bioinformatic analysis. Initially, WGCNA was used to screen 82 genes that showed an inverse correlation with 28-day mortality. The PPI network analysis reduced the number of potential target genes from 82 to 14. Then, the 14 genes were subjected to univariate Cox regression analysis and LASSO regression analysis. As a result, a prognostic evaluation model for 28-day mortality of sepsis consisting of 3 genes (MPEG1, CX3CR1, and HLA-DRB1) was successfully constructed. The internal validation demonstrates that the models exhibit good robustness. To enhance comprehension and comparison of the models’ performance, a potential clinical research direction for future steps would be to test the model’s robustness in septic patients with varying comorbidities and disease severities.

The study found that the low mortality risk group had significantly higher expression of the 3 genes than the high mortality risk group, suggesting a negative association between the 3-gene prognostic model and the 28-day mortality of sepsis. We used the CIBERSORT algorithm to examine the connection between immune cells and gene expression profiles. It has been demonstrated that disparities in model gene expression may exert an influence on the differentiation process of leukocytes. Furthermore, alterations in model gene expression are associated with the differentiation of leukocytes into subtype cell populations. Our analysis revealed a difference in the infiltration of immune cells between the low-risk and high-risk groups. The low-risk group had a higher proportion of neutrophils, monocytes, activated NK cells, activated dendritic cells, and memory B cells compared to the high-risk group. The expression of 3 genes, MPEG1, CX3CR1, and HLA-DRB1, was significantly and positively associated with monocytes. These 3 genes seem significantly impact monocytic functions, which may play a critical role in the sepsis immune response. Monocytes play a crucial role in the pathogenesis of sepsis as one of the main circulating cells of the innate immune system.^[17–19] Previous studies have investigated the correlations between MPEG1, CX3CR1, HLA-DRB1, and monocytes in the context of infection or sepsis.

MPEG1 expression was discovered in leukocytes and macrophages.^[20] Our research suggests that the downregulation of MPEG1 expression is associated with a high risk of mortality in sepsis patients. MPEG1, as a key effector molecule of innate immunity, is regulated by inflammatory signals such as lipopolysaccharide (LPS), TNF-α, and IFN-γ.^[21,22] It accelerates the activation of the JAK-STAT pathway and induces the expression of IFN-stimulated gene factors (ISGs), such as ISG15 and Mx1 through the NF-κB pathway and glycosylation modifications at extracellular domain sites (such as N185, N375) that pre-bind to type I interferon receptor (IFNAR1/2), amplifying antibacterial and antiviral effects.^[23] The dual regulatory role of MPEG1 between pro-inflammatory and homeostasis suggests that its excessive activation may exacerbate inflammatory damage in endotoxin shock mice, while deficient mice show resistance to LPS-induced endotoxin shock.^[23] In antimicrobial infection, MPEG1 directly disrupts the membrane structure of bacteria (such as MRSA, Mycobacterium tuberculosis, Enteropathogenic Escherichia coli) by forming 100Å transmembrane pores through the membrane attack complex/perforin domain, and enhances bactericidal efficiency in conjunction with reactive oxygen species/reactive nitrogen species and lysosomal enzymes, with its absence leading to uncontrolled bacterial spread in mice.^[20] In antiviral responses, it inhibits the replication of viruses such as vesicular stomatitis virus expressing green fluorescent protein by enhancing type I IFN signaling, while phosphorylation of STAT1/2 and ISG activation are impaired in deficient cells.^[23] Pathogens can inhibit NEDD8 modification, block MPEG1 transport, or modify membrane lipids to escape by secreting CIF proteins.^[24] The host responds by upregulating MPEG1 through membrane-anchored MPEG1a and secreted MPEG1b subtypes, as well as M1 polarized macrophages in the placental decidua.^[25] In summary, MPEG1 defends against infections through dual mechanisms of pore formation and IFN signaling regulation, indicating that MPEG1 may play an important role in sepsis by modulating immunity and participating in the defense against pathogenic microorganisms.

CX3CR1 is a transmembrane-spanning G protein-coupled receptor that is primarily expressed in monocytes, NK cells, and lymphocytes.^[26,27] CX3CR1 mediates the recruitment and activation of monocytes.^[26] In sepsis and inflammatory responses, the expression changes of CX3CR1 present a dynamic and complex pattern, playing a key role in the progression of sepsis. During sepsis, the mRNA and protein expression of CX3CR1 on the surface of monocytes is significantly downregulated, with a more pronounced and sustained downregulation in non-surviving patients.^[28,29] Consistent with previous research, our study found that CX3CR1 had a protective effect on the mortality risk of sepsis. This downregulation can be induced in vitro by LPS, whole bacteria (such as Escherichia coli and Staphylococcus aureus), and corticosteroids, while IL-10 and soluble CX3CL1 (sCX3CL1) do not have this effect.^[30] An increase in the concentration of sCX3CL1 in the serum of sepsis patients suggests that membrane-bound CX3CL1 on endothelial cells is cleaved to increase its soluble form.^[30] CX3CR1 mediates the adhesion and migration of inflammatory monocytes (such as Ly6Chigh monocytes) to sites like renal vasculature by binding to CX3CL1, exerting a tissue-protective effect, and its absence exacerbates renal injury and mouse mortality.^[31] CX3CR1 activation can regulate the adhesion capacity of monocytes, influence the production of anti-inflammatory cytokines such as IL-1ra, and inhibit negative regulatory factors associated with the NF-κB pathway (such as suppressors of cytokine signaling 1 and A20-binding inhibitor of NFkappa-B activation 3) that may be involved in its expression regulation.^[32] Additionally, the I249 allele of human CX3CR1 is associated with a reduced incidence of acute kidney injury in sepsis patients, indicating that enhanced adhesion function may have a protective effect.^[31] Therefore, the downregulation of CX3CR1 expression exacerbates immune suppression and organ damage in sepsis by weakening monocyte function.

HLA-DR is an HLA class II antigen closely associated with antigen presentation.^[33] HLA-DR are transmembrane glycoproteins present in monocytes, DCs, and B cells. The changes in the expression of HLA-DRB1 during sepsis and inflammatory responses are closely related to its core role in immune regulation. The expression of HLA-DRB1 molecules on the surface of monocytes in sepsis patients is significantly downregulated.^[34,35] This downregulation involves not only transcriptional suppression (such as reduced activity of the MHC-II transactivator class II transcription factor and enhanced binding of the CCCTC-binding factor in the MHC-II region, which inhibits HLA-DRB1 promoter activity) but is also associated with post-translational modifications and endocytosis.^[34] Genetic polymorphism studies have shown that alleles such as HLA-DRB1*04:01, 07:01, and 08 are significantly more frequent in sepsis patients, possibly exacerbating immune paralysis and tissue damage by affecting antigen presentation efficiency and the release of pro-inflammatory cytokines (such as IL-6 and MIF).^[36] The downregulation of HLA-DRB1 leads to a defect in the antigen presentation function of monocytes, rendering them unable to effectively activate CD4⁺ T cells, while also causing abnormalities in the TLR signaling pathway, which dulls the monocyte response to secondary infection stimuli.^[34] Furthermore, during sepsis, HLA-DRB1 interacts with the network of inflammatory factors: anti-inflammatory factors such as IL-10 can induce the endocytosis of HLA-DRB1, while the absence of HLA-DRB1 further weakens the ability of monocytes to clear pathogens, creating a vicious cycle.^[35] In the current study, HLA-DRB1 expression was also significantly lower in patients dying of sepsis. Notably, HLADRB1 alleles are associated with complications of sepsis (such as acute kidney injury), and patients carrying specific alleles (like 04:01) are more likely to experience organ dysfunction, potentially related to immune damage mediated by HLA-DRB1 in renal vascular endothelial cells.^[37] HLA-DRB1 participates in the immune dysregulation of sepsis through dual mechanisms of genetic variation and epigenetic regulation.^[34] Its abnormal expression is not only a marker of immune paralysis but also plays a key role in the progression of sepsis by affecting antigen presentation, cytokine networks, and organ damage.^[34] Targeting HLA-DRB1 and its regulatory pathways may provide new strategies for personalized treatment of sepsis.

To our knowledge, the combination of MPEG1, CX3CR1, and HLA-DRB1 has not been used before to predict 28-day mortality in sepsis patients. The combination of markers could potentially overcome the instability of a single marker. However, it is important to note that this research has certain limitations. Firstly, the absence of pertinent clinical data, including but not limited to comorbidities, treatment regimens, APACHE II score, SOFA score, immunosuppression, organ failure, pathogen type, procalcitonin, and lactic acid, in the dataset precluded our ability to comprehensively interpret the influence of comorbidities, treatment regimens, and other variables on the expression of differentially expressed genes and the prognosis of patients. Additionally, further exploration into the correlations between the genetic model and the APACHEII score, SOFA score, procalcitonin, serum lactate levels, immunosuppression, organ failure, pathogen type, and the value differences in predicting the prognosis of patients was not possible. Secondly, in this study, the CIBERSORT analysis relied on the whole blood transcriptome, which might not accurately reflect the tissue-specific immune response. We observed a correlation between the 3 genes and immune cells such as monocytes and conducted preliminary work. It is important to conduct further experiments to assess the robustness of the prognostic risk model. In the future, more works need to be done to investigate how they are involved in the regulation of monocytes and how the prognostic model could be integrated with clinical practice.

5. Conclusion

In conclusion, we have developed and validated a robust prognostic risk model that includes 3 genes: MPEG1, CX3CR1, and HLA-DRB1. The expression of the 3 genes had a significant negative relationship with the 28-day mortality of sepsis. Additionally, the potential impact on immune function, specifically the regulation of monocytes, necessitates further investigation through in vitro and animal experimentation to confirm and elucidate the underlying mechanisms.

Author contributions

Conceptualization: Yiqian Zeng, Suna Peng.

Data curation: Yutian Liao, Suna Peng.

Formal analysis: Yiqian Zeng, Yutian Liao, Yang Wang, Suna Peng.

Funding acquisition: Yiqian Zeng.

Investigation: Yiqian Zeng.

Methodology: Yiqian Zeng, Suna Peng.

Software: Yiqian Zeng, Yutian Liao.

Supervision: Suna Peng.

Validation: Yang Wang.

Visualization: Yutian Liao, Yang Wang, Suna Peng.

Writing — review & editing: Suna Peng.

Writing – original draft: Yiqian Zeng, Yutian Liao, Yang Wang, Suna Peng.