Establishment of a Predictive Model for Mortality in Sepsis Patients Using WGCNA and Machine Learning Algorithms

Abstract

Background

Predicting mortality in sepsis patients remains a challenge in clinical settings. This study aimed to identify genes associated with sepsis-related mortality and determine the predictive value of a model using these death-related genes.

Methods

We employed Weighted Gene Co-expression Network Analysis (WGCNA), differential expression analysis, and machine learning (ML) algorithms to identify genes related to sepsis mortality by analyzing the GSE185263 dataset. A predictive model was constructed base on the genes related to sepsis mortality and validated using external cohorts. Blood samples from 120 sepsis patients were collected to test the expression of key genes. The predictive value of the key genes was verified in clinical samples.

Results

WGCNA results revealed that the turquoise module was related to sepsis mortality. Thirty-five death-related genes were identified by intersecting module genes with differentially expressed genes between survivor and non-survivor samples. Three ML algorithms identified TMTC1, AMOTL1, and PTGFR as key genes. The model constructed with these three genes showed moderate predictive value. External cohorts with a total of 523 samples validated the predictive value of the model at moderate to high levels. Clinical sample results indicated that all three genes were elevated in deceased sepsis patients, and the predictive value was moderate. Combining the model with the APACHE II score achieved high predictive accuracy.

Conclusion

This study identified and validated three genes related to sepsis mortality and established a predictive model that performed well in distinguishing sepsis patients at high risk of death.

Introduction

Sepsis, a severe and life-threatening condition triggered by the body’s response to infection, presents a significant risk of mortality.¹ Various factors have been documented to contribute to the likelihood of death in sepsis patients, including advanced age, pre-existing chronic conditions (such as diabetes, cancer, and liver disease), a compromised immune system, and the presence of invasive medical devices.² Additionally, genetic predispositions and gene expression patterns are related to sepsis mortality. Mechanistically, aberrant expression of key genes exacerbates dysregulation of immune response, inflammation, coagulation, and cellular metabolism, resulting in abnormal cytokine and chemical production, and pathogen recognition.^3,4 Therefore, timely identification and management of these risk factors are crucial for reducing sepsis-associated mortality.

Currently, clinical indicators such as C-reactive protein (CRP), lactate, procalcitonin (PCT), and scoring systems like the Acute Physiology and Chronic Health Evaluation II (APACHE II) and Sequential Organ Failure Assessment (SOFA) are routinely used to predict sepsis mortality. However, their predictive accuracy varies significantly across studies, limiting their universal applicability.^5–7 For instance, PCT exhibits suboptimal sensitivity and specificity for differentiating sepsis from other causes of systemic inflammatory responses.⁸ Fundamentally, these traditional indicators primarily reflect established organ dysfunction or systemic inflammation at a single time point, often lagging behind the underlying molecular pathophysiology. Furthermore, sepsis is a highly heterogeneous syndrome, and conventional markers may lack sufficient specificity for precise risk stratification across diverse patient subpopulations. In contrast, novel gene-based signatures hold promise as complementary prognostic tools because they can capture early host response dynamics before overt clinical deterioration, reflect specific molecular mechanisms driving disease progression, and potentially offer improved predictive performance in heterogeneous cohorts. Identifying robust gene signatures is therefore critical for advancing precision medicine in sepsis management.

Recognizing this urgent need, researchers have increasingly explored both individual biomarkers and computational models to improve prognostication. For example, pentraxin 3 (PTX3) has demonstrated moderate predictive capability for sepsis mortality with an area under the curve (AUC) of 0.73,⁹ and machine learning algorithms utilizing clinical parameters have achieved an AUC of 0.79 for predicting 28-day mortality in sepsis-associated liver injury.¹⁰ More specifically, given the critical role of aberrant gene expression in sepsis progression, several transcriptomic signatures have been developed. A five-gene panel (NCOA4, GABARAPL1, GJA1, CISD1, CP) has shown strong predictive potential for sepsis-induced cardiomyopathy,¹¹ while signatures comprising 37 genes¹² or 7 genes (AUC = 0.89)¹³ have been reported for general sepsis survival prediction. Despite these advances, many existing signatures suffer from excessive gene numbers that hinder clinical translation, or remain unvalidated in independent cohorts, leaving a substantial gap in identifying concise, reliable, and broadly applicable prognostic biomarkers.

To address this gap, this study aimed to screen for key genes associated with sepsis mortality and construct a concise, robust gene signature for prognostic prediction. By integrating weighted gene co-expression network analysis (WGCNA) with multiple machine learning algorithms, we identified a novel three-gene signature and rigorously validated its predictive value and robustness using independent clinical samples. Our findings may facilitate the development of targeted therapies and personalized treatment strategies, ultimately contributing to improved survival rates and patient outcomes in sepsis.

Materials and Methods

Dataset Selection and Data Preprocessing

For this research, we sourced three sepsis-related datasets from the GEO database, each containing information on survival and deceased patients. Specifically, the GSE54514 dataset comprised 36 healthy controls, 31 non-survivors, and 96 survivors; the GSE236713 dataset included 91 non-survivors and 356 survivors; and the GSE63042 dataset contained 28 non-survivors and 68 septic patients. Samples associated with Systemic Inflammatory Response Syndrome (SIRS) were excluded from our analyses. Collectively, we analyzed data from 1,779 subjects, with their baseline characteristics detailed in Supplementary Table S1. Prior to conducting differential expression analysis, all datasets underwent preprocessing, which involved background correction using the RMA algorithm. To identify differentially expressed genes (DEGs) within the GSE54514 cohort, we employed the “limma” package, comparing gene expression profiles among healthy controls, non-survivor patients, and survivor patients. Ethics committee was not necessary because the data were freely provided by the GEO database.

Weighted Gene Co-Expression Network Analysis (WGCNA)

To uncover significant modules and genes linked to non-survivor septic patients, we applied WGCNA to the GSE54514 dataset. Initially, we established a soft-thresholding power (beta value) based on the scale-free topology fit index to build a correlation-based adjacency matrix. Subsequently, we utilized the dynamic tree cut method for module identification. During the selection process, the adjacency matrix was transformed into a topological overlap matrix (TOM), and modules were delineated through hierarchical clustering. We also performed sample clustering to explore the association between gene expression patterns and clinical traits. From the modules that exhibited strong correlations with clinical features and possessed high intramodular connectivity, we extracted trait-associated genes. These genes represent potential biomarkers or therapeutic targets for sepsis, particularly in non-survivor cases.

Gene Function Enrichment Analysis

To elucidate the biological roles of status-related genes, we performed Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analyses. The GO analysis encompassed three categories: Biological Process (BP), Molecular Function (MF), and Cellular Component (CC). KEGG analysis was employed to pinpoint significantly enriched pathways. We utilized the “clusterProfiler” package for conducting both GO and KEGG analyses, with a threshold of p-value < 0.05 set to denote statistically significant enrichment.

Screening and Validation of Key Genes Using Multiple Machine Learning Algorithms

For the identification of key genes, we applied a multi-step approach using the GSE54514 dataset. After excluding healthy controls, samples were randomly divided into training and testing sets in a 7:3 ratio. Three machine learning algorithms, including Support Vector Machine (SVM), Decision Tree (DT), and Extreme Gradient Boosting (XGBoost), were implemented via the e1071, rpart, and xgboost R packages, respectively, to identify the most important genes. The intersecting genes from these three algorithms were selected as common markers. These common genes were then cross-referenced with DEGs identified among healthy controls, non-survivor patients, and survivor patients to isolate key genes associated with sepsis survival. A multivariate logistic regression model was subsequently constructed using these key genes to develop a predictive signature for sepsis outcome.

Patient Selection and Clinical Data

This study recruited consecutive sepsis patients from the First and Tenth Affiliated Hospitals of Guangxi Medical University between January 2023 and January 2024. Ethical approval was obtained from the Ethics Committee of the Tenth Affiliated Hospital of Guangxi Medical University (No. 201612), and all procedures adhered to the ethical guidelines outlined in the Declaration of Helsinki. Sample size calculations were based on statistical power considerations, ensuring a minimum of 50 patients per group. The diagnosis of septic shock was established according to the definition of the Third International Consensus Definitions for Sepsis and Septic Shock (Sepsis-3 definition).¹⁴

Inclusion criteria included adult patients (≥18 years) admitted to the ICU with a diagnosis of sepsis, for whom only data from their first ICU stay were collected, and those with an ICU stay of at least 24 hours. Exclusion criteria comprised pregnant women, patients in an agonizing state, and those under limited therapeutic efforts. Samples meeting only SIRS criteria without confirmed or suspected infection were excluded from the analysis. SIRS criteria exhibit limited specificity for sepsis, as they can be triggered by various non-infectious conditions. Including SIRS-positive but non-septic samples would introduce substantial confounding heterogeneity into the transcriptomic dataset, potentially diluting true sepsis-specific gene expression signals and compromising the validity and generalizability of the identified prognostic signature. Ultimately, our study cohort consisted of 60 surviving and 60 non-surviving septic patients. Demographic information, including age, sex, laboratory test results, APACHE II scores, and mortality rates, was collected for all participants. In our study, the usage of antibiotics only referred to those new antibiotic prescriptions after ICU admission, excluding patients already receiving antibiotics prior to enrollment, which accounts for the low usage rate of antibiotics. The clinical characteristics of the patient cohort are summarized in Table 1.

Table 1.

Baseline of the Septic Patients

Variables	Total	Dead (N=60)	Survival (N=60)	P Value
Age (years)	68.67±15.03	68.67±15.89	68.68±14.25	0.995
Gender				0.429
Female	37 (30.83%)	16 (26.7%)	21 (35%)
Male	83 (69.17%)	44 (73.3%)	39 (65%)
Comorbidity				0.620
COPD	4 (3.33%)	1 (1.7%)	3 (5%)
Diabetes	39 (32.50%)	20 (33.3%)	19 (31.7%)
Heart disease	27 (22.50%)	15 (25%)	12 (20%)
Hypertension	28 (23.33%)	16 (26.7%)	12 (20%)
Liver failure	5 (4.17%)	1 (1.7%)	4 (6.7%)
Trauma	4 (3.33%)	2 (3.3%)	2 (3.3%)
Tumor	13 (10.83%)	5 (8.3%)	8 (13.3%)
APACHE II score	29.21±9.81	31.63±9.21	26.78±9.86	0.006
WBC (10×9)	12.89±7.23	13.29±8.60	12.49±5.60	0.551
Platelet (10×9)	177.94±132.04	153.52±141.71	202.37±117.76	0.042
Surgery	107 (89.17%)	53 (88.3%)	54 (90%)	1.000
Infection site				0.074
Abdomen	11 (9.17%)	5 (8.3%)	6 (10%)
Blood	29 (24.17%)	20 (33.3%)	9 (15%)
Other	19 (15.83%)	5 (8.3%)	14 (23.3%)
Respiratory tract	57 (47.50%)	28 (46.7%)	29 (48.3%)
Urethra	4 (3.33%)	2 (3.3%)	2 (3.3%)
Treatment				0.004
Antibiotic	5 (4.17%)	5 (8.3%)	0 (0%)
Dialysis	7 (5.83%)	7 (11.7%)	0 (0%)
Mechanical ventilation	83 (69.17%)	36 (60%)	47 (78.3%)
Vasopressor boosting	25 (20.83%)	12 (20%)	13 (21.7%)
Bacteria*				0.483
Fungus	18 (15.00%)	9 (15%)	9 (15%)
Gram negative	60 (50.00%)	33 (55%)	27 (45%)
Gram positive	42 (35.00%)	18 (30%)	24 (40%)
Total bilirubin (mg/dl)	44.42±73.58	51.95±82.45	36.90±63.29	0.265
Blood sugar (mg/dl)	11.28±9.20	11.65±10.69	10.91±7.49	0.662
ScVo2 (%)	75.56±16.16	73.03±14.35	78.09±17.55	0.087
CRP (mg/L)	65.09±47.98	87.63±50.75	42.55±32.17	<0.001
Procalcitonin (ng/mL)	19.40±41.52	16.34±26.55	22.46±52.48	0.422

Notes: *Bacteriological results were derived from the following culture sources: blood cultures, sputum cultures, and urine cultures.

Abbreviations: APACHE II, Acute Physiology and Chronic Health Evaluation II; COPD, Chronic obstructive pulmonary disease; WBC, white blood cells; CRP, C-reactive protein.

Peripheral Venous Blood Sample Collection

Peripheral venous blood samples, each 10 mL in volume, were collected into tubes containing EDTA as an anticoagulant. The samples were processed for the isolation of peripheral blood mononuclear cells (PBMCs), which were subsequently stored at −80°C until further analysis.

Real-Time Quantitative Polymerase Chain Reaction (RT-qPCR) Assay

To validate the findings from our bioinformatics analysis, we assessed the expression levels of key genes in PBMCs using RT-qPCR. Total RNA was extracted from tissues and PBMCs using TRIzol reagent (Thermo Fisher Scientific, USA) following the manufacturer’s protocol. Gene-specific primers for RT-qPCR are listed in Supplementary Table S2. The RT-qPCR reactions were performed using the SYBR® Premix Ex Taq kit (Takara, Dalian, China). Relative gene expression levels were quantified using the 2^−ΔΔCT method.

Statistical Analysis

All statistical analyses were conducted using R software (Version 4.1.2). Continuous data between two groups were compared using either the independent Student’s t-test or the Wilcoxon rank-sum test, depending on the data distribution. Predictive models were developed using logistic regression for binary outcomes or Cox proportional hazards regression for survival analysis. The performance of these models was evaluated using receiver operating characteristic (ROC) curves, with the area under the curve (AUC) serving as a quantitative measure of predictive accuracy. A p-value < 0.05 was considered statistically significant.

Results

Screen Hub Genes Associated with Mortality in Sepsis Patients Through WGCNA Method

The study design, data preparation, and analysis workflow are illustrated in Figure 1. The GSE54514 dataset underwent analysis utilizing the WGCNA method to uncover modules and hub genes linked to mortality in sepsis patients. By establishing the beta value as six based on scale-free topology (Figure 2A), five distinct modules emerged through gene profile clustering (Figure 2B). Subsequently, integration of clinical features with these modules revealed the correlation of clinical attributes with specific modules, with a notable emphasis on the turquoise module in relation to sepsis patient status (Figure 2C). Correlation analysis substantiated a significant relationship between the genes within this module and clinical outcomes (Figure 2D). Consequently, the 594 genes contained in this module were identified as most closely associated with the status of sepsis patients. Subsequent steps involved the identification of DEGs between deceased patients and survivors, resulting in the recognition of 208 DEGs (Figure 2E). Lastly, overlaying the 594 module-related genes with the 208 DEGs led to 35 overlapping genes (Figure 2F). These 35 genes were deemed predominantly associated with sepsis patient mortality, thereby serving as pivotal components for subsequent analyses. Notably, SOFA scores did not consistently align with ultimate survival outcomes in this cohort. This discrepancy likely arises because SOFA captures acute organ dysfunction at a single time point, whereas mortality is determined by a complex interplay of longitudinal disease trajectory, comorbidities, therapeutic interventions, and host resilience that a static snapshot cannot fully encompass.

Figure 1.

The flowchart of present study.

Figure 2.

Screen hub genes associated with death of septic patients by WGCNA method. (A) The soft-thresholding powers and the scale-free fit index. (B) Dynamic dendrogram of all genes clustered based on a dissimilarity measure. (C) Heatmap of the correlation between module eigengenes and clinical features of sepsis. (D) Scatter plot of the correlation of gene significance with turquoise module in sepsis. (E) Volcano of differentially expressed genes (DEGs) between survival and death sepsis patients. (F) Venn plot of intersection of the DEGs and the turquoise module genes.

Functional Enrichment Analysis for Death-Related Genes

The GO analysis revealed that the 35 death-related genes of sepsis were mainly involved in the BP of positive regulation of protein phosphorylation, positive regulation of penile erection, and positive regulation of MAP kinase activity. The CC component is enriched in presynaptic active zone membrane, plasma membrane, and ciliary transition zone. The MF is enriched in peptide hormone binding (Figure 3A). The KEGG pathway analysis revealed that these genes mainly enriched in the TGF-beta signaling pathway, small cell lung cancer, p53 signaling pathway and Oocyte meiosis (Figure 3B).

Figure 3.

Functional enrichment analysis for death-related genes. (A) GO analysis for the death-related genes of sepsis. (B) KEGG enrichment analysis for the death-related genes of sepsis.

Screening Key Genes Related to Sepsis Mortality Using ML Algorithms

To identify key genes associated with sepsis mortality, the GSE185263 dataset underwent analysis utilizing three ML algorithms (LASSO, SVM, and XGBoost) after excluding healthy samples. Subsequently, sepsis survival and death samples were randomly split into training and test sets at a 7:3 ratio. Incorporating the 35 septic status-related genes into each ML algorithm, LASSO, SVM, and XGBoost revealed 5, 27, and 23 genes, respectively, linked to sepsis mortality in the training sets. The ROC curve illustrated that all models exhibited moderate predictive efficacy in distinguishing sepsis mortality from survival cases within the testing sets (Figure 4A–C). By intersecting the genes identified by the algorithms, it was observed that three genes (TMTC1, AMOTL1, and PTGFR) were common across all three algorithms (Figure 4D). Subsequently, a multivariate logistic regression analysis was employed to develop a predictive model utilizing these three genes. The optimal cutoff value of 0.186 for the model discriminating between the two groups indicated that patients exceeding this threshold were categorized as part of the death group; otherwise, they were classified under the survival group. The ROC curve showcased an AUC value of 0.719, with a sensitivity and specificity of 0.788 and 0.615, respectively (Figure 4E). Moreover, within the GSE185263 dataset, the Sequential Organ Failure Assessment (SOFA) score of the samples was available. Evaluation of the predictive value of SOFA yielded an AUC value of 0.699 (Figure 4F), akin to that of the three-gene model. These findings underscored that the predictive model constructed with these three genes effectively discerns sepsis patients at high risk of mortality.

Figure 4.

Screen key genes related to septic survival by machine learning (ML) algorithms. (A–C) ROC analysis for the predictive value in testing sets of GSE185263 dataset by SVM, XGBoost, and LASSO algorithms. (D) Venn plot of the intersection of genes from the SVM, XGBoost, and LASSO algorithms. (E) ROC analysis for the predictive value of the combination of three genes in predicting death of sepsis patient by logistic regression analysis in GSE185263 dataset. (F) ROC analysis for the predictive value of SOFA score in predicting death of sepsis patient in GSE185263 dataset.

Validation of the Predictive Model in External Cohorts

To validate the predictive capacity of the model established with the three genes for sepsis patients at risk of mortality, we utilized three external cohorts - namely, the GSE54514, GSE236713, and GSE63042 datasets. Samples not classified as sepsis were excluded, and the remaining samples were categorized into survival and death groups. Initially, we assessed the expression levels of the three genes in the survival and death groups and subsequently evaluated the predictive efficacy of the gene combination. Figure 5A–C revealed that the expression patterns of TMTC1, AMOTL1, and PTGFR varied across the three datasets. In the GSE54514 dataset, all three genes displayed significantly increased expression levels in the death group compared to the survival group. However, the differences in expression were less pronounced in the GSE236713 and GSE63042 datasets. ROC analysis indicated that the predictive value of the three-gene combination was moderate, yielding AUC values of 0.912, 0.645, and 0.681 in the respective datasets (Figure 5D–F). In summary, these findings suggest that the expression profiles of the three genes exhibit variations across different cohorts. Furthermore, the predictive ability of the gene combination for identifying sepsis patients at risk of mortality was deemed moderate, underscoring the potential utility of these genes in prognostic assessments.

Figure 5.

Validation of the predictive model in external cohorts. (A–C) Comparison of TMTC1, AMOTL1, and PTGFR expression between survival and death sepsis patients in GSE54514, GSE236713, and GSE63042 datasets. (D–F) ROC analysis for the predictive value in predicting death of sepsis patients in GSE54514, GSE236713, and GSE63042 datasets. ****p<0.0001.

Abbreviation: ns, not significant.

Association of the Predictive Model with the Sepsis Clinical Parameters and Survival Outcome

The expression of the three genes (TMTC1, AMOTL1, and PTGFR) was evaluated using RT-qPCR assay in PBMCs from 60 sepsis survivors and 60 non-survivors, and a predictive model was established. Figure 6A showed that the expression of the three genes was significantly higher in non-survival septic patients compared with survival septic patients (p<0.001).

Figure 6.

Association of the predictive model with the sepsis clinical parameters and survival outcome. (A) Comparison of TMTC1, AMOTL1, and PTGFR expression between survival and death sepsis patients in clinical samples; (B) Correlation analysis among predictive model with laboratory test indexes; (C) Comparison of high- and low- risk groups of model regarding septic patients’ gender, comorbidity, surgery history, infection site, treatment modalities and infection types; (D) ROC analysis for the predictive value in predicting death of sepsis patients of each gene; (E) ROC analysis for the predictive value in predicting death of sepsis patients of predictive model, APACHE score and the combination. ****p<0.0001.

Subsequently, a predictive model was constructed using the three genes based on a logistic regression model. Through correlation analysis, no significant correlation was identified between the model and the laboratory test indexes (Figure 6B). We then divided the patients into high-risk and low-risk groups based on the median value of the model. No significant discrepancies were observed between the two groups concerning gender, comorbidities, site of infection, treatment modalities, and infection types (Figure 6C), indicating an association of high-risk sepsis patients with unfavorable outcomes.

We then determined the predictive value of each gene for patient mortality utilizing ROC analysis. The results showed that all three genes presented moderate predictive value for the mortality of septic patients (TMTC1: AUC=0.737; AMOTL1: AUC=0.766; PTGFR: AUC=0.795; Figure 6D). We also examined the predictive value of the model and the APACHE score, as well as the combination of the model and APACHE score. The results showed that the predictive value of the model (AUC=0.838) was much higher than that of the APACHE score (AUC=0.629), and the combination of the model and APACHE score yielded an AUC value of 0.860, indicative of a high predictive capability (Figure 6E). Taken together, these findings suggest that the predictive model incorporating the three genes exhibited superior performance in predicting survival outcomes for sepsis patients.

Discussion

The identification of sepsis patients at high risk of adverse outcomes poses a significant challenge due to the substantial clinical and biological heterogeneity among individuals, alongside the intricate dynamics of the host immune response. In this study, we initiated our investigation by employing the WGCNA algorithm to screen genes associated with mortality in sepsis patients utilizing extensive gene datasets. Through this process, we pinpointed a module directly linked to sepsis-related mortality. Subsequently, by overlapping these module-related genes with DEGs between surviving and deceased patients, we identified genes specifically associated with mortality. Following this step, we utilized three machine learning algorithms to identify key genes and successfully screen three genes crucial for predicting outcomes. Leveraging these genes, we developed a predictive model that demonstrated moderate to high predictive accuracy when validated across three external cohorts. Furthermore, assessment of gene expression and predictive performance using clinical samples yielded results consistent with those obtained from the initial gene datasets, highlighting a better predictive value compared to the APACHE II score. This underscored the model’s effectiveness in discerning sepsis patients at heightened risk of mortality.

Although the addition of APACHE II score to the gene-based model yielded a numerically higher AUC (0.860 vs 0.838), the difference was not statistically significant. Rather than indicating redundancy, this finding carries important clinical implications. The three-gene signature alone achieves prognostic accuracy comparable to a composite model incorporating a well-established clinical scoring system. This suggests that the gene signature provides robust, objective molecular information that is largely independent of subjective clinical assessment. In clinical settings where comprehensive APACHE II scoring is impractical, delayed, or subject to inter-rater variability, the gene-based model offers a standardized and readily quantifiable alternative. We therefore propose that this gene signature serves as a complementary tool to the existing clinical prognostic frameworks.

Existing evidence underscores the pivotal role of differential gene expression in influencing pathogen recognition efficiency, clearance mechanisms, and the body’s capacity to repair infection-induced tissue damage.¹⁵ Thus, identification of these genes could help in developing better therapeutic interventions to manage and treat sepsis. Identifying such genes holds promise for advancing therapeutic strategies in sepsis management. Notably, numerous genes have been implicated in sepsis development, progression, and outcomes. For instance, S100A9 exacerbates sepsis-induced acute lung injury via the IL17-NFkappaB-caspase-3 pathway.¹⁶ Additionally, C3AR1 serves as a robust biomarker indicating poorer outcomes in sepsis patients, associated with immune and coagulation cascades.¹⁷ Various genes like TNF-α, TLR, IL1RN, and GLCCI1 have demonstrated either protective or deleterious roles in septic patients.¹⁸ Moreover, certain genes or molecular pathways play critical roles in mediating treatment responses in sepsis, including the C/EBPbeta-Dependent MAPK pathway,¹⁹ SIRT1-DRP1/PGC-1alpha pathway.²⁰

In this study, we employed WGCNA in conjunction with three machine learning algorithms to identify key genes associated with sepsis-related mortality. This approach offers a more robust and reliable method compared to traditional differential expression analysis.²¹ WGCNA stands out for its ability to construct a scale-free network that delineates gene modules and their correlations with phenotypic traits. By highlighting highly connected hub genes within biologically significant modules, WGCNA provides a solid foundation for unraveling the genetic underpinnings of complex traits.²² The utilization of machine learning algorithms further enhances our capacity to pinpoint key genes by effectively managing large datasets and offering predictive insights that might not be readily discernible through conventional methodologies. This leads to a more precise and comprehensive selection of genes.^23,24

Notably, TMTC1, AMOTL1, and PTGFR have not been previously implicated in sepsis within existing literature. PTGFR has been observed to promote PTGS-2 and growth factors in bovine endometrial epithelial cells,^25,26 which might critical for orchestrating the host inflammatory response during sepsis. AMOTL1 enhances YAP1 stability and promotes YAP1-driven gastric oncogenesis²⁷ as well as mediating the sequestration of the Hippo effector YAP1 downstream of Fat4 to restrict heart growth,²⁸ its dysregulation may contribute to the capillary leak and microvascular dysfunction hallmark of septic shock. In contrast, TMTC1 has been linked to promoting invasiveness in ovarian cancer cells²⁹ and serves as a repeat-containing adapter protein involved in calcium homeostasis.³⁰ These findings underscore the potential significance of these three genes in the context of sepsis, particularly concerning mortality. Further research into the roles of TMTC1, AMOTL1, and PTGFR in sepsis is warranted, as they represent promising candidates for exploring novel mechanisms underlying sepsis pathogenesis and outcomes.

The current evidence regarding genes or signatures associated with sepsis mortality remains limited. Our study sought to address this gap by employing multiple algorithms to identify key genes linked to sepsis-related mortality, leveraging large datasets that were validated across external cohorts and clinical samples. This rigorous approach ensures the robustness of our findings and unveils new potential markers for prognostic indicators in sepsis management. Nevertheless, it is crucial to acknowledge several limitations within our study. Firstly, the heterogeneity presents in the included datasets, encompassing variations in sepsis diagnosis criteria, sites of infection, types of infection, or treatment regimens, may introduce variability in gene expression profiles, potentially impacting result robustness. Secondly, the absence of certain information, such as lactate levels, prevented us from comparing the predictive value with our identified gene signature. Thirdly, while we identified key genes associated with sepsis mortality, the underlying mechanisms of these genes in relation to sepsis require further exploration through in vitro and in vivo experiments.

Conclusions

The present study identified and validated three genes linked to sepsis mortality, establishing a novel three-gene predictive model for assessing mortality risk among sepsis patients. This model exhibited strong performance in distinguishing sepsis patients at high risk of death, showcasing its potential utility in clinical settings for prognostic purposes.

Funding Statement

This study was partially supported by research funding from the Science and Technology Projects of Guangxi (No. Z20251098).

Data Sharing Statement

The datasets used and/or analyzed in the study are available from the corresponding author on reasonable request.

Ethics Approval Statement

This study was conducted in accordance with the principles of the Declaration of Helsinki. The study protocol was approved by the both of Ethics Committee of The First People’s Hospital of Qinzhou (Approval No. KY-20250711). All the participants provided written informed consent. The blood samples were obtained from patients as part of routine hospital procedure, the ethical approval was obtained for the use of these samples. GEO database belongs to public databases. The patients involved in the database have obtained ethical approval. Users can download relevant data for free for research and publish relevant articles.

Author Contributions

Zhao-yin Fu: Conceptualization, Methodology, Data curation, Visualization, Funding acquisition, Writing - original draft, Writing- review & editing.

Yu-long Bai: Formal analysis, Software, Writing - original draft, Writing -review & editing.

Yu-biao Xu: Investigation, Data curation, Writing - review & editing.

Shi-hui Tan: Formal analysis, Software, Writing - original draft.

Yi-sheng Huang: Data curation, Formal analysis, Writing - review & editing.

Xiao-xiang Huang: Supervision, Methodology, Writing – review & editing, Resources.

All authors made a significant contribution to the work reported, whether that is in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work. Yu-long Bai and Yu-biao Xu are co-first authors and have contributed equally to this work.

Disclosure

The authors declare that they have no conflict of interest.