Integrated bioinformatics analysis unravels mitochondrial-immune crosstalk and infiltration dynamics in sepsis progression

Abstract

Background

Sepsis is a critical illness, and mitochondrial dysfunction is associated with its progression. However, the classification of mitochondrial-related differentially expressed genes (MitoDEGs) in sepsis and the immune infiltration characteristics have not been thoroughly investigated. This study aimed to explore the relevant content.

Methods

Gene expression data were obtained from the Gene Expression Omnibus (GEO), while mitochondrial-related genes were sourced from the MitoCarta3.0 database. We applied Weighted Gene Co-expression Network Analysis (WGCNA) to identify Sepsis-related MitoDEGs (Se-MitoDEGs), and utilized unsupervised clustering analysis to categorize sepsis samples into distinct clusters. Machine learning algorithms identified hub Se-MitoDEGs, and a validation set and a nomogram for sepsis diagnosis were established. The CIBERSORT algorithm was employed to investigate immune infiltration characteristics in sepsis and their association with hub Se-MitoDEGs. The expression levels of relevant genes were evaluated in peripheral blood samples from septic patients and normal controls through quantitative real-time reverse transcription polymerase chain reaction (qRT-PCR). Associated transcription factors, miRNAs, and drugs were constructed into a diagram via NetworkAnalyst and Comparative Toxicogenomics Database (CTD).

Results

15 Se-MitoDEGs exhibited differential expression between septic and normal samples. Immune infiltration analysis demonstrated significant increases in neutrophils, activated mast cells, and M0 macrophages among septic patients compared to control subjects. We categorized sepsis samples into two clusters; most hub genes in cluster 2 (C2) were highly expressed, exhibiting low immune infiltration and immune score. Some differences were observed in the pathways between the two clusters. By utilizing machine learning techniques and the validation set, MSRB2, TSPO, and BLOC1S1 were identified, and a nomogram of the three genes exhibited a substantial area under the curve (AUC) of 0.886, and the AUC for the validation set was recorded at 0.866, highlighting the robustness of our predictive model. Survival analysis found that low expression of TSPO and high expression of MSRB2 in peripheral blood were negatively correlated with the 28-day survival rate of septic patients. qRT-PCR validation indicated that the expression levels of these three hub genes are consistent with our bioinformatics analysis results. Associated small molecules, including Estradiol, pirinixic acid, and Valproic acid, are potential therapeutic drugs for sepsis.

Conclusion

By integrating bioinformatics with machine learning models, we identified three mitochondrial and immune-related biomarkers (MSRB2, TSPO, and BLOC1S1) with diagnostic value for sepsis. These biomarkers provide new insights into subtype stratification, immune infiltration characteristics, and targeted therapy in sepsis.

Supplementary Information

The online version contains supplementary material available at 10.1186/s40001-025-03142-w.

Introduction

Sepsis is a critical medical condition associated with organ dysfunction and high mortality. Annually, about 49 million people worldwide are impacted, leading to nearly 11 million deaths and accounting for approximately 20% of all global deaths [1]. The mortality rate of severe sepsis ranges from 25 to 30% [2]. Traditional diagnostic methods primarily depend on clinical indicators such as the systemic inflammatory response syndrome (SIRS) and the Sequential Organ Failure Assessment (SOFA) score. However, these often lack specificity and may fail to accurately reflect the underlying pathophysiological mechanisms [3, 4]. Emerging research underscores the fact that molecular biomarkers and unique gene expression profiles enhance sepsis diagnosis and management [5]. Mitochondrial dysfunction is widely recognized to be pivotal in terms of sepsis development. Mitochondria are the primary sites of cellular energy production, and play crucial roles in immune cell activation, proliferation, and differentiation [6]. Mitochondrial dysfunction is a key component of the immune dysregulation associated with sepsis, characterized by a reduced mitochondrial membrane potential, diminished respiratory chain activity, and decreased reactive oxygen species (ROS) production [7–10]. These changes trigger an energy deficiency in immune cells, impairing cell functions. The resulting immunosuppression elevates the infection risk [11]. Furthermore, ROS and cytokines released from mitochondria exacerbate inflammatory responses, ultimately worsening the condition [12]. Researchers studying other diseases, such as cancer and neurodegenerative disorders, have used mitochondrial gene expression profiles to identify disease-specific biomarkers and explore potential therapeutic targets [13]. Changes in mitochondrial-related gene expression play pivotal roles in the immune dysregulation associated with sepsis [14]. These genetic alterations not only reflect the impact of sepsis on mitochondria but may also serve as potential biomarkers of the condition and as clinically therapeutic targets. An in-depth investigation of mitochondrial gene expression patterns in sepsis would significantly advance our understanding of the immunological pathogenesis. However, few studies have investigated the patterns of mitochondria-related gene expression and the associations thereof with immune infiltration in the context of sepsis [15]. Interventions that counter mitochondrial dysfunction might improve the immune function and prognosis of sepsis patients.

Here, we used microarray data from the Gene Expression Omnibus (GEO) database to identify differentially expressed genes (DEGs) in sepsis patients compared to healthy controls. The DEGs were next cross-referenced with mitochondrial genes cataloged in the MitoCarta3.0 database; this identified MitoDEGs. We used protein–protein interaction (PPI) networks and weighted gene co-expression network analysis (WGCNA) to understand interactions among these genes and to derive co-expression patterns. We measured the immune cell infiltrations when determining correlations between MitoDEG expression and the immune cell populations in the context of sepsis and classified the mitochondrial gene expression patterns of sepsis patients. We finally used machine learning algorithms to derive the areas under the curve (AUCs) of critical genes in the validation set. This identified hub genes. We used these data to develop a predictive model, the effectiveness of which was subsequently validated using a nomogram and an integrated external dataset. Our findings offer new insights into the molecular mechanisms of sepsis, aiding the identification of biomarkers that may facilitate early diagnosis and targeted treatment to enhance patient outcomes. The entire workflow is illustrated in Fig. 1.

Fig. 1.

A detailed workflow about the study of MitoDEGs in sepsis

Materials and methods

Identification of DEGs

We downloaded the gene expression profile microarrays (GSE32707, GSE69528, GSE9960, GSE13904, GSE54514, GSE95233) of normal and septic blood samples from the GEO (https://www.ncbi.nlm.nih.gov/geo/). The specific details are in Additional file 1, Table 1. The R ‘sva’ package analyzes high-dimensional data, especially those of gene expression studies. The package removes unwanted variations, enhancing the accuracy of subsequent analyses [16]. We used this package (version 3.52.0, accessed June 24, 2024) to eliminate batch effects. The GSE32707 and GSE69528 datasets were combined to form a sepsis training set (Additional file 2: Fig. S1A-D). The amalgamated GSE9960 and GSE13904 datasets were the sepsis validation set (Additional file 3: Fig. S2A-D). The Limma package [17] of R (version 3.60.4, accessed June 24, 2024) enables simple analysis of differences and visualization of the results, and was employed to ascertain DEGs in septic compared to normal samples of the training set. The cutoff thresholds were p < 0.05 and |log2FC (fold-change) |> 0.5 (Additional file 2: Fig. S1E).

Identification of MitoDEGs

We accessed the MitoCarta3.0 mitochondrial protein database (http://www.broadinstitute.org/mitocarta) [18], which contains information on 1,136 genes associated with mitochondrial functions. MitoDEGs were determined by overlapping the DEGs of the training dataset with these 1,136 mitochondrial genes, and a Venn diagram was employed for visualization. The R ‘ComplexHeatmap’ [19] package (version 0.95, accessed June 24, 2024) was used to generate a heat map and R ‘ggplot2’ [20] (version 3.5.1, accessed June 24, 2024) to create a volcano plot.

PPI and correlation analyses of MitoDEGs

We used the STRING database (https://string-db.org/) in conjunction with R Cytoscape (version 3.9.0, accessed June 25, 2024) for PPI and correlation analyses. The software contains functions that greatly aid the creation, analysis, and visualization of biological networks [21]. The PPIs of MitoDEGs were examined.

WGCNA

The R ‘WGCNA’ package constructs gene co-expression networks and identifies gene modules [22]. The package was employed for weighted analysis that revealed co-expression modules associated with sepsis. Initially, we used the optimal soft threshold to formulate an adjacency matrix via computation and transformed this into a topological overlap matrix. We then created various modules using a hierarchical clustering method and assigned random hues to each module. A color variation indicated a significant difference. The genes in the most significant sepsis block subsequently underwent correlation analysis.

Sepsis-related MitoDEG extraction and GO and KEGG enrichment analysis

Genes that were both in the sepsis-related module and among the MitoDEGs were termed Se-MitoDEGs. The ‘corrplot’ package of R (version 0.95, accessed June 26, 2024) quickly and intuitively visualizes correlations [23]. Correlations among Se-MitoDEGs were assessed using the Spearman correlation coefficient. The Se-MitoDEGs were subjected to genetic enrichment analyses using the Gene Ontology database (GO, https://geneontology.org/ accessed July 1, 2024) and the Kyoto Encyclopedia of Genes and Genomes (KEGG, https://www.kegg.jp/ accessed July 1, 2024) of the R ‘clusterProfiler’ package (version 4.12.6, accessed July 1, 2024), which identifies gene sets associated with specific biological processes, molecular functions, or cellular components [24]. The significance threshold was set to p < 0.05.

Immune infiltration

The R ‘cibersort’ package (version 1.04, accessed July 1, 2024) quantifies the immune cell proportions in septic and control samples; the cell distributions are distinct [25]. We used this package to explore immune cell infiltration. Bar graphs present the proportions of all immune cell types in different samples and facilitate comparative analysis of the cells of septic and normal samples. A heatmap was created using the R ‘corrplot’ package (version 0.95, accessed July 1, 2024) to depict correlations between immune infiltrating cells and the 15 Se-MitoDEGs.

Construction of unsupervised septic clusters

The R ‘ConsensusClusterPlus’ package (version 1.2.0, accessed July 1, 2024) [26] was used for unsupervised clustering of Se-MitoDEGs that exhibited distinct patterns of mitochondria-related genes expressed in sepsis patients. The optimal number of subtypes (denoted k) was determined by evaluating the characteristics of the cumulative distribution function (CDF) curve, including curve inclination and smoothness, and the consensus score and consensus matrix.

Gene set variation analysis

The ‘c5.go.symbols’ and ‘c2.cp.Kegg.symbols’ files of gene set variation analysis (GSVA) were obtained from the MSigDB database (https://www.gsea-msigdb.org/gsea/msigdb, version 7.4, accessed July 1, 2024) [27]. The R package ‘GSVA’ was used to investigate the modified pathways and biological functions associated with the various Se-MitoDEG clusters.

Selection of hub Se-MitoDEGs via LASSO and random forest

We employed the least absolute shrinkage and selection operator (Lasso) [28] (version 4.1–8, accessed July 1, 2024) and random forest (version 4.7–1.2, accessed July 1, 2024) [29] algorithms to identify hub MitoDEGs, and constructed a Venn diagram of feature gene overlap. We used the R ‘pROC’ package (version 0.7.7, accessed July 1, 2024) to generate receiver operating characteristic (ROC) curves and to calculate the areas under the curves (AUCs) that revealed the predictive utilities of genes in a validation set. The R ‘rms’ package (version 6.2.0, accessed July 1, 2024) was used for regression analysis, and to build and evaluate a model via simple function calls [30]. Finally, we developed a nomogram using the R ‘survival’ package [31] (version 3.5.7, accessed August 16, 2025) and the GSE54514 (31 sepsis nonsurvivors vs. 96 survivors) and GSE95233 (34 nonsurvivors vs. 68 survivors) datasets to integrate 28-day survival status with peripheral blood gene expression data from septic patients. We evaluated the prognostic utility of each gene using the Cox method.

Clinical data and qRT-PCR

This study involved 12 participants, including six sepsis patients hospitalized at the Second Affiliated Hospital of Soochow University within 48 h of diagnosis and six healthy controls, with no significant differences in between the groups. The expression levels of MSRB2、TSPO and BLOC1S1 were assessed using qRT-PCR. The sepsis criteria fulfilled the diagnostic requirements outlined in Sepsis 3.0 [32]. Prior to data collection, all participants provided written informed consent, and the study protocol was approved by the Ethics Committee of the Second Affiliated Hospital of Soochow University (ethical approval number: JD-LK2023025-I01). Peripheral blood was collected in citrate-containing tubes, and RNA was subsequently extracted from the blood samples using the TRIzol reagent (Invitrogen, CA, USA), adhering to the manufacturer's instructions, and the RNA concentration was measured using NanoDrop2000 (Thermo Fisher, USA). The isolated RNA was then reverse-transcribed into complementary DNA using a reverse transcription kit with SynScript®III RT SuperMix from Tsingke in Peking, China. Following this, qRT-PCR was performed with the ArtiCanCEO SYBR qPCR Mix kit from Tsingke in Peking, China, utilizing the ABI Prism 7500 Sequence Detection System from Applied Biosystems in the USA. To assess the relative changes in mRNA expression levels of the target genes across various samples, the 2^−△△Ct method was employed. The internal control for mRNA was β-actin. Each sample underwent triplicate analysis to guarantee accuracy and reproducibility. The primers utilized for qRT-qPCR in this study were designed and validated by AZENTA Biotech (AZENTA Biotech, Soochow, China), detailes were in Additional file 4: Table 2. Statistical significance was assessed with a threshold of p < 0.05. The clinical data of sepsis patients were collected, including age, sex, inflammatory index, SOFA score, Acute Physiology and Chronic Health Evaluation II (APACHEIIscore), lymphocyte typing and 28-day survival. The relationships between the expression levels of the three genes and clinical characteristics in septic patients were presented using a correlation heatmap (Additional file 5: Table 3).

Single-gene GSEA

Single-gene gene set enrichment analysis (GSEA) (version 1.42.0, accessed July 1, 2024) was used to divide samples into high-expression (≥ 50%) and low-expression (< 50%) groups based on the median levels of diagnostic genes. To investigate the underlying molecular mechanisms associated with the phenotypes, we obtained the ‘c2.cp.kegg.v7.4.symbols.gmt’ subset of the Molecular Signatures Database (MSigDB; https://www.gsea-msigdb.org/gsea/msigdb, version 7.4, accessed July 1, 2024). The minimum gene set size was 5, and the Maximum 5,000; we performed 1,000 resampling iterations. A p-value < 0.05 was deemed statistically significant.

Correlations among immune-infiltrating cell numbers and the expression of signature genes

The correlations between Se-MitoDEG expression levels and the numbers of immune infiltrating cells were calculated, followed by Spearman rank correlation analysis (version 4.1.1, accessed July 4, 2024) to examine the associations between immune cells and hub genes. Lollipop plots were generated using the R ‘ggplot’ package (version 3.5.1).

Construction of regulatory networks

The regulatory structures of diagnostic biomarkers among microRNAs (miRNAs) and transcription factors (TFs) derived from hub genes were established with the aid of NetworkAnalyst (https://www.networkanalyst.ca) and the network visualized employing Cytoscape (version 3.9.0, accessed July 4, 2024). We used the comparative troxicogenomics database (CTD) (http://ctdbase.org/) to identify non-biomarker materials associated with sepsis. Finally, we created a Sankey diagram illustrating the various chemical-gene interactions.

Statistical analysis

Statistical evaluations were conducted utilizing R software (version 4.3.1), SPSS Statistics 23 (IBM Corporation, NY, USA), and GraphPad Prism 10.0 (GraphPad Software, CA, USA). The determination of significant differences between the two groups was carried out employing either non-parametric tests or t-tests, contingent upon the distribution characteristics of the data. All statistical tests were performed as two-tailed, with a significance threshold set at p < 0.05.

Results

Identification of MitoDEGs and WGCNA

Batch effects initially present in GSE32707 and GSE69528 were removed (Additional file 2: Fig. S1A-D). The integrated sepsis training datasets revealed 2,460 DEGs, of which 1,243 were downregulated and 1,217 upregulated in septic samples (N = 162) compared to control samples (N = 62) (Additional file 2: Fig. S1E). After cross-referencing these 2,460 DEGs with 1,136 genes associated with mitochondrial function, we identified the 145 overlapping genes shown in the Venn diagram of Fig. 2A. The heat map (Fig. 2B) and volcano plot (Fig. 2C) illustrate the top 30 MitoDEGs that distinguish septic from normal samples. PPI network analysis of these MitoDEGs revealed that GFM1 exhibited the most interactions with other genes, followed by MTIF2 and NDUFS8 (Fig. 2D). The sepsis training dataset (merged datasets GSE32707 and GSE69528) contained 62 normal and 162 septic samples, which were selected for clustering after exclusion of low-quality samples employing a predetermined threshold. The soft threshold power was set to 7 when establishing a scale-free network; the R² value was 0.86 (Fig. 2E–F). Dynamic TreeCut analysis identified 20 distinct modules (Fig. 2G). Correlations among these modules were next sought (Fig. 2H). Notably, the midnight blue module correlated most strongly with sepsis [correlation coefficient 0.5 and a highly significant p-value of 1.5e-15 (Fig. 2I)]. Therefore, the 1,009 genes of this module were subjected to further analysis, which confirmed a significant positive correlation between the midnight blue module and associated genes (Fig. 2J).

Fig. 2.

MitoDEGs between normal and sepsis groups, and Construction and module analysis of WGCNA. A A Venn diagram illustrates the overlap between mitochondria-related genes and DEGs in the sepsis and normal groups. B A heatmap displays the expression patterns of the 30 most significant MitoDEGs, identified based on their lowest p-values, in both groups. C A volcano plot depicts the distribution of MitoDEGs between sepsis and normal samples. D An interaction network diagram highlights the protein–protein interactions among the 145 MitoDEGs. E and F Selection of a soft threshold of β = 7 and evaluation of the scale-free topology fitting index (R²). G Dendrogram of co-expressed genes, where different colors represent distinct gene co-expression modules. H Collinearity heatmap of module feature genes, with red indicating high correlation and blue indicating the opposite trend. I Heatmap displaying module-feature relationships, including associated p-values and correlation coefficients for each pair. J Scatter plot illustrating the gene correlation of the midnight blue module with the sepsis group. MitoDEGs, mitochondrial-related differentially expressed genes; DEGs, differentially expressed genes; WGCNA, weighted gene co-expression network analysis

Identification of Se-MitoDEGs

Sepsis-specific WGCNA revealed 15 genes that were both module-associated genes and MitoDEGs (Se-MitoDEGs; Fig. 3A). Specifically, the expression levels of PGS1, ECHDC3, TSPO, MSRB2, RAB24, BLOC1S1, MSRA, ACAA1, CYP27A1, SLC25A44, LYRM1, SFXN5, PDK3, and SPTLC2 were significantly elevated in sepsis samples, and that of THNSL1 reduced (Fig. 3B). In the Figure, *** denotes a p-value < 0.001, ** a p-value < 0.01, and * a p-value < 0.05. The heatmap visually depicts the differences in expression in whole bloods of septic patients and normal individuals (Fig. 3C). We found various correlations among the differentially expressed MitoDEGs (Fig. 3D). KEGG pathway enrichment analysis indicated that the top five pathways enriched in Se-MitoDEGs were cholesterol metabolism, the PPAR signaling pathway, metabolic pathways, primary bile acid biosynthesis, and alpha-linolenic acid metabolism (Fig. 3E). GO enrichment analysis highlighted that Se-MitoDEGs were significantly enriched in mitochondrial regulation, components, and membranes (Fig. 3F).

Fig. 3.

Identification and functional enrichment analysis of Se-MitoDEGs. A Identification of Se-MitoDEGs using a Venn diagram. B Comparison of expression levels of the 15 Se-MitoDEGs between the sepsis and control groups. C Heatmap illustrating the differences in Se-MitoDEGs between the sepsis and control groups. D Correlation analysis among the differentially expressed MitoDEGs. E KEGG pathway analysis for Se-MitoDEGs. (F) GO biological process analysis for Se-MitoDEGs. Se-MitoDEGs, Sepsis-related mitochondrial-related differentially expressed genes; MitoDEGs, Mitochondrial-related differentially expressed genes; KEGG, Kyoto Encyclopedia of Genes and Genomes; GO, Gene Ontology

Evaluation of immune cell infiltration

The levels of 22 infiltrating immune cell types differed between the sepsis and control groups (Fig. 4A). Sepsis was associated with significantly more infiltration of neutrophils, activated mast cells, M0 macrophages, monocytes (all p < 0.001); and gamma-delta T cells (p < 0.01). In contrast, the normal group exhibited significantly higher levels of eosinophils, resting mast cells (both p < 0.05); M2 macrophages, follicular helper T cells, naive CD4 T cells (all p < 0.01); and resting NK cells, memory resting CD4 T cells, CD8 T cells, and naive B cells (all p < 0.001) (Fig. 4B). Furthermore, correlation analysis revealed a complex relationship between immune cell numbers and Se-MitoDEGs (Fig. 4C), suggesting that Se-MitoDEGs significantly influenced the immune infiltration landscape of sepsis.

Fig. 4.

Overview of immune infiltration in sepsis. A Relative abundance of 22 infiltrating immune cells between sepsis and normal groups. B Bar chart showing the difference in immune infiltration between sepsis and normal groups. C Correlation analysis of 22 differentially expressed Se-MitoDEGs with infiltrating immune cells. *p < 0.05, **p < 0.01, ***p < 0.001. Se-MitoDEGs, Sepsis-related mitochondrial-related differentially expressed genes

Mitochondria-related clusters in sepsis

We used a consensus clustering algorithm to classify 162 sepsis samples by the expression levels of the 15 Se-MitoDEGs (Fig. 5A). We varied k from 2 to 10, and found that, when k = 2, the consensus index of the CDF curve exhibited minimal fluctuations but a relatively high consensus score (Fig. 5B). Consequently, we classified all sepsis patients into two clusters: Cluster 1 (C1) with 83 and Cluster 2 (C2) with 79 and analyzed the expression levels of the 15 Se-MitoDEGs. C2 exhibited significantly higher expression of ACAA1, BLOC1S1, ECHDC3, LYRM1, MSRA, MSRB2, PDK3, PGS1, RAB24, SFXN5, SLC25A44, SPTLC2, and TSPO; and C1 increased expression of CYP27A1 and THNSL1 (Fig. 5C–D). The immune cell infiltrations differed significantly between the two clusters: C1 exhibited more immune infiltration and a higher immune score and C2 the opposite (Fig. 5E–G). GSVA demonstrated that C1 was significantly associated with the p53 signaling pathway, and that C2 was predominantly linked to asthma (Fig. 5H). GSVA indicated that C1 exhibited a strong correlation with interleukin 1 receptor binding. Conversely, C2 was primarily associated with tumor necrosis factor receptor binding and natural killer cell cytokine production (Fig. 5I).

Fig. 5.

Identification and evaluation of molecular clusters associated with Se-MitoDEGs in sepsis. A Consensus clustering matrix at k = 2. B CDF curves representing k values of 2–10, respectively. C Boxplot of differentially expressed Se-MitoDEGs between mitochondria‐related clusters. D Heatmap of differentially expressed Se-MitoDEGs between mitochondria‐related clusters. E Relative abundance of 22 types of infiltrating immune cells in mitochondria‐related clusters. F Boxplot of immune‐related cells between mitochondria‐related clusters. G Boxplot of immune score between mitochondria‐related clusters. H GSVA results of the KEGG gene set between two mitochondria‐related clusters were plotted in the bar graph. I GSVA results of the GO set between two mitochondria‐related clusters are plotted in the bar graph. *p < 0.05, **p < 0.01, ***p < 0.001,****p < 0.0001. Se-MitoDEGs Sepsis-related mitochondrial-related differentially expressed genes, CDF Cumulative distribution function, GSVA Gene Set Variation Analysis, KEGG Kyoto Encyclopedia of Genes and Genomes, GO Gene Ontology

Identification of hub Se-MitoDEGs via LASSO and random forest

We used two algorithms to identify biomarkers among the 15 MitoDEGs that might predict sepsis onset. Initially, we employed the Lasso-Cox method for regression analysis, using a tenfold cross-validation technique to determine the optimal model. The lambda value was 0.0313105716925459. This identified six genes: ACAA1, BLOC1S1, MSRB2, RAB24, THNSL1, and TSPO (Fig. 6A–B). Next, the random forest algorithm identified all genes of the 15 with relative significance scores ≥ 2 (Fig. 6C–D). ACAA1, BLOC1S1, MSRB2, RAB24, THNSL1, and TSPO converged in both machine learning models (Fig. 6E). The ROC curves for the six genes that diagnosed sepsis in both the training and validation sets (Fig. 6F–G) indicate that MSRB2, TSPO, and BLOC1S1 were significantly more predictive in both datasets. In the validation set, both the bar charts and a heat map show that the expression levels of these three biomarkers were higher in sepsis patients (Fig. 6H-I).

Fig. 6.

Identification and evaluation of the signature genes. A Ten cross-validations of adjusted parameter selection in the LASSO model. Each curve corresponds to one gene. B LASSO coefficient analysis. Vertical dashed lines are plotted at the best lambda. C Relationship between the number of random forest trees and error rates. D Ranking of the relative importance of genes. E Venn diagram showing the feature genes shared by LASSO and random forest. F ROC curves illustrate the diagnostic performance of the six signature genes in the training set (GSE32707 and GSE69528). G ROC curves depict the diagnostic efficacy of the six signature genes as validated in the validation set (GSE9960 and GSE13904). In both (F) and (G), MSRB2, TSPO, and BLOC1S1 demonstrated notably high predictive values. H and I The expression levels of MSRB2, TSPO, and BLOC1S1 in sepsis versus control samples from the validation set are illustrated in bar charts and a heatmap. LASSO least absolute shrinkage and selection operator, RF random forest. ROC Receiver Operating Characteristic

Development of a nomogram and evaluation of the diagnostic utility thereof

Nomograms that diagnosed sepsis incorporated three specific genes (Fig. 7A). ROC curve analysis revealed good performance; the AUC was 0.886 (Fig. 7B). The calibration curve (Fig. 7C) confirmed accuracy. Decision curve analysis demonstrated that the nomogram could aid the management of sepsis patients (Fig. 7D). The validation datasets further supported reliability; the AUC was 0.866 (Fig. 7E). The diagnostic accuracy was remarkable; the nomogram is clinically applicable. Survival analysis indicated that low TSPO expression in peripheral blood (GSE54514) and high MSRB2 expression (GSE95233) negatively correlated with 28-day survival (Fig. 7F–G). No significant correlation was found between BLOC1S1 expression and 28-day survival in the GSE95233 dataset (Fig. 7H). The clinical data further validated that the sepsis group exhibited higher expression levels of MSRB2, TSPO, and BLOC1S1 (Fig. 7I-K). Analysis of the clinical data of 12 patients revealed that the AUC of our predictive model was 0.967 (95% confidence interval [CI] 0.874–1.000), and that of the sepsis 3.0 standard 1.000 (95% CI 1.000–1.000) (Fig. 7L). Our nomogram is almost as accurate as the current diagnostic standard. However, given the small sample size, the results may be subject to certain biases. The correlation thermogram showed that the genes were closely related, and the expression levels correlated to various extents with the clinical indicators of sepsis (Fig. 7M).

Fig. 7.

Development and validation of nomogram of hub genes, and relevant clinical data. A Nomogram for assessing the risk of sepsis development. B Nomogram’s ROC curve for diagnosing sepsis. C Nomogram’s calibration curve. D Decision curve analysis with nomogram model. These findings showed the potential utility of the selected genes as biomarkers for sepsis and underscore the efficacy of the proposed predictive model in clinical diagnosis. E ROC curves from the validation set. F Analysis of peripheral TSPO expression and 28-day survival in septic patients in GSE54514 dataset. G–H Analysis of peripheral MSRB2 and BLOC1S1 expression and 28-day survival in septic patients in GSE95233 dataset, respectively. I–K qRT-PCR of MSRB2, TSPO and BLOC1S1 in sepsis patients were significantly upregulated. All data were presented as the mean ± SD of triplicate experiments. L Comparison of AUC in diagnosing sepsis: Nomogram’s model vs. SOFA score ≥ 2 (M) The expression levels of MSRB2, TSPO and BLOC1S1 have some correlation with the clinical data of sepsis patients. *p < 0.05, **p < 0.01, ***p < 0.001. ROC Receiver Operating Characteristic

The possible biological roles of diagnostic Se-MitoDEGs

GSEA was used to seek potential signaling pathways linked to the signature genes. The phenotype characterized by high-level MSRB2 expression was significantly enriched in the NFKB pathway and TLR4 signaling, among other biological processes (Fig. 8A). The low-TSPO expression phenotype was linked to the CD28-dependent PI3K AKT and CTLA4 pathways, and the high-TSPOi phenotypic enrichment of HIF1A and PPARG involved the glycolytic, STAT3, and IL1 pathways (Fig. 8B). The low-BLOC1S1 expression phenotype was enriched in pathways associated with CD28-dependent PI3K AKT signaling and apoptosis (Fig. 8C). These findings yield crucial insights into mechanisms that may contribute to sepsis.

Fig. 8.

GSEA identifies signaling pathways, immune infiltration and construction of the regulatory network of the hub genes. A–C Significantly enriched pathways for high or low expression of the signature genes: (A) MSRB2, (B) TSPO, and (C) BLOC1S1. (D–F) Correlation analysis of immune infiltration with signature gene expression in the combined GSE32707 and GSE69528 datasets: (D) MSRB2, (E) TSPO, and (F) BLOC1S1. Dot size indicates the strength of correlation with immune cells, with larger dots reflecting stronger correlations. Dot color represents p-values, with bluer shades denoting lower p-values. Red numbers indicate statistical significance, with p < 0.05 considered statistically significant. G–H Construction of the regulatory network. G A miRNA–hub gene–TF regulatory network was established using Cytoscape, where the red circles represent hub genes, light blue diamonds indicate miRNAs, and orange triangles denote TFs. H The drug-gene interaction network related to sepsis was developed based on the Comparative Toxicogenomics Database (CTD), identifying 34 chemicals associated with the key genes. GSEA Gene Set Enrichment Analysis, TFs transcription factors, CTD Comparative Toxicogenomics Database

Immune infiltration correlation analyses of three genes

We sought correlations between the expression levels of three diagnostic genes in the training dataset, and immune cell infiltrations. MSRB2 was positively correlated with infiltration of neutrophils, M0 macrophages, monocytes, activated mast cells, M1 macrophages, and regulatory T cells (Tregs); but negatively correlated with infiltration of eosinophils, memory B cells, follicular helper T cells, M2 macrophages, resting NK cells, resting CD4 memory T cells, and CD8 T cells (Fig. 8D). TSPO was positively correlated with infiltration of M0 macrophages, monocytes, neutrophils, activated mast cells, and Tregs; but negatively correlated with infiltration of resting mast cells, plasma cells, activated memory CD4 T cells, helper follicular T cells, M2 macrophages, memory B cells, resting memory CD4 T cells, resting NK cells, and CD8 T cells (Fig. 8E). BLOC1S1 was positively correlated with infiltration of monocytes, neutrophils, activated mast cells, and Tregs; but negatively correlated with infiltration of activated dendritic cells, memory B cells, helper follicular T cells, naive CD4 T cells, M2 macrophages, resting NK cells, resting memory CD4 T cells, and CD8 T cells (Fig. 8F). In conclusion, these critical diagnostic genes may play roles in the immune regulatory mechanisms of sepsis development.

Construction of regulatory networks

To investigate the regulatory mechanisms of hub genes, we predicted interacting TFs and miRNAs. We then applied a filter > 1 to refine the selection. The interaction network (Fig. 8G) revealed five transcription factors (MAZ, GLIS2, GTF2E2, ZBTB11, and KLF8), three Hub genes, and 21 miRNAs. We used Cytoscape to create a regulatory network among these components to better understand the transcriptional mechanisms that controlled the Hub genes. We also built a sepsis drug-gene interaction network using the CTD data and identified 34 chemicals associated with essential genes (Fig. 8H), of which several, including estradiol, pirinixic acid, and valproic acid, targeted more than two common genes.

Discussion

Sepsis is a life-threatening emergency and the pathogenesis is complex [33]. Mitochondrial dysfunction plays roles in sepsis development and progression [34, 35]. Mitochondria regulate the inflammatory responses and immune cell actions [6, 11]. This study employed microarray data and advanced bioinformatics techniques to identify MitoDEGs linked to sepsis. By integrating PPI networking, WGCNA, immune infiltration profiles, and MitoDEGs-based classification, we elucidated the relationships between mitochondrial function and the immune response of sepsis, offering new insights in terms of sepsis detection and treatment.

Fifteen Se-MitoDEGs were identified. PGS1, ECHDC3, TSPO, MSRB2, RAB24, BLOC1S1, MSRA, ACAA1, CYP27A1, SLC25A44, LYRM1, SFXN5, PDK3, and SPTLC2 were all significantly elevated in septic compared to control patients, but THNSL1 expression was notably decreased, highlighting the importance of MitoDEGs in terms of sepsis development. Furthermore, infiltrations of neutrophils, activated mast cells, and M0 macrophages were significantly increased in patients with sepsis. Healthy individuals exhibited more eosinophils, resting mast cells, and M2 macrophages, consistent with previous studies [36, 37]. Immune cell dysregulation is important in the sepsis context. Neutrophils trigger hyperinflammation and tissue damage [36], and the increased levels of undifferentiated M0 macrophages suggest disruption of the normal macrophage polarization that resolves inflammation and promotes tissue repair [37].

Unsupervised clustering of the 15 genes identified two distinct groups: C1 (n = 83) and C2 (n = 79). GSVA showed that C1 was closely associated with the p53 signaling pathway, which plays a crucial role in apoptosis by suppressing the autophagic protein Bcl-2 [38]. GSVA indicated that C2 was closely associated with binding of the tumor necrosis factor receptor and the MHC protein complex, and natural killer cell cytokine production. A previous study found that TRAF6-induced inflammation contributed to lipopolysaccharide-induced acute kidney injury [39]. Feng [40] reported a negative correlation between the NK cell proportion among lymphocytes and 28-day mortality of septic patients. Sepsis-related phenotypic changes, including increases in the levels of inhibitory receptors and reduced NK cell effector function, may trigger immunosuppression. The number of NK cells in C2 was lower than that in C1 (Fig. 7D), suggesting that distinct mitochondrial-related gene patterns affect NK cell function during sepsis. A focus on mitochondria-related genes and the interactions thereof with immune cells may improve sepsis treatment. Transcriptomics (analysis of gene expression profiles) lays the foundation for molecular subtyping and precise stratification of sepsis. A prospective cohort study using whole-blood transcriptomic data classified sepsis patients into four subtypes (Mars 1 to 4). Patients with the Mars 1 subtype were at a significantly higher risk of mortality than others (hazard ratio = 1.86). The present study identified a further 140 key genes (including BPGM and TAP2) that may serve as subtype biomarkers aiding the early identification of high-risk patients [41]. Single-cell RNA sequencing (scRNA-seq) enhances the resolution of traditional transcriptomics; changes in gene expression are apparent at the single-cell level. Yao et al. [42] identified an HLA-DR^lowS100A^high monocyte subpopulation via cross-species analysis, highlighted the central role thereof in the immune suppression associated with sepsis, and identified novel cellular molecular markers for dynamic monitoring of immune status. Notably, different septic sites may exhibit distinct gene enrichments. Transcriptomic studies of sepsis caused by fecal peritonitis or community-acquired pneumonia revealed that the transcriptomic responses were generally unrelated to the source of infection, Rather reflecting the immune response status and patient prognosis. Among different sources of infection, 263 genes were differentially regulated, particularly those associated with interferon-mediated signal transduction and antigen presentation [43]. Our findings are consistent with those of a previous study [44], reinforcing the suggestion that transcriptomic analysis of gene expression profiles may aid the classification of molecular sepsis subtypes and guide personalized treatments.

The current Sepsis 3.0 Guidelines view sepsis as an acute organ dysfunction syndrome attributable to dysregulation of the host immune response after infection. These Guidelines recommend use of the SOFA to assess end-organ injury; an increase of 2 indicates life-threatening dysfunction [45]. We included six sepsis patients and six normal controls; the sepsis inclusion criterion was a SOFA score ≥ 2. The clinical data revealed that the AUC of our predictive model was 0.967 (95% CI 0.874–1.000), and that of the sepsis 3.0 standard 1.000 (95% CI 1.000–1.000). The accuracy of our nomogram is very similar to that of the current diagnostic standard. However, given the small sample size, certain biases May be in play. The Sepsis 3.0 diagnostic criteria quantify the organ dysfunctions associated with sepsis, but lack biomarkers of infection, diagnosis, and prognosis. The gold standard infectious diagnosis is pathogen identification, but the time required for culture is often long (at least 2–3 days) and the positivity rates are low. As early specific diagnostic markers of sepsis are lacking, some patients may be diagnosed late, associated with suboptimal treatment outcomes and poor prognosis. Early diagnostic biomarkers are urgently required. Biomarkers aid diagnosis of a disease or condition, assist disease monitoring or progression, reveal therapeutic or pharmacodynamic responses, predict or stratify patient prognosis, and identify patient subgroups that respond differently to interventions [46]. Ideal biomarkers are rapidly analyzed, reproducible, cost-effective, and differentiable. Most sepsis biomarkers are one of two types: (1) identify pathogens or (2) detect host response dysregulation and predict prognosis [47]. Sepsis is characterized by high-level heterogeneity, and the complex immune mechanisms in play pose significant challenges to both basic researchers and clinicians but also offer opportunities to explore new therapies. In future, precision medicine and personalized immune interventions may offer hope. Novel biomarkers would be of great assistance.

Based on our bioinformatics data and clinical validation, elevated MSRB2, TSPO, and BLOC1S1 expression are diagnostic of sepsis, and these molecules may play roles in the complex pathogenesis. Survival analysis indicated that low-level TSPO expression (GSE54514) and high-level MSRB2 expression (GSE95233) in peripheral blood were negatively correlated with 28-day survival (Fig. 7F–G). No significant correlation was found between BLOC1S1 expression and 28-day survival in the GSE95233 dataset. MSRB2 encodes methionine sulfoxide reductase B2 that protects cells against oxidative stress induced by sepsis, associated with cell damage and inflammation. MSRB2 upregulation mitigates ROS accumulation and preserves mitochondrial integrity [48, 49]. Low expression of MSRB2 has been linked to Alzheimer’s disease, Parkinson’s disease, cardiovascular conditions, and malignant tumors [50]. TSPO of the outer mitochondrial membrane is involved in steroid synthesis, programmed cell death, and cellular stress responses [51]. Significant associations have been found between low TSPO expression and Alzheimer’s disease, mental health disorders, cardiovascular conditions, and diabetes [52–58]. TSPO deficiency exacerbated NLRP3 inflammasome-driven pyroptosis and sepsis-induced acute lung injury [59]. BLOC1S1 is involved in intracellular vesicle transport and signal transduction, playing a key role in protein acetylation within the electron transport chain, in turn directly affecting mitochondrial oxygen consumption and ATP production levels [60]. BLOC1S1 levels are abnormal in patients with schizophrenia, Th2 cell-related inflammatory disorders, and high-grade serous ovarian cancer [61–63]. Notably, BLOC1S1 expression is increased in individuals who succumb to sepsis [64]. We found that, compared to the control group, MSRB2, TSPO, and BLOC1S1 were expressed at significantly higher levels in the sepsis cohort, and MSRB2 and TSPO were significantly associated with prognosis, indicating that these genes may be promising therapeutic targets.

GSEA of these hub genes indicates that they are involved in various biological processes and signal transduction pathways, such as the NFKB pathway, EPO receptor signaling, and CD28-dependent PI3K/AKT signaling. The NFκB pathway is crucial for regulating the immune response and inflammation; however, its disruption during sepsis may result in heightened inflammation and tissue injury [65, 66]. Enhancing the EPO receptor signal may involve regulating inflammatory response and providing protection against tissue damage during sepsis [67]. The signal pathway involving CD28-dependent PI3K/AKT plays a vital role in the activation and survival of T cells, thus emphasizing the significance of immune regulation in sepsis [68]. Our investigation revealed particular Se-MitoDEGs linked to immune cell infiltration, highlighting the interplay between mitochondrial impairment, immune infiltration, and inflammation in sepsis. The identified pathways, especially the NFκB signaling pathway, can be a potential therapeutic direction for mitochondrial-related genes to regulate immune response and improve prognosis in sepsis patients. However, further experiments are needed to confirm the regulatory interaction between these hub genes and explore the mechanisms of various signal pathways related to sepsis.

Furthermore, we extensively analyzed these characteristic genes, emphasizing their correlation with immune cell infiltration and investigating their interaction networks with microRNAs (miRNAs), transcription factors (TFs), and pharmacological regulation. Our results showed that the following chemicals had more than two common target genes: bisphenol A, Benzo(a)pyrene, Carbon Tetrachloride, Doxorubicin, Estradiol, pirinixic acid, Tetrachlorodibenzodioxin, Thioacetamide, and Valproic Acid. Consistent with previous studies [69–71], our results show that small molecular compounds related to hub genes, such as Estradiol, pirinixic acid, and valproic acid, may be potential therapeutic drugs for sepsis. However, the exact interaction mechanism between these hub genes and the above drugs needs further exploration.

Despite employing a comprehensive bioinformatics approach, this study has limitations. First, it remains unclear whether gene expression levels vary among individuals from different regions or ethnic backgrounds. Second, integrating multiple data sets enhances the robustness of results but increases the risk of batch effects, which may not be fully mitigated by the'sva'package, potentially impacting the identification of DEGs. Thirdly, the small number of clinical patients included in our study requires future studies to evaluate further these genes'diagnostic and prognostic value in septicemia patients with larger sample sizes or multi-center clinical studies. Finally, this study is based on bioinformatics analysis, but lacks a survey of relevant mechanisms. We will conduct in vivo animal studies and in vitro cell experiments to elucidate MSRB2, TSPO, and BLOC1S1 expression patterns and their mechanisms in immune regulation during sepsis.

Conclusion

This study identified several MitoDEGs associated with sepsis, detailed the expression patterns, examined the roles played in terms of immune infiltration, and explored the diagnostic potential. Sepsis patients could be divided into two clusters, with the C2 cluster exhibiting high-level expression of hub genes, low immune infiltration, and reduced immune scores. Estradiol, pirinixic acid, and valproic acid are promising sepsis therapeutics. This study enhances our understanding of sepsis pathophysiology and offers insights into potential therapeutic drugs and immunotherapeutic strategies.

Abbreviations

MitoDEGs

Mitochondrial-related differentially expressed genes

GEO

Gene expression omnibus

WGCNA

Weighted gene co-expression network analysis

Se-MitoDEGs

Sepsis-related mitochondrial-related differentially expressed genes

qRT-PCR

Quantitative real-time reverse transcription polymerase chain reaction

CTD

Comparative toxicogenomics database

C2

Cluster 2

MSRB2

Methionine sulfoxide reductase B2

TSPO

Transporter protein

BLOC1S1

Biogenesis of lysosomal organelles complex 1 subunit 1

AUC

Area under the curve

SIRS

Systemic inflammatory response syndrome

SOFA

Sequential organ failure assessment

ROS

Reactive oxygen species

PPI

Protein–protein interaction

TOM

Topological overlap matrix

GO

Gene ontology

KEGG

Kyoto encyclopedia of genes and genomes

CDF

Cumulative distribution function

GSVA

Gene set variation analysis

LASSO

Least absolute shrinkage and selection operator

ROC

Receiver operating characteristic

TFs

Transcription factors

SOFA

Sequential organ failure assessment

APACHE

Acute physiology and chronic health evaluation II

GSEA

Gene set enrichment analysis

miRNAs

MicroRNAs

Tregs

Regulatory T cells

C1

Cluster 1

AD

Alzheimer’s disease

PD

Parkinson’s disease

Author contributions

Fanjian Meng: Writing – original draft, Visualization, Validation, Software, Methodology, Data curation. Anyuan Zhong: Writing – original draft, Visualization, Software, Investigation, Data curation. Ting Li: Writing – original draft, Visualization, Software. Yun Yang: Writing – review & editing, Validation. Chen Chen: Writing – review & editing, Data curation. Yongkang Huang: Writing – review & editing, Software. Tong Zhou: Writing – review & editing, Software. Yongjian Pei: Writing– review & editing, Supervision, Funding acquisition, Conceptualization. Minhua Shi: Writing – review & editing, Investigation, Funding acquisition, Conceptualization. All authors reviewed the manuscript.

Funding

This work was supported by the Science and Education Strengthening Health Project of Suzhou, China (grant no. QNXM2024019); This work was supported by the Collaborative Research Project of Shanghai Xinxin Medical Science and Technology Development Foundation, China (grant no. ZYY20241074).

Data availability

All data in this study are included in this article and its supplementary information fles.

Declarations

Ethics approval and consent to participate

This study were approved by the Ethics Committee of the Second Affiliated Hospital of Soochow University (ethical approval number: JD-LK2023025-I01), conducted in accordance with ethical principles for medical research involving human subjects reported in the Declaration of Helsinki and its later amendments. All the sequencing data used in this article were obtained from public databases, and initial ethical approval was provided in the original article.

Consent for publication

All authors know and approve the publication of this manuscript.

Competing interests

The authors declare no competing interests.