Identification and exploration of novel biomarkers and potential therapeutic agents for the progression of sepsis to septic ARDS

Abstract

Patients with sepsis complicated by acute respiratory distress syndrome (ARDS) face a significantly increased risk of in-hospital death. This study aimed to identify sepsis-associated genes involved in ARDS pathogenesis and discover candidate biomarkers for its diagnosis. Gene expression profiling data from the Gene Expression Omnibus database were analyzed to identify key septic ARDS genes using differential expression analysis and weighted gene co-expression network analysis. Functional enrichment and connectivity map analyses were performed to explore underlying mechanisms and potential therapeutic drugs. Machine learning algorithms were applied to screen biomarkers and construct a diagnostic nomogram. Receiver operating characteristic, calibration, and decision curve analyses evaluated its diagnostic performance. Single-sample gene set enrichment analysis and peripheral blood mononuclear cell sequencing were used to assess immune cell infiltration. We identified 49 key genes linked to septic ARDS, enriched in inflammatory response regulation, reactive oxygen biosynthesis, and immune suppression. Connectivity map analysis suggested etoposide as a potential ARDS treatment. Three hub genes, ATPase plasma membrane Ca²⁺ transporting 1 (ATP2B1), retinol binding protein 7 (RBP7), and absent in melanoma 2 (AIM2), were selected as candidate biomarkers for nomogram construction, demonstrating ideal diagnostic performance. Immune cell infiltration analysis revealed immune downregulation in ARDS, with ATP2B1, RBP7, and AIM2 significantly associated with immune dysregulation. This study uncovered inflammatory immune pathways in ARDS and developed an ATP2B1/RBP7/AIM2-based nomogram for septic ARDS diagnosis, offering new insights for diagnosis and therapeutic interventions.

1. Introduction

Sepsis is a common cause of poor prognosis in intensive care unit (ICU) patients, and sepsis-related deaths account for 19.7% of all deaths worldwide.^[1] The development of acute respiratory distress syndrome (ARDS) among ICU patients with sepsis confers an increased risk of ICU and in-hospital mortality.^[2] However, the diagnosis of ARDS mainly relies on functional diagnosis, such as SpO₂/FiO₂, oxygenation index, or physician judgment with limited resources (the new global definition in 2023),^[3] and there is still a lack of early accurate diagnostic markers based on pathogenesis.

Abnormalities in various immune cells are critical contributors to autoimmune disruption in sepsis-induced ARDS. We found that dendritic cells accumulate and activate in the lungs, ultimately leading to acute lung injury in mice after intra-airway injection of lipopolysaccharide (LPS).^[4,5] Monocytes, macrophages, and neutrophils have also been shown to be closely related to the development of septic ARDS.^[6–8] However, the mechanism of ARDS is not fully understood. Exploring the key immune regulatory molecules may provide important insights for the early diagnosis and treatment of sepsis-related ARDS.

Predictive biomarker screening of respiratory critical illness based on gene networks has had some good results in studies in the biomedical field,^[9,10] such as those for distinct chronic obstructive pulmonary disease subtypes in former smokers revealed by gene network perturbation analysis by Buschur et al. Machine learning can be utilized to detect hard-to-recognize patterns from large, noisy, or complex datasets and is particularly suited for data analysis applications in the medical field; machine learning has often been used in recent years for ARDS diagnosis and subtype analysis.^[11] The emergence of modern computer-assisted medical science, such as semisupervised deep learning, harmony search, AdaBoost, and Directionality Measure in ARDS diagnosis, has brought much guidance and hope for the treatment of previously untreatable diseases.^[12–14] To satisfy the early-diagnosis requirement, there have been many attempts to develop a new method using deep learning-based analysis.^[15] ARDS affects the prognosis of patients with sepsis, so clinically accurate diagnoses for septic ARDS is undoubtedly the most important application of machine learning.

In this study, we first performed differential expression analysis between sepsis and septic ARDS samples, used weighted gene co-expression network analysis (WGCNA) to correlate sepsis and ARDS samples to identify gene modules associated with ARDS, and the features extracted by least absolute shrinkage and selection operator (LASSO) regression, support vector machine recursive feature elimination (SVM-RFE), and the random forest (RF) algorithm were combined to assess their ability to diagnose septic ARDS. We validated the expression patterns of hub genes and evaluated the diagnostic efficiency of the constructed nomogram. Finally, we explored the immune cell signatures of septic ARDS to uncover the association of the hub genes with the immunological landscape. In addition, we characterized the relationship between immune cell infiltration and potential diagnostic biomarkers of septic ARDS and further validated the results using single-cell transcriptome data.

2. Materials and methods

2.1. Data collection and download

ARDS gene expression datasets containing GSE66890 and GSE32707 were downloaded from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/geo/). The GSE66890 dataset (28 sepsis samples and 29 sepsis-related ARDS samples) was derived from the GPL6244 platform of [HuGene-1_0-st] Affymetrix Human Gene 1.0 ST Array. The GSE32707 dataset (28 sepsis samples and 13 sepsis-related ARDS samples) was derived from the GPL10558 platform of the Illumina HumanHT-12 V4.0 expression beadchip.

2.2. Differentially expressed gene (DEG) analysis

Background correction, normalization, and gene symbol conversion were performed on the septic ARDS dataset (GSE66890). GSE66890 datasets were run through the R software packages “limma” and “sva,” the batch effect was removed by combat function, and the data was adjusted by the “FDR” method as the training set; GSE32707 was used as the validation dataset to confirm the analysis results. DEGs in the septic ARDS dataset were screened with the thresholds of P value <0.05 and |log₂(fold change)| > 0.5. Subsequently, the expression patterns of DEGs were visualized in the form of volcano plots and heatmaps with the “ggplot2” package and the “pheatmap” package in R software, respectively.

2.3. WGCNA and key module gene identification

The gene expression matrix of the GSE66890 dataset was used as input data for WGCNA, and genes with small fluctuations in all samples were removed from the data used in gene coexpression network analysis with the WGCNA package in R. The network construction function of the “WGCNA” package was employed to construct a scale-free coexpression gene network. Meanwhile, the appropriate soft threshold power (β = 6) was taken as the weight value in this experiment. After obtaining the modules, the different module eigengenes (MEs) were obtained based on the first principal component of the module expression, while the module-trait relationships were evaluated in line with the association between MEs and clinical characteristics. Modules were screened based on the thresholds of moduleTraitPvalue < 0.05 and |moduleTraitCor| > 0.3. Among them, the modules with the most significant positive and negative correlations of module-trait relationships were selected. Then, the module membership and gene significance scores in modules were also evaluated to determine the module significance.

2.4. Functional enrichment analysis

The Metascape (http://metascape.org) database allows for enrichment analysis using Gene Ontology (GO) to further understand the potential biological significance of the common genes in the “key module genes of ARDS from WGCNA and sepsis-ARDS DEGs.” Enrichment analysis was performed using the hallmark gene set of the Molecular Signature Database and the Kyoto Encyclopedia of Genes and Genomes pathway. Gene set enrichment analysis (GSEA) was used to analyze DEGs, and the standard was a P value < .05, using the “GSEABase” and “cluster Profiler” packages.

2.5. Identification of candidate small molecules

Connectivity map (cMAP) (https://clue.io)^[16] is a gene expression profile (GEP) database based on gene expression signatures that reveals the relationships between diseases, genes, and small molecule compounds. In this study, the common genes with the most significant positive and negative correlations in the WGCNA module-trait relationship and DEGs were run through the cMAP online database to identify potential septic ARDS small molecule drug treatments. Finally, the top 10 compounds with the highest enrichment scores were identified.

2.6. Molecular docking verification

We performed molecular docking verification of the small molecule drugs predicted in cMAP and the potential target proteins of ARDS and judged the reliability of the drug treatment of ARDS by the value of the binding energy. Mol2 file format structures of the compounds were obtained from the PubChem database, and the crystal structures of the core targets were collected from the Research Collaboratory for Structural Bioinformatics Protein Data Bank (http://www.rcsb.org/). First, target proteins were dehydrated and ligand-removed using PyMOL 2.3.0 software (Schrödinger, LLC, New York) and stored in Protein Data Bank format. The processed target protein was then imported into AutoDock Tools 1.5.6 software (The Scripps Research Institute, La Jolla) for hydrogenation and charge calculation and stored in PDBQT format. Mol2 files of small molecule drugs were imported into AutoDock Tools 1.5.6 software, the total charge was determined, the charge was assigned, and the flexible rotatable bonds were viewed and saved in PDBQT format. Grid box data for the proteins of interest were obtained. Finally, AutoDock Vina 1.1.2 software (The Scripps Research Institute, La Jolla) was run for molecular docking. Molecular docking results were visualized using PyMOL 2.3.0 software.

2.7. Machine learning

Three machine learning algorithms, the LASSO algorithm, the SVM-RFE algorithm and the RF algorithm, were used in this study to screen for significant diagnostic or prognostic variables. LASSO was performed using the “glmnet” package in R and is a regression analysis algorithm that applies regularization for variable selection.^[17] SVM-RFE is a widely used supervised machine learning protocol for classification and regression, which uses the “e1071” package. The SVM-RFE algorithm was used to identify genes with high discriminant ability.^[18] The RF algorithm integrates multiple trees through ensemble learning for better accuracy and uses the “RF” package to narrow down the candidate biomarkers.^[19] The overlapping genes obtained by the 3 machine learning algorithms were defined as the key genes for establishing sepsis-related ARDS diagnostic models.

2.8. Construction of the nomogram and assessment of the diagnostic marker prediction model

The nomogram was constructed based on the 3 hub genes by using the “rms” package.^[20] The area under the receiver operating characteristic (ROC) curve was drawn to evaluate the performance of each hub gene and the nomogram in the diagnosis of septic ARDS. Furthermore, an ROC curve was generated to determine whether the nomogram-based decision was conducive to the diagnosis of aortic valve sclerosis. Finally, calibration curves and decision curve analysis (DCA) were carried out to assess the nomogram predictive efficiency in septic ARDS.

2.9. Immune infiltration analysis

Normalized septic ARDS GEP data were compared with the gene set using “GSVA” (R package). Single-sample gene set enrichment analysis classifies gene sets with common biological functions, chromosomal localization, and physiological regulation.^[21] The gene sets included 782 genes for predicting the abundance of 28 tumor-infiltrating immune cells in individual tissue samples. The normalized septic ARDS GEP data were compared with the gene set to evaluate the enrichment of the 28 tumor-infiltrating immune cells in the blood of patients with septic ARDS.

2.10. Single-cell transcriptome data processing and analysis

GSE151263 raw data were downloaded from the GEO database, and the dataset (4 sepsis samples and 3 sepsis-related ARDS samples) was derived from the GPL20301 Illumina HiSeq 4000 platform (Homo sapiens). In the process of single-cell transcriptome data processing, we carried out the normalization, scaling, and clustering of cells and identified 10 main cell types in R v4.2.2 using Seurat v4.3.0. The single-cell extraction standard was nFeature_RNA in the range of 200 to 3000, and percent.mt < 20% was performed to remove double and dead cells. We then normalized the filtered gene barcode matrix using the “NormalizeData” function. The top 2000 highly variable genes were found using the “vst” method via the “FindVariableFeatures” function; the highly variable genes were previously centered and scaled using “ScaleData.” We then performed a principal component analysis based on these 2000 highly variable genes and reduced dimensionality using the Harmony package to remove the batch effect; then, Seurat’s “FindNeighbors” and the “FindClusters” and “runTSNE” functions were used to display dimensionally reduced clusters on a 2D map generated by t-distributed t-SNE. The Kruskal–Wallis test was used to estimate differences in gene expression levels.

3. Results

3.1. Screening of DEGs in sepsis and septic ARDS

The bioinformatics analysis strategy is shown in Figure 1A. Data on sepsis ARDS samples were collected from GSE66890, including 29 specified samples in the septic ARDS group and 28 control samples in the sepsis group. Differential analysis of the sepsis and septic ARDS samples revealed 102 DEGs, containing 59 upregulated and 43 downregulated genes. Volcano plots and DEG heatmaps were utilized to depict the expression pattern of DEGs (upregulated and downregulated genes each displaying the top 6) in the septic ARDS dataset (Fig. 1B and C).

Figure 1.

Screening of DEGs in sepsis and septic ARDS. (A) Flowchart of the study. (B) Volcano plot of DEGs constructed using the fold change values and P-value; red dots represent upregulated differential genes, gray dots represent nonsignificant genes, and blue dots represent downregulated differential genes. (C) Heatmap plot of DEGs (upregulated and downregulated genes each displaying the top 6). Different colors represent the trend of gene expression in different tissues. ARDS = acute respiratory distress syndrome, cMAP = connectivity map, DCA = decision curve analysis, DEGs = differentially expressed genes, LASSO = least absolute shrinkage and selection operator, PBMC = peripheral blood mononuclear cell, ROC = receiver operating characteristic, SVM-RFE = support vector machine recursive feature elimination, WGCNA = weighted gene co-expression network analysis.

3.2. Construction of a weighted gene coexpression network and identification of key modules in ARDS

To further explore the key genes responsible for septic ARDS in sepsis, we used WGCNA to identify the gene modules associated with septic ARDS in sepsis and ARDS septic samples. According to the scale independence and average connectivity, the soft threshold power was selected as 6 (Fig. 2A). In total, 15 modules were generated using that power, and the cluster dendrogram of the modules is presented in Figure 2B. The clustering of ME is displayed in Figure 2C.

Figure 2.

Construction of a weighted gene coexpression network and identification of key modules in ARDS. (A) The scale‐free topology model was utilized to identify the best β value, and β = 6 was chosen as the soft threshold based on the average connectivity and scale independence. (B) The network heatmap showing the gene dendrogram and module eigengenes. (C) The cluster dendrogram presenting module eigengenes. (D) The heatmap revealing the relationship between module eigengenes and status of septic ARDS. The correlation (upper) and P‐value (bottom) of module eigengenes and status of ARDS were presented. The salmon and brown modules correlated with ARDS exhibited the highest and lowest correlation coefficients, respectively, which were identified as the key modules in ARDS. (E) The correlation plot between the salmon (or brown) module membership and the gene significance of genes in the salmon (or brown) module. (F) A total of 258 key genes in ARDS were identified by taking the intersection between key modules genes and DEGs via the venn diagram. ARDS = acute respiratory distress syndrome, DEGs = differentially expressed genes.

Furthermore, this study explored the correlation between septic ARDS and gene modules (Fig. 2D). These data showed that the salmon module had the highest positive correlation with ARDS (53 genes, R = 0.40, P = .006), whereas the brown module had the most negative correlation with ARDS (205 genes, R = −0.38, P = .008). On this basis, the salmon and brown modules were considered the key modules for subsequent analysis. Moreover, we found a strong association between module membership and gene significance in the salmon (R = 0.64, P = 3.5e−07) and brown modules (R = 0.44, P = 4.1e−11) (Fig. 2E). Therefore, 258 crucial genes that were significantly associated with ARDS were identified in the salmon and brown modules. In addition, we further screened for common genes in the DEG dataset and crucial genes from the WGCNA of septic ARDS samples to identify the key genes in ARDS; a total of 49 hub genes were obtained and subjected to further analysis (Fig. 2F).

3.3. Functional analysis of potential genes

GO and pathway analyses were performed with the Metascape database to determine the biological function of potential genes. Potential genes were mainly involved in the positive regulation of the inflammatory response, reactive oxygen biosynthesis, or cell migration and the negative regulation of immune system processes or intracellular signal transduction (Fig. 3A). In addition, these genes were mainly enriched in the regulation of leukocyte and macrophage activation and tumor necrosis factor production. Figure 3B shows the relationships between the enriched terms. Table 1 lists the top 10 representative clusters of enrichment terms.

Figure 3.

Functional enrichment analysis of DEGs. (A) Bar plot of DEGs functional enrichment terms. (B) Network relationship plots among all enriched terms. Colored by P-value, where terms containing more genes tend to have a more significant P-value. GSEA enrichment analysis results in (C) ARDS samples and (D) sepsis samples. ARDS = acute respiratory distress syndrome, DEGs = differentially expressed genes, GSEA = gene set enrichment analysis.

Table 1.

Top 10 clusters with their representative enriched terms (one per cluster).

GO	Description	Count	%	Log10 (P)	Log10 (q)
GO:0050729	Positive regulation of inflammatory response	6	15	−6.3	−2.2
GO:0032680	Regulation of tumor necrosis factor production	6	15	−5.8	−2.1
GO:1903555	Regulation of tumor necrosis factor superfamily cytokine production	6	15	−5.8	−2.1
GO:0043030	Regulation of macrophage activation	4	9.8	−5.2	−1.7
GO:0032103	Positive regulation of response to external stimulus	8	20	−4.9	−1.5
GO:1903428	Positive regulation of reactive oxygen species biosynthetic process	3	7.3	−4.8	−1.5
GO:0031349	Positive regulation of defense response	7	17	−4.8	−1.5
GO:0043032	Positive regulation of macrophage activation	3	7.3	−4.5	−1.3
GO:0002694	Regulation of leukocyte activation	8	20	−4.5	−1.3
GO:0050865	Regulation of cell activation	8	20	−4.3	−1.1

Moreover, GSEA results showed that chemokines, Leishmania infection, natural killer (NK) cell-mediated cytotoxicity, taste transduction, and Toll-like receptors were mainly enriched in sepsis samples (Fig. 3C). The cell cycle, glycolytic gluconeogenesis, oocyte meiosis, proteasome, systemic lupus, and erythematosus were enriched in septic ARDS samples (Fig. 3D).

3.4. Identification of candidate small molecule compounds for ARDS treatment

To further investigate the potential small molecule drugs that might have a therapeutic effect in sepsis-related ARDS patients, 49 hub genes in ARDS samples from sepsis-related pathogenic genes were imported into the cMAP database to predict small molecule compounds that could reverse the altered expression of sepsis-related pathogenic genes in ARDS. Following this inquiry, the top 10 compounds, including JAK3-inhibitor-VI, CD-437, THM-I-94, pyrvinium-pamoate, puromycin, SCH-79797, etoposide, amacrine, digoxin, and mitomycin-c, with the highest negative scores were considered potential pharmacological therapeutic agents for the treatment of septic ARDS (Fig. 4A). The descriptions of the target pathways and chemical structures of these 10 compounds are displayed in Figure 4B.

Figure 4.

Screening of the potential small‐molecular compounds for the treatment of septic ARDS via cMAP analysis. (A) The heatmap presenting the top 10 compounds with the most significantly negative enrichment scores in 10 cell lines based on cMAP analysis, and descriptions of the top 10 compounds. (B) The chemical structures of those 10 compounds were shown. ARDS = acute respiratory distress syndrome, cMAP = connectivity map.

3.5. Molecular docking verification

Among the 3 candidate hub genes, retinol binding protein 7 (RBP7) is an important anti-inflammatory and antioxidant molecule, and its expression was decreased in the ARDS group; thus, this gene is more likely to be used as a therapeutic target.^[22,23] Using AutoDock Vina 1.1.2 software, the screened small molecule drugs were docked to the core target RBP7. Studies have shown that the lower the binding energy is, the more stable the binding conformation and the greater the likelihood of binding.^[24]

As shown in Figure 5A, the minimum binding energy between the ligand and the receptor was mostly less than −8.5 kcal·mol⁻¹, indicating that the target protein has a good affinity for the active ingredient, and the small molecule drugs are likely to act on the target RBP7. Small molecule drug docking targets with the lowest binding energy were selected for docking visualization (Fig. 5B). The dotted lines in the figure are hydrogen bonds. For example, amacrine exerts its biological efficacy most likely by binding to RBP7 and forming hydrogen bonds with the 2 amino acids Asn103 and Arg104 near the active site.

Figure 5.

Docking diagram of small molecule drugs with targets. (A) Line diagram of the lowest binding energy for molecular docking. (B) Docking diagram of JAK3-inhibitor-VI, CD-437, THM-I-94, pyrvinium-pamoate, puromycin, SCH-79797, etoposide, amacrine, digoxin, and mitomycin-c docked to RBP7, respectively. RBP7 = retinol binding protein 7.

3.6. Screening of hub genes with diagnostic value via machine learning

Since the common DEGs between septic ARDS and sepsis samples may play a key role in sepsis-related ARDS, 49 common genes were found to be both ARDS key modules and sepsis-ARDS DEGs; these genes were included in the subsequent construction of the ARDS diagnostic model. The LASSO regression algorithm was applied to identify 16 potential candidate genes from 49 hub genes that have a significant impact on the diagnosis of sepsis-related ARDS patients (Fig. 6A). The RF machine learning algorithm was also used to identify 49 common genes according to the variable importance of each gene with a MeanDecreaseGini > 2 (Fig. 6B). To further narrow down the range of diagnostic biomarkers, the SVM-RFE algorithm was used to classify 16 features among the 49 hub genes and identify a subset of 28 significant features (Fig. 6C). Interestingly, after comparing the 5 potential genes from Lasso and RF and the 28 significant genes from SVM-RFE, only 3 hub genes overlapped in 3 subsets, including RBP7, ATPase plasma membrane Ca²⁺ transporting 1 (ATP2B1), and absent in melanoma 2 (AIM2) (Fig. 6D).

Figure 6.

Screening of hub genes with diagnostic value via machine learning. (A) Sixteen diagnostic markers were screened by the least absolute shrinkage and selection operator (LASSO) logistic regression algorithm. (B) Five diagnostic markers were screened by the random forest (RF) algorithm. (C) Twenty-eight diagnostic markers were screened by a support vector machine recursive feature elimination (SVM-RFE) algorithm. (D) Venn diagram of 3 variables including RBP7, ATP2B1, and AIM2 intersected by LASSO, RF, and SVM-RFE algorithms. AIM2 = absent in melanoma 2, ATP2B1 = ATPase plasma membrane Ca²⁺ transporting 1, RBP7 = retinol binding protein 7.

3.7. Construction and validation of a diagnostic model for septic ARDS

For better diagnosis and prediction, a nomogram was constructed on the basis of 2 hub genes by logistic regression analysis (Fig. 7A). ROC curves were used to evaluate the values of each hub gene, and the area under the nomogram curve (AUC) was used to determine the sensitivity and specificity of these genes in the diagnosis of septic ARDS. As expected, the AUC values of the 3 hub genes were above 0.75, and the nomogram AUC values were higher than those of each hub gene, suggesting that the nomogram may have a strong diagnostic value for septic ARDS (Fig. 7B). The correction curve shows that the prediction probability of the constructed modal graph diagnosis model is almost the same as that of the ideal model (Fig. 7C). In addition, we performed a nomogram DCA, and the results showed that decision making based on the nomogram model may be helpful in the diagnosis of septic ARDS (Fig. 7D). The clinical impact curve on the ground of the DCA curve was further plotted to evaluate the clinical effects of the nomogram more intuitively. The “number high risk” curve was close to the “number high risk with event” curve at a high-risk threshold from 0.2 to 1, which demonstrated that the nomogram had powerful predictive ability (Fig. 7E). Sepsis is the main cause of ARDS. The nomogram also demonstrated an ideal predictive value among sepsis patients with ARDS in the GSE32707 dataset from the GEO database, which includes blood samples from 18 patients with ARDS and 30 patients with sepsis (Fig. 7F), implying that the nomogram model exhibited good diagnostic efficacy for early ARDS sepsis patients.

Figure 7.

Construction and validation of a diagnostic model for septic ARDS. (A) The nomogram of diagnostic biomarkers to predict the occurrence of septic ARDS. (B) The ROC curve for the diagnostic performance of each candidate biomarker including ATP2B1, RBP7, and AIM2, and the nomogram model constructed for sepsis‐related ARDS. (C) The calibration curve to assess the predictive power of the nomogram model. (D) The DCA curve to evaluate the clinical application value of nomogram model. (E) Clinical impact curves of the nomogram model. (F) The ROC curve for the diagnostic performance of our nomogram model in predicting patients with septic ARDS in the GSE32707 dataset. AIM2 = absent in melanoma 2, ARDS = acute respiratory distress syndrome, ATP2B1 = ATPase plasma membrane Ca²⁺ transporting 1, DCA = decision curve analysis, RBP7 = retinol binding protein 7, ROC = receiver operating characteristic.

3.8. Immune cell infiltration and correlation of hub genes with invading immune cells in septic ARDS

We found that the function and pathway analysis of sepsis-related pathogenic ARDS genes show that they are closely related to inflammation and immune processes. Single-sample gene set enrichment analysis was used to determine the characteristics of immune cells and investigate the correlation between immune regulation, diagnostic biomarkers, and immune cell infiltration in septic ARDS. A heatmap was created to describe the patterns between immune cell landscapes in the sepsis microenvironment. Figure 8A shows the proportion of 28 types of immune cells in each sample, with significant differences in 5 subpopulations of immune cells when comparing the septic ARDS sample and the sepsis group. Compared with those in the sepsis group, the proportions of neutrophils, myeloid-derived suppressor cells (MDSCs), macrophages, CD56-bright NK cells and eosinophils in the septic ARDS group were lower (Fig. 8B). In addition, correlation analysis of 28 immune cells showed that monocytes were significantly positively correlated with macrophages, MDSCs, neutrophils, CD56-bright NK cells, and central memory CD8 T cells and negatively correlated with type 2 T helper cells and activated CD4 T cells (Fig. 8C). In addition, we further investigated the relationship between the expression of the 3 hub genes and the proportion of different infiltrating immune cell types. As shown in Figure 8D, the central genes ATP2B1, RBP7, and AIM2 were significantly correlated with the accumulation of immune cells in ARDS.

Figure 8.

Immune cell infiltration analysis in septic ARDS. (A) Stacked histogram displaying the immune cell proportions in each sample. (B) Violin plot showing the comparison of 28 kinds of immune cells between septic ARDS and sepsis groups. Red and blue represent ARDS group and sepsis group, respectively, with statistical P-values expressed at the top. (C) The heatmap revealing the correlation of 28 kinds of immune cells infiltration upon the threshold of P < .05. (D) The correlation map representing the association of the differentially infiltrated immune cells with 3 hub genes upon the threshold of P < .05. ARDS = acute respiratory distress syndrome.

3.9. Expression levels of ATP2B1, RBP7, and AIM2 in peripheral blood mononuclear cell (PBMC) single-cell transcriptome data

Although there was no significant difference in monocytes between the 2 groups, a previous study reported that monocytes play an important role in the progression of septic ARDS.^[25] Due to the small number of mononuclear cells in peripheral blood, it is difficult to obtain objective statistical evidence using conventional chip analysis. The PBMC single-cell sequencing dataset GSE151263 was downloaded from the NCBI GEO database for normalization, scaling, clustering, and high-variant gene screening. The dimensionality reduction cluster was then displayed on a 2D plot generated based on the T-distribution t-SNE (principal component analysis) of these 2000 highly variable genes (Fig. 9A). ATP2B1, RBP7, and AIM2 are shown in the cell cluster expression diagram (Fig. 9B and C). Interestingly, the most important PBMC types, CD14 monocytes (ATP2B1, RBP7), CD16 monocytes (ATP2B1, RBP7), and B cells (ATP2B1, AIM2), were expressed at the highest levels among the 9 major cell types. The Kruskal–Wallis test showed that the expression of RBP7 and ATP2B1 in CD14M and CD16M cells was significantly upregulated (Fig. 9C). To validate ATP2B1, RBP7, and AIM2 as septic ARDS processing genes, based on their increasingly reduced expression, we divided the single-cell transcriptomic data into roughly 2 groups, consistent with previous analyses. Compared with that in the sepsis group, the expression of ATP2B1 and RBP7 in the septic ARDS group was significantly decreased, but that of AIM2 was not (Fig. 9D).

Figure 9.

Hub genes in PBMC single-cell transcriptome data. (A) The t-distributed stochastic neighbor embedding (t-SNE) plot of the 10 identified main cell types among single-cell transcriptome dataset GSE151263. (B) t-SNE map highlighting the expression of genes ATP2B1, AIM2, and RBP7. (C) Bubble plot showing the expression of the ATP2B1, AIM2, and RBP7 related different cell types. The size of each dot represents the percent expressed; average expression is shown by color. (D) Single-cell transcriptome data revealed differences in the expression of 3 hub genes in mononuclear cells. AIM2 = absent in melanoma 2, ATP2B1 = ATPase plasma membrane Ca²⁺ transporting 1, PBMC = peripheral blood mononuclear cell, RBP7 = retinol binding protein 7.

4. Discussion

ARDS is a common clinical problem affecting the prognosis of critically ill patients, and sepsis and infectious factors account for 78.5% of ARDS etiology.^[26,27] Once sepsis is combined with ARDS, the mortality rate more than doubles.^[28] Therefore, exploring marker genes for the development of septic ARDS will undoubtedly provide a broader clinical picture for ARDS diagnosis and treatment. Based on this, we compared sepsis and septic ARDS samples to search for more candidate diagnostic biomarkers and variations among immune cells during ARDS.

This study identified sepsis-related causative genes and elucidated the association between sepsis and subsequent ARDS by applying comprehensive bioinformatics analysis approaches. It was found that inflammatory and immune processes, together with signaling pathways, including “cell cycle,” “glycolytic gluconeogenesis,” and “proteasome,” might be part of the potential mechanism underlying sepsis-related ARDS. Moreover, 3 immune-related hub genes, ATP2B1, RBP7, and AIM2, were employed to develop a diagnostic nomogram model to predict the risk of ARDS by machine learning approaches. According to our results, these 3 hub genes displayed ideal predictive performance for ARDS, as assessed by the ROC curve. Finally, through the external validation of the GSE32707 dataset, the expression patterns of ATP2B1, RBP7, and AIM2 were consistent with the obtained datasets, and the diagnostic nomogram models based on ATP2B1, RBP7, and AIM2 levels performed well in differentiating ARDS sepsis and sepsis patients.

Alveolar epithelial injury and vascular endothelial injury are the basic pathological changes in ARDS and are driven by the abnormal regulation of immune cells. In our previous study, we found that dendritic cells in the early stage of ARDS accumulate and mature in the lung, which may be an important cause of ARDS.^[29] Zhang et al found that activating adenosine monophosphate-activated protein kinase in macrophages or neutralizing high mobility group box 1 in bronchoalveolar lavage fluid could improve efferocytosis and neutrophil extracellular trap (NET) clearance.^[30] In LPS-induced ARDS mice, the infiltration of neutrophils and monocytes into the lung aggravated lung injury and increased the amount of reactive oxygen species.^[7] Therefore, the immune inflammatory response is the key link leading to septic ARDS. In this study, we found that genes causative of sepsis-associated ARDS were mainly enriched in inflammatory and immune-related pathways by GO and pathway analyses and GSEA, suggesting that the inflammatory-immune pathway may be a potential pathogenetic mechanism of sepsis-associated ARDS.

Effective pharmacotherapy for the treatment of ARDS is still lacking, so there is an urgent need to explore potential drugs. Over the past few years, many important breakthroughs have been made in identifying small molecule compounds that have potential to treat a variety of diseases.^[31] Small molecule compounds have the advantages of high tissue penetration, adjustable half-life, and oral bioavailability, leading to better multitherapeutic effects.^[32] β-Nitrostyrene derivatives (BNSDs) elicit protective effects on LPS-induced paw edema and acute lung injury via the inhibition of neutrophil accumulation, pro-inflammatory mediator release, platelet aggregation, myeloperoxidase activity, and NET release.^[33] Protein arginine methyltransferase 4 inhibition with a small molecule compound attenuated lymphocyte death in complementary models of sepsis.^[34] However, no previous studies have identified potential small molecule compounds based on septic ARDS gene expression features for therapeutic application through high-throughput screening. This study provides a new perspective on linking genes associated with sepsis through cMAP analysis to identify different compounds targeting ARDS. The pathogenic genes associated with septic ARDS were analyzed by cMAP, and 10 small molecule compounds (JAK3-inhibitor-VI, CD-437, THM-I-94, pyrvinium-pamoate, puromycin, SCH-79797, etoposide, amacrine, digoxin, and mitomycin-c) were selected as candidate compounds. Notably, in the cMAP analysis, JAK3-inhibitor-VI showed the highest negative enrichment score, which means that it maximally reversed the expression of sepsis-related teratogenic genes in ARDS. Bhavsar et al reported that JAK3 upregulates the activity of sodium-glucose cotransporter 1 by increasing the abundance of carrier proteins in the cell membrane, thereby promoting glucose uptake by cells and transfer to activated lymphocytes, thus promoting the immune response, while JAK3-inhibitor-VI inhibitors impede this process.^[35] Therefore, we speculate that the early application of a JAK3-inhibitor-VI may inhibit the occurrence and development of septic ARDS by preventing an excessive immune response.

Once patients with sepsis are complicated with ARDS, their mortality risk and other prognostic indicators become significantly worse.^[2] To date, there is still a lack of biomarkers for the early prediction of ARDS, and the diagnosis of ARDS still mainly relies on the functional index of the PaO₂/FiO₂ ratio or the respiratory support parameter positive end-expiratory pressure. Although a new consensus has been established in recent years,^[3] it has not changed this predicament. At present, there is still a lack of routine serum biomarkers for the early prediction of ARDS, and the diagnosis of ARDS still mainly relies on functional indicators such as the PaO₂/FiO₂ ratio. This study included the development of a more comprehensive diagnostic nomogram model based on 2 pivotal genes, which had a higher diagnostic value for sepsis-associated ARDS than independent biomarkers. Furthermore, the nomogram model was effective in diagnosing patients with ARDS, suggesting that this diagnostic nomogram model is also effective in predicting ARDS in the early stages. In addition, validation from the GSE32707 dataset showed decreased levels of ATP2B1, RBP7, and AIM2 mRNA in the blood of the ARDS group compared to that of the sepsis group. Our constructed diagnostic nomogram was capable of significantly distinguishing between sepsis and septic ARDS samples.

The change in hub gene expression may cause dysfunctional immune regulation. ATP2B1 has been reported to promote the immune response.^[36,37] The upregulation of ATP2B1 expression in immune cold tumors increases the level of immune cell infiltration, further activating immune signals and inducing the immune response.^[36] In this study, the expression of ATP2B1 in the ARDS group was decreased, which may reflect immunosuppression in the process of reorganization. In addition, the decreased expression of the ATP2B1 gene was related to impaired NOS activity and nitric oxide production in endothelial cells,^[37] which may be the mechanism underlying the disruption of the gas-blood barrier in ARDS. Peroxisome proliferator-activated receptor gamma (PPARγ) also protects vascular endothelial cells by increasing nitric oxide bioavailability and preventing oxidative stress. RBP7 is a PPARγ target gene enriched in vascular endothelial cells, which may be a target for PPARγ to exert its cardioprotective effect.^[38] RBP7 is involved in a feed-forward mechanism facilitating PPARγ activity in endothelial cells in response to pro-oxidant stressors, which presumably induce production of a PPARγ ligand, further leading to the induction of a subset of PPARγ target genes that provide antioxidant protection. Loss of RBP7 selectively impairs PPARγ activity, leading to oxidative stress and endothelial dysfunction.^[22,23] In addition, promoter methylation of RBP7 is involved in its gene silencing in breast cancer, thus regulating the occurrence and development of estrogen receptor-positive breast cancer through the PPAR and PI3K/AKT pathways.^[39] The decreased expression of RBP7 is thought to be an important mechanism of endothelial injury leading to the breakdown of the blood gas barrier and may be a potential marker of sepsis complicated with ARDS. LPS causes NETs to induce alveolar macrophage pyroptosis, and silencing of the AIM2 gene protects against alveolar macrophage pyroptosis.^[40] Although AIM2 is generally accepted as an inflammasome effector in myeloid cells, Chou et al’s findings demonstrate a T-cell-intrinsic role of AIM2 in restraining autoimmunity by reducing AKT-mTOR signaling and altering immune metabolism to enhance the stability of Treg cells.^[41] Therefore, AIM2 may be an important biomarker of ARDS in sepsis. In conclusion, these findings all suggest that downregulation of ATP2B1, RBP7, and AIM2 may lead to immunosuppression.

Relative immunosuppression may be related to ARDS caused by sepsis. In the analysis of immune cell infiltration, the accumulation of various types of immune cells has been confirmed to exist in patients with ARDS of different etiologies, and there are also differences in septic and nonseptic ARDS.^[42] Neutrophil influx is a hallmark of ARDS and is associated with the release of tissue-destructive immune effectors. However, in this study, significant differences in the infiltration of immune cells were identified between the sepsis and septic ARDS groups, with lower abundances of neutrophils, macrophages, MDSCs, and eosinophils and higher abundances of CD56-bright NK cells in the septic ARDS group. It has been well established that monocytes, macrophages, and neutrophils are essential for host control of infection and that the presence of leukopenia and granulocytopenia is an independent risk factor for mortality in sepsis.^[43,44] Selective depletion of alveolar macrophages in multi-microbial sepsis increases lung injury and mortality.^[45] Neutrophil apoptosis was closely related to lung injury in a mouse model of sepsis.^[8] A growing body of data suggests that eosinophils play a protective role in the innate immune response to sepsis, circulating in the blood as mature cells that can be rapidly absorbed into infected tissues via pattern recognition receptors, pro-inflammatory cytokines, antibacterial proteins, and DNA traps to fight pathogens and promote an effective immune response.^[46] However, we did not observe the expansion of MDSCs in immunosuppressive cells in the sepsis-ARDS group. According to previous reports, the expansion of MDSCs associated with COVID-19 is directly related to lymphocytopenia, and lymphocytes in this group did not greatly differ between the 2 groups, which may be the reason for the nonsignificant expansion of MDSCs.^[47] Monocytes have been reported to play a central role in the pathogenesis of sepsis through immunomonitoring and treatment.^[48] In PBMC single-cell sequencing data, the ATP2B1, RBP7, and AIM2 genes were found among 2000 highly variable genes, and further analysis showed that they were highly expressed in CD14+ and CD16+ monocytes or B cells. We found that the proportions of neutrophils and eosinophils reflecting immune activation in the ARDS group were significantly reduced, and these immune cells were positively correlated with ATP2B1, RBP7, and AIM2. We speculate that altering the expression of these hub genes promotes the suppression of immune function, which may be an important mechanism of the occurrence of septic ARDS.

5. Conclusion

In summary, by comparing sepsis and septic ARDS samples, we found many different genes involved in the inflammation pathway and the potential diagnostic markers ATP2B1, RBP7, and AIM2, thereby providing more options for clinical diagnosis. Meanwhile, when sepsis progressed to septic ARDS, immune cells such as neutrophils and macrophages and hub genes such as ATP2B1, RBP7, and AIM2, all moved toward immunosuppressive phenotypes; these results provide new insights into sepsis-associated ARDS.

Acknowledgments

The authors thank Dr Tianfeng Wang from the University of Science and Technology of China for his guidance on this article, and AJE (part of Springer nature) for providing native language revisions to this article.

Author contributions

Conceptualization: Zhonghua Lu.

Data curation: Hu Chen, Mei Liu, Lu Fu.

Formal analysis: Tongyue Du, Ping Li.

Investigation: Yan Tang, Weili Yu.

Methodology: Yun Sun.

Resources: Yuli Cao.

Software: Mei Liu.

Supervision: Tongyue Du, Zhonghua Lu.

Validation: Lu Fu.

Visualization: Yun Sun.

Writing – original draft: Hu Chen.