Transcriptome and Experimental Verification Identified Candidate Biomarkers Related to Mitochondrial Metabolism in Sepsis-Associated Encephalopathy

Abstract

Background:

Mitochondrial metabolism (MM) abnormalities have been implicated in multiple diseases, but their contribution to sepsis-associated encephalopathy (SAE) remains insufficiently understood. The purpose of this research was to explore the candidate biomarkers associated with MM in SAE and the underlying mechanisms.

Methods:

The relevant transcriptome data were acquired from the public databases. MM-related genes were scoured from published articles. Candidate biomarker identification integrated differential expression profiling, protein-protein interaction network construction, machine learning algorithms, receiver operating characteristic curve evaluation, and expression quantification. Immune infiltration, multilayer perceptron network, Gene Set Enrichment Analysis, molecular regulatory network, drug prediction, and molecular docking were applied to probe the mechanisms of candidate biomarkers in SAE. The expression levels of candidate biomarkers were further validated in the brain tissues of the lipopolysaccharides -induced SAE mouse model.

Results:

Four MM-RGs-INSIG1, SREBF1, CIDEC, and PNPLA3-were identified as candidate biomarkers, and their expression patterns in the rodent model were consistent with bioinformatics predictions. Within the GSE135838 dataset, the multilayer perceptron model demonstrated good performance in distinguishing SAE from control samples. Differential immune cells included resting mast cells, naive B cells, and activated natural killer cells. Gene Set Enrichment Analysis indicated significant enrichment in the allograft rejection pathway. The regulatory network involved transcription factors (e.g., THRB), microRNAs (miRNAs) (e.g., hsa-mir-29c-3p), and long noncoding RNAs (e.g., KCNQ1OT1). Furthermore, experimental validation confirmed that these candidate biomarkers were significantly downregulated in SAE samples.

Conclusion:

This study identified and experimentally validated four MM-related candidate biomarkers in SAE, revealing their potential roles in immune dysregulation. These candidate biomarkers require further validation in independent cohorts and in a more diverse range of animal models before clinical translation.

BACKGROUND

Sepsis is life-threatening organ dysfunction caused by a dysregulated host response to infection and systemic inflammation (1). As a critical healthcare challenge, sepsis and septic shock affect millions globally each year. It ranks among the leading causes of intensive care unit (ICU) admissions, accounting for 24.7% of such cases, and imposes considerable physiological, social, and economic burdens on patients and their families (2,3). The clinical manifestations and symptoms of sepsis are heterogeneous and closely associated with the specific organ systems affected and the severity of complications (3,4). Sepsis-associated encephalopathy (SAE), a life-threatening central nervous system (CNS) complication occurring in 39%–68% of ICU sepsis cases, demonstrates the vulnerability of the CNS to septic insults (5,6). SAE is identified as a key predictor of prolonged hospital stays and elevated mortality risk in septic patients. The pathogenesis of SAE is complex and may involve damage to the blood–brain barrier (7), changes in neurotransmitter release (8), dysregulated iron metabolism (9), overactivation of proinflammatory cytokines, mitochondrial dysfunction, and oxidative stress (10,11). Bustamante et al. performed transcriptomic profiling of parietal cortex gray matter from patients who died of sepsis and generated the GSE135838 dataset. Their analysis revealed a large gene module strongly associated with sepsis, which was enriched in genes involved in cytokine signaling, immune function, and regulation (12). However, the precise mechanisms remain incompletely understood. Studies have reported that patients with a history of sepsis exhibit a significantly increased incidence of seizures (13,14). Septic systemic inflammation may trigger sustained neuroinflammation, ultimately resulting in neurodegeneration and irreversible CNS impairment (15). Therefore, early identification and timely intervention in patients with SAE are essential for improving outcomes and reducing mortality in sepsis (13). Currently, the diagnosis of SAE primarily relies on a combination of clinical, electrophysiological, and neuroimaging approaches. These include clinical assessment tools such as the Glasgow Coma Scale and the Confusion Assessment Method for the ICU, electrophysiological assessments like electroencephalography, and neuroimaging techniques such as magnetic resonance imaging. Nevertheless, there remains an urgent clinical need for more sensitive and specific diagnostic modalities to enable earlier detection and more targeted therapeutic interventions.

Elevated inflammation and oxidative stress are critically associated with SAE progression. Systemic inflammation induces sustained microglial and astrocytic activation in the CNS, driving excessive proinflammatory factors and reactive oxygen species release, resulting in chronic neuroinflammation and neurodegeneration, though precise mechanisms remain incompletely understood (16,17). Mitochondrial dysfunction critically promotes neuroinflammation and oxidative stress (18). During inflammatory responses, activated immune cells like macrophages undergo metabolic transformations, including enhanced glycolysis, tricarboxylic acid cycle remodeling, and increased one-carbon metabolism to meet the biochemical and energy requirements of pathogen defense (19). Generated metabolites, including succinate, promote cytokine release and activate innate immune receptors (Toll-like, NOD-like, and RIG-I-like receptors) and downstream inflammatory pathways (TNFR-NF-κB and NLRP3-caspase-1). Mitochondrial metabolic imbalances may disrupt cellular homeostasis and exacerbate inflammation (19). Proinflammatory factors and reactive oxygen species directly induce mitochondrial DNA damage. As key damage-associated molecular patterns, mitochondrial DNA release amplifies immune responses by activating inflammasomes and promoting oxidative stress (20,21). Recent Alzheimer’s disease research demonstrates that imbalanced astrocytic mitochondrial fatty acid metabolism leads to acetyl-CoA accumulation, STAT3 acetylation, and astrocyte reactivity, enhancing neuronal oxidative stress and microglial activation via interleukin-3 signaling (22). These findings highlight mitochondrial inflammatory pathways’ critical role in immune regulation, suggesting targeting mitochondrial metabolism (MM) may offer promising therapeutic strategies for neuroinflammation.

Transcriptomic analysis conducted by Bustamante et al. confirmed the central role of neuroinflammation and immune activation in SAE. However, their work did not address the potential contribution of mitochondrial dysfunction on neuroinflammation. To further elucidate the mechanism of SAE and to identify a candidate biomarker, the present study employed machine learning methods to screen for MM-related genes (MM-RGs) associated with SAE. The identified candidate biomarkers underwent comprehensive analyses, including multilayer perceptron (MLP) network analysis, immune infiltration profiling, and regulatory factor prediction. Combined with experiments in a mouse model, this study provides preliminary evidence supporting the potential of these MM-related genes to distinguish SAE and explores their underlying molecular mechanisms.

MATERIALS AND METHODS

Data collection

Public SAE transcriptome data were obtained from the Gene Expression Omnibus database (https://www.ncbi.nlm.nih.gov/geo/). The dataset GSE135838 (12) (Sequencing method: High-throughput sequencing), sourced from platform GPL20301, included 12 SAE tissue samples and 12 matched control tissues. In the data preprocessing stage, data from non-protein-coding genes were excluded, and redundant records where multiple genes mapped to a single row were removed. Missing values and duplicate entries were also eliminated. Following this processing, a total of 19,416 genes were retained for subsequent differential expression analysis. The 1,234 MM-RGs were obtained from the reference (23) (Supplementary Table 1, https://links.lww.com/SHK/C853). All subsequent analyses (differential expression, candidate gene screening, machine learning, and MLP network) were performed on this same cohort of 24 samples (12 SAE, 12 controls). No a priori sample size calculation or power analysis was performed, as this study is exploratory and hypothesis-generating in nature, aiming to identify candidate biomarkers that require validation in future independent cohorts.

Analysis of differential expression patterns

We detected differentially expressed genes (DEGs) in GSE135838 by comparing SAE with control samples using DESeq2 (v 1.40.2) (24), applying thresholds of |log₂FC| > 0.5 and P < 0.05. DESeq2 was performed using default parameters. When extracting differentially expressed results, only the comparison group (“SAE” vs. “control”) was specified, without adjusting the significance level (alpha, default value 0.1) or the log2-fold change threshold (lfcThreshold, default value 0). Volcano plots were generated using ggplot2 (v 3.4.1) (25) to visualize the distribution of all DEGs after sorting by log₂FC, highlighting the top 10 up- and down-regulated genes. Meanwhile, Heatmaps were generated using the pheatmap R package (v 1.0.12) (26) to illustrate the expression patterns of the top 10 up- and down-regulated DEGs between SAE and control samples.

Functional analysis and systematic screening of candidate genes

To acquire MM-related DEGs in SAE, namely candidate genes, the intersection between DEGs and MM-RGs was computed utilizing the ggvenn package (v 1.7.1) (27). Moreover, Gene Ontology (GO) (adj.P < 0.05) and Kyoto Encyclopedia of Genes and Genomes (KEGG) (adj.P < 0.05) analyses were conducted to probe the functions and pathways of candidate genes utilizing the clusterProfiler package (v 4.2.2) (28). GO enrichment analysis was performed using the human genome annotation database (org.Hs.eg.db). The analysis employed gene symbols as identifiers and applied the false discovery rate method for multiple testing correction. In KEGG pathway analysis, gene identifiers were first converted from gene symbols to ENTREZ IDs and were then enriched using the species code “hsa.” P values were similarly corrected using the false discovery rate method. The top 5 biological process (BP) terms, the top 5 molecular function (MF) terms, and all cellular component (CC) terms in the GO analysis, and the top 5 significant pathways in the KEGG analysis were presented.

Identification of hub genes and machine learning

To investigate protein interactions among candidate gene products, we submitted the candidate genes to the STRING database (https://string-db.org/; interaction score ≥0.4) and constructed the protein-protein interaction network using Cytoscape (v 3.8.2) (29). Following the removal of isolated targets, the remaining candidate genes were sorted based on their degree of connectivity. The top 10 genes with the highest Degree were designated as hub genes for subsequent analysis. In cases where genes shared an identical Degree of connectivity, all such genes were included in the final hub gene set.

Furthermore, the three types of machine learning algorithms were, respectively, applied to acquire core genes based on the hub genes in all samples of GSE135838. First, the Least Absolute Shrinkage and Selection Operator (LASSO) algorithm was performed on the hub gene expression matrix (SAE = 1, Control = 0) using the glmnet package (v 1.1.10) (30). The model employed L1-regularized logistic regression with the elastic net mixing parameter α set to 1. The optimal regularization strength was determined via 10-fold cross-validation using the cv.glmnet function. Finally, genes with non-zero coefficients (excluding the intercept) were selected as feature genes, using the lambda.min value that minimized the cross-validation error as the criterion. Then, the support vector machine-random feature elimination (SVM-RFE) algorithm was implemented using the e1071 package (v 1.7-16) (31) for feature selection. The process involved iterative evaluation of feature importance through 5-fold cross-validation, with the least important features being successively eliminated. The feature gene was ultimately determined as the feature subset corresponding to the minimum cross-validation error rate. Additionally, the Boruta algorithm was executed using the Boruta R package (v 8.0.0) (32). This algorithm calculated feature importance Z-scores based on random forests and, through comparison with randomly generated shadow features, defined the features classified as “Confirmed” as feature genes. Finally, the feature genes obtained by each algorithm were intersected utilizing the ggvenn package (v 1.7.1), and the genes in the intersected part were outlined as core genes.

Assessments of receiver operating characteristic (ROC) and gene expression levels

Next, we assessed the diagnostic performance of core genes in distinguishing SAE from controls using ROC analysis (GSE135838 cohort) implemented with the pROC package (v 1.18.5) (33). Core genes whose area under the curve (AUC) values go beyond 0.7 were considered to have good diagnostic ability for SAE and were defined as preliminary candidate biomarkers. Additionally, differences in the expression levels of preliminary candidate biomarkers between SAE samples and the control group were evaluated using the Wilcoxon test (with P < 0.05 as the threshold), and genes with significant expression changes were defined as the candidate biomarkers.

Construction of the neural network

To investigate the weight of candidate biomarkers in the diagnosis of SAE, an MLP network with a single hidden layer containing 2 neurons was constructed via the neuralnet package (v 1.44.2) (34) for all samples from the GSE135838 dataset. The model was optimized with the following parameters: the resilient backpropagation algorithm (algorithm = “rprop+”) was used for training, the activation function was set to the logistic sigmoid (act.fct = “logistic”), and the output layer was configured for nonlinear classification (linear.output = FALSE). Additionally, a neural interpretation diagram of the network was plotted with the assistance of the NeuralNetTools package (v 1.5.3) (35). Notably, this study not only analyzed the importance of candidate biomarkers in the MLP network using the Olden method but also evaluated the model’s classification performance based on a confusion matrix. Due to the limited sample size (n = 24), the MLP model was trained on the full dataset without data partitioning or cross-validation. The reported performance (confusion matrix), therefore, represents apparent accuracy (within-sample performance) and is intended for descriptive illustration of candidate biomarker weights rather than for deriving generalizable predictive metrics. The simple network architecture (single hidden layer with 2 neurons) was designed to mitigate overfitting while capturing nonlinear relations between key genes and the phenotype. Nevertheless, readers are cautioned that the within-sample accuracy estimate may still be overly optimistic, and external validation is needed.

Immune infiltration analysis

We employed CIBERSORT to map 22 immune cell infiltration patterns comparing the SAE and control groups in GSE135838. Notably, the Wilcoxon test was employed to compare differences in immune cell infiltration between the SAE group and the control group. The test utilized cell abundance values as input and was performed independently for each cell type under default parameters (two-tailed, unpaired). Immune cells with P values below 0.05 were defined as differentially infiltrating immune cells. The above results were displayed utilizing the ggplot2 package (v 3.4.1). Moreover, GSE135838’s complete sample set underwent candidate biomarker-immune cell correlation profiling via psych (v 2.1.6) (36), with significant associations defined by |r| > 0.3 and P < 0.05. The correlation threshold of |r| > 0.3 was selected based on commonly accepted criteria for moderate correlation strength in biomedical research (37,38). Additionally, a stricter significance level of P < 0.05 was applied to minimize false positive associations given the exploratory nature of the analysis.

Gene set enrichment analysis (GSEA)

To examine the pathways in which the candidate biomarkers were involved and the potential biological mechanisms in SAE, GSEA was performed on all samples from GSE135838 using the “c2.cp.kegg_legacy.v2024.1.Hs.symbols.gmt” from the MSigDB (https://www.gsea-msigdb.org/gsea/msigdb/) as the reference gene set. Initially, we applied the psych package (v 2.1.6) to carry out Spearman correlations between every candidate biomarker and the remaining genes. Notably, individual genes were sorted in the order of decreasing correlation coefficient values. Subsequently, GSEA was performed using the clusterProfiler package (v 4.2.2), with significance thresholds set as [|Normalized Enrichment Score (NES) | > 1, adj.P < 0.05]. The top five enriched pathways were subsequently visualized through enrichplot (v 1.18.3) (39) and the Cytoscape software (v 3.8.2).

Functions of candidate biomarkers

First, the psych package (v 2.1.6) was applied to probe the correlations among candidate biomarkers in all samples of GSE135838 (|cor| > 0.3, P < 0.05). Second, the GOSemSim package (v 2.33.0) (40) was applied to analyze whether the functions of candidate biomarkers were similar by calculating the semantic similarity of candidate biomarkers in the three categories of BP, CC, and MF of GO analysis (score > 0.5). At last, the GeneMANIA database was implemented to predict the top 20 genes with functions similar to those of candidate biomarkers and the top 5 pathways in which these genes were involved.

Molecular regulatory networks

We predicted regulatory interactions linking candidate biomarkers to transcription factors (TFs), miRNAs, and long noncoding RNAs (lncRNAs) using public databases, subsequently constructing molecular networks. The ENCODE database (https://www.encodeproject.org/) in the NetworkAnalyst platform (https://www.networkanalyst.ca/NetworkAnalyst/) was employed to predict TFs. TF-mRNA regulatory networks were constructed in Cytoscape (v 3.8.2) using TFs linked to >2 candidate biomarkers. The miRTarBase database (https://mirtarbase.cuhk.edu.cn/) and TarBase database (https://dianalab.e-ce.uth.gr/tarbasev9) in the NetworkAnalyst platform were applied to predict miRNAs. The miRNAs jointly predicted by the 2 databases were defined as key miRNAs. Moreover, the miRNet platform (https://www.mirnet.ca/miRNet/) identified lncRNAs targeting miRNAs. The top three lncRNAs showing the highest number of interactions with miRNAs were used to build the lncRNA–miRNA–mRNA network using Cytoscape software (v 3.8.2).

Drug prediction and molecular docking

To obtain drugs that might act on candidate biomarkers, potential therapeutic drugs targeting candidate biomarkers were obtained based on the DGIdb (https://dgidb.org/), and a drug–candidate biomarker interaction network was built via the Cytoscape software (v 3.8.2). Based on the interaction strength scores with candidate biomarkers, the top five drugs with 2D structures were selected for molecular docking with the candidate biomarkers. The 3D structures of the candidate biomarkers were collected and retrieved from the AlphaFold database (https://alphafold.ebi.ac.uk/), while the 3D structures of the drugs were obtained from the PubChem database (https://pubchem.ncbi.nlm.nih.gov/). At last, molecular docking was performed using CB-Dock software (v 2.0.0) (41).

Animal treatment

Specific pathogen-free male C57BL/6J mice were obtained from Henan Skobes Biotechnology Co., Ltd. All animal experiments were approved by the Animal Experiment Ethics Committee of Yunnan University of Chinese Medicine and were conducted in accordance with the provided guidelines (Approval No R062023054). A single systemic injection of lipopolysaccharides (LPS) (15 × 10⁶ EU/kg, i.p., Escherichia coli0111: B4, Sigma-Aldrich, St. Louis, MO) was administered to 7- to 12-week-old C57BL/6J as LPS-induced SAE models (N = 5). We administered saline (5 mL/kg, i.p.) to vehicle control mice (N = 5). After 4 weeks postinjection, mice were euthanized for brain tissue collection and analysis. All mice were housed with 12 hours of artificial light–dark cycle and provided fresh deionized or acidified water.

Reverse transcription quantitative PCR (RT-qPCR)

RT-qPCR was utilized to confirm the expression levels of candidate biomarkers between SAE and control brain samples. The total RNA was extracted from 50 mg aliquots of frozen mice brain tissue samples, SAE group (n = 5) and control group (n = 5), using TRIzol reagent (Invitrogen). Each sample was homogenized in 1 mL TRIzol reagent and incubated on ice for 10 minutes. After adding 300 µL chloroform, samples were shaken for 30 seconds, incubated at room temperature for 10 minutes, and centrifuged at 12,000 × g for 15 minutes at 4°C. The upper aqueous phase containing RNA was carefully transferred. RNA was precipitated with an equal volume of ice-cold isopropanol (10 minutes, room temperature), centrifuged (12,000 × g, 10 minutes, 4°C), washed twice with 75% ethanol (1 mL), and dissolved in RNase-free water (20 µL). Then, the RNA concentrations of the samples were computed by NanoPhotometer N50. Additionally, we performed mRNA reverse transcription using a commercial cDNA synthesis kit (Yeasen, Shanghai, China). Finally, the RT-qPCR was conducted following the reagents, conditions, and primers in the attached table (Supplementary Table 2, https://links.lww.com/SHK/C854). Notably, the expression levels of candidate biomarkers were calculated by 2^−ΔΔCt. The internal reference gene was GAPDH, which was applied to normalize the results. The results were determined with GraphPad Prism 10. Intergroup comparisons were performed using the independent samples t-test, with P < 0.05 as the threshold for statistical significance.

RESULTS

Candidate genes and hub genes in SAE

In the GSE135838, compared with the control group, 547 DEGs were acquired in the SAE group (|log₂FC| > 0.5 and P < 0.05), comprising 342 up-regulated DEGs and 205 down-regulated DEGs. Based on log2FC values, all DEGs and the top 10 names of up- and down-regulated DEGs were displayed and marked by the volcano plot (Fig. 1A). Meanwhile, the top 10 DEGs (up/down-regulated) showed distinct profiles in SAE versus controls, visualized by heatmap (Fig. 1B). After intersecting DEGs and MM-RGs, 34 candidate genes were acquired in SAE (Fig. 1C). Notably, GO analysis revealed that the candidate genes were primarily enriched in 177 terms, which included 148 BP terms (e.g., alcohol metabolic process), 1 CC term (e.g., lipid droplet), and 28 MF terms (e.g., iron ion binding) (adj.P < 0.05) (Fig. 1D, Supplementary Table 3, https://links.lww.com/SHK/C855). The candidate genes also showed significant enrichment in 17 KEGG pathways, including the PPAR signaling pathway (adj.P < 0.05) (Fig. 1E, Supplementary Table 4, https://links.lww.com/SHK/C856).

Fig. 1.

Acquisition of candidate and hub genes in SAE. A, Volcano plot of DEGs. Red and blue dots denote significantly up- and down-regulated genes (|Log₂FC| ≥ 0.5, P < 0.05), respectively. Gray dots represent non-significant genes. Top 10 up- and down-regulated genes (by |Log₂FC|) are labeled. B, Heatmap of expression for the top 20 DEGs. Red indicates high expression; blue indicates low. Top annotation: yellow for disease, green for control. Density plot shows expression distribution per sample. C, Venn diagram shows overlap between DEGs and target gene set, indicating candidate genes. D, GO enrichment analysis of candidate genes. E, KEGG enrichment analysis of candidate genes. Bubble color depth corresponds to significance; x-axis shows gene ratio. (f) PPI network of candidate genes. Nodes represent proteins; edges indicate interactions. G, Bar plot shows connectivity degree of 27 genes in PPI network. Top 10 genes (by degree, in red) are defined as hub genes for further analysis. PPI, protein-protein interaction network.

Additionally, 34 candidate genes were uploaded to the STRING database for protein-protein interaction network analysis, with a threshold of an integrated score ≥0.4. After removing isolated targets, interactions were identified among proteins encoded by 27 candidate genes (e.g., patatin-like phospholipase domain-containing protein 3 [PNPLA3]), suggesting their potential key roles in SAE (Fig. 1F). Based on Degree, 11 candidate genes—including insulin-induced gene 1 (INSIG1), G6PC2, and CYP2C9 (all with Degree of 5)—were identified as hub genes for subsequent analysis (Fig. 1G).

Candidate biomarkers in SAE

Next, based on the hub genes identified through the above screening process, multiple machine learning methods were further employed to select feature genes. LASSO regression analysis determined the optimal regularization parameter (lambda.min = 0.012) by minimizing cross-validation error, thereby identifying nine feature genes (Fig. 2, A and B). The prediction error rate of the SVM-RFE gene screening model reached its minimum value (0.13) during iteration, at which point 10 feature genes were selected (Fig. 2, C and D). Features classified as “confirmed” through the Boruta algorithm were defined as feature genes, with a total of four ultimately identified (Fig. 2E). After taking the intersection of the feature genes obtained by the above three algorithms, four feature genes (INSIG1, sterol regulatory element-binding TF 1 [SREBF1], cell death-inducing DFFA-like effector C [CIDEC], and PNPLA3) were defined as core genes (Fig. 2F). It is worth noting that the AUC values of the ROC curves for all core genes exceeded 0.7, suggesting that these genes have potential diagnostic value for SAE; they were therefore identified as preliminary candidate biomarkers (Fig. 2G, Supplementary Table 5, https://links.lww.com/SHK/C857). Specifically, the AUC value for INSIG1 was 0.799 (95% confidence interval [CI] 0.619–0.978), SREBF1 was 0.847 (95% CI 0.66–1), CIDEC was 0.889 (95% CI 0.758–1), and PNPLA3 was 0.868 (95% CI 0.726–1). Moreover, SAE samples exhibited significantly lower expression of all preliminary candidate biomarkers versus controls (P < 0.0001) (Fig. 2H). Thus, INSIG1, SREBF1, CIDEC, and PNPLA3 were considered as candidate biomarkers in SAE.

Fig. 2.

Screening and analysis of core genes in SAE. A, LASSO penalty coefficient spectrum. Note: The x-axis represents the natural logarithm of Lambda, where Lambda is a regularization parameter in Lasso regression. The larger its value, the stronger the regularization intensity, and the coefficients in the model tend to be closer to zero. The y-axis reflects the magnitude of the regression coefficients. B, Cross-validation error plot. Note: The x-axis represents the regularization parameter log(λ), where the smaller the value of λ, the stronger the regularization intensity. The y-axis represents the mean squared error; the smaller its value, the better the predictive performance of the model. C, Number of features and cross-validation error of the SVM-RFE algorithm. D, Number of features and accuracy of the SVM-RFE algorithm. Note: The x-axis represents the number of features; the y-axis reflects the cross-validation error (left) or accuracy (right). E, Feature selection of the Boruta algorithm. Note: The x-axis represents different gene names, and the y-axis reflects the importance scores of these genes (importance). F, Core genes are obtained by taking the intersection of genes derived from the three algorithms. G, ROC analysis of core genes. Note: The x-axis represents the false positive rate (False Positive Rate), and the y-axis represents the true positive rate (True Positive Rate). The closer the ROC curve is to the upper left corner, the stronger the classification ability of the model. H, Expression status of core key genes in the training set. Note: Yellow represents the disease group, and green represents the control group. The P values of the Wilcoxon rank-sum test are labeled in the figure.

The weight of candidate biomarkers in the diagnosis of SAE

Candidate biomarkers were input into the MLP network, which could well distinguish SAE samples from control samples (Fig. 3A). CIDEC and PNPLA3 had relatively high weights in the MLP network and played an important role in the diagnosis of SAE (Fig. 3b). Within the training dataset, the confusion matrix showed that the MLP model’s predictions for the SAE and control samples were consistent with the actual classifications, achieving an apparent accuracy of 100% for both groups (Fig. 3c). However, given the small sample size, this perfect within-sample performance should be interpreted with caution, as it may reflect overfitting. Overall, the MLP model demonstrated the ability to separate SAE from control samples within the discovery cohort, but external validation is necessary to assess its generalizability.

Fig. 3.

Multilayer perceptron network analysis to evaluate the diagnostic ability of key genes. A, Neural network interpretation plot of key genes in the training set. Note: Positive weights between layers are plotted as red lines, and negative weights as gray lines. The thickness of the lines is proportional to the relative magnitude of each weight. B, Bar chart of the relative importance of key genes for disease prediction in the training set. C, Confusion matrix plot of the model’s disease prediction performance in the training set. Note: Rows represent actual classes, and columns represent model-predicted classes.

Immune microenvironment in SAE and pathways enriched with candidate biomarkers

The stacked diagram illustrated the abundances of 22 immune cells in both the SAE and control samples (Fig. 4A). Relative to the control group, the SAE group exhibited a significant reduction in the infiltration levels of naive B cells and activated natural killer cells (P < 0.05); in contrast, the infiltration level of resting mast cells in the SAE group was significantly increased (P < 0.05). Therefore, these three types of cells were defined as differential immune cells (Fig. 4B). Among the aforementioned cells, PNPLA3 exhibited the most significant positive correlation with naive B cells, with a correlation coefficient (cor) of 0.47 and P < 0.05 (Fig. 4C, Supplementary Table 6, https://links.lww.com/SHK/C858). Additionally, INSIG1 was observed to exhibit the most notable negative correlation with resting mast cells (cor = −0.57, P < 0.01) (Fig. 4C, Supplementary Table 6, https://links.lww.com/SHK/C858). These differential immune cells might affect the development of SAE. Not only that, but also the pathways of candidate biomarker enrichment might also affect the development of SAE.

Fig. 4.

Immune infiltration and pathway enrichment analysis in SAE. A, Stacked plot of infiltrating immune cells. Note: The x-axis represents different samples, and the y-axis shows the percentage of various types of immune cells. The legend at the bottom indicates the cell types represented by different colors. B, Violin plot of the infiltration status of immune cells. Note: SAE group (orange) and control group (blue). C, Correlation between key genes and differential immune cells. Note: Red areas represent positive correlations, and blue areas represent negative correlations. Significantly correlated immune cell-gene pairs are marked with asterisks. D, GSEA enrichment analysis of the key gene INSIG1. Note: The x-axis represents gene ranking, and the y-axis is divided into two parts: the upper part shows the enrichment score, and the lower part displays the ranking distribution of all genes. E, GSEA enrichment analysis of the key gene SREBF1. F, GSEA enrichment analysis of the key gene CIDEC. G, GSEA enrichment analysis of the key gene PNPLA3. H, Key gene-pathway network. Note: Key genes are represented by red nodes, and pathways by green nodes.

GSEA analysis suggested that INSIG1 was remarkably enriched in 23 pathways, such as the spliceosome and ubiquitin-mediated proteolysis (|NES| > 1, adj.P < 0.05) (Fig. 4D, Supplementary Table 7, https://links.lww.com/SHK/C859). Three pathways (e.g., the ribosome and spliceosome) exhibited significant enrichment of SREBF1, where the screening criteria of |NES| > 1 and adj.P < 0.05 were satisfied. Details can be found in Figure 4E and Supplementary Table 8, https://links.lww.com/SHK/C860. CIDEC was notably enriched in eight pathways, including glycosaminoglycan biosynthesis chondroitin sulfate and allograft rejection, meeting the criteria of |NES| > 1 and adj.P < 0.05 (Fig. 4F, Supplementary Table 9, https://links.lww.com/SHK/C861). PNPLA3 was significantly enriched in 13 pathways such as allograft rejection and ubiquitin-mediated proteolysis (|NES| > 1, adj.P < 0.05) (Fig. 4G, Supplementary Table 10, https://links.lww.com/SHK/C862). Furthermore, the candidate biomarker–pathway network revealed that multiple genes were co-enriched within the same pathway—with allograft rejection as an example—which suggests these pathways may exert a key role in SAE progression (Fig. 4H).

Functions and molecular regulatory network of candidate biomarkers

Among candidate biomarkers, INSIG1 had the greatest remarkable positive link with PNPLA3 (cor = 0.58, P < 0.01) (Fig. 5A, Supplementary Table 11, https://links.lww.com/SHK/C863). Besides, all candidate biomarkers had lower similarity in terms of BP and MF, while they had higher similarity in terms of CC (score > 0.5) (Fig. 5B). Furthermore, the gene–gene interaction network demonstrated the interactions between candidate biomarkers and genes with similar functions, such as HMGCR and SCAP. These genes were mainly enriched in biological functions like the sterol biosynthetic process (Fig. 5C).

Fig. 5.

Analysis of the functions and molecular regulatory networks of biomarkers. A, Correlation analysis of key genes. Note: Red areas represent positive correlations, and blue areas represent negative correlations. Significantly correlated ones are marked with asterisks. B, Semantic similarity analysis of key genes. Note: The y-axis represents different GO groups, and the x-axis represents semantic similarity scores. C, GeneMANIA network graph. Note: Nodes represent key genes and related genes, and edges represent relations between genes. D, TF-mRNA network graph. Note: Red nodes represent key genes, and green nodes represent TFs targeting key genes. E, mRNA–miRNA–lncRNA molecular regulatory network graph. Note: Red nodes represent key genes, green nodes represent miRNAs, and purple nodes represent lncRNAs.

Subsequently, the online analysis platform NetworkAnalyst was used to query the ENCODE database, and 34, 49, 11, and 93 TFs that targeted CIDEC, INSIG1, PNPLA3, and SREBF1, respectively, were identified (Supplementary Table 12, https://links.lww.com/SHK/C864). Notably, some TFs were associated with multiple candidate biomarkers, such as THRB and ZFP2 (Fig. 5D). Additionally, the online analysis platform NetworkAnalyst was utilized to search for miRNAs targeting the candidate biomarkers in the miRTarBase and TarBase databases. There were 3, 32, 2, and 14 miRNAs predicted to target CIDEC, INSIG1, PNPLA3, and SREBF1, respectively, resulting in a total of 44 miRNAs, such as hsa-mir-29c-3p (Supplementary Table 13, https://links.lww.com/SHK/C865). In addition, prediction using the miRNet database indicated that these 44 miRNAs were associated with 32 lncRNAs, such as KCNQ1OT1 (Fig. 5E, Supplementary Table 14, https://links.lww.com/SHK/C866).

The binding ability between candidate biomarkers and drugs

To identify potential drugs targeting candidate biomarkers, a search was conducted in the DGIdb database. Results showed that PNPLA3 was predicted to be associated with 12 drugs, such as pegaspargase. SREBF1 was predicted to be related to two drugs, such as recombinant neurotrophic factor (Fig. 6A). Except for pegaspargase, asparaginase, recombinant neurotrophic factor, and insulin, the 2D structures of the corresponding drugs for the others could be retrieved from the PubChem database. The top 5 drugs in terms of the interaction strength score with PNPLA3 were prednisolone, mercaptopurine, thioguanine, cyclophosphamide anhydrous, and vincristine. These five drugs were used for molecular docking with PNPLA3 (Table 1). SREBF1 could not perform molecular docking with the predicted drugs. A relatively strong binding energy was observed between PNPLA3 and vincristine, with a value of −9.4 kcal/mol (Fig. 6B).

Fig. 6.

Drug prediction and molecular docking analysis of key genes. A, Key gene-drug network graph. Note: Red nodes represent key genes, and green nodes represent drugs targeting key genes. B, Molecular docking 3D structure diagram. Note: Blue dashed lines represent hydrogen bonds; light blue represents weak hydrogen bonds; gray dashed lines represent hydrophobic interactions; cyan represents halogen bonds; yellow represents ionic interactions; orange represents cation–π interactions; green represents π–π stacking.

Table 1.

Interaction scores between drugs and key genes and molecular docking binding energy

Drug	Interaction score	ChEMBLID	Vina score
Prednisolone(Prednisolone)	0.466105344	CHEMBL131	−8.9
Prednisolone(Prednisolone)	1.466105344	CHEMBL132	−7.9
Prednisolone(Prednisolone)	2.466105344	CHEMBL133	−6.9
Prednisolone(Prednisolone)	3.466105344	CHEMBL134	−5.9
Cyclophosptamide anhydrous (Anhydrous cyclophosphamide)	0.214830447	CHEMBL88	−6.2

The expression level of candidate biomarkers in animal samples

To validate the expression of candidate biomarkers, RT-qPCR validation was performed on animal samples. The expression levels of INSIG1 (P < 0.001) (Fig. 7A), SREBF1 (P < 0.0001) (Fig. 7B), CIDEC (P < 0.05) (Fig. 7C), and PNPLA3 (P < 0.001) (Fig. 7D) in the SAE group were significantly inferior to those in the control group. The results suggested that the results of bioinformatics analysis were reliable.

Fig. 7.

Expression levels of candidate biomarkers in LPS-induced SAE mouse model measured by RT-qPCR (n = 5 per group). A, The expression levels of INSIG1. B, The expression levels of SREBF1. C, The expression levels of CIDEC. D, The expression levels of PNPLA3. Data were analyzed using an unpaired Student’s t-test. *P < 0.05, ***P < 0.001, ****P < 0.0001. LPS, lipopolysaccharides.

DISCUSSION

SAE is recognized as one of the most serious complications of sepsis, which notably contributes to prolonged hospital stays and higher mortality rates among patients with sepsis (42,43). Early recognition of SAE and further exploration of its pathogenesis are essential for improving patient prognosis. Based on the analysis of transcriptomic data from public databases, this study identified 547 DEGs. Through the application of LASSO, SVM-RFE, and Boruta algorithms, four MM-related genes—INSIG1, SREBF1, CIDEC, and PNPLA3—were identified as potential diagnostic biomarkers for SAE.

INSIG1 encodes for an endoplasmic reticulum (ER) membrane protein that is involved in regulating lipid, cholesterol, and glucose metabolism (44), and has been implicated in immune regulation (45,46). Although primarily localized to the ER, INSIG1 maintains functional connections with MM via the SREBP signaling pathway. Under sterol-rich conditions, INSIG1 retains SREBF1 in the ER, thereby limiting its activation (46). Studies have demonstrated that SREBF1 is a conserved regulator of PTEN-induced kinase 1–PARK2-mediated mitophagy, and its deficiency may impair mitochondrial quality control by reducing PTEN-induced kinase 1 stability on the outer mitochondrial membrane (47). Therefore, dysregulation of INSIG1 may disrupt the activation of SREBF1, potentially leading to impaired mitophagy and the accumulation of dysfunctional mitochondria. Consistently, our results showed that INSIG1 expression was positively correlated with SREBF1 (Fig. 5A), and both were significantly downregulated in SAE samples, suggesting a coordinated suppression of the ER–mitochondria axis. In addition to its metabolic role, INSIG1 has been reported to inhibit inflammasome activation through promoting NLRP3 ubiquitination and to modulate T cell differentiation, highlighting its involvement in immune regulation (48). Aberrant activation of inflammatory pathways, lipid metabolism disorders, and cholesterol metabolism abnormalities are commonly observed in neurodegenerative diseases such as Parkinson’s disease, Huntington’s disease, and amyotrophic lateral sclerosis (49). Therefore, the coordinated downregulation of INSIG1 and SREBF1 in SAE may contribute to neuroinflammation through two complementary mechanisms: first, by impairing mitophagy and promoting the accumulation of dysfunctional mitochondria; and second, by modulating immune cell activation, thereby enhancing inflammatory responses and disrupting lipid metabolism.

CIDEC was originally recognized as a protein associated with lipid droplets, participating in lipolysis and lipid homeostasis within adipocytes; its function is particularly well-defined in the process of 3T3-L1 cell differentiation (50). Although the role of CIDEC in the CNS remains unclear, evidence indicates that its expression correlates with inflammation (51). High expression of CIDEC inhibits AMP-activated protein kinase (AMPK) activity and activates the NF-κB pathway in renal tubular epithelial cells, thereby promoting the release of pro-inflammatory cytokines and cell apoptosis (51). In the present study, CIDEC was significantly downregulated in SAE samples, suggesting a potential disruption of lipid droplet-associated metabolic regulation. Given that CIDEC has been reported to modulate AMPK activity, its reduced expression may relieve AMPK inhibition and subsequently modulate NF-κB signaling, although this regulatory direction in the CNS requires further validation. Additionally, CIDEC expression was positively correlated with INSIG1 (Fig. 5A), indicating that lipid droplet-associated proteins may coordinately participate in the metabolic–inflammatory network in SAE. Collectively, these findings suggest that CIDEC downregulation may contribute to neuroinflammatory regulation via the AMPK/NF-κB axis in SAE, potentially acting in concert with INSIG1.

PNPLA3 is a member of the PNPLA family and functions as a lipid droplet-associated enzyme involved in triglyceride metabolism (52). It is highly expressed in the liver and found in the skin, retina, kidney, and brain (53). PNPLA3 variants have been associated with magnetic resonance imaging markers of cerebral small vessel disease in nonalcoholic fatty liver disease patients, including white matter hyperintensities, cerebral microbleeds, and covert brain infarctions. The proposed mechanism is that hepatic inflammation and oxidative stress may promote CNS inflammation and white matter injury (54). This study found that PNPLA3 was significantly downregulated in SAE samples. Given its established role in lipid metabolism, reduced PNPLA3 expression may disrupt lipid homeostasis and exacerbate metabolic stress, thereby contributing to neuroinflammation. Furthermore, PNPLA3 and CIDEC, both lipid droplet-associated proteins, showed a positive correlation (Fig. 5A), further implicating coordinated dysregulation of lipid metabolism in the pathogenesis of SAE.

In the present study, naive B cells, activated natural killer cells, and resting mast cells showed differential abundances between the SAE and control groups. Such an imbalance in immune cell proportions and dysregulated immune responses may contribute to the onset and progression of SAE. Notably, INSIG1 expression was significantly negatively correlated with resting mast cells, whereas PNPLA3 was positively correlated with naive B cells, indicating that these genes may modulate immune cell recruitment and activation through lipid metabolism-related pathways. Previous studies have indicated that mast cells can contribute to blood–brain barrier disruption and cognitive impairment in SAE through histamine-dependent mechanisms (55), while inhibition of mast cell activation alleviates neuroinflammation (10). Collectively, these findings support a potential interaction between lipid metabolism-related genes and immune cells in regulating neuroinflammation in SAE. Furthermore, GSEA analysis revealed that four candidate biomarkers were significantly enriched in spliceosome and allograft rejection pathways. Spliceosome pathway enrichment has been associated with impaired mitophagy and neuroinflammation in a mouse model of Parkinson’s disease (56). Significant enrichment of the allograft rejection pathway was observed and correlated with microglial activation and neuroinflammation onset in a mouse model of multiple sclerosis (57). These results suggest that aberrant expression of candidate biomarkers may contribute to neuroinflammation in SAE through these pathways.

At the upstream regulatory level, our analysis identified 44 miRNAs and 32 lncRNAs targeting the candidate biomarkers, including hsa-mir-29c-3p. Previous studies have shown that downregulation of mir-29c-3p increases β-secretase (BACE1) expression, potentially promoting β-amyloidogenesis and neuroinflammation (58). In contrast, upregulation of mir-29c-3p has been reported to attenuate Aβ-induced microglial apoptosis and reduce proinflammatory cytokine production in in vitro models of Alzheimer’s disease (59). These findings suggest that hsa-mir-29c-3p and its associated lncRNAs may contribute to the pathogenesis of SAE by regulating candidate biomarker expression.

The integration of machine learning algorithms consistently identified four core genes (INSIG1, SREBF1, CIDEC, and PNPLA3) as candidate biomarkers for SAE. The convergence of multiple algorithms reduces model-specific bias, and RT-qPCR validation independently confirmed the significant downregulation of these genes in SAE samples. The consistency between bioinformatics predictions and experimental results supports the reliability of this analytical strategy. The marked suppression of these biomarkers suggests their potential involvement in SAE pathogenesis, highlighting their potential as diagnostic indicators and possible therapeutic targets for sepsis-associated neurological complications. However, several limitations should be acknowledged. First, all analyses were based on a single GEO dataset with a limited sample size (n = 24), without data partitioning or external validation, which may introduce population heterogeneity and overfitting bias. The ratio of variables to samples was substantially imbalanced (547:24 in the DEG stage and 34:24 in the candidate gene selection stage), neither meeting the recommended 1:10 standard, thereby increasing the risk of overfitting. Although LASSO and SVM-RFE employed cross-validation for feature selection, the MLP model was trained on the full dataset without partitioning, and its 100% classification accuracy represents only within-sample performance, limiting its generalizability. Future studies should employ large-scale, multicenter SAE cohorts with rigorous cross-validation and external validation strategies to confirm the diagnostic value of these biomarkers. Second, experimental validation was performed exclusively in an LPS-induced SAE mouse model, which does not fully recapitulate the clinical and etiological heterogeneity of human SAE (e.g., polymicrobial sepsis and chronic comorbidities), and protein-level validation was not conducted. Subsequent studies should incorporate multiple animal models (e.g., cecal ligation and puncture model) and multicenter clinical samples (peripheral blood or cerebrospinal fluid) for cross-validation to enhance the generalizability of the results and evaluate the potential of these four genes. Third, biomarker validation was conducted only at a relatively late time point (4 weeks post-LPS), reflecting a chronic neuroinflammatory stage and limiting evaluation of their utility for early diagnosis. Future studies integrating in vivo and in vitro approaches are warranted to further confirm these findings and elucidate the underlying mechanisms. Investigating the temporal dynamics of these biomarkers during the acute phase will be essential to clarify their diagnostic and therapeutic potential. Finally, no function experiments were performed, and the causal relations between the identified genes and SAE pathophysiology remain unclear; and the current findings are primarily associative. Further in vivo and in vitro functional studies are required to elucidate the underlying molecular mechanisms. In summary, this study provides evidence for MM–related biomarkers in SAE; however, their clinical translation requires validation in large-scale cohorts and deeper mechanistic investigations.

CONCLUSIONS

This study identified INSIG1, SREBF1, CIDEC, and PNPLA3 as candidate biomarkers for SAE, constructed a classification model based on them, and preliminarily validated these findings in an LPS-induced SAE mouse model. Through in-depth analysis of their associations with the immune microenvironment, transcriptional regulatory network, and pharmaceuticals, the study provided candidate molecular targets for future mechanistic studies and preliminary clues for the early identification of SAE. However, their diagnostic or therapeutic value remains to be determined in large-scale, prospective clinical cohorts and more diverse animal models.

ACKNOWLEDGMENTS

We would like to express our sincere gratitude to all individuals and organizations who supported and assisted us throughout this research. We acknowledge support from the Yunnan Provincial Clinical Center for Respiratory Diseases (Grant No. ZX2019-01-03) for providing platform resources.