Machine Learning-Guided Multi-Omics Integration Identifies UGCG as a Candidate Lipid Metabolic Biomarker and Potential Therapeutic Target in Sepsis

Abstract

Background

Sepsis is a highly lethal organ dysfunction syndrome characterized by dysregulation of the host immune response. This study aims to identify robust lipid metabolism-related genes (LMRGs) as candidate diagnostic biomarkers and to elucidate their underlying immunometabolic mechanisms.

Methods

We integrated two sepsis transcriptomic cohorts to screen for differentially expressed LMRGs and identified metabolic-immune subtypes through consensus clustering. Subsequently, we constructed diagnostic models using three machine learning algorithms to identify core genes. We further performed single-cell RNA sequencing (scRNA-seq) analysis, AI-driven drug screening, and in vitro experimental validation combined with lipidomics to explore the biological functions of the core targets.

Results

GSEA results indicated synergistic overexpression of lipid metabolism and innate immune pathways in sepsis. LMRG-based clustering analysis successfully identified two molecular subtypes, with lipid metabolism activity highly correlated with characteristic patterns of immune infiltration. Machine learning algorithms precisely identified three core genes (ALDH9A1, MBOAT2, and UGCG). The diagnostic model demonstrated excellent diagnostic performance in both the training set (AUC = 1.000) and four external validation sets (AUCs ranging from 0.943 to 0.998). BRD-K525600704 was predicted to reverse the transcriptomic signature of sepsis and exhibited favorable binding affinity for UGCG and MBOAT2. scRNA-seq revealed a profound myeloid skew in the sepsis microenvironment. In cell models, knockdown of UGCG significantly suppressed LPS-induced upregulation of inflammatory cytokines. Lipidomics further showed that UGCG knockdown attenuated pathological ceramide accumulation, suggesting partial restoration of cellular sphingolipid metabolic homeostasis.

Conclusion

This study developed a candidate sepsis diagnostic model comprising three LMRGs. Through integrated multi-omics analysis and in vitro validation, our findings suggest that UGCG is closely associated with lipid metabolic remodeling and myeloid-mediated inflammatory responses in sepsis. However, large-scale prospective clinical validation is currently lacking and warrants future investigation.

Introduction

Sepsis is an organ dysfunction syndrome caused by a dysregulated response of the human body to infection.^1,2 Septic shock, as well as multiple organ dysfunction and failure, is one of the potentially fatal consequences of this emergency condition.^3–6 Currently, clinical diagnosis and severity assessment of sepsis largely rely on conventional indicators, including PCT, CRP, serum lactate, and the Sequential Organ Failure Assessment (SOFA) score.² Although PCT and CRP are widely used inflammatory biomarkers, their specificity remains insufficient for accurately distinguishing sepsis from non-infectious systemic inflammation.^7,8 Serum lactate, a key marker of tissue hypoperfusion and disease severity, can be influenced by factors such as hepatic dysfunction and altered metabolic states, complicating its interpretation.⁹ Moreover, the SOFA score primarily reflects established organ dysfunction and may not adequately capture early immune dysregulation preceding overt organ failure.¹⁰ Therefore, identifying novel, mechanism-driven biomarkers is essential for the timely diagnosis of sepsis, enhancing treatment precision, and ultimately improving prognosis.

In recent years, research progress in the field of immune metabolism has brought new perspectives to elucidate the pathogenesis of sepsis.¹¹ While unbiased pan-transcriptomic analyses have extensively cataloged the downstream inflammatory cascades in sepsis, they frequently mask the upstream metabolic vulnerabilities that drive these immune responses. Recent multi-omics and metabolomic studies have highlighted that severe infections and sepsis trigger profound systemic metabolic shifts. These alterations, including dysregulated glycolysis, altered amino acid routing, and perturbed lipid homeostasis, closely dictate the trajectory of host immune responses and organ dysfunction.^12,13 In the context of severe infections, abnormal expression of key lipid metabolic enzymes may disrupt lipid homeostasis and promote the accumulation of bioactive lipid intermediates.^14,15 Among these lipid species, sphingolipids have emerged as critical regulators of immune activation, with ceramides serving as potent signaling molecules that amplify inflammatory cascades and contribute to cellular dysfunction.¹⁶ The balance between ceramide accumulation and its conversion into downstream glycosphingolipids is tightly regulated by metabolic enzymes and represents an important metabolic axis involved in cellular stress responses and immune homeostasis. UGCG, the rate-limiting enzyme that catalyzes the conversion of ceramide to glucosylceramide, serves as a key metabolic checkpoint in sphingolipid metabolism.¹⁷ However, although sphingolipid dysregulation has been implicated in inflammatory disorders, whether UGCG-mediated ceramide remodeling contributes to immune dysfunction, inflammatory responses, and organ injury during sepsis remains largely unknown. Moreover, the comprehensive landscape and regulatory roles of lipid metabolism-related genes (LMRGs) in sepsis have not been fully elucidated.

In this study, we first identified differentially expressed LMRGs and revealed molecular subtype characteristics associated with lipid metabolism. Subsequently, we applied multiple machine learning algorithms to construct diagnostic models, yielding a three-gene signature comprising ALDH9A1, MBOAT2, and UGCG. These genes represent distinct aspects of lipid metabolic regulation: ALDH9A1 functions as an aldehyde dehydrogenase involved in the detoxification of lipid peroxidation-derived aldehydes;¹⁸ MBOAT2 regulates phospholipid remodeling through the Lands cycle and contributes to membrane lipid homeostasis,¹⁹ and UGCG serves as a critical regulator of sphingolipid metabolism by controlling the generation of glucosylceramide.¹⁷ Finally, we validated the biological functions of the core genes in sepsis using in vitro cell models combined with lipidomics.

Methods

Data Acquisition and Preprocessing

The datasets were obtained from the GEO, which included six sepsis datasets GSE57065,²⁰ GSE95233,²¹ GSE69528,²² GSE65682,²³ GSE26440,²⁴ and GSE13904.²⁵ Supplementary Table 1 summarizes the clinical characteristics of these datasets. GSE57065 and GSE95233 were combined as a training set. Batch effect correction was performed using the “ComBat” algorithm of the “sva” package.²⁶ To validate the effectiveness of batch removal, Principal Component Analysis (PCA) was performed before and after the correction. Successful batch effect elimination was confirmed by the homogeneous overlapping of samples from different datasets in the PCA scatter plots, ensuring that downstream analyses were driven by biological differences rather than technical dataset biases. The datasets GSE69528, GSE65682, GSE26440, and GSE13904 were used as independent external test sets. Additionally, 175 LMRGs were extracted from the MSigDB database.

Identifying the DEGs of Sepsis

Differentially expressed genes (DEGs) in sepsis were identified according to established criteria (|log2FC| > 1 and adjusted p value < 0.05). To reveal alterations in the overall signaling pathways under disease conditions, we performed GSEA using the “clusterProfiler” package.²⁷

Consensus Clustering Analysis of Sepsis

Based on the expression profile of DE-LMRGs, unsupervised consensus clustering was applied to sepsis samples using the “ConsensusClusterPlus” R software package to explore the molecular subtyping features related to lipid metabolism in sepsis.²⁸ PCA was used to verify the stability and reliability of clustering results. The differences in biological functions and immune characteristics of different subtypes of samples were compared using GSVA and CIBERSORT.^29,30

Machine Learning-Based Feature Selection Workflow

To identify the most robust and diagnostically valuable feature genes and prevent model overfitting, we designed an ensemble feature selection workflow incorporating three independent machine learning algorithms. The integrated batch-corrected dataset was used entirely as the training cohort, while four independent datasets served as external testing cohorts to ensure the generalizability of the results. First, the LASSO regression model was constructed using the “glmnet” package. To strictly prevent overfitting, the optimal penalty parameter (λ) was tuned via 10-fold cross-validation, and the lambda.1se criterion was selected to extract the most regularized feature genes. Second, the SVM-RFE algorithm was implemented using the “e1071” package. SVM-RFE was executed with 10-fold cross-validation, and the optimal feature subset was determined at the point where the cross-validation generalization error reached its minimum. Third, the RF algorithm was deployed using the “randomForest” package. To optimize model parameters and reduce redundancy, we initially grew 500 trees and tracked the Out-Of-Bag (OOB) error rate. The model was then fine-tuned by selecting the optimal number of trees that yielded the minimum OOB error. Genes with a MeanDecreaseGini importance score greater than 5 in the optimized RF model were selected as critical features. Ultimately, to maximize diagnostic robustness across different mathematical assumptions, only the genes identified by the intersection of all three algorithms were selected as the final hub genes.³¹ Following feature selection, a multivariable Logistic Regression classifier was ultimately used to integrate the expression profiles of these identified hub genes and build the final diagnostic model.

Functional Validation, Immune Correlation and Transcriptional Regulation Analysis of Core Genes

To explore the potential molecular mechanisms of model genes in sepsis, we performed GSVA using the “GSVA” package.³⁰ Based on the CIBERSORT results, we conducted an in-depth investigation into the correlations between key genes and immune cells in sepsis samples. To elucidate the potential transcriptional regulatory mechanisms of core genes, we utilized the TFTF platform³² including FIMO_JASPAR, ChIP_Atlas, ENCODE, GTRD to predict upstream transcription factors (TFs) targeting these model genes.

Single-Cell RNA Sequencing Analysis

The sepsis scRNA-seq dataset GSE167363 was evaluated with the aid of the R package “Seurat”.³³ Prior to downstream analysis, rigorous quality control was performed. Genes expressed in fewer than 5 cells were removed to reduce background noise. To filter out low-quality cells and dying cells, cells with fewer than 100 detected genes (nFeature_RNA < 100) or a mitochondrial gene expression ratio exceeding 10% (percent.mt > 10%) were excluded. Subsequently, the filtered gene expression matrix was normalized using the “LogNormalize” method. The “SingleR” package was used to automatically perform cell annotation.³⁴ Visualizations of cell clusters, cell proportions, and marker gene distributions were generated using the R package “scRNAtoolVis”. Cellular communication networks were analyzed with the “CellChat” package to examine interaction patterns between key and other cell types, evaluating communication quantity, strength, and ligand-receptor expression.³⁵ Finally, to evaluate the systemic transcriptomic impact and regulatory function of the core gene UGCG, the gene knockout simulation was conducted using the “scTenifoldKnk” R package.³⁶

AI-Powered Drug Screening and Molecular Docking

DrugRefLector is a phenotype-based drug screening method capable of identifying compounds that induce the desired changes in cellular transcriptomic profiles.³⁷ We used DrugRefLector to identify potential therapeutic candidates for sepsis. We obtained protein structures from the PDB database and retrieved chemical structures of active ingredients from the PubChem database. Molecular docking was then performed by CB-Dock2.³⁸

Cell Culture and Model Construction

The cell line RAW264.7 was purchased from the American Type Culture Collection (VA, USA), which was cultured in DMEM supplemented with 10% FBS. To establish the in vitro sepsis model, RAW264.7 was treated with LPS (MCE, USA) at a concentration of 1 μg/mL for 24 hours.³⁹ Inflammatory cytokine levels were used to confirm the establishment of the in vitro model and to compare the trends in gene expression changes in the model following LPS-induced inflammation. si-Ugcg was purchased from Hema Biology (Huzhou, China). After 24 hours of transfection, the culture medium was replaced with fresh medium containing 1 μg/mL LPS, and the cells were incubated for an additional 24 hours for subsequent functional and lipidomic analyses.

Lipid Extraction and LC-MS/MS Analysis

Cellular pseudo-targeted lipidomics was performed following a previously established protocol.⁴⁰ Briefly, cell pellets were lysed via ultrasonication, and class-specific stable isotope-labeled internal standards were spiked into each lysate for relative quantification. Lipids were extracted using a double-step MTBE (methyl tert-butyl ether) method. The organic phase was collected, vacuum-dried, and reconstituted in a solvent mixture of isopropanol/acetonitrile/methanol (5:5:2, v/v/v). Chromatographic separation was achieved on an ACQUITY HSS C18 column (1.8 µm, 2.1×100 mm; Waters) at 40°C. The mobile phases consisted of (A) 60:40 ACN/H2O and (B) 90:10 IPA/ACN, both containing 10 mM ammonium acetate. A 17.5-min gradient (20% to 95% B) was applied at 0.2 mL/min on a 4500 triple quadrupole mass spectrometer (Sciex). Given the small cohort size (n = 12), all samples were prepared and acquired in a single continuous batch to minimize experimental variation. Data were acquired in scheduled SRM mode and processed using MultiQuant MD 3.0.3. The raw peak areas of lipids were normalized against their respective internal standards to obtain relative abundances, with all peaks undergoing rigorous manual inspection to ensure accuracy. Lipids were identified using the LipidMaps and HMDB databases, and data analysis was performed using MetaboAnalyst 6.0.

Statistical Analysis

All analyses were performed using R (version 4.1.3) and GraphPad Prism (version 10.1.2). Sample sizes for all in vitro experiments (n ≥ 3) were empirically determined based on standard laboratory practices. Data are presented as mean ± SD. Two-group and multi-group comparisons were evaluated using the Student’s t-test and One-way ANOVA followed by Tukey’s post-hoc test, respectively. To rigidly adjust for multiple comparisons, raw p-values in both high-throughput transcriptomic analyses and the entire lipidomic dataset were corrected using the Benjamini-Hochberg False Discovery Rate (FDR) method. Diagnostic performance was assessed using ROC curves, reporting the Area Under the Curve (AUC) and 95% Confidence Intervals (CIs). Statistical significance was defined as p < 0.05.

Results

Transcriptomic and Functional Characteristics of Sepsis

First, GSE57065 and GSE95233 were served as the training cohorts. Before correction, PCA revealed a distinct separation of samples driven by their dataset origins, with the first two principal components (PC1 and PC2) explaining 60.8% and 9.7% of the total variance, respectively (Figure 1A). Following adjustment, the samples from the two cohorts overlapped evenly, and the variance explained by PC1 and PC2 shifted to 24.9% and 9.1%, visually validating the successful elimination of technical batch effects (Figure 1B). Based on the batch-corrected integrated dataset, we identified a set of DEGs, comprising 420 upregulated and 349 downregulated genes (Figure 1C). GSEA analysis revealed that sepsis patients were significantly enriched in lipid metabolism and inflammation-related upregulated pathways, such as the PPAR signaling pathway (NES = 1.84, adjusted p < 0.001) and the complement and coagulation cascade (NES = 2.13, adjusted p < 0.001) (Figure 1D and E). In contrast, the antigen processing and presentation (NES = −2.32, adjusted p < 0.001) and primary immunodeficiency pathways (NES = −2.15, adjusted p < 0.001) were significantly downregulated (Figure 1F and G), reflecting the immunological characteristics of sepsis, in which lipid metabolism disorders are accompanied by innate immune activation and adaptive immune paralysis. Further analysis of immune cell infiltration revealed that, compared with the normal control group, there were highly significant differences in the infiltration proportions of key immune cells, such as macrophages and T cell subsets, in the sepsis samples (Figure 1H and I).

Figure 1.

The transcriptomic landscape in sepsis. (A and B) PCA plots before and after batch effect correction. (C) Volcano plot exhibiting the DEGs. (D–G) GSEA plots demonstrating the metabolic-immune dysregulation in sepsis. (H and I) Immune cell infiltration analysis comparing the distribution of 22 immune cell subsets between control and sepsis groups. Statistical significance was determined using the Wilcoxon rank-sum test. * p < 0.05, ** p < 0.01, *** p < 0.001.

Cluster Characteristics Related to Lipid Metabolism in Sepsis

Given the critical role of lipid metabolism dysregulation in sepsis, we intersected the collected LMRGs with DEGs to identify 13 differentially expressed LMRGs (DE-LMRGs) (Figure 2A). Subsequently, we performed consensus clustering analysis on the corrected integrated dataset based on the DE-LMRGs. We classified the samples into two clusters, C1 and C2, and PCA validated the reliability of this clustering (Figure 2B and C). Figure 2D illustrates the expression levels of the 13 DE-LMRGs across the two subtypes. GSVA analysis (Figure 2E and F) revealed that core lipid metabolism pathways—including lipophagy, sphingolipid metabolism, and glycerophospholipid metabolism—were significantly upregulated in cluster C2, whereas pathways related to cell cycle regulation, pyrimidine metabolism, and signal transduction were downregulated. Further analysis revealed significant differences in the infiltration rates of various immune cell subsets between the C1 and C2 clusters (Figure 2G and H), further supporting the synergistic regulatory relationship between lipid metabolism reprogramming and immune microenvironment remodeling in sepsis.

Figure 2.

Consensus clustering and molecular characteristics of sepsis based on DE-LMRG expression. (A) Consensus matrix defining two stable clusters of sepsis samples. (B) PCA plot verifying the reliability of C1/C2 clustering. (C) Proportions of DEGs and LMRGs in C1 and C2 clusters. (D) Expression level of DE-LMRGs in two clusters. (E and F) GSVA showing differentially activated pathways between C2 and C1. (G and H) Immune cell infiltration profiles (G) and statistical comparison (H) of immune cell subsets between C1 and C2 clusters. Statistical significance was determined using the Wilcoxon rank-sum test. * p < 0.05, *** p < 0.001.

Screening of the Diagnostic Model for Sepsis

First, LASSO regression analysis identified 6 feature genes (Supplementary Figure 1A and B). Recursive feature elimination based on the SVM-RFE algorithm yielded 9 feature genes (Supplementary Figure 1C and D). Using the RF algorithm to rank genes by importance, 5 genes (importance > 5) were selected as feature genes (Supplementary Figure 1E and F). Finally, the Venn diagram revealed three common hub genes shared by all three algorithms: ALDH9A1, MBOAT2, and UGCG (Figure 3A). The chromosomal distribution of these hub genes is displayed in the Circos plot (Figure 3B). Subsequently, we constructed ROC curves for three independent genes and their combined models to evaluate their diagnostic performance. The results showed that all three genes demonstrated good diagnostic performance (Figure 3C). More importantly, the three-gene diagnostic model demonstrated outstanding performance in both the training set and four external validation sets (Figure 3D–H). The detailed performance metrics across all cohorts are comprehensively summarized in Supplementary Table 2. MBOAT2 and UGCG were markedly upregulated in the sepsis group, whereas ALDH9A1 showed a significant downregulation (Figure 3I). Correlation analyses revealed low multicollinearity among these genes, implying that each of them could provide distinct diagnostic information (Figure 3J). In addition, we validated the expression levels of the core genes in four external validation sets. The results showed that their expression trends in the validation sets were consistent with those in the training set, namely, MBOAT2 and UGCG were upregulated in sepsis, while ALDH9A1 was downregulated in sepsis (Supplementary Figure 2A–D). Finally, we established an in vitro cell model for further validation (Figure 3K). The results showed that inflammatory factors were upregulated in RAW264.7 following LPS stimulation, confirming the successful establishment of the in vitro model. In the LPS-stimulated group, MBOAT2 and UGCG were upregulated, while ALDH9A1 was downregulated, further validating the accuracy of the bioinformatics analysis.

Figure 3.

Construction of the diagnostic model. (A) Venn diagram displaying the three hub genes (ALDH9A1, MBOAT2, UGCG) identified by the intersection. (B) The chromosomal distribution of three model genes. (C and D) ROC curves of each model gene (C) and three-gene diagnostic model (D) in the training set. (E–H) Validation of ROC curves in four independent external validation sets. (I) The expression levels of three genes in sepsis and the control group (Wilcoxon rank-sum test). (J) Correlation analysis of three genes. (K) The expression levels of inflammatory factors and three model genes in RAW264.7 after LPS stimulation. The in vitro experiments were performed with three independent biological replicates (n = 3). Data are presented as mean ± standard deviation (SD). Statistical significance between two groups was determined by Student’s t-test. * p < 0.05, ** p < 0.01, *** p < 0.001.

Analysis of Pathway Regulation Characteristics of Core Genes

To explore the molecular mechanisms of the core genes, we further performed GSVA analysis. ALDH9A1 was primarily negatively correlated with inflammatory processes associated with sphingolipid synthesis, steroid biosynthesis and lipid metabolism pathways, and chemokine signaling pathways (Figure 4A and B). MBOAT2 was positively correlated with immune-inflammatory responses associated with sphingolipid synthesis, glutathione metabolism, and antigen processing and presentation (Figure 4C and D). UGCG was significantly positively correlated with sphingoglycolipid synthesis, steroid biosynthesis, and lipid metabolism (Figure 4E and F).

Figure 4.

Single-gene GSVA of the model genes. (A-F) GSVA revealing the specific signaling pathways associated with each model gene. Red bars indicate activated pathways, and green bars indicate suppressed pathways.

Correlation Between Core LMRGs and Immune Cell Infiltration in Sepsis

We then further analyzed the correlation between the expression levels of the core LMRGs and immune cell infiltration in sepsis. Downregulated ALDH9A1 showed a significant and strong negative correlation with immune cells such as eosinophils (R = −0.47, p = 1.1e-11) and M0 macrophages (R = −0.36, p = 6.0e-07) (Figure 5A–C). Upregulated MBOAT2 showed a significant positive correlation with key inflammatory cells such as neutrophils (R = 0.27, p = 2e-04) and M2 macrophages (R = 0.27, p = 0.00026), making it a crucial gene in regulating inflammatory immune cell infiltration (Figure 5D–F). UGCG exhibited a very strong positive correlation with key sepsis-associated inflammatory cells, including M0 macrophages (R = 0.52, p = 4.3e-14) and M1 macrophages (R = 0.49, p = 1e-12) (Figure 5G–I).

Figure 5.

Immune analysis of three model genes in sepsis. (A) Correlations between different types of immune cells and ALDH9A1. (B and C) Analysis of the association between ALDH9A1 and eosinophils (B) and M0 macrophages (C). (D) Correlations between different types of immune cells and MBOAT2. (E and F) Analysis of the association between MBOAT2 and neutrophils (E) and M2 macrophages (F). (G) Correlations between different types of immune cells and UGCG. (H and I) Analysis of the association between UGCG and M0 macrophages (H) and M1 macrophages (I). Correlation analyses were performed using Spearman’s rank correlation coefficient. Red numbers denote statistically significant correlations (p < 0.05).

TFs Prediction and AI-Driven Drug Screening

To elucidate upstream regulatory networks, we predicted TFs targeting the three central genes using four databases (FIMO_JASPAR, ChIP_Atlas, ENCODE, GTRD) (Supplementary Figure 2E–G). Next, we cross-referenced the potential transcription factors for the three core genes. A Venn diagram revealed that CTCF, E2F6, MAZ, SP1, and PBX3 can simultaneously target all three model genes (Figure 6A).

Figure 6.

TFs Prediction and AI-Driven Drug Screening. (A) The common TFs for the three core genes. (B) Drug candidates for sepsis predicted by DrugRefLector. (C) Molecular docking of MBOAT2 and BRD-K52560704. (D) Molecular docking of UGCG and BRD-K52560704.

To further explore potential therapeutic agents for sepsis, we analyzed transcriptomic data using DrugRefLector. Figure 6B showed the top five drugs that may reverse sepsis. To assess whether the signature genes could be considered targets for these drugs, we performed molecular docking validation on the top five drugs and two target genes that are upregulated in sepsis (Supplementary Table 3). The results indicated that BRD-K52560704 exhibited good binding affinities with both MBOAT2 and UGCG (−10.4 kcal/mol and −9.3 kcal/mol) (Figure 6C and D).

Single-Cell Transcriptomic Landscape and in Silico Knockout of UGCG

After quality control, normalization, and dimensionality reduction, we visualized the single-cell data using t-SNE (Supplementary Figure 3A). The cells were well-separated into distinct major lineages, including 9 cell subsets (Figure 7A). Analysis of cell proportions revealed that the proportions of neutrophils and monocytes were significantly elevated in the sepsis group, while the proportions of subpopulations such as platelets, CD8+ T cells, and CD4+ T cells were significantly reduced (Figure 7B). Further cell-cell communication analysis revealed that monocytes acted as highly active communication hubs (Figure 7C and D). Pathway classification indicated that the communications were predominantly driven by Secreted Signaling and Cell-Cell Contact, mapping mainly to KEGG pathways (Supplementary Figure 3B). Furthermore, detailed ligand-receptor pair analysis highlighted specific signaling axes, such as MIF-(CD74+CD44), MDK-LRP1, and various LGALS9-related interactions (Supplementary Figure 3C).

Figure 7.

Single-cell transcriptomic landscape, intercellular communication, and in silico knockout of UGCG. (A) The t-SNE plot visualizing the 9 distinct immune cell subsets identified in the scRNA-seq data. (B) Bar chart illustrating the proportional changes of each cell subpopulation between the control and sepsis groups. (C and D) Cellular communication networks inferred by CellChat, displaying the number of interactions (C) and strength (D) among different cell types. (E) Volcano plot displaying the systemic transcriptomic changes upon UGCG perturbation. (F) Bar chart illustrating the top 20 DEGs with the highest fold changes (FC). (G and H) Bubble plots showing the GO functional enrichment analysis (G) and KEGG pathway enrichment analysis (H).

Among the three identified diagnostic hub genes, UGCG was selected for further experimental validation due to its close involvement in sphingolipid metabolism and strong relevance to the lipid metabolism focus of this study. In addition, UGCG showed relatively high diagnostic performance across datasets and the most pronounced expression alteration following LPS stimulation in our preliminary cellular models. The virtual knockout of UGCG triggered significant transcriptome changes, with the 20 genes showing the most significant differences as shown in Figure 7E and F. GO enrichment analysis showed that significantly enriched biological processes were concentrated in T cell differentiation, lymphocyte differentiation, and T cell receptor signaling pathways (Figure 7G). KEGG pathway analysis revealed a high enrichment of Th1, Th2, and Th17 cell differentiation, as well as the T cell receptor signaling pathway (Figure 7H).

In vitro Validation of UGCG and Lipidomic Analysis

We then conducted more in-depth in vitro experiments on UGCG. First, the knockdown efficiency of Ugcg in RAW 264.7 cells was confirmed by RT-qPCR. Compared with the negative control, Ugcg mRNA expression was significantly reduced in cells transfected with si-Ugcg (p < 0.05) (Figure 8A). Functional assays demonstrated that si-Ugcg was able to inhibit the release of LPS-induced inflammatory cytokines (p < 0.001) (Figure 8B). To elucidate the underlying metabolic mechanisms, we conducted lipidomic profiling, in which a total of 527 lipid species were reliably quantified across all samples. PCA revealed distinct lipidomic signatures among the NC, LPS-stimulated, and si-Ugcg groups (Figure 8C). A heatmap of the top 50 differentially regulated lipid species further illustrated distinct lipid abundance patterns among the three groups (Figure 8D), and all 50 lipid species displayed had adjusted q-values < 0.01 after Benjamini–Hochberg correction. Notably, Ugcg knockdown markedly attenuated LPS-induced lipid dysregulation. Specifically, the pathological accumulation of highly toxic upstream ceramides, such as Cer(d18:1/23:0) and Cer(d18:2/16:0), was significantly reduced (Figure 8E–F). This was accompanied by a profound, blockade-induced depletion of downstream hexosylceramides (eg, HexCer(d18:1/22:1)) (Figure 8G). We also observed a compensatory metabolic shunting upon UGCG inhibition, evidenced by the marked accumulation of neutral lipids, including triglycerides (TG) and diglycerides (DG) (Figure 8H and I). These experimental results provided further support for our bioinformatics predictions, suggesting that UGCG may contribute to myeloid-associated inflammatory responses through its potential involvement in intracellular lipid homeostasis regulation.

Figure 8.

In vitro validation of UGCG function and lipidomic remodeling. (A) RT-qPCR validation of Ugcg knockdown efficiency in RAW264.7 macrophages using three distinct siRNAs. (B) Relative mRNA expression levels of IL-1β, IL-6, and TNF-α in the Control, LPS+si-NC, and LPS+si-Ugcg groups. (C) PCA score plot of the lipidomic profiles. (D) Heatmap illustrating the distinct abundance patterns of differentially regulated lipid species across the three groups. All 50 lipid species displayed in the heatmap had an adjusted q-value < 0.01 after Benjamini–Hochberg correction. (E and F) Relative abundance of representative upstream neurotoxic ceramides: Cer(d18:1/23:0) (E) and Cer(d18:2/16:0) (F). (G) Relative abundance of the downstream product HexCer(d18:1/22:1). (H and I) Relative abundance of representative neutral lipids, TG(48:0-FA16:0) (H) and DG(18:0/18:2) (I). All in vitro experiments and lipidomic analyses were performed with at least three independent biological replicates. Data are presented as mean ± SD. Statistical significance between two groups was determined by Student’s t-test. For lipidomic data, differences among the three groups were assessed by one-way ANOVA, followed by Benjamini–Hochberg correction for multiple comparisons. ** p (or q) < 0.01, *** p (or q) < 0.001.

Discussion

Sepsis is a highly lethal organ dysfunction syndrome characterized by dysregulation of the host immune response.^41,42 Research has shown that sepsis is typically accompanied by profound systemic metabolic reprogramming.^12,43 Among these changes, lipid metabolism dysfunction is now considered a primary cause of immune dysregulation.⁴⁴ Therefore, studying LMRGs offers a highly promising avenue for identifying candidate diagnostic biomarkers and potential therapeutic targets.

In this study, we first used RNA-seq data to demonstrate that metabolic and immune dysregulation are closely linked in sepsis. GSEA analysis revealed that sepsis is characterized by the concurrent upregulation of innate inflammatory cascades and lipid metabolism pathways, as well as severe suppression of adaptive immune responses such as antigen presentation. To investigate these distinct metabolic changes, we classified sepsis patients into two groups based on LMRG. Group C2 exhibited very high activity in baseline lipid metabolism, particularly in phospholipids, sphingolipids, and glycerophospholipids. This metabolic overactivation is intricately linked to distinct patterns of immune cell infiltration. This demonstrates that changes in lipid metabolism in sepsis are closely associated with the immune environment.

Subsequently, we integrated multiple machine learning algorithms to identify robust lipid metabolism-related diagnostic markers for sepsis. These findings are particularly relevant in the context of sepsis research, as prior transcriptomic studies have extensively characterized generalized immune suppression and cytokine storm signatures in sepsis,^45,46 while the upstream metabolic drivers underlying these immune phenotypes often remain underexplored. By specifically extracting a lipid-driven transcriptomic signature, our approach uncovers a metabolic dimension that has been largely implicit in previous transcriptomic investigations. Second, our findings align closely with the growing body of sepsis metabolomics research. Robust metabolomics-based studies have consistently demonstrated that sepsis is characterized by profound systemic metabolic reprogramming and widespread lipid alterations.^13,47 The hub genes identified in the present study provide a critical upstream transcriptional framework that helps explain the aberrant lipid fluxes and lipotoxicity documented in these metabolomic profiles.

From a clinical translation perspective, the three-gene signature shows potential value for clinical application. Unlike whole-blood transcriptomic sequencing or high-throughput mass spectrometry, which are resource-intensive and time-consuming, evaluating a three-gene panel via RT-qPCR is readily achievable in standard clinical laboratories. The assay turnaround time is rapid, typically yielding quantitative results within 2 to 4 hours. Furthermore, the low cost and technical accessibility of PCR implementation allow it to be seamlessly integrated into routine ICU workflows. Upon ICU admission, this targeted panel could be run alongside standard blood draws, providing clinicians with a rapid, metabolism-driven risk stratification tool to guide timely therapeutic interventions.

ALDH9A1 is a member of the ALDH gene superfamily, which consists of 19 functionally active genes that oxidize aldehydes to their corresponding carboxylic acids.¹⁸ MBOAT2 is a membrane-bound acyltransferase belonging to the MBOAT family, primarily involved in the acylation modification of lipid molecules.¹⁹ It participates in cell membrane lipid remodeling and the generation of signaling molecules by transferring fatty acids to substrates.⁴⁸ UGCG is an enzyme that catalyzes the initial step of glycosphingolipid synthesis by converting ceramide into glucosylceramide.¹⁷ The glycosphingolipids generated by UGCG are essential components of the cell membrane, contributing to the stabilization and amplification of T cell receptor signaling, thereby supporting T cell activation and effector functions.⁴⁹ UGCG is also involved in cell proliferation, differentiation, and signal transduction, and its dysregulation is associated with various pathological processes, including tumors and inflammation.^50,51 Our findings indicate that MBOAT2 and UGCG are significantly and strongly positively correlated with the infiltration of key myeloid cells, particularly neutrophils and macrophages. GSVA analysis suggested that this lipid-driven immune expansion is closely associated with transcriptomic signatures of membrane remodeling and metabolic hyperactivation. Concurrently, the altered expression of these two enzymes was enriched in pathways related to severe cellular stress, such as mitochondrial cytochrome c release and apoptotic DNA fragmentation. Overall, we hypothesize that MBOAT2 and UGCG may be involved in the tissue infiltration and pro-inflammatory polarization of macrophages and neutrophils, and may be related to the apoptotic phenotype of these immune cells during sepsis progression. In addition, we analyzed intercellular interactions and the dynamic immune microenvironment in sepsis using scRNA-seq. The virtual knockout simulation suggested that UGCG perturbation significantly altered transcriptional programs associated with T-cell receptor signaling and helper T-cell differentiation. Therefore, UGCG in myeloid cells may disrupt membrane organization by regulating metabolism, thereby affecting their interaction with T cells and leading to immune dysregulation during sepsis. However, these potential mechanisms remain speculative and warrant careful experimental confirmation in future studies.

Translating these molecular findings into therapeutic strategies is a key objective of this study. Interestingly, our molecular docking analysis indicated that BRD-K52560704 exhibits favorable binding affinity with both UGCG and MBOAT2. However, it is crucial to emphasize that our DrugRefLector predictions and molecular docking results are strictly exploratory and hypothesis-generating. Therefore, rigorous empirical in vitro and in vivo experiments are strictly required in future studies to ascertain the true therapeutic efficacy of BRD-K52560704.

To complement our bioinformatics analysis, we conducted preliminary in vitro experiments using a macrophage model. We first confirmed that the expression trends of ALDH9A1, MBOAT2, and UGCG following LPS stimulation were consistent with our transcriptomic predictions. We found that knocking down Ugcg significantly attenuated the LPS-induced release of pro-inflammatory cytokines. Subsequent lipidomics indicated that this anti-inflammatory effect was closely associated with an altered intracellular balance between toxic ceramides and downstream glycosphingolipids. By mitigating the accumulation of upstream lipotoxic mediators, Ugcg silencing appeared to dampen the inflammatory signaling cascade. These preliminary in vitro findings indicate that UGCG may be involved in lipid-mediated inflammatory responses in sepsis, supporting its value as a candidate therapeutic target for further investigation.

However, several limitations should be acknowledged in this study. First, the heterogeneity among datasets, including differences in patient characteristics, infection sources, disease severity, and sequencing platforms, may introduce potential biases despite the use of standardized bioinformatic approaches and batch correction. Second, the diagnostic model requires further validation in larger prospective clinical cohorts to confirm its clinical applicability. Third, our experimental validation was limited to macrophage models, and the effects of UGCG in other immune cell types or in vivo systems remain to be investigated. Finally, further studies are needed to elucidate the detailed molecular mechanisms underlying the role of UGCG in sepsis progression.

Conclusion

In summary, we have developed a candidate three-gene diagnostic model for sepsis. Through multi-omics integration analysis, we have highlighted a potential mechanism by which UGCG may contribute to the innate immune dysregulation in sepsis through sphingolipid reprogramming. These findings provide a candidate diagnostic biomarker model and potential therapeutic targets for combating sepsis.

Funding Statement

The authors declare that no funding was received for this study.

Data Sharing Statement

The datasets analyzed in this work can be found in the GEO database (https://www.ncbi.nlm.nih.gov/gds). The data and codes used in this study are available from the corresponding author upon reasonable request.

Ethics Approval and Consent to Participate

Human-related sequencing data analyzed in this study were retrieved from the public Gene Expression Omnibus (GEO) database. All included data were de-identified to protect individual privacy, with original collection and release approved by the relevant Ethical Review Committee (ERC) as stated in the database. In line with Articles 32 (1) and 32 (2) of the Measures for Ethical Review of Life Science and Medical Research Involving Human Subjects (National Health Commission, China, 2023), this study was exempt from additional ethical review, as it only involved bioinformatics analysis of existing public data without new human subjects, biological sample collection, or personal information disclosure.

Author Contributions

Fengchen Lv; Conceptualization, Investigation, Writing – original draft; Xin Wang; Conceptualization, Investigation, Writing – original draft; Li Hu; Data curation, Formal analysis, Writing – review & editing. All authors gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.

Disclosure

The authors report no conflicts of interest in this work.