Integration of single-cell and bulk transcriptomics with machine learning identifies LDHA as a lactate-related diagnostic and prognostic biomarker in sepsis

Abstract

Background

Sepsis-induced metabolic dysregulation, particularly abnormal lactate metabolism, is closely associated with disease severity and mortality. This study aimed to systematically identify lactate-related genes involved in sepsis and evaluate their diagnostic and prognostic value using integrated transcriptomic analyses.

Methods

Single-cell RNA sequencing data were analyzed using AUCell, singscore, and ssGSEA algorithms combined with correlation analysis to identify lactate-related genes. A comprehensive machine learning framework incorporating 131 algorithms was applied for feature selection and construction of a diagnostic model. Survival analyses were performed to assess the prognostic significance of candidate genes. The expression of key genes was further validated by qPCR and western blotting in clinical samples and a sepsis animal model.

Results

Sepsis patients were stratified into distinct molecular subgroups exhibiting significant differences in prognosis, clinical characteristics, pathway enrichment, immune infiltration, and immune checkpoint gene expression. Single-cell analysis revealed that MARCO⁺ macrophages exhibited relatively high lactate-related activity compared with most immune populations, while plasma cells showed comparable levels. Sixteen genes were identified at the intersection of macrophage differentially expressed genes and lactate-related genes. Machine learning analysis further identified six core diagnostic genes. Survival analysis demonstrated that high LDHA expression was significantly associated with poor prognosis in sepsis patients. Consistently, elevated LDHA expression was observed in septic animals compared with controls.

Conclusion

By integrating single-cell and bulk RNA-seq data, we developed and validated a novel diagnostic model for sepsis, termed the Lacty Model, and identified LDHA as a key biomarker associated with poor prognosis. These findings highlight the potential clinical relevance of lactate metabolism–related genes in sepsis and provide a foundation for future mechanistic and therapeutic studies.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12879-026-12897-4.

Introduction

Sepsis is a life-threatening organ dysfunction caused by a dysregulated host response to infection and remains a leading cause of mortality in intensive care units (ICUs) worldwide [1–3]. Owing to its complex and heterogeneous nature, patients with sepsis often present with similar clinical manifestations while exhibiting markedly distinct underlying pathophysiological mechanisms [4–6]. This heterogeneity contributes to variable treatment responses and outcomes, underscoring the urgent need to identify reliable biomarkers and mechanistic pathways that reflect disease severity and prognosis.

Emerging evidence indicates that epigenetic regulation plays a central role in sepsis-associated transcriptional reprogramming, particularly within immune cells [2]. Immune effector functions are closely coupled to metabolic remodeling, and alterations in cellular metabolism are now recognized as key drivers of immune dysregulation in sepsis. Therefore, elucidating immune–metabolic pathways may provide novel insights into sepsis pathogenesis and identify potential therapeutic targets for restoring immune homeostasis [7, 8].

Lactate, long considered a metabolic byproduct, is now recognized as an active signaling metabolite involved in immune regulation. Lactate-derived lysine lactylation, a recently discovered post-translational modification, occurs on both histone and non-histone proteins and has been detected across multiple cellular compartments [9]. Accumulating studies suggest that lactylation contributes to inflammatory regulation, immune imbalance, and organ dysfunction in sepsis [9]. Clinically, elevated serum lactate levels are well-established prognostic markers and are strongly associated with increased mortality in sepsis [10]. Clinically, elevated serum lactate levels are well-established prognostic markers and are strongly associated with increased mortality in sepsis [11]. According to the Sepsis-3 definition, persistent serum lactate levels > 2 mmol/L despite adequate fluid resuscitation constitute a key diagnostic criterion for septic shock [12].

Although biomarkers such as procalcitonin (PCT) and C-reactive protein (CRP) are widely used in sepsis management, their diagnostic and prognostic value remains limited and context-dependent [13]. Advances in omics technologies, particularly single-cell RNA sequencing (scRNA-seq), have enabled unprecedented resolution of immune cell heterogeneity and cell-type–specific transcriptional programs in sepsis [14]. Integrating single-cell analysis with bulk transcriptomic data and machine learning approaches offers a promising strategy for identifying robust biomarkers linked to immune–metabolic dysregulation.

In this study, we systematically integrated scRNA-seq and bulk RNA-seq datasets to identify lactate-associated immune cell populations and construct a lactate-related diagnostic model (“Lacty Model”) for sepsis. Using a comprehensive machine learning framework, we identified key lactate-associated genes with diagnostic and prognostic relevance. We focused exclusively on sepsis-related transcriptomic analyses to ensure disease specificity. Our findings highlight LDHA as a potential biomarker associated with sepsis severity and prognosis.

Methods

Data collection

All transcriptomic data were obtained from the Gene Expression Omnibus (GEO) database. scRNA-seq datasets included GSE167363 and GSE175453. Bulk RNA-seq datasets comprised GSE69063 (training cohort), GSE28750, GSE54514, and GSE69528 (validation cohorts). Survival analyses were conducted using GSE65682 (n = 479), while survival status validation used GSE26378, GSE26440, and GSE95233. All sepsis and septic shock cases in this study were defined based on the current sepsis related criteria used in the sequencing dataset. Patients were classified as sepsis cases according to the diagnostic criteria reported in the original studies. Historical datasets followed their original diagnostic criteria rather than Sepsis-3 [15]. To reduce biological heterogeneity and demographic confounding, adult and pediatric cohorts were analyzed independently. Pediatric datasets were used only for external validation rather than direct integration.

Data processing

For single cell data, the analysis is based on Seurat4.2.2 version 4.2.2. Specifically, after reading each data set, quality control is carried out to remove cells with nFeature_RNA <200 or > 5000 or mitochondrial ratio greater than 10. The SCTransform function completes data standardization and normalization. The first 15 principal components were selected for dimensionality reduction clustering, RunUMAP function was used to calculate dimensionality reduction clustering results, and DoubletFinder package was used to remove double cells. Harmony algorithm is used to remove batch effect between data sets and samples. SingleR and previous literature are used for cell annotation. FindAllMarkers function was used to analyze the marker gene of each cell type. The ggplot2 package is used to display umap results.

For bulk RNA data, limma package is used for data standardization and normalization. Survival package and survminar package are used for survival curve analysis and drawing. Survival analyses were limited by retrospective design, heterogeneous clinical annotations, and incomplete covariates, and therefore should be interpreted with caution.

Ro/e analysis

The Ro/e (Ratio of observed to expected) value quantifies the enrichment of a cell cluster in a specific tissue type by comparing the observed cell count to the count expected by chance [16]. The expected cell number of each combination of cell clusters and tissues was obtained by chi-square test. Different from chi-square value, chi-square value is defined as and can only indicate the difference between the observed results and random expectations, and chi-square value defined by Ro/e can indicate whether the cells of a specific cell cluster are enriched or depleted in a specific tissue. For example, if Ro/e > 1, it indicates that the cells of a given T cell cluster are observed more frequently than randomly expected, that is, enriched. If Ro/e < 1, it indicates that the frequency of observing cells of a given cell cluster in a specific tissue is lower than random expectation, that is, exhausted. Through Ro/e calculation, we can effectively quantify the organizational preference of cell clusters [17].

Cell communication

The “Cell Chat” R software package (version 1.5.0) is used to reveal the potential mechanism of cell-to-cell communication within a single cell [18]. Use the createCellChat function to construct the cellchat object, and use the aggregateNet function to describe the signals emitted from each cell type. NetVisual_circle function is used to show the number and weight of communication between cells, and netAnalysis_computeCentrality function is used to infer the input and output weights of specific signal paths. Macrophage subtypes (Mac_CD74+, Mac_MARCO+, Mac_CD83+, Mac_IFIT3+, and Mac_S100A12+) were annotated based on canonical marker genes and reference-based strategies reported in previously published single-cell transcriptomic studies. Detailed subtype definitions followed established literature rather than predefined expression thresholds [19].

Gene set enrichment analysis

To robustly quantify lactate-related metabolic activity at the single-cell level, three complementary gene set–based scoring approaches—AUCell, singscore, and single-sample gene set enrichment analysis (ssGSEA)—were applied. Lactate-related gene sets were curated from published literature and pathway databases associated with lactate metabolism and protein lactylation [20–22]. The complete gene lists are provided in the Supplementary Materials.

All analyses were implemented using the irGSEA R package. Briefly, normalized gene expression matrices were used as input to calculate enrichment scores for each cell based on the relative expression of genes within the predefined signatures. AUCell estimates the area under the recovery curve of gene rankings, singscore computes rank-based single-sample scores, and ssGSEA evaluates enrichment using empirical cumulative distribution functions. These complementary methods were applied to improve robustness and reduce method-specific bias. Scores were calculated for each cell using the irGSEA.score function and subsequently integrated across cell populations with irGSEA.integrate. Visualization was performed using irGSEA.halfvlnplot.

Differential expression analysis and enrichment analysis

Differential expression analysis was conducted using limma, with |log2FC| >1 and adjusted P < 0.05. Gene set enrichment analyses were performed using GSEA and clusterProfiler for GO, KEGG, and Reactome pathways.

Machine learning and diagnostic model construction

To identify robust diagnostic biomarkers and construct a lactate-related diagnostic model, twelve machine learning algorithms—including least absolute shrinkage and selection operator (LASSO), elastic net (ENet), stepwise generalized linear model (Step GLM), support vector machine (SVM), gradient boosting generalized linear model (GLMBoost), extreme gradient boosting (XGBoost), ridge regression, partial least squares regression for generalized linear models (PLSRGLM), random forest, gradient boosting machine (GBM), linear discriminant analysis (LDA), and naïve Bayes—were systematically benchmarked.

To minimize overfitting and improve generalizability, model development was conducted using a nested cross-validation framework. Specifically, the outer loop was used to estimate model performance, whereas the inner loop was applied for hyperparameter optimization using grid/random search. All feature selection and model tuning steps were performed exclusively within the training folds to prevent information leakage.

A total of 131 algorithm–parameter combinations were evaluated. Model performance was assessed using the mean area under the receiver operating characteristic curve (AUC) across folds. The best-performing model was subsequently trained on the full training cohort and independently validated in external datasets. The final model output was defined as the lactate-related diagnostic score (lacty_score), calculated as the weighted combination of selected features according to the optimized model coefficients. Given the exploratory nature of machine learning, results were interpreted cautiously and considered hypothesis-generating rather than causal.

Immune correlation analysis

The relative enrichment scores of 29 kinds of immune cells and immune processes were calculated by using GSVA and GSEABase packages. Based on ssGSEA strategy, sepsis patients were divided into high and low lacty score groups according to the median lacty score, and the enrichment scores of immune cells and immune processes were compared. The expression of chemokine family and TNF family molecules was used to analyze the correlation with the lacty core gene of sepsis, and ggplot2 drew the correlation heat map.

Animals specimens

This study used sixteen healthy, six-week-old male C57BL/6 mouse with body weight 20 ± 2 g obtained from the Guangdong yaokang biology science and technology co., ltd. After all mice were assigned to either the control group (n = 8), lipopolysaccharide (LPS, Escherichia coli O111:B4, Sigma-Aldrich, St. Louis, MO, USA; catalog number L2630) sepsis model group (n = 8) into two groups using a random number table. In this study, a LPS inflammation model was established by intraperitoneal injection of LPS from Escherichia coli O55:B5. The control group was given intraperitoneal injection of 5 ml/kg saline, LPS group was given intraperitoneal injection of endotoxin (LPS) 15 µg/g group. This dose was aimed at inducing a moderate sepsis phenotype, with the severity precisely controlled at a level slightly below 50% mortality (approximately 40%). After successfully constructing the model, blood was collected from the animals and peripheral blood mononuclear cells (PIMCs) were isolated. PIMCs were isolated from anticoagulated whole blood using Ficoll density-gradient centrifugation. Briefly, blood samples were diluted 1:1 with phosphate-buffered saline (PBS), carefully layered onto Ficoll-Paque PLUS (GE Healthcare), and centrifuged at 400 × g for 30 min at room temperature. The mononuclear cell layer was collected, washed twice with PBS, and immediately processed for RNA and protein extraction. Detect the expression of LDHA through subsequent experiments. Euthanize mice 72 h after LPS injection. The study was approved by Guangzhou laian technology co., ltd (Number: G2025004).

Western blot and qPCR

For quantitative real-time PCR (qRT-PCR), total RNA was extracted using TRIzol reagent (Invitrogen, USA) and reverse-transcribed using a PrimeScript RT reagent kit (Takara, Japan). qPCR was performed with SYBR Green Master Mix (Applied Biosystems) on an ABI 7500 Real-Time PCR System. Relative gene expression levels were calculated using the 2^−ΔΔCt method with GAPDH as the internal reference.

For Western blotting, total protein was extracted using RIPA lysis buffer supplemented with protease inhibitors. Equal amounts of protein (20–30 µg) were separated by SDS-PAGE and transferred onto PVDF membranes. Membranes were blocked with 5% non-fat milk for 1 h and incubated overnight at 4 °C with primary antibodies against LDHA (1:1000, Cell Signaling Technology) and GAPDH (1:2000). After incubation with HRP-conjugated secondary antibodies, signals were detected using enhanced chemiluminescence (ECL) reagents and quantified using ImageJ software. Each primer sequence is presented in Table 1.

Table 1.

The each primer sequence information of real-time qPCR

Gene	Forward (5’-3’)	Reverse (5’-3’)
LDHA	TGTGTGGAGTGGTGTGAATG	CTTGATCACCTCGTAGGCAC
GAPDH	GGCCTCCAAGGAGTAAGAAA	GCCCCTCCTGTTATTATGG

Statistical analysis

All statistical analyses were performed using R software (version specified in the Methods). Two-group comparisons were conducted using the Wilcoxon rank-sum test unless otherwise stated. Multiple testing correction was applied using the Benjamini–Hochberg method. Statistical significance was defined as an adjusted P value < 0.05. These details have been added to the Methods section and relevant figure legends.

Result

Single-cell transcriptomic landscape of sepsis

To systematically characterize immune cell heterogeneity in sepsis, two independent scRNA-seq datasets were integrated after batch correction using the Harmony algorithm. UMAP (uniform manifold approximate projection) diagram showed that the samples were well integrated, including 21 samples (Fig. 1A). After quality control, 87,155 cells were retained, including 33,674 cells from controls and 53,481 cells from patients with sepsis. Using an optimized clustering resolution, cells were grouped into 24 clusters (clusters 0–23; Fig. 1B) and subsequently annotated into eight major immune cell types based on canonical marker genes: T cells, monocytes, macrophages, neutrophils, NK cells, B cells, plasma cells, and platelets (Fig. 1C–E).

Fig. 1.

Single-cell landscape and intercellular communication analysis in sepsis. (A) UMAP plot showing integration of two sepsis scRNA-seq datasets after Harmony batch correction across 21 samples. (B) Unsupervised clustering identified 24 cell clusters. (C) Clusters were annotated into eight major cell types using canonical markers. (D–E) Marker gene expression for T cells, monocytes, platelets, NK cells, B cells, macrophages, neutrophils, and plasma cells. After quality control, 87,155 cells were retained (33,674 controls; 53,481 sepsis). (F–G) Ro/e enrichment analysis showed monocytes and T cells enriched in controls, whereas most immune cells—especially neutrophils—were expanded in sepsis. (H) Cell–cell communication analysis revealed strong outgoing signals from monocytes, B cells, and plasma cells, with prominent activation of MIF and CD99 pathways in monocytes and CD23 signaling in B cells

In order to evaluate the cell types that play a key role in the occurrence and development of sepsis, we adopted the R o/e algorithm. R o/e is a commonly used method to evaluate organizational preference, which has been reported in many previous literatures [17, 23]. Monocytes and T cells were relatively enriched in control samples, whereas most immune cell populations—particularly neutrophils—were enriched in sepsis samples, reflecting broad immune activation during sepsis (Fig. 1F–G). Cell–cell communication analysis further indicated that monocytes, B cells, and plasma cells exhibited prominent outgoing signaling activity. Pathway-level analysis showed marked activation of MIF and CD99 signaling in monocytes, while CD23 signaling was predominant in B cells (Fig. 1H).

Identification of lactate-associated immune cell populations

Lactate is an important and concerned pathophysiological process [24–26]. Given the importance of lactate metabolism in sepsis, three complementary scoring methods (AUCell, singscore, and ssGSEA) were applied to evaluate lactate-related activity at the single-cell level. Although lactate-related genes were broadly expressed across multiple immune cell types, macrophages consistently exhibited among the highest lactate-related activity scores, showing significantly higher levels than most cell populations (adjusted P < 0.01), while differences compared with plasma cells were not statistically significant (Fig. 2A–C). Key lactate metabolism–related genes showed cell type–specific patterns: LDHA, which primarily mediates lactate production, was relatively enriched in macrophages, whereas LDHA expression was comparatively higher in T cells (Fig. 2D), indicating distinct metabolic roles across immune subsets rather than uniformly elevated activity in a single population.

Fig. 2.

Lactate score of immune cells and analysis of related gene functions. A-C: Comparison of lactate modification levels in immune cells under different algorithms. D: Heatmap showing the expression profile of lactate-related genes. E-G: Functional enrichment analysis of genes related to lactate score in macrophages

Functional enrichment analysis of macrophages revealed significant upregulation of immune effector process, complement cascade, antigen processing and presentation, and, innate immune system (Fig. 2E–G), suggesting that macrophages exhibit inflammatory activation, whereas the relationship with lactate-related metabolism remains correlative rather than causal.

Subclassification of macrophages and lactate-related gene selection

To further refine macrophage heterogeneity, macrophages were subdivided into five distinct subpopulations characterized by high expression of CD74, MARCO, CD83, IFIT3, and S100A12, respectively (Fig. 3A–B). Proportion analysis demonstrated that MARCO⁺, IFIT3⁺, and S100A12⁺ macrophages were significantly expanded in sepsis samples compared with controls (Fig. 3C). Among these expanded subsets, MARCO⁺ macrophages exhibited the highest lactate-related activity scores (Fig. 3D).

Fig. 3.

Classification of macrophages. A-B: genes with high expression in five subgroups. C: cell percentage diagram. D: lactate score. E: the intersection of differentially expressed genes and lactate genes. F: expression of intersection genes. *P < 0.05, **P < 0.01, P < 0.001, P < 0.0001

Sixteen lactate-related genes were identified by intersecting differentially expressed genes from MARCO⁺ macrophages with curated lactate-associated gene sets (Fig. 3E). Violin plot analysis showed that all 16 genes were significantly upregulated in sepsis compared with controls, indicating a consistent association with the septic state rather than cell-type–exclusive expression (Fig. 3F).

Machine learning–based identification of core diagnostic genes

To prioritize candidate genes with diagnostic potential, a comprehensive machine learning framework incorporating 131 algorithmic combinations was applied. The optimal performance was achieved using glmBoost for feature selection combined with Stepglm (forward) modeling, yielding a mean area under the curve (AUC) of 0.876 (Fig. 4A). Confusion matrix analyses showed prediction accuracies exceeding 80% in both training and validation cohorts, indicating robust classification performance (Fig. 4B–E).

Fig. 4.

Machine learning of 131 combinations constructed by 12 algorithms. A: the AUC value of machine learning model. The correct prediction rate of disease group and control group was verified by training set and verification set (B: GSE69063 dataset, C: GSE28750 dataset, D: GSE54514 dataset, E: GSE69528 dataset). F: the gene expression of six diagnostic core genes screened by the model

Six core diagnostic genes—COX8A, IRAK1, LDHA, NDUFA11, NDUFB11, and NDUFB8—were retained in the final model. Among these, COX8A, LDHA, and NDUFB11 were upregulated in sepsis, whereas the remaining genes showed higher expression in control samples (Fig. 4F).

Construction of the lactate-related diagnostic score

Based on the six core diagnostic genes, a composite diagnostic score (lacty_score) was calculated for each patient. Patients were stratified into high- and low-lacty_score groups according to the median value. Differential expression and Gene Ontology biological process analyses demonstrated that high-lacty_score patients exhibited enhanced activation of myeloid cell–related processes, acute inflammatory responses, and complement pathways (Fig. 5A). KEGG pathway analysis further showed enrichment of Toll-like receptor signaling, complement cascades, and oxidative phosphorylation pathways in the high-lacty_score group (Fig. 5B).

Fig. 5.

Functional enrichment analysis of diagnostic core genes. A: GO function analysis of Lacty_score group, B: enrichment analysis of KEGG pathway in Lacty_score group, C-D: Hallmark analysis

Hallmark pathway analysis revealed that inflammatory pathways, including TNFα- and IL-6–related signaling, were upregulated in high-lacty_score patients, whereas pathways associated with cell cycle regulation and differentiation, such as WNT and E2F signaling, were relatively downregulated (Fig. 5C–D).

Immune landscape associated with lacty_score

Immune infiltration and immune-related functional signatures were evaluated using ssGSEA. Patients with low lacty_score showed higher enrichment of multiple immune cell populations, including B cells, CD8⁺ T cells, and immature dendritic cells. In contrast, macrophages and regulatory T cells (Tregs) were significantly enriched in the high-lacty_score group (Fig. 6A). Correlation analyses revealed heterogeneous relationships between core diagnostic genes and chemokines, as well as TNF family members, suggesting distinct immunometabolic association patterns rather than uniform immune regulation (Fig. 6B–C).

Fig. 6.

Immune correlation analysis. A: ssGSEA analysis, B: correlation analysis of chemokines, C: Molecular correlation analysis of TNF family

Prognostic relevance of LDHA expression

Survival analyses were performed for the six core diagnostic genes. Among them, only LDHA expression showed a consistent and significant association with patient survival across datasets. Patients with high LDHA expression exhibited significantly poorer survival compared with those with low expression (Fig. 7A–F). LDHA expression was further elevated in patients with septic shock, indicating an association with disease severity (Fig. 7G–I). Across three independent cohorts, non-survivors consistently showed higher LDHA expression than survivors, supporting its prognostic relevance in sepsis (Fig. 7J–L).

Fig. 7.

Lactate analysis. A-F: six diagnostic core genes were analyzed for survival, G-I: the expression of LDHA in three datasets, J-L: The expression of LDHA in non-survival patients and survival groups of three datasets

LDHA-associated transcriptional and immune features

To explore transcriptional patterns associated with LDHA expression, patients were stratified into high- and low-LDHA groups. Genes upregulated in the high-LDHA group included sepsis-associated genes such as MMP8 and ARG1 (Fig. 8A). Hallmark enrichment analysis showed predominant activation of inflammatory pathways, consistent with patterns observed in the high-lacty score group (Fig. 8B). Immune correlation analysis demonstrated that LDHA expression was negatively correlated with the enrichment scores and estimated abundance of several adaptive immune cell populations, including T helper cells and B cells, while showing positive correlations with macrophages, Tregs, and immune co-inhibitory signatures (Fig. 8C). These association patterns suggest that elevated LDHA expression may reflect an altered immune landscape or potential immune dysregulation in sepsis, rather than direct immune cell–specific regulation.

Fig. 8.

Mechanism analysis of LDHA in sepsis. A: gene related to LDHA; B: HALLMARK enrichment analysis; C: correlation analysis between LDHA expression and immune process enrichment scores as well as immune cell subset abundance in sepsis

Preliminary validation of LDHA expression

To provide preliminary experimental support for the transcriptomic findings, LDHA expression was evaluated in PBMCs isolated from septic and control mice. Western blotting demonstrated significantly higher LDHA protein levels in the sepsis group compared with controls (Fig. 9A–B). Consistently, RT-qPCR analysis showed increased LDHA mRNA expression in PBMCs from septic mice (Fig. 9C). These results corroborate the bioinformatic analyses at the expression level.

Fig. 9.

Animal experiment verification. A-B: Western blooting, B: PCR analysis

Discussion

Lactate, once thought to be a metabolic waste, has been increasingly discovered in recent years to play an important role as a signaling molecule, energy substrate, and immune regulatory molecule. Its important role in regulating metabolism in the body is gradually being discovered [27, 28]. Elevated lactate levels are important biomarkers of sepsis and positively correlated with sepsis related mortality [29]. However, the specific mechanism of action of lactate in sepsis patients is currently unclear. Lactic exerts its biological activities, in part, through a process termed lactylation modification. This newly discovered post-translational modification targets specific lysine residues, altering the function of both histone and non-histone proteins, and has been shown by recent studies to play an important role in diverse pathological and physiological processes [27, 30].

By integrating scRNA-seq and bulk transcriptomic analyses, our study demonstrates that macrophages—particularly MARCO⁺ macrophages—exhibit the highest lactate-associated activity in sepsis. This finding aligns with previous reports indicating that lactate modulates macrophage polarization and contributes to the transition from hyperinflammation to immunosuppression [31, 32]. The enrichment of lactate-related genes in MARCO⁺ macrophages suggested a potential link between metabolic remodeling and macrophage-driven immune dysregulation.

Using a robust machine learning framework, we identified six lactate-associated genes (COX8A, IRAK1, LDHA, NDUFA11, NDUFB11, and NDUFB8) that collectively formed a diagnostic model with strong predictive performance. Among these, LDHA emerged as the only gene significantly associated with patient survival, highlighting its potential prognostic relevance. LDHA catalyzes the conversion of pyruvate to lactate and has been implicated in aerobic glycolysis, a metabolic feature associated with poor outcomes in both sepsis and cancer [33–35].

To translate these transcriptional features into clinically meaningful indicators, we implemented a comprehensive machine learning framework to prioritize diagnostic genes. By benchmarking multiple algorithms within a nested cross-validation design, we aimed to minimize overfitting and improve generalizability. Nevertheless, machine learning models derived from retrospective transcriptomic data remain inherently susceptible to dataset bias and should be considered hypothesis-generating. The identification of LDHA as the only gene consistently associated with survival across cohorts supports its robustness but does not establish direct pathogenic involvement. Prospective validation and mechanistic studies are required to determine whether LDHA represents a driver of disease progression or simply a surrogate marker of metabolic stress.

Several methodological considerations further contextualize our findings. First, publicly available datasets vary in demographic characteristics, geographic origins, and clinical definitions of sepsis. To reduce potential confounding, adult and pediatric cohorts were analyzed separately and used primarily for independent validation rather than direct integration. Second, survival analyses were limited by incomplete clinical covariates and retrospective design. Third, the LPS-induced mouse model captures aspects of systemic inflammation but does not fully recapitulate the complexity of human polymicrobial sepsis. Therefore, experimental results should be interpreted cautiously.

Despite these limitations, our study highlights several potential clinical implications. Transcriptome-derived metabolic signatures may complement traditional biomarkers such as lactate concentration, procalcitonin, and C-reactive protein by providing mechanistic insight into immune status. In particular, LDHA expression may serve as a candidate marker for risk stratification or patient subgroup identification, which could facilitate precision medicine approaches in sepsis. Future studies integrating metabolomics, proteomics, and functional perturbation experiments will be essential to clarify the causal role of lactate metabolism and evaluate whether targeting LDHA-related pathways offers therapeutic benefit.

In summary, this integrative analysis delineates the immunometabolic landscape of sepsis and identifies LDHA as a lactate-associated biomarker linked to adverse outcomes. Rather than defining a single pathogenic mechanism, our findings provide a transcriptomic framework for understanding how metabolic remodeling intersects with immune dysfunction in sepsis and offer a foundation for future mechanistic and translational investigations.

Conclusion

By integrating single-cell and bulk transcriptomic analyses with machine learning, we constructed a robust lactate-related diagnostic model for sepsis and identified LDHA as a key gene associated with disease severity and prognosis. These findings highlight the potential of LDHA as a biomarker for risk stratification and provide a foundation for future studies exploring immune–metabolic interventions in sepsis.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

NK

Natural killer cell

PLS

Partial least squares regression

LDA

Latent Dirichlet Allocation

PBMC

Peripheral blood mononuclear cell

Author contributions

Conception and design of the work: ZYL, HPS, XHC and CLY. Acquisition, analysis and interpretation of data: ZYL and CLY. Drafting of the manuscript: ZYL. Critical revision for important intellectual content: ZYL. All authors approved the final version of this manuscript.

Funding

Not applicable.

Data availability

The data sets used during the current study are available from the corresponding author on a reasonable request. The datasets generated for this study can be found at (https://www.ncbi.nlm.nih.gov). All public datasets were obtained from open-access repositories, and animal experiments were conducted in accordance with institutional ethical guidelines.

Declarations

Ethics approval and consent to participate

This study was approved by the Ethics Committee of Jiangxi Provincial People’s Hospital. This study abides by the Helsinki Declaration. This study confirms that all methods are carried out according to the ARRIVE guidelines.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.