Single-cell transcriptomic analysis deciphers heterogeneity and transcriptional regulatory programs of sepsis with different prognosis

Abstract

Sepsis has a high mortality rate, yet the cellular heterogeneity and transcriptional regulatory programs associated with divergent clinical outcomes remain incompletely understood. Here, we integrated two peripheral blood single-cell RNA-seq cohorts to explore prognosis-associated immune remodeling patterns in sepsis. Using reference-based annotation, pySCENIC-inferred regulon activity, pathway enrichment, and CellChat based ligand-receptor inference, we observed broad differences in cell composition, transcriptional programs, and intercellular communication between survivors and non-survivors. Non-survivors exhibited relative decrease of monocytes, B-cells, NK cells, and CD4/CD8 T cells, together with relative platelet expansion. cDC2 and plasmablasts showed relatively large transcriptional disturbance compared to other cell types. In cDC2, poor outcome was associated with increased TNF-α and NF-κB related regulon activity, which linked to AP-1 transcription factors (JUN, FOSL2, CEBPB, NFIL3, KLF6, and FOSB), together with reduced STAT1 and STAT2-associated interferon signaling. Gene regulatory network analysis highlighted cell type-specific transcription factor and target gene relationships. CellChat suggested that cDC2 may occupy a more connected position in survivors, whereas connectivity appeared reduced in non-survivors. Independent bulk transcriptomic validation supported increased FOSL2 and CEBPB expression, and a multivariable eight-transcription-factor model showed preliminary discriminatory performance. Overall, this study suggests that poor-outcome sepsis may be associated with altered cDC2 regulatory states and reduced cDC2 centered intercellular coordination.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13062-026-00836-x.

Introduction

Sepsis is a severe organ dysfunction caused by an abnormal host response to infection that poses a significant global threat to healthy individuals [1]. Early identification and diagnosis are crucial for improving outcomes, facilitate timely administration of antibiotics and supportive therapies [2]. Despite progress in critical care, sepsis remains a major cause of mortality and morbidity worldwide [3], mainly because of the heterogeneous nature of patients, complex pathological mechanisms and limited predictive value of conventional biomarkers such as C-reactive protein (CRP), procalcitonin (PCT), interleukin-6 (IL-6), and lactate, which often lack sufficient sensitivity or specificity for individual outcome prediction [4]. Emerging data indicate that sepsis involves simultaneous and dynamic processes of increased inflammation and immune suppression [5]. Nonetheless, the precise regulatory mechanisms of cell types and changes in intercellular communication remain poorly understood.

Single-cell transcriptomics enables detailed exploration of immune heterogeneity; however, most previous studies have emphasized changes in cell composition or marker expression, whereas the regulatory programs that shape cellular states, and the inferred communication networks linking these states, remain less well characterized [6–8]. In particular, identifying transcription factor-associated regulatory programs in specific immune subsets, and understanding how these subsets may differ in their inferred interaction patterns across clinical outcomes, could provide clinically relevant insight into prognosis-associated immune remodeling.

This study combines high-resolution single-cell regulatory mapping with intercellular communication inference and independent bulk validation to address the current gaps. We assessed regulatory disturbances at the cellular level, identified cell type-specific regulons, and decomposed the transcription factor variance into elements associated with cellular identity and clinical consequences. Parallel ligand-receptor network analysis elucidated the alteration of signaling links in relation to prognosis, defined as 28-day all-cause mortality. We further evaluated the critical regulatory markers in an independent bulk cohort to assess their potential prognostic relevance. These findings reveal cell-type-specific regulatory changes and disruption of intercellular coordination in fatal sepsis, suggesting possible biomarkers and potential therapeutic targets for further investigation.

Methods

scRNA data collection and basic analysis

Two public single-cell RNA sequences datasets were obtained from the Gene Expression Omnibus database (GEO, https://www.ncbi.nlm.nih.gov/geo/) with the accession numbers GSE167363 [6] and GSE175453 [9]. These datasets comprised samples from sepsis patients and healthy controls. For integrated analysis, the integrated dataset included 7 healthy controls (7 samples), 7 survivors (10 samples), and 2 non-survivors (4 samples). As some patients contributed longitudinal samples across timepoints, the number of samples exceeded the number of unique patients in the survivor and non-survivor groups. All analyses were performed in R (v.4.3.2) using Seurat (v.4.4.0), including quality control, clustering, reference mapping, and differential gene expression analysis. To remove low-quality cells, empty droplets, doublets, and erythrocyte-contaminated cells, the following filtering criteria were applied: nFeature.RNA > 200, nCount.RNA < 40,000, percent.mt < 15%, percent.ribo < 40%, and percent.hb < 25%. These thresholds were selected based on established best practices in single-cell immune profiling [10, 11] and the quality distributions observed in our datasets (see Supplementary Fig. 1A–C), thereby ensuring the removal of low-quality cells, while preserving biologically relevant immune populations.

The gene expression matrix was log-normalized using NormalizeData. The top 2000 highly variable genes were identified utilizing FindVariableFeatures function with the variance-stabilizing transformation method, and the data was scaled employing ScaleData. Principal component analysis (PCA) was then performed on the highly variable genes to reduce dimensionality. To mitigate inter-sample batch effects, Harmony (v1.2.1) was applied to the PCA embeddings using sample identity as the batch variable. The Harmony-corrected coordinates were then used for downstream clustering and UMAP visualization. Cell clustering was conducted utilizing the FindNeighbors and FindClusters functions at a resolution of 0.5, and the results were visualized with the RunUMAP function.

Reference mapping

The traditional approach of manual categorization and annotation in single-cell analysis is labor-intensive and subjective, diminishing the reproducibility of outcomes [12]. With the progression of data integration methodologies and the availability of extensive body- and organ-scale reference atlases, researchers have proposed mapping novel datasets onto well-curated reference atlases as a method for data integration [12]. In this study, cell types were labeled by reference mapping utilizing the MapQuery function, which aligns query data with reference atlas [13] using anchors determined by the FindTransferAnchors function. The predicted.score parameter was used as a metric to assess the quality of the reference mapping for each cell. Cells with a predicted mapping score < 0.4 were excluded, as this threshold was chosen to remove low-confidence annotations [13]. To further improve annotation robustness, we employed a three-tiered validation strategy: (i) identification of cluster-specific marker genes using FindAllMarkers; (ii) comparison with canonical markers reported in the original datasets; and (iii) and experimentally curated markers from the CellMarker 2.0 database. We confirmed consistent expression patterns and alignment with reference atlas projections.

pySCENIC

To infer transcription factor-associated regulatory programs, we applied the SCENIC framework [14], which comprises co-expression analysis, cis-regulatory motif enrichment and regulon activity scoring. Due to the high sparsity and large number of cells in single-cell data, we first generated metacell using a graph-based clustering approach, requiring a minimum metacell size of 10 cells [15], with a resolution of 50 was used to reduce cell numbers, eliminate noise, and improve computational efficiency. This yielded metacell with a median size of ~ 50 cells (range: 10 ~ 200), which balances dropout reduction with preservation of biological resolution of distinct immune states. Lower resolutions merged biologically relevant subsets, whereas higher resolutions produced metacell too small for robust regulon inference. GENIE3 was used to infer co-expression modules, each consisting of TF and its co-expressed candidate targets. Subsequently, these modules were filtered with RcisTarget to retain only those in which the TF’s binding motif is significantly enriched; the TF together with its filtered targets was then defined as a regulon. Regulon activity was then quantified in each cell using AUCell, generating a regulon activity score (RAS) matrix. To rank TFs by cell type specificity, we used the Regulon Specificity Score (RSS), calculated as: RSS = 1 - Jensen Shannon Divergence (regulon_AUC, cell_type_indicator), where: regulon_AUC is the vector of regulon activity scores across all cells; cell_type_indicator is a binary vector (1 for target cell type, 0 otherwise). The resulting activity profiles were used for downstream clustering and cell state characterization based on shared regulatory subnetworks.

Disturbance evaluation of cells

To quantify transcriptional perturbation at the cell-type level, we defined a cellular disturbance score based on distances in PCA space derived from the regulon activity score (RAS) matrix. For each cell within a given cell type, the disturbance score was defined as the Euclidean distance from the reference centroid in PCA space. The reference state was defined as the centroid of cells from the reference group within that cell type. Specifically, for each cell type, we first computed a PCA embedding using the normalized expression matrix. We then defined the reference state as the centroid of cells from the reference group within that cell type in the leading principal components. To reduce the influence of unequal cell numbers across groups, we additionally performed repeated subsampling. For each comparison, 20 cells were randomly sampled per group, repeated 20 times, and pairwise distances were computed within the same PCA space. The median distance for each subsample was calculated, and the average of these median distances was used as a summary estimate of disturbance.

Variance decomposition of regulon activity

To identify transcription factors whose regulatory activity was explained more variance by clinical outcome than cell type identity, we performed variance decomposition using a mixed linear model implemented using the lme4::lmer function with the model: regulon activity ~ (1|celltype) + (1|survival_status). For each regulon, we partitioned its activity variance into components attributable to ‘cell type’ and ‘survival status’, which were treated as random effects to estimate their respective contributions to total variance. TFs with more than 20% of variance explained by survival status were considered prognosis-associated.

Pathway enrichment analysis and gene regulatory network

Differentially expressed genes between survivors and non-survivors within specific cell types were identified using FindMarkers. Gene set enrichment analysis (GSEA) was conducted to examine the significantly enriched pathways in MSigDB Hallmark gene sets, using the GSEA function in the clusterProfiler package (v.4.14.4). Pathways with Benjamini–Hochberg-adjusted P values less than 0.05 were considered significantly enriched. To further identify TFs linked to pathways of interest, the enricher function in clusterProfiler was executed with the parameters TERM2GENE = regulons, minGSSize = 10, and maxGSSize = 5000. Subsequently, the gene regulatory network of the relevant pathways was inferred using the transcription factors, target genes, and pySCENIC results.

Cell-cell interaction analysis

To explore the potential intercellular communication patterns, we constructed separate CellChat objects for survivors and non-survivors using CellChat (v2.2.0). The CellChatDB.human dataset was used, and the analysis was restricted to the secreted signaling category. The highly expressed ligands and receptors were identified using the identifyOverExpressedInteractions function. Communication probabilities were inferred using the computeCommunProb function with default parameters. The non-survival and survival cellchat objects were compared using the compareInteractions function to summarize differences in the number and inferred intensity of interactions between cell populations.

Bulk-RNA seq analysis

The GSE95233 bulk RNA-seq dataset was used for external validation of expression levels. The expression of each TF was first evaluated in the validation cohort. To assess the combined prognostic value of the eight transcription factors (TFs), we constructed a multivariable model using the bulk RNA-seq dataset. For each sample, normalized expression values of the eight TFs (CEBPB, FOSB, FOSL2, JUN, KLF6, NFIL3, STAT1, and STAT2) were used as predictors. A logistic regression model was fitted with survival status (survivor vs. non-survivor) as the outcome variable. Model-predicted probabilities were used to generate receiver operating characteristic (ROC) curves using the pROC package (v. 1.18.5) to assess model performance. The single sample gene set enrichment analysis (ssGSEA) approach was employed to evaluate the relative abundance of 28 unique immune cell types using the GSVA package (v.2.0.5), with the gsva function with method = “ssgsea” and kcdf = “Gaussian”.

Results

Heterogeneity landscape of Sepsis with different prognostics

To investigate the heterogeneous landscape of sepsis with differing prognoses at single-cell level, we performed an initial analysis of the integrated single-cell RNA sequencing data from two datasets, GSE175453 (HC: n = 5; sepsis: n = 4) and GSE167363 (HC: n = 2; sepsis: n = 5). The GSE167363 data originated from sepsis patients at the time of sepsis detection and those diagnosed within 6 h, and two of the non-survivor samples were exclusively present in GSE167363. GSE175453 originated from late-stage sepsis patients identified 14–21 days after the onset of sepsis. Therefore, pooled survivor/non-survivor differences may partially reflect disease stage and cohort-specific sampling structure.

We conducted quality control procedures and employed the fundamental Seurat methodology to remove low-quality cells and minimize RNA contamination (Fig. 1A, sFig.1 A, B). Subsequently, a reference mapping approach was employed to annotate cell cluster in accordance with a predefined cell atlas (sFig.1 C). Cells with predicted scores less than 0.4 were considered low quality and were omitted from subsequent analysis (Fig. 1B, sFig.1D). In total, 77,331 cells were assessed in this study (HC, 30,329 cells; sepsis survivors, 36,898 cells; sepsis non-survivors, 10,104 cells). Based on the specified cell types from the referenced data, these cells were categorized into 31 distinct cell types (Fig. 1C, sFig.1B). The dot plot illustrates the expression of the markers of different cell types (Fig. 1F).

Fig. 1.

Single-cell analysis revealed heterogeneity of sepsis. (A) UMAP for each sample and distinct group. (B) Quality evaluation of reference mapping after to filtering with Predicted.score ≥ 0.4. (C) Density plot for the query data projected onto the reference atlas. (D) Bar chart depicting the proportion of cells in each sample. (E) Alluvial plot of different group. (F) Characteristic markers for various clusters

The density of multiple cell subsets was mapped onto the cell embedding of the reference data to to visualize differences across different sepsis prognoses. It revealed that monocytes, B intermediate cells, B naïve cells, NK cells, CD4 T cells, and CD8 T cells were reduced in non-surviving sepsis patients relative to surviving patients and healthy controls, whereas platelets were relatively increased in non-surviving patients, accompanied by a relatively decreased in the percentage of platelets in surviving patients (Fig. 1C, D, E). Heterogeneities in cellular components were observed among all three groups (Fig. 1E), and cellular composition within the same patient at various time points also showed substantial variability (Fig. 1D).

Cell disturbance and candidate regulon programs in cDC2 and plasmablasts

The state and role of the cell are intricately connected to the function of TFs. To examine regulatory-state variation across groups, we projected the regulon activity score (RAS) matrix into PCA space. PCA results of RAS revealed distinct cell type differentiation in the plot; variations in RAS among patients with various prognoses were also apparent (Fig. 2A). To quantify the changes in transcriptional disturbance among the different groups, we then quantified transcriptional disturbance using the PCA-distance-based metric. The results from 20 random subsampling iterations, revealed that cDC2 showed the highest disturbance estimates, followed by plasmablasts (Fig. 2B). It should be note that cDC2 were highly imbalanced between groups, with substantial representation in healthy controls and survivors but extremely sparse representation in non-survivors (HC = 792, S = 321, NS = 10).

Fig. 2.

Regulon activity. (A) UMAP for principal component analysis with regulon activity score (RAS). (B) Violin plot of cellular disturbance assessed by RAS. (C/D) The top five transcription factors specific to cDC2 and plasmablasts are labeled among various groups. (E/F) Regulon exhibiting significant activity among various groups in cDC2 and plasmablasts

We then examined candidate regulons with high cell-type specificity in cDC2 and plasmablasts across healthy controls, survivors, and non-survivors. In cDC2 cells, RFX5 and USF1 showed high regulon specificity across all groups, suggesting that these TF-associated regulatory programs may be linked to maintenance of cDC2 identity regardless of disease status (Fig. 2C, E). This evidence indicates that RFX5 and USF1 are crucial for maintaining the cellular identity of cDC2, regardless of the health status or disease conditions. HESX1 was observed in both survivor and non-survivor sepsis groups, consistent with possible involvement in sepsis-associated cDC2 remodeling (Fig. 2C, sFig.2), suggesting a potential role in the pathogenesis of sepsis. POU2F1 appeared enriched in non-survivor cDC2 cells (Fig. 2C, E), indicating a association between POU2F1 and an unfavorable sepsis prognosis.

In plasmablasts, MIXL1, BHLHE41, and MEOX1 were shared across healthy controls, survivors, and non-survivors, suggesting that these regulons may contribute to preservation of plasmablast identity across conditions (Fig. 2D, sFig.2), indicating that these factors are required to preserve the cellular identity of plasmablasts, irrespective of health state or illness condition. ELF3 and TCF3 showed higher regulon specificity in non-survivors, indicating possible association with adverse prognosis-related plasmablast states (Fig. 2D, F), suggesting their potential involvement in adverse prognostic states of the condition.

Pathway regulation of cDC2 and plasmablast cells

Analysis of transcriptional disturbance revealed differences in TF activity between survivors and non-survivors, especially among cDC2 and plasmablast cells. We further examined pathway-level changes in these two cell populations.

In cDC2 cells, GSEA results revealed that the majority of HALLMARK pathways were activated, whereas interferon alpha and gamma, MYC targets, and allograft rejection pathways appeared relatively suppressed (Fig. 3A). In active pathways, TNF-α signaling via NF-κB represented the largest gene ratio in non-survivors, whereas the interferon alpha response showed the most substantial gene ratio in repressed pathways (Fig. 3A). We identified the TFs associated with TNF-α signaling via NF-κB and interferon alpha response pathways by both co-expression and significant motif enrichment in cis-regulatory regions. Regulon-based enrichment suggested that JUND, CEBPB, FOSL2, JUNB, FOSB, JUN, KLF6, NFIL3, RELB, and BHLHE40 were associated with the TNF-α/NF-κB-related regulatory program (Fig. 3B), whereas IRF1, STAT1, IRF7, STAT2, ETV7, IRF9, IRF8, RELB, ATF5, and GABPB1 were associated with the interferon-α response program (Fig. 3C).

Fig. 3.

Enrichment analysis. (A)Dot plot shows the activated and suppressed HALLMARK pathways in cDC2. (B) Gene set enrichment analysis (GSEA) indicates significant activation of TNF-α signaling via NF-κB, whereas the accompanying dot plot illustrates transcription factors that regulate this pathway in cDC2. (C) Interferon-α response is significantly suppressed according to GSEA, and the right dot plot shows transcription factors regulate this pathway in cDC2. (D) Dot plot shows the suppressed HALLMARK pathways in plasmablast. (E/F) MYC targets pathway and MTORC1 signaling are significantly suppressed, and the right dot plot shows the transcription factors regulate this pathway in plasmablast

In plasmablast cells, GSEA revealed a broader downregulation of HALLMARK pathways in non-survivors compared to survivors (Fig. 3D). The MYC-targeted pathway represented the largest proportion of the gene ratios and was similarly suppressed in cDC2 cells. The mTORC1 signaling pathway was inhibited in plasmablasts, whereas it was activated in cDC2 cells (Fig. 3A). Regulon-based enrichment analysis revealed that ILF2, BCLAF1, IKZF1, YY1, TFDP1, XBP1, CREB3, and NFYC were associated with the inhibition of MYC targeting pathway (Fig. 3E), whereas SPI1, ILF2, CREB3, XBP1, MLX, CREB3L2, and IRF4 are associated with the suppression of mTORC1 signaling pathway (Fig. 3F). We examined the differential expression patterns of the mTORC1 signaling pathway in cDC2 and plasmablast cells, revealing that the pathway in cDC2 cells was primary associated with BCLAF1, JUN, KLF6, FOSL2, XBP1, CEBPB, ETS2, TFDP1, NFIL3, and KLF4 (sFig.3 A, B). These findings indicate that the transcription factors of the mTORC1 signaling pathway show distinct differences between cDC2 and plasmablasts.

Prognosis-associated candidate regulons in cDC2

Variance decomposition using a mixed-effects model identified 29 regulons for which more than 20% of explained variance was attributable to survival status (Fig. 4A). Among these, six were linked to the TNF-α signaling via NF-κB program and two to the interferon-α response program (Fig. 4B). A Violin plot depicted the activity of these TFs, demonstrating that the activities of JUN, FOSL2, NFIL3, KLF6, CEBPB, and FOSB were enhanced in non-surviving sepsis compared t with survivors and healthy controls (Fig. 4C, D; sFig.4), whereas STAT1 and STAT2 were markedly reduced (Fig. 4E, F). Analysis of RNA expression patterns showed that, except for CEBPB, which showed a large increase in expression in non-survivors, other TFs did not exhibit substantial alterations at the single-cell RNA level (Fig. 4G, H; sFig.4).

Fig. 4.

Regulon associated with survival and not survival. (A) Intercellular perturbation-related transcription factors are decomposed into cell type-related and survival-related transcription factors. (B) Venn diagram of the intersection of survival-related transcription factors and those regulating the TNF-a and IFN-a pathways. (C/D) Highly active regulon in the TNF-α pathway associated with survival. (E/F) Highly suppressed regulon in the INF-α pathway associated with survival. (G/H) UMAP of regulon activity and RNA expression level among different groups

Gene regulatory network associated with cDC2 and plasmablast pathway programs

A gene regulatory network was inferred to illustrate the transcriptional regulatory relationships between the pathways, genes, and pySCENIC-derived regulons. The gene regulatory network of TNF-α signaling via NF-κB pathway suggests that the six TFs (JUN, FOSL2, NFIL3, KLF6, CEBPB, and FOSB) may act in a coordinated manner, collectively associated with modulation of target genes within the pathway (Fig. 5A). In the interferon-α response pathway, STAT1 and STAT2 exhibit overlapping targets, with STAT1 predominantly regulating the majority of the downstream target genes (Fig. 5B). The mTORC1 signaling pathway shows distinct activation patterns in cDC2 and plasmablast cells. We observed differences in the TFs associated with the mTORC1 signaling pathway in cDC2 and plasmablast cells. The mTORC1 signaling pathway is regulated by FOSL2, JUN, CEBPB, NFIL3, KLF6, ETS2, KLF4, BCLAF1, and XBP1 in cDC2 cells, whereas it is modulated by ILF2, CREB3, XBP1, MLX, and SPI1 in plasmablasts. Although XBP1 was shared between the two cell types, the downstream target genes are largely distinct in various cell types (Fig. 5C, D).

Fig. 5.

Gene regulate network (GRN). (A) GRN of TNF-α signaling via NF-κB in cDC2. (B) GRN of interferon-α response in cDC2. (C) GRN of MTOCR1 signaling in cDC2. (D) GRN of MTOCR1 signaling in plasmblast

Validation with bulk-RNA

We evaluated the expression patterns of the eight TFs at the gene level in the bulk RNA datasets. Our results indicated that FOSL2 and CEBPB are significantly overexpressed in non-survivors, although the expression levels of the other six TFs did not show substantial changes (Fig. 6A), consistent with the single-cell RNA analyses. We also evaluated the potential of these eight TFs as prognostic markers for survival. The area under the curve for each transcription factor was as follows: CEBPB = 0.711, FOSB = 0.512, FOSL2 = 0.739, JUN = 0.535, KLF6 = 0.562, NFIL3 = 0.607, STAT1 = 0.543, and STAT2 = 0.600 (Fig. 6B). A multivariable logistic regression model incorporating the eight TFs showed improved apparent discriminatory performance compared with individual TFs in the available dataset (AUC = 0.888) (Fig. 6C). Because formal internal resampling procedures such as cross-validation or bootstrap optimism correction were not incorporated into the current modeling framework, the AUC should be interpreted as an apparent performance estimate in the available validation dataset rather than a fully overfitting-corrected measure of generalizable predictive accuracy. These results suggest that individual TF have limited prognostic value for survival in sepsis. However, the combined use of the eight TFs exhibited improved predictive value.

Fig. 6.

Bulk-RNA analysis. (A) RNA expression of eight transcription factors between survival and not survival. (B) The receiver operating characteristic curves (ROC) and the area under the curves (AUC) of RNA expression for each transcription factor. (C) The ROC and AUC of the combined eight transcription factors. (D) immune infiltration between survival and not survival. (E) The correlation among different cell groups in immune infiltration

The relative abundance of 28 distinct immune cell types was assessed using the bulk RNA data. The findings demonstrated that the percentage of activated dendritic cells was markedly diminished in non-survivors, and showed positive correlations with NK cells, regulatory T cells, neutrophils, macrophages, γδT cells, mast cells based on spearman correlation analyses (Fig. 6D, E).

Cell communication networks in survival and not survival patients

Cell-cell interactions were based on the expression of ligand-receptor genes in individual cells to clarify intercellular relationships among patients with different prognoses. At the pooled group level, non-survivors showed a reduced number of inferred interactions and overall interaction strength than survivors (Fig. 7A). Heatmap suggested that cDC2 cells participated in fewer and weaker inferred interactions in non-survivors than in survivors (Fig. 7B). Similarly, among the six most transcriptionally disturbed cell populations, both incoming and outgoing communication appeared reduced in non-survivors (Fig. 7C). The primary observation was that cDC2 appeared to form a relatively highly connected population in the survival group, displaying numerous incoming and outgoing signals, whereas reduced inferred connectivity was observed in non-survivors. The two-dimensional graph of incoming and outgoing signals illustrates the differences in the transmission and reception of signals among various cell populations (Fig. 7D). At the pathway level, the stacked plot indicated reduced information flow for multiple signaling pathways in non-survivors (Fig. 7E). We further examined cDC2-centered interactions and found that many inferred ligand–receptor interactions were reduced in non-survivors (Fig. 7F). This results suggested that cDC2 may have severed their connections with other cells in non-survivors. However, these observations should be interpreted with caution, as differences in cell abundance, particularly the extremely limited representation of cDC2 cells in non-survivors, may influence the inferred interaction patterns.

Fig. 7.

Cell communication. (A) Number and strength of inferred interaction between survival and not survival. (B) Differential number and strength of interaction between various clusters in non-survival patients compared with survival. (C) Number of interactions for the top six disturbance clusters. (D) Interaction strength of incoming and outcoming. (E) The information flow of different signals. (F) Signaling from cDC2 to other clusters

Discussion

This study integrated two single-cell RNA sequencing datasets to characterize differences in cell composition, transcriptional regulatory programs, and inferred intercellular communication patterns between sepsis survivors and non-survivors, with complementary validation using the bulk transcriptomic data. Overall, the pooled analysis suggested four major features: (i) a widespread alteration of circulating immune populations, characterized by the depletion of various lymphoid and myeloid subsets and an increase in platelets; (ii) pronounced perturbation of inferred regulatory states in cDC2 cells and plasmablasts; (iii) a cDC2-associated program characterized by stronger TNF-α/NF-κB related regulatory activity and weaker interferon/STAT associated regulatory activity in non-survivors; and (iv) a combined multi-TF expression signature with preliminary discriminatory performance in bulk RNA data. These observations extend current understanding of immune heterogeneity in sepsis and identify testable hypotheses regarding prognosis-associated immune dysfunction.

We observed relative reductions in monocytes, B cells, NK cells, and CD4/CD8 T cells in non-survivors, whereas platelet proportion was elevated. These findings are broadly consistent with prior reports linking sepsis-associated lymphopenia, lymphocyte apoptosis, and platelet activation to disease severity and mortality [16, 17]. Our single-cell mapping further suggests that these alterations are heterogeneous across cell states and patients, supporting the view that hyperinflammatory responses and immunosuppressive features can coexist in severe sepsis, and which may help explain why therapeutic strategies focused solely on suppressing inflammation have often shown limited efficacy [18].

Among the examined cell populations, cDC2 and plasmablasts showed relatively high disturbance estimates in the pooled analysis, suggesting that these populations may be particularly vulnerable to prognosis-associated immune remodeling. Previous studies have reported dendritic cell depletion and dysfunction in sepsis, including impaired antigen presentation and reduced expression of costimulatory molecules [19, 20]. Our analysis is consistent with this literature and further suggests that cDC2 may occupy a comparatively more connected position in the inferred communication network of survivors, whereas their inferred connectivity appears reduced in non-survivors. The potential impairment of the cDC2-associated communication patterns may be relevant to T cell priming and the coordination of innate and adaptive responses, potentially leading to subsequent infections and adverse outcomes in severe sepsis [20]. Potential reasons for cDC2 reduction may include cellular apoptosis, altered migratory patterns, and downregulation of chemokine receptors or ligand expression [21, 22]. The reduced cDC2 interactions may result in impaired co-stimulation, decreased cross-presentation, and weakened orchestration of adaptive responses [23, 24]. Distinguishing these possibilities necessitates longitudinal and spatially resolved data together with validation at the protein level.

In cDC2, the pooled analysis suggested a regulatory pattern characterized by increased TNF-α signaling via NF-κB and decreased interferon-α/γ-associated activity in non-survivors. The “high inflammatory/low interferon” pattern aligns with prior clinical and experimental evidence suggesting that robust pro-inflammatory signaling can coexist with compromised antiviral and antigen presentation mechanisms [25–27], and that diminished interferon signaling correlates with immune paralysis and viral reactivation in critically ill patients [27–29]. Our regulon-based analyses further suggested that AP-1 family members and myeloid-associated regulators, including JUN, FOSL2, CEBPB, KLF6, NFIL3, and FOSB, were linked to the TNF-α/NF-κB-associated program; whereas STAT1 and STAT2 regulon activities, which were associated with the interferon-α/γ pathway, were decreased in non-survivors. Importantly, the differences in STAT1 and STAT2 regulon activity were supported by both cell-level and sample-level statistical comparisons. These trends align with the literature indicating that inhibition of STAT1/2-dependent signaling diminishes antiviral effects [30, 31] and correlates with poorer outcomes in sepsis and critical illness [32].

In plasmablasts, the observed quantity reduction together with suppression of MYC target and mTORC1 signaling pathways is consistent with impaired adaptive immune function in poor-outcome sepsis. MYC regulates B-cell proliferation and antibody synthesis [33], and its inhibition corresponds to clinical findings in sepsis [34]. At the same time, mTORC1 pathway was suppressed in plasmablasts but was activated in cDC2. Although XBP1 was shared as an upstream regulator, their downstream targets exhibit substantial differences between cDC2 and plasmablasts, highlighting that the same regulator may drive divergent transcriptional programs based on cellular context [35]. This fact has practical implications, therapies that modify a specific TF or pathway may have varied, even contradictory, effects across different cell types, advocating for cell-selective or pathway-balanced therapeutic techniques instead of broad immune suppression or stimulation.

Several prognosis-associated TFs showed higher inferred activity in non-survivors without corresponding large changes at the transcript level, whereas CEBPB and FOSL2 were also increased in the bulk RNA dataset. This observation highlights a inportant point: TF activity inferred from SCENIC can capture post-transcriptional modulation, nuclear localization, or signaling-dependent activation, which may not be captured by transcript abundance alone [36, 37]. AP-1 components (FOS/JUN family) and CEBPB have established roles in myeloid cell activation and emergency myelopoiesis [38, 39]; NFIL3 and KLF6 have been linked to myeloid differentiation and inflammatory responses [40, 41]. Reduced STAT1/2 activity is associated with impaired interferon responsiveness [42]. Taken together, the TF signature is consistent with a potential shift toward maladaptive inflammatory programs and away from interferon-mediated host defenses in fatal sepsis. The combined eight-TF signature showed stronger discriminatory performance than any individual TF in the bulk RNA cohort, suggesting that multi-factor regulatory readouts may provide complementary information compared to single-gene markers. However, the reported AUC of 0.888 should be interpreted cautiously. In the current study, the 8-TF model was not fully accompanied by formal internal resampling-based overfitting assessment, and its performance was therefore an apparent estimate in the available dataset rather than a fully optimism-corrected measure. Accordingly, this signature should be regarded as a preliminary prognostic model requiring further internal validation and independent external testing before clinical interpretation.

Our analyses emphasize the restoration of cDC2 and the rebalancing of the inflammatory and interferon pathways as potential therapeutic hypotheses. Restoration of the cDC2 hub functionality may enhance antigen presentation and T cell activation, including targeted cytokine therapy, dendritic cell adoptive transfer, or drugs that safeguard dendritic cell function [43, 44]. Notably, cDC2 cells were extremely sparse in the non-survivor group, which limits the robustness of inferences for this subset. The antagonistic regulation of the TNF-α/NF-κB and interferon pathways may suggest that precise modulation, mitigating harmful hyperinflammation while selectively reinstating interferon signaling, may represent a more targeted therapeutic direction.

Several limitations should be acknowledged in our research. First, the integrated analysis combined two sepsis scRNA-seq datasets generated under different sampling contexts, including early-stage and later-stage sepsis, and therefore some pooled survivor/non-survivor differences may be confounded by disease stage, sampling time, or cohort-specific structure. Second, the number of independent patients was limited, particularly in the non-survivor group (2 patients contributing 4 samples), and repeated sampling does not increase the number of independent biological units. This limitation is particularly relevant for sparse subsets such as cDC2. Third, SCENIC-based regulon activities and CellChat-based communication patterns were computationally inferred from transcriptomic data and do not directly measure protein activation, ligand availability, receptor occupancy, phosphorylation, or true physical cell–cell communication. Fourth, several key analyses, including pathway comparisons and CellChat summaries, were based on pooled cell-level data and should therefore be interpreted as exploratory or hypothesis-generating rather than definitive donor-level inferences. Finally, functional causality remains unresolved: whether cDC2 remodeling, altered inferred communication, or prognosis-associated regulons contribute directly to adverse outcome remains to be tested.

Conclusion

The integrated analysis indicates dendritic-cell-associated immune dysfunction and imbalance between inflammatory and interferon- associated regulatory programs as features associated with poor outcome in sepsis and identifies candidate molecular signatures for future validation in larger, prospectively designed cohorts.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Yuanming Yang performed the bioinformatic analysis and wrote the paper. Yiwei Hua and Suyi Yang provided assistance with data collection and paper writing. Nan Liu and Jun Li conceptualized and supervised this research.

Funding

National Natural Science Foundation of China (82474409); 2024 Scientific Research and Development Cultivation Project of Guangdong Provincial Laboratory of Traditional Chinese Medicine (Hengqin Laboratory, HQL2024PZ004); National Interdisciplinary Innovation Team for Traditional Chinese Medicine (ZYYCXTD-D202406).

Data availability

All data reported in this paper are available upon reasonable request. The public scRNA-seq datasets used in this research were obtained from the Gene Expression Omnibus (GSE175453 and GSE167363).

Declarations

Ethics approval and consent

Not applicable. This study relies on publicly available datasets and in silico analyses; no human subjects, animals, or primary samples were involved. Therefore, ethical approval and consent are not required.

Competing interests

The authors declare no competing interests.