Abstract

   Sepsis is a condition resulting from the uncontrolled immune response
   to infection, leading to widespread inflammatory damage and potentially
   fatal organ dysfunction. Currently, there is a lack of specific
   prevention and treatment strategies for sepsis across different age
   groups. Programmed Cell Death (PCD) can regulate the enrichment of
   effector immune cells or regulatory immune cells, providing a new
   perspective for immunotherapy. Within the framework of computational
   biology and machine learning strategies, and against the backdrop of
   global multicenter sepsis cohort data, this study aims to deeply mine
   and screen specific biomarkers related to the immune microenvironment
   and programmed cell death in populations across different life stages
   (neonates, children, and adults). This will provide foundational data
   for precision treatment and drug development in artificial
   intelligence-assisted sepsis diagnosis and treatment management. Gene
   expression data from sepsis patients across global multicenter
   populations, including China, Europe, and the United States, were
   obtained from the Gene Expression Omnibus (GEO) database, and
   differentially expressed genes (DEGs) were identified. A literature
   review was conducted to obtain 18 PCD-related genes, which were
   intersected with DEGs to identify DEGs associated with specific types
   of PCD. Nine machine learning algorithms (Logistic Regression LR,
   Decision Tree DT, Gradient Boosting Machine GBM, K-Nearest Neighbors
   KNN, LASSO, Principal Component Analysis PCA, Random Forest RF, Support
   Vector Machine SVM, and XGBoost) were applied to training and testing
   datasets with 10-fold cross-validation to select three optimized
   algorithm models. The SHAP algorithm was further used to quantify the
   contribution of each gene based on cell death features to the
   prediction of sepsis. Key PCD patterns were identified based on model
   evaluation metrics (Accuracy, Precision, Recall, F1 score, and Receiver
   Operating Characteristic Curve ROC), and their associated DEGs were
   obtained through intersection, followed by immune-related analysis of
   DEGs. The study included a total of 1507 sepsis cases and 484 controls
   globally, with 90 neonatal cases and 95 controls, 527 children cases
   and 101 controls, and 890 adult cases and 288 controls. The best model
   for predicting sepsis across different populations was GBM.The key PCD
   patterns selected by machine learning for different age groups were
   Pyroptosis (neonates), Ferroptosis (children), and Autophagy (adults).
   (1) In neonatal sepsis, the models constructed by GBM, XGBoost, and RF
   algorithms performed the best, and identified 5 key DEGs associated
   with Pyroptosis (CHMP7, NLRC4, AIM2, GZMB, PRKACA), with NLRC4 showing
   the best predictive ability (AUC = 0.902, P < 0.05), significantly
   positively correlated with neutrophils and negatively correlated with
   CD8 + T cells. (2) In the children sepsis population, models
   constructed using the Gradient Boosting Machine (GBM), Support Vector
   Machine (SVM), and Least Absolute Shrinkage and Selection Operator
   (LASSO) algorithms demonstrated the best performance. Six key DEGs
   associated with Ferroptosis were identified (AKR1C3, GCLM, PEBP1, CARS,
   MAP1LC3B, SCL11A2), among which MAP1LC3B, playing a role in
   mitochondrial reactive oxygen species energy metabolism, showed the
   strongest predictive ability (AUC = 0.883, P < 0.05). It was
   significantly positively correlated with M0-type macrophages and
   significantly negatively correlated with activated CD4 + memory T
   cells. (3) In the adult sepsis population, models constructed using
   GBM, SVM, and LASSO algorithms showed the best performance. Three key
   DEGs associated with Autophagy were identified (TSPO, HTRA2, USP10),
   with TSPO, which mediates oxidative stress regulation, iron
   homeostasis, and cholesterol transport, showing the strongest
   predictive ability (AUC = 0.825, P < 0.05). It was significantly
   positively correlated with M1-type macrophages and significantly
   negatively correlated with CD8 + T cells. This study, through the
   integrated application of computational biology and machine learning
   algorithms, discovered biomarkers of PCD patterns that affect cytokine
   storm-mediated inflammation and immunosuppressive effects in sepsis
   populations across different age groups (neonates, children, and
   adults). These findings have specific clinical application and drug
   development value, providing a scientific basis for the global
   application of artificial intelligence-assisted sepsis diagnosis and
   treatment management.

Supplementary Information

   The online version contains supplementary material available at
   10.1038/s41598-025-14600-0.

   Keywords: Sepsis, Machine learning, Cell death, Immunosuppression,
   Biomarkers

   Subject terms: Infectious diseases, Neonatal sepsis, Biomarkers,
   Diagnostic markers, Immunology, Cell death and immune response

Introduction

   Sepsis is a syndrome of life-threatening organ dysfunction caused by
   dysregulated immune responses and extensive inflammatory damage
   triggered by infection, and it is one of the leading causes of death
   among critically ill patients^[42]1–[43]3. In the early stages of
   systemic infection, a rapid cascade of pro-inflammatory and
   anti-inflammatory cytokine production is triggered, a phenomenon known
   as the “cytokine storm”^[44]4. Thanks to the timely advancements in
   intensive care, most sepsis patients now survive the acute phase of
   sepsis, which mitigates the damage of the cytokine storm to tissues and
   organs^[45]5. However, following the cytokine storm, patients may
   experience transient lymphopenia and long-term immune dysfunction, a
   condition referred to as “immune paralysis.” This makes surviving
   patients more susceptible to secondary infections and viral
   reactivation, resulting in a significantly lower 5-year survival rate
   compared to non-infected individuals^[46]6–[47]8. According to a report
   from the World Health Organization, there are approximately 49 million
   cases of sepsis worldwide each year, with as many as 11 million related
   deaths^[48]9. A retrospective analysis indicated that the in-hospital
   mortality rate for sepsis is 17%, and for severe sepsis, it is
   26%^[49]10. These figures highlight the severity of sepsis as a risk
   factor for patient mortality. Despite this, current preventive and
   therapeutic options for sepsis remain limited, with treatment choices
   mainly confined to general intensive care unit care and a lack of
   specific treatments for sepsis. This is primarily due to our limited
   understanding of the immunophysiology associated with the
   disease^[50]7. Therefore, the search for effective therapeutic targets
   and strategies has become particularly urgent and crucial.

   In multicellular organisms, cell death pathways are diverse and
   flexible, and molecular regulation exhibits high plasticity. For a long
   time, it has been believed that various modes of cell death participate
   in the regulation of effector immune cells or the enrichment of
   regulatory immune cells through different pathways, and are involved in
   maintaining homeostasis and disease development. Studies have shown
   that sepsis is closely related to multiple cell death pathways. For
   example, review studies have pointed out that various forms of PCD are
   involved in regulating the immune response in sepsis patients,
   including pyroptosis, apoptosis, necroptosis, PANoptosis, NETosis,
   autophagy, ferroptosis, etc.^[51]11. However, the current limitation of
   research is that it focuses only on one type of cell death in a
   specific age group. There are significant differences in immune
   responses between neonates, children, and adults, and these differences
   affect the clinical manifestations and pathological processes of
   sepsis. The immune system of neonates is not yet fully mature, and
   their immune cells may not respond effectively to pathogens, leading to
   uncontrolled inflammatory responses^[52]12. The immune system of
   children gradually matures during growth, but their immune regulatory
   mechanisms may still not be as stable as those of adults^[53]13. In
   comparison, the immune system of adults is more mature and stable, but
   with increasing age, immune function may gradually decline, affecting
   the ability to defend against infections^[54]14. Therefore, the purpose
   of this study is to conduct an in-depth analysis of sepsis patients
   across different life stages through multiple centers worldwide, to
   identify the key PCD patterns among the 18 types of programmed cell
   death that play crucial roles in neonates, children, and adults, and to
   screen for key genes, thereby laying a theoretical foundation for the
   development of precision treatment plans.

   In recent years, complex deep learning models such as Graph Attention
   Mechanism-Based Graph Neural Network (GAMB-GNN)^[55]15, Multi-scale
   Unsupervised Self-embedding Graph Neural Network (Muse-GNN)^[56]16,
   Integrative Clustering Framework (iCluF)^[57]17 and Autoencoder^[58]18
   have demonstrated remarkable potential in the field of biomedical data
   analysis, particularly in capturing complex nonlinear relationships and
   graph-structured information. However, given the relatively limited
   sample size of the omics data available for this study, as well as the
   potential overfitting risks and computational resource demands
   associated with overly complex models, we have chosen to prioritize the
   evaluation of a suite of classical and computationally efficient
   machine learning methods (including LR, DT, GBM, KNN, LASSO, PCA, RF,
   SVM, and XGBoost). These methods have been proven to exhibit robust
   performance on datasets of similar scale^[59]19–[60]21, and their
   interpretability facilitates our preliminary exploration of the
   relationships between omics features and sepsis prediction. In the
   future, systematic exploration and comparison of these advanced models
   will be of great significance when larger-scale datasets and more
   abundant computational resources become available.

Methods

Data sources

   Using “sepsis” as the keyword, we downloaded microarray datasets from
   the GEO database for human subjects required for this study, obtaining
   a total of 20 datasets with 1507 cases and 484 controls. Neonatal
   group: [61]GSE25504 and [62]GSE69686. Children group: [63]GSE4607,
   [64]GSE9692, [65]GSE13904, [66]GSE26378, and [67]GSE26440. Adult group:
   [68]GSE10361, [69]GSE10474, [70]GSE13176, [71]GSE28750, [72]GSE40012,
   [73]GSE54514, [74]GSE57065, [75]GSE66890, [76]GSE67530, [77]GSE69528,
   [78]GSE95233, [79]GSE32707, and [80]GSE65682. Seventy percent were
   randomly selected as the training set and 30% as the testing set.
   Eighteen PCD-related regulatory genes were obtained through literature
   review^[81]22.

Differential expression gene processing

   We utilized the limma package in R software to perform differential
   expression gene analysis on sepsis samples and healthy control samples
   from the microarrays. DEGs were identified based on an absolute LogFC
   value > 1 and an adjusted P-value (FDR) < 0.05.

Immune-related analysis

Immune cell infiltration assessment

   We employed the IOBR package in the R language environment, integrating
   the advanced CIBERSORT algorithm, to conduct a comprehensive immune
   cell infiltration analysis on the collected samples. The CIBERSORT
   algorithm is capable of estimating the relative abundance of 22 major
   immune cell subsets in samples based on gene expression data and
   revealing differences in immune cell composition among various samples.

Association study between DEGs and immune factors

   This study employed Pearson correlation analysis to explore the
   potential connections between DEGs in the dataset and specific immune
   cell subsets. These immune cell subsets include, but are not limited
   to, plasma cells and T cells, which play crucial roles in immune
   responses and disease progression.

Machine learning algorithms

   To establish a consensus of PCD-related genes with high accuracy and
   stability, we applied nine machine learning algorithms (Logistic
   Regression LR, Decision Tree DT, Gradient Boosting Machine GBM,
   K-Nearest Neighbors KNN, LASSO, Principal Component Analysis PCA,
   Random Forest RF, Support Vector Machine SVM, and Distributed Gradient
   Boosting Library XGBoost) to the training and testing datasets with
   tenfold cross-validation to select three optimized algorithm models.
   The Evaluation Indicators of Machine Learning Model were including
   F1-score, Recall, Precision, AUPRC and AUROC. The SHAP algorithm was
   further used to quantify the contribution of each gene based on cell
   death features to the prediction of sepsis outcomes.

Functional enrichment analysis

   Functional annotation and KEGG pathway enrichment analysis of the three
   groups of DEGs were performed using R software^[82]23. The “ggplot2”
   and “clusterProfiler” packages were used to analyze and visualize the
   results.

Results

Identification of DEGs in Sepsis patients across different age groups

   The flowchart of this study is depicted in Fig. [83]1. We collected
   transcriptomic data and clinical information of sepsis patients across
   different age groups compared to normal tissues from the GEO database.
   For the neonatal group, the [84]GSE25504 and [85]GSE69686 datasets were
   analyzed. After processing, merging, and removing duplicate tissues, an
   expression matrix consisting of 12,442 genes was constructed. Principal
   component analysis (PCA) was performed to remove batch effects, and we
   successfully eliminated batch influence from the dataset (Fig. [86]2A),
   further identifying 5839 DEGs between sepsis and normal tissues
   (Fig. [87]2B). The children group was based on five datasets:
   [88]GSE4607, [89]GSE9692, [90]GSE13904, [91]GSE26378, and [92]GSE26440.
   After processing, an expression matrix of 22,876 genes was formed.
   Following PCA analysis to remove batch effects (Fig. [93]2C), 14,662
   DEGs were identified (Fig. [94]2D). The adult group analyzed a total of
   13 datasets: [95]GSE10361, [96]GSE10474, [97]GSE131761, [98]GSE28750,
   [99]GSE40012, [100]GSE54514, [101]GSE57065, [102]GSE66890,
   [103]GSE67530, [104]GSE69528, [105]GSE95233, [106]GSE32707, and
   [107]GSE65682. After processing, an expression matrix of 6,115 genes
   was obtained. After PCA to remove batch effects (Fig. [108]2E), we
   identified 5,149 DEGs (Fig. [109]2F). Additionally, 18 genes related to
   programmed cell death pathways were obtained through literature search
   (Table [110]S1)^[111]22.

Fig. 1.

   [112]Fig. 1
   [113]Open in a new tab

   A graphic abstract of this study.

Fig. 2.

   [114]Fig. 2
   [115]Open in a new tab

   Analysis of Age-Group-Specific DEGs in Sepsis. (A) Neonatal PCA without
   and with batch effect removal. (B) Differential gene expression profile
   of neonates. (C) Children PCA without and with batch effect removal.
   (D) Differential gene expression profile of children. (E) Adult PCA
   without and with batch effect removal. (F) Differential gene expression
   profile of adults.

ML algorithm screening and evaluation of key PCD patterns in different age
groups

   The DEGs from different age groups were intersected with genes related
   to the 18 programmed cell death pathways to obtain DEGs associated with
   different modes of death. Subsequently, nine machine learning
   algorithms were trained (LR, DT, GBM, KNN, LASSO, PCA, RF, SVM,
   XGBoost), and tenfold cross-validation was performed to select three
   optimized algorithm models. The results indicated that different ML
   algorithms showed competitiveness in risk assessment across different
   age groups. The Evaluation Indicators of Machine Learning Model across
   different age groups were described in Supplementary Table 2 (Table
   [116]S2). Specifically, the models constructed by GBM, XGBoost, and RF
   algorithms performed better in neonatal sepsis (Fig. [117]3A), while
   the models constructed by GBM, SVM, and LASSO algorithms performed
   better in children sepsis (Fig. [118]3B), and in adult sepsis, the
   models constructed by GBM, SVM, and LASSO algorithms also performed
   better (Fig. [119]3C). Furthermore, the SHAP algorithm was used to
   quantify the contribution of each gene based on cell death features to
   the prediction of sepsis outcomes in different age groups
   (Fig. [120]4A–C). The results showed that the key PCD patterns selected
   by machine learning for different age groups were Pyroptosis
   (neonates), Ferroptosis (children), and Autophagy (adults).

Fig. 3.

   [121]Fig. 3
   [122]Open in a new tab

   ML Algorithm Performance in Sepsis Risk Assessment (A) Neonatal sepsis
   model comparison (B) Children sepsis model comparison (C) Adult sepsis
   model comparison.

Fig. 4.

   [123]Fig. 4
   [124]Open in a new tab

   SHAP Analysis of PCD-Related Genes (A) Contribution in neonatal sepsis
   (B) Contribution in Children sepsis (C) Contribution in adult sepsis.

Enrichment analysis of key PCD patterns in different age groups

   In the enrichment analysis of key cellular programmed death patterns
   among different age groups, we observed the following patterns: In the
   neonatal population, GO enrichment analysis revealed that DEGs
   associated with pyroptosis were primarily involved in the regulation of
   enzymatic activities, including positive regulation of cysteine-type
   endopeptidase activity, positive regulation of endopeptidase activity,
   and positive regulation of peptidase activity (Fig. [125]5A). KEGG
   pathway analysis indicated that these pyroptosis-related DEGs were
   closely related to lipid metabolism and immune response, with
   involvement in pathways such as Lipid and atherosclerosis, NOD-like
   receptor signaling pathway, and Shigellosis (Fig. [126]5B). In the
   children population, GO enrichment analysis identified DEGs associated
   with ferroptosis as being primarily enriched in response to metal ion,
   response to oxidative stress, and long-chain fatty acid metabolic
   process (Fig. [127]5C). Further KEGG pathway analysis showed that these
   ferroptosis-related DEGs were enriched in lipid metabolism pathways,
   such as Fatty acid biosynthesis, Mineral absorption, and Fatty acid
   metabolism (Fig. [128]5D). In the adult population, DEGs related to
   Autophagy were primarily involved in the regulation of autophagic
   processes, including regulation of autophagy, macroautophagy, and
   regulation of macroautophagy (Fig. [129]5E). KEGG pathway analysis
   revealed that these Autophagy-related DEGs were enriched in pathways
   such as Autophagy-animal, Shigellosis, and Longevity regulating pathway
   (Fig. [130]5F).

Fig. 5.

   [131]Fig. 5
   [132]Open in a new tab

   Enrichment Analysis of PCD Patterns. Pyroptosis GO Enrichment (A) and
   KEGG Pathways (B) in newborn. Ferroptosis GO Enrichment (C) and KEGG
   Pathways (D) in Children. Autophagy GO Enrichment (E) and KEGG Pathways
   (F) in Adults.

Immune cell correlation analysis in different age groups

   In the comparative analysis of immune cells across different age
   groups, we identified the following characteristics: In the neonatal
   population, significant differences in immune cell levels were observed
   between sepsis patients and normal controls for B cell naïve, T cells
   CD4 memory resting, T cells CD4 memory activated, Tregs, macrophages
   M0, and neutrophils (P < 0.05) (Fig. [133]6A). Correlation analysis
   revealed significant correlations between T cells CD4 memory resting
   and Tregs, T cells CD8 and NK cells activated, T cells CD8 and
   neutrophils, and NK cells activated and neutrophils (Fig. [134]6B). In
   the children population, significant differences in immune cell levels
   were noted between sepsis patients and normal controls for mast cells
   resting and neutrophils (P < 0.05) (Fig. [135]6C). Further correlation
   analysis indicated a positive correlation between T cells CD8 and T
   cells CD4 memory activated (Fig. [136]6D). In the adult population,
   significant differences in immune cell levels were observed between
   sepsis patients and normal controls for B cell naïve, T cells CD8, T
   cells CD4 naïve, T cells CD4 memory resting, Tregs, NK cells resting,
   monocytes, macrophages M0, dendritic cells resting, mast cells resting,
   and neutrophils (P < 0.05) (Fig. [137]6E). However, correlation
   analysis did not identify significant correlations among immune cells
   (Fig. [138]6F).

Fig. 6.

   [139]Fig. 6
   [140]Open in a new tab

   Immune Cell Analysis Across Age Groups. Neonatal Immune Cell
   Differences (A) and Immune Cell Correlations (B). Children Immune Cell
   Differences (C) and Immune Cell Correlation (D). Adult Immune Cell
   Differences (E) and Immune Cell Correlations (F).

Identification of key genes in PCD patterns and their immune correlation
analysis in different age groups

   In the neonatal population, the intersection of sepsis DEGs and
   pyroptosis-related genes yielded 5 genes (CHMP7, NLRC4, AIM2, GZMB,
   PRKACA), among which CHMP7 and GZMB were highly expressed in sepsis
   patients, while NLRC4, PRKACA, and AIM2 were lowly expressed
   (Fig. [141]7A). In the children population, the intersection of sepsis
   DEGs and Ferroptosis-related genes resulted in 6 genes (AKR1C3, GCLM,
   PEBP1, CARS, MAP1LC3B, SCL11A2), with AKR1C3 and PEBP1 being highly
   expressed in sepsis patients, and GCLM, CARS, MAP1LC3B, and SCL11A2
   being lowly expressed (Fig. [142]7B). In the adult population, the
   intersection of sepsis DEGs and Autophagy-related genes identified 3
   genes (TSPO, HTRA2, USP10), with TSPO and HTRA2 being highly expressed
   in sepsis patients, and USP10 being lowly expressed (Fig. [143]7C).

Fig. 7.

   [144]Fig. 7
   [145]Open in a new tab

   Key PCD Gene Expression in Sepsis (A) Neonatal Pyroptosis Genes (B)
   Children Ferroptosis Genes (C) Adult Autophagy Genes.

   The ROC curve indicated that NLRC4 had the strongest predictive power
   in the neonatal group (AUC = 0.902, P < 0.05) (Fig. [146]8A–E), which
   was significantly positively correlated with neutrophils and
   significantly negatively correlated with CD8 + T cells
   (Fig. [147]9A,B). In the children group, MAP1LC3B had the strongest
   predictive power (AUC = 0.883, P < 0.05) (Fig. [148]8F–K), which was
   significantly positively correlated with M0-type macrophages and
   significantly negatively correlated with activated CD4 + memory T cells
   (Fig. [149]9C,D). In the adult group, TSPO had the strongest predictive
   power (AUC = 0.825, P < 0.05) (Fig. [150]8L–N), which was significantly
   positively correlated with M1-type macrophages and significantly
   negatively correlated with CD8 + T cells (Fig. [151]9E,F).

Fig. 8.

   [152]Fig. 8
   [153]Open in a new tab

   ROC Analysis of Key Genes (A–E) ROC analysis of neonatal sepsis key
   genes (F–K) ROC analysis of children sepsis key genes (L–N) ROC
   analysis of adult sepsis key genes.

Fig. 9.

   [154]Fig. 9
   [155]Open in a new tab

   Immune Cell Correlations with Key Genes (A,B) Neonatal NLRC4
   correlations (C,D) Children MAP1LC3B correlations (E,F) Adult TSPO
   correlations.

Discussion

   Recent years have witnessed significant breakthroughs in the
   application of machine learning to sepsis research, particularly in
   early prediction model optimization and clinical decision support
   system development. Regarding predictive models, a systematic analysis
   integrating 28 studies encompassing 130 models demonstrated that
   machine learning models can accurately predict the onset of sepsis in
   advance^[156]24. Another study analyzing data from 2312 patients
   diagnosed with sepsis successfully established a machine learning-based
   predictive model, confirming a strong association between stress
   hyperglycemia ratio and 28-day all-cause mortality in critically ill
   septic patients^[157]25. In clinical translation, existing research
   teams have utilized machine learning methods to optimize treatment
   protocols for sepsis and septic shock patients, effectively reducing
   the 28-day mortality rate^[158]26. In contrast, this study innovatively
   integrates multiple machine learning algorithms across global
   multicenter cohorts to conduct in-depth analysis of sepsis patients at
   different life stages. We identified crucial PCD patterns from 18 PCD
   modalities that play pivotal roles in neonatal, pediatric, and adult
   populations, screening key genes to establish a theoretical foundation
   for precision therapy. The study successfully identified distinct
   dominant PCD patterns across different age groups: pyroptosis (key gene
   NLRC4) predominates in neonates, ferroptosis (key gene MAP1LC3B) in
   children, and autophagy (key gene TSPO) in adults. These findings not
   only reveal the complex relationship between sepsis and PCD but also
   emphasize the necessity for personalized treatment strategies targeting
   specific PCD patterns across different life stage populations.

   Firstly, this study found that pyroptosis may serve as the dominant
   mode of PCD in the course of sepsis in the neonatal population.
   Pyroptosis is a form of programmed cell death triggered by inflammasome
   activation, characterized by the activation of caspase-1, which in turn
   leads to cell membrane rupture and the release of inflammatory
   factors^[159]27. In neonatal sepsis, pyroptosis may drive the
   bidirectional dynamic evolution of immune dysfunction through a
   mechanism mediated by endogenous host defense peptides: in the early
   stage, due to the immaturity of the neonatal immune system, endogenous
   host defense peptides can trigger excessive inflammatory responses
   (such as cytokine storms) by inducing pyroptosis in
   macrophages^[160]28; as the disease progresses, the important
   regulatory component of the defense peptide system, LL-37, tends to
   become depleted^[161]29. It is worth noting that multiple studies have
   confirmed that LL-37 has an inhibitory effect on cell
   pyroptosis^[162]30,[163]31, and this regulatory mechanism may help
   maintain immune homeostasis during the compensatory phase. Based on the
   above evidence, we speculate that as LL-37 becomes progressively
   depleted, its inhibitory effect on cell pyroptosis gradually weakens,
   leading to abnormal activation of the pyroptosis pathway. This
   pathological shift may exacerbate immune paralysis through a dual
   mechanism: on one hand, persistent cell pyroptosis may accelerate the
   exhaustion of immune cells (especially macrophages with
   antigen-presenting functions); on the other hand, damage-associated
   molecular patterns released by pyroptotic cells may continuously
   activate immune cells through pattern recognition receptors, inducing
   them into an exhausted state, ultimately leading to a vicious cycle of
   decreased immune responsiveness and immune tolerance.

   NLRC4, as an intracellular pattern recognition receptor, can sense
   bacterial components such as flagellin and type III secretion systems,
   triggering inflammasome formation and activation. This process
   subsequently activates Caspase-1 and induces pyroptosis^[164]32. In
   sepsis, this mechanism is particularly critical, as NLRC4 activation
   may lead to excessive immune cell activation and massive release of
   inflammatory cytokines, thereby exacerbating the
   disease^[165]33–[166]37. However, previous studies have not
   sufficiently explored this phenomenon in neonatal populations.This
   study found that in neonatal sepsis, NLRC4 is highly expressed, with
   its expression levels significantly positively correlated with the
   number of neutrophils and significantly negatively correlated with the
   number of CD8 + T cells. Neutrophils, as key members of the innate
   immune system, play an important role in sepsis, participating in
   inflammatory responses and pathogen clearance. The study found that
   compared with healthy controls, neonates with sepsis have a higher
   neutrophil ratio^[167]38, a finding that is highly consistent with the
   conclusions drawn in this study. In addition, CD8 + T cells play a key
   role in protecting the host from intracellular infections. Research has
   pointed out that compared with surviving neonates with sepsis, deceased
   neonates with sepsis have lower levels of CD8 + T lymphocytes^[168]39.
   This study found that children with sepsis have lower CD8 + T
   lymphocytes than healthy children. This result may imply that in the
   pathological process of neonatal sepsis, the function of CD8 + T cells
   may be suppressed, and the reduction in their numbers may be closely
   related to the severity of the disease and may even have a potential
   link to the poor prognosis of the children. Studies indicate that NLRC4
   activation promotes neutrophil recruitment and activation^[169]40, with
   neutrophil pyroptosis serving as the primary source of IL-1β secretion.
   This process recruits additional neutrophils to infection sites through
   a positive feedback amplification loop, exacerbating inflammatory
   responses^[170]41. Therefore, we hypothesize that during the early
   stages of sepsis, NLRC4 activation induces neutrophil pyroptosis,
   further amplifying inflammation. Concurrently, as the disease
   progresses, excessive NLRC4 inflammasome activation-driven pyroptosis
   may lead to an overactive immune response^[171]42,[172]43, triggering
   immunosuppressive mechanisms to prevent excessive damage. These
   mechanisms include the activation of regulatory T cells (Tregs) or the
   secretion of immunosuppressive cytokines (e.g., CCL2, CCL3, CXCL2),
   which suppress CD8 + T cell function^[173]44. Consequently, we propose
   that in the late stages of sepsis, NLRC4 impairs effective
   anti-infective CD8 + T cell responses, resulting in immune paralysis.
   Given the immature neonatal immune system and their weaker capacity to
   regulate inflammatory responses, neonates may exhibit heightened
   vulnerability to sepsis, rendering them more susceptible to severe
   outcomes.

   Secondly, in pediatric sepsis, this study reveals that ferroptosis may
   serve as a significant mechanism as the dominant mode of PCD.
   Ferroptosis may drive sepsis-related immune imbalance through a dual
   mechanism: on one hand, the oxidized phospholipids (such as POVPC)
   released by ferroptotic cells activate the TLR4 signaling pathway in
   endothelial cells^[174]45, triggering a storm of pro-inflammatory
   cytokines such as IL-6 and TNF-α^[175]45. This finding provides a new
   perspective for explaining the occurrence of cytokine storms in the
   early stages of sepsis. On the other hand, the reduction of
   mitochondrial cristae and the decrease in membrane potential during
   ferroptosis lead to metabolic disorders in immune cells (such as
   dendritic cells^[176]46), which in turn cause impaired
   antigen-presenting ability and chemotactic dysfunction.In addition, as
   the disease progresses, the level of free iron in the blood
   significantly increases^[177]47. Iron overload directly damages the
   lysosomal function of macrophages^[178]48 and interferes with T-cell
   receptor signaling transduction^[179]49, ultimately leading to a
   decline in antimicrobial efficacy. This dynamic evolution process may
   explain the pathological characteristics of the transition from early
   excessive inflammatory response to late immune suppression/immune
   paralysis in sepsis.

   MAP1LC3B (microtubule-associated protein 1 light chain 3 beta), a
   critical protein in autophagy, plays a crucial role in autophagosome
   formation^[180]50. Studies have shown that autophagy can promote
   ferroptosis by selectively degrading anti-ferroptosis regulatory
   factors^[181]51–[182]53. Notably, MAP1LC3B has also been implicated in
   ferroptosis regulation, serving as a biomarker for ferroptosis in
   gastric cancer^[183]54, ischemic stroke^[184]55, and
   osteoarthritis^[185]56. These findings suggest that MAP1LC3B may play a
   significant role in ferroptosis regulation, although its precise
   mechanisms require further investigation.In the immune regulation of
   sepsis, MAP1LC3B also plays an important role. The study by Di et al.
   identified it as a key gene in the immune regulation of sepsis^[186]57,
   implying its involvement in the regulation of sepsis. Our study further
   revealed that in pediatric sepsis patients, MAP1LC3B is significantly
   overexpressed. It shows a significant positive correlation with
   unactivated M0 macrophages and a significant negative correlation with
   activated CD4 + memory T cells. M0 macrophages, as precursor cells, can
   polarize into the pro-inflammatory M1 or anti-inflammatory M2
   phenotypes, but their specific functions in pediatric sepsis have not
   yet been clarified. The results of this study provide a new direction
   for research in this area. On the other hand, activated CD4 + T cells
   are responsible for rapidly recognizing and eliminating pathogens that
   reinvade in adaptive immunity. However, it has been reported in
   previous studies that these cells are significantly reduced in the
   peripheral blood of pediatric sepsis patients and are closely related
   to immune dysfunction^[187]54, which is consistent with the negative
   correlation results of this study. Although no studies have directly
   investigated whether MAP1LC3B participates in the polarization process
   of M0 macrophages, influences their pathogen phagocytosis and
   degradation capabilities, or regulates activated CD4 + T cells,
   existing research has shown significant correlations between MAP1LC3B
   and other immune cells^[188]54,[189]58,[190]59. This suggests that
   MAP1LC3B may possess potential regulatory roles in immune cell
   function.In the adult population, our research findings reveal
   autophagy as a key mode of PCD. Autophagy may exhibit dynamic dual
   roles in different stages of sepsis by regulating immune cell
   functions. On one hand, autophagy can activate the NLRP3 inflammasome
   pathway in macrophages^[191]60, and NLRP3 can significantly exacerbate
   the overproduction of pro-inflammatory cytokines (such as IL-1β, IL-18,
   etc.)^[192]61, which in turn may drive the pathological process of
   cytokine storm. This may explain why autophagy promotes the sepsis
   cytokine storm in the early stages of sepsis. On the other hand, as the
   disease progresses, persistent autophagy activation can lead to the
   degradation of pathogen-associated molecular patterns and
   damage-associated molecular patterns required for inflammasome
   activation, thereby inhibiting inflammasome activation^[193]62. In
   addition, autophagy also promotes the internalization of MHC class I
   molecules, thereby suppressing the antiviral response of CD8 + T
   cells^[194]63,[195]64, ultimately leading to immune cell exhaustion and
   an immunoparalysis state. TSPO (Translocator protein), a transporter
   protein localized to the mitochondrial outer membrane, plays a critical
   role in regulating cellular stress responses and maintaining
   mitochondrial function. Current studies suggest that TSPO may
   indirectly influence the autophagy process by modulating mitochondrial
   dynamics and oxidative stress levels^[196]65,[197]66. Although no
   research has directly elucidated the mechanism of TSPO in
   sepsis-associated autophagy, existing evidence indicates its
   involvement in regulating sepsis-related pathological damage. For
   instance, research by Koumine et al. found that modulating TSPO
   expression can influence M1/M2 polarization phenotypes of macrophages,
   significantly alleviating liver injury in septic mice^[198]67,
   suggesting TSPO may play an important role in immune regulation during
   sepsis.

   This study found that the expression level of TSPO in the peripheral
   blood of adult sepsis patients is significantly elevated, and it is
   significantly positively correlated with the activation of M1
   macrophages and significantly negatively correlated with the activation
   of CD8 + T cells. M1 macrophages, as pro-inflammatory phenotype
   macrophages, are closely related to the progression of sepsis when
   over-activated. Studies have shown that the over-activation of M1
   macrophages and their pro-inflammatory cytokine secretion in adult
   sepsis patients are key factors in triggering systemic inflammatory
   response syndrome and multiple organ dysfunction^[199]68. It is worth
   noting that during the course of sepsis, platelets mediate the
   suppression of CD8 + T cell function through MHC class I
   molecules^[200]69, which echoes the negative correlation between TSPO
   and CD8 + T cell activation found in this study. Existing evidence
   suggests that the upregulation of TSPO expression may exacerbate the
   release of inflammatory cytokines and their mediated organ damage by
   promoting the activation of M1 macrophages^[201]67. On the other hand,
   CD8 + T cells play an important role in the immune response against
   intracellular pathogens. Although there are currently no reports on
   TSPO directly regulating CD8 + T cells in adult sepsis, multiple
   studies have shown that TSPO can induce T cell immune paralysis and
   lead to functional impairment^[202]70–[203]72. This suggests that TSPO
   may participate in the pathological process of immune imbalance in
   sepsis by regulating the function of CD8 + T cells, and this regulatory
   mechanism needs to be further explored in subsequent studies.

   In summary, our study indicates that in neonatal sepsis, pyroptosis is
   the key PCDmode, with NLRC4 being the critical DEG, which is
   significantly positively correlated with neutrophils and significantly
   negatively correlated with CD8 + T cells. In the pediatric sepsis
   population, ferroptosis is the key PCD mode, with MAP1LC3B as the
   critical DEG, which is significantly positively correlated with M0
   macrophages and significantly negatively correlated with activated
   CD4 + memory T cells. In the adult sepsis population, autophagy is the
   key PCD mode, with TSPO as the critical DEG, which is significantly
   positively correlated with M1 macrophages and significantly negatively
   correlated with CD8 + T cells.

Conclusion

   Our study, through the integrated application of computational biology
   and machine learning algorithms, has identified biomarkers of PCD
   patterns that influence cytokine storm-mediated inflammation and
   immunosuppressive effects in sepsis populations across different age
   groups (neonates, children, and adults). These findings hold specific
   value for clinical applications and drug development, providing a
   scientific basis for the global implementation of artificial
   intelligence-assisted diagnosis and treatment management of sepsis.

Supplementary Information

   Below is the link to the electronic supplementary material.
   [204]Supplementary Material 1^ (49.3KB, xlsx)
   [205]Supplementary Material 2^ (9.7KB, xlsx)

Author contributions

   J.Y. and F.Y. Ou conceived and supervised the entire project. B.B. Li
   and L.X. Zeng wrote the code and drafted the manuscript Q.L. Chen, H.Y.
   Gan and J.N. Yu processed the data and analyzed the results. Q.G, J.H.
   Feng and J.F Zhang. improved the algorithms and reviewed the
   manuscript. The authors wish to acknowledge that Jie Yang, Fanyan Ou
   and Binbin Li contributed equally to this work and are considered
   co-first authors. Jihua Feng and Jianfeng Zhang are co-corresponding
   authors. All authors discussed, revised, and proofread the manuscript.

Funding

   Guangxi Natural Science Foundation (No.2021GXNSFBA196017) and National
   Natural Science Foundation of China (No.82302461 and No.82360374).

Data availability

   The datasets presented in this study can be found in online
   repositories. The names of therepository/repositories and accession
   number(s) can be found in the article.

Declarations

Competing interests

   The authors declare no competing interests.

Footnotes

   Publisher’s note

   Springer Nature remains neutral with regard to jurisdictional claims in
   published maps and institutional affiliations.

   Jie Yang, Fanyan Ou, and Binbin Li contributed equally to this work.

Contributor Information

   Jihua Feng, Email: fengjihua@gxmu.edu.cn.

   Jianfeng Zhang, Email: zhangjianfeng@gxmu.edu.cn.

References