Sepsis-associated gene expression datasets were obtained from the Gene Expression Omnibus (GEO) database (ncbi.nlm.nih.gov/geo/) searching the keywords ‘sepsis’ and ‘Homo sapiens’. The transcriptomic datasets that were selected were required to conform to the following criteria (11): i) Samples had been collected from whole blood; ii) samples contained healthy control and septic patients (in a state of either sepsis or septic shock); and iii) the sample size was >50. Ultimately, four datasets (dataset nos. GSE185263, GSE57065, GSE131761 and GSE95233) were downloaded from the GEO database. Of these datasets, GSE185263 was the only one generated by a high-throughput sequencing platform, the other datasets were generated by a microarray platform.

The batch effect from experimental restrictions in data batches often impedes subsequent analysis with statistical heterogeneity (12), and the dataset bias from the significant difference in the principle and measurement of microarray data (the continuous measurement on the signal sensitivity of probes) (13) and high-throughput data (integer counts on the copy number of transcripts) (14) causes the incomparability of transcriptomic profiling (15). The design of two previous septic studies were consulted (11,16) to avoid the batch effect and dataset bias simultaneously. The authors had selected one high-throughput dataset and multiple microarray datasets and separately used each dataset by using the high-throughput dataset as a training cohort and microarray datasets as independent external validations. Following the aforementioned precedents, GSE185263 (high-throughput dataset) was assigned as the training cohort and the three other microarray datasets (GSE57065, GSE131761, and GSE95233) were assigned as the independent external validation cohorts.

GSE185263, the training cohort, generated by the GPL16791 high-throughput sequencing platform, consists of 348 septic and 44 control samples (17). The GRCh38 Human Genome Assembly (ncbi.nlm.nih.gov/datasets/genome/GCF_000001405.40/) was downloaded, the raw counts of the RNA-sequencing dataset were annotated and the data were normalized as transcripts per kilobase million (18).

The three other microarray datasets, GSE57065, GSE131761 and GSE95233, were annotated by matching series matrices and corresponding annotation files, and the data were uniformly normalized using robust multiarray averaging, with log2 transformation. GSE57065, the test cohort, generated by the GPL570 microarray platform, included 82 septic and 25 control samples (19), whereas GSE131761 (validation cohort A), generated by the GPL13497 microarray platform, contained 81 septic and 15 control samples (20) and GSE95233 (validation cohort B), generated by the GPL570 microarray platform, comprised 102 septic and 22 control samples (21).

In the training cohort, DEGs were screened using the R package ‘edgeR’ version 4.6.2 (bioconductor.org/packages/release/bioc/html/edgeR.html) with a threshold of |log fold change|≥1 and a false discovery rate <0.05 (22). DEGs were then visualized using the R packages ‘ggplot2’ and ‘pheatmap’.

Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses of DEGs were performed using the R package ‘clusterProfiler’ (23); gene set variation analysis (GSVA) was performed to reveal the status of signaling pathways between groups using the R packages ‘GSVA’ (24) and ‘DOSE’ (25). All results were filtered through a threshold of P≤0.05, and data were visualized using the R package ‘enrichplot’.

In the present study, FRDEGs were defined as the overlapping genes of DEGs and ferroptosis-related genes (FRGs). The latter were separately obtained from the FerrDb database (26) (n=484) and the GeneCards database (27) (n=92), genes with a relevance score >3 were retained. After combining the two lists of FRGs, duplicated genes were removed and a set of unique FRGs (n=506) was generated. The intersection of DEGs and FRGs was visualized using the R package ‘ggvenn’ and the intersecting genes were identified as FRDEGs.

PPINs have been applied to identify key regulators in biological processes (28). In the present study, a PPIN was used to identify key regulators in sepsis-associated ferroptosis, and these key regulators were defined as hub-FRDEGs. To build the PPIN, FRDEGs were imported into the Search Tool for the Retrieval of Interacting Genes/Proteins (STRING) online database (29), the PPINs were then imported into Cytoscape (30) for visualization.

To select hub-FRDEGs from the raw PPIN, ‘cytoHubba’, a Cytoscape plugin for network ranking, was deployed in order to rank the topological importance of each PPIN node using different algorithms (31). In the present study, five CytoHubba algorithms, Maximum Neighborhood Component, Density of Maximum Neighborhood Component, Maximal Clique Centrality, Edge Percolated Component and Degree, were used to rank the PPIN nodes, and the top 50 nodes from each algorithm were chosen, intersected and pinpointed as hub FRDEGs. In addition, chromosomal locations of hub FRDEGs were recognized and portrayed using the R package ‘RCicos’.

Machine learning algorithms enable the number of candidate hub-FRDEGs to be reduced using the modelling fitting of the transcriptomic profile. Conforming to specific key parameters, the hub-FRDEGs selected by the machine learning algorithms were deemed to be the hub-FRDEGs that were most relevant to sepsis. In the present study, three machine learning algorithms, namely the least absolute shrinkage and selection operator (LASSO), support vector machine-recursive feature elimination (SVM-RFE) and random forest, were used to select the hub-FRDEGs that were most relevant to sepsis.

For the LASSO algorithm, the key parameter was the λ value determining the binomial deviance of model fitting, and the λ value responding to the minimum binomial deviance was deemed as optimal (32). Under the optimal λ, all hub-FRDEGs retained in the LASSO fitting model were selected using the R package ‘glmnet’ (33). Regarding the random forest algorithm, the key parameter was the Gini index, and all hub-FRDEGs were rated and selected below the specific Gini index using the R package ‘randomForest’ (34). Considering the SVM-RFE algorithm, realized using the R package ‘e1071’, all the hub-FRDEGs that minimized the 10-fold cross-validation error rate were recommended (35). Overlapping hub-FRDEGs, as selected by the algorithms, were deemed to be the hub-FRDEGs for further study, and their closeness with other genes was analyzed and visualized using the GeneMANIA database (36).

In the training cohort, logistic regression analysis was performed to build a predictive model of sepsis using the R package ‘rms’. A diagnostic nomogram was subsequently constructed according to the logistic regression formula, and its diagnostic accuracy was evaluated using the calibration curve, the confusion matrix, the decision curve and the receiver operating characteristic (ROC) curve. The differences in expression were then identified by comparing the hub-FRDEGs between the control and sepsis samples, and these were subsequently visualized using heatmaps and box-plots, whereas their diagnostic accuracy was assessed using the ROC curve. The calibration curve with Homer-Lemeshow test, decision curve, heatmap, ROC curve and box plots were constructed using the R packages ‘caret’, ‘rmda’, ‘pheatmap’, ‘pROC’ and ‘ggplot2’, respectively. Finally, the confusion matrix was calculated using the R package ‘caret’, and subsequently visualized for presentation.

To validate the diagnostic accuracy of the hub-FRDEGs and nomogram, the changes in expression of hub-FRDEGs under septic conditions were quantified, the area under the curve (AUC) values in the ROC curves of the hub-FRDEGs and nomogram were calculated, and all results in the test cohort (GSE57065), validation cohort A (GSE131761) and validation cohort B (GSE95233) were visualized. The changes in hub-FRDEGs resulting from sepsis were visualized using box plots and heatmaps, whereas the diagnostic accuracies of hub-FRDEGs and nomograms were visualized using a confusion matrix and calibration curve, respectively.

Consensus clustering analysis revealed the distinct ferroptotic patterns of sepsis samples using the R package ‘ConsesusClusterPlus’. According to the change in area under the cumulative distribution function curve, the optimal number of clusters (with k ranging from 2–10) was determined, and septic samples were classified based on the expression level of the ferroptotic gene signature. Principal component analysis (PCA) of septic clusters was subsequently performed to reveal the disparities in the septic data using the R package ‘factoextra’. The transcriptomic differences of the hub-FRDEGs between clusters were subsequently evaluated and visualized using box plots and heatmaps. The immunocyte abundance between septic clusters was measured after analyzing the immune microenvironment in all the samples. DEGs between septic subgroups were screened using the R package ‘edgeR’, and enrichment analyses (GO and KEGG) were then performed to reveal their metabolic differences.

In the training cohort, 22 types of immune cellular components and degrees were measured using the R package ‘CIBERSORT’ (37). Septic changes of the immunocytes were calculated and visualized using box plots, and the correlation between immunocyte abundance and the hub FRDEG transcriptomic profile was quantified and plotted. The correlation between infiltrating immunocytes was also measured using Pearson's correlation coefficient and visualized using the R package ‘corrplot’. Furthermore, the infiltration score of 28 immunocytes (38) was computed using single-sample gene set enrichment analysis (ssGSEA). The ssGSEA algorithm was used to calculate the proportion of immunocytes using the R-package ‘GSVA’ (24) and the metagene list of 28 immunocytes (38). Results were expressed in the form of an infiltration score between 0 and 1. The ssGSEA algorithm and CIBERSORT were often revealed to be mutually complementary to each other with respect to evaluating the immune infiltration (39–41). Using the ssGSEA algorithm, the immune cell proportions comparing between samples that exhibited a differential expression of hub-FRDEGs were compared and visualized.

All reagents and resources used in the present study are presented in Table SI. The human monocyte cell line THP-1 was obtained from Wuhan Pricella Biotechnology Co., Ltd ., and the cells were cultured in RPMI-1640 medium containing 10% fetal bovine serum and 0.05 mM β-mercaptoethanol in an incubator at 37°C in an atmosphere containing 5% CO₂.

THP-1 cells were treated with 100 nM phorbol myristate acetate (PMA) for 24 h to induce their differentiation into macrophage-like cells at 37°C. The treated cells were further cultured in medium without PMA for 24 h to enable them to stabilize and mature into M0 macrophages (42) at 37°C. To simulate the sepsis microenvironment, lipopolysaccharide (LPS; 1 µg/ml) was added following the induction of THP 1 differentiation, and the incubation was allowed to continue for 24 h at 37°C (43).

Total RNA was isolated from 2.5×10⁶ adherent cells using TRIzol^® reagent (cat. no. 15596026; Invitrogen; Thermo Fisher Scientific, Inc.). and reverse-transcribed into cDNA using PrimeScript™ RT reagent Kit with gDNA Eraser (Takara, catalog number RR047A) at 37°C for 15 min), step 2 (85°C at 5 sec), and step 3 (4°C). SYBR Green qPCR Master Mix solution (MedChemExpress, catalog number HY-K0501), was used for PCR following the manufacturer's instructions.

Primers were obtained from the Beijing Tsingke Biotech Co., Ltd.. RT-qPCR analysis was performed using a CFX96 Touch Real-Time PCR Detection System (C1000Touch™; Bio-Rad Laboratories, Inc.). Thermocycling conditions were as follows: Initial denaturation at 95°C for 30 sec, followed by 40 cycles of 5 sec at 95°C and 30–60 sec at 60°C. ACTB (β-actin) was amplified as the internal reference gene. The 2^−ΔΔCq method was applied for quantification of the data (44).

Cells were lysed in RIPA lysis buffer containing PMSF, protease and phosphatase inhibitor cocktail for 30 min at 4°C, and then centrifuged at 12,000 × g for 15 min in 4°C. The protein supernatant was concentrated, and the protein concentration was measured using a BCA protein assay kit. Subsequently, 10% SDS-PAGE electrophoresis was used to separate 10 µl protein samples at the concentration of 2 µg/µl. Subsequently, the gels containing the separated proteins were cut horizontally according to the corresponding molecular weights of the protein, and the gel slices were transferred to polyvinylidene fluoride membranes, followed by blocking with 5% skimmed milk for 1 h at room temperature. The membranes were incubated overnight at 4°C with primary antibodies (1:1,000 dilution; diluted according to the instructions of the manufacturer). The membranes were then incubated with horseradish peroxidase-conjugated secondary anti-rabbit or anti-mouse IgG (1:10,000 dilution) for 1 h at room temperature. Finally, the levels of the proteins were detected using a chemiluminescence reagent kit, and the protein expression density was quantified using Quantity One Software (Bio-Rad Laboratories, Inc.).

Thioredoxin (TXN)-targeting siRNA (si-TXN; 5′-UUUGACUUCACACUCUGAAGC-3′) or non-targeting control siRNA (si-NC; 5′-CCUACGCCACCAAUUUCGU-3′) were synthesized by the Beijing Tsingke Biotech Co., Ltd. Induced 1.5×10⁶ M₀ macrophage cells in the exponential growth phase were seeded into 6-well plates. These cells were subsequently transiently transfected with either 10 µl si-TXN or si-NC (100 nmol/l) using Advanced Transfection Reagent (Zeta Life, cat. no. AD600150). The duration of transfection was 24 h at 37°C, and the duration between the end of transfection and subsequent experiments was 24 h.

MDA content was measured following the manufacturer's protocol. In brief, fresh THP-1-derived macrophages were collected and lysed on ice for 20 min to prepare samples. After the MDA solution was added to each sample or standard, the mixture was heated at 100°C for 15 min and subsequently centrifuged at 1,000 × g for 10 min to harvest the supernatant at 4°C. The absorbance of the supernatant was then measured at optical density (OD)= 532 nm using a colorimetric method, and MDA levels were expressed as nmol/mg.

GSH/GSSG detection was performed according to the manufacturer's instructions. Briefly, fresh THP-1-derived macrophages were collected. Each sample was divided into two aliquots: One for measuring the GSSG content following treatment with the GSH clearing reagent, and the other for measuring the total GSH content without treatment. The OD values were measured at 412 nm using a Multiskan™ FC Microplate Photometer (cat. no. 51119100; Thermo Fisher Scientific, Inc.), and the contents of total GSH, GSSG and reduced GSH were calculated.

Fe²⁺ levels were evaluated using the FerroOrange staining method. After removing the previous culture medium, the cells were washed twice in a serum-free RPMI-1640 medium. A working solution of l µmol/l FerroOrange fluorescent probe was prepared in a serum-free RPMI-1640 medium before being added to the cells. The cells were incubated at 37°C in a dark environment for 30 min. Subsequently, the nuclei were labelled with Hoechst 33342. The cells were then observed and images were captured using a Leica DMi8 fluorescence microscope. Finally, the fluorescence intensity due to FerroOrange staining was quantified using ImageJ software version 1.54m (National Institutes of Health).

The bioinformatical data obtained from the various analyses underwent statistical evaluation using R software (version 4.2.2). The Shapiro-Wilk test was deployed to assess the normality of the data, and the unpaired Student's t-test or Wilcoxon rank sum test was used to compare the differences in transcriptomics and immune infiltration. Comparisons between multiple groups were made by one-way analysis of variance, followed by the Bonferroni post hoc test. Data analysis and the construction of graphs were performed using GraphPad Prism 10 (Dotmatics). P<0.05 was considered to indicate a statistically significant difference.

The overall design of the present study is presented in Fig. 1. Datasets acquired are summarized in Table I. A total of 4,410 DEGs between the sepsis and control groups were obtained from the training cohort (Table SII). The Volcano plot (Fig. 2A) and heatmap (Fig. 2B) reveal the transcriptomic distribution of DEGs. FRGs (n=506; Table SIII) were obtained and aggregated from the FerrDb and GeneCards database (Fig. 2C), and overlapping genes (n=94; Fig. 2D) between the DEGs (n=4,410) and FRGs were identified as FRDEGs for the subsequent analytic steps.

FRDEGs were loaded into the STRING database in order to construct the PPIN, and hub-FRDEGs (n=84) were selected as the nodes (Fig. 2E). Hub-FRDEGs were ranked according to five internal algorithms of CytoHubba in the Cytoscape desktop utility, and the top 50 hub-FRDEGs from diverse rankings were then intersected (Table SIV and Fig. 2F). Shared hub-FRDEGs (n=38) were identified as candidates for determining machine learning algorithms, and their chromosomal positions are shown in Fig. 2G. LASSO was employed as the machine learning algorithm, and 18 hub-FRDEGs were selected for the minimum binomial deviance (Table SV; Fig. 3A and B); upon employing random forest, three hub-FRDEGs below the cut-off Gini of 6 were selected (Table SVI; Fig. 3C-D); upon using SVM-RFE, the lowest 10-fold error rate was attained when considering 38 of the hub-FRDEGs (Table SVII and Fig. 3E). Intersection among the three machine learning algorithms determined the hub-FRDEGs (n=3) that were most relevant to sepsis, namely ataxia telangiectasia mutated (ATM), dipeptidyl peptidase 4 (DPP4) and TXN (Fig. 3F), with the predicted interaction network shown in Fig. 3G.

In the training cohort, a binary logistic model was purposed for classifying the status of patients into either of the control or sepsis groups according to the ATM, DPP4 and TXN expression levels. The logistic formula applied was as follows: Logit (P=Sepsis) =0.2244-(0.0704 × ATM expression levels)+(0.0755 × TXN expression levels)-(0.0840 × DPP4 expression levels). The diagnostic nomogram was subsequently plotted (Fig. 4A). The decision curve revealed the acceptable clinical utility of this nomogram to be 0–1 when predicting the septic status (Fig. 4B).

The transcriptomic septic changes of the hub-FRDEGs were initially examined in the training cohort (Fig. 4C and Table SVIII), the test cohort (Fig. 4D and Table SIX), validation cohort A (Fig. 4E and Table SIX) and validation cohort B (Fig. 4F and Table SIX), which revealed that the expression levels of both ATM and DPP4 were significantly decreased in the sepsis group when compared with control. TXN also revealed a transcriptomic elevation when examining the septic status. Subsequently, the ROC curves of biomarkers and the nomogram in the training cohort, test cohort, validation cohort A and validation cohort B showed the diagnostic accuracy of both hub-FRDEGs and nomogram was high in all cohorts(Fig. 4G-J), and these data are summarized in Table II. Furthermore, the confusion matrix of this nomogram in the training cohort (Fig. 4K) revealed that the model had an acceptable specificity (65.91%), high sensitivity (98.56%) and high diagnostic accuracy (94.90%), and the confusion matrix of this nomogram in the test cohort, validation cohort A and validation cohort B (Fig. 4L-N) revealed comparable or elevated values for the specificity, sensitivity and accuracy. Moreover, the unsupervised hierarchical clustering of the three hub-FRDEGs revealed their strength in distinguishing septic from control status (Fig. 4O); in addition, their strong performance in classifying sepsis in the test cohort, validation cohort A and validation cohort B was also confirmed (Fig. 4P and R). Finally, the calibration curve of the nomogram in the training cohort was drawn with the P-value of the Homer-Lemeshow test >0.05, which revealed a satisfactory performance in modelling fitness (Fig. 4S); parallel fitness of this diagnostic logistic model was also demonstrated by the calibration curve in the test cohort, validation cohort A and validation cohort B (Fig. 4T-V).

GO functional enrichment analysis of DEGs from the training cohort consists of biological process (BP), cellular component (CC) and molecular function (MF) categories (Table SX and Fig. 5A and B). The BP category included processes such as the ‘production of molecular mediators of the immune response’, ‘humoral immune response’ and ‘immunoglobulin production’. The CC category contained processes such as the ‘immunoglobulin complex’, the ‘external side of the plasma membrane’ and the ‘immunoglobulin complex, circulating’. Finally, the MF category featured processes such as ‘antigen binding, ‘immunoglobulin receptor binding’ and ‘structural constituents of chromatin’. KEGG pathway enrichment analysis of the DEGs mainly revealed processes such as ‘neutrophil extracellular trap formation’, ‘systemic lupus erythematosus’, transcriptional mis-regulation in cancer’ and ‘alcoholism’ (Table SXI and Fig. 5C and D). The results from GSVA revealed that septic progression primarily activates inflammatory and metabolic signaling pathways, including ‘cysteine-type endopeptidase inhibitor activity’, ‘antimicrobial humoral immune response mediated by an antimicrobial peptide’, ‘fatty acid binding’, ‘defense response to fungus’ and ‘endopeptidase regulator activity’, and cellular axes such as tissue morphogenesis and transformation were found to be deactivated, including ‘embryonic appendage morphogenesis’, ‘lung saccule development’, ‘regulation of animal organ morphogenesis’ and ‘negative regulation of vascular-associated smooth muscle cell differentiation’ (Table SXII and Fig. 5E).

Using the transcriptomic profile of three hub-FRDEGs as a suitable base, two distinct septic patterns were identified by consensus clustering, which were termed cluster A and cluster B (Table SXIII and Fig. 6A-D). A PCA plot of the septic clusters revealed their ferroptotic disparity (Fig. 6E). Compared with cluster A, the TXN and ATM expression levels were reduced in cluster B, and DPP4 revealed a comparable distinction between the clusters (Fig. 6F and G). After the subsequent analysis of the septic immune microenvironment, the abundance of 22 immunocytes between the two clusters was discerned using CIBERSORT (Fig. 6H). Analysis revealed ~50% of the immunocytes were differentially distributed, including M₀ macrophages, CD4⁺ memory resting T Cells, γδT Cells, regulatory T Cells (Tregs), monocytes, resting natural killer (NK) cells and neutrophils. Differential expression analysis of the clusters was subsequently performed, followed by GO and KEGG enrichment analyses. GO enrichment analysis of the clusters revealed that the BP category mainly contained ‘immunoglobulin production’, ‘complement activation, classical pathway’; the CC category predominantly included ‘immunoglobulin complex’, ‘immunoglobulin complex, circulating’ and ‘hemoglobin complex’; and the MF category included processes such as ‘antigen binding’, ‘immunoglobulin receptor binding’ and ‘haptoglobin binding’ (Fig. 6I). By contrast, KEGG enrichment analysis of the clusters only revealed an association with the ‘malaria’ pathway (Fig. 6J).

Septic changes of the immune microenvironment were investigated using the CIBERSORT and ssGSEA algorithms, respectively. Initially, the quantitative differences of the 22 immunocytes between the control and sepsis groups were analyzed (Table SXIV; Fig. 7A and B). The majority of the immunocytes were revealed to be septically modulated, as exemplified by the aggravated infiltration of neutrophils, monocytes and M₀ macrophages, and the sepsis-induced decreases in the populations of CD4⁺ memory resting T Cells, CD8⁺ T Cells and resting NK cells. The correlation between the transcriptomics of the hub-FRDEGs and the abundance of immunocytes was subsequently visualized by heatmap analysis (Fig. 7C). Both ATM and DPP4 were revealed to be positively associated with CD4⁺ memory resting T Cells, CD8⁺ T Cells, resting NK cells and naive B cells. TXN was negatively correlated with neutrophils, T follicular helper cells, M₀ macrophages and γδT Cells. The correlation between immunocytes was also visualized(Fig. 7D) . Furthermore, the immune scores comparing between samples with differential expression of hub-FRDEGs were compared using ssGSEA (Table SXV and Fig. 7E-G). Samples with a high expression of ATM were revealed to have increased infiltrated levels of activated B cells, activated CD4⁺ T Cells and activated CD8⁺ T Cells, although the opposite findings were obtained for activated dendritic cells, macrophages, monocytes and type 2 T helper cells (Fig. 7E). Similar results were observed in samples with high expression of DPP4, except for monocytes and type 2 T helper cells (Fig. 7F). However, samples with high expression of TXN exhibited reduced proportions of activated CD8⁺ T Cells, CD56⁺ NK cells, CD4⁺ central memory T Cells and CD8⁺ effector memory T Cells, whereas the oppose degree of infiltration was observed for activated CD4⁺ T Cells and γδT Cells (Fig. 7G).

To validate the expression of key genes in sepsis-induced ferroptosis, THP-1-derived macrophages were treated with LPS for 24 h to mimic this process and to establish an in vitro model. Findings of RT-qPCR and western blotting data were consistent with those of the bioinformatics data analysis. Primer sequences are presented in Table III. Compared with the control group, the relative mRNA expression levels of DDP4 (Fig. 8A) and ATM (Fig. 8B) were significantly decreased, whereas that of TXN (Fig. 8C) increased in the model group when compared with control. The results of the western blotting experiments revealed that, compared with the control group, the expression levels of ATM and DPP4 protein in the model group were significantly decreased, whereas that of TXN was increased (Fig. 8D and E).

TXN is a key antioxidant and regulator of ferroptosis (45). TXN was demonstrated to alleviate ferroptosis by the upregulation of GPX4 in a mouse model of Parkinson's disease (46), and has been identified as a hub ferroptotic regulator of proliferative diabetic retinopathy (47). Furthermore, TXN has been identified as the downstream target of retinoic acid receptor α, and as an inhibitor of ferroptosis in lung adenocarcinoma (48). The present study revealed that the AUC value of TXN was the highest among the ROC curves constructed for the three hub-FRDEGs. Therefore, TXN was further investigated to validate these findings. Reduced TXN expression was revealed in samples treated with si-TXN when compared with control (Fig. 8F-G). Ferroptosis-associated markers in TXN-inhibited and control macrophages were examined following LPS treatment. Western blotting experiments revealed significantly reduced expression levels of the proteins solute carrier family 7 member 11, GPX4 and ferritin heavy chain 1, which act as defenders against ferroptosis, in the LPS group compared with the control group. Compared with the si-NC+LPS and LPS groups, their expression was further decreased in the si-TNX+LPS group, compared with the LPS only treated group. Furthermore, the level of the lipid peroxidation-promoting protein ACSL4 was increased following LPS treatment compared with control, and treatment with si-TXN led to a further elevation of its expression level when compared with LPS only treated group (Fig. 8H and I). The FerroOrange staining experiments further suggested that LPS treatment caused an increase in Fe²⁺ levels in macrophages, and subsequent treatment with si-TXN augmented this increase (Fig. 8J). Treatment with LPS led to a decrease in the GSH/GSSG ratio of macrophages compared with control, and subsequent si-TXN treatment caused a further decrease in this ratio compared with LPS only treated samples (Fig. 8K). In addition, the MDA detection experiments suggested that treatment with LPS led to an increase in the lipid peroxidation level of the macrophages compared with control, and subsequent si-TXN treatment caused a further increase compared with the LPS only treated group (Fig. 8L). Taken together, these results reveal that treatment of the cells with TXN resulted in an increased level of ferroptosis in macrophages following sepsis.