When selecting datasets related to ARDS with whole blood or peripheral blood mononuclear cell samples, the dataset should include a sufficient sample size and provide raw or processed gene expression data. Datasets with insufficient sample size or poor quality control will be excluded. ARDS-related datasets were obtained from the Gene Expression Omnibus (GEO) database (http://www.ncbi.nlm.nih.gov/geo/). The GSE32707 (GPL10558) dataset included 31 whole blood samples from patients with sepsis-induced ARDS and 34 whole blood samples from individuals without sepsis, systemic inflammatory response syndrome (SIRS), or ARDS (referred to as ARDS and control samples, respectively), which served as training set 1. The GSE243066 (GPL30209) dataset consisted of 34 whole blood samples from patients with ARDS and 15 healthy controls, serving as training set 2. The GSE76293 (GPL570) dataset included 12 blood polymorphonuclear neutrophil samples from patients with ARDS and 12 healthy controls, serving as the validation set. Additionally, the GSE180578 (GPL24676) dataset included single-cell RNA sequencing (scRNA-seq) data from peripheral blood mononuclear cells (PBMCs) collected from 4 patients with ARDS and 4 healthy controls. A total of 1,804 metabolism reprogramming-related genes (MRRGs) were retrieved from the literature (27) ( Supplementary Table 1 ).

DEGs1 between ARDS and control samples in training set 1 were identified using the limma (v 3.54.0) package (28), applying a threshold of |log₂ fold-change (FC)| > 1 and P < 0.05. The signal intensity of chip data approximately follows a normal distribution. Limma, based on linear models and empirical Bayes methods, is specifically designed for such data and can effectively handle technical noise. For lowly expressed genes, Limma’s voom transformation enhances the detection ability of lowly expressed genes by applying weighted root mean square standard deviations. Similarly, DEGs2 between ARDS and control samples in training set 2 were identified using the Differential Expression Sequencing analysis (DESeq2, v1.42.0) package (29) with the same criteria. GSE243066 used a high-throughput sequencing platform, and the reads count of RNA-seq data followed a negative binomial distribution. DESeq2 directly models this distribution using a generalized linear model, avoiding information loss during the normalization process. DESeq2 uses the median ratio method for normalization, which is insensitive to extreme values and is suitable for handling outliers in sequencing data. Volcano plots for DEGs1 and heatmaps for DEGs2 were generated using the ggplot2 (v 3.5.1) package (30) and the ComplexHeatmap (v 2.14.0) package (31), respectively.

To evaluate the infiltration of 64 immune cell types in training set 1, relative abundance was calculated using the xCell (v 1.1.0) package (32), and the proportional distribution of immune cells was visualized using ggplot2 (v 3.5.1). Differences in immune cell infiltration between ARDS and control samples in training set 1 were assessed using the Wilcoxon test, identifying immune cell types with significant differences in infiltration (P < 0.05), referred to as differential immune cells. Subsequently, WGCNA was performed on training set 1 using the WGCNA (v 1.71) package (33). Hierarchical clustering was applied to ARDS and control samples in training set 1 to identify and remove outlier samples. The optimal soft threshold (β) was selected when the scale-free fit index (R²) exceeded 0.9, and the average connectivity was near 0. The adjacency between genes was then calculated, with each module containing at least 50 genes. Co-expression modules were identified, and a hierarchical clustering tree was generated. Using differential immune cell scores as phenotypic features, the correlation between these scores and co-expression modules was calculated (|correlation coefficient (cor)| > 0.30, P < 0.05) (34). The co-expression modules with the highest positive and negative correlations with differential immune cell scores were selected as key modules, and the key module genes were subsequently identified.

The intersection of DEGs1, DEGs2, MRRGs, and key module genes was determined using the ggvenn (v 0.1.10) package (35), and the overlapping genes were designated as candidate genes. Subsequently, Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses were performed on the candidate genes using the clusterProfiler (v 4.7.1.003) package (36, 37)(P < 0.05). The enrichment results from both GO and KEGG analyses were visualized with the GOplot (v 1.0.2) package (38). A protein-protein interaction (PPI) network (interaction score > 0.15) was constructed using the STRING database (https://www.string-db.org) and visualized using Cytoscape (v 3.7.1) software (39).

For candidate genes, Least Absolute Shrinkage and Selection Operator (LASSO) analysis was performed using the glmnet (v 4.1.4) package (40), with genes not penalized to zero selected for further analysis. Simultaneously, the caret (v 6.0-93) package (41) was employed to conduct Support Vector Machine Recursive Feature Elimination (SVM-RFE) analysis. The final feature genes were identified by intersecting the results from LASSO and SVM-RFE using the ggvenn (v 0.1.10) package. The expression patterns of these feature genes were then assessed in training set 1, training set 2, and the validation set (P < 0.05) to identify potential biomarkers. Finally, receiver operating characteristic (ROC) curves for these feature genes were plotted using the partial ROC (v 1.18.5) package (42), and genes with an area under the curve (AUC) > 0.7 were considered potential biomarkers across the three datasets.

To further validate the reliability of these biomarkers in predicting ARDS, In the GSE32707 dataset, key genes were selected as features, and all features were normalized using min-max scaling, adjusting the value range to (0, 1). Regarding the model architecture, the input layer consisted of 3 neurons, corresponding to the 3 feature genes, with a hidden layer containing 5 neurons and an output layer of 2 neurons for the binary classification task (ARDS and control). The activation function used the hyperbolic tangent function (tanh) for the hidden layer, and the Sigmoid function was applied to the output layer. The loss function chosen was cross-entropy (Cross Entropy), as it more effectively measures the difference between predicted and true labels in classification tasks. During training, the backpropagation algorithm was applied, with a learning threshold set at 0.1 and a random seed of 17 to ensure reproducibility. The optimization goal was to minimize the cross-entropy loss. Finally, the network structure was visualized using the plot() function from the neuralnet (v 1.44.2) package (43). ROC curves of the ANN model were then plotted for training set 1, training set 2, and the validation set, using the pROC (v 1.18.5) package, to assess the model’s performance (AUC > 0.7).

The chromosomal localization of biomarkers was determined using the RCircos (v 1.2.2) package (44). Gene set enrichment analysis (GSEA) was performed on the biomarkers using the c2.cp.kegg.v7.4.symbols.gmt file obtained from the Molecular Signatures Database (MSigDB) (https://www.gsea-msigdb.org/gsea/msigdb/) as the background gene set. Spearman correlations between the biomarkers and other genes were calculated using the psych (v 2.2.9) package (45). Subsequently, GSEA was performed for each biomarker using the clusterProfiler (v 4.7.1.003) package, with the criteria |normalized enrichment score (NES)| > 1 and P < 0.05. The top 5 signaling pathways, ranked by descending P-value, were presented using the enrichplot (v 1.18.0) package (46). Additionally, genes associated with biomarker functions and their respective roles were predicted from the GeneMANIA database (http://genemania.org), and a gene-gene interaction (GGI) network was constructed.

To predict miRNAs targeting the biomarkers, the DIANA-microT database (http://www.microrna.gr/microT) was used. The starBase database (https://rnasysu.com/encori/) was then employed to identify lncRNAs upstream of miRNAs (clipExpNum > 4), facilitating the construction of an lncRNA-miRNA-biomarker network. Transcription factors (TFs) targeting biomarkers were predicted using the NetworkAnalyst database (https://networkanalyst.ca/NetworkAnalyst/), resulting in the construction of a TF-biomarker network. Cytoscape (v 3.7.1) software was used to visualize these regulatory networks.

The DrugBank database (https://go.drugbank.com/) was utilized to predict potential drugs targeting the biomarkers. The Cytoscape (v 3.7.1) software was again used to visualize the drug-biomarker network. Based on the highest-scoring drugs targeting the biomarkers, molecular docking was performed using the CB-Dock2 online tool (https://cadd.labshare.cn/cb-dock2/php/index.php) to assess the binding ability of the biomarkers to these drugs (binding energies ≤ -5 kcal/mol) (47). The three-dimensional structures of the drugs were obtained from the PubChem database (https://pubchem.ncbi.nlm.nih.gov/), and the biomarkers were imported into the Protein Data Bank (PDB) database (https://www.rcsb.org/) to retrieve their protein structures.

For single-cell analysis, the GSE180578 dataset was processed using the Seurat (v 5.0.1) package (48). Cells and genes of low quality were filtered out (min.cells=3), and the remaining cells and genes were further selected based on stringent criteria (200 < nFeature_RNA < 3000, 200 < nCount_RNA < 15000, percent.mt < 20%). The filtered single-cell data were integrated using the Harmony function and normalized using the LogNormalize function. The top 2000 highly variable genes (HVGs) were identified using the vst method of the FindVariableFeatures function. Principal component analysis (PCA) was performed based on the HVGs, and a scree plot was generated using the Elbowplot function. The appropriate principal components (PCs) for downstream analysis were selected using the JackStraw function. Finally, high-quality cells were separated into distinct clusters using the FindNeighbors and FindClusters functions, with the uniform manifold approximation and projection (UMAP) clustering method. Marker genes for different clusters were identified using the FindAllMarkers function for further annotation (49–59) ( Supplementary Table 2 ).

The ARDS and control sample data from the GSE180578 dataset were used to analyze cellular communication networks between different cell types using the CellChat (v 1.6.1) package (60). The interaction relationships of receptors and ligands for each cell type were determined. Additionally, the Wilcoxon test was applied to assess differences in the expression of biomarkers across different cell types in ARDS and control samples (P < 0.05) to identify key cells.

Key cells were selected from the GSE180578 dataset for further dimensionality reduction and clustering. The identified key cells were then reclustered and categorized into distinct cell subpopulations. Following this, to investigate the potential differentiation or activation trajectories of key cells during the pathogenesis of ARDS, cell pseudo-time trajectory analysis was conducted using the Monocle (v 2.30.0) package (61), and the cell differentiation trajectories were visualized using the DDRTree (v 0.1.5) package (62).

Twenty whole blood samples were collected from Zhongda Hospital of Southeast University, consisting of 10 samples from patients with ARDS and 10 from healthy controls. The study was approved by the Ethics Committee of Zhongda Hospital Southeast University (approval number: 2022ZDSYLL402-P01), and informed consent was obtained from all participants. Total RNA was extracted from the 20 samples using TRIzol reagent (Ambion, Cat#15596026, USA) following the manufacturer’s protocol. RNA concentration was measured using the NanoPhotometer N50. cDNA synthesis was carried out via reverse transcription using the SweScript RT II First Strand cDNA Synthesis Kit (Servicebio Cat#G3333,China), and the process was performed with the S1000TM Thermal Cycler (Bio-Rad, USA). Primer sequences for RT-qPCR are provided in Supplementary Table 3 . RT-qPCR was performed using the CFX Connect Real-Time Quantitative Fluorescence PCR Instrument (Bio-Rad, USA), with β-actin used as the internal reference gene. The RT-qPCR results were analyzed using the 2-ΔΔCT method, exported to Excel, and then imported into GraphPad Prism 10.1.2 for statistical analysis and visualization (P < 0.05).

Bioinformatics analyses were performed using R (v 4.2.2). This study first identified differentially expressed genes (DEGs1) in training set 1 using the limma package (v 3.54.0) with the criterion of |log2fold-change (FC)| > 1 and P < 0.05. Using the same threshold criteria, DEGs2 were identified in training set 2 with the DEseq2 package (v 1.42.0). Based on these candidate genes, LASSO regression analysis was performed using the glmnet package (v 4.1.4) in R software, with the family parameter set to binomial and ten-fold cross-validation to select the feature genes. Subsequently, the caret package (v 6.0-93) in R was used to further screen feature genes using the random forest method, and recursive feature elimination (RFE) was applied to iteratively remove unimportant genes, ultimately obtaining the selected feature genes. Next, the Wilcoxon test was used to compare the expression differences of key genes between ARDS and control samples in the main training set (GSE32707), auxiliary training set (GSE243066), and validation set (GSE76293), with FDR correction applied for multiple testing and a significance threshold of p < 0.05. Additionally, the Wilcoxon test was used to analyze immune cell infiltration differences between disease and normal sample groups, with the Benjamini-Hochberg method for multiple testing correction (α=0.05).

A total of 523 dDEGs1 were identified in training set 1 through differential expression analysis, with 286 up-regulated and 237 down-regulated genes. In training set 2, 4,813 DEGs2 were detected, including 1,795 up-regulated and 3,018 down-regulated genes. The top 10 up-regulated and down-regulated DEGs and their expression profiles were displayed on volcano plots and heatmaps, respectively ( Figures 2A-D ). In both ARDS and control samples, the infiltration levels of 64 immune cell types were presented in a stacked plot. Supplementary Table 5 displays the top 10 immune cells with the highest infiltration abundance, with neutrophils exhibiting the most significant infiltration ( Figure 3A , Supplementary Table 5 ). It is worth noting that some of the control samples exhibited a relatively high proportion of immune cells. This phenomenon may have reflected individual differences during sample collection, such as potential subclinical infections, inflammatory responses, or other undefined physiological states, and these factors may have caused the non-specific activation and recruitment of immune cells in the control group. A total of 35 immune cells showed significant differences in infiltration ( Figure 3B , Supplementary Table 6 ). Following this, WGCNA was conducted using differential immune cell scores as traits. No abnormal samples were found in training set 1 ( Figure 3C ). The β value was determined to be 9 ( Figure 3D ). A co-expression matrix was then constructed, and 14 gene modules were identified ( Figure 3E ). Among these, MEbrown (cor=0.95) and MEturquoise (cor=-0.94), which showed the largest positive and negative correlations with differential immune cell scores, respectively, were considered as key modules ( Figure 3F ). These modules contained a total of 6,601 key module genes.

After overlapping DEGs1, DEGs2, key module genes, and MRRGs, 27 candidate genes were identified ( Figure 4A ). Enrichment analysis revealed that the 27 candidate genes were significantly associated with 408 GO terms, including “structural constituent of ribosome” ( Figure 4B , Supplementary Table 7 ). Additionally, KEGG analysis identified 101 enriched terms, primarily linked to amino acid biosynthesis and metabolic pathways, such as those involving arginine, proline, and glutamate ( Figure 4C , Supplementary Table 8 ). These results suggest a strong association between the candidate genes and ribosomal function, as well as amino acid metabolism. In the PPI network, two discrete targets were removed, and genes like NME1 and ASS1 showed strong connectivity with other genes, highlighting their potential relevance ( Figure 4D ).

Through LASSO analysis, 10 feature genes were selected, including RPL37A, VNN2, RPS28, NME1, TCN1, RPL14, SMARCD3, ASS1, KYNU, and CKB ( Figures 4E, F ). The SVM-RFE analysis identified 22 feature genes, including SMARCD3, TCN1, UPP1, RPL37A, HK3, KYNU, VNN1, NUP214, RPL14, PGS1, RPS28, RPL10A, ARG1, RPL23A, PFAS, CKB, ITPR3, ASS1, ALOX5, ALOX5AP, SAT1, and FASN ( Figure 4G ). A final set of 8 feature genes was derived ( Figure 4H ).

Expression analysis indicated that CKB and RPL14 were significantly down-regulated in ARDS samples across training set 1, training set 2, and the validation set. Conversely, SMARCD3 and TCN1 were up-regulated in ARDS samples across all datasets. The expression patterns of ASS1 and KYNU were inconsistent, and no significant expression differences for RPL37A and RPS28 were observed ( Figures 5A-C ). Consequently, CKB, RPL14, SMARCD3, and TCN1 were identified as candidate biomarkers. Furthermore, the AUC values for RPL14, SMARCD3, and TCN1 exceeded 0.7 across the three datasets ( Figures 5D-F ), supporting their potential as reliable biomarkers.

To further validate these biomarkers, RT-qPCR analysis was performed using human whole-blood samples. The results confirmed a significant upregulation of SMARCD3 and TCN1 in ARDS samples (P < 0.05) ( Figure 5G ), consistent with the findings from training set 1, training set 2, and the validation set. While RPL14 showed a downward trend in ARDS samples, the difference was not statistically significant (P=0.44).

Based on these biomarkers, an ANN model was constructed to evaluate their predictive power in ARDS onset ( Figure 6A ). The ROC curve analysis showed that the AUC values for training set 1, training set 2, and the validation set were 0.97 (specificity: 0.9706, sensitivity: 0.9677), 0.80 (specificity: 0.8670, sensitivity: 0.7330), and 0.75 (specificity: 0.9170, sensitivity: 0.5830), respectively ( Figures 6B-D ).

RPL14, SMARCD3, and TCN1 were mapped to chromosomes 3, 7, and 7, respectively ( Figure 6E ), suggesting that these three biomarkers may have distinct biological functions. GSEA revealed that RPL14, SMARCD3, and TCN1 were enriched in 82, 89, and 69 pathways, respectively ( Supplementary Tables 9 – 11 ). Notably, the ribosome pathway was significantly associated with both RPL14 and SMARCD3, while chemokine, B cell receptor, and T cell receptor signaling pathways were uniquely linked to both SMARCD3 and TCN1 ( Figures 6F-H ). These findings indicate that these biomarkers may contribute to ARDS by modulating T/B cell immune responses, in addition to the inflammatory response.

The GGI network constructed revealed the top 20 genes associated with the function of the biomarkers, such as CBLIF, which are involved in water-soluble vitamin metabolic processes ( Figure 6I ).

A total of 44 miRNAs were predicted to target RPL14 and SMARCD3, while no miRNA targeting TCN1 was identified. Based on a threshold of clipExpNum > 4, 107 lncRNAs were predicted by the 20 miRNAs. Consequently, a lncRNA-miRNA-mRNA network involving 2 biomarkers, 20 miRNAs, and 107 lncRNAs was constructed ( Figure 7A ). Additionally, 62 TFs targeting RPL14, SMARCD3, and TCN1 were identified, and the TF-biomarker network was generated ( Figure 7B ). Among these, KLF9 was found to target both RPL14 and SMARCD3, suggesting common transcriptional regulation in ARDS for these biomarkers.

The drug-biomarker network revealed that 41 drugs targeted RPL14, 4 targeted SMARCD3, and 6 targeted TCN1. Notably, tetradioxin (CTD 00006848) was identified as a drug targeting both SMARCD3 and TCN1, while selenium (CTD 00006731) and CsA (CTD 00007121) were found to target both RPL14 and SMARCD3 ( Figure 7C , Supplementary Table 12 ). Molecular docking was then performed using the highest-scoring drugs targeting the biomarkers. Mesalazine (MCF7 UP) and arbutin (CTD 00005438) showed the highest scores for RPL14 and TCN1, respectively. For SMARCD3, although selenium (CTD 00006731) and methaneseleninic acid (CTD 00000412) had relatively high scores, their 3D structures were unavailable, so tetradioxin (CTD 00006848) was used instead. The binding energies of the complexes formed by RPL14 and mesalazine ( Figure 7D ), SMARCD3 and tetradioxin ( Figure 7E ), and TCN1 and arbutin ( Figure 7F ) were -4.5, -6.8, and -7.1 kcal/mol respectively, with the binding centers located at (-4, -3, -1), (-22, 38, 4), and (-15, -3, 10) ( Supplementary Table 13 ). The binding energy of RPL14 and mesalazine did not reach -5 kcal/mol, indicating a relatively low affinity between the ligand and receptor, suggesting that the complex formed might be easily dissociated. Overall, these results highlight the potential therapeutic relevance of RPL14, SMARCD3, and TCN1, particularly in the context of ARDS treatment.

In the GSE180578 dataset, prior to quality control, a total of 57,811 cells and 19,704 genes were identified. After quality control, 53,816 cells and 19,704 genes were retained ( Figures 8A, B ). The top 2,000 HVGs and the top 30 PCs were then used for UMAP clustering ( Figures 8C, D ). This process led to the classification of all high-quality cells into 16 distinct clusters ( Figure 8E ). The cell clusters were annotated, revealing 8 cell types, including B cells, endothelial cells, macrophages, natural killer cells, neutrophils, plasma cells, red blood cells, and T cells ( Figure 8F ). The marker genes displayed high specificity across different cell clusters ( Figure 8G ). Notably, T cells comprised the highest proportion among all cell types. A comparison between ARDS and control samples revealed that macrophages, endothelial cells, and neutrophils were significantly more abundant in ARDS samples ( Figures 8H, I ).

The cell-cell communication network analysis showed that, in ARDS samples, macrophages and neutrophils exhibited a greater number of interactions with other cell types compared to control samples ( Figures 9A, C ). This suggests that macrophages and neutrophils may play more pivotal roles in ARDS than other cell types. In ARDS samples, the interaction intensity of Neutrophils with other cells was stronger than that in control samples ( Figures 9B, D ). Figures 9E, F display the receptor-ligand pairings between different cell types in ARDS and control samples. Among these pairings, MIF-(CD74+CXCR4) and MIF-(CD74+CD44) were the core regulatory pathways for intercellular communication in both types of samples, and they could collectively regulate a variety of intercellular interactions, including T cell→Macrophage, Plasma cell→B cell, and Plasma cell→Macrophage. Furthermore, the expression patterns of RPL14, SMARCD3, and TCN1 at the cellular level were examined. It was found that RPL14 and SMARCD3 were expressed in multiple cell types, whereas TCN1 was not expressed ( Figure 10A ). Among these, RPL14 and SMARCD3 showed significant differential expression in macrophages and neutrophils ( Figures 10B, C ), making them key cells for further investigation.

Secondary dimensionality reduction clustering analysis was performed on the key cells. It was observed that macrophages were divided into 16 subtypes, and neutrophils were divided into 9 subtypes ( Figures 11A, B ). These subtypes were then placed along a differentiation trajectory based on differentiation time sequence, with darker shades of blue corresponding to earlier stages of differentiation. Macrophages were found to have 7 differentiation states, with State 1 representing the earliest and most specific stage ( Figure 11C ). Neutrophils exhibited 3 differentiation states, with State 1 also being the earliest and most specific ( Figure 11D ).

As macrophages differentiated, the expression of RPL14 increased initially and then decreased, while the expression of SMARCD3 followed an opposite trend ( Figure 11E ). In neutrophils, RPL14 expression generally trended upward, whereas SMARCD3 expression initially increased and then decreased ( Figure 11F ). These results suggest that the expression trends of RPL14 and SMARCD3 during differentiation of key cells were heterogeneous, potentially reflecting their distinct roles in cell differentiation. RPL14 and SMARCD3 likely function in different capacities during this process, influencing the differentiation trajectories of macrophages and neutrophils in ARDS.

Building upon these results, we next sought to determine whether these effects could be recapitulated in vivo.C57BL/6J mice were stratified into control and LPS-induced acute lung injury groups, and the validity of the model was validated using pathological morphology and inflammation/oxidative stress indicators. Histopathological examination of lung tissues via HE staining revealed pronounced structural disruption of alveoli in murine model of acute lung injury, characterized by widespread thickening of alveolar septa ( Figure 12A ). Further quantification by ELISA showed markedly increased concentrations of proinflammatory cytokines IL-6 and TNF-α, coupled with a significant reduction in GSH levels in lung tissue homogenates from LPS-induced acute lung injury mice relative to controls ( Figure 12B ). Subsequent RT-qPCR analysis demonstrated that RPL14 mRNA expression remained unchanged in both lung tissue homogenates and peripheral blood neutrophils of LPS-induced acute lung injury mice. In contrast, SMARCD3 and TCN1 transcript levels were significantly elevated in both lung tissue ( Figure 12C ) and neutrophils ( Figure 12D ) compared to control.

To decipher the cellular mechanisms underlying the observed lung injury, we next performed in vitro experiments. THP-1 cells were first transfected with siRNA targeting RPL14, SMARCD3, or TCN1, then stimulated with LPS ( Figure 13A ). ROS detection ( Figure 13B ) revealed that RPL14 silencing did not affect mitochondrial membrane potential and concomitantly showed no change in JC-1 staining, with TUNEL assay further confirming no significant effect on cellular apoptosis. In contrast, both SMARCD3 and TCN1 silencing led to reduced ROS production and increased mitochondrial membrane potential, accompanied by suppressed apoptotic activity. Compared to the control group, GSH levels were significantly increased by SMARCD3 and TCN1 silencing after LPS stimulation ( Figures 13C-E ). RT-qPCR analysis revealed that RPL14 interference did not affect IL-6 or TNF-α mRNA expression, whereas SMARCD3 and TCN1 knockdown significantly reduced both cytokines ( Figure 13F ). Consistently, ELISA measurements showed RPL14 silencing had no notable effects on IL-6/TNF-α secretion; however, SMARCD3 and TCN1 silencing markedly inhibited IL-6/TNF-α secretion ( Figure 13G ). In terms of alterations in glucose metabolism, silencing SMARCD3 or TCN1, but not RPL14, decreased lactate production, glucose uptake and glucose consumption compared to the control group. ( Figures 13H-J ).

To further investigate the role of CsA in LPS-induced macrophages, we differentiated THP-1 cells into macrophages with PMA (63). We measured cytokine expression by RT-qPCR, the result showed that LPS stimulation significantly upregulated IL-6 and TNF-α mRNA, while medium/high-dose CsA attenuated this upregulation ( Figure 14A ). Notably, the inhibitory effects of CsA extended beyond cytokine modulation to transcriptional regulators. LPS-treated macrophages exhibited unchanged RPL14 mRNA but significantly increased SMARCD3 expression. The LPS+CsA group showed no alteration in RPL14 versus LPS-only group, but displayed significant SMARCD3 downregulation ( Figure 14B ). These findings indicate that LPS stimulation enhances SMARCD3 expression in macrophages, while CsA administration reverses this LPS-induced SMARCD3 upregulation.

JC-1 ( Figure 14C ) and TUNEL assay ( Figure 14D ) revealed that LPS stimulation significantly reduced mitochondrial membrane potential and induced apoptosis in macrophages, effects that were dose-dependently reversed by medium/high-concentration CsA treatment. Concurrently, LPS triggered pro-inflammatory cytokine overproduction (IL-6, TNF-α) and depleted intracellular GSH, all of which were attenuated by CsA in a concentration-dependent manner. Notably, all CsA-treated groups exhibited significantly reduced cytokine secretion and restored GSH levels versus LPS-only controls, with maximal efficacy at medium/high doses ( Figure 14E ). These data demonstrate that CsA mitigates LPS-induced macrophage injury, and its protective effects are associated with downregulation of SMARCD3.