Identification and Validation of Key Genes Related to Arginine Methylation Modification in Sepsis Using Transcriptome Combined with Mendelian Randomization Analysis

Abstract

Background:

Patients with sepsis often exhibit a decrease in lymphatic numbers, which can be facilitated by protein arginine methyltransferase (PRMT). However, it is unclear how PRMT contributes to lymphopenia in sepsis.

Methods:

This study employed the sepsis-related datasets (GSE65682 and GSE134347) and nine PRMT genes. First, we intersected the differentially expressed genes with the weighted gene co-expression network analysis module genes to identify differentially expressed PRMT-related genes. Thereafter, candidate key genes were obtained after Mendelian randomization analysis and machine learning screening. Eventually, we subjected key genes identified by expression analysis and receiver operating characteristic curves to gene set enrichment analysis, immune infiltration analysis, immune checkpoint analysis, molecular docking, regulatory networks construction, and nomogram development.

Results:

We first intersected 4,246 differentially expressed genes with 1,884 PRMT scoring module genes to obtain 969 differentially expressed PRMT-related genes. Further Mendelian randomization analysis and machine learning jointly identified five candidate genes (CRLF3, ELAC2, PBX2, MCTP2, and EMB). Among these, ELAC2, PBX2, MCTP2, and EMB demonstrated consistent expression trends, with the area under the curve values of the receiver operating characteristic curve exceeding 0.7 in GSE65682 and GSE134347. Therefore, they were defined as key PRMT-related genes. The gene set enrichment analysis showed enrichment in cytoplasmic translation (ELAC2, MCTP2), noncoding RNA metabolism (EMB), and metabolic processes (PBX2). The immune infiltration analysis revealed a significant correlation between PBX2 and neutrophils, as well as between ELAC2/MCTP2/EMB with activated natural killer cells, CD8+ T cells.

Conclusion:

In this study, ELAC2, PBX2, MCTP2, and EMB were identified as key genes related to PRMT for sepsis, which provided a theoretical basis for the study of sepsis.

INTRODUCTION

Sepsis, a pathological condition triggered by microorganism infection, leads to a dysregulated host immune response (1,2). It is a global health crisis that affects millions of people worldwide annually, with its epidemiology underscoring high incidence and mortality rates, marking it as a major public health concern (3). Sepsis is characterized by widespread inflammation and blood clotting, which can result in organ damage and failure (4–6). Pathologically, sepsis involves the activation of multiple inflammatory pathways (7), endothelial dysfunction (8), and impaired perfusion (9), all of which contribute to multiorgan dysfunction syndrome (MOD) (10). Major risk factors for sepsis include advanced age, compromised immune systems, chronic illnesses such as diabetes, cancer, and liver disease, as well as recent surgery or long hospitalization (11). Given that sepsis involves complex pathological mechanisms and multiple risk factors, current research is still exploring effective treatment options. Therefore, further basic research and clinical trials are urgently needed to gain a deeper understanding of its pathogenesis, explore new therapeutic avenues, and ultimately improve patient prognosis.

Protein arginine methyltransferases (PRMTs) are a class of enzymes that can perform monomethyl and dimethyl modifications on the nitrogen atom of the guanidine group of arginine. Nine family members have been identified to date, collectively mediating extensive arginine methylation modifications within cells (12). PRMTs play an important role in regulating signal transduction, DNA damage repair, and RNA metabolism (6,7). In terms of immune regulation, PRMTs are involved in inflammatory responses by modulating the expression of inflammatory factors and the function of immune cells. For instance, PRMT1 can influence the expression of inflammation-related genes by regulating the phosphorylation levels of signal transducer and activator of transcription proteins (12). Multiple studies have demonstrated that PRMTs play crucial roles in the activation and functional execution of immune cells such as T cells and macrophages (12,13). Particularly in macrophages, alterations in the expression and activity of PRMTs can modulate their polarization state, thereby influencing the progression of inflammatory responses (14).

Importantly, PRMTs contribute significantly to the pathogenesis of sepsis by participating in inflammatory and immunoregulatory processes. A clinical study has shown that reduced lymphocyte numbers are commonly observed in sepsis patients, with PRMT4 implicated in this pathological process (9). Additionally, PRMT2 promotes interferon-β production through the TLR4/IRF3 signaling pathway, suggesting its potential role in regulating endotoxemia (15). Further research has revealed that PRMT2 can influence macrophage polarization, providing preliminary insights into its mechanism in immune regulation (16). Recent studies indicate that PRMTs affect the expression of sepsis-related inflammatory factors through epigenetic mechanisms. For example, PRMT1 enhances the transcription of proinflammatory cytokines such as tumor necrosis factor-α and interleukin-6 by methylating key proteins in the NF-κB signaling pathway (17). Moreover, PRMT1 was found to activate glutaminase 2 by methylating early growth response 1, thereby driving sepsis-related acute lung injury, suggesting therapeutic potential for targeting PRMT1 in sepsis (18). Nevertheless, the precise molecular mechanisms of arginine methylation in sepsis pathogenesis remain incompletely elucidated, and its potential value as a therapeutic target requires further in-depth investigation.

Transcriptomic analysis employs high-throughput technologies (e.g., RNA sequencing) to profile genome-wide RNA expression dynamics, uncovering molecular mechanisms, differentially expressed genes (DEGs), and regulatory networks underlying diseases or phenotypes (19). This approach aids in identifying biomarkers or therapeutic targets. Mendelian randomization (MR) uses genetic variants as instrumental factors to assess causal exposure-outcome connections, mitigating confounding and reverse causation biases inherent in observational studies (20,21). Integrating transcriptomic data with MR prioritizes causal genes by linking genetically predicted expression to disease risk, bridging molecular insights and population-level evidence to enhance causal inference in complex diseases.

Here, we hypothesize that PRMT-related genes exhibit a specific expression pattern in sepsis and participate in the pathogenesis and progression of sepsis by regulating key pathways such as immune-inflammatory responses and cellular metabolism. Furthermore, it is hypothesized that an integrated multiomics strategy combining transcriptomics and MR analysis can systematically identify critical gene targets with central regulatory functions in sepsis. Therefore, our study used the data from sepsis-related public databases and comprehensive bioinformatics methods to systematically investigate the possible impact of arginine methylation in the progression of sepsis and its associated key genes. This study reveals the genes and their biological pathways associated with arginine methylation and sepsis, providing a novel conceptual framework for a better comprehension of the molecular mechanisms of sepsis.

MATERIALS AND METHODS

Data source

In this study, sepsis-related datasets were derived from the Gene Expression Omnibus (GEO) database (https://www.ncbi.nlm.nih.gov/gds). Among them, GSE65682 included blood samples from 760 sepsis patients and 42 controls as a training set. While GSE134347 served as a validation set, including 156 blood specimens of patients with sepsis and 83 controls. All blood samples were collected in the early stage of sepsis diagnosis (within 24 hours of admission). The detailed clinical information of the two datasets is presented in Tables S1 and S2, https://links.lww.com/SHK/C651. Moreover, the ieu-b-4980 was downloaded from the Open Genome-Wide Association Study (GWAS) database (https://gwas.mrcieu.ac.uk/), including 11,643 samples of sepsis, 474,841 samples of control, and 12,243,539 single nucleotide polymorphisms (SNPs). Besides, PRMTs were collected based on Maron et al. (22), mainly comprising PRMT1, PRMT2, PRMT3, PRMT4, PRMT5, PRMT6, PRMT7, PRMT8, and PRMT9. All septic patients included in the dataset were diagnosed in accordance with the Sepsis-3 criteria. These criteria defined sepsis as life-threatening organ dysfunction caused by a dysregulated host response to infection, with a Sequential Organ Failure Assessment score of ≥2 (23). The control group consisted of healthy volunteers with no history of infection, no chronic organ dysfunction, and no autoimmune diseases. To reduce nonbiological variations caused by differences in batches, platforms, or experimental conditions, this study used the ComBat method in the sva R package (v 3.36.0) (24) to perform batch correction on the training set and validation set, respectively.

Weighted gene co-expression network analysis (WGCNA)

First, we organized the PRMT genes into gmt expression profile files, using the single-sample gene set enrichment analysis (ssGSEA) algorithm as the background for ssGSEA, and calculated the PRMT scores of all the samples in GSE65682. We then used the rank-sum test to determine if the sepsis and control groups’ PRMT scores differed from one another. Subsequently, to find out the pivotal module genes associated with PRMT score, we carried out WGCNA analysis on all specimens of the GSE65682 via WGCNA package (v 1.69) with PRMT score as a trait (25). Specifically, we first clustered the samples to understand their correlation. We then determined a soft threshold for the data to maximize the conformance of gene interactions to the scale-free distribution. The modules were separated using the hybrid dynamic tree cutting technique, which established a minimum of 100 genes in each gene module. Moreover, according to the correlation between modules and traits, key modules were identified with nongrey modules, P value <0.05, and |cor| > 0.3 as screening conditions. Finally, we screened key module genes on the basis of |Gene significance (GS)| > 0.2 and |Module membership (MM)| > 0.6, respectively.

Acquisition and functional enrichment analysis of differentially expressed PRMT-related genes (DE-PRMT-RGs)

To identify DEGs among different samples, differential expression analysis was performed on the samples in GSE65682 through the limma package (v 3.44.3) using adjusted P < 0.05 and |log₂fold change| > 0.5 as the screening criteria (26). And the DE-PRMT-RGs were obtained by taking intersections of DEGs and key module genes, which were subjected to establish protein–protein interaction (PPI) network upon the STRING site (https://string-db.org) to explore their interactions (confidence = 0.9). Cytoscape (v 3.10.1) was used to view the PPI network (27). To explore the DE-PRMT-RGs’ biological roles, Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses (adjusted P < 0.05) were carried out via clusterProfile package (v 3.16.0) (28).

MR analysis

The publicly available GWAS database (https://gwas.mrcieu.ac.uk/) provided the OpenGWAS-ids of DE-PRMT-RGs. These OpenGWAS-ids were analyzed for exposure variables, with sepsis serving as an outcome for MR analysis. The MR analysis rests upon three fundamental premises: 1) a notable correlation was established between instrumental variables (IVs) and exposure factors; 2) IVs were devoid of any confounding factors; 3) IVs could only influence outcomes through exposure factors and not through other pathways. The exposure factors were initially read and the IVs (SNPs) were identified by the extractc_instruments function of the TwoSampleMR package (v 0.5.6) (20). The screening criterion was defined as P < 5 × 10⁻⁸, which identified IVs (SNPs) that were significantly associated with the exposure factors. To circumvent potential bias due to linkage disequilibrium, IVs (SNPs) with linkage disequilibrium were excluded utilizing the parameters “clump = TRUE, r2 = 0.001, and kb = 10,000” (29). The filtered SNPs were then subjected to calculating an F value (F = beta²/SE²), which was generally considered to be a strong IV if the F value was larger than 10 (30). Afterward, the MR analysis was carried out by five methods: MR Egger (31), weighted median (32), inverse variance weighted (IVW) (33), simple mode (20), weighted mode (34), among which the IVW method played a decisive role in the results (P < 0.05). We presented the MR results in funnel plots, forest plots, and scatter plots. Consequently, heterogeneity test, horizontal pleiotropy assays, and Leave-One-Out analyses were included in sensitivity analysis to determine the reliability of the MR analysis results. The MRPRESSO R package was used to perform outlier SNP tests on exposed parrots. A Global Test P value <0.05 indicated the detection of significant horizontal pleiotropy, implying the presence of potential outlier SNPs and suggesting that there might be bias among IVs (SNPs), which affected the reliability of causal inference. A Global Test P value ≥0.05 indicated that no significant pleiotropy was detected, and the IVs were relatively “clean,” making the inference more reliable. At last, according to the genes with significant causality as the exposure factor and sepsis as the outcome, a Steiger test was conducted to determine whether the MR analysis would interfere with the reverse causality (correct_causal_direction = 1, steiger_test_ad j < 0.05). We referred to the genes screened by the above analysis as candidate genes.

Screening critical genes by machine learning

In this study, candidate genes were further screened by the least absolute shrinkage and selection operator (LASSO) (family = binomial), support vector machine–recursive feature elimination (SVM-RFE), and Boruta analysis, which were executed through the glmnet package (v 4.0-2), e1071 package (v 1.7-3), and RFB package (v 7.0.0), respectively (35–37)). Furthermore, we assessed the ability of the LASSO and Boruta models to discriminate between samples using receiver operating characteristic (ROC) curves. The potential critical genes were derived from the intersection of characteristic genes regarding the three machine learning techniques. Further, candidate key genes were used for expression analysis and ROC curve construction in GSE65682 and GSE134347, and genes with consistent and significant expression trends, and area under the ROC curve (AUC) values for both datasets larger than 0.7 were considered essential genes for further examination.

Gene set enrichment analysis of key genes

For the purpose of gaining insight into the regulatory pathways and biological functions related to key genes, GSEA analyses were conducted through the clusterProfiler package (v 3.16.0) (28). The correlation coefficients of each of the important genes were initially calculated and ranked individually. Subsequently, GSEA was conducted according to a screening condition of normalized enrichment score > 1, adjusted P < 0.05. Meanwhile, the background gene set for GSEA was downloaded from the Molecular Signatures Database (https://www.gsea-msigdb.org/).

Immune infiltration analysis

To explore the infiltration of immune cells in samples from GSE65682, the proportion of 22 immune cell types in various specimens was calculated based on the CIBERSORT algorithm (v 1.03) and the LM22 gene marker (38). Furthermore, the content of each type of cell in the tissue was analyzed by the linear support vector regression method for the back convolution of the expression matrix. Moreover, the immune cells’ differences between sepsis and control groups were examined using the rank-sum test, and the association between differential immune cells and key genes was investigated by Spearman analysis.

Construction of regulatory networks for key genes

We constructed the relevant regulatory networks to further explore the molecular regulatory mechanisms of key genes. First, micro ribonucleic acids (miRNAs) of key genes were predicted based on the miRWalk website (http://mirwalk.umm.uni-heidelberg.de/), which was filtered by an energy parameter less than −30 to get key miRNAs. Besides, key long-noncoding RNAs (lncRNAs) for key miRNAs were screened based on the starBase website (https://starbase.sysu.edu.cn/), and the condition that the lncRNA expression level was greater than 1. And the competitive endogenous RNAs (ceRNAs) network diagram was visualized using Cytoscape software (v3.10.1). In addition, the TRRUST database on the NetworkAnalyst website (https://www.networkanalyst.ca/) was used to identify potential transcription factors (TFs) for key genes, and the results were presented in the mulberry diagrams. To further explore which drugs would affect key genes, we employed the Comparative Toxicogenomics Database (https://ctdbase.org/) to figure out their potential target drugs. We visualized drug pairs with increases and decreases in expression with key genes using the Cytoscape software (v3.10.1).

Immune checkpoints analysis

To understand the associations between immune checkpoints and the occurrence of sepsis, 32 immune checkpoints were obtained. Then, based on all samples in GSE65682, the distinctions in the 32 immune checkpoints between the sepsis samples and controls were compared through the Wilcoxon test (P < 0.05), and differentially expressed immune checkpoints were obtained.

Construction and evaluation of nomogram

To evaluate how key genes differentiate between sepsis samples and control samples, a nomogram model was constructed using all samples from the GSE65682 dataset with the rms R package (v 6.5.0) (39). The calibration curve was assessed, and slope proximity to 1, ROC curve (1 > AUC > 0.7), and decision curves (the net income value >0) of the nomogram were drawn using the rms R packages, pROC R packages (v 1.18.5) (40), and ggDCA R packages (v 1.2) (41), respectively, to assess the predictive capability of the nomogram model.

Molecular docking

To define the binding ability between the drug and the key genes, molecular docking was performed; the selection of the top one drug was based on the interaction score for each key gene. The three-dimensional molecular structures of the corresponding drug were obtained using PubChem database (https://pubchem.ncbi.nlm.nih.gov/), and the protein structures of key genes were obtained through the PDB database (https://www.rcsb.org/). AutoDock software (v 4.2.6) (42) was employed to conduct molecular docking of the key genes as well as the intersection drugs, and the free binding energy data were obtained through molecular docking simulations. The Pymol software (v.2.5.7) (43) was used to display the molecular docking results (binding energy <−5 kcal/mol indicated high binding affinity).

Quantitative reverse-transcription polymerase chain reaction (qRT-PCR) verification

We gathered five pairs of frozen whole blood specimens in our department and extracted total RNA to confirm the repeatability and dependability of the four genes in sepsis patients. All samples were collected in the early stage of sepsis diagnosis (within 24 hours of admission). The detailed clinical information of these five pairs of clinical verification samples is shown in Table S3, https://links.lww.com/SHK/C651. The First Affiliated Hospital of Fujian Medical University’s Institutional Review Board and Ethical Committee authorized our study (Approval No.: MRCTA, ECFAH Of FMU [2021]679), and each patient gave written informed consent. The primer sequences were as follows: GAPDH: forward 5′-CGAAGGTGGAGTCAACGGATTT-3′, reverse 5′-ATGGGTGGAATCATATTGGAAC-3′; EMB: forward 5′-AATCCATCCCATTGCCTCCC-3′, reverse 5′-AGAAAGTGGACTCAGGCAGC-3′; MCTP2: forward 5′-CTCGCACAGATACCCAGACA-3′, reverse 5′-ATGGCTGCATGAAGACCCAA-3′; ELAC2: forward 5′-ACACGGAAGCAAGCAGATGA-3′; reverse 5′-GCAGCAGGAGGAAGCAAAAG-3′; PBX2: forward 5′-AGCTATGGACGAACGGCTAC-3′, reverse 5′-AGAGAGCAGGCTTCATTCGG-3′. In accordance with the manufacturer’s instructions, their total RNA was extracted using TRIzol reagent (Invitrogen, Carlsbad, California). The RNA purity was detected using NanoDrop 2000 spectrometer (Thermo Fisher Scientific, Waltham, Massachusetts). Reverse-transcription reactions using StarScript II First-strand cDNA Synthesis Mix (GenStar BioSolutions, Beijing, China) were conducted in two steps according to its specification. After the manufacturer’s instructions, qRT-PCR was carried out using the Step One PlusTM PCR System (Applied Biosystems, SantaClara, California) using RealStar Green Fast Mixture (GenStar BioSolutions) to detect the expression levels of the four genes. The results were adjusted based on GAPDH expression. Using the 2^−∆∆Ct technique, which is ∆Ct = Ct_genes – Ct_GAPDH and ∆∆Ct = ∆Ct_{experimental group} – ∆Ct_{control group}, the qRT-PCR data were computed.

Statistical analysis

GraphPad 6.0 (GraphPad Software, San Diego, California) and SPSS version 23.0 (SPSS, Chicago, Illinois) were used for statistical analyses. The Student t test was employed to compare the two groups. Paired t test was used to compare the paired samples. P < 0.05 was regarded as statistically significant.

RESULTS

There were 1,884 key module genes in yellow, red, tan, turquoise, and green modules

The effectiveness of the correction was verified by principal component analysis—after correction, the samples were clustered more clearly according to the “disease/control” groups, and batch interference was effectively controlled (Figure S1, https://links.lww.com/SHK/C651). Based on the ssGSEA algorithm, the sepsis group’s PRMT scores were substantially lower than the control group’s (Fig. 1A). We excluded 2 outlier samples (Fig. 1B) from GSE65682 by setting the “distance to hclust clustering” threshold at 130 (Fig. 1, C and D). The mean of the adjacency function was near zero, and the R2 was around 0.85 when the ideal soft threshold was 7 (Fig. 1, E and F). Thereafter, 21 gene modules were delineated, among which yellow, red, tan, turquoise, and green modules were considered to be the key modules (Fig. 1, G and H and Table S4, https://links.lww.com/SHK/C651). We obtained a total of 1,884 key module genes related to PRMT score using the criterion of | GS| > 0.2 and |MM| > 0.6 (Fig. 1, I–M and Table S4, https://links.lww.com/SHK/C651).

Fig. 1.

1,884 key module genes associated with PRMT scores were identified. A, The PRMT scores were calculated for all samples in the GSE65682 dataset using the ssGSEA algorithm. B–G, WGCNA was performed on all samples in the GSE65682 dataset to aggregate modules associated with PRMT scores. H, Correlation of 21 gene modules with PRMT scores. I–M, Five crucial gene modules (yellow, red, tan, turquoise, green) were screened with the conditions of nongrey modules, P value <0.05, and correlation |cor| > 0.3, and a total of 1,884 key module genes associated with PRMT scores were obtained. PRMT, protein arginine methyltransferase; ssGSEA, single-sample gene set enrichment analysis; WGCNA, weighted gene co-expression network analysis.

A total of 969 DE-PRMT-RGs were significantly enriched in ribosome, protein targeting to ER, and other processes

After differential analysis, there were 4,246 DEGs in normal and sepsis samples, of which 1,341 were upregulated genes and 2,905 were downregulated genes (Fig. 2A). The heatmap displayed the expression of the top 20 upregulated and downregulated DEGs (Fig. 2B). The DEGs were overlapped with key module genes to identify 969 DE-PRMT-RGs (Fig. 2C). In the PPI network, there were 411 DE-PRMT-RGs forming 1,632 interaction relationships, such as NAGK-GNPNAT1, PGM2-HK3, etc. (Fig. 2D). Based on the KEGG database, the top five signaling pathways significantly enriched by DE-PRMT-RGs included ribosome, Th1 and Th2 cell differentiation, and so on (Fig. 2E). Similarly, in GO enrichment analysis, the top five entries significantly enriched for DE-PRMT-RGs contained cotranslational protein targeting to membrane, protein targeting to ER, etc. (Fig. 2F).

Fig. 2.

A total of 969 DE-PRMT-RGs were significantly enriched in ribosome, protein targeting to ER, and other processes. A, Volcano plot for differential expression analysis based on samples from the GSE65682 dataset. B, Heatmap based on differential expression analysis of samples from the GSE65682 dataset, showing the expression of the top 20 upregulated and downregulated DEGs, respectively. C, The integration analysis of DEGs with key modular genes based on the GSE65682 dataset, identifying 969 DE-PRMT-RGs. D, PPI analysis of DE-PRMT-RGs using the STRING website. E, KEGG pathway analysis of these 969 DE-PRMT-RGs. F, GO functional enrichment analysis of these 969 DE-PRMT-RGs. DEGs, differentially expressed genes; DE-PRMT-RGs, differentially expressed PRMT-related genes; GO, Gene Ontology; KEGG, Kyoto Encyclopedia of Genes and Genomes; PPI, protein–protein interaction; STRING, search tool for the retrieval of interacting genes.

A total of 17 candidate genes were identified by MR analysis

We performed the MR analysis using OpenGWAS-id of 969 DE-PRMT-RGs as the exposure element, and sepsis as the result (Table S5, https://links.lww.com/SHK/C651). The findings demonstrated that 18 genes in all were causally linked to sepsis (P < 0.05). Of these, 10 genes (ELAC2, IL18R1, ITPA, PIK3AP1, EMB, CRLF3, GPR141, LONP1, PBX2, and TNFRSF25) were protective factors for sepsis (odds ratio <1) and 8 genes (ANAPC4, SDR39U1, ELP3, MCTP2, CD96, TUFM, FHIT, and SFT2D1) were risk factors for sepsis (odds ratio >1) (Fig. 3 and Table S6, https://links.lww.com/SHK/C651). In the scatter plots, the slopes of the IVW method lines were negative for the 10 genes and positive for the 8 genes (Figure S2, https://links.lww.com/SHK/C651). In the forest plots, the overall effect of the 10 genes corresponding to SNPs on sepsis was less than 0, while the overall effect of the 8 genes corresponding to SNPs on sepsis was greater than 0 (Figure S3, https://links.lww.com/SHK/C651). These results further suggested that the 10 genes were protective factors, whereas the 8 genes were risk factors for sepsis. The funnel plot showed a roughly symmetric distribution of SNPs from left to right, suggesting that MR followed Mendel’s second law of random grouping (Figure S4, https://links.lww.com/SHK/C651).

Fig. 3.

Univariate Cox regression analysis of genes causally associated with sepsis obtained from MR analysis was performed and displayed in a forest plot. MR, Mendelian randomization.

In sensitivity analysis, the Q–P values for the heterogeneity test and the P values for the horizontal pleiotropy assay were both larger than 0.05, suggesting that the MR analysis lacked both heterogeneity and horizontal pleiotropy (Tables S7 and S8, https://links.lww.com/SHK/C651). Furthermore, in the Leave-One-Out analyses, each SNP was gradually eliminated, and the remaining SNPs had little effect on outcome (Figure S5, https://links.lww.com/SHK/C651). Results from MRPRESSO showed that there were no outlier SNPs among the 18 exposure factors (Table S9, https://links.lww.com/SHK/C651).

These results further demonstrated the reliability of MR analysis. In Steiger analysis, except for FHIT, the causal relationship of the remaining 17 genes to sepsis was effective and not affected by reverse causality (Table S10, https://links.lww.com/SHK/C651). Therefore, we selected them as candidate genes for further analysis.

ELAC2, PBX2, MCTP2, and EMB were key genes for sepsis by machine learning algorithms

In LASSO regression analysis, 10 feature genes were screened when lambda was close to 0.0019, namely CRLF3, ELAC2, PBX2, CD96, TNFRSF25, ITPA, IL18R1, SFT2D1, MCTP2, and EMB (Fig. 4, A and B). In the SVM-RFE model, 10-fold cross-validation was used to screen out 6 feature genes (CRLF3, ELAC2, EMB, PBX2, LONP1, and MCTP2) (Fig. 4C), while in Boruta, 17 feature genes all passed the test (Fig. 4D). The ROC curves’ AUC values for the LASSO and Boruta models were 1 and 0.99, respectively, demonstrating that the two models performed exceptionally well (Figure S6, https://links.lww.com/SHK/C651). After intersecting the characteristic genes obtained from these three machine learning techniques, five candidate key genes (CRLF3, ELAC2, PBX2, MCTP2, and EMB) were identified (Fig. 4E). Among them, in GSE65682 and GSE134347, ELAC2 and PBX2 were significantly underexpressed in sepsis, whereas EMB and MCTP2 were upregulated (Fig. 4, F and G). Similarly, in the ROC curve, the AUC value of CRLF3 in GSE134347 was less than 0.6, while the AUC values of ELAC2, PBX2, MCTP2, and EMB in both datasets were greater than 0.7, so they were considered key genes in sepsis (Fig. 4, H and I).

Fig. 4.

Multiple machine learning screening of ELAC2, PBX2, MCTP2, and EMB as key genes of sepsis. A, Curve of MSE with Log(Lambda) in LASSO regression analysis. B, Curve of regression coefficients with Log(Lambda) in LASSO regression analysis. C, Plot of generalization error with the number of features obtained based on 10-fold cross-validation in SVM-RFE analysis. D, Plot of significance of variables versus genes in Boruta analysis, showing that all 17 feature genes passed the test. E, The integrated analysis of the above three machine learning predicted genes identified five sepsis-related key genes. F, Boxplot of key genes’ expression in the GSE65682 dataset. G, Boxplot of key genes’ expression in the GSE134347 dataset. H, ROC curves for key genes in the GSE65682 dataset. I, ROC curves for key genes in the GSE134347 dataset. LASSO, least absolute shrinkage and selection operator; MSE, mean squared error; ROC, receiver operating characteristic; SVM-RFE, support vector machine–recursive feature elimination.

Gene Set Enrichment Analysis showed that key genes were enriched in many different signaling pathways

GSEA enrichment analysis further explored the biological functions of key genes. First, based on the GO database, the functional entries that were significantly enriched for both ELAC2 and MCTP2 included cytoplasmic translation, noncoding RNA (NcRNA) metabolic process, and NcRNA processing (Fig. 5, A and B). EMB significantly enriched entries included the establishment of organelle localization, Golgi vesicle transport, etc. (Fig. 5C), while PBX2 played an important role in Chromatin Organization, Chromatin Binding, and so on (Fig. 5D). In KEGG enrichment analysis, both ELAC2 and MCTP2 were involved in the ribosome signaling pathway (Fig. 5, E and F). The signaling pathways significantly enriched for PBX2 included natural killer (NK) cell-mediated cytotoxicity, chemokine signaling pathway, etc. (Fig. 5G). Besides, the signaling pathway involved in EMB was pathogenic Escherichia coli infection, starch and sucrose metabolism, and so on (Fig. 5H).

Fig. 5.

GSEA revealed that key genes were enriched in many different signaling pathways. A, GO enrichment results of ELAC2 single-gene GSEA (top 10). B, GO enrichment results of MCTP2 single-gene GSEA (top 10). C, GO enrichment results of EMB single-gene GSEA (top 10). D, GO enrichment results of PBX2 single-gene GSEA (top 10). E, KEGG enrichment results of ELAC2 single-gene GSEA (top 10). F, KEGG enrichment results of MCTP2 single-gene GSEA (top 10). G, KEGG enrichment results of PBX2 single-gene GSEA (top 10). H, KEGG enrichment results of EMB single-gene GSEA (top 10). GSEA, Gene Set Enrichment Analysis; GO, Gene Ontology; KEGG, Kyoto Encyclopedia of Genes and Genomes.

Neutrophils, activated NK cells, and CD8 T cells were significantly associated with key genes

The infiltration abundance of 22 different types of immune cells in each sample was investigated based on the CIBERSORT algorithm (Fig. 6A), and the CD8 T cells and activated NK cells had the strongest correlation (cor = 0.64) (Fig. 6, B and C). Moreover, there were substantial differences between the sepsis and control groups in 17 immune cells (such as naive B cells, memory B cells, etc.) (Fig. 6D), among which neutrophils were significantly correlated with PBX2, while activated NK cells and CD8 T cells were significantly related to the remaining three key genes (ELAC2, MCTP2, and EMB) (Fig. 6E).

Fig. 6.

Key genes were significantly associated with neutrophils, activated NK cells, and CD8 T cells. A, Infiltration abundance of 22 immune cells in each sample according to the CIBERSORT algorithm. B, Correlation heatmap of immune cell ratios. C, Heatmap of correlation of P values of immune cell ratios. D, Differences in the percentage of immune cells between control and sepsis groups. E, Heatmap of the correlation between immune cells and key genes.

A total of 25 TFs, 7 miRNAs, 188 lncRNAs, and 70 sensitive drugs interacting with key genes were predicted

Based on miRWalk and starbase databases, we predicted 7 miRNAs and 188 lncRNAs of 3 key genes (MCTP2, ELAC2, and PBX2) to construct ceRNA networks, and the regulatory relationships they formed included NORAD-hsa-miR-345-3P-ELAC2, LINC01355-hsa-miR-493-3p-MCTP2, etc. (Fig. 7A). After that, 25 TFs of four key genes were also predicted, which contained ELAC2-E2F1, EMB-ARID3A, and so on (Fig. 7B). At last, a total of 70 drugs (e.g., D000082, D008727, etc.) were identified as sensitive drugs for key genes (Fig. 7C).

Fig. 7.

A total of 25 TFs, 7 miRNAs, 188 lncRNAs, and 70 sensitive drugs interacting with key genes were predicted. A, The mRNA–miRNA–lncRNA network map of key genes. Red circles signify mRNAs, green triangles signify lncRNAs, and blue diamonds signify miRNAs. B, The mRNA-TF Sankey diagram of key genes. The liners indicate pairs of relationships between key genes and TFs. C, Network diagram of key genes with predicted sensitive drugs. Red circles signify mRNAs, and green diamonds signify drugs. Red lines depict upregulation and blue lines depict downregulation. TFs, transcription factors; miRNAs, micro ribonucleic acids; lncRNAs, long-noncoding ribonucleic acids; mRNA, messenger ribonucleic acids.

Immune checkpoints analysis

A total of 25 immune checkpoints were differentially expressed between sepsis samples and control samples. Among them, CD44, ADORA2A, CD276, PDCD1, CD274, and PDCD1LG2 were upregulated in the sepsis group (Fig. 8), indicating that these genes might be a potential target for sepsis immunotherapy (P < 0.05).

Fig. 8.

A total of 25 immune checkpoints were differentially expressed.

Nomogram with excellent performance for diagnosis of sepsis

The nomogram was developed on the basis of ELAC2, EMB, MCTP2, and PBX2 (Fig. 9A). The probability of having sepsis increased with a higher score. The calibration curve of the nomogram showed a slope close to 1, which proved that the nomogram prediction accuracy was excellent (Fig. 9B). Besides, the ROC curve revealed an AUC value of 0.995 for the nomogram (Fig. 9C), and the decision curve revealed that the nomogram demonstrated a high clinical benefit rate (Fig. 9D). This nomogram, incorporating ELAC2, EMB, MCTP2, and PBX2, provided a reliable tool for predicting sepsis risk.

Fig. 9.

A nomogram was developed. A, The nomogram was developed on the basis of ELAC2, EMB, MCTP2, and PBX2. B, The calibration curve of the nomogram. C, The ROC curve revealed an AUC value of 0.995 for the nomogram. D, The decision curve of the nomogram. AUC, area under the ROC curve; ROC, receiver operating characteristic

Molecular docking of key genes in sepsis

For the key gene ELAC2, the top one interaction score drug was identified as baclofen, with a binding energy of −5.5 kcal/mol (Fig. 10A). For the key gene EMB, the top one drug was rifampicin, with a binding energy of −7.8 kcal/mol (Fig. 10B). For the key gene MCTP2, the top one interaction score drug was identified as emetine, with a binding energy of −6.3 kcal/mol (Fig. 10C). For the key gene PBX2, the top one drug was methazolamide, with a binding energy of −5.8 kcal/mol (Fig. 10D). These results indicated that these drugs may have potential applications in the management of disorders associated with the corresponding genes.

Fig. 10.

Molecular docking of key genes in sepsis. A, Molecular docking of ELAC2 and baclofen. B, Molecular docking of EMB and rifampicin. C, Molecular docking of MCTP2 and emetine. D, Molecular docking of PBX2 and methazolamide.

Confirmation of the four genes in the independent cohort with qRT-PCR

qRT-PCR was used to evaluate the four genes’ messenger RNA (mRNA) expression levels. The results showed that the sepsis group had greater levels of MCTP2 and EMB expression than the control group (Fig. 11, A and B), while the expression level of ELAC2 and PBX2 was lower (Fig. 11, C and D). The results showed that the four genes were considered key genes related to PRMT in sepsis, which provided a theoretical basis for the study of sepsis.

Fig. 11.

Validation of the expression of four genes in an independent cohort using qRT-PCR. A, The mRNA expression of MCTP2 in five pairs of frozen whole blood samples. B, The mRNA expression of EMB in five pairs of frozen whole blood samples. C, The mRNA expression of ELAC2 in five pairs of frozen whole blood samples. D, The mRNA expression of PBX2 in five pairs of frozen whole blood samples. mRNA, messenger ribonucleic acids; qRT-PCR, quantitative reverse-transcription polymerase chain reaction.

DISCUSSION

Sepsis and septic shock represent an increasingly severe and damaging systemic inflammatory response to infection, posing substantial clinical challenges in critical care medicine (44,45). Recent advances in data-driven methodologies, including machine learning and multiomics integration, have accelerated the discovery of sepsis-specific molecular signatures and biomarkers (46–48). Arginine methylation, a crucial post-translational modification mediated by PRMTs, is implicated in diverse diseases and biological processes (49–51). Its dysregulation offers potential diagnostic and therapeutic clues for sepsis, though the precise molecular mechanisms involved remain incompletely elucidated (15).

Our integrated transcriptome and MR analysis identified four PRMT-RGs—ELAC2, PBX2, MCTP2, and EMB—as critical regulators in sepsis pathogenesis. ELAC2 functions in tRNA processing and maturation, microtubule interaction, and transforming growth factor-β (TGF-β) signaling pathway modification (52,53). Notably, the TGF-β pathway not only possesses significant immunoregulatory functions, but recent studies also suggest its potential as a therapeutic target in sepsis patients (54). This indicates ELAC2 may engage in the immune dysregulation process of septic shock via this pathway. PBX2 belongs to the PBX transcription factor family. In the context of septic shock, it may influence multiple organ dysfunction by modulating immune and inflammatory responses. MCTP2, a calcium signaling regulator, may promote organ injury through calcium dysregulation and stress responses (55,56). EMB, with immune functional redundancy, aligns with late-stage immune paralysis (57). Collectively, these genes may form a regulatory network affecting immune function, metabolism, and cellular stress, driving the progression of septic shock and providing novel mechanistic and therapeutic insights.

Additionally, the present study revealed that these genes may converge on sepsis pathogenesis through immune-mediated dysregulation. MCTP2 and ELAC2 co-regulate the ribosome signaling pathway, which was recently shown to modulate T-cell survival through NUFIP1-mediated ribophagy during sepsis (58). PBX2 significantly enhanced the NK cell-mediated cytotoxicity and the chemokine signaling pathway, which is highly similar to Mac-1-dependent neutrophil aggregation in septic pulmonary microcirculatory failure (59). EMB is closely linked to pathogen-sensing and immunometabolic pathways (starch/sucrose metabolism), paralleling the inhibition of NF-κB-driven inflammation by agmatine through ribosomal S6 kinase regulation (60). These findings indicate that the four PRMT-RGs may influence sepsis progression through aberrant immunological mechanisms induced by immune cells. Targeting these genes may recalibrate immune homeostasis and attenuate multiorgan dysfunction.

Further ceRNA network analysis identified miR-345-3p and miR-493-3p as upstream regulators of MCTP2 and ELAC2, respectively. Notably, miR-345-3p has been shown to drive macrophage polarization (M1 to M2) via MAP3K1/NF-κB suppression (61) and to attenuate cellular stress by inhibiting BAK1-mediated apoptosis (62), suggesting its dual capacity to regulate inflammatory and metabolic homeostasis. Similarly, miR-493-3p counteracts fibrotic remodeling by mitochondrial stabilization (63) and suppresses oncogenic Wnt/β-catenin signaling (64). These pathways represent core drivers of sepsis-induced multiorgan dysfunction. Taken together, their upstream positioning in PRMT-RG networks, coupled with demonstrated therapeutic efficacy in modulating analogous pathways (e.g., NF-κB inhibition, mitochondrial protection), positions these miRNAs as master regulators of sepsis pathophysiology and promising candidates for targeted intervention.

Here, we for the first time identified four important genes linked to arginine methylation in sepsis (ELAC2, PBX2, MCTP2, and EMB), which showed stable expression patterns and an AUC >0.7, underscoring their biomarker potential for sepsis. Mechanistically, these genes appear to exert their effects by regulating immune infiltration and signaling pathways. Additionally, ceRNA network construction, transcription factor prediction, and molecular docking analysis further elucidate their upstream regulatory mechanisms and identify potential drug targets.

However, this study has several limitations. First, the analysis primarily relied on public data, making it difficult to reveal the dynamic expression of these genes during sepsis progression. Meanwhile, the heterogeneity in data sources, coupled with the absence of certain clinical covariates (such as age, sex, and Sequential Organ Failure Assessment scores), may compromise the generalizability and reproducibility of the findings. Second, the small sample size for qRT-PCR validation (only five pairs) reduces statistical power, and the use of bulk peripheral blood transcriptomes precludes cell-type-specific or tissue-specific mechanistic insights. Finally, functional validation of predicted networks and drug candidates was not performed. Therefore, future studies will expand multicenter cohorts, incorporate longitudinal sampling, and apply single-cell sequencing and animal models to clarify spatiotemporal gene dynamics and molecular mechanisms. Advancing preclinical evaluation of targeted drugs will provide a foundation for precise classification and personalized treatment of sepsis.

CONCLUSION

Our study pioneers the integration of transcriptome, MR analysis, and experimental validation to identify ELAC2, PBX2, MCTP2, and EMB as core arginine methylation regulators in sepsis, establishing the first mechanistically defined PRMT-RGs. This work addresses critical gaps in understanding how protein arginine methylation coordinates immune-metabolic dysregulation during sepsis progression. Taken together, these findings highlight the potential of PRMT-RGs characterization to enhance clinical diagnosis and management of sepsis. However, it should be noted that the current findings are predominantly based on bioinformatics analyses and require further experimental validation. We intend to conduct more in-depth investigations into these aspects in subsequent studies.

Acknowledgments

We thank GEO, GWAS, STRING site, MSigDB, miRWalk website, Starbase website, NetworkAnalyst website, and Comparative Toxicogenomics Database for providing data. We would also like to express our sincere gratitude to the Startup Fund for Scientific Research of Fujian Medical University (Grant No. 2022QH1061).