Mechanistic Insights into Maxing Shigan Decoction in Community-Acquired Pneumonia Based on Bioinformatics and Experimental Validation

Abstract

Purpose

To explore the mechanisms of Maxing Shigan Decoction (MXSGD) in community-acquired pneumonia (CAP) via network pharmacology, machine learning, molecular docking, and experimental validation.

Methods

MXSGD active compounds and targets were collected from TCMSP and TCM-ID; CAP-related genes were obtained from GEO datasets, and integrated with targets from GeneCards, OMIM, DisGeNET, and TTD. Overlapping targets were subjected to PPI, GO, and KEGG analyses. LASSO and SVM-RFE were used to select feature genes, and six core targets with robust diagnostic performance (AUC > 0.8 in both training and validation sets) were identified after ROC evaluation. Immune infiltration, GSEA, molecular docking, and 100 ns molecular dynamics simulations were performed. In vivo, the efficacy of MXSGD was assessed in a clinically relevant influenza A virus–Streptococcus pneumoniae co-infection mouse model by measuring lung index, histopathology, serum IL-6 and TNF-α levels by ELISA, and mRNA expression by qRT-PCR.

Results

We obtained 139 active compounds, 867 MXSGD targets, and 4,444 CAP genes, yielding 262 shared targets enriched in immune-inflammation pathways. Six core targets with high diagnostic accuracy were identified: ABCB1 (AUC = 0.9072), KIT (AUC = 0.8072), PLA2G7 (AUC = 0.8391), SLC2A3 (AUC = 0.9372), PRF1 (AUC = 0.8082), and GPR18 (AUC = 0.8414). Immune infiltration revealed eight differentially abundant immune cell types. Molecular docking and molecular dynamics supported the stable binding of glabridin to KIT and SLC2A3. In vivo, MXSGD significantly ameliorated lung pathological injury, reduced lung index and serum IL-6 and TNF-α levels (FDR-adjusted P < 0.001), and partially reversed mRNA expression changes of KIT, GPR18, and SLC2A3 (FDR-adjusted P < 0.05), while changes in ABCB1, PLA2G7, and PRF1 showed a trend without statistical significance.

Conclusion

These results suggest that MXSGD may exert immunomodulatory and anti-inflammatory effects in CAP, potentially by regulating ABCB1, KIT, PLA2G7, SLC2A3, PRF1, and GPR18; however, only three of these (KIT, GPR18, and SLC2A3) showed statistically significant mRNA changes. These findings provide a multi-component, multi-target rationale requiring further experimental and clinical confirmation.

Introduction

Community-acquired pneumonia (CAP) is a pulmonary parenchymal infectious disease acquired outside of hospitals, commonly presenting with symptoms such as fever, cough, sputum production, and chest pain.¹ CAP exhibits high global incidence and mortality rates, making it a leading cause of hospitalization and imposing significant economic burdens on individuals and societies. According to the Global Burden of Disease Study (GBD) 2021, in 2021 there were approximately 344 million incident cases of lower respiratory infections (predominantly pneumonia) worldwide, corresponding to an incidence rate of 4,350 per 100,000 population, and resulting in 2.18 million deaths.² Beyond mortality, CAP generates substantial public health costs through prolonged hospital stays, intensive care unit admissions, and lost productivity, with a disproportionate economic impact on older adults, immunocompromised individuals, and resource-limited settings.³

Pathogenically, CAP is caused by a broad range of pathogens, including bacteria such as Streptococcus pneumoniae (Spn), Haemophilus influenzae, and Staphylococcus aureus, as well as respiratory viruses such as influenza A virus (IAV) and respiratory syncytial virus.⁴ Traditional research has predominantly focused on single-pathogen infections. However, clinically, a substantial proportion of severe CAP cases, particularly during epidemic seasons, are closely associated with virus-bacteria co-infections. Studies indicate that over 30% of seasonal and pandemic IAV infections involve co-infections.^5,6 Epidemiological studies confirm that influenza virus infection increases the risk of secondary bacterial pneumonia by nearly 100-fold, with Spn being the most commonly isolated strain. This co-infection is clinically relevant because IAV damages the respiratory epithelium, impairs mucociliary clearance, and dysregulates innate immune responses, thereby facilitating Spn colonization, invasion, and dissemination to the lower respiratory tract; co-infected patients also tend to have more severe disease, with higher rates of intensive care admission and mortality. A synergistic pathogenic mechanism exists between the two pathogens, rather than a simple additive effect.⁷

The cornerstone of effective CAP management continues to be antibiotic therapy tailored to the specific causative pathogen.⁸ However, due to the lack of rapid diagnostic methods for identifying pathogens, CAP treatment relies heavily on empirical antibiotic use.⁹ This empiric strategy, however, frequently results in patient adverse effects and drives the development of antibiotic-resistant strains, thereby contributing to major challenges in the clinical management of CAP.^10,11 Immune responses participate in pathogen or toxin recognition and clearance to prevent excessive tissue damage. Immune cells, epithelial cells, and stromal cells interact to jointly initiate and coordinate inflammatory responses.^12,13 Therefore, exploring adjunctive therapeutic strategies that modulate host immune and inflammatory responses while synergizing with antibiotics is crucial for enhancing efficacy or reducing antibiotic dependence.

Extensive clinical research confirms that TCM can serve as an effective adjunct therapy for community-acquired pneumonia, reducing antibiotic usage while enhancing treatment effectiveness.^14–16 Maxing Shigan Decoction (MXSGD) is a principal formula derived from traditional Chinese medicine for pneumonia treatment. Its composition contains four herbs: Ephedra sinica (Mahuang), Semen armeniacae amarum (Kuxingren), Gypsum Fibrosum (Shigao), and Glycyrrhiza uralensis (Gancao). MXSGD is considered suitable for CAP because of its significant antiviral effects, its ability to improve microvascular permeability, and its anti-inflammatory and immunomodulatory activities.^17–19 Moreover, clinical studies have shown that MXSGD combined with antibiotics effectively alleviates clinical symptoms in CAP patients, improves pulmonary infection status, and enhances clinical recovery rates without increasing adverse reaction risks.^20,21 Since 1949, numerous TCM formulas applied during four major Chinese epidemics—including COVID-19—have been developed based on MXSGD’s formulation framework. These include Qingfei Paidu Decoction, Qingfei Huatan Decoction, and Lianhua Qingwen Capsules, whose efficacy and safety have been validated by multiple clinical studies.^22,23

Although the role of MXSGD in the treatment of CAP has been confirmed, its core therapeutic targets and regulatory mechanisms remain unclear. Previous studies have mainly adopted single-pathogen infection models, whereas severe CAP often involves viral–bacterial co-infections. Moreover, prior research has largely focused on clinical efficacy observations or specific signaling pathways, and the multi-component, multi-target nature of MXSGD remains poorly understood.

Unlike previous pharmacological studies that have focused on single pathogens or limited targets, this study integrated a suite of bioinformatics approaches, including network pharmacology, machine learning (LASSO and SVM-RFE), immune infiltration analysis, molecular docking, and molecular dynamics simulations, to systematically predict and validate the key targets of MXSGD in the treatment of CAP. More importantly, we employed a clinically relevant IAV–Spn co-infection mouse model for in vivo validation, which better recapitulates the pathogenesis of severe CAP. This multi-level research strategy, spanning computational prediction, simulation-based validation, and experimental verification, aims to provide robust mechanistic evidence for MXSGD in the treatment of CAP.

Materials and Methods

Bioinformatics Prediction Analysis

Screening of MXSGD Active Components and Targets and Network Construction

The four herbal compounds of MXSGD were retrieved from the TCMSP (https://www.tcmsp-e.com/index.php, accessed January 2026) and TCMID (https://www.bidd.group/TCMID/, accessed January 2026) databases. To capture any potentially overlooked constituents, supplementary searches were conducted in the PubMed, CNKI, and Wanfang databases. Active ingredients of MXSGD compounds were screened based on oral bioavailability (OB) and drug-likeness (DL) parameters. Specifically, compounds with OB ≥ 30% and DL ≥ 0.18 were considered as candidate active ingredients.²⁴ The chemical structures of the identified compounds were obtained from the PubChem database (https://www.ncbi.nlm.nih.gov/, accessed January 2026). These structural data were subsequently submitted to the SwissTargetPrediction platform (http://www.swisstargetprediction.ch/, accessed January 2026) for the prediction of potential protein targets. The resulting target gene names were standardized using the UniProt database (https://www.uniprot.org/, accessed January 2026), yielding a refined list of targets limited to the species “Homo sapiens”. The workflow of the entire study is shown in Figure 1.

Figure 1.

Flowchart of the entire study.

Disease Targets Associated with CAP

To identify potential targets associated with community-acquired pneumonia (CAP), we first searched the OMIM, DisGeNET, GeneCards, and TTD databases using the keyword “Community-acquired pneumonia” and retained genes from GeneCards with a relevance score > 10.²⁵ All database searches for CAP-related targets were performed on January 2026. Two datasets, GSE103119 (GPL10558) and GSE196399 (GPL24676), were downloaded from the GEO database.^26,27 GSE103119 was used as the training set, comprising gene expression data from 20 healthy controls and 152 CAP samples. GSE196399 was used as the validation set, comprising data from 21 healthy controls and 56 CAP samples. The normalized expression matrices were directly obtained from GEO; the raw fluorescence intensity data had undergone background correction and quantile normalization in Illumina GenomeStudio software.²⁸ Moreover, all samples in the datasets were processed in a single experimental batch, eliminating the need to merge multiple batches. Therefore, no additional batch effect correction was required in the present study. Differential expression analysis was performed on the normalized expression matrix of GSE103119 using the limma package in R.²⁹ The resulting P-values were adjusted using the Benjamini–Hochberg method to control the false discovery rate (FDR). Differentially expressed genes (DEGs) were identified with the criteria of |log₂ fold change (FC)| > 1 and FDR < 0.05. Subsequently, the targets derived from the database search and the DEGs were merged and deduplicated to serve as the CAP disease targets.

PPI Network Construction

The intersection between the drug target genes and disease-associated genes was computed using Venny 2.1.0, and the resulting gene set was defined as the potential therapeutic targets of MXSGD for CAP. To elucidate the functional relationships among these candidate targets, a protein-protein interaction (PPI) network was generated employing the STRING database. The organism was set to “Homo sapiens”, the minimum required interaction score was set to 0.40 (medium confidence). The architecture of this PPI network was ultimately visualized and subjected to topological analysis using Cytoscape 3.9.1.

GO and KEGG Pathway Enrichment Analysis

GO analysis was performed to systematically evaluate the functional attributes of the targets, covering biological processes (BP), cellular components (CC), and molecular functions (MF). Concurrently, KEGG enrichment analysis was conducted to identify potential pathways linked to the shared targets of MXSGD and CAP. The set of intersecting genes was submitted to the DAVID database for both GO and KEGG enrichment analyses. Graphical representation of the acquired data was achieved employing bar and bubble plots on the bioinformatics platform.

Machine Learning (ML) Analysis Based on Intersecting Genes

The intersecting gene set was subjected to further screening via two distinct machine learning methodologies: Least Absolute Shrinkage and Selection Operator (LASSO) and Support Vector Machine-Recursive Feature Elimination (SVM-RFE). The objective was to discern feature genes that differentiate CAP patients from healthy controls. In the LASSO regression, performed with the “glmnet” package, the optimal penalty parameter (lambda) was identified through five-fold cross-validation, and candidate genes were selected based on their non-zero coefficients. For SVM-RFE, executed using the “e1071” and “caret” packages, the model was tuned via five-fold cross-validation, and the core gene set was defined by the feature number associated with the minimal error rate. The subsequent intersection of the candidate genes from both algorithms pinpointed the key targets for MXSGD in treating CAP, a result illustrated in a Venn diagram via Venny2.1.0. The discriminatory power and expression levels of these key targets were first assessed in the training set using receiver operating characteristic (ROC) curves and box plots, and subsequently externally validated in the independent dataset GSE196399, all generated in GraphPad Prism 10.

Immune Infiltration Analysis

The CIBERSORT deconvolution algorithm was applied to the GSE103119 dataset to characterize the relative proportions of 22 immune cell subsets. This approach was taken to assess whether the key targets in CAP patients were associated with specific modifications in the composition of the immune microenvironment. Samples with a deconvolution output P value < 0.05 were retained for comparison of the relative abundances of 22 immune cell types between the healthy control group and the CAP group. Simultaneously, Pearson correlation analysis was employed to dissect interactions among different immune cell types. Furthermore, by integrating Pearson correlation analysis with Mantel tests, we investigated potential associations between core gene expression and the 22 immune cell types from both linear and matrix perspectives, providing a more comprehensive and in-depth theoretical basis for CAP immune mechanism research.

GSEA Analysis

GSEA was implemented utilizing the “clusterProfiler” package in R. Pathways with a Normalized Enrichment Score (NES) absolute value greater than 1, an adjusted p-value (FDR) below 0.25, and a nominal p-value less than 0.05 were deemed statistically significant. The subsequent data visualization was carried out in R (v4.1.2).

Determining the Relationship Between TFs and miRNAs with Key Targets

Using the NetworkAnalyst platform, we screened for upstream regulators of our key target genes by querying two integrated databases: the JASPAR database was used to identify significant TFs, while Tarbase9.0 was utilized to extract experimentally validated miRNAs. The resulting TF-gene and miRNA-gene interaction networks were then mapped and visualized in Cytoscape.

Molecular Docking

Protein structures were acquired from the RCSB PDB (https://www.rcsb.org) and AlphaFold databases (https://www.rcsb.orghttps://alphafold.com/). Ligand two-dimensional structures, retrieved from TCMSP, were converted into three-dimensional conformations via ChemDraw. Prior to docking, the protein structures were preprocessed using MGLTools. The molecular docking procedure was executed in AutoDock Vina to predict the binding pose between the core active ingredient (ligand) and the core protein target (receptor). Final visualization and interaction profiling were conducted with Discovery Studio.

To validate the reliability and rationality of the docking parameters and protocol, a re-docking procedure was carried out. The co-crystallized ligand was extracted from the protein crystal structure and then re-docked into the predefined active site of the target protein. The root-mean-square deviation (RMSD) between the re-docked pose and the original ligand conformation (crystal structure) was calculated. An RMSD value of less than 2.0 Å was considered acceptable, indicating that the established docking protocol and parameters possess good reliability and can accurately reproduce the protein–ligand binding mode.

Molecular Dynamics Simulation

All molecular dynamics (MD) simulations were performed with GROMACS 2022. The protein and the ligand were described by the CHARMM36 and CGenFF force fields, respectively. Each system was solvated in a TIP3P water box, neutralized with counterions, and subjected to energy minimization. Subsequently, three independent 100 ns MD simulations were performed for each protein–ligand complex, each initiated with different random initial velocities. All simulations were carried out in the NPT ensemble at 310 K with a 1 fs time step, and long-range electrostatic interactions were treated using the particle mesh Ewald (PME) method. Key trajectory parameters, including root-mean-square deviation (RMSD), radius of gyration (Rg), solvent-accessible surface area (SASA), hydrogen bonds, and root-mean-square fluctuation (RMSF), were analyzed using the built-in modules of GROMACS, and the free energy landscape (FEL) was constructed. The mean ± standard deviation of each parameter over the equilibrated phase of the triplicate simulations was calculated for subsequent statistical analysis.

In vivo Experiment Validation

Experimental Animals

Eighteen 6–8 week-old SPF-grade Balb/c mice (female: male = 1:1, 20±2g) were provided by Liaoning Changsheng Biotechnology Co., Ltd. [Certificate No.: SCXK(Liao)2020–0001] and housed at the Animal Experiment Center of Liaoning University of Traditional Chinese Medicine [Certificate No.: SYXK(Liao)2024–0006]. Under strictly controlled and stabilized conditions (23 ± 2°C, 50 ± 5% humidity, 12-h light/dark cycle), all animals were maintained with a standard diet and water available ad libitum for the study’s duration. All animal experimental procedures were conducted in accordance with the Guide for the Care and Use of Laboratory Animals, and this manuscript was prepared in compliance with the ARRIVE guidelines,³⁰ and the completed ARRIVE checklist is provided as Supplementary Material. The Animal Ethics Committee of Liaoning University of Traditional Chinese Medicine approved all procedures (Approval No.: 21,000,042,023,152).

Bacterial and Viral Strains

The influenza virus mouse lung-adapted strain A/FM1/47 (H1N1) was provided by the Virology Laboratory of the Affiliated Hospital of Liaoning University of Traditional Chinese Medicine. The Streptococcus pneumoniae colonizing type 19F strain (ATCC 49619) was provided by the American Type Culture Collection (ATCC) and stored frozen in the Virology Laboratory of the Affiliated Hospital of Liaoning University of Traditional Chinese Medicine.

Preparation of MXSGD

MXSGD consists of Ephedra sinica 15g, Semen armeniacae amarum 15g, Gypsum Fibrosum 30g, and Glycyrrhiza uralensis 10g. This composition and dosage ratio are derived from the classic Chinese medical text Shanghan Lun (Treatise on Cold Pathogenic Diseases) and have been widely adopted in modern clinical and experimental studies.³¹ All herbs were purchased from the Outpatient Department of Liaoning University of Traditional Chinese Medicine Affiliated Hospital (Shenyang, China). Add 3000 mL of distilled water (10 times the total weight of herbs) to the herbs and soak for 30 minutes. Bring to a boil and decoct for 30 minutes. After decoction, filter the first decoction of MXSGD. For the second decoction, add 2100mL distilled water to the residue (7 times the residue volume). Bring to a boil and simmer for 20 minutes. After decoction, filter the second decoction of MXSGD. Finally, combine both decoctions of MXSGD and concentrate to a final concentration of 1 g crude drug/mL.³²

Animal Model Establishment and Intervention

Mice were randomly assigned to three groups: normal control (n = 6), model (n = 6), and MXSGD treatment (n = 6). Randomization was performed by an independent statistician not involved in the subsequent experiment, using a computer-generated random number sequence (SPSS). Specifically, all animals were first sorted by body weight and given sequential identification numbers. A random number sequence was then generated using SPSS, and the mice were allocated to groups according to the rank of these random numbers. The allocation sequence was sealed in sequentially numbered, opaque envelopes to ensure allocation concealment, and the envelopes were opened only by a designated researcher at the time of nasal inoculation. Because the herbal decoction had a distinctive color and odor, blinding of the personnel performing gavage was not feasible; however, all outcome assessments were conducted under strictly blinded conditions.

The model and MXSGD treatment groups received intranasal inoculation of 15LD50 H1N1 influenza virus at a total volume of 50μL. Forty-eight hours later, mice in these groups received intranasal inoculation of 10⁵ CFU Spn mouse-adapted strain in 50μL. Normal control mice received 100μL saline solution per mouse following the same procedure.³³ Twenty-four hours after intranasal inoculation, the MXSGD treatment group mice received MXSGD (1.80 g·kg⁻¹·d⁻¹/1 mL) via oral gavage, while normal control and model group mice received an equal volume of saline via oral gavage for 5 consecutive days; the doses of MXSGD were calculated as 9.1 times the human clinical dose (assuming a 70 kg body weight and converted by body surface area). Daily observations included recording body weight and general vital signs. On day 7, mice were anesthetized by intraperitoneal injection of 3% pentobarbital (150 mg/kg). Blood samples were collected from the retro-orbital venous sinus, after which the mice were euthanized by cervical dislocation. Total lung tissues were harvested and weighed. The blood samples were allowed to stand at room temperature for 30 min and then centrifuged (3,000 rpm, 15 min, 4°C) to obtain the supernatant serum. The serum was aliquoted and stored at −80°C until analysis. The left lung was fixed in 4% paraformaldehyde, followed by routine paraffin embedding and sectioning; the right lung was stored at −80°C until analysis.

Lung Index Assessment

The lung index was employed as a semi-quantitative indicator for evaluating pulmonary edema and inflammatory exudation, a method that has been validated in a mouse model of acute lung injury.³⁴ Immediately after euthanasia, intact lung tissues were carefully dissected, blotted dry with filter paper to remove surface blood and fluid, and weighed on an analytical balance. The lung index was calculated using the following formula: lung wet weight (g)/body weight (g) × 100%. An elevated lung index reflects more severe pulmonary congestion, edema, and inflammatory cell infiltration, and shows a good correlation with histopathological severity scores.

Hematoxylin and Eosin (H&E) Staining

Left lung tissues fixed in 4% paraformaldehyde were dehydrated through a graded ethanol series, embedded in paraffin, and cut into 5-μm-thick serial sections. After deparaffinization and rehydration, the sections were stained using a H&E Staining Kit (G1120, Beijing Solarbio Science and Technology Co., Ltd., China), followed by dehydration, clearing, and mounting with neutral balsam. Six representative pathological images from each group were examined under a light microscope (200× magnification) to assess the extent of pathological damage and inflammatory cell infiltration in the lung tissue.

The degree of lung injury was evaluated using a semi-quantitative scoring system based on four readily identifiable pathological processes: (1) alveolar congestion, (2) hemorrhage, (3) infiltration or aggregation of neutrophils in the airspace or the vessel wall, and (4) thickness of the alveolar wall (with or without hyaline membrane formation). Each parameter was graded on a five-point scale from 0 to 4 according to the extent of pathological involvement within each high-power field: 0, no detectable injury (0%); 1, minimal injury (<25% of the field); 2, mild injury (25%–50% of the field); 3, moderate injury (50%–75% of the field); and 4, severe/maximal injury (>75% of the field). The total lung injury score for each animal was calculated as the sum of the four individual parameter scores, yielding a possible range of 0 to 16.³⁵

Enzyme-Linked Immunosorbent Assay (ELISA)

Serum concentrations of IL-6 and TNF-α were determined using Mouse IL-6 ELISA Kit (HJ182, Shanghai Epizyme Biomedical Technology Co., Ltd., China) and Mouse TNF-α ELISA Kit (HJ207, Shanghai Epizyme Biomedical Technology Co., Ltd., China). Serum samples collected from each group of mice were assayed in strict accordance with the manufacturer’s instructions. Briefly, after the steps of coating, blocking, and sample addition, the plates were incubated with enzyme-conjugated detection antibodies, and color development was carried out by adding the chromogenic substrate. The absorbance was measured at 450 nm using an ELX-800 microplate reader. All standards and samples were assayed in duplicate. The concentrations of TNF-α and IL-6 were calculated from the respective standard curves, and the results were expressed in pg/mL.

Quantitative Reverse Transcription Polymerase Chain Reaction (qRT-PCR)

Total RNA was extracted from mouse right lung tissue homogenates using TRIzol reagent, and the RNA concentration was determined with a nucleic acid/protein analyzer. cDNA was subsequently synthesized according to the manufacturer’s instructions for the reverse transcription kit. qRT-PCR was performed using the 2× Universal Blue SYBR Green qPCR Master Mix (G3326, Servicebio Technology Co., Ltd., China), and the primer sequences are provided in Supplementary Table S1. The thermal cycling program consisted of an initial denaturation at 98°C for 2 min, followed by 30 cycles of denaturation at 98°C for 20s and annealing/extension at 55°C for 20s. All samples were run in triplicate. The relative expression levels of six core genes (ABCB1, KIT, PLA2G7, SLC2A3, PRF1, and GPR18) were calculated using the 2 ^−ΔΔCt method, with β-actin serving as the internal reference gene.

Statistical Analysis

Data meeting normal distribution requirements were expressed as mean ± standard deviation (SD). Data analysis was performed using SPSS 27.0 software. One-way ANOVA was performed to assess overall differences among groups. When ANOVA was significant, pairwise comparisons were conducted using Fisher’s LSD test. Raw P values from all pairwise tests were adjusted for multiple testing using the Benjamini-Hochberg FDR procedure. Results with FDR-adjusted P < 0.05 were considered statistically significant.

To evaluate the statistical power for the primary efficacy endpoint of this experiment, a post hoc power analysis was performed using G*Power 3.1 software. Lung index, lung histopathological score, and two cytokines were treated as the primary outcomes, and overall differences among groups were assessed by one-way ANOVA. The effect size, Cohen’s f, was calculated from the actual mean, standard deviation, and sample size of each group. The F-test was employed, with α set at 0.05, and the number of groups and sample size were entered to compute the achieved statistical power (1 - β).

Results

Bioinformatics Prediction Analysis Results

Construction of MXSGD-Active Ingredient-Target Network

First, all components of the four herbs included in MXSGD were retrieved from TCMSP and TCMID. Active ingredients of MXSGD compounds were screened based on OB and DL parameters. Additionally, databases such as PubMed, CNKI, and Wanfang were utilized to supplement other omitted components. As a result, we obtained a total of 139 active components: 23 from Ephedra, 19 from bitter almond, 92 from licorice, and 5 from gypsum (Supplementary Table S2). The structures of these active components were imported into the SwissTargetPrediction database to facilitate further target identification. After removing duplicates, a total of 867 MXSGD-related targets were generated, and gene names were added for all targets (Supplementary Table S3). Additionally, we constructed an MXSGD-active ingredient-target network to provide a systematic and holistic view (Figure 2A). Network nodes were color-rendered based on their degree values, with darker colors and larger node sizes indicating a greater degree of freedom and potential biological functions.

Figure 2.

Network pharmacology analysis of MXSGD against CAP. (A) MXSGD’s active components-targets network. (B) Heatmap of the top 60 DEGs in gene expression profiles between the normal group and CAP group. (C) Venn Diagram of MXSGD Drug Targets and CAP Disease Target Genes. (D) Presents the PPI network based on intersecting targets and the top 20 intersecting target networks selected by degree value. (E) GO analysis of the top 20 cross-targets ranked by degree value. (F) KEGG pathway enrichment analysis of the top 20 cross-targets ranked by degree value.

Identification of CAP-Related Target Genes

The GSE103119 dataset from the GEO database was selected as the training set. This dataset comprises 172 samples, of which 20 are normal samples and 152 are community-acquired pneumonia (CAP) samples. By comparing gene expression profiles between CAP and normal samples, we identified 412 DEGs, including 87 genes upregulated in CAP and 325 downregulated (Supplementary Figure  S1, Supplementary Table S4). A heatmap was generated to visualize the expression patterns of these DEGs (Figure 2B). Additionally, we integrated disease targets from GeneCards, OMIM, DisGenet, and TTD databases—718 from GeneCards, 3679 from OMIM, 30 from DisGenet, and 1 from TTD (Supplementary Table S5). After merging and deduplicating these disease targets with the GEO DEGs, we ultimately identified 4444 CAP disease targets.

Construction of Protein-Protein Interaction (PPI) Networks Based on Overlapping Targets in MXSGD and CAP

We identified 262 intersections between MXSGD drug targets and CAP-associated genes. These overlapping genes were identified as potential targets for MXSGD treatment of CAP and visualized using a Venn diagram (Figure 2C), where MXSGD targets are represented in blue and CAP targets in orange. These overlapping targets were then uploaded to the STRING database to construct a PPI network (Figure 2D). This network comprised 259 nodes (3 genes excluded due to out-of-range values), forming 4,782 edges. Nodes were color-rendered based on Degree values: darker shades and larger node sizes indicate higher degree values, signifying greater biological functionality potential. By screening Degree values, the top 20 cross-targets were identified and a target network was constructed (Figure 2D). This diagram clearly demonstrates that targets such as IL6, TNF, AKT1, ALB, and EGFR play crucial roles within the PPI network.

GO and KEGG Pathway Enrichment Analysis

To further investigate the molecular mechanisms underlying MXSGD treatment for CAP, we conducted comprehensive GO and KEGG enrichment analyses on the top 20 cross-targets ranked by degree. Results were visualized using bar charts and Sankey bubble diagrams, highlighting the top 30 most significant pathways, each with an adjusted p-value below 0.05. The GO enrichment analysis yielded three distinct categories: BP, MF, and CC. Specifically, 259 BP terms, 65 MF terms, and 33 CC terms were enriched. Similarly, the KEGG pathway enrichment analysis identified 133 signaling pathways.

BP terms were primarily associated with inflammatory response, cellular response to viruses, miRNA transcriptional regulation, apoptosis processes, and positive regulation of cytokine production involved in inflammatory response. Additionally, they participated in the epidermal growth factor receptor signaling pathway and the interleukin-6-mediated signaling pathway. CC results indicated associations with protein complexes, RNA polymerase II transcription regulatory complexes, and transcription regulatory complexes, while MF analysis highlighted roles in transcription coactivator binding, protein kinase activity, and protein tyrosine kinase activity (Figure 2E, Supplementary Table S6). KEGG pathway analysis showed significant enrichment in pathways including HIF-1 signaling, C-type lectin receptor signaling, TNF signaling, and IL-17 signaling (Figure 2F). These GO and KEGG pathway enrichment analyses collectively underscore the impact of these cross-targets on immune regulation and inflammatory responses.

Screening of Core Genes Involved in MXSGD Treatment of CAP

To identify the core genes involved in MXSGD treatment of CAP, we employed two machine learning algorithms, LASSO regression and SVM-RFE, to select from the candidate intersection genes using GSE103119 as the training set. LASSO analysis identified 15 feature genes: GPR84, ABCB1, EPHX2, TGFBR2, MMP8, TXK, PLA2G7, KIT, ECE1, GPR18, CASK, INSR, SLC2A3, ARG1, and PRF1 (Figures 3A and B). The SVM-RFE algorithm selected 22 feature genes, including ABCB1, NR3C1, KIT, TRPA1, DYRK1B, AURKA, HIF1A, MAPK14, GPR18, SLC2A3, TLR9, NFKB1, EP300, CXCR4, PLA2G7, CASR, CCR3, PGK1, PRF1, CASK, TGFBR2, and RAC1 (Figure 3C). Intersection analysis of the results derived from the two machine learning methods yielded eight overlapping genes: ABCB1, TGFBR2, PLA2G7, KIT, GPR18, CASK, SLC2A3, and PRF1 (Figure 3D).

Figure 3.

Machine learning screening and validation of core genes in CAP. (A) Coefficient plot generated by LASSO analysis; (B) Regularization plot generated by LASSO analysis. The red curve shows the cross‑validated binomial deviance as a function of log(λ). The optimal λ is selected at the point where this red curve reaches its minimum value. (C) Error rate curve obtained through five-fold cross-validation using the SVM-RFE algorithm. The blue dot represents the best performance (lowest RMSE) under SVM-RFE, corresponding to the 15 feature genes. (D) Intersection analysis of core genes screened by the two ML algorithms. (E) ROC curve analysis of the eight overlapping genes based on the GSE103119 dataset. (F) ROC curve analysis of the six core genes based on the GSE196399 dataset. The expression levels of hub genes in GSE103119 (G) and GSE196399 (H) datasets. **** P <0.0001.

Assessment of the Diagnostic Value of Core Genes in CAP

To evaluate the diagnostic value of the eight overlapping genes for CAP, we analyzed their performance using ROC curves in the training set GSE103119 and the external validation set GSE196399. With an area under the curve (AUC) greater than 0.8 adopted as the criterion for high diagnostic efficacy, CASK and TGFBR2 in the training set GSE103119 yielded AUC values of 0.773 and 0.577, respectively, both below 0.8, suggesting unstable diagnostic performance; thus, they were excluded. After removal of CASK and TGFBR2, the remaining six genes (ABCB1, PLA2G7, KIT, GPR18, SLC2A3, and PRF1) all achieved AUC values exceeding 0.8 in both the training and validation sets, demonstrating robust diagnostic efficacy and promising predictive potential for CAP progression (Figures 3E and F). Consequently, these six genes were identified as the core genes involved in MXSGD treatment of CAP. We further analyzed the expression differences of the six core genes between healthy controls and CAP patients. The results revealed consistent expression trends for the six core genes across the two datasets; compared with the healthy control group, the expression level of SLC2A3 was significantly upregulated in the CAP group, whereas the expression levels of ABCB1, PLA2G7, KIT, GPR18, and PRF1 were all significantly downregulated (Figures 3G and H).

Immune Cell Infiltration Analysis

To investigate differences in immune cell infiltration between CAP and healthy control samples, we performed infiltration analysis on GSE103119 data using the “CIBERSORT” package in R. This revealed the distribution of 22 immune cell types across samples (Figure 4A, Supplementary Table S7). Subsequently, box plots were used to compare immune cell differences between CAP and healthy control samples (Figure 4B). Significant differences were observed in eight immune cell types between CAP and healthy control samples. Compared to healthy controls, CAP samples exhibited reduced estimated proportions of T cells CD8, T cells CD4 memory resting, T cells CD4 memory activated, NK cells resting, and Dendritic cells resting, suggesting a relative reduction in these adaptive and antigen-presenting cell populations, which may reflect a shift in the immune landscape. Conversely, Monocytes, Macrophages M0, and Neutrophils showed higher estimated proportions in CAP samples, consistent with a pronounced innate immune signature commonly observed in acute infectious conditions.

Figure 4.

Immune microenvironment, functional pathways, and transcriptional regulatory network analysis of core genes in CAP. (A) Stacked bar charts of different immune cells in healthy control and CAP samples from the GSE103119 dataset. (B) Box plots showing differences in immune cell proportions between healthy control and CAP samples. (C) Heatmap showing Pearson correlations between core gene expression and immune cell types. (D) Mantel correlation heatmap between core genes and immune cell subsets globally. (E) GO enrichment analysis associated with core genes. (F) Signaling pathways of core genes identified by gene set enrichment analysis. (G) Interaction network between TFs and core genes. (H) Interaction network between miRNAs and core genes. *P < 0.05; **P < 0.01; ***P < 0.001; ****P < 0.0001; ns, not significant.

Correlation Analysis Between Core Genes and Infiltrating Immune Cells

We analyzed and visualized Pearson correlation heatmaps (Figure 4C) to demonstrate the expression correlations of six core genes across different immune cell types. T cells CD8, T cells CD4 memory resting, NK cells resting, and T cells CD4 memory activated exhibited significant associations with all core genes. Furthermore, PLA2G7, PRF1, GPR18, KIT, and ABCB1 showed positive correlations with numerous adaptive immune cell types, including T cells CD4 memory resting, T cells CD8, and NK cells resting. Conversely, these genes exhibited negative correlations with innate immune cells such as Monocytes, Macrophages M0, and Neutrophils. In stark contrast, SLC2A3 exhibited the opposite pattern, showing positive correlations with myeloid-derived cells and negative correlations with lymphocyte subsets.

To explore the significant correlation patterns between the six core genes and distinct immune cell subsets, we employed Mantel tests and Pearson correlation analyses, revealing that these associations form a correlation network within the immune microenvironment (Figure 4D, Supplementary Table S8). The results indicate that PLA2G7, PRF1, and ABCB1 show significant positive correlations with adaptive immune cells, such as T cells CD8 and T cells CD4 memory resting, while negatively correlating with myeloid cells and neutrophils; SLC2A3 exhibited the opposite pattern, positively correlated with myeloid cells and negatively correlated with lymphoid cells. Notably, GPR18 displayed distinct dual correlation patterns, being positively correlated with M1 macrophages and activated dendritic cells, yet negatively correlated with M2 macrophages and resting mast cells. These differentiated correlation patterns point to distinct association profiles for each gene: PLA2G7, PRF1, and ABCB1 were positively correlated with lymphocyte subsets, indicating potential links with adaptive immune cell proportions; SLC2A3 showed positive correlations with myeloid cells, suggesting an association with innate immune infiltration; and GPR18’s dual correlations indicate context-dependent associations with different immune cell types. All these observations are correlative in nature and should be interpreted as hypothesis-generating for future functional studies.

GSEA Analysis of Core Genes

To further elucidate the biological functions and potential pathways of the six core genes, we performed GSEA to identify biological functions or regulatory pathways significantly associated with core gene expression. Based on GSEA results (Figures 4E and F), in terms of cellular components, the six core genes were primarily associated with mitochondrial inner membrane protein complex, mitochondrial protein-containing complex, and respiratory chain complex. Regarding molecular functions, they were mainly linked to growth factor activity, oxygen binding, oxygen carrier activity, lysozyme activity, and oxidoreduction-driven active transmembrane transporter activity. In biological processes, they were primarily associated with immune response regulation, including regulation of inflammatory response, cellular response to molecule of bacterial origin, activation of innate immune response, negative regulation of viral entry into host cells, regulation of cytokine production, cellular response to lipopolysaccharide, ribonucleoprotein complex biogenesis, phagocytosis, and neutrophil activation involved in immune response. Additionally, in terms of enriched biological signaling pathways, reduced expression levels of ABCB1, KIT, PLA2G7, PRF1, and GPR18 are primarily associated with Staphylococcus aureus infection, neutrophil extracellular trap formation, leukocyte migration through endothelial cells, cAMP signaling pathways, TNF signaling pathways, chemokine signaling pathways, Fc gamma R-mediated phagocytosis, C-type lectin receptor signaling pathway, bacterial invasion of epithelial cells, and the PPAR signaling pathway. Elevated expression levels of core genes are primarily associated with ribosomes and the COVID-19 signaling pathway, whereas SLC2A3 exhibits expression patterns opposite to those of the aforementioned enriched genes. The GSEA analysis results indicate that the six core genes exert immune regulatory effects through multiple cellular components, biological processes, and signaling pathways.

Transcriptional Regulation Mechanisms of Core Genes

To identify potential transcriptional regulatory mechanisms for the six core genes, we identified 26TFs using the JASPAR database and constructed an interaction network between these TFs and the core genes (Figure 4G). Among these TFs, 10 (FOXC1, MEF2A, POU2F2, HINFP, GATA2, FOXL1, STAT1, CREB1, TEAD1, TFAP2C) had a Degree value of 2 or higher. Additionally, using the TarBase database, we predicted and mapped the interaction network between miRNAs and core genes (Figure 4H). We predicted a total of 262 potential miRNAs and identified 99 miRNAs with a Degree value ≥ 2. These findings indicate that all six core genes are regulated by the same transcription factors and other shared regulatory processes.

Molecular Docking Results

To investigate potential binding interactions between core targets and the top bioactive compounds, we performed molecular docking. Six core targets were selected as receptors. The top six compounds ranked by Degree value in the “MXSGD - Active Ingredient - Target” network analysis (quercetin, kaempferol, glabridin, naringenin, licochalcone B, and mairin) were chosen as ligands. Ultimately, 36 receptor-ligand pairs were subjected to docking, and binding energy heatmaps were generated (Figure 5A). The grid box parameters are detailed in Supplementary Table S9. Generally, a binding energy lower than −5.0 kcal/mol is considered to indicate a certain predicted binding affinity, while a value lower than −7.0 kcal/mol is considered to indicate high predicted binding affinity; lower binding energies correlate with more stable interactions in the docking simulations. The docking results showed that the binding energies of all core target–compound complexes were below −5 kcal/mol, and approximately 70% of these complexes exhibited binding energies below −7 kcal/mol. The six pairs with the lowest binding energies were selected for molecular docking model construction (Figure 5B–G). The two combinations with the lowest binding energies were KIT-Glabridin (−10.0 kcal/mol) and SLC2A3-Glabridin (−9.8 kcal/mol).

Figure 5.

Molecular docking of core targets with active ingredients and MD stability assessment of their complexes. (A) Heatmap of docking binding energies. (B) KIT-Glabridin docking simulation. (C) SLC2A3-Glabridin docking simulation. (D) ABCB1-Quercetin docking simulation. (E) ABCB1-Glabridin docking simulation. (F) KIT-Quercetin docking simulation. (G) KIT-Kaempferol docking simulation. (H) RMSD values of the protein-ligand complex over time. (I) Rg values of the protein-ligand complex over time. (J) SASA values of the protein-ligand complex over time. (K) Time-dependent HBonds values of the protein-ligand complex. (L) RMSF values of the KIT-Glabridin complex. (M) RMSF values of the SLC2A3-Glabridin complex. (N) Gibbs free energy landscape of the KIT-Glabridin complex. (O) Gibbs free energy landscape of the SLC2A3-Glabridin complex.

In the redocking validation of co-crystallized protein–ligand complexes, two core proteins, GPR18 and PRF1, were not subjected to conventional re-docking evaluation because experimentally determined ligand-bound conformations are unavailable for RMSD calculation. For the remaining four core proteins, the ligand RMSD values for the ABCB1, KIT, PLA2G7, and SLC2A3 systems were 1.391 Å, 0.150 Å, 1.910 Å, and 0.626 Å, respectively (Supplementary Figure  S2). All RMSD values were below 2.0 Å, indicating that the docking protocol can reliably reproduce experimental ligand binding poses and is suitable for the present virtual screening. These computational results only predict binding potential and do not directly confirm target activation or inhibition in a biological context.

Molecular Dynamics Simulation Results

We performed three independent 100 ns molecular dynamics simulations for each of the two complexes exhibiting the best binding activity, KIT-Glabridin and SLC2A3-Glabridin. RMSD directly reflects the overall conformational stability of the system. In all three replicates, both complex systems reached dynamic equilibrium after 60 ns of simulation. The average RMSD of the SLC2A3-Glabridin complex over the last 40 ns was 0.1704 ± 0.01075 nm, whereas that of the KIT-Glabridin complex was 0.2558 ± 0.0416 nm. The SLC2A3-Glabridin complex exhibited a lower baseline RMSD and a narrower fluctuation range (P < 0.05). Nonetheless, the trajectories of both complexes remained stable without large drift, confirming that Glabridin can stably bind to both target proteins (Figure 5H). Rg characterizes the overall compactness of the protein and reflects spatial conformational changes during the simulation. After reaching equilibrium, the mean Rg of SLC2A3-Glabridin was 2.2665 ± 0.0067 nm, and that of KIT-Glabridin was 1.9667 ± 0.0183 nm. The Rg curves of both complexes remained relatively flat throughout the simulations, with no obvious expansion or contraction (Figure 5I). The KIT protein itself adopts a more compact spatial structure, resulting in an overall lower Rg and greater structural compactness upon ligand binding, whereas SLC2A3 displays a more extended conformation and a relatively higher Rg (P < 0.05). SASA reflects the solvent exposure of the protein surface and the degree of binding pocket occlusion. After equilibration, the mean SASA values of SLC2A3-Glabridin and KIT-Glabridin were 195.5362 ± 3.28922 nm² and 151.7007 ± 3.15246 nm², respectively. The SASA of both complexes showed no obvious increasing or decreasing trend across the triplicate runs and remained within a stable range, indicating that the insertion of Glabridin into the active pockets of the two proteins did not alter the hydrophilic/hydrophobic distribution on the protein surface (Figure 5J). Compared with the SLC2A3-Glabridin complex, the KIT-Glabridin complex exhibited a lower SASA (P < 0.05). Hydrogen bond analysis revealed that the number of hydrogen bonds between Glabridin and each complex ranged from 0 to 4, with approximately 2 hydrogen bonds stably maintained over most of the simulation time, indicating that Glabridin forms persistent and effective hydrogen bond interactions with polar amino acid residues of both target proteins (Figure 5K). RMSF analysis assessed residue-level flexibility. The mean RMSF of the KIT-Glabridin complex was 0.1451 ± 0.1059 nm, whereas that of the SLC2A3-Glabridin complex was 0.1068 ± 0.0698 nm. For both complexes, the RMSF values of the vast majority of residues remained below 0.3 nm, indicating overall low flexibility and high structural rigidity, with only minor motions observed in terminal flexible loop regions (Figures 5L and M). Residues in the functional domains of SLC2A3 exhibited smaller fluctuations and greater rigidity, whereas residues in local active regions of KIT showed slight flexible fluctuations, reflecting a somewhat higher overall structural flexibility that may facilitate conformational adjustment to accommodate the ligand (P < 0.05). FEL analysis showed that both complexes exhibited distinct blue regions representing low free energy basins, from which the thermodynamically most stable conformations could be identified (Figures 5N and O). For SLC2A3-Glabridin, the low-energy conformations were more narrowly distributed with greater conformational uniformity, suggesting low conformational heterogeneity and a propensity to adopt the energetically most favorable state. In contrast, KIT-Glabridin displayed a somewhat wider distribution of stable conformations, indicating better conformational adaptability. Collectively, comparison of these dynamic parameters demonstrates that Glabridin can form stable and reliable complexes with both SLC2A3 and KIT, with favorable overall binding. The SLC2A3-Glabridin complex exhibits greater overall structural rigidity, higher conformational uniformity, and superior stability.

In vivo Experimental Verification Results

MXSGD Attenuates Body Weight Loss and Pulmonary Edema in Mice with IAV–Spn Coinfection

The workflow of the entire in vivo experimental validation design is illustrated in Figure 6A. Following coinfection, mice in the model group exhibited progressive body weight loss over time (Figure 6B). At the experimental endpoint, the percentage of initial body weight in the model group was significantly lower than that in the control group (control: 136.22 ± 4.86%; model: 81.97 ± 5.57%; FDR-adjusted P < 0.001). The reduction in body weight was markedly attenuated in the MXSGD-treated group at the endpoint (MXSGD: 92.56 ± 1.44% vs model; FDR-adjusted P < 0.001). In addition, The lung index, an indirect indicator of pulmonary edema severity, was significantly elevated in the model group compared with the control group (model: 2.02 ± 0.14%; control: 0.83 ± 0.06%; FDR-adjusted P < 0.001), and was significantly decreased in the MXSGD group relative to the model group (MXSGD: 1.45 ± 0.20%; FDR-adjusted P < 0.001; Figure 6C). Consistent with these findings, ex vivo gross observation of the lungs revealed pronounced congestion and diffuse edema in the model group, whereas the extent of pulmonary edema and hemorrhage appeared reduced in MXSGD-treated mice (Figure 6D).

Figure 6.

In vivo protective effects of MXSGD and core gene expression validation in co-infected mice. (A) In vivo Experimental Validation Design. (B) Changes in body weight among mouse groups. (C) Lung index of mice in each group. (D) Gross structure of lungs. (E) Histopathological examination of lungs after HE staining (200× magnification). (F) Pathological scores of mice in each group. (G) Serum concentrations of IL-6 in each group of mice. (H) Serum concentrations of TNF-α in each group of mice. (I) Expression of six core genes in lung tissues of mice from each experimental group. * FDR-adjusted P < 0.05; ** FDR-adjusted P < 0.01; *** FDR-adjusted P < 0.001; ns, not significant.

MXSGD Ameliorates Lung Histopathological Injury Induced by IAV–Spn Coinfection

Examination of H&E-stained lung tissue sections revealed severe pathological changes in the model group, including marked thickening of bronchiolar and vascular walls, massive infiltration of inflammatory cells into the alveolar spaces, peribronchiolar and perivascular areas, and pronounced thickening of the alveolar septa (Figure 6E). These alterations were substantially ameliorated in the MXSGD treatment group. Semi-quantitative histopathological scoring confirmed that the lung injury score in the model group was significantly higher than that in the control group (model: 13.00 ± 1.41; control: 0.83 ± 0.75; FDR-adjusted P < 0.001). Compared with the model group, the histopathological score was significantly reduced in the MXSGD group (MXSGD: 9.00 ± 0.89; FDR-adjusted P < 0.001; Figure 6F).

MXSGD Reduces Serum Concentrations of IL-6 and TNF-α in Mice with IAV–Spn Coinfection

To further evaluate systemic inflammation, serum concentrations of interleukin-6 (IL-6) and tumor necrosis factor-α (TNF-α) were measured using ELISA. The levels of both cytokines were significantly increased in the model group compared with the control group (IL-6: model, 122.37 ± 7.01 pg/mL vs control, 54.07 ± 3.84 pg/mL; TNF-α: model, 158.07 ± 6.10 pg/mL vs control, 98.29 ± 6.74 pg/mL; FDR-adjusted P < 0.001 for both). Compared with the model group, serum IL-6 and TNF-α concentrations were significantly decreased in the MXSGD treatment group (IL-6: MXSGD, 102.45 ± 5.55 pg/mL; TNF-α: MXSGD, 137.74 ± 6.19 pg/mL; FDR-adjusted P < 0.001 for both; Figures 6G and H).

Validation of Core Gene Expression in Mouse Lung Tissue

To validate transcriptional changes in core genes identified in previous analyses, the mRNA levels of ABCB1, KIT, PLA2G7, PRF1, GPR18, and SLC2A3 in lung tissue were determined by qRT-PCR. The expression of ABCB1, KIT, PLA2G7, PRF1, and GPR18 was significantly downregulated in the model group compared with the control group (FDR-adjusted P < 0.05 for each gene), whereas SLC2A3 expression was significantly upregulated (FDR-adjusted P < 0.05). Following MXSGD intervention, the expression levels of KIT and GPR18 were significantly upregulated compared with the model group (KIT: MXSGD, 0.27 ± 0.05 vs model, 0.11 ± 0.09; GPR18: MXSGD, 0.59 ± 0.18 vs model, 0.34 ± 0.02; FDR-adjusted P < 0.05 for both), while SLC2A3 expression was significantly downregulated (SLC2A3: MXSGD, 1.69 ± 0.50 vs model, 3.43 ± 1.41; FDR-adjusted P < 0.05). After FDR correction, only a trend without statistical significance was observed for the expression differences in ABCB1, PLA2G7, and PRF1 between the model and MXSGD groups (Figure 6I). Taken together, these in vivo data provide partial experimental validation of the computationally predicted core gene signatures, with significant changes observed for KIT, GPR18, and SLC2A3, whereas ABCB1, PLA2G7, and PRF1 showed only non-significant trends.

Post Hoc Statistical Power

Post hoc power analysis was conducted for the primary efficacy endpoints—lung index, lung histopathological score, and serum IL-6 and TNF-α concentrations—using the actual means, standard deviations, and sample sizes of each group. The achieved statistical power (1 − β) for all primary outcome measures exceeded 0.80 (lung index: Cohen’s f = 3.53, power > 0.999; histopathological score: Cohen’s f = 4.78, power > 0.999; serum IL-6: Cohen’s f = 5.11, power > 0.999; serum TNF-α: Cohen’s f = 3.91, power > 0.999). These results confirm that the sample size of the present study was sufficient to detect the observed effect sizes for these key indicators.

Discussion

CAP is a major global health problem and the third leading cause of death worldwide after cardiovascular and cerebrovascular diseases, accounting for approximately 3 million deaths annually.³⁶ This burden is particularly pronounced in aging populations and in regions with limited access to healthcare services. As antimicrobial resistance continues to undermine the efficacy of conventional antibiotics, there is an increasing need for adjunctive immunomodulatory or host-directed strategies.³⁷ As a traditional Chinese medicine formula, MXSGD may represent a relatively accessible and low-cost complementary option for CAP management. Furthermore, the core targets identified in this study may provide a basis for the identification of prognostic biomarkers for CAP and for the development of individualized treatment strategies.

Previous studies indicate that MXSGD possesses multiple pharmacological activities. Both MXSGD and oseltamivir, used alone or in combination, can shorten the duration of fever in patients infected with the H1N1 influenza virus.³⁸ Furthermore, MXSGD exerts anti-Streptococcus pneumoniae effects by effectively inhibiting pneumococcal hemolysin.³⁹ Previous UPLC-Q-TOF/MS studies of MXSGD identified key active components, including ephedrine, pseudoephedrine, quercetin, naringin, amygdalin, kaempferol, and glycyrrhizic acid. These compounds effectively dilate bronchial smooth muscle, mitigate tissue hypoxia damage, and suppress excessive T-lymphocyte activation and proliferation.⁴⁰ In this study, network analysis of the Chinese herbal medicine targets for MXSGD’s active components revealed that quercetin, kaempferol, glabridin, naringenin, licochalcone B, and mairin interact with multiple targets within the network. Among these, glabridin and licochalcone B are noted for their anti-inflammatory and antioxidant properties.^41,42 Quercetin and kaempferol have been reported to exhibit antiviral and lung-protective effects.^43,44 Mairin, a steroidal saponin, has been shown to mitigate acute lung injury by regulating the MAPK/NF-κB signaling pathway and NLRP3 inflammasome.⁴⁵ Molecular docking and dynamics simulations further suggested that glabridin forms stable complexes with KIT and SLC2A3, indicating a potential contribution to the therapeutic effects of MXSGD, although these computational predictions require experimental confirmation.

Immune responses are crucial for the initiation, amplification, and persistence of CAP. The loss of immune tolerance leads to increased secretion of multiple proinflammatory mediators by various cell types, thereby triggering inflammation.⁴⁶ In this study, we constructed a PPI network using MXSGD-related drug targets and disease targets as cross-targets, followed by GO and KEGG analyses. Results revealed that the functions of these cross-targets were primarily associated with immune regulation, suggesting their potential role in modulating CAP progression. Furthermore, a comprehensive analysis integrating the GSE103119 training set and the GSE196399 validation set using machine learning techniques identified six core genes: ABCB1, KIT, PLA2G7, SLC2A3, PRF1, and GPR18.

At the levels of immune initiation and cell formation, KIT, as a receptor for stem cell factor (SCF), functions as an upstream regulator. Through its signaling pathway, it drives hematopoietic stem cell differentiation and mast cell activation, providing the cellular foundation for innate and adaptive immune responses.⁴⁷ During early infection, the KIT/SCF signaling pathway can be upregulated to release large numbers of immune cells. However, when inflammation becomes excessive, sustained activation of the KIT pathway may lead to overproliferation and activation of myeloid cells, exacerbating the cytokine storm and causing more severe tissue damage. Regarding energy metabolism in immune regulation, SLC2A3—a high-affinity glucose transporter primarily expressed in immune cells like NK cells, macrophages, and neutrophils—significantly upregulates its encoded GLUT3 transporter. This rapidly meets the heightened energy demands of activated immune cells, providing essential metabolic support for immune regulation.⁴⁸ During the immune effector phase, PRF1, a key cytotoxic effector molecule, is expressed in activated cytotoxic T lymphocytes (CTLs) and NK cells, directly mediating target cell apoptosis. This constitutes a core mechanism for eliminating pathogens or abnormal cells.⁴⁹ PLA2G7, a vital lipid hydrolase, whose functional deficiency leads to abnormal accumulation of the proinflammatory mediator platelet-activating factor (PAF) and oxidized phospholipids. This, in turn, exacerbates vascular permeability and neutrophil infiltration by activating endothelial and immune cells. Concurrently, the vicious cycle of oxidative stress and disruption of lipid signaling balance during late-stage inflammation further impedes normal resolution of inflammation.⁵⁰ At the level of immune microenvironment regulation and homeostasis maintenance, GPR18 senses lipid-derived ligands (eg, Resolvin D2) to guide chemotactic migration of immune cells to inflammatory sites. In the late phase of the response, it participates in regulating inflammation resolution by influencing immune cell retention and clearance.⁵¹ ABCB1, as a multidrug resistance protein, maintains the internal environment stability of effector cells by effluxing immune metabolites, toxins, and drugs. This enhances their survival and functional persistence under stress conditions, thereby ensuring the effective execution of immune responses at the systemic level.⁵² In summary, these six core genes sequentially play pivotal roles across four hierarchical levels—immune initiation, energy metabolism, effector killing, and microenvironment regulation—forming an immunoregulatory network with clear hierarchical relationships. This network not only deepens our understanding of immunopathological mechanisms but also provides multi-level targets for CAP immunotherapy strategies.

For the elucidation of core gene-immune cell interactions, immune infiltration analysis was carried out. The results revealed that the six core genes exhibit distinct correlation patterns within the immune microenvironment. PLA2G7, PRF1, KIT, ABCB1, and GPR18 correlated positively with adaptive immune components such as T cells and NK cells, whereas SLC2A3 showed a positive correlation with myeloid cells like monocytes/macrophages and neutrophils. Notably, GPR18 displayed dual correlations, being positively associated with M1 macrophages and activated dendritic cells, and negatively associated with M2 macrophages, suggesting context-dependent regulatory roles. Thus, immune infiltration analysis reveals a multi-gene synergistically regulated immune cell interaction network, providing a systems-level basis for elucidating the immunopathological mechanisms underlying MXSGD’s intervention in CAP. However, CIBERSORT is constrained by its dependence on the predefined LM22 signature matrix, which impedes the accurate discrimination of phenotypically similar or tissue-specific immune subsets. Moreover, this method exhibits substantial estimation errors for low-abundance cells, potentially yielding false-negative results that necessitate further validation through subsequent experiments.

Additionally, we performed GSEA enrichment and transcriptional analysis to further elucidate the intricate relationship between core genes and immune-inflammatory regulatory mechanisms. GSEA results revealed that, in terms of cellular components, these genes showed a strong association with the mitochondrial inner membrane and respiratory chain complexes, further supporting the pivotal role of immune metabolic reprogramming in the development of CAP. Functionally, these genes were significantly enriched in key biological processes, including inflammatory response regulation, innate immune activation, cytokine production, and neutrophil-mediated immune responses. At the molecular mechanism level, they were closely associated with multiple immune-related pathways such as Staphylococcus aureus infection, neutrophil extracellular trap formation, leukocyte migration, TNF signaling, chemokine signaling, and PPAR signaling. These findings indicate that the six core genes not only extensively participate in activating and regulating innate and adaptive immune responses but also interact through multiple signaling pathways, collectively forming a multi-layered immune regulatory system.

TFs and miRNAs serve as primary regulators of gene expression, generating feedback regulation in biological processes to jointly determine cell fate and govern life processes. Our analysis identified FOXC1 and hsa-miR-124-3p as potential key regulators in the network governing core gene expression. FOXC1 may influence macrophage and neutrophil polarization and recruitment by modulating downstream TGF-β and NF-κB signaling pathways, thereby regulating pulmonary inflammatory intensity.⁵³ Furthermore, FOXC1 expression in alveolar epithelial cells could contribute to barrier function maintenance and repair by influencing epithelial-mesenchymal transition (EMT) and cell adhesion molecule synthesis.⁵⁴ In pulmonary microvascular endothelial cells, FOXC1 dysfunction may induce increased vascular permeability and abnormal inflammatory exudation, participating in acute lung injury development.⁵⁵ hsa-miR-124-3p has been reported to negatively regulate excessive immune responses by inhibiting the TLR/NF-κB signaling pathway and promoting macrophage M2 polarization, thereby mitigating lung tissue injury and promoting repair.⁵⁶

Molecular docking results indicated that the binding energies for all ligand-receptor combinations were below −5.0 kcal/mol, and approximately 70% of the complexes achieved binding energies below −7.0 kcal/mol, suggesting favorable and, in many cases, high predicted binding affinity between key active components and core targets. Redocking validation of the co-crystallized complexes yielded ligand RMSD values below 2.0 Å for the four evaluable targets, confirming the reliability of the docking protocol employed in this study. The strongest binding energies were observed for KIT-Glabridin and SLC2A3-Glabridin. To ensure the robustness of these findings, three independent 100 ns molecular dynamics simulations were performed for each complex. Both complexes maintained stability across all replicates, with the SLC2A3-Glabridin complex exhibiting consistently lower RMSD, narrower fluctuation ranges, higher conformational uniformity, and superior overall stability compared to KIT-Glabridin. These computational findings support the possibility that glabridin may contribute to the therapeutic effects of MXSGD on CAP through binding to SLC2A3 and KIT. However, these computational predictions reflect thermodynamic and structural plausibility, not biological efficacy. Whether glabridin acts as an agonist, antagonist, or allosteric modulator of KIT and SLC2A3 remains unknown. Direct experimental validation is lacking; future studies employing binding assays (eg, SPR or MST) and cellular functional assays are needed to determine glabridin’s pharmacological mode of action and translate these computational hypotheses into testable biological conclusions.

Using both the training dataset GSE103119 and the independent external validation dataset GSE196399, we evaluated the expression levels of the six core genes. In both datasets, SLC2A3 was consistently upregulated in CAP patients, while ABCB1, PLA2G7, KIT, GPR18, and PRF1 were consistently downregulated, demonstrating the robustness of these gene signatures across independent cohorts. To further validate the expression of these six key genes in CAP, we established a mouse model of IAV–Spn co-infection pneumonia closely related to the clinical phenotype of human CAP. Compared to normal controls, this model exhibited significant weight loss in mice, along with marked pulmonary congestion and edema. Additionally, this model induced typical pathological features of pneumonic injury. Detection of the expression levels of the six core genes in lung tissue from this constructed model revealed that, compared to the Control group, SLC2A3 was upregulated and ABCB1, PLA2G7, KIT, GPR18, and PRF1 were downregulated in the Model group, consistent with the bioinformatics analysis results.

To clarify MXSGD’s actual efficacy against CAP in physiological systems and its impact on core gene expression levels, intervention with MXSGD significantly reduced weight loss compared to the Model group. Pulmonary congestion and edema were alleviated, with marked improvement in lung pathology. Furthermore, MXSGD treatment significantly reduced serum concentrations of the pro-inflammatory cytokines IL-6 and TNF-α compared to the Model group, indicating an attenuation of systemic inflammation. Simultaneously, detection of expression levels for the six core genes in lung tissue revealed that, compared to the Model group, KIT and GPR18 expression levels were significantly upregulated in the MXSGD group, while SLC2A3 expression levels were significantly downregulated. Changes in ABCB1, PLA2G7, and PRF1 did not reach statistical significance after correction, suggesting partial modulation. These findings collectively suggest that MXSGD mitigates pneumonia-induced injury during CAP progression, and this protective effect was accompanied by changes in the expression of a subset of core genes. However, given that only three of the six genes were significantly regulated after FDR correction, these data provide partial transcriptional support rather than definitive mechanistic proof.

However, the present study has several limitations. First, the sample size of the animal experiments was relatively small and only a single dose of MXSGD was administered; therefore, the dose–response relationship and the reproducibility of the observed effects remain unclear. Second, an independent survival observation cohort was not established; all animals were euthanized at predetermined endpoints, and survival curve data are therefore lacking. Third, pathogen burden was not assessed, precluding determination of whether the beneficial effects of MXSGD are attributable to direct pathogen clearance, immunomodulation, or both. Fourth, in the transcriptional validation of the six core genes, after FDR correction for multiple comparisons, only KIT, GPR18, and SLC2A3 showed statistically significant changes in response to MXSGD intervention, whereas ABCB1, PLA2G7, and PRF1 showed only nonsignificant trends. Therefore, mechanistic conclusions should be limited to the significantly regulated genes, and the roles of ABCB1, PLA2G7, and PRF1 remain suggestive. This may be attributable to the relatively small sample size, the low expression abundance of some genes, and the fact that measurements were performed at only a single time point. Fifth, core target expression was assessed only at the mRNA level, and protein-level validation is lacking. In addition, functional experiments have not yet been conducted to confirm the roles of glabridin or the six candidate targets in CAP. Direct binding assays, pharmacological inhibition or overexpression/knockdown studies, and cell-based functional experiments are required to establish causality. Therefore, future studies should incorporate independent survival experiments, larger sample sizes, multiple MXSGD doses, serial pathogen burden measurements, dynamic assessments at multiple time points, and protein-level and functional validation to more comprehensively evaluate the pharmacodynamic effects and molecular mechanisms of MXSGD. Furthermore, pharmacokinetic–pharmacodynamic studies of MXSGD components, particularly glabridin, are needed to assess in vivo target engagement. Finally, well-designed prospective clinical studies are still required to determine whether these preclinical findings can be translated into public health benefits.

Conclusion

In conclusion, this study explored the mechanisms of MXSGD in CAP through an integrated strategy combining network pharmacology, machine learning, molecular docking, and experimental validation, thereby providing a multi-level predictive framework supported by preliminary experimental validation. Six core targets (ABCB1, PLA2G7, KIT, GPR18, SLC2A3, and PRF1) were identified with robust diagnostic performance. Immune infiltration analysis and gene set enrichment analysis revealed significant associations between these genes and specific immune cell subsets as well as inflammatory pathways. Molecular docking and dynamics simulations provided computational evidence for stable binding between the core targets and key active components of MXSGD. In the influenza A virus–Streptococcus pneumoniae sequential co-infection mouse model, MXSGD significantly attenuated pulmonary injury and systemic inflammation. Among the six core targets, only KIT, GPR18, and SLC2A3 exhibited significant transcriptional changes after MXSGD treatment, whereas ABCB1, PLA2G7, and PRF1 showed only non-significant trends. These findings offer a predictive foundation for potential therapeutic targets in CAP, pending rigorous clinical validation.

Funding Statement

This research was supported by the National Natural Science Foundation of China (NO.8197152087) and the Liaoning Provincial Department of Education Project (NO. 2022SK 2014).

Data Sharing Statement

The data analyzed in this study are publicly available from the GEO database. Specifically, the GSE103119 dataset was generated by Wallihan et al. and is cited as reference [26]; the GSE196399 dataset was generated by Zhang et al. and is cited as reference [27].

The data analyzed in this study are included in the websites mentioned above. The original contributions from this study can be found in the article and the Supplementary Material section. For further information or inquiries, please contact the corresponding author.

Ethics Approval and Consent to Participate

All animal experimental procedures were conducted in strict accordance with the Guide for the Care and Use of Laboratory Animals and were approved by the Animal Ethics Committee of Liaoning University of Traditional Chinese Medicine (Approval No.: 21000042023152). This manuscript was prepared in compliance with the ARRIVE guidelines for the reporting of in vivo experiments.

The GEO is a public database, and the patients included in this database have provided ethical approval for their data usage. Users can freely download relevant data for research purposes and publish related articles. Consequently, the Human Research Ethics Committee at Liaoning University of Traditional Chinese Medicine has reviewed and waived the requirement for ethical approval for this study.

Author Contributions

All authors made a significant contribution to the work reported, whether that is in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work.

Disclosure

The authors report no conflicts of interest in this work.