Integrative single-cell and bulk transcriptomic analyses identify IGFBP2 within a peri-anesthetic stress-related gene framework linked to injury/transitional remodeling of alveolar type II cells in idiopathic pulmonary fibrosis

Abstract

Background

In idiopathic pulmonary fibrosis (IPF), damage to alveolar type II (AT2) cells and their aberrant reprogramming are regarded as key events in disease progression. Apart from anesthetic exposure, the peri-anesthetic period is also characterized by additional stressors, such as mechanical ventilation-induced stretch. Here, peri-anesthetic stress-related genes (PSRGs) were treated as a literature-derived, biologically relevant prior gene framework to prioritize candidate genes associated with AT2 state remodeling in IPF.

Methods

Single-cell and bulk transcriptomic datasets from IPF and control lungs were analyzed. PSRGs were intersected with concordant differentially expressed genes to identify candidate hub genes, which were further prioritized using an integrated machine-learning framework. IGFBP2 expression in AT2-related states was analyzed in GSE128033 and validated in GSE135893, with complementary pseudo-bulk, correlation, covariate, CellChat, scTenifoldKnk, and DrugCLIP analyses.

Results

Two AT2 states, mature and injury/transitional, were identified. IGFBP2 was prioritized for state-focused analysis and showed reproducible enrichment in injury/transitional AT2 cells in both GSE128033 and GSE135893. IGFBP2 expression also increased along pseudotime, whereas COL1A1 did not show a clear trajectory-related pattern. Cell-cell communication analysis suggested prominent microenvironmental interactions involving injury/transitional AT2 cells. Computational perturbation analysis further implicated epithelial secretory and mucosal defense-related programs within the IGFBP2-associated network.

Conclusions

IGFBP2 was identified as a reproducibly enriched candidate in injury/transitional AT2 cells across discovery and validation cohorts. These findings suggest that IGFBP2 may be associated with AT2 state remodeling and altered epithelial microenvironmental communication in IPF, although functional and protein level validation remains necessary.

Highlight box.

Key findings

• IGFBP2 was prioritized within a peri-anesthetic stress-related gene (PSRG) framework and was linked to injury/transitional remodeling of alveolar type II (AT2) cells in idiopathic pulmonary fibrosis (IPF).

• IGFBP2 was enriched in injury/transitional AT2 cells, increased along pseudotime, and was associated with inferred microenvironmental interaction patterns and epithelial secretory/barrier-related programs.

What is known and what is new?

• Aberrant AT2 reprogramming is a central event in IPF progression, and peri-anesthetic stress may be relevant to epithelial injury in susceptible lungs.

• This study integrates bulk and single-cell transcriptomic analyses with a literature-derived PSRG framework used for candidate prioritization and identifies IGFBP2 as a lead candidate associated with injury/transitional AT2 remodeling in IPF.

What is the implication, and what should change now?

• These findings support further mechanistic investigation of IGFBP2 in AT2 injury-associated remodeling and fibrotic microenvironmental change in IPF.

• Future work should validate the role of IGFBP2 in independent cohorts and experimental systems, particularly in relation to epithelial state transition, protein-level localization, and peri-anesthetic stress-associated biological contexts.

Introduction

Idiopathic pulmonary fibrosis (IPF) is a progressive form of interstitial lung disease characterized by irreversible disruption of alveolar architecture, excessive extracellular matrix (ECM) accumulation, and persistent loss of lung function (1). As a key stem cell population in the alveolar epithelium, alveolar type II (AT2) cells are responsible for epithelial regeneration and repair after injury. The abnormal reprogramming of these cells is regarded as a significant event in the pathological development of IPF (2-4). Clinical observations indicate that patients with IPF are at increased risk of postoperative acute exacerbation (AE-IPF) and pulmonary complications after surgery under general anesthesia (5). Stretch induced by mechanical ventilation, medications administered during the peri-anesthetic period, and associated tissue injury could be potential triggering factors (6,7). Since acute exacerbation is usually accompanied by significant epithelial injury and inflammatory dysregulation (8), peri-anesthetic stress may be involved in the disease process through effects on alveolar epithelial cell states.

Whether peri-anesthetic stress-related genes (PSRGs), treated here as a literature-derived, biologically relevant prior gene framework, can help identify candidate genes associated with AT2 state remodeling in IPF remains unclear. In this study, the PSRG framework was not intended to suggest that peri-anesthetic stress directly causes IPF. Instead, PSRGs were used as a stress response gene set to prioritize candidates that may be relevant to both acute perioperative stress and chronic epithelial remodeling in fibrotic lung tissue. This strategy differs from screening based solely on differential expression because it introduces prior biological information related to epithelial stress, mechanical stimulation, and injury-related signaling before integrating bulk and single-cell transcriptomic evidence.

In this research, the potential relevance of PSRGs to AT2 state remodeling in IPF was examined via an integrated multi-layer analytical strategy. Single-cell and bulk transcriptomic data were analyzed together with PSRGs derived from previous literature, including anesthetic-related drug targets and mechanical stimulation-related genes. Candidate hub genes were subsequently prioritized, and their relevance to cell-state transitions as well as epithelial-immune interactions was evaluated. Machine-learning, in silico virtual perturbation, and molecular interaction analyses were further incorporated to explore the potential regulatory and binding-related features (Figure 1). Collectively, these analyses facilitated a cross-layer evaluation of PSRG-associated candidates in relation to the injury/transitional AT2 state and provided a basis for subsequent mechanistic studies. We present this article in accordance with the TRIPOD reporting checklist (available at https://jtd.amegroups.com/article/view/10.21037/jtd-2026-0896/rc).

Figure 1.

Overall study workflow. AT2, alveolar type II.

Methods

Data sources and preprocessing

Bulk transcriptomic datasets of IPF lung tissues, based on microarray technology, were retrieved from the GEO database. GSE35145 and GSE10667 were utilized as the training sets, while GSE24206 and GSE53845 served as external validation sets. Batch effects in the combined training datasets were rectified using the ComBat algorithm (sva package, parametric prior). The performance of this correction was evaluated via principal component analysis (PCA) and relative log expression (RLE) plots. The external validation datasets were normalized using the same procedure but analyzed separately. Differential expression analysis was carried out within the limma linear modeling framework, applying empirical Bayes moderation. Genes with |log₂fold change (FC)| ≥0.585 and false discovery rate (FDR) <0.05 were considered differentially expressed.

The single-cell transcriptomic dataset GSE128033 was retrieved from GEO and analyzed using Seurat (v4.0) (9). Cells that met the following quality-control criteria were retained: 500< nFeature_RNA <6,000, nCount_RNA >1,000, and percent.mt <10%. After normalization with the LogNormalize method, 2,000 highly variable genes were selected, and then PCA was performed for dimensionality reduction. The first eight principal components were retained according to the elbow plot. Harmony was then applied to reduce batch effects across samples. A nearest-neighbor graph was constructed from the corrected low-dimensional embeddings for cell clustering, with the resolution set to 0.8 after evaluation by clustree. Uniform manifold approximation and projection (UMAP) was used for visualization. Cell-type annotation was determined through the integration of SingleR-based automated annotation, AUCell marker scoring, and canonical marker-gene expression patterns. Marker genes were defined mainly in accordance with previously published human lung single-cell reference atlases and related studies (10-12). The full list is provided in Table S1. Subsequently, cells were assigned to the major lung cell types, and AT2 cells were isolated for downstream analyses. Cell-type proportions were computed at the sample level and contrasted between groups by applying the Wilcoxon rank-sum test with Benjamini-Hochberg correction for multiple testing. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.

AT2 pseudo-bulk differential expression analysis

After cell-type annotation, AT2 cells were extracted, and raw UMI counts were aggregated per sample to generate a pseudo-bulk expression matrix (13). After lowly expressed genes were filtered using filterByExpr, differential expression analysis was carried out with edgeR. This analysis included Trimmed Mean of M-values (TMM) normalization, fitting of the generalized linear model, and evaluation of between-group differences based on the quasi-likelihood F-test. Multiple testing was regulated using the Benjamini-Hochberg method. Genes with |log₂FC| ≥0.585 and FDR <0.05 were regarded as differentially expressed.

Construction of PSRGs, candidate gene screening, and enrichment analysis

Peri-anesthetic drug-target genes reported in previous studies (14) and mechanical stimulus-related genes (15) were merged and deduplicated to construct the PSRGs (Table S2). Differentially expressed genes identified from the IPF bulk transcriptomic and AT2 pseudo-bulk transcriptomic datasets were first classified by direction of change (upregulated or downregulated). The intersections of concordantly upregulated genes and concordantly downregulated genes were then determined separately across the two datasets and combined to create a set of consistently differentially expressed genes. The resulting gene set was subsequently intersected with the PSRGs to identify candidate hub genes. In this study, PSRGs were considered as a predefined, biologically relevant prior gene framework for candidate prioritization, instead of as direct transcriptomic signatures of peri-anesthetic exposure in IPF. Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analyses were performed using the clusterProfiler package. An FDR adjusted by the Benjamini–Hochberg method and less than 0.05 was considered significant.

Machine-learning model construction and interpretation

To prioritize candidate genes with cross-cohort robustness, machine-learning models were built using genes shared between the training set and the external validation cohorts. After centering and scaling the expression matrix, a multi-algorithm framework incorporating least absolute shrinkage and selection operator (LASSO), random forest, and support vector machine (SVM) was established. Feature selection was initially performed in the training set, and the selected features were then used for predictive model construction. Each candidate model was trained by cross-validation in the training set and further evaluated in the external cohorts for discriminative performance and cross-cohort robustness. Receiver operating characteristic (ROC) curves were utilized for performance assessment, with the area under the curve (AUC) serving as the primary metric. Based on the results from the training set and external cohorts, candidate models were compared, and the model with the optimal overall cross-cohort performance was selected for subsequent analyses. For the selected model, the optimal classification threshold was determined in the training set by applying the Youden index. This threshold was then used to calculate accuracy, sensitivity, and specificity. The positions of the selected core genes in the differential-expression landscape were also visualized in a volcano plot. SHapley Additive exPlanations (SHAP) was applied to interpret model predictions by quantifying the contribution of individual features. In addition, a logistic regression model based on the chosen core genes was constructed and presented as a nomogram. Calibration curves were utilized to assess predictive agreement, and decision curve analysis (DCA) was carried out as a supplementary evaluation of model behavior across threshold probabilities.

AT2 subclustering and state annotation

AT2 cells were extracted after initial cell-type annotation and underwent subclustering analysis. Following normalization, highly variable genes were identified and PCA was performed for dimensionality reduction. A nearest-neighbor graph was constructed using the first 15 principal components for clustering. UMAP was used for visualization. After eliminating low-quality cells, DoubletFinder-positive cells, and clusters demonstrating expression of non-AT2 lineage markers, dimensionality reduction and clustering were reiterated to acquire a refined clean AT2 subset. Module scores were calculated from mature AT2 marker genes (SFTPC, SFTPB, SFTPA1, SFTPA2, ABCA3, NAPSA, and SLC34A2) and injury/transitional-state marker genes (KRT8, KRT18, KRT19, CLDN4, and KRT17) (12,16,17). Cells were then categorized into mature and injury/transitional states according to their relative scores. Subcluster markers were identified based on FindAllMarkers, and subcluster composition as well as state proportions were quantified across samples. Proportion estimation and between-group comparisons were carried out as described above. Considering the possibility for minor contamination within the initially extracted AT2 population, the core candidate genes were further verified for robustness in the clean-AT2 subset. Particular attention was paid to their consistency with mature and injury/transitional state classification to support subsequent biological interpretation.

Independent single-cell RNA sequencing (scRNA-seq) validation and cross-cohort reproducibility analysis

For independent validation, GSE135893 was processed using the same scRNA-seq workflow and AT2-state marker framework as the discovery cohort. IGFBP2 expression was summarized at the sample-state level using pseudo-bulk log1p[counts per million (CPM)] and mean normalized expression. Mature and injury/transitional AT2 states were compared using paired Wilcoxon tests, with Cohen’s dz and bootstrap 95% confidence intervals used for effect-size estimation. An AT2 transition score was defined as the mean injury/transitional marker score minus the mean mature AT2 marker score. Spearman correlation was used to assess the association between sample-state mean IGFBP2 expression and the transition score in both cohorts, with Benjamini-Hochberg adjustment applied to P values. To assess potential technical confounding in GSE135893, sample-state mean IGFBP2 expression was further correlated with median nCount_RNA, median nFeature_RNA, and median percent.mt.

AT2 pseudotime trajectory inference

The reclustered AT2 cells were subjected to pseudotime analysis using monocle3 (18). A cell_data_set was constructed from the Seurat expression matrix. Highly variable genes were employed as ordering genes, and core genes were preserved in the analysis gene set. Pseudotime ordering was performed with mature AT2 cells designated as the primary root. As a sensitivity analysis, SFTPC-high cells were utilized as an alternative root to evaluate the consistency of pseudotime results under different root settings. The pseudotime trajectory was depicted in UMAP space, and pseudotime distributions were contrasted between AT2 states using density plots and the Wilcoxon rank-sum test. The dynamic expression patterns of core genes along pseudotime were illustrated with LOESS-smoothed curves and binned mean trajectories. Correlations with pseudotime were further evaluated by Spearman analysis. Subsequent mechanistic analyses were focused primarily on the core gene showing a trend consistent with AT2 state transition, namely IGFBP2.

Cell-cell communication analysis

CellChat was applied to infer the ligand-receptor communication network among cells. Based on the normalized expression matrix and cell-type annotations, AT2 cells were further subdivided into AT2_mature and AT2_injury/transitional subgroups, whereas all other cell types retained their original annotations. The analysis was conducted using the human CellChat database, with the secreted signaling and ECM-receptor categories selected to concentrate on secreted signals and ECM-related intercellular communication. The cell-cell communication network was reconstructed in the full cellular context, and ligand-receptor communication probabilities were computed. Interaction number and interaction strength were further contrasted among different cell types at the network level. At the pathway level, communication probabilities were aggregated to assess the activity of relevant signaling pathways. CellChat outputs were also queried at the ligand-receptor level by ranking source-target-specific communication pairs according to inferred communication probability, with the IGFBP signaling category screened for IGFBP2-containing interactions. Subsequent analyses focused on the incoming and outgoing signaling patterns between AT2_injury/transitional cells and other cell types, together with the associated signaling pathways (19).

Computational perturbation analysis of IGFBP2

To evaluate the potential regulatory role of the core gene IGFBP2 in injury/transitional AT2 cells, a computational perturbation analysis was conducted using scTenifoldKnk v1.0.3. (20). Multiple gene regulatory networks (GRNs) were constructed from the expression matrix of the target subpopulation by means of random subsampling. Before network construction, 5,000 highly variable genes were selected after LogNormalize normalization, with IGFBP2 retained in the input gene set. Genes expressed in fewer than 30 cells or showing near-zero variance were removed. scTenifoldKnk was run with 25 reconstructed networks and up to 1,000 cells per network. Network alignment was then employed to evaluate differential regulatory effects before and after IGFBP2 virtual perturbation. Genes showing significant network-inferred regulatory changes after IGFBP2 virtual perturbation were identified based on differential regulatory scores and statistical significance (FDR <0.05). To further contextualize the knockout-inferred genes within AT2 remodeling, their expression patterns were projected back onto the clean-AT2 UMAP. Comparisons were made between mature and injury/transitional AT2 states. Their expression dynamics along pseudotime were further examined using state-level expression plots, feature plots, LOESS-smoothed trajectories, and binned mean curves. In addition, sample-level distributions within injury/transitional AT2 cells were descriptively assessed.

Pathway analysis of core genes

Within the IPF group, samples were stratified into high- and low-IGFBP2 expression groups according to the median IGFBP2 level. Differential expression analysis was then performed between these two groups, and genes were ordered according to their log₂FC values to generate the ranked list for downstream enrichment analysis. Gene set enrichment analysis (GSEA) was performed using the clusterProfiler R package with gene sets derived from KEGG. Benjamini-Hochberg-adjusted P values were used for significance assessment, with FDR <0.05 as the cutoff. Representative positively and negatively enriched pathways were displayed according to the normalized enrichment score (NES). To further evaluate pathway activity at the individual-sample level, single-sample gene set enrichment analysis (ssGSEA) was carried out using the gene set variation analysis (GSVA) package in R. The resulting pathway scores were compared between the IGFBP2-high and IGFBP2-low groups. Group differences were tested using two-sided t-tests, and the corresponding P values were further adjusted by the Benjamini-Hochberg procedure to account for multiple comparisons.

Virtual screening and exploratory structural assessment of IGFBP2

Structure-based virtual screening was carried out on the DrugCLIP platform (www.drugclip.com) to identify small-molecule compounds with potential binding affinity for IGFBP2 (Enamine_screening_collection_sdf_202504 library). DrugCLIP was used as a hosted web-screening platform, and no local model training, hyperparameter tuning, or re-training was performed; therefore, local training parameters were not applicable. The highest-scoring protein binding pocket was chosen as the docking center. Candidate compounds were first ranked according to the DrugCLIP score and further refined using the docking score. Toxicity and drug-likeness were predicted with ADMETlab 2.0, and structural modifications were introduced to reduce potential toxicity and improve drug-likeness. The IGFBP2 structure (AF-P18065-F1-v6) was retrieved from the AlphaFold database (www.alphafold.com). The top 10 candidate compounds, ranked by DrugCLIP score, were then subjected to molecular docking with AutoDock Vina. A grid box, centered at (−6.99, −2.38, 12.33) with dimensions of 47×47×47 Å, exhaustiveness =32, and seed =20,250,101, was used. The center coordinates were defined according to the predicted center of the optimal binding pocket identified by DrugCLIP. The pose with the lowest Vina score was selected for molecular dynamics simulation. Molecular dynamics simulations were conducted using GROMACS 2024.4 with a 200-ns production run. The system was parameterized with the Amber99-ILDN force field and the TIP3P water model, and Na⁺/Cl⁻ ions were added for charge neutralization. Energy minimization was performed first, followed by NVT and NPT equilibration. Trajectories were saved every 10 ps, and complex stability was evaluated by analyzing root mean square deviation (RMSD), root mean square fluctuation (RMSF), radius of gyration (Rg), and solvent-accessible surface area (SASA).

Statistical analysis

All statistical analyses and visualization were performed using R software packages (v4.0 or later). For bulk transcriptomic and pseudo-bulk data, differential expression analyses were executed using the empirical Bayes moderation framework within the limma package or the generalized linear model with quasi-likelihood F-test within the edgeR package. Group-level comparisons of cell-type proportions, cluster distributions, and pathway scores were performed using the Wilcoxon rank-sum test or two-sided t-tests. For paired sample-state comparisons in the validation cohort, the paired Wilcoxon test was applied, with effect sizes estimated using Cohen’s dz and bootstrap 95% confidence intervals. Continuous variables, trajectory positions, and gene expression dynamics across pseudotime were evaluated using Spearman correlation analysis. Feature selection and machine learning model performance were assessed using ROC curves and the AUC metric. Multiple testing corrections across differential expression, correlation, and pathway enrichment analyses were strictly regulated using the BH procedure. A two-sided FDR <0.05 was considered statistically significant.

Results

Bulk transcriptomic differential analysis and single-cell clustering

After integration and batch correction of the training datasets GSE35145 and GSE10667 (Figure 2A-2D, Figure S1A,S1B), differential expression analysis uncovered a series of genes significantly dysregulated between the IPF and control groups, as illustrated by the volcano plot and heatmap (Figure 2E,2F). As for the single-cell dataset, after quality control and preprocessing, it was retained for downstream analysis (Figure S1C-S1J). The integrated UMAP showed good mixing of cells from different samples, indicating that batch effects had been effectively removed (Figure 3A). Clustering analysis identified 24 cell clusters with distinct transcriptional features (Figure 3B, Figure S1K). Based on marker-gene expression patterns, these clusters were assigned to nine major cell types (Figure 3C). Expression of representative marker genes further validated the annotation results (Figure 3D). Analysis of cell-type composition across samples unveiled distinct distribution patterns among the major cell populations (Figure 3E). Several cell types also showed marked differences in proportion between the Control and IPF groups (Figure 3F).

Figure 2.

Defferential gene expression analysis of IPF bulk transcriptomes after barch correction. (A,B) PCA before and after batch correction; (C,D) boxplots before and after batch correction; (E) volcano plot of differentially expressed genes between IPF and control lung tissues; (F) heatmap showing the expression patterns of representative differentially expressed genes. FC, fold change; IPF, idiopathic pulmonary fibrosis; PCA, principal component analysis.

Figure 3.

Single-cell Atlas of IPF lung identifies major cell types. (A) UMAP visualization of all cells before and after Harmony correction. (B) UMAP showing 24 transcriptionally distinct cell clusters identified at a Seurat resolution of 0.8. (C) Manual annotation of nine major lung cell types. (D) Dot plot of representative marker genes across annotated cell types. Dot size indicates the percentage of expressing cells, and color intensity indicates average expression level. (E) Sample-wise distribution of major cell populations. (F) Comparison of cell-type proportions between control and IPF samples. AT2, alveolar type II; IPF, idiopathic pulmonary fibrosis; UMAP, uniform manifold approximation and projection.

Transcriptional alterations in AT2 cells revealed by pseudo-bulk analysis

Pseudo-bulk differential expression analysis revealed significant transcriptional differences in AT2 cells between the IPF and control groups. Multidimensional scaling (MDS), PCA, biological coefficient of variation (BCV), M versus A (MA), and quantile-quantile (QQ) plots supported the overall quality and statistical validity of the data (Figure S2A-S2E). Under the predefined thresholds, both upregulated and downregulated genes were recognized, as shown in the volcano plot and heatmap (Figure 4A,4B).

Figure 4.

Intergrated bulk and single-cell analysis identifies PSRG-related hub genes. (A) Volcano plot of differentially expressed genes identified in the AT2 pseudo-bulk analysis; (B) heatmap of the top differentially expressed genes in AT2 pseudo-bulk samples from control and IPF lungs; (C,D) venn diagrams showing concordantly upregulated and downregulated genes shared between the bulk IPF dataset and the AT2 pseudo-bulk dataset; (E) overlap between PSRGs and concordantly dysregulated genes, yielding six candidate hub genes; (F) GO enrichment analysis of the candidate hub genes; (G) KEGG pathway enrichment analysis of the candidate hub genes. AT2, alveolar type II; DEG, differentially expressed gene; ECM, extracellular matrix; FDR, false discovery rate; GO, Gene Ontology; IPF, idiopathic pulmonary fibrosis; KEGG, Kyoto Encyclopedia of Genes and Genomes; NS, not significant; PSRG, peri-anesthetic stress-related gene.

Prioritization of candidate genes through PSRGs overlap and functional enrichment analyses

A total of 326 PSRGs were obtained by integrating gene sets reported in previous studies. Afterwards, 163 genes exhibiting concordant expression trends in both the bulk and AT2 pseudo-bulk transcriptomes were identified (Figure 4C,4D). On this basis, overlap analysis with the PSRGs further yielded six candidate hub genes, including MMP14, MDK, COL1A1, IGFBP2, KRT5, and HPN (Figure 4E). GO enrichment analysis indicated that these candidate genes were primarily associated with response to mechanical stimulus, epithelial-to-mesenchymal transition, ECM, collagen trimer, and growth factor binding (Figure 4F). KEGG analysis further showed that they were mainly enriched in the ECM-receptor interaction and AGE–RAGE signaling pathways (Figure 4G). To further assess the specificity of AT2-derived signals, the key candidate genes identified here were subsequently examined in a refined clean-AT2 subset after removal of contaminating non-AT2 clusters (see section “AT2 subclustering and state annotation”).

Machine-learning-assisted prioritization of candidate genes and interpretation

Satisfactory predictive performance was achieved by the model built using the multi-algorithm combinatorial framework in both the training set and the external validation cohorts (Figure 5A). Among the candidate models, the optimal model was found to yield relatively high AUC values across multiple datasets (Figure 5B-5D). Its stable classification performance was further supported by the confusion matrices in the three datasets (Figure 5E-5G). Further support for the discriminative potential of the core genes was offered by single-gene ROC analysis (Figure 5H). The accuracy, F1 score, precision, sensitivity, and specificity metrics are shown in Figure 5I. Significant differences in the expression of COL1A1 and IGFBP2 were observed between the control and IPF groups (Figure 5J-5L). Meanwhile, only weak correlations were detected between these two genes, suggesting the absence of appreciable collinearity (Figure 5M-5O). Both IGFBP2 and COL1A1 were also highlighted in the differential-expression landscape (Figure 6A). SHAP analysis further showed that both genes made important contributions to the model predictions, supporting their initial prioritization at the screening stage (Figure 6B,6C). The SHAP dependence plot and individual-sample interpretation results further characterized the influence of feature expression levels on model predictions (Figure 6D and Figure S2F-S2G). In addition, the logistic regression model built based on the core genes was presented as a nomogram (Figure 6E). The calibration curve demonstrated good concordance between the predicted and observed probabilities (Figure 6F). DCA was generally consistent with the overall model performance (Figure 6G) and served as a supplementary evaluation of model behavior across threshold probabilities.

Figure 5.

Machine learning identifies IGFBP2 and COLIA1 as key diagnostic genes. (A) Heatmap summarizing the performance of candidate machine-learning models across datasets; (B-D) ROC curves of the optimal model in the training cohort, GSE24206, and GSE53845; (E-G) confusion matrices of the optimal model in the training cohort, GSE24206, and GSE53845; (H) single-gene ROC curves for COL1A1 and IGFBP2; (I) classification metrics of the optimal model across datasets at the selected threshold; (J) expression of COL1A1 and IGFBP2 in the training cohort; (K) expression of COL1A1 and IGFBP2 in GSE24206; (L) expression of COL1A1 and IGFBP2 in GSE53845; (M) correlation analysis between COL1A1 and IGFBP2 in the training cohort; (N) correlation analysis between COL1A1 and IGFBP2 in GSE24206; (O) correlation analysis between COL1A1 and IGFBP2 in GSE53845. **, P<0.01; ***, P<0.001. AUC, area under the curve; CI, confidence interval; GBM, glioblastoma; IPF, idiopathic pulmonary fibrosis; LASSO, least absolute shrinkage and selection operator; LDA, linear discriminant analysis; RF, random forest; ROC, receiver operating characteristic; SVM, support vector machine.

Figure 6.

Model interpretation and clinical validation of IGFBP2/COL1A1 signature. (A) Volcano plot highlighting IGFBP2 and COL1A1 in the differential-expression landscape; (B) SHAP summary plot showing the distribution of SHAP values for COL1A1 and IGFBP2 across samples; (C) mean absolute SHAP values of COL1A1 and IGFBP2; (D) SHAP dependence plots illustrating the influence of feature expression on model output; (E) nomogram based on COL1A1 and IGFBP2; (F) calibration curve of the nomogram; (G) decision curve analysis. SHAP, SHapley Additive exPlanations.

AT2 subclustering and state annotation

After reclustering the annotated AT2 cells, the initial results showed that these cells could be divided into multiple subpopulations with distinct transcriptional features (Figure S3A,S3B). Further evaluation of marker expression patterns revealed that some subclusters expressed markers of non-AT2 lineages, suggesting contaminated clusters (Figure S3C). After removing the contaminating clusters, a marked improvement was observed in the AT2 marker expression pattern (Figure S3D), and a more purified clean-AT2 population was obtained (Figure 7A). Based on the expression patterns of mature AT2 marker genes and injury/transitional state markers, the clean-AT2 population could be further annotated into two states, namely mature AT2 and injury/transitional AT2 (Figure 7B). Notably, after removing the contaminating clusters, both IGFBP2 and COL1A1 remained highly expressed in injury/transitional AT2 cells. IGFBP2 showed a more prominent signal, suggesting a clearer association with the injury/transitional AT2 state (Figure 7C). Additional comparison of state composition between groups revealed that the proportion of injury/transitional AT2 cells was relatively higher in the IPF group (Figure 7D).

Figure 7.

AT2 cell subtypes and IGFBP2 dynamics in IPF. (A) UMAP of the refined clean-AT2 subset after removal of contaminating non-AT2 clusters; (B) annotation of mature and injury/transitional AT2 states based on state marker expression, the lower panel shows marker-gene expression patterns across the two AT2 states; (C) expression of IGFBP2 and COL1A1 in mature and injury/transitional AT2 cells; (D) comparison of mature and injury/transitional AT2 state proportions between control and IPF samples; (E) pseudotime trajectory of clean-AT2 cells inferred by monocle3; (F) projection of mature and injury/transitional AT2 states onto the pseudotime trajectory; (G) density distribution of mature and injury/transitional AT2 cells along pseudotime; (H) dynamic expression patterns of COL1A1 and IGFBP2 along pseudotime. ***, P<0.001. AT2, alveolar type II; IPF, idiopathic pulmonary fibrosis; UMAP, uniform manifold approximation and projection.

Pseudotime analysis of AT2 state transitions and progressive IGFBP2 upregulation

Pseudotime analysis showed that AT2 cells formed a continuous trajectory in UMAP space, stretching from mature AT2 to injury/transitional AT2 (Figure 7E). This pattern suggested that the inferred trajectory might reflect transcriptional continuity associated with changes in AT2 cell state. A similar distribution trend was observed when the trajectory was colored by cluster identity or sample group (Figure S3E,S3F). When AT2 states were projected onto the trajectory, mature AT2 cells were predominantly found at earlier pseudotime stages, while injury/transitional AT2 cells were more frequently distributed at later stages (Figure 7F). This distribution was consistent with the inferred direction of AT2 state transitions. Pseudotime density analysis further showed that mature AT2 cells were predominantly concentrated at earlier pseudotime stages, whereas injury/transitional AT2 cells exhibited a broader distribution extending into later stages (Figure 7G). Differences in pseudotime density profiles among sample groups are shown in Figure S3G. Analysis of IGFBP2 expression dynamics along the pseudotime further revealed that its expression gradually increased as the pseudotime progressed, suggesting a potential association with changes in the AT2 cell state (Figure 7H). Higher expression levels of the mature AT2 markers (SFTPC and ABCA3) were observed at earlier pseudotime stages, meanwhile gradual upregulation of the injury-associated markers (KRT8 and CLDN4) was detected at later stages, further supporting the use of mature AT2 as the trajectory root (Figure S3H-S3J).

Independent scRNA-seq validation and cross-cohort reproducibility of IGFBP2 enrichment

In the independent validation cohort GSE135893, canonical marker patterns supported the major cell annotations and the mature versus injury/transitional AT2 state assignment (Figure S4A,S4B). IGFBP2 was preferentially enriched in the injury/transitional AT2 compartment, spatially aligning with the KRT8-high state rather than the SFTPC-high mature AT2 state (Figure S4C,S4D). This enrichment was consistent across GSE128033 and GSE135893. Sample-level pseudo-bulk, mean-expression, and transition-score analyses all supported higher IGFBP2 expression in injury/transitional AT2 cells in both cohorts (Figure S4E-S4G; Table S3), indicating that the observed pattern was reproducible rather than dataset-specific. In GSE135893, IGFBP2 expression showed no significant association with sample-state median nCount_RNA, nFeature_RNA, or percent.mt (Figure S5; Table S4). This negative-control analysis suggests that the IGFBP2 enrichment pattern was unlikely to be primarily explained by sequencing depth, detected gene number, or mitochondrial content.

CellChat-based cell-cell communication analysis

CellChat analysis was conducted to characterize ligand-receptor interactions among different cell types. Mature AT2 and injury/transitional AT2 were analyzed as separate states within the AT2 population. Widespread ligand-receptor interactions were inferred across multiple cell types in the overall communication network. Moreover, relatively prominent interaction activity was identified between injury/transitional AT2 cells and several other cell populations (Figure 8A,8B). The signaling database composition and global ligand-receptor landscape are shown in Figure S6A,S6B. At the signaling pathway level, a prominent incoming pattern from Mesenchymal cells to injury/transitional AT2 cells was inferred. Enrichment was observed mainly in ECM-related pathways, including COLLAGEN, LAMININ, FN1, and THBS, along with growth factor-like signaling such as HGF and FGF (Figure 8C,8D). Outgoing signaling analysis further suggested that injury/transitional AT2 cells may participate in multiple signaling interactions with Endothelial cells and Macrophages. Interactions with Endothelial cells were found to be mainly enriched in pathways such as VEGF, IGFBP, and CXCL (Figure 8E). In comparison, interactions with Macrophages involved inflammatory pathways including MIF, COMPLEMENT, and VISFATIN, as well as inflammation- and stress-related signals such as ANNEXIN and CypA (Figure 8F). Key inferred ligand-receptor pairs and matched mature AT2 comparator interactions are summarized in Tables S5-S10. Comparative ligand-receptor analysis indicated selective remodeling rather than uniform enhancement of injury/transitional AT2 communication. The IGFBP pathway screen did not identify a robust IGFBP2-specific ligand-receptor pair; the detected IGFBP pathway signal mainly involved IGFBP3-TMEM219 (Table S11). Therefore, these CellChat results were interpreted as inferred microenvironmental communication patterns associated with the injury/transitional AT2 state, rather than direct evidence of IGFBP2-mediated signaling to fibroblasts, macrophages, or endothelial cells.

Figure 8.

Cell-cell communication analysis reveals AT2 interaction changes in IPF. (A) Global intercellular communication network showing the number of inferred ligand-receptor interactions among major cell populations; (B) global intercellular communication network showing the overall interaction strength among major cell populations; (C) bubble plot of inferred incoming ligand-receptor interactions from mesenchymal cells to injury/transitional AT2 cells; (D) pathway-level summary of incoming signaling from mesenchymal cells to injury/transitional AT2 cells; (E) pathway-level summary of outgoing signaling from injury/transitional AT2 cells to endothelial cells; (F) pathway-level summary of outgoing signaling from injury/transitional AT2 cells to macrophages. AT2, alveolar type II.

Computational perturbation analysis of IGFBP2 in injury/transitional AT2 cells

Following scTenifoldKnk-based virtual perturbation of IGFBP2 in injury/transitional AT2 cells, five genes showed significant network-inferred regulatory changes, including SCGB1A1, SCGB3A1, BPIFB1, SLPI, and PIGR (Figure 9A,9B). Most of these genes were related to epithelial secretory function and mucosal defense. At the transcriptome-wide level, only a small number of genes reached the adjusted P value threshold (Figure 9C), suggesting that the inferred IGFBP2-associated perturbation was relatively restricted rather than broadly distributed across the global transcriptional landscape. Detailed scTenifoldKnk parameters and significant knockout-inferred genes are summarized in Tables S12,S13.

Figure 9.

IGFBP2 perturbation identifies epithelial secretory gene networks. (A) Differential regulatory scores of genes inferred after scTenifoldKnk-based virtual perturbation of IGFBP2; (B) significant network-inferred genes identified after IGFBP2 virtual perturbation; (C) transcriptome-wide distribution of inferred regulatory changes following IGFBP2 virtual perturbation; (D) expression of IGFBP2 and representative knockout-inferred genes in mature and injury/transitional AT2 cells within the clean-AT2 subset; (E) UMAP projection of IGFBP2 and representative knockout-inferred genes in the clean-AT2 subset; (F) dynamic expression patterns of representative knockout-inferred genes along pseudotime. AT2, alveolar type II; FC, fold change; UMAP, uniform manifold approximation and projection.

These knockout-inferred genes were evaluated in the clean-AT2 subset. Higher expression of IGFBP2, SCGB1A1, SCGB3A1, and BPIFB1 was observed in injury/transitional AT2 cells than in mature AT2 cells. SLPI showed a broader epithelial pattern with relative enrichment in the injury/transitional state, and PIGR showed greater heterogeneity (Figure 9D). On the clean-AT2 UMAP, SCGB1A1, SCGB3A1, and BPIFB1 were mainly localized to the injury/transitional region and partially overlapped with IGFBP2 (Figure 9E). Along pseudotime, a subset of knockout-inferred genes, particularly SCGB1A1, SCGB3A1, and BPIFB1, showed higher expression in the mid-to-late stages and remained more strongly expressed along the injury/transitional trajectory than along the mature trajectory (Figure 9F). Overall, these network-inferred genes were preferentially associated with an injury/transitional AT2-related epithelial secretory/barrier program (additional validations are shown in Figure S6C-S6F).

Pathway analysis associated with IGFBP2 expression

GSEA revealed that the IGFBP2 high-expression group was mainly enriched in pathways related to xenobiotic metabolism and glutathione-associated metabolic processes, including drug metabolism-cytochrome P450, glutathione metabolism, and metabolism of xenobiotics by cytochrome P450 (Figure 10A). By contrast, the IGFBP2 low-expression group was predominantly enriched in multiple immune-related pathways, including antigen processing and presentation, chemokine signaling pathway, cytokine-cytokine receptor interaction, and natural killer cell mediated cytotoxicity (Figure 10B). GSVA further demonstrated that several immune- and inflammation-related pathways, such as chemokine signaling, NOD-like receptor signaling, MAPK signaling, and leukocyte transendothelial migration, were more active in the IGFBP2 low-expression group. Higher activity of metabolism-related pathways, including glutathione metabolism and xenobiotic metabolism, was observed in the IGFBP2 high-expression group (Figure 10C). These findings indicate that IGFBP2 expression is associated with distinct metabolic and immune-related pathway activity patterns within the IPF transcriptomic context.

Figure 10.

Functional and structural analysis of IGFBP2 using pathway and molecular modeling. (A) GSEA results for the IGFBP2 high-expression group; (B) GSEA results for the IGFBP2 low-expression group; (C) ssGSEA-based comparison of pathway activity between IGFBP2-high and IGFBP2-low samples; (D) predicted toxicity and drug-likeness before and after structural optimization of the top candidate compound; (E) predicted binding mode of the optimized compound within the putative IGFBP2 binding pocket; (F) RMSF profiles of apo IGFBP2 and the IGFBP2-ligand complex; (G) Rg of apo IGFBP2 and the IGFBP2-ligand complex; (H) RMSD profiles of apo IGFBP2 and the IGFBP2-ligand complex; (I) SASA profiles of apo IGFBP2 and the IGFBP2-ligand complex. GSEA, gene set enrichment analysis; KEGG, Kyoto Encyclopedia of Genes and Genomes; Rg, radius of gyration; RMSD, root mean square deviation; RMSF, root mean square fluctuation; SASA, solvent-accessible surface area; ssGSEA, single-sample gene set enrichment analysis.

Exploratory structural assessment of potential small-molecule accessibility of IGFBP2

Virtual screening based on the DrugCLIP platform identified multiple candidate compounds having potential affinity for IGFBP2, among which the parent scaffold with the highest DrugCLIP score exhibited favorable predicted binding but relatively high potential toxicity (Table 1 and Table S14). After structural optimization, an improved candidate was obtained. For this candidate, reduced predicted toxicity and retained drug-likeness were observed (Figure 10D and Table S15). The optimized compound was predicted to occupy the putative DrugCLIP-defined binding pocket of IGFBP2 and to interact with several key residues (Figure 10E). Molecular dynamics simulations further suggested that the IGFBP2-ligand complex could be maintained under the simulated conditions. Compared with the apo protein, the complex showed reduced residue-level fluctuations and a lower radius of gyration, indicating decreased local flexibility and a more compact overall conformation (Figure 10F,10G). RMSD and SASA profiles further supported structural compatibility of the complex under the simulated conditions, although they did not indicate uniformly stronger global stabilization (Figure 10H,10I). Detailed information on the DrugCLIP web-screening procedure, model principle, compound prioritization, docking settings, ADMET evaluation, and molecular dynamics parameters is provided in Tables S16-S18.

Table 1. Top candidate compounds identified by DrugCLIP screening against IGFBP2.

Mol ID	Library	SMILES	DrugClip score	Docking score
FCG3595354617	FCHGroup_SC_202007	Cc1ccc(o1)-c2nc(C)c(s2)C(=O)N3C[C@@H](O)[C@@H](O)CO3	2.47	−6.095
FCG2915212305	FCHGroup_SC_202007	OC(C1)CN(C1(C)CO)C(=O)c(c2C)sc(n2)-c3ccc(Cl)cc3	2.43	−6.568
Z2958564350	Enamine_screening_collection_sdf_202504	Clc1ccc(Cl)cc1-c(nc2C)sc2C(=O)N3C[C@@H](O)C[C@H]3CO	2.35	−6.548
Z4424122789	Enamine_screening_collection_sdf_202504	OC[C@@H](C1)O[C@@H](CO)CN1C(=O)c2ccc(cc2)-c(ccc3)c(c34)nccc4	2.3	−7.147
FCG3595351950	FCHGroup_SC_202007	CCc1ccc(cc1)-c2nc(cs2)C(=O)N3C[C@@H](O)[C@@H](O)CO3	2.29	−7.03
OSSM_226728	Princeton_BioMolecular_Research	OCC(O)(C)C(=O)/C=C/c1cc(ccc1)OCc2ccccc2	2.24	−6.531
FCG3595351583	FCHGroup_SC_202007	O[C@@H]1CON(C[C@@H]1O)C(=O)c2ccc(o2)-c3ccccc3Cl	2.24	−6.325
FCG2915208098	FCHGroup_SC_202007	OC(C1)CN(C1(C)CO)C(=O)c(c2C)sc(n2)-c3ccc(F)c(F)c3	2.24	−6.967
Z4424111760	Enamine_screening_collection_sdf_202504	OC[C@H](C1)O[C@@H](CO)CN1C(=O)c2ccc(cc2)-c(ccc3)c(c34)nccc4	2.24	−6.267
Z1601657767	Enamine_screening_collection_sdf_202504	COc(c1)c(OC)cc(c12)c(N)nc(n2)N3CCN([C@@H](C)C3)S(=O)(=O)C	2.23	−5.793

SMILES, Simplified Molecular Input Line Entry System.

Discussion

IPF is a progressive interstitial lung disease with a poor prognosis, and its pathogenesis has been linked in part to aberrant AT2 reprogramming (21,22). Considering the high in-hospital mortality associated with AE-IPF (23), the potential link between PSRGs and AT2 cell injury warrants further attention. In this study, PSRGs were used as a literature-derived prioritization framework rather than as evidence that peri-anesthetic stress directly causes IPF or that the analyzed datasets represent peri-anesthetic exposure. Through integrated analysis of bulk and single-cell transcriptomic data, IGFBP2 was identified as a core candidate gene associated with AT2 state transitions in IPF. Its enrichment in injury/transitional AT2 cells was reproduced in both the discovery cohort GSE128033 and the independent validation cohort GSE135893. Sample-level expression analyses showed consistent directionality across cohorts, and IGFBP2 expression was positively associated with the AT2 transition score in both datasets. These findings support IGFBP2 as a reproducible injury/transitional AT2-associated candidate rather than a signal restricted to a single scRNA-seq dataset.

Both IGFBP2 and COL1A1 were identified as core candidate genes during the machine learning-based screening process, suggesting that both may participate in the transcriptional remodeling associated with IPF. COL1A1 is more commonly associated with the ECM-producing fibroblast/myofibroblast program rather than serving as a canonical marker of AT2 identity (24,25). Recent mechanistic studies have shown that, during the pathological progression of IPF, injured AT2 cells may be characterized by impaired AT2-to-AT1 differentiation, accumulation of transitional-state cells, and aberrant expression of ECM-related genes and proteins. Through dysregulated epithelial-fibroblast interactions, fibroblast activation, myofibroblast differentiation, and ECM deposition may be further promoted, thereby contributing to amplification of the fibrotic microenvironment (26,27). In this context, the COL1A1 signal remained biologically interpretable within the AT2-centered framework of the present study. Rather than reflecting canonical AT2 identity, it was more likely to represent an ECM-related transcriptional context accompanying the injury/transitional AT2 state. This interpretation is consistent with recent evidence that fibrotic AT2 cells may actively contribute to ECM remodeling and collagen crosslinking through epithelial YAP-TEAD/LOX signaling (28). After the contaminating clusters had been removed, COL1A1 expression was still detected in injury/transitional AT2 cells within the clean-AT2 subset. This finding suggests that the observed signal was not entirely derived from contaminating cells. Nevertheless, subsequent analyses focused on IGFBP2, as its association with AT2 state changes was more consistent across state stratification, pseudotime, and clean-AT2 robustness analyses.

To further interpret the pathological context surrounding injury/transitional AT2 cells, their potential relationship with the fibrotic lung microenvironment was considered at the tissue level. Cell-cell communication analysis showed that the inferred incoming signals from Mesenchymal cells to injury/transitional AT2 cells were mainly enriched in ECM-related pathways, including COLLAGEN, LAMININ, FN1, and THBS. These ECM-associated signals have been reported to mediate epithelial stress responses through receptors such as integrins or CD44, and may also be functionally linked to the IGF/IGFBP signaling axis (29-31). Accordingly, mesenchymal cell-derived ECM signals may be associated with the transcriptional context of the injury/transitional AT2 state. Inferred outgoing signaling from injury/transitional AT2 cells to endothelial cells and macrophages involved VEGF, CXCL, IGFBP, and inflammation-related pathways, suggesting that the AT2 injury state may be associated with changes in the local vascular and immune microenvironment (32,33). Consistently, bulk pathway analysis indicated that IGFBP2 expression was associated with metabolism- and immune-related transcriptomic contexts. Previous studies have suggested that disrupted oxidative stress homeostasis and metabolic reprogramming are important components of chronic epithelial injury and aberrant repair in IPF, with immune and inflammatory responses also playing key roles in the establishment and maintenance of the fibrotic microenvironment (34). Therefore, the IGFBP2-associated expression pattern may reflect stress-, metabolism-, and immune-related transcriptomic contexts within IPF tissue. However, because bulk transcriptomic signals are derived from whole-tissue samples and may be influenced by cellular composition, they should not be interpreted as direct evidence of cell-type-specific regulatory mechanisms.

These communication patterns suggest that injury/transitional AT2 cells may occupy a microenvironmental interface between epithelial injury, ECM remodeling, vascular responses, and immune regulation (35). Mesenchymal-derived ECM-related signals may contribute to a fibrotic niche that influences epithelial stress responses and impaired repair (36). This finding is supported by recent studies on spatially organized fibrotic niches and ECM-associated epithelial remodeling. In parallel, outgoing signals from injury/transitional AT2 cells toward endothelial and macrophage populations may reflect vascular remodeling and immune-inflammatory regulation, consistent with reported roles of endothelial cells and macrophage crosstalk in the pulmonary fibrotic microenvironment. These findings should not be interpreted as direct IGFBP2-mediated signaling. No robust IGFBP2-specific ligand-receptor pair was detected toward fibroblast/mesenchymal or macrophage populations, and the inferred IGFBP pathway signal was mainly represented by IGFBP3-TMEM219. Thus, the current analysis does not demonstrate that IGFBP2 directly acts on fibroblasts or macrophages or promotes collagen deposition. Instead, these findings suggest that IGFBP2 enrichment marks an injury/transitional AT2 state embedded in a broader epithelial-mesenchymal-immune communication program during fibrotic remodeling, rather than a single IGFBP2-driven ligand-receptor mechanism.

In this tissue-level context, the potential biological role of IGFBP2 in injury/transitional AT2 deserves closer consideration. At the single-cell level, IGFBP2 was enriched in injury/transitional AT2 cells and increased progressively along pseudotime, supporting a dynamic association with this cell state. The biological significance of this enrichment may be related to the transitional nature of this epithelial state, which may reflect a balance between attempted epithelial repair and maladaptive remodeling. Recent evidence also supports the relevance of KRT8-positive transitional AT2 cell accumulation in pulmonary fibrosis (37). In this setting, increased IGFBP2 expression may represent a state-associated epithelial response program linked to epithelial stress adaptation, epithelial-matrix interaction, metabolic adjustment, and tissue-repair signaling. Previous studies have shown that IGFBP2 is an important regulator of the IGF signaling pathway. By modulating IGF bioavailability and functionally interacting with the ECM, integrins, and other cell-surface signaling components, it may participate in multiple processes, including cell survival, migration, metabolic regulation, and tissue repair (38,39). In pulmonary disease and fibrosis-related settings, IGFBP2 has also been implicated in inflammatory responses, tissue remodeling, and fibrosis-associated processes, although its expression pattern and functional effects appear to depend on cell type and pathological context (40). Loss of IGFBP2 has been reported to promote AEC2 senescence and aggravate experimental pulmonary fibrosis, while restoration of IGFBP2 appears to confer partial protection (41). These observations suggest that the role of IGFBP2 in pulmonary fibrosis is unlikely to be explained by a simple pro-fibrotic or protective classification. In our study, IGFBP2 may be better understood as part of a state-associated response program, rather than as a passive marker of fibrosis alone. However, given that the injury/transitional state may itself encompass distinct biological implications, including adaptive repair and pathological persistence, this interpretation still requires further functional validation (42).

Virtual perturbation analysis further suggested that IGFBP2 may be associated with specific functional programs within injury/transitional AT2 cells. When mapped back to the clean-AT2 state space, the knockout-inferred genes showed preferential association with the injury/transitional AT2 state. This pattern was most evident for SCGB1A1, SCGB3A1, and BPIFB1. By comparison, SLPI showed a broader epithelial stress/barrier-associated pattern, and PIGR displayed greater heterogeneity. These findings support the interpretation that IGFBP2-associated perturbation is linked to a restricted epithelial secretory/barrier-related program associated with injury/transitional AT2 remodeling (43-45), rather than to a diffuse transcriptome-wide shift. Previous studies have shown that secretory and barrier-associated epithelial molecules, including SCGB family proteins and SLPI, are closely involved in immune regulation, barrier function, and inflammatory control in the airway epithelium (46). These findings may also be viewed alongside the inferred ECM remodeling, inflammatory signaling, and vascular-related communication patterns identified by CellChat. These findings do not indicate direct mediation of specific ligand–receptor interactions. Instead, they suggest that altered microenvironmental interactions of injury/transitional AT2 cells may occur in parallel with remodeling of AT2-intrinsic secretory, barrier, and inflammatory-response programs associated with IGFBP2. Because these findings were inferred from a single-cell regulatory network, the specific relationships between the affected genes and the corresponding functional modules still require further experimental validation.

In addition to the transcriptomic findings, the potential small-molecule accessibility of IGFBP2 was also preliminarily explored at the structural level. The virtual screening results suggested that IGFBP2 may contain an accessible binding pocket. The relatively high predicted toxicity of the initially top-ranked candidate indicated that docking performance alone was insufficient for compound prioritization. After structural optimization, a more favorable balance was achieved between predicted toxicity and computational binding performance. Docking and molecular dynamics analyses further suggested that the optimized compound could be accommodated within the predicted pocket and retain a certain degree of conformational stability under the simulated conditions. Overall, this part of the study should be regarded primarily as an exploratory assessment of potential small-molecule interaction with IGFBP2, providing only preliminary computational clues to its possible structural accessibility. Given that the precise functional role of IGFBP2 in this context remains incompletely understood, these observations should not be interpreted as direct support for therapeutic development. Further biochemical and functional studies are required before any translational implications can be considered.

Integrated multi-layer transcriptomic analyses revealed potential links among IGFBP2, the injury/transitional AT2 state, and microenvironmental interactions. Nevertheless, several limitations should be acknowledged. This study relied primarily on computational analyses of public transcriptomic datasets, and no dedicated peri-anesthetic exposure cohort or experimental model was included. The PSRG framework should be interpreted as a literature-derived prior framework for candidate prioritization rather than as direct exposure-derived transcriptomic evidence. Differences in sample sources and technical platforms across datasets may have influenced the results. Cell-cell communication, virtual perturbation, and structural analyses were inference-based and therefore require further experimental validation. In addition, IGFBP2 protein-level expression and spatial localization in IPF patient tissues were not validated by immunostaining, spatial transcriptomics, or proteomic assays. The robustness of single-cell state annotation, pseudotime analysis, and microenvironmental interaction patterns should be further assessed in independent cohorts and experimental systems. Although IGFBP2 was associated with injury/transitional AT2-related transcriptional programs and microenvironmental alterations in IPF, its precise functional role remains unclear. Future studies integrating independent validation with in vitro and in vivo experiments are needed to clarify its role in AT2 injury-associated remodeling and fibrotic microenvironmental change.

Conclusions

This study identified IGFBP2 as a core candidate gene in IPF within a PSRG framework through integrated bulk and single-cell transcriptomic analyses. IGFBP2 was enriched in injury/transitional AT2 cells and increased along pseudotime, supporting its association with dynamic AT2 state remodeling. Cell-cell communication analysis suggested altered microenvironmental communication patterns associated with injury/transitional AT2 cells, whereas scTenifoldKnk-based virtual perturbation suggested a restricted epithelial secretory/barrier-related regulatory program associated with IGFBP2. Overall, these findings support further investigation of the role of IGFBP2 in injury/transitional AT2 remodeling in IPF, although experimental validation remains necessary.

Supplementary

The article’s supplementary files as

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments.