Exposure-informed lung transcriptomic analysis links predicted NNK targets to cell type-specific remodeling programs in idiopathic pulmonary fibrosis Authors: Kairui Meng, Lu Wang, Xueqing Gong, Wenjun Tang, Guobing Jia, Yanmei Wang, Chengshi He Published: 2026 ABSTRACT Idiopathic pulmonary fibrosis (IPF) is associated with cigarette smoking, yet the relationship between the tobacco-specific nitrosamine nicotine-derived nitrosamine ketone (NNK) and IPF-associated lung transcriptional remodeling remains incompletely understood. Here, we developed an exposure-informed computational framework integrating multi-database target prediction, lung single-cell and single-nucleus transcriptomic analysis, co-expression network analysis, bulk lung cohort projection, and exploratory structure-based modeling. Putative human protein targets of NNK were predicted from ChEMBL, PharmMapper, and SwissTargetPrediction, yielding 2,505 nonredundant targets. These targets were intersected with IPF-associated intramodular hub genes identified from cell type-specific weighted gene co-expression network analysis, defining a focused 42-gene ExposureA-core gene set. Projected ExposureA-core scores showed the clearest IPF-control differences in endothelial and epithelial pseudo-bulk profiles. Functional annotation of the training-derived epithelial ExposureA-core Top30 signature highlighted MAPK and p38 MAPK signaling, PI3K-AKT signaling, angiogenesis or vasculature regulation, cell–substrate adhesion, and membrane-, adhesion-, and cytoskeleton-related cellular components, suggesting remodeling- and adhesion-related epithelial transcriptional features in IPF. The fixed epithelial ExposureA-core Top30 signature remained detectable in independent bulk lung transcriptomic cohorts without gene re-selection, coefficient fitting, or score optimization, with exploratory ROC analyses showing apparent IPF-control separation in GSE110147 and GSE92592. Exploratory docking and 100-ns molecular dynamics simulations of selected epithelial Top30-encoded candidates showed that modeled ECE1–NNK, MMP7–NNK, and TGM2–NNK complexes reached dynamic equilibrium, with relatively stable RMSD, radius of gyration, solvent-accessible surface area, residue-level fluctuation, and low-energy conformational states, supporting their structural plausibility as candidate modeled complexes. Overall, this study defines a focused exposure-informed IPF-associated transcriptional framework and prioritizes epithelial remodeling-related candidate features for future experimental validation. These findings support hypothesis-generating computational prioritization rather than direct evidence that NNK drives IPF pathogenesis.Supplementary InformationThe online version contains supplementary material available at 10.1007/s10142-026-01980-3. Introduction Idiopathic pulmonary fibrosis (IPF) is a fatal interstitial lung disease of unknown etiology, characterized by irreversible progression of lung fibrosis and poor prognosis (Harada et al. 2025). The median survival after diagnosis is only 3 to 5 years, although outcomes vary substantially among patients (Bonella et al. 2023, Podolanczuk et al. 2023). Patients typically present with progressive dyspnea, restrictive ventilatory impairment, and reduced diffusing capacity, with many ultimately requiring lung transplantation or succumbing to respiratory failure. The clinical course is highly heterogeneous, and currently available clinical and physiological indices capture this variability only partially, underscoring the need for biologically informed molecular features that more accurately reflect the underlying pathogenic processes (Bonella et al. 2023). Environmental and lifestyle exposures are increasingly recognized as important determinants of IPF risk and disease progression. Among these, cigarette smoking is one of the most consistently identified risk factors. Smoking-related molecular injury is thought to intersect with age-related susceptibility and may contribute to chronic epithelial damage, aberrant tissue repair, and excessive fibrogenesis (Bae et al. 2022, Andersson et al. 2021, Moss et al. 2022). However, cigarette smoke is a complex mixture containing thousands of chemicals, and the specific components and molecular targets linking smoking-related exposure to fibrotic lung remodeling remain incompletely defined. This knowledge gap limits etiologic interpretation and the development of exposure-informed strategies for identifying high-risk biological pathways. Nicotine-derived nitrosamine ketone (NNK; 4-(methylnitrosamino)-1-(3-pyridyl)-1-butanone) is a prototypical tobacco-specific nitrosamine generated during tobacco curing and combustion. NNK is one of the most abundant and potent tobacco-specific nitrosamines in cigarette smoke and smokeless tobacco products and has been classified by the International Agency for Research on Cancer as carcinogenic to humans (Akopyan and Bonavida 2006, Smokeless 2007). Experimental studies have shown that NNK and its major metabolite NNAL are pulmonary carcinogens capable of inducing genetic and epigenetic alterations and promoting lung tumorigenesis in multiple model systems (Boo et al. 2023, Aggarwal et al. 2025). Beyond cancer, NNK exposure has been linked to inflammatory and immune dysregulation, but its relationship with chronic fibrotic lung diseases, including IPF, remains poorly explored (Yalcin and Monte 2016, Shim and Lim 2025). In particular, it remains unclear whether predicted NNK-related molecular targets converge with IPF-associated transcriptional programs in specific lung cell compartments. IPF pathogenesis involves complex, cell type-specific interactions among alveolar epithelial cells, fibroblasts, endothelial cells, and immune populations within the lung microenvironment. Single-cell RNA sequencing studies have revealed profound heterogeneity across these compartments, including aberrant epithelial cell states, activated fibroblast subsets, and dysfunctional immune cells associated with disease severity and progression (Habermann et al. 2020, Jin et al. 2024, Adams et al. 2020). However, most transcriptomic studies have focused on disease-associated cell states or gene signatures without explicitly integrating information about upstream environmental exposures or predicted exposure-related molecular targets. As a result, the extent to which tobacco-derived toxicants such as NNK overlap with cell type-specific IPF transcriptional networks remains insufficiently understood. Advances in systems toxicology and network biology provide a framework for addressing this gap. Large-scale target prediction resources, including ChEMBL, PharmMapper, and SwissTargetPrediction, enable systematic inference of putative protein targets for small molecules such as NNK across the human proteome (Wang et al. 2017, Daina et al. 2019, Mendez et al. 2019). When combined with patient-derived lung transcriptomes and co-expression network analysis, these predicted targets can be intersected with IPF-associated network hubs to identify focused exposure–disease overlap gene sets enriched for both disease relevance and network centrality (Taboureau et al. 2020, Audouze et al. 2010). In parallel, independent transcriptomic projection can assess whether a training-derived signal remains detectable across external tissue datasets, while exploratory structure-based modeling can provide structural context for prioritized candidate proteins without establishing direct binding or functional regulation (Huang et al. 2025, Ganesan et al. 2017, Liu et al. 2018). In this study, we developed an exposure-informed computational framework to examine the relationship between predicted NNK targets and IPF-associated lung transcriptional remodeling. We first predicted putative human protein targets of NNK using multiple databases and then integrated these targets with cell type-specific IPF co-expression networks derived from lung single-cell and single-nucleus transcriptomic datasets. This strategy defined a focused ExposureA-core gene set and prioritized epithelial remodeling-related features for downstream functional annotation, independent bulk-cohort projection, and exploratory structure-based modeling. By combining exposure-informed target prediction with lung cell type-resolved transcriptomics, this study provides a hypothesis-generating framework for prioritizing IPF-associated transcriptional features and candidate epithelial proteins for future experimental validation. Methods Study design This study was designed as an exposure-informed computational analysis integrating chemogenomic target prediction, lung single-cell transcriptomics, bulk lung transcriptomics, and structure-based modeling. Training single-cell cohorts were used to define IPF-associated co-expression modules, construct the ExposureA-core signature, and derive a fixed epithelial ExposureA-core Top30 signature. Independent bulk lung cohorts were used to assess the external detectability of the fixed epithelial signature. The overall analytical workflow is summarized in Fig. 1. Fig. 1Exposure-informed analytical framework for identifying and prioritizing transcriptional features linked to predicted NNK targets in idiopathic pulmonary fibrosis. The workflow comprised three analytical lanes. Lane 1, discovery and convergence: (A1) Predicted human targets of 4-(methylnitrosamino)-1-(3-pyridyl)-1-butanone (NNK) were integrated from ChEMBL, PharmMapper, and SwissTargetPrediction, yielding 2, 505 nonredundant predicted targets. Because smoking status and pack-year data were inconsistently available across cohorts, the framework represents computational convergence with predicted NNK targets rather than measured NNK exposure. (B1) Lung single-cell and single-nucleus RNA-sequencing datasets (GSE279404, GSE128033, and GSE283885) were annotated and summarized as sample-by-cell-type pseudo-bulk profiles. (B2) Cell type-specific weighted gene co-expression network analysis identified IPF-associated modules (|R| ≥ 0.50, FDR < 0.05) and intramodular hub genes (|MM| ≥ 0.70, |GS| ≥ 0.30). (C1) The intersection of predicted NNK targets with IPF-associated intramodular hub genes defined the 42-gene ExposureA-core. Lane 2, cell-type activity and epithelial prioritization: (D1) Single-sample gene set enrichment analysis characterized ExposureA-core activity across cell types, with the clearest IPF–control differences in endothelial and epithelial profiles. (D2) A fixed, training-derived epithelial Top30 signature was generated using expression-detection filtering and ranking weighted by effect size and statistical significance. (D3) Gene Ontology and Hallmark enrichment analyses provided supportive functional interpretation and did not contribute to ExposureA-core definition. Lane 3, external projection and structural plausibility: (E1) The fixed epithelial Top30 signed score was projected into the external bulk lung cohorts GSE110147 and GSE92592 without gene re-selection, coefficient fitting, or score optimization. (F1) Exploratory NNK docking and 100-ns molecular dynamics simulations of the ECE1–NNK, MMP7–NNK, and TGM2–NNK complexes assessed structural plausibility only. (H1) The integrated workflow prioritized epithelial remodeling-related candidate features and generated hypotheses requiring experimental validation. Solid arrows indicate the primary analytical flow, whereas dashed arrows indicate supportive connections. Colors distinguish the major analytical components. Pseudo-bulk differential-expression analysis contextualized IPF-associated transcriptional remodeling but was not used to define the ExposureA-core. FDR, false discovery rate; GO, Gene Ontology; GS, gene significance; IPF, idiopathic pulmonary fibrosis; MD, molecular dynamics; MM, module membership; NNK, 4-(methylnitrosamino)-1-(3-pyridyl)-1-butanone; ssGSEA, single-sample gene set enrichment analysis; WGCNA, weighted gene co-expression network analysis Exposure-informed analytical framework for identifying and prioritizing transcriptional features linked to predicted NNK targets in idiopathic pulmonary fibrosis. The workflow comprised three analytical lanes. Lane 1, discovery and convergence: (A1) Predicted human targets of 4-(methylnitrosamino)-1-(3-pyridyl)-1-butanone (NNK) were integrated from ChEMBL, PharmMapper, and SwissTargetPrediction, yielding 2, 505 nonredundant predicted targets. Because smoking status and pack-year data were inconsistently available across cohorts, the framework represents computational convergence with predicted NNK targets rather than measured NNK exposure. (B1) Lung single-cell and single-nucleus RNA-sequencing datasets (GSE279404, GSE128033, and GSE283885) were annotated and summarized as sample-by-cell-type pseudo-bulk profiles. (B2) Cell type-specific weighted gene co-expression network analysis identified IPF-associated modules (|R| ≥ 0.50, FDR < 0.05) and intramodular hub genes (|MM| ≥ 0.70, |GS| ≥ 0.30). (C1) The intersection of predicted NNK targets with IPF-associated intramodular hub genes defined the 42-gene ExposureA-core. Lane 2, cell-type activity and epithelial prioritization: (D1) Single-sample gene set enrichment analysis characterized ExposureA-core activity across cell types, with the clearest IPF–control differences in endothelial and epithelial profiles. (D2) A fixed, training-derived epithelial Top30 signature was generated using expression-detection filtering and ranking weighted by effect size and statistical significance. (D3) Gene Ontology and Hallmark enrichment analyses provided supportive functional interpretation and did not contribute to ExposureA-core definition. Lane 3, external projection and structural plausibility: (E1) The fixed epithelial Top30 signed score was projected into the external bulk lung cohorts GSE110147 and GSE92592 without gene re-selection, coefficient fitting, or score optimization. (F1) Exploratory NNK docking and 100-ns molecular dynamics simulations of the ECE1–NNK, MMP7–NNK, and TGM2–NNK complexes assessed structural plausibility only. (H1) The integrated workflow prioritized epithelial remodeling-related candidate features and generated hypotheses requiring experimental validation. Solid arrows indicate the primary analytical flow, whereas dashed arrows indicate supportive connections. Colors distinguish the major analytical components. Pseudo-bulk differential-expression analysis contextualized IPF-associated transcriptional remodeling but was not used to define the ExposureA-core. FDR, false discovery rate; GO, Gene Ontology; GS, gene significance; IPF, idiopathic pulmonary fibrosis; MD, molecular dynamics; MM, module membership; NNK, 4-(methylnitrosamino)-1-(3-pyridyl)-1-butanone; ssGSEA, single-sample gene set enrichment analysis; WGCNA, weighted gene co-expression network analysis Smoking metadata assessment We reviewed publicly available GEO series matrix files, sample annotations, available SRA metadata, and supplementary materials for smoking-related variables, including smoking status, tobacco exposure, cigarette use, and pack-year information. Because these variables were not consistently reported across the included single-cell and single-nucleus cohorts, smoking-stratified analyses could not be performed. Accordingly, ExposureA-core was interpreted as a computationally inferred convergence between predicted NNK targets and IPF-associated lung transcriptional network hubs, rather than as a measured NNK exposure signature. Dataset-level metadata and analytical roles are summarized in Supplementary Table S1. IPF transcriptomic cohorts IPF-related transcriptomic datasets were downloaded from NCBI GEO (Edgar et al. 2002, Barrett et al. 2013), with original study records and associated publications reviewed where available. Datasets were screened by tissue source, sequencing modality, disease annotation, smoking metadata availability, and suitability for cell type-resolved or bulk lung transcriptomic analysis. The main single-cell framework was restricted to human lung-derived samples or lung-resident cell populations; peripheral blood and other non-lung datasets were reviewed for metadata completeness but excluded from the main lung-cell analyses. The training single-cell set included harmonized pseudo-bulk matrices from GSE279404, GSE128033, and GSE283885, retaining samples with complete availability across the required cell type-specific matrices. GSE110147 and GSE92592 served as the main independent bulk lung cohorts for projection of the fixed epithelial ExposureA-core Top30 signature, while GSE32537 and GSE24206 were retained as exploratory sensitivity cohorts. Dataset metadata, analytical roles, and pseudo-bulk availability are summarized in Supplementary Tables S1 and S3. Acquisition of NNK properties and predicted targets NNK physicochemical information and its canonical two-dimensional structure were retrieved from PubChem (Kim et al. 2016). Candidate human protein targets were predicted using ChEMBL (Mendez et al. 2019), SwissTargetPrediction (Daina et al. 2019), and PharmMapper (Wang et al. 2017) with queries restricted to Homo sapiens. Nonredundant predicted targets were merged for integration with IPF-associated transcriptomic networks and are listed in Supplementary Table S4. Single-cell processing and pseudo-bulk construction Raw or processed single-cell and single-nucleus RNA-seq count matrices from eligible cohorts were processed using Seurat and SingleCellExperiment (Stuart et al. 2019, Amezquita et al. 2020) workflows. Cell-level quality control was performed before normalization and annotation. Cells with low feature complexity, extreme feature counts, or high mitochondrial transcript proportions were excluded according to dataset-specific QC summaries. When explicit doublet annotations were available from the original study or accompanying metadata, annotated doublets were removed; otherwise, doublet removal was not inferred de novo. For cell type annotation, count matrices were converted to SingleCellExperiment (Amezquita et al. 2020) objects and log-normalized using logNormCounts. Highly variable genes were selected using scran, and the top 1,000 highly variable genes were used for dimensionality reduction (Lun et al. 2016). Principal component analysis was performed using irlba, retaining 30 principal components. A cosine-distance Annoy K-nearest-neighbor graph was constructed with k = 10, followed by Louvain clustering. Cluster-level cell type annotation was performed using SingleR (Aran et al. 2019) with the Human Primary Cell Atlas reference, and pruned cluster labels were propagated to individual cells where available. Pseudo-bulk expression profiles were generated by summing raw counts within each donor–cell type combination to preserve donor-level biological replication for downstream differential-expression analysis (Squair et al. 2021). Donor–cell type combinations with fewer than 20 cells were excluded to reduce instability from sparse pseudo-bulk profiles. The resulting gene-by-sample pseudo-bulk matrices were accompanied by sample-level annotations, including sample identifier, disease condition, dataset, batch, and cell type. Dataset-level QC settings, cell type availability, SingleR annotation summaries, and pseudo-bulk sample distributions are provided in Supplementary Tables S2 and S3. Differential expression in pseudo-bulk data For each eligible cell type, pseudo-bulk profiles were normalized using TMM in edgeR (Robinson et al. 2010) and transformed using voom (Law et al. 2014). Within each dataset, IPF-control differential expression was estimated using limma linear models with empirical Bayes moderation (Ritchie et al. 2015). When multiple datasets were available for a given cell type, shared-gene results were combined using fixed-effect inverse-variance meta-analysis. Differential-expression results included log fold changes, test statistics, P values, and Benjamini–Hochberg false discovery rates (Benjamini and Hochberg 1995). These analyses were used to describe cell type-specific IPF-associated transcriptional remodeling and were not used as the primary criterion for defining ExposureA-core or selecting downstream model features. The epithelial ExposureA-core Top30 signature was selected separately within epithelial ExposureA-core genes, as described below. Summary results are provided in Supplementary Table S5. WGCNA and definition of ExposureA-core In the training set, WGCNA was performed separately for each eligible lung cell type using pseudo-bulk voom-transformed log2 counts per million matrices (Langfelder and Horvath 2008). Up to 7,000 highly variable genes expressed in at least 10% of samples were retained for network construction. Signed co-expression networks were constructed using biweight midcorrelation-based adjacency, with soft-thresholding powers selected according to scale-free topology criteria, and modules were identified using blockwiseModules. Module eigengenes were correlated with IPF status, and strongly IPF-associated modules were defined as those with an absolute module eigengene correlation of |R| ≥ 0.50 and a Benjamini–Hochberg false discovery rate < 0.05 (Benjamini and Hochberg 1995). Within strongly IPF-associated modules, intramodular hub genes were defined as genes with absolute module membership ≥ 0.70 and absolute gene significance ≥ 0.30. ExposureA-core was then defined as the intersection between predicted NNK targets and IPF-associated intramodular hub genes. Genes that did not meet both the module-level and hub-gene criteria were not included in the main ExposureA-core definition. The resulting ExposureA-core gene set was used for cell type-resolved ssGSEA scoring and epithelial signature derivation. WGCNA hub-gene filtering results and the final ExposureA-core gene list are provided in Supplementary Tables S6 and S7. Cell type-resolved scoring of ExposureA-core ExposureA-core ssGSEA scores were calculated in eligible lung cell types using GSVA-based single-sample GSEA scoring (Hänzelmann et al. 2013), with only detected ExposureA-core genes in each corresponding pseudo-bulk matrix included. Cell types with fewer than three detected ExposureA-core genes were excluded from visualization and group comparison. IPF-control differences were assessed using Wilcoxon rank-sum tests with Benjamini–Hochberg false discovery rate correction (Benjamini and Hochberg 1995). Supportive exploratory feature-space analysis Supportive exploratory feature-space analyses were performed to assess whether the training-derived ExposureA-core features showed non-random IPF-control separability. Penalized logistic-regression specifications, label-permutation testing, repeated nested cross-validation, and sparse-model feature-stability summaries were used only to characterize the training-derived feature space (Friedman et al. 2010, Varma and Simon 2006). Detailed procedures and summary outputs are provided in the Supplementary Methods SM4, Supplementary Table S9, and Supplementary Figure S1. These analyses were considered supportive only and were not used for diagnostic classifier validation, external model evaluation, candidate selection, or downstream biological interpretation. Derivation of the training-derived epithelial ExposureA-core Top30 signature Candidate genes were restricted to ExposureA-core genes detected in the training epithelial pseudo-bulk matrix. Genes were retained if they were detected in at least 70% of training samples and had expression values available in at least two IPF and two control samples. For each retained gene, the mean IPF-control expression difference, direction of change, Wilcoxon rank-sum P value, and Benjamini–Hochberg false discovery rate (Benjamini and Hochberg 1995) were calculated. Retained genes were ranked using a prespecified effect-size- and significance-weighted score, defined as the absolute mean IPF-control expression difference multiplied by −log10(P value). The top 30 genes were selected as the epithelial ExposureA-core Top30 signature and assigned as up-in-IPF or down-in-IPF according to the training-set expression direction. After ExposureA-core restriction, no additional hard effect-size or false discovery rate cutoff was imposed, and genes were not prioritized by expression abundance alone. Functional enrichment of epithelial and endothelial ExposureA-core signatures Gene Ontology Biological Process, Molecular Function, and Cellular Component enrichment analyses were performed using clusterProfiler with Gene Ontology annotations to provide functional context for the training-derived epithelial ExposureA-core Top30 signature and endothelial ExposureA-core genes (Yu et al. 2012, The 2021). For the epithelial analysis, the Top30 genes were used as the input gene set. For the endothelial analysis, endothelial-detected ExposureA-core genes were used as the input gene set. P values were adjusted using the Benjamini–Hochberg method (Benjamini and Hochberg 1995). Enriched terms were summarized to support biological interpretation of cell type-specific ExposureA-core programs. Epithelial enrichment results are provided in Supplementary Table S11 and Supplementary Figure S7, and supportive endothelial enrichment results are shown in Supplementary Table S12 and Supplementary Figure S2. Exploratory Hallmark GSEA was additionally performed using MSigDB Hallmark gene sets to provide pathway-level context for ExposureA-core-associated transcriptional remodeling in macrophage, epithelial, and endothelial pseudo-bulk profiles (Liberzon et al. 2015). These analyses were used only for supportive biological interpretation and were not used to define ExposureA-core, derive the epithelial Top30 signature, fit predictive models, or prioritize structure-based candidates. Detailed procedures and results are provided in Supplementary Methods SM6, Supplementary Table S17, and Supplementary Figures S4–S6. Molecular docking and molecular dynamics simulation Structure-based analyses were performed as exploratory assessments of the structural plausibility of modeled NNK–protein complexes with selected candidate proteins from the training-derived epithelial ExposureA-core Top30 signature. Candidates were prioritized based on signature membership, ranking evidence, IPF-associated direction of change, structural availability, and feasibility for docking and molecular dynamics simulation. NNK was retrieved from PubChem (Kim et al. 2016) and energy-minimized before docking. Candidate protein structures were obtained from the Protein Data Bank (Berman et al. 2000) or curated structural resources (UniProt: 2019), when available, and were prepared by removing crystallographic waters where applicable and adding hydrogens before docking with NNK using AutoDock Vina (Trott and Olson 2010). Top-ranked docking poses were carried forward for molecular dynamics simulation when applicable. Molecular dynamics simulations were performed in GROMACS 2022 (Abraham et al. 2015) using the CHARMM36/CHARMM36m force field (Huang et al. 2017), with CGenFF (Vanommeslaeghe et al. 2010) parameters assigned to NNK. Protein–NNK complexes were solvated, energy-minimized, equilibrated under NVT and NPT ensembles, and simulated for 100 ns under NPT conditions. Complex stability was assessed using root mean square deviation, root mean square fluctuation, radius of gyration, solvent-accessible surface area, intermolecular hydrogen bonds, and free energy landscape analysis. External bulk-cohort projection of the training-derived epithelial ExposureA-core Top30 signature The fixed training-derived epithelial ExposureA-core Top30 signature was projected into independent bulk lung cohorts without gene re-selection, coefficient fitting, or score optimization. GSE110147 and GSE92592 were used as the main bulk-projection datasets. For GSE110147, normalized microarray data were annotated to gene symbols and collapsed by averaging probes mapping to the same gene. For GSE92592, gene-level counts were processed using edgeR, filtered for low expression, normalized using TMM, and transformed to log2 counts per million (Robinson et al. 2010). For each cohort, matched Top30 genes were standardized to z scores across samples. A signed signature score was calculated as the mean z score of up-in-IPF genes minus the mean z score of down-in-IPF genes. IPF-control differences were assessed using two-sided Wilcoxon rank-sum tests. Exploratory ROC analysis was performed using the fixed score, with AUC values and 95% confidence intervals estimated by the DeLong method using pROC (DeLong et al. 1988, Robin et al. 2011). The same scoring procedure was applied to GSE32537 and GSE24206 as exploratory sensitivity analyses. Results Computational prediction of NNK protein targets for IPF network-overlap analysis The molecular structure of NNK was retrieved from PubChem (Fig. 2A). Candidate human protein targets of NNK were predicted using three complementary resources—ChEMBL, PharmMapper, and SwissTargetPrediction—and nonredundant predicted targets were merged across databases. This strategy identified 2,505 unique predicted targets (Fig. 2B; Supplementary Table S4). These targets were then intersected with IPF-associated intramodular hub genes from cell type-specific co-expression analyses to define the ExposureA-core gene set for downstream analyses. IPF-associated transcriptional remodeling across four analyzable lung cell types Using pseudo-bulk expression matrices across 12 lung cell types, we assessed IPF-associated transcriptional remodeling at cell type resolution. Four cell types had sufficient IPF-control representation for cell type-specific differential-expression analysis: endothelial cells, epithelial cells, macrophages, and tissue stem/progenitor cells. After false discovery rate correction at adjusted P < 0.05, 283 differentially expressed genes were identified in endothelial cells, 2,228 in epithelial cells, 2,510 in macrophages, and 5 in tissue stem/progenitor cells. Macrophages and epithelial cells showed the largest numbers of detected IPF-associated differentially expressed genes, whereas tissue stem/progenitor cells showed only minimal differential-expression signal. This analysis provided a descriptive cell type-resolved map of IPF-associated transcriptional alterations and contextualized the subsequent network-based ExposureA-core analysis (Fig. 2C–F; Supplementary Table S5). Fig. 2NNK structure, predicted target overlap, and IPF-associated cell type-specific transcriptional remodeling. (A) Chemical structure of NNK. (B) Venn diagram illustrating the overlap of predicted NNK protein targets obtained from ChEMBL, PharmMapper, and SwissTargetPrediction. Numbers and percentages indicate unique and shared predicted targets across databases. (C) Volcano plot showing IPF-associated differential gene expression in macrophages based on pseudo-bulk analysis. Significantly upregulated genes in IPF versus control samples are shown in red, significantly downregulated genes are shown in blue, and non-significant genes are shown in gray. Significance was defined as an adjusted P value < 0.05, with the direction of change determined by log₂ fold change. Dashed horizontal lines indicate the adjusted P-value threshold, and dashed vertical lines indicate log₂ fold-change reference lines. The x-axis represents log₂ fold change, and the y-axis represents − log₁₀ adjusted P value. (D) Volcano plot for tissue stem/progenitor cells, as described in (C). (E) Volcano plot for epithelial cells, as described in (C). (F) Volcano plot for endothelial cells, as described in (C). IPF, idiopathic pulmonary fibrosis; NNK, nicotine-derived nitrosamine ketone NNK structure, predicted target overlap, and IPF-associated cell type-specific transcriptional remodeling. (A) Chemical structure of NNK. (B) Venn diagram illustrating the overlap of predicted NNK protein targets obtained from ChEMBL, PharmMapper, and SwissTargetPrediction. Numbers and percentages indicate unique and shared predicted targets across databases. (C) Volcano plot showing IPF-associated differential gene expression in macrophages based on pseudo-bulk analysis. Significantly upregulated genes in IPF versus control samples are shown in red, significantly downregulated genes are shown in blue, and non-significant genes are shown in gray. Significance was defined as an adjusted P value < 0.05, with the direction of change determined by log₂ fold change. Dashed horizontal lines indicate the adjusted P-value threshold, and dashed vertical lines indicate log₂ fold-change reference lines. The x-axis represents log₂ fold change, and the y-axis represents − log₁₀ adjusted P value. (D) Volcano plot for tissue stem/progenitor cells, as described in (C). (E) Volcano plot for epithelial cells, as described in (C). (F) Volcano plot for endothelial cells, as described in (C). IPF, idiopathic pulmonary fibrosis; NNK, nicotine-derived nitrosamine ketone Co-expression network analysis identifies IPF-associated hub genes in lung cell types retained by stringent criteria Among the lung cell types evaluated, weighted gene co-expression network analysis identified strongly IPF-associated modules in endothelial cells, epithelial cells, and macrophages under the prespecified criteria of |R| ≥ 0.50 and FDR < 0.05. Two modules were retained in endothelial cells, whereas epithelial cells and macrophages each contributed one module (Fig. 3A; Supplementary Table S6). Hub-gene filtering within these modules identified 216 intramodular hub genes in endothelial cells, 60 in epithelial cells, and 2 in macrophages (Fig. 3B; Supplementary Table S6). These IPF-associated intramodular hub genes were subsequently intersected with predicted NNK targets to define the ExposureA-core gene set for downstream analyses (Supplementary Table S7). Fig. 3Cell type-specific co-expression network analysis identifies IPF-associated modules and intramodular hub genes. (A) Module eigengene correlations with IPF status in endothelial cells, epithelial cells, and macrophages. Each row represents a WGCNA co-expression module, and the color scale indicates the direction and magnitude of the correlation coefficient. Asterisks indicate modules meeting the prespecified criteria for strong IPF association (|R| ≥ 0.50 and FDR < 0.05). (B) Module membership (MM) versus gene significance (GS) plots for modules meeting the strong IPF-association criteria. Only modules marked by asterisks in (A) are shown. Each point represents a gene, colored by module. MM reflects intramodular connectivity, whereas GS reflects association with IPF status. Dashed vertical and horizontal lines indicate the prespecified hub-gene thresholds for |MM| and |GS|, respectively; genes with |MM| ≥ 0.70 and |GS| ≥ 0.30 were prioritized as IPF-associated intramodular hub genes Cell type-specific co-expression network analysis identifies IPF-associated modules and intramodular hub genes. (A) Module eigengene correlations with IPF status in endothelial cells, epithelial cells, and macrophages. Each row represents a WGCNA co-expression module, and the color scale indicates the direction and magnitude of the correlation coefficient. Asterisks indicate modules meeting the prespecified criteria for strong IPF association (|R| ≥ 0.50 and FDR < 0.05). (B) Module membership (MM) versus gene significance (GS) plots for modules meeting the strong IPF-association criteria. Only modules marked by asterisks in (A) are shown. Each point represents a gene, colored by module. MM reflects intramodular connectivity, whereas GS reflects association with IPF status. Dashed vertical and horizontal lines indicate the prespecified hub-gene thresholds for |MM| and |GS|, respectively; genes with |MM| ≥ 0.70 and |GS| ≥ 0.30 were prioritized as IPF-associated intramodular hub genes Cell type-specific activity of the ExposureA-core gene set in IPF lungs To define a focused computational exposure–disease overlap, predicted NNK targets were intersected with IPF-associated intramodular hub genes identified by WGCNA. This analysis identified a 42-gene ExposureA-core gene set that integrated predicted NNK target status, intramodular network centrality, and IPF association (Fig. 4A; Supplementary Table S7). ExposureA-core scores were quantified in cell type-specific pseudo-bulk profiles using ssGSEA. Among scored cell types, endothelial and epithelial cells showed the clearest IPF-control differences and were therefore displayed in the main cell type-resolved visualization. Group-wise comparison showed higher ExposureA-core ssGSEA scores in IPF samples in endothelial cells (FDR = 0.033, Δ = 0.42) and epithelial cells (FDR = 0.044, Δ = 0.42), where Δ represents the mean difference between IPF and control samples (Fig. 4B; Supplementary Table S8). Z-scored ssGSEA heatmaps also showed higher projected ExposureA-core scores in IPF samples in both endothelial and epithelial profiles (Fig. 4C). Overall, endothelial and epithelial cells showed the clearest projected ExposureA-core score differences among the scored cell types. Functional annotation of the epithelial ExposureA-core Top30 signature Given the projected ExposureA-core score difference observed in epithelial cells, we examined the functional context of the training-derived epithelial ExposureA-core Top30 signature. Gene Ontology biological process enrichment highlighted MAPK and p38 MAPK signaling, PI3K-AKT signaling, angiogenesis or vasculature regulation, and cell–substrate adhesion (Fig. 4D; Supplementary Table S11). Additional Gene Ontology molecular function and cellular component enrichment results are provided in Supplementary Table S11 and Supplementary Figure S7. Expression profiling of the epithelial ExposureA-core Top30 genes showed distinct distributional shifts between control and IPF epithelial pseudo-bulk samples, thereby defining the fixed epithelial signature used for subsequent external projection analyses (Fig. 4E; Supplementary Table S10). Fig. 4ExposureA-core construction, cell type-resolved projected scores, and epithelial Top30 signature characterization. (A) Venn diagram showing the overlap between predicted NNK protein targets and IPF-associated intramodular hub genes identified by cell type-specific WGCNA. The intersection defined the 42-gene ExposureA-core gene set used for downstream analyses. (B) Group-wise comparison of ExposureA-core ssGSEA scores between control and IPF samples in endothelial and epithelial cells. Points represent individual pseudo-bulk samples, and boxplots summarize the distribution within each condition. FDR values were calculated using Benjamini–Hochberg correction, and Δ indicates the mean difference in ExposureA-core ssGSEA score between IPF and control samples, calculated as mean(IPF) − mean(control). (C) Heatmap of z-scored ssGSEA scores showing projected ExposureA-core scores in endothelial and epithelial cell pseudo-bulk profiles. Rows represent cell types, and columns represent individual samples grouped by condition. The color scale indicates lower and higher relative ExposureA-core scores. (D) Gene Ontology biological process enrichment analysis of the training-derived epithelial ExposureA-core Top30 genes, highlighting terms related to MAPK/p38 MAPK signaling, PI3K-AKT signaling, angiogenesis or vasculature regulation, and cell–substrate adhesion. Dot size indicates gene count, and color indicates the Benjamini–Hochberg adjusted P value. (E) Expression distribution of the epithelial ExposureA-core Top30 genes in epithelial pseudo-bulk profiles from control and IPF samples. Each row represents one gene, and boxplots summarize scaled expression within each condition. FDR, false discovery rate; GO, Gene Ontology; IPF, idiopathic pulmonary fibrosis; NNK, nicotine-derived nitrosamine ketone; ssGSEA, single-sample gene set enrichment analysis; WGCNA, weighted gene co-expression network analysis ExposureA-core construction, cell type-resolved projected scores, and epithelial Top30 signature characterization. (A) Venn diagram showing the overlap between predicted NNK protein targets and IPF-associated intramodular hub genes identified by cell type-specific WGCNA. The intersection defined the 42-gene ExposureA-core gene set used for downstream analyses. (B) Group-wise comparison of ExposureA-core ssGSEA scores between control and IPF samples in endothelial and epithelial cells. Points represent individual pseudo-bulk samples, and boxplots summarize the distribution within each condition. FDR values were calculated using Benjamini–Hochberg correction, and Δ indicates the mean difference in ExposureA-core ssGSEA score between IPF and control samples, calculated as mean(IPF) − mean(control). (C) Heatmap of z-scored ssGSEA scores showing projected ExposureA-core scores in endothelial and epithelial cell pseudo-bulk profiles. Rows represent cell types, and columns represent individual samples grouped by condition. The color scale indicates lower and higher relative ExposureA-core scores. (D) Gene Ontology biological process enrichment analysis of the training-derived epithelial ExposureA-core Top30 genes, highlighting terms related to MAPK/p38 MAPK signaling, PI3K-AKT signaling, angiogenesis or vasculature regulation, and cell–substrate adhesion. Dot size indicates gene count, and color indicates the Benjamini–Hochberg adjusted P value. (E) Expression distribution of the epithelial ExposureA-core Top30 genes in epithelial pseudo-bulk profiles from control and IPF samples. Each row represents one gene, and boxplots summarize scaled expression within each condition. FDR, false discovery rate; GO, Gene Ontology; IPF, idiopathic pulmonary fibrosis; NNK, nicotine-derived nitrosamine ketone; ssGSEA, single-sample gene set enrichment analysis; WGCNA, weighted gene co-expression network analysis Molecular dynamics simulation of modeled ECE1–NNK, MMP7–NNK, and TGM2–NNK complexes To assess the dynamic behavior of selected modeled protein–ligand complexes, 100-ns molecular dynamics simulations were performed for modeled ECE1–NNK, MMP7–NNK, and TGM2–NNK complexes.The RMSD trajectories showed that the ECE1–NNK complex reached equilibrium after approximately 35 ns and fluctuated around 0.21 nm. The MMP7–NNK and TGM2–NNK complexes reached equilibrium after approximately 70 ns and 45 ns, with RMSD values stabilizing around 0.30 nm and 0.38 nm, respectively (Fig. 5A). The Rg curves remained relatively stable for the three complexes, indicating no marked global expansion or contraction during the simulations (Fig. 5B). RMSF analysis showed limited residue-level fluctuations across most regions, with only localized higher-flexibility peaks (Fig. 5C). SASA analysis also showed only minor fluctuations across the complexes (Fig. 5D). Free energy landscape analysis identified low-energy conformational basins for the ECE1–NNK, MMP7–NNK, and TGM2–NNK complexes, consistent with relatively stable conformational states during the simulations (Fig. 5E). Corresponding three-dimensional FEL surface plots are provided in Supplementary Figure S8, with summary results in Supplementary Tables S13 and S14. Overall, the MD simulations indicated that the ECE1–NNK, MMP7–NNK, and TGM2–NNK modeled complexes showed broadly stable conformational behavior during the simulations. ECE1–NNK and MMP7–NNK showed relatively stronger RMSD-based conformational stability, whereas TGM2–NNK also reached equilibrium but exhibited a slightly higher RMSD level. Fig. 5Exploratory molecular dynamics behavior of selected modeled NNK–protein complexes. (A) Root mean square deviation (RMSD) trajectories of modeled TGM2–NNK, MMP7–NNK, and ECE1–NNK complexes over 100 ns. (B) Radius of gyration (Rg) profiles showing global compactness of the three modeled complexes. (C) Root mean square fluctuation (RMSF) profiles showing residue-level flexibility for TGM2, MMP7, and ECE1. (D) Solvent-accessible surface area (SASA) trajectories showing changes in solvent exposure during the simulations. (E) Two-dimensional free energy landscape (FEL) projections for TGM2–NNK, MMP7–NNK, and ECE1–NNK based on principal components. Overall, the modeled complexes showed broadly stable conformational behavior after equilibration, supporting exploratory structural plausibility of these selected candidates. These results do not establish direct NNK binding, target engagement, or functional regulation. Abbreviations: ECE1, endothelin converting enzyme 1; FEL, free energy landscape; MD, molecular dynamics; MMP7, matrix metalloproteinase 7; NNK, nicotine-derived nitrosamine ketone; Rg, radius of gyration; RMSD, root mean square deviation; RMSF, root mean square fluctuation; SASA, solvent-accessible surface area; TGM2, transglutaminase 2 Exploratory molecular dynamics behavior of selected modeled NNK–protein complexes. (A) Root mean square deviation (RMSD) trajectories of modeled TGM2–NNK, MMP7–NNK, and ECE1–NNK complexes over 100 ns. (B) Radius of gyration (Rg) profiles showing global compactness of the three modeled complexes. (C) Root mean square fluctuation (RMSF) profiles showing residue-level flexibility for TGM2, MMP7, and ECE1. (D) Solvent-accessible surface area (SASA) trajectories showing changes in solvent exposure during the simulations. (E) Two-dimensional free energy landscape (FEL) projections for TGM2–NNK, MMP7–NNK, and ECE1–NNK based on principal components. Overall, the modeled complexes showed broadly stable conformational behavior after equilibration, supporting exploratory structural plausibility of these selected candidates. These results do not establish direct NNK binding, target engagement, or functional regulation. Abbreviations: ECE1, endothelin converting enzyme 1; FEL, free energy landscape; MD, molecular dynamics; MMP7, matrix metalloproteinase 7; NNK, nicotine-derived nitrosamine ketone; Rg, radius of gyration; RMSD, root mean square deviation; RMSF, root mean square fluctuation; SASA, solvent-accessible surface area; TGM2, transglutaminase 2 Bulk-cohort projection of the training-derived epithelial ExposureA-core signature Finally, we examined whether the epithelial ExposureA-core signature defined in the training set could be detected in independent bulk lung transcriptomic cohorts. Using the fixed training-derived epithelial ExposureA-core Top30 signature, we calculated a signed signature score in GSE110147 and GSE92592 without gene re-selection, coefficient fitting, or score optimization. In both cohorts, the fixed signature score was higher in IPF samples than in controls. In GSE110147, IPF samples showed higher signature scores than controls (n = 33 samples; Wilcoxon P = 6.59 × 10 − 5) (Fig. 6A). Exploratory ROC analysis showed apparent IPF-control separation, with an AUC of 0.934 (95% CI, 0.854–1.000) (Fig. 6B). Similar findings were observed in GSE92592, where the signature score was elevated in IPF samples (n = 39 samples; Wilcoxon P = 5.69 × 10 − 6) (Fig. 6C), with an AUC of 0.926 (95% CI, 0.851–1.000) (Fig. 6D). These findings indicate that the fixed training-derived epithelial ExposureA-core signature remained detectable in independent bulk lung transcriptomic data (Supplementary Table S15). As exploratory sensitivity analyses, the same fixed signature was additionally projected into GSE32537 and GSE24206. These cohorts showed weaker or less consistent IPF-control separation and are therefore reported in Supplementary Table S16 and Supplementary Figure S3. Fig. 6External bulk-cohort projection of the training-derived epithelial ExposureA-core Top30 signature in idiopathic pulmonary fibrosis cohorts. (A, C) Distribution of the signed z-score signature score in independent bulk lung transcriptomic cohorts. The fixed training-derived epithelial ExposureA-core Top30 signature was projected into bulk lung datasets without gene re-selection, coefficient fitting, or score optimization. Box-and-point plots summarize signature-score distributions between control and IPF lung tissue samples, with overlaid points representing individual samples. Group differences were assessed using the Wilcoxon rank-sum test. (A) GSE110147 cohort: control vs. IPF, n = 33 samples, P = 6.59 × 10⁻⁵. (C) GSE92592 cohort: control vs. IPF, n = 39 samples, P = 5.69 × 10⁻⁶. (B, D) Exploratory receiver operating characteristic analysis of the projected signature score. ROC curves show the apparent separation between control and IPF samples based on the fixed signature score. (B) GSE110147 cohort: area under the curve = 0.934, 95% CI, 0.854–1.000. (D) GSE92592 cohort: area under the curve = 0.926, 95% CI, 0.851–1.000. AUC, area under the curve; CI, confidence interval; IPF, idiopathic pulmonary fibrosis; ROC, receiver operating characteristic External bulk-cohort projection of the training-derived epithelial ExposureA-core Top30 signature in idiopathic pulmonary fibrosis cohorts. (A, C) Distribution of the signed z-score signature score in independent bulk lung transcriptomic cohorts. The fixed training-derived epithelial ExposureA-core Top30 signature was projected into bulk lung datasets without gene re-selection, coefficient fitting, or score optimization. Box-and-point plots summarize signature-score distributions between control and IPF lung tissue samples, with overlaid points representing individual samples. Group differences were assessed using the Wilcoxon rank-sum test. (A) GSE110147 cohort: control vs. IPF, n = 33 samples, P = 6.59 × 10⁻⁵. (C) GSE92592 cohort: control vs. IPF, n = 39 samples, P = 5.69 × 10⁻⁶. (B, D) Exploratory receiver operating characteristic analysis of the projected signature score. ROC curves show the apparent separation between control and IPF samples based on the fixed signature score. (B) GSE110147 cohort: area under the curve = 0.934, 95% CI, 0.854–1.000. (D) GSE92592 cohort: area under the curve = 0.926, 95% CI, 0.851–1.000. AUC, area under the curve; CI, confidence interval; IPF, idiopathic pulmonary fibrosis; ROC, receiver operating characteristic Discussion In this study, we developed an exposure-informed computational framework integrating multi-database target prediction, lung single-cell transcriptomic analysis, co-expression network analysis, bulk lung cohort projection, and exploratory structure-based modeling. By intersecting 2,505 predicted NNK protein targets with IPF-associated intramodular hub genes identified from cell type-specific WGCNA, we defined a focused 42-gene ExposureA-core gene set. This gene set represents a computationally inferred overlap between predicted exposure-related targets and IPF-associated lung transcriptional network hubs, rather than a measured NNK exposure signature. Projected ExposureA-core scores were most evident in endothelial and epithelial pseudo-bulk profiles. Functional annotation of the training-derived epithelial ExposureA-core Top30 signature highlighted MAPK and p38 MAPK signaling, PI3K-AKT signaling, angiogenesis or vasculature regulation, cell–substrate adhesion, and membrane-, adhesion-, and cytoskeleton-related cellular components. The fixed epithelial signature was also detectable in independent bulk lung transcriptomic cohorts without gene re-selection, coefficient fitting, or score optimization. Together, these findings prioritize exposure-informed IPF-associated transcriptional features and candidate epithelial features and proteins for future experimental validation, while remaining descriptive and hypothesis-generating rather than evidence that NNK exposure directly drives IPF pathogenesis. ExposureA-core gene set highlights IPF-associated programs in structural lung cell compartments Integrating predictions from ChEMBL, PharmMapper, and SwissTargetPrediction, we identified 2,505 nonredundant human protein targets of NNK and intersected them with IPF-associated intramodular hub genes identified under prespecified cell type-specific WGCNA criteria. This yielded a focused 42-gene ExposureA-core gene set, representing computational convergence between predicted NNK targets and IPF-associated network hubs. In line with prior single-cell studies showing coordinated disease-associated transcriptional remodeling across IPF lung cell populations (Habermann et al. 2020, Adams et al. 2020), and with toxicogenomics network approaches linking environmental chemicals to disease-related biological targets (Audouze et al. 2010), our findings support a network-level interpretation of exposure-informed IPF-associated transcriptional features. Projected ExposureA-core signal showed the clearest IPF-control differences in endothelial and epithelial pseudo-bulk profiles, consistent with the involvement of structural lung cell compartments in fibrotic remodeling (Mei et al. 2021, Confalonieri et al. 2022, Lu et al. 2025). These findings indicate where the training-defined ExposureA-core signal was most detectable, rather than demonstrating direct NNK sensitivity of these cell types. Epithelial cells were therefore prioritized for downstream Top30 signature derivation and functional annotation, given the clear projected epithelial signal and the established relevance of epithelial remodeling and repeated epithelial injury in IPF (Moss et al. 2022, Confalonieri et al. 2022). Epithelial ExposureA-core Top30 signature highlights remodeling- and adhesion-related programs in IPF At the functional annotation level, the training-derived epithelial ExposureA-core Top30 signature highlighted MAPK and p38 MAPK signaling, PI3K-AKT signaling, angiogenesis or vasculature regulation, cell–substrate adhesion, and membrane-, adhesion-, and cytoskeleton-related cellular components. These findings point to an epithelial remodeling- and adhesion-related program, but are not sufficient to define a broadly EMT-centered response. The prominence of the epithelial signal is consistent with the central role of repetitive epithelial injury in the initiation and progression of IPF (Confalonieri et al. 2022). Projected ExposureA-core scores were also evident in endothelial pseudo-bulk profiles, suggesting that epithelial and endothelial structural lung compartments were the main compartments in which this exposure-informed IPF-associated signal was detectable (Lu et al. 2025, Bian et al. 2023). Together, these patterns indicate coherent remodeling-related transcriptional features, while remaining descriptive rather than evidence of coordinated NNK-driven fibrotic progression. Biologically, the enrichment of MAPK/PI3K-AKT signaling, vascular regulation, adhesion, and cytoskeleton-related components provides a focused context for epithelial remodeling and cell–matrix interaction in IPF (Moss et al. 2022, Confalonieri et al. 2022). These findings are compatible with remodeling-associated epithelial transcriptional changes rather than a fully defined causal pathway. Although endothelial remodeling and barrier alteration are recognized contributors to the fibrotic lung microenvironment (Lu et al. 2025, Bian et al. 2023), the present analysis primarily supports an epithelial ExposureA-core Top30 context rather than a comprehensive pathway map across all lung cell types. Overall, these remodeling- and adhesion-related epithelial ExposureA-core features should be interpreted as hypothesis-generating candidates for future experimental validation, not as evidence of EMT activation, metabolic suppression, or NNK-driven pathway reprogramming in IPF. Exploratory structure-based modeling provides structural plausibility for selected epithelial candidate proteins As an exploratory extension of the epithelial ExposureA-core Top30 analysis, we performed molecular docking and 100-ns molecular dynamics simulations to assess whether selected Top30-encoded proteins could form computationally plausible modeled complexes with NNK. This analysis was designed to provide structural context for candidate proteins prioritized from the training-derived epithelial signature, rather than to validate direct biochemical binding or functional regulation. Docking and molecular dynamics simulations described modeled complex behavior, including root mean square deviation, radius of gyration, solvent-accessible surface area, residue-level fluctuation, free energy landscape features, and protein–ligand hydrogen-bond patterns. The molecular dynamics results were consistent with the structural plausibility of the ECE1–NNK, MMP7–NNK, and TGM2–NNK modeled complexes. All three complexes reached equilibrium during the simulations, with relatively stable RMSD trajectories, limited changes in radius of gyration and solvent-accessible surface area, and limited residue-level fluctuations across most regions. Free energy landscape analysis further suggested the presence of low-energy conformational basins for these complexes, consistent with relatively stable modeled conformational states. Among the three candidates, ECE1–NNK and MMP7–NNK showed relatively stronger RMSD-based conformational stability, whereas TGM2–NNK also stabilized but displayed a slightly higher RMSD level, indicating relatively greater conformational fluctuation. These structure-based findings should be interpreted as hypothesis-generating evidence rather than as biochemical proof of direct NNK–protein interaction. Molecular docking and molecular dynamics simulations can describe modeled binding poses and conformational behavior, but they cannot establish binding affinity, target engagement, exposure-response relationships, or downstream functional effects in lung epithelial cells (Ganesan et al. 2017, Liu et al. 2018). Therefore, the present results do not demonstrate that NNK directly disrupts epithelial homeostasis or initiates pro-fibrotic signaling through ECE1, MMP7, or TGM2. Nevertheless, the convergence of epithelial ExposureA-core membership, IPF-associated expression shifts, prior evidence linking MMP7 (Rosas et al. 2008, Pardo et al. 2016), TGM2 (Olsen et al. 2011, Fell et al. 2021), and the endothelin/ECE1 (Saleh et al. 1997, Swigris and Brown 2010) axis to IPF-related epithelial remodeling, extracellular-matrix regulation, or fibrotic signaling, and favorable modeled complex stability supports these proteins as structurally plausible candidates for future validation. Bulk-cohort projection supports external detectability of the fixed epithelial ExposureA-core signature An additional analysis of this work was the projection of a training-derived epithelial ExposureA-core Top30 signature into independent bulk lung transcriptomic cohorts. This analysis used a fixed signed z-score signature score without gene re-selection, coefficient fitting, or score optimization, and was designed to assess tissue-level detectability rather than diagnostic or clinical classification performance. This distinction is important given prior machine-learning studies reporting high-performing molecular classifiers in IPF (Ding et al. 2024, Guo et al. 2025, Li et al. 2023) and broader concerns regarding transparency and reproducibility in artificial intelligence-based biomedical prediction (Haibe-Kains et al. 2020). In both GSE110147 and GSE92592, the fixed epithelial signature score was higher in IPF samples than in controls, and exploratory ROC analysis showed apparent IPF-control separation. These observations suggest that the training-derived epithelial ExposureA-core signal can be detected across independent bulk lung datasets. However, because bulk tissue profiles lack explicit cell type resolution and are influenced by cellular composition, these results should be interpreted as evidence of external detectability rather than validation of a cell-resolved classifier or risk-stratification model (Aran et al. 2017). The cross-cohort preservation of this epithelial signal is nevertheless consistent with the idea that disease-associated transcriptional programs can remain detectable at the tissue-transcriptome level. Taken together, these findings support the cross-cohort detectability of the training-derived epithelial ExposureA-core Top30 signature as an IPF-associated transcriptional feature in independent bulk lung cohorts. They do not establish diagnostic utility, prognostic value, or clinical stratification performance. Future studies using larger independent cohorts with harmonized clinical phenotypes, smoking or exposure annotations, and ideally paired cell type-resolved data will be needed to determine whether exposure-informed epithelial signatures contribute to clinically relevant IPF subtyping or outcome prediction (Sadeleer et al. 2022). Strengths and limitations This study has several strengths. We integrated predicted NNK protein targets with lung cell type-specific IPF co-expression networks using prespecified WGCNA criteria, thereby refining a broad computational target space into a focused 42-gene ExposureA-core gene set. The main cell type-resolved framework was restricted to lung-derived single-cell and single-nucleus transcriptomic datasets, and the training-derived epithelial ExposureA-core Top30 signature was projected into independent bulk lung cohorts without gene re-selection, coefficient fitting, or score optimization. This design allowed us to assess whether an exposure-informed epithelial transcriptional signal remained detectable across datasets without relying on test-set-derived feature selection or diagnostic model validation. Several limitations should be considered. First, all transcriptomic analyses were based on observational public datasets, and residual confounding cannot be excluded. Smoking status, pack-year exposure, and other tobacco-related metadata were not consistently available across the included single-cell and single-nucleus cohorts, preventing smoking-stratified analyses. Accordingly, ExposureA-core should be interpreted as a computationally inferred convergence between predicted NNK targets and IPF-associated lung transcriptional network hubs, not as a measured NNK exposure signature. Second, NNK targets were inferred from chemogenomic and ligand- or structure-informed prediction resources, which may include false-positive and false-negative targets. The true in vivo target spectrum of NNK in human lung tissue remains unknown. Third, pseudo-bulk aggregation improves stability for differential-expression and network analyses but masks intra-cell-type heterogeneity, disease-relevant subclusters, and rare cell states. WGCNA modules and intramodular hubs should therefore be viewed as dominant co-expression programs rather than a complete map of cell state-specific transcriptional responses. Fourth, the supportive feature-space analyses were used only to assess whether the training-derived ExposureA-core feature space showed non-random IPF-control separability. They were not used for diagnostic classifier validation, external model evaluation, target prioritization, or downstream biological interpretation. Similarly, independent bulk-cohort projection supports the external detectability of the fixed epithelial ExposureA-core Top30 signature at the tissue-transcriptome level, but does not establish diagnostic utility, prognostic value, or clinical risk-stratification performance. Bulk lung tissue also lacks explicit cell type resolution and may be influenced by cellular composition (Aran et al. 2017). Finally, molecular docking and molecular dynamics simulations provide only exploratory structural plausibility for modeled NNK–protein complexes. They do not establish direct binding, target engagement, functional regulation, or causal involvement of NNK in IPF pathogenesis (Ganesan et al. 2017, Liu et al. 2018). Experimental validation using biochemical binding assays, controlled exposure-response models, and functional studies in relevant lung-cell systems will be required to determine whether the prioritized epithelial candidate proteins have biological roles in epithelial injury, aberrant repair, or fibrotic remodeling. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 1 Supplementary Material 2 Supplementary Material 2 Supplementary Material 3 Supplementary Material 3 Supplementary Material 4 Supplementary Material 4 Supplementary Material 5 Supplementary Material 5 Supplementary Material 6 Supplementary Material 6 Supplementary Material 7 Supplementary Material 7 Supplementary Material 8 Supplementary Material 8