Fibroblast-derived HSD11B1 deficiency drives lung cancer in never-smokers through steroid metabolic imbalance

Abstract

Lung cancer is one of the leading causes of cancer-related death worldwide, yet its development in people who have never smoked remains poorly understood. This form of cancer is thought to differ from smoking-related disease in both clinical behavior and molecular features, but its causes are still unclear. Growing evidence suggests that changes in the way cells process and regulate metabolites may play an important role. Here we show that a specific steroid-regulating enzyme, HSD11B1, acts as a protective factor against lung cancer in people who have never smoked. Using genetic analyses, single-cell profiling and functional studies, we found that loss of this enzyme in fibroblasts alters the local balance of steroid metabolites, which in turn promotes the transformation and growth of lung epithelial cells. These findings identify fibroblast-derived HSD11B1 as a gatekeeper of metabolic balance in the lung, with its loss linked to early disease progression and poorer outcomes in patients. By uncovering a metabolic mechanism that is distinct from smoking-related pathways, this work provides fresh insight into why never-smokers develop lung cancer and suggests potential new avenues for prevention and treatment.

Introduction

Lung cancer in never-smokers (LCINS), predominantly manifesting as lung adenocarcinoma, has emerged as the most prevalent subtype of lung cancer globally and ranks among the leading causes of cancer-related mortality¹. In contrast to smoking-related lung cancers, LCINS exhibits distinct histopathological, molecular, and clinical features, suggesting a divergent etiology. However, despite increasing recognition of its unique biological identity, the causal factors driving LCINS remain poorly defined¹.

Studies have implicated a complex interplay of environmental exposures, inherited genetic variants, hormonal influences and inflammatory disease in LCINS susceptibility^1,2. Yet, most prior evidence has been correlational, lacking the analytical rigor to establish causal inference, thereby limiting our mechanistic understanding and therapeutic targeting of LCINS.

Metabolites, the intermediates and end products of cellular metabolism, are increasingly recognized as critical mediators of disease initiation and progression³. Circulating metabolite levels reflect an integrated output of host genetics, environment, diet, and microbiota, offering a unique window into disease-relevant biochemical pathways^4,5. While metabolomics has yielded biomarkers for various malignancies, the specific metabolites causally linked to LCINS have not been elucidated.

Human genetics offers a powerful approach to uncover causal relationships between metabolites and disease. Given that many metabolites exhibit high heritability, Mendelian randomization (MR)—a method leveraging genetic variants as instrumental variables—enables the evaluation of putative causal effects between exposures and outcomes. By mimicking randomized controlled trials through the random allocation of alleles at conception, MR reduces the risk of confounding and reverse causation, thus providing robust causal inference⁶. Given the high heritability of many metabolite levels, MR is particularly well-suited to uncover causal metabolic drivers of disease⁷.

In this study, we applied an integrative multi-omics framework to investigate the metabolic underpinnings of LCINS. By combining bidirectional MR with single-cell RNA sequencing (scRNA-seq) and functional validation, we aimed to identify metabolites that causally influence LCINS risk and to delineate their cell-type-specific mechanisms of action. This approach not only strengthens etiological insight into LCINS but also uncovers potential metabolic vulnerabilities amenable to therapeutic intervention.

Results

Study design and overview

This study employed an integrative multi-omics approach to identify causal metabolic risk factors and dissect their cell-specific mechanisms in LCINS. We combined bidirectional Mendelian randomization (MR), Bayesian colocalization, scRNA-seq, and functional assays. The overall workflow is illustrated in Fig. 1.

Fig. 1. Flowchart of study design.

Exposure data from the Canadian Longitudinal Study on Aging (CLSA; 1091 metabolites/309 ratios, European cohort) and outcome data from the McKay and TRICL studies were analyzed using Mendelian randomization (MR) with sensitivity and colocalization tests. Single-cell and functional assays (proliferation, stemness), as well as proteomic, transcriptomic, and survival analyses, were performed for validation.

Identification of causal metabolites associated with LCINS via MR

GWAS metabolite data from the Canadian Longitudinal Study on Aging (CLSA) were utilized, comprising 1091 individual metabolites and 309 metabolite ratios measured in 8299 European participants⁸. LCINS GWAS summary statistics were sourced from studies by James D. McKay et al. and the Transdisciplinary Research Into Cancer of the Lung (TRICL) consortium^9,10. In these cohorts, never-smokers were strictly defined as individuals who had smoked fewer than 100 cigarettes during their lifetime. A total of 23 metabolites were identified as being associated with LCINS (Fig. 2a). Following sensitivity analyses and reverse MR, 11 metabolites were excluded due to heterogeneity, pleiotropy, or reversed causality, leaving 12 significant metabolites (Supplementary Data 1).

Fig. 2. Mendelian randomization and colocalization analyses identify LCINS-associated metabolites.

a, b Forest plots showing causal effects of candidate metabolites on LCINS risk in the discovery (McKay) (a) and validation (TRICL) (b) cohorts. Effect estimates (odds ratios, 95% CI) were calculated using IVW, weighted median, and MR-Egger methods. c The Venn diagram shows that the discovery cohort and the validation cohort share seven positive metabolites. d Bayesian colocalization plots showing shared genetic association signals for the seven metabolites and LCINS. SNPs are colored by linkage disequilibrium (R²) with the lead variant. P-16:0_18:1-GPC: 1-(1-enyl-palmitoyl)-2-oleoyl-GPC (P-16:0/18:1); PC(P-16:0/16:1(9Z)): 1-(1-enyl-palmitoyl)-2-palmitoleoyl-GPC (P-16:0/16:1); GPC(18:1/22:6): 1-oleoyl-2-docosahexaenoyl-GPC (18:1/22:6); PC(P-16:0/16:0): 1-(1-enyl-palmitoyl)-2-palmitoyl-GPC (P-16:0/16:0); Ac-DMLys: N2-acetyl,N6,N6-dimethyllysine; Ac-MMLys: N2-acetyl,N6-methyllysine; C0/C3: Carnitine to propionylcarnitine (C3) ratio; Sia/GlcNAc/GalNAc: N-acetylneuraminate to N-acetylglucosamine to N-acetylgalactosamine ratio; Glu/Cer(d18:1/18:0): Glucose to N-stearoyl-sphingosine (d18:1 to 18:0) ratio; Andro-G/Etio-G: Androsterone glucuronide-to- etiocholanolone glucuronide ratio.

The validation cohort employed GWAS data from TRICL, encompassing 9298 samples (2303 cases and 6995 controls). Applying the same Mendelian randomization framework to this independent dataset identified 16 metabolites associated with LCINS risk (Fig. 2b and Supplementary Data 3). Notably, cross-cohort comparison revealed seven metabolites that were consistently associated with LCINS across both the discovery and validation cohorts (Fig. 2c).

Specifically, the metabolites positively associated with LCINS risk included 1-(1-enyl-palmitoyl)-2-palmitoleoyl-GPC (P-16:0/16:1), 1-(1-enyl-palmitoyl)-2-palmitoyl-GPC (P-16:0/16:0), and N2-acetyl, N6, N6-dimethyllysine. The metabolites inversely associated with LCINS risk included 2-ketocaprylate, X-11444, the androsterone glucuronide-to-etiocholanolone glucuronide ratio, and X-11787, suggesting their protective role in LCINS development.

To further enhance the robustness of these MR-derived associations and determine whether shared causal variants underlie both metabolite levels and LCINS risk, we performed Bayesian colocalization analysis for the seven overlapping metabolites (Fig. 2c). A posterior probability (PH4) >0.8 was considered strong evidence for colocalization. Among the seven metabolites, only the androsterone glucuronide:etiocholanolone glucuronide ratio demonstrated significant colocalization with LCINS (PH4 = 0.872; reduced LCINS risk) (Fig. 2d), indicating that genetic variants influencing this metabolite are likely to also modulate LCINS susceptibility. These results provide compelling evidence that this specific steroid metabolite is closely linked to the development of LCINS.

LCINS-specificity of metabolic associations validated across 12 common cancer types

To evaluate whether the identified association was specific to LCINS rather than other cancer types, we performed MR analyses of the colocalized metabolite across 12 common cancers, including lung cancer in smokers.

In all cases, the estimated ORs were close to 1 (Table 1), suggesting no significant associations with other malignancies. This specificity further underscores the unique etiological relevance of the androsterone-to-etiocholanolone glucuronide ratio in LCINS, where this metabolite exerts a protective effect against LCINS risk.

Table 1.

Mendelian randomization estimates of metabolite effects across common cancer types

Exposure	Outcome	Method	P	OR (95% CI)
Androsterone glucuronide-to-etiocholanolone glucuronide ratio	Liver cancer	IVW	0.14	0.99(0.99,1.00)
	Skin cancer	IVW	0.02	0.99(0.99,1.00)
	Esophageal cancer	IVW	0.92	0.99(0.81,1.20)
	LUSC	IVW	0.65	1.02(0.92,1.14)
	LUAD	IVW	0.94	0.99(0.92,1.08)
	Cervical carcinoma	IVW	0.06	1.10(0.99,1.20)
	Colon cancer	IVW	0.14	1.00(0.99,1.00)
	Pancreatic cancer	IVW	0.19	1.25(0.90,1.73)
	Thyroid cancer	IVW	0.32	0.88(0.68,1.13)
	Ovarian cancer	IVW	0.76	1.02(0.87,1.20)
	Rectal cancer	IVW	0.41	0.99(0.99,1.00)
	Breast cancer	IVW	0.37	1.04(0.95,1.15)
	LCIS	IVW	0.83	0.99(0.92,1.07)

Results show causal estimates for the androsterone glucuronide-to-etiocholanolone glucuronide (Andro-G/Etio-G) ratio on 13 common cancers, including liver, skin, and colorectal cancer, using the inverse-variance weighted (IVW) method. Displayed are P values, odds ratios (ORs), and 95% confidence intervals (CIs) for each outcome, assessing the disease specificity of the metabolite-LCINS association.

LUSC lung squamous cell carcinoma, LUAD lung adenocarcinoma, LCIS lung cancer in smoker.

Single-cell transcriptomics highlight SRD5A1 and HSD11B1 as differentially expressed steroid regulators in early-stage LCINS

To further elucidate the potential biological relevance of the identified metabolite in LCINS, we queried the Human Metabolome Database (HMDB) to investigate its associated metabolic pathways and biological functions. Specifically, the androsterone-to-etiocholanolone glucuronide ratio demonstrated consistent positive associations with LCINS in both Mendelian randomization validation and colocalization analyses. Previous studies have established that this trait is embedded within the androstenedione metabolism pathway, a well-characterized steroidogenic process (Fig. 3a)^11,12. This metabolic ratio is typically regulated by four key enzymes: SRD5A1 and AKR1D1, which are responsible for the conversion of androstenedione into androsterone glucuronide and etiocholanolone glucuronide, respectively; and CYP11B1 and HSD11B1, which modulate androstenedione availability and downstream flux¹³. These enzymes collectively influence the metabolic balance between the two glucuronidated androgens and may act as upstream regulators linking steroid metabolism to the pathogenesis of LCINS. Based on the MR analysis, an increased ratio of this metabolite is considered to reduce the risk of LCINS (Fig. 2a).

Fig. 3. Single-cell analysis of the androgen metabolism pathway in lung adenocarcinoma and fibroblast-specific gene expression in LCINS.

a Metabolic pathway of androstenedione. The metabolites are shown in gray, and the enzymes regulating metabolism are highlighted in orange. b UMAP plot showing clustering of major lung cell types identified across integrated single-cell RNA-seq data from normal lung tissue, smoking-associated lung adenocarcinoma, and LCINS. c UMAP plots displaying sample origin of cells (colored by group: normal [N], smoker-associated tumor [Smoker], and never-smoker tumor [No-Smoker]) and illustrating expression patterns of SRD5A1, HSD11B1, and AKR1D1 across all cell populations, where color intensity reflects gene expression levels. d UMAP plot showing subclustering of fibroblasts into antigen-presenting CAFs (apCAFs), matrix CAFs (mCAFs) and inflammatory CAFs (iCAFs) based on marker gene expression. e Dot plot summarizing expression of canonical marker genes across fibroblast subtypes; dot size denotes percentage of expressing cells and color indicates average expression level. f UMAP visualization of HSD11B1 expression within fibroblast populations, demonstrating subtype specificity. g Inferred intercellular communication network among fibroblast subtypes and immune cells, highlighting ligand-receptor interactions predicted by CellChat analysis. h Heatmap depicting gene set variation analysis (GSVA)-based pathway enrichment across fibroblast subtypes, showing differential enrichment of hallmark and ECM-related pathways.

To identify key enzymes potentially involved in the steroidogenic process underlying LCINS pathogenesis, we analyzed the expression and abundance of four enzyme-encoding genes by integrating scRNA-seq data from 29 human lung samples: 11 normal lung tissues, 7 lung cancer tissues from smokers, and 11 LCINS tissues at very early stages (preinvasive lung adenocarcinoma: adenocarcinoma in situ of the lung and minimally invasive adenocarcinoma of the lung)^14,15. Following dimensionality reduction, clustering, and cell annotation, 23 cell clusters were identified (Supplementary Fig. 1a), categorized into eight major types: epithelial cells (EPCAM, KRT18, and CDH1), B cells (CD79A, CD79B, and CD19), myeloid cells (LYZ, MARCO, and CD68), fibroblasts (COL1A2, DCN, and COL1A1), T cells (CD3G and CD3D), endothelial cells (CLDN5 and PECAM1), NK cells (NKG7 and GNLY), and mast cells (TPSAB1, GATA2, and KIT) (Fig. 3b and Supplementary Fig. 1b)^16,17.

Among the four genes, CYP11B1 exhibited minimal expression across all samples and was therefore excluded from downstream analyses. UMAP visualization of gene expression distribution revealed that SRD5A1 and AKR1D1 were broadly expressed across multiple cell types, whereas HSD11B1 expression was highly specific to fibroblasts (Fig. 3c and Supplementary Fig. 1c). Furthermore, comparative analyses across smoker, and non-smoker tumor samples only revealed differential expression of SRD5A1 and HSD11B1 between smoking and non-smoking patients (Supplementary Fig. 1e, f), implying the engagement of these two genes in the tumorigenesis of LCINS.

Cell type specificity of HSD11B1 to fibroblasts from early-stage LCINS samples prioritizes this gene and this type of cells for further investigations. Fibroblasts isolated from early-stage LCINS samples were further subclustered into six distinct populations (Supplementary Fig. 1g). Based on marker gene expression, these were classified into three fibroblast subtypes: Matrix CAFs (FN1, COL6A3, and COL5A2) (Cluster 1, 2), antigen-presenting fibroblasts (HLA-DRA and HLA-DRB1) (Cluster 0, 3, 4), and immune-interacting fibroblasts (CCL5 and CXCR4) (Cluster 5) (Fig. 3d, e)^18,19. HSD11B1 expression was predominantly confined to Matrix CAFs (Fig. 3f). Cell-cell communication analysis revealed that matrix fibroblasts predominantly interact with epithelial cells, supporting their potential interaction with cancer cells in LCINS (Fig. 3g). Pseudotime analysis revealed that matrix fibroblasts are located at the starting point of differentiation trajectories for other subpopulations, suggesting that they reside in primitive differentiation state within LCINS samples and may play a pivotal role in initiating or driving the differentiation or transformation process (Supplementary Fig. 1h). Gene set variation analysis (GSVA) of Matrix CAFs demonstrated significant enrichment (FDR <0.05) in hallmark pathways, including estrogen metabolism, mitotic signaling, and extracellular matrix organization (Fig. 3h). Collectively, these data suggested matrix CAFs as the dominant source of HSD11B1 in the LCINS tumor microenvironment.

HSD11B1 inhibition in fibroblasts enhances stemness and tumor-initiating capacity in lung epithelial and cancer cells

The observed differential expression of the SRD5A1 and HSD11B1 genes between smokers and non-smokers promoted us to investigate the role of these two genes in the malignant transformational potential of normal bronchial epithelial cells and tumor initiation capacity (TIC) potential of cancer cells, based on classical cancer stem cell (CSC) marker CD44 expression, spheroid formation, soft agar colony formation, and proliferation assays.

Given that SRD5A1 is mainly expressed in lung epithelial cells and its expression in tumor cells is higher than in normal epithelial cells (Supplementary Fig. 1c, i), we selected normal bronchial epithelial cells (BEAS-2B) and lung cancer cell lines with high SRD5A1 expression, namely SW1573 and H2030. SiRNA-mediated knockdown of SRD5A1 in BEAS-2B dramatically inhibited its proliferation, without affecting that of lung cancer cells (Supplementary Fig. 2a, c). Besides, SRD5A1 knockdown significantly downregulated the CSC marker CD44 across normal and cancer cells (Supplementary Fig. 2b). In line with this, the transformational potential of BEAS-2B and the TIC potential of lung cancer cells (SW1573 cell line) were significantly inhibited. These findings contradict our Mendelian results but agree with the literature evidence²⁰.

Since HSD11B1 is exclusively expressed in fibroblasts from LCINS, we utilized normal human lung fibroblasts (HFL-1) and cancer-associated fibroblasts (CAFs) derived from LCINS, to evaluate their effects on normal bronchial epithelial cells (BEAS-2B) and lung cancer cells with differential HSD11B1 expression (HCC827 with high expression and A549 with low expression) (Fig. 4a, b). Conditioned media from HSD11B1-knockdown fibroblasts (HF group) were co-cultured with target epithelial cells, with untreated controls (Ctrl group) and normal fibroblast media controls (NF group).

Fig. 4. Functional validation of fibroblast-derived HSD11B1 in regulating epithelial transformation and tumor-initiating capacity.

a Experimental overview. Normal lung fibroblasts (HFL-1) and cancer-associated fibroblasts (CAFs) from a non-smoker lung adenocarcinoma patient were treated with HSD11B1 siRNA or the selective inhibitor PF-915275. Conditioned media (CM) were applied to BEAS-2B, A549, and HCC827 cells for phenotypic assays. b Relative HSD11B1 expression across cell types (BEAS-2B, CALU-1, A549, H2030, H1299, HCC827, HFL-1, and CAFs), normalized to ACTIN. c, d Spheroid formation assays in BEAS-2B, A549, and HCC827 cells cultured with CM from HFL-1s (c) or CAFs (d) under Ctrl, NF, and HF conditions. Spheroid numbers were quantified per group. Statistical significance was determined using one-way analysis of variance (ANOVA). *p < 0.05, ****p < 0.0001. e CD44 expression in BEAS-2B, A549, and HCC827 cells after co-culture with HFL-1 or CAFs under different CM conditions (Ctrl, NF, HF), normalized to ACTIN. Statistical significance was determined using ANOVA. *p < 0.05 and ***p < 0.001. f Proliferation of BEAS-2B, A549, and HCC827 cells co-cultured with HFL-1s and exposed to CM from control (Ctrl), non-targeting (NF), or HSD11B1-knockdown (HF) fibroblasts. OD 450 nm measured at 24, 48, and 72 h. Statistical significance was determined using one-way analysis of variance ANOVA. **p < 0.01 and ***p < 0.001.

In the HFL-1 co-culture system, conditioned media from the HF group significantly increased spheroid numbers in BEAS-2B, A549, and HCC827 cells (Fig. 4c). Consistently, in the CAF co-culture system, HF-derived media enhanced spheroid formation, with a more pronounced effect in lung cancer cells (A549 and HCC827) (Fig. 4d). To test whether this phenotype reflected the acquisition of stem-like features, we assessed CD44 expression as a surrogate marker. In both HFL-1 and CAF settings, fibroblast HSD11B1 knockdown significantly increased CD44 mRNA levels in BEAS-2B and A549 cells (Fig. 4e), supporting a molecular basis for the observed increase in tumor-initiating capacity. In contrast, effects on cell growth diverged across conditions, with distinct proliferative patterns observed. Specifically, in the HFL-1 co-culture system, BEAS-2B cells exhibited reduced proliferation in the NF group relative to the untreated control (Ctrl). Exposure to HF-conditioned media partially alleviated this growth suppression, resulting in significantly higher proliferation than in the NF condition, although it did not fully restore proliferation to the Ctrl baseline. (Fig. 4f and Supplementary Fig. 3a). In contrast, in both A549 and HCC827 cells, HF-conditioned media reproducibly increased proliferation compared with the NF condition (Fig. 4f and Supplementary Fig. 3a). In the CAF model, HF media enhanced the proliferation of A549 and HCC827 cells, while showing no statistically significant difference in BEAS-2B cells compared to controls (Fig. 4f and Supplementary Fig. 3c, d). Collectively, these results indicate a context-dependent response to HF-conditioned media. In non-malignant BEAS-2B cells, HF-conditioned media was associated with increased stem-like features while concomitantly restraining proliferative capacity. In contrast, in lung cancer cells, HF-conditioned media enhanced both stemness-associated phenotypes and cell proliferation.

Because conditioned media contain both extracellular vesicles (EVs) and soluble (non-EV) factors, we next asked which fraction mediates the observed effects. Conditioned media generated from HF were depleted of EVs using a commercial exosome isolation workflow followed by ultracentrifugation. Notably, whereas intact HF-conditioned media robustly increased cell proliferation, this pro-proliferative effect was lost after EV depletion (HF-depl) in both HFL-1 cells (Supplementary Fig. 4a) and CAFs (Supplementary Fig. 4b). In parallel, the HF-induced enhancement of spheroid formation was similarly abrogated by EV depletion in both fibroblast models (Supplementary Fig. 4c, d). Collectively, these results indicate that vesicular components, rather than the soluble fraction or nonspecific changes in media composition, are necessary for the phenotypes observed with HF-conditioned media. To corroborate our genetic findings with pharmacological evidence, we employed the specific HSD11B1 inhibitor PF-915275²¹. We selected the 20 μM dose as optimal because it inhibited enzymatic activity (Fig. 5a) without affecting fibroblast viability (Fig. 5b). Using this dose, media from PF-915275-treated HFL-1 (Fig. 5c) and CAFs (Fig. 5d) significantly enhanced the colony formation efficiency of BEAS-2B, A549, and HCC827 cells. Furthermore, proliferation assays (Fig. 5e, f and Supplementary Fig. 3b, e) recapitulated the siRNA results: in the HFL-1 model, inhibitor treatment resulted in BEAS-2B proliferation levels that were higher than the vehicle control (NF) but lower than the untreated Ctrl, while promoting cancer cell growth; in the CAF model, it selectively enhanced cancer cell proliferation.

Fig. 5. Functional assays evaluating the effects of conditioned media from PF-915275-treated HFL-1 and CAFs on tumor cell growth and colony formation.

a Cortisol secretion in PF-915275–treated HFL-1 and CAFs measured by ELISA (mean ± SEM). b Proliferation of HFL-1 and CAFs treated with PF-915275 (2.5–40 μM) for 24–72 h, assessed by OD 450 nm. Statistical significance was determined using one-way analysis of variance (ANOVA). **p < 0.01 and ***p < 0.001. c, d Colony formation of BEAS-2B, A549, and HCC827 cells cultured with CM from Ctrl, NF, HF, or PF-915275–treated HFL-1s (c) or CAFs (d). Statistical significance was determined using ANOVA. *p < 0.05, **p < 0.01, and ***p < 0.001. e, f Proliferation of BEAS-2B, A549, and HCC827 cells co-cultured with HFL-1 (e) or CAFs (f) and exposed to CM from Ctrl, NF, or PF-915275–treated fibroblasts. Statistical significance was determined using ANOVA. **p < 0.01, ***p < 0.001, and ****p < 0.0001.

Collectively, these results establish that HSD11B1 downregulation in fibroblasts—whether via genetic knockdown or pharmacological inhibition—dramatically enhances the transformational potential of normal bronchial epithelial cells and tumor initiation capacity of lung cancer cells, positioning it as a key driver of LCINS pathogenesis (Fig. 6f).

Fig. 6. Differential expression and survival analysis of HSD11B1 in LCINS.

a Workflow summarizing datasets included in the analysis (TCGA-LUAD, Jun-Yu Xu et al., S. Satpathy et al., and CPTAC-LUAD). Transcriptomic, proteomic, and prognostic data from early-stage (stage I/II) LCINS were analyzed, adjusting for smoking history and pathological stage. b Violin plots of HSD11B1 mRNA expression in LCINS and normal lung tissues across cohorts. Statistical significance was evaluated using two-tailed unpaired t-tests (****p < 0.0001 for TCGA-LUAD and S. Satpathy et al.; ns indicates no significance for CPTAC-LUAD). c Violin plots showing differences in protein expression of HSD11B1 across different cohorts. Statistical significance was determined using two-tailed unpaired t-tests (**p < 0.01 for Jun-Yu Xu et al.; ****p < 0.0001 for S. Satpathy et al. and CPTAC-LUAD). d Kaplan–Meier survival curves for high vs. low HSD11B1 expression in TCGA-LUAD. Low HSD11B1 predicted worse overall survival. Survival differences were compared using the log-rank test (p < 0.05 for TCGA-LUAD). Hazard ratios were calculated using the Cox proportional hazards regression model (p < 0.05 for TCGA-LUAD). e Independent validation of HSD11B1-survival association in S. Satpathy et al. and Jun-Yu Xu et al. Survival differences were assessed using the log-rank test (p < 0.05 for Jun-Yu Xu et al.). f Proposed model: Fibroblast-specific loss of HSD11B1 in early-stage LCINS disrupts local steroid metabolic balance, producing a pro-transformation microenvironment that enhances epithelial stemness and tumor initiation. This fibroblast-epithelial metabolic axis provides a mechanistic explanation for LCINS development independent of smoking exposure.

HSD11B1 expression as an independent prognostic marker in early-stage LCINS

To investigate the clinical significance of HSD11B1 in early-stage LCINS, we analyzed transcriptomic, proteomic, and prognostic data from multiple independent early-stage LUAD cohorts, including TCGA-LUAD¹⁶, CPTAC-LUAD²², and two additional high-quality datasets (refs. ^23,24) (Fig. 6a). At both the mRNA level (Fig. 6b) and the protein level (Fig. 6c), HSD11B1 expression in tumor tissue was broadly significantly lower than that in normal lung tissue. Survival analyses across multiple independent datasets consistently demonstrated that high HSD11B1 expression was associated with longer overall survival in early-stage LCINS patients (Fig. 6d, e). Moreover, multivariate Cox regression analysis demonstrated that, after adjusting for age, sex, and tumor stage, HSD11B1 expression remained an independent prognostic factor in this patient subgroup (Fig. 6d).

To determine whether the prognostic association of HSD11B1 is specific to non-smoking status, we performed an analogous analysis in the smoker subset of the TCGA-LUAD cohort. In contrast to the pattern observed in never-smokers, Kaplan–Meier analysis showed no significant difference in overall survival between HSD11B1-high and HSD11B1-low groups among smokers (Supplementary Fig. 4a). Consistently, multivariable Cox regression did not identify HSD11B1 expression as an independent prognostic factor in smokers, whereas pathologic stage remained significantly associated with outcome (Supplementary Fig. 4b). Moreover, within the smoker cohort, HSD11B1 expression did not differ significantly between tumor and matched normal tissues (Supplementary Fig. 5c). Collectively, these null associations in smokers support the interpretation that the clinical relevance of fibroblast-derived HSD11B1 is context dependent and preferentially linked to the LCINS etiology.

Discussion

In this study, we employed an integrative multi-omics strategy to unravel the causal metabolic underpinnings of LCINS. Bidirectional MR identified seven circulating metabolites with consistent associations across independent cohorts, including four protective and three risk-related metabolites. Bayesian colocalization analysis further pinpointed the key metabolic trait—androsterone glucuronide-to-etiocholanolone glucuronide ratio—as uniquely associated with LCINS but not the other cancer types. By mapping these metabolites to steroidogenic pathways and analyzing scRNA-seq data, we identified fibroblast-derived HSD11B1 as a plausible modulator of dysregulated steroid metabolism in LCINS. This enzyme, responsible for reversible glucocorticoid activation and broader steroid substrate conversion, was selectively downregulated in CAFs within early-stage LCINS but not in normal tissues or smoking-related lung cancers, suggesting a fibroblast-specific metabolic reprogramming during tumor initiation. Functionally, both genetic silence and pharmacological inhibition of HSD11B1 enhanced the stemness and proliferative capacity of bronchial epithelial and cancer cells, reinforcing a tumor-suppressive role of HSD11B1 in the fibroblast niche. Survival analysis showed that high expression of HSD11B1 was associated with better prognosis, and high HSD11B1 expression is an independent prognostic factor for OS in LCINS. Collectively, these findings highlight a previously unrecognized fibroblast-epithelial metabolic axis driven by HSD11B1 dysregulation and underscore its potential as a context-dependent gatekeeper of epithelial transformation in LCINS.

HSD11B1 is expressed across multiple tumor types and regenerates active glucocorticoids from inactive metabolites via 11β-hydroxysteroid dehydrogenase type 1 (11β-HSD1)²⁵. By modulating local glucocorticoid synthesis, HSD11B1 shapes the tumor immune microenvironment, promoting regulatory T cell-mediated suppression and inhibiting CD8⁺ T cell activation, thereby facilitating tumor growth and immune evasion. Accordingly, HSD11B1 inhibition may restore anti-tumor immunity and enhance therapeutic efficacy. HSD11B1 exhibits tissue- and context-dependent functions in cancer. For instance, specific polymorphisms increase breast cancer risk in postmenopausal women by influencing drug metabolism and immune regulation, whereas its inhibition in renal cancer enhances immunotherapy response by reversing glucocorticoid-driven immunosuppression. These findings highlight HSD11B1 as a context-specific regulator of tumor progression and therapy response^26–28.

Steroid hormones themselves show substantial interindividual variability influenced by sex, age, health status, and lifestyle factors such as smoking²⁹. These hormones play pivotal roles in tumorigenesis, as demonstrated in hormone-sensitive cancers like prostate, breast, and high-grade serous ovarian cancer^30,31. The interaction of steroid hormone receptors (SHRs) with DNA damage response (DDR) pathways profoundly affects gene expression programs, contributing to tumor development and progression. Importantly, steroid-metabolizing enzymes significantly modulate these effects, suggesting potential therapeutic targets in both hormone-related and non-hormone-related malignancies.

These results gain broader significance when considered alongside fibroblast-mediated mechanisms in other malignancies. In colorectal cancer, Lepr+ fibroblasts promote tumor microenvironment formation and carcinogenesis initiation through TGF-β-induced MCAM expression, which activates the NF-κB-IL34/CCL8 signaling axis to recruit tumor-associated macrophages³². Similarly, in prostate cancer, heterogeneous CAFs reprogram androgen metabolism via HSD11B1 and alter the local microenvironment through NF-κB-driven inflammatory signaling, thereby stimulating malignant transformation of epithelial cells³³. This comparative perspective reinforces the emerging paradigm of fibroblasts as master regulators of tissue-specific tumor initiation through metabolic and inflammatory pathway modulation.

Several limitations should be considered when interpreting these findings. First, regarding the causal inference methodology, the MR analyses revealed nominal associations with LCINS (P < 0.05). Given the large number of metabolites screened and the highly correlated nature of metabolic traits, a stringent Bonferroni correction may be overly conservative and could suppress biologically meaningful causal signals. Therefore, rather than relying solely on metabolome-wide statistical significance, our emphasis was placed on cross-cohort reproducibility, strong colocalization evidence, and convergent consistency across multiple analytical approaches. Although the MR framework reduces the risk of reverse causality, residual pleiotropy cannot be entirely excluded due to the inherent complexity of metabolic regulation. Additionally, it is important to acknowledge that our colocalization analysis assumes a single causal variant per locus. While this approach effectively identifies shared signals, it may not fully capture the complexity of loci harboring multiple causal variants.

Second, concerning the mechanistic link between genetics and gene expression, while our study nominates HSD11B1 as a key regulator supported by functional validation, the direct regulation of lung HSD11B1 expression by the identified genetic variants remains inferred. We did not observe direct colocalization with lung tissue eQTLs, likely due to the limited gene coverage of current lung-specific eQTL datasets. Thus, HSD11B1 is proposed as a plausible effector pending formal genetic mediation analysis in future lung tissue cohorts.

Third, regarding the extrapolation of systemic metabolites to local tissues, our MR analyses relied on circulating (serum) steroid glucuronide ratios as genetic proxies. While these systemic markers are robustly associated with LCINS risk, they may not fully capture the specific steroid dynamics within the local lung interstitium. Future studies utilizing interstitial fluid profiling are warranted to validate the correspondence between systemic proxies and the local metabolic microenvironment. Related to this, a limitation is the inability to directly quantify the androsterone-to-etiocholanolone glucuronide ratio in our cell culture models, as this requires specialized isotope-labeled standards not feasible in our current setting. Consequently, cortisol was used as a surrogate readout. Although cortisol regeneration is a well-established proxy for HSD11B1 activity, it does not fully substitute for direct glucuronide-based quantification. Fourth, regarding data generalization and specific analyses, several constraints must be acknowledged. The lack of multi-ethnic metabolomic datasets limits the generalizability of our conclusions, which were derived primarily from European populations; future work incorporating diverse ancestries will be essential. While the cross-cancer MR analysis serves as a useful contextual control, the observed null associations should be interpreted as an absence of current evidence rather than a definitive absence of causal effect, potentially due to limited statistical power. In our single-cell analysis, although we identified a robust enrichment of HSD11B1 in matrix CAFs (mCAFs), these findings rely on aggregated datasets where donor-level heterogeneity and potential sample-composition biases were not explicitly modeled. Consequently, the observed mCAF specificity should be interpreted as a strong, predominant association rather than an absolute exclusivity. Furthermore, it is important to note that HSD11B1 expression is notably enriched in fibroblasts. As our survival models relied on bulk transcriptomic data and were not uniformly adjusted for tumor purity or stromal content, the observed prognostic associations may be partially confounded by variations in the stromal proportion of the tumor microenvironment.

Finally, a major limitation is the absence of in vivo models that faithfully represent the tumorigenic process of LCINS. Current spontaneous, carcinogen-induced, or genetically engineered mouse models do not capture the etiological or biological features of LCINS. Consequently, in vivo validation could not be performed within this study. Future work will focus on developing or employing more suitable models for mechanism-based validation.

Together, by coupling causal inference with cellular and functional validation, our study uncovers fibroblast-derived HSD11B1 as a putative metabolic modulator in LCINS initiation. These findings provide a new conceptual framework for understanding never-smoker lung cancer and open avenues for developing metabolism-oriented preventive and therapeutic strategies.

Methods

Cell lines and fibroblast isolation

Human bronchial epithelial cells (BEAS-2B) and lung adenocarcinoma (LUAD) cell lines—including HCC827, A549, SW1573, and H2030—were maintained in our laboratory. Human lung fibroblasts (HFL-1) were kindly provided by Prof. Hao Zhang (Xuzhou Medical University). Cancer-associated fibroblasts (CAFs) were isolated from fresh non-smoking LUAD tumor specimens obtained from surgical resections at Shanghai Chest Hospital under approved protocols.

Briefly, tumor tissues were washed with cold phosphate-buffered saline (PBS), mechanically minced, and enzymatically digested using a collagenase-containing dissociation buffer in conjunction with the gentleMACS™ Dissociator (program h_tumor_01_01), with intermittent incubation at 37 °C. The resulting cell suspension was sequentially filtered through 100 μm and 40 μm strainers, subjected to red blood cell (RBC) lysis, and centrifuged to obtain single-cell suspensions. Cell viability was assessed using Trypan Blue or Türk’s solution, and viable cells were aliquoted for downstream applications.

All cell lines were cultured in RPMI-1640 medium supplemented with 10% fetal bovine serum (FBS) and 1% penicillin-streptomycin and maintained at 37 °C in a humidified incubator with 5% CO₂.

Fibroblast co-culture and conditioned media preparation

HFL-1 and CAFs were transfected with siRNA targeting HSD11B1 using Lipofectamine RNAiMAX. To ensure standardization across conditions, cells were seeded at a fixed density of 7 × 10⁵ cells per 10 cm dish. Following transfection (or inhibitor treatment), cells were washed with PBS and supplied with fresh medium to eliminate residual transfection reagents or inhibitors. After 72 h of incubation, conditioned media (CM) were collected and sequentially centrifuged at 300×g for 10 min, 2000×g for 10 min, and 15,000×g for 30 min to remove debris and apoptotic bodies. The supernatants were passed through 0.45-μm PES filters (Millipore). To distinguish specific paracrine signaling from nonspecific medium compositional changes, a subset of the CM was subjected to extracellular vesicle (EV) removal. This was performed using an Exosome Isolation Kit coupled with ultracentrifugation, yielding EV-depleted supernatants (designated as HF-depl). The processed CM samples were stored at −80 °C until use. Repeated freeze–thaw cycles were strictly avoided to preserve sample integrity. The collected CM (including complete and HF-depl formulations) were subsequently used to culture BEAS-2B, A549, and HCC827 cells in functional assays assessing proliferation, stemness, and metabolic responses.

Enzyme-linked immunosorbent assay (ELISA)

Given the lack of commercially available assays to directly quantify the androsterone:etiocholanolone glucuronide ratio, we measured cortisol levels in fibroblast supernatants as a surrogate readout of HSD11B1 activity, leveraging the well-established role of cortisol as a functional indicator of HSD11B1 enzymatic function. To evaluate the enzymatic activity of HSD11B1, we quantified cortisol concentrations in the culture supernatants of HFL-1 and CAFs, with or without treatment by the selective HSD11B1 inhibitor PF-915275 (20 μM, 48 h). Data presented represent the mean of three independent biological replicates (n = 3). Cell culture supernatants were harvested, centrifuged at 2000–3000 rpm for 20 min to eliminate debris, and stored at −80 °C.

Cortisol levels were measured using a commercially available ELISA kit according to the manufacturer’s protocol. Serial two-fold dilutions were used to generate standard curves. All samples and standards were assayed in triplicate to ensure technical precision. Fifty microliters of either standards or samples were added to each well, followed by incubation at 37 °C for 50 min. After washing with 1x wash buffer, biotin-conjugated antibody and streptavidin-biotin complex (SABC) working solutions were added sequentially, with incubations at 37 °C for 50 and 30 min, respectively. Color development was initiated by TMB substrate at 37 °C in the dark for 10–20 min and stopped with 50 μL stop solution. Absorbance was read at 450 nm within 30 min.

Cell counting and CCK-8 proliferation assay

For direct cell counting assays, 5 × 10⁴ cells were seeded into six-well plates. Cells were trypsinized and manually counted every 48 h. Experiments were performed using three independent biological replicates $\left(n=3\right)$, each consisting of three technical replicates. For CCK-8 proliferation assays, 1 × 10³ cells were seeded into 96-well plates, and absorbance at 450 nm was measured at 6, 24, 48, and 72 h using the Cell Counting Kit-8.

In parallel, 1 × 10⁴ cells were seeded into low-attachment 6-well plates and cultured in serum-free F12 medium for 6 days to assess sphere-forming capacity. Tumorspheres were defined as spheroids >50 μm in diameter for A549 cells and >30 μm for BEAS-2B and HCC827 cells. Spheres were imaged and quantified under an inverted microscope.

Sphere-formation and soft agar colony formation assays

For the tumorsphere assay, 1 × 10⁴ cells were seeded into ultra-low-attachment six-well plates and maintained in serum-free F12 medium supplemented with EGF and bFGF for 6 days. This assay was repeated in three independent biological experiments. Formed spheres were quantified under a microscope using predefined size criteria as stated above. Spheres were imaged and quantified under an inverted microscope in a blinded manner. No specific exclusion criteria were applied to the images. For the soft agar colony formation assay, 24-well plates were prepared with a base layer of 0.6% agar in culture medium. A top layer consisting of 0.3% agar mixed with 1 × 10³ cells was overlaid and solidified at room temperature. Plates were incubated at 37 °C in a 5% CO₂ humidified incubator for 14 days. Experiments were conducted with three biological replicates, each containing triplicate wells. Colonies containing more than 50 cells were manually counted under a microscope in at least five random fields per well by an investigator blinded to the experimental groups. No fields were excluded from the analysis unless a technical failure was observed.

Quantitative polymerase chain reaction (qPCR)

Total RNA was extracted using the UNIQ-10 Column Trizol Kit, reverse-transcribed with HiScript II Q RT SuperMix, and subjected to qPCR on an ABI ViiA™ 7 using ChamQ Universal SYBR Master Mix with ACTIN as the internal control. Primers: SRD5A1 (F: TCAGACGAACTCAGTGTACGG, R: CGTAGTGGACGAGGAACATGG), HSD11B1 (F: CTCAGTTACGTGGTCCTGACT, R: GAGGAGACGACAACAATGCTT), CD44 (F: CTGCCGCTTTGCAGGTGTA, R: CATTGTGGGCAAGGTGCTATT).

Mendelian randomization

To identify causal metabolic drivers of LCINS, we conducted Mendelian randomization (MR) analyses on 1,400 circulating metabolites, based on genome-wide association study (GWAS) data from 8299 individuals⁵. Single-nucleotide polymorphisms (SNPs) significantly associated with metabolite levels (P < 5 × 10⁻⁸) were selected as instrumental variables (IVs). SNPs in linkage disequilibrium (LD) were pruned at R² < 0.001. Only IVs with F-statistics >10 were retained to minimize weak instrument bias. Metabolite-SNP associations were harmonized with outcome datasets prior to analysis³⁴.

For metabolites with a single SNP instrument, causal effects on LCINS risk were estimated using the Wald ratio method. For metabolites with multiple independent SNPs, we applied the inverse-variance weighted (IVW) approach to compute odds ratios (ORs) and 95% confidence intervals³⁵.

MR analyses were conducted using the TwoSampleMR R package (v0.6.18)³⁶. Harmonization of genetic instruments was performed to ensure alignment of effect alleles between exposure and outcome datasets. Strand alignment was inferred based on effect allele frequencies (EAF), and ambiguous palindromic SNPs were excluded. For each metabolite, the primary causal estimate was derived via IVW or Wald ratio, depending on the number of available IVs. Heterogeneity among instruments was assessed using Cochran’s Q statistic, and Steiger directionality tests were used to validate the direction of causality. Sensitivity analyses were performed using weighted median, weighted mode, and MR-Egger regression to evaluate the robustness of causal inferences and detect horizontal pleiotropy³⁷. MR-Egger estimates were interpreted only when the intercept suggested the presence of directional pleiotropy. Bidirectional MR was conducted to confirm causal directionality between metabolites and LCINS³⁸. All analyses were performed using R 4.3.3.

Colocalization analysis

To assess whether metabolite-LCINS associations were driven by shared genetic loci, we performed Bayesian colocalization analysis using the coloc R package. For each locus, the analysis was conducted within a genomic window of ±50 kb around the lead SNP. We calculated the posterior probability of hypothesis 4 (PPH4), which indicates a shared causal variant. Default prior probabilities were applied: $p_1=1\times {10}^{-4}$ (probability of a causal variant for trait 1), $p_2=1\times {10}^{-4}$ (probability of a causal variant for trait 2), and $p_{12}=1\times {10}^{-5}$ (probability of a shared causal variant). A PPH4 >0.8 was considered strong evidence for colocalization. The linkage disequilibrium (LD) structure was modeled using the 1000 Genomes Project Phase 3 European reference panel to ensure ancestry matching with the GWAS summary statistics. A PPH₄ >0.8 was considered strong evidence of colocalization³⁹.

Single-cell expression analysis

To further assess cell-type-specific expression of candidate genes associated with LCINS, scRNA-seq data from LUAD tissues of never-smokers and smokers were obtained from the Gene Expression Omnibus (GEO, GSE189357¹⁴, and GSE131907¹⁵). Data preprocessing and transformation were performed using the “Seurat” package (version 4.3.0), where genes with fewer than three counts per cell and cells with fewer than 200 unique features were removed¹⁴. Following quality control, data normalization was performed using the NormalizeData function, and the top 2000 highly variable features were identified via the FindVariableFeatures function using the “vst” selection method. To correct for batch effects and integrate data from different datasets/donors, we applied the canonical correlation analysis (CCA) integration workflow provided by Seurat.The integrated data were scaled using ScaleData, and principal component analysis (PCA) was performed. The first 20 principal components (PCs) were selected for downstream analysis based on the ElbowPlot. A K-nearest neighbor (KNN) graph was constructed using the FindNeighbors function based on these 20 PCs. Clusters were identified using the FindClusters function with a resolution of 0.4 (unless otherwise specified for subclustering). Cell-type annotation was conducted using the “SingleR” package (version 2.0.0)⁴⁰. The same preprocessing pipeline was applied to all publicly available scRNA-seq datasets used in our study.

Differential expression analysis

Differentially expressed genes (DEGs) between clusters were identified using the FindAllMarkers function with a log-fold change threshold of 0.25 and filtering for positive markers only. The Wilcoxon rank-sum test was used to compare gene expression levels across clusters, and p values were adjusted for multiple testing using the Bonferroni correction.

Cell-cell communication inference

To infer intercellular communication networks, we used CellChat (v1.5.0). Ligand-receptor pairs involved in secreted signaling, ECM-receptor interactions, and direct cell-cell contact were curated from the CellChat database⁴¹. Communication probabilities were computed using the computeCommunProb, computeCommunProbPathway, and aggregateNet functions. Visualization was performed using netVisual_circle and netVisual_bubble.

Trajectory analysis

Fibroblast differentiation trajectories were reconstructed using Monocle 2. Dimensionality reduction was performed using the reduceDimension function with the DDRTree method. Pseudotime-associated gene expressions were analyzed using differentialGeneTest, and branch-dependent fate decisions were evaluated using BEAM (Branch Expression Analysis Modeling). Gene expression dynamics across pseudotime were visualized with plot_genes_branched_pseudotime⁴².

Gene set enrichment analysis (GSEA)

GSEA was performed using the clusterProfiler (v4.6.0) and GSEABase (v1.60.0) packages in R. DEGs identified from scRNA-seq or bulk RNA-seq comparisons were subjected to enrichment against the MSigDB Hallmark and KEGG gene sets⁴³. Enrichment was assessed using default parameters, with significance thresholds set at nominal p < 0.05, adjusted p < 0.25, and normalized enrichment score (NES) >1.

Clinical multi-omics data analysis

Transcriptomic and proteomic data were obtained from multiple independent LCINS cohorts, including TCGA-LUAD¹⁶, CPTAC-LUAD²², and the studies by refs. ^23,24. The mRNA levels and protein expression levels of HSD11B1, were compared between normal lung tissue and tumor tissue from LCINS patients.

Survival analysis

Patient information was listed in Supplementary Data 2. To ensure data quality, a complete-case analysis strategy was adopted; patients with missing survival information (time or status) or incomplete key clinical covariates (e.g., pathological stage) were excluded from the final cohort. Patients were subsequently divided into two groups based on the optimal cutoff value of HSD11B1 levels, determined using the surv_cutpoint function in the survminer package (v0.4.9). Kaplan–Meier methods from the survival package (v3.3-1) were used to estimate survival curves for the two patient groups, and the statistical significance was assessed using the log-rank test. A multivariate Cox proportional hazards model was used to evaluate the factors (pathological stage, sex, and age) independently affecting survival outcomes.

Ethical statement

The application of lung cancer tissue was sanctioned by the Institutional Review Board of the Shanghai Lung Tumor Clinical Medical Center, Shanghai Chest Hospital (Approval No.: KS(Y)23072), conforming to all pertinent ethical norms, such as the Helsinki Declaration as revised in 2013. Ahead of their participation in the research, comprehensive consent was secured from all the subjects involved.

Statistical analysis

Statistical details for each figure are provided in the corresponding legends. Data distributions were assumed to be normal, though not formally verified. All experiments yielded consistent and reproducible results. Statistical analyses were performed using GraphPad Prism 9.0 and R (v4.2.0). Significance was assessed by unpaired, two-tailed Student’s t-tests or two-way ANOVA, with p < 0.05 considered statistically significant. Significance levels are denoted as: *P < 0 .05, **P < 0.01, *** P < 0.001, and **** P < 0.0001.

Author contributions

The project conception and planning were overseen by Haitang Yang, Hao Zhang, and Feng Yao. Zesen Luo was responsible for developing the bioinformatics pipeline, conducting all experiments, and performing data analysis. Figures and schematic diagrams were created by Jiaying Jia and Zesen Luo. Human LUAD samples were provided by Haitang Yang, and Feng Yao. The manuscript was written by Zesen Luo and Haitang Yang.

Data availability

Genome-wide association study (GWAS) summary statistics for Mendelian Randomization analyses are publicly available as follows: Lung cancer and other malignancies: [IEU GWAS Database](https://gwas.mrcieu.ac.uk) and [GWAS Catalog](https://www.ebi.ac.uk/gwas); Metabolite quantitative trait loci (mQTL): [GWAS Catalog](https://www.ebi.ac.uk/gwas) under accession numbers GCST90199621 to GCST90201020. scRNA-seq data of lung cancer tissues were deposited in the GEO repository under accession ([https://www.ncbi.nlm.nih.gov/gds], GSE189357 and GSE131907). Our transcriptomic data, proteomic data, and prognostic data were sourced from TCGA-LUAD, CPTAC-LUAD(PDC000153), Jun-Yu Xu et al., and S. Satpathy et al.'s study.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41698-026-01445-7.