circNOX4 activates an inflammatory fibroblast niche to promote tumor growth and metastasis in NSCLC via FAP/IL-6 axis

Tumor samples were collected from a cohort of 84 NSCLC patients at the Fourth Hospital of Hebei Medical University (Shijiazhuang, China) between August 2018 and March 2019 with the approval of the Medical Ethics Committee and informed consent from the corresponding participants (2018MEC005). All samples were histologically confirmed to have NSCLC. Patients’ clinical information was collected and stored in a database, which was updated every 3 months by telephone follow-up. Overall survival (OS) was defined as the time interval from diagnosis to the date of cancer-related death or the last follow-up (January 2024).

IHC staining was performed according to standard protocols as previously described [18]. The staining was evaluated based on the intensity of FAP (ab207178, 1:250, Abcam, CA, USA) by two independent pathologists, and all scoring was completed blinded to the clinical information. The intensity of FAP was classified as 0, no staining; 1, weak intensity; 2, moderate intensity; and 3, high intensity. Scores of 0 and 1 were considered low expression, while scores of 2 and 3 were considered high expression.

Five paired CAFs and NFs were isolated from human NSCLC tissues and normal tissues (more than 5 cm away from the tumor margin). Freshly isolated surgical specimens were acquired from patients with the approval of The Fourth Hospital of Hebei Medical University (2018MEC005). The patient’s basic information is listed in Additional file: Table S1. For fibroblast isolation, tissue was initially stored and washed in PBS with penicillin/streptomycin (100 U/100 μg, BI, Israel). The tissue was then physically minced and digested enzymatically in a solution containing 2 mg/mL DNase (Gibco, Waltham, MA, USA), 5 mg/mL collagenase type I (Gibco, Waltham, MA, USA), 100 U/mL penicillin, and 100 μg/mL streptomycin at 37 °C for 2 h. After a series of filtration and centrifugation steps, cells were subsequently cultured in 15% fetal bovine serum (FBS, Gibco, Waltham, MA, USA), DMEM (Gibco, Waltham, MA, USA), 20 ng/mL EGF (PeproTech, Suzhou, China) and penicillin/streptomycin (100 U/100 μg) at 37 °C in humidified air with 5% CO₂ for subsequent expansion. Fibroblast populations were identified by morphology, western blotting and immunofluorescence staining (fibroblast markers α-SMA and FAP; epithelial cell marker EpCAM). Fibroblasts were subcultured at 80% confluence and used in experiments at passages 4–10. Antibody information is provided in Additional file: Table S2.

Human lung cancer cell lines (A549, PC9, H226, and H1581) and human bronchial epithelial cells (BEAS-2B) were purchased from Procell Life Science & Technology Co., Ltd. (Wuhan, China) and maintained in RPMI-1640 medium (Gibco, Waltham, MA, USA) containing 10% FBS, 100 U/ml penicillin, and 100 μg/ml streptomycin at 37 °C and 5% CO₂ under saturated humidity.

Both complementary DNA (cDNA) and genomic DNA (gDNA) extracted from CAFs and NFs were amplified with circNOX4 primers (both convergent and divergent primers). Agarose gel electrophoresis was applied to analyze the RT–PCR products. The back-splice junction site of circNOX4 was verified by Sanger sequencing. Extracted RNAs were treated with RNase R (Geneseed, 3 U/μg, 37 °C, 20 min) to detect the stability of circRNA.

In brief, total RNA was isolated using the E.Z.N.A. Total RNA Kit (OMEGA, Japan). cDNA was synthesized using HiScript III RT SuperMix (Vazyme, Nanjing, China) and amplified using a Light Cycler 480 II Real-Time PCR System (Roche, Basel, Switzerland) with ChamQ Universal SYBR qPCR Master Mix (Vazyme, Nanjing, China). GAPDH and U6 were used as internal controls via the classical ΔΔCt method. Primers were designed and synthesized by RiboBio (Guangzhou, China) and Tsingke (Wuhan, China). The primer sequences are listed in Additional file: Table S3.

FISH assays were performed on CAFs/NFs and NSCLC tissues as previously described [19]. FISH probes were designed and synthesized by Servicebio (Wuhan, China). Dig-labeled probes specific for circNOX4 and biotinylated locked nucleic acid miR-329-5p probes were used during hybridization. The targeted sequences of the probes are provided in Additional file: Table S4. The FISH score for circNOX4 in NSCLC tissues was determined based on the integrated optical density values scanned by AIPATHWELL software (Servicebio Technology Co., Wuhan, China). The optimal cut-off value of 0.1 was determined by X-tile software. Based on this threshold value, patients were categorized into circNOX4 high-expression or low-expression subgroups for further analysis.

The fibroblasts were embedded in three-dimensional collagen matrices in 24-well plates with Rat Tail Collagen I (Corning, NY, USA) at a final concentration of 2 mg/mL collagen and 5 × 10⁵ cells per mL. A 500 μL aliquot of this mixture was dispensed into each well and allowed to solidify for 30 min at 37 °C. Gels were freed from the wells using a pipette tip after polymerization, and complete culture medium was added to the wells, followed by incubation for 12 h. ImageJ software was employed to measure the contractile capacity, which is defined by the following equation: [Area (well)-Area (gel)/Area (well)] × 100%.

Standard western blotting was carried out by a wet transfer system (Bio-Rad Laboratories). Briefly, cells were lysed in lysis buffer (Solarbio, Beijing, China) supplemented with protease inhibitors. Total proteins were separated on SDS–PAGE gels and then transferred onto PVDF membranes (Millipore, Billerica, MA, USA). After blocking the membranes with 5% milk-TBST, they were incubated with appropriate dilutions of specific primary antibodies against FAP (ab207178, 1:1000, Abcam) and α-SMA (ab124964, 1:1000, Abcam). The blots were incubated with HRP-conjugated secondary antibodies and visualized using the ECL system (GE, Boston, MA, USA).

Indicated small interfering RNAs (siRNAs) were transfected into primary fibroblasts or cancer cells using Lipofectamine 3000 reagent (Invitrogen, Carlsbad, CA, USA). For stable transfection, the lentiviral vector (pGLV3/GFP/Puro) containing short hairpin RNAs (shRNAs) was designed based on the si-circNOX4#1 sequence and packaged by GenePharma (Shanghai, China). The circNOX4 sequence was packaged into the pGLV5/GFP/Puro vector. Fluorescent signals were observed using a fluorescence microscope (Olympus, Japan). The miRNA mimics, inhibitors and corresponding negative controls (NC) for miR-329-5p and miR-624-5p were synthesized and purchased from RiboBio (Guangzhou, China). The sequences used are listed in Additional file: Table S5.

The cell viability of fibroblasts was measured by a cell counting kit-8 (CCK-8) assay (Solarbio, Beijing, China). A total of 2000 cells were seeded in 96-well plates per well, and 10 μl of CCK-8 reagent was added and incubated for 2 h at 37 °C. The absorbance was detected using a microplate reader (Tecan, Switzerland) at a wavelength of 450 nm.

For coculture of CAFs and cancer cells, NSCLC cells were plated in the upper chamber of transwell apparatus (0.4 μm insert; Corning, Corning, NY, USA), and CAFs were cultured in the lower chamber. After incubation for 48 h, supernatants were collected for further measurements. For CM preparation, indicated fibroblasts were cultured to reach 80% confluence and then replaced with serum-free DMEM for 48 h. The supernatants were collected and centrifuged to remove cell pellets. The CM was subjected to cytological experiments or stored at -80 °C.

For the cell migration assay, NSCLC cells were cultured in six-well plates and scraped linearly to create an artificial wound. Then, NSCLC cells were incubated in CAF-CM or NF-CM for 48 h. For the cell invasion assay, the upper chambers of the Transwell plates (8 μm insert) were precoated with Matrigel (dilution, 1:8; BD Biosciences, Franklin Lakes, NJ, USA) at 37 °C for 30 min. NSCLC cells (5 × 10⁴/well) suspended in serum-free medium were seeded into the upper chambers. Indicated fibroblasts (3 × 10⁴/well) were seeded into the lower chambers. Subsequent procedures were performed according to previous studies [20].

Wild-type (WT) or mutant (MUT) pmiRGlo-circNOX4 and pmiRGlo-FAP-3′UTR dual luciferase reporter vectors incorporating predicted binding sites for miR-329-5p and miR-624-5p were constructed by GenePharma (Shanghai, China). The sequences are provided in Additional file: Table S5 and Additional file: Table S6. Cells were seeded on 12-well plates at a density of 60% and then co-transfected with WT or MUT plasmids with miRNA mimics or a negative control (NC) using Lipofectamine 3000 reagent. After 48 h of incubation, the cells were collected and tested for luciferase activity with a Dual-Luciferase Assay System kit (Promega, Madison, WI, USA) according to the manufacturer’s instructions.

RIP was performed with an RNA immunoprecipitation kit (Geneseed, Guangzhou, China). CAFs were transfected with miR-NC or miR-329-5p for 48 h. Then, the cells were lysed with RIP lysis buffer and incubated with magnetic beads precoated with antibodies against Ago2 (ab186733, 1:50, Abcam) or IgG (ab172730, 1:50, Abcam). After the antibody was recovered by protein A/G beads and purification was completed, qRT-PCR was performed to detect the enrichment of circNOX4 and miR-329-5p in the precipitates.

A human cytokine array (AAH-CYT-G5; RayBiotech, Norcross, GA, USA) was used to measure the secretion levels of 80 cytokines in the CM of sh-circNOX4 CAFs and sh-NC CAFs according to the manufacturer’s instructions. Quantitative array analysis was performed using the ImageQuant LAS4000 Scanner (GE, Boston, MA, USA). Cytokines were screened using the following integrated conditions: Fold change ≥ 1.2 and fluorescence intensity values > 300.

Cytokines in the supernatant of fibroblasts or cancer cells were measured using ELISA kits (IL-6, CCL2: Abcolonal, Wuhan, China; TGF-β2, IGFBP-4: RayBiotech, Norcross, GA, USA) following the manufacturer’s instructions.

The animal experiments were conducted in accordance with national guidelines and approved by the Animal Research Committee of the Fourth Hospital of Hebei Medical University (SYXK2022-011). BALB/c nude mice (female, age 5 weeks) were purchased from Beijing HFK Bio-Technology Co., Ltd.

For the subcutaneous xenograft model, 5 × 10⁶ A549 cells alone or in a 3:1 mixture with stably transfected fibroblasts were subcutaneously implanted into the axillae of mice (n = 6–8 mice per group). For the anti-IL-6 experiment group, neutralizing anti-IL-6 antibody (10 μg/ml, R&D, Minneapolis, MN, USA) was administered via intratumoral injection twice a week after 14 days of model construction [21]. Tumor volume was calculated by using the formula: Volume = (length × width²)/2. IHC analysis was performed on the xenografts to evaluate protein expression utilizing antibodies against FAP (ab207178, 1:250, Abcam), CD31 (28083–1-AP, 1:200, Proteintech), N-cadherin (Proteintech, 1:100, Servicebio), Vimentin (10366–1-AP, 1:400, Proteintech), matrix metalloproteinase 2 (MMP2) (ab86607, 1:200, Abcam), matrix metalloproteinase 9 (MMP9) (ab76003, 1:200, Abcam) and IL-6 (GB11117, 1:200, Servicebio).

For the lung metastasis model, 2 × 10⁵ A549 cells in 100 μL PBS were injected into the tail vein of BALB/c nude mice. The mice were randomly assigned to one of three groups: (1) untreated A549 cells, (2) A549 cells "primed" in vitro for 72 h in a transwell coculture model with CAFs transduced with empty vector (sh-NC), or (3) CAFs transduced with circNOX4 shRNA (sh-circNOX4). After 5 weeks, the mice were sacrificed and dissected, and the lungs were collected. Hematoxylin and eosin (HE) staining was used to observe the metastatic foci.

The preparation of ⁶⁸ Ga-Labeled FAPI-04 solutions was performed by previous method [22]. Briefly, ⁶⁸ Ga was generated by a ⁶⁸Ge–⁶⁸ Ga generator and eluted with a solution of 0.1 M hydrochloride. The mixture of the ⁶⁸ Ga solution (74 MBq), 1.0 M sodium acetate (95 μL) and 1.15 mM FAPI-04 (20 μL) were reacted at 95 °C for 10 min. ⁶⁸ Ga FAPI-04 PET imaging was performed up to 60 min after intravenous injection of 11.1 MBq of ⁶⁸ Ga-FAPI-04 in tumor xenograft mice (n = 3) using a micro-PET/SPECT/CT machine (Novel Medical Equipment Ltd., China) under isoflurane anesthesia. PET images were reconstructed with the comprehensive image analysis software Pmod v4.201 (PMOD Technologies LLC, Switzerland) and were converted to SUV images. Quantification was performed using a region-of-interest (ROI) technique and expressed as SUVmax.

Given the supporting role of CAFs in tumor growth and metastasis [7], we first confirmed the contribution of FAP⁺CAFs in NSCLC. We queried FAP mRNA expression using the TCGA database and found that FAP was upregulated in lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC) compared with normal lung tissues (Fig. 1A). The data also showed the copy number and mutation status of FAP gene were not correlated with FAP mRNA expression, indicating that the high expression level of FAP was potentially regulated by epigenetic mechanisms (Fig. S1A). Next, we analyzed the spatial cellular expression pattern of FAP across different datasets and cell types in the TME and confirmed that FAP was specifically enriched in fibroblast lineages (Fig. S1B). Moreover, we analyzed the GEO profile GSE127465, which revealed that fibroblasts were distributed into five clusters in NSCLC. We found that FAP was enriched in the C1 and C2 subpopulations, indicating the heterogeneous nature of CAFs in the TME (Fig. 1B, C). To uncover the biological function of FAP⁺CAFs in development trajectories, we constructed developmental pseudotime and discovered that FAP expression was relatively higher along the branch 1 cellular state, which was accompanied by increased protumor activities such as angiogenesis, inflammation, and metastasis (Fig. 1D-F). These findings revealed that FAP exhibits cell lineage- and developmental stage-dependent expression patterns while undergoing protumor activities.

To investigate the clinical relevance between fibroblast activation and the prognosis of NSCLC patients, we collected clinical samples from 84 NSCLC patients and performed IHC staining. The result showed that FAP was expressed in the stromal compartment in NSCLC tissues (Fig. 1G). Analysis of clinicopathological variables based on FAP expression level revealed that high expression of FAP was positively correlated with high stromal fraction and multiple distant metastases (Fig. 1H, I), indicating that FAP is closely related to the process of metastasis. Moreover, FAP expression was inversely correlated with overall survival in NSCLC patients (P = 0.023; Fig. 1J), in line with the clinical outcome data from the TCGA database (Fig. 1K). Taken together, FAP⁺CAFs are populated in metastatic tumors with inferior clinical outcomes in NSCLC patients.

To gain insight into the contribution of circRNAs to aberrant fibroblast activation in the TME, we first isolated primary CAFs and NFs from human NSCLC tissues and adjacent normal tissues and identified the characteristics of fibroblasts (Fig. S2A-C). By performing circRNA microarray screening of CAFs and NFs, we identified 62 upregulated circRNAs in CAFs (cutoff value: |Fold change|> 1.5, P values < 0.05, Mean > 7, Fig. 2A and Fig. S3A). Then, we selected the top 5 upregulated circRNAs as candidate circRNAs based on a competing endogenous RNA (ceRNA) network and detected 2 circRNAs (circ_0064142 and circ_0023988) that were enriched in CAFs through gel electrophoresis (Fig. S3A-C). However, Sanger sequencing of the reverse transcription product failed to confirm the circular splicing of circ_0064142, so we chose circ_0023988 (circNOX4) for further identification. Then, qRT-PCR analysis showed that circNOX4 was significantly upregulated in CAFs compared to NFs, human NSCLC cell lines (A549, PC9, H226 and H1581) and bronchial epithelial cells (BEAS-2B), indicating a potential role in fibroblast activation (Fig. 2B, C). Next, we identified the origin of circNOX4, which was generated from exons 6–11 of the NOX4 transcript (chr11:89,155,069–89185063) with a length of 476 nt. The back-splice junction site of circNOX4 was confirmed by Sanger sequencing (Fig. 2D). To further verify the circular property of circNOX4, the back-splice junction site of circNOX4 was amplified by RT–PCR using convergent and divergent primers, which only amplified a specific divergent band in cDNA but not in gDNA (Fig. 2E). Additionally, circNOX4 was RNase R-resistant, whereas the control linear NOX4 and GAPDH mRNAs were not (Fig. 2F). These results indicate that circNOX4 is a bona fide circRNA highly expressed in CAFs.

To determine whether circNOX4 was clinically relevant in NSCLC patients, we performed fluorescence in situ hybridization (FISH) in 84 NSCLC specimens. The results showed that circNOX4 was expressed in the intratumoral stroma but not in cancer cells (CK19-positive areas) (Fig. 2G). Additionally, a higher circNOX4 expression level was correlated with distant metastasis (Fig. 2G, Table 1). Next, we investigated the correlation between circNOX4 expression and patient survival. The overall survival of NSCLC patients with high circNOX4 expression was significantly shorter than that of patients with low circNOX4 expression (P < 0.0001; Fig. 2H). Univariate and multivariate Cox regression models revealed that high circNOX4 expression and distant metastasis were independent prognostic factors for NSCLC (Fig. 2I, J). Collectively, our findings suggest that a CAF-specific circRNA, circNOX4, correlates with metastatic progression and may be a prognostic marker in NSCLC patients.

To delineate the functional role of circNOX4 in the intrinsic characteristics of fibroblasts, we knocked down circNOX4 in CAFs or overexpressed circNOX4 in NFs without altering the expression of the parental gene NOX4 (Fig. S4A-C). Then, we speculated that circNOX4 could regulate the phenotype of fibroblasts. The CCK-8 and collagen contraction assays exhibited that circNOX4 overexpression significantly enhanced the proliferation and contraction capabilities of fibroblasts, while circNOX4 knockdown exerted the opposite effects (Fig. 3A, B). Afterwards, we examined how circNOX4 affected the expression of CAF markers and regulators, including FAP, MMP2, MMP14, collagen type 1 alpha 1 (COL1A1) and PDGFRβ. Interestingly, qRT-PCR suggested a reversion of the above gene expression profiles. Notable among them was FAP, which displayed the greatest reduction following circNOX4 knockdown (Fig. 3C). In parallel, circNOX4 overexpression upregulated FAP expression while circNOX4 knockdown attenuated FAP expression, as shown by western blotting and qRT-PCR analysis (Fig. 3D and Fig. S4D). To extend our understanding of circNOX4 in CAF activation, we detected circNOX4 expression in NFs after stimuli of TGF-β1, a critical mediator of fibroblast activation and cancer-stroma bidirectional communication within the TME (Fig. S5A). Consequently, the expression of circNOX4 and FAP increased in a dose- and time-dependent manner upon exposure to TGF-β1 (Fig. S5B, C). Collectively, these findings support a potent effect of circNOX4 on fibroblast activation.

As the activation of fibroblasts is linked to protumor properties, we explored the role of circNOX4 in supporting cancer cell migration and invasion. NSCLC cell lines were treated with fibroblast-derived conditioned medium (CAF-CM and NF-CM). As expected, the wound healing assays showed that medium from sh-circNOX4 CAFs drastically decreased the migration ability of A549 and PC9 cells. In stark contrast, overexpressing circNOX4 increased the migration ability of cancer cells (Fig. 3E, F). Meanwhile, the transwell coculture system demonstrated that sh-circNOX4 CAFs reduced the invasion of A549 and PC9 cells, while circNOX4-overexpressing NFs exerted the opposite effects (Fig. 3G, H). Overall, circNOX4 facilitates NSCLC migration and invasion by activating CAFs and maintaining a protumor phonotype.

To dissect the molecular mechanisms by which circNOX4 induces fibroblast activation, we detected its intracellular distribution. The results of FISH and subcellular fractionation assays showed that circNOX4 was mainly localized in the cytoplasm of CAFs and NFs (Fig. 4A, B). It has been reported that cytoplasmic circRNAs function by sponging miRNAs or encoding peptides [23]. We next observed that circNOX4 had a limited protein-coding potential with no predicted protein features using circRNADb software (R Score < 1.6, Fig. S6A). Therefore, we propose that miRNA sponge activity could be a possible mechanism for its functional effect.

To determine the miRNAs interacting with circNOX4, we overlapped the results obtained from two different algorithms (miRanda and RNAhybrid) predicting miRNAs that could bind with circNOX4 and screened 47 candidate miRNAs (Fig. 4C). Then, through prediction algorithms of the TargetScan database, only miR-329-5p and miR-624-5p were predicted to target both circNOX4 and FAP (Fig. 4D). We therefore explored the ceRNA network including circNOX4, miR-329-5p/miR-624-5p and FAP. The qRT-PCR results showed that miR-329-5p and miR-624-5p were downregulated in CAFs compared with paired NFs (Fig. 4E and Fig. S6B). Moreover, the expression of miR-329-5p and miR-624-5p was increased by knocking down circNOX4 (Fig. 4F and Fig. S6C), suggesting that circNOX4 could negatively regulate the expression of miR-329-5p and miR-624-5p.

Next, to validate whether circNOX4 sponges miR-329-5p and miR-624-5p, we constructed luciferase reporters harboring wild-type and mutant forms of circNOX4 according to the predicted binding sites on miR-329-5p or miR-624-5p, respectively (Fig. 4G, H and Fig. S6D). Consequently, the miR-329-5p mimic could specifically target the wild-type form of circNOX4, resulting in markedly decreased luciferase activity, but not the mutant form (Fig. 4H and Fig. S6E). RIP assays further confirmed the interaction of miR-329-5p with circNOX4 in CAFs (Fig. 4I). Moreover, RNA-FISH revealed that circNOX4 and miR-329-5p co-localized in the cytoplasm of CAFs (Fig. 4J). Unfortunately, due to the low abundance of circNOX4, AGO2 was unable to accumulate circNOX4 in NFs (Fig. S6F). In addition, the miR-624-5p mimic had no effect on the luciferase activity of the circNOX4 wild-type reporter (Fig. S6G). These results support that miR-329-5p directly binds to circNOX4 in CAFs, whereas miR-624-5p does not.

Furthermore, similar experiments were conducted to confirm the binding of miR-329-5p with the 3′-untranslated region (3′UTR) of FAP. As shown in Fig. 4K and Fig. 4L, we observed a high interaction probability between the miR-329-5p RNA sequence and FAP protein by the RPISeq prediction algorithm (https://​www.​pridb.​gdcb.​iastate.​edu/​RPISeq/​), and the 3′UTR of FAP harbored specific sequences complementary to miR-329-5p. Dual-luciferase reporter assays showed decreased luciferase activity of the wild-type reporter of the FAP 3′UTR with the miR-329-5p mimic, while no significant differences were found with the mutant reporter (Fig. 4L). Additionally, the expression of FAP was significantly decreased with miR-329-5p mimic transfected into CAFs (Fig. 4M). The above evidence confirms that miR-329-5p binds to the 3' UTR of FAP in CAFs.

To verify whether circNOX4 regulated the expression of FAP through miR-329-5p, we performed rescue assays. Knockdown of circNOX4 significantly decreased the expression of FAP, while miR-329-5p inhibitor co-transfection neutralized this downregulation (Fig. 4N). Conversely, overexpression of circNOX4 significantly increased the expression of FAP, while miR-329-5p mimic co-transfection abrogated this alteration (Fig. 4O), indicating that the regulation of FAP by circNOX4 was dependent on miR-329-5p. In addition, we conducted collagen contraction assays to determine whether miR-329-5p was involved in circNOX4-induced fibroblast activation. The results showed that the inhibitory effect of circNOX4 knockdown was reversed by the miR-329-5p inhibitor (Fig. 4P). In summary, our results suggest that circNOX4 sponges miR-329-5p to upregulate FAP and induce fibroblast activation.

To evaluate how circNOX4 contributes to the TME and cancer progression in vivo, we constructed two different mouse models: the xenograft tumor model and the lung metastasis model. For xenograft tumor model establishment, A549 cells were subcutaneously co-transplanted with or without CAFs isolated from NSCLC tissue samples (Fig. 5A). We found that xenografts derived from co-transplantation with CAFs resulted in enhanced tumor volume and tumor weight compared with xenografts derived from A549 cells alone, suggesting that CAFs promoted NSCLC progression (Fig. 5B, and Fig. S7A). More strikingly, sh-circNOX4 CAFs significantly stunted tumorigenesis compared with A549 cells co-transplanted with shNC-CAFs (Fig. 5B, and Fig. S7A), indicating that knocking down circNOX4 efficiently disabled CAFs and impeded the development of invasive tumors.

Next, to intuitively visualize the expression of FAP in vivo, we utilized micro-positron emission tomography (PET)/CT imaging with a specific nuclide molecular probe (⁶⁸ Ga-FAPI-04). Micro-PET/CT revealed tracer uptake in the tumor xenografts, as shown in the 3D MIP images (Fig. 5C). The quantitative results of FAPI-PET imaging suggested that SUV uptake in the xenografts was significantly reduced in the sh-circNOX4 group compared to the control group (Fig. 5D), indicating that knocking down circNOX4 could downregulate FAP expression and reverse fibroblast activation in vivo. Furthermore, we examined the effect of circNOX4 on the expression of FAP, the endothelial marker CD31, mesenchymal markers (N-cadherin and Vimentin), and metastasis-related genes (MMP2 and MMP9). All these factors are critically involved in angiogenesis and epithelial-mesenchymal transformation (EMT), which contribute to the process of metastasis [24, 25]. The IHC results showed that CAFs significantly increased the expression levels of FAP, CD31, N-cadherin, Vimentin, and MMPs. However, these effects were reversed by transfecting sh-circNOX4 into CAFs, implying that circNOX4 provides a pro-tumorigenic niche that facilitates metastasis in vivo (Fig. 5E and Fig. S7B, C).

To further substantiate the effect of circNOX4, we constructed lung metastasis mouse models. In brief, A549 cells were either injected through the tail vein or "primed" in a coculture system with CAFs for 48 h in vitro prior to injection (Fig. 5F). As shown in Fig. 5G, more lung metastatic foci were detected in mice that had been injected with A549 cells "primed" with fibroblasts, compared with mice injected with the control A549 cells. Notably, a lower number of lung metastatic foci was presented in the sh-circNOX4-CAF group than in the shNC-CAF group, as evidenced by reduced HE-stained tumors in the lungs (Fig. 5G, H). All these data thus suggest that circNOX4 potentiates tumor metastasis by inducing fibroblast activation in vivo.

The activation of fibroblasts coincides with paracrine signaling changes in the TME. To examine whether circNOX4 promotes metastasis in a paracrine manner, we collected supernatants from sh-NC CAFs and sh-circNOX4 CAFs and adopted a cytokine antibody array. The results showed that sh-circNOX4 CAFs secreted a decreasing panel of cytokines, including interleukin-6 (IL-6), chemokine C–C motif ligand 2 (CCL2), TGF-β2 and insulin-like growth factor binding protein 4 (IGFBP-4), which are reported as niche characteristic genes that facilitate cancer cell migration, invasion, and colonization in the metastatic process [26] (Fig. 6A and Fig. S8A). Additionally, Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis showed enrichment of inflammatory signaling pathways and cytokine–cytokine receptor interaction pathways (Fig. 6B). Further validation by ELISA indicated that IL-6 was the most dramatically downregulated cytokine in sh-circNOX4 CAFs (Fig. 6C). Previous studies have revealed that IL-6, as one of the key pro-inflammatory cytokines in the TME, plays a pivotal role in initiating pro-metastatic inflammatory responses and immune cell recruitment. These results encouraged us to focus on IL-6 for further investigation.

To confirm whether IL-6 is mainly secreted by the CAFs, we initially detected the mRNA level of IL-6 in both CAFs and cancer cells. The results revealed significantly higher expression levels of IL-6 in CAFs compared to A549 and PC9 cells (Fig. S8B). Subsequently, we knocked down IL-6 in CAFs or cancer cells separately and established a transwell coculture system to measure IL-6 secretion level (Fig. 6D and Fig. S8C). ELISA analysis indicated that knockdown of IL-6 in CAFs drastically reduced its secretion within the coculture system; however, knockdown of IL-6 in A549 or PC9 only slightly lowered its secretion within the coculture system (Fig. 6E, F). We also found that IL-6 was primarily derived from the tumor stroma and fibroblast subsets in both the human and murine TMEs through bioinformatics analysis (Fig. S8D-F). Thus, these findings support that CAFs are the primary source for secreting IL-6.

Then, we determined the role of the FAP/IL-6 axis in CAF activation. Single-cell analysis from the TISCH database indicated that both FAP and IL-6 were enriched in subsets of fibroblasts (Fig. S8G). Moreover, IL-6 was positively correlated with FAP expression in the TCGA database of LUAD and LUSC cohorts (Fig. 6G). Gene set enrichment analysis (GSEA) showed substantial enrichment signatures of IL-6 signaling in NSCLC patients with high FAP expression (Fig. 6H). Importantly, the IL-6 secretion level in circNOX4-overexpressing NFs was significantly increased compared with the corresponding control, while a FAP inhibitor (linagliptin, 5 nm) hindered the CAF-secreted IL-6 level upon circNOX4 overexpression, suggesting that circNOX4 induced IL-6 secretion via FAP (Fig. 6I).

To examine whether the pro-tumoral function of circNOX4 on cancer cells was mediated by stimulating paracrine IL-6, rescue experiments involving circNOX4 and IL-6 were carried out. The wound healing assay and transwell assay together revealed that neutralizing IL-6 in a conditioned medium from circNOX4-overexpressing NFs greatly abrogated the migration and invasion capacity of NSCLC cells (Fig. 6J, K). In contrast, the exogenous addition of recombinant human IL-6 (rhIL-6) partially reversed the effects of circNOX4 knockdown on the ability of NSCLC cells to invade (Fig. 6K). Collectively, these results demonstrate that circNOX4 elicits a pro-metastatic inflammatory microenvironment by accelerating IL-6 secretion within the fibroblast niche and consequently boosts NSCLC cell migration and invasion.

To investigate the potential therapeutic role of the circNOX4/IL-6 axis, we established xenograft tumor models and performed experimental therapy with IL-6 antibodies (Fig. 7A). A549 cells were co-injected with CAFs into the axilla of nude mice. After fourteen days after implantation, the mice received an intratumoral injection of IL-6 neutralizing antibody (2 mg/kg) twice a week. Conceivably, overexpression of circNOX4 enhanced CAF-mediated tumor growth compared to the control group. Of note, injection of the anti-IL-6 neutralizing antibody led to a reduction in tumor volume and weight (Fig. 7B-D). In further IHC analysis, the expression levels of IL-6 were increased in circNOX4-overexpressing xenograft tumors, accompanied by the increased expression of FAP, CD31, N-cadherin, Vimentin and the matrix metalloproteinases MMP2 and MMP9. However, the expression of the metastasis-related factors was partly reversed by treatment with anti-IL-6 antibody (Fig. 7E and Fig. S9A, B), indicating that IL-6 acted as a soluble mediator of remodeling the local inflammatory niche and metastatic microenvironment. In conclusion, we demonstrate that circNOX4 activates an inflammatory fibroblast niche to promote tumor progression. Our findings emphasize that the circNOX4/FAP/IL-6 axis could be exploited as a therapeutic target for CAF-based cancer therapy (Fig. 8).