CircSMAD4 shapes matrix-remodeling TAMs in lung adenocarcinoma

Abstract

Lung adenocarcinoma (LUAD) progression is strongly shaped by tumor-associated macrophages (TAMs), yet the post-transcriptional mechanisms that sustain matrix-remodeling TAM states remain incompletely understood. Here, circRNA profiling of LUAD TAMs versus normal tissue-resident macrophages identified circSMAD4 (hsa_circ_0047713) as a consistently TAM-enriched circRNA associated with advanced clinicopathological features and unfavorable survival. circSMAD4 exhibited canonical circular properties, including a validated back-splice junction, RNase R resistance, and enhanced transcript stability. Functionally, circSMAD4 knockdown in human and murine macrophages attenuated tumor education, shifted macrophages away from an M2-like phenotype, and weakened their ability to promote LUAD-cell proliferation, invasion, and EMT-like changes in co-culture. In syngeneic orthotopic lung and experimental metastasis models, circSMAD4-depleted macrophages restrained tumor growth and reduced metastatic burden. Mechanistically, cytoplasmic circSMAD4 acted as a ceRNA to sequester miR-562 and relieve repression of COL4A1. In parallel, circSMAD4 formed a specific ribonucleoprotein complex with the m6A reader IGF2BP2, facilitating IGF2BP2 association with COL4A1, ACTA2, and SPI1 transcripts and enhancing their m6A-dependent stability. Together, these dual branches converge on a matrix-remodeling output, positioning circSMAD4 as a post-transcriptional hub that reinforces protumor TAM programs in LUAD and a potential target for microenvironment-directed therapy.

1. Introduction

Lung adenocarcinoma (LUAD) is the most common histological subtype of non-small cell lung cancer and remains a leading cause of cancer-related mortality worldwide [[1], [2], [3], [4]]. Although targeted therapy and immunotherapy provide substantial benefits to subsets of patients, relapse, metastasis, and therapy resistance are still frequent, highlighting the persistent, driving influence of the tumor microenvironment (TME) on disease progression [[5], [6], [7]]. Defining the key cellular populations and regulatory networks within the TME is therefore essential for understanding LUAD malignancy and identifying new therapeutic vulnerabilities.

Tumor-associated macrophages (TAMs) represent a highly abundant and exceptionally plastic immune population in LUAD, and are often associated with immunosuppression, poor therapeutic responses, and unfavorable prognosis [[8], [9], [10]]. Macrophage activation spans a continuum, for which the M1/M2 framework is commonly used as an operational descriptor; in solid tumors, TAMs frequently adopt immunosuppressive and tissue-repair programs with substantial spatial and temporal heterogeneity [[11], [12], [13]]. Among TAM outputs, matrix remodeling is particularly relevant [14,15]. Extracellular matrix (ECM) deposition and remodeling can reshape tissue mechanics and signaling, facilitate tumor cell invasion and migration, and promote immune exclusion and suppressive niches, thereby accelerating metastasis and resistance [13,16,17]. Understanding how TAMs acquire and maintain a matrix-remodeling phenotype is thus central to linking immune regulation with metastatic biology.

At the molecular level, post-transcriptional regulation is increasingly recognized as a key layer controlling TAM state maintenance and transitions [[18], [19], [20], [21], [22]]. Circular RNAs (circRNAs) are covalently closed transcripts with enhanced stability and cell-type/context specificity across tissues and disease states [[23], [24], [25]]. Functionally, circRNAs can act as competitive endogenous RNAs (ceRNAs) that sequester miRNAs to de-repress miRNA-targeted mRNAs [[24], [25], [26]]. In addition, circRNAs can serve as scaffolds to assemble RNA-binding proteins (RBPs) into ribonucleoprotein (RNP) complexes, influencing RNA processing, localization, stability, and translation [[27], [28], [29]]. However, it remains unclear whether distinct lung macrophage subsets (e.g., normal tissue-resident macrophages versus TAMs) exhibit systematic circRNA remodeling in LUAD, and how circRNAs mechanistically integrate miRNA and RBP pathways to reinforce protumor programs in macrophages.

N6-methyladenosine (m6A) is the most prevalent internal modification on mammalian mRNAs and regulates RNA fate through writer–eraser–reader systems [[30], [31], [32]]. At the reader level, multiple classes of RNA-binding proteins can recognize m6A-marked transcripts or assemble functional complexes on these RNAs, thereby modulating transcript stability and translation and contributing to oncogenic gene programs [32,33]. In macrophages, an important open question is whether circRNAs can engage specific RBPs to reshape RBP–mRNA interactions and post-transcriptional stability control, ultimately reinforcing immunosuppressive or matrix-remodeling outputs. The IGF2BP family represents one group of reported m6A readers that can enhance the stability and translation of selected targets [30,32,[34], [35], [36]], yet how such m6A reader–centered mechanisms operate within protumor TAM programs remains to be systematically defined.

Here, we performed circRNA profiling of LUAD TAMs versus normal tissue-resident macrophages (NTRMs) and identified circSMAD4 as a TAM-enriched circRNA. We validated its expression features and prognostic association in clinical specimens and characterized its canonical circular properties, including the back-splice junction, RNase R resistance, and increased stability [25,28,37]. Functionally, circSMAD4 silencing in human CD14⁺ monocyte–derived macrophages and murine BMDMs, combined with tumor–cell co-culture assays, demonstrated that circSMAD4 promotes tumor-educated M2-like polarization and enhances tumor-cell proliferation, invasion, and EMT-associated phenotypes. In orthotopic lung and tail-vein models, circSMAD4 depletion markedly suppressed tumor growth and reduced metastasis.

Mechanistically, we show that cytoplasmic circSMAD4 acts as a ceRNA that binds miR-562 and de-represses COL4A1, supported by candidate prioritization, AGO2 dependency, reporter assays, and functional rescue experiments [[38], [39], [40]]. In parallel, RNA pull-down and RIP assays reveal a specific circSMAD4–IGF2BP2 RNP complex. Integrative transcriptomic analyses further indicate that IGF2BP2 enhances the stability of COL4A1, ACTA2 (α-SMA), and SPI1 transcripts, thereby driving an M2-like, matrix-remodeling TAM program [[41], [42], [43]]. Collectively, our findings define a circSMAD4-centered, dual-branch post-transcriptional framework that links miRNA and RBP/m6A pathways to reinforce matrix-remodeling TAM outputs in LUAD.

2. Materials and methods

2.1. Cell lines and cell culture

A549 (human lung adenocarcinoma) cells were obtained from the Cell Bank of the Chinese Academy of Sciences (Shanghai, China). Lewis lung carcinoma (LLC, mouse lung cancer) cells were obtained from ATCC. A549 and LLC cells were cultured in high-glucose DMEM (MeilunBio, MA0212) supplemented with 10% FBS (MeilunBio, PWL217) and penicillin/streptomycin solution (MeilunBio, MA0110) at 37 °C in a humidified incubator with 5% CO₂. A stable luciferase-expressing LLC line (LLC-Luc) was established for in vivo bioluminescence imaging. All cell lines were routinely tested negative for mycoplasma contamination, and experiments were performed using cells within 15 passages.

2.2. Primary macrophage differentiation

Human monocyte-derived macrophages (hMDMs) were generated from peripheral blood mononuclear cells of healthy donors with informed consent and ethical approval. CD14⁺ monocytes were isolated and differentiated in RPMI-1640 medium (MeilunBio, MA0215) containing 10% FBS (MeilunBio, PWL217) and human M-CSF (MCE, HY-P7050) for 7 days. Mouse bone marrow–derived macrophages (BMDMs) were generated from C57BL/6 mice (6–8 weeks old) by culturing bone marrow cells in high-glucose DMEM (MeilunBio, MA0212) supplemented with 10% FBS (MeilunBio, PWL217) and mouse M-CSF (MCE, HY-P7085) for 7 days. Tumor-conditioned macrophages (TC-hMDMs and TC-BMDMs) were generated using the Transwell co-culture system described below.

2.3. Clinical tissue dissociation and magnetic bead–based macrophage enrichment (MACS)

Human LUAD tumors and paired adjacent normal lung tissues were collected from patients undergoing surgery at the Department of Cardiothoracic Surgery, Xinhua Hospital Affiliated to Shanghai Jiao Tong University School of Medicine, with written informed consent and approval from the institutional ethics committee. Fresh tissues were minced and digested in RPMI-1640 (MeilunBio, MA0215) containing Collagenase IV (Sigma, C5138; 1 mg/mL) and DNase I (Selleck, E7018; 100 μg/mL) for 30 min at 37 °C. Cell suspensions were passed through a 70-μm strainer and red blood cells were removed using RBC Lysis Buffer (biogems, 64010-00).

Macrophages were enriched by magnetic activated cell sorting (MACS). For adjacent normal tissues, cells were labeled with PE anti-human CD11b (BioLegend, Cat# 379903) followed by Anti-PE MicroBeads (Miltenyi Biotec, 130-048-801), and CD11b⁺ NTRMs were collected using a MACS separator and columns. For tumor tissues, cells were labeled with PE anti-human CD163 (BioLegend, Cat# 333606) followed by Anti-PE MicroBeAds (Miltenyi Biotec, 130-048-801), and CD163⁺ TAMs were collected in parallel. Enrichment efficiency was evaluated by flow cytometry and analyzed with FlowJo.

Purified TAMs and NTRMs were used immediately for RNA extraction/sequencing or cultured short-term in RPMI-1640 supplemented with 10% FBS (MeilunBio, PWL217), penicillin/streptomycin (MeilunBio, MA0110), and human M-CSF (MCE, HY-P7050).

2.4. shRNA-mediated knockdown

Gene silencing was performed using lentiviral shRNAs. shRNAs targeting the human circSMAD4 back-splice junction (and the murine circSmad4 ortholog, avoiding linear Smad4) as well as IGF2BP2 were cloned into the pLKO.1-puro vector (Addgene, #8453). A non-targeting shRNA (shNC) served as the control. Lentiviruses were produced in 293T cells by co-transfecting the shRNA plasmid with psPAX2 (Addgene, #12260) and pMD2.G (Addgene, #12259) using Lipofectamine 3000 (Invitrogen, Cat# L3000008). Viral supernatants were harvested at 48 h, filtered (0.45 μm), and used to infect target cells in the presence of Polybrene (Beyotime, Cat# C0351; 8 μg/mL). After infection, cells were selected with puromycin (InvivoGen, Cat# ant-pr-1; 2 μg/mL) for 5–7 days. Knockdown efficiency was verified by qRT–PCR and, where applicable, by Western blotting.

2.5. Plasmid overexpression and RNA transfection

For circSMAD4 overexpression, the full-length circSMAD4 sequence was cloned into a circRNA mini-vector. IGF2BP2 expression plasmids (full-length and truncation mutants) were purchased from Shanghai BioeGene Co., Ltd. Transient plasmid transfections were performed using Lipofectamine 3000 (Invitrogen, Cat# L3000008) according to the manufacturer's instructions. miR-562 mimic/inhibitor and corresponding negative controls were transfected using Lipofectamine RNAiMAX (Invitrogen, Cat# 13778150). The sequences of shRNAs, miRNA oligonucleotides, and primers are listed in Supplementary Table S1.

2.6. Tumor cell–macrophage co-culture

To generate tumor-educated macrophages in vitro, indirect co-culture was performed using Transwell inserts with 0.4 μm pores (Corning, Cat# 3450) to permit soluble-factor exchange without direct cell contact. Macrophages (human monocyte-derived macrophages or murine BMDMs) were seeded in the upper inserts (2 × 10^5 cells/insert), while lung cancer cells were plated in the lower wells (A549 for human setting; LLC for mouse setting; 1 × 10^5 cells/well). Co-cultures were maintained in complete medium for 48 h.

After co-culture, macrophages were collected for downstream analyses, and supernatants were harvested for cytokine measurements. Human IL-10 was quantified using IL-10 DuoSet ELISA (R&D Systems, Cat# DY217B), and human TGF-β1 was measured using the Human TGF beta-1 ELISA Kit (Invitrogen, Cat# BMS249-4). For mouse experiments, mouse IL-10 was measured using the Mouse IL-10 ELISA Kit (R&D Systems, Cat# M1000B), and mouse TGF-β1 was measured using the Mouse TGF beta-1 ELISA Kit (Invitrogen, Cat# BMS608-4), following the manufacturers’ instructions.

2.7. Cell proliferation and viability assays

Tumor-cell proliferation was measured using Cell Counting Kit-8 (CCK-8; MedChemExpress, Cat# HY-K0301). Tumor cells (2 × 10^3 per well) were seeded in 96-well plates and cultured in macrophage-conditioned medium. CCK-8 reagent was added at the indicated time points, and absorbance at 450 nm was recorded after 2 h to estimate viable cell numbers.

For colony formation, tumor cells (1000 per well) were seeded in 6-well plates and cultured for 10–14 days in conditioned medium. Colonies were fixed with 4% paraformaldehyde, stained with 0.5% crystal violet, and colonies containing ≥50 cells were counted.

Patient-derived LUAD organoids were cultured in organoid medium and treated with 50% macrophage-conditioned medium. Organoid viability was quantified using CellTiter-Glo 3D Cell Viability Assay (Promega, Cat# G9681) according to the manufacturer's instructions.

2.8. Transwell migration and invasion assays

Tumor-cell migration and invasion were assessed using 24-well Transwell inserts (Corning, Cat# 3422). Tumor cells (5 × 10^4) were suspended in 200 μL serum-free medium and seeded into the upper chamber. For invasion assays, inserts were pre-coated with Matrigel basement membrane matrix (BD Biosciences, Cat# 354234) and allowed to polymerize before cell seeding. The lower chamber was filled with 600 μL macrophage-conditioned medium or co-culture supernatant as the chemoattractant. After 24 h, cells on the lower surface of the membrane were fixed with 4% paraformaldehyde, stained with 0.1% crystal violet, and quantified in five random microscopic fields per insert.

2.9. Animal models

C57BL/6 mice (6–8 weeks old) were maintained under specific pathogen-free conditions. All animal experiments were approved by the institutional ethics committee, and mice were randomly allocated to groups.

For the orthotopic lung tumor model, LLC cells were mixed 1:1 with syngeneic BMDMs (2 × 10^5 cells each in 50 μL PBS) immediately before implantation. Mice were anesthetized with isoflurane, and the cell suspension was injected into the left lung lobe via an intercostal approach. At 4 weeks after implantation, mice were euthanized and tumors were collected for tumor-burden assessment (tumor weight). In a separate cohort (n = 6 per group), overall survival was monitored and analyzed by Kaplan–Meier curves.

For the experimental metastasis model, mice received a tail-vein injection of a 1:1 mixture of LLC-Luc cells and syngeneic BMDMs (2 × 10^5 cells each in 50 μL PBS). Metastatic burden was evaluated by bioluminescence imaging. Briefly, D-luciferin (Yeasen, Cat# 40901ES02; 150 mg/kg, i.p.) was administered 10 min before imaging, and signals were acquired on a Tanon ABL X5 In Vivo Optical Imaging System. Mice were euthanized at humane endpoints, and tissues were collected for downstream analyses.

2.10. Histology and immunohistochemistry (IHC)

Tissue samples were fixed in 10% neutral buffered formalin, paraffin-embedded, and sectioned at 4 μm. Sections were deparaffinized and rehydrated, followed by heat-induced antigen retrieval using citrate antigen-retrieval buffer (pH 6.0; Servicebio, Cat# G1203). Endogenous peroxidase activity was quenched with 3% H₂O₂ (Abcarta, Cat# PK001), and nonspecific binding was blocked with normal goat serum blocking solution (CST, Cat# 5425S). Sections were incubated with primary antibodies against Ki-67 (Servicebio, Cat# GB111499), E-cadherin (Proteintech, Cat# 20874-1-AP), and Vimentin (Proteintech, Cat# 10366-1-AP). After washing, an HRP-conjugated secondary antibody was applied (Abcarta, Cat# PK001), and signals were developed with DAB substrate (Abcarta, Cat# PK001). Nuclei were counterstained with hematoxylin (Servicebio, Cat# G1004). IHC signals were evaluated in multiple random fields per sample and scored independently by two blinded observers.

2.11. Immunofluorescence (IF)

Paraffin sections were deparaffinized and rehydrated, followed by antigen retrieval with citrate antigen-retrieval buffer (pH 6.0; Servicebio, Cat# G1203) and blocking with normal goat serum blocking solution (CST, Cat# 5425S). Human LUAD sections were stained with anti-CD68 (ABclonal, Cat# A23205), and murine lung tumor sections were stained with anti-F4/80 (ABclonal, Cat# A23788). After primary antibody incubation, HRP-conjugated secondary antibody was applied (Abcarta, Cat# PK001) and signals were amplified using TSA dyes (Runnerbio, Cat# bry-0067-050K). Nuclei were counterstained with DAPI (ABclonal, Cat# RM02978). Images were acquired using a Zeiss confocal microscope.

2.12. RNA in situ hybridization (ISH)

circSMAD4 signals were detected on paraffin-embedded sections using an RNA fluorescence in situ hybridization kit (Servicebio, RNASweAMI™ ISH CY3 Detection Kit, Cat# GF002-50T) with a custom-designed probe targeting the circSMAD4 back-splice junction. Sections were processed following the manufacturer's protocol, including endogenous enzyme blocking, heat-mediated target retrieval, and protease treatment. Hybridization was carried out at 40 °C for 2 h, followed by signal amplification and CY3-based fluorescence detection. For combined ISH/IF analyses, immunofluorescence staining was performed after completion of the ISH procedure. Images were acquired using a Zeiss confocal microscope. Probe information is provided in Supplementary Table S1.

2.13. RNA extraction and qRT-PCR

Total RNA was extracted using the SteadyPure RNA Extraction Kit (Accurate Biology, Cat# AG21024) and treated for genomic DNA removal (Vazyme gDNA Wiper). Reverse transcription for mRNA/circRNA was performed with HiScript III (Vazyme, Cat# R312-01), and miRNA reverse transcription was performed using the Vazyme miRNA RT system. qPCR was carried out using ChamQ SYBR qPCR Master Mix (Vazyme, Cat# Q331) on a qTOWER3 84 G Real-Time PCR Thermal Cycler. circSMAD4 was quantified using divergent primers spanning the back-splice junction, whereas linear transcripts were detected using convergent primers. GAPDH or β-Actin served as internal controls for mRNA/circRNA, and U6 served as the control for miRNA; relative expression was calculated using the 2^-ΔΔCt method. For circRNA validation, RNase R digestion was performed using RNase R (Beyotime, Cat# R7092L) prior to qPCR to confirm RNase R resistance. For RNA stability assays, cells were treated with actinomycin D (Sigma, Cat# A9415; 5 μg/mL), and total RNA was collected at the indicated time points for RT–qPCR analysis. For mRNA decay assays (e.g., COL4A1, ACTA2, and SPI1), samples were harvested from 0 to 24 h. For circSMAD4 stability assays, samples were harvested from 2 to 24 h, and transcript half-lives were estimated by one-phase decay fitting. All primer sequences are provided in Supplementary Table S1.

2.14. Western blotting

Cells or tissues were lysed in RIPA Lysis Buffer (Epizyme, Cat# PC102) supplemented with a protease and phosphatase inhibitor cocktail (MedChemExpress, Cat# HY-K0013). Protein concentrations were measured using a BCA Protein Assay Kit (Vazyme, Cat# E112-02). Equal amounts of protein were separated by SDS–PAGE, transferred onto a PVDF transfer membrane (LIAODA, Cat# YQPVDF223Z), blocked with 5% non-fat milk, and incubated with primary antibodies at 4 °C overnight. Membranes were then incubated with an HRP polymer–conjugated goat anti-mouse/rabbit IgG secondary antibody (ABclonal, Cat# AS080), and signals were developed using an ECL substrate (Meilunbio, Cat# MA0186-2). Band intensities were quantified by densitometry using ImageJ. Primary antibodies included IGF2BP2 (ABclonal, Cat# A5189), COL4A1 (ABclonal, Cat# A25131), E-cadherin (Proteintech, Cat# 20874-1-AP), N-cadherin (ABclonal, Cat# A0433), Vimentin (Proteintech, Cat# 10366-1-AP), α-SMA/ACTA2 (ABclonal, Cat# A7248), and β-Actin (ABclonal, Cat# AC026).

2.15. Flow cytometry analysis

Macrophages were harvested, washed, and stained with the viability dye Zombie NIR™ (BioLegend, Cat# 423105) according to the manufacturer's instructions to exclude dead cells, and then stained for surface markers with fluorophore-conjugated antibodies on ice for 30 min. Cells were then fixed and permeabilized using Fixation/Permeabilization Buffer (Thermo Fisher eBioscience, Cat# 88-8824-00) and stained for intracellular markers on ice. Samples were acquired on a BD LSRFortessa flow cytometer. Compensation was performed using single-stained controls, and fluorescence-minus-one (FMO) controls were included for gate setting. Data were analyzed using FlowJo (v10.8).

For TC-hMDMs (human), surface staining used BV510 anti-human CD45 (BioLegend, Cat# 368525), BV785 anti-human CD68 (BioLegend, Cat# 333825), BV421 anti-human HLA-DR (BioLegend, Cat# 307635), APC anti-human CD86 (BioLegend, Cat# 305411), PE anti-human CD206 (BioLegend, Cat# 321105), and PerCP/Cy5.5 anti-human CD163 (BioLegend, Cat# 333607), and intracellular staining used Alexa Fluor 488 anti-iNOS/NOS2 (Novus Biologicals, Cat# NBP2-22119AF488) and PE-Cy7 anti-ARG1 (Invitrogen/eBioscience, Cat# 25-3697-82).

For TC-BMDMs (mouse), surface staining used PE anti-mouse CD45 (BioLegend, Cat# 157603), PerCP/Cy5.5 anti-mouse CD68 (BioLegend, Cat# 137009), BV421 anti-mouse I-A/I-E (BD Biosciences, Cat# 562564), BV605 anti-mouse CD86 (BioLegend, Cat# 105037), APC anti-mouse CD206 (BioLegend, Cat# 141707), and BV711 anti-mouse CD163 (BioLegend, Cat# 155325), and intracellular staining used Alexa Fluor 488 anti-iNOS/NOS2 (Novus Biologicals, Cat# NBP2-22119AF488) and PE-Cy7 anti-ARG1 (Invitrogen/eBioscience, Cat# 25-3697-82).

2.16. RNA pull-down assay

Biotin-labeled RNA pull-down was performed to identify miRNAs and RNA-binding proteins interacting with circSMAD4. Biotinylated circSMAD4 sense RNA (with antisense RNA as a negative control) was generated by in vitro transcription using Biotin-16-UTP (Beyotime, Cat# D7336), incubated with macrophage lysates, and captured using streptavidin beads (Invitrogen, Cat# 88816). RNA recovered from bead-bound complexes was analyzed by RT–qPCR to quantify candidate miRNAs. For protein discovery, bead-bound proteins were analyzed by SDS–PAGE followed by LC–MS/MS, and key candidates were validated by Western blot using an anti-IGF2BP2 antibody (ABclonal, Cat# A5189). To map the IGF2BP2-binding region, biotinylated circSMAD4 fragments were generated and tested in parallel pull-down assays, and IGF2BP2 binding was evaluated by immunoblotting.

2.17. RNA immunoprecipitation (RIP) and m6A-RIP (MeRIP)

RIP was performed to validate circSMAD4-associated RNP complexes. Cell lysates were incubated with an anti-AGO2 antibody (Abcam, Cat# ab156870) or control IgG, and immune complexes were captured using Protein A/G Magnetic Beads (Thermo Fisher Scientific, Cat# 88802). Co-precipitated RNAs were extracted and analyzed by RT–qPCR to quantify enrichment of circSMAD4 and candidate miRNAs/mRNAs relative to IgG controls. IGF2BP2-RIP was performed similarly using an anti-IGF2BP2 antibody (ABclonal, Cat# A5189) to assess enrichment of circSMAD4 and candidate mRNAs. For IGF2BP2 truncation mapping, Flag-tagged IGF2BP2 constructs were immunoprecipitated using an anti-Flag antibody (ABclonal, Cat# AE092), and the co-precipitated circSMAD4 was quantified by RT–qPCR (anti-Flag RIP–qPCR). For MeRIP, fragmented total RNA was immunoprecipitated using an anti-m6A antibody (Abcam, Cat# ab208577) following the Magna MeRIP m6A Kit workflow (Sigma, Cat# 17-10499), and enrichment of m6A-containing regions in target transcripts was quantified by RT–qPCR. Primer information is provided in Supplementary Table S1.

2.18. Dual-luciferase reporter assays

Dual-luciferase reporter assays were performed to evaluate the circSMAD4–miR-562 interaction and miR-562 targeting of COL4A1. A ∼1.4 kb fragment of the MIR562 promoter was cloned into the pGL4.10 [luc2] vector (Promega, Cat# E6651), and a Renilla luciferase plasmid was co-transfected as an internal control. The circSMAD4 fragment containing the predicted miR-562 binding site and the COL4A1 3′UTR were cloned into the pmirGLO Dual-Luciferase vector (Promega, Cat# E1330) or the psiCHECK-2 vector (Promega, Cat# C8021). Corresponding seed-site mutants and m^6A-site mutants were generated using the Mut Express II Fast Mutagenesis Kit V2 (Vazyme, Cat# C214). Cells were co-transfected with reporter plasmids together with miR-562 mimics/inhibitor or control oligonucleotides, and/or circSMAD4 or IGF2BP2 expression plasmids according to the experimental design. At 48 h post-transfection, luciferase activities were measured using the Dual Luciferase Reporter Assay Kit (Vazyme, Cat# DL101-01), and relative luciferase activity was calculated by normalizing Firefly to Renilla signals. All assays were performed in technical triplicates and repeated in at least three independent experiments.

2.19. RNA sequencing and bioinformatic analysis

mRNA-seq was performed on tumor-conditioned macrophages transduced with sh-circSMAD4 versus sh-NC and sh-IGF2BP2 versus sh-NC (n = 3 biological replicates per group). After rRNA depletion, strand-specific libraries were prepared (NEB) and sequenced on an Illumina NovaSeq platform. Reads were quality-filtered with Trimmomatic, aligned to hg38 using HISAT2, and quantified with featureCounts. Differential expression was analyzed in R using DESeq2 (adjusted P < 0.05; fold-change cutoff as defined in the study). To identify shared downstream targets, differentially expressed genes from the two knockdown datasets were intersected with IGF2BP2 target predictions from ENCORI/starBase, yielding common candidates including COL4A1, ACTA2, and SPI1.

For circRNA profiling, circRNA-seq was performed on sorted human TAMs and NTRMs (n = 3 donors). Back-splice junction reads were identified using algorithms such as find_circ and CIRI2, circRNA abundance was quantified based on junction reads, and differential circRNA expression between TAMs and NTRMs was determined using standard statistical workflows in R.

2.20. Statistical analysis

Data are presented as mean ± SEM. Statistical analyses and graphing were performed using GraphPad Prism 9.0, and R (v4.2.0) was used for RNA-seq analyses, differential expression, and visualization. Flow cytometry data were analyzed with FlowJo v10, and image quantification was performed using ImageJ; figures were assembled in Adobe Illustrator. Two-group comparisons were performed using two-tailed Student's t-test, and multi-group comparisons were analyzed using one-way or two-way ANOVA with appropriate post hoc tests. Kaplan–Meier survival curves were compared using the log-rank test, and correlations were assessed using Pearson's correlation analysis. In all analyses, P < 0.05 was considered statistically significant.

3. Results

3.1. circSMAD4 is markedly upregulated in TAMs and its high expression correlates with poor prognosis

To systematically define circRNA remodeling in LUAD-associated macrophages, single-cell suspensions were prepared from LUAD tumors and adjacent non-cancerous tissues. We profiled circRNAs from CD11b⁺ (NTRMs) and CD163⁺ cells (TAMs) to identify key pro-tumorigenic candidates (Fig. 1A). Hierarchical clustering revealed a clear separation of TAMs and NTRMs based on circRNA expression patterns (Fig. 1B), and volcano-plot analysis indicated extensive circRNA remodeling between the NTRMs and TAMs (Fig. 1C). circSMAD4 (hsa_circ_0047713) displayed a robust and consistent upregulation in TAMs (Fig. 1B and C), and was prioritized for downstream validation and mechanistic studies.

Fig. 1.

circSMAD4 is enriched in LUAD TAMs and is associated with advanced disease and poor prognosis.

(A) Workflow for isolating human LUAD TAMs and paired normal tissue-resident macrophages (NTRMs) for circRNA profiling.

(B) Heatmap of the top differentially expressed circRNAs between TAMs and NTRMs (z-score logCPM).

(C) Volcano plot of circRNA-seq (TAMs vs NTRMs) showing differentially expressed circRNAs. Differential-expression categories were defined as follows: Up (red), log2FC ≥ 1 and FDR <0.05; Down (blue), log2FC ≤ −1 and FDR <0.05; all remaining circRNAs were classified as Normal (gray).

(D) RT–qPCR validation of selected circRNA candidates in TAMs versus NTRMs.

(E) Independent cohort validation showing higher circSMAD4 expression in TAMs than in NTRMs.

(F) Representative images of combined CD68 immunofluorescence (green) and circSMAD4 ISH (red) in LUAD tumor and adjacent normal tissues; nuclei were counterstained with DAPI (blue). Scale bar, 50 μm.

(G) Kaplan–Meier analysis of overall survival stratified by circSMAD4 expression in TAMs (high vs low).

(H) Schematic of circSMAD4 genomic origin and back-splice junction validation by Sanger sequencing.

(I) Divergent-primer PCR showing circSMAD4 detection in cDNA but not gDNA; GAPDH serves as a linear control.

(J, L) RNase R digestion assay showing resistance of circSMAD4 relative to linear SMAD4 mRNA in patient-derived TAMs (J) and TC-BMDMs (L).

(K, M) Actinomycin D chase assay showing greater stability of circSMAD4 than SMAD4 mRNA in patient-derived TAMs (K) and TC-BMDMs (M). Half-life estimated by one-phase decay (Y0 = 1, Plateau = 0).

∗P < 0.05; ∗∗P < 0.01; ∗∗∗P < 0.001; ∗∗∗∗P < 0.0001; ns, not significant.

Among the shortlisted candidates, circSMAD4 showed the most robust and significant enrichment in TAMs compared with NTRMs in the initial validation set (Fig. 1D), and this difference was further confirmed in an independent TAM–NTRM cohort (Fig. 1E). Having verified the enrichment at the sorted-cell level, we extended these observations to tissue specimens. Combined CD68 immunofluorescence and circSMAD4 in situ hybridization demonstrated markedly stronger circSMAD4 signals within CD68-positive regions in tumors than in adjacent normal tissues, indicating macrophage-enriched expression in the tumor microenvironment (Fig. 1F). Clinicopathological stratification further showed higher circSMAD4 levels in cases with nodal involvement (N0 versus Nx; Fig. S1A) and distant metastasis (M0 versus M1; Fig. S1B), as well as in advanced-stage specimens relative to early-stage specimens (Fig. S1C–D). Consistently, Kaplan–Meier analysis showed an association between high circSMAD4 expression and poorer overall survival (Fig. 1G), suggesting clinical relevance.

To confirm the circular identity of circSMAD4 and minimize potential confounding from linear transcripts, we characterized its biogenesis and molecular features. Genomic annotation indicated that circSMAD4 is generated from specific SMAD4 exons via a canonical back-splicing junction, which was validated by Sanger sequencing (Fig. 1H). Using divergent primers, circSMAD4 was amplified from cDNA but not from gDNA, and this cDNA specificity was observed in both patient-derived TAMs and tumor-conditioned BMDMs (TC-BMDMs) (Fig. 1I). In RNase R assays, circSMAD4 exhibited resistance to exonuclease digestion, whereas linear SMAD4 mRNA was markedly reduced in patient-derived TAMs (Fig. 1J) and TC-BMDMs (Fig. 1L), consistent with the covalently closed structure of circRNAs [25,37]. Moreover, following transcriptional blockade by actinomycin D, circSMAD4 decayed more slowly than SMAD4 mRNA in patient-derived TAMs (Fig. 1K) and TC-BMDMs (Fig. 1M), in line with the generally longer half-life of circRNAs [25,28,44].

Collectively, Fig. 1 and Fig. S1 identify circSMAD4 as a TAM-enriched circRNA in LUAD, with localization to tumor CD68-positive macrophage regions and associations with tumor stage and overall survival.

3.2. circSMAD4 depletion restrains tumor-educated M2-like macrophage polarization and suppresses tumor growth and metastasis

The macrophage-enriched upregulation of circSMAD4 and its association with clinical outcome suggest a functional role in shaping protumor macrophage states. Because tumor microenvironments often skew macrophages toward immunosuppressive, tissue-repair/remodeling programs that facilitate tumor growth, invasion, and metastasis [13,45,46], and because macrophage activation exists along a continuum for which the M1/M2 framework is commonly used as an operational description [11,47], we performed integrated in vitro and in vivo assays to determine whether circSMAD4 regulates macrophage polarization and downstream malignant phenotypes.

We established circSMAD4-silencing models in two complementary macrophage systems and mimicked tumor education (Fig. 2A). In the human system, peripheral-blood CD14-positive monocytes were differentiated into macrophages (hMDMs); in parallel, murine bone marrow-derived macrophages (BMDMs) were generated. In both systems, cells were transduced with shNC or two independent shRNAs targeting circSMAD4 (sh1/sh2 circSMAD4), followed by 48-h Transwell co-culture with tumor cells to obtain tumor-conditioned macrophages (TC-hMDMs and TC-BMDMs) (Fig. 2A).

Fig. 2.

circSMAD4 drives tumor-educated M2-like polarization of macrophages and promotes tumor-cell aggressiveness. (A) Workflow for generating TC-hMDMs and TC-BMDMs, circSMAD4 knockdown, and downstream functional assays. (B) RT–qPCR analysis of M1-associated markers (MHC-II [HLA-DRA in TC-hMDMs; H2-Ab1 in TC-BMDMs], NOS2, and CD86) and M2-associated markers (CD163, CD206, and ARG1) in TC-hMDMs and TC-BMDMs. (C) Representative flow-cytometry histograms for HLA-DR, iNOS, CD86, CD163, CD206, and ARG1 in TC-hMDMs. Gating strategy and marker thresholds were defined based on FMO controls (see Fig. S2). (D) Flow-cytometry quantification of marker-positive cells in TC-hMDMs and TC-BMDMs. (E) ELISA of IL-10, TGF-β, and iNOS in culture supernatants. (F) CCK-8 assays of A549 and LLC cells. (G) Colony-formation assays of A549 and LLC cells with quantification. (H) Bioluminescence-based growth readouts of patient-derived LUAD organoids (PDO #1 and PDO #2) after co-culture with TC-hMDMs. (I) Immunoblot analysis of EMT-related proteins (E-cadherin, N-cadherin, Vimentin) in A549 and LLC cells. (J) Transwell migration and invasion assays of A549 and LLC cells with quantification. Scale bar, 50 μm.

∗P < 0.05; ∗∗P < 0.01; ∗∗∗P < 0.001; ∗∗∗∗P < 0.0001; ns, not significant.

At the macrophage level, RT–qPCR consistently showed that circSMAD4 depletion increased M1-associated markers—MHC-II (HLA-DRA in TC-hMDMs and H2-Ab1 in TC-BMDMs), NOS2/iNOS, and CD86—and decreased M2-associated markers, including CD163, CD206, and ARG1, in both TC-hMDMs and TC-BMDMs (Fig. 2B). Flow cytometry confirmed these changes at the population level, with increased fractions of MHC-II- (HLA-DR in TC-hMDMs and I-A/I-E in TC-BMDMs), iNOS-, and CD86-positive cells and decreased CD163-, CD206-, and ARG1-positive cells upon circSMAD4 knockdown (Fig. 2C and D). Interpreted within the M1/M2 operational framework for a continuous activation spectrum [11], these results support a role for circSMAD4 in promoting an immunosuppressive, repair-like TAM polarization bias. Consistently, measurements in supernatants indicated reduced IL-10 levels and increased TGF-β and iNOS levels upon circSMAD4 depletion (Fig. 2E), suggesting that circSMAD4 amplifies TAM-like functional programs beyond marker expression.

To assess tumor cell phenotypes, we found that tumor cells exposed to circSMAD4-silenced macrophage conditions exhibited reduced proliferation in A549 and LLC cells (CCK8) (Fig. 2F) and impaired clonogenicity (Fig. 2G). Patient-derived organoids (PDOs) from two independent cases also showed decreased growth/viability (Fig. 2H), indicating robustness across clinically relevant models. EMT-related patterns were reversed, with increased E-cadherin and decreased N-cadherin and Vimentin (Fig. 2I), and Transwell assays demonstrated reduced migration and invasion (Fig. 2J). These findings align with established roles of TAMs in promoting invasion and metastatic dissemination [[48], [49], [50]], and indicate that circSMAD4 enhances malignant tumor phenotypes by driving protumor macrophage polarization.

Physiological relevance was further evaluated in vivo (Fig. 3A). circSMAD4-silenced BMDMs were mixed with LLC cells to establish an orthotopic lung tumor model, and a separate tail-vein model was performed to assess metastatic dissemination. Compared with controls, circSMAD4 knockdown reduced tumor burden, reflected by smaller lung nodules (Fig. 3B) and decreased tumor weight (Fig. 3C), and conferred a survival benefit (Fig. 3D). Immunofluorescence revealed reduced circSMAD4 signals within F4/80-positive macrophage regions in tumors (Fig. 3E), confirming effective macrophage-associated circSMAD4 suppression in vivo. Ki67 staining and IHC scores were decreased (Fig. 3F and G), consistent with reduced proliferation. In the metastasis setting, in vivo imaging demonstrated diminished metastatic signals upon circSMAD4 depletion (Fig. 3H), accompanied by improved survival (Fig. 3J). IHC analyses further supported EMT reversal, with enhanced E-cadherin and reduced Vimentin (Fig. 3K–M), consistent with the in vitro observations.

Fig. 3.

circSMAD4 depletion in macrophages restrains LUAD growth and metastasis in vivo. (A) Schematic of orthotopic lung implantation and experimental metastasis models using LLC cells mixed with BMDMs expressing shNC or sh-circSMAD4. (B) Representative images of orthotopic lung tumors. (C) Tumor weight of orthotopic implants. (D) Overall survival of mice bearing orthotopic tumors. (E) Immunofluorescence showing F4/80 and circSMAD4 signals in tumor tissues. Scale bar, 50 μm. (F, G) Representative Ki-67 IHC staining and quantification in orthotopic tumors. Scale bar, 50 μm. (H) Representative bioluminescence images of lung tumor burden in the metastasis model. (I) Tumor weight in the metastasis model. (J) Overall survival of mice in the metastasis model. (K–M) Representative IHC staining and quantification of E-cadherin and vimentin in tumors. Scale bar, 50 μm.

∗P < 0.05; ∗∗P < 0.01; ∗∗∗P < 0.001; ∗∗∗∗P < 0.0001; ns, not significant.

These results indicate that circSMAD4 promotes tumor-educated M2-like macrophage polarization and amplifies protumor outputs, thereby enhancing tumor cell proliferation and invasive behaviors. Targeting circSMAD4 attenuates both macrophage protumor programs and tumor malignant phenotypes, resulting in suppressed tumor growth and reduced metastasis in vivo.

3.3. Cytoplasmic circSMAD4 acts as a ceRNA to sponge miR-562 and de-repress COL4A1, forming a functionally rescuable regulatory axis

circSMAD4 exerts clear protumor outputs in macrophages, yet the underlying mechanism required clarification. Because many circRNAs are enriched in the cytoplasm and can function as competitive endogenous RNAs (ceRNAs) by sequestering miRNAs to de-repress target mRNAs [24,27,28], we tested whether circSMAD4 mediates its effects through a miRNA-centered regulatory axis.

We first examined the subcellular distribution of circSMAD4 using nuclear/cytoplasmic fractionation. circSMAD4 predominantly localized to the cytoplasmic compartment in tumor-conditioned human macrophages and patient-derived TAMs (Fig. 4A). RNA-FISH further revealed prominent cytoplasmic circSMAD4 signals in macrophages (Fig. 4B), providing a spatial basis for circSMAD4 to engage cytoplasmic miRNA-mediated regulation.

Fig. 4.

circSMAD4 functions as a cytoplasmic ceRNA to sequester miR-562 and de-repress COL4A1.

(A) Subcellular distribution of circSMAD4 in TC-hMDMs and patient-derived TAMs assessed by nuclear/cytoplasmic fractionation.

(B) Representative immunofluorescence/ISH images showing circSMAD4 signals in macrophages (CD163) with nuclear counterstaining (DAPI). Scale bar, 50 μm.

(C) Venn diagram of predicted circSMAD4-interacting miRNAs from circInteractome and circBank, yielding a shortlist including miR-562.

(D) miR-562 levels following circSMAD4 knockdown in TC-hMDMs.

(E–G) pri-miR-562, pre-miR-562, and miR-562 promoter reporter activity after circSMAD4 overexpression.

(H) AGO2-RIP enrichment of circSMAD4 and miR-562 relative to IgG in TC-hMDMs.

(I) AGO2 immunoblotting after circSMAD4 sense/antisense RNA pull-down in TC-hMDMs.

(J) Predicted pairing between miR-562 and circSMAD4 (WT) and the corresponding mutant design. Mutations were introduced within the predicted miR-562 seed-matching region using transition substitutions (A↔G, C↔U) to disrupt miRNA–target pairing while minimizing changes in sequence composition and local RNA structure.

(K) Dual-luciferase assays for circSMAD4-WT/MUT reporters in the presence of miR-562 mimics or inhibitor.

(L) Intersection of miRNA target predictions (miRTarBase, miRmap, TargetScan, and miRDB) identifying candidate miR-562 targets.

(M) COL4A1 mRNA levels after miR-562 mimics or inhibitor in TC-hMDMs.

(N) Predicted miR-562 binding site within the COL4A1 3′UTR (WT) and mutant design. Mutations were introduced within the predicted miR-562 seed-matching region using transition substitutions (A↔G, C↔U) to disrupt miRNA–target pairing while minimizing changes in sequence composition and local RNA structure.

(O) Dual-luciferase assays for COL4A1 3′UTR WT/MUT reporters with miR-562 mimics or inhibitor.

∗P < 0.05; ∗∗P < 0.01; ∗∗∗P < 0.001; ∗∗∗∗P < 0.0001; ns, not significant.

We then predicted putative miRNA interactors of circSMAD4 using circInteractome and circBank and prioritized overlapping candidates [51,52], yielding a small set including miR-361-3p, miR-562, miR-626, and miR-647 (Fig. 4C). To experimentally prioritize functionally relevant candidates, we measured their responses to circSMAD4 knockdown. In patient-derived TAMs, circSMAD4 knockdown selectively increased miR-562, whereas other candidates were largely unchanged; conversely, circSMAD4 overexpression reduced miR-562 (Fig. S3A). Consistently, circSMAD4 knockdown increased miR-562 in TC-hMDMs (Fig. 4D). In primary cells, miR-562 levels were lower in TAMs than in NTRMs (Fig. S3B) and inversely correlated with circSMAD4 expression (Fig. S3C). Together, these data prioritized miR-562 as the most relevant candidate and supported a sequestration-consistent regulatory direction.

To distinguish miRNA sequestration from altered miRNA biogenesis, we assessed miR-562 precursors and promoter activity. circSMAD4 overexpression did not change pri-miR-562 or pre-miR-562 levels (Fig. 4E and F), nor did it alter miR-562 promoter-driven luciferase activity (Fig. 4G), favoring a post-transcriptional binding/decoy mechanism.

We next validated direct circSMAD4–miR-562 interaction at the complex and sequence levels. AGO2-RIP showed significant enrichment of circSMAD4 and miR-562 in AGO2 immunoprecipitates compared with IgG controls (Fig. 4H), implicating an AGO2-dependent miRNA effector complex. A sense/antisense probe pull-down further enriched AGO2 preferentially in the sense condition (Fig. 4I), supporting circSMAD4 engagement with RISC-related complexes. Sequence specificity was confirmed using luciferase reporters containing the predicted binding site: miR-562 mimics suppressed the circSMAD4-WT reporter, whereas the binding-site mutant reporter was largely resistant (Fig. 4J and K).

Having established circSMAD4–miR-562 binding, we identified key downstream mRNA targets of miR-562. By intersecting predictions from miRTarBase, miRmap, TargetScan, and miRDB [[53], [54], [55], [56]], a small set of candidates was obtained (NRIP1, CBY1, COL4A1, ZNF449) (Fig. 4L). Screening in patient-derived TAMs revealed that COL4A1 was significantly decreased by miR-562 mimics and increased by a miR-562 inhibitor, whereas the other candidates showed no significant changes (Fig. S4A), nominating COL4A1 as the priority functional target. This regulation was further confirmed in TC-hMDMs (Fig. 4M). A COL4A1 3′UTR luciferase assay demonstrated direct targeting: miR-562 mimics suppressed the COL4A1-WT reporter but had a markedly attenuated effect on the mutant reporter (Fig. 4N and O).

Rescue experiments further supported causality: circSMAD4 knockdown reduced COL4A1 expression, which was restored by miR-562 inhibition at both the mRNA and protein levels (Fig. S4B–C). Clinically, COL4A1 was elevated in TAMs relative to NTRMs (Fig. S4D), higher COL4A1 in TAMs was associated with poorer overall survival (Fig. S4E), and COL4A1 expression correlated negatively with miR-562 (Fig. S4F) but positively with circSMAD4 (Fig. S4G), supporting a coherent circSMAD4–miR-562–COL4A1 cascade in patient macrophages.

Finally, functional rescue experiments established causality. In circSMAD4-silenced TC-hMDMs, addition of a miR-562 inhibitor partially restored polarization-associated outputs, reversing marker changes by qPCR and flow cytometry (Fig. S3E–F) and restoring IL-10, while reducing the increased TGF-β and iNOS levels (Fig. S3G). In tumor cells, miR-562 inhibition rescued the circSMAD4 knockdown–induced reductions in proliferation and clonogenicity (Fig. S3H–I), reversed EMT-associated protein changes (Fig. S3K), and restored migration/invasion capacities (Fig. S3L). Collectively, these results support a model in which cytoplasmic circSMAD4 functions as a ceRNA to sequester miR-562, thereby de-repressing COL4A1 and promoting protumor macrophage programs and downstream malignant phenotypes.

3.4. circSMAD4 forms a specific RNP complex with IGF2BP2 to promote TAM protumor programs

circSMAD4 is predominantly cytoplasmic in macrophages and can regulate downstream genes via a miRNA axis. Because circRNAs also frequently exert functions through interactions with RNA-binding proteins (RBPs) [57,58], we next profiled circSMAD4-associated proteins in tumor-conditioned macrophages to define additional mechanistic layers.

Biotinylated circSMAD4 RNA pull-down followed by mass spectrometry identified multiple candidate binders, with IGF2BP2 among the most significantly enriched proteins (Fig. 5A). This interaction was validated in TC-hMDMs, where the circSMAD4 sense probe selectively enriched IGF2BP2, whereas the antisense control showed minimal binding (Fig. 5B). Reciprocally, IGF2BP2-RIP robustly enriched circSMAD4 (Fig. 5C), supporting a stable circSMAD4–IGF2BP2 ribonucleoprotein (RNP) complex in tumor-educated macrophages.

Fig. 5.

circSMAD4 physically associates with IGF2BP2 in macrophages.

(A) LC–MS/MS summary of proteins enriched by circSMAD4 RNA pull-down.

(B) Western blot validation of IGF2BP2 in circSMAD4 sense (vs antisense) pull-down from TC-hMDMs.

(C) IGF2BP2 RIP–qPCR showing circSMAD4 enrichment over IgG in TC-hMDMs.

(D–E) catRAPID prediction and ViennaRNA RNAfold secondary-structure modeling indicating multiple candidate IGF2BP2-binding regions on circSMAD4.

(F) Western blot of IGF2BP2 after pull-down with circSMAD4 fragments (1#–3#).

(G) Schematic of IGF2BP2 domain architecture and the Flag-tagged truncation/deletion constructs used for mapping circSMAD4 interaction (designed based on catRAPID prediction and annotated RRM/KH domain boundaries).

(H) Anti-Flag RIP–qPCR showing circSMAD4 enrichment precipitated by the indicated Flag-tagged IGF2BP2 truncation/deletion constructs (presented as % input and normalized to IgG).

(I) Nuclear–cytoplasmic fractionation followed by RT–qPCR showing circSMAD4 distribution and the effect of IGF2BP2 knockdown on the nuclear-to-cytoplasmic ratio of circSMAD4 in TC-hMDMs. Fractionation quality was validated using nuclear/cytoplasmic marker transcripts/proteins.

(J) Representative immunofluorescence/ISH images showing circSMAD4 signals and IGF2BP2 staining in macrophages (CD163) with nuclear counterstaining (DAPI). Scale bar, 50 μm.

(K–N) qPCR and Western blot showing no reciprocal change in expression between circSMAD4 and IGF2BP2 upon knockdown/overexpression.

∗P < 0.05; ∗∗P < 0.01; ∗∗∗P < 0.001; ∗∗∗∗P < 0.0001; ns, not significant.

We then mapped the interaction interfaces. Using catRAPID to predict protein–RNA interaction propensity and RNAfold Web Server to model circSMAD4 secondary structure [59,60], we observed multiple putative IGF2BP2-binding regions distributed along circSMAD4 (Fig. 5D and E). Pull-down assays using circSMAD4 fragments (1#–3#) indicated preferential enrichment of IGF2BP2 by fragment 2# (Fig. 5F), nominating this region as a major interaction hotspot. For IGF2BP2 domain mapping, we adopted a prediction-guided and domain-constrained truncation strategy. We first used catRAPID prediction to nominate a focused interaction hotspot and then designed truncation constructs according to the annotated RRM/KH domain architecture of IGF2BP2, placing breakpoints at domain boundaries to preserve domain integrity and reduce misfolding artifacts. A panel of Flag-tagged truncation mutants spanning the RRM- and KH-containing regions (e.g., 1–251, 1–302, 303–599, and 251–599), together with a focused internal deletion (Δ251–302) and a minimal-fragment construct (251–302), was generated to map the circSMAD4-binding determinant (Fig. 5G). Using anti-Flag RIP–qPCR to quantify circSMAD4 enrichment precipitated by each construct, we found that full-length IGF2BP2 (1–599 aa) robustly enriched circSMAD4, whereas deletion of aa 251–302 markedly reduced this enrichment; notably, the 251–302 aa segment alone retained substantial binding capacity (Fig. 5H), indicating that this region is necessary and sufficient for efficient circSMAD4–IGF2BP2 interaction.

Given a recent report that IGF2BP2 can regulate nuclear export of circRNAs [61], we next examined whether IGF2BP2 affects the subcellular distribution of circSMAD4. Nuclear–cytoplasmic fractionation followed by RT–qPCR revealed that circSMAD4 remained predominantly cytoplasmic, and IGF2BP2 knockdown did not appreciably alter the nuclear-to-cytoplasmic ratio of circSMAD4 (Fig. 5I), indicating that IGF2BP2 is dispensable for circSMAD4 nuclear export in TC-hMDMs.

Immunofluorescence also revealed prominent spatial proximity and signal overlap between circSMAD4 and IGF2BP2 within cells (Fig. 5J), consistent with a predominantly cytoplasmic interaction mode.

Importantly, the circSMAD4–IGF2BP2 association was not attributable to altered expression of either partner. circSMAD4 knockdown or overexpression did not change IGF2BP2 mRNA or protein levels (Fig. 5K–L), and IGF2BP2 perturbation did not affect circSMAD4 abundance (Fig. 5M−N), supporting a physical interaction rather than transcriptional regulation.

Given that IGF2BP2 is a well-established RBP and has been reported as an m6A reader that enhances target mRNA stability and translation [33], we assessed whether IGF2BP2 contributes to macrophage polarization and protumor outputs. Similar to circSMAD4 silencing, IGF2BP2 knockdown increased M1-associated markers and reduced M2-associated markers in both TC-hMDMs and TC-BMDMs (Fig. S5B), with concordant shifts by flow cytometry (Fig. S5C). ELISA measurements showed reduced IL-10 and increased TGF-β and iNOS levels (Fig. S5D). In tumor cells, IGF2BP2 knockdown conditions suppressed proliferation (CCK8), reduced clonogenicity (Fig. S5E–F), decreased PDO viability (Fig. S5G), reversed EMT-associated patterns (Fig. S5H), and impaired migration/invasion (Fig. S5I–J). Together, these data indicate that IGF2BP2 is required for tumor-educated M2-like macrophage outputs and protumor phenotypes, paralleling the functional directionality of circSMAD4.

Collectively, circSMAD4 forms a specific RNP complex with IGF2BP2 in tumor-conditioned macrophages, dependent on the IGF2BP2 aa 251–302 region and primarily mediated by a defined circSMAD4 segment. The phenotypic convergence between IGF2BP2 and circSMAD4 perturbations suggests that circSMAD4 may amplify protumor macrophage programs through IGF2BP2-linked post-transcriptional regulation, potentially involving mRNA stability control.

3.5. circSMAD4 recruits the m6A reader IGF2BP2 to stabilize COL4A1/ACTA2/SPI1 transcripts and shape an M2-like, matrix-remodeling TAM program

The stable interaction between circSMAD4 and IGF2BP2, together with their convergent protumor phenotypes, suggested that circSMAD4 may amplify TAM programs via IGF2BP2-mediated post-transcriptional regulation. Because IGF2BP2 can recognize m6A-modified transcripts and enhance their stability and translation [33], we next defined the key downstream transcripts governed by the circSMAD4–IGF2BP2 axis and evaluated whether they are linked to a matrix-remodeling TAM program.

To prioritize candidate targets at a transcriptome-wide level, we performed mRNA-seq in shNC versus shIGF2BP2 macrophages and in shNC versus shcircSMAD4 macrophages, and integrated these differentially expressed genes with ENCORI (starBase) predictions of IGF2BP2-bound targets [62]. This integration was intended to nominate circSMAD4-responsive and IGF2BP2-dependent transcripts, rather than to infer m6A modification directly. Intersection analysis yielded a small set of shared candidates, including COL4A1, SPI1, ACTA2, and SALL1 (Fig. 6A). We next validated these candidates by RT–qPCR in TC-hMDMs. circSMAD4 knockdown consistently reduced COL4A1, SPI1, and ACTA2, whereas SALL1 showed no reproducible change (Fig. S6A). Similarly, IGF2BP2 knockdown decreased COL4A1/SPI1/ACTA2 with minimal impact on SALL1 (Fig. S6B). Therefore, SALL1 was deprioritized and excluded from subsequent mechanistic analyses. Among the validated targets, COL4A1 encodes a type IV collagen component of basement-membrane/ECM modules, and prior work has suggested that TAMs can actively contribute to tumor ECM remodeling rather than being passive bystanders [63]. SPI1 encodes PU.1, a core myeloid transcription factor, and PU.1/SPI1 activity has been implicated in shaping protumor macrophage programs in tumor contexts [64]. ACTA2 encodes α-SMA, a cytoskeletal remodeling marker that is closely associated with tissue remodeling and profibrotic macrophage states. Notably, α-SMA–positive macrophage populations have been reported during macrophage-to-myofibroblast transition after injury and are linked to fibrosis-related programs [65].

Fig. 6.

circSMAD4 facilitates IGF2BP2-dependent stabilization of m6A-marked transcripts.

(A) Venn diagram intersecting ENCORI-predicted IGF2BP2 targets with DEGs from shIGF2BP2 versus shNC and shcircSMAD4 versus shNC mRNA-seq, identifying shared candidates.

(B) MeRIP–qPCR showing m6A enrichment on COL4A1, SPI1, and ACTA2 candidate regions (CRDs) in shNC and shIGF2BP2 cells.

(C) IGF2BP2-RIP–qPCR showing IGF2BP2 binding to COL4A1, SPI1, and ACTA2 CRDs in shNC + Vector, shcircSMAD4 + Vector, shNC + IGF2BP2, and shcircSMAD4 + IGF2BP2 groups.

(D) Biotin-circSMAD4 pull-down followed by qPCR showing enrichment of COL4A1, SPI1, and ACTA2 CRDs in Vector + shNC, circSMAD4 + shNC, Vector + shIGF2BP2, and circSMAD4 + shIGF2BP2 groups.

(E–G) Schematics of m6A-site mutations introduced into COL4A1, SPI1, and ACTA2 reporters.

(H–J) Dual-luciferase assays for CRD reporters (WT and m6A-mutant) in Vector, circSMAD4, and IGF2BP2 groups.

(K–M) MeRIP–qPCR for WT and m6A-mutant CRD reporters in Vector, circSMAD4, and IGF2BP2 groups.

(N–P) mRNA decay assays of endogenous COL4A1, SPI1, and ACTA2 following circSMAD4 knockdown with Vector or IGF2BP2 overexpression. Half-life estimated by one-phase decay (Y0 = 1, Plateau = 0).

(Q–S) mRNA decay assays of endogenous COL4A1, SPI1, and ACTA2 following circSMAD4 overexpression with shNC or shIGF2BP2. Half-life estimated by one-phase decay (Y0 = 1, Plateau = 0).

∗P < 0.05; ∗∗P < 0.01; ∗∗∗P < 0.001; ∗∗∗∗P < 0.0001; ns, not significant.

We then tested whether these transcripts are controlled through an m6A-linked IGF2BP2 pathway. m6A-RIP (anti-m6A) showed robust enrichment of key regions from COL4A1, SPI1, and ACTA2 in controls, whereas IGF2BP2 knockdown markedly reduced these signals (Fig. 6B), consistent with an essential role of IGF2BP2 in maintaining these m6A-associated transcripts. IGF2BP2-RIP (anti-IGF2BP2) further showed that circSMAD4 knockdown markedly reduced IGF2BP2 association with the COL4A1-, SPI1-, and ACTA2-CRD regions, and this reduction persisted even upon IGF2BP2 overexpression (Fig. 6C). These anti-IGF2BP2 RIP–qPCR results indicate that circSMAD4 is required for efficient IGF2BP2 binding to the COL4A1/SPI1/ACTA2 regions, supporting a cooperative mode of action between circSMAD4 and IGF2BP2. Concordantly, biotinylated circSMAD4 pull-down enriched these mRNAs, and this enrichment shifted upon IGF2BP2 perturbation (Fig. 6D), supporting a model in which circSMAD4 acts as a scaffold to facilitate an efficient IGF2BP2–mRNA regulatory RNP unit.

To address how the m6A signal on these transcripts is supported, we examined the requirement of the canonical m6A writer complex. Depletion of METTL3 or METTL14 markedly reduced m6A enrichment on the COL4A1-, SPI1-, and ACTA2-CRD regions (Fig. S6C), consistent with METTL3/METTL14 as the core writer complex [66]. Moreover, METTL3/14 depletion also reduced IGF2BP2 association with these CRD regions (Fig. S6D), consistent with IGF2BP proteins functioning as m6A readers whose binding depends on m6A marks [67]. Collectively, these data support that circSMAD4–IGF2BP2 acts primarily at the “reader/scaffold” level to utilize writer-installed m6A marks, rather than directly depositing m6A.

To establish m6A-site dependence, we generated wild-type and m6A-site mutant reporters for predicted m6A motifs in COL4A1, SPI1, and ACTA2 (Fig. 6E–G). circSMAD4 and IGF2BP2 increased reporter activity for the wild-type constructs, whereas the enhancement was markedly attenuated or lost in the m6A-mutant reporters (Fig. 6H–J). Consistently, m6A-RIP readouts indicated that the promoting effect of circSMAD4/IGF2BP2 depended on intact m6A sites (Fig. 6K–M), supporting an m6A-dependent post-transcriptional mechanism.

Next, mRNA stability assays under transcriptional blockade revealed that circSMAD4 prolonged the decay kinetics of COL4A1, SPI1, and ACTA2, whereas IGF2BP2 knockdown accelerated their degradation; importantly, the stabilizing effect of circSMAD4 was largely blunted in the absence of IGF2BP2 (Fig. 6N–S). In line with a writer–reader dependency, METTL3 or METTL14 knockdown similarly shortened the half-lives of COL4A1, SPI1, and ACTA2 transcripts (Fig. S6E–G), supporting that their stabilization requires the canonical m6A-writing machinery upstream of IGF2BP2.

Because COL4A1 is a convergence point of the miR-562 arm and the IGF2BP2 arm, we further tested combined perturbation. Simultaneous IGF2BP2 knockdown and miR-562 mimics suppressed COL4A1 more strongly than either manipulation alone (Fig. S6H), supporting additive contributions of the two branches to this shared output.

Functionally, silencing COL4A1, SPI1, or ACTA2 in TC-hMDMs shifted macrophage marker profiles away from an M2-like remodeling program (Fig. S6I–K), consistent with their contribution to the circSMAD4-driven, matrix-remodeling TAM phenotype.

Together, these results indicate that circSMAD4 recruits IGF2BP2 to selectively stabilize COL4A1/ACTA2/SPI1 transcripts in an m6A-dependent manner, thereby providing a post-transcriptional basis for an M2-like, matrix-remodeling TAM signature. Given that tumor-associated ECM remodeling supports invasion, immune evasion, and metastatic dissemination, this mechanism offers a direct molecular explanation for how circSMAD4 reinforces protumor TAM programs.

4. Discussion

Our study defines a post-transcriptional framework by which a TAM-enriched circRNA reinforces protumor macrophage programs in LUAD. circSMAD4 is enriched in TAMs and associated with adverse clinical outcome; its silencing reprograms tumor-conditioned macrophages away from an M2-like bias, attenuates tumor-cell malignant phenotypes, and suppresses tumor growth and metastasis in vivo. Mechanistically, circSMAD4 engages two complementary branches: a ceRNA arm in which cytoplasmic circSMAD4 sequesters miR-562 to de-repress COL4A1, and an RBP arm in which circSMAD4 forms a specific RNP complex with the m6A reader IGF2BP2 to stabilize COL4A1, ACTA2, and SPI1 transcripts, thereby promoting an M2-like, matrix-remodeling TAM program. Together, these findings connect circRNA biology, miRNA-mediated de-repression, and m6A-RBP–dependent RNA stabilization into a coherent mechanism for sustaining protumor TAM outputs [[24], [25], [26], [27], [28], [29], [30], [31], [32], [33]].

Biologically, TAM plasticity and heterogeneity are central to immunosuppression and metastatic progression in solid tumors [6,[8], [9], [10], [11], [12], [13],45,46,50]. While the M1/M2 framework is operationally useful, TAM states exist along a continuum and often integrate immunosuppressive, tissue-repair, and matrix-remodeling features [[11], [12], [13],47]. The circSMAD4-driven program identified here aligns with ECM remodeling as a functional axis that facilitates invasion, supports suppressive niches, and promotes dissemination [13,16,17]. The convergence on COL4A1 suggests that basement-membrane/ECM modules can be actively reinforced within tumor-educated myeloid cells, not solely by tumor cells or fibroblasts [16,17]. Stabilization of SPI1 (PU.1), a core myeloid regulator, may further consolidate a durable transcriptional identity that sustains these outputs, whereas ACTA2-associated stabilization may reflect enhanced cytoskeletal and remodeling capacity consistent with a matrix-remodeling phenotype.

Mechanistically, the dual-branch architecture is notable. The ceRNA branch provides a de-repression route via miR-562, whereas the IGF2BP2 branch offers an m6A-dependent stabilization mechanism that can sustain and amplify selected transcript outputs over time [[30], [31], [32], [33], [34], [35], [36],41,42]. Their convergence on COL4A1 provides a plausible explanation for the persistence of remodeling programs under complex tumor-conditioning cues. More broadly, the data support a view of circRNAs as post-transcriptional hubs that integrate RISC-associated miRNA availability with RBP-centered RNA fate control, converting transient signals into stable cell-state programs. Beyond miRNA sequestration and RBP scaffolding, circRNAs have also been reported to regulate gene expression through additional modes. For example, certain nuclear circRNAs can modulate transcription in cis/trans by engaging transcriptional machinery or chromatin-associated regulators [68,69]. CircRNAs have also been implicated in genome stability control, including potential roles in DNA damage/repair-associated processes or RNA–DNA hybrid (circR-loop/ciR-loop)-linked genome instability [70,71]. In addition, an increasing number of circRNAs have been shown to harbor coding potential and can be translated into functional peptides/proteins in specific contexts [72,73]. In the present study, our data primarily support a cytoplasmic, post-transcriptional mechanism for circSMAD4 in TAM-like macrophages; however, whether circSMAD4 may additionally participate in these alternative regulatory modes under distinct microenvironmental conditions remains an interesting question for future work.

Translationally, circSMAD4 enrichment in TAMs and its prognostic association motivate evaluation as a microenvironment-linked biomarker in larger cohorts. Therapeutically, targeting circSMAD4 itself or key nodes in its network (miR-562, IGF2BP2) may provide an avenue to reprogram TAM function [6,45], potentially complementing existing immunotherapies that are limited by suppressive microenvironments [6,7]. ECM-related outputs (e.g., COL4A1-linked remodeling) may serve as pharmacodynamic readouts of effective TAM reprogramming in metastasis-relevant niches [16,17].

Several limitations merit consideration. First, although circSMAD4 is consistently enriched in LUAD TAMs, we did not define the upstream mechanisms responsible for this upregulation. Tumor-derived soluble factors, microenvironmental stresses (e.g., hypoxia), epigenetic remodeling, or altered activity of circRNA biogenesis regulators that control back-splicing may all contribute to circSMAD4 accumulation in TAMs. Dissecting these upstream cues and identifying the key regulators will be an important direction for future studies, and may further enable therapeutic strategies that prevent the establishment of the circSMAD4-driven, matrix-remodeling TAM program. TAM state diversity extends beyond M1/M2 markers, and single-cell and spatial profiling will be important to define circSMAD4-high TAM subsets and niches [11,12,47]. ceRNA effects can be sensitive to stoichiometry and competing endogenous targets; although binding, reporter, and rescue experiments support causality, quantitative in vivo dynamics of miR-562 availability and competition remain to be elucidated [25,26,37]. The IGF2BP2 arm implicates m6A-dependent stabilization, but upstream changes in writer/eraser systems and genome-wide target selection would benefit from CLIP-seq/Ribo-seq-level resolution [[30], [31], [32],62]. The biological interpretation of ACTA2 in macrophages requires careful validation, ideally with lineage controls and in vivo tracing. Finally, while BMDM-based in vivo models establish functional directionality, macrophage-specific delivery or conditional perturbation will strengthen causal inference in more physiological contexts.

In summary, our work identifies a circSMAD4-centered post-transcriptional network in TAMs that integrates miR-562 sequestration with IGF2BP2-mediated, m6A-dependent mRNA stabilization to promote a matrix-remodeling, M2-like TAM state and drive LUAD progression and metastasis (Fig. 7).

Fig. 7.

Proposed model: circSMAD4 drives matrix-remodeling TAM programs in LUAD.

Schematic summary illustrating that circSMAD4 in tumor-associated macrophages promotes a matrix-remodeling, M2-like state through two post-transcriptional routes: (i) circSMAD4 sequesters miR-562 in an AGO2-dependent manner to relieve repression of COL4A1 mRNA; (ii) circSMAD4 associates with IGF2BP2 to enhance the stability of m6A-marked transcripts, including COL4A1, SPI1, and ACTA2 (α-SMA). These combined outputs reinforce extracellular matrix remodeling within the LUAD tumor microenvironment.

CRediT authorship contribution statement

Zhengwei Yu: Writing – original draft, Visualization, Methodology, Investigation. Xinyue Wang: Visualization, Investigation. Yiqian Zheng: Investigation. Yifan He: Resources. Jiayu Lin: Formal analysis, Data curation. Yue Xiao: Formal analysis, Data curation. Bin Mo: Resources. Haoyu Xie: Resources. Sitong Hang: Resources. Xia Gao: Resources. Pei Xu: Resources. Yihao Liu: Writing – review & editing, Funding acquisition, Conceptualization. Haibo Xiao: Supervision, Resources, Project administration, Funding acquisition, Conceptualization.

Ethics approval and consent to participate

This study was approved by the Ethics Committee of Xinhua Hospital, Affiliated to Shanghai Jiao Tong University School of Medicine. All procedures involving human participants were conducted in accordance with the ethical standards of the institutional and/or national research committee and with the 1964 Declaration of Helsinki and its later amendments or comparable ethical standards. Written informed consent was obtained from all participants.

Consent for publication

The authors confirm that all participants provided written informed consent for publication.

Availability of data and material

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Declaration of generative AI and AI-assisted technologies in the manuscript preparation process

During the preparation of this work the authors used ChatGPT in order to refine and optimize the language of the manuscript. After using this tool/service, the authors reviewed and edited the content as needed and take full responsibility for the content of the published article.

Funding

This work was supported by the National Natural Science Foundation of China (Grant Nos. 82173382, 82473433, and 82503169), the China Postdoctoral Science Foundation (Certificate No. 2025M782141), and the Shanghai Sailing Program (Grant No. 24YF2726700).

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

The following is the supplementary data to this article: