Programmable iPSC-derived CAR-NK vesicles remodel the immune microenvironment and eradicate tumors

Summary

Chimeric antigen receptor (CAR) cell therapy transforms hematologic cancer treatment but remains limited in solid tumors due to stromal barriers and an immunosuppressive tumor microenvironment that restricts immune cell infiltration. To address these barriers, we develop a cell-free therapeutic platform based on CAR-engineered induced pluripotent stem cell (iPSC)-derived natural killer (NK) extracellular vesicles (CAR-iNEVs), which retain tumor-targeting capability without reliance on live-cell delivery. CAR-iNEV demonstrates potent antitumor activity and excellent tolerability across multiple xenograft and patient-derived models. Mechanistically, CAR-iNEV directly eliminates tumor cells and remodels the tumor microenvironment by promoting pro-inflammatory macrophage polarization, thereby enhancing host innate antitumor immunity. CAR-iNEV also functions cooperatively with immune checkpoint blockade, and combined treatment with CAR-iNEV and CD47 inhibition increases tumor clearance and induces long-term immunological memory in surviving mice. These findings support the therapeutic potential of CAR-iNEV for solid tumors through coordinated tumor targeting and immune microenvironment modulation.

Graphical abstract

Highlights

• •CAR-iNEV therapy targets solid tumors and modulates the immune microenvironment
• •iPSC-derived CAR-NK vesicles function as a flexible and scalable immunotherapy tool
• •CAR-iNEV activates macrophage NOS2 signaling to reprogram the suppressive TME
• •CAR-iNEV and CD47 blockade work together to increase tumor clearance and immunity

Hao Zhang et al. develop CAR-iNEV, a cell-free therapy from engineered iPSC-NK vesicles, which targets solid tumors, modulates the microenvironment, and activates macrophages, acting alone or with CD47 blockade to strengthen antitumor responses and showing efficacy in diverse tumor models.

Introduction

Chimeric antigen receptor (CAR) therapies have demonstrated remarkable success in hematologic malignancies, but their application in solid tumors remains limited by cytokine release syndrome, neurotoxicity, and the immunosuppressive tumor microenvironment (TME).^1,2,3 Natural killer (NK) cells offer a promising alternative due to their major histocompatibility complex (MHC)-independent cytotoxicity and low risk of graft-versus-host disease.^4,5,6,7 They also secrete chemokines such as XCL1 and CCL5 to recruit cDC1s and support a proinflammatory TME.^8,9,10 However, conventional CAR-NK approaches using NK92 or peripheral blood NK cells suffer from poor tumor persistence, limited penetration, and expansion inefficiency.^11,12 Induced pluripotent stem cell (iPSC)-derived NK cells (iNKs) provide a renewable, standardized cytotoxic source suitable for clinical-grade production^11,13,14 and thus were selected as the platform for further development.

Despite these advantages, key challenges such as immune rejection, limited stromal penetration, and inefficient in vitro expansion hinder the translation of CAR-NK therapies.^15,16,17 To address these challenges, we propose an innovative therapeutic strategy based on CAR-iNK extracellular vesicles (EVs) (CAR-iNEVs). In this study, we selected iNKs as the parental source of CAR-iNEVs, due to their unlimited expansion potential, batch-to-batch consistency, and genetic engineering compatibility. Their nanoscale size enables CAR-iNEVs to efficiently penetrate the dense stromal barrier of solid tumors and exert antitumor effects deep within the tumor core.^18,19 In addition to favorable dimensions for passive diffusion, EVs can actively remodel the TME. CAR-iNEVs express ECM-modulating proteins such as MMP1, MMP14, and CD147, which promote collagen degradation and stromal loosening, and can modulate CAF activity to reduce ECM deposition and matrix stiffness.²⁰ These properties facilitate deeper infiltration into tumor tissue.^21,22 Collectively, these mechanisms support the selection of EVs over cell-based platforms for overcoming stromal barriers in solid tumors.

CD133 is a well-established CSC marker that is highly expressed in multiple malignancies, including glioblastoma,²³ colorectal cancer,²⁴ prostate cancer,²⁵ hepatocellular carcinoma (HCC),²⁶ and pancreatic cancer.^27,28,29 Upregulation of CD133 expression is closely associated with tumor aggressiveness, chemoresistance, and recurrence.³⁰ To improve tumor targeting, we designed a high-affinity CD133 nanobody (Nb) and used it to replace the conventional single-chain variable fragment (scFv) in CAR constructs. Compared with scFvs, Nbs exhibit superior structural stability and reduced immunogenicity.^18,31

In this study, we developed CD133-modified CAR-iNEVs (CAR^CD133-iNEVs) that selectively eliminate tumor cells and reprogram the immune microenvironment by promoting macrophage and neutrophil infiltration and activation. In humanized mice, CAR^CD133-iNEVs suppressed tumor progression and significantly prolonged survival. When combined with anti-CD47 therapy, they further enhanced tumor regression and induced immune memory, resulting in durable tumor control. These results establish CAR^CD133-iNEVs as a cell-free CAR immunotherapy capable of overcoming key limitations of conventional CAR-T and CAR-NK cell therapies, highlighting their translational potential for solid tumor treatment.

Results

Selection and functional validation of a high-affinity CD133 nanobody

To develop an optimized CD133-targeting Nb, we established a prokaryotic expression system for recombinant human and murine CD133 proteins and performed phage display screening combined with ELISA to identify high-affinity Nb candidates (Figures 1A and 1B). The top 10 candidate clones identified in the initial selection step were expressed and purified in a prokaryotic expression system, and their expression was validated using hemagglutinin (HA)-tag and VHH-specific antibodies (Figure 1C; Table S1). ELISA analysis confirmed that Nb63 exhibited the highest binding affinity toward both the human and murine CD133 proteins (Figure 1D). Surface plasmon resonance (SPR) analysis was conducted to determine the binding kinetics of Nb-CD133 (Nb63). The dissociation constants (Kds) for Nb63 binding to human and murine CD133 were 1.97 nM and 2.51 nM, respectively (Figure 1E).

Figure 1.

Screening of CD133 nanobodies and generation of engineered iPSC-derived CAR-NK cells

(A) Schematic representation of the CD133 nanobody screening process.

(B) Coomassie blue staining of purified human (H) and murine (M) CD133 proteins.

(C) Coomassie blue staining of purified nanobodies (15–20 kDa), and confirmation of their expression by western blotting with anti-VHH and anti-HA antibodies.

(D) ELISA analysis of the binding affinities of different nanobody variants (Nb13, Nb22, Nb26, Nb34, Nb63, Nb89, Nb132, Nb224, Nb381, and Nb592) to human and murine CD133 proteins (n = 3 biological replicates).

(E) Binding kinetics of nanobodies to human and murine CD133 proteins determined by SPR (n = 3 biological replicates).

(F) Schematic representation of CAR constructs containing different signaling domains.

(G) Flow cytometric analysis of CAR-iNK cells after cell sorting, using antibodies against the VHH domain to confirm post-sort CAR positivity (n = 3 biological replicates).

(H) Flow cytometric analysis of CAR-iNK cells using antibodies against CD16, NKP46, FasL, and TRAIL (n = 3 biological replicates). The data are presented as the means ± SDs.

Design and optimization of iPSC-derived CAR-NK cells

Optimizing CAR design is critical for enhancing the therapeutic function of CAR-iNK cells against solid tumors. In this study, we engineered an improved CAR construct by incorporating a high-affinity Nb, Nb-CD133 (Nb63), in place of the conventional scFv. Furthermore, to optimize CAR signaling, we designed seven distinct NK-specific CAR architectures, each of which contained a transmembrane domain, one or two NK-associated costimulatory domains, and a CD3ζ signaling chain (Figure 1F). The CAR constructs were subcloned and inserted into the pCD513B-CopGFP lentiviral vector, and Nb-CD133-expressing iNK cells were sorted by flow cytometry for subsequent functional assessment (Figure 1G). Notably, the third-generation CAR construct, which included CD3ζ and NKG2D-2B4-DAP10 costimulatory domains, exhibited the greatest cytotoxic potential while maintaining NK cell functional integrity, as reflected by sustained expression of effector molecules (FasL, TRAIL, perforin, granzyme B) and activating receptors (CD16, NKp46; Figures 1H and S1A). To further assess the influence of different costimulatory domains on CAR-iNK cell cytotoxicity, we established CD133-knockout (CD133-KO) Colon26 tumor cells. In vitro cytotoxicity assays revealed that, regardless of the CD133 expression status, CAR-iNK cells that expressed NKG2D-2B4-DAP10-CD3ζ CAR presented the highest tumor-killing activity (Figures S1B and S1C). Following the functional screening, we selected the NKG2D-2B4-DAP10-CD3ζ construct as the optimal CAR backbone due to its superior cytotoxicity and preserved NK cell phenotype. To further characterize its functional activity, we measured CD107a expression and intracellular interferon (IFN)-γ production in CAR-iNK cells by flow cytometry. In the absence of the CD133 target, CAR-iNK cells exhibited a baseline level of activation; however, upon engagement with CD133⁺ tumor cells, both CD107a and IFN-γ levels were markedly elevated. This antigen-dependent activation pattern was consistently observed across multiple tumor cell lines, including wild-type and CD133-KO variants of H22, Panc02, and Colon26 (Figure S1D). These results demonstrate that CAR-iNK cells expressing the NKG2D-2B4-DAP10-CD3ζ construct exhibit significantly increased antitumor activity and enhanced ability to selectively target CD133⁺ tumor cells.

Preparation and characterization of CAR-iNEVs

Uniform particle size is critical for achieving the enhanced permeability and retention effect. Therefore, we employed tangential flow filtration (TFF) to isolate CAR-iNEVs, which, compared with ultracentrifugation, offers superior control over size distribution, yield, and purity, thereby enhancing tumor accumulation and in vivo efficacy.^32,33 To enhance the targeting specificity of CAR-iNEVs, a CD133 Nb was incorporated into the CAR construct (CAR^CD133-iNEV), while CAR-iNEVs modified with a non-targeting control Nb sequence (CAR^Ctrl-iNEV) and unmodified iNEVs without CAR (Blank-iNEV) served as controls (Figure 2A). Nanoparticle tracking analysis (NTA) revealed that the mean particle diameters of CAR^CD133-iNEVs and CAR^Ctrl-iNEVs were 113 ± 16 nm and 116 ± 22 nm, respectively, with zeta potentials of −14.2 ± 1.5 mV and −13.73 ± 0.49 mV, respectively (Figures 2B and 2C). Western blot analysis confirmed the successful purification of CAR-iNEVs, as indicated by the presence of the EV markers CD9, CD81, CD63, and Alix, while the endoplasmic reticulum marker calnexin was absent (Figure 2D). Transmission electron microscopy (TEM) further validated the morphology of the CAR-iNEVs, revealing spherical or elliptical vesicle structures with diameters ranging from 50 to 120 nm, consistent with the NTA measurements. Additionally, immunogold labeling combined with HA-magnetic bead enrichment confirmed the presence of HA-tagged CD133 Nbs on the surfaces of the CAR-iNEVs (Figure 2E). Flow cytometry analysis further supported these findings: compared with unmodified iNEVs, CAR-iNEVs presented a significantly stronger fluorescein isothiocyanate (FITC) fluorescence signal upon anti-HA-FITC antibody staining, confirming the effective surface presentation of the CAR structure (Figure 2F). ELISA-based quantification revealed that each CAR-iNEV carried approximately 86 Nb molecules (Figure S2A). Stability testing showed that CAR-iNEVs maintained a consistent size distribution and particle concentration for up to 60 days at 4°C (Figure S2B). In parallel, we assessed EV markers, functional proteins (perforin and Granzyme B), NK cell surface receptors (NKG2D and NKp46), and cytotoxic activity, all of which remained unchanged relative to day 0 (Figures S2C and S2D). Together, these results demonstrate that CAR-iNEV preserves both structural and functional integrity during refrigerated storage, underscoring their stability. Additionally, western blot analysis showed that CAR-engineered NK cells and their derived EVs from iNK, PB-NK, and NK-92 sources expressed comparable levels of cytotoxic proteins and activating receptors (Figure S2E). Given the superior expandability, stability, and production consistency of iNKs, we selected them as the optimal source for CAR-iNEVs. Previous studies have demonstrated that NK cell-derived EVs retain many of the effector molecules of their parental cells. To validate this in our system, we performed a comprehensive comparative analysis of key functional components in CAR^CD133-iNEVs and CAR^Ctrl-iNEVs. Quantitative assays revealed that both CAR^CD133-iNEVs and CAR^Ctrl-iNEVs contain high levels of cytotoxic proteins (perforin and Granzyme B) and proinflammatory factors (tumor necrosis factor [TNF]-α and IFN-γ; Figures S2F and S2G).

Figure 2.

Preparation and characterization of CAR-iNEVs

(A) Schematic illustration showing the process used to generate and isolate CAR-iNEVs.

(B and C) NTA measuring the particle size distributions (left) and zeta potentials (right) of different types of CAR-iNEVs.

(D) Western blot analysis of EV markers in CAR-iNK cells and isolated CAR-iNEVs.

(E) Representative TEM (top, scale bar, 200 nm), cryo-electron microscopy (cryo-EM) (middle, scale bar, 50 nm), and immunogold labeling (IEM) images of CAR-iNEVs (bottom, scale bar, 50 nm). The blue arrows indicate HA-tag-conjugated beads attached to CAR-iNEVs, and the red arrows mark CAR-iNEVs. Bead size: 10 nm.

(F) Flow cytometric analysis of nanobody surface expression in engineered CAR-iNEVs using FITC-labeled anti-HA antibodies.

(G) Immunofluorescence analysis showing the uptake of PKH67-labeled CAR-iNEVs (green) by CD133-expressing cancer cells (Panc02, H22, and Colon26; scale bars, 10 μm).

(H) Cellular internalization of CAR-iNEVs (green) in cancer cells expressing or lacking CD133 (scale bar, 10 μm).

(I) ELISA quantification of the binding affinity of CAR-iNEVs for human (H) and murine (M) CD133 proteins (n = 3 biological replicates).

(J) Cell-based ELISA confirming the interaction between CAR-iNEVs and CD133-expressing cancer cells (n = 3 biological replicates).

(K) Cell-based ELISA evaluating the binding of CAR-iNEVs to CD133-KO Panc02 cells (n = 3 biological replicates).

(L) Luciferase-based cytotoxicity assay assessing the cytotoxic effects of CAR-iNEVs on wild-type and CD133-KO cancer cells (Panc02-Luc, H22-Luc, and Colon26-Luc) (n = 3 biological replicates). The data are presented as the means ± SDs. Statistical significance was determined using two-way ANOVA with Tukey’s post hoc test (I–L) or one-way ANOVA with Tukey’s post hoc test (G). H, human; M, mouse. ∗∗∗p < 0.001.

To visually evaluate the tumor-targeting ability of CAR-iNEVs, we labeled the vesicles with PKH67 fluorescent dye and incubated them with fixed tumor cells. Confocal microscopy imaging revealed that CAR^CD133-iNEVs were highly enriched on CD133-expressing tumor cell membranes, whereas almost no binding of CAR^Ctrl-iNEVs was observed (Figure 2G). Next, we investigated the process through which live cells internalize CAR-iNEVs. Tumor cells were coincubated with PKH67-labeled CAR-iNEVs, and cellular uptake of the CAR-iNEVs was monitored. Compared with the CAR^Ctrl-iNEV group, the CAR^CD133-iNEV group showed rapid internalization of the labeled CAR-iNEVs within 12 h, as indicated by the significantly greater fluorescence intensity of the cells in that group. Moreover, the increase in fluorescence intensity observed in the CAR^CD133-iNEV group was abolished upon CD133 KO (Figure 2H). To assess the link between CAR-iNEV uptake and cytotoxicity, we performed time-course assays in which CAR-iNEVs were co-cultured with tumor cells for different time points. Tumor cell death increased with co-culture time and plateaued at about 12 h. Blocking EV internalization with the dynamin inhibitor Dynasore markedly reduced CAR-iNEV-mediated killing (Figures S2H and S2I). Together, these results support a mechanistic model where CAR-iNEVs must first be efficiently taken up by target tumor cells, allowing for the release of its cytotoxic effector payload inside the cells to trigger tumor cell death.

To further verify the CD133 specificity of CAR-iNEVs, we performed ELISA and cell-based ELISA. The ELISA results confirmed that CAR^CD133-iNEVs bound strongly to both human and murine recombinant CD133 proteins (Figure 2I). Additionally, a cell-based ELISA demonstrated that CAR^CD133-iNEVs effectively recognize and bind to endogenous CD133 on the surfaces of gastric, liver, and pancreatic cancer cells (Figure 2J). Notably, binding of CAR^CD133-iNEVs was almost absent in CD133-KO cells (Figure 2K). Finally, we assessed the cytotoxic activity of CAR-iNEVs in vitro, CAR^Ctrl-iNEVs and Blank-iNEVs served as controls. Although CAR^Ctrl-iNEVs and Blank-iNEVs exhibited baseline cytotoxic effects, CAR^CD133-iNEVs induced significantly increased tumor cell lysis in all three tumor cell lines, whereas this effect was not observed in CD133-KO cells (Figure 2L). The increased antitumor activity of CAR^CD133-iNEVs can be attributed to their superior targeting affinity and internalization efficiency. Taken together, these results show that we successfully engineered and purified CAR^CD133-iNEVs that exhibited exceptional CD133 specificity, robust tumor cell binding, and potent cytotoxic activity.

In vivo safety and tumor targeting of CAR-iNEVs

Safety remains a critical consideration in the clinical application of CAR-T/CAR-NK cell therapy. In this study, we systematically evaluated the biosafety profile of CAR-iNEVs and compared their therapeutic efficacy and toxicity with those of CAR-iNK cells. To rigorously assess potential toxicity, mice were administered injections of CAR-iNK cells and CAR-iNEVs. According to previous reports, the effective circulating dose of CAR-NK cells is approximately 1 × 10⁷ cells, which informed the design of our dosing regimen.^11,34,35 In the H22-induced HCC mouse model, mice were administered CAR-iNEVs at high doses (1 × 10¹² particles) or low doses (1 × 10¹¹ particles) or CAR-iNK cells (1 × 10⁷ cells) (Figure 3A). The physiological conditions of the mice were closely monitored, and humane euthanasia was performed if the animals’ body weight loss exceeded 15% or if severe distress was observed. The results revealed that no mortality was observed in any of the CAR-iNEV-treated groups, regardless of the administered dose (Figure 3B). To further assess systemic toxicity, we performed hematological and biochemical analyses of the treated mice. The mice treated with CAR^CD133-iNK cells exhibited severe adverse effects, including weight loss, convulsions, and mortality, accompanied by abnormal blood cell counts. In contrast, the CAR^CD133-iNEV-treated groups displayed no overt toxic reactions, indicating that CAR-iNEV therapy has a wider therapeutic window and reduced toxicity risk compared to CAR-iNK therapy (Figures 3C and S3A). Furthermore, H&E staining of the major organs of high- or low-dose CAR^CD133-iNEV-treated mice revealed no signs of tissue damage, further supporting the excellent biocompatibility and safety profile of CAR^CD133-iNEVs (Figure S3B). At either dose, CAR^CD133-iNEV treatment significantly reduced tumor volume and weight, whereas the CAR^Ctrl-iNEV group exhibited no notable tumor suppression (Figure 3D). Of the tested conditions, treatment with CAR^CD133-iNEVs at high doses resulted in the most potent tumor inhibition, with no observable toxicity, and the high dose was thus used in the subsequent experiments. These results establish CAR^CD133-iNEV as a well-tolerated and effective immunotherapeutic candidate and provide strong support for its clinical translation.

Figure 3.

Tumor-targeting efficacy and safety profile of CAR-iNEV-based therapy

(A) Schematic representation of the CAR-iNK and CAR-iNEV treatment regimens. Murine-derived CAR-iNK cells and CAR-iNEVs were utilized. At the designated time points, H22 tumor-bearing mice were intravenously injected with PBS (200 μL), CAR^CD133-iNK (1 × 10⁷ cells/injection), CAR^Ctrl-iNEV-high (1 × 10¹² particles/injection), CAR^Ctrl-iNEV-low (1 × 10¹¹ particles/injection), CAR^CD133-iNEV-high (1 × 10¹² particles/injection), or CAR^CD133-iNEV-low (1 × 10¹¹ particles/injection).

(B) Kaplan-Meier survival analysis of mice intravenously injected with PBS, CAR^CD133-iNK-low, or various doses of CAR-iNEVs (n = 5 biological replicates).

(C) Incidence of severe adverse reactions (cachexia, seizures, diarrhea, and mortality) in mice intravenously injected with CAR-iNK or various doses of CAR-iNEVs (n = 5 biological replicates).

(D) Growth curves of tumors in the different treatment groups (n = 5 biological replicates).

(E) Representative images of in vivo tumor imaging were used to evaluate the tumor-targeting efficiency of CAR^Ctrl-iNEVs and CAR^CD133-iNEVs. Tumor-bearing mice were intravenously injected with CAR^Ctrl-iNEVs (1 × 10¹² particles/injection/200 μL) or CAR^CD133-iNEVs (1 × 10¹² particles/injection/200 μL; n = 5 biological replicates).

(F) Quantification of in vivo tumor imaging signals from (E).

(G) Ex vivo imaging of tumors and major organs collected from the mice in (E) to assess biodistribution (n = 5 biological replicates).

(H) Quantification of ex vivo imaging signals from (G).

(I) Immunohistochemical staining of tumor and normal organ tissues with anti-VHH antibodies (n = 5 biological replicates; scale bars: 20 μm in the and 5 μm in the right). Red arrows indicate positive VHH staining in tumor tissues.

(J) Flow cytometric analysis of CAR^CD133-iNEV binding to various normal cell types in the subcutaneous H22 cell-derived mouse model, showing minimal interaction compared with strong binding to CD133⁺ tumor cells.

(K) Quantitative biodistribution analysis of ⁸⁹Zr-labeled CAR^CD133-iNEVs in tumor and major organs at 24 h post-injection in the H22 cell-derived subcutaneous mouse model.

(L) Time course of ⁸⁹Zr-CAR^CD133-iNEV accumulation in CD133⁺ tumors following systemic administration. The data are presented as the means ± SDs. Statistical significance was determined using two-way ANOVA (F, I, and L), one-way ANOVA with Tukey’s post hoc test (K), or the Mantel-Cox test for survival analysis (B). ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001.

To assess the tumor-targeting ability of CAR^CD133-iNEVs and track their biodistribution in vivo, we employed an in vivo imaging system (IVIS) for live-animal imaging. When the animals’ tumor volumes reached ∼150 mm³, the animals were injected with CAR^Ctrl-iNEVs or CAR^CD133-iNEVs, and fluorescence signals were recorded at 4, 8, and 24 h post-injection. Compared with the CAR^Ctrl-iNEV group, the CAR^CD133-iNEV group presented significantly greater fluorescence signals at the tumor site, and these signals persisted for at least 24 h, indicating sustained accumulation of CAR^CD133-iNEVs at the tumor site (Figures 3E, 3F, and S3C). Ex vivo imaging further confirmed the preferential enrichment of CAR^CD133-iNEVs within tumor tissues. Fluorescence analysis of tumor sections revealed intense signal accumulation in the CAR^CD133-iNEV group, with minimal signal accumulation in major organs other than the liver (Figures 3G and 3H). Immunohistochemical staining further confirmed that CAR^CD133-iNEVs selectively accumulated in tumors, where they yielded Nb signals that were significantly greater than those in the CAR^Ctrl-iNEV group. Importantly, no substantial signal accumulation was detected in other organs, confirming the tumor specificity of CAR^CD133-iNEVs (Figure 3I). To further quantify their tumor-targeting capability, we incorporated additional in vivo flow cytometry and zirconium-89 (⁸⁹Zr) radiotracer imaging analyses. These data demonstrated that CAR^CD133-iNEVs preferentially accumulated in CD133⁺ tumor tissues, with minimal circulation levels and low uptake in major normal organs (Figures 3J–3L and S3D). This tumor-enriched biodistribution underpins their favorable safety profile by limiting off-target exposure.

CAR^CD133-iNEVs induce direct tumor cell killing and reprogram the tumor immune microenvironment

To further validate the therapeutic efficacy of CAR-iNEVs, we established xenograft tumor models for HCC (H22), pancreatic cancer (Panc02), and colorectal cancer (Colon26) in mice. The model mice were intravenously injected with CAR^CD133-iNEVs at a high dose (1 × 10¹² particles) three times per week, and tumor volume and body weight were monitored regularly (Figure 4A). The results revealed that treatment with CAR^CD133-iNEVs significantly inhibited tumor growth in all the xenograft tumor models (Figure 4B). No significant changes in body weight and no hematological abnormalities were observed during the treatment period, indicating that CAR^CD133-iNEVs have an excellent safety profile, even at high doses (Figures S4A and S4B). To further mimic the native TME, we established an orthotopic HCC model using luciferase-labeled Hep53.4 cells. IVIS live-animal imaging demonstrated that treatment with CAR^CD133-iNEVs effectively suppressed tumor growth and significantly prolonged survival in tumor-bearing mice (Figures 4C–4E and S4C). Next, we evaluated the clinical relevance of CAR^CD133-iNEV therapy using humanized immune system (HIS) mice (huHSC-NCG-M) engrafted with patient-derived xenografts (PDXs) of liver (n = 4), pancreatic (n = 3), and colorectal (n = 3) cancers. CAR-iNEVs (human) were generated from human iPSC-derived NK cells and validated for CAR loading by western blotting, flow cytometry, electron microscopy, and NTA (Figures S5A–S5D). Notably, CAR^CD133-iNEV (human) treatment significantly suppressed tumor progression in 7 of 10 CD133⁺ PDX models (Figures 4F–4H). Furthermore, no significant weight loss or treatment-related adverse effects were observed in HIS mice, further supporting the favorable safety and translational potential of CAR^CD133-iNEV therapy (Figures 4I and 4J).

Figure 4.

CAR-iNEV therapy potently eliminates tumor cells in mice

(A) Schematic showing the treatment regimens for liver cancer (H22), colon cancer (Colon26), and pancreatic cancer (Panc02) xenograft tumor models using intravenous injection of CAR^Ctrl-iNEVs (1 × 10¹² particles/injection/200 μL) and CAR^CD133-iNEVs (1 × 10¹² particles/injection/200 μL) (n = 10 biological replicates).

(B) Tumor growth curves following treatment with intravenous injection of CAR^Ctrl-iNEVs (1 × 10¹² particles/injection/200 μL) or CAR^CD133-iNEVs (1 × 10¹² particles/injection/200 μL) in various tumor models (n = 10 biological replicates).

(C) Establishment of an orthotopic HCC model using Hep53.4-Luc cells. Mice were intravenously injected with PBS (200 μL), CAR^Ctrl-iNEVs (1 × 10¹² particles/injection/200 μL), or CAR^CD133-iNEVs (1 × 10¹² particles/injection/200 μL). Tumor size was monitored via in vivo imaging and ex vivo liver analysis (n = 5 biological replicates).

(D) Quantification of tumor weights and fluorescence intensity from (C) (n = 5 biological replicates).

(E) Kaplan-Meier survival curves of orthotopic HCC-bearing mice with intravenously injected of CAR^Ctrl-iNEVs (1 × 10¹² particles/injection) or CAR^CD133-iNEVs (1 × 10¹² particles/injection) (n = 5 biological replicates).

(F) Schematic showing the treatment regimen used in PDX tumor-bearing mice.

(G and H) (G) IHC-confirmed CD133 expression levels in PDX tissues from liver, pancreatic, and colorectal cancers. (H) Tumor growth curves following intravenous injection of CAR^CD133-iNEVs or CAR^Ctrl-iNEVs (n = 10 biological replicates; scale bars, 200 μm).

(I) Hematological parameters of mice showed no significant changes after blood samples were collected from all animals when the first mouse reached a tumor volume of 1,500 mm³ at day 20 of treatment (n = 10 biological replicates).

(J) Body weight of mice remained stable throughout the treatment period (n = 10 biological replicates). The data are presented as the means ± SDs. Statistical significance was determined using a two-way ANOVA (B, H, and J), a one-way ANOVA (D), an unpaired two-tailed t test (I), or the Mantel-Cox test for survival analysis (E). ∗∗p < 0.01, ∗∗∗p < 0.001.

CAR^CD133-iNEV-driven activation of macrophages and neutrophils reshapes the tumor immune microenvironment

To determine whether CAR^CD133-iNEVs can modulate the TME, we performed single-cell RNA sequencing on CD45⁺ immune cells isolated from tumor tissues (Panc02) collected 2 weeks post-treatment. Preliminary analysis revealed significant alterations in the composition of tumor-infiltrating immune cell subsets in the CAR^CD133-iNEV treatment group, suggesting that the treatment had a profound reprogramming effect on immune cells (Figures S6A–S6C). The macrophages found in the tissues were classified into 11 distinct clusters; a substantial increase in the proportion of proinflammatory macrophages and a marked reduction in the proportion of tumor-promoting C1q⁺ macrophages were observed following CAR^CD133-iNEV treatment (Figures 5A, 5B, S7A, and S7B). Gene enrichment analysis revealed that the key upregulated genes in macrophages in the treatment group were associated with proinflammatory responses (Figure S8A). Additionally, Uniform manifold approximation and projection (UMAP) analysis identified seven neutrophil subsets (Clusters 0–7) within the tumor tissues (Figure S9A). CAR^CD133-iNEV treatment markedly increased neutrophil abundance and enriched subsets (Clusters 0, 1, 3, 6, and 7) associated with chemotaxis, inflammatory activation, and innate immune responses, consistent with antitumor N1-like polarization. In contrast, Clusters 5 and 6 were more prevalent in the control group and linked to pathways involved in tissue development and growth regulation (Figures S9B and S9C). CAR-iNEV-treated neutrophils also displayed elevated expression of CCL3 and CCL4, hallmark features of N1 polarization, further supporting their role in promoting an antitumor neutrophil phenotype within the TME (Figure S9D).

Figure 5.

The antitumor effects of CAR^CD133-iNEVs depend on activation of macrophage NOS2

(A) UMAP visualization of the distribution of macrophages in tumors from CAR^CD133-iNEV-treated (intravenous injection, 1 × 10¹² particles, 200 μL) and CAR^Ctrl-iNEV-treated (intravenous injection, 200 μL) Panc02 established xenograft tumor model mice (n = 3 biological replicates).

(B) Proportional distribution of major immune cell types in the CAR^CD133-iNEV and PBS treatment groups (n = 3 biological replicates).

(C) Pseudotime trajectory analysis of macrophages from the two groups. Macrophage subtypes are color coded on the basis of Seurat clustering, treatment group, pseudotime state, and pseudotime value (left to right).

(D) Branched expression analysis modeling (BEAM) in Monocle2 revealed distinct gene expression patterns during macrophage fate transitions. Differentially expressed genes in each cluster were enriched using Gene Ontology Biological Processes.

(E) Expression of macrophage marker genes in the two branches.

(F) Metabolomic profiling of glycolytic intermediates in macrophages from CAR^CD133-iNEV-treated (intravenous injection, 200 μL, 1 × 10¹² particles) and PBS-treated (200 μL) tumors using liquid chromatography-mass spectrometry.

(G) ELISA quantification of lactate production and glucose consumption by macrophages in the tumor microenvironment following treatment with CAR^CD133-iNEVs (1 × 10¹² particles/injection) (left). ECAR measurements of aerobic glycolysis in macrophages (right).

(H) Schematic representation of the treatment regimens used in Panc02 tumor-bearing mice; the treatments included intravenous injection of CAR^CD133-iNEVs (1 × 10¹² particles/200 μL/injection), anti-Ly6G (0.5 mg/200 μL/injection), and clodronate (200 μg/200 μL/injection) (left). Tumor growth curves observed after various treatments (right) (n = 5 biological replicates).

(I) Experimental design for the CAR^CD133-iNEV (1 × 10¹² particles/200 μL/injection) intravenous injection treatment of macrophage-specific Nos2^cKO (F4/80-Cre; Nos2^flox/flox) tumor-bearing mice (left). Tumor growth curves of wild-type (WT) and Nos2^cKO mice after various treatments (right) (3 times/week, 200 μL/injection) (n = 5 biological replicates). Statistical significance was calculated using two-way ANOVA with Tukey’s post hoc test (H and I) or an unpaired two-tailed t test (G). ns indicates no significance. ∗p < 0.05, ∗∗p < 0.01.

Given the pronounced functional remodeling of macrophages upon CAR^CD133-iNEV treatment, we further conducted pseudotime trajectory analysis using Monocle2. Cluster 3 cells were identified as classical monocytes, and State 2, which mainly involved Cluster 3, was designated as the starting point of the trajectory (Figure 5C). As pseudotime progressed, macrophages diverged at branchpoint 2, differentiating into two distinct lineages. One branch, which was predominantly composed of Cluster 1 and 2 cells from the CAR^CD133-iNEV group, exhibited a proinflammatory phenotype, whereas the other branch, which was mainly composed of Cluster 0 cells from the control group, retained an immunosuppressive signature (Figure 5D). Analysis of macrophage gene expression revealed that Ifitm2, Nos2, and Cxcl2 were highly expressed along the first branch, with their expression levels increasing over pseudotime. In contrast, the second branch displayed high expression of immunosuppressive macrophage markers, including C1qa, C1qb, and C1qc (Figure S10A). Furthermore, BEAM algorithm-based branch-dependent gene expression analysis identified four distinct gene clusters associated with branchpoint 2. Before branchpoint 2, the macrophage clusters (Clusters 2 and 4) presented signatures associated with oxidative phosphorylation and mRNA processing. However, macrophages differentiating toward the CAR^CD133-iNEV group-enriched state (Cluster 3) were strongly enriched in cytokine-mediated signaling pathways and immune response-activating pathways, indicating an increase in immune effector functions. Conversely, macrophages that differentiated toward the PBS group state (Cluster 1) were enriched in chromatin remodeling and cell cycle regulation processes, resulting in significantly reduced immune activation compared with that at branchpoint 2 (Figures 5E and S10B). Collectively, these findings demonstrate that CAR^CD133-iNEV treatment effectively stimulates macrophage immune activation, shifting the TME toward a more immunostimulatory and proinflammatory state.

Previous studies have demonstrated that increased glycolytic activity in macrophages and neutrophils is typically associated with a high activation state; this plays key roles in inflammation, immune responses, and the TME.^36,37,38 To investigate the metabolic status of these immune cells in the context of CAR^CD133-iNEV treatment, we isolated macrophages and neutrophils from tumor tissues by flow cytometry-based sorting. Metabolomic analysis revealed a significant increase in aerobic glycolysis in both cell populations. These findings were further validated by ELISA and by measurement of the extracellular acidification rate (ECAR), confirming a metabolic shift toward glycolysis in CAR^CD133-iNEV-activated macrophages and neutrophils (Figures 5F, 5G, and S11A–S11C). Co-culture experiments revealed that direct exposure of macrophages or neutrophils to CAR^CD133-iNEVs had minimal phenotypic effects, whereas conditioned media from CAR-iNEV-treated tumor cells markedly upregulated M1 macrophage markers (CD86 and MHC II) and induced an N1-like neutrophil phenotype with elevated TNF-α and interleukin-12 (Figures S11D and S11E). These findings indicate that CAR-iNEV-mediated activation of macrophages and neutrophils occurs indirectly via tumor cell modulation and is insufficient to cause excessive neutrophil activation.

To determine whether CAR^CD133-iNEV-mediated tumor suppression is dependent on macrophages or neutrophils, we selectively depleted these immune cells in vivo using antibodies against Ly6G (for neutrophil depletion) or clodronate liposomes (for macrophage depletion). The results demonstrated that depletion of either neutrophils or macrophages significantly impaired the therapeutic efficacy of CAR-iNEVs (Figure 5H). Notably, macrophage depletion almost completely abrogated the antitumor effects of CAR^CD133-iNEVs, suggesting that tumor suppression by CAR^CD133-iNEVs relies primarily on macrophage activation. To assess the role of T cells in CAR-iNEV-mediated tumor killing, we evaluated their efficacy in Rag1-KO (T-cell-deficient) mice and nude mice. Strikingly, CAR^CD133-iNEVs remained effective in both models, demonstrating that their tumor-killing ability is independent of T cell involvement (Figures S11F–S11H).

The production of reactive oxygen species is a well-established mechanism by which macrophages eliminate pathogens and tumor cells. Given that macrophages in CAR^CD133-iNEV-treated tumors presented elevated levels of inducible nitric oxide synthase (NOS2), we generated macrophage-specific NOS2 conditional KO (F4/80-Cre; NOS2^flox/flox) mice for further investigation. The results revealed that CAR^CD133-iNEV-mediated tumor suppression was significantly diminished in NOS2^cKO mice (Figure 5I). These findings suggest that CAR^CD133-iNEV exerts its antitumor effects through activation of macrophages and neutrophils, primarily by inducing macrophage NOS2 upregulation.

Synergistic inhibition of metastatic cancer and induction of immunological memory by combination therapy

Given the profound role of CAR-iNEVs in remodeling the TME, we further explored their potential in combination therapies. Immune checkpoint inhibitors (ICIs), such as antibodies against PD-1, CTLA-4, and CD47, are essential strategies in cancer immunotherapy. We next explored the therapeutic synergy between CAR^CD133-iNEVs and ICIs. Our results showed that monotherapy with either CAR^CD133-iNEVs or PD-1/CTLA-4/CD47 antibodies partially suppressed tumor growth, whereas treatment with CAR^CD133-iNEVs combined with a CD47 antibody completely blocked tumor progression (Figures 6A and 6B).

Figure 6.

CAR^CD133-iNEVs synergize with anti-CD47 therapy to suppress tumor progression and induce durable immune memory

(A) Schematic diagram showing the treatment regimen used in HCC tumor-bearing mice receiving intravenous injection of CAR^CD133-iNEVs (1 × 10¹² particles/injection) combined with PD-1 (10 mg/kg), CTLA-4 (10 mg/kg), or an antibody against CD47 (10 mg/kg) 3 times/week for 3 weeks (n = 5 biological replicates).

(B) Tumor growth curves of HCC-bearing mice treated with CAR^CD133-iNEVs alone or in combination with immune checkpoint inhibitors (n = 5 biological replicates).

(C) Representative bioluminescence imaging of pulmonary metastatic tumors with Panc02-Luc cells after various treatment strategies (n = 5 biological replicates).

(D) Representative bioluminescence imaging of abdominal metastatic tumors with Panc02-Luc cells after various treatment strategies (n = 5 biological replicates).

(E) Tumor growth curves of PDX-bearing mice resistant to PD-1 blockade and treated with intravenous injection of CAR^CD133-iNEV + anti-CD47 (10 mg/kg) (n = 5 biological replicates).

(F) Tumor growth curves of PDX-bearing mice resistant to chemotherapy/radiotherapy and treated with intravenous injection of CAR^CD133-iNEVs (1 × 10¹² particles/injection) + anti-CD47 (10 mg/kg) (n = 5 biological replicates).

(G) Tumor growth curves of 5 mice that were randomly selected from the 7 previously cured mice rechallenged with the same Panc02-Luc cells at 120 days post-treatment compared with those of age-matched control mice (n = 5 biological replicates).

(H) Flow cytometry analysis of effector memory T cells (CD62L⁻ CD44⁺) in the spleens of rechallenged or control mice at the experimental endpoint (n = 5 biological replicates).

(I) Flow cytometry analysis of effector memory T cells (CD62L⁻ CD44⁺) in the spleens of rechallenged or without rechallenge mice (n = 3 biological replicates). Statistical significance was calculated using two-way ANOVA with Tukey’s post hoc test (B–G) or a two-tailed unpaired t test (H and I). ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001.

To assess the efficacy of CAR^CD133-iNEVs and CD47 antibodies in clinically relevant metastatic models, we established peritoneal and pulmonary metastasis models in mice. Luciferase-labeled Hep53.4 cells were injected intraperitoneally or intravenously into mice, and bioluminescence imaging was used to monitor tumor progression. In the pulmonary metastasis model, monotherapy with CAR^CD133-iNEVs or with an antibody against CD47 only modestly delayed tumor progression, and the mice ultimately succumbed to the tumor burden. In contrast, combination therapy significantly suppressed metastatic lesions and prolonged survival, with 80% (4/5) of the mice achieving complete tumor clearance. Similarly, in the peritoneal metastasis model, combination therapy with CAR^CD133-iNEVs and CD47 blockade led to complete tumor regression in 60% of the mice, significantly extending their survival (Figures 6C, 6D, and S11I). These data suggest that treatment with CAR^CD133-iNEVs combined with CD47 blockades effectively eradicates metastatic tumors and markedly improves survival outcomes.

To assess the therapeutic potential of CAR^CD133-iNEVs and CD47 blockade in PD-1-resistant tumors, we established HIS-PDX models. The tumors were derived from five pancreatic cancer patients in whom neoadjuvant PD-1 therapy had failed, and the tumor tissues were implanted into HIS mice. The results revealed that CD47 blockade alone failed to significantly suppress tumor progression, whereas CAR^CD133-iNEV monotherapy partially delayed tumor growth but was insufficient to achieve complete remission. However, treatment with CAR^CD133-iNEVs combined with anti-CD47 therapy significantly suppressed tumor growth and led to nearly complete tumor clearance (Figure 6E). We also included five PDX samples from colon cancer patients who had failed chemotherapy or radiotherapy. Combination therapy with CAR^CD133-iNEVs and CD47 blockade demonstrated comparable efficacy across these PDXs and significantly inhibited tumor growth (Figure 6F).

Given the superior efficacy of CAR^CD133-iNEVs together with CD47 blockade in eliminating metastatic tumors, we next examined whether this therapy could induce durable immunological memory. At 120 days post-tumor clearance, the tumor-free mice were rechallenged with the same tumor cells (Panc02-Luc cells). No tumor recurrence was observed in the combination therapy group, whereas tumors developed rapidly in age-matched control mice (Figure 6G). Flow cytometry analysis of the splenic T cell subsets present in the mice revealed a significant increase in the proportion of effector memory T cells (CD62L⁻CD44⁺) within the CD4⁺ and CD8⁺ T cell populations after subcutaneous inoculation with Panc02 cells (Figure 6H). Moreover, we demonstrated that this increase was not due to residual T cell activation, as effector memory T cells remained at baseline levels in mice without rechallenge (Figure 6I). These results suggest that combined treatment with CAR^CD133-iNEVs and CD47 blockade not only achieves long-term tumor eradication but also induces robust T-cell-mediated immunological memory, providing long-term protection against tumor recurrence.

Discussion

As a cell-free vesicular therapy, CAR^CD133-iNEV therapy offers distinct advantages over conventional CAR-T and CAR-NK cell therapies in the highly immunosuppressive TME. These engineered nanovesicles can more effectively penetrate dense tumor stromal barriers than can the cells used in the other therapies and are inherently less susceptible to TME-derived immunosuppressive signals. Unlike CAR^CD133-iNK cells, which retain innate cytotoxicity through natural activating receptors, CAR-iNEVs rely almost exclusively on CAR-mediated antigen recognition for cytotoxic activity. This reliance confers greater target selectivity and may reduce off-target effects in clinical applications. The enhanced tumor-penetrating ability of CAR-iNEVs is not solely attributable to their nanoscale size. In addition to favorable dimensions for passive diffusion, EVs can actively remodel the TME. CAR-iNEVs express extracellular matrix (ECM)-modulating proteins such as MMP1, MMP14, and CD147, which promote collagen degradation and stromal loosening, and can modulate CAF activity to reduce ECM deposition and matrix stiffness.²⁰ These properties facilitate deeper infiltration into tumor tissue, consistent with prior observations that NK cell-derived EVs penetrate more effectively into tumor spheroids than their parental NK cells.²¹ Collectively, these mechanisms support the selection of EVs over cell-based platforms for overcoming stromal barriers in solid tumors.

Studies have shown that CAR-T-derived vesicles can deliver stimulatory RNAs that activate type I IFN signaling, thereby enhancing CD8⁺ T cell priming and promoting tumor clearance.³⁹ Similarly, EVs released by innate immune cells naturally contain cytotoxic proteins and immunomodulatory cytokines, facilitating crosstalk with other immune cells within the TME.^40,41,42 Notably, CAR^CD133-iNEV treatment led to a marked increase in M1-like macrophages and N1-like neutrophils, and this pro-inflammatory shift was strongly associated with tumor volume reduction. These actions not only directly target tumor cells but also remodel the TME by amplifying inflammatory cues and suppressing immunosuppressive components, establishing a self-reinforcing cycle of anti-tumor immunity. Together, our findings suggest that CAR-iNEVs promote tumor regression via a dual mechanism: direct tumor cell killing and reprogramming of the TME through activation of pro-inflammatory M1 and N1 subsets, resulting in enhanced therapeutic efficacy.

In addition to their immunotherapeutic properties, CAR^CD133-iNEVs can be synergistically combined with ICIs to maximize therapeutic efficacy. In this study, CAR^CD133-iNEV treatment combined with CD47 immune checkpoint blockade significantly increased tumor clearance and induced long-term immunological memory in surviving mice. This synergistic effect underscores the ability of CAR^CD133-iNEVs to be integrated with complementary immunotherapies, effectively engaging both innate and adaptive immune responses to achieve greater and more durable antitumor efficacy than can be achieved through the use of monotherapy.

In summary, CAR^CD133-iNEV represents not only a safe and potent stand-alone therapeutic modality but also a versatile platform for integration into combinatorial immunotherapies, enabling personalized and precision treatment of solid tumors. To address the quantity and quality requirements for clinical-scale CAR^CD133-iNEV production, we propose TFF as a core platform, given its superior yield, purity, and batch-to-batch consistency.³² Recent studies using hollow-fiber bioreactors combined with TFF and chromatography have achieved ∼10¹³ EV particles per batch with preserved cytotoxic activity and up to 100-fold lower contaminant protein level,^43,44 providing a feasible roadmap for large-scale, high-quality CAR^CD133-iNEV manufacturing. Future studies should focus on validating the efficacy and safety of their use in advanced preclinical and clinical trials.

Limitations of the study

Despite their numerous advantages, CAR^CD133-iNEVs still faces several challenges that must be addressed to facilitate their clinical translation. First, the large-scale production, purification, and characterization of CAR^CD133-iNEVs remain unstandardized. Given the potential heterogeneity of EV preparations, stringent quality control measures and manufacturing protocols are essential to ensure consistency, reproducibility, and therapeutic efficacy. Second, the therapeutic efficacy and safety of CAR^CD133-iNEVs have been evaluated primarily in murine tumor models. To increase their translational potential, additional studies conducted in large animal models are necessary to further validate their long-term efficacy, optimize dosing regimens, and assess potential toxicities.

Resource availability

Lead contact

Requests for further information, resources, and reagents should be directed to and will be fulfilled by the lead contact, Peng Xia (pengxia@uchicago.edu).

Materials availability

All unique/stable reagents generated in this study are available from the lead contact with a completed materials transfer agreement.

Data and code availability

• •High-throughput sequencing data are available through the GEO database via the accession number GSE293929.
• •This paper does not report original code.
• •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgments

This research was financially supported by the National Natural Science Foundation of China (82403916), the Chicago Fellows Program, the Cancer Center Support Grant (P30 CA014599), Natural Science Foundation of Shandong Province (ZR2024QH227), and Natural Science Foundation of Liaoning Province (2025JH2/101800397). We thank the Center for Research Informatics of the University of Chicago for the use of the Gardner High-Performance Computing cluster.

Author contributions

Conceptualization, H.Z., B.J., and P.X.; methodology, H.Z., S.L., C.L., Y.J., X.N., Y.G., and P.X.; visualization, C.L., X.G., C.-W.J., W.J., and X.Z.; validation, S.L., H.L., Y.J., and X.N.; funding acquisition, P.X. and H.Z.; data curation, P.X.; software, C.L., H.L., and W.J.; project administration, B.J. and P.X.; supervision, X.Z., B.J., and P.X.; writing – original draft, H.Z., S.L., H.L., and Y.J.; writing – review & editing, C.L., Y.G., and P.X.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Rabbit anti-Camelid VHH Antibody	GenScript	Cat#A01861; RRID: AB_3083750
Rabbit anti-Camelid VHH Cocktail	GenScript	Cat#A02019; RRID: AB_3083750
488 anti-HA.11 Epitope Tag Antibody	BioLegend	Cat#901509; RRID: AB_2565072
Anti-HA-Tag(26D11) mAb	Abmart	Cat#M20003; RRID: AB_2864345
Anti-CD9 antibody	Abcam	Cat#ab236630; RRID: AB_2922400
Anti-ALIX antibody	Abcam	Cat#ab275377; RRID: AB_3644262
Anti-CD63 antibody	Abcam	Cat#ab134045; RRID: AB_2800495
Anti-CD81 antibody	Abcam	Cat#ab79559; RRID: AB_1603682
Anti-TSG101 antibody	Abcam	Cat#ab125011; RRID: AB_10974262
Anti-Calnexin antibody	Abcam	Cat#ab133615; RRID: AB_2864299
Anti-M13 antibody	Abcam	Cat#ab235228; RRID: AB_10790812
Anti-Granzyme B antibody	Abcam	Cat#ab255598; RRID: AB_2860567
Anti-NKG2D antibody	Abcam	Cat#ab36136; RRID: AB_776794
Anti-NCR1 antibody	Abcam	Cat#ab199128; RRID: AB_2890127
Anti-beta Actin antibody	Abcam	Cat#ab8226; RRID: AB_306371
Anti-CD133 antibody	CST	Cat#64326; RRID: AB_3069353
Anti-Perforin antibody	CST	Cat#44865; RRID: AB_10772474
Anti-Granzyme B antibody	CST	Cat#46890; RRID: AB_10784284
Anti-mouse CD45-APC/Cyanine7	Biolegend	Cat#103116; RRID: AB_312981
Anti-mouse CD8a-FITC	Biolegend	Cat#100705; RRID: AB_312744
Anti-mouse CD8a-PE	Biolegend	Cat#100708; RRID: AB_312747
Anti-mouse Ly-6C-FITC	Biolegend	Cat#128006; RRID: AB_1186135
Pacific Blue™ anti-mouse/human CD11b	Biolegend	Cat#101224; RRID: AB_755986
Anti-mouse/human CD11b-PE	Biolegend	Cat#101208; RRID: AB_312791
Anti-mouse IFN-γ-APC	Biolegend	Cat#505809; RRID: AB_315403
Anti-mouse CD11c-APC	Biolegend	Cat#117310; RRID: AB_313779
Anti-mouse CD11c-FITC	Biolegend	Cat#117306; RRID: AB_313775
Pacific Blue™ anti-mouse I-A/I-E	Biolegend	Cat#107620; RRID: AB_493527
Alexa Fluor® 700 anti-mouse Ly-6G-	Biolegend	Cat#127622; RRID: AB_10643269
Anti-mouse Ly-6G-PE	Biolegend	Cat#127608; RRID: AB_1186099
Anti-mouse F4/80-FITC	Biolegend	Cat#123108; RRID: AB_893502
Anti-mouse CD4-FITC	Biolegend	Cat#100405; RRID: AB_312690
Anti-mouse CD4-PE	Biolegend	Cat#100407; RRID: AB_312692
Anti-human CD206-PE/Cyanine5	Biolegend	Cat#321108; RRID: AB_571919
Anti-mouse CD86-APC	Biolegend	Cat#159216; RRID: AB_3106041
Anti-human CD107a-PE	Biolegend	Cat#328607; RRID: AB_1186062
Anti-mouse CD44-AF700	Biolegend	Cat#103025; RRID: AB_493712
Anti-mouse CD62L-PE	Biolegend	Cat#104407; RRID: AB_313094
Anti-Human CD16-FITC	BD Pharmingen	Cat#561308; RRID: AB_10644202
Anti-Human CD178-APC	BD Pharmingen	Cat#564262; RRID: AB_2738714
Anti-Human CD335-PE	BD Pharmingen	Cat#557991; RRID: AB_1833753
Anti-Human CD253-APC	BD Pharmingen	Cat#563642; RRID: AB_2738341
Spark Blue™ 574 anti-mouse NK1.1 Antibody	Biolegend	Cat#285313; RRID: AB_3683243
Ultra-LEAF™ Purified anti-mouse IFN-γ Antibody	Biolegend	Cat#631054; RRID: AB_3097543
Brilliant Violet 421™ anti-mouse CD62L Antibody	Biolegend	Cat#104435; RRID: AB_10900082
Alexa Fluor® 700 anti-mouse CD44 Antibody	Biolegend	Cat#156009; RRID: AB_3068217
Anti-mouse CD8β (Lyt 3.2)	Selleckchem	Cat#A2137; RRID: AB_2687706
InVivoMAb anti-mouse PD-L1	BioXcell	Cat#BE0101; RRID: AB_10949073
InVivoMAb anti-mouse CD47	BioXcell	Cat#BE0270; RRID: AB_2687793
InVivoMAb anti-mouse CTLA-4	BioXcell	Cat#BE0131; RRID: AB_10950184
Bacterial and virus strains
BL21(DE3)	Sigma-Aldrich	69450
T7 phage	Novagen (EMD MIllipore)	70010–3
Chemicals, peptides, and recombinant proteins
Puromycin Dihydrochloride	Gibco	Cat#A1113803
Mouse LIF Recombinant Protein	Gibco	Cat#PMC9484
Human SCF Recombinant Protein	Gibco	Cat#PHC2111
MethoCult™ GF M3434	Fisher scientific	Cat#NC9099909
Mouse SCF Recombinant Protein	PeproTech	Cat#250-03-10UG
Mouse IL-3 Recombinant Protein	PeproTech	Cat#213-13-10UG
Human IL-6 Recombinant Protein	PeproTech	Cat#200-06-20UG
Human IL-2 Recombinant Protein	PeproTech	Cat#200-02-50UG
Mouse IL-15 Recombinant Protein	PeproTech	Cat#210-15-10UG
Human VEGF-165 Recombinant Protein	PeproTech	Cat#100-20-10UG
Human BMP-4 Recombinant Protein	PeproTech	Cat#120-05-5UG
2-Deoxy-D-glucose	MCE	Cat#HY-13966
PKH 67	MCE	Cat#HY-D1421
DiR	MCE	Cat#HY-D1048
Matrigel Matrix	Corning	Cat#356234
Donkey serum	Sigma-Aldrich	Cat#D9663
Imidazole	Sigma-Aldrich	Cat#I202
Ni-NTA Agarose	Qiagen	Cat#30210
Imidazole Wash Buffer	Bioworld	Cat#42320000
RIPA buffer	Beyotime	Cat#P0013E
Lenti-X concentrator	Takara	Cat#631231
methylcellulose	Methocult	Cat#M3434
Purified CD133 protein	This paper	N/A
89Zr	Qilu hospital	N/A
Critical commercial assays
Annexin V-FITC/PI Apoptosis Kit	Biolegend	Cat#640914
PKH67GL-1KT	Sigma-Aldrich	Cat#PKH67GL
TUNEL Staining Kit	Sigma-Aldrich	Cat#APT110
L-Lactate Assay Kit	Sigma-Aldrich	Cat#MAK329
Glucose Uptake Assay Kit	Abcam	Cat#ab136956
Cell Counting Kit	Yeasen	Cat#40203ES60
BCA assay kit	Beyotime	Cat#P0009
Mouse IFN-γ ELISA Kit	Beyotime	Cat#PI508
Mycoplasma PCR Detection Kit	Beyotime	Cat#C0301S
10X Genomics Chromium NextGEM Single Cell 3′ Kit	10× Genomics	Cat#PN-1000121
Luciferase Assay System	Promega	Cat#E1500
Mouse Ccl3 ELISA Kit	JONLNBIO	Cat#JL20414
Mouse Ccl5 ELISA Kit	JONLNBIO	Cat#JL17335
Mouse Cxcl1 ELISA Kit	JONLNBIO	Cat#JL20150
Mouse Cxcl5 ELISA Kit	JONLNBIO	Cat#JL28925
Mouse IL-1α ELISA Kit	JONLNBIO	Cat#JL13044
Mouse IL-1β ELISA Kit	JONLNBIO	Cat#JL18442
Mouse IL-6 ELISA Kit	JONLNBIO	Cat#JL20268
Mouse Tnf-α ELISA Kit	JONLNBIO	Cat#JL10484
Mouse Csf3 ELISA Kit	JONLNBIO	Cat#JL10697
Mouse Cxcl2 ELISA Kit	Abnova	Cat#KA1505
Deposited data
Single-cell sequencing data	This paper	GEO: GSE293929
Experimental models: Cell lines
HEK 293T/17	ATCC	Cat#CRL-11268
K562	ATCC	Cat#CCL-243
Panc02	ATCC	Cat#CRL-2553
Mouse iPSC	Alstem	Cat#iPS02M
Human iPSC	ATCC	Cat#ACS-1027
H22	Ubigene Biosciences	Cat#YC-C053
Colon26	Creative biogene	Cat#CSC-RR0461
Hep-53.4	CLS Cell Lines Service	Cat#400200
Panc02-Luc	This paper	N/A
Colon26-Luc	This paper	N/A
Hep-53.4-Luc	This paper	N/A
Experimental models: Organisms/strains
F4/80-Cre mice	The Jackson laboratory	N/A
huHSC-NCG-M	Gempharmatech	Cat#T057533
BALB/c Mice	Charles	Cat#028
C57BL/6 Mice	Charles	Cat#027
C57BL/6Smoc-Nos2em1(flox)Smoc	Shanghai Model Organisms	Cat#NM-CKO-226494
Macrophage-specific Nos2-KO mice	This paper	N/A
Recombinant DNA
pCDH-CMV-MCS-EF1α-GreenPuro	System Biosciences	Cat#CD513B-1
psPAX2	Addgene	Cat#12260
pMD2.G	Addgene	Cat#12259
pLVX-mbIL-21-puro	Addgene	Cat#125839
Luciferase lentivirus	Applied Biological Materials	Cat#LVP010079
pET-14B-CD133	This paper	N/A
Software and algorithms
ImageJ	Schneider et al.45	https://imagej.net/ij/
Graphpad Prism version 10.0	GraphPad software	https://www.graphpad.com
FlowJo 10.0	FlowJo, LLC	https://www.flowjo.com/
Living Image software v4.5	PerkinElmer	N/A
SPSS 20.0	IBM	https://www.ibm.com/it-it/products/spss-statistics
Biacore S200 Evaluation software	Cytiva	https://www.cytivalifesciences.com/en/us/support/software/biacore-downloads/
Biorender	Biorender	https://app.biorender.com/
Other
Nanobodies targeting CD133	This paper	N/A

Experimental model and study participant details

Mice

Immunocompetent male C57BL/6 (Cat. 027) and male BALB/c (Cat. 028) mice, aged 8 weeks, were acquired commercially (Charles River Laboratories, IL, USA). Macrophage-specific NOS2 KO mice were generated by crossing Nos2-flox (C57BL/6Smoc-Nos2^{em1(flox)Smoc}) mice (Shanghai Model Organisms Center, Inc., Shanghai, China, Cat. NM-CKO-226494) with F4/80-Cre mice (The Jackson Laboratory). The resulting offspring were genotyped and male Nos2 conditional knockout mice (8 weeks old) were used for experiments. Male huHSC-NCG-M mice (8 weeks old; Cat. T057533) were purchased from Gempharma Technology Company (Cat. T057533). All mice were housed under standard institutional conditions, and group sizes are specified in the figure legends. No statistical methods were used to predetermine sample sizes. All animal experiments were conducted following protocols approved by the Institutional Animal Care and Use Committee (IACUC) of the University of Chicago (71829-11) and the Animal Use and Care Committee of the Second Qilu Hospital of Shandong University (KYLL-2023-412).

Cell lines and cell culture

The K562 (Cat. CCL-243), HEK 293T/17 (Cat. CRL-11268), Panc02(Cat. CRL-2553) and human iPSC (Cat. ACS-1027) cell lines were obtained from ATCC (Manassas, VA, USA). The mouse iPSC line was purchased from Alstem (Cat. iPS02M). H22 (Cat. YC-C053) cell lines was purchased from Ubigene Biosciences (Ubigene Biosciences Co., Ltd., Guangzhou, China). Colon26 cell line was purchased from Creative biogene (Shirley, NY, USA, Cat. CSC-RR0461). The Hep53.4 cell line was purchased from CLS-Cell Lines Service (CLS Cell Lines Service GmbH, Eppelheim, Germany, Cat. 400200). All cell lines were authenticated by the vendors through STR profiling (human lines) or genetic validation (murine lines). All cell lines were routinely tested for mycoplasma contamination every 4 weeks using a PCR-based detection kit (Beyotime, C0301S) and were consistently negative before experiments.

The Panc02, H22 and Hep 53.4 cell lines were transfected with a lentivirus encoding the luciferase gene (Applied Biological Materials, Richmond, BC, Canada, Cat. LVP010079) to construct Panc02-Luc, H22-Luc and Hep 53.4-Luc cell lines to allow for live animal tumor imaging. Cell culture conditions were as follows: H22 cells were maintained in RPMI 1640 medium (Gibco, USA) supplemented with 10% fetal bovine serum (FBS) and 2 mM L-glutamine. Panc02 and Colon26 cells were cultured in high-glucose Dulbecco’s Modified Eagle’s Medium (DMEM; Gibco, USA) with the same supplements. iPSCs were cultured in the iPSC medium (Oricell, Cyagen Biosciences, Cat. HUIPS-90011). iPSC cultures were dissociated using Accutase (Innovative Cell Technologies, Inc., AT104). All cell lines were incubated at 37°C in a humidified atmosphere with 5% CO₂.

Human samples

Human tumor specimens used in this study were obtained from patients undergoing surgical resection at the Second Qilu Hospital of Shandong University with approval from the institutional ethics committee (KYLL-2023-413). Written informed consent was obtained from all participants. A total of ten tumor samples were included, comprising hHCC (n = 4), colorectal cancer (n = 3), and pancreatic cancer (n = 3). Each tumor specimen was cut into small pieces and implanted subcutaneously to establish PDX models. After successful engraftment, PDX-bearing mice were assigned to the experimental group or the control group according to whether they received CAR^Ctrl-iNEV treatment or CAR^CD133-iNEV treatment. Clinical information including sex, race, ancestry, and ethnicity was not provided by the hospital and therefore could not be incorporated into downstream analyses, which is acknowledged as a limitation.

Method details

CD133 nanobody library screening

Nanobodies targeting CD133 were identified from a VHH phage display library (Shenzhen Kangtai Biotechnology) containing approximately 4 × 10⁹ unique clones, according to previously established protocols.⁴⁶ For selection, immunotubes coated with CD133 protein was used for three rounds of panning. A total of 731 individual clones were selected and analyzed by phage ELISA using a horseradish peroxidase (HRP)-conjugated anti-M13 secondary antibody. The reaction was developed with TMB substrate solution, and the absorbance at 450 nm was measured within 30 min of the addition of stop solution. Clones that tested positive in all three rounds of phage ELISA were subjected to sequencing for further characterization.

Plasmid construction and lentivirus production

Various CAR constructs were designed to contain the following domains in sequence: the CD8 signal peptide, VHH (variable heavy-chain domain of heavy-chain-only antibody), the CD8α hinge region, either the NKp46 or NKG2D extracellular domain, the 2B4 or DAP10 intracellular signaling domain, and the CD3ζ activation domain. In addition, the vector carried a puromycin resistance gene and an HA tag. These CAR sequences were codon-optimized for human expression and synthesized de novo by System Biosciences (Palo Alto, CA, USA). The CAR constructs were cloned into the pCDH-CMV-MCS-EF1α-GreenPuro lentiviral backbone vector under the control of the EF1α promoter. A bicistronic cassette was engineered to co-express the CAR and the transduction reporter CopGFP, separated by a self-cleaving T2A peptide to ensure equimolar expression of both proteins. For lentiviral vector production, HEK293T/17 cells (ATCC, Cat. CRL-11268) were seeded in 10 cm culture dishes at approximately 70% confluence one day before transfection. Transient transfection was performed using the Lipofectamine 3000 reagent (Thermo Fisher Scientific, Cat. L3000015). Briefly, plasmid DNA encoding the CAR construct (15 μg), the packaging plasmid psPAX2 (10 μg; Addgene, Cat. 12260), and the envelope plasmid pMD2.G (5 μg; Addgene, Cat. 12259) were mixed in Opti-MEM (Gibco, Cat. 31985070) and incubated with the transfection reagent to form DNA complexes. These complexes were added dropwise to the HEK293T/17 cells in serum-free medium. After 8 h, the medium was replaced with fresh complete DMEM. Viral supernatants were harvested at 48- and 72-h post-transfection, collected into sterile tubes, and centrifuged at 500×g for 10 min at 4°C to remove cell debris. The clarified viral-containing medium was concentrated using Lenti-X concentrator (Takara, USA, Cat. 631231). Viral supernatant was mixed with Lenti-X concentrator at a 3:1 ratio and incubated overnight at 4°C, followed by centrifugation at 1500×g for 45 min at 4°C. The viral pellet was resuspended in sterile PBS and aliquoted for storage at −80°C.

Generation of CAR-iPSC cells

For lentiviral transduction, iPSCs at approximately 60–70% confluency were incubated with concentrated lentiviral particles encoding the CAR constructs in the presence of 5μg/ml polybrene (Sigma-Aldrich, Cat. TR-1003) to enhance viral entry. The cells were gently rocked every 15 min during the 6-h transduction period at 37°C. Following transduction, the viral supernatant was replaced with fresh medium, and the cells were allowed to recover for 48 h. Subsequently, puromycin (Gibco, Cat. A1113803) was added at a final concentration of 2.5 μg/mL to selectively enrich for CAR-expressing cells. The selection continued for 10 days with daily medium changes containing puromycin to remove non-transduced cells. To verify surface expression of the CAR, the antibiotic-resistant iPSCs were harvested using Accutase (Gibco, Cat. A1110501) and washed twice with ice-cold staining buffer (PBS supplemented with 2% fetal bovine serum and 2 mM EDTA). Cells were incubated on ice for 30 min with fluorophore-conjugated anti-VHH primary antibodies. After staining, cells were washed thrice with staining buffer and filtered through a 40 μm cell strainer to remove aggregates. Flow cytometric analysis and fluorescence-activated cell sorting (FACS) were performed using a Sony SH800ZBP cell sorter. Live single cells were gated based on forward and side scatter profiles, and CAR-positive cells with HA tag were sorted with a purity threshold set above 95%. Following sorting, CAR-expressing iPSCs were plated onto fresh Matrigel-coated plates and cultured continuously without puromycin. Periodic validation of CAR expression by flow cytometry was performed to ensure stable transgene expression.

Differentiation of CAR-iPSCs into NK cells

For NK cell differentiation, mice iPSCs were dissociated using 0.25% trypsin-EDTA and resuspended at a density of 2 × 10³ cells/ml in Iscove’s modified Dulbecco’s medium (IMDM) supplemented with 15% FBS, 1% methylcellulose (Methocult, Cat. M3434) and 2mM-glutamine. Recombinant mouse stem cell factor (SCF, PeproTech, 50 ng/mL), interleukin-3 (IL-3, PeproTech, 30 ng/mL), and human interleukin-6 (IL-6, PeproTech, 30 ng/mL) were added to promote hematopoietic progenitor induction. The cells were seeded in 35-mm low-attachment dishes and cultured at 37°C in 5% CO₂ for 8 days to allow embryoid body (EB) formation. On day 8, the EBs were harvested, washed, and dissociated into single-cell suspensions using 0.25% trypsin-EDTA. For NK cell induction, HPCs were reseeded at a density of 2 × 10⁴ cells per well in α-MEM supplemented with 10% FBS, 1000 U/ml recombinant human IL-2 (PeproTech) and 5 ng/mL mouse IL-15 (PeproTech). The cultures were maintained at 37°C, with half-medium changes every 3–4 days. The NK cells were harvested after 10–14 days.

Differentiation of human iPSCs was induced according to a published protocol.⁴⁷ First, 5000 TrypLE-adapted iPSCs were seeded per well in round-bottom 96-well plates. The cells were cultured in APEL medium supplemented with 40 ng/mL human SCF (Gibco, Cat. PHC2111), 20 ng/mL human vascular endothelial growth factor (VEGF), and 20 ng/mL recombinant human bone morphogenetic protein 4 (BMP-4). On day 11 of differentiation, the resulting EBs were transferred directly to uncoated 24-well plates and allowed to continue to differentiate under NK cell culture conditions. The cultures were maintained for an additional 28–32 days in NK cell differentiation medium containing 10 ng/mL (IL-15, 20 ng/mL IL-7, 20 ng/mL SCF, and 10 ng/mL FLT3 ligand. Notably, IL-3 at 5 ng/mL was included only during the first week of this period to support early differentiation. When the NK cells had matured, they were harvested and cocultured with irradiated K562-mbIL-21-4-1BB artificial antigen-presenting cells (aAPCs) to stimulate expansion. The expansion culture was conducted in RPMI 1640 medium supplemented with 10% FBS and 50 U/ml human IL-2.

Tangential flow filtration (TFF) and isolation of extracellular vesicles

The cell-conditioned media were clarified and concentrated using a Minimate tangential flow filtration capsule (Pall Corporation) equipped with an Omega membrane to remove detached cells, large debris, and particulate contaminants. Filtration was performed using a MASTERFLEX Console Analog L/S Pump Drive (Model 77521-47), with flow rates adjusted to maintain the following pressures: feed pressure (P Feed): 30 psi; retentate pressure (P Retentate): 25 psi; and filtrate pressure (P Filtrate): 0 psi. Cell-conditioned media were introduced into the reservoir and concentrated at the indicated flow rates until a final volume of 10 mL was reached. After concentration, buffer exchange and dialysis filtration were performed using an amount of PBS equivalent to the original volume of medium. PBS was added to the reservoir and recirculated for 10 min to remove residual impurities and extracellular vesicles trapped in the filter membrane. The sample was then concentrated to a volume of 10 mL, and the recirculation/dialysis step was repeated three times. The final 10-mL sample underwent an additional 10-min recirculation, followed by collection from the reservoir. To ensure thorough removal of residual contaminants, PBS (10 mL) was added, and the mixture was recirculated for 10 min and collected. This PBS recirculation step was repeated three times in total. Further concentration of the final 30-mL volume was performed using a Centricon Plus-70 centrifugal filter (Ultracel-PL membrane, 100 kDa) with centrifugation at 3000 × g at 4°C in an Allegra X-15R centrifuge. The concentrate was recovered by reverse centrifugation at 1000 × g for 2 min at 4°C. To maximize throughput, 1 mL of PBS was added to the Centricon Plus-70 filter, and the sample was re-concentrated by centrifugation at 3000 × g for 2 min at 4°C, followed by reverse centrifugation at 1000 × g for 2 min.

Protein purification

For target protein expression and purification, pET-14B-CD133 (human and mouse) recombinant plasmids were transformed into BL21(DE3) cells. Bacterial cultures were grown at 37°C and 225 rpm until reaching an OD600 of 0.6, then induced with 0.2 mM IPTG at 16°C overnight. Cells were harvested by centrifugation at 8000 × g for 15 min at 4°C and resuspended in lysis buffer (300 mM NaCl, 50 mM NaH₂PO₄, 10 mM imidazole, pH 8.0, 1 mM PMSF). High-pressure homogenization was performed at low temperature for three cycles, followed by centrifugation at 12,000 × g for 45 min to remove cell debris. The supernatant was loaded onto a gravity column containing 1 mL Ni-NTA agarose resin (Qiagen, Germany). The bound proteins were sequentially washed with 50 mL Wash Buffer I (20 mM imidazole, pH 8.0) and 50 mL Wash Buffer II (40 mM imidazole, pH 8.0), then eluted with 25 mL Elution Buffer (250 mM imidazole, pH 8.0). The eluate was further purified via Superdex-150 gel chromatography using an AKTA Pure System (GE Healthcare Life Sciences, USA) in 1× PBS. The purified proteins were analyzed by SDS-PAGE, rapidly frozen in liquid nitrogen, and stored at −80°C. Nanobody purification followed the same Ni-NTA affinity chromatography protocol. The purified nanobodies were characterized by western blot using 6×His tag, HA tag, and anti-VHH antibodies.

Enzyme-linked immunosorbent assay (ELISA)

Mouse liver tissues were harvested and homogenized. Levels of inflammatory cytokines and chemokines in liver homogenates were quantified using commercially available ELISA kits according to the manufacturers’ protocols. The following kits from JONLNBIO were employed: Mouse Ccl3 ELISA Kit (Cat. JL20414), Mouse Ccl5 (Cat. JL17335), Mouse Cxcl1 (Cat. JL20150), Mouse Cxcl5 (Cat. JL28925), Mouse IL-1α (Cat. JL13044), Mouse IL-1β (Cat. JL18442), Mouse IL-6 (Cat. JL20268), Mouse TNF-α (Cat. JL10484), and Mouse Csf3 ELISA Kit (Cat. JL10697). Additionally, Mouse Cxcl2 levels were measured using the ELISA kit from Abnova (Cat. KA1505). All assays were performed following the respective manufacturer’s instructions, and absorbance was measured to determine the concentration of each inflammatory mediator in the liver samples. Tissue samples for ELISA analysis were prepared as homogenates using the following procedure: target tissues were placed on ice and rinsed with pre-cooled phosphate-buffered saline (PBS, 0.01 M, pH 7.4) to remove residual blood, then weighed and reserved. Tissue homogenization was performed on ice using a lysis buffer composed of 50 mM Tris, 0.9% NaCl, and 0.1% SDS at pH 7.3, with the volume of lysis buffer added proportional to tissue weight, typically 9 mL per 1 g of tissue fragments, supplemented with 1 mM phenylmethylsulfonyl fluoride (PMSF) as a protease inhibitor. Homogenates were further processed by ultrasonic disruption under ice-cooling to prevent overheating. The prepared homogenates were then centrifuged at 5,000 × g for 5 min at 4°C, and the supernatants were collected for ELISA. Prior to analysis, total protein concentrations of the tissue homogenates were quantified using the bicinchoninic acid (BCA) assay, and samples were adjusted to a final protein concentration of 2 mg/mL for ELISA detection.

Surface plasmon resonance (SPR) analysis

SPR experiments were performed on a Biacore S200 instrument to characterize the binding kinetics of nanobodies (Nbs) with human and mouse CD133. Prior to immobilization, the S-CM5 sensor chip (Cytiva) was preconditioned by sequential injections of 10 mM glycine-HCl (pH 1.5) and running buffer to stabilize the surface. Human and mouse CD133 proteins were immobilized onto separate flow cells of the chip via standard amine coupling chemistry. Briefly, the chip surface was activated by injecting a freshly prepared mixture of 0.05 M N-hydroxysuccinimide (NHS) and 0.2M 1-ethyl-3-(3-dimethylaminopropyl) carbodiimide (EDC) for 7 min at a flow rate of 10 μL/min. CD133 proteins, diluted to 30 μg/mL in 10 mM sodium acetate buffer (pH 5.0), were injected over the activated surface until immobilization levels reached approximately 2000 response units (RU). A reference flow cell was subjected to the same activation and blocking procedure without protein coupling to control for nonspecific binding and bulk refractive index changes. All binding assays were conducted at 25°C using a running buffer composed of 10 mM HEPES (pH 7.4), 150 mM NaCl, 3 mM EDTA, and 0.05% (v/v) Tween 20, which was filtered and degassed prior to use. Serial 2-fold dilutions of each nanobody were injected over the sensor surface at a flow rate of 30 μL/min for an association phase of 120 s, followed by a 180-s dissociation phase with running buffer flow. Regeneration of the chip surface between cycles was achieved by injecting 10 mM glycine-HCl (pH 2.0) for 120 s, ensuring complete dissociation of bound Nbs while preserving the integrity of immobilized CD133. Sensorgrams were double-referenced by subtracting signals from the reference cell and blank buffer injections to correct for nonspecific binding and instrumental drift. Binding kinetic parameters, including association rate constant, dissociation rate constant, and equilibrium dissociation constant, were derived by fitting the data to a 1:1 Langmuir binding model using Biacore S200 evaluation software. The quality of fit was assessed based on residuals and chi-square values.

Luciferase-based cytotoxicity assay

Cytotoxic activity was evaluated using a luciferase-based assay. Panc02, Colon26, and H22 tumor cell lines stably expressing firefly luciferase were used as target cells. Target cells were seeded in 96-well plates and co-incubated with CAR-iNK cells carrying different CAR constructs at the indicated effector-to-target (E/T) ratios. After 2 h of incubation at 37°C, cell lysis was quantified by measuring residual luciferase activity using a standard luciferase assay system (Promega) according to the manufacturer’s instructions. Luminescence was detected with a microplate reader, and cytotoxicity was calculated as the percentage reduction in luminescence relative to untreated control wells.

For Figure S1C, target cells were incubated with CAR-iNK for 4 h prior to measurement.

For Figure S3D, target cells were incubated with CAR^CD133-iNEV or CAR^Ctrl-iNEV for 2 h prior to measurement.

In vivo liver orthotopic cancer model

Luciferase-expressing Hep53.4 tumor cells (Hep53.4-Luc) were harvested at 70–80% confluence using 0.05% trypsin-EDTA, washed twice with sterile PBS, and resuspended at a concentration of 2 × 10⁷ cells/200μL in cold PBS immediately prior to implantation. 6-8-week-old female C57BL/6 mice (weight 18–22 g) were anesthetized using isoflurane (induction at 3–4%, maintenance at 1.5–2%). Following aseptic preparation of the abdominal area, a small subcostal incision was made to expose the left lateral liver lobe. A suspension containing 1 × 10⁶ Hep53.4-Luc cells in 10 μL PBS was gently mixed with 20 μL of Matrigel at a 1:2 (v/v) ratio immediately prior to injection. The 30 μL mixture was carefully injected under the liver capsule to establish an orthotopic liver tumor. Tumor burden was longitudinally monitored by the IVIS Spectrum system (PerkinElmer). At each imaging time point, mice received an intraperitoneal injection of D-luciferin (3 mg in 100 μL sterile PBS). Four minutes post-injection, animals were anesthetized with isoflurane and images were acquired. Quantification of signals was performed using Living Image software version 4.5, defining regions of interest (ROIs) over the liver area and calculating total photon flux (photons/sec/cm²) as a measure of tumor burden. Mice were randomized into different groups. Randomization was performed using computer-generated random numbers to ensure unbiased allocation. Treatment regimens were initiated according to the study design, as detailed in the flowchart, with investigators blinded to group assignments to minimize bias. Throughout the study, mice were monitored daily for signs of distress or adverse effects, and humane endpoints were strictly observed.

In vivo lung metastasis model

For generation of the lung metastasis model, Panc02-Luc cells were prepared as described above and resuspended in sterile PBS at 1 × 10⁷ cells/ml. Six-to eight-week-old female C57BL/6 mice were restrained and warmed to dilate the tail vein, then injected with 1 × 10⁶ Panc02-Luc cells in 100 μL PBS via the lateral tail vein. Successful intravenous administration was confirmed by the lack of swelling or leakage at the injection site. Mice were closely monitored for respiratory distress and general health throughout the study. Tumor growth and metastasis were quantified by BLI following intraperitoneal injection of D-luciferin. Baseline BLI signals were used to randomize mice into treatment groups, consistent with the orthotopic model protocol.

In vivo subcutaneous model

Tumor cell lines Hep53.4, Panc02, H22, and Colon26 were cultured and harvested using 0.05% trypsin-EDTA at 70–80% confluence. Cells were washed twice with sterile PBS and resuspended at 1 × 10⁷ cells/ml. For subcutaneous tumor implantation, 1 × 10⁶ Hep53.4 and Panc02 cells were injected into the right flank of syngeneic female C57BL/6 mice, whereas 1 × 10⁶ H22 and Colon26 cells were injected subcutaneously into the right flank of syngeneic female BALB/c mice, respectively. Tumor growth was monitored every 2 days by caliper measurements. Tumor volume was calculated using the formula: (length × width × width)/2. Mice were randomly grouped when tumor volume reached 150 mm³. Mice were euthanized when tumor volume reached 1500 mm³ or when predefined humane endpoints were observed.

In vivo patient-derived xenograft (PDX) model

The clinical patient specimens were obtained with authorization and approval from the Research Ethics Committee of the Second Qilu Hospital of Shandong University (KYLL-2023-413). The informed consent forms were obtained from patients and their families. Fresh tumor tissues were transported on ice in sterile RPMI-1640 supplemented with 1% antibiotic-antimycotic solution. Within 2 h of resection, tumors were minced into approximately 3 × 3 × 3 mm³ fragments using sterile scalpels under aseptic conditions. Fragments were implanted subcutaneously into the right rear flank of huHSC-NCG-M mice under isoflurane anesthesia. Mice received analgesics post-operatively as per institutional protocols. Tumor growth was monitored via caliper measurements, calculating tumor volume using the formula: (length × width × width)/2. Once tumors reached approximately 150 mm³, mice were randomized into different groups. Antitumor efficacy and survival were assessed. Mice were euthanized when tumors exceeded 1500 mm³ or if they exhibited signs of distress, following humane endpoints.

In vivo flow cytometry

Mice were administered CAR-iNEVs at 1 × 10¹² particles per animal via the tail vein. 24h post-injection, animals were euthanized under isoflurane anesthesia and perfused via the left ventricle with 10 mL ice-cold PBS to clear circulating vesicles. Organs (heart, liver, spleen, lung, kidney) as well as harvested tumors were excised, placed in ice-cold HBSS and immediately processed to single-cell suspensions. Liver tissue was minced and digested in HBSS containing Collagenase IV (0.5 mg/mL) and DNase I (50 U/mL) for 30 min at 37°C with gentle rotation, then filtered through a 70 μm cell strainer, followed by RBC lysis (ACK buffer, 2–3 min at RT) and washed twice with FACS buffer (PBS with 2% FBS and 2 mM EDTA). Lung tissue was digested similarly using Collagenase D (1.0 mg/mL) and DNase I (50 U/mL) for 30 min at 37°C, passed through a 70 μm strainer, RBC-lysed and washed in FACS buffer. Spleens were gently dissociated by mechanical disruption through a 70 μm strainer, RBC-lysed, then washed in FACS buffer. Kidney and heart tissues were minced and digested in HBSS containing Collagenase II (1.0 mg/mL) plus DNase I (50 U/mL) for 30 min at 37°C, filtered through a 70 μm strainer and washed in FACS buffer. Tumor tissue was excised, minced and digested in HBSS containing Collagenase IV (0.5 mg/mL) and DNase I (50 U/mL) for 30 min at 37°C, filtered, RBC-lysed and washed. To identify tumor cells vs. tumor-associated fibroblasts (TAFs) further staining for lineage markers (EpCAM for tumor cells, PDGFRα/αSMA for fibroblasts) was performed. After dissociation, viable cells were assessed by trypan-blue exclusion and viable cells were counted. Approximately 1×10⁶ viable cells per sample were transferred into 100 μL FACS buffer for staining. Surface staining was performed by adding fluorochrome-conjugated surface markers for the cell subsets (CD45, EpCAM, PDGFRα, αSMA) at pre-optimized dilutions and incubated for 30 min at 4°C in the dark.

Flow cytometry for cell uptake experiments, cells were washed twice with FACS buffer, then fixed with 100 μL of 4% paraformaldehyde in PBS for 15 min at RT in the dark. After fixation, cells were washed once in FACS buffer and then permeabilized by adding 200 μL of permeabilization buffer (PBS containing 0.1% saponin and 2% FBS) and incubated for 10 min at RT. Without additional wash, intracellular staining was performed by adding fluorochrome-conjugated anti-HA (1:100) and incubated for 30 min at RT in the dark in permeabilization buffer. Cells were then washed twice in FACS buffer, resuspended in 100 μL FACS buffer. Data were acquired on a flow cytometer with identical voltage settings across all organs and samples.

Labeling CAR-iNEV with ⁸⁹Zr-DFO

To enable surface radiolabeling of CAR^Ctrl-iNEV or CAR^CD133-iNEV with ⁸⁹Zr, a two-step chelation approach was used. Initially, exosomes were functionalized with the chelator desferrioxamine (DFO). Briefly, p-NCS-Bn-DFO (Future Chem, Seoul, Korea), dissolved in DMSO at a concentration of 5–20 mg/mL, was added to the CAR^Ctrl-iNEV or CAR^CD133-iNEV suspension at a DFO-to-exosomal protein weight ratio of 1:10. After incubation, exosomes labeled with DFO were purified using a dextran desalting column (molecular weight cut-off (MWCO) 5000, Pierce, WI, USA), with 10 mM PBS as the eluent. Subsequently, the DFO-modified CAR^Ctrl-iNEV or CAR^CD133-iNEV were labeled with ⁸⁹Zr. The ⁸⁹Zr⁴⁺, supplied in 1.0 M oxalic acid, was first neutralized to a pH between 6.8 and 7.5 using 1.0 M Na₂CO₃. The neutralized ⁸⁹Zr solution was then combined with the DFO-exosome conjugates, and the mixture was incubated at room temperature for 1 h under gentle shaking (550 rpm). Labeled exosomes were subsequently purified by size exclusion chromatography using a dextran desalting column (MWCO 50,000), and 10 mM PBS (pH 7.4, without Ca²⁺ and Mg²⁺) was used as the elution buffer to ensure removal of unbound radionuclide. The final radiolabeled exosome preparation was stored at 4°C until use.

Ex vivo biodistribution

A total of 40 male BALB/c mice were used for ex vivo biodistribution analysis. Mice were anesthetized with isoflurane (2%, 1 L/min) and intravenously injected with ⁸⁹Zr-CAR^Ctrl-iNEV or ⁸⁹Zr-CAR^CD133-iNEV at a fixed dose of 1 × 10¹² particles per animal via the tail vein. At predetermined time points (1, 4, 8, 12, and 24 h) after a single administration, four mice per group were euthanized under deep anesthesia. Major organs (blood, colon, stomach, kidneys, lungs, spleen, liver, heart and tumor) were harvested, rinsed with PBS, blotted dry, and weighed. Radioactivity in each sample was measured using a γ-counter system (1480 Wizard, PerkinElmer, MA, USA) with an energy window centered on the 909 keV peak of ⁸⁹Zr. A standard containing a known fraction of the injected dose was prepared in parallel and counted under identical conditions to enable activity calibration. All measurements were corrected for background and physical decay to the time of injection. The percentage of injected dose per gram of tissue (%ID/g) was calculated as: %ID/g = (organ activity/injected activity) × 100/organ weight.

In vivo treatments

To evaluate the therapeutic efficacy of CAR^CD133-iNK cells and CAR^CD133-iNEV, subcutaneous tumor models were established using H22 cells. When tumors reached approximately 150 mm³, tumor-bearing mice (BALB/c, n = 5 per group) were randomized into seven treatment groups: (1) PBS control (200 μL), (2) high-dose CAR^CD133-iNK (1 × 10⁷ cells per injection), (3) low-dose CAR^CD133-iNK (1 × 10⁶ cells per injection), (4) high-dose CAR^Ctrl-iNEV (1 × 10¹² particles per injection), (5) low-dose CAR^Ctrl-iNEV (1 × 10¹¹ particles per injection), (6) high-dose CAR^CD133-iNEV (1 × 10¹² particles per injection), and (7) low-dose CAR^CD133-iNEV (1 × 10¹¹ particles per injection). All treatments were administered via tail vein injection in a total volume of 200 μL every other day. Tumor volumes were measured every other day using calipers and calculated.

To assess the antitumor activity of CAR^CD133-iNEV across multiple tumor models, subcutaneous models using H22, Panc02, and Colon26 cells were employed. Tumor-bearing mice (n = 5 per group) were randomized into three groups and treated intravenously with: (1) PBS Ctrl (200 μL), (2) CAR^Ctrl-iNEV (1 × 10¹² particles per injection), or (3) CAR^CD133-iNEV (1 × 10¹² particles per injection). Treatments were administered three times per week, each injection delivered in 200 μL volume. Tumor sizes were measured every other day as described above.

For combination immunotherapy studies in PDX models, mice bearing established tumors (∼150 mm³) were randomized as follows: In the experiment shown in Figure 6E, four groups (n = 5 per group) received (1) PBS (200 μL), (2) anti-CD47 antibodies (10 mg/kg), (3) CAR^CD133-iNEV (1 × 10¹² particles per injection), or (4) combination therapy with CAR^CD133-iNEV (1 × 10¹² particles per injection) plus anti-CD47 antibodies (10 mg/kg). In the experiment depicted in Figure 6F, PDX-bearing mice were randomized into two groups (n = 5 per group) and treated with PBS (200 μL) or CAR^CD133-iNEV (1 × 10¹² particles per injection) combined with anti-CD47 antibodies (10 mg/kg). Mice were euthanized upon tumor volumes exceeding 1500 mm³ or following humane endpoints. Survival was monitored for up to 60 days. Similarly, CDX-bearing mice were randomized into four groups (n = 5 per group) receiving (1) PBS (200 μL), (2) either anti-PD-1 antibodies (12.5 mg/kg) or anti-CD47 antibodies (10 mg/kg), (3) CAR^CD133-iNEV (1 × 10¹² particles per injection), or (4) CAR^CD133-iNEV combined with either anti-PD-1 (12.5 mg/kg) or anti-CD47 antibodies (10 mg/kg). Treatments were administered via tail vein injection once weekly for three doses (200 μL per injection). Mice were euthanized when tumors reached 1500 mm³ or at experimental endpoint (32 days).

For metastasis models, mice bearing Panc02-Luc lung or peritoneal metastases were randomized into four groups (n = 5 per group) and treated intravenously with (1) PBS (200 μL), (2) anti-CD47 antibodies (10 mg/kg), (3) CAR^CD133-iNEV (1 × 10¹² particles per injection), or (4) combination therapy with CAR^CD133-iNEV (1 × 10¹² particles per injection) plus anti-CD47 antibodies (10 mg/kg). All treatments were administered three times weekly via tail vein injection (200 μL per injection). Tumor burden was monitored every five days using BLI. Mice were monitored for survival and euthanized upon meeting humane endpoints or natural death. The experimental endpoint was set at 40 days post-treatment initiation.

To assess the longevity of antitumor immune memory, tumor rechallenge experiments were performed approximately 120 days after the final CAR^CD133-iNEV + anti-CD47 antibody treatment. This interval was considered adequate for clearance of residual therapeutic agents and resolution of acute immune responses. Five mice, randomly selected from the seven mice previously cured by combination therapy (three from the lung metastasis model and two from the peritoneal metastasis model), were rechallenged subcutaneously with panc02-Luc cells (as previously described for the establishment of the lung and peritoneal metastasis models). Age-matched mice were used as controls. Tumor growth and survival were monitored to evaluate the persistence of antitumor immunity in the long-term survivors.

To investigate the roles of macrophages and neutrophils in tumor progression, Colon26 tumor-bearing mice were randomly assigned to four groups. As shown in Figure 5H, mice were randomized into four groups (n = 5 per group) receiving: (1) PBS (200 μL), (2) CAR^CD133-iNEV (1 × 10¹² particles per injection), (3) CAR^CD133-iNEV plus clodronate (200 μg per injection), or (4) CAR^CD133-iNEV plus anti-Ly6G antibodies (0.5 mg per injection; clone 1A8, Bio X Cell). In Figure 5I, Colon26 tumor-bearing wild-type (WT) and macrophage-specific Nos2 knockout (Nos2-KO) mice (n = 5 per group) were treated with (1) WT + PBS (200 μL), (2) WT + CAR^CD133-iNEV (1 × 10¹² particles per injection), (3) Nos2-KO + PBS (200 μL), or (4) Nos2-KO + CAR^CD133-iNEV (1 × 10¹² particles per injection). All treatments were administered via tail vein injection three times per week (200 μL per injection). Mice were euthanized when tumor volumes reached 1500 mm³ or upon meeting humane endpoints.

Hematoxylin and eosin (H&E) staining

For histopathological evaluation, major organs (heart, liver, spleen, lung, kidney, and colon) were collected from tumor-bearing mice following treatment. Tissues were rinsed briefly in cold PBS to remove residual blood and fixed in 4% paraformaldehyde (PFA) at 4°C for 24 h. Fixed tissues were dehydrated through graded ethanol (70%, 80%, 90%, 95%, and 100%), cleared in xylene, and embedded in paraffin. Paraffin blocks were sectioned at 4 μm thickness using a rotary microtome and mounted on positively charged glass slides. Sections were deparaffinized in xylene (2 × 10 min), rehydrated through descending ethanol concentrations to distilled water, and stained with Mayer’s hematoxylin for 5 min. Slides were rinsed in running tap water for 10 min, differentiated in 1% acid alcohol for 30 s, blued in 0.1% ammonia water for 1 min, counterstained with eosin for 2 min, and dehydrated through ascending ethanol series. After clearing in xylene, slides were mounted with neutral resin. Images were captured under bright-field illumination.

IHC for HA-tagged CAR-iNEV detection in mouse tissues

To assess the biodistribution of CAR-iNEV in vivo, IHC staining was performed on heart, liver, spleen, lung, kidney, and tumor tissues harvested from treated mice. Tissues were fixed in 4% PFA overnight at 4°C, dehydrated, embedded in paraffin, and sectioned at 4 μm. After deparaffinization and rehydration, endogenous peroxidase activity was quenched with 3% H₂O₂ for 10 min at RT. Antigen retrieval was performed in 10 mM sodium citrate buffer (pH 6.0) at 95°C for 20 min. Non-specific binding was blocked with 5% normal donkey serum in PBS for 30 min at RT. Sections were incubated overnight at 4°C with anti-HA antibody (Abmart, Cat#M20003; 1:200 dilution) diluted in blocking buffer. After washing with PBS (3 × 5 min), slides were incubated with HRP-conjugated secondary antibody for 30 min at RT. Color development was achieved using 3,3′-diaminobenzidine (DAB) substrate, and sections were counterstained with hematoxylin for 30 s, dehydrated, cleared, and mounted with neutral resin.

IHC staining for CD133 detection in human tumor samples

Human hepatocellular carcinoma (n = 4), colorectal cancer (n = 3), and pancreatic cancer (n = 3) specimens were obtained from patients undergoing surgical resection at the Second Qilu Hospital of Shandong University. Tissues were fixed in 4% PFA for 24 h, embedded in paraffin, and sectioned at 4 μm. After deparaffinization and rehydration as described above, endogenous peroxidase activity was blocked using 3% H₂O₂ for 10 min, and antigen retrieval was conducted in 10 mM sodium citrate buffer (pH 6.0) at 95°C for 20 min. Non-specific binding was blocked with 5% normal donkey serum for 30 min at RT. Sections were incubated overnight at 4°C with anti-CD133 antibody (CST, Cat#64326; 1:200 dilution) diluted in PBS containing 1% BSA. Slides were washed and incubated with HRP-linked anti-rabbit secondary antibody for 30 min at RT. Staining was developed using DAB substrate, counterstained with hematoxylin, dehydrated, cleared, and mounted.

Quantification and statistical analysis

Unless otherwise specified, data are presented as Mean ± Standard Deviation (SD). Statistical analyses were conducted as follows: Unpaired two-tailed Student’s t test was used for two-group comparisons, while one-way ANOVA was applied for multi-group comparisons, including tumor weights and toxicological parameters. two-way ANOVA was employed for time-course analyses of in vitro cell viability and tumor growth curves. Survival analysis was conducted using the log rank test. A p-value <0.05 was considered statistically significant, with significance levels indicated as follows: ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001. Graphpad Prism v10 (GraphPad Software, http://www.graphpad.com) and SPSS v20 (statistical package for Windows SPSS Inc., https://www.ibm.com/it-it/products/spss-statistics) were used for data management, statistical analysis and graph generation.

Published: February 5, 2026