Engineering bispecific exosome activators of T cells to target immune checkpoint inhibitor-resistant metastatic melanoma

Abstract

Cancer immunotherapy with immune checkpoint inhibitors (ICIs) is often limited by an immunosuppressive tumor microenvironment (TME). Simultaneous targeting of the TME and immune checkpoints is a promising approach to address this limitation. Here we develop an inhalable exosome system that enables co-display of two inhibitory ligands and apply it to treat lung metastases of ICI-resistant melanoma. As immune exclusion in this context is often mediated by Wnt/β-catenin signaling, we harnessed the Alix sorting domain for tandem display of PD-1 and FZD8 to block PD-L1 and Wnt7b, which is overexpressed in ICI-resistant melanoma. This technology, called bispecific exosome activator of T cells (BEAT), enables uniform 1:1 co-display of two proteins on the exosome surface. We show that BEAT concurrently recruits and activates CD8⁺ T cells to reprogram the TME, yielding robust antitumor activity in ICI-resistant melanoma mouse models. Inhaled BEAT outperforms linked dual antibody targeting PD-L1 and Wnt7b in vivo. This approach to tandem protein display may be applicable to diverse ICI-resistant cancers.

Immune checkpoint inhibitors (ICIs) have transformed melanoma treatment by extending survival, but they remain ineffective in a substantial subset of individuals, with nonresponse rates approaching 40% (refs. 1-5), and their systemic administration is associated with immune-related adverse effects^6-9. Resistance commonly arises from multifactorial immunosuppression within the tumor microenvironment (TME) that is unlikely to be reversed by inhibition of a single pathway. These constraints motivate coordinated delivery of multiple immunomodulators (for example, pairing a checkpoint antagonist with an immunostimulatory agent) to elicit synergistic antitumor activity. However, dual-protein regimens based on monoclonal or bispecific antibodies face practical barriers, including immunogenicity, inflexible stoichiometric control and heightened off-tumor toxicities from systemic exposure¹⁰. To mitigate toxicity, local ICI administration using microneedle patches^11,12, intratumoral injection¹³ or nanoparticle-enabled inhalation has been explored¹⁴. Although promising, these strategies are constrained by limited access to deep-seated or disseminated lesions, undefined orientation of delivered proteins, rapid clearance by phagocytes and difficulty achieving synchronized multitarget engagement. Overcoming these challenges requires innovative delivery platforms capable of site-specific, ratio-defined co-presentation of multiple proteins while minimizing systemic exposure.

A major barrier to efficacy is the immunosuppressive TME, which prevents sufficient infiltration of CD8⁺ T cells into the tumor, a phenomenon closely linked to activation of the Wnt/β-catenin signaling pathway^15,16. Despite the rarity of mutations in this pathway, alterations in Wnt ligands and their receptors appear to play a critical role in facilitating immune evasion. Targeting paracrine Wnt/β-catenin signaling has therefore emerged as a promising strategy to overcome ICI resistance. Efforts to overcome immune resistance driven by Wnt/β-catenin signaling have shown that combining Wnt pathway inhibitors with PD-1 blockade can improve antitumor responses^17,18. However, most Wnt-targeted agents, such as the tankyrase inhibitor G007-LK or the Frizzled-targeting antibody OMP-18R5, lack tissue specificity, and their systemic delivery raises safety concerns owing to the pathway’s role in intestinal homeostasis, bone remodeling and hematopoiesis^19-21. These challenges underscore the need for more localized and tolerable strategies to modulate the Wnt/β-catenin pathway and enhance ICI efficacy.

Melanoma is prone to metastasizing to the lungs, the most common nonskin metastasis site^5,22. We have developed exosome-mediated inhalation therapy for pulmonary diseases, including coronavirus disease 2019 and lung cancer^23-27. Advantages of exosomes include their natural origin, biocompatibility and safety profile^28,29. However, traditional chemical methods of conjugating targeting proteins to exosomes often disrupt their native proteome and lead to unstable protein configurations³⁰. Additionally, although current strategies for engineering exosomal protein display have increased enrichment efficiency³¹, substantial scope for further optimization remains. In particular, existing approaches have yet to rectify the heterogeneity that arises when co-displaying two proteins or to confer demonstrable synergistic effects between them. This highlights a need for new techniques that preserve exosome integrity while enhancing therapeutic payload delivery and simultaneously enabling the coordinated display of multiple proteins.

In this study, we engineered a bispecific exosome system that co-displays two inhibitory ligands for lung metastases of ICI-resistant melanoma therapy. We use Alix as a sorting domain to enrich the tandem display of FZD8 and PD-1 on the exosome surface. Notably, the tandem expression of PD-1 and FZD8 on the same vesicle ensures efficient internalization and degradation of both PD-L1 and Wnt ligands, preventing their repeated engagement and maximizing therapeutic benefit. Exosomal FZD8 acts as a decoy for Wnt7b, a ligand overexpressed in ICI nonresponders, thereby intercepting tumor-derived immune exclusion signals and remodeling the ICI-resistant TME. Concurrently, exosomal PD-1 binds PD-L1 on tumor cells, restoring CD8⁺ T cell activity and promoting tumor eradication. We name this approach bispecific exosome activators of T cells (BEAT). BEAT can be efficiently delivered via the inhalation route, offering a noninvasive strategy for pulmonary targeting. BEAT shows robust antitumor activity in four different ICI-resistant melanoma mouse models, including a humanized patient-derived xenograft (PDX) mouse model. We find that compared to systemically delivered linked dual antibody targeting PD-L1 and Wnt7b, inhaled BEAT exhibits enhanced pulmonary retention and superior antitumor efficacy. In addition, BEAT demonstrates tumor-suppressive capacity in an ICI-resistant liver metastasis mouse model. Our bispecific exosome-based therapeutic strategy leverages a highly efficient Alix-guided protein display technology for exosome surface engineering and offers potential for applications in diverse ICI-resistant cancers.

Results

Systemic comparison of engineering strategies for exosome surface display of therapeutic proteins

To establish an effective exosome surface display strategy for therapeutic proteins, we devised a range of display strategies using exosomal proteins, which are not only widely recognized as exosome markers but also notably enriched within exosomes^32-35 (Supplementary Fig. 1a,b). As a model for the display of therapeutic proteins on exosomes, we engineered a fusion of the signaling-incompetent PD-1, incorporating its ligand-binding and intramembrane domains with exosome sorting domains. This design aims to decoy PD-L1 and concurrently enrich PD-1 presentation on exosomes. An array of genetic constructs was designed using exosomal sorting domains derived from exosome markers CD9, CD81, Alix and syntenin (Fig. 1a,b). The C-terminal regions of CD9 and CD81 were used due to the demonstrated sorting potential of the tetraspanin protein family and their binding affinity to syntenin^31,36, which plays a pivotal role in exosome sorting and biogenesis. Similarly, the functional domain of Alix, known for its interaction with syntenin, was used in this study.

Fig. 1 ∣. Systematic screening of multiple endogenous exosome display strategies.

a, Schematic showing the design of the endogenous exosome display platform; aa, amino acids. b, Schematic showing the design of the construct plasmids. c, Immunoblot analysis of Flag-tagged exosomes and parental cells. Samples were loaded based on equal protein mass. Actin was used as the loading control for cells, whereas TSG101 served as the loading control for exosomes. Duplicate assays were performed with similar results; Ctrl, control. d, PD-1 expression in exosomes (n = 3 independent experiments); OD₄₅₀, optical density at 450 nm. e, Cryo-electron microscopy imaging of WT-Exo and PD-1–Alix-Exo. The black arrow indicates the PD-1 corona; scale bar, 50 nm. Triplicate assays were performed with similar results. f, Immunoblotting of the pulldown assay. Duplicate assays were performed with similar results; IB, immunoblot. g, Flow cytometry showing the percentage of recombinant mouse PD-1-Fc⁺ tumor cells after exosome treatment (n = 3 independent experiments). See Supplementary Fig. 17 for gating strategies. h, Schematic illustrating the assessment of antitumor efficacy following treatment. Intratumoral injections of exosomes and anti-PD-L1 (at a molecular amount equivalent to PD-1 in PD-1–Alix-Exo) were initiated 1 week after tumor inoculation and administered every 3 days for a total of 19 days; s.c., subcutaneous. i, Tumor weights 19 days after treatment. j, Tumor morphologies 19 days after treatment; n = 5 independent mice for each group in h–j. P values were determined by one-way analysis of variance (ANOVA) with a post hoc Tukey’s multiple comparisons test (d and g) and two-tailed Student’s t-test (i) using GraphPad PRISM software. Exact P values are indicated. Results are presented as means ± s.d. The schematics in a and b and mouse cartoon were created with BioRender.com.

Human embryonic kidney (HEK293T) cells were infected with lentivirus generated from the plasmids encoding the display constructs. Engineered exosomes were subsequently concentrated and purified from the supernatants of stable cell cultures using tangential flow filtration and ultracentrifugation. Exosomal proteins were characterized by western blotting (Extended Data Fig. 1a). Furthermore, the exosomes demonstrated similar distributions at the nanoscale (Extended Data Fig. 1b), and their morphology was validated through transmission electron microscopy (Extended Data Fig. 1c). The potency of the exosomal PD-1 displayed was assessed by western blotting of Flag tag fused to the C terminus of the constructs, indicative of functional PD-1 expression. Each construct displays a clear and even band in the parental cells, whereas in exosomes, the constructs included the C-terminal sorting domain derived from Alix, a protein that is implicated in the sorting of protein cargo into exosomes^37,38, that exhibits the most potent ability to incorporate PD-1 into exosomes (Fig. 1c,d). In an enzyme-linked immunosorbent assay (ELISA), PBS was consistently used as the solvent, ensuring the preservation of exosome membrane integrity. PD-1 was observed to be not only enriched within the exosomes but also expressed on their outer membranes. These findings were further validated through surface plasmon resonance experiments (Extended Data Fig. 1d) and cryo-electron microscopy (Fig. 1e), confirming the membrane-associated, corona-like expression of PD-1. Single-vesicle flow cytometry analysis demonstrated that over 90% of CD63⁺ exosomes derived from PD-1–Alix fusion-expressing cells (PD-1–Alix-Exo) were PD-1⁺ (Extended Data Fig. 2a-c). Additionally, dSTORM imaging revealed a relatively uniform distribution of PD-1 on exosomes (Extended Data Fig. 2d).

Next, we used a pulldown western blotting assay to validate that the engineered PD-1-displaying exosomes could functionally bind to PD-L1. The results showed that the bead–PD-L1 complex could bind the engineered exosomes, and PD-1–Alix-Exo showed the most potency (Fig. 1f). We next tested whether the interaction between exosomal PD-1 and PD-L1 could prevent the binding of PD-L1 to nonexosomal PD-1. Compared to wild-type exosomes (WT-Exo), PD-1–Alix-Exo significantly attenuated the binding of recombinant mouse PD-1-Fc to the cell surface (Fig. 1g), and this inhibition occurred in a dose-dependent manner (Extended Data Fig. 1e).

To examine the potential effect of exosomal PD-1 in cancer, a subcutaneous Braf^V600E/Pten^−/− (BP) melanoma model was used (Fig. 1h). PD-1–Alix-Exo exhibited the most substantial tumor-suppressing effect (Fig. 1i,j and Supplementary Fig. 2a-c). To access the antitumor function of exosomal PD-1, PD-L1⁺ tumor cells were co-incubated with PD-1–Alix-Exo. PD-1–Alix-Exo bound to PD-L1 and transported it into cells, inhibiting the functional activity of PD-L1 on the cell surface and eventually leading to its degradation (Supplementary Fig. 3). Thus, Alix can serve as an effective sorting domain, efficiently enriching functional PD-1 on the outer surface of exosomes.

PD-1–Alix-Exo inhalation inhibits the growth of lung metastatic melanoma

The distribution of Cy5-labeled PD-1–Alix-Exo after inhalation was evaluated to assess the efficiency of exosome delivery. Ex vivo imaging demonstrated a marked preference for lung enrichment of PD-1–Alix-Exo, observed at 24 and 48 h after inhalation, with a concurrent reduction in distribution to off-target organs, notably the liver and kidneys (Fig. 2a,b). To compare the efficacy of PD-1–Alix-Exo, anti-PD-L1 and PD-1 protein under inhalation administration, we used a lung-colonized melanoma model. Flow cytometry analysis demonstrated that following inhalation, PD-1–Alix-Exo exhibited the highest binding efficiency to tumor cells, surpassing anti-PD-L1, PD-1 protein and even systemically administered anti-PD-L1 (Fig. 2c and Extended Data Fig. 3a). Ex vivo fluorescence imaging demonstrated that inhaled exosomes showed greater accumulation in the lungs and lower systemic distribution than clinically used systemically delivered antibodies (Extended Data Fig. 3b,c). Inhaled PD-1–Alix-Exo also effectively promoted the degradation of PD-L1 on tumor cells (Supplementary Fig. 4a,b). Furthermore, under inhalation conditions, ex vivo imaging confirmed that PD-1–Alix-Exo displayed markedly enhanced lung retention relative to both anti-PD-L1 and PD-1 protein (Fig. 2d).

Fig. 2 ∣. PD-1–Alix-Exo inhalation therapy reduces the growth of lung-colonizing tumors.

a,b, Representative ex vivo images (a) and quantitative analysis (b) of major mouse organs that received fluorescence-labeled PD-1–Alix-Exo before and 24 and 48 h after inhalation; n = 3 independent mice for each group in a and b. c, Flow cytometry showing the percentage of PD-1–Alix-Exo⁺, anti-PD-L1⁺ and PD-1 recombinant protein⁺ tumor cells at 0, 0.5, 4 and 24 h after inhalation (n = 3 independent mice for each group). Flag positivity identified PD-1–Alix-Exo, rat Fc positivity identified anti-PD-L1 presence, and His tag positivity identified recombinant PD-1 protein localization. All panels were conducted simultaneously, sharing the same 0-h control. See Supplementary Fig. 18 for gating strategies. d, Statistical lung distribution at 48 h after each treatment. The normalized rate was determined by comparing the fold changes over 0 h (C_t/C₀), as characterized by the fluorescent intensity from the lungs; n = 3 independent mice for each group. e, Schematic showing the establishment of the melanoma lung-colonized tumor model; i.v., intravenous. f, Statistical lung-colonized tumor growth over time in each treatment group. The normalized tumor growth rate was calculated by comparing the fold changes over the initial tumor burden, as characterized by the bioluminescence intensity from tumor cells. g, Ex vivo imaging of mouse lungs after 24 days of inhalation treatment; n = 5 independent mice for each group in e–g. h, Quantification of the bioluminescence density in ex vivo mouse lungs (g). i, Lung weights 24 days after inhalation treatment; n = 4 independent mice for PBS, WT-Exo and PD-1 protein groups; n = 5 independent mice for the other groups in h and i. P values were determined by one-way ANOVA with a post hoc Tukey’s multiple comparisons test (b–d and i) and two-tailed Student’s t-test (f and h) using GraphPad PRISM software. Exact P values are indicated. Results are presented as means ± s.d. The mouse cartoon was created with BioRender.com.

After confirming the preferential localization and prolonged retention of PD-1–Alix-Exo in the lungs, we assessed its antitumor efficacy in a lung-colonized BP melanoma model, comparing inhaled PD-1–Alix-Exo to anti-PD-L1 and PD-1 protein (depicted in Fig. 2e). After inhalation therapy, PD-1–Alix-Exo dramatically suppressed tumor growth in the lungs, exhibiting superior antitumor efficacy compared to anti-PD-L1 or PD-1 protein (Fig. 2f-h and Supplementary Fig. 5). With diminished tumor burden, PD-1–Alix-Exo-treated mice had correspondingly decreased lung weights compared to anti-PD-L1- or PD-1 protein-treated counterparts (Fig. 2i). Notably, PD-1–Alix-Exo inhalation exhibited greater antitumor efficacy than systemically administered anti-PD-L1 at a clinically relevant dose (Supplementary Fig. 6a,b). Moreover, the tumor-suppressive effect of PD-1–Alix-Exo followed a clear dose-dependent pattern (Supplementary Fig. 6c,d). Thus, inhalation-based delivery of PD-1–Alix-Exo showed enhanced lung targeting and antitumor efficacy in a lung-colonized melanoma, out-performing anti-PD-L1 and PD-1 protein while reducing systemic exposure and tumor burden.

Wnt7b–FZD8 interaction leads to ICI resistance

Despite the advancements in ICI therapy for metastatic melanoma, the overall effectiveness remains limited, with response rates of up to 60% (refs. 4,39,40). Therapeutic efficacy is preferentially observed in individuals exhibiting pre-existing tumor-specific T cell responses, indicated by baseline CD8⁺ T cell infiltration in the TME^41,42. It was identified that tumor-intrinsic active Wnt/β-catenin signaling results in T cell exclusion and resistance to ICI therapy¹⁶. To examine the role of paracrine and autocrine Wnt ligand signaling in ICI resistance in individuals with melanoma, we reanalyzed a previously published RNA-sequencing (RNA-seq) dataset of individuals with metastatic melanoma who had undergone ICI therapy (Gene Expression Omnibus (GEO): GSE78220)⁴³. Among genes associated with the Wnt signaling pathway, WNT7B emerged as the most markedly upregulated gene, followed by FZD8 (Fig. 3a).

Fig. 3 ∣. Wnt7b interaction with FZD8 decreases CD8⁺ T cell infiltration in the TME.

a, Reanalysis of RNA-seq data from biopsies of responding (red) and nonresponding (blue) individuals with metastatic melanoma focused on the Wnt signaling pathway (GSE78220); FC, fold change. b, Immunoblotting of immunoprecipitation assay. Duplicate assays were performed with similar results. c, TCF/LEF luciferase reporter assay in WT or shFzd8 tumor cells co-transfected with Wnt7b–FZD8 or control for 48 h (n = 4 independent experiments). d, mRNA expression in WT or shFzd8 tumor cells treated with Wnt7b or PBS for 24 h (n = 3 independent experiments). e, mRNA expression in sorted CD45⁺CD103⁺ cells from tumors (n = 3 independent experiments). f, Statistical lung-colonized tumor growth over time after anti-PD-L1 treatment. The normalized tumor growth rate was calculated by comparing the fold changes over the initial tumor burden, as characterized by the bioluminescence intensity from tumor cells. g,h, Quantified flow cytometry analysis of CD8⁺ T cells (CD45⁺CD3⁺CD8⁺; g) and CD103⁺ DCs (CD45⁺CD11c⁺CD103⁺; h) infiltrated in WT and Wnt7b–FZD8 tumors; n = 6 independent mice for each group in f–h. See Supplementary Fig. 19 for gating strategies. P values were determined by one-way ANOVA with a post hoc Tukey’s multiple comparisons test (c and d) and two-tailed unpaired Student’s t-test (e and f–h) using GraphPad PRISM software. Exact P values are indicated. Results are presented as means ± s.d.

Given that Wnt and FZD function as ligands and receptors, we conducted a pulldown assay to assess the interaction between Wnt7b and FZD8, which confirmed a direct binding between them (Fig. 3b). Suppressing FZD8 expression markedly diminished Wnt7b-induced β-catenin activation (Fig. 3c and Supplementary Fig. 7a). Previous reports had suggested that Wnt/β-catenin signaling induces expression of the transcriptional repressor ATF3 and that ATF3 suppresses CCL4 (refs. 44,45). CCL4 has been shown to correlate positively with the infiltration of CD103⁺ dendritic cells (DCs) into tumors, a key process for the recruitment of CD8⁺ T cells^16,46. Fzd8 knockdown in tumor cells reduced Wnt7b-activated Atf3 transcription and concurrently upregulated CCL4 expression (Fig. 3d). CXCL9 and CXCL10 are principal chemokines secreted by DCs to facilitate the recruitment of CD8⁺ T cells^47-49. Tumor-infiltrating CD103⁺CD11c⁺ DCs express substantially higher levels of Cxcl9 and Cxcl10 than CD103⁻CD11c⁺ DCs (Fig. 3e).

To investigate whether the interaction between Wnt7b and FZD8 leads to resistance in ICI therapy, we overexpressed Wnt7b and FZD8 in BP cells and established a lung-colonized tumor model. Treatment with anti-PD-L1 significantly inhibited the growth of WT BP tumors, whereas tumors with Wnt7b and FZD8 overexpression exhibited rapid growth (Fig. 3f and Supplementary Fig. 7b). Tumors overexpressing Wnt7b and FZD8 exhibited a markedly lower infiltration of CD103⁺ DCs and CD8⁺ T cells than WT tumors (Fig. 3g,h). Together, these data showed that elevated expression of Wnt7b and FZD8 contributes to resistance to ICI therapy, positioning Wnt7b as a promising target to counteract resistance.

Tandem presentation of FZD8 and PD-1 on BEAT

Given Wnt7b’s potential as a target for overcoming ICI resistance, we designed BEAT for tandem expression of the Wnt-binding CDR domain of FZD8 and signaling-incompetent PD-1 on the outer membrane via the sorting domain of Alix (Fig. 4a). This configuration enables FZD8 and PD-1 to act as simultaneous decoys for Wnt7b and PD-L1, respectively. We also constructed the Wnt7b-binding CDR domain of FZD8-expressing exosomes via the Alix sorting domain (FZD8–Alix-Exo) as a control. The expression of FZD8 and PD-1 on exosomes was confirmed by western blotting (Fig. 4b and Supplementary Fig. 7c). ELISAs, using PBS as the solvent, verified the expression of FZD8 and PD-1 on the external surface of BEAT (Fig. 4c). To further confirm the dual expression of FZD8 and PD-1 on BEAT, we conducted assays using Wnt7b protein to bind BEAT and detected with anti-PD-1 or PD-L1 protein for BEAT binding, followed by detection with anti-FZD8. The results verified the simultaneous expression of FZD8 and PD-1 on the external surface of BEAT (Fig. 4d and Supplementary Fig. 8a-c). Super-resolution microscopy confirmed the coexpression of FZD8 and PD-1 on an individual exosome (Supplementary Fig. 9).

Fig. 4 ∣. Strategies for the co-presentation of PD-1 and FZD8 in exosomes.

a, Schematics showing the design of the tandem protein exosome display platform (left) and the design of the construct plasmids (right). b,c, FZD8 and PD-1 expression on exosomes determined by immunoblotting (b) and ELISA (c). d, FZD8 expression on Wnt7b- and PD-L1-interacted exosomes (left) and PD-1 expression on Wnt7b- and PD-L1-interacted exosomes (right); n = 3 independent experiments for each group in c and d. e, TCF/LEF luciferase reporter assay on exosome-treated tumor cells for 48 h (n = 6 independent experiments). f,g, mRNA expression in exosome-cocultured tumor cells; n = 3 independent experiments for each group in f and g. P values were determined by one-way ANOVA, followed by a post hoc Tukey’s multiple comparisons test using GraphPad PRISM software. Exact P values are indicated. Results are presented as means ± s.d. Schematics in a and b were created with BioRender.com.

PD-L1 has been reported to be present not only on the surface of cells but also on extracellular vesicles (EVs)^50-52. To assess whether BEAT can bind to EV-associated PD-L1, we used a fluorescence quenching assay. The results showed that BEAT specifically interacted with PD-L1-presenting exosomes, as indicated by a reduced enhanced green fluorescent protein (eGFP) signal in the BEAT-treated group but not in the control group (Supplementary Fig. 10). These findings highlight BEAT’s potential for selectively targeting PD-L1-expressing exosomes.

To validate that the tandem fusion of PD-1 and FZD8 does not impede their ability to bind their respective ligands due to potential steric hindrance, we performed a PD-L1–His pulldown assay. The results confirmed that the interaction between PD-L1 protein and exosomal PD-1 remains unaffected by the presence or absence of Wnt7b (Extended Data Fig. 4a). Furthermore, PD-1 on exosomes facilitates targeted delivery to PD-L1-overexpressing tumor cells, promoting endocy-tosis and cargo depletion, including PD-L1 itself (Extended Data Fig. 4b). This enables FZD8 on the exosomes to not only block Wnt7b but also promote its internalization and degradation within tumor cells (Extended Data Fig. 4c). Given that the binding of Wnt ligands to FZD receptors is reversible rather than permanent⁵³, merely blocking Wnt7b may not be sufficient to achieve optimal therapeutic efficacy. These findings highlight the advantage of coexpressing PD-1 and FZD8 on the same exosome, ensuring efficient dual-target modulation.

To ascertain if exosomal FZD8 can inhibit the tumor Wnt signaling pathway, T cell factor/lymphoid enhancer factor (TCF/LEF) dual-luciferase reporter assays were performed. Our findings demonstrated that coculturing FZD8–Alix-Exo or BEAT with reporter-transfected BP cells reduced luciferase activity, a suppression not evident with WT-Exo and PD-1–Alix-Exo (Fig. 4e and Supplementary Fig. 11). Transcription of Atf3 was inhibited by coculturing with FZD8–Alix-Exo or BEAT, which in turn promoted the expression of Ccl4 (Fig. 4f,g). Collectively, these findings indicate that exosomal surface expression of FZD8 effectively suppresses the Wnt/β-catenin signaling pathway in tumor cells, thereby positioning FZD8- and PD-1-coexpressing BEAT as a promising strategy to overcome resistance to ICI therapy.

BEAT inhalation suppresses the growth of ICI-resistant melanoma

To verify the efficacy of BEAT against tumors resistant to ICIs, we established a lung colonization tumor model using BP cells overexpressing Wnt7b and FZD8 (Fig. 5a). Inhaled BEAT notably suppressed the growth of tumors resistant to ICIs, whereas WT exosomes and exosomes expressing either PD-1 or FZD8 alone did not impact tumor proliferation (Fig. 5b and Supplementary Fig. 12a,b). Ex vivo lung imaging at the endpoint of the study showed that treatment with BEAT significantly reduced tumor burden and decreased lung weights, indicating potent tumor suppression (Fig. 5c,d). Moreover, the tumor-inhibitory effect of BEAT was dose dependent (Extended Data Fig. 5a-c). The administration of a 1:1 mixture of FZD8-Exo and PD-1-Exo was less effective than BEAT (Extended Data Fig. 4d,e), further emphasizing the advantage of the tandem bispecific design. BEAT also demonstrated strong antitumor efficacy in a spontaneous lung metastasis model (Extended Data Fig. 6a-c). When benchmarked against a 1:1 linked dual antibody targeting PD-L1 and Wnt7b, BEAT elicited stronger tumor inhibition (Extended Data Fig. 5d-f). To further evaluate the efficacy of BEAT in treating larger solid tumors, we used a subcutaneous ICI-resistant tumor model. Treatment with intratumoral BEAT injection was initiated once the tumor volume reached approximately 200 mm³ (Supplementary Fig. 13a). Even under these stringent conditions, BEAT demonstrated potent tumor-suppressive activity, further supporting its therapeutic potential in established solid tumors (Supplementary Fig. 13b,c).

Fig. 5 ∣. Inhalation of BEAT suppresses the growth of tumors resistant to checkpoint inhibitor therapy.

a, Schematic showing the establishment of the melanoma lung-colonized tumor model. b, Statistical lung-colonized tumor growth over time in each treatment group. The normalized tumor growth rate was calculated by comparing the fold changes over the initial tumor burden, as characterized by the bioluminescence intensity; n = 6 independent mice for each group in a and b. c, Ex vivo imaging of mouse lungs after 21 days of inhalation treatment (left) and quantification of the bioluminescence density in ex vivo mouse lungs (right). d, Lung weights 21 days after inhalation treatment; n = 4 independent mice for PBS and WT-Exo groups; n = 5 independent mice for PD-1–Alix-Exo and FZD8–Alix-Exo groups; n = 6 for the FZD8–PD-1–Alix-Exo (BEAT) group in c and d. e, Representative immunostaining images of tumor-bearing mouse lungs. Flag was stained with anti-Flag (gray or red, as indicated in the images). PD-L1 was stained with anti-PD-L1 (red), and Wnt7b was stained with anti-Wnt7b (gray). Nuclei were stained with DAPI (blue); scale bar, 15 μm. Images are representative of six mice. f, Relative mRNA expression in sorted tumor cells (n = 3 independent experiments). g,h, Flow cytometry analysis of CD103⁺ DC (CD45⁺CD11c⁺CD103⁺; g) and CD8⁺ T cell (CD45⁺CD3⁺CD8⁺; h) infiltration in tumors. i, Flow cytometry analysis of effector CD8⁺ T cells (CD45⁺CD3⁺CD8⁺granzyme B⁺Lag3⁻). j, Flow cytometry analysis of exhausted CD8⁺ T cells (CD45⁺CD3⁺CD8⁺PD-1⁺Lag3⁺); n = 6 independent mice for each group in g–j. See Supplementary Fig. 19 for gating strategies. P values were determined by one-way ANOVA with a post hoc Tukey’s multiple comparisons test using GraphPad PRISM software. Exact P values are indicated. Results are presented as means ± s.d. The mouse cartoon was created with BioRender.com.

Subsequently, we demonstrated that, following 24 h of inhalation treatment, BEAT binds to Wnt7b and PD-L1 in the TME, as evidenced by immunofluorescence staining (Fig. 5e). In tumor cells sorted from BEAT- and FZD8–Alix-Exo-treated tumors, Atf3 expression was markedly decreased, accompanied by an upregulation of Ccl4 (Fig. 5f). BEAT treatment significantly enhanced the infiltration of CD103⁺ DCs and CD8⁺ T cells into tumors (Fig. 5g,h), potentially elucidating the observed mitigation in tumors resistant to ICIs. The PD-1–PD-L1 interaction reduces CD8⁺ T cell activity by limiting their production of critical cytolytic agents like interferon-γ (IFNγ) and granzyme B and leading to metabolic exhaustion, weakening their defense against tumors^54-56. Treatment with BEAT led to a marked increase in effector CD8⁺ T cells and a significant reduction in the population of exhausted CD8⁺ T cells (Fig. 5i,j). BEAT therapy also elicited immune memory development. Six weeks of BEAT treatment revealed a surge in efficacious CD8⁺ memory T cells (Supplementary Fig. 14). Together, these results indicated that inhalation of BEAT effectively countered ICI-resistant lung-colonized tumors by enhancing immune infiltration and bolstering the immune response, underscoring its therapeutic promise.

Beyond the skin, the liver is the second most common site of distant metastases in melanoma, following the lungs. To evaluate BEAT’s efficacy in treating metastases beyond the lungs, we developed a melanoma liver metastasis model with Wnt7b- and FZD8-overexpressing BP cells. BEAT exhibited strong liver tropism following intravenous injection (Extended Data Fig. 7a,b). BEAT was then administered to tumor-bearing mice every 3 days over a 19-day period. The results demonstrated that BEAT effectively inhibited the growth of liver-colonized tumors (Extended Data Fig. 7c,d), highlighting its potential for broad therapeutic applications.

BEAT counteracts ICI resistance in a humanized PDX mouse model

To more accurately reflect tumor growth and pharmacological responses in humans, while maintaining the original tumor’s heterogeneity and complexity, we used a PDX subcutaneous model in mice with humanized immune systems. Considering the potential influence of CD34⁺ cell donors on treatment outcomes, we injected immunodeficient NOD/SCID mice with CD34⁺ cells from three distinct donors: donor 1321, donor 1323 and donor 1401. After 12 weeks, these mice successfully harbored a fully developed human immune system (Supplementary Fig. 15a-d). Patient-derived metastatic melanoma resistant to pembrolizumab treatment was used in this model (Fig. 6a). We used the same methodology as in mice to construct the human Wnt7b-binding CDR domain of FZD8 and human signaling-incompetent PD-1 on BEAT. In humanized mice derived from donors 1323 and 1401, BEAT intratumor injection completely inhibited tumor growth, achieving complete regression in some cases. For mice derived from donor 1321, although BEAT did not halt tumor progression, it markedly reduced tumor growth compared to controls. Across all humanized mice derived from the three donors, pembrolizumab failed to mitigate tumor growth (Fig. 6b-d). Within the TME, the interaction between BEAT and both PD-L1 and Wnt7b was confirmed via immunofluorescence staining (Fig. 6e). These interactions facilitate an increased infiltration of CD103⁺CD11c⁺ cells compared to control and pembrolizumab treatment, subsequently recruiting CD8⁺ T cells into the TME (Fig. 6f,g). Additionally, CD8⁺ T cells in tumors treated with BEAT exhibited decreased PD-1 expression, indicative of diminished T cell dysfunction (Fig. 6h)⁵⁴. Furthermore, FoxP3⁺ regulatory T cell populations were lower following BEAT treatment than following pembrolizumab or control treatment (Fig. 6i).

Fig. 6 ∣. BEAT inhibits the growth of patient-derived tumors resistant to checkpoint inhibitor therapy in humanized mice.

a, Schematic showing the establishment of the PDX model in humanized CD34 mice. b, Image of excised tumors 21 days after treatment in humanized CD34 mice. c, Tumor weights 21 days after treatment. d, Statistical tumor growth over time in each treatment group. The normalized tumor growth rate was calculated by comparing the fold changes over the initial tumor volume; n = 3 independent mice for each group in a–d. e, Representative immunostaining images of PDX tumors after treatment. Flag was stained with anti-Flag (gray or red, as indicated in the images). PD-L1 was stained with anti-PD-L1 (red), and Wnt7b was stained with anti-Wnt7b (gray). Nuclei were stained with DAPI (blue); scale bar, 15 μm. Images are representative of six mice. f–i, Quantified flow cytometry analysis of CD8⁺ T cells (CD45⁺CD3⁺CD8⁺; f), CD103⁺ DCs (CD45⁺CD11c⁺CD103⁺; g), PD1⁺ CD8⁺ T cells (CD45⁺CD3⁺CD8⁺PD-1⁺; h) and regulatory T cells (T_reg; CD45⁺CD4⁺Foxp3⁺; i) infiltrated in the PDX tumors; n = 9 independent mice for each group in f–i. See Supplementary Fig. 19 for gating strategies. j, Schematic showing the establishment of the PDX lung-colonized model in humanized CD34 mice. k, Representative hematoxylin and eosin staining of the lungs after 21 days of inhalation treatment (n = 3 independent mice for each group). l, Statistical analysis of tumor nodules in k; n = 9 independent mice for each group. m, Graphical summary. P values were determined by one-way ANOVA with a post hoc Tukey’s multiple comparisons test using GraphPad PRISM software. Exact P values are indicated. Results are presented as means ± s.d. The mouse cartoon was created with BioRender.com.

To further assess the efficacy of BEAT against lung-colonized tumors, primary tumor cells were isolated from patient-derived tumors xenografted in the subcutaneous model, and a lung colonization tumor model was established in humanized mice (Fig. 6j). Inhalation of BEAT significantly suppressed lung tumor growth, surpassing the efficacy of pembrolizumab, as evidenced by markedly reduced tumor sizes and a substantial decrease in tumor nodules (Fig. 6k,l). We conclude that inhalation of BEAT specifically targets the lung area, significantly diminishing ICI-resistant lung metastasis of melanoma by inhibiting the Wnt/β-catenin pathway and enhancing the recruitment and activation of CD8⁺ T cells (Fig. 6m).

Toxicity study of BEAT

To assess the toxicity and safety of BEAT administered via inhalation, serum samples from humanized mice were analyzed at the study’s endpoint. Systemic administration of pembrolizumab is associated with a notable risk of inducing hepatitis and acute kidney injury^6,7,9. By contrast, our findings revealed that aspartate aminotransferase and alanine aminotransferase levels remained within physiological norms following the administration of BEAT by inhalation (Supplementary Fig. 16a,b). Additionally, serum urea nitrogen and creatinine levels were comparable to those observed in the control group (Supplementary Fig. 16c,d), indicating no significant renal impact. Additionally, hematoxylin and eosin staining indicated an absence of severe lesions in primary organs after treatment with BEAT via inhalation (Supplementary Fig. 16e). All evaluated parameters remained within normal ranges.

To directly compare the immune-related adverse effects of BEAT therapy and ICIs, we conducted a study using healthy C57BL/6 mice (Extended Data Fig. 8a). The mice were administered anti-PD-L1 intravenously and BEAT via inhalation twice a week for up to 8 weeks. At the endpoint of the study, the results showed that mice treated with anti-PD-L1 exhibited elevated blood glucose levels, likely due to autoimmune attacks on pancreatic beta cells, whereas BEAT-treated mice did not (Extended Data Fig. 8b-d). Additionally, serum aspartate aminotransferase levels were significantly increased in the antibody-treated group, suggesting potential liver damage, whereas BEAT treatment did not induce such effects (Extended Data Fig. 8e). Histological analysis of the liver, lungs, heart, spleen and kidneys showed no signs of inflammation or damage in either treatment group (Extended Data Fig. 8f). These results highlight that BEAT therapy demonstrates a more favorable safety profile than systemic anti-PD-L1 treatment, with reduced immune-related adverse effects, likely due to its localized biodistribution.

Discussion

Conventional methods for the endogenous incorporation of proteins into exosomes have consistently shown suboptimal efficiency, particularly in scenarios requiring the coexpression of two distinct proteins within the same exosome⁵⁷. The endeavor to colocalize two proteins on a single exosome frequently reduces loading efficiencies. Chemical modification to improve efficiency has inadvertently altered the exosomal surface proteome³⁰. Recent work by Gupta et al. demonstrated the potential of engineering exosomes to display therapeutic proteins by fusing decoy receptors to the N terminus of syntenin for systemic inflammation treatment³¹. Our BEAT platform introduces four distinct relevant advances. First, we use an Alix-based sorting system rather than syntenin, which markedly improves loading efficiency of therapeutic cargo onto exosomal membranes. Second, BEAT uses a tandem fusion of PD-1 and FZD8, ensuring that every engineered exosome simultaneously displays both proteins at a 1:1 ratio. This overcomes a major limitation of separate coexpression, which leads to population-level heterogeneity. The tandem design improves exosome uniformity, enhances reproducibility and provides a more reliable basis for clinical translation. Third, both proteins share the same transmembrane anchor and exosomal loading signal, minimizing competition for sorting machinery and ensuring consistent orientation and surface presentation. Fourth, tandem fusion fixes the spatial geometry between PD-1 and FZD8 at the nanoscale, increasing the likelihood of simultaneous engagement with PD-L1 and Wnt ligands, thereby enhancing cooperative internalization and degradation. Collectively, BEAT supports coordinated loading, preserves native exosomal properties and promotes enhanced internalization and degradation of both PD-L1 and Wnt ligands within the TME, thereby maximizing therapeutic efficacy.

Inhalation therapy has the potential to improve management of lung metastases²⁵. Our comparative analysis reveals the advantages of inhaled PD-1–Alix-Exo over anti-PD-L1 and PD-1 protein. Notably, PD-1–Alix-Exo demonstrates lung localization with minimal extrapulmonary dispersion, significantly boosting antitumor efficacy compared to systemic antibodies. This advantage is largely attributed to the exosomes’ extended pulmonary retention, which avoids hepatotoxicity and nephrotoxicity.

Efficacy of ICIs in melanoma is observed in less than 60% of treated individuals^4,39,40. The Wnt signaling pathway is instrumental in cancer progression through its roles in promoting tumor proliferation, resisting apoptosis and fostering immunosuppressive TME formation^58,59. Notably, activation of the Wnt pathway has been linked to diminished CD8⁺ T cell infiltration, further exacerbating resistance to ICI therapy¹⁶. We therefore engineered BEAT to co-display FZD8 and PD-1, remodeling the ICI-resistant TME and activating suppressed CD8⁺ T cells. This approach demonstrated efficacy both in C57BL/6 mice and in humanized immune system mouse PDX models, suppressing the growth of human-derived metastatic melanoma resistant to pembrolizumab in both subcutaneous and lung colonization contexts. In this study, we used CD34⁺ cells from healthy donors to establish humanized mice, a widely accepted approach in preclinical research. Healthy donor-derived CD34⁺ cells offer superior viability, proliferation and differentiation capacity, ensuring robust and reproducible immune reconstitution. Although patient-derived CD34⁺ cells may better capture disease-specific immune alterations, their variability can limit experimental consistency. Despite not fully replicating patient-specific immune dysfunctions, the use of healthy donor cells provides a reliable and standardized model for evaluating therapeutic effects. Our results support the potential of inhalation therapy using engineered exosomes as a strategy to counteract ICI resistance in melanoma.

Despite the strong preclinical efficacy demonstrated by the BEAT platform, several limitations should be acknowledged. First, although inhaled delivery achieves localized targeting and minimizes systemic toxicity, its applicability may be limited to tumors with lung involvement. Notably, intravenous administration of BEAT also exhibited efficacy in a model of liver metastasis, suggesting potential flexibility in delivery routes depending on metastatic patterns. Second, the mechanisms underlying ICI resistance are diverse. Although BEAT effectively reverses resistance mediated by Wnt/β-catenin signaling, its efficacy may be reduced in tumors driven by alternative resistance pathways. Further exploration of combination strategies or biomarker-guided patient selection may help address this limitation. Last, although results in humanized PDX models are encouraging, clinical translation will require a comprehensive evaluation of long-term safety and dosing in larger animal models and ultimately in human trials.

Methods

Cell culture

HEK293T cells were obtained from ATCC (ACS-4500), and BP cells were a generous gift from the laboratory of B. Hanks (Duke University Medical Center). The cells were cultured at 37 °C and 5% CO₂ in Dulbecco’s modified Eagle’s medium with 10% fetal bovine serum and 1% penicillin–streptomycin. HEK293T cells and BP cells were transfected using Lipo-fectamine 2000 (Thermo Fisher Scientific). Infection of HEK293T cells was performed using the standard lentiviral infection protocol⁶⁰.

Exosome isolation and electron microscopy

Exosomes were isolated from filtered (0.22 μm, Corning) secretomes using tangential flow filtration (KR2i, Repligen), followed by ultra-centrifugation (130,000g; Beckman Coulter Ultracentrifuges)⁶¹. Concentrations of exosomes were quantified by NanoSight (NS3000, Malvern Panalytical). Successful exosome isolation was confirmed by immunoblotting for exosome markers CD9, TSG101 and Alix. Exosome morphology was confirmed by transmission electron microscopy (HT7800, RuliTEM). The protein crown on the exosome surface was confirmed by cryo-electron microscopy (Talos Arctica G3, Thermo Fisher Scientific).

Animal models

All mice were housed in a controlled environment with a 12-h light/12-h dark cycle, ambient temperature maintained at 20–24 °C and relative humidity of 45–55%. C57BL/6J mice (6 weeks old, female) were purchased from Charles River Laboratories. Humanized CD34 mice with PDX tumors (TM00702) were purchased from The Jackson Laboratory⁶². Briefly, female NOD/SCID mice were injected with human hematopoietic stem cells (hCD34⁺). Engraftment of mature human white blood cells (hCD45⁺) was confirmed 12 weeks after injection. Any mouse with more than 25% hCD45⁺ cells was considered successfully humanized. Tumor (TM00702) was subcutaneously engrafted into the right flank of the humanized mice. All animal studies complied with the regulations and guidelines relating to the use of animals in the Institutional Animal Care and Use Committee of Columbia University (protocol number AC-AABY3662) and North Carolina State University (protocol number 22-411).

Tumor inoculation and drug delivery

Subcutaneous melanoma models were established through the subcutaneous injection of 1 × 10⁶ BP cells at the right flank. In total, 1 × 10¹¹ exosomes or an equivalent molar amount of anti-PD-L1 (matching PD-1–Alix-Exo) were intratumorly injected after 1 week of tumor inoculation and repeated every 3 days for a total of 2–3 weeks. The tumor size was calculated by (length × width × height) / 2. Lung metastatic melanoma tumor models were established through intravenous injection of 5 × 10⁵ luciferase-expressing tumor cells. Tumor signals were detected via the luciferase activity of intraperitoneally injected D-luciferin bioluminescent substrate (PerkinElmer), followed by quantification using an IVIS Spectrum imaging system. Inhalation treatments started 1 week after tumor injection and were administrated every 3 days. The dose was standardized by the exosome particle number (1 × 10¹¹ particles per mouse). The dose of antibodies and proteins was equivalent to exosomal PD-1 (1.2 μg). Systemic antibody delivery was performed via tail vein injection at a dose of 200 μg per mouse, administered every 3 days. For the PDX model, the tumor was finely minced, and 40 μl of tumor tissue was subcutaneously engrafted into the right flank of each mouse. Treatment was started 1 week after tumor inoculation. The tumor size was calculated by (length × width × height) / 2. For the spontaneous lung metastasis model, 1 × 10⁶ tumor cells were subcutaneously injected into the right flank of the mice, and primary tumors were surgically removed after reaching 1,000 mm³. BEAT inhalation (1 × 10¹¹ exosomes per dose) was initiated 4 weeks after tumor removal and was administered every 3 days. For the liver metastasis model, 1 × 10⁶ tumor cells were injected through the portal vein. BEAT treatment (1 × 10¹¹ exosomes per dose) was administered via intravenous injection starting 1 week after tumor inoculation and continued every 3 days. For the distribution study, animals were killed 24 h after inhalation of NHS-rhodamine-labeled samples, followed by IVIS analysis of rhodamine signal.

Plasmid construction and generation of stable cell lines

Generation of the plasmids was described previously⁶³. The plasmids were constructed by inserting the sequences of interest into the pLenti-C-Myc-DDK-IRES-Puro lentiviral gene expression vector (Origene). The sequence information is shown in Supplementary Table 1. Stable expression of the constructs in HEK293T cells was achieved by lentiviral infection, followed by puromycin selection. Fzd8 short hairpin RNA (shRNA) plasmid was generated by inserting 5′-CCGGGCAAAGGCATCGGTTACAACTCTCGAGA GTTGTAACCGATGCCTTTGCTTTTTG-3′ into pLV-RNAi-Vector (Bioset-tia). The TCF/LEF dual-luciferase reporter plasmid was purchased from System Biosciences (R413PA-P).

Duel-luciferase reporter assay

Dual-luciferase reporter, Renilla control vector and shRNA or expression plasmids were co-transfected into BP cells. Exosomes were added after 24 h of transfection. Luciferase levels were measured after 48 h using a Dual Luciferase Assay kit according to the manufacturer’s instructions (Promega). Corrected values were obtained according to the following formula: (luciferase sample – luciferase control) / (Renilla sample – Renilla control), where control values were obtained from untransfected cells. Relative luciferase values were defined as the corrected value of the samples divided by the corrected value of the controls⁶⁴.

Immunoblotting

Western blotting was performed using a standard protocol as described previously⁶⁵. Cells were lysed in NETN buffer (20 mM Tris-HCl (pH 8.0), 100 mM NaCl, 0.5 mM EDTA and 0.5% NP-40) containing protease inhibitors, aprotinin (4 μg μl⁻¹), PMSF (1 mM) and phosphatase inhibitors (Thermo Fisher Scientific, 88667). Cell lysates were boiled in 2× SDS loading buffer and subjected to SDS–PAGE. The commercial antibodies used in this study are listed in the Supplementary Table 2.

Pulldown assay

Exosomes were incubated with recombinant His-tagged PD-L1 protein (9048-B7-100, R&D) overnight at 4 °C, followed by a 2-h incubation with His-tag Dynabeads (10103D, Thermo Fisher Scientific). After washing, western blots were performed to detect Flag-tagged PD-1. Myc-tagged Wnt7b and HA-tagged FZD8 were overexpressed in BP cells. Cells were lysed 48 h after transfection, and the lysates were incubated with HA tag magnetic beads (88836, Thermo Fisher Scientific) for 2 h at 4 °C. After washing, western blots were performed to detect Myc-tagged Wnt7b. Recombinant Wnt7b (CUSABIO) and His-tagged PD-L1 protein were incubated with BEAT overnight at 4 °C, followed by a 2-h incubation with His tag Dynabeads. After thorough washing, western blotting was performed to detect Flag-tagged PD-1.

Immunofluorescence

Cells were seeded on coverslips and fixed with cold methanol for 15 min at −20 °C. Frozen tissue slides were fixed with 4% paraformaldehyde (15710, Electron Microscopy Sciences) and permeabilized and blocked using Dako protein blocking solution (X0909, Dako) containing 0.1% saponin (47036, Sigma-Aldrich). The slides were then washed with PBS with 0.1% Tween-20, followed by blocking with 5% normal goat serum diluted in PBS. Primary antibodies were diluted in 5% normal goat serum to the appropriate concentrations and incubated with cells overnight at 4 °C. The slides were then washed with PBS with 0.1% Tween-20 and subsequently incubated with diluted secondary antibodies at room temperature for 1 h. After washing and staining the nuclei with DAPI, the slides were sealed with a quenching-preventive mounting medium. Images were acquired using a confocal microscope. Hematoxylin and eosin staining (HSS16 and 318906, Sigma-Aldrich) was performed on paraffin-embedded, formalin-fixed tissues from major mouse organs (spleen, heart, liver, lung and kidney).

For LysoTracker staining, LysoTracker Red (MCE, HY-D1300) was added after 6 h of incubation with DiD-labeled exosomes. Following a 30-min incubation with LysoTracker, Hoechst was added, and the cells were analyzed using confocal microscopy.

For single-exosome imaging, exosomes were stained using an ONI EV Profiler kit with anti-PD1 and anti-FZD8. The stained samples were observed under an ONI Nanoimager and analyzed using the CODI platform.

Quantitative real-time PCR

Total RNA was extracted using TRIzol reagent (Invitrogen) following manufacturer’s instructions. cDNA was then synthesized using a cDNA Synthesis kit (Bio-Rad). Quantitative real-time PCR was performed in a CFXTM real-time thermal cycler (Bio-Rad) using a Universal SYBR Green Supermix (Bio-Rad). Data analysis was performed with the comparative change in cycling threshold (ΔC_t) method using Gapdh as an internal control. The sequences of the primers used in this study are listed in Supplementary Table 3.

Flow cytometry and cell sorting

For flow cytometry with cells, cells were collected and stained with the indicated markers. For flow cytometry with tissues, after tumor-bearing mice were killed, lungs or tumors were sliced and digested using a mixture of collagenase IV (1 g 100 ml⁻¹ HBSS; Sigma-Aldrich, C-5138), hyaluronidase (100 mg 100 ml⁻¹ HBSS; Sigma-Aldrich, H-6254) and DNaseI (20,000 U 100 ml⁻¹ HBSS; Sigma-Aldrich, D-5025) in serum-free RPMI (Sigma-Aldrich, R8758) for 30 min at 37 °C with 250 rpm rotation. These digested tumor tissues were then filtered using 70-μm nylon strainers and ground into single-cell suspensions. After removing red blood cells with ACK Lysis Buffer (Gibco), cell viability was distinguished using a Fixable Near-IR Dead Cell Stain kit (Invitrogen, L10119), and Fc was blocked with anti-CD16/CD32 (101320, Biolegend). The cells were then stained. Intracellular staining of FoxP3 and granzyme B was performed using a FoxP3 Transcription Factor Buffer Set (00-5523-00, eBioscience). Stained cell samples were analyzed using a BD LSR II, and data were analyzed using FlowJo software. Cell sorting was performed using an Aria II cell sorter. We used the following markers to identify cell populations: CD8⁺ T cells (CD45⁺CD3⁺CD8⁺), CD4⁺ T cells (CD45⁺CD3⁺CD4⁺), regulatory T cells (CD45⁺CD4⁺FoxP3⁺), effector CD8⁺ T cells (CD45⁺CD3⁺CD8⁺granzyme B⁺Lag3⁻), exhausted CD8⁺ T cells (CD45⁺CD3⁺CD8⁺PD-1⁺Lag3⁺) and DCs (CD45⁺CD11c⁺). For single-vesicle flow cytometry, DiD-labeled exosomes were incubated with fluorophore-conjugated anti-CD63 and anti-PD-1 at room temperature for 30 min. The stained exosomes were then analyzed using an ImageStreamX-MKII flow cytometer.

Fluorescence quenching-based exosomal binding assay

PD-L1–eGFP exosomes were generated by infecting HEK293 cells with a lentivirus encoding PD-L1–eGFP, followed by exosome isolation. To introduce a quencher (BHQ1), WT-Exo and BEAT were incubated with BHQ1-DNA-cholesterol (Eurofins) at 37 °C for 60 min. BHQ1-modified WT-Exo and BEAT were then incubated with PD-L1–eGFP exosomes at 37 °C for 30 min. Fluorescence scanning was performed on a TECAN INFINITE M PLEX plate reader.

Construction of dual antibody

Rat anti-PD-L1 (Bioxcell, BE0101) and ipafricept (MCE, HY-P99667) were incubated with TCO-PEG-NHS (NANOCS, PG2-NSTC-5k) and tetrazine-PEG-NHS (NANOCS, PG2-NSTZ-5k) at a 1:1 molar ratio at 4 °C for 24 h. Unreacted TCO-PEG-NHS and tetrazine-PEG-NHS were removed by dialysis using Slide-A-Lyzer MINI Dialysis Units (10,000 molecular weight cutoff) and centrifugation with Amicon Ultra-0.5 filters (100 kDa). Equimolar amounts of purified anti-PD-L1–TCO and ipafricept–tetrazine were then mixed and incubated at 4 °C for 48 h to allow conjugation via the TCO–tetrazine linkage. The final dual antibody was concentrated using ultrafiltration.

Toxicity study

Healthy C57BL/6 mice received intravenous anti-PD-L1 or BEAT inhalation every 3 days for up to 8 weeks. Additionally, complete Freund’s adjuvant (InvivoGen) was administered subcutaneously on days 35 and 56 following the initial ICI injection to simulate immune activation⁸. At the end of the study, blood samples were analyzed for metabolic panels, and major organs were collected for histological examination of inflammation. Blood biochemistry tests were conducted by North Carolina State University Veterinary Hospital’s clinical pathology laboratory or Columbia University Institute of Comparative Medicine.

Statistics and reproducibility

All quantitative data are presented as means ± s.d. A two-tailed unpaired Student’s t-test was used to analyze differences between two groups. One-way ANOVAs were used for comparisons among three or more groups, supplemented with Tukey’s multiple comparisons tests. Differences were analyzed using GraphPad PRISM software, and exact P values are documented in the figures. The exact replication numbers are stated in the figure legends. No exclusion criteria were incorporated in the design of the experiments for this study.

Extended Data

Extended Data Fig. 1 ∣. Characterization of engineered exosomes.

a, Immunoblotting of exosome markers in engineered exosomes. b, NanoSight size distribution analysis of the engineered exosomes. c, Representative TEM image of the isolated engineered exosomes. Scale bar, 100nm. d, SPR analysis on the affinity between PD-L1 and engineered exosomes. e, Flow cytometry showing the percentage of rmPD1-Fc positive tumor cells post PD1-Alix-Exo treatment (n = 3 independent experiments). P values were determined by one-way ANOVA post-Tukey’s multiple comparisons test using GraphPad PRISM software. Exact P values are indicated. Results are presented as means ± s.d.

Extended Data Fig. 2 ∣. Single vesicle analysis of engineered exosomes.

a, Single exosome flow cytometry analysis of PD1 positive exosomes after engineering. b, Representative single exosomes in (a). c, Gating strategy for (a). d. Single exosome imaging for engineered PD1 exosomes generated by ONI Nanoimager.

Extended Data Fig. 3 ∣. Comparison of Inhaled PD1-Alix-Exo and Intravenously Administered Anti-PD-L1 Antibody.

a, Flow cytometry showing the percentage of anti-PD-L1 antibody positive tumor cells at 0 h, 0.5 h, 4 h, 24 h and 48 h post i.v. injection (n = 5 independent mice for each group). b, Quantification of Ex vivo fluorescent images of lungs after Cy7-labeled PD1-Alix-Exo inhalation and Cy7-labeled anti-PD-L1 antibody i.v. injection at 0 h, 0.5 h, 24 h, 48 h, 96 h, and 168 h. The normalized rate was determined by comparing the fold changes over pre-delivery. c, Quantification of Ex vivo fluorescent imaging of blood after Cy7-labeled PD1-Alix-Exo inhalation and Cy7-labeled anti-PD-L1 antibody i.v. injection at 0 h, 0.5 h, 24 h, 48 h, and 96 h. The normalized rate was determined by comparing the fold changes over pre-delivery. P values were determined by one-way ANOVA post-Tukey’s multiple comparisons test in (a) and by two-tailed unpaired Student’s t-test in (b and c). P values were determined using GraphPad PRISM software. Exact P values are indicated. Results are presented as means ± s.d.

Extended Data Fig. 4 ∣. Synergistic Targeting of BEAT.

a, Immunoblotting of immunoprecipitation assay. b, Western blot of tumor cell PD-L1 with the indicated treatment. c, Representative immunostaining images of BEAT treated tumor cells. Lysosomes were stained with lysotracker (red), BEAT were stained with DiD (gray), and nuclear were stained with Hoechst (blue). d, Schematic illustrating the assessment of antitumor efficacy following treatment. e, Statistical lung colonized tumor growth over time post BEAT and FZD8-Exo plus PD1-Exo treatment. The normalized tumor growth rate was calculated by comparing the fold-changes over the initial bioluminescence intensity from tumor cells. Experiments were conducted simultaneously with Extended Data Fig. 5f, sharing the same PBS and BEAT groups. n = 5 independent mice for each group. P values were determined by two-tailed unpaired Student’s t-test using GraphPad PRISM software. Exact P values are indicated. Results are presented as means ± s.d.

Extended Data Fig. 5 ∣. BEAT therapy exhibits dose-dependent tumor suppression and outperforms systemically delivered dual antibody.

a, Schematic illustrating the assessment of antitumor efficacy following treatment. b, Normalized luminescent intensity of tumors at the endpoint of the study. n = 5 independent mice for each group c, Dose reponse curve. n = 5 independent mice for each group. d, Confirmation of linked dual antibody. e, Schematic illustrating the assessment of antitumor efficacy following treatment. f, Statistical tumor growth over time post BEAT and dual antibody treatment. The normalized tumor growth rate was calculated by comparing the fold-changes over the initial bioluminescence intensity from tumor cells. Experiments were conducted simultaneously with Extended Data Fig. 4e, sharing the same PBS and BEAT groups. n = 5 independent mice for each group. P values were determined by two-tailed unpaired Student’s t-test using GraphPad PRISM software. Exact P values are indicated. Results are presented as means ± s.d.

Extended Data Fig. 6 ∣. BEAT inhibits the tumor growth of spontaneous lung metastasize.

a, Schematic illustrating the assessment of antitumor efficacy following treatment in a checkpoint inhibitor therapy–resistant tumor model established by subcutaneous injection of 1×10⁶ cells. b, Representative H&E staining of lungs with spontaneous metastasize after BEAT treatment. c, Statistical analysis of tumor nodules in (a). P values were determined by two-tailed unpaired Student’s t-test using GraphPad PRISM software. Exact P values are indicated. Results are presented as means ± s.d.

Extended Data Fig. 7 ∣. BEAT inhibits the melanoma liver metastasis.

a, Ex vivo imaging of tumor bearing mouse liver after 24h of Cy7-labeled BEAT intravenous injection. b, Quantification of the fluorescence density in ex vivo mouse livers (a). c, Schematic illustrating the assessment of antitumor efficacy following treatment. d, Statistical liver colonized tumor growth over time post BEAT treatment. The normalized tumor growth rate was calculated by comparing the fold-changes over the initial bioluminescence intensity from tumor cells. n = 5 independent mice for each group. P values were determined by two-tailed unpaired Student’s t-test using GraphPad PRISM software. Exact P values are indicated. Results are presented as means ± s.d.

Extended Data Fig. 8 ∣. Toxicity study of BEAT therapy in healthy mice.

a, Schematic illustrating the assessment of antitumor efficacy following treatment. b, Serum glucose levels in healthy mice after BEAT and antibody therapy. n=5 in each treatment group. c, Representative H&E staining of the pancreas after BEAT and antibody treatment. d, Representative immunostaining images of BEAT and antibody treated mouse pancreas. Insulin was stained with anti-insulin antibody (red), cleaved caspase3 was stained with anti-cleaved caspase 3 antibody (green) and and nuclear were stained with DAPI (blue). e, Serum AST levels in healthy mice after BEAT and antibody therapy. n=5 in each treatment group. f, Representative H&E staining of mouse major organs at the study endpoint. P values were determined by one-way ANOVA post-Tukey’s multiple comparisons test using GraphPad PRISM software. Exact P values are indicated. Results are presented as means ± s.d.

Supplementary Material

Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41587-025-02890-8.

Data availability

Sequences for all plasmids and primers are provided in Supplementary Information. The RNA-seq data for Fig. 3a were reanalyzed from the dataset deposited in GEO (GSE78220)⁶⁶. Source data are provided with this paper.