This study was designed as an in vitro proof-of-concept investigation to evaluate the efficacy of PSMA-targeted EVs for the delivery of ADAM9-targeting siRNA and to investigate their effects on prostate cancer cell growth, migration, and resistance-associated signaling. Experiments were conducted using both hormone-sensitive and CRPC cell models. Functional assays, molecular analyses, and EV characterization were performed to assess delivery efficiency and biological effects.

Human prostate cancer cell lines LNCaP [CRL-1740; research resource identifier (RRID): CVCL_0395], CWR22Rv1 (CRL-2505; RRID:CVCL_1045), PC3 (CRL-1435; RRID:CVCL_0035), DU145 (HTB-81; RRID:CVCL_0105), and C4-2 (CRL-3314; RRID:CVCL_4782) were obtained from the American Type Culture Collection (ATCC, Manassas, USA) and maintained in Roswell Park Memorial Institute medium (RPMI-1640). A reference short tandem repeat (STR) profile is available via the ATCC STR database; no additional STR profiling was performed in our laboratory for this study. An immortalized human bone marrow-derived MSC line, previously established and kindly provided by Dr. Shih-Chieh Hung^[14], was cultured in low-glucose Dulbecco’s Modified Eagle Medium (DMEM). HEK293FT and HEK293A cells (Invitrogen, Waltham, USA) were maintained in high-glucose DMEM. All media were supplemented with 10% fetal bovine serum (FBS) and 1% penicillin-streptomycin-neomycin unless otherwise specified. Androgen-targeted therapy was carried out by culturing prostate cancer cells in phenol red-free RPMI-1640 with 5% charcoal-stripped serum (CSS) to eliminate exogenous steroid hormones and/or treating them with 10 µM abiraterone (Cat. #SML1527, Sigma-Aldrich, St. Louis, USA) to inhibit endogenous androgen synthesis. For sphere formation, cells were cultured in DMEM/F12 (1:1) supplemented with B27 (Cat. #17504044, ThermoFisher Scientific, Waltham, USA), 20 ng/mL basic fibroblast growth factor (Cat. #3718-FB, R&D Systems, Minneapolis, USA), and 20 ng/mL epidermal growth factor (Cat. #236-EG, R&D Systems, Minneapolis, USA). Unless otherwise specified, all media and reagents were obtained from Cytiva (Marlborough, USA). All cell lines were routinely tested for mycoplasma contamination using a Mycoplasma polymerase chain reaction (PCR) detection kit (Cat. #G238, Applied Biological Materials, Richmond, BC, Canada), and were generally used within approximately 30 passages of receipt.

A pUC57 plasmid encoding a PSMA-targeting scFv (PSMA-scFv) was constructed by Genomics BioSci & Tech (New Taipei City, Taiwan) using the variable regions of the J591 antibody (GenBank: FR853148.1 and FR853149.1) linked by a (Gly₄Ser)₃ peptide linker. The PSMA-scFv cassette was cloned into a lysosome-associated membrane protein 2B (LAMP2B) expression vector (Cat. #RC200456; OriGene, Rockville, USA) to enable EV surface display. For initial validation of the fusion construct, the PSMA-scFv-LAMP2B plasmid was transiently transfected into HEK293A cells using Maestrofectin Transfection Reagent (Cat. #MF002; Omics Bio, Taipei, Taiwan) according to the manufacturer’s protocol. Following validation, the PSMA-scFv-LAMP2B fusion construct was subsequently subcloned into a lentiviral vector (pLAS3w.Pbsd) using an In-Fusion Snap Assembly Kit (Cat. #638947; Takara Bio, Kusatsu, Japan) according to the manufacturer’s instructions. A LAMP2B-only lentiviral construct was generated in parallel as a control.

The PSMA-scFv-LAMP2B and LAMP2B-only lentiviral constructs were used to generate engineered MSCs for EV production. In parallel, a lentiviral vector encoding human FOLH1 (Cat. #20736061; Applied Biological Materials, Richmond, Canada) was used to establish PSMA-expressing prostate cancer cell models. Lentiviral particles were produced by transfecting HEK293FT cells using Maestrofectin. Viral supernatants were collected and used to transduce target cells to establish stable cell lines following protocols provided by the RNAi Core Facility (Academia Sinica, Taipei, Taiwan).

EVs were isolated from conditioned media of parental and engineered MSCs cultured in serum-free conditions for 48 h. The media were filtered through a 0.22 µm pore-size membrane filter (Cat. #SLGVR33RS; Merck Millipore, Burlington, MA, USA), and subsequently concentrated using Vivaspin concentrators (Cat. #VS15T12, Sartorius, Göttingen, Germany). EVs were isolated from the media using the ExoEasy Maxi kit, according to the manufacturer’s protocol (Cat. #76064, Qiagen, Hilden, Germany), followed by buffer exchange to phosphate-buffered saline (PBS) using Vivaspin (Cat. #VS04T9, Sartorius, Göttingen, Germany) centrifugal units.

EV size distribution and particle concentration were analyzed using a NanoSight NS300 system (Malvern Panalytical, Malvern, UK). EV samples were diluted in PBS to achieve a particle concentration within the optimal detection range (1 × 10⁷-1 × 10⁹ particles/mL). For each sample, three 60-s videos were captured at camera level 13 and detection threshold 5 under constant temperature control (25 °C). Mean particle size and concentration (particles/mL) were determined from three independent measurements.

Morphology and structural features of EVs were assessed by TEM. A small volume of EV suspension (10 µL) was applied onto a formvar/carbon-coated TEM grid and allowed to adsorb for 2 min. Excess liquid was carefully removed using filter paper. The grid was then stained by adding 10 µL of 2% uranyl acetate solution for 1 min. After removing excess stain with filter paper, the grid was air-dried for 10 min prior to imaging under HT-7700 TEM (Hitachi High-Technologies, Tokyo, Japan).

siRNAs targeting ADAM9 (siADAM9 #1 and #2), PSMA-targeting siRNA (siPSMA), and negative control siRNA (siNC) were chemically synthesized (sequences provided in Supplementary Table 1) by GenePharma (Shanghai, China). siRNAs (10 µM, 1.204 × 10¹¹ molecules per 1 μL) were loaded into EVs using the Exo-Fect siRNA/miRNA Transfection Kit (Cat. #EXFT200A-1; System Biosciences, Palo Alto, USA), according to the manufacturer’s instructions. To eliminate residual unencapsulated siRNA, the reaction mixture was treated with ribonuclease (RNase) at 37 °C for 1 h, and then subjected to EVs isolation as prescribed, and used for subsequent assays.

Total RNA was extracted from EVs using the Total Exosome RNA and Protein Isolation Kit (Thermo Fisher Scientific, Waltham, USA) according to the manufacturer’s instructions. The quantity of encapsulated siRNA was determined by reverse transcription quantitative polymerase chain reaction (RT-qPCR) (primer information shown in Supplementary Table 2) using a standard curve generated from serially diluted synthetic siRNA standards. The absolute copy number of siRNA was calculated based on the standard curve, and normalized to EV particle numbers determined by NTA.

Flow cytometric data were acquired using a BD FACS Canto II flow cytometer (BD Biosciences, New Jersey, USA) and analyzed with FlowJo software (version 10.8.1) (FlowJo, LLC, Ashland, USA). Appropriate gating strategies were applied to exclude debris and doublets.

Surface expression of PSMA was assessed by flow cytometry. Cell suspensions were incubated with an anti-PSMA primary antibody (1 µg/mL; Cat. #GTKUX80149, GenTex, Irvine, USA) or an isotype-matched immunoglobulin G (IgG) control (0.5 µg/mL; Cat. #MBAB002; R&D Systems, Minneapolis, USA), followed by incubation with an Alexa Fluor 488-conjugated anti-mouse IgG secondary antibody (2 µg/mL; Cat. #A21202, Invitrogen, Thermo Fisher Scientific). After washing to remove unbound antibodies, cells were subjected to flow cytometric analysis.

To evaluate surface display of scFv-engineered EVs, flow cytometric analysis was performed using the Exosome-Human CD63 Isolation/Detection Reagent (Cat. #10606D, Thermo Fisher Scientific), following the manufacturer’s protocol with minor modifications. Briefly, EVs were captured using CD63 antibody-coated Dynabeads and washed with isolation buffer [PBS containing 0.1% bovine serum albumin (BSA)]. Bead-bound EVs were incubated with phycoerythrin (PE)-labeled recombinant Protein L (Cat. #RPL-PP2H2; ACRO Biosystems, Newark, USA) at room temperature for 30 min to detect scFv binding. Following incubation, samples were washed, magnetically separated to remove unbound reagents, and resuspended in isolation buffer for analysis.

Unless otherwise specified, prostate cancer cells were treated with EVs at a standardized EV-to-cell ratio of 1 × 10⁴ particles per cell in all cellular experiments. EV particle numbers were determined by NTA before treatment. For siRNA-loaded EVs, the corresponding siRNA dose was estimated based on the experimentally determined siRNA loading per EV. Treatment duration and frequency are specified for each individual assay.

Parental, LAMP2B, or PSMA^scFv EVs were fluorescently labeled using ExoGlow (Cat. #EXOGP100A-1, System Biosciences) or ExoSparkler (Cat. #EX02; Dojindo Laboratories, Kumamoto, Japan) membrane-labeling kits according to the manufacturer’s instructions. Approximately 5 × 10¹⁰ EV particles were used for each labeling reaction. Following labeling, free dye was removed, and the labeled EVs were washed and resuspended in PBS. EV particle concentrations were subsequently re-determined by NTA before use. Labeled EVs (1 × 10⁹ particles) were incubated with 1 × 10⁵ prostate cancer cells for 1 h at room temperature in the dark. After washing to remove unbound EVs, cells were subjected to flow cytometric analysis. EV binding was quantified using FlowJo software based on the fluorescence intensity of cell-associated labeled EVs.

LNCaP (1.5 × 10⁵) and PC3 (2.5 × 10⁵) were seeded on glass coverslips and incubated overnight. Cells were stained with CellBrite Cytoplasmic Membrane Dye (Cat. #30021; Biotium, Fremont, USA), according to the manufacturer’s instructions, followed by incubation with fluorescently labeled parental, LAMP2B, or PSMA^scFv EVs. After washing, cells were fixed with 4% paraformaldehyde and mounted with 4’,6-diamidino-2-phenylindole (DAPI)-containing medium (Cat. #sc-24941; Santa Cruz Biotechnology, Dallas, USA). Images were acquired using a Stellaris confocal microscope (Leica Microsystems, Wetzlar, Germany). EV-associated fluorescence intensity was quantified using Fiji/ImageJ software. Three to four fields of view were analyzed per sample, and fluorescence intensities were quantified and averaged for comparison among EV treatment groups. For assessment of intracellular localization, confocal z-stack images were acquired and reconstructed to generate orthogonal views, allowing EV-associated fluorescence to be evaluated relative to the membrane-defined cellular boundaries.

Protein extracts from cells and EVs were prepared and separated by sodium dodecyl sulfate-polyacrylamide gel electrophoresis (SDS-PAGE). Proteins were transferred onto membranes, which were subsequently blocked and incubated with primary antibodies (typically 1:1,000 dilution) at 4 °C overnight. Membranes were then incubated with appropriate horseradish peroxidase-conjugated secondary antibodies (1:10,000 dilution). Protein bands were detected using an enhanced chemiluminescence detection kit (Cat. # JT96-K002, T-Pro Biotechnology) and imaged using an Amersham Imager 600 system (Cytiva, Marlborough, USA). Detailed information on primary antibodies is provided in Supplementary Table 3.

Total RNA was extracted from cells using NC RNA extraction reagent (Cat. #MRE-N3200; EBL Biotechnology, Taipei, Taiwan) according to the manufacturer’s instructions. Complementary DNA (cDNA) was synthesized from total RNA using the PrimeScript RT reagent kit (Cat. # RR037A, Takara Bio). Quantitative PCR was performed using gene-specific primers (listed in Supplementary Table 4) and the QuantiNova SYBR Green PCR kit (Cat. # 208052, Qiagen) on a LightCycler 96 system (Roche, Basel, Switzerland), following the manufacturers’ protocols. Relative gene expression levels were calculated using the 2^-∆∆Ct method, with HSPCB serving as the internal reference gene for normalization.

Cell migration was assessed using 8-µm pores Transwell inserts (Cat. #MCEP24H48, Merck Millipore), with three biological replicates per condition. The lower chamber was filled with 800 µL of RPMI-1640 supplemented with 1% FBS to serve as a chemoattractant. EV-treated cells (5 × 10⁴) suspended in a 200 µL serum-free medium were seeded into the upper chambers and incubated under standard culture conditions. Following incubation, non-migrated cells on the upper surface of the membrane were removed. Cells that had migrated to the lower surface were fixed with 10% neutral-buffered formalin, stained with crystal violet for 30 min at room temperature, and rinsed with distilled water to remove excess stain. Membranes were air-dried, and migrated cells were imaged in randomly selected fields using a light microscope. For quantitative analysis, retained crystal violet was solubilized in a solution containing distilled water, ethanol, and acetic acid (1:1:2, v/v/v). Absorbance was measured at 590 nm using a microplate reader, and cell migration was quantified based on the absorbance of solubilized crystal violet.

EV-treated cells were seeded into culture insert compartments (Cat. # 80466, Ibidi, Gräfelfing, Germany) at densities of 2.5 × 10⁴ cells/100 µL and maintained under standard culture conditions. After cell attachment, inserts were removed to generate a defined cell-free gap, and wells were gently washed with PBS to remove unattached cells. The resulting cell-free gap or wound area was imaged at 0, 24, and 48 h microscopically. The wound area at each time point was quantified using ImageJ software (National Institutes of Health, USA), and wound closure at 24 and 48 h was calculated relative to the initial wound area at 0.

Cell proliferation was assessed using a WST-8 assay. EV-treated cells were seeded at a density of 5 × 10³ cells per well in 96-well plates and incubated overnight under standard culture conditions. At the indicated time points, 10 µL of WST-8 reagent (Cat. #11644807001; Roche, Basel, Switzerland) was added to each well, followed by incubation for 4 h. Absorbance was measured at 440/630 nm using a microplate reader (CLARIOstar Plus, BMG Labtech, Ortenberg, Germany).

Sphere formation assay was performed under non-adherent conditions using serum-free sphere-forming medium. EV-treated cells were seeded at a density of 2.5 × 10³ cells per well in ultra-low attachment plates and maintained in suspension culture from day 0. EV-siADAM9 treatment was initiated on day 3 and repeated every 3 days thereafter to ensure continuous exposure over 7-14 days. At the endpoint, spheres were imaged for assessment of number and size, and subsequently collected for downstream RNA and protein analyses.

Unless otherwise specified, experiments were performed in triplicate and repeated independently at least three times. Data are presented as mean ± standard error of the mean (SEM) unless otherwise stated. Statistical analyses were conducted using GraphPad Prism software (GraphPad, San Diego, CA, USA). Comparisons between two groups were performed using Student’s t-test. Comparisons involving three or more groups were analyzed using one-way or two-way analysis of variance (ANOVA), as appropriate to the experimental design, followed by multiple-comparison tests as specified in the corresponding figure legends. A P-value < 0.05 was considered statistically significant.

PSMA is highly elevated in advanced and CRPC, making it an attractive target for tumor-selective delivery^[15]. To establish a PSMA-targeted EV platform, we engineered EV-producing cells to express a PSMA-targeting scFv (PSMA-scFv) fused to LAMP2B, enabling surface display of PSMA-scFv on secreted vesicles^[16] [Figure 1A]. The expression of the resulting PSMA-scFv-LAMP2B construct was initially validated in HEK293A cells. Western blot analysis demonstrated robust expression of LAMP2B in both whole-cell lysates and corresponding EV fractions of PSMA-scFv-LAMP2B- and LAMP2B-engineered samples. A distinct band at the expected molecular weight of the fusion protein (~150 kDa) was detected in PSMA-scFv-LAMP2B-transfected samples but not in parental and LAMP2B controls, confirming successful incorporation of PSMA-scFv into EVs [Figure 1B]. MSCs are increasingly recognized for their intrinsic therapeutic properties, including immunomodulatory capacity and tumor-homing potential. We next established stable expression of PSMA-scFv-LAMP2B in MSCs, the intended EV-producing cell source. Consistent with the HEK293A validation, the PSMA-scFv-LAMP2B fusion protein was readily detected in both MSC lysates and EV [Figure 1C] fractions, indicating efficient expression and vesicular incorporation of the engineered construct.

Having established stable expression and EV incorporation of PSMA-scFv, we next evaluated the structural and functional characteristics of the engineered EVs. To confirm EV identity and purity, we performed comprehensive protein characterization of EVs that were derived from parental, LAMP2B-expressing (LAMP2B), and PSMA-scFv-LAMP2B-expressing MSCs in accordance with the Minimal Information for Studies of Extracellular Vesicles (MISEV) 2023 guidelines^[17]. EV marker analysis confirmed the presence of canonical EV proteins, including CD63, CD9, CD81, TSG101, β-actin and HSP70, and the absence of cellular contaminants, such as the endoplasmic reticulum marker calnexin and nuclear marker lamin A/C, consistent with MISEV 2023 guidelines. In addition, the MSC marker CD90 was detected, supporting the cellular origin of EVs, while BSA was not detected in EV fractions, indicating minimal contamination from culture media [Figure 1C]. NTA demonstrated that MSC-derived EVs, including EV-Parental, EV-LAMP2B, and EV-PSMA^scFv, exhibited comparable size distributions, predominantly within the 50-400 nm range, with a peak around ~100 nm [Figure 1D]. TEM further confirmed the characteristic spherical morphology and bilayer membrane structure of EVs across all groups, indicating that genetic modification did not alter vesicle integrity [Figure 1E]. To determine whether PSMA-scFv was correctly displayed on the EV surface, we performed Protein L binding assays [Supplementary Figure 1A]. Flow cytometric analysis revealed a marked increase in PE-labeled Protein L binding in EV-PSMA^scFv compared with EV-parental and EV-LAMP2B controls, as evidenced by a rightward shift in fluorescence intensity and a significantly higher proportion of Protein L-positive events [Figure 1F]. Quantitative flow cytometry [Supplementary Figure 1B] and optical density (OD) [Supplementary Figure 1C] analyses further demonstrated significantly greater PE-Protein L binding in EV-PSMA^scFv than in EV-parental and EV-LAMP2B controls. Given that Protein L specifically binds to the κ light chain of antibody fragments, these findings confirm successful surface localization and proper extracellular orientation of the PSMA-scFv on engineered EVs and siRNA encapsulated EVs. Collectively, these results demonstrate that PSMA-scFv engineering and cargo loading do not alter EV structural integrity while enabling functional surface display of the targeting ligand.

To evaluate targeting specificity across clinically relevant prostate cancer subtypes, we selected cell line models representing distinct stages of disease progression. LNCaP, which is an androgen receptor (AR)-positive and androgen-sensitive cell line originally derived from lymph node metastases of prostate cancer patients^[18], was used as a model of high-risk HSPC. CWR22Rv1 cells, which retain AR signaling but exhibit androgen-independent growth due to constitutively active AR variants^[19], were used to represent AR-positive CRPC. In contrast, PC3 and DU145 cells, which lack functional AR signaling and are androgen-independent, were used as models of AR-negative CRPC.

Flow cytometric analysis first confirmed differential PSMA surface expression among these prostate cancer models. LNCaP cells exhibited high PSMA expression, whereas CWR22Rv1 cells showed moderate expression. In contrast, PSMA expression was undetectable in PC3 and DU145 cells [Figure 2A and B]. We next evaluated whether PSMA-scFv functionalization enhanced EV binding in a PSMA-dependent manner. Baseline signal intensity of labeled EVs was confirmed to be comparable across preparations prior to cell exposure, as indicated by similar absorbance values [Supplementary Figure 2A]. Following incubation with EV-PSMA^scFv, PSMA-expressing LNCaP and CWR22Rv1 cells exhibited significantly increased EV binding compared with EV-Parental and EV-LAMP2B controls [Supplementary Figure 2B]. In contrast, PSMA-negative PC3 and DU145 cells showed no difference in EV binding among the three types of EVs. Consistently, flow cytometry analysis, which independently assessed EV binding at the single-cell level, demonstrated significantly higher populations of EV-positive cells and increased mean fluorescence intensity (MFI) in LNCaP and CWR22Rv1 cells following exposure to EV-PSMA^scFv, whereas no significant enhancement was observed in PC3 or DU145 cells [Figure 2C and D]. Notably, measurable basal EV binding was observed in the PSMA-negative PC3 and DU145 cells across all EV preparations, indicating that PSMA-scFv functionalization does not eliminate the intrinsic PSMA-independent interactions of MSC-derived EVs with tumor cells. The magnitude of this basal EV association varied among prostate cancer cell lines, suggesting cell context-dependent differences in intrinsic MSC-EV interactions. Nevertheless, PSMA-scFv functionalization provides an additional PSMA-dependent enhancement of EV binding, as supported by the relationship between EV binding and relative PSMA surface expression across the prostate cancer cell lines. To directly determine whether EV binding is dependent on PSMA expression, we next employed both loss- and gain-of-function approaches in prostate cancer cell models. Flow cytometric analysis confirmed efficient modulation of PSMA expression. In LNCaP cells, PSMA surface expression was markedly reduced following siRNA-mediated knockdown (siPSMA) compared with non-targeting control (siNC) and mock-transfected cells, as evidenced by decreased fluorescence intensity [Figure 3A]. Conversely, stable expression of PSMA in PC3 cells resulted in elevated PSMA surface expression relative to wild-type controls [Figure 3B]. We next assessed the binding of EV-PSMA^scFv to these PSMA-modified cells. In LNCaP cells, PSMA knockdown significantly impaired EV binding, as demonstrated by reduced EV positive populations and lower MFI compared with mock and siNC controls [Figure 3C and D]. In contrast, PSMA overexpression in PC3 cells significantly enhanced EV binding relative to wild-type cells [Figure 3E and F]. Collectively, these findings demonstrate that binding of PSMA^scFv-engineered EVs is directly dependent on surface PSMA expression, establishing PSMA as a functional determinant of EV targeting specificity.

To determine whether PSMA-dependent binding translates into enhanced cellular uptake, we next evaluated the internalization of engineered EVs in prostate cancer cell models. Intracellular distribution of EVs was assessed by confocal microscopy following incubation with ExoSparkler-labeled EVs. In PSMA-positive LNCaP cells, EV-PSMA^scFv exhibited markedly increased intracellular accumulation with predominantly cytoplasmic localization compared with EV-parental and EV-LAMP2B controls [Figure 4A]. Quantitative analysis demonstrated an approximately two-fold increase in mean EV-associated fluorescence intensity per cell following exposure to EV-PSMA^scFv relative to control EV groups [Figure 4B]. In contrast, PSMA-negative PC3 cells exhibited minimal differences in EV internalization among these three EV preparations, with no significant change in EV-associated fluorescence intensity per cell [Figure 4C and D].

To further establish the role of PSMA in EV uptake, confocal microscopy was performed using complementary PSMA loss- and gain-of-function models. PSMA knockdown in LNCaP cells reduced EV-PSMA^scFv-associated intracellular fluorescence [Supplementary Figure 3A and B], whereas ectopic PSMA expression in PC3 cells enhanced intracellular EV-PSMA^scFv fluorescence compared with the corresponding controls [Supplementary Figure 4A and B]. Confocal z-stack imaging and orthogonal reconstruction further demonstrated EV-associated fluorescence within the cell, supporting intracellular localization of EV-PSMA^scFv [Supplementary Figure 5]. These findings demonstrate that PSMA^scFv engineering enhances EV uptake in a PSMA-dependent manner, linking receptor-specific binding to efficient intracellular delivery in prostate cancer cells.

Having established PSMA-dependent binding and uptake of engineered EVs, we next evaluated their capacity to encapsulate and protect therapeutic siRNA cargo. EV-PSMA^scFv were loaded with siADAM9 (two independent sequences), and encapsulation efficiency was assessed by quantitative PCR analysis. Encapsulation efficiency was assessed by quantitative PCR analysis. The analysis demonstrated that approximately 250-300 siRNA copies per EV particle were encapsulated [Figure 5A], corresponding to an estimated 10% of the input siRNA being successfully encapsulated [Figure 5B]. Moreover, compared with naked siRNA, EV-encapsulated siADAM9 exhibited markedly enhanced resistance to RNase-mediated degradation, as evidenced by significantly lower Cq values following RNase exposure [Figure 5C]. These findings indicate that EV encapsulation effectively protects siRNA cargo from extracellular degradation. To determine whether siRNA loading affects the targeting specificity of EV-PSMA^scFv, we evaluated the binding of siADAM9-loaded EV-PSMA^scFv using the PSMA loss- and gain-of-function models described above. PSMA knockdown significantly reduced the binding of siADAM9-loaded EV-PSMA^scFv in LNCaP cells compared with mock and siNC controls [Supplementary Figure 6A and B]. Conversely, ectopic PSMA expression markedly enhanced EV binding in PC3 cells relative to wild-type (WT) cells [Supplementary Figure 6C and D], indicating that PSMA-dependent targeting is retained following siADAM9 loading. These results demonstrate that PSMA-dependent targeting is retained following siADAM9 loading, indicating that therapeutic cargo incorporation does not compromise the receptor-specific binding of EV-PSMA^scFv.

To further assess the reproducibility of the engineered EV platform, an independently prepared batch of siADAM9-loaded EV-PSMA^scFv was characterized. The immunoblotting confirmed incorporation of the PSMA-scFv-LAMP2B fusion protein [Supplementary Figure 7A]. The EV proteomic characterization was also carried out in the siADAM9 encapsulated samples [Supplementary Figure 7B]. NTA demonstrated comparable particle size characteristics [Supplementary Figure 7C and D], while Protein L binding analysis further demonstrated surface-accessible scFv display, consistent with that observed in the original EV preparation [Supplementary Figure 8A-D]. Together, these findings support the reproducibility of PSMA-scFv incorporation and surface display across independent EV preparations.

We next evaluated whether EV-PSMA^scFv-mediated delivery of siADAM9 resulted in functional gene silencing in recipient prostate cancer cells. Treatment with EV-PSMA^scFv loaded with siADAM9 significantly reduced ADAM9 messenger RNA (mRNA) expression in both LNCaP and CWR22Rv1 cells compared with unloaded EVs, mock-treated cells, and EVs loaded with non-targeting siRNA (siNC) [Figure 5D]. Consistently, immunoblot analysis confirmed substantial suppression of ADAM9 protein expression following treatment with siADAM9-loaded EV-PSMA^scFv in both prostate cancer models [Figure 5E]. Notably, the two independent ADAM9-targeting siRNAs, siADAM9#1 and siADAM9#2 (sequences provided in Supplementary Table 1), exhibited comparable silencing activity, reducing ADAM9 protein expression to similar levels in both LNCaP and CWR22Rv1 cells. EV-PSMA^scFv efficiently encapsulates and protects siRNA cargo while enabling functional intracellular delivery and robust target gene silencing in prostate cancer cells.

To determine whether EV-PSMA^scFv-mediated delivery of siADAM9 affects cell migratory behavior, LNCaP and CWR22Rv1 cells were treated with EV-PSMA^scFv loaded with siADAM9, siNC, or unloaded EVs. Wound healing assays demonstrated that EV-mediated ADAM9 silencing significantly reduced wound closure in both prostate cancer cell lines within 24-48 h, indicating impaired migratory capacity following siADAM9 delivery [Figure 6A and B]. Consistent with these findings, Transwell assays revealed a significant reduction in the number of migrated cells following siADAM9-loaded EV-PSMA^scFv treatment compared with control groups [Figure 6C and D]. In contrast, siNC-loaded and unloaded EVs showed minimal effects on cell migration. To further investigate whether the reduced migratory phenotype was associated with epithelial-mesenchymal transition (EMT) remodeling, we next examined the expression of key EMT- and extracellular matrix-associated markers following treatment with siADAM9-loaded EV-PSMA^scFv. In both LNCaP and CWR22Rv1 cells, EV-PSMA^scFv-mediated siADAM9 silencing increased expression of the epithelial marker E-cadherin while reducing expression of the mesenchymal markers such as vimentin and N-cadherin. In parallel, expression of the matrix remodeling proteins MMP2 and MMP9 was decreased, whereas the matrix metalloproteinase inhibitor TIMP3 was increased following siADAM9 delivery, indicating suppression of extracellular matrix remodeling programs associated with tumor progression and migratory potential [Figure 6E and F].

Given the critical role of stemness-associated programs in prostate cancer progression, including their contributions to cellular plasticity, metastatic potential, and therapeutic resistance, we next investigated the effects of EV-mediated ADAM9 silencing on these aggressive tumor phenotypes. Cell proliferation assays first demonstrated that treatment with siADAM9-loaded EV-PSMA^scFv significantly reduced cell growth in both LNCaP and CWR22Rv1 cells under monolayer culture conditions compared with control treatments [Supplementary Figure 9]. Notably, significant suppression of cell proliferation became evident at later time points, suggesting that the early reduction in wound closure was unlikely to be attributable solely to reduced proliferation. To further evaluate tumor stemness-associated properties, sphere formation assays were performed under three-dimensional culture conditions. Treatment with siADAM9-loaded EV-PSMA^scFv significantly reduced both the number and size of tumor spheres in LNCaP and CWR22Rv1 cells, indicating impaired sphere-forming and tumor-propagating capacity following ADAM9 silencing [Figure 7A and B]. Consistent with these phenotypic changes, expression of key stemness-associated regulators, including SOX2, OCT4 and BMI1, was markedly reduced at both the mRNA and protein levels in response to siADAM9 delivery compared with control groups [Figure 7C-F]. Notably, efficient suppression of ADAM9 expression was maintained under sphere culture conditions, indicating that EV-PSMA^scFv-mediated siRNA delivery remains effective within structurally complex three-dimensional tumor-like environments. Collectively, these findings demonstrate that EV-PSMA^scFv-mediated ADAM9 silencing suppresses aggressive tumor phenotypes, including migration and tumor propagation, while reversing EMT- and stemness-associated molecular programs in both hormone-sensitive and CRPC cells.

In current clinical practice, metastatic HSPC (HSPC) is commonly treated with ADT (ADT) in combination with androgen receptor pathway inhibitors; however, therapeutic resistance frequently develops. Based on our previous findings demonstrating that the ADAM9/UBN2/AKR1C3 axis contributes to the transition of androgen-dependent prostate cancer cells toward an androgen-independent state^[7], we next investigated whether EV-mediated ADAM9 silencing could enhance responsiveness to androgen-targeted therapy in LNCaP cells. Cell viability assays demonstrated that treatment with siADAM9-loaded EV-PSMA^scFv further reduced cell survival compared with ADT monotherapy or combined ADT and abiraterone treatment alone [Figure 8A]. Correspondingly, although ADT alone or in combination with abiraterone induced marked upregulation of AKR1C3 relative to pre-treatment levels, EV-PSMA^scFv-mediated ADAM9 silencing significantly suppressed expression of ADAM9, UBN2, and AKR1C3 compared with siNC-loaded controls under androgen-targeted treatment conditions [Figure 8B and C]. In addition, the increased full-length androgen receptor (AR-FL) protein expression observed following androgen-targeted therapy was attenuated by siADAM9-loaded EV-PSMA^scFv, whereas androgen receptor splice variant 7 (AR-V7) remained undetectable under all conditions [Figure 8C].

To further evaluate therapeutic relevance in CRPC, where AR pathway inhibitors are commonly used as first-line systemic therapy, we next examined the effects of EV-PSMA^scFv-mediated ADAM9 silencing in CWR22Rv1 cells treated with abiraterone alone or combined ADT and abiraterone. Similar to the LNCaP model, addition of siADAM9-loaded EV-PSMA^scFv significantly reduced cell survival under androgen-targeted treatment conditions [Figure 8D]. Both siADAM9#1 and siADAM9#2 produced comparable enhancement of the treatment response, with no consistent difference in therapeutic effects between the two siRNAs. However, in contrast to the HSPC model, EV-mediated ADAM9 silencing in CWR22Rv1 cells exerted minimal effects on UBN2 and AKR1C3 expression despite partial induction of AKR1C3 following androgen-targeted therapy [Figure 8E and F]. Furthermore, AR-FL and AR-V7 protein levels remained elevated across all treatment conditions [Figure 8F], suggesting that ADAM9-mediated regulation of the UBN2/AKR1C3 axis preferentially operates within an androgen-responsive cellular context and is not maintained in androgen-independent CRPC cells. Notably, siADAM9 treatment was associated with increased p21 expression in both LNCaP and CWR22Rv1 cells, suggesting that modulation of cell cycle-associated pathways may represent a broader response to ADAM9 suppression rather than a mechanism specific to androgen-independent prostate cancer [Figure 8C and F]. These findings suggest that EV-PSMA^scFv-mediated ADAM9 silencing may enhance therapeutic responsiveness through distinct context-dependent mechanisms in HSPC and CRPC.