VesicleVoyager: In Vivo Selection of Surface Displayed Proteins That Direct Extracellular Vesicles to Tissue‐Specific Targets

ABSTRACT

The development of technologies for screening proteins that bind to specific tissues in vivo and facilitate delivery of large cargos remains challenging, with most approaches limited to cell culture systems that often yield clinically irrelevant hits. To overcome this limitation, we developed a novel molecular screening platform using an extracellular vesicle (EV) display library. EVs are natural molecular carriers capable of delivering diverse cargos, which can be engineered to enhance specificity and targeting through surface modifications. We constructed an EV‐display library presenting monobody repertoires on EV surfaces, with genetic cargo inside the EVs corresponding to the displayed proteins. These libraries were screened for tissue specific delivery through serial passage in mice via sequential intravenous administration in and recovery of tissue‐selected EVs and amplification of their encapsulated monobody genes at each passage. Our results demonstrated successful selection of tissue‐specific targeting proteins, as revealed by fluorescence and bioluminescence imaging followed by DNA sequencing. To understand the stochastic relationship between displayed proteins and packaged genes, we developed a mathematical analysis that revealed the complex selection dynamics and demonstrated successful enrichment despite the imperfect correlation between phenotype and genotype. This EV‐based monobody screening approach, combined with mathematical modelling, is a significant advancement in targeted drug delivery by leveraging the natural capabilities of EVs with the selection of targeting proteins in a physiologically relevant environment.

1. Introduction

The ability to precisely target therapeutic cargo to specific tissues and cells remains one of the greatest challenges in drug delivery. Despite decades of research, less than 2% of systemically administered drugs are estimated to reach their intended targets, highlighting the critical need for innovative targeting solutions (Danesh‐Meyer et al. 2016). While numerous approaches exist, we present a platform that leverages the intrinsic properties of extracellular vesicles (EVs) to create an alternative to traditional display technologies that is performed in vivo. We sought to develop our VesicleVoyager system as an in vivo selection strategy to identify targeting molecules in the context of living organ systems with an intact circulatory system providing the selection pressure. As such, targeting molecules are screened in the native context in which they will be used for drug delivery. This is a fundamental shift from conventional screening methods and addresses the critical limitations of other approaches through the strategic integration of extracellular vesicle biology and directed evolution.

Small membrane‐bound EVs represent nature's own delivery system that is capable of transporting bioactive molecules between cells in vivo and delivering functional cargo. (Herrmann et al. 2021) These natural nanoparticles shuttle proteins, nucleic acids, and lipids between cells, serving as critical mediators of intercellular communication (Raposo and Stahl 2019). The molecular signature of EVs have been used as a diagnostic indicators, and EVs have been manipulated for the purpose of developing therapeutic delivery tools (Ciferri et al. 2021; Murphy et al. 2019). Beyond their diagnostic potential as molecular fingerprints of health and disease, EVs offer the potential of transformative therapeutic delivery capabilities that synthetic alternatives cannot, at this time, match (Ciferri et al. 2021; Murphy et al. 2019).

EV engineering has largely involved the addition of single, or small numbers of, surface modifications and the addition of therapeutic cargo (Komuro et al. 2022). Importantly, we have demonstrated that DNA transfection of cells produces EVs that display plasmid‐encoded surface proteins and maintain association with the exogenously introduced plasmid DNA (pDNA) (Kanada et al. 2015; Komuro et al. 2021; Kawai‐Harada et al. 2024). Much like EV encapsulated RNA, the encapsulated DNA is protected from degradation, which is a critical feature for effective delivery vehicles (Kanada et al. 2015; Komuro et al. 2021; Kawai‐Harada et al. 2024). This was demonstrated with size exclusion chromatography in which the plasmid DNA co‐purified with EVs, and this co‐purified DNA was protected from DNase digestion (Kawai‐Harada et al. 2024). The size of supercoiled plasmid DNA has been proposed to be smaller than the size of EVs (Zakharova et al. 2002; Smrek et al. 2021; Irobalieva et al. 2015; Li et al. 2016), this and the DNase protection data indicates encapsulation of nucleic acids. These findings lead to the possibility that EV‐display libraries could be created, similar to phage display libraries, with the distinct advantages of being compatible with mammalian systems and being screened in vivo with functional readouts.

Phage display technology, pioneered by George P. Smith over three decades ago, remains a cornerstone in discovering biological knowledge and developing pharmaceutical interventions (Smith 1985; Alfaleh et al. 2020; Wu et al. 2016). While alternative screening platforms (such as yeast display, mammalian cell display, bacterial display and ribosome display) have emerged, phage display remains the most widely used method for screening antibodies and peptides because of its ease of use, low cost and high complexity (Frenzel et al. 2016; Jaroszewicz et al. 2022). Phage display screen leverages the unique characteristic of bacteriophages, bacteria‐infecting viruses, to selectively amplify antigen‐specific antibodies by displaying foreign amino acids on their surface while packaging the corresponding coding gene. In addition, bacteriophage replicate to large number and in large quantities (Smith 1985). Despite its pioneering status and widespread use, phage display faces limitations in mammalian systems. While invaluable for in vitro screening, in vivo screens for cell and tissue targeting remain challenging due to unusual clearance mechanisms and concerns regarding immunogenicity and toxicity (Asar et al. 2024; Saw and Song 2019; Holman et al. 2017). The complexity of mammalian tissues (with their structural variations, diverse post‐translational modifications and intricate molecular interactions) presents challenges that conventional display technologies cannot readily overcome (Frenzel et al. 2016).

EVs are a heterogeneous population of lipid‐bound nano‐size particles secreted by all cell types that offer distinct advantages over conventional library screens and position them for development as a superior screening tool (Yáñez‐Mó et al. 2015; Valadi et al. 2007; Raposo et al. 1996). They mediate intercellular communication by transferring genetic materials, lipids and proteins, with diverse roles documented in immune regulation (Théry et al. 2002), antigen presentation (Théry et al. 2002), tissue repair and regeneration (Lai et al. 2013), cancer (Xu et al. 2018) and neurodegenerative diseases (Hill 2019). This role in communication suggests a number of functional readouts that could be built into EV display screens where effective delivery of specific types of cargo results in a detectable signal. EVs demonstrate unique kinetics, biodistribution patterns, and functionality and clearance pathways that synthetic lipids and nanocarriers cannot replicate (Ciferri et al. 2021; de Freitas et al. 2021). Most importantly, they avoid the toxicity and immunogenicity associated with lipid‐based nanocarriers and viral delivery systems (Herrmann et al. 2021; EL Andaloussi et al. 2013). While some EVs naturally exhibit tissue tropism, controlling and enhancing this targeting specificity has remained challenging. By adding affinity molecules (such as antibody mimetics or peptides) to the EV surface, cellular affinity and targeted delivery have been altered both in cultured cells and in vivo (Murphy et al. 2019; Komuro et al. 2021; Komuro et al. 2022; Kooijmans et al. 2016; Ye et al. 2018; Wang et al. 2017).

Several approaches have been developed to engineer EVs for targeted delivery, each with distinct advantages and limitations. Genetic engineering strategies have focused on fusing targeting peptide or proteins domains to EV‐enriched membrane proteins such as Lamp2b or tetraspanins including CD63, enabling display of targeting moieties on EV surfaces (Kooijmans et al. 2016; Ohno et al. 2013). For example, Lamp2b‐RVG fusions have been used to target EVs to neuronal tissues by exploiting the rabies virus glycoprotein's affinity for nicotinic acetylcholine receptors (Alvarez‐Erviti et al. 2011). Similarly, transmembrane domain fusions using platelet‐derived growth factor receptor (PDGFR) transmembrane regions have enabled surface display of targeting antibodies and nanobodies (Hung and Leonard 2015). Chemical conjugation approaches have employed post‐isolation modification strategies, including click chemistry and bioorthogonal reactions, to attach targeting molecules to EV surfaces after isolation (Smyth et al. 2014; Nakase et al. 2015). While these methods offer flexibility in targeting molecule selection, they often suffer from reduced EV integrity, altered biodistribution properties and challenges in maintaining genotype–phenotype relationships necessary for iterative optimization (Luan et al. 2017). Electroporation‐based loading has been used to introduce both therapeutic cargo and targeting molecules, though this approach can compromise EV membrane integrity and reduce targeting efficiency (Haney et al. 2015; Fuhrmann et al. 2015).

Previous EV‐display studies have focused primarily on displaying pre‐selected targeting molecules rather than screening combinatorial libraries for novel binders (Jang et al. 2013; Tian et al. 2014; Wiklander et al. 2024). Although these strategies demonstrate that EVs can be engineered for targeting, they do not provide a platform for functional in vivo selection. To our knowledge, systematic screening of combinatorial EV libraries in vivo has not been reported. In contrast, the VesicleVoyager platform enables iterative in vivo selection of surface‐displayed monobody libraries while maintaining genotype‐phenotype linkage, allowing the recovery of EVs that successfully traffic to specific tissues under physiological conditions. This inherently accounts for factors such as circulation dynamics, tissue barriers and protein corona formation, which are absent in conventional in vitro approaches. By providing scalable functional screening in a native biological context, VesicleVoyager complements existing EV display methods and addresses a critical gap in the field.

Despite significant advances in EV research, a systematic approach for screening and selecting molecules that can direct EVs to specific cellular targets in vivo has not been described, hindering the development of precisely targeted therapeutic delivery systems. Building on our previous work, we established the technological foundation for such a strategy that is possible due to two key innovations: (1) the development of SLiCE technology for creating recombinant DNA libraries in mammalian cells with exceptional efficiency (Kawai‐Harada et al. 2024), and (2) the engineering of EVs displaying monobodies with nanomolar affinity through lactadherin fusion constructs, demonstrating enhanced cellular targeting capabilities (Komuro et al. 2022).

The VesicleVoyager platform is based on engineering EVs to display monobodies (small, single‐domain binding proteins) on their surface while packaging the encoding DNA within. This creates a self‐contained screening system where the phenotype (surface‐displayed monobody) and genotype (encoding DNA) remain physically linked. Through iterative rounds of selection and amplification, we can identify specific monobodies that direct EVs to desired tissues, enabling the enrichment of EV populations carrying specific targeting molecules. Using both fluorescence and bioluminescence readouts of selection with sequence validation, we demonstrated successful in vitro and in vivo screening of these EV‐displayed monobodies and identification of unique tissue‐targeting monobodies.

The in‐situ optimization approach employed by VesicleVoyager offers several critical advantages over conventional strategies that separately optimize binding proteins and subsequently load them onto EVs. First, membrane‐anchored monobodies exhibit fundamentally different binding kinetics, conformational stability and avidity effects compared to their soluble counterparts (Xie et al. 2014; Hu et al. 2013). The transmembrane anchor constrains protein orientation and creates multivalent binding interactions that cannot be recapitulated by post‐production loading of pre‐optimized soluble binders (Oostindie et al. 2022; Vauquelin and Charlton 2013). Second, the unique EV membrane microenvironment, enriched in cholesterol, sphingomyelin and specific lipid rafts, influences protein folding and stability. Monobodies optimized within this native context maintain superior binding properties during EV storage, circulation and target engagement (Haraszti et al. 2016; Skotland et al. 2017). Third, our approach enables simultaneous co‐optimization of both target binding affinity and EV delivery characteristics, including circulation stability, tissue‐specific uptake and cargo delivery efficiency, a multi‐parameter optimization impossible through sequential approaches. Finally, in situ optimization maintains perfect genotype–phenotype linkage throughout iterative selection cycles, essential for recovering genetic information from selected EVs and enabling the evolutionary optimization process.

To understand the stochastic relationship between displayed monobodies and their encoding DNA within EVs, we developed a mathematical framework to analyze the selection dynamics. This analysis provides insights into the efficiency of each selection step and explains how successful targeting molecule enrichment occurs despite the imperfect correlation between phenotype and genotype. Here we detail the development, validation, applications, and mathematical modelling of this novel platform. This innovative approach addresses a critical need for targeted delivery systems that can navigate the complexities of mammalian biology, and has transformative potential for drug discovery, diagnostics, and personalized medicine applications where precise tissue targeting is essential for therapeutic success.

2. Materials and Methods

2.1. Reporter Construct

The cloning methods and the backbone constructs were adopted from our previous work (Murphy et al. 2019). Synthetic double‐stranded DNA coding for NanoLuc (Gaspar et al. 2021), CD63 (Gene bank accession number: NP_001771) and ThermoLuc (Koksharov and Ugarova 2011) were purchased from Twist Bioscience to generate pcS‐NanoLuc‐C1C2 and pcS‐ThermoLuc‐CD63 (Komuro et al. 2022, 2021). Briefly, we amplified the backbone sequence from pcS‐mCherry‐C1C2 (Addgene #178425) and inserted the codon humanized NanoLuc, CD63 and ThermoLuc fragments.

2.2. Library DNA Preparation

The methods and the backbone constructs were adopted from our previous work (Komuro et al. 2022, 2021). All primer sequences used in this study are listed in Table S1. Briefly, the library backbone was amplified from template pcS‐RDG‐C1C2 (Addgene #200163) using primers Lib‐BB‐F and Lib‐BB‐R, followed by the restriction digest by DpnI, to eliminate template DNA. Monobody library fragments coding for variable loop regions from Hydrophilic Fibronectin Library second generation (Woldring et al. 2015) containing 44 bp at the 3′‐end and 45 bp overlap at the 5′‐end were amplified using primers Lib‐IN‐F and Lib‐IN‐R (Woldring et al. 2015) and fused to the backbone using Seamless cloning Ligation Cell Extract (SLiCE) method (Komuro et al. 2022, 2021). Following the cleanup using QIAquick PCR purification kit (QIAGEN) and the concentration measurement by the Qubit dsDNA BR Kit (Invitrogen), the assembled DNA was electroporated into electrocompetent E. coli cells (NEB) and pre‐cultured at 37°C for 1 h without antibiotics, then cultured in the Lauria Broth (LB) containing 100 µg/mL ampicillin for 8 h at 37°C in a shaking incubator. The library DNA was extracted using the Midiprep kit (QIAGEN), and concentration was determined by NanoDrop (Thermo Fisher Scientific).

2.3. Cell Culture and Treatment

Cell lines were obtained from American Type Culture Collection (ATCC) and routinely tested for mycoplasma contamination. Lines included: HEK293T (Human Embryonic Kidney cell line), A431 (Human carcinoma cell line) and MCF‐7 (human breast cancer cell line). Cells were cultured in high‐glucose DMEM (Gibco) supplemented with 100U/mL penicillin, 100 µg/mL streptomycin and 10% (v/v) fetal bovine serum (FBS, Gibco), and maintained in a humidified incubator with 5% CO₂ at 37°C.

For EV production, HEK293T cells were seeded at 1.5 × 10⁶ in 10 cm tissue culture dishes 24 h prior to transfection. For transfection, 10 µg DNA was mixed with in‐house PEI (polyethylenimine) reagent prepared from PEI (Sigma 408727) at a concentration of 1 mg/mL [pH7.0] at DNA:PEI ratio of 1:2.5 (µg:µL) in non‐supplemented DMEM, pulse‐vortexed for 30 s, incubated at room temperature for 10 min and added to the cells (Komuro et al. 2021). After 24 h of incubation with 5% CO₂ at 37°C, cells were rinsed once with PBS, and culture media was replaced with 20 mL of DMEM supplemented with Insulin‐Transferrin‐Selenium (ITS) (Corning), 100 U/mL penicillin and 100 µg/mL streptomycin (conditioned media) and incubated for another 24 h for library EV generation. EVs were co‐labelled with imaging molecules by co‐transfecting 5 µg of monobody‐display plasmid and 5 µg of imaging plasmid (pcS‐NanoLuc‐C1C2, pcS‐mCherry‐C1C2, pcS‐ThermoLuc‐CD63) per dish.

2.4. EV Isolation

EVs were purified from conditioned media by differential centrifugation. Briefly, culture media was centrifuged at 600 × g for 30 min to remove cells and cellular debris, and the supernatant was further centrifuged at 2000 × g for 30 min to remove apoptotic bodies. The supernatant was then ultracentrifuged in PET Thin‐Walled ultracentrifuge tubes (Thermo Scientific 75000471) at 12,000 × g with a Sorvall WX+ Ultracentrifuge equipped with an AH‐629 rotor (k factor = 242.0) for 90 min at 4°C to pellet the EVs. The pellet containing EVs was resuspended in 100 µL EV storage buffer (Kawai‐Harada et al. 2023). EVs were also isolated by Tangential Flow Filtration. The supernatant was filtered by MICROKROS 20CM 0.65UM MPES (Repligen, C06‐E65U‐07‐S), and concentrated by MICROKROS 20CM 0.05UM PS (Repligen, C02‐S05U‐05‐S) (Kawai‐Harada et al. 2024). Concentrated EVs were rinsed with EV storage buffer and collected in 1 mL of EV storage buffer (Kawai‐Harada et al. 2023).

2.5. Nanoparticle Tracking Analysis (NTA)

The particle size and concentration were measured using a ZetaView (Particle Metrix) Nanoparticle Tracking Analyzer following the manufacturer's instruction. The following parameters were used for measurement: (Post Acquisition parameters (Min brightness: 22, Max area: 800, Min area: 10, Tracelength: 12, nm/Class: 30, and Classes/Decade: 64)) and Camera control (Sensitivity: 85, Shutter: 250, and Frame Rate: 30)). EVs were diluted in PBS between 20‐ and 200‐fold to obtain a concentration within the recommended measurement range (0.5 × 10⁵ to 10¹⁰ per mL).

2.6. DNase I Treatment of EVs

A total of 5 µL of Monobody Library EVs were incubated at room temperature for 15 min with 1 U of DNase I (Zymo Research) and DNA Digestion Buffer. The plasmid DNA was then isolated from the EVs using Qiamp Miniprep kits and quantified by qPCR.

2.7. Quantitative Real‐Time Polymerase Chain Reaction (qPCR)

qPCR was performed using Dream Taq DNA polymerase (ThermoFisher). Each reaction contained 200 µM dNTPs, 500 nM each of forward/reverse primer, 400 nM probe (Table S1), 0.5 U DreamTaq DNA polymerase, 1x Dream Taq buffer A and 1 µL sample DNA in a total reaction volume of 10 µL using CFX96 Touch Real‐Time PCR Detection System (BIO‐RAD). The PCR amplification cycle was as follows: 95°C for 2 min; 40 cycles of 95°C for 20 s, 65°C for 30 s. The pDNA copy number were determined by absolute quantification using the standard curve method, and the copy number of EV‐encapsulated pDNA per vesicles was calculated based on nanoparticle tracking analysis (NTA) and qPCR results.

2.8. Western Blotting

EVs were denatured at 70°C for 10 min in 1x NuPAGE LDS Sample Buffer (Thermo Fisher Scientific), separated on 4%–20% Mini‐PROTEAN TGX Precast Protein Gels (BioRad), and transferred to Polyvinylidene fluoride or polyvinylidene difluoride (PVDF) membranes using CAPS‐based transfer buffer. Membranes were blocked with EveryBlot Blocking Buffer (BioRad) for 2 h and then incubated with primary antibodies at 4°C overnight. Following three washes with TBS containing 0.1% Tween 20 (TBST), membranes were incubated with horseradish peroxidase‐conjugated secondary antibodies for 2 h at room temperature. After three additional TBST washes, protein bands were visualized using SuperSignal West Pico PLUS chemiluminescent substrate (Thermo Scientific) and imaged with ChemiDoc Imaging System (BioRad). Primary antibodies used: anti‐HA (Sigma Aldrich, H3663, RRID:AB_262051), anti‐CD63 (Thermo Fisher, 10628D, RRID:AB_2532983), anti‐ALIX (Proteintech, 12422‐1‐AP, RRID:AB_2162467), TSG101 (Abcam, ab125011, RRID:AB_10974262), Calnexin (Abcam, ab133615, RRID:AB_2864299) and GAPDH (Cell Signaling Technology, #2118, RRID:AB_561053). Secondary antibodies used: Goat anti‐Mouse IgG (H+L) Highly Cross‐Adsorbed Secondary Antibody, HRP (Invitrogen, A16078, RRID:AB_2534751) and Goat anti‐Rabbit IgG (H+L) Highly Cross‐Adsorbed Secondary Antibody, HRP (Cell Signaling Technology, A16110, RRID:AB_2534783).

2.9. In Vitro EV Library Screening

For in vitro screening, A431 cells were seeded at 0.3 × 10⁶ cells/well in 6‐well plates 24 h prior to EV treatment. Cells were treated with 2.0 × 10⁷ library EVs in 2 mL media for 30 min at 37°C. Following PBS washing to remove residual EVs, cells were harvested using trypsin. Plasmid DNA was isolated using a modified protocol for plasmid isolation from organ homogenates using QIAprep Spin Miniprep Kit (Isolation of plasmid DNA 1995), and used as template for the next round of library DNA preparation. This screening process was repeated for five rounds to enrich targeting monobody sequences.

2.10. In Vitro Bioluminescence Assay

A431 cells were seeded at 0.02 × 10⁶ cells/well in 96‐well plates (UV‐Star Microplate, 96 well, COC, F‐Bottom (Chimney Well), uClear, Clear; Greiner Bio‐one) 24 h prior to EV treatment. For the assay, 5 × 10⁶ of NanoLuc co‐labelled EVs (generated by co‐transfection with EV‐displayed NanoLuc constructs) were added to wells in triplicate. After incubation at 37°C, cells were washed twice with PBS to remove unbound EVs. A total of 50 µL of 1 µg/mL Coelenterazine‐H (CTZ; Regis Technologies) was added to each well immediately before imaging. Luminescence was recorded using an in vivo imaging system (IVIS; Spectrum Perkin Elmer) and the particle numbers emitting equal amounts of luminescence/radiance (photons/s/cm²/sr) were calculated.

2.11. Confocal Microscopy

Co‐labelled engineered EVs (eEVs) were prepared by co‐transfection as described above. A total of 1 × 10⁵ each of A431 and MCF‐7 cells were mixed and seeded in 8‐well chamber slide (0030742036, Eppendorf, Germany) 24 h before EV treatment. These co‐cultured cells were incubated with 2.0 × 10⁷ monobody‐mCherry co‐labelled eEVs for 10 min, followed by three PBS washes to remove unbound EVs. Cells were fixed with 4% PFA at room temperature for 15 min, washed with PBS containing 0.1% Tween20 three times, and blocked with blocking buffer (1% BSA in PBS) for 60 min. Cells were then incubated with primary antibody (CST 4267T, RRID:AB_2246311, Cell Signaling Technology, Danvers, MA, USA) in a humidified chamber for 60 min at room temperature. After three 5‐min PBS washes, cells were incubated with secondary antibody (CST 4412, RRID:AB_1904025, Cell Signaling Technology, Danvers, MA, USA) for 1 h at room temperature in the dark. After three additional PBS washes, cells were incubated with DAPI for 10 min at room temperature. Slide were mounted with mounting medium (00‐4958‐02 Fisher Scientific, USA) and fluorescence images were taken at 60× magnification using confocal laser scanning microscopy (A1 HD25/A1R HD25 confocal microscope, Nikon, Japan).

2.12. Binding Specificity Quantification

Confocal microscopy images were analyzed by manual cell counting to assess binding selectivity. EGFR‐positive cells (bright green fluorescence) and EGFR‐negative cells (minimal/no green fluorescence) were identified and scored for the presence of co‐localized mCherry‐positive EVs (red fluorescence). A minimum of 2–3 representative fields per condition were analyzed, with approximately 15–25 cells counted per field. Binding frequency was calculated as the percentage of cells showing clear EV association within each population.

2.13. Super Resolution Microscopy

Isolated EVs were analyzed with the EV Profiler V2 Kit for Nanoimager (ONI) according to the manufacturer's protocol. Super‐resolution images were analyzed by batch analysis of Clustering & Counting using CODI software (https://alto.codi.bio/). A cluster was considered positive as a single EV particle when more than 3 individual antibody signals were detected in the same channel within the detected cluster. For each EV cluster, the HA signal number was quantified, and the number of monobody molecules per individual EV was calculated from EV cluster counting data, as previously described (Xu et al. 2024). Three fields of view were acquired from each sample for statistical analysis.

2.14. Next‐Generation Sequencing

The sample for next‐generation sequencing (Illumina MiSeq) were prepared by PCR amplification using a primer pair (CS1‐LibHA‐F, CS2‐G4S‐R) with sequencing indices. PCR products were quantified using Qubit before sequencing. All samples were normalized to the same concentration, and agarose gel electrophoresis confirmed product size. Sequencing was performed at the MSU Genomics Core facility using MiSeq Reagent Kit v3 for 250 bp paired‐end reads. The generated FASTQ files were extracted, processed, and clustered by sequence similarity using custom software (Woldring et al. 2016).

2.15. In Vivo EV Library Screening

In this study, 8‐ to 12‐week‐old female Balb/cJ mice from Jackson Laboratories were used for animal experiments. Animals were housed in the University Laboratory Animal Resources Facility, and all procedures were performed with approval from the Institutional Animal Care and Use Committee of Michigan State University. Approximately 1 × 10⁹ monobody library EVs in EV storage buffer (Kawai‐Harada et al. 2023) were injected into mice intravenously (IV). One hour after administration, mice were sacrificed, and visceral organs (heart, lung, liver, kidney, pancreas and spleen) were excised and homogenized using Bulk Ceramic Beads 2.8 mm (Fisher Scientific) and BeadBug 6 Microtube Homogenizer (Benchmark Scientific). Plasmid DNA was isolated from organ homogenates using QIAprep Spin Miniprep Kit using modified protocols for plasmid isolation from mammalian cells (Komuro et al. 2021; Isolation of plasmid DNA 1995), and used as template for the next round of library preparation. This screening process was repeated for five rounds to enrich targeting monobody sequences. The enriched variants were re‐cloned into the EV display construct for individual monobody characterization.

2.16. In Vivo and Ex Vivo Imaging

To quantify relative detection levels in ex vivo studies, bioluminescence per 1 × 10⁸ particles was measured for each EV sample. Approximately 1 × 10⁹ monobody EVs co‐labelled with NanoLuc were injected intravenous into Balb/c mice (n = 3). One hour after administration, mice received an intraperitoneal injection of CTZ (10 µg/g), 10 min prior to bioluminescence imaging (BLI) using an In Vivo Imaging System (IVIS; Revvity, previously PerkinElmer). Mice were anesthetized with isoflurane and imaged intact. Following in vivo imaging, the mice were sacrificed and visceral organs (heart, lungs, liver, kidneys, pancreas and spleen) were dissected and imaged using an IVIS. Bioluminescence in each organ was quantified, and the amount of detection was normalized as the relative detection value to the initially introduced bioluminescent material. For in vivo imaging with ThermoLuc, approximately 2 × 10¹⁰ Monobody EVs co‐labelled with ThermoLuc were injected intravenous into Balb/c mice (n = 3).

2.17. Stochastic Analysis of the Selection Process

To analyze the relationship between EV surface‐displayed monobodies and their packaged plasmid DNA, we developed a mathematical framework to understand the selection dynamics. Key parameters were estimated from experimental data as follows:

System parameters:

• Monobody display efficiency (p_display ) was calculated from super‐resolution microscopy analysis, showing approximately 80% of EVs displayed monobodies

• DNA packaging efficiency (p_packaging ) was derived from DNase‐resistant plasmid DNA quantification, with an average of 1.5 DNA copies per EV

• Binding probabilities (p_binding ) were estimated from time‐course experiments with A431 cells

• DNA recovery efficiency (p_recovery ) was calculated from qPCR quantification of DNA recovered from target organs

Genotype–phenotype correlation analysis: Each EV displays approximately 20 monobody molecules but contains an average of only 1.5 copies of plasmid DNA. Under conservative assumptions about library complexity and protein distribution, the probability of strong correlation between displayed monobodies and packaged DNA in individual EVs is low. However, even modest correlations can provide selective advantages during iterative screening.

Enrichment modelling: For enrichment simulations, we used a competitive binding model that incorporated selection pressure and initial variant frequency. For a monobody variant with initial frequency f ₀, the frequency after multiple rounds was modelled using:

$$fi=\frac{f_{i-1}·e_{factor}}{1+f_{i-1}·(e_{factor}-1)}$$

where $e_{factor}$ represents the enrichment advantage of the variant over background sequences.

2.18. Statistical Analysis

All experiments were performed with a minimum of n = 3 biological replicates unless otherwise specified. Biological replicates represent independent cell cultures, EV preparations or animal cohorts performed on different days. Technical replicates represent multiple measurements from the same biological sample and were performed in triplicate for in vitro assays.

Sample sizes:

• ‐In vitro binding assays: n = 3 biological replicates with technical triplicates
• ‐In vivo EV screening: n = 3 mice per group across multiple independent experiments
• ‐Confocal microscopy: minimum five fields per experiment, >100 cells analyzed per condition
• ‐qPCR analysis: technical triplicates for each biological replicate

Statistical tests:

• ‐Two‐group comparisons: unpaired t‐tests (normally distributed data) or Mann–Whitney U tests (non‐normal distribution)
• ‐Multiple group comparisons: one‐way ANOVA followed by Tukey's post‐hoc test
• ‐Time course data: two‐way ANOVA with Bonferroni correction for multiple comparisons
• ‐Normality was assessed using Shapiro–Wilk tests
• ‐Equal variance was tested using F‐tests

Data presentation: Data are presented as mean ± standard error of the mean (SEM). Individual data points are shown where possible. Statistical significance was set at p < 0.05 (*p ≤ 0.05, **p ≤ 0.01, ***p ≤ 0.001, ****p ≤ 0.0001).

All statistical analyses were performed using GraphPad Prism version 9.0 (GraphPad Software, La Jolla, CA).

3. Results

3.1. EV‐Based Monobody Display Screening Strategy and Design

Our EV‐monobody screening platform design builds on the method of decorating EV surfaces using EV‐surface display constructs (Komuro et al. 2022, 2021). We used a yeast monobody library (Woldring et al. 2015) to generate the DNA for an EV‐display monobody library with a diversity of 4.2 × 10⁹. Negative depletion magnetic bead sorting was done against streptavidin‐coated beads removed non‐selective binders (approximately 1 × 10⁷ to 1 × 10⁸ monobodies). DNA from the remaining 10⁹ monobody variants was cloned into the EV display vector using SLiCE, transfected into HEK293 cells, and EVs were recovered as the monobody‐displaying library. After isolating the EV‐monobody library from conditioned media, we quantified vesicle numbers and assessed plasmid DNA loading prior to screening in cultured cells or animals.

For repeated selection, cells or organs were harvested, DNA extracted, monobody‐coding regions amplified by PCR, and re‐cloned into the EV‐display backbone. The process was repeated 5 times to ensure co‐selection of enriched monobodies and their encoding DNA (Figure 1).

FIGURE 1.

Schematic illustration of monobody‐EV library screening strategy. (A) Overview of EV‐based monobody screening processes, involving monobody library pDNA generation, pDNA transfection to EV donor cells (HEK293T), EV isolation from the cell culture media, EV treatment or administration, DNA extraction from cell or organ, monobody amplification and re‐cloning to generate enriched monobody library pool. This image was created with BioRender.com. (B) Structural model of the EV‐displayed monobody construct. Three‐dimensional structure of the engineered monobody‐lactadherin fusion protein displayed on EV surfaces. The construct consists of an N‐terminal signal peptide (purple) for membrane targeting, followed by the fibronectin‐based monobody scaffold (red) containing randomized binding loops (yellow), a flexible G4S linker (gray), a PAS domain (orange) and the lactadherin C1C2 domain (purple) for EV membrane integration. The randomized regions (yellow) in the monobody scaffold represent the variable loops that provide binding diversity in the library. Structural modelling was performed using Phyre2 server, and the final visualization was generated using ChimeraX software.

3.2. Monobody Library EVs From HEK293T Cells Display Monobody Proteins While Packaging and Safeguarding pDNA

Large and small EVs can be separated using differential centrifugation (Szatanek et al. 2015). Engineered EVs (eEVs) contain the pDNA used for transfection (Komuro et al. 2021). To identify an EV population with higher pDNA content, we isolated large EVs (lEVs) at 12,000 × g and small EVs (sEVs) at 100,000 × g for 90 min from the supernatant after centrifugation of lEVs (Komuro et al. 2021). Most pDNA was present in the lEVs separated at 12,000 × g (Figure S1), confirming previous reports (Kanada et al. 2015). We therefore focused on lEVs for our study.

For quality control, each EV library batch was characterized using nanoparticle tracking analysis (NTA) and quantitative PCR (qPCR). NTA consistently revealed peak EV sizes of 120–140 nm (Figure 2A,B). To confirm that pDNA was protected by EVs, we treated samples with DNase I, which degrades free‐floating DNA to undetectable levels, then quantified pDNA (Komuro et al. 2021). Figure 2C illustrates the distribution of monobody‐coding pDNA per EV, with an average packaging efficiency of approximately 1.5 copies per particle. Western blot analysis demonstrated enrichment of EV markers (CD63, TSG101 and Alix), successful monobody display (HA), and absence of the cell‐specific marker calnexin (Figure 2D) (Welsh et al. 2024). Super‐resolution microscopy confirmed monobody protein display through co‐localization with EV markers (CD63 and CD9) and HA‐tag (Figure 2E, H), with approximately 20 monobody molecules expressed per EV particle (Figure 2I).

FIGURE 2.

EV characterization showing successful EV isolation, pDNA loading, and surface display of monobody proteins. (A) Size distribution, (B) peak size distribution of multiple samples, (C) amount of pDNA loading in eEV, (D) immunoblot analysis, (E) representative image of eEV before screening and (F) population analysis, and (G–H) after five rounds of screening, respectively. (I) Number of monobody molecules expressed on the EV surface.

While we could not establish a direct one‐to‐one link between specific surface‐displayed proteins and their corresponding packaged plasmid DNA within individual EVs, our iterative selection process demonstrated that targeting monobodies were successfully enriched through multiple rounds of screening. This suggests that despite the stochastic nature of the initial library, the selection process effectively identified and amplified functional targeting molecules.

3.3. In Vitro Enrichment of Cell Type‐Specific Monobody Sequences

We conducted time‐course experiments to determine the optimal treatment duration for in vitro screening. A431 EGFR+ cells were treated with co‐labelled EVs displaying either EGFR‐specific monobody (E626) (Hackel et al. 2012) or non‐binder control (RDG), and the results are shown in Figure S2. We observed time‐dependent increases in pDNA uptake with both monobodies, but cells exhibited significantly higher pDNA uptake when treated with EVs displaying anti‐EGFR monobody E626 compared to non‐binder RDG or co‐labelled EVs. This difference was evident at 15 min and more pronounced at 30 min.

Based on these results, we initially selected a 30‐min treatment duration as optimal. However, since 15‐min treatments also showed significant differences, we compared both timepoints directly. To assess enrichment of high‐binder monobody sequences, we spiked 1% each of E626 (positive control) and RDG (negative control) into the monobody‐displaying library. After three screening rounds, E626 was enriched at both treatment durations, but more efficiently with 30‐min treatments (Figure S3). The complete list of sequence data is available on GitHub https://github.com/HaradaLabMSU/EV‐Library.

We performed five rounds of screening with 30‐min EV treatments. qPCR analysis demonstrated that E626 was significantly enriched in rounds one through three and stabilized in rounds four and five (Figure 3A). The occurrence frequency of each unique sequence from NGS analysis was tabulated for each experimental replicate and panning round (p1‐p5). The ratio between E626 and RDG sequences is plotted in Figure 3B. An increase in the E626/RDG ratio—indicating enrichment of E626 and depletion of RDG—occurred in five out of six independent screening experiments by the third round. The decrease in unique sequences and increase in prevalent‐to‐rare sequence ratios suggest library collapse during iterative panning (Figure S4). This library collapse is expected in successful screening, where high‐performing sequences should appear with increasing frequency (Figure 3B).

FIGURE 3.

Enrichment of target sequences on in vitro EV‐Monobody library screening. (A) Validation of fold change of RDG and E626 in A431 cells by qPCR. (B) The ratio of E626 (positive control) to RDG (negative control) sequences quantified for each of five passages across six separate campaigns. (C) Enrichment growth rate of each variant calculated using: Final count = initial count × Exp (enrichment growth rate × time). (D) Phylogenetic tree showing relatedness between positive control (E626), negative control (RDG) and novel lead monobody variants EG64, EG100, EG130, EG142 and EG165.

Furthermore, novel variants were enriched from a naïve monobody library. In Figure 3C, each dot represents a unique monobody variant, with dot size proportional to sequence frequency across all replicates. The phylogenetic tree (Figure 3D) shows relatedness between positive control (E626), negative control (RDG) and novel lead monobody variants (named EG64, EG100, EG130, EG142 and EG165). NGS revealed 562 unique protein variants, which were clustered based on amino acid sequence similarity using CD‐Hit (90% threshold). Sequences from the resulting 25 clusters were used to generate the phylogenetic tree. We identified E626 variants (with mutations likely introduced during PCR or cellular replication) and novel clones, suggesting competitive selection of high‐affinity monobodies. A panel of variants showing the highest enrichment over multiple passages was selected for further characterization.

Mathematical analysis of the selection dynamics revealed that despite the imperfect correlation between phenotype (displayed proteins) and genotype (encoding DNA), the competitive advantage of high‐affinity binders during multiple rounds of selection enables successful enrichment of target‐binding sequences. This analysis accurately predicted the enrichment trajectory of E626, closely matching our experimental observations of increasing E626/RDG ratios over five rounds of selection.

3.4. Enhanced Monobody Sequences Exhibit Strong Binding Affinity to Target Cells

We investigated the binding properties of enriched sequences to target cells by re‐cloning five high‐binder candidates (EG64, EG100, EG130, EG142 and EG165) into an EV‐display backbone (Table S2). A bioluminescence‐based binding assay using EVs displaying monobodies and NanoLuc revealed that two monobodies (EG130 and EG142) and the positive control (E626) exhibited significantly higher binding relative to controls and other binders (Figure 4A).

FIGURE 4.

Evaluation of selected clones using bioluminescence imaging (BLI). (A) A431 and MCF7 cells were treated with monobody‐displayed EVs co‐labelled with NanoLuc. Total photon flux (p/s) from EVs bound to cells was quantified using IVIS. Values represent mean ± SD (n = 3). Two‐way ANOVA was used to assess time course effects. Significance against 0 min is expressed as: *p ≤ 0.05, **p ≤ 0.01, ***p ≤ 0.001 and ****p ≤ 0.0001. NP: non‐peptide, non‐binder negative control. (B) A431 and MCF‐7 cells were co‐cultured and treated with Monobody‐mCherry co‐labelled EVs for 10 min. Binding was assessed by confocal laser scanning imaging of EVs (red), anti‐EGFR antibody (green) and DAPI nuclear staining (blue). Scale bar: 25 µm. (C) Quantitative image analysis of confocal images. The individual fluorescence intensities of EGFR‐positive and ‐negative cells were quantified by using ImageJ. (D) Quantitative analysis of EV binding selectivity from confocal microscopy images through systematic cell counting. Data represent individual cell counts from representative fields. Statistical analysis by Fisher's exact test comparing EGFR+ vs. EGFR‐ binding frequencies. Green shading indicates selective binders (p < 0.01), red shading indicates non‐selective controls.

To evaluate specific binding capability at the single‐cell level, we used a co‐culture system for confocal microscopy. We generated mCherry‐coated EVs co‐labelled with monobodies and treated co‐cultured A431 and MCF7 cells for 10 min, followed by anti‐EGFR antibody and DAPI staining. E626 and high‐binder candidates selected by in vitro screening demonstrated specific binding to A431 cells (Figures 4B and S5). No specific binding was observed with RDG or NP (negative controls).

Quantitative analysis of confocal microscopy images confirmed binding selectivity for EGFR‐positive cells across all enriched monobody variants. Regions of interest (ROIs) were drawn around all EGFR‐positive and ‐negative cells in the image, and the fluorescence intensity of each cell was quantitatively analyzed. E626, EG130, EG142 and EG165 showed statistical significance, confirming the bulk binding analysis (Figure 4C). Manual cell counting also revealed that EG64, EG100 and EG165 demonstrated binding frequencies of 70% (7/10), 80% (8/10) and 63% (5/8) to EGFR+ cells, respectively, compared to only 15% (2/13), 8% (1/13) and 14% (2/14) to EGFR‐cells, corresponding to selectivity ratios of 4.7‐, 10.0‐, and 4.5‐fold, respectively. Similarly, EG130 and EG142 showed strong selectivity with 82% (9/11) and 75% (6/8) binding to EGFR+ cells versus only 7% (1/15) and 7% (1/14) to EGFR‐ cells (11.7‐ and 10.7‐fold selectivity). In contrast, negative controls RDG and NP showed minimal selectivity, with binding ratios of 1.9‐ and 1.7‐fold, respectively. The positive control E626 demonstrated the highest selectivity at 14.8‐fold (89% vs. 6% binding). These results confirm that despite not showing statistical superiority in the bulk IVIS assay (Figure 4A), EG64, EG100 and EG165 exhibit significant binding specificity at the single‐cell level, validating their enrichment through the screening process (Figure 4D).

These findings align with the binding assay results and confirm selection of biomolecules specific to recipient cells through in vitro EV library screening.

3.5. In Vivo Screening of EV‐Based Monobody Libraries

To determine optimal circulation time for in vivo screening, we administered library EVs to mice via tail vein and euthanized animals at different timepoints (1, 4 and 24 h). We isolated pDNA from various organs (liver, kidney, pancreas and spleen) to determine which duration produced the highest pDNA yield. Recovery varied by organ and timepoint, with average recovery decreasing from 1 to 4 h or 24 h across all organs (Figure S6A). Based on these findings, we used a 1‐h circulation time for subsequent library screening.

We confirmed successful pDNA recovery and monobody fragment amplification from these organs (Figure S6B). Previous studies demonstrated that pDNA introduced via eEVs can be detected in heart and lungs (Komuro et al. 2021). Therefore, we selected six key organs (heart, liver, lung, pancreas, kidney and spleen) to test for specific EV accumulation.

The monobody‐displaying EV library was injected into mice via tail vein and allowed to circulate for 1 h. After sacrifice, target organs were excised, pDNA extracted from each organ, and organ‐enriched monobody sequences amplified and re‐cloned into the EV‐display backbone. Individual EV libraries were prepared from each organ, then pooled for subsequent screening rounds (Figure S7). Three independent rounds of in vivo screening were conducted, and library DNA from the fourth and fifth rounds underwent next‐generation sequencing, revealing several enriched monobody sequences within each organ's DNA library (Table S3).

3.6. Pancreas‐Enriched Monobody Library EVs Exhibit Targeted Accumulation

The pancreas is a challenging organ to target therapeutically, making it an important focus in this research. To validate in vivo enrichment of pancreas‐targeting monobodies, we compared biodistribution of the initial monobody library EVs (P0) with pancreas‐enriched monobody library EVs (Pan‐P5). The Pan‐P5 library consisted of pooled monobody‐library pDNA after five independent in vivo screening rounds.

Both P0 and Pan‐P5 monobody library EVs were co‐labelled with NanoLuc and administered intravenously. Sequential imaging over 30 min revealed similar initial biodistribution patterns (Figure S8). However, at the 1‐h timepoint, ex vivo imaging demonstrated significantly increased Pan‐P5 monobody library EV accumulation within the pancreas compared to P0, providing robust evidence of pancreas‐specific monobody enrichment through in vivo screening (Figure 5A).

FIGURE 5.

Five rounds of in vivo screening created a library with high pancreatic accumulation. EVs were co‐labelled with NanoLuc and each Library DNA (P0, Pan‐P5), high binder candidate P1316 or non‐binder RDG. Equal numbers of particles were introduced into mice by IV injection, and ex vivo imaging was performed after 1 h of circulation. (A) Ex vivo images of each organ, comparison of percentage of detected BLI from injected EVs, (B) ex vivo image of pancreases, accumulation degree of pancreatic library relative to control (original library). (C) Biodistribution of P1316‐EVs co‐labelled with ThermoLuc‐CD63. (D) Ex vivo image of each organ, comparison of percentage of detected BLI from injected EVs with NP (no‐surface engineering), RDG (no‐binder monobody control) or P1316 (Pancreas enriched monobody). Unpaired t‐test was used to evaluate the enrichment in each organ. Significance is expressed as follows: *p ≤ 0.05, **p ≤ 0.01, ****p ≤ 0.0001.

To quantify this enrichment, we normalized signals from each organ to the input. The Pan‐P5 library EV signals were approximately twice as high as those of P0. In a direct pancreas‐to‐pancreas comparison, the Pan‐P5 signal was approximately four times higher than that of the P0 library (Figure 5B).

Next, we evaluated the ability of individual monobody sequences (whose enrichment was confirmed by sequence analysis) to accumulate in the pancreas. Four high‐binder candidates were cloned into an EV‐display construct and preliminarily assessed for pancreatic binding capacity. These high‐binder candidate EVs, co‐labelled with NanoLuc protein by co‐transfecting EV‐displayed Nanoluc pDNA with the EV‐displayed monobody pDNA into producer cells. The high‐binder, Nanoluc EVs were introduced into mice via tail vein injection, and organ accumulation was evaluated through ex vivo imaging. Among the candidates, P1316 demonstrated particularly high pancreatic accumulation (Figure S9).

We further evaluated P1316 using EVs co‐expressing the monobody and a ThermoLuc‐CD63 fusion protein, whose expression is specific for intracellular delivery. P1316‐EVs showed a signal vicinity to the pancreas region compared to the control, NP (Figures 5C and S10). Furthermore, ex vivo evaluation using NanoLuc confirmed significant accumulation in the pancreas (Figure 5D), supporting the successful enrichment of organ‐specific biomolecules through our in vivo screening approach. To address concerns about non‐specific targeting and background accumulation, we included additional controls comparing P1316‐displaying EVs with both non‐surface engineered EVs (NP) and negative control monobody‐displaying EVs (RDG). This analysis demonstrated several key findings: (1) Selective enhancement with proper controls: P1316‐displaying EVs showed statistically significant enhancement specifically in pancreas compared to all control conditions (4‐fold vs. NP, p < 0.0001; 2‐fold vs. RDG, p < 0.01), while other organs showed minimal differences between groups. (2) Validation of targeting specificity: Both non‐engineered EVs and negative control monobody (RDG) showed minimal pancreatic accumulation (∼0.5 × 10⁻³ % and ∼1.0 × 10⁻³ %, respectively), confirming that the observed enhancement with P1316 (∼2.0 × 10⁻³ %) results from specific monobody‐mediated targeting rather than non‐specific EV accumulation. (3) Organ selectivity: P1316 demonstrated enhancement only in pancreas, with accumulation in other organs (heart, lung, liver, spleen and kidney) remaining comparable to or lower than control conditions, confirming true tissue specificity rather than general increased uptake.

3.7. Mathematical Analysis Reveals Selection Dynamics of EV‐Displayed Monobodies

To understand the stochastic relationship between surface‐displayed monobodies and their packaged DNA, we mathematically analyzed the selection process (Figure 6A). This model enabled us to quantify the efficiency of each step in the selection cycle and predict enrichment rates for specific monobody variants.

FIGURE 6.

Mathematical analysis of EV‐displayed monobody selection process. (A) Schematic representation of the selection process, showing key stages and the stochastic relationship between surface protein and packaged DNA. (B) Experimental validation of selection efficiency through population analysis. (C) Enrichment of E626 monobody frequency over five rounds of selection, demonstrating effective selection despite imperfect genotype–phenotype linkage. Error bars represent standard error across multiple experimental campaigns (n = 6). (D) Comparison of organ distribution profiles between the initial library (P0) and pancreas‐enriched library (Pan‐P5) Error bars represent standard error of the mean (n = 3).

Our analysis revealed that while individual EVs exhibit imperfect genotype–phenotype correlation, population‐level selection successfully enriches target‐binding sequences. Each EV displays approximately 20 monobody molecules on its surface, it contains on average of only 1.5 copies of plasmid DNA. This creates a complex stochastic relationship where the correlation between any displayed monobody and its encoding DNA varies significantly among individual particles.

Despite this imperfect correlation, our experimental observations demonstrated successful enrichment of target‐binding sequences. Simulation of E626 enrichment (initially present at 1% of the library) showed progressive increases through five rounds of selection, closely matching our experimental observations of increasing E626:RDG ratios (Figure 6C). This confirms that iterative selection can overcome the stochastic nature of monobody display and DNA packaging.

For organ‐specific targeting, our analysis predicted that enhanced pancreatic accumulation (as observed for Pan‐P5 compared to P0) would shift the organ distribution profile, increasing pancreatic targeting from 4% to 14%, while proportionally decreasing targeting to other organs (Figure 6D). These predictions correlated well with our ex vivo imaging results for P1316, validating the model's utility of mathematical modelling in understanding organotropic targeting.

3.8. Key Finding

The successful experimental enrichment despite low theoretical individual‐particle correlation suggests that EV‐based selection is more robust to stochastic variation than initially anticipated, with important implications for the design of particle‐based screening systems.

4. Discussion

In the development of new drugs and targeted therapies, relevant screening techniques for selecting tissue and cell targeting molecules from a combinatorial library will enable directed delivery and reduce off‐target toxicity. Screens that inherently use particles in the selection process should select for targeting proteins that can direct nanoparticles and other large cargos to target cells and tissues. While phage display (Smith and Petrenko 1997), yeast display libraries (Boder and Wittrup 1997) and various screening methods have successfully demonstrated biomolecule selection (Saw and Song 2019), our EV‐display library screens extend peptide and protein screening to in vivo assays with greater relevance to targeted drug and nanoparticle delivery.

The key challenge in our approach was establishing the correlation between the displayed surface proteins and the packaged DNA within individual EVs. Our mathematical analysis quantified this relationship, revealing that each EV displays approximately 20 monobody molecules but contains an average of only 1.5 copies of plasmid DNA. Under realistic assumptions about library complexity and protein distribution, the probability of strong genotype–phenotype correlation in individual EVs is low. However, our experimental success demonstrates that even modest correlations provide a sufficient selective advantage for effective enrichment between a displayed monobody and its encoding DNA in the original EV library, with greater probabilities expected in sequentially selected pools as the library complexity collapses. Library convergence to fewer unique sequences increases the frequency of identical sequences, where higher frequencies indicate more effective selection. The distribution of these sequence frequencies provides quantitative data for modelling selection dynamics and assessing screening efficiency. Although we could not demonstrate a direct one‐to‐one relationship between surface‐displayed monobodies and their corresponding encoding DNA in single EVs, the iterative selection process effectively enriched targeting monobodies. This suggests that despite the initially stochastic nature of protein display and DNA packaging, the selection pressure applied through multiple rounds of screening successfully identifies functional targeting molecules.

Screening display libraries in relevant mammalian ecosystems using naturally occurring bionanoparticles (EVs) reveals targeting peptides that direct particles, and perhaps small molecules, to target tissues. The EV structure, a lipid bilayer membrane encapsulating various biomolecules, allows cells to be used as packaging factories to produce engineered particles with defined cargo and specific targeting molecules on their surface (Herrmann et al. 2021). In this way, the natural function of EVs for intercellular communication can be adapted for use as delivery vehicles. Our approach takes advantage of modified cargo (pDNA) to correlate with modified EV surfaces, enabling effective in vivo screening. Our analysis demonstrates that this stochastic approach is robust, with the iterative selection process effectively enriching target‐binding sequences despite imperfect individual particle correlation of the library while enriching target‐binding sequences.

Various methods for modifying the EV surface have been reported (Komuro et al. 2022). In our previous studies, we demonstrated targeting by EV surface display using the C1C2 domain of human lactadherin (Komuro et al. 2022, 2021). Our engineered EVs (eEVs) shield at least one transfected pDNA molecule per particle from external environmental factors. Additionally, they express an average of 20 monobody molecules on their surface. Using these characteristics, we designed an EV library screening methodology similar to phage display screens but adapted to the stochastic relationship between DNA cargo and surface markers.

To our knowledge, this is the first report of EVs being used in a combinatorial biomolecule screening platform, either in culture or in vivo, rather than simply displaying pre‐selected molecules, either in culture or living organisms. As an initial demonstration, we performed in vitro screening using A431 cells. Both qPCR and NGS evaluations confirmed enrichment of the anti‐EGFR monobody E626, which was spiked in as a positive control. This proof‐of‐principle illustrates that the platform can identify functional binders even in the context of stochastic cargo‐display correlation.

The improved efficiency may relate to the EV type selected. Characterization of the library EVs showed an average particle size of approximately 130 nm containing, on average, 1.5 molecules of pDNA per particle. The detection of typical EV markers (CD63, ALIX and TSG101) along with the engineered HA marker confirmed that the collected particles were EVs, with 80% displaying monobodies on their surface. While there are few reports quantitatively evaluating EV labelling efficiency, our results may serve as a guide for future research (Thane et al. 2019; Liu et al. 2019; Corso et al. 2019). Our single EV imaging studies confirmed that monobodies can be displayed on the EV surface in multiple copies, which may account for our better‐than‐expected results and provide a significant advantage over phage display, where typically only five copies of the same peptide are present on each particle (Saw and Song 2019).

The process by which pDNA is packaged into eEVs is likely due to cytoplasmic DNA not being tolerated by cells, with EV packaging serving as one mechanism to remove this DNA (Rädler et al. 2023). The relationship between the expression of EV‐targeted proteins (the lactadherin fusion proteins), biogenesis of eEVs and packaging of the pDNA cargo is likely stochastic. The stochastic relationship we observed between surface‐displayed proteins and packaged plasmid DNA aligns with recent findings by Fordjour et al. (2022), who demonstrated that EV cargo proteins can vary by approximately 100‐fold from one vesicle to another through a shared stochastic mechanism (Fordjour et al. 2022). In biological selection systems, iterative processes can effectively enrich for functional variants despite imperfect initial correlations between phenotype and genotype. While each individual EV may contain variable combinations of surface proteins and internal DNA, our repeated rounds of selection effectively collapsed this probabilistic distribution towards EVs containing both the targeting monobody on the surface and its encoding DNA inside. This progressive enrichment underlines the ability of the VesicleVoyager platform to identify tissue‐targeting EVs even when single‐particle correlations are modest.

While we could not establish a direct correlation between displayed proteins and packaged DNA within individual EVs, our results demonstrate that through iterative rounds of selection, we successfully enriched targeting monobodies. The observed tissue enrichment in vivo was significant compared to background controls, though modest, highlighting that further validation across additional cell types and tissues will strengthen the platform. Protein corona formation and EV surface properties, such as zeta potential, likely influence tissue‐specific uptake. By performing selection directly in vivo, VesicleVoyager inherently accounts for these factors, providing physiologically relevant readouts that pre‐selected in vitro screens cannot replicate.

Among the five high‐binder candidate sequences newly identified by in vitro screening, EG130 and EG142 showed significant affinity to A431 cells in bulk assays. Confocal microscopy confirmed that all sequences were specific to A431 cells, possibly due to the sensitivity of the assay. The pancreatic library that underwent five rounds of in vivo screening showed higher affinity for the pancreas compared to the starting P0 library, and the high‐affinity candidates obtained by NGS analysis also demonstrated affinity for the pancreas. This confirms that our in vivo EV library screening successfully selects for tissue‐targeting monobodies. However, we observed some non‐specific delivery to other organs, suggesting that further optimization is needed to enhance specificity for therapeutic applications. The inclusion of non‐surface‐engineered EV and negative control monobody comparisons in our pancreatic targeting studies provides rigorous validation of specificity. The 4‐fold enhancement over non‐engineered EVs and 2‐fold enhancement over negative control monobodies, combined with the lack of enhancement in non‐target organs, demonstrates that our screening approach successfully identifies sequences that confer genuine tissue‐specific targeting properties rather than artifacts of the experimental system.

The identification of clinically relevant targeting monobodies presents challenges that can be addressed in several ways. Compared to in vivo screening of phage display libraries, EV library screening avoids potential bystander effects on the mammalian microbiome and reduces risks to the immune system because it utilizes a universal mammalian mechanism rather than a bacterial virus. Once a targeting monobody has been identified, its target protein and epitope can be characterized, and if homologous to human proteins, translation to clinical applications is straightforward.

Alternatively, monobody libraries can be pre‐selected against human targets and then screened in vivo to increase the likelihood of finding clinically relevant monobodies. EV libraries could also be screened in humanized mice to identify relevant monobodies. In this study, we used HEK293T cells to produce eEVs, but the intrinsic targeting properties of EVs vary depending on the cell type used for production. Therefore, careful consideration of the cells used as EV factories is important (Edelmann and Kima 2022; Yáñez‐Mó et al. 2015). Furthermore, using disease models as screening hosts could potentially select molecules that directly target disease sites.

5. Conclusion

We have developed a screening system that utilizes engineered EVs as a peptide display platform to select molecules that direct nanoparticles to target cells and tissues. If the identified monobodies are used on other nanoparticles, some of the off‐target delivery observed in our study would likely be reduced. Alternatively, if synthetic EVs are used, or if natural EVs can be “hardened” to prevent non‐specific delivery to non‐target sites, monobody‐coated nanoparticles could achieve highly selective drug delivery to specific cells and tissues. This technology opens an exciting new approach to drug discovery that combines the advantages of natural EVs with the power of directed molecular evolution. Our mathematical analysis provides insights into the selection dynamics of EV‐displayed monobodies. The successful enrichment despite stochastic relationship between displayed proteins and packaged DNA demonstrates the robustness of iterative selection processes and has important implications for particle‐based screening systems.

Author Contributions

Yuki Kawai‐harada: writing ‐ original draft, methodology, validation, investigation, formal analysis, visualization. Mehrsa Mardikoraem: software, data curation. Ashley V. Makela: investigation, writing ‐ review and editing. Katherine Lauro: investigation. Jeannie Lam: investigation. Christopher H. Contag: conceptualization, funding acquisition, writing ‐ review and editing. Daniel Woldring: writing ‐ review and editing, software, supervision. Masako Harada: conceptualization, investigation, resources, supervision, writing ‐ review and editing, funding acquisition, project administration.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Supplementary Material: jev270184‐sup‐0001‐SuppMat.docx

Kawai‐Harada, Y. , Mardikoraem M., Makela A. V., et al. 2025. “VesicleVoyager: In Vivo Selection of Surface Displayed Proteins That Direct Extracellular Vesicles to Tissue‐Specific Targets.” Journal of Extracellular Vesicles 14, no. 11: e70184. 10.1002/jev2.70184

Data Availability Statement

The datasets generated and/or analyzed in this study are available in the HaradaLabMSU/EV‐Library GitHub repository, https://github.com/HaradaLabMSU/EV‐Library. GitHub repository is available under the Creative Commons Attribution‐NonCommercial‐NoDerivatives 4.0 International Public License (“Public License”) at https://github.com/HaradaLabMSU/EV‐Library.