Abstract Extracellular vesicles (EVs) are emerging as a promising drug delivery vehicle as they are biocompatible and capable of targeted delivery. However, clinical translation of EVs remains challenging due to the lack of standardized and scalable manufacturing protocols to consistently isolate small EVs (sEVs) with both high yield and high purity. The heterogenous nature of sEVs leading to unknown composition of biocargos causes further pushback due to safety concerns. In order to address these issues, we developed a robust quality‐controlled multi‐stage process to produce and isolate sEVs from human embryonic kidney HEK293F cells. We then compared different 2‐step and 3‐step workflows for eliminating protein impurities and cell‐free nucleic acids to meet acceptable limits of regulatory authorities. Our results showed that sEV production was maximized when HEK293F cells were grown at high‐density stationary phase in semi‐continuous culture. The novel 3‐step workflow combining tangential flow filtration, sucrose‐cushion ultracentrifugation and bind‐elute size‐exclusion chromatography outperformed other methods in sEV purity while still preserved high yield and particle integrity. The purified HEK293F‐derived sEVs were thoroughly characterized for identity including sub‐population analysis, content profiling including proteomics and miRNA sequencing, and demonstrated excellent preclinical safety profile in both in‐vitro and in‐vivo testing. Our rigorous enrichment workflow and comprehensive characterization will help advance the development of EVs, particularly HEK293F‐derived sEVs, to be safe and reliable drug carriers for therapeutic applications. Keywords: drug delivery, exosomes, extracellular vesicles, HEK293, preclinical safety, purification 1. INTRODUCTION Extracellular vesicles (EVs) are nanosized vesicles bound by lipid bilayer membrane and secreted from virtually all cells. They can be classified into exosomes, microvesicles, and apoptotic bodies based on how they are formed; or into small Evs (sEVs)—less than 200 nm in diameter, and large EVs based on size (EL Andaloussi et al., [46]2013; Yáñez‐Mó et al., [47]2015). In addition to the great potential for therapeutic and diagnostic applications, sEVs are actively studied as an alternative drug delivery system due to the intrinsic biocompatibility, ability to bypass biological barriers and particularly to deliver cargos to specific cell populations (EL Andaloussi et al., [48]2013; Yáñez‐Mó et al., [49]2015). However, on‐going clinical trials investigating sEVs as a drug delivery platform accounts for less than 10% of total EV‐based studies (Rezaie et al., [50]2022). The superiority of sEVs over synthetic systems such as lipid nanoparticles and the benefit‐risk ratio remain debatable (Rezaie et al., [51]2022). The major challenge in developing sEV‐based therapeutics is the lack of standardization to isolate and purify high quality sEVs on a large scale. Several methods have been developed to purify sEVs based on size, density, charge and immunocapture; each with their own drawbacks for yield, purity, quality, cost and scalability (Herrmann et al., [52]2021). Common methods that are scalable and compatible with good manufacturing practices (GMP) are tangential flow filtration (TFF); ultracentrifugation (UC) including direct UC, UC with density gradient or sucrose cushion; and size exclusion chromatography (SEC) (Corso et al., [53]2017; Stam et al., [54]2021). Combining multiple sEV purification methods has consistently been shown to outperform single method in sEV purity regardless of starting materials (Bellotti et al., [55]2021; Stam et al., [56]2021; Visan et al., [57]2022). Majority of these studies involved 2‐step workflows combined in different orders while very few examined three or more methods (Stam et al., [58]2021; Visan et al., [59]2022; Zhang et al., [60]2020). While adding more steps in the purification process is likely to further improve purity, the yield and intactness of sEVs at the end could be compromised. Another challenge in developing sEVs as a drug delivery vehicle is the source of sEVs that is scalable for high‐yield production, together with excellent reproducibility and well‐documented safety profile. Mesenchymal stem cells (MSCs) and cancer cell lines have been extensively used in research and even clinical trials to deliver synthetic cargos (Rezaie et al., [61]2022). Since the sEVs from these cells also carry biocargos reflecting the biological state and function of their parental cells, concerns for long‐term safety have been raised for potential carcinogenicity and impact on immune system upon repeated dosing. Particularly with MSCs, high batch‐to‐batch inconsistency, contamination with animal‐derived proteins from serum‐containing media, and cost‐effective scaling‐up remain problems to be addressed (Herrmann et al., [62]2021; Rezaie et al., [63]2022). On the other hand, a promising source of sEVs for drug delivery is the human embryonic kidney HEK293F cells, which have rapid growth in suspension using chemically‐defined serum‐free media. This is the only well‐studied, immortalized human cell line not derived from cancer. However, sEVs produced by HEK293F cells have not been fully characterized compared to those from MSCs or its variant HEK293T cells (Kugeratski et al., [64]2021; Li et al., [65]2016). In this study, we cultured HEK293F cells in stirred‐tank bioreactor (STR) to produce sEVs. We thoroughly examined different 2‐ and 3‐step workflows and optimized a novel robust protocol that combined TFF, UC with sucrose cushion, followed by bind‐elute SEC, to greatly improve purity while still maintaining a high yield of sEVs. The purified HEK293F‐derived sEVs were then extensively characterized for identity, content profiling as well as both in‐vitro and in‐vivo safety and immunogenicity. 2. MATERIALS AND METHODS 2.1. Cell culture FreeStyle HEK293F cells (ThermoFisher, USA) were cultured in Hyclone CDM4HEK293 medium (Cytiva, USA) supplemented with 1% penicillin—streptomycin (ThermoFisher, USA) and 4 mM L‐glutamine (Cytiva, USA) in shake flasks at 70 rpm orbital shaking, with 5% CO[2] and 95% relative humidity. For comparison of sEV production, cells at passage P10‐20 were inoculated at a density of 3 × 10^5 cells/mL in either shake flasks or the BioFlo 120 bioreactor system (Eppendorf, USA) with controlled parameters: agitation at 100 rpm, temperature at 37°C, pH at 7.2 and dissolved oxygen at 40%. To examine the sEV yield in different growth phases, sEVs were collected when cell density reached either 1–2 × 10^6 cells/mL at Day 3 (exponential growth phase) or 3–4 × 10^6 cells/mL at Day 4 (stationary phase). In STR batch cultures, starting volume was 1 L and sEVs were all harvested when cell density reached 3–4 × 10^6 cells/mL at Day 4. In semi‐continuous cultures, starting volume was the same but each time the cell density reached 3–4 × 10^6 cells/mL, only 500 mL of the media was harvested while 500 mL of fresh media was added. The semi‐continuous cultures were continuously harvested at Day ‐4, ‐6 and ‐8. Cell density and viability were monitored daily by haemocytometer with trypan blue staining (ThermoFisher, USA). 2.2. Isolation and enrichment of sEVs 2.2.1. Prefiltration and benzonase treatment Conditioned medium (CM) of HEK293F cells was centrifuged first at 300 × g for 5 min to remove cells, then at 3000 × g, 4°C for 30 min to remove cell debris. The clarified CM was filtered through a 0.22‐μm polyethersulfone (PES) membrane (Merck, Germany) to remove large EVs. For subsequent cell‐free nucleic acid digestion, 20 U of Benzonase nuclease, ultrapure (Sigma, USA) was added to every 100 mL of the CM, which was gently stirred at 37°C for 2 h. 2.2.2. Tangential flow filtration (TFF) CM was next concentrated using the Pellicon cassettes and Biomax 300 kDa membrane (Merck, Germany) for 10–15× concentration, followed by two rounds of diafiltration with an equal volume of phosphate‐buffered saline (PBS) each time. 2.2.3. Sucrose cushion ultracentrifugation (US) Sucrose (Sigma‐Aldrich, Germany) was dissolved in sterile PBS to form 30% sucrose solution. After 20 mL of the concentrated media was added to ultracentrifuge tube, 5 mL of 30% sucrose solution was loaded slowly underneath the media layer. The tubes were ultracentrifuged at 120,000 × g, 4°C for 4 h or overnight. The supernatant was then discarded; ∼8 mL of the bottom sucrose fraction and media was diluted in PBS and concentrated by either direct ultracentrifugation (UC) or Amicon ultra centrifugal filter 10 kDa molecular weight cut off (MWCO) (Merck, Germany). Three rotors were tested for this US step: type 70.1 Ti, 50.2 Ti, SW 55 Ti with k‐factors of 36, 69, 48 respectively, using the ultracentrifuge Optima XE‐90 (Beckman Coulter, USA). 2.2.4. Bind‐elute size exclusion chromatography (BE‐SEC) BE‐SEC was performed using Capto Core 700 resins (Cytiva, USA) with multimodal, octylamine ligands that give the resins dual functionality in both SEC and binding to contaminants. The column was connected to ÄKTA Pure chromatography system (Cytiva, USA) and was loaded with sEV‐containing solution, followed by elution with PBS. The purified sEV fraction was collected based on the peak of UV[280] and then concentrated by either direct UC or Amicon filter as above. 2.2.5. Ultracentrifugation (UC) All direct UC steps were performed at 120,000 × g, 4°C for 3 h, using Type 50.2 Ti fixed‐angle rotor. 2.2.6. Storage sEV samples were stored in three buffers: PBS, Ringerfundin (RF), RF with 0.2% human serum albumin, and examined for stability. Briefly, sEV samples were aliquoted into 5 μg of total proteins in 10 μL buffer per tube for storage. The whole tube content was then lysed in the same tube and used for immunoblotting. RF was chosen as the storage buffer in this study. Samples were stored at −80°C for long term storage and slowly thawed at 4°C for use. 2.3. Immunoblotting Immunoblotting was performed as previously described (Tran et al., [66]2023). Briefly, exosome samples were lysed in the radioimmunoprecipitation assay (RIPA) lysis buffer and the total protein content was determined using the Pierce rapid gold BCA protein assay kit (Thermo Fisher, USA). Samples with equal protein loading were separated by SDS‐PAGE and then transferred semidry to a PVDF membrane using the Power Blotter system (Thermo Fisher, USA). The membranes were then blocked with 5% (w/v) bovine serum albumin (BSA) in tris‐buffered saline with the addition of 0.1% Tween‐20 (TBST) for 1 h before incubation with primary antibodies in the blocking buffer at 4°C overnight. The primary antibodies used were CD81 (#sc‐166029, Santa Cruz Biotechnology, CA, USA), CD9 (#ab236630, Abcam, Cambridge, UK), TSG101 (#sc‐7964, Santa Cruz Biotechnology), β‐actin (#sc‐47778, Santa Cruz Biotechnology), Calnexin (#2679, Cell Signaling), β‐Tubulin (#2128, Cell Signaling). After washing with TBST, the membranes were incubated with secondary antibody at room temperature for 1 h. Immunoblots were captured and quantified using the iBright imaging system (Thermo Fisher Scientific). 2.4. Field emission scanning electron microscopy Purified sEVs in PBS were fixed with 4% paraformaldehyde for 5 min at room temperature, then diluted in 0.9% sodium chloride. A drop was immobilized on a pure copper sheet and air dried. Scanning electron micrographs (SEM) of sEVs were obtained under LA‐BSE mode, 5 kV accelerating voltage with a FESEM SU8010 microscope (Hitachi High‐Technologies, Japan). 2.5. sEV size, concentration and sub‐population analysis Samples were diluted in particle‐free PBS and clarified immediately prior to use with a 0.22 μm PES filter. Particle concentration and size distribution were measured using either ZetaSizer Ultra machine (Malvern, UK), Nanoparticle tracking analysis (NTA) (Nanosight NS300), or Leprechaun system (Unchained Labs, UK). For ZetaSizer, samples were run in standard settings: temperature 25°C, refractive index 1.33, viscosity 0.89, dielectric constant 78.5 and analysed by ZS Xplorer software. For NTA, samples were analysed at 25°C, using 488 nm laser, sCMOS camera level 13–14, detection threshold 4, triplicate videos of 30 s analysed by NTA 2.3 software. For the Leprechaun system, immunocapture of CD81‐, CD9‐ and CD63‐positive EVs was performed using the Leprechaun Human Tetraspanin Kit (Unchained Labs, UK) as previously described (Nguyen et al., [67]2024). Briefly, about 10^6 particles in 50 μL was loaded on the background‐scanned chips and incubated for 1 h at room temperature. The chips were scanned in the Leprechaun instrument both in the interferometric microscopy (IM) and in the fluorescence channels corresponding to CD81, CD9 and CD63 signals. Acquired images were analysed using the Leprechaun Analysis 1.1. software for the particle concentration and size. For sub‐population analysis, sEVs were classified into seven sub‐populations based on the expression of their tetraspanins: CD9^+, CD81^+, CD63^+, CD9^+CD81^+, CD9^+CD63^+, CD63^+CD81^+, CD9^+CD81^+CD63^+. The structure of sEV populations was further visualized by t‐distributed Stochastic Neighbour Embedding (tSNE) method (perplexity = 30) using fluorescent intensity data of each particle. sEV particle quantification reported in the study was performed using the Leprechaun system and the term ‘particles’ referred to those positive for at least one of the 3 markers CD81, CD9 and CD63. 2.6. Proteome profiling sEV samples were lysed in RIPA buffer and total proteins were digested using S‐TRAP Micro column. Each sample (1 μg total protein) was loaded on nanoEase M/Z C18 5‐μm trap column, eluted on C18 1.7‐μm column (Waters, USA) and analysed using LC‐MS/MS on the TripleTOF 6600 system (Sciex, Canada). Peptide identification was performed as previously described (Nguyen et al., [68]2024). This analysis was performed by the Mass Spectrometry facility in the Protein and Proteomics Centre, National University of Singapore, Singapore. Peptides identified with a confidence level ≥95% were used for further analysis. Total proteins were analysed for Gene Ontology (GO) enrichment of biological process, molecular function, and cellular component using the DAVID Bioinformatics Resources (Sherman et al., [69]2022). Pathway enrichment analysis was performed using g:Profiler (Kolberg et al., [70]2023). Proteomic data were deposited in the Vesiclepedia database (Chitti et al., [71]2024), project #3590. 2.7. RNA extraction and miRNA sequencing Total RNAs including miRNAs were extracted from 10 μg of EVs using the miRNeasy Micro Kit (Qiagen, Germany). 10 ng of total RNA was then used to prepare sequencing libraries for small RNAs using the small RNA‐seq Library Prep Kit (Lexogen, Italy). Libraries were pooled and sequenced on the Nextseq 2000 (Illumina, USA). The sequencing data was analysed by nf‐core/smrnaseq—a bioinformatics best‐practice analysis pipeline for Small RNA‐Seq (Ewels et al., [72]2020) as previously described (Nguyen et al., [73]2024). miRNA counts were normalized by counts per million (CPM) and miRNAs with CPM less than 5 were removed from further analysis. Other quality control for sequencing data was performed by miRTrace (Kang et al., [74]2018). miRNA targets predicted by miRNet 2.0 (Chang et al., [75]2020) were analysed for pathway enrichment using g:Profiler (Kolberg et al., [76]2023). 2.8. Macrophage and monocyte activation assay Adherent monocytes from human cord blood mononuclear cells were cultured with 50 ng/mL GM‐CSF (Peprotech, USA) for 7 days to differentiate into macrophages, which were confirmed for CD11b expression by flow cytometry. Macrophages were then seeded at 300,000 cells/well in 12‐well plates and treated with PBS (vehicle control), 1 ng/mL lipopolysaccharides LPS (positive control) (Sigma‐Aldrich, Germany) or sEVs at concentration of 5 × 10^8 particles/mL. After 4 h of treatment, total RNA was isolated using the Quick‐RNA Microprep Kit (Zymo Research, USA), followed by cDNA synthesis using the High‐Capacity cDNA Reverse Transcription Kit (Applied Biosystems, USA). Real‐time PCR was performed by SYBR Green detection method using validated primer sequences (Table [77]S1). In a similar experiment, human monocyte THP‐1 cells were seeded at 1,000,000 cells/well in 12‐well plates, treated with two doses of sEVs at 5 × 10^7 and 5 × 10^8 particles/mL for 6 h and processed as above. All gene expressions were normalized to the internal control gene RPLP0; relative quantification of fold‐change was performed using the 2^−ΔΔCt method. 2.9. Animal study The animal study was performed following the protocols approved by the Animal Ethics Committee of University of Science, Ho Chi Minh City, Vietnam (approval number 1232B/KHTN‐ACUCUS). All the male and female BALB/c mice were purchased from BioLASCO (Taiwan Republic of China). To examine safety profile of sEV repeated dosing, 16‐week‐old mice were randomly divided into two groups to receive either Ringerfundin as the vehicle control (n = 6), or sEVs produced from two batches (n = 7) at approximately 10^11 particles per dose, equivalent to ∼10 μg. All doses with volume of 100 μL each were administered intravenously by tail vein injection, with the frequency of 3 times a week for 2 weeks. Animals were monitored during the entire study for body weight, visible changes in behaviour and posture, and then sacrificed 3 days after the last dose. Whole blood was collected by cardiac puncture and analysed for complete blood count by haematology analyser MINDRAY BC‐2800 (Mindray, China) and for biochemical profiling by Pointcare chemistry analyser PV3 (Mnchip, China). A portion of whole blood samples was centrifuged at 2000 × g, 4°C for 10 min to separate the serum. The buffy coat layer was collected and red blood cells (RBC) were then lysed using 1× RBC Lysis Buffer (Biolegend, USA). The remaining leukocytes were subjected to total RNA extraction and real‐time PCR as described above. Animal organs including liver, spleen, and kidneys were collected and fixed in 10% neutral buffered formalin for 24 h. After paraffin embedding, 4‐μm thick sections were cut and stained with haematoxylin and eosin (H&E) and evaluated by a certified pathologist from the Department of Anatomic Pathology, University of Medicine and Pharmacy, Ho Chi Minh City, Vietnam. To examine biodistribution of sEVs, 100 μg of purified sEVs were stained with DiR (Thermo Scientific, USA) at final concentration of 20 μM for 30 min at room temperature. The staining solution was ultracentrifuged at 110,000 × g, 4°C for 70 min and sEV pellet was resuspended gently with 0.9% sodium chloride. The DiR‐stained sEVs were injected intravenously in the tail vein of 10 week‐old mice (n = 3), that were imaged using the IVIS spectrum in‐vivo imaging system (Perkin Elmer, USA) at different time points: 5, 10, 15, 30, 60 min after injection. After mice were sacrificed, organs including heart, lungs, liver, spleen, and kidneys were obtained for ex‐vivo imaging using the same machine. 2.10. Other safety tests Mycoplasma testing was performed using the MycoAlert Mycoplasma Detection Kit (Lonza, USA). Level of endotoxin was quantified by the kinetic turbidimetric test (Associates of Cape Cod, USA). Bacterial contamination was determined by the direct inoculation method. These tests were performed for both cells and purified sEV samples by the Institute of Drug Quality Control (Ho Chi Minh City, Vietnam). Cell‐free DNA (cfDNA) and cell‐free RNA (cfRNA) were measured using Quantus fluorometer (Promega, USA). The in‐vitro safety assessment was performed as previously described (Tran et al., [78]2023). Briefly, four cancer cell lines A549, MDA‐MB‐231, MKN‐45, MCF‐7 were used. A total of 300,000 cells were seeded in 12‐well plates for 18 h and then treated with either PBS (vehicle control), sEVs at concentrations of 5 × 10^8 and 5 × 10^10 particles/mL or doxorubicin (positive control) for 24 h. After treatment, cells were trypsinized and incubated with 500 nM propidium iodide for 15 min and then analysed by flow cytometry using the BD Accuri C6 Plus flow cytometer (BD Bioscience, USA) (PE channel). 2.11. EV‐TRACK We submitted all relevant data of our experiments to the EV‐TRACK knowledgebase (EV‐TRACK ID: [79]EV230996). 2.12. Statistical analysis Student's t‐test was used for comparison of two groups; one‐way ANOVA with Tukey's post‐hoc test was used for comparison of more than two groups. All statistical tests were performed using GraphPad Prism. Bar graphs indicated mean ± SEM; p < 0.05 was considered significant. 3. RESULTS 3.1. Complete workflow of sEV production and enrichment This study developed a standardized multi‐stage production and enrichment workflow that started with the culture of HEK293F cells in a stirred‐tank bioreactor (STR) to generate sEVs. Conditioned media (CM) was then centrifuged at different speeds to remove cells and debris; filtered through 0.22‐μm PES membrane to remove large EVs; followed by treatment with benzonase for nucleic acid digestion (Figure [80]1a). Subsequently, the core isolation and enrichment process combined three techniques: TFF, sucrose cushion ultracentrifugation (US) and BE‐SEC using Capto Core 700 multimodal resins, to obtain sEVs of high yield and high purity (Figure [81]1a). At each stage, quality control measures were taken, using defined methods and criteria (Table [82]S2). The resulting sEVs were subjected to in‐depth characterization for identity analysis, content profiling and safety studies (Figure [83]1b), using standardized methods and acceptance criteria (Table [84]1). FIGURE 1. FIGURE 1 [85]Open in a new tab Workflow of sEV production, enrichment and characterization. (a) HEK293F cells were scaled up from shake flask to stirred‐tank bioreactor for sEV production. Conditioned media were subjected to prefiltration and nucleic acid digestion, followed by a 3‐step isolation and enrichment process including tangential flow filtration, sucrose‐cushion ultracentrifugation and bind‐elute size exclusion chromatography using Capto Core 700 multimodal resins. Purified sEVs were resuspended in storage buffer and stored at −80^oC. QC measurements were applied at each step. (b) Comprehensive characterization of purified sEVs included identification, content profiling, and safety evaluation. QC, quality control; sEV, small EVs. TABLE 1. Criteria for sEV characteristics. Parameter Acceptance criteria Technology Identity Purity >1 × 10^10 particles/μg BCA, Leprechaun Particle concentration >1 × 10^13 particles/L Leprechaun Particle Size 30–200 nm NTA, Leprechaun Morphology Round and intact SEM Surface markers CD9 ^+ , CD63 ^+ , CD81 ^+ , TSG101 ^+ , CANX ^‐ Western blot, Leprechaun Content profiling Sub‐populations ∼50% CD81 ^+ CD9 ^+ CD63 ^+ Leprechaun Total RNA Fluorometry miRNA profile miRNA‐seq Proteomic profile LC‐MS/MS Safety cfDNA, cfRNA >99.9% removal Fluorometry Endotoxin <5 EU/mL Kinetic turbidimetric Mycoplasma Negative Fluorometry Sterility Negative Direct inoculation Immunogenicity Absence In‐vitro and in‐vivo [86]Open in a new tab Abbreviations: cfDNA, cell‐free DNA; cfRNA, cell‐free RNA; sEV, small extracellular vesicles. 3.2. Optimized sEV production HEK293F cells were grown in chemically‐defined, serum‐free media in either shake flasks or STR at the same starting cell density of 3 × 10^5 cells/mL. Cell cultures were controlled for doubling time, viability, contamination and culture conditions with defined methods and release criteria (Figure [87]2a, Table [88]S2). When cells were in the log phase with exponential growth, sEVs were harvested at the same time for both cultures in the shake flask and STR, and the sEV yield per cell was not different between the two groups (Figure [89]2b–c). However, when cells were grown for longer time to achieve high density of 3–4 × 10^6 cells/mL and entered the stationary phase, the sEV yields per cell and per day were both 2–4 times higher for the STR and not for the shake flask (Figure [90]2b–c). At this high density, we observed a rapid decline in cell viability for the shake flask, while cell viability was not affected in the STR (Figure [91]2b). Specifically, for the STR, the duration that cells stayed in the stationary phase was prolonged when they were grown in the semi‐continuous mode compared to batch mode for the same period of time (Figure [92]2d). Cell morphology and viability were not affected in the semi‐continuous mode (Figure [93]2e–f). As a result, the particle yield per day was significantly higher in the semi‐continuous culture (Figure [94]2d), which was then selected as the method for sEV production in this study. Perfusion or continuous culture was not available in our STR model. FIGURE 2. FIGURE 2 [95]Open in a new tab Optimization of upstream sEV production. (a) HEK293F cells when scaled up from shake flask to STR were closely monitored for cell growth, viability, contamination and culture conditions. (b) Relationship between particle yield, cell viability and HEK293F cell density in different growth phases and in either shake flask or STR (n = 3–5 replicates/group). (c) Particle yield was highest when cells were cultured at high density in the STR (n = 3–5 replicates/group). (d) High cell density in stationary phase was achieved for a longer time duration in semi‐continuous culture compared to batch culture. (e) Cell morphology was not affected in semi‐continuous culture. (f) Comparison of cell viability between batch and semi‐continuous cultures. (g) Particle yield per day was significantly higher in the semi‐continuous culture than batch culture while purity was not significantly different (n = 4–6 replicates/group). All sEV samples were purified by the complete workflow in Figure 1a. *p < 0.05, Student's t‐test (b), (g) and one way ANOVA with Tukey's post‐hoc test (c). STR, stirred‐tank bioreactor. 3.3. Optimized sEV isolation and enrichment We first optimized each of the isolation and enrichment step separately. For TFF, casein and other proteins from skim milk were spiked in the media to optimize key filtration parameters. Coomassie‐stained gels showed that TFF run with transmembrane pressure (TMP) of 0.69 or 0.39 (bar) for concentration factors of 10 and 15 times, which are the common factors used in the literature, reduced casein content to ∼52% (Figure [96]S1a). Increasing number of diafiltration rounds with PBS also helped decreasing the casein content in the retentate but the third round did not remove as much as the first and second rounds (Figure [97]S1b). Therefore, concentration factors of 10–15 times followed by diafiltration twice with PBS were chosen for the standardized TFF protocol hereafter, since it was efficient for protein removal and time saving. For US, we loaded a layer of sucrose cushion underneath the CM, and first compared two types of available rotors: fixed angle and swinging bucket rotors (Figure [98]S2a). The particle purity and yield from using fixed angle‐ were non‐inferior to swinging bucket rotor (Figure [99]S2b), so the fixed angle rotor was chosen for its scalability, endurance and better pelleting efficiency in the later UC step. We then optimized the sucrose layer concentration and compared between 30% and 40%, the commonly used sucrose cushions. We obtained significantly higher sEV purity when 30% sucrose was used (Figure [100]S2c). After US, sEVs remain in the sucrose layer to be harvested. Since there could be mixing at the interface between the sucrose and media layers, we hypothesized that the optimal fraction of sEVs to be collected might be more than the initial 5 mL of sucrose. Our result indeed showed that collecting a fraction of 8 mL recovered ∼10% extra sEVs than 5 mL while the purity was not affected (Figure [101]S2d). For the washing step after US, direct UC was selected as it showed the same rate of particle recovery but was more scalable than the Amicon ultra centrifugal filter 10 kDa MWCO (Figure [102]S2e). The final assembled protocol for the US step was illustrated in Figure [103]S2a. For BE‐SEC using Capto Core 700 multimodal resins, sEVs were detected in the flow‐through fraction at the same peak by all UV wavelengths 280, 260 and 230 nm (Figure [104]S3a). We then determined the enrichment capacity of a column packed with ∼5 mL resins to be less than 1.0 × 10^13 particles per 5 mL loading, as it showed significantly better purity and particle recovery compared to higher loads of 2.5 × 10^13 particles (Figure [105]S3b). This was then set as the maximal sEV loading at one time for the BE‐SEC column. After BE‐SEC, concentrating the eluate by direct UC had significantly higher particle recovery and purity than the Amicon filter (Figure [106]S3c) and hence was chosen for the final workflow. For storage, PBS is the most common buffer for sEVs but since it is not compatible with repeated transfusion, we aimed to test sEV samples resuspended in two other buffers: Ringerfundin (RF) and RF with 0.2% human serum albumin (RF+HSA). Immunoblot analysis for sEV markers showed that the stability of sEVs for short‐term storage at 4°C for 1 and 2 weeks, as well as long‐term storage at −80°C for 3 months in the RF and RF+HSA buffers was not inferior to that in the PBS (Figure [107]S4a–c). After 3 months of storage in RF buffer at −80°C, we thawed sEV samples in RF buffer and determined that the size distribution (Figure [108]S4d) and subpopulations of stored sEVs (Figure [109]S4e–f) were both comparable with freshly isolated samples. Since RF is clinically used for intravenous infusion, it was chosen as the storage buffer in the final workflow. We next aimed to compare the sEV yield and purity of four enrichment methods: (1) TFF only, (2) TFF+US, (3) TFF+US+BE‐SEC, and (4) TFF+BE‐SEC. Direct UC was performed at the end of all methods. sEV isolation by the 3‐step method TFF+US+BE‐SEC had the lowest protein contamination (Figure [110]3a) and highest purity (> 5 × 10^10 particles/μg) (Figure [111]3b) compared to sEVs isolated by other methods. The 3‐step workflow also reduced particle yield to ∼10% compared to the starting media (Figure [112]3c). Levels of cfDNA and cfRNA was the lowest after 3 steps of enrichment (Figure [113]3d). Particle size analysis by the Leprechaun system demonstrated uniform size distribution and the mean size of sEVs purified by four methods was all ∼45–47 nm, similar to sEVs in the CM before enrichment (Figure [114]3f). Specifically for the 3‐step method, NTA analysis of purified sEVs showed a mean size of 164.3 ± 84.2 nm (Figure [115]3E); SEM analysis showed intact and round morphology of sEVs (Figure [116]3g). The discrepancy in particle size measurement by different technologies was further evaluated for this same sample (Figure [117]S5a). Overall, the size of sEVs measured by ZetaSizer and NTA were comparable (>100 nm), and larger than the size estimated using SEM images and by Leprechaun system (∼40–60 nm). Furthermore, western blot analysis demonstrated that positive markers for sEVs: CD81, CD9 and TSG101 were the most enriched in the sEV preparation by 3‐step method TFF+US+BE‐SEC, while negative marker Calnexin was undetected in all preparations (Figure [118]3h). To summarize, the standardized workflow combining TFF, US using 30% sucrose and BE‐SEC using Capto Core 700 resins effectively removed protein impurities and residual cell‐free DNA/RNA with high particle recovery. On average, the yield we obtained was ∼0.5–1 mg/L; purity was ∼5 × 10^10−2 × 10^11 particles/μg; residual DNA was 0.6 ng per 10^11 particles. FIGURE 3. FIGURE 3 [119]Open in a new tab Comparing performance of four isolation and enrichment workflows. TFF; US, BE‐SEC; UC was performed at the end of all workflows. sEV isolation by the 3‐step method TFF+US+BE‐SEC had (a) lowest protein contamination, (b) highest purity (>5 × 10^10 particles/μg), (c) reduced sEV yield, (d) lowest cfDNA and cfRNA compared to other workflows (n = 3–8 replicates/group). Size distribution of sEVs as measured by (e) NTA or (f) Leprechaun system. (g) Representative SEM showed round and intact morphology of sEVs isolated by the 3‐step method TFF+US+BE‐SEC. (H) Immunoblots of sEV positive markers: CD81, CD9 and TSG101; negative marker CANX, and other potential protein impurities TUBB and ACTB. *p < 0.05, one way ANOVA with Tukey's post‐hoc test. BE‐SEC, bind‐elute size exclusion chromatography; cfDNA, cell‐free DNA; cfRNA, cell‐free RNA; NTA, nanoparticle tracking analysis; SEM, scanning electron micrograph; TFF, tangential flow filtration; UC, direct ultracentrifugation; US, sucrose cushion ultracentrifugation. 3.4. Comprehensive characterization of HEK293F‐derived sEVs To understand how different enrichment workflows could affect sub‐populations of sEVs, particles were first captured on the CD81 lane of the Leprechaun chip and analysed for expression of tetraspanin markers. While the 3‐step method seemed to enrich the CD63^+CD81^+CD9^+ sEVs slightly more than others, the overall compositions of sub‐populations from the 4 methods were relatively similar, with the CD63^+CD81^+CD9^+ population being the most abundant (∼50%), followed by CD63^+CD81^+ sEVs (Figure [120]4a). The fluorescence intensity, which semi‐quantitatively reflects the amount of each tetraspanin, was also comparable in sEV preparations from different enrichment methods (Figure [121]4b). There was no difference in size distribution of sEVs expressing CD63, CD81 or CD9 (Figure [122]4c). The analysis was repeated for sEVs captured on the CD63‐ and CD9‐ lanes of the chip and showed similar results (Figure [123]S5b–d). FIGURE 4. FIGURE 4 [124]Open in a new tab Sub‐population analysis of sEVs purified by different workflows. TFF; US; BE‐SEC; UC was performed at the end of all workflows. (a) Visualization of sub‐populations based on expression intensity of sEV surface markers CD63, CD81 and CD9. CD63^+CD81^+CD9^+ sEVs were the most abundant, followed by CD63^+CD81^+ sEVs in all sEV preparations. (b) Analysis of fluorescence intensity showed relatively equal expression levels of CD9, CD63 and CD81 in sEVs isolated by different methods. (c) Correlation of size and expression levels of CD9, CD63 and CD81 on sEVs purified by the 3‐step method TFF+US+BE‐SEC. BE‐SEC, bind‐elute size exclusion chromatography; TFF, tangential flow filtration; UC, direct ultracentrifugation; US, sucrose cushion ultracentrifugation. Subsequently, we performed proteomic profiling by LC‐MS/MS to examine the proteome of HEK293F‐derived sEVs isolated by the standardized 3‐step workflow. The total number of proteins found in sEV samples produced in two separate batches were similar (batch 1: 2592 proteins, batch 2: 2656 proteins) and consistent with HEK293T‐derived sEVs in previous proteomic studies (Figure [125]5a). sEVs from both batches shared 1749 proteins with those reported in the ExoCarta database (Figure [126]5b). The top 25 most abundant proteins were consistent between the two batches, including largely tubulins and chaperones (Figure [127]5c). Also based on the abundance, the proteins were categorized into four groups: high abundance (95th percentile: 121 proteins), medium abundance (50th–95th percentile: 1169 proteins), low abundance (5th–50th percentile: 709 proteins) and extremely‐low abundance (5th percentile: 593 proteins) (Figure [128]5d). The top biological processes associated with proteins in the high abundance group mostly involved cell cycle, protein folding, and organelle organization (Figure [129]5e). Analysis for the medium and low abundance groups also gave similar results (Figure [130]S6). FIGURE 5. FIGURE 5 [131]Open in a new tab Proteomic analysis of HEK293F‐derived sEVs produced by the optimized workflow. (a) Total number of proteins profiled in two batches of sEV production, in comparison with previous studies (Kugeratski et al., Nature Cell Biology, [132]2021; Li et al., PLoS One, [133]2016). (b) Comparison of proteins identified in our two batches with those identified in the ExoCarta database. (c) Identity of the 25 most abundant proteins in the two batches of sEVs. (d) Proteins were classified into high‐, medium‐, low‐ and extremely low‐ abundance groups. (e) GO:BP and Reactome pathway analysis of proteins in the High abundance group. GO:BP, gene ontology biological processes. We then extracted total RNA from sEVs and consistently obtained ∼3.8 ng RNA per μg protein or 5.3 × 10^−11 ng RNA per particle in different batches (Figure [134]S7a). Since miRNAs are important constituents of sEVs that can potentially impact the transcriptome of target cells, we performed next generation sequencing to profile miRNAs in sEVs produced by three separate batches (quality controls of sequencing in Figure [135]S7b–c). The total numbers of miRNAs found in three batches were similar (86, 84, 92 miRNAs) and consistent with a previous study (Figure [136]6a). Among the three batches, 65 miRNAs were found identical (Figure [137]6b) and the 20 most abundant miRNAs were also matching (Figure [138]6c). Based on the abundance, miRNAs were categorized into three groups of high‐, medium‐ and low‐ abundance and the top three were miR‐182‐5p, miR‐3960 and miR‐10b‐5p (Figure [139]6d). The target genes predicted to be regulated by these three miRNAs were shown in Figure [140]S7d and the biological processes associated with these targets primarily involved gene transcription (Figure [141]6e). The analysis for the miRNAs in the medium abundance group showed enriched pathways in metabolism and cell cycle (Figure [142]S7e). FIGURE 6. FIGURE 6 [143]Open in a new tab miRNA profiling of HEK293F‐derived sEVs produced by the optimized workflow. (a) Total number of miRNAs profiled in three batches of sEVs, in comparison with previous study (Li et al., PLoS One, [144]2016). (b) Comparison of miRNAs shared among our three batches of sEVs. (c) Identity of the 20 most abundant miRNAs identified in the three batches. (d) miRNAs were classified into high‐, medium‐, and low‐ abundance groups. (e) GO:BP and Reactome pathway analysis of target genes predicted to be regulated by the miRNAs in the High abundance group. GO:BP, gene ontology biological processes. In addition to the consistent data from proteomics and miRNA‐seq, reproducibility of the workflow was also illustrated from production phase (Figure [145]S8a) to sub‐population analysis of final sEV samples across three batches (Figure [146]S8b–d). 3.5. Safety profile of HEK293F‐derived sEVs To evaluate the safety and immunogenicity of sEVs purified by the optimized workflow, we first treated A549, MDA‐MB‐231, MKN‐45, MCF‐7 cells with different doses of sEVs and observed no apoptotic cell death in all groups (Figure [147]S9a). We next treated macrophages derived from human cord blood mononuclear cells with sEVs at the dose of 5 × 10^8 particles per mL for 4 h; LPS was used as a positive control. No upregulation of inflammatory cytokines IL‐1β, IL‐6, TNF‐α and IL‐8 was observed in the sEV‐treated group (Figure [148]7a). Human monocyte THP‐1 cells were also treated with purified sEVs at different doses for 6 h and did not upregulate any of the cytokines IL‐1β, IL‐6, TNF‐α, IL‐8 and IL‐10 (Figure [149]S9b). FIGURE 7. FIGURE 7 [150]Open in a new tab No immunogenicity of HEK293F‐derived sEVs. (a) Macrophages derived from human cord blood mononuclear cells were treated with vehicle (PBS), purified sEVs and LPS for 4 h. Expression levels of IL‐1β, IL‐6, TNF‐α, and IL‐8 were not upregulated in the sEV‐treated group (n = 3–5 replicates/group). (b) The dosing schedule of intravenous tail vein injections in immuno‐competent BALB/c mice for the vehicle group and sEV‐treated group. (c) Mouse body weights were similar between the two groups throughout the treatment. (d) mRNA levels of Il‐1β, Il‐6 and Tnf‐α in the peripheral blood leukocytes were not different between the two groups. (e) Hematoxylin and eosin staining of liver, kidney and spleen tissues from vehicle‐ and sEV‐treated mice showed no sign of inflammation. LPS, lipopolysaccharides; PBS, phosphate‐buffered saline. For in vivo examination, immune‐competent BALB/c mice were injected intravenously with either sEVs (∼10^11 particles/dose) or Ringerfundin vehicle, every other day for total six doses (Figure [151]7b). During treatment, mice injected with sEVs did not display any sign of stress or behavioural changes, and their body weights were not different from those treated with vehicle (Figure [152]7c). After sacrifice, blood samples were analysed for complete blood count and biochemistry analysis. There was no statistically significant difference between the two groups, indicating no systemic inflammation or toxicity (Tables [153]2 and [154]3). Peripheral blood leukocytes were also freshly isolated from these mice and qPCR analysis confirmed no upregulation of inflammatory cytokines Il‐1β, Il‐6 and Tnf‐α in mice repeatedly dosed with sEVs (Figure [155]7d). When mice were injected with DiR‐labeled sEVs in a separate experiment, we observed majority of distribution in the liver and spleen (Figure [156]S9c). Therefore, we examined the liver, spleen and kidney of sEV‐treated mice by H&E staining and found for no sign of inflammation or necrosis (Figure [157]7e, S9D). TABLE 2. Complete blood count results for mice injected with either vehicle or sEVs. Parameters Vehicle (n = 6) sEVs (n = 7) WBC 4.7 ± 0.8 4.7 ± 0.8 LYMPH# 3.5 ± 0.6 3.7 ± 0.6 MON# 0.2 ± 0.0 0.1 ± 0.0 GRAN# 1.1 ± 0.2 0.9 ± 0.2 LYMPH% 72.2 ± 2.7 78.4 ± 1.7 MON% 4.1 ± 0.3 3.3 ± 0.2 GRAN% 23.7 ± 2.5 18.3 ± 1.7 RBC 10.3 ± 0.9 7.2 ± 1.1 HGB 124.8 ± 10.9 103.1 ± 13.1 HCT 47.7 ± 4.1 31.2 ± 4.9 MCV 46.3 ± 0.5 45.5 ± 0.5 MCH 12.8 ± 0.6 14.6 ± 0.6 MCHC 274.7 ± 11.5 321.7 ± 11.8 RDW 14.6 ± 0.5 14.3 ± 0.5 PLT 791.2 ± 167.2 824.9 ± 179.0 MPV 6.3 ± 0.1 6.1 ± 0.1 [158]Open in a new tab Note: Data are shown as mean ± SEM. Abbreviations: GRAN#, granulocyte count; GRAN%, % granulocytes; HCT, haematocrit; HGB, haemoglobin; LYMPH#, lymphocyte count; LYMPH%, % lymphocytes; MCH, mean corpuscular haemoglobin; MCHC, mean corpuscular haemoglobin concentration; MCV, mean corpuscular volume; MON#, monocyte count; MON%, % monocytes; MPV, mean platelet volume; PLT, platelet count; RBC, red blood cells; RDW, red distribution width; WBC, white blood cells. TABLE 3. Blood biochemistry results for mice injected with either vehicle or sEVs. Parameters Vehicle (n = 6) sEVs (n = 7) ALB (g/L) 36.6 ± 1.7 37.9 ± 1.7 ALT (U/L) 164.5 ± 36.0 84.6 ± 30.6 TBIL (umol/L) 4.1 ± 0.6 3.3 ± 0.5 ALP (U/L) 75.8 ± 9.5 39.7 ± 7.6 BUN (mmol/L) 11.2 ± 0.7 9.7 ± 0.6 AMY (U/L) 816.7 ± 53.4 647.1 ± 69.4 CRE (umol/L) 53.0 ± 11.0 61.4 ± 10.9 BUN/CRE 56.7 ± 10.5 45.3 ± 8.2 Ca (mmol/L) 0.9 ± 0.2 1.0 ± 0.2 GLU (mmol/L) 14.0 ± 0.8 13.3 ± 1.2 P (mmol/L) 3.0 ± 0.2 2.8 ± 0.1 CHOL (mmol/L) 9.4 ± 0.4 9.6 ± 0.6 CK (U/L) 3000.0 ± 85.4 2759.0 ± 132.9 [159]Open in a new tab Note: Data are shown as mean ± SEM. Abbreviations: ALB, albumin; ALP, alkaline phosphatase; ALT, alanine aminotransferase; AMY, amylase; BUN, blood urea nitrogen; Ca, Calcium; CK, creatin‐kinase; CHOL, Cholesterol; CRE, Creatinine; GLO, globulins; GLU, glucose; P, phosphorus; TBIL, total bilirubin; TP, total protein. 4. DISCUSSION Standardizing sEV production and enrichment workflows has been the major challenge in the field, hindering the clinical translation of sEV‐based therapeutics. In addition, the heterogenous nature leading to unknown and unpredictable molecular composition of sEVs causes further pushback due to safety concerns. In order to address these issues, we meticulously optimized a robust quality‐controlled multi‐stage process, starting with HEK293F cell culture for sEV production to a novel 3‐step enrichment workflow. The purified sEVs were thoroughly characterized and demonstrated excellent preclinical safety profile. Firstly, maximizing sEV yield is becoming increasingly important, particularly to reduce cost at an industrial scale. Strategies to boost sEV secretion vary, including physical, chemical or biological means (Debbi et al., [160]2022). In this study, we aimed to enhance secretion capacity of HEK293F cells by fine‐tuning culture conditions instead of physical stimulation, genetic engineering or addition of chemicals. Our findings showed that the number of sEV particles released per cell increased by ∼4 times as the cells entered the stationary phase, characterized by a ∼2‐fold reduction in growth rate. The maximum density for the stationary phase was 3–4 × 10^6 cells/mL, as recommended by the cell‐line manufacturer and previous study (Hein et al., [161]2023). Such high density was not well‐supported in shake flasks, probably due to inadequate aeration and mixing. Hence, STR was chosen for cultures at high‐cell density and in large volume for sEV production. In the STR, semi‐continuous culture that prolonged stationary phase enhanced sEV production efficiency during the same time span compared to batch culture. Our result was consistent with a study by Paganini et al., in which the same culture mode boosted the space‐time yield of sEVs by 3 times (Paganini et al., [162]2023). We speculated that perfusion or continuous culture would achieve similar or better efficiency as high‐density cultivation of HEK293F cells in perfusion culture was also reported to increase the yield of recombinant proteins by 3 times (Schwarz et al., [163]2020). Secondly, for sEV isolation and enrichment, many optimized 2‐step workflows demonstrated superior performance compared to single methods (Bellotti et al., [164]2021; Stam et al., [165]2021; Visan et al., [166]2022), but the exact methods and their combined order also varied from one study to another (Table [167]S3). In some studies, SEC was recommended as the first step, followed by UC (Stam et al., [168]2021) or affinity chromatography (Bellotti et al., [169]2021), while in some others, SEC was chosen as a second step, after either TFF or UC (Visan et al., [170]2022), or ultrafiltration (Tsai et al., [171]2021). When assembling our workflows, we decided to use TFF as the default first step, largely for concentrating purpose. The 10–15‐fold reduction in volume tremendously helped disburden subsequent steps like UC and SEC that both have lower loading capacity than TFF, and hence allowed processing a large volume of starting media. TFF systems that are GMP‐compliant and capable of filtering hundreds of litres are already in use (Gimona et al., [172]2017). Moreover, TFF also allowed high particle recovery, as found in our study (up to 90%) and other studies (Busatto et al., [173]2018; Visan et al., [174]2022), compared to 50% recovery rate using dead‐end filtration (Tsai et al., [175]2021). The pore size selected for TFF was 300 kDa MWCO, as optimized previously (Corso et al., [176]2017; Visan et al., [177]2022). Since the operation of TFF is simple, the process is time‐ and cost‐efficient, we suggest TFF be a staple in any sEV isolation and enrichment workflow. For the second step after TFF, we found that UC with sucrose cushion was a better choice than chromatography as it could purify samples 2–3 times better and potentially accommodate higher volume input even though it could be less scalable. Furthermore, among different types of UC, iodixanol density gradient UC was a powerful and robust method (Duong et al., [178]2019; Zhang et al., [179]2020) but limited by a small amount of sample to layer on top of the gradient (Duong et al., [180]2019). The enrichment efficiency of density gradient UC was found comparable to SEC in a study by Lobb et al. (Lobb et al., [181]2015). Since our sEVs do not require separation from a complex matrix contaminated with animal serum proteins or plasma lipoproteins (Zhang et al., [182]2020), we decided to use US rather than gradient UC to be more time‐efficient, scalable and yet effective at removing impurities (Bordas et al., [183]2020; Gupta et al., [184]2018; Webber & Clayton, [185]2013). The type of rotors did not affect the performance of US but instead, we found that the density of sucrose cushion and the volume of fraction to retrieve sEVs were the key parameters to ensure maximal purity and recovery. This US step was the primary isolation and enrichment step in our whole workflow, and the samples isolated by TFF‐US alone achieved purity of at least 1 × 10^10 particles/μg. While there has been no consensus for purity criteria of sEVs yet, Webber et al. used the ratio of 2 × 10^10 particles/μg and below to represent ‘low’ purity (Webber & Clayton, [186]2013), which occurred in some batches using TFF‐US method. Furthermore, the level of residual cfDNA in these batches, on average, remained at 10–50 μg/L, or 5–12 ng per 10^11 particles. This exceeded the acceptable residual limits set by regulatory authorities for biological products: 10 ng/dose by World Health Organization and European Union (Knezevic et al., [187]2010) or 100 pg/dose by the US Food and Drug Administration (U.S. Food and Drug Administration, [188]1997), assuming 1 dose contains 10^11 sEV particles. Therefore, it was necessary to add a third step in the workflow to further remove impurities. The Capto Core 700 chromatography was selected because in addition to size exclusion, its multimodal resins have hydrophobic and positively‐charged octylamine ligands to bind small protein impurities, negatively‐charged oligonucleotide fragments and retain them in the column. This technology not only improved enrichment efficiency, as previously demonstrated (Corso et al., [189]2017; Nordin et al., [190]2019) but also enabled much higher sample load compared to conventional SEC (up to 100 times, according to manufacturer). Meanwhile, it still preserved the flexibility of SEC that there was no restriction in elution buffer composition like in ion exchange‐ or affinity‐ chromatography. Thus, our final workflow combined TFF, US and BE‐SEC to achieve the ultimate purity and yield of sEVs. A major limitation of this 3‐step workflow was that it did significantly reduce the sEV yield by ∼10 folds compared to the starting CM, and by 2.5–5 folds compared to other 2‐step workflows. Whether such reduction in yield could be outweighed by the improvement in purity depends on the downstream clinical application. While the highest purity is desired for intravenous infusion, the requirement for other routes of administration such as topical application or subcutaneous injection could be less stringent. Moreover, substantial removal of protein and nucleic acid contaminants could impact certain physicochemical properties of sEVs such as tendency to aggregate and plastic adsorption; as well as composition of the protein corona that ultimately influences in vivo biodistribution and cellular uptake efficiency (Görgens et al., [191]2022; Heidarzadeh et al., [192]2023). Whether these aspects are negatively affected and the solutions to improve them have to be addressed in future studies. Next, our study was the first to thoroughly characterize HEK293F‐derived sEVs. The particles remained individual, round and smooth, indicating minimal impact on integrity. CD81 was the most common tetraspanin and CD81^+CD63^+CD9^+ sEVs were the most abundant, accounting for 40%–50% of captured sEVs. This sub‐population analysis by the Leprechaun system surprisingly agreed with previous studies using fluorescence and atomic force microscopy (Cavallaro et al., [193]2021) or fluorescence‐activated cell sorting analysis (Kugeratski et al., [194]2021). We observed marginal shifts in sub‐populations among enrichment workflows and also among different batches; but we expect negligible impact on the performance of sEVs as drug nanocarriers. Proteomics and miRNAseq analysis revealed stable content of sEVs isolated in independent batches, and there were no proteins or miRNAs associated with acute toxicity or direct immuno‐modulation. Compared to previous profiling using HEK293T‐derived sEVs, we also identified low amount of epidermal growth factor receptor and Raf kinase, but not Src kinase in the proteome (Li et al., [195]2016). These molecules might be pathogenic if given at high dose, high frequency or over a long period of time; and hence should be taken into account when used in clinical practice. The overall GO and Reactome pathway analysis for both proteomics and miRNAseq shared comparable results to published literature using HEK293T cells (Kugeratski et al., [196]2021; Li et al., [197]2016). HEK293F‐derived sEVs have been shown to be safe in‐vitro (Tran et al., [198]2023; Tsai et al., [199]2021), in‐vivo by intra‐muscular injection (Tsai et al., [200]2021) and already used in human clinical trials (Estes et al., [201]2022). sEVs from HEK293T cells were also found safe with repeated dosing of 10^10 particles in mice for 3 times a week, but the first dose was administered intravenously followed by two doses of intraperitoneal route (Zhu et al., [202]2017). In this study, we injected a dose of 10^11 particles in mice for 3 times a week, all intravenously into the tail vein. The results demonstrated that sEVs isolated by our workflow did not cause any toxicity nor induce any observable immune response in BALB/c mice following sustained intravenous administration. This result, together with the content profiling, strongly highlights the potential of using HEK293F‐derived sEVs as drug delivery vehicle in clinical settings. In conclusion, our multi‐stage workflow allowed robust and reproducible production of ultrapure sEVs with stable characteristics from batch to batch and highly adaptable to mass scale production. Although this protocol would need adjustments to scale up and comply with GMP standards, the transition would be accelerated given our rigorous quality control criteria in each isolation and enrichment steps, as well as characterization criteria for sEVs that match those proposed by the International Society for Extracellular Vesicles (Lotvall et al., [203]2014). This standardized workflow would hopefully contribute to the on‐going efforts to develop sEVs as safe and reliable drug carriers for therapeutic applications. AUTHOR CONTRIBUTIONS Nhan Vo: Data curation; formal analysis; methodology. Chau Tran: Data curation; formal analysis; methodology. Nam H. B. Tran: Data curation; formal analysis; methodology; visualization. Nhat T. Nguyen: Data curation; formal analysis; methodology. Thieu Nguyen: Data curation; formal analysis. Duyen T. K. Ho: Methodology. Diem D. N. Nguyen: Data curation. Tran Pham: Data curation; methodology. Tien Anh Nguyen: Data curation; software. Hoa T. N. Phan: Writing—original draft; writing—review and editing. Hoai‐Nghia Nguyen: Supervision; writing—review and editing. Lan N. Tu: Conceptualization; investigation; supervision; writing—original draft; writing—review and editing. CONFLICT OF INTEREST STATEMENT LNT is a stockholder of NexCalibur Therapeutics, Corp. Other authors declare no conflict of interest statement. Supporting information Supporting Information [204]JEV2-13-e12454-s001.pdf^ (2.1MB, pdf) ACKNOWLEDGEMENTS