Development of autologous in vitro vascular models for use in preclinical biotherapeutic development

Key Points

• •Single‑donor human whole blood and blood fraction models of blood vessels were established for early phase drug safety assessments.
• •Models were applied in both 2D plate-based and 3D vascular structure formats for tailoring of data output to research requirements.

Visual Abstract

Adapted from figure created in BioRender. Lund, E. (2025) https://BioRender.com/phgimlr

Abstract

Following previous failures to predict drug-induced adverse immune reactions in clinical trials, for example in cases with preclinical species differences or poorly indicative in vitro assays, there has been an emphasis on developing improved preclinical hazard identification tools. Concurrently, there is a regulatory agency-backed responsibility to reduce reliance on preclinical animal models, highlighted by the Food and Drug Administration (FDA) Modernization Act 2.0 and the FDA's 2025 announcement to phase out animal testing for specific compounds. Traditional in vitro cytokine release assays utilize plastic-based formats of antibody presentation to blood cell fractions, and, although biologically simple to run, they do not accurately recapitulate in vivo blood vessel physiology. Including endothelial cells improves physiological relevance by representing the internal vascular wall, enabling cell-cell interactions, compound presentation, and cellular responses from endothelial cells alongside blood cells. Here, endothelial cells outgrown from healthy donors were cocultured with their blood cells to model the immune response to compounds. Building on existing endothelial assays cocultured with blood cell fractions, we established the model using whole blood as an alternative format. We then transferred both formats from 2-dimensional (2D) 96-well plates into a 3D microfluidics system, further mimicking the dynamics and structural microenvironment of a blood vessel. We used these human vasculature models to recapitulate the expected cytokine response to existing compounds and highlight the additional preclinical safety end points that can be investigated by using a 3D vessel, such as vascular leak. This proof-of-concept study demonstrates foundations for a scalable, physiologically relevant method for preclinical testing while reducing reliance on animal models.

Introduction

Biological therapeutics remain at the forefront of pharmaceutical product development, with the majority of therapeutics having immune-related targets.^1,2 Such immune interaction inherently carries a risk of inducing immune-related adverse outcomes (IRAOs), which may be increased with complex modalities.³ Detecting toxicities such as the cytokine release syndrome (CRS) and capillary leak often associated with immunotherapies is therefore key prior to first-in-human dosing. However, preclinical animal and in vitro safety testing does not always accurately predict the clinical response, with TGN1412 often cited as a notable example.4, 5, 6

Differential target expression and compound avidity contributed to TGN1412 species-specific responses, illustrating the lack of translatability of animal studies with first-in-human trials.7, 8, 9 Alongside physiological differences, ethical considerations equally warrant transitioning to human-based models. Although having been acknowledged previously, only recently has this gained significant regulatory support.^10,11 In 2025 the US Food and Drug Administration announced that monocloncal antibody preclinical testing no longer requires animal testing, specifically mentioning cytokine release assays (CRAs) as a crucial safety net during development.¹² This unprecedented regulatory backing reinforces the drive toward using new approach methodologies (NAMs), and the importance of sharing knowledge within the field to enable this. This work therefore seeks to provide a foundational data set to aid in such discussions at a critical timepoint within the changing climate.

Alongside animal studies, the original in vitro CRA format using human peripheral blood mononuclear cells (PBMCs) similarly failed to identify TGN1412’s CRS risk,⁷ as it does not reflect the in vivo T-cell receptor clustering involved in drug and target interactions.^13,14 Alternative CRAs have been developed which more accurately recapitulate this (Figure 1), several of which successfully detect TGN1412’s CRS risk.^7,15 This highlights the importance of understanding compound mechanism of action (MoA) and interaction with surrounding biology when designing preclinical studies. With increasingly complex therapeutic modalities being developed, combining this tailored approach with continual advancement of human in vivo recapitulation is key to generating predictive preclinical results.

Figure 1.

Increasing physiological relevance in CRA formats. Traditional 96-well plate-based CRA formats comprise immune cells (WB or PBMCs) presented with compound in solution (format 1) or immobilized onto the base of the well through dry or wet coating (format 2).⁷ Preculturing PBMCs at a higher density prior to plating in the CRA allows for a “primed” state, mimicking the involvement of lymph nodes in the immune response.¹⁵ Inclusion of an endothelial cell monolayer provides representation of the blood vessel wall, with BOECs (format 5) offering a fully autologous version of the heterologous assay achieved when using HUVECs (format 4).^7,16 Transfer of the endothelial CRA from a 2D setting onto a 3D, perfusable vessel-on-a-chip system (format 6) provides a dynamic microenvironment more recapitulative of the in vivo vessel setting. The dashed arrow represents assays compatible with PBMCs only. The full arrows represent assays compatible with WB or PBMCs. Figure created with biorender.com. Lund E. (2025) https://biorender.com/ocn3l45.

Endothelial cells are a crucial component of the immune system, secreting key CRS cytokines such as interleukin-6 (IL-6)^9,17 and displaying immune-modulatory properties.¹⁸ Their inclusion in CRAs, for example using human umbilical vein endothelial cells (HUVECs),⁷ recapitulates the vascular wall, compound presentation on endothelial Fc receptors and facilitates cell-cell interactions through expression of adhesion molecules. Reed et al¹⁶ demonstrate that blood outgrowth endothelial cells (BOECs)¹⁹ can be used to give an autologous endothelial CRA. They report a more defined cytokine response when compared to the HUVEC:PBMC CRA, citing a tissue-mismatch response when using heterologous systems,^16,20 demonstrated elsewhere when culturing endothelial cells with heterologous PBMCs.^21,22 The tissue source of endothelial cells used in in vitro assays should be carefully considered, as it affects function.²³ HUVECs upregulate proliferative genes and downregulate tight-junction proteins vs BOECs,²⁴ suggesting BOECs may be more reflective of peripheral vascular endothelial phenotypes. Utilizing BOECs also enables disease state models to be generated by using blood from patient populations to outgrow endothelial cells with disease-state phenotypes.25, 26, 27 Further, including multiple BOEC donors as opposed to a single HUVEC culture allows for potential donor variability caused by the endothelial cells themselves,²⁴ given their active role in response as previously described^9,18 and the donor-specific nature of CRS.²⁰

Building on existing BOEC models, here we establish the 2-dimensional (2D) autologous PBMC:BOEC assay in a commercial laboratory and develop a new whole blood (WB):BOEC alternative for applications requiring the full blood fraction. We transfer both assays to a perfused 3D microfluidics platform, using BOECs to generate vessel structures within the MIMETAS OrganoPlate 2-lane 96 (Figure 1). This new vessel-on-a-chip CRA model is also used to study compound-induced vascular leak, highlighting the deeper insight available when using a 3D vs the standard plate–based model, providing proof-of-concept data to support the drive toward NAM implementation in drug development.

Methods

Human blood collection

Human blood was collected at Labcorp, Harrogate (Integrated Research Application System ID 282700, Research Ethics Committee reference 21/YH/0129) into heparin vacutainers from healthy, consenting donors, in line with the UK Human Tissue Act 2004. This study was conducted in accordance with the Declaration of Helsinki.

BOEC generation

BOECs were outgrown using established methods^16,19 (Figure 2A). Lonza EGM-2 medium was prepared as per manufacturer’s instructions except for fetal bovine serum (FBS), using EBM-2 basal medium (catalog no. CC-3156, Lonza) and EGM-2 SingleQuots (catalog no. CC-4176, Lonza). FBS from CC-4176 was replaced with 10% HyClone FBS (catalog no. SH30071.03, Cytiva). Briefly, blood was collected into sodium heparin vacutainers, processed for PBMCs via density centrifugation, counted, diluted, and plated on 6-well 50 μg/mL rat tail collagen type 1–coated plates (catalog no. 354236, Corning). Media was replaced every 2 to 3 days, and colonies removed using TrypLE (catalog no. 12605-010, Gibco). Cells were expanded in T25 then T75 flasks prior to cryopreservation at passage 3 in liquid nitrogen using CryoSFM medium (catalog no. C-29912, PromoCell). Cells were thawed and cultured up to a maximum of passage 10 to maintain endothelial function,²⁷ with media replacement or passage as required every 2 to 3 days.

Figure 2.

BOEC isolation and characterization. (A) Schematic of BOEC outgrowth. (B) Representative histograms of flow cytometry BOEC panel showing unstained (gray) and stained (color) cells for CD31/PECAM1, CD34, CD144/VE-cadherin, CD146/MCAM, CD201/EPCR and CD45. (C) Immunocytochemistry on an example vessel on a chip, stained for VWF (green), VE-cadherin (red), and DAPI (blue). Image acquired using a Molecular Devices ImageXpress Micro XLS Widefield High-Content Analysis System. Figure 2A created with biorender.com. Lund E. (2025) https://biorender.com/gif4tfu.

BOEC characterization

For flow cytometry immunophenotyping, BOECs isolated from 6 donors were analyzed for endothelial markers on a BD FACSCanto II using the following antibodies: PerCP/Cyanine5.5 anti-human CD31/PECAM1 (catalog no. 303132, BioLegend), APC anti-human CD34 (catalog no. 343608, BioLegend), PE/Cyanine7 anti-human CD144/VE-cadherin (catalog no. 348516, BioLegend), Brilliant Violet 510 anti-human CD146/MCAM (catalog no. 361022, BioLegend), PE anti-human CD201/EPCR (catalog no. 351904, BioLegend), and AmCyan anti-human CD45 (catalog no. 339192, BD). Cells were stained for 30 minutes at room temperature in the dark, then washed and resuspended in autoMACS buffer (catalog no. 130-091-221, Miltenyi Biotec) for reading. Data was analyzed using De Novo Software FCS Express 7 IVD Edition. For immunocytochemistry, cells were fixed for 15 minutes in 3.7% formaldehyde (catalog no. 252549, Sigma), washed twice, permeabilized for 10 minutes, washed, and blocked for 30 to 45 minutes. Primary antibody was incubated on the MIMETAS rocker for 1 to 2 hours at room temperature, washed twice, and secondary antibody was added for 30 minutes in the dark at room temperature. Cells were washed twice, stained with DAPI (4′,6-diamidino-2-phenylindole), and washed, and phosphate-buffered saline (PBS) was added prior to image capture on the Molecular Devices ImageXpress Micro XLS Widefield High-Content Analysis System at MIMETAS. The permeabilization buffer was 0.3% Triton X-100 (catalog no. T8787, Sigma). The blocking solution consisted of 2% FBS (catalog no. A13450, Gibco), 2% bovine serum albumin (catalog no. A2153, Sigma), and 0.1% Tween 20 (catalog no. P9616, Sigma) in PBS. The washing solution was 4% FBS (catalog no. A13450, Gibco) in PBS. VE-cadherin staining was done with anti–VE-cadherin (primary, catalog no. MAB9381, Bio-Techne) and NL557–anti-mouse IgG (secondary, catalog no. NL007, Bio-Techne). Von Willebrand factor (VWF) staining was done with anti-VWF (primary, catalog no. PA560551, Invitrogen) and Alexa Fluor 488–anti-rabbit IgG (secondary, catalog no. A11070, Invitrogen). DAPI staining was performed with MBD0015 (Sigma). Antibodies were prepared using the manufacturer’s recommended dilutions.

Vessel formation

MIMETAS OrganoPlate 2-lane 96 plates were prepared as per the MIMETAS OrganoPlate 2-lane tubule seeding protocol. Channels were loaded with 4 mg/mL rat collagen-I (catalog no. 3447-020-01, R&D Systems) in 37 mg/mL pH 9.5 NaHCO₃ and HEPES (N-2-hydroxyethylpiperazine-N′-2-ethanesulfonic acid): 2 μL per extracellular matrix (ECM) channel. Plates were incubated at 37°C, 5% CO₂ for 15 minutes for collagen polymerization. BOEC cell suspension (2 μL at 2500 cells per μL) was added to media inlets followed by 50 μL of 10% FBS EGM-2. Plates were placed at ∼75° at 37°C, 5% CO₂ for 2 to 5 hours to allow cell attachment. Media was then added to media outlets (50 μL) and the plate placed on the MIMETAS bidirectional plate rocker with an 8-minute cycle to 7° inclination each way (Figure 3A). Vessels were cultured to confluency over 4 to 6 days, with complete media replacements every 2 to 3 days.

Figure 3.

Vessel-on-a-chip tubule formation. (A) Schematic of BOEC vessel-on-a-chip formation. (B-C) Vessel on a chip stained for VE-cadherin (red) and DAPI (blue), demonstrating longitudinal vessel formation (B) and cross section of a vessel showing complete vessel lumen (C). Images acquired using Molecular Devices ImageXpress Micro XLS Widefield High-Content Analysis System. (D) Representative images of dextran leakage after 2 hours using 150-kDa TRITC-dextran. Scale bar, 800 μm. Images acquired using a Sartorius Incucyte S3 Live-Cell Analysis Instrument (4× objective) at 37°C and 5% CO₂ using 2019B Rev2 software, with dextrans resuspended in culture medium. Images were exported from the Incucyte software and analyzed using Fiji ImageJ 1.47T for quantification of signal intensity. (E) Barrier integrity assessment confirming expected leak of 150-kDa TRITC-dextran from channels containing ECM only (no cells) over time, vs its containment in channels seeded with cells to form vessels. Error bars denote standard deviation of the mean. (F-G) Comparison of mean ratio of fluorescence signal intensity following 1 hour (F) and 2 hours (G) in ECM channel/vessel for fluorescent dextrans in a complete vessel vs empty channel with ECM only. The ratio value is proportional to barrier leak. An unpaired t test was performed to establish P value ∗P < .05, ∗∗P < .01, ∗∗∗ P < .001 & ∗∗∗∗P < .0001. Donor n = 4, using a minimum of 3 replicate vessels per condition per donor. Error bars denote standard deviation of the mean. Figure 3A created with biorender.com. Lund E. (2025) https://biorender.com/ruuopc7.

Cytokine release coculture assays

Vessels were cultured as described earlier. The day prior to stimulations, 96-well flat-bottom plates were seeded with 15 000 BOECs per well (same donors as used for vessels) and incubated statically overnight at 37°C, 5% CO₂ to form a confluent monolayer. All BOECs used in 2D and 3D assays were used from passage 6 to 10 inclusive to maintain endothelial function,²⁷ with 2D assays having a maximum passage increase of 3 vs 3D assays due to later seeding. On the day of stimulation, the same donors were recalled and WB collected, with an aliquot being processed to isolate PBMCs. Each vessel and 2D well was stimulated in a total volume of 200 μL, using 200 000 autologous PBMCs or 190 μL autologous WB with compounds or negative controls as per the diluent used in each assay format (10% heat inactivated human AB serum and 1% penicillin/streptomycin RPMI 1640 medium for PBMC conditions, PBS for WB conditions). Compounds assessed were 10 μg/mL anti-CD3 OKT3 (catalog no. 317347, BioLegend), 10 μg/mL anti-CD28 ANC28.1 (catalog no. 217669, EMD Millipore), 10 μg/mL anti-CD52 alemtuzumab (Campath) analog (catalog no. 15/178, National Institute for Biological Standards and Control), 10 μg/mL trastuzumab (50242-056-56, Genentech), and 10 μg/mL isotype controls (catalog no. 15/198 and 15/218, National Institute for Biological Standards and Control). Conditions were plated in a minimum of biological duplicate, using 4 independent donors across multiple occasions. Plates were incubated at 37°C, 5% CO₂ statically (2D) or on a bidirectional rocker (3D) on an 8-minute cycle (7° inclination each way). After 24 hours, plates were centrifuged at 18 000g for 5 minutes. Supernatants were harvested and frozen at ˗80°C. On thawing, supernatants were centrifuged to remove debris (3000 rpm for 10 minutes). For IL-2, IL-10, interferon gamma (IFN-γ), and tumor necrosis factor α (TNF-α) analysis, supernatants were diluted 1 in 2 using sample diluent (catalog no. SD13, Bio-Techne) and analyzed with a Simple Plex Cartridge following the manufacturer’s instructions (catalog no. SPCKE-PS-005701, Bio-Techne) using the ProteinSimple Ella Automated Immunoassay System, Runner Software version 3.9.0.28. For IL-6 analysis, supernatants were diluted 1 in 20 using calibrator diluent and analyzed using a Human Magnetic Luminex Performance High Sensitivity Base Kit following the manufacturer’s instructions (catalog no. LHSCM000, Bio-Techne) on the Luminex 200 (Bio-Plex Manager version 6.2). Data were exported to Excel (Microsoft Office Professional Plus 2016) for tabulation. Values above or below the limits of quantification were replaced with limit values, accounting for sample dilution. Data was visualized and statistical analysis performed using GraphPad Prism 9.0.1. An ordinary 2-way analysis of variance (ANOVA) of means for each condition was performed, followed by Dunnett’s multiple comparison for all treatments vs the unstimulated control.

Vessel permeability assay

Vessels were cultured as described earlier. All BOECs were used from passage 7 to 9 inclusive to maintain endothelial function.²⁷ On reaching confluency, a barrier integrity assay was performed as per the MIMETAS 2023_Protocol_Barrier Integrity Assay_V3.1 method using 150-kDa tetramethylrhodamine isothiocyanate (TRITC)–dextran (catalog no. T1287, Sigma). Images were captured on a Sartorius Incucyte S3 Live-Cell Analysis Instrument using 2019B Rev2 software (plates static). Channels not seeded with BOECs but including collagen in the ECM channel provided a positive control. Images were exported and analyzed using Fiji ImageJ 1.47T using the MIMETAS_BI_Acquisition_and_Analysis protocol. Mean signal intensity of the ECM channel was divided by the mean signal intensity of an equal area of the vessel channel to calculate a ratio proportional to the gradient of dextran between the 2 channels. Data was visualized and statistical analysis performed using GraphPad Prism 9.0.1. For verification of tubule formation, images were captured at 30-minute intervals up to 2 hours. To assess leak between complete vessels and ECM-only channels at 1- and 2-hour time points, unpaired t tests were performed to establish P value: ∗P < .05, ∗∗P < .01, ∗∗∗P < .001 and ∗∗∗∗P < .0001. Assays were performed using cells from 4 independent donors, with a minimum of 3 replicate vessels per assay condition.

Compound-induced vascular leak assay

Vessel integrity was assessed prior to compound exposure using the permeability assay described earlier. An ordinary 1-way ANOVA of ratio means after 1 hour of imaging was performed for vessels assigned to each condition, followed by Dunnett’s multiple comparisons test for conditions vs ECM only to establish P value: ∗P < .05, ∗∗P < .01, ∗∗∗P < .001 and ∗∗∗∗P < .0001. Following tubule verification, dextrans were removed, a media wash performed, and a stimulation carried out using previously banked PBMCs (200 000 per vessel in a total of 200 μL) in the presence of 10 μg/mL OKT3 (catalog no. 317347, BioLegend) and ANC28 (catalog no. 217669, EMD Millipore); 100 ng/mL IL-1β (catalog no. 201-LB-005, Bio-Techne) and TNFα (catalog no. 210-TA-005, Bio-Techne) or 10% heat-inactivated human AB serum and 1% penicillin/streptomycin RPMI 1640 medium only. Conditions were plated in a minimum of biological triplicates, using 4 independent donors across multiple occasions. Plates were incubated at 37°C, 5% CO₂ on a bidirectional rocker on an 8-minute cycle of 7° inclination each way. After 24 hours, media was removed and stored at ˗80°C for cytokine analysis. The barrier integrity assay was repeated, as previously described. An ordinary 2-way ANOVA of means for each condition 1 hour before and after stimulation followed by Bonferroni’s multiple comparisons test for before vs after stimulation within each condition, was performed to establish P value: ∗P < .05, ∗∗P < .01, ∗∗∗P < .001 and ∗∗∗∗P < .0001. Analysis of supernatants for IL-2, IL-6, IL-10, TNF-α, and IFN-γ was performed as per the method described for IL-6 in the cytokine release coculture assay, except supernatants were diluted 1 in 3. Data were exported to Excel for tabulation.

Results

Flow cytometry immunophenotyping confirmed the expected endothelial phenotype. BOECs showed positive staining for CD31/PECAM1, CD34, CD144/VE-cadherin, CD146/MCAM, and CD201/EPCR and demonstrated low expression levels of the hematopoietic marker CD45 (Figure 2B), consistent with literature,28, 29, 30, 31 although there was some variation in expression levels between donors. Immunocytochemistry (Figure 2C) similarly demonstrated expression of 2 characteristic endothelial markers: VWF in diffuse cytoplasmic foci, consistent with expression in Weibel-Palade bodies, and VE-cadherin localized to the cell surface.³⁰

Confocal imaging of vessels confirmed formation of intact tubule structures with a central lumen (Figure 3B-C; supplemental Material 1). Vascular integrity was confirmed with a permeability assay using 150-kDa TRITC-labeled dextran (Figure 3D-G). A 150-kDa molecule is large enough that it should be confined to the lumen if vessels are integral.³² By determining the ratio of fluorescence in the ECM channel vs the vessel channel, we demonstrated significantly less leak of confluent vessels vs conditions with collagen added to the ECM channel but no BOECs seeded into the vessel channel for the 2 hours that vessels were assessed (Figure 3E-G).

Graphical representation of the data in Figure 4 illustrates the similar overall pattern of response across 2D and 3D platforms for each condition within each matrix (PBMC and WB), although note that P <.05 for comparisons of 2D and 3D for PBMC assays looking at IL-2, IL-10, and IL-6 concentrations. For PBMC conditions in both 2D and 3D formats, elevated IL-2, IL-10, IFN-γ, and TNF-α responses were detected upon stimulation with anti-CD3, anti-CD28, and anti-CD52, with reduced responses from isotype and unstimulated negative controls. Although overall response patterns were conserved between 2D and 3D assays, statistical significance of positive controls was achieved more frequently in the 2D format. Trastuzumab, a compound associated with only low level or rare CRS, did not elicit significant cytokine responses in either 2D or 3D, as expected.^16,20 Elevated IL-6 background levels were observed in unstimulated conditions in both 2D and 3D formats, with 2D PBMC assays generating a statistically significant response to anti-CD52 and anti-CD3 and 3D to anti-CD3.

Figure 4.

Cytokine release assessment. Analysis of IFN-γ, IL-10, IL-2, TNFα and IL-6 (picograms per milliliter) in human PBMC cell culture supernatant (A) or WB plasma (B) following a 24-hour incubation with compounds. Both PBMC and WB assays were run in 2D 96-well plate and 3D vessel-on-a-chip CRA formats using the same donors to give a total of four distinct CRAs (2D PBMC; 3D PBMC; 2D WB; 3D WB). All 4 CRA formats were incubated for 24 hours with the same compounds: 10 μg/mL anti-CD28 ANC28; 10 μg/mL anti-CD3 OKT3; 10 μg/mL anti-CD52 alemtuzumab analog; 10 μg/mL immunoglobulin G1 (IgG1) isotype control; 10 μg/mL IgG2 isotype control; unstimulated negative controls (cell culture media for PBMC assays; PBS for WB assays). Donor responses n = 4. Statistical analysis: a 2-way ANOVA of means for each condition was performed to establish P value, ∗P < .05; ∗∗P < .01; ∗∗∗ P < .001; ∗∗∗∗P < .0001, followed by Dunnett’s multiple comparison for conditions to unstimulated controls. Data are mean ± standard deviation.

In line with the literature surrounding existing WB assay formats, the WB:BOEC coculture assay successfully detected statistically significant responses to anti-CD52 in IFN-γ and TNF-α but not IL-2 or IL-10 in both 2D and 3D formats.^33,34 Conversely our data did not detect a significant response to anti-CD52 in IL-6, however this was likely due to the high IL-6 background response observed in unstimulated conditions for all assay formats tested. No significant response to anti-CD3 was detected in WB assays, in line with previous work.³⁴ Statistically significant IL-2 and IL-10 responses to anti-CD28 were observed in 2D WB assays, but not for IFN-γ, IL-6, or TNF-α, demonstrating variable consistency with previous data.^33,34 However, this may be due to differing antibody clones, assay format differences and/or detection methods.

Following confirmation of intact vessel culture for use in drug-induced vascular leak assays (Figure 5A), a significant increase in leakage was observed in vessels exposed to both OKT3/ANC28 and IL-1β/TNF-α over a 24-hour stimulation (Figure 5B). In contrast, no significant change in permeability was observed for positive controls (ECM only, no endothelial barrier present) or negative controls (confluent vessel exposed to PBMCs and media only) as expected. Cytokine analysis of supernatants from chips used in the barrier integrity assay confirmed higher levels of IL-2, IL-6, IL-10, IFN-γ, and TNF-α in conditions that induced vascular leak in comparison to ECM only and confluent vessel controls (Figure 5C).

Figure 5.

Assessment of drug-induced vascular leak in vessel-on-a-chip model. (A) Confirmation of vessel integrity after 1 hour of 150-kDa TRITC-dextran addition to vessel lumens prior to compound exposure (x-axis labels indicate the vessels used for downstream stimulation with corresponding compounds). Statistical analysis: a 1-way ANOVA of means for each “condition” followed by Dunnett’s multiple comparisons test was performed to establish P value, ∗P < .05; ∗∗P < .01; ∗∗∗ P < .001; ∗∗∗∗P < .0001. Error bars denote standard deviation of the mean. (B) Comparison of barrier integrity before and after a 24-hour stimulation using anti-CD28 (ANC28) with anti-CD3 (OKT3), IL-1β with TNFα and controls (confluent vessel and ECM only channels). Statistical analysis: a 2-way ANOVA of means for 1 hour prestimulation and poststimulation ratios within each condition followed by Bonferroni’s multiple comparisons test was performed to establish P value, ∗P < .05; ∗∗P < .01; ∗∗∗ P < .001; ∗∗∗∗P < .0001. Ratio for A and B calculated as mean 150-kDa TRITC fluorescent signal intensity of the ECM channel divided by mean signal intensity of the vessel channel. The ratio value is proportional to barrier leak. Donor responses n = 4, using a minimum of 3 replicate vessels per condition per donor. Error bars denote standard deviation of the mean. (C) Cytokine analysis of poststimulation supernatants from vessels used in drug-induced vascular leak assay. Color scale applied to illustrate high (red), mid (yellow) and low (green) values.

Discussion

This work demonstrates the potential of using 2D and 3D in vitro vascular models, in both WB and PBMC formats, for highlighting potential risk of IRAOs. The consistent pattern of cytokine response between 2D and 3D offers flexibility in platform choice, depending on context of studies. For example, where a compound carries potential to induce vascular leak, the vessel on a chip could assess this alongside cytokines, providing a more informative data set. Capillary leak syndrome (CLS) gained attention as a significant side effect of the COVID-19 vaccine Vaxzevria³⁵ and is associated with cancer treatments commonly screened for CRS due to their immune-modulatory properties.³⁶ Detecting CLS indication early in development would therefore be useful, either as a standalone or secondary endpoint. Considering the drug failure rate of >80%,³⁷ detecting IRAOs earlier would reduce wasted time, cost and use of animals. The vessel on a chip also enables study of phenomena requiring a dynamic environment, for example, cell behaviors³⁸ or compound interactions affected by fluid flow.^39,40 Conversely, where the goal of an in vitro study is purely to derisk a compound for potential CRS (where the MoA is not expected to be dependent on a dynamic 3D environment), the 2D CRA could offer a more cost-effective approach.

With 2D and 3D platforms offering end point flexibility, matrix choice could be determined by compound MoA. WB offers an option for compounds interacting with components of the full blood fraction, for example complement or platelet activation. WB is arguably more physiologically relevant than PBMCs; however, WB CRAs have historically lacked sensitivity³⁴ and relied on compound manipulation.⁷ The WB:BOEC coculture assay circumvents this through physiological compound presentation. As a new assay, to the best of our knowledge, detection of a statistically significant cytokine response to anti-CD52 suggests the WB:BOEC assay may provide a valuable platform for relevant compounds. Of note, endothelial-free, plate-based WB assays require a sample of 8 to detect a statistically significant response to anti-CD52,³⁴ suggesting our WB:BOEC assay may be a valuable method for researchers to consider using after detecting significance from a sample of 4.

Despite consistency of results within donors and detection of statistically significant responses, variability between donors combined with the low sample size used here limits power for wider statistical significance. As is the case in the clinic, age, sex, and individualized immune responses may contribute to variability, although this would need a larger data set to investigate. Growth rates of BOECs varied between donors, and whereas only confluent vessels were used, replicate tubule formation varied within a donor. High variability was also noted with OrganoPlate 2-lane 96 used as a liver microphysiological system, with authors recognizing model functionality but the requirement for large data sets to compensate for variability,⁴¹ in line with our findings. To account for varying basal cytokine levels, it can be informative to use a fold change (vs negative control) cut-point to determine positivity, for example twofold⁴² or threefold over baseline.²⁰ This would identify more compounds as positive here; however, the risk of false positives is increased. Estimates of sample sizes required to obtain statistical significance for specific compound and assay formats highlight the need for large data sets,³⁴ which may not be financially feasible or in keeping with timelines. The typically accepted sample size in many regulatory submissions is 10 to 20,⁴³ although this varies with context of use of data, and interpretation of CRA data remains inconsistent.

Insufficient statistical power may be of greater concern for end points with lower clinical incidence. Vascular leak is associated with severe but not all cases of CRS, therefore detection in vitro may require larger data sets.⁴⁴ Similarly, the timing of vascular leak may vary between compound identity, concentration,^45,46 and likely donor. Relying on imaging-based approaches before and after a predetermined stimulation time therefore limits the ability to detect potential CLS, as vessels can recover after initial impact.⁴⁶ Real-time analysis would be advantageous, for example by using transendothelial electrical resistance (TEER) approaches successfully used elsewhere.⁴⁷ This would also resolve issues with image analysis associated with fluorescence-based approaches faced here, such as obtaining incomplete, misaligned, or out-of-focus images. Although the area of fluorescence intensity compared within each vessel and ECM channel was consistent and a ratio was calculated to normalize results, image inconsistencies between wells/time points prevented full channels from being analyzed. TEER or electric cell-substrate impedance sensing would resolve this, while also providing data throughout assay incubations and avoiding any impact to vessel integrity caused by the repeated solution changes involved in the fluorescence-based method. Due to the limitations of the fluorescence-based method, to confirm whether results were robust, we ran the same assay in a TEER-compatible 3D BOEC model (supplemental Material 2). Although the vessel structures established in the commercially available AKITA platform may not be as physiologically recapitulative as in the MIMETAS OrganoPlate, the model is similar, as it utilizes bidirectional rocking for gravity-induced fluid flow. This confirmed that the proof-of-concept data generated using the fluorescence-based approach is reproducible and demonstrated that BOEC vessel-on-a-chip models are suitable for compound-induced vascular leak studies. However, future validation would require transitioning to a platform that supports physiological vessel formation alongside more reliably quantifiable detection methods for robustness.

Although this work has generated foundational data, further work is essential to expand the data set, test robustness, and investigate potential additional applications. We noted high background IL-6 particularly in 3D, possibly due to mechanical sheer stress. Calculating cytokine response as fold-change over background can normalize this; however, the sensitivity may still be lost. Elevated background IL-6 may also have contributed to higher basal levels of vascular leak,⁴⁸ potentially causing the ratio in negative control vessels to be ∼0.5 as opposed to nearer 0 as is expected with a fully leak-tight vessel. Although a ratio of ∼0.5 enabled detectable compound-induced changes, decreasing this toward 0 would increase sensitivity.

Although the current vessel-on-a-chip model features perfusion of PBMCs or WB through the lumen, replacing bidirectional with unidirectional flow would be beneficial, for example by using MIMETAS OrganoPlate UniFlow.⁴⁹ This would potentially induce endothelial cell alignment, improving the physiological recapitulation of the model.⁵⁰ Other groups note the advantage of using BOECs vs induced pluripotent stem cell–derived autologous endothelial cells for such a response.⁵¹ Increasing shear stress through active pumping as opposed to passive gravitational flow would further recapitulate the in vivo setting. Arterial shear stress ranges from 1 to 7 Pa with veins from 0.1 to 0.6 Pa,⁵² and although the level achieved here varies due to viscosity (both PBMC vs WB, but also WB between donors due to hydration levels), it falls at the bottom end of this range. Alternative platforms offer both unidirectional and tunable flow rates⁵³; however, each platform has its own compromise. Ultimately the choice is dependent on model context of use.

It would be valuable to assess the ability of the vessel on a chip to recapitulate further healthy or disease-state clinical phenomena and investigate additional end points, for example looking at endothelial cell dysfunction, thrombosis, flow cytometry activation analysis or RNA sequencing. The latter could identify upregulated genes in response to stimuli, potentially identifying early biomarkers of adverse outcomes. Introducing compounds used to treat CRS, such as tocilizumab, would assess whether the model responds as expected beyond initial cytokine release. Using disease-state samples could recapitulate clinical phenotypes of both blood and endothelial cells, an advantage of using peripheral blood–derived BOECs from a range of donors vs the more limited source of HUVECs, providing disease-relevant models for future work.25, 26, 27^,54 Including additional channels or tissues such as tumor cells or organoids would also enable studying of cell migration, molecule diffusion and development of vascularized tissues. This would support studying of microenvironments currently difficult to model preclinically, for example chimeric antigen receptor T-cell applications. These often utilize immune-depleted animal models which are subsequently poorly predictive of immunotoxicology.

Ultimately, the translational potential of NAMs will require rigorous testing. Although validation is not a current regulatory requirement for immunostimulatory assays,¹⁰ larger data sets and robustness testing are required to ensure methods are fit for purpose. Our hope is that this initial data will support discussion of and transition to NAM implementation, which will require open dialect and collaboration between stakeholders. Organizations such as the Innovative Health Initiative, the Health and Environmental Sciences Institute and BioSafe provide effective forums for model and drug developers to discuss approaches in a collaborative environment. The European Medicines Agency Innovative Task Force and Food and Drug Administration Innovative Science and Technology Approaches for New Drugs Progam should also be used to gain acceptance of NAMs, backed by increasing regulatory support. In this way, the proof-of-concept vascular models developed here provide a foundation for commercially viable, animal-free tools for derisking immune-modulatory therapeutics, alongside further potential applications in drug development and wider industries.

Conflict-of-interest disclosure: All authors are or were employed by Labcorp and received partial funding for this project from an internal Labcorp Science Council. Labcorp has rights to intellectual property that is relevant to the blood outgrowth endothelial cell cytokine release assay described in the publication.