The microfluidic platform was fabricated by polystyrene injection molding at R&D factory (Korea). This process involved machining and polishing steps for an aluminum alloy mold, and injection molding was performed with a clamping force of 130 tons and 55 bar injection pressure for 15 s, while the nozzle temperature was 220°C. The injection-molded microfluidic platforms underwent surface oxygen plasma treatment at 5.00e−1 Torr and 50 W for 3 min and were then bonded to a pressure sensitive adhesive film.

Human umbilical endothelial cells (HUVECs; Lonza, Switzerland) were cultured in endothelial growth medium 2 (EGM-2; Lonza, Switzerland), and passages 4 and 5 were used. Lung fibroblasts (LFs; Lonza, Switzerland) were cultured in fibroblast growth medium 2 (FGM-2; Lonza, Switzerland), and passages 6 and 7 were used. All CRC cell lines were cultured in Roswell Park Memorial Institute (RPMI)-1640 medium (Gibco, USA) supplemented with 10% fetal bovine serum (FBS; HyClone, USA) and 1% penicillin-streptomycin (PS; Gibco, USA). Primary NK cells were expanded from peripheral blood mononuclear cells (PBMCs) by co-culturing with 100 Gy-irradiated K562-OX40L-membrane-bound (mb) IL-18/21 feeder cells (23). In particular, human PBMCs were isolated from healthy adult donors using density-gradient centrifugation with Ficoll–Hypaque (d = 1.077, LymphoprepTM; Axis-Shield, Oslo, Norway) and washed twice with phosphate-buffered saline (PBS) (Welgene, USA). PBMCs were co-cultured with feeder cell (6:1 ratio) in a 24-well-plate with RPMI 1640 culture medium supplemented with 10% fetal bovine serum (FBS) (Gibco), 100 U/ml penicillin, 100 μg/ml streptomycin, and 4 mmol/L L-glutamine (Thermo Fisher) containing 10 U/ml recombinant human IL-2 (Peprotech). After 1 week of culture, the concentration of IL-2 was increased to 100 U/ml, and 5 ng/ml of soluble IL-15 (Peprotech) was added to the fresh medium to induce vigorous proliferation. The medium was replaced every 2–3 days. Expanded NK cells from day 14–21 were cryopreserved and used for NK cell killing assay with live imaging when needed. No data were used for personal identification of human PBMC. The use of human primary NK cells was approved by the Institutional Review Board in Samsung Medical Center (No. SMC 2020-08-033-002).

HUVECs, LFs, and CRC cells were resuspended in EGM-2 and mixed with fibrinogen solution (final concentration of 2.5 mg/ml; Sigma-Aldrich, USA) and bovine thrombin (0.5 U/ml, Sigma-Aldrich, USA) immediately before cell seeding on the platform. The final cell concentrations of HUVECs, LFs, and CRC cells were 6 × 10⁶ cells ml⁻¹, 2 × 10⁶ cells ml⁻¹, and 0.1 to 0.3 × 10⁶ cells ml⁻¹, respectively. Of this acellular mixture, 0.9 µl was injected into the central channel and incubated for 6 min for cross-linking. Then, the side channel was filled with 3 µl of HUVEC suspension at a concentration of 3 × 10⁶ cells ml⁻¹, and the platform was tilted 90° for 15 min to allow HUVECs to cover the side of the fibrin gel. This process is repeated for the opposite side channel. The medium reservoirs were filled with 200 µl of EGM-2 supplemented with 10% or 20% FBS, and the platform was incubated at 37°C with 5% CO₂. The medium was changed every day for 5 days. The experiment under the same culture conditions were repeated twice on different days, and a total of 14 experimental samples were collected for each cancer cell line from every experiment.

The cells were fixed with 4% (w/v) paraformaldehyde (Biosesang, Korea) for 15 min and permeabilized with 0.2% Triton X-100 (Sigma-Aldrich, USA) for 15 min. EC-specific staining was performed using fluorescein-labeled Ulex Europaeus Agglutinin I (1:1,000; Vector, UK) and DyLight 594-conjugated Ulex Europaeus Agglutinin I (1:1,000; Vector, UK). Alexa Fluor 488 tagged antiepithelial cell adhesion molecule (EpCAM; Biolegend, USA) was used at a dilution of 1:200 in bovine serum albumin (BSA) for tumor-specific surface staining. Unlike other CRC cell lines, EC-specific staining used in this experiment was also positive in LoVo cells. Their merged images (green and red) show co-stained sections in LoVo cells in yellow. All the samples were maintained in PBS at 4°C until imaging.

Vascular permeability was measured at day 5. Firmly perfusable vessels were labeled with lectin (red, 1:2,000 dilution; Vector, UK), and CRC cells were labeled with EpCAM (green, 1:500 dilution; Biolegend, USA) before measurement. Regions of interest (ROIs) were selected based on the regions where both perivascular and intravascular regions were evenly distributed. The cell culture medium was removed from each well, and a solution containing fluorescence day was added to the medium reservoirs. Cascade blue (MW = 3 kDa) was used here to measure the endothelial barrier permeability. The samples were imaged every 20 s for 10 min in a live-cell imaging chamber. The vascular permeability coefficient was calculated based on the equation derived from previous research, and the final value was the average of six samples for each condition.

Time-lapse live-cell imaging was performed using a Ti2-Eclipse inverted microscope with NIS elements software (Nikon, Japan). After immunofluorescence staining of HUVECs and CRC cells for at least 3 h prior to imaging, NK cells were labeled with CellTrace™ Far Red Cell Proliferation Kit (Thermo Fisher, C34572) at a dilution of 1:1,000 in serum-free RPMI-1646 medium for 30-min incubation. The same amount of serum-included RPMI-1646 medium was added to the serum-free RPMI-1642 medium and incubated for another 5 min. The concentrations of NK cells at 0.05 × 10⁶ cells ml⁻¹ and 0.1 × 10⁶ cells ml⁻¹ were used for NK cell cytotoxicity assays. Once NK cells were introduced into the vessels by gravity-driven flow, time-lapse live-cell imaging was performed for 24 h, while the samples were maintained in a live-cell imaging chamber at 37°C with 5% CO₂.

Confocal images were analyzed using Image J. After 3D images were converted to 2D images by z-projection, the region of interest was cropped and converted to a binary mask obtained using thresholding. All image-based data analysis including the total areas of blood vessels and cancer cells was performed automatically, and the number of junctions in tumor vasculature was analyzed using Angiotool (National Cancer Institute).

Statistical comparisons were performed using GraphPad Prism software. Ordinary one-way ANOVA with multiple comparisons was used to obtain the statistical value. The p-value thresholds for statistical significance were set: *p < 0.1. **p < 0.01; ***p < 0.001; ****p < 0.0001; and ns (not significant). The standard error of the mean was presented in error bars.

Our injection-molded platform comprising 28 wells was designed for a straightforward hydrogel patterning technique and a high degree of compatibility with general laboratory equipment. The platform provides rapid and robust hydrogel patterning, which takes approximately 30 s to complete the entire row with a high success rate ( Figure 1A ). These characteristics allow the platform to be suitable for high-throughput experimentation with a variety of applications (24). The platform has three parallel microchannels as fluid guide rails designed to pattern different types of hydrogels or cell suspensions. The rails work as guiding structures, which allow fluid to flow spontaneously into the channels due to capillary forces produced on hydrophilic surfaces ( Figure 1B ) (25). The mixture of HUVECs, LFs, and CRC cells with the hydrogel solution was injected into the central channel and cultured for 5 days until constructing firmly perfusable vessels. NK cells were then introduced into the vessels through the side channel, and cellular responses in the model were monitored by time-lapse live-cell imaging for 24 h ( Figure 1C ).

High-throughput experimentation using our platform allowed us to quickly determine the optimal conditions for culturing in vitro tumor vasculature, each induced by CRC cell lines with distinct subtypes ( Supplementary Table S1 ). The sufficient amount of additional FBS in the culture media could satisfy both CRC cells and ECs to construct robust blood vessels ( Figure 2A ). The total areas of blood vessels and cancer regions were increased in all CRC cell lines when the serum concentration increased to 10% ( Figure 2B ). HT29, SW480, HCT116, and SW48 showed significant changes in the blood vessel area, which were increased by 3.09%, 7.71%, 10.43%, and 12.85%, respectively. There was also an increasing tendency of the vessel area when the serum concentration in the culture medium increased to 20% ( Supplementary Figures S1A–C ). SW480 and HCT116 showed increased total cancer areas of 5.03% and 4.57%, indicating higher proliferation rates compared to the other CRC cell lines ( Figure 2C ). Since single cancer cells in the TME continuously proliferate and form individual clusters as they grow, this result might come not only from differences in proliferation rates but also their distinctive ways to form clusters ( Supplementary Figures S2A–C ). Additionally, the number of junctions indicated connections between the branches and was higher in most cases under the condition with additional 10% FBS ( Figure 2D ). EGM-2 supplemented with 10% FBS was used as the optimal cell culture media for the rest of our experiments to obtain high reproducible and perfusable tumor vasculature in our model.

Since cancer cells produce a lot of metabolic wastes such as lactic acid (26), cancer cell concentration should be optimized in order to preserve the formation of blood vessels in a microscale channel. We first tested cancer cell concentrations ranging from 0.1 × 10⁶ to 0.3 × 10⁶ cells ml⁻¹, considering high proliferation rates of cancer cells ( Figure 3A ). As a result, there was a small difference in the total vessel area when the cancer cell concentration increased from 0.1 × 10⁶ to 0.2 × 10⁶ cells ml⁻¹. SW480 and HCT116 have vessel areas increased by 1.94% and 2.75%, respectively, at the cancer cell concentration of 0.2 × 10⁶ cells ml⁻¹ ( Figure 3B ). On the other hand, the total vessel area decreased in most CRC cell lines when the cancer cell concentration was at 0.3 × 10⁶ cells ml⁻¹. Especially, SW480 and HCT116 showed approximately 4.69% and 9.00% decreases in the total vessel area compared to the cancer cell concentration of 0.2 × 10⁶ cells ml⁻¹. The total cancer area gradually increased when cancer cells were initially seeded at high concentrations, and there was a dramatic increase in SW480 and HCT116 ( Figure 3C ). This result can indicate either the proangiogenic effect of CRC cells or the adverse effects of acidic wastes on blood vessels. We hypothesized that growth factors and cytokines such as VEGF-A secreted from cancer cells enhanced angiogenic sprouting (27), whereas excess acidic wastes accumulated in the culture medium overwhelmed the formation of new blood vessels within our platform (28). In other words, the effect of proangiogenic factors could be dominant up to a certain point, and thereafter effects of wastes secreted from cancer cells prevailed; this hypothesis was not confirmed in this study since our goal here was to maximize cancer cell cluster sizes and thoroughly preserve blood vessels in our model. The cancer cell concentration of 0.2 × 10⁶ cells ml⁻¹ was used as our optimal cell culture condition.

Each CRC cell line co-cultured with HUVECs and LFs formed morphologically and functionally different TMEs at day 5 after cell seeding on the platform. CRC cell lines have different proliferation rates and form unique aggregate shapes, which vary from spherical to star-like shapes (29). We determined the size of individual cancer cell clusters and observed that HT29, SW620, and SW48 formed relatively small clusters in near-spherical shapes ( Supplementary Figures S2A, B ). Even though the concentration of FBS in the culture medium increased to 20%, there were no significant differences in the distribution of the cluster sizes ( Supplementary Figure S2C ). HCT116 and SW480 formed wide and sparse star-shaped aggregates, and their sizes are larger than those of other CRC cell lines. The interaction between the size of cancer cell clusters and lymphocyte extravasation will be briefly discussed later in this paper on whether larger cancer cell clusters get more chances to induce lymphocyte extravasation from peripheral blood vessels to tumor sites. In addition, we evaluated vascular permeability coefficients of tumor vasculature co-cultured with each CRC cell line. Figures 4A, B show time-series confocal images of tumor vasculature, and cascade blue as a fluorescent dye was used to analyze intensity changes in perivascular regions for permeability measurement. The fluorescence intensity slightly increased at the perivascular region due to fluorescence dye leakage from blood vessel barriers. The permeability coefficient P was calculated based on the equation below:

where I_w is the length of the vessel wall that separates the perivascular and intravascular regions, I_j is the mean intensity in the intravascular region, and I is the total intensity in the perivascular region. More detailed derivation of this equation is described in our previous research (30). Fluorescence dyes started to leak from weak tight junctions of the vessels, marked by dashed white circles ( Figure 4B ). A total of six vascularized tumor models for each CRC cell line were used for quantitative analysis, and Figure 4C shows their relative permeability coefficients. Blood vessels induced by HT29 and SW480 are more hyperpermeable compared to normal vasculature without cancer cells and other cell lines, demonstrating differences in vascular permeability between CRC cell lines.

Using our tumor vasculature model, we demonstrated NK cell cytotoxicity assays according to the subtypes of CRC. Since cytotoxic lymphocytes such as NK cells play a critical role in immune surveillance and continuously circulate in the blood to monitor and destroy cancer cells (31), in vitro validation of their cytotoxicity should be performed in the complex 3D TME including blood vessels. We compared the cytotoxicity of primary NK cells on a conventional 2D cell culture dish to our 3D culture platform. Primary NK cells used in our assays were expanded from PBMCs using genetically engineered feeder cells developed in previous research (23). After NK cells were added to the 2D cell culture dish, most cancer cells were dead within 4 h ( Supplementary Figure S3A ), while NK cells in our 3D tumor vasculature model had relatively slower cytotoxic effects, which involved additional processes including NK cell extravasation and intravasation ( Supplementary Figures S3B–D ). NK cells migrated toward the tumor site across the vascular barrier and killed CRC cells, and it took approximately 3.5 h from extravasation to intravasation in this assay ( Figure 5A ). Then, we observed distinct cytotoxic activities of NK cells according to the CMS classification. CRC cell lines with CMS1, CMS3, and CMS4 were used in this assay ( Supplementary Table S1 ). Blue fluorescence indicating dead cells appeared on cancer cells immediately after the attacks by NK cells, and the cytotoxic effect of NK cells was barely observed toward endothelial cells and fibroblasts ( Figure 5B and Supplementary Videos S1 – S6 ). As a result, NK cells showed much higher cytotoxic activity against LoVo and SW48 (CMS1) compared with other CRC cell lines. There were no significant differences between CMS3 and CMS4 CRC cell lines in the cytotoxic activity. SW620 slightly induced NK cell cytotoxicity; however, a notable result was not found ( Figures 5B, C ). In addition, we used 0.05 × 10⁶ cells ml^-1 and 0.1 × 10⁶ cells ml⁻¹ of NK cells to validate their cytotoxicity according to the ratio of NK cells to cancer cells ( Figure 5B ). Under the condition with the concentration of 0.05 × 10⁶ cells ml⁻¹, LoVo cells were mostly killed after 24 h ( Figure 5B ), while NK cells at the concentration of 0.1 × 10⁶ cells ml⁻¹ required a relatively short time (i.e., <10 h) to eliminate LoVo cells; however, their cytotoxicity against HT29 remained low ( Figure 5B ). This result shows that a high ratio of NK cells to cancer cells provides more rapid cytotoxic responses to cancer cell, but the trends of NK cell cytotoxicity against different subtypes of CRC cells are similar regardless of the ratio. Intriguingly, among the number of extravasated NK cells, not all NK cells were involved in the cytotoxic activity; only a low number of NK cells resulted in the cytotoxic effect and low viability in cancer cells. The number of extravasated NK cells is plotted with the number of cancer cells killed by NK cells for HT29 (CMS3), SW48, and LoVo (CMS1) ( Figures 6A–C ). The results confirmed that CRC cells with CMS3 induce fewer chances of NK cell extravasation compared with other subtypes, and even though NK cells extravasated into the extracellular matrix around CRC cells, their killing capacity varied with the subtypes and resulted in different cytotoxic effects of NK cells against CRC cells as shown in the time period between 10 and 20 h on the graph. Both the number of extravasated NK cells and the death rate of CRC cells for SW48 and LoVo increased exponentially, but in the case of HT29, the number of extravasated NK cells slightly increased over time, while the death rate of CRC cells showed no significant changes ( Figures 6A–C ).