Pooled human umbilical vein endothelial cells (HUVECs) were purchased from Thermo Fisher Scientific. HUVECs were maintained in an Endothelial Growth Medium (EGM-2, Lonza) containing 2% fetal bovine serum (FBS). Normal adult dermal fibroblasts (Thermo Fisher Scientific) were grown in a human fibroblast expansion basal medium supplemented with a low serum growth supplement (Thermo Fisher Scientific). All primary cells were used between passages 4 and 8. Two adenocarcinoma breast cancer cell lines: MCF7 and MDA-MB-231, were obtained from the ATCC. Both cell lines were grown in DMEM media with 10% FBS and 1% Penicillin-Streptomycin mixture. All cell cultures were maintained in an incubator with 37°C/5% CO₂. We used Accutase^® solution (Sigma) for all cell expansion procedures.

The targeted anti-cancer therapy drug library (HY-L080) was purchased from MedChemExpress and comprised 99 targeted therapy drugs used in targeted cancer therapy. The compounds were arrayed in 96-well plates at a concentration of 10 mM in DMSO solution. Compounds were diluted to obtain a final concentration of 100 μM in an intermediate 96-well plate. The compounds were diluted in EGM-2 media to a final concentration of 5 μM for screen use. DMSO 0.05% solution was used as a vehicle.

The GFP was expressed at the whole cell level, allowing the tumor cells to be labeled for cell-tracking studies. Both tumor cell lines: MCF7 and MDA-MB-231, stably expressed GFP by transduction of a pre-made eGFP-puro-CMV-lentiviral vector (GeneCopoeia). Cell lines were incubated with puromycin every 3–4 days until drug-resistant colonies became dominant.

To establish a perfused microvessel in culture, we used the OrganoPlate^®-3-lane (Mimetas), a microfluidic organ-on-chip platform established on a 384-well plate format with 40 or 64 microfluidic chips. Briefly, 2.2 μl of type I collagen with a final concentration of 5 mg/ml (stock solution of 10 mg/ml rat tail type I collagen (Corning) was neutralized with 10% 37 g/l Na₂CO₃ (pH 9.5) and 10% 1 M HEPES buffer) and was loaded in the middle channel of the chip. The plate was incubated for 15 min at 37°C, 5% CO₂, until the polymerization of collagen gel. Next, 40 μl of fibronectin in PBS with a final concentration of 10 μg/ml was added into the inlets of both the perfusion channels. The next day, endothelial cells were resuspended in EGM-2 media at 10,000 cells/μl. After the washing step with PBS, 50 μl of the EGM-2 media was added to the outlet of perfusion channels. Next, 2 μl of cell suspension was dispensed into the inlet of one of the perfusion channels (microvessel perfusion channel). The plate was then tilted for 2-3 hours in a humidified incubator before adding 50 μl of EGM-2 media to the perfusion inlet wells. Plates were moved to the Mimetas rocker platform in the cell culture incubator, where they were exposed to fluid shear stress induced by a bi-directional pulsatile flow generated by leveling (31). The leveling was set at a 14° angle and 8 min interval, corresponding to a pulsatile flow with 2.5 dyne/cm². These settings were maintained during the experiment period. The EGM-2 growth media was changed every 48 hours. OrganoPlates^® were incubated for 4 days until a monolayer endothelial tube (microvessel) formed. A mixture of cell suspension of GFP-labelled cancer cell line (4x10⁶ cells per ml) and adult dermal fibroblasts (4x10⁶ cells per ml) were prepared in DMEM media. 3 μl of cell suspension mixture was added into the inlet of the opposite perfusion channel (tumor cell channel). OrganoPlates were then incubated for 2-3 hours at 37°C, 5% CO₂ until complete cell attachment. Then, 50 μl of DMEM growth media was added into the inlet and outlet of tumor perfusion channel. Plates were then moved to the Mimetas rocker platform in the cell culture incubator. HUVEC grown in the microvessel perfusion channel were maintained using the EGM2 media, whereas the MDA-MB-231/fibroblasts grown in the tumor perfusion channel were maintained using DMEM media.

Drug screening was conducted using the optimized metastasis-on-chip assay. One day after tumor cells/fibroblasts seeding, compounds were added at a final concentration of 5 μM into the inlet and outlet microvessel perfusion channels. The drugs were replenished once after three days. According to the manufacturer’s recommendation, cell culture in the OrganoPlate^® requires medium culture refreshment every 2-3 days. Besides, several previous studies that used OrganoPlates^®, the media refreshment interval ranged from two days up to six days. Six days after tumor cell seeding, the chips were fixed and stained for cell imaging.

All cell staining procedures in OrganoPlate^® were performed per the manufacturer’s instructions (32). Briefly, OrganoPlates^® were fixed using 4% paraformaldehyde in PBS for 15 min and washed three times with PBS for 5 min. OrganoPlates^® were incubated with blocking and permeabilization buffer containing 3% Bovine Serum Albumin and TWEEN 0.03% in PBS for 30 min. Next, the chips were incubated with the primary antibody, mouse anti-human VE-Cadherin (1:30, Santa Cruz) overnight at 4°C. The chips were washed twice for 5 min with PBS and then incubated with the secondary antibody: donkey anti-mouse Alexa Fluor 568 (Thermo Fisher Scientific) for 1 hour at room temperature. Hoechst 33342 (Thermo Fisher Scientific) was used for nuclear staining, and Deep Red Cell Mask (Thermo Fisher Scientific) was used to delineate the cell body.

OrganoPlates^® were imaged with a 10× objective using the CLS operetta high content microscopy system (PerkinElmer). Images of chip channels were taken under spinning disk confocal mode with 40 z-steps with 5 µm spacing. Four adjacent view fields were acquired to cover chip channels, and four others were acquired to cover the microvessel perfusion inlet and outlet. The chip confocal image tiles were stacked and stitched to perform the image segmentation and quantification using the built-in Harmony^® image analysis software ( Figure 1E ). For screen data analysis, all quantified data were normalized to the negative controls of each plate.

GFP-labelled MDA-MB-231 cells were seeded at a density of 5000 cells in 96-well format plates. The next day, cells were incubated with the compound library at a final concentration of 5 μM for 48 hours. We used DMSO (vehicle) and 5 μM Sunitinib (TOCRIS) as negative and positive controls, respectively. The screen was conducted in duplicate. Screen plates were imaged immediately (t0), 24 h, and 48 h after compound addition using a controlled environment chamber of the CLS Operetta high content imaging system (PerkinElmer). Nine view fields were acquired for each well using a 10x objective. Based on GFP cell labelling, cell number was quantified using Harmony^® image analysis software. The duplicate data were averaged and then normalized to their respective controls at 0h ( Supplementary Figure 2 ). Next, each drug’s data was further normalized to the mean value of the vehicle (DMSO) negative control of the respective plate.

GFP-labelled MDA-MB-231 cells were seeded at 18,000 cells per well in 96-well plates fitted with stoppers (Oris™ Cell Migration Assay, Platypus Technologies). Cells were incubated overnight at 37°C and 5% CO₂ before removing the stoppers. After removing the stoppers, the library compounds were added at a final concentration of 5 μM for 48 h. DMSO (vehicle) and 5 μM ROCK inhibitor (GSK429286, TOCRIS) were used as negative and positive controls, respectively. The screen was conducted in duplicate. The plates were imaged at hour 0 and hour 48 upon removal of stoppers using the CLS Operetta high content imaging system (PerkinElmer). Nine view fields were acquired for each well using a 10x objective. The view field image tiles were stitched into one image for quantification. The number of migrated cells into the gap was quantified using Harmony^® image analysis software and NIH Image J software ( Supplementary Figure 3 ). Next, each compound’s data value was normalized to the mean value of the vehicle (DMSO) negative control of the respective plate.

We employed the uniform manifold approximation and projection (UMAP) method to visualize and characterize clusters of drug treatments based on the following normalized phenotypic parameters: intravasation, invasion, proliferation, migration, endothelial cell number, and tumor cell number in tumor channels. UMAP-based dimension reduction to two dimensions was carried out in R-studio using the open-source ‘umap’ package (https://cran.r-project.org/web/packages/umap/vignettes/umap.html) (33). Following uploading the dataset into R-studio, we tuned several parameters from the default settings of the ready-to-use algorithm to determine optimal settings. We found that selecting a low value for the minimum distance as 0.1 and the number of neighbors as 6 generated clear separations between clusters. Both parameters are related to two matrices: one matrix links each data point to a fixed number of nearest neighbors, and the second matrix determines the distances between the neighboring data points. The UMAP results were then plotted in Prism GraphPad version 9. Next, the normalized phenotypic parameters of drugs in each cluster were averaged to create an illustrative heatmap for the cluster phenotypic features.

To evaluate the integrity of endothelial cell monolayer, trans-endothelial electrical resistance (TEER) was measured for HUVEC seeded on fibronectin-coated 24-well transparent transwell inserts with a 0.4 µm pore size, as previously described (34). The Millicell^® ERS (Electrical Resistance System, Merck) was used to measure total resistance (Ω). HUVEC were seeded at a density of (5 × 10⁴ cells) into the insert (upper chamber), and 1 ml of medium per well was added in the lower chamber. Cells were grown until confluency, confirmed by observing the cobblestone-like appearance using phase contrast microscopy (4 days after seeding). A mixture of MDA-MB-231/fibroblasts (5 × 10⁴ cells for each cell type) was seeded in the bottom chamber. Two days later, the electrical resistance was measured (in triplicate) per well. The mean value was then used for the final data calculation. The values obtained from a blank insert (no cells) were subtracted to calculate TEER to give the net resistance. The resulting value was normalized by multiplying it by the surface area of the insert (0.32 mm²).

HUVEC were seeded in the perfusion inlet of the OrganoPlate^®-3-lane (Mimetas) to form an endothelial tube as described above. Cells were maintained in the EGM-2 media for four days until the formation of a monolayer endothelial tube. At day 5, EGM-2 media was removed, and Imatinib (5 μM) or Vehicle (DMSO) was added and incubated for 2 hours at 37°C with 5% CO₂. 50 µL of GFP-labelled MDA-MB-231 cell suspension with a density of 100,000 cells per ml was added into the inlet and outlet of the microvessel perfusion channel, followed by incubation for 30 minutes on the Mimetas rocker platform. The chips were washed with EGM-2 media, and then fresh EGM-2 media with Imatinib (5 μM) or Vehicle (DMSO) were added to the inlet and outlet of the microvessel perfusion channel. The OrganoPlate^® was then incubated at 37°C with 5% CO₂ on a rocking platform for three days. Next, cells were fixed and stained with anti-VE-cadherin and Hoechst 33342.

Statistics were carried out using Prism GraphPad version 9. Two treatment groups were compared by unpaired t-test. Multiple group comparisons were analyzed by one-way or two-way analysis of variance (ANOVA) with post hoc Bonferroni’s test. Three to six independent experiments were performed to guarantee the reproducibility of the findings. To determine the quality and robustness of the high throughput functional assays, we will calculate their robust Z’ factor statistical parameter (35) using positive and negative controls. We applied Fisher’s exact test to determine whether a molecular target was selectively enriched in a drug cluster. A p-value of < 0.05 was considered to be statistically significant.

To model cancer metastasis with TME pathophysiologic relevance ( Figure 1A ), we used OrganoPlate^® 3-lane plate from Mimetas, a high-throughput platform containing 64 individual chips. This microfluidics system allows continuous perfusion of nutrient-rich media at a shear stress of around 2.5 dyne/cm², sufficient to mimic the microcirculation of capillaries at tumor intravasation sites (36). In this chip design, when the middle channel was dispensed with type I collagen gel via its inlet, two lateral channels were formed due to the phase guide-controlled capillary forces (37). The two compartments simulated the extracellular matrix and blood vessel compartments surrounding tumors ( Figure 1B ). The collagen gel represents the extracellular matrix (ECM) of tumor fibrotic stroma. We created a microvessel by seeding HUVEC in one of the formed perfusion channels coated by fibronectin. Stromal fibroblasts play a critical role in the TME, influencing tumor invasiveness and metastasis, and are the most abundant cell component of TME (38). Thus, we seeded a mixture of adult stromal fibroblasts and GFP-labelled MDA-MB-231, an aggressive metastatic breast cancer cell line in the opposite perfusion channel following the formation of complete endothelial monolayer tube (microvessel). We found that the mixture of adult fibroblasts and MDA-MB-231 with a (1:1) ratio produced the maximal response of MDA-MB-231 for both cell invasion and intravasation process ( Supplementary Figure 1 ). Over five days, GFP-labelled tumor cells invaded the ECM gel along with stromal fibroblasts toward the microvessel wall. At day 3 or 4 after tumor cell seeding, several GFP-labelled tumor cells were observed to intravasate across the microvessel wall. Additionally, we could detect the intravasated tumor cells entered the hollow lumen of the microvessel and traveled within the media flow through the channel to adhere to the HUVEC monolayer at the inlet and outlet surface of the microvessel perfusion channel ( Figure 1C ). To confirm the formation of a continuous 3D endothelial barrier, we used the vascular endothelial (VE)-cadherin antibody to visualize the endothelial cell junctions. We found that endothelial cells formed a confluent monolayer covering the whole perfusion channel and the ECM gel-microvessel interface in the presence of invading and intravasating tumor cells. Z-stack images of the ECM gel/microvessel interface were 3D reconstructed to monitor the tumor cells transmigrating the endothelial monolayer ( Figure 1D , Supplementary Movie 1 ). The chips were imaged to quantify tumor cell intravasation and invasion using the high-content screening system. All nuclei were segmented based on the Hoechst channel. Next, total tumor cells were identified based on their positive GFP fluorescence. The different tumor cell populations were selected based on their position into invaded tumor cells, intravasated tumor cells, and tumor cells in the tumor channel. The endothelial cells of microvessels were also segmented based on their positions in the chip and the absence of cellular GFP fluorescence. The intravasated tumor cells that traveled and seeded in the microvessel perfusion inlets and outlets were identified based on the nuclei staining channel and their positive GFP fluorescence ( Figure 1E ). Thus, four phenotypic parameters were possible to be calculated, including intravasated tumor cell number, invaded tumor cell number, endothelial cell number, and total tumor cell number in the tumor channels.

To assess the effect in our model, we perform a time course experiment to monitor the interaction between fibroblasts and tumor cells in the metastasis-on-chip assay. When cultured alone or co-cultured with fibroblasts, we found that MDA-MB-231 cells invaded the ECM gel towards the microvessel. Nevertheless, in the case of co-culture with fibroblasts, tumor cell invasion was enhanced as assessed by two parameters: invaded cell number and mean cell invasion distance all over the five days of the assay. The invasion of stromal fibroblasts preceded the tumor cell invasion, indicating that fibroblasts facilitate and lead the tumor cell invasion ( Figure 2A ). MDA-MB-231 invaded the ECM with a collective type of migration. However, when MDA-MB-231 co-cultured with fibroblasts, they invaded the ECM with a single-cell type of migration as visualized by their distribution of invaded tumor cells within the Z-depth of the ECM gel, and by the increase of the SD of cell invasion distance ( Figure 2B ). We detected a few tumor cells (1-3 cells per chip) intravasated the microvessel after three days, reaching 10% of the total invaded cell number at day 5. In contrast, we did not detect any tumor cell intravasation when MDA-MB-231 was cultured alone during this time frame ( Figure 2A ). Notably, the co-cultured fibroblasts were observed to reach the microvessel wall edge on day 2 before MDA-MB-231 cells. However, we did not detect any fibroblasts intravasated inside the microvessel or on the inlet/outlet of the microvessel channel over the experiment period. Next, to assess the model’s selectivity of tumor cells’ metastatic nature, we used the MCF7 breast cell line, which is known as a non-invasive human breast cancer cell line that expressed estrogen and progesterone receptors. After five days of tumor cell seeding, we found that MCF7 remain incapable of invading the ECM gel. The co-culture with fibroblasts did not promote significantly the MCF7 invasion of ECM as compared to MDA-MB-231 ( Figure 3 ).

To determine whether our metastasis-on-chip model represents a suitable preclinical platform to test drug sensitivity against cancer metastatic activity, we applied our model to a high content screen of a compound library for annotated targeted anti-cancer therapy. First, as part of the screen workflow, we screen the drug library at a final concentration of 5 μM in live tumor cell proliferation assay ( Supplementary Figure 2 ) and 2D migration assay ( Supplementary Figure 3 ). The drugs that reduced the tumor cell proliferation over 48 hours > 60% were excluded from testing in the subsequent metastasis-on-a chip screen. Since the Rho small GTPases regulate many cellular processes, including cytoskeletal dynamics, cell motility, and cell polarity, which are essential for the metastatic potential of tumor cells (39, 40), and previous studies have proposed the potential of ROCK-targeted therapy for cancer metastasis (41), we used a selective ROCK inhibitor (GSK429286) as positive drug control in the metastasis-on-a chip screen. To confirm the selected ROCK inhibitor efficacy, we evaluated its incubation in the microvessel perfusion or tumor opposite channel on the MDA-MB-231 cell invasiveness. GSK429286 reduced both tumor cell invasion and intravasation when it was added in the microvessel perfusion channel, which would be relevant to the in vivo conditions ( Supplementary Figure 4 ). Having established that the positive compound targeting ROCK1/2 activity known to be important in metastasis can also reduce MDA-MB-231 invasiveness in our model, we conducted a screen of the selected 86 drugs. HUVEC were seeded in the perfusion channels of the OrganoPlates^®, in which they formed a microvessel over four days. Subsequently, a mixture of GFP-labelled MDA-MB-231 and fibroblasts with a ratio of 1:1 were added into the opposite channels and incubated for 24 h. The selected library drugs (86 compounds) were then added at a single concentration of 5 μM to the microvessel perfusion channels and incubated for three days. Next, the culture media and drugs were replenished and incubated for additional two days. ROCK1/2 inhibitor (positive control, 5 μM) and DMSO (negative control, 0.05%) were added to each screening plate. The chips were then fixed, stained and imaged using the CLS Operetta high content imaging system. Image data were processed and quantified with the built-in Harmony^® image analysis software as described above. Four phenotypic parameters were extracted from the images of each chip: invaded tumor cells, intravasated tumor cells, total tumor cells in the tumor perfusion channels, and the total endothelial cells in the microvessel channels. The invaded and intravasated tumor cell numbers were normalized to the total tumor cell number of the tumor channels for each chip. All parameters were then normalized to the negative control mean (vehicle, DMSO) for each plate. We calculated an average Z′ factor of 0.2 between negative and positive controls. The screen identified 30 drugs inhibiting the MDA-MB-231 cell intravasation < 30% relative to the negative controls, which is equivalent to [mean_vehicle-4xSD_vehicle] ( Figures 4A, B ). The identified drugs are targeting BCR-ABL, MEK1/2, RAF1, PDGFR, VEGFR, KIT, and PARP ( Figure 4C ).

The detected effect of drugs on tumor cell intravasation could also be a secondary effect to other cellular mechanisms of metastatic cascade, such as tumor cell proliferation, motility, and invasion. In addition, we observed that a subset of drug library showed substantial cytotoxicity against endothelium. Therefore, we sought to conduct a clustering analysis to characterize different drug categories based on their phenotypic features resulting from the tumor cell proliferation, 2D migration, and metastasis-on-chip screens. We used UMAP clustering analysis that combined alterations in the following phenotypic descriptors: tumor cell proliferation, tumor cell 2D migration, 3D tumor cell invasion, tumor cell number, tumor cell intravasation, and endothelial cell number. This multi-parametric phenotypic analysis resulted in clustering the drugs into eight distinct groups ( Figure 4D ). The vehicle (negative control) was located in the cluster 2. The influence of compounds in this cluster on tumor cell proliferation and metastatic characteristics was indistinguishable from negative controls. The cluster included many hormone drugs such as Tamoxifen, Fulvestran, and Anastrozole. The compounds in clusters 1 & 3 moderately affected both proliferative and metastatic activity for MDA-MB-231. Indeed, most EGFR and HER2 inhibitors were clustered into clusters 1, 2 & 3. In contrast, compounds in cluster 8 caused the decrease of intravasation, which was associated with a reduction of tumor cell invasion. In this cluster, the reduction of tumor intravasation appeared to be not due to a direct effect on endothelium/tumor cell interaction but somewhat secondary to the cell invasion decrease. The molecular target MEK was enriched in this cluster. We further detected that while the compounds in cluster 7 inhibited the tumor cell intravasation and invasion similarly to the compounds in cluster 8, they also appeared to have a substantial effect on cell proliferation and migration. Both clusters showed substantial effects of endothelial cell toxicity. VEGFR and SRC drug targets were enriched in cluster 8. The compounds in cluster 4 showed reduced metastatic activity, including tumor cell intravasation, invasion, and 2D migration, mildly affecting tumor cell proliferation. ROCK1/2 inhibitor (positive control) was detected in this cluster. In contrast, the compounds in cluster 5 showed around an 80% reduction of tumor cell intravasation with limited effect on tumor cell proliferation, migration, and invasion. Also, these compounds showed a slight increase in the endothelial cell number of the microvessel compared to negative controls ( Figures 4D–F ). Since improving the endothelial barrier function and inhibiting tumor cell-endothelium interactions are the key cellular events to target for preventing cancer metastasis, we focused on cluster 5 for the follow-up studies due to their potential mechanism of action related to endothelium/tumor cell interaction. The molecular targets: BCR/ABL, KIT and PDGFR were enriched in this cluster. We retested three drugs in five biological replicates from this cluster to validate their response. Such compounds were Sorafenib, Imatinib, and Abiraterone acetate. While Sorafenib and Abiraterone significantly reduced both tumor cell intravasation and invasion, Imatinib significantly reduced only the tumor cell intravasation, indicating Imatinib’s anti-metastatic effect could be related to the interaction between tumor cells and endothelium ( Figure 5 ). Therefore, we investigated whether Imatinib affects the vascular mechanisms that are involved in cancer metastasis.

Abnormal endothelial barrier function promotes endothelium permeability. Invaded tumor cells tend to intravasate at the sites where endothelial permeability is increased (42). It has been demonstrated in previous studies that Imatinib decreases vascular permeability (43, 44). Since we found that Imatinib reduced the MDA-MB-231 cell intravasation without significant effect of cell invasion, we investigate whether Imatinib can enhance HUVEC monolayer integrity and permeability when they are co-cultured with MDA-MB-231 cells/fibroblasts. To evaluate the effect of Imatinib on endothelial barrier function, we measured transendothelial electrical resistance (TEER) in a secondary assay using the co-culture transwell system. The co-culture of endothelial cell monolayer with the mixture of MDA-MB-231/fibroblasts significantly decreased the endothelial barrier function compared to endothelial cells cultured alone. The co-culture treatment with Imatinib (5 μM) prevents the reduction of endothelial barrier function caused by the co-culture with tumor cells/fibroblasts ( Figure 6A ). VEGF is a potent permeability factor. It increases VE-cadherin phosphorylation leading to its internalization and, subsequently, the loss of the cell-cell contact (45). Confluent HUVEC monolayers were treated with vehicle or VEGF at 5, 10, and 20 ng/ml concentrations. At 48 hours, the TEER measurement showed a significant reduction (20%) from the concentration of 10 ng/ml, compared with controls ( Supplementary Figure 5 ). The treatment of confluent HUVEC monolayer with 10 ng/ml of VEGF and 5 μM Imatinib were not significantly different from those of controls, indicating that Imatinib could stabilize HUVEC monolayer and inhibit of VEGF-induced permeability effect ( Figure 6A ). The circulating tumor cells interact with the endothelium, destabilizing the endothelial barrier, which is required to accomplish tumor cell extravasation (12). Therefore, we tested whether Imatinib prevents endothelial cell monolayer disruption by seeding MDA-MB-231 tumor cells into the microvessel perfusion channel with a confluent monolayer of HUVECs in OrganoPlate^®. Tumor cells adhered to the endothelium and progressively disrupted the endothelial barrier by dislodging endothelial cells and cell expansion over 72 hours, as previously described (46). We found that the endothelial cell number and endothelium area were preserved more in the Imatinib-treated chips than in controls ( Figure 6B ), and many tumor cells remained on the endothelium surface in Imatinib-treated chips. These findings imply that Imatinib reduced the disruption of the endothelial cell monolayer, which is an important effect in preventing the intravasation and extravasation of tumor cells.