Systematic Evaluation of Gastric Tumor Cell Index and Two-Drug Combination Therapy via 3-Dimensional High-Throughput Drug Screening

Tumor heterogeneity greatly limits personalized treatment of cancer. Patient-derived tumor cell (PDC) models precisely recapitulate the molecular properties and biology of the disease, making them effective preclinical tools for assessing anti-cancer drug activities. Accurate estimation of tumor purity is essential for performing high-throughput drug screening (HTS). In the present study, we measured and predicted the tumor population index in PDC models for two-drug combinational strategies using HTS system. Gastric cancer cell-lines and PDCs were subjected to multi-color immunofluorescence analysis against EpCAM and vimentin to evaluate the tumor cell index based on EpCAM expression levels. We generated a tumor purity prediction model using five different gastric cancer cell-lines (AGS, KATO-III, MKN-45, NCI-N87, SNU-216) with fluorescence intensity-based techniques. Afterwards, stage IV gastric cancer PDC models were evaluated using a micropillar/microwell chip-based HTS system. HER2/CCNE1-amplified PDCs were considerably resistant to an HER2 inhibitor, while combinational treatment consisting of an HER2 inhibitor with anti-WEE1 compound substantially suppressed tumor cellular growth. Moreover, PDCs with BRCA1/2 mutations were synergistically sensitive to HER2 and PARP inhibition therapy. Finally, somatic mutations in TP53 and CDKN2A with MYC amplification rendered PDCs susceptible to the drug combination of WEE1 and HER2. Collectively, our systematic method of high-throughput drug sensitivity screening is an integral pre-clinical platform for evaluating potential two-drug combinational approaches for personalized treatment of cancer.


INTRODUCTION
Gastric cancer is the third leading cause of cancer-related mortality. The current standard-of-care for patients with gastric cancer provides only palliative treatment despite the availability of curative surgery (1). Although previous studies performed comprehensive molecular characterization of gastric adenocarcinomas based on four specific molecular subtypes (2,3), the clinical application potential of molecular targeted therapy for personalized treatment remains unclear. Based on the clinical success of trastuzumab in patients with HER2-overexpressing advanced gastric cancer (4), numerous phase III clinical trials, concomitant with other targeted agents, have been initiated, but have shown disappointing outcomes (5)(6)(7)(8)(9)(10)(11). An essential contributing factor to such termination is the extensive molecular and transcriptional heterogeneity of gastric cancer.
Conventional pre-clinical tools for evaluating pharmacological drug responses primarily rely on twodimensional cultured cancer cell-lines or patient-derived xenograft (PDX) models. While both models have been used widely and extensively for translational cancer research, several challenges limit their direct clinical utility. First, traditional cancer cell-lines may not recapitulate the unique genetic background of each patient. Second, while PDX systems retain the genomic characteristics of the parental tumors in situ, generation of PDX models is often resource-and timeconsuming (12). Patient-derived tumor cell (PDC) models provide unique opportunities for prospective preclinical research. Moreover, we previously demonstrated that PDCs exhibit high degrees of genetic similarity compared to the parental tumors both molecularly and phenotypically (13,14).
Drug combination therapies can enhance therapeutic efficacy, decrease toxicity, and circumvent both inherent and acquired resistance to standard treatments. Although previous studies revealed dynamic pharmacogenomic interactions across a broad range of tumor types, major gene-drug associations were primarily based on single target agents (13,(15)(16)(17). Furthermore, most drug combination suggestions were primarily based on conceptions and algorithms, lacking direct experimental evidences that support such speculations. To this end, we have developed a systematic platform for evaluation of two-drug combinations in 3-dimensionally cultured PDCs (18). While PDC models serve as reliable proxies for examining anti-cancer drug activities, precise estimation of tumor cell populations is also required to predict the patient response within the clinical framework. Therefore, in this study, we performed immunofluorescence-based image analysis to measure and predict the tumor population index in PDC models for twodrug combinational strategies using an HTS system. We found that our systematic platform could identify potential synergistic responses in individual patients with gastric cancer in situ.

Cell Lines and Patient-Derived Cell (PDC) Culture
Human gastric cancer cell lines, AGS, KATOIII, and NCI-N87, were purchased from American Type Culture Collection (ATCC, Manassas, VA, USA), and MKN-45 and SNU-216 were purchased from the Korean Cell Line Bank (Seoul, South Korea). All cell lines were maintained in RPMI 1640 medium supplemented with 1% antibiotic-antimycotic and 10% fetal bovine serum (Gibco, Grand Island, NY, USA). The primary human dermal normal fibroblast adult (HDFa) cell line was purchased from ATCC and maintained in fibroblast basal media supplemented with fibroblast growth kit compounds (ATCC). Surgically removed tumor tissue, biopsy tissue, or malignant ascites were collected from patients with gastric cancer. The protocol was approved by the Institutional Review Board at the Samsung Medical Center. Collected tissue was minced and dissociated enzymatically, and the collected effusions (500-1,000 mL) were divided into 50-mL tubes, centrifuged at 1,700 rpm for 10 min, and washed twice with Dulbecco's phosphate-buffered saline. Cell pellets were added to 75-cm 2 flasks containing RPMI 1640 medium supplemented with 10% fetal bovine serum, 1% antibiotic-antimycotic (Gibco), 0.5 g/mL hydrocortisone (Sigma Aldrich, St. Louis, MO, USA), 5 µg/mL insulin, and 5 ng epidermal growth factor (Peprotech, Rocky Hill, NJ, USA). The cell lines and PDCs were maintained at 37 • C in a humidified atmosphere in a 5% CO 2 incubator and passaged using TrypLE-Express (Gibco) to detach the cells when they reached 80-90% confluence.

Immunofluorescence Staining in Micropillar Chip
The microwell chip composed of polystyrene has 532 complementary microwells and the micropillar composed of poly (styrene-co-maleic anhydride) contains 532 micropillars. The detailed protocol was described in our previous papers (19)(20)(21). Fist, 950 nL 3D culture medium was automatically dispensed onto a microwell chip and incubated in a gaspermeable chamber with water in a 37 • C incubator. Next, 50-nL spots containing a 1:1 mixture of 70-100 cells and 0.75% alginate were dispensed onto the micropillar chip using ASFA TM Spotter ST (Medical and Bio Device, Suwon, South Korea). After cell dispensing, the micropillar chip was stamped on top of the microwell chip and incubated in the chamber in a 37 • C incubator for 3-5 days. The cells cultured under 3D conditions on the micropillar chip were fixed with 4% paraformaldehyde solution (Biosesang, Sungnam, South Korea). After fixation, the micropillar chips were permeabilised with 1% bovine serum albumin in water containing 0.3% Triton-X for 1 h. Each micropillar chip was incubated overnight at 4 • C with primary antibodies and secondary antibodies at room temperature for 3 h. The chips were washed in staining buffer and dried in the dark. The following antibodies were used: anti-EpCAM monoclonal antibody (fluorescein isothiocyanate-conjugated, 1:100, Invitrogen, Carlsbad, CA, USA), anti-vimentin (SP20, 1:200, Abcam, Cambridge, UK), anti-HER2 (3B5, 1:100, Invitrogen), anti-MET (3D4, 1:100, Invitrogen), anti-mouse Alexa Fluor 350 (1:300, Thermo Fisher Scientific, Waltham, MA, USA), anti-rabbit Alexa Fluor 594 (1:300, Thermo Fisher Scientific), and Hoechst 33342 (1:1,000, Thermo Fisher Scientific). The stained chips were scanned using an optical scanner device (ASFA TM Scanner ST, Medical and Bio Device) and scanned images were evaluated with image analysis software (S+ analysis, Samsung Electro-Mechanics Co. Ltd., Suwon-si, South Korea).

Combination Drug Screening Using 384-Well Pillar Plates
Combination drug screening using 384-pillar plates was performed by dispensing compounds into a 384-well plate and then sandwiching the 384-pillar plate on the 384-well plate. The drugs were used: herceptin, AZD2281, AZD1775, AZD6738, LEE011, palbociclib, and staurosporin (positive control). The maximum drug concentration was 20 µM (dissolved in dimethyl sulfoxide) and the doses of drug is six with 3-fold dilution ratio. Two-drug-combination effect was estimated by a Combination Index for the Loewe Additivity (22,23). The CI 50 is defined by: where A and B are the IC 50 values of drugs A and B, respectively. The IC 50 is the drug concentration at which cell viability is 50% following single drug treatment. a and b are the concentrations of drugs A and B at which cell viability is 50% when A and B are treated in combination. In practice, a CI 50 < 1 indicates that the doses of a and b producing a given effect in combination are lower than the expected dose from additive effects and can thus be directly interpreted as synergy. Similarly, a CI 50 >1 indicates that the doses of a and b producing a given effect in combination are superior to those of the expected doses from additive effects and can thus be directly interpreted as antagonism.

Targeted Sequencing
Tumors were subjected to target exome sequencing which covers a range of exonic regions of specific genes that are associated with cancer progression. Genomic DNA was shared using a Covaris S220 (Covaris, Woburn, MA) to construct a sequencing library using the SureSelect XT Reagent Kit, HSQ (Agilent Technologies) on target genes. A paired-end sequencing library was purified and amplified with a barcode tag, and the library quality and quantity were determined. Sequencing was carried out using the 100-bp paired-end mode of the TruSeq Rapid PE Cluster kit and TruSeq Rapid SBS kit on HiSeq 2500 sequencing platform (Illumina, San Diego, CA, USA).

Mutation Calls
The sequenced reads in the FASTQ files were aligned to the human genome assembly (hg19) using the Burrows-Wheeler aligner (BWA). The initial alignment BAM files were subjected to sorting (SAMtools), removing duplicated reads (Picard), locally realigning reads around potential small insertion/deletion and recalibrating base quality score [Genome Analysis Toolkit (GATK)]. We used MuTect to generate high-confidence predictions on mutation calls. Variant Effector Predictor was used to annotate the called mutations.

Statistical Analysis
Statistical analysis was performed with GraphPad Prism 5.0 software (GraphPad, Inc., San Diego, CA, USA) based on the fluorescence intensity of EpCAM and vimentin, IC 50 values, and tumor purity. T-tests (and non-parametric tests) were used to compare the mean fluorescence intensity of EpCAM and vimentin from gastric cell lines and normal dermal fibroblasts, and to determine actual purity/predicted purity. Statistically significant mean differences between EpCAM/vimentin intensity and tumor purity were indicated as * * * p < 0.0001.

Prediction of Tumor Purity in Gastric Cancer Cell Lines and PDCs
To establish a systematic HTS platform for evaluating the tumor cell index and two-drug combinational strategy in gastric cancer, we generated a library of PDCs derived from surgically resected gastric tumor specimens or ascites-derived tumor cells (Figure 1A). We have previously demonstrated establishment of 3D cell-based immunostaining protocol. In the present study, the 3D cell-based immunostaining platform has been applied to evaluate gastric cancer purity (19). Multicolor immunofluorescence analyses of EpCAM, vimentin, and DAPI were performed by measuring the fluorescence intensities of respective target molecules in 3D-cultured human gastric cancer cell-lines (AGS, KATOIII, NCI-N87, MKN-45, and SNU-216) on a micropillar chip. Normal dermal fibroblasts were used as a control for detecting non-malignant cells ( Figure 1B). Fluorescence intensity analysis showed that all five gastric cancer cell lines were marked by global expression levels of EpCAM, while normal fibroblasts exhibited upregulation of vimentin expression ( Figure 1B). Consistently, immunoblot analysis revealed a significant difference between the protein expression levels of EpCAM and vimentin in both gastric cancer cell lines and fibroblasts. Using the differential intensity levels of EpCAM and vimentin, we formulated an image-based tumor purity estimation to measure the tumor cell index. Notably, when we co-cultured NCI-N87 gastric cancer cells with normal fibroblasts at various cell-to-cell concentrations, we observed a significant correlation between EpCAM and vimentin fluorescent intensity levels (Figure 2A). EpCAM and vimentin expression levels of biological replicates from the mixture of NCI-N87 cancer cells with fibroblasts at different ratios showed significant correlations with minimal variations (Figure 2B). To investigate the minimal requirement for the tumor cellular index to evaluate the appropriate drug response, we seeded a mixture of HER2-positive gastric cancer cells with non-neoplastic cells at various concentrations (from 1 to 90%) and treated the cells with lapatinib. Notably, >30% tumor purity was sufficient for evaluating the therapeutic vulnerability of HER2-positive tumor cells to lapatinib (Figure 2C). To further evaluate the two-drug combinational approach in PDC models, we first determined the tumor cellular index in 5 HER2-positive and 3 MET-positive PDCs ( Table 1). Immunofluorescence analysis of EpCAM and vimentin revealed that tumor cells constitute more than 50% of the total cell populations in all 8 gastric PDCs, making them suitable proxies for comprehensive pharmacological analysis (Figures 3A,B).

Molecular-Guided Two-Drug Combination Treatment in PDC Models
To identify ideal two-drug combination effects for individual patients, 15 gastric PDCs and 2 cancer cell lines were subjected to lapatinib treatment with olaparib, AZD1775, AZD6738, palbociclib, savolitinib, or staurosporin at various concentrations ( Figure 4A). To quantify the degree of twodrug synergistic or antagonistic effects, tumor cellular viabilities were assessed against the expected combination response under non-interaction assumptions using various reference models (24)(25)(26). Using the SynergyFinder algorithm, we calculated the synergistic scores for each dose-response matrix among various two-drug combinations (27). Interestingly, although most two-drug combination effects at various concentrations were either additive or synergistic, some combination effects were antagonistic at varying doses ( Figure 4B), suggesting that the results of the two-drug combinational approach should be interpreted with caution. Overall, we observed a wide range of drug sensitivities among different drug combinations, demonstrating the highly heterogeneous nature of gastric PDCs ( Figure 4C).
As the distinct molecular background across each gastric PDC may explain the dynamic drug profiles, we conducted targeted massively parallel sequencing to identify somatic genomic alterations, including single-nucleotide variations, small insertions/deletions, and copy number variations. Mutations with variant allele frequencies of >5% and >20 reads were considered. Interestingly, most PDCs harbored one or more genomic alterations in key cancer-driver genes, including TP53, BRCA1, BRCA2, ERBB2, CCNE1, and CDKN2A ( Figure 4D). As genomic variations continue to be employed as reliable biomarkers for predicting the clinical response to molecular-guided targeted therapy (28)(29)(30), we analyzed pharmacogenomic interactions in gastric PDCs to identify molecular links that dictate the synergistic response to twodrug combination effects. Interestingly, we discovered that tumors with mutations in either BRCA1 or BRCA2 were therapeutically susceptible to lapatinib and olaparib combination treatment ( Figure 4E). PARP inhibition therapy has shown significant therapeutic success in patients diagnosed with advanced ovarian adenocarcinomas with germline BRCA1/2 mutations. We also found that MET-mutant tumors were synergistically sensitive to the combination of savolitinib and lapatinib. Furthermore, somatic mutations in TP53 and CDKN2A with MYC genomic amplification showed high sensitivity to the WEE1 inhibitor AZD1775. As p16, p53, and c-Myc are essential regulators of the cell cycle program, deficiency of these key molecules rendered cells more dependent on the Wee1-mediated checkpoint. Consistently, a phase II, single-arm study of AZD1775 monotherapy was conducted (NCT02688907) to evaluate anti-Wee1 treatment in relapsed small cell lung FIGURE 1 | Overview of systematic platform for prediction of tumor purity from patient tumor-derived cells (PDCs) and 3D-based high-throughput drug screening for two-drug combination therapy (A) Schematic representation of the generation of patient-derived tumor cell models from tumor tissue or malignant ascites from patients with stage IV gastric cancer. Two-dimensional (2D) cultured monolayer PDCs were seeded with 3D-culture medium. Multi-color antibodies including EpCAM, vimentin, and DAPI were used and fluorescence intensity in various gastric cancer cell lines and PDCs was measured. Tumor purity was predicted. Using PDCs with a proper amount of tumor purity, high-throughput monotherapy, or combinatorial drug screening was performed in a micropillar/microwell chip screening assay. (Continued)

DISCUSSION
With recent advancements in sequencing technology, molecular characterization of tumors has been widely conducted to facilitate personalized treatment (2,31,32). However, designing effective therapies based on computational analysis alone is confounded by tumor-inherent characteristics, including genomic complexity and intra-tumoral heterogeneity. Molecular and transcriptional heterogeneity of gastric cancer has been characterized using various approaches, led by The Cancer Genome Atlas (TCGA) consortium and multiple global collaborative efforts (2,3). Furthermore, the HTS system provides additional opportunities for exploring the biological behaviors of cancer cells at the cellular level. Combined, these methods can be used to systematically identify genomic biomarkers and appropriate patient stratifications that can guide development of novel compounds for pre-clinical trials. While these studies have provided comprehensive insights into the dynamic pharmacogenomic interactions across a wide spectrum of cancer types, most gene-drug associations have primarily relied on single agent analysis. Combination therapy has gained considerable attention in the field of oncology in recent years, with numerous studies demonstrating its significant advantage over monotherapies (33)(34)(35)(36).
In the present study, we established a systematic method for precisely estimating the tumor cell index to aid in evaluating two-drug combination therapy. Using an immunofluorescencebased approach, we employed numerous gastric cancer cell lines and PDCs to assess tumor cell populations within each given tumor by analyzing the intensity of EpCAM and vimentin expression. It is easy and simple way to estimate the tumor cell population in PDCs by analyzing EpCAM and vimentin immunofluorescence intensity. Furthermore, we could examine the proportion of specific biomarker-expressing tumor cells in each PDC, such as HER-2, EGFR, and MET etc. before anticancer drug-sensitivity test.
We discovered that all PDCs harbored >50% tumor cell populations across multiple biological replicates and a minimum cellular index of 30% was required to evaluate reliable drug activities. Moreover, two-drug combination treatments exhibited various drug-drug interactions, varying from synergistic to antagonistic effects. Notably, BRCA1/2-mutant tumors were synergistically more susceptible to lapatinib and olaparib combinations, while somatic mutations in MET conferred increased sensitivity to savolitinib and olaparib treatment. Interestingly, we showed that tumors harboring genomic alterations in cell cycle-encoding genes, including MYC, MAP2K1, and CCNE1 were synergistically vulnerable to the CDK4/6 inhibitor palbociclib. Finally, MYC-amplified tumors with both TP53 and CDKN2A mutations showed an increased response to AZD1775, suggesting the therapeutic benefits of Wee1-mediated therapy in patients with genomic ablations in the cell cycle program. Because biomarkers including genomic alterations and/or protein expression could be changed as a part of drug-resistant mechanism during and after treatment, promising results from in vitro testing do not always translate into in vivo efficacy. And limited by small samples of PDCs and available drugs, further drug-combination strategies via optimal biomarker matched needs to be warranted.   We are convinced that our study makes a significant contribution to the literature because we established a systematic method for precisely estimating the tumor cell index to aid in evaluating two-drug combination therapy. Furthermore, our platform provides a real-time relevant tool for personalized treatment through the use of mixed cell populations that are derived from patients without considerable in vitro culture. We found that our systematic platform could identify potential synergistic responses in individual patients with gastric cancer in situ.

CONCLUSIONS
Collectively, our systematic two-drug HTS platform is integral for addressing current clinical needs to facilitate precision oncology therapy.

DATA AVAILABILITY STATEMENT
The raw data supporting the conclusions of this manuscript will be made available by the authors, without undue reservation, to any qualified researcher.

ETHICS STATEMENT
This investigation was conducted in accordance with the ethical standards of the Declaration of Helsinki and national and international guidelines and was approved by the Institutional Review Board at Samsung Medical Center in Seoul, Korea.