Angiogenesis gene signatures in patient-derived tumor spheroids for genetic and tumor angiogenesis profiling

Abstract

Background: gastric cancers are highly vascular tumors, with elevated pro-angiogenic factors correlating with a poor prognosis. Despite advancements in precision medicine, there remains a critical need for platforms capable of identifying patient-specific therapeutic vulnerabilities. In this study, we present a 3D-printed patient-specific tumor angiogenesis chip that integrates genetic data to evaluate the molecular and functional characteristics of tumor angiogenesis in tumor spheroids derived from patients with gastric cancer. Results: whole-transcriptome analysis classified tumor spheroids into high- and low-angiogenesis-related gene signatures groups. Tumors with high angiogenesis-related gene signatures exhibited significantly enhanced blood vessel formation and tumor growth on the 3D tumor angiogenesis chips compared to those with low gene signatures [vessel density: 1.091 vs. 0.7538; vessel length: 1.070 vs. 0.8344; and angiogenic sprouting number: 1.184 vs. 0.6541]. The platform also enabled quantitative drug response assessments, providing a robust framework for evaluating treatment efficacy. Conclusion: these 3D-printed tumor angiogenesis chips, leveraging genomic profiling and patient-specific tumor characteristics, offer a powerful tool for advancing personalized medicine in gastric cancer.

---

Background

Advanced gastric cancer (GC) is a highly aggressive malignancy characterized by extensive heterogeneity, high metastatic potential, and limited therapeutic option.^1,2 Despite a growing awareness of its global burden, translational models capable of capturing the genetic and functional complex of GC remained limited. A deeper understanding of tumor-specific biology is essential to support precision medicine, accelerate drug development, and enable timely interventions that can reduce progression to advanced, life-threatening disease.^3–6

Tumor-on-a-chip platforms have emerged as powerful tools for modeling cancer biology under physiologically relevant conditions. By combining microfluidic systems with three-dimensional (3D) cell cultures, these platforms recreate key features of the tumor microenvironment (TME), including spatial tissue organization, interstitial flow, and dynamic cell–cell interactions.^7–10 While traditional 2D or 3D cancer models offer structural insights into tumor tissue,¹¹ they fail to mimic dynamic microenvironments.^12,13 By contrast, microfluidic systems enable the recreation of vascular–tumor interfaces in vitro, providing a platform to study angiogenesis and lymphangiogenesis.^14–16 While these platforms have provided valuable insights into how tumor cells modulate endothelial cell behavior and promote vascular remodeling—critical steps in cancer progression and metastasis^17–19—most existing models suffer from significant limitations that hinder clinical translation. Current approaches predominantly rely on cancer cell lines and lack integration of patient-specific genetic contexts.^10,14,16,19 More critically, there is a substantial gap in platforms that systematically integrate patient-derived genetic profiling with functional angiogenesis assessment. These interconnected limitations manifest in three key areas: the absence of standardized methods for correlating transcriptomic data with functional angiogenic readouts, insufficient capability to stratify patients based on molecular signatures for predicting therapeutic responses, and the lack of patient-derived models that incorporate genetic features from bulk clinical tumor samples to predict angiogenic potential. Addressing these challenges is essential for translating tumor-on-a-chip technology from research tools to clinically relevant platforms for precision oncology applications.

To address this gap, we developed a high-throughput microfluidic chip capable of modeling patient-derived tumor spheroids (PDTS) and their associated angiogenic responses. Using transcriptome profiling, we defined an angiogenesis gene signature (AGS) score to classify tumors based on the expression of curated angiogenesis-related gene sets. This AGS-based stratification was used to evaluate functional angiogenesis in PDTS models from 45 GC patients. The integration of image-based vascular metrics with AGS classification enabled us to correlate molecular subtypes with angiogenic phenotypes. This integrated approach bridges the gap between genomic profiling and functional TME assessment, providing a robust platform for advancing precision medicine in GC through the combination of transcriptomic signatures with quantitative angiogenesis modeling.

Methods

Patient enrollment and ethical approval

This study included 45 patients with advanced gastric cancer (GC) who provided biospecimens after signing consent forms approved by the Institutional Review Board (IRB) of Samsung Medical Center (IRB #2021-09-052). All procedures adhered to the principles of the Declaration of Helsinki and Good Clinical Practice guidelines (ClinicalTrials.gov identifier: NCT02589496). Biospecimens included 26 primary tumors obtained from surgeries and biopsies, and 19 samples collected from peritoneal dialysis fluid (Table 1 and Supplementary data, ESI†).

Table 1 Baseline characteristics of patients with advanced gastric cancer

Contents	Total (n = 45)
Age (years)	58 (28–80)
Sex
Male	31 (68.9%)
Female	14 (31.1%)
Race
Asian	45 (100%)
Sampling
Ascites	19 (42.2%)
Primary tumor	26 (57.8%)
TCGA
GS	35 (77.8%)
CIN	7 (15.6%)
EBV	2 (4.4%)
MSI histopathology	1 (2.2%)
Tubular adenocarcinoma	18 (40%)
Signet ring cell carcinoma	4 (8.9%)
Mucinous adenocarcinoma	1 (2.2%)
Cytopathology
Atypical	10 (22.2%)
Molecular target biomarker
HER2 positivity	9 (20.0%)
c-MET positivity	7 (15.6%)
FGFR2 positivity	3 (6.7%)
HER2 & c-MET positivity	1 (2.2%)

---

Sample preparation for sequencing

Tumor tissues, malignant cells from ascites, and matched peripheral blood samples were collected. Tumor DNA and RNA were isolated from samples with tumor cellularity >40%, as determined by pathological and transcriptomic reviews. DNA and RNA extraction was performed using the QIAamp mini kit (Qiagen, Hilden, Germany), with RNase digestion during DNA preparation (RNase A, CAT #19101; Qiagen). DNA concentrations and purity (260/280 and 260/230 ratios) were assessed using a NanoDrop 1000 spectrophotometer (NanoDrop Technologies LLC; Thermo Fisher Scientific, Waltham, MA, USA). Quantification of DNA and RNA was performed using a Qubit fluorometer (Life Technologies, Carlsbad, CA, USA).

Whole-exome and whole-transcriptome sequencing

Genomic DNA was isolated using the QIAamp DNA blood kit (Qiagen) and prepared for sequencing with the Agilent SureSelect XT Human All Exon V6 probe set (Agilent Technologies, Santa Clara, CA, USA). DNA was sheared to ∼150 bp fragments using a LE220 focused-ultrasonicator (Covaris Inc., Woburn, MA, USA), and sequencing libraries were constructed with TruSeq Rapid Cluster and TruSeq Rapid SBS kits (Illumina, San Diego, CA, USA). Sequencing was performed on an Illumina HiSeq 2500 platform using paired-end reads (2 × 100 bp). RNA libraries were prepared using the TruSeq RNA Access Library Prep kit (Illumina). RNA integrity was assessed using Quant-IT RiboGreen (Invitrogen, Waltham, MA, USA) and Agilent TapeStation analysis with RNA Screen Tape (CAT #5067-5582; Agilent). Sequencing libraries were quantified using the KAPA Library Quantification kit for Illumina platforms (CAT #KK4854; Kapa Biosystems, Wilmington, MA, USA) and sequenced on the Illumina HiSeq 2500 platform.

Variant calling and filtering

Sequenced reads were aligned to the human reference genome (GRCh37, hg19) using the Burrows–Wheeler aligner maximal exact matches algorithm.²⁰ Duplicate marking, indel realignment, and base recalibration were performed using the Genome Analysis Toolkit (version 4.1.1.0).²¹ Variants were called using Mutect2 (version 2.0)²² and annotated using dbSNP (version 138),²³ 1000 Genomes Project (phase I),²⁴ and HapMap (phase III)²⁵ datasets. Variants with a depth <4 or fewer than 2 alternative alleles were excluded. Annotation was conducted with the Ensembl Variant Effect Predictor (release 87)²⁶ using the GRCh37 database.²⁷

Molecular subtype classification

Molecular subtypes were determined using an EBV detection assay (PathSeq algorithm) to identify samples with high EBV burden.^28,29 MSI-high subtypes were detected using the MSI sensor,³⁰ while the remaining subtypes were distinguished by somatic copy number aberrations. A genomic instability index was used to quantify chromosomal alterations, including deletions and insertions.³¹

Transcriptome data analysis

RNA sequences were aligned to the GRCh38 reference genome using Spliced Transcripts Alignment to a Reference software (STAR, version 2.6.1)³² and annotated with Ensembl (version 98). RNA expression was quantified as transcripts per million using RNA-Seq with expectation–maximization software (version 1.3.1).³³ Gene signaling pathways were analyzed using gene set variation analysis.³² The three core angiogenesis pathways (angiogenesis pathway, VEGF signaling pathway, and Hedgehog signaling pathway) used for AGS classification were previously validated in an independent clinical cohort by our group.³ In that study, transcriptomic analysis of gastric cancer patients demonstrated that these pathways effectively stratified patients into high-AGS and low-AGS groups, with significant correlation to ramucirumab treatment response versus non-response in clinical settings. Based on this prior validation, we applied the same three-pathway AGS classification system to evaluate functional angiogenesis in the current PDTS cohort. Angiogenesis-related gene sets were obtained from the pathway interaction database (PID), focusing on pathways known to influence anti-angiogenic therapeutic responses. The complete gene list is provided in Table 3. To classify patients into high- and low-angiogenesis gene signature (AGS) groups, we performed unsupervised hierarchical clustering of the transcriptomic data using Euclidean distance method, based on three angiogenesis-related gene expression profiles (angiogenesis pathway, VEGF signaling pathway, and Hedgehog signaling pathway). Tumor microenvironment subtypes were classified into immune-depleted, fibrotic, immune-enriched, or immune-enriched/fibrotic categories using molecular functional portraits (MFP).³³

Preparation of PDCs from primary solid tissues

Tissues were placed in a Petri dish containing fresh sterile Dulbecco's phosphate-buffered saline (DPBS) supplemented with 1% penicillin/streptomycin. After rinsing two to three times, the tissues were transferred to an e-tube containing 0.5–1.0 mL of tissue lysis buffer composed of RPMI 1640 (Gibco, Waltham, MA, USA) supplemented with collagenase (1.0 mg mL⁻¹; CAT #07912; Stemcell Technologies, Vancouver, Canada), dispase (0.5 mg mL⁻¹; CAT #07923; Stemcell Technologies), DNase (0.15 mg mL⁻¹; CAT #07469_C; Stemcell Technologies), and 5% fetal bovine serum (FBS; Gibco). Tissues were minced into small pieces using scissors before being transferred to a 15-mL tube containing 1–2 mL of lysis buffer. The tube was incubated at 37 °C for 0.5–1.0 h while vortexing at 70 rpm. After incubation, the tube was centrifuged at 300 × g for 5 min at 21–23 °C, the supernatant was discarded, and the pellet was washed with PBS. Subsequently, the mixture was incubated with tumor spheroid culture medium, following the procedure for ascites, and tumor selection was performed.

Preparation of PDCs from ascites

Approximately 1 L of ascites was drained from each patient to isolate tumor spheroids. The ascite sample was dispensed into 50-mL conical tubes and centrifuged at 300 × g for 5 min at 4 °C. After removing the supernatant, the cells were resuspended in PBS, and centrifugation was repeated. To filter the cell suspension, a 70-μm cell strainer (CAT #CLS431751; Corning, USA) was coupled with a fresh 50-mL conical tube, and the suspension was passed through it. Red blood cell (RBC) lysis buffer (CAT #158904; Qiagen) was added to the cell suspension along with PBS (PBS : RBC lysis buffer = 1 : 3, v/v), and the mixture was incubated for 30–60 min at 21–23 °C to lyse RBCs. RBC lysis was repeated until the red color disappeared from the cell pellet. The supernatant was discarded, and the cells were resuspended in PBS. The cell concentration was adjusted to the target value (typically 10⁷ cells per 150-mm dish) and cultured in tumor spheroid medium for two to three passages. The tumor spheroid medium consisted of RPMI 1640 (Gibco) supplemented with 10% FBS (Gibco), 1% antibiotic–antimycotic (CAT #CA00; GenDEPOT), epidermal growth factor (EGF; 5 ng mL⁻¹; PeproTech, Waltham, MA, USA), insulin (5 μg mL⁻¹; Sigma-Aldrich), and hydrocortisone (0.5 μg mL⁻¹; Sigma-Aldrich).^34–36 The absence of mycoplasma contamination was confirmed using a universal mycoplasma detection kit (CAT #30-1012K; American Type Culture Collection).

3D-printed microfluidic device fabrication

The chip development process combined 3D printing for prototype fabrication and injection molding for mass production. Chip prototypes were designed using SolidWorks (Dassault Systèmes, Vélizy-Villacoublay, France) and fabricated with a 3D printer. Upon completion, prototypes were rinsed with isopropyl alcohol for 15 min to remove residual resin, followed by immersion in water for ultraviolet light post-curing at 385 nm for 30 min. To enhance surface smoothness and ensure biocompatibility, poly(c-xylene) was deposited via plasma-enhanced chemical vapor deposition (Lavida, Femto Science, South Korea).

Primary cell line culture

Human umbilical vein endothelial cells (HUVECs; CAT #C2519A; Lonza, Basel, Switzerland) were cultured in endothelial cell growth medium-2 (EGM2 BulletKit: CAT #CC-3162; Lonza), using passages 4 and 5 for experiments. Lung fibroblasts (CAT #CC-2512; Lonza) were cultured in fibroblast growth medium-2 (FGM2 BulletKit: CAT #CC-3132; Lonza), with passages 5 and 6 used in experiments. Cells were incubated at 37 °C with 5% CO₂ for 2–3 days before loading into the device. Cell lines were authenticated via short tandem repeat DNA profiling (SMC Basic Research Support Center, Seoul, South Korea), tested for mycoplasma contamination, and maintained according to manufacturer's protocols.

Tumor spheroid formation

Tumor spheroids were prepared by dispensing a predetermined amount of the patient-derived cell (PDC) suspension into U-bottom-shaped 96-well plates (CAT #MS-9096UZ; Sumitomo Bakelite, Japan). Cellular niches were established by adding a 1% volume ratio of Matrigel (CAT #35623; Corning) to 150 μL of medium containing 10 000 PDCs and 500 fibroblasts. Spheroids were cultured for 2–5 days and subsequently introduced to the 3D tumor angiogenesis model.

Tumor spheroid-induced angiogenesis assay

Prior to cell and hydrogel patterning, microfluidic devices were treated with oxygen plasma (CUTE; Femto Science) for 3 min to ensure cleanliness and induce hydrophilicity. Hydrogels containing tumor spheroids were prepared by mixing tumor-spheroid-containing media with fibrinogen (CAT #F8630-10G; Sigma-Aldrich) at a 3 : 1 volume ratio (final fibrin hydrogel concentration: 2.5 mg mL⁻¹). Thrombin (0.8 μL; CAT #T7201-100UN; Sigma-Aldrich) was added to the mixture, and 7.5 μL of the hydrogel containing tumor spheroids was injected into the microfluidic channel. After priming the hydrogel for 7 min, 50 μL of endothelial cell suspension (40 000 cells) was injected into the “endothelialized channel.” Culture medium containing 5 μg mL⁻¹ insulin (PeproTech) was added to 250 μL in the opposite reservoir. Samples were maintained in an incubator (37 °C, 5% CO₂) with medium replacement every 2 days according to the experimental timeline. Ramucirumab (0.1 μM) was introduced on day 2, and the concentration was maintained throughout the culture period.

Immunostaining and image acquisition

Samples were fixed with 4% (w/v) paraformaldehyde (Biosesang, Yongin, South Korea) in DPBS (WELGENE, Gyeongsan, South Korea) for 15 min, followed by permeabilization in 0.20% Triton X-100 (Sigma-Aldrich) for 20 min. To block non-specific binding, samples were treated with 3% bovine serum albumin (Sigma-Aldrich) for 24 h. Fluorescence staining was performed using the following reagents: fluorescein-labeled Ulex Europaeus agglutinin I (UEA I) (1 : 500; CAT #FL-1061-2; Vector Laboratories, Newark, CA, USA), Alexa Fluor 594 anti-human CD31 antibody (1 : 200; CAT #303126; Biolegend, San Diego, CA, USA), Alexa Fluor 594 anti-human CD326 (EpCAM) antibody (1 : 300; CAT #324228; Biolegend), and Hoechst 33342 (1 : 1000; CAT #62249; Thermo Fisher Scientific).

Statistical analyses

All statistical analyses were conducted using the Wilcoxon concordant pairwise signed-rank test, unpaired t-test, Mann–Whitney U test, and two-way analysis of variance (ANOVA), with statistical significance set at p < 0.05. No statistical methods were employed to pre-determine the sample size, blinding, or randomization. The average sprouting length values for all items were calculated and plotted based on measurements taken within the total cell culture area (Fig. 3d). Statistical significance (p-value) was determined using paired t-tests comparing the drug-treated and control groups across all patients. Drug responsiveness was analyzed as the percentage change in vessel length induced by drug treatment relative to the endpoint, with data plotted for each patient (Fig. 3). The Wilcoxon signed-rank test was used to evaluate differences in drug responsiveness between high- and low-angiogenesis gene signature (AGS) groups. Tumor area values were defined as regions showing simultaneous expression of EpCAM and nuclei (Fig. 4a and b), and drug condition values were normalized against control condition values. Unpaired t-tests were performed to compare tumor areas between drug-treated and control conditions.

Key reagents and chemicals used in this study, including CAS numbers, purity specifications, and supplier information, are presented in Table 2.

Table 2 Key reagents and chemicals used in this study with CAS numbers, purity specifications, and supplier information

Key reagent and chemicals	Brand	Cat. no.	CAS no.	Purity
Cell culture and sample preparation
RPMI 1640 medium	Gibco	11875093	N/A, complex medium	Sterile, cell culture teste
Fetal bovine serum (FBS)	HyClone	SH30919.03	N/A, characterized grade	Triple filtered
Antibiotic–antimycotic	GenDEPOT	CA00	Antibiotic mixture	Cell culture tested
Epidermal growth factor	PeproTech	AF-100-15	62253-63-8	≥98% (HPLC)
Insulin	Sigma-Aldrich	I1507	11061-68-0	≥98% (HPLC)
Hydrocortisone	Sigma-Aldrich	H0888	50-23-7	≥98% (HPLC)
Tissue processing
Collagenase	Stem cell Technologies	07912	9001-12-1	3000 U mL^−1
				Collagenase, 1000 U mL^−1 hyaluronidase,
				1.0 mg mL^−1
Dispase	Stemcell technologies	07923	42613-33-2	1 Unit per mL
DNase I	Stemcell technologies	07469	9003-98-9	≥2000 units per mg
RBC lysis buffer	Qiagen	158904	N/A	N/A
3D culture and angiogenesis assay
Matrigel	Corning	356234	N/A	N/A
Fibrinogen	Sigma-Aldrich	F8630-10G	9001-32-5	≥75% of protein is clottable
Thrombin	Sigma-Aldrich	T7201-	9002-04-4	≥2000 NIH units per mg
Aprotinin	Sigma-Aldrich	100UNA1153-	9087-70-1	3–8 TIU per mg solid
Dulbecco's phosphate-buffered saline	Welgene	10MGLB001-01	N/A	Cell culture grade
Cell lines
Human umbilical vein endothelial cells	Lonza	C2519A	N/A	Quality assured
Lung fibroblasts	Lonza	CC-2512	N/A	Quality assured
EGM2 BulletKit	Lonza	CC-3162	N/A	Quality assured
FGM2 BulletKit	Lonza	CC-3132	N/A	Quality assured
Immunostaining
Paraformaldehyde	Biosesang	P2031	30525-89-4	≥95–98%
Triton X-100	Sigma-Aldrich	X-100	9036-19-5	High-purity
Bovine serum albumin	Sigma-Aldrich	A9418	9048-46-8	≥96%
Alexa Fluor 488 anti-human CD31 antibody	Biolegend	303110	N/A	N/A
Alexa Fluor 594 anti-human CD326 (EpCAM) antibody	Biolegend	324228	N/A	N/A
Hoechst 33342	Thermo Fisher Scientific	62249	N/A	≥99% (HPLC)

---

Table 3 Core angiogenesis gene signature (AGS) pathways used for patient stratification

Pathway interaction database	Pathway summary	Contents
PID_ENDOTHELIN_ PATHWAY	Endothelins/angiogenesis	PLCB3,MAPK14,MMP1,ADCY9,SLC9A1,GNA L,FOS,EDNRB,SRC,PTK2B,COL3A1,CYSLTR 2,PRKCQ,ADCY8,ADCY4,EDN3,PLA2G4A,B CAR1,GNAZ,ADCY7,COL1A2,GNA12,GNAQ
PID_VEGF_VEGFR_ PATHWAY	VEGF signaling network	FLT1,VEGFA,PGF,VEGFD,NRP2,KDR,FLT4,VEGFB,NRP1,VEGFC
PID_HEDGEHOG_2P ATHWAY	Signaling events mediated by the Hedgehog family	DHH,HHIP,CDON,STIL,PIK3CA,HHAT,AKT1,BOC,PTHLH,GLI2,LRPAP1,PTCH2,IHH,ARR B2,TGFB2,GAS1,SHH,GRK2,SMO,LRP2,PTC H1,PIK3R1

---

Results

Modeling tumor-specific angiogenesis using a microfluidic high-throughput 3D cell culture

Tumor-induced angiogenesis plays a critical role in tumor survival, progression, and evasion of immune surveillance. This process is primarily mediated by vascular endothelial growth factor (VEGF) secreted by tumor cells, which binds to VEGF receptors (VEGFR) on endothelial cells (ECs), promoting angiogenic sprouting and vascularization.^37–39 Antiangiogenic therapies, such as using Ramucirumab, a VEGFR2 antagonist, effectively inhibit this signaling pathway and reduce tumor angiogenesis (Fig. 1a).

Fig. 1 Development of a 3D-printed platform for assessing tumor invasion and angiogenesis in microfluidic PDTS models. (a) Schematic representation of GC characteristics, highlighting tumor angiogenesis driven by VEGF binding to VEGFR. VEGF secretion from tumor cells activates autocrine and paracrine signaling, promoting tumor invasion and angiogenesis. Ramucirumab, a VEGFR-2 antagonist, inhibits VEGF-A, -C, and -D signaling, thereby suppressing tumor progression. (b) Time-lapse images showing tumor spheroid invasion and angiogenic sprouting analyzed in the 3D microfluidic platform (left). Early endothelial cell activation occurs by 24 hours, progressing to multicellular, lumenized tubular structures. Right panels show high-magnification confocal images and Z-stack cross-sectional views confirming lumen formation (green: CD31). Scale bar, 200 μm.

To investigate patient-specific tumor angiogenesis, we developed a patient-derived tumor spheroid (PDTS) model to induce tumor angiogenesis using a microfluidic high-throughput 3D cell culture chip (Fig. 1b). This chip allows parallel culture of eight independent tumor models and supports image-based quantification of tumor-induced angiogenesis. Time-lapse imaging revealed the dynamic progression of endothelial sprouting in response to tumor growth. By 24 hours, ECs exhibited early activation and directional migration. At later timepoints, sprouts extended further and formed multicellular, lumenized tubular structures, as confirmed by Z-stack confocal imaging and cross-sectional views (Fig. 1b, right). These angiogenic sprouts reflect early vessel morphogenesis rather than single-cell collective invasion. The ability to visualize and quantify these 3D structures under controlled microenvironmental conditions underscores the utility of the platform for modeling tumor-induced angiogenesis in a patient-specific context.

Angiogenesis-related gene signatures reveal molecular subtypes and proangiogenic profiles in GC

Between December 2014 and January 2021, 45 patients with advanced GC were enrolled in this study (Table 1). Tumor samples were collected from either primary tumor tissues (26) or malignant ascitic fluid (19). Patient-derived tumor cells were purified and subjected to whole-exome and transcriptome sequencing to characterize their molecular profiles. According to The Cancer Genome Atlas (TCGA) classification, 35 samples were categorized as genomically stable (GS), seven as chromosomal instability (CIN), two as Epstein–Barr virus (EBV), and one as microsatellite instability (MSI) (Fig. 2a). Tumor organoids derived from patient samples retained high genomic similarity to primary tumors,⁴⁰ even when cultured under low-passage conditions (0–3 passages) (Fig. S1a, ESI†).³⁴ Despite the viability limitations of PDTS during prolonged culture, this low-passage tumor organoid system demonstrated consistency with the genomic profiles of ascitic and primary tumor samples. Across subgroups, mutations were frequently detected in AT-rich interaction domain 1A (ARID1A), cadherin 1 (CDH1), mucin 6 (MUC6), ERBB2, and ring-finger protein 43 (RNF43), with VEGF-A amplification observed in five cases: one EBV, one MSI, and three CIN subtypes (Fig. 2a, b).⁴¹ To further explore angiogenesis-related molecular characteristics, unsupervised clustering was performed on whole-transcriptome sequencing data, focusing on three previously validated angiogenesis-related gene sets (angiogenesis, VEGF signaling, and Hedgehog signaling pathways) (Table 3). These three pathways were reported to play crucial roles in tumor angiogenesis and progression, and were also identified as key determinants of therapeutic response to second-line chemotherapy with ramucirumab plus paclitaxel in patients with metastatic gastric cancer who had failed first-line chemotherapy.³ The cohort was divided into two groups based on angiogenesis gene signature (AGS): high-AGS (n = 24, elevated expression in angiogenesis-related pathways) and low-AGS (n = 21, lower expression in these pathways). Differential expression of genes in the VEGF signaling pathway significantly distinguished these groups (p = 5.2 × 10⁻¹³, Wilcoxon signed-rank test) (Fig. 2c). Additionally, the high-AGS group exhibited elevated expression levels in the cell migration-related gene set including VEGFR3 signaling in lymphatic endothelium, beta-3 integrin cell-surface interactions, beta-1 integrin cell-surface interactions, integrins in angiogenesis, and Syndecan-1-mediated signaling (Fig. 2d and Table 4). Furthermore, the high-AGS group also showed a higher frequency of oncogenic mutations, suggesting a potential link between angiogenesis-related transcriptional programs and the underlying genomic alterations driving tumor aggressiveness (Fig. 2e). To investigate whether this approach could reflect the angiogenic landscape of primary tumors, we estimated four distinct tumor microenvironment subtypes (immune-depleted, fibrotic, immune-enriched, or immune-enriched/fibrotic) using the MFP algorithm. The TME profiles of the two spheroids were highly concordant with those of their parental tumors. Notably, a strong correlation was observed between fibrotic TMEs²⁷ and high-AGS expression levels (Fig. S1b and c, ESI†).

Fig. 2 Genomic landscape of the GC patient cohort. (a) Petal plot illustrating the distribution of GC molecular subtypes in the cohort (n = 45), including genomically stable (GS), chromosomal instability (CIN), Epstein–Barr virus (EBV), and microsatellite instability (MSI). Primary tumor tissues and malignant ascites were used as sample sources. (b) Landscape of genomic alterations across patient-derived samples, categorized by molecular subtypes and mutations types (e.g., indels, missense, nonsense). (c) Unsupervised clustering of angiogenesis-related gene sets, stratified by the GC subtype, sampling method, and gene set variant analysis scores. (d) Differential expression of cell migration-related genes between high-AGS and low-AGS groups. Violin plots display median and interquartile ranges; p-values were calculated using the Wilcoxon signed-rank test. (e) Oncoplot highlighting mutation frequencies between high-AGS (n = 24) and low-AGS (n = 21) groups.

Table 4 Secondary enriched pathways identified in high-AGS patients through gene set variation analysis (GSVA)

Pathway interaction database	Pathway summary	Contents
PID_LYMPH_ANGIO GENESIS_PATHWAY	VEGFR3 signaling in lymphatic endothelium	MAPK1,ITGA5,VEGFC,MAPK3,MAPK11,ITGA1,CR K,PIK3R1,COL1A2,FLT4,COL1A1,SHC1,ITGB1,MA P2K4,PIK3CA,MAPK14,ITGA4,CREB1,ITGA2,SOS1,VEGFD,GRB2,AKT1,FN1,RPS6KA1
PID_INTEGRIN3_PAT HWAY	Beta3 integrin cell surface interactions	COL4A5,VEGFA,LAMB1,L1CAM,ITGA2B,CCN1,C OL4A6,FGG,LAMA4,COL4A1,EDIL3,F11R,IBSP,PD GFB,FBN1,PDGFRB,FGB,FN1,COL4A3,SPHK1,TGF BR2,ITGB3,SPP1,COL1A1,COL1A2,LAMC1,TGFBI,TNC,PECAM1,PLAUR,VTN,PLAU,KDR,ITGAV,CO L4A4,CD47,HMGB1,FGA,THY1,THBS1,SDC4,PVR,SDC1
PID_INTEGRIN1_PAT HWAY	Beta1 integrin cell surface interactions	ITGA3,ITGA9,COL4A4,IGSF8,COL5A1,COL11A1,IT GA7,CD81,NID1,VTN,TGFBI,PLAU,LAMB1,COL4A 1,ITGB1,COL7A1,LAMA4,LAMA1,ITGA5,COL6A2,THBS2,PLAUR,COL2A1,COL4A5,ITGA10,COL1A1,LAMA2,COL5A2,SPP1,COL18A1,LAMC1,ITGA4,C OL4A3,VCAM1,CD14,COL3A1,THBS1,ITGA11,COL 1A2,LAMC2,LAMA3,CSPG4,FGA,COL6A1,LAMB3,FBN1,MDK,FN1,ITGA1,VEGFA,ITGAV,COL4A6,L AMA5,FGB,COL6A3,TGM2,F13A1,ITGA2,LAMB2,J AM2,FGG,ITGA6,COL11A2,TNC,ITGA8,NPNT
PID_AVB3_INTEGRIN_PATHWAY	Integrins in angiogenesis	ANGPTL3,PXN,ITGAV,PIK3R1,COL1A2,COL5A2,S RC,MAPK3,MAPK1,CDKN1B,COL11A2,HSP90AA1,CBL,PI4KB,COL5A1,TGFBR2,COL16A1,PTK2B,CO L8A2,SDC1,ITGB3,RAC1,COL8A1,COL6A3,PTPN11,COL2A1,COL4A3,TLN1,MFGE8,COL13A1,COL10A 1,FGF2,VAV3,COL6A1,RPS6KB1,EDIL3,COL11A1,CSF1R,COL4A5,ROCK1,PIK3C2A,COL15A1,IRS1,PI 4KA,VTN,COL9A2,PIK3CA,AKT1,COL4A1,PTK2,C OL17A1,COL7A1,COL1A1,IGF1R,CASP8,COL9A3,COL6A2,RHOA,VEGFA,SPP1,COL9A1,COL4A6,F11 R,BCAR1,CSF1,COL14A1,COL3A1,FN1,ILK,KDR,C OL12A1,ADGRA2,COL4A4,VCL
PID_SYNDECAN_1_ PATHWAY	Syndecan-1- mediated signaling events	COL9A3,COL11A2,COL6A3,COL8A2,SDCBP,PRKA CA,MET,COL13A1,MAPK1,MAPK3,SDC1,COL7A1,MMP1,COL11A1,COL1A2,COL2A1,HPSE,COL5A2,PPIB,COL4A3,COL16A1,COL12A1,COL10A1,CASK,COL5A1,COL4A5,COL3A1,COL4A4,COL14A1,BSG,TGFB1,COL8A1,CCL5,MMP9,COL4A1,COL6A2,M MP7,COL15A1,COL6A1,LAMA5,COL9A1,HGF,COL 9A2,COL1A1,COL17A1,COL4A6

---

Collectively, these analyses indicate that a significant proportion of patients with GC exhibit high AGS, which is closely associated with fibrotic TME subtypes and implication of clinical response to ramucirumab in AGC patients (Fig. S1d and e, ESI†).

High-AGS predicts enhanced tumor angiogenesis in patient-specific 3D models

To investigate ex vivo tumor-specific angiogenesis, we established PDTS from 45 patients with GC and performed next-generation sequencing to obtain transcriptomic profiles. Tumor spheroids were derived from 26 primary tumor tissues and 19 malignant ascitic fluid samples. Primary tumor specimens were collected during surgery or endoscopy as part of standard patient care, while malignant ascites were obtained from patients with peritoneal metastases, where fluid accumulation reflects dysfunctional and leaky vasculature (Fig. 3a).^42,43

Fig. 3 Angiogenesis mapping in patient-specific GC models using a transcriptome-guided tumor angiogenesis chip. (a) Schematic of the experimental workflow from clinical sample acquisition (ascites or primary tumor) to tumor spheroid formation and vascular co-culture in the microfluidic platform. (b) Representative confocal images showing tumor spheroid-induced angiogenesis within the reconstructed TME (green: CD31; red: EpCAM; blue: nuclei). (c) Tumor angiogenesis profiles across 45 patient-derived models stratified by angiogenesis gene signature (AGS). RAM-X indicates a ramucirumab-treated tumor spheroid model derived from patient sample X (RAM-46 to RAM-56 do not include genomic data). Horizontal white-dotted lines indicate intervals of 500 μm (green: CD31; red: EpCAM; blue: nuclei). (d) Quantitative morphometric analysis comparing vascular-vessel density, angiogenic sprout number, and average sprouting length- between high- (red) and low- (blue) AGS groups. p-Values were calculated using the Mann–Whitney U test (n = 24 for high-AGS; n = 21 for low-AGS). (e) VEGF-A secretion levels in high-AGS and low-AGS groups, analyzed quantitatively (Mann–Whitney U test: n = 12 for each group). AGS was defined based on transcriptome profiling of curated angiogenesis-related gene sets. Scale bar, 1 mm.

Fig. 4 Correlation of clinical genomic data with the 3D PDTS-induced angiogenesis model outputs. (a) Scatter plots showing the correlation between angiogenesis-related parameters (e.g., vessel density and angiogenic sprouting length) derived from the 3D PDTS-induced angiogenesis model and whole-transcriptome sequencing data, analyzed using Pearson's correlation tests. The VEGF signaling pathway demonstrated a significant positive correlation with vessel density and sprouting length, while Hedgehog signaling showed a weaker association. (b) Heatmap integrating patient-specific data, including TCGA subtypes, sampling methods, angiogenesis metrics from the 3D PDTS model, and gene expression profiles. High- and low-AGS groups are highlighted, demonstrating their distinct angiogenesis patterns and genomic characteristics.

The formation of PDTS was initiated using 10 000 PDCs and 500 human fibroblasts containing 1% Matrigel. After two days of spheroid formation, the aggregates were introduced into the tumor angiogenesis chip and cultured for five days. Fluorescence confocal imaging revealed the spatial relationship between the tumor spheroid and the surrounding angiogenic sprouts (Fig. 3b). A 3D vascular network emerged in response to tumor-derived proangiogenic signals, enabling image-based analysis of angiogenesis under physiologically relevant conditions.

Tumor models were stratified by AGS, a transcriptome-derived score based on curated angiogenesis-related gene sets. High-AGS tumors showed markedly elevated angiogenic activity compared to low-AGS tumors (Fig. 3c). The term “RAM-X” in Fig. 3 refers to ramucirumab-treated tumor spheroid models, with “X” denoting an anonymized patient ID. Quantitative analysis demonstrated significant increases in angiogenesis metrics including vessel density, total vessel length, angiogenic sprout number, and sprout length in the high-AGS group. Compared to PDTS-free controls, tumor spheroids significantly enhanced sprouting length and vessel density (mean sprouting length: 7.63 × 10²vs. 7.61 × 10¹, p = 0.0286; vessel density: 1.224 vs. 7.450 × 10⁻³, both p = 0.0286). Stratified analysis revealed that high-AGS tumors (n = 24) exhibited higher angiogenesis indices than low-AGS tumors (n = 21), including median relative vessel density (1.072 vs. 0.7982), sprouting length (1.091 vs. 0.7538), total vessel length (1.070 vs. 0.8344), and sprouting number (1.184 vs. 0.6541) (Fig. 3d).

Consistent with these findings, VEGF-A secretion was significantly higher in the high-AGS group, suggesting that elevated VEGF signaling drives the enhanced angiogenic phenotype (Fig. 3e). These results validate the use of AGS as a functional biomarker and demonstrate the utility of this 3D platform for linking transcriptomic profiles to quantifiable vascular phenotypes.

Correlation between VEGF signaling and PDTS-induced angiogenesis parameters

The analysis revealed a strong correlation between the genomic profiles of PDTS and angiogenesis parameters quantified from the chip-generated data. Specifically, vascular morphological scores, including vessel density and angiogenic sprouting length, were significantly associated with VEGF signaling pathway activation. The VEGF signaling pathway exhibited a robust correlation with relative vessel density (Pearson's correlation coefficient, r = 0.5096; 95% CI: 0.254–0.699, p = 0.0003) and relative angiogenic sprouting length (r = 0.51494; 95% CI: 0.267–0.705, p = 0.0003) (Fig. 4a). These results highlight the pivotal role of VEGF signaling in regulating PDTS-induced angiogenesis in patient-derived models.

Notably, additional correlations were assessed to evaluate the contributions of Hedgehog signaling and other angiogenesis-related pathways to vascular characteristics. Although these pathways demonstrated weaker correlations, angiogenesis signaling showed a modest but statistically significant association with angiogenic sprouting length (r = 0.3334, p = 0.0252) (Fig. 4a). In contrast, no significant association was observed between the angiogenesis signaling and vessel density (r = 0.1469, p = 0.3357) (Fig. 4a).

A comprehensive heatmap was generated to compare genomic subtypes, VEGF signaling activity, and angiogenesis indices across the cohort of 45 patients (Fig. 4b). Patients with high VEGF signaling activity, predominantly observed in the high-AGS group, exhibited elevated vessel density, total vessel length, and angiogenic sprout numbers. Cancer invasiveness scores, derived from angiogenesis indices, further emphasized the proangiogenic environment driven by VEGF in the high-AGS group. Conversely, the low-AGS group demonstrated reduced angiogenesis metrics, consistent with lower VEGF signaling activity and less invasive tumor phenotypes.

Collectively, these findings underscore the successful development of a robust angiogenesis screening platform capable of quantifying functional angiogenesis parameters. The platform's ability to correlate PDTS-induced angiogenesis indices with genomic characteristics, particularly VEGF pathway activation, aligns with proangiogenic signaling profiles observed in primary tumor samples. This highlights its potential utility in evaluating angiogenesis-targeted therapies and advancing precision oncology.

High angiogenesis signature predicts enhanced response to VEGFR2-targeted therapy

To investigate the effectiveness of antiangiogenic drugs, we tested the reponse of the 3D PDTS-induced angiogenesis model to ramucirumab, a VEGFR2-specific monoclonal antibody. Ramucirumab (1 μM) was introduced on day 2 of culture and maintained throughout the experimental period (Fig. 5a). As a VEGFR2 antagonist, ramucirumab binds directly to VEGFR2, blocking VEGF-mediated signaling and suppressing tumor-driven angiogenesis (Fig. 5b).

Fig. 5 Morphological features of 3D PDTS-induced angiogenesis in patients with GC responding to ramucirumab. (a) Schematic representations of the 3D tumor angiogenesis model used to evaluate the antiangiogenic effects of ramucirumab. Ramucirumab targets VEGFR2 expressed on tumor and endothelial cells co-cultured within the chip. (b) Schematic of EC morphological changes following ramucirumab treatment, with white arrows indicating terminal vascular regions (scale bar, 10 μm). (c) Maximum intensity Z-projection fluorescence images comparing control and ramucirumab-treated samples (green: CD31; red: EpCAM; blue: nuclei). (d) Quantitative assessment of drug response across patient-specific samples, comparing average sprouting length between control (plain bars) and ramucirumab-treated (patterned bars) groups. Statistical significance was determined using paired t-tests (n = 7 for control and n = 6 for treatment). Scale bar, 1 mm.

Compared to the untreated control group, the ramucirumab-treated group demonstrated significantly reduced angiogenic sprouting, as evidenced by both visual and quantitative analyses on the platform. Specifically the median angiogenic sprouting length in the ramucirumab-treated group was substantially reduced (control vs. ramucirumab-treated: 1.10 × 10³vs. 6.23 × 10²; p < 0.0001) (Fig. 5c and d). Vessel density and total vessel length were also markedly lower in the treated samples, highlighting the potent antiangiogenic effects of ramucirumab.

Further stratification of the data revealed that the high-AGS group exhibited a more pronounced response to ramucirumab compared to the low-AGS group. Patients with high-AGS tumor spheroids showed significantly greater reductions in angiogenesis indices, including sprouting length, vessel density, and total vessel length, following ramucirumab treatment (p = 2.42 × 10⁻², Wilcoxon signed-rank test) (Fig. 5d). These findings suggest that the high-AGS group is more dependent on VEGF signaling for tumor-driven angiogenesis and, therefore, more susceptible to VEGFR2-targeted therapy.

Notably, among the 12 patients in the high-AGS cohort, ramucirumab treatment led to significant tumor regression, further corroborating the drug's efficacy in suppressing angiogenesis and its downstream effects. Confocal microscopy images visually highlighted these differences, with the ramucirumab-treated group exhibiting drastically reduced vascular networks and tumor invasiveness compared to controls.

These results underscore the potential of the 3D PDTS-specific angiogenesis model as a robust platform for evaluating patient-specific responses to antiangiogenic therapies. By accurately measuring functional indices of tumor angiogenesis and stratifying patients based on AGS expression, this platform offers a promising tool for precision oncology, enabling targeted and personalized treatment strategies for patients with GC.

Discussion

Research on angiogenesis using 3D chip-based tumor models plays a pivotal role in advancing our understanding of the fundamental processes driving cancer progression and metastasis. Such insights are essential for developing targeted therapies that can disrupt angiogenic pathways and offer more effective treatment options for patients.^14–16,44 Numerous studies have developed 3D tumor angiogenesis models for various cancers, including hepatocellular carcinoma,⁴⁵ human glioblastoma,⁴⁶ renal cell cancer,⁴⁷ lung cancer,⁴⁸ and GC.³⁵ These models have demonstrated the reciprocal interactions between tumor and stromal cells that promote tumor growth, invasion, and EC proliferation. However, most of these studies have relied on cancer cell lines to simulate the TME,^10,14,16,19 while only a few have utilized PDCs to examine tumor-induced angiogenesis in a 3D TME.^35,47

In this study, we developed a high-throughput tumor spheroid-induced angiogenesis model incorporating clinical samples (n = 45), using genetic classifications to guide experimental design. The image-based quantification of tumor progression and angiogenesis in our model demonstrated a strong correlation with clinical information, emphasizing the ability of our 3D PDTS platform to mimic patient-specific tumor conditions accurately.

TCGA categorizes GCs into four molecular subtypes—GS, CIN, EBV, and MSI—based on whole-exome and transcriptome analyses.²⁸ Previous studies, including TCGA, have identified recurrent VEGF-A amplifications predominantly in the CIN subtype, limiting the identification of angiogenesis-related differences between GC subtypes,^49,50 and our results revealed significant distinctions when stratifying patients by AGS. Specifically, the high-AGS group exhibited significantly enhanced angiogenic sprouting and vessel extension compared to the low-AGS group.

While our platform represents a significant advancement in patient-specific angiogenesis modeling, several limitations should be acknowledged. First, our model primarily captures tumor-endothelial interactions and would benefit from incorporation of additional cellular components such as immune cells, cancer-associated fibroblasts, and pericytes to more fully recapitulate the native tumor microenvironment complexity. Second, our approach focused on transcriptomic-level assessment of angiogenesis-related genes, while tumor angiogenesis involves complex protein-level interactions and post-transcriptional regulation that could influence therapeutic responses beyond what gene expression profiles can capture. Third, the 5-day culture period, although sufficient for angiogenic sprouting assessment, may not capture long-term adaptive responses or chronic therapeutic effects. Future research directions include expanding cellular complexity to create more physiologically relevant tumor ecosystems, incorporating multi-pathway angiogenesis profiling beyond VEGF signaling, and developing extended culture protocols for comprehensive therapeutic validation. Despite these limitations, our integrated genomic-functional approach provides immediate translational value for precision oncology. This study highlights the promise of 3D PDTS models as robust tools for personalized oncology, bridging the gap between genomic profiling and functional assessments of tumor-specific angiogenesis.

Given the high reproducibility of the patient-specific TME in our 3D platform, these findings highlight its utility for predicting drug responses across diverse tumor types. Furthermore, the platform holds potential for studying the mechanisms of action of novel pharmaceuticals targeting tumor-induced angiogenesis. This study highlights the promise of 3D PDTS models as robust tools for personalized oncology, bridging the gap between genomic profiling and functional assessments of tumor-specific angiogenesis.

Conclusions

This study presents a 3D cell culture model for large-scale patient screening and drug response prediction, integrating genetic data for enhanced clinical relevance. Unlike previous proof-of-concept studies, our platform is scalable and reproducible, with potential for real-world applications and multi-omics integration. It enables patient-specific microenvironmental investigations, with the incorporation of stromal or immune cells offering further physiological relevance. Expanding this model to other cancer types could uncover shared and unique angiogenic mechanisms. The microfluidic-based endothelialized microchannels replicate in vivo angiogenesis, allowing precise assessments of angiogenic capacity in gastric cancer patients. By integrating transcriptome data, the platform provides crucial insights into patient-specific angiogenesis, supporting diagnosis and personalized therapy. Bridging genomic profiling with functional angiogenesis assessments, this platform represents a significant advancement in translational cancer research and a promising tool for personalized treatment strategies and understanding tumor angiogenesis.

Ethics approval and consent to participate

All patients who participated in this study and provided biospecimens signed consent forms approved by the Institutional Review Board (IRB) of Samsung Medical Center (SMC) (IRB #2021-09-052). The study was conducted in accordance with the principles of the Declaration of Helsinki and the Guidelines for Good Clinical Practice (ClinicalTrials.gov identifier: NCT02589496).

Author contributions

Conceptualization: SH, JK, and JL; methodology: SH, JK, MA, and JL; investigation/data analysis: SH, JK, STK, SHP, JYH, SHL, KMK, and JL; visualization: SH, JK, and MA; funding acquisition: JL; supervision: JL; writing original draft: all authors; and writing – review and editing: all authors.

Conflicts of interest

JL has served as a consultant/advisor in Mirati, Oncxerna, Seattle Genetics, Turning Point Therapeutics, and Astra Zeneca. The authors declare that they have no competing interests.

Abbreviations

AGS	Angiogenesis gene signature
ARID1A	AT-rich interaction domain 1A
bp	Base pair
CDH1	Cadherin 1
CIN	Chromosomal instability
DPBS	Dulbecco's phosphate-buffered saline
EC	Endothelial cell
EGM2	Endothelial cell growth medium-2
EGF	Epidermal growth factor
EBV	Epstein–Barr virus
GC	Gastric cancer
gDNA	Genomic DNA
GS	Genomically stable
IRB	Institutional review board
MSI	Microsatellite instability
MUC6	Mucin 6
PDC	Patient-derived cell
PDTS	Patient-derived tumor spheroid
PCR	Polymerase chain reaction
RBC	Red blood cell
RNF43	Ring-finger protein 43
SMC	Samsung medical center
TCGA	The cancer genome atlas
TME	Tumor microenvironment
UEA I	Ulex europaeus agglutinin I
VEGF	Vascular endothelial growth

Data availability

All data generated or analyzed during this study are included in this published article and its ESI.†

Acknowledgements

This research was supported by the Bio&Medical Technology Development Program of the National Research Foundation, funded by the Korean Government (MSIT) (No. RS-2023-00222838). Figures were produced using BioRender (https://biorender.com).

References

1. E. Morgan, M. Arnold, M. C. Camargo, A. Gini, A. T. Kunzmann, T. Matsuda, F. Meheus, R. H. A. Verhoeven, J. Vignat, M. Laversanne, J. Ferlay and I. Soerjomataram, EClinicalMedicine, 2022, 47, 101404.
2. M. Sekiguchi, I. Oda, T. Matsuda and Y. Saito, Digestion, 2022, 103, 22–28.
3. S. T. Kim, J. K. Sa, S. Y. Oh, K. Kim, J. Y. Hong, W. K. Kang, K. M. Kim and J. Lee, Genome Med., 2021, 13, 11.
4. B. Li, F. Zhang, Q. Niu, J. Liu, Y. Yu, P. Wang, S. Zhang, H. Zhang and Z. Wang, Mol. Ther. Nucleic Acids, 2023, 31, 224–240.
5. Q. Zhou, Y. Yuan, H. Lu, X. Li, Z. Liu, J. Gan, Z. Yue, J. Wu, J. Sheng and L. Xin, J. Transl. Med., 2023, 21, 80.
6. Z. Li, X. Gao, X. Peng, M. J. May Chen, Z. Li, B. Wei, X. Wen, B. Wei, Y. Dong, Z. Bu, A. Wu, Q. Wu, L. Tang, Z. Li, Y. Liu, L. Zhang, S. Jia, L. Zhang, F. Shan, J. Zhang, X. Wu, X. Ji, K. Ji, X. Wu, J. Shi, X. Xing, J. Wu, G. Lv, L. Shen, X. Ji, H. Liang and J. Ji, Sci. Adv., 2020, 6, eaay4211.
7. M. R. Carvalho, D. Barata, L. M. Teixeira, S. Giselbrecht, R. L. Reis, J. M. Oliveira, R. Truckenmuller and P. Habibovic, Sci. Adv., 2019, 5, eaaw1317.
8. K. Hiratsuka, T. Miyoshi, K. T. Kroll, N. R. Gupta, M. T. Valerius, T. Ferrante, M. Yamashita, J. A. Lewis and R. Morizane, Sci. Adv., 2022, 8, eabq0866.
9. W. Li, Z. Zhou, X. Zhou, B. L. Khoo, R. Gunawan, Y. R. Chin, L. Zhang, C. Yi, X. Guan and M. Yang, Adv. Healthcare Mater., 2023, 12, e2202609.
10. V. Surendran, D. Rutledge, R. Colmon and A. Chandrasekaran, Biofabrication, 2021, 13.
11. S. Li, K. Yang, X. Chen, X. Zhu, H. Zhou, P. Li, Y. Chen, Y. Jiang, T. Li, X. Qin, H. Yang, C. Wu, B. Ji, F. You and Y. Liu, Biofabrication, 2021, 13.
12. X. Liu, J. Fang, S. Huang, X. Wu, X. Xie, J. Wang, F. Liu, M. Zhang, Z. Peng and N. Hu, Microsyst. Nanoeng., 2021, 7, 50.
13. P. Ramos, M. R. Carvalho, W. Chen, L. P. Yan, C. H. Zhang, Y. L. He, R. L. Reis and J. M. Oliveira, Biofabrication, 2023, 15, 032001.
14. X. Cao, R. Ashfaq, F. Cheng, S. Maharjan, J. Li, G. Ying, S. Hassan, H. Xiao, K. Yue and Y. S. Zhang, Adv. Funct. Mater., 2019, 29, 1807173.
15. Y. Cho, K. Na, Y. Jun, J. Won, J. H. Yang and S. Chung, Front. Bioeng. Biotechnol., 2021, 9, 697657.
16. S. R. Lee, Y. Kim, S. Kim, J. Kim, S. Park, S. Rhee, D. Park, B. Lee, K. Baek, H. Y. Kim and N. L. Jeon, Microsyst. Nanoeng., 2022, 8, 126.
17. M. B. Chen, J. A. Whisler, J. Frose, C. Yu, Y. Shin and R. D. Kamm, Nat. Protoc., 2017, 12, 865–880.
18. L. Neufeld, E. Yeini, N. Reisman, Y. Shtilerman, D. Ben-Shushan, S. Pozzi, A. Madi, G. Tiram, A. Eldar-Boock, S. Ferber, R. Grossman, Z. Ram and R. Satchi-Fainaro, Sci. Adv., 2021, 7, eabi9119.
19. H. F. Wang, R. Ran, Y. Liu, Y. Hui, B. Zeng, D. Chen, D. A. Weitz and C. X. Zhao, ACS Nano, 2018, 12, 11600–11609.
20. K. Cibulskis, M. S. Lawrence, S. L. Carter, A. Sivachenko, D. Jaffe, C. Sougnez, S. Gabriel, M. Meyerson, E. S. Lander and G. Getz, Nat. Biotechnol., 2013, 31, 213–219.
21. S. T. Sherry, M. H. Ward, M. Kholodov, J. Baker, L. Phan, E. M. Smigielski and K. Sirotkin, Nucleic Acids Res., 2001, 29, 308–311.
22. C. Genomes Project, A. Auton, L. D. Brooks, R. M. Durbin, E. P. Garrison, H. M. Kang, J. O. Korbel, J. L. Marchini, S. McCarthy, G. A. McVean and G. R. Abecasis, Nature, 2015, 526, 68–74.
23. C. International HapMap, D. M. Altshuler, R. A. Gibbs, L. Peltonen, D. M. Altshuler, R. A. Gibbs, L. Peltonen, E. Dermitzakis, S. F. Schaffner, F. Yu, L. Peltonen, E. Dermitzakis, P. E. Bonnen, D. M. Altshuler, R. A. Gibbs, P. I. de Bakker, P. Deloukas, S. B. Gabriel, R. Gwilliam, S. Hunt, M. Inouye, X. Jia, A. Palotie, M. Parkin, P. Whittaker, F. Yu, K. Chang, A. Hawes, L. R. Lewis, Y. Ren, D. Wheeler, R. A. Gibbs, D. M. Muzny, C. Barnes, K. Darvishi, M. Hurles, J. M. Korn, K. Kristiansson, C. Lee, S. A. McCarrol, J. Nemesh, E. Dermitzakis, A. Keinan, S. B. Montgomery, S. Pollack, A. L. Price, N. Soranzo, P. E. Bonnen, R. A. Gibbs, C. Gonzaga-Jauregui, A. Keinan, A. L. Price, F. Yu, V. Anttila, W. Brodeur, M. J. Daly, S. Leslie, G. McVean, L. Moutsianas, H. Nguyen, S. F. Schaffner, Q. Zhang, M. J. Ghori, R. McGinnis, W. McLaren, S. Pollack, A. L. Price, S. F. Schaffner, F. Takeuchi, S. R. Grossman, I. Shlyakhter, E. B. Hostetter, P. C. Sabeti, C. A. Adebamowo, M. W. Foster, D. R. Gordon, J. Licinio, M. C. Manca, P. A. Marshall, I. Matsuda, D. Ngare, V. O. Wang, D. Reddy, C. N. Rotimi, C. D. Royal, R. R. Sharp, C. Zeng, L. D. Brooks and J. E. McEwen, Nature, 2010, 467, 52–58.
24. W. McLaren, L. Gil, S. E. Hunt, H. S. Riat, G. R. Ritchie, A. Thormann, P. Flicek and F. Cunningham, Genome Biol., 2016, 17, 122.
25. M. A. Walker, C. S. Pedamallu, A. I. Ojesina, S. Bullman, T. Sharpe, C. W. Whelan and M. Meyerson, Bioinformatics, 2018, 34, 4287–4289.
26. B. Niu, K. Ye, Q. Zhang, C. Lu, M. Xie, M. D. McLellan, M. C. Wendl and L. Ding, Bioinformatics, 2014, 30, 1015–1016.
27. R. Shen and V. E. Seshan, Nucleic Acids Res., 2016, 44, e131.
28. N. Cancer Genome Atlas Research, Nature, 2014, 513, 202–209.
29. R. Cristescu, J. Lee, M. Nebozhyn, K. M. Kim, J. C. Ting, S. S. Wong, J. Liu, Y. G. Yue, J. Wang, K. Yu, X. S. Ye, I. G. Do, S. Liu, L. Gong, J. Fu, J. G. Jin, M. G. Choi, T. S. Sohn, J. H. Lee, J. M. Bae, S. T. Kim, S. H. Park, I. Sohn, S. H. Jung, P. Tan, R. Chen, J. Hardwick, W. K. Kang, M. Ayers, D. Hongyue, C. Reinhard, A. Loboda, S. Kim and A. Aggarwal, Nat. Med., 2015, 21, 449–456.
30. A. Dobin, C. A. Davis, F. Schlesinger, J. Drenkow, C. Zaleski, S. Jha, P. Batut, M. Chaisson and T. R. Gingeras, Bioinformatics, 2013, 29, 15–21.
31. B. Li and C. N. Dewey, BMC Bioinf., 2011, 12, 323.
32. S. Hanzelmann, R. Castelo and J. Guinney, BMC Bioinf., 2013, 14, 7.
33. A. Bagaev, N. Kotlov, K. Nomie, V. Svekolkin, A. Gafurov, O. Isaeva, N. Osokin, I. Kozlov, F. Frenkel, O. Gancharova, N. Almog, M. Tsiper, R. Ataullakhanov and N. Fowler, Cancer Cell, 2021, 39, 845–865.e847.
34. J. K. Lee, Z. Liu, J. K. Sa, S. Shin, J. Wang, M. Bordyuh, H. J. Cho, O. Elliott, T. Chu, S. W. Choi, D. I. S. Rosenbloom, I. H. Lee, Y. J. Shin, H. J. Kang, D. Kim, S. Y. Kim, M. H. Sim, J. Kim, T. Lee, Y. J. Seo, H. Shin, M. Lee, S. H. Kim, Y. J. Kwon, J. W. Oh, M. Song, M. Kim, D. S. Kong, J. W. Choi, H. J. Seol, J. I. Lee, S. T. Kim, J. O. Park, K. M. Kim, S. Y. Song, J. W. Lee, H. C. Kim, J. E. Lee, M. G. Choi, S. W. Seo, Y. M. Shim, J. I. Zo, B. C. Jeong, Y. Yoon, G. H. Ryu, N. K. D. Kim, J. S. Bae, W. Y. Park, J. Lee, R. G. W. Verhaak, A. Iavarone, J. Lee, R. Rabadan and D. H. Nam, Nat. Genet., 2018, 50, 1399–1411.
35. S. Y. Lee, S. Byeon, J. Ko, S. Hyung, I. K. Lee, N. L. Jeon, J. Y. Hong, S. T. Kim, S. H. Park and J. Lee, Cancer Med., 2021, 10, 7253–7262.
36. J. H. Cho, J. S. Kim, S. T. Kim, J. Y. Hong, J. O. Park, Y. S. Park, D. H. Nam, D. W. Lee and J. Lee, PLoS One, 2019, 14, e0215080.
37. R. Lugano, M. Ramachandran and A. Dimberg, Cell. Mol. Life Sci., 2020, 77, 1745–1770.
38. X. Jiang, J. Wang, X. Deng, F. Xiong, S. Zhang, Z. Gong, X. Li, K. Cao, H. Deng, Y. He, Q. Liao, B. Xiang, M. Zhou, C. Guo, Z. Zeng, G. Li, X. Li and W. Xiong, J. Exp. Clin. Cancer Res., 2020, 39, 204.
39. M. Nilsson and J. V. Heymach, J. Thorac. Oncol., 2006, 1, 768–770.
40. A. S. Crystal, A. T. Shaw, L. V. Sequist, L. Friboulet, M. J. Niederst, E. L. Lockerman, R. L. Frias, J. F. Gainor, A. Amzallag, P. Greninger, D. Lee, A. Kalsy, M. Gomez-Caraballo, L. Elamine, E. Howe, W. Hur, E. Lifshits, H. E. Robinson, R. Katayama, A. C. Faber, M. M. Awad, S. Ramaswamy, M. Mino-Kenudson, A. J. Iafrate, C. H. Benes and J. A. Engelman, Science, 2014, 346, 1480–1486.
41. T. J. Giordano, Endocr. Pathol., 2014, 25, 362–365.
42. A. Gupta, R. Sedhom and M. S. Beg, JAMA Oncol., 2020, 6, 308.
43. M. K. Kilgour, S. MacPherson, L. G. Zacharias, J. LeBlanc, S. Babinszky, G. Kowalchuk, S. Parks, R. D. Sheldon, R. G. Jones, R. J. DeBerardinis, P. T. Hamilton, P. H. Watson and J. J. Lum, Nat. Protoc., 2022, 17, 2668–2698.
44. J. Ko, J. Ahn, S. Kim, Y. Lee, J. Lee, D. Park and N. L. Jeon, Lab Chip, 2019, 19, 2822–2833.
45. Y. Qiu, N. Wang, T. Guo, S. Liu, X. Tang, Z. Zhong, Q. Chen, H. Wu, X. Li, J. Wang, S. Zhang, Y. Ou, B. Wang, K. Ma, W. Gu, J. Cao, H. Chen and Y. Duan, Biomater. Sci., 2021, 9, 6064–6085.
46. S. Kim, H. Lee, M. Chung and N. L. Jeon, Lab Chip, 2013, 13, 1489–1500.
47. C. P. Miller, C. Tsuchida, Y. Zheng, J. Himmelfarb and S. Akilesh, Neoplasia, 2018, 20, 610–620.
48. D. Kim, K. S. Hwang, E. U. Seo, S. Seo, B. C. Lee, N. Choi, J. Choi and H. N. Kim, Adv. Healthcare Mater., 2022, 11, e2102581.
49. S. K. Garattini, D. Basile, M. Cattaneo, V. Fanotto, E. Ongaro, M. Bonotto, F. V. Negri, R. Berenato, P. Ermacora, G. G. Cardellino, M. Giovannoni, N. Pella, M. Scartozzi, L. Antonuzzo, N. Silvestris, G. Fasola and G. Aprile, World J. Gastrointest. Oncol., 2017, 9, 194–208.
50. G. Aprile, E. Ongaro, M. Del Re, S. E. Lutrino, M. Bonotto, L. Ferrari, K. Rihawi, G. G. Cardellino, N. Pella, R. Danesi and G. Fasola, Crit. Rev. Oncol. Hematol., 2015, 95, 165–178.

---

Footnotes

† Electronic supplementary information (ESI) available. See DOI: https://doi.org/10.1039/d5tb00577a

‡ These authors contributed equally to this study.

---