Establishment, biobanking, and multi-omics characterization of 41 patient-derived colorectal cancer organoids: MYC/PRC classification

Highlights

• •Integrated multi-omics profiling of colorectal cancer PDOs reveals a MYC–PRC transcriptional axis reflecting stem-like and differentiated tumor states.
• •The Stem Maturation Index (SMI) stratifies CRC organoids into biologically distinct MYC and PRC subtypes.
• •MYC-type organoids display proliferative signaling programs and increased sensitivity to MEK/EGFR-targeted therapies.
• •The MYC–PRC axis is reproducible across large scale PDO cohorts and the TCGA-COAD dataset.
• •This framework links tumor differentiation states with therapeutic vulnerabilities in colorectal cancer.

Abstract

Purpose

Colorectal cancer (CRC) exhibits marked genetic, transcriptomic, and phenotypic heterogeneity, limiting the robustness of existing molecular classification systems. We aimed to apply and validate an SMI-based MYC/PRC classification framework in CRC patient-derived organoids and to characterize its molecular, pharmacologic, and clinical relevance across independent cohorts.

Methods

We established 41 CRC PDOs from 28 patients and performed whole-exome sequencing, RNA sequencing, and high-throughput drug screening. SMI scores (−1 to 1) were used to classify PDOs as MYC-type (stem-like) or PRC-type (differentiated). A concordantly classified subset of 25 PDOs underwent integrative multi-omics analyses. Findings were validated in an expanded cohort of 181 CRC PDOs and in TCGA-COAD. We compared pathway activities, mutational profiles, and drug responses, and performed archetypal modeling (K = 2) to position samples along a MYC–PRC transcriptional continuum.

Results

MYC-type PDOs exhibited enrichment of cell-cycle and proliferation-related programs, whereas PRC-type PDOs showed differentiation-associated transcriptional programs, including coagulation and NF-κB signaling. Drug screening showed greater sensitivity of MYC-type PDOs to MEK/EGFR-targeted agents, whereas PRC-type PDOs displayed more heterogeneous responses. Archetypal modeling preserved the α(PRC)–α(MYC) trade-off and sample ordering across feature sets and larger cohorts. Validation in the expanded PDO dataset and TCGA-COAD reproduced subtype-specific transcriptional patterns and supported their clinical relevance.

Conclusions

SMI-based MYC/PRC classification captures biologically coherent tumor states with distinct molecular features and therapeutic sensitivities. This transcriptome-based framework complements existing classification systems and may support subtype-informed therapeutic prioritization in CRC.

Introduction

Colorectal cancer (CRC) is a biologically and clinically heterogeneous malignancy driven by diverse genetic, transcriptomic, epigenetic, and microenvironmental alterations. Over the past decade, several molecular classification systems have been proposed to stratify CRC into biologically meaningful subtypes such as Consensus Molecular Subtypes (CMS) [1], CRCAssigner [2], and Colorectal Cancer Intrinsic Subtypes (CRIS) [3]. More recently, pathway-level classification, termed pathway-derived subtypes (PDS), has been proposed. PDS leverages transcriptome-derived pathway activity to stratify tumors into PDS1–PDS3, revealing a functional axis that cuts across existing taxonomies; notably, it identifies within CMS2 a slow-cycling, differentiation-leaning subset (PDS3) that is associated with worse outcomes in locally advanced disease [4]. Recent reviews have synthesized advances in CRC molecular taxonomy, integrating bulk and single-cell RNA analyses and highlighting the tumor microenvironment (TME) as a key prognostic component [5]. While these frameworks have advanced our understanding of CRC biology and treatment response, a substantial proportion of tumors remain unclassified or show overlap between categories due to the dynamic plasticity of cancer cells and intra-tumoral heterogeneity [6]. These limitations underscore the need for a classification system that not only reflects static molecular features but also captures functional states associated with tumor aggressiveness and therapy response. To address this gap, we applied the Stem Maturation Index (SMI), a transcriptome-based metric derived from MYC/PRC target activity and implemented in the PDS Classifier R package, to stratify colorectal cancer PDOs along a stemness-differentiation continuum, yielding reproducible MYC/PRC subtypes associated with distinct pathways and drug responses. The SMI framework originates from the PDS classification system originally developed and validated in bulk tumor data; notably, Malla et al. explicitly observed that no existing preclinical model faithfully recapitulates the PRC-like (PDS3) biological state, highlighting a gap between transcriptomic classification and functional preclinical modeling. The present study addresses this gap by demonstrating that SMI-based MYC/PRC classification is reproducible in an epithelial-only PDO system- a context in which TME-dependent immune and stromal signals central to CMS and CRIS are absent and by directly linking subtype membership to differential pharmacologic vulnerability through systematic drug screening. This functional integration of transcriptomic subtyping with drug response profiling represents the primary added value of our approach relative to CMS, CRIS, and the original PDS framework. Although this binary classification simplifies the biological complexity of CRC, it offers a functionally grounded and complementary perspective to existing systems and may be particularly informative for identifying subtype-informed druggable vulnerabilities. In this study, we analyzed a cohort of 41 patient-derived CRC organoids (PDOs), a model system known to preserve the genetic and phenotypic heterogeneity of primary tumors [7,8]. By applying the SMI framework, we classified PDOs along the stemness–differentiation axis and restricted integrative analyses to a concordantly classified subset of 25 PDOs for multi-omics profiling, integrating transcriptomic, mutational, and pharmacological data [9,10]. To assess whether this selection introduced systematic bias, we compared clinical and genomic characteristics between the 25 included and 16 excluded PDOs (Supplementary Table 7 and Supplementary Fig. 1). No significant differences were observed in age (p = 0.125), sex (p = 0.344), AJCC stage (p = 0.747), MSI status (p = 0.143), KRAS mutation frequency (p = 1.00), tumor location (p = 0.225), or MYC/PRC subtype distribution (p = 0.268). Importantly, the 16 excluded PDOs did not represent a biologically intermediate subtype but rather showed discordant MYC/PRC assignments between the 41 PDO and 181 PDO cohort classifications, reflecting transcriptomic borderline scores rather than a distinct biological state. We validated the reproducibility and molecular relevance of the MYC/PRC classification using an expanded dataset of 181 CRC PDOs and TCGA-COAD cohorts. Our analyses revealed distinct molecular signatures, pathway activities, and drug sensitivity profiles between MYC and PRC subtypes, highlighting the potential utility of SMI-based classification in precision oncology. Patient-derived colorectal cancer (CRC) organoids generated in this study are available through the Korean Organoid Biobank (https://organoid.snu.ac.kr) upon request and in accordance with institutional and biobank policies.

Materials and methods

Ethics statement

The research protocol was reviewed and approved by the Institutional Review Board of Seoul National University Hospital (IRB No 1102–098–357). This study was conducted in accordance with the principles of the Declaration of Helsinki. Written informed consent was obtained from all patients enrolled in this study.

MSS/MSI analysis

Microsatellite instability (MSI) status was evaluated using the MSI Analysis System Kit (Promega) according to the manufacturer's protocol. The analysis involved Multiplex PCR amplification of seven markers, including five mononucleotide repeat markers (BAT-25, BAT-26, NR-21, NR-24, and MONO-27) and two pentanucleotide repeat markers (Penta C and Penta D), and the amplified products were analyzed by capillary electrophoresis in a single injection using the ABI 3730 DNA Analyzer (Applied Biosystems). Detailed results for each sample are provided in Supplementary Table 1. Data analysis was performed using the GeneMapper V5.0 software. Samples showing instability in two or more markers were classified as MSI-high (MSI-H), whereas those without detectable alterations were classified as microsatellite stable (MSS).

Sample collection and preparation

A total of 41 multifocal specimens of human colorectal tumors were collected from 28 CRC patients who underwent radical resection at Seoul National University Hospital (Seoul, Korea). The tumor tissues were histologically diagnosed as adenocarcinomas by a pathologist. Surgically resected tumor tissues were submerged in Opti-MEM supplemented with 1% penicillin/streptomycin (GIBCO, Cat#15140–122) and transferred to an ice box directly from the operating room to the laboratory. The resected colon tissue was unfolded laterally to expose the neoplastic gland. Detailed information regarding the tumor size and location is summarized in Supplementary Table 2. The starting material information is indicated by the last characters of the sample names: TO for Patient-Derived Tumor Organoids, NT for Normal Tissue. Genomic DNA was extracted from resected tumor tissues and paired adjacent normal mucosa using the AllPrep DNA/RNA Mini Kit (Qiagen, Hilden, Germany), according to the manufacturer's protocol. We performed CODIS STR Analysis or Extended CODIS STR Analysis to assess the genetic profiles of 41 colorectal cancer (CRC) organoids along with their matched normal and tumor tissues. Subsequently, among the 41 established PDOs, we generated whole-exome sequencing (WES; n = 37) and RNA-seq (n = 39) libraries. All libraries passed quality control (read quality, mapping rate, and duplicate metrics) and were used for analyses. For RNA-seq, three outlier samples identified after normalization were excluded, yielding 36 samples for transcriptomic analyses.

Tumor isolation, organoid culture, and passaging

Primary tumor tissues were photographed, labeled according to regional origin, and washed five times with cold PBS supplemented with penicillin/streptomycin (GIBCO, Cat#15140–122). Tissues were minced on sterile ice-cold dishes and enzymatically digested at 37 °C for 30–60 min in a medium containing collagenase II (1.5 mg mL⁻¹, GIBCO, Cat#17101–015), hyaluronidase (20 µg mL⁻¹, Sigma-Aldrich, Cat#H3506), and Y-27632 (10 µM, Selleckchem, Cat#S1049). The cell suspension was filtered through a 100 µm strainer (SPL Life Sciences, Cat#93100), centrifuged at 1000 rpm for 3 min, washed twice in basal medium (DMEM/F12 supplemented with 1% penicillin/streptomycin, GlutaMAX 10 mM, and 10% FBS), and subjected to red blood cell lysis (Sigma-Aldrich, Cat#R7757) for 1–5 min. Cells were embedded in Basement Membrane Extract (BME, GIBCO, Cat#A14132–02) and plated as 50 µL domes (∼20,000 cells mL⁻¹) in T-25 flasks (Corning, Cat#353108). After polymerization at 37 °C for 10 min, 5 mL of Human Intestinal Stem Cell (HISC) medium was added, and cultures were maintained at 37 °C in 5% CO₂. HISC medium consisted of basal medium and L-WRN conditioned medium supplemented with B27, human EGF (50 ng mL⁻¹), human FGF-10 (10 ng mL⁻¹), nicotinamide (10 mM), N-acetylcysteine (1.25 mM), A83–01 (500 nM), SB202190 (3 µM), prostaglandin E2, and primocin (100 µg mL⁻¹). For passaging, BME domes were mechanically disrupted and incubated with TrypLE Express (GIBCO, Cat#12604–021) at 37 °C for 5–10 min, centrifuged at 1000 rpm for 3 min, resuspended in fresh BME, replated as domes, and overlaid with HISC medium.

Organoid seeding and drug treatment

Drug screening was conducted in two independent biological experiments across 25 PDOs (MYC, n = 15; PRC, n = 10), with technical triplicates included for each condition in each experiment. Organoids were enzymatically and mechanically dissociated into single cells by incubating them in TrypLE Express solution (GIBCO) for 5 to 10 min, followed by pipetting. The mixture was then centrifuged at 1500 rpm for 3 min. After removing the BME, the cell pellet was re-suspended in a 1:1 mixture of HISC medium and BME (GIBCO). A homogeneous BME mixture (40 µL, 20,000 organoids/mL) was plated in clear-bottomed white-walled 96-well plates (Corning, Cat# 3903) using E1-ClipTip Electronic Pipettes (Thermo Fisher Scientific). Plates were incubated at 37 °C with 5% CO2 for 15 min to solidify the BME before adding pre-warmed HISC medium (40 µL) to each well using an EpMotion 5070 (Eppendorf Corporation). Drug solutions (20 µL) were added to each well at 96 h after seeding. Drugs were serially diluted in DPBS at a 1:3 ratio from the maximum concentration to prepare six concentration points. Control wells were treated with a mixture of HISC medium and drug solvent solution. Drug max concentrations are summarized in Supplementary Table 3.

Drug sensitivity assay and analysis

After drug treatment, cell viability was quantified using ATP-based luminescence assays CellTiter-Glo 3D reagent (Promega, #G968B), according to the manufacturer’s instructions. Briefly, an equal volume of reagent was added to each well, followed by orbital shaking to ensure complete cell lysis. After incubation at room temperature to stabilize the luminescent signal, ATP-dependent luminescence was measured using a Luminoskan Ascent plate reader (Thermo Scientific). Luminescence values were normalized to vehicle-treated (DMSO) controls maintained in parallel under identical culture conditions for each individual PDO to calculate relative viability. Critically, DMSO-treated control wells were seeded and maintained under the same conditions as drug-treated wells for each PDO, such that the luminescence signal from vehicle-treated wells captures the baseline proliferative activity of each organoid at the time of measurement. Normalization of drug-treated luminescence values to these PDO-specific DMSO controls therefore helps reduce the influence of inter-organoid differences in baseline proliferative activity, allowing the resulting relative viability values and derived AUC metrics to better reflect drug-specific cytotoxic or cytostatic effects under matched experimental conditions. Drug response curves were generated from dose-response series, and global sensitivity was quantified using the area under the curve (AUC). Nonlinear regression modeling was performed in GraphPad Prism (version 7, GraphPad Software). To compare drug responses between MYC and PRC subtypes, two-sided Wilcoxon rank-sum tests were applied to AUC values, with Benjamini-Hochberg FDR correction across drugs. For dose-response curves, differences between MYC and PRC at each drug concentration were assessed using two-sided Welch’s t-test. When multiple doses were tested, AUC values provided a global summary of sensitivity, while per-dose p-values were reported for visualization in the plots. For the representative dose-response plots (Fig. 3B, C), PDOs were selected to illustrate characteristic subtype-associated response profiles based on proximity to the median Trametinib AUC within each subtype. To address the potential influence of patient-level clustering, representative PDOs were restricted to those derived from independent patients, with no more than one PDO per patient included in the representative curves. Although multifocal PDOs from the same patient were treated as biologically independent samples in the full 25-PDO cohort analysis—given their derivation from spatially distinct tumor sites and distinct transcriptomic profiles—the representative curves were constructed using PDOs from entirely independent patients to provide the most conservative illustration of subtype-associated drug response differences.

Fig. 3.

Differential drug response patterns in MYC/PRC organoid types reveal type-specific therapeutic sensitivities.

(A) Heatmap of drug response across PDOs grouped by MYC/PRC subtype. Drug response is summarized as area under the dose–response curve (AUC) and drugs are annotated by therapeutic class (e.g., EGFR inhibitors, MEK/EGFR signaling inhibitors, DNA repair and cell-cycle modulators, cytotoxic agents). (B) Dose–response curves for trametinib (MEK inhibitor) and afatinib (EGFR inhibitor) across representative MYC and PRC PDOs; dose-specific statistical comparisons are indicated (p-values shown). (C) Dose–response curves for additional agents (olaparib, AZD2014, oxaliplatin, MK-5108) illustrating heterogeneous response patterns across PDOs.

Confocal microscopy

Organoids were recovered from BME with Cell Recovery Solution, fixed, blocked in 0.1% BSA/PBS-T, incubated overnight at 4 °C with anti-CD133 antibody, and stained with secondary antibody, TRITC-phalloidin, and Hoechst 33342 (10 μg/mL, Thermo Fisher Scientific). Samples were mounted on chamber slides and imaged using a Leica Stellaris 5 confocal microscope.

Whole-exome sequencing

Genomic DNA from organoids was extracted using the QIAamp DNA Mini Kit (Qiagen, Hilden, Germany) whereas DNA from tissue was extracted using a QIAamp Fast DNA Tissue Kit (Qiagen). Whole-exome capture (SureSelect Human All Exon V5, Agilent) was applied to all samples, including normal colon mucosa, and libraries were sequenced on a HiSeq 2500 platform (Illumina) with 100 bp paired-end reads for organoids and 200 bp for tissues. Reads were aligned to hg19 using BWA (v0.7.12), followed by duplicate removal (Picard-tools v1.130) and local realignment and base quality score recalibration with GATK. Variant calling was performed via HaplotypeCaller (GATKv3.4.0), and somatic mutations (SNVs and short indels) were identified using MuTect (GATK v3.8.0) in the tumor-normal mode. Annotations were performed with SnpEff (v4.1 g), dbSNP (v142), 1000 Genomes, ESP6500, ClinVar, and dbNSFP 2.9. Matched normal tissues from 20 patients were used to exclude variants found in the normal samples, thereby identifying somatic mutations (Supplementary Table 4). The somatic mutation oncoplot for the MYC and PRC groups is presented in Fig. 1C.

Fig. 1.

Histopathological and immunohistochemical validations of patient-derived colorectal organoids demonstrating morphological and marker concordance with original tumor tissue.

(A) Alluvial plot integrating Consensus Molecular Subtypes (CMS1–4), Colorectal Cancer Intrinsic Subtypes (CRIS-A–E), and epithelial cell-state programs (Enterocyte, Goblet-like, Inflammatory, and Transit-Amplifying [TA]) across the CRC PDO biobank. (B) Representative histology and phenotypic validation of PDOs. Top: H&E-stained FFPE sections showing preservation of glandular epithelial architecture. Bottom: Confocal immunofluorescence images of SNU-4707S2-TO, SNU-4707S3-TO, and SNU-4707S4-TO organoids demonstrating preserved cytoskeletal organization and CD133 expression in culture. Color channels: F-actin (red), CD133 (green), Hoechst nuclear stain (blue). Scale bars as indicated. (C) Oncoprint summarizing recurrent somatic alterations identified by whole-exome sequencing across PDOs, sorted by MYC/PRC subtype classification. Bars denote alteration frequencies and colors indicate variant classes. MYC and PRC subtype assignments are indicated in the annotation bar above the oncoprint. Supplementary Fig. 3 shows matched tumor-PDO oncoprints for 16 paired samples demonstrating a mean driver mutation concordance rate of 94.44% across 32 cancer-related genes.

RNA sequencing

Total RNA was extracted from the cell lysates using TRIzol (Qiagen, Hilden, Germany) in combination with the Qiagen RNeasy Kit (Qiagen, Hilden, Germany). Paired-end sequencing reads of cDNA libraries (101 bp) generated using a NovaSeq 6000 instrument were evaluated for sequence quality using FastQC v0.11.7 [11]. Trimmomatic v0.38 was used for preprocessing, trimming low-quality bases and adapter sequences from the reads [12]. Cleaned reads were aligned to the human genome (UCSC hg19) using HISAT v2.1.0 [13], and transcript assembly was performed using StringTie v1.3.4 [14]. Based on these results, transcript and gene expression levels were quantified as read counts or transcripts per million mapped read (TPM). Differential expression analysis was performed using limma package in R [15].

To visualize transcriptomic differences, t-SNE (t-distributed Stochastic Neighbor Embedding) was applied using the Rtsne package [16]. Prior to t-SNE, we performed a Wilcoxon rank-sum test for each gene and defined differentially expressed genes as those with Benjamini–Hochberg (BH)–adjusted p values < 0.05. Using this criterion, 2716 genes were selected for dimensionality reduction. This dimensionality reduction was used for visualization purposes only and was not used to define subtype membership. Pathway enrichment analysis was conducted using Gene Set Enrichment Analysis (GSEA) with the clusterProfiler package in R, which identified statistically significant gene sets distinguishing MYC and PRC subtypes [17,18]. Gene Set Variation Analysis (GSVA) was applied to compute sample-wise pathway enrichment scores, enabling continuous assessment of pathway activity across the PDO samples (Supplementary Table 5, Supporting Information) [19]. All statistical tests were two-sided. p-values were adjusted for multiple testing using the Benjamini–Hochberg method where applicable. Statistical analyses were conducted in R and significance was defined as q < 0.05 or p < 0.05 unless otherwise stated.

FFPE block production and H&E staining

The gel dome containing organoids was scraped from the flask using a cut pipette tip. PBS was added, and the mixture was transferred to a 1.5 mL tube. After a short centrifugation, the supernatant was aspirated to remove as much of the gel as possible, leaving only the organoids. The collected pellets were embedded in 2% agarose gel (iNtRON Biotechnology, Seongnam-si, Korea). The solidified agarose gel was fixed with 10% paraformaldehyde (PFA) for 30 min at room temperature (RT). Agarose blocks were processed before embedding in paraffin. Sections (4 μm in thickness) of the organoids were subjected to routine hematoxylin and eosin (H&E) staining.

Stem maturation index (SMI) based molecular classification and cross-cohort validation of PDOs

Transcriptomic profiles of colorectal cancer patient-derived organoids (PDOs) were analyzed using the Stem Maturation Index (SMI), a transcriptome-derived metric that quantifies the relative position of a sample along a stemness–differentiation continuum. SMI scores were calculated using the calculateSMI function implemented in the PDS Classifier R package, based on predefined gene expression signatures reflecting MYC/PRC-associated transcriptional programs. For binary subgroup assignment, zero was prespecified as the decision boundary, and samples were classified according to the sign of the score (SMI < 0, MYC; SMI > 0, PRC). Samples with an exact SMI value of 0, if present, were considered unassigned. To minimize classification ambiguity in downstream integrative analyses, we defined a concordant subset consisting of PDOs whose MYC/PRC assignment remained unchanged when the same zero-centered classification rule was applied to both the original 41 PDO cohort and the expanded 181 PDO organoid compendium. Detailed SMI scores and corresponding MYC/PRC subtype assignments for the 25 PDOs are provided in Supplementary Table 6. To evaluate potential selection bias introduced by this concordance-based exclusion criterion, clinical and genomic characteristics of the 25 included and 16 excluded PDOs were compared. Fisher's exact test was applied to categorical variables (sex, AJCC stage, MSI status, KRAS mutation status, tumor location, and MYC/PRC subtype distribution), and the independent samples t-test was applied to continuous variables (age) added in Supplementary Table 7 and Supplementary Fig. 1. While the SMI provides a linear scalar index reflecting the relative balance of MYC and PRC transcriptional programs and was used for binary subtype assignment, archetypal analysis was employed as a complementary multi-dimensional approach to identify extreme transcriptional states and confirm the geometric stability of the MYC–PRC axis. These two metrics serve distinct but complementary roles: the SMI enables straightforward sample classification, whereas archetypal modeling validates the robustness of the underlying transcriptional continuum across gene sets and cohort sizes.

Survival analysis in the TCGA-COAD cohort

To evaluate the biological relevance of this classification despite differences between primary tumor tissues from patients and PDO models, Kaplan-Meier survival analysis was performed using the survival and survminer packages in R to determine whether MYC and PRC subtypes were associated with significant differences in patient survival [20].

Multi-Omics integration

Multi-omics integration was performed using the DIABLO framework implemented in the mixOmics R package (version 6.22.0) [21,22]. RNA-seq data, drug response profiles, and pathway activity scores were normalized and filtered using modality-specific preprocessing pipelines prior to integration. The final input matrix consisted of three blocks: (i) gene expression (591 features), (ii) drug response data (16 features), and (iii) hallmark pathway activity scores (30 features), estimated using the ssGSEA method implemented in the GSVA package.

Archetypal modeling

We modeled the MYC–PRC axis using archetypal analysis (K = 2) on gene-wise z-scored log2(TPM+1) RNA-seq matrices. Three feature sets were used for the same cohort: (i) whole transcriptome after low-abundance filtering, (ii) the published PDS MYC/PRC signature, and (iii) a discriminator set (∼2000 genes) derived from the 25 profiled PDOs. The models were fitted with stepArchetypes (R archetypes v2.2–0.2; 25 random starts, 1000 iterations; Frobenius loss), and the minimum-RSS run was retained [23]. For each sample, coefficients α(MYC) and α(PRC) were non-negative and summed to 1; subtype was assigned by argmax(α) (high-confidence: α(MYC) ≥ 0.60 or α(MYC) ≤ 0.40 PRC-proximal). Full details are provided in Supplementary Fig. 2.

Results

Histopathological and immunohistochemical validations of patient-derived colorectal organoids demonstrating morphological and marker concordance with original tumor tissue

A living biobank of 41 patient-derived colorectal cancer organoids (PDOs) from 28 patients was generated and subjected to multi-omics profiling, including RNA sequencing, whole-exome sequencing (WES), and functional drug screening. To assess whether the biobank captures established CRC biological diversity, transcriptome-based subtype annotations were performed using multiple CRC classification frameworks (Fig. 1A). All major CMS categories were represented, and structured correspondences were observed upon alignment to CRIS intrinsic subtypes. Integration with epithelial cell-state programs further indicated that the PDOs span multiple differentiation states, including prominent representation of the transit-amplifying (TA) program together with enterocyte-like, goblet-like, and inflammatory programs (Fig. 1A), supporting broad coverage of CRC epithelial phenotypes.

Histologic and phenotypic fidelity to the originating tumors was evaluated next. Representative H&E-stained FFPE sections of PDOs derived from the same patient (SNU-4707) showed preservation of glandular epithelial architecture under culture conditions (Fig. 1B, top). Representative confocal immunofluorescence images of SNU-4707S2-TO, SNU-4707S3-TO, and SNU-4707S4-TO organoids stained for F-actin (red), CD133 (green), and Hoechst (blue) demonstrated preserved cytoskeletal organization and CD133 expression in culture (Fig. 1B, Bottom). To assess genomic fidelity between primary tumor tissues and corresponding PDOs, matched driver mutation concordance was evaluated across 16 tumor-PDO pairs spanning 32 cancer-related genes. A mean concordance rate of 94.44% was observed, where concordance was defined as the proportion of assessed genes showing matching mutation status per tumor–PDO pair averaged across all 16 pairs. Key CRC driver mutations (APC, TP53, KRAS) were consistently preserved in PDOs relative to their matched tumor tissues (Supplementary Fig. 3). These findings support the genomic fidelity of the established PDO models. In addition, the mutational landscape across the PDO cohort included alterations in canonical CRC drivers together with recurrent events involving chromatin regulation and DNA repair pathways (Fig. 1C), underscoring genetic diversity across the PDO resource.

SMI-based MYC/PRC classification reveals distinct pathway activation and stem maturation profiles

t-SNE visualization of transcriptome profiles in the 25 PDOs was performed using genes differentially expressed between MYC and PRC subtypes to illustrate subtype-discriminative transcriptional patterns (Fig. 2A), concordant with divergence along the SMI scores (Fig. 2B). Box plot showing the Stem Maturation Index (SMI) scores for each subgroup. The PRC subgroup exhibited significantly higher SMI scores compared with the MYC subgroup, suggesting more pronounced differentiated features in PRC-type samples. Also, differential expression analyses identified subtype-associated markers: MYC PDOs exhibited increased expression of proliferation-associated genes such as DTL, KIFC1, and WDR34, whereas PRC PDOs showed higher expression of differentiation and developmental regulators including HOXA6, HOXB8, and TBX3 (Fig. 2C).

Fig. 2.

SMI-based MYC/PRC classification reveals distinct pathway activation and stem maturation profiles.

(A) t-SNE embedding of RNA-seq profiles for 25 PDOs colored by SMI-derived subtype (MYC vs PRC), with marginal density distributions for each group. (B) Distribution of SMI scores in MYC and PRC PDOs (individual points overlaid). (C) Expression of representative subtype marker genes (MYC: DTL, KIFC1, WDR34; PRC: HOXA6, HOXB8, TBX3); p-values are shown in each panel. (D) Functional enrichment analysis of MYC-associated genes highlighting cell-cycle and DNA metabolism programs (dot size indicates gene count; color indicates adjusted p-value). (E) Representative GSEA plots showing hallmark pathways enriched in MYC (G2M checkpoint, E2F targets) and PRC (coagulation, TNFα signaling via NF-κB). Statistical parameters for each gene set are indicated within the plots: HALLMARK_G2M_CHECKPOINT (NES=3.75, FDR q = 1.15 × 10⁻⁹); HALLMARK_E2F_TARGETS (NES=3.83, FDR q = 1.15 × 10⁻⁹); HALLMARK_COAGULATION (NES=−1.82, FDR q = 0.0077); HALLMARK_TNFA_SIGNALING_VIA_NFKB (NES=−1.83, FDR q = 0.012).

Pathway-level analyses reinforced these biological differences. MYC-associated genes were enriched for cell-cycle and DNA metabolism programs (e.g., cell-cycle checkpoints, DNA replication, Fanconi anemia pathway, homologous recombination and mismatch repair–related processes), consistent with a proliferative, stem-like state (Fig. 2D). Gene set enrichment analyses further indicated preferential enrichment of hallmark proliferation programs (G2M checkpoint, E2F targets) in MYC PDOs, whereas PRC PDOs showed relative enrichment of programs including coagulation and TNF-α signaling via NF-κB (Fig. 2E). We note that these gene sets, while nominally classified as inflammatory or stromal in bulk tissue contexts, here reflect epithelial cell-intrinsic transcriptional activity: NF-κB target genes identified among the leading edge of HALLMARK_TNFA_SIGNALING_VIA_NFKB, including NR4A1, EGR3, and CDKN1A, are expressed by tumor epithelial cells independent of stromal or immune cell populations, consistent with their established roles in differentiation-associated epithelial programs. Collectively, these data indicate that SMI-derived MYC/PRC classification captures divergent functional states with coherent transcriptional and pathway-level signatures.

Differential drug response patterns in MYC/PRC organoid types reveal type-specific therapeutic sensitivities

The relationship between MYC/PRC tumor states and therapeutic vulnerability was evaluated using high-throughput drug screening. AUC-based response profiling revealed both shared and subtype-skewed response patterns across the 25 PDOs (Fig. 3A). Targeted agents within MEK/EGFR signaling showed prominent subtype-associated effects. To provide the most conservative assessments of subtype-associated drug response differences, representative PDOs were selected from entirely independent patients based on proximity to the median Trametinib AUC within each subtype (MYC: SNU-4707S2-TO, SNU-4717S1-TO, SNU-4718S3-TO; PRC: SNU-4834S2-TO, SNU-4983AS3-TO, SNU-5048S1-TO). Dose-response analyses confirmed greater sensitivity of MYC PDOs to the MEK inhibitor trametinib and the EGFR inhibitor afatinib compared with PRC PDOs, with the significant differences maintained at three dose points for trametinib and two dose points for afatinib (Fig. 3B; Asterisks indicate MYC–PRC differences (two-sided Welch's t-test at each dose): *p < 0.05; **p < 0.01; ***p < 0.001. Exact p-values are shown in the inset tables), consistent with increased dependence of MYC-high states on growth factor-linked signaling circuits. In contrast, several additional agents including olaparib, AZD2014, oxaliplatin, and MK-5108 exhibited heterogeneous responses with less consistent separation by subtype across the PDOs (Fig. 3C). Between-subtype differences in AUC were assessed using two-sided Wilcoxon rank-sum tests with Benjamini–Hochberg FDR correction across compounds.

Subtype-specific transcriptional correlates of MEK/EGFR inhibitor response

Given the differential sensitivity to trametinib and afatinib, associations between drug response and pathway activity were examined within each subtype. Within MYC PDOs, drug response (AUC) showed strong relationships with proliferation-linked signatures. Trametinib AUC inversely correlated with the MYC targets V1 score (R = –0.71, p = 0.003) in MYC PDOs but not in PRC PDOs (R = 0.05, p = 0.886) (Fig. 4A). Afatinib AUC inversely correlated with the WNT/β-catenin signaling score (R = –0.70, p = 0.004) in MYC PDOs, with no significant association in PRC PDOs (Fig. 4B). Afatinib response was also associated with MYC targets V1 activity in MYC PDOs (R = –0.66, p = 0.008) (Fig. 4C), and trametinib response inversely correlated with E2F targets activity (R = –0.65, p = 0.008) in MYC PDOs (Fig. 4D). These findings support subtype-specific coupling between proliferative transcriptional programs, pathway context, and MEK/EGFR inhibitor responsiveness. Correlation analyses were performed using Spearman’s rank correlation, and p-values were adjusted for multiple testing within each analysis panel where applicable (Benjamini–Hochberg). To summarize higher-order relationships among drug response, pathway activity, and gene expression, an association network was constructed using a stringent correlation threshold (|r| ≥ 0.8). The resulting chord diagram highlighted strong links connecting trametinib/afatinib response to proliferation-associated pathways (E2F targets, G2M checkpoint) and selected expression markers (including HMMR, ESPL1, and HSPA1A), with directionality reflected by positive versus negative correlation edges (Fig. 4E).

Fig. 4.

Subtype-specific transcriptional correlates of MEK/EGFR inhibitor response.

(A-C) Correlation analyses between pathway/signature scores and drug response (AUC) performed within MYC and PRC types: (A) MYC targets V1 score vs trametinib AUC; (B) WNT/β-catenin signaling score vs afatinib AUC; (C) MYC targets V1 score vs afatinib AUC; (D) E2F targets score vs trametinib AUC. Spearman correlation coefficients and p-values are displayed in each panel. (E) DIABLO circos plot integrating drug response, pathway activity scores, and gene expression features. Variables selected by the DIABLO model are grouped by data block: Drug (red), Gene Expression (blue), and Hallmark Pathway activity scores (green). Links denote variable pairs with an absolute Spearman correlation coefficient exceeding the predefined cutoff (|r| ≥ 0.8), irrespective of link thickness. Red and blue links indicate negative and positive correlations, respectively. Outer class traces depict subtype-specific patterns for MYC (blue) and PRC (orange) subtypes.

External validation of MYC/PRC subtypes in TCGA-COAD reveals distinct pathway activities and prognostic associations

Reproducibility of the MYC/PRC framework was assessed in an expanded CRC PDO transcriptome dataset (n = 181). Pathway activity profiling recapitulated separation between MYC and PRC states, with MYC samples enriched for proliferative hallmarks and PRC samples enriched for mesenchymal/differentiation-associated programs (Fig. 5A; Wilcoxon rank-sum test, all p < 1 × 10⁻¹⁷). Violin plots confirmed higher activity of HALLMARK_MYC_TARGETS_V1/V2 and HALLMARK_E2F_TARGETS in MYC, whereas PRC exhibited relative enrichment for signatures such as HALLMARK_MESENCHYMAL_TRANSITION, HALLMARK_PANCREAS_BETA_CELLS, and HALLMARK_MYOGENESIS (Fig. 5B).

Fig. 5.

External validation of MYC/PRC subtypes in TCGA-COAD reveals distinct pathway activities and prognostic associations.

(A) Heatmap of pathway activity profiles across an expanded CRC PDO cohort (n = 181) annotated by MYC/PRC subtype. (B) Violin plots of representative hallmark gene-set scores showing enrichment of proliferative programs (HALLMARK_MYC_TARGETS_V1/V2, HALLMARK_E2F_TARGETS) in MYC and mesenchymal/differentiation-associated programs (HALLMARK_MESENCHYMAL_TRANSITION, HALLMARK_PANCREAS_BETA_CELLS, HALLMARK_MYOGENESIS) in PRC; p-values are shown. (C) Kaplan–Meier overall survival analysis of TCGA-COAD patients (n = 349) stratified by MYC/PRC signature; log-rank p value is indicated. (D) Multivariable Cox proportional hazards model including MYC vs PRC and clinical covariates; hazard ratios with 95% confidence intervals and p-values are shown.

Clinical relevance was evaluated in TCGA-COAD (n = 349). Kaplan–Meier analysis demonstrated a significant overall survival difference between groups (log-rank p = 0.018), with the MYC subtype associated with poorer outcomes (Fig. 5C). In multivariable Cox proportional hazards modeling incorporating clinicopathologic covariates, MYC versus PRC remained significantly associated with hazard (Fig. 5D; model p = 0.034 for MYC vs PRC).

Archetype-based, MYC–PRC classification for CRC: reproducible subtyping aligned with MEK/EGFR sensitivity and compatible with clinical linkage

Archetypal modeling (K = 2) was used to project colorectal cancer (CRC) PDOs onto a MYC–PRC continuum to assess the robustness of subtype structure across feature sets and sample size. Using the same cohort but different input gene sets (A) the whole-transcriptome matrix, (B) the PDS MYC/PRC signature, and (C) the top ∼2000 discriminator genes obtained from a t-SNE–based separation of the 25 profiled PDOs all three analyses produced an essentially identical one-dimensional trade-off between alpha(PRC) and alpha(MYC) (≈ y = 1 − x) with consistent ordering of samples along the continuum (Fig. 6A–C). The 25 highlighted PDOs maintained their relative positions irrespective of the gene set, and subtype calls were stable across analyses. This cross gene-set reproducibility indicates that the MYC–PRC axis reflects a biologically stable subtype dimension rather than an artifact of a particular signature or of limited sample size. Importantly, PDOs near the MYC end of the continuum correspond to heightened sensitivity to MEK/EGFR-pathway inhibitors (e.g., trametinib, afatinib; Fig. 3), whereas PRC-proximal samples display more heterogeneous responses. Thus, the discriminator gene set derived from the 25 PDOs (Fig. 6C) can be used prospectively to stratify samples and prioritize MYC-targeted therapeutics in colorectal cancer, providing an actionable subtype framework. Application of the archetype model to TCGA-COAD (n = 319) reproduced a continuous MYC–PRC transcriptional axis in bulk tumors. Samples align along the same convex mixture constraint (α(MYC) + α(PRC) = 1), supporting the generalizability of the MYC–PRC continuum beyond PDOs and reinforcing its compatibility with patient-level clinical association analyses (Supplementary Fig. 2).

Fig. 6.

Archetype-based, MYC–PRC classification for CRC: reproducible subtyping aligned with MEK/EGFR sensitivity and compatible with clinical linkage.MYC–PRC archetype continuum reproduced across gene sets (A) Continuum obtained from the whole-transcriptome expression matrix (181 PDOs; 25 highlighted and labeled). (B) Continuum computed using the PDS study MYC/PRC gene set. (C) Continuum computed using the top ∼2000 genes that separate MYC and PRC groups in a t-SNE–based analysis of the 25 profiled PDOs, then applied to the full cohort. In all panels, the x-axis shows alpha(PRC archetype) and the y-axis shows alpha(MYC archetype) (convex-combination constraint; dashed line denotes y = 1 − x). Points are colored by subtype (MYC, pink; PRC, teal). The stable ordering and end-member enrichment across (A–C) demonstrate that the MYC–PRC axis is robust to the choice of feature set and to increased sample size, may provide a framework for subtype-informed therapeutic prioritization (Fig. 6A-C).

Discussion

Colorectal cancer (CRC) exhibits substantial intratumoral heterogeneity (ITH), contributing to variability in treatment response and clinical outcomes. Although organoid-based studies have characterized molecular diversity, few have directly linked heterogeneity to pharmacologic vulnerability and survival, underscoring the need for a classification framework applicable across both preclinical models and patient cohorts.

In this study, we established a biologically informed MYC/PRC classification using patient-derived organoids (PDOs) and integrative multi-omics analysis. Guided by the Stem Maturation Index (SMI), CRC PDOs were stratified into MYC and PRC subtypes with distinct molecular and therapeutic profiles. MYC-type organoids exhibited proliferative and cell cycle–associated programs, whereas PRC-type organoids were enriched for differentiation-associated transcriptional programs, including coagulation and NF-κB signaling. As PDOs are composed exclusively of tumor epithelial cells, these enrichment patterns reflect intrinsic epithelial transcriptional states rather than contributions from immune or stromal cell populations. Notably, leading edge NF-κB target genes including NR4A1, EGR3, and CDKN1A are well-established regulators of epithelial differentiation and stress response; the gene set nomenclature derived from bulk tumor analyses should be interpreted accordingly in the PDO context. From 181 CRC PDOs, a concordantly classified 25 PDO subset was used for integrative profiling, and the resulting MYC/PRC signature was applied to the TCGA-COAD cohort without modification. Subtype-specific molecular patterns and survival differences were reproduced, supporting the robustness and transferability of this framework despite a modest discovery set. Archetypal modeling further confirmed that the MYC/PRC axis represents a stable transcriptional continuum across alternative gene sets. MYC-proximal samples demonstrated greater sensitivity to MEK/EGFR-targeted agents, whereas PRC-proximal samples displayed more heterogeneous responses, suggesting that the transcriptome-derived MYC/PRC axis may serve as a practical stratification layer for subtype-aware therapeutic prioritization. In TCGA-COAD, the PRC subtype was significantly associated with improved overall survival compared with MYC in multivariable analysis (HR=0.092, 95% CI: 0.010–0.821, p = 0.032), corresponding to an approximately 10.9-fold higher hazard of death in MYC-type tumors. To assess independence from established CRC classification systems, we performed an additional multivariable Cox regression incorporating CMS 1–4 as a covariate (Supplementary Fig. 4). The PRC subtype retained a consistent direction of effect after CMS adjustment (HR=0.137, 95% CI: 0.015–1.28, p = 0.081), though statistical significance was attenuated, most likely due to the limited number of death events (n = 13) and restriction of the analysis to patients with available RNA sequencing data, yielding an EPV of approximately 1.4 with 9 covariates—substantially below the recommended threshold of 10. The alluvial plot further showed that neither MYC nor PRC mapped exclusively onto a single CMS class, with both subtypes represented across all four CMS categories, although their distribution differed significantly by CMS class (Chi-square p < 0.001). These findings indicate that the two classification systems are statistically associated but not redundant, supporting the complementary nature of MYC/PRC classification relative to CMS boundaries. Formal demonstration of CMS-independent prognostic value will require validation in a larger cohort with sufficient event numbers. Integrative correlation analysis linking subtype-specific genes with E2F/G2M pathway activity and MEK/EGFR inhibitor response further supports the biological coherence of this framework. Several limitations should be acknowledged. PDOs do not fully recapitulate the tumor microenvironment, and the limited number of discovery samples constrains captured diversity. The concordance-based selection of 25 PDOs from the original 41 excludes approximately 40% of the established biobank; however, comparison of clinical and genomic features between included and excluded samples revealed no statistically detectable differences in any assessed variable, although the modest sample sizes preclude exclusion of small differences (Supplementary Table 7 and Supplementary Fig. 1). Accordingly, these findings are associative and hypothesis-generating pending prospective validation. The heterogeneous drug responses observed in PRC-type PDOs may reflect a relatively quiescent or dormant transcriptional state, consistent with differentiation-associated gene expression programs enriched in this subtype. This biological state may lack a dominant oncogenic dependency, suggesting that PRC-type tumors may require alternative therapeutic strategies beyond MEK/EGFR pathway inhibition. Future studies with broader drug panels may identify subtype-specific vulnerabilities in PRC-type CRC. Regarding the potential confounding effect of basal proliferation rates on drug sensitivity interpretation, we note that drug sensitivity in this study was quantified as AUC derived from dose-response curves normalized to DMSO controls maintained in parallel under identical culture conditions for each individual PDO. Because DMSO control wells were maintained in parallel under identical culture conditions for each PDO, normalization to vehicle-treated controls partially accounts for differences in baseline proliferative activity across organoids. Therefore, the derived AUC values are expected to better reflect drug-specific effects under matched experimental conditions, although formal adjustment for doubling time could not be performed due to lack of systematic doubling-time measurements for all PDOs. Furthermore, the subtype-specific nature of the observed correlations—significant within MYC-type PDOs (R=−0.71, n = 15, p = 0.003 for trametinib; R=−0.70, n = 15, p = 0.004 for afatinib) but absent in PRC-type PDOs(R = 0.05, n = 10, p = 0.886) — argues against a simple proliferation-driven explanation. Fisher’s r-to-z transformation confirmed that the MYC and PRC correlations differed significantly (z=−1.97, p = 0.049), supporting a transcriptionally driven rather than proliferation-driven mechanism. Although this p-value is borderline, the magnitude of the correlation difference (ΔR=0.76 between MYC and PRC subtypes) and the biological coherence of the subtype-specific dissociation provide additional support for this interpretation. Because the zero-centered classification rule reflects the intrinsic symmetry of the SMI scale, the dichotomization was not derived from outcome optimization but from the sign of a continuous biological metric. Finally, the 41 CRC PDOs established in this study are available through the Korean Organoid Biobank to facilitate independent validation and future translational research integrating transcriptome-based stratification with therapeutic decision-making.

Data availability

Raw sequencing datasets in this study have been deposited in the Korea BioData Station (K-BDS, https://kbds.re.kr) under accession KAP241792 (direct link: https://kbds.re.kr/acc/KAP241792). Drug screening data are available from the corresponding author upon reasonable request. TCGA data used for validation are available via the GDC portal. Patient-derived organoids are available through the Korean Organoid Biobank (https://organoid.snu.ac.kr) upon request and in accordance with institutional and biobank policies. Analysis code is available at https://github.com/youtruth/MYC-PRCclassification and contains the core analysis workflows and scripts used in this study.

Additional information

Ethics declarations

Ethics approval and consent to participate

This study was approved by the Institutional Review Board of Seoul National University Hospital (IRB No 1102–098–357).

Funding

This research was funded and supported by the Korean Cell Line Research Foundation, and National Research Foundation of Korea (NRF) grant funded by the Korean government (Grant no RS-2020-NR119695, RS-2022-NR070833).

CRediT authorship contribution statement

You Jin Lee: Writing – review & editing, Writing – original draft, Visualization, Validation, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Jae Hyun Park: Investigation. Hee Ju Nam: Investigation. Soon-Chan Kim: Investigation, Formal analysis. Min Jung Kim: Resources. Seung-Yong Jeong: Resources, Project administration. Ji Won Park: Supervision, Resources, Project administration. Ja-Lok Ku: Supervision, Resources, Project administration.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.