Organoid morphology-guided classification for oral cancer reveals prognosis

Summary

Oral cancer is an aggressive malignancy with a survival rate below 50% in advanced stages due to low mutation rates, lack of molecular subtypes, and limited treatment targets. This study presents a pioneering approach to classifying oral cancer subtypes based on the morphology of patient-derived organoids (PDOs) and proposes a therapeutic strategy. We establish 76 cancer and 81 normal PDOs. For cancer PDOs, both manual classification and AI-based scoring are utilized to categorize them into three distinct subtypes: normal-like, dense, and grape-like. These subtypes correlate with unique transcriptomic profiles, genetic mutations, and clinical outcomes, with patients harboring dense and grape-like organoids exhibiting poorer prognoses. Furthermore, drug response assessments of 14 single agents and cisplatin combination therapies identify a synergistic treatment approach for resistant subtypes. This study highlights the potential of integrating morphology-based classification with genomic and transcriptomic analyses to refine oral cancer subtyping and develop effective treatment strategies.

Graphical abstract

Highlights

• •Development of an extensive patient-derived organoid repository for oral cancer research
• •Organoid morphology-driven refined classification and prognosis of oral cancer
• •Correlation of morphology-based subtype with genomic and transcriptomic profiling
• •Exploration of therapeutic strategies using oral cancer organoids

Lee et al. propose a therapeutic strategy for oral cancer by classifying patient-derived organoids (PDOs) into three morphological subtypes: normal-like, dense, and grape-like. Their findings reveal significant differences in genetic mutations and tumor mutation burden, paving the way for more tailored treatment strategies based on organoid classification.

Introduction

Oral cancer originates in the lip, gum, tongue, mouth, and palate, with oral squamous cell carcinoma (OSCC) being the most prevalent subtype. OSCC is characterized by a high rate of local recurrence, and, in advanced stages, the 5-year survival rate is a mere 23%–30%.^1,2 Despite its aggressive nature, treatment options for OSCC patients remain limited.³ Although human papillomaviruses (HPVs) are an important cause of oropharyngeal squamous cell carcinoma and a predictor of radiotherapy response, their role in OSCC remains unclear.⁴ Currently, the tumor, node, metastasis (TNM) staging system, which assesses tumor size and lymph node metastasis, is the sole marker used to determine prognosis and treatment strategies in oral cancer. No molecular markers are available to capture the biological diversity among patients.⁵ This unique characteristic, combined with the low incidence and heterogeneous mutational landscape of oral cancer, poses significant challenges in identifying therapeutic targets or prognostic markers. Because of the low mutation rate and high tumor heterogeneity, identifying optimal targeted drugs, driver mutations, and molecular classifications for oral cancer has proven difficult, even through extensive multi-omics analysis of patient cohorts.^6,7 Consequently, treatment strategies heavily rely on surgical resection, and patients at high risk of recurrence have limited effective options, primarily radiotherapy or combined chemotherapy and radiation therapy.⁸ Notably, oral cancer is a rare and intractable disease with an incidence of approximately 6 cases per 100,000 people in the Republic of Korea.² This limited number of cases hampers the ability to screen various anticancer drugs through large-scale clinical trials, resulting in a small pool of models for preclinical research.

From this perspective, patient-derived organoids (PDOs) serve as an “all-in-one” model suitable for molecular characterization and target validation, which can overcome the limitations of oral cancer research.⁹ PDOs recapitulate tissue architecture, genetic profiles, and therapeutic responses observed in solid tumors such as colon, pancreatic, gastric, biliary tract, bladder, breast, ovarian, and head and neck cancers.^{10,11,12,13,14,15,16,17,18,19} Researchers have utilized organoids for disease modeling, studying tumor mechanisms, drug screening, testing cancer immunotherapies, and developing personalized medicine approaches.²⁰ A distinguishing feature of organoids is their ability to develop diverse morphologies under three-dimensional culture conditions, reflecting the integration of genetic mutations and transcriptional characteristics.^{21,22,23,24,25,26} Notably, morphological transformations have been observed through genetic manipulation of genes such as TP53, APC, and CDH1 using CRISPR-Cas9 technology in gastric, pancreatic, and colon cancer organoids.^12,27

In this study, we hypothesized that it would be possible to classify molecular subtypes of oral cancers and suggest treatment strategies from the classification of organoid morphology. We established the largest oral cancer research platform, including early- and late-stage oral cancer organoids. Furthermore, we propose three morphological subtypes of oral cancer. Our findings suggest that these subtypes correlate not only with mutational characteristics but also with distinct transcriptional profiles and biological functions, potentially paving the way for innovative anticancer strategies. Additionally, these subtypes are indicative of recurrence-free survival, providing insights into patient prognosis. We have also evaluated the sensitivity of these organoids to a combination of cytotoxic agents and targeted drugs, previously untested in oral cancer, to delineate the most effective treatments for each morphological subtype. Particularly for the dense subtypes associated with poor prognosis, a combination therapy with a synergistic effect was identified through biological pathway analysis. Finally, we predict that this perspective on organoid morphology will open possibilities for treatment strategies in oral cancer.

Results

Establishment of a PDO library for oral cancer

For human oral normal and cancer organoid library, we established organoids based on protocols described for sigmoid organoids in previous studies, with modifications to the culture conditions and growth media for optimization.^17,18 A typical feature of normal oral organoids is keratinization starting from the seventh day of growth at the organoid center. However, this keratinized zone was not entirely dissociable by trypsin and remained solid during subsequent passages, posing a challenge for long-term culture (Figures 1A and S1). To address this issue, we strictly controlled the subculture period within 7–10 days and optimized the culture medium to boost the stability of our banking platform. Essential niche elements were identified by comparing organoid morphology and viability upon the addition of components like Noggin, WNT-3a, CHIR99021, N-2 supplement, or insulin-transferrin-selenium (ITS) to a minimal medium base containing R-spondin 1, epidermal growth factor (EGF), fibroblast growth factor 10 (FGF10), and forskolin. The growth-promoting effects of Noggin and WNT-3a contrasted with the inhibitory impact of the Glycogen Synthase Kinase 3 (GSK-3) inhibitor CHIR99021 at concentrations exceeding 1 μM, guiding our choice of optimal culture conditions (Figures 1B–1D).

Figure 1.

Generation of oral cancer organoids from patient tumors and tumor-adjacent non-malignant tissues

(A) Summary of the modified organoid subculture process.

(B and C) Optimization of organoid growth media. The basic conditions in which R-spondin 1, EGF, and FGF10 were added to advanced DMEM F/12 medium containing B-27 complement, HEPES, glutamine, and N-acetyl cysteine, termed the basal medium. 1% Noggin conditioned media (CM), 100 ng/mL recombinant Wnt-3a (rWNT), 3 mM CHIR99021, 1x N-2 supplement, or ITS were added to basal media. After culturing for 7 days, organoid images were taken at 5× and 20× magnifications, and the ATP assay was performed using CellTiter-Glo (3D) reagent to compare cell viability relative to the basal medium. Scale Bar, 100 μm. Each experiment was performed in biological triplicates. Statistical significance levels: ∗p < 0.05, ∗∗p < 0.01, and ∗∗∗p < 0.001.

(D) Representative images of organoids treated in a concentration-dependent manner for CHIR99021. Scale Bar, 50 μm.

(E) H&E and IHC staining of tumor tissue and matching PDO. Scale Bar, 50 μm. H&E, hematoxylin and eosin; IHC, immunohistochemistry; PDO, patient-derived organoid.

See also Figure S1 and Tables S1, S2, and S3.

Under modified culture conditions, we observed histological characteristics in both normal and cancer organoids. Normal organoids were established with well-separated basal and prickle layers. Cells positive for Pan-CK (Pan-Cytokeratin; an epithelial cell marker) were located internally, whereas cells positive for P63 (basal cell marker) were located in the external layer, which was opposite to the actual normal tissue. In the case of tumor, organoids were representative of the patients’ tumor tissue, not only in H&E staining but also in Pan-CK, P63, Ki-67, and P53 staining (Figure 1E). As a result, we established 76 cancer and 81 normal organoids from 180 and 186 patients, respectively, excluding those with fibroblast overgrowth, low cellularity, or bacterial contamination. These organoids were derived from primary and metastatic OSCC, mucoepidermoid carcinoma, melanoma, and rare types like clear-cell odontogenic carcinoma. Normal organoids had similar establishment success rates regardless of their origin (Table S2). The stably established oral cancer organoids included tissue from a patient diagnosed with not only OSCC but also the less frequently diagnosed types of mucoepidermoid carcinoma, adenocarcinoma, malignant melanoma, clear-cell carcinoma, primary intraosseous carcinoma, large cell neuroendocrine carcinoma, peripheral ameloblastoma, and sarcomatoid carcinoma (Figure 2A). The majority of these organoids maintained histological similarities with the patient’s tissue, with higher success rates observed in samples from poorly differentiated or extranodal extension (ENE)-positive tissues (Figure S1B and Table S3). These cancer organoids successfully recovered even after cryopreservation and grew to the point where they could be passaged within 7–10 days on average. Collectively, we successfully established the oral normal and cancer organoid platform that faithfully recapitulates the histological features of the original tumors, providing a reliable preclinical model.

Figure 2.

Classification and characterization of organoid morphology

(A) Composition of the oral cancer organoid library. Pie chart indicates the histological diagnosis of the tumor organoids.

(B) Representative organoid morphology for each subtype. The illustration summarizes the appearance and cross-sectional characteristics of each morphology. The white outlined box shows the distinguishing features at the edges of normal-like and dense organoids at high magnification. White arrows point to daughter organoids that have escaped from the main organoid body.

(C) Cell density based on organoid morphology. Cell density was calculated by counting the number of hematoxylin-stained nuclei within a circle with a radius of 50 μm from the center of the organoid section in the H&E image.

(D) A comparison of growth rates based on organoid morphology. Growth rates were calculated as mean values between passages 4 and 8 using the number of cells harvested per incubation period for each passage.

(E) Kaplan-Meier curves of 2 years recurrence-free survival according to organoid morphology.

(F) Circularity and solidity score based on organoid morphology. A circularity value of 1.0 indicates a perfect circle. A solidity value of 1.0 indicates that the shape is completely filled, resembling the reference sphere. Each score was calculated by selecting one from each of the 68 OSCC organoids Expanded ranges of normal-like and dense scores are shown on the right. Comparing solidity or circularity as a single factor (gray dotted line) results in overlapping range, whereas considering both factors together (red dotted line) can be clearly distinguished.

(G) The morphology score was calculated by randomly selecting 2 to 5 images for each of the 68 OSCC organoids, for a total of 336 images. The score was defined as 0–10.

Statistical significance levels: ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, and ∗∗∗∗p < 0.001. See also Figure S2 and Table S4.

Morphology-guided classification and prognosis of oral cancer organoids

Once the organoid proliferation reached a stable phase, we found that cancer organoids appeared in three distinct morphologies during growth: normal-like, dense, and grape-like. These morphologies were clearly distinguished by observing differential interference contrast images of organoids. Normal-like and dense both appear to be round, but normal-like is closer to a sphere, and dense is more of a distorted ellipse. The critical difference between these two subtypes is shown in the enlarged images shown in the white box of Figure 2B. Normal-like has a clear band at the border of the circle and a smooth edge. On the other hand, dense has no distinct border band and has jagged edges. Grape-like organoids showed an irregular shape, completely losing their roundness. Some highly invasive cells escaped from these organoids, invaded the basement membrane extract, and proliferated into multiple daughter organoids. Their strong invasiveness was evident, as indicated by the white arrows in Figure 2B.

Further detailed histological analysis and exclusive focus on OSCC organoids allowed us to classify 68 samples into the three aforementioned morphological subtypes: 32 normal-like, 24 dense, and 12 grape-like, based on visual inspection under light microscopy (Figures 2A and S2). Moreover, morphological characteristics were further detailed through histological analysis, providing a deeper understanding of the organoid structures. The cell density within the organoids varied according to their morphology. Normal-like organoids contained fewer nuclei compared to other organoid types. The relatively bright center of normal-like organoids in light microscope images, despite their larger size, was attributed to the high concentration of cells at the edges, with a lower cell density in the center (Figure 2C). This morphological variation significantly impacted on the growth rates of the organoids, with dense and grape-like types proliferating faster than their normal-like counterparts, correlating with their respective Ki-67 staining patterns where proliferation markers were more abundantly expressed in the dense and grape-like types (Figures 1E and 2D).

Importantly, organoid morphology was strongly correlated with prognosis. Although organoid morphology was not correlated with clinical factors associated with tumor progression, such as TNM stage or differentiation grade, patients with dense or grape-like organoids had a higher risk of recurrence and significantly lower recurrence-free survival compared to patients with normal-shaped organoids. (Figure 2E and Table 1). These findings suggested the potential of using organoid morphology as a prognostic tool in oral cancer, emphasizing its role in overcoming previously unclear subtype classifications and guiding treatment strategies.

Table 1.

Clinical characteristics by organoid morphologies

		Normal-like		Dense		Grape-like		p value
Characteristics		N	%	N	%	N	%
Age
	Mean ± SD	61.0	(±13.3)	62.3	(±13.3)	59.3	(±13.3)	–
	20–39	3	(60%)	1	(20%)	1	(20%)	0.75
	40∼	29	(46%)	23	(37%)	11	(17%)
Gender
	Male	18	(43%)	16	(38%)	8	(19%)	0.68
	Female	14	(54%)	8	(31%)	4	(15%)
Smoking
	Yes	16	(47%)	12	(35%)	6	(18%)	>0.99
	No	16	(47%)	12	(35%)	6	(18%)
T stagea
	T(1,2)	13	(48%)	9	(33%)	5	(19%)	0.96
	T(3,4)	19	(46%)	15	(37%)	7	(17%)
N stagea
	N(0)	17	(47%)	13	(36%)	6	(17%)	0.97
	N(1,2,3)	15	(47%)	11	(34%)	6	(19%)
TNM stagea
	Stage I, II	9	(47%)	6	(32%)	4	(21%)	0.87
	Stage III, IV	23	(47%)	18	(37%)	8	(16%)
Differentiation
	Well	10	(50%)	6	(30%)	4	(20%)	0.23
	Moderate	15	(45%)	15	(45%)	3	(9%)
	Poorly	6	(50%)	2	(17%)	4	(33%)
	Unknown	1	(33%)	1	(33%)	1	(33%)	–
Extranodal extension (ENE)
	Yes	7	(37%)	6	(32%)	6	(32%)	0.17
	No	25	(51%)	18	(37%)	6	(12%)
Recurrence
	Yes	4	(22%)	9	(50%)	5	(28%)	0.04
	No	28	(56%)	15	(30%)	7	(14%)

^aAJCC stage.

Morphology scoring for morphology-guided organoid classification

As described earlier, organoid morphology predicts both clinical prognosis and mutational landscape, establishing it as a robust and reliable classification method. Expert classification readily distinguishes the three morphological subtypes (normal-like, dense, and grape-like) under the microscope. The morphological subtype identification scheme is summarized in Figure S2B. In brief, normal-like organoids exhibit a near-perfect circular shape with low cell density at the center, while dense organoids are slightly oval, with high central cell density. The grape-like subtype is characterized by its irregular shape, making it easily recognizable. To capture this expertise and quantify morphological differences, an image-based scoring algorithm was developed. Shape-related parameters were prioritized, as they remain consistent despite segmentation variability and image quality differences. Using a training dataset of 68 organoid images, the final morphology score model was optimized to classify organoid subtypes based on two shape-related parameters: circularity and solidity (Figure 2F). The resulting morphology score function is

$$Morphologyscore=4.221\times circularity+5.779\times solidity$$

with threshold values of 9.50 to distinguish normal-like and dense subtypes and 9.03 to differentiate dense and grape-like subtypes.

To validate the accuracy of the morphology score, an additional 336 OSCC organoid images from 68 organoids were analyzed. The morphology scores were compared with expert classification, yielding an overall accuracy of 0.82 (Figures 2G and S2C). The model demonstrated the highest accuracy in classifying the grape-like subtype, whereas the dense subtype was more challenging to classify. This was expected, as the model primarily captures organoid shape rather than cell density, which is a key distinguishing feature between normal-like and dense organoids.

Genomic characterization of organoids by morphological subtype

For further biological characterization, we performed whole-exome sequencing on our classified OSCC organoids, revealing an average of 73.5 mutations per organoid and a total of 4,075 somatic variants across 3,122 genes. These variants comprised 3,563 missense mutations, 185 nonsense mutations, 103 frameshift deletions, 59 frameshift insertions, 50 in-frame deletions, 19 in-frame insertions, and 15 non-stop mutations. The most common variant type was SNPs, and the most common SNV class was the C-to-T variant (Figure S3A). Importantly, TP53 mutations emerged as the most prevalent, detected in 48% of organoids, followed by those in TTN (30%), NPF1 (22%), CASP8 (20%), FAT1 (20%), HOXA3 (20%), MUC16 (20%), and NOTCH1 (20%). The median tumor mutation burden (TMB) was 4.16 mutations/Mb (range, 0.48–40.33) (Figure S3B). We summarized the variation in cancer-related genes associated with oral cancer in an oncoplot sorted by morphological subtype (Figures 3A and S3C). Normal-like organoids detected relatively low mutations in tumor suppressor genes (TSGs) or oncogenes. TP53 mutations were detected at a high frequency (87%) in both dense (10/13) and grape-like (9/11) organoids, whereas they were not detected in normal-like organoids (0/19). With the accumulation of NOTCH1, CASP8, or FAT1 mutations, a greater number of organoids lost their original shape, exhibited condensed nuclei, or became invasive. TMB was significantly increased in dense and grape-like organoids. Through mutually exclusive or simultaneous genetic analyses using somatic interaction functions, it was observed that the grape-like organoid exhibited concurrent somatic interactions involving CASP8 and KMT2D, whereas variations in TP53 and KRT7 were mutually exclusive interactions in the dense organoid (Figure S3D). As commonly observed, the C>G mutation was the most prevalent SNV class in whole oral cancer organoids. Specifically, the T>G variant was increased in normal-like organoids, while the T>G variant was increased in grape-like organoids (Figures 3B–3D). These results connect the morphological subtype with characteristic mutation profiles, which may influence organoid morphology and may provide guidance for oral cancer molecular subtyping.

Figure 3.

Genomic characterization by organoid morphology

(A) Oncoplot displaying the somatic landscape of oral cancer organoid. Genes are arranged by their mutation frequency. TMB is shown as a bar plot (top). Stacked bar plot (bottom) shows distribution of mutation spectra for every sample in the mutatin annotation format (MAF) file. SNV class of organoid with normal-like (B), dense (C), or grape-like (D) morphology. Statistical significance levels: ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, and ∗∗∗∗p < 0.0001. TMB, tumor mutation burden. See also Figure S3.

Transcriptomic characterization of organoids by morphological subtype

To characterize the transcriptional landscapes of OSCC organoid, we performed bulk RNA sequencing (RNA-seq). Principal-component analysis of global gene expression profiles revealed distinct clustering of the three morphological subtypes, indicating substantial transcriptional divergence (Figure 4A). Leveraging these transcriptional differences, we classified The Cancer Genome Atlas (TCGA) OSCC patients to assess the clinical relevance of this classification (Figure 4B). To identify functional differences among the subtypes, we performed single-sample gene set enrichment analysis (ssGSEA) on organoids using Gene Ontology: Biological Process (GOBP) pathways. A total of 43 pathways were identified as significantly enriched across the subtypes (Figure S4). Heatmap analysis revealed a strong anti-correlation between the normal-like and grape-like subtypes, whereas the dense subtype consistently exhibited an intermediate enrichment profile (Figure 4C). Only one GOBP pathway, associated with the positive regulation of receptor binding, exhibited maximal enrichment in the dense subtype.

Figure 4.

Transcriptional characteristics of OSCC organoids and application of morphological subtypes to TCGA

(A) Dimension-reduction plot of 40 organoids, showing that each morphological subtype is well clustered.

(B) Schematic representation of the morphological subtype classification process for TCGA patients, based on morphology-specific transcriptional features (see STAR Methods for details).

(C) Heatmap of 43 Gene Ontology: Biological Process (GOBP) pathways showing significant subtype-specific enrichment (left), along with three representative pathways per subtype (right). Normal-like and grape-like subtypes exhibit multiple enriched pathways, while the dense subtype shows an intermediate enrichment profile. Pairwise statistical significance was assessed using the Kruskal-Wallis test.

(D) Autocorrelation heatmap (left) and UMAP representation of morphology-specific features (right) of TCGA patients, filtered based on statistically significant correlations with morphological subtypes. Three distinct clusters are evident in both the heatmap and UMAP representation.

(E) Kaplan-Meier plot of 2-year overall survival analysis of TCGA patients stratified by morphological subtype. The normal-like subtype demonstrates the most favorable prognosis, followed by the dense and grape-like subtypes.

Statistical significance levels: ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, and ∗∗∗∗p < 0.0001. See also Figures S4 and S5.

To determine whether the transcriptional profiles specific to organoid morphology are reflected in patient tumors, we applied ssGSEA to bulk RNA-seq data from TCGA OSCC patients (N = 338) using pathways enriched in each morphological subtype. An autocorrelation-based clustering approach was applied to identify functionally distinct OSCC patient subclusters. Pearson correlation analysis was then performed to compare each subcluster’s mean pathway profile with those of organoid-derived subtypes, determining their closest morphological subtypes (Figure S5, see STAR Methods for details). Ultimately, 184 patients were assigned to morphological subtypes; including 57 normal-like, 50 dense, and 77 grape-like cases. Patients within the same subtype exhibited strong self-similarity, as demonstrated by autocorrelation analysis and uniform manifold approximation and projection (UMAP) dimension-reduction analysis (Figure 4D). To assess the clinical significance of these subtypes, a two-year survival analysis was performed. The subtypes demonstrated statistically significant prognostic differences (log rank test, p < 0.05). Kaplan-Meier survival analysis revealed that the normal-like subtype was associated with a good prognosis, whereas the dense and grape-like subtypes were associated with significantly poorer prognosis (Figure 4E, hazard ratio = 2.619/2.592, p < 0.05). This trend was consistent with the recurrence-free survival analysis, where patients with normal-like organoids demonstrated the most favorable prognosis (Figure 2E). Collectively, these findings indicate that the morphological subtypes identified through organoid models exhibit distinct transcriptional and pathway-specific features, enabling the classification of OSCC patients in a clinically meaningful manner. The observed associations between organoid morphology and patient outcomes suggest that the proposed classification system has potential applications in precision oncology, highlighting the relevance of organoid-based morphological characterization in cancer research.

Exploring potential anticancer drug using oral cancer organoids

The development of targeted therapies for oral cancer remains challenging due to the limited molecular targets and models that effectively bridge clinical and preclinical research. In this regard, organoid models have the potential to serve as a useful platform, given their ability to reflect patient tissues with high concordance and their suitability for multi-omics analysis. To evaluate the therapeutic potential of inhibitors targeting dysregulated signaling pathways in oral cancer, we assessed the efficacy of 14 compounds in organoid models. OSCC arises from genetic alterations, epigenetic modifications, and dysregulation of the tumor microenvironment, driven by chronic exposure to tobacco use, alcohol consumption, malnutrition, and immune suppression.²⁸ These factors contribute to the aberrant activation of oncogenic pathways, including epidermal growth factor receptor (EGFR), Wnt/β-catenin, Janus kinase/signal transducers and activators of transcription (JAK/STAT), Notch, Phosphoinositide 3-Kinase/protein kinase B/mechanistic target of rapamycin (PI3K/AKT/mTOR), and rat sarcoma virus oncogene/rapidly accelerated fibrosarcoma/mitogen-activated protein kinase (RAS/RAF/MAPK), while simultaneously disrupting tumor-suppressive pathways such as TP53/RB and p16/Cyclin D1/Rb, thereby promoting OSCC progression.²⁹ To systematically target these dysregulated pathways and explore their therapeutic relevance, we selected 10 targeted inhibitors based on their established involvement in oral cancer and prior validation in preclinical or clinical studies: Tipifarnib (HRAS inhibitor), EPZ015666 (PRMT5 inhibitor), AZD4547 (FGFR inhibitor), niraparib (PARP inhibitor), alpelisib (PI3K inhibitor), sapitinib (EGFR inhibitor), AZD7762 (CHK1/2 inhibitor), everolimus (mTOR inhibitor), BAY1895344 (ATR inhibitor), and Nutlin-3a (MDM2 inhibitor). Additionally, to provide a comprehensive comparison with standard chemotherapy, we included four widely used first-line chemotherapeutic agents—cisplatin, 5-FU, docetaxel, and paclitaxel—to evaluate their efficacy alongside targeted therapies within the organoid models (Figure S6A). Each organoid’s response was quantified using area under the curve (AUC) measurements, and a standardized Z score was applied to categorize responses, revealing varied sensitivities that aligned with the genetic and morphological characteristics of the organoids (Figures S6B and S6C). The accuracy of our organoid drug evaluation system was validated by comparing the response to targeted drugs and genetic variation. Three PDOs with HRAS mutations belonged to the moderate responder or responder group for tipifarnib. Most organoids with TP53 mutations were resistant to Nutlin-3a (an MDM2 inhibitor) (Figure S6D). Although the number of cases available for comparing drug response with clinical prognosis in organoids was limited, we found that some drugs exhibited subtype-specific growth-inhibitory effects. Consistent with the higher frequency of TP53 mutations observed in dense and grape-like organoids, Nutlin-3a was selectively effective against normal-like organoids. Grape-like organoids exhibited a drug response profile similar to that of normal-like organoids. In contrast, dense organoids demonstrated resistance to most targeted therapies, including EPZ015666, AZD4547, niraparib, and everolimus, as well as mild resistance to other drugs such as alpelisib, sapitinib, and BAY1895344 (Figures 5A and S6E).

Figure 5.

Exploring targeted drugs for combination treatment using an organoid model

(A) Comparison of drug response by organoid morphology.

(B) Comparison of synergy scores for cisplatin by target drug. The Bliss score was calculated using SynergyFinder to evaluate the synergistic effects of BAY1895344, AZD7762, EPZ015666, or everolimus in combination with cisplatin across 27 OSCC organoids, comprising 10 normal-like, 8 dense, and 9 grape-like subtypes.

(C) Representative images of Bliss score according to the combination treatment. Normal-like: OC122T-O, dense: SOC34T-O, and grape-like: SOC22T-O.

(D) Mean value of Bliss score by morphology. Each experiment was performed in biological quadruplicates.

Statistical significance levels: ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, and ∗∗∗∗p < 0.0001 See also Figures S6 and S7 and Table S5.

Considering these results, the high drug tolerance of the dense subtype led us to explore combination treatment strategies, whereas the grape-like subtype, despite its worse prognosis compared to normal-like organoids, exhibited sensitivity to single-drug treatment. Given the standard use of cisplatin in oral cancer treatment, we aimed to identify drugs that exhibit synergistic effects when combined with cisplatin. Based on commonly activated pathways in oral cancer, we selected four candidate drugs. These drugs have been previously evaluated in clinical trials for other cancers and were chosen for their potential efficacy in enhancing cisplatin sensitivity. First, we hypothesized that targeting the DDR (DNA damage response) pathway would induce excessive DNA damage, leading to increased cell death triggered by cisplatin. To investigate potential combination therapies for advanced oral cancer, we examined BAY1895344 and AZD7762. EPZ015666 is a selective inhibitor of EZH2 that interferes with the function of the PRC2 complex by blocking the formation of H3K27me3. By inhibiting EZH2, we expected to interfere with the DNA damage repair process and thereby increase the sensitivity of cancer cells to cisplatin-induced apoptosis, especially in response to cisplatin-induced DNA double-strand breaks.³⁰ And, the synergistic effect of the combination treatment of cisplatin and everolimus has been previously documented in preclinical studies.³¹ Co-treatment of OSCC organoids with BAY1895344 and cisplatin demonstrated strong synergism across all subtypes, with Bliss scores ranging from 19.57 to 50.94. The optimal synergy was observed at a combination of 1 μM cisplatin and 0.01 μM BAY1895344 (Figures 5B–5D, S7A, and S7B). According to RNA-seq data, dense organoids exhibited an upregulation of the E2F targets gene set and the G2M checkpoint gene set, which are associated with cell-cycle regulation. The observed enhancement in cisplatin and BAY1895344 combination therapy efficacy may be attributed to these molecular features. BAY1895344 disrupts DDR signaling, exacerbating DNA damage induced by cisplatin. The elevated expression of E2F targets and G2M checkpoint genes suggests a heightened dependency on DDR mechanisms, making these organoids more susceptible to ATR inhibition and subsequent apoptosis (Figure S7C). Our results demonstrate that PDOs serve as a suitable model for evaluating treatment response, particularly in low-incidence cancers such as oral cancer.

This study demonstrates the potential of oral cancer organoid models in analyzing drug responses based on morphological subtypes and identifying targeted therapies. We found that dense-subtype organoids exhibit general resistance to common anticancer treatments, which can be overcome by combination therapy with cisplatin and BAY1895344, demonstrating previously unreported synergistic effects. Our findings suggest that organoid-based drug screening can serve as a valuable preclinical tool for OSCC, offering insights into personalized treatment strategies.

Discussion

In this study, we demonstrated the potential of organoid morphology-based classification in oral cancer, revealing its strong correlation with molecular and clinical characteristics. While organoid morphology has traditionally been used to assess histological similarities with corresponding tumor tissues,²² our findings highlight its broader significance in reflecting the underlying molecular biology of cancer. This was achieved through the establishment of one of the largest organoid libraries dedicated to oral cancer, encompassing over 70 cancers and organoids derived from patients with HPV-negative oral cancer. Previously, only dense morphology was reported.^17,18 Here, we identified three distinct morphological subtypes, systematically classified organoids based on morphology, and integrated multi-omics and clinical data within the same cohort. Moreover, we developed a morphological score equation, defined by morphological roundness and robustness metrics. This model had the ability to classify three subtypes by objectively determining morphological subtypes. Above all, the morphology-based subtypes showed statistically significant associations with the clinical prognosis of the patients from whom they were derived. Notably, patients with dense or grape-like organoids had poor prognosis, reinforcing the prognostic relevance of organoid morphology in oral cancer and highlighting its potential utility in patient stratification. This is consistent with findings in other cancers, where grape-like morphologies are found in organoids derived from metastatic cancer tissue, ascites, or pleural fluid.³² Such observations support the clinical utility of morphology-based classification in predicting disease progression.

A key finding of our study is that TP53 mutation serves as a primary determinant distinguishing normal-like from dense and grape-like morphologies. The loss of TP53 function led to nuclear atypia and hyperproliferation, disrupting the basal layer, a characteristic feature of normal-like organoids, and subsequently giving rise to dense and grape-like morphologies. This observation aligns with previous reports in pancreatic and colorectal organoids, where CRISPR-Cas9-mediated driver gene mutations (e.g., KRAS, TP53, CDKN2A, and SMAD4) induced severe nuclear atypia, structural deformation, and tumor budding phenotypes.^23,24 Moreover, the morphological classification of organoids was also applicable to tumor tissues. To further validate the OSCC morphological subtypes, we applied biological pathway-based label transfer to publicly available datasets. This confirmed that each subtype was associated with consistent clinical characteristics. Distinct transcriptional and functional profiles of the subtypes were identified across 43 GOBP pathways. These functional characteristics were also reflected in the bulk transcriptomics of patient tumors, which were then used for patient subtyping. Notably, patients with the normal-like OSCC subtype had the best overall survival prognosis compared to the dense and grape-like subtypes, aligning with the progression-free survival analysis from our cohort.

Our findings also emphasize the potential of organoid models in personalized therapy, particularly highlighted by the ongoing organoid-guided clinical trials across various cancer types.³³ Within our study, we explored personalized treatment strategies by assessing the response of organoid models to 14 therapeutic agents, including a combination of four targeted therapies with cisplatin In particular, BAY1895344 showed strong synergism with cisplatin, especially in dense organoids that typically exhibit higher resistance to anticancer treatments. The DDR is one of the key resistance mechanisms to cisplatin and is considered a promising target in cancer therapy.³⁴ When DNA is damaged by cisplatin, the ATR-CHK1 signaling pathway is activated, halting cell-cycle progression at the G2/M checkpoint to allow sufficient time for DNA repair.³⁵ In this context, DDR inhibitors can effectively induce apoptosis in cancer cells through a synthetic lethality mechanism.³⁶ Our findings indicate that dense organoids exhibit increased expression of E2F target genes and G2/M checkpoint genes compared to normal-like organoids. While further validation is needed, these results suggest that the enhanced efficacy of ATR inhibitors in combination with cisplatin may be particularly pronounced in dense organoids. Hence, the inhibition of the ATR pathway can be considered an attractive combination treatment strategy for patients with poor prognosis and dense organoids resistant to targeted anticancer drug treatment alone.

Collectively, we established a large-scale organoid library specializing in oral cancer and proposed an oral cancer classification system guided by organoid morphology that has bridged the gap between cancer genotypes and phenotypes, significantly enhancing the translational pathway from preclinical studies to clinical applications. These advances provide insights into morphology-based therapeutic strategies, offering promising directions for personalized cancer treatment.

Limitations of the study

While our study provides valuable insights into the morphological subtyping of oral cancer organoids, several challenges must be addressed to enhance their clinical applicability. First, oral cancer exhibits a high degree of mutational heterogeneity and lacks major driver mutations, except for TP53. Therefore, a larger cohort is required to accumulate and validate data for statistical significance, ensuring a clearer distinction between the clinical and genetic characteristics of the three organoid subtypes. Expanding the study sample size will also strengthen the robustness of our findings. Additionally, using PDOs to further validate the combined effects of cisplatin and BAY1895344 in more cases may offer therapeutic insights into targeting DDR through ATR inhibition and cisplatin combination therapy. Second, while we demonstrated that organoid morphological subtypes correlate with distinct mutational and transcriptomic profiles, additional validation is necessary to identify the direct molecular mechanisms underlying these morphological variations. Furthermore, as our experiments were conducted using tumor-cell-enriched organoid models, future studies should incorporate tumor microenvironment components to better validate our findings in a more physiologically relevant context. By addressing these challenges, future research can further refine organoid-based subtyping and advance precision medicine approaches for oral cancer treatment.

Resource availability

Lead contact

Requests for further information and resources and reagents should be directed to the lead contact, Prof. Yun-Hee Kim (sensia37@ncc.re.kr).

Materials availability

Further requests for PDOs generated in this study should be directed to the lead contact, Prof. Yun-Hee Kim (sensia37@ncc.re.kr).

Data and code availability

• •The data reported in this paper will be shared by the lead contact upon request. The RNA-seq dataset generated during this study is available at GEO: GSE293370. The exome sequencing dataset generated during this study is also available at GEO: GSE293371.
• •This paper does not report original code.
• •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

Acknowledgments

This study was supported by grants from the NCC (grant number: NCC-2210980 and NCC-2410780) and the National Research Foundation of Korea funded by the Korean government (MIST) (grant number: 2020M3A9A5036362, 2022R1A2C1012450, and 2023R1A2C1006760).

This work was supported by the Research Core Center in NCC of Korea.

Author contributions

M.R.L. performed investigation and formal analysis and wrote the original draft of the manuscript. S.M.K. contributed to investigation, formal analysis, and validation. Jonghyun Lee contributed to formal analysis and data curation and wrote the original draft of the manuscript. Y.-S.L., H.W.S., and G.K. participated in the investigation. Jiyoung Lee contributed to formal analysis. S.M.Y., D.W.K., J.Y.P., and S.M.K. provided resources. S.-Y.K., W.C., and Jonghyun Lee conceptualized the study. D.S. supervised the study, contributed to methodology and writing – review and editing, and provided resources. I.-J.K. and S.-W.C. supervised the study, provided resources, and contributed to writing – review and editing. Y.-H.K. managed the project, supervised the work, acquired funding, and contributed to writing – review and editing.

Declaration of interests

The authors declare no competing interests.

STAR★Methods

Key resources table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
Anti-pan Cytokeratin antibody [C-11]	Abcam	Cat#ab7753; RRID:AB_306047
CONFIRM anti-Ki-67 (30-9) Rabbit Monoclonal Primary Antibody	VENTANA	Cat#05278384001; RRID:AB_2631262
CONFIRM anti-p53 (DO-7) Primary Antibody	VENTANA	Cat#05278775001; RRID:AB_2892528
VENTANA anti-p63 (4A4) Mouse Monoclonal Primary Antibody	VENTANA	Cat#05867061001; RRID:AB_2335989
Biological samples
Human oral cancer sample	National cancer center, Republic of korea	N/A
Chemicals, peptides, and recombinant proteins
Cultrex Reduced Growth Factor Basement Membrane Extract, Type 2, Pathclear	R&D system	Cat#3533-010-02
Advanced DMEM/F-12	Gibco	Cat#12634-028
B-27™ supplement (50X), serum free	Gibco	Cat#17504-044
N-acetyl-L-cysteine	Sigma	Cat#A9165
Animal-Free Recombinant Human EGF	Peprotech	Cat#AF-100-15
Recombinant Human FGF-10	Peprotech	Cat#100-26
Recombinant human R-spondin 1	Qkine	Cat#006
Noggin-Fc Fusion Protein Conditioned Medium	IPA	Cat#N002
WNT Surrogate-Fc Fusion Protein	IPA	Cat#N001
Nicotinamide	Sigma	Cat#N0636
A83-01	Tocris	Cat#2939
Forskolin	Tocris	Cat#1099
GlutaMAX™ Supplement	Gibco	Cat#35050061
Primocin®	Invivogen	Cat#ant-pm-05
HEPES	Gibco	Cat#15630056
5-fluorouracil	Chemscene	Cat#CS-0993
Cisplatin	Selleckchem	Cat#S-1166
Paclitaxel	Chemscene	Cat#CS-1145
Docetaxel	Selleckchem	Cat#S1148
Tipifarnib	Chemscene	Cat#CS-0475
EPZ015666	Chemscene	Cat#CS-3995
AZD4547	Chemscene	Cat#CS-0971
Niraparib	Chemscene	Cat#CS-0780
Alpelisib	Chemscene	Cat#CS-0663
Nutlin-3a	Chemscene	Cat#CS-0296
Sapitinib	Chemscene	Cat#CS-3951
AZD7762	Chemscene	Cat#CS-0025
Everolimus	Chemscene	Cat#CS-0064
Bay1895344 hydocholoride	Medchemexpress	Cat#HY-101566
Critical commercial assays
CellTiter-Glo® 3D Cell Viability Assay	Promega	Cat# G9681
Tumor Dissociation Kit, human	Miltenyi Biotec	Cat#130-095-929
Deposited data
Raw and analyzed data	This paper	GEO: GSE293370
Raw and analyzed data	This paper	GEO: GSE293371
Software and algorithms
GraphPad prism8.4.3	GraphPad	https://www.graphpad.com
Segment Anything Model (SAM), v1.0	Meta AI	https://github.com/facebookresearch/segment-anything
R version 4.3.1	R Core Team	https://cran.r-project.org/
TCGAbiolinks, v2.30.0	R/Bioconductor	https://bioconductor.org/packages/release/bioc/html/TCGAbiolinks.html
ssGSEA2.0 v2.0	Broad Institute	https://github.com/broadinstitute/ssGSEA2.0
Python version 3.9	Python Software Foundation	https://www.python.org/
scikit-image v.0.19.3	Python Software Foundation	https://scikit-image.org/
Mutect (version 2)	Broad Institute	https://gatk.broadinstitute.org/
SnpEff	Cingolani P et al.37	https://github.com/pcingola/SnpEff
Maftools	Mayakonda et al.38	https://www.bioconductor.org/packages/release/bioc/html/maftools.html
Other
TCGA Head and Neck Squamous Cell Carcinoma (HNSC)	The Cancer Genome Atlas (TCGA)	https://portal.gdc.cancer.gov/
Gene Ontology: Biological Process (GOBP) gene sets	Molecular Signatures Database (MSigDB)	https://www.gsea-msigdb.org/gsea/msigdb

Experimental models and study participant details

Patient tumor specimens and ethics statement

We prospectively enrolled 204 patients with oral cancer, including oral squamous cell carcinoma (OSCC), mucoepidermoid carcinoma, adenocarcinoma, large cell neuroendocrine carcinoma, and clear cell odontogenic carcinoma, during surgery or biopsy, who visited the Oral Oncology Clinic, Center for Rare Cancers, National Cancer Center and the Department of Oral and Maxillofacial Surgery, Seoul National University Dental Hospital, Republic of Korea from March 2021 to June 2023. Written informed consent was obtained from all patients. The study protocol was approved by the Institutional Review Board of the National Cancer Center of Korea (Approval No.: NCC2020-0290). Detailed clinical information is available in Table S1. The specimens were transferred to tissue storage solution (Miltenyi) after collection. All procedures, including specimen isolation, collection, and cell dissociation, were completed within 2 h.

Method details

Organoid cultures

For organoid culture, cells were isolated from a viable portion of tumor tissue. Single-cell suspensions were obtained using a combination of mechanical dissociation and enzymatic degradation of the extracellular matrix (gentle MACS Dissociators; Miltenyi Biotec). PDOs were dissociated and mixed with 40 μL of basement membrane extract (BME) containing 30,000 cells per well in a 24-well plate. After the Matrigel hardened, organoid growth media with advanced DMEM/F12 medium, containing B27, 1.25 mM N-acetyl-L-cysteine, 50 ng/mL EGF, 100 ng/mL FGF-10, 50 ng/mL R-spondin 1, 10 mM nicotinamide, 5 μM A83-01, 10 μM forskolin, 1% Noggin-conditioned media, 100 ng/mL Wnt surrogate-Fc fusion protein, 1% GlutaMAX, 10 mM HEPES, and 1% primocin, were added.

The growth rate of the organoids was calculated using the following equation:

$$Organoidgrowthrate=\frac{{\mathit{log}}_2\left(\frac{Numberofharvestedsinglecells}{Numberofsinglecellsinitiallyseeded}\right)}{Daysofcultureduringonepassage}$$

Histopathological analysis

To embed organoids with diameters of 100–200 μm in paraffin blocks, the medium was removed, and the cells were washed with PBS. Agarose was added at a concentration of 20%, and the mixture was incubated for 20 min at room temperature. Hardened organoid-agarose gels were fixed in 10% formaldehyde for 24 h and embedded in paraffin. Paraffin blocks were sectioned at 4 μm thickness, stained with hematoxylin and eosin (H&E), or subjected to immunohistochemistry (IHC). Briefly, the slides were heated for 1 h in a dry oven at 60°C and rinsed in xylene for deparaffinization. Slides were rehydrated with increasing concentrations of alcohol and washed with PBS. Endogenous peroxidase was blocked with hydrogen peroxide and incubated at high pressure for 20 min in citrate buffer for antigen retrieval. The prepared slides were then overnight incubated at 4°C with primary antibodies against pan-Cytokeratin (CK), Ki-67, P53, and P63. After washing with PBS, the specimens were incubated with mouse or rabbit secondary antibodies at room temperature for 30 min and then stained with 3, 3′-diaminobenzidine (DAB). Images were captured using the Vectra Polaris imaging system (PerkinElmer Informatics).

Morphology scoring

Brightfield microscope images were segmented using the Segment Anything Model (SAM) to isolate organoids from the background (https://github.com/facebookresearch/segment-anything). Two segmentation approaches were applied: (1) center-based segmentation, where the model selected the primary object at the center of the image, and (2) full-field segmentation, where all objects in the image were segmented, followed by filtering to retain only organoids that met predefined morphological criteria. Segmentation masks were manually evaluated, and the most representative organoid mask was selected for each image. Morphological features were quantified using circularity and solidity, extracted via the regionprops function in the skimage package (Python 3.9, Python Software Foundation, Wilmington, DE, USA). These metrics were normalized such that a perfect circle without indentations or internal gaps scored 1.0. A morphology score function was derived through optimization, ensuring that the minimum score of normal-like samples exceeded the maximum score of dense samples. Optimization was performed using the optim function with default parameters in R 4.3.1 (R Core Team, Vienna, Austria). To validate the morphology score classification, an additional 336 independent images (not used in score derivation) were evaluated using the optimized parameters. Morphology scores were calculated, and a confusion matrix was generated to assess model performance in subtype classification (Figure S2B).

Somatic single nucleotide polymorphism (SNP)/insertion and deletion (INDEL) calling and annotation

We collected aligned binary alignment map (BAM) files for each normal-tumor pair for somatic mutation analysis and processed them using the somatic analysis tool Mutect2. Mutect2 identifies somatic point mutations with high confidence through statistical analysis and detects short indels by assessing read counts and aligning quality-related statistics around potential indels. Mutations were filtered out false-positive somatic mutations using the Mutect2 germline resource. Filtered variants were annotated using another program called SnpEff and filtered using dbSNPs and SNPs from the 1000 Genomes Project. The final product was in the.vcf format. Subsequently, an in-house program and SnpEff were used to annotate additional databases, including ESP6500, ClinVar, dbNSFP, and ACMG. To identify the mutational landscape, we employed the “maftools” package in R (R Foundation for Statistical Computing, Vienna, Austria).³⁸

Tumor mutation burden (TMB) calculation

The following formula was used to calculate the TMB: TMB = Number of somatic single-nucleotide variants (SNVs) and INDELs within the exonic region/length of the exonic region (Mbp). The length of the exonic region was determined using positional data from refGene data sourced from the UCSC database. The exon region for a gene was identified by combining the exonic regions of all its isoforms. Non-coding genes lacking amino acid sequences were excluded from calculating exonic region lengths. Only non-synonymous somatic SNVs within the exonic region were counted to calculate the TMB. We screened variants from the passed variants, limited to targeted regions, using an in-house script to exclude those with 1,000 genome allele frequencies above 0.05 or below 0.02 and a read depth (DP) less than 30. The filtered variants were saved as “Final.filter.vcf.”

ssGSEA Analysis and Subtype Classification

Gene Ontology: Biological Process (GOBP) gene sets were obtained from the Molecular Signatures Database (MSigDB). The latest head and neck squamous cell carcinoma (HNSC) dataset from The Cancer Genome Atlas (TCGA) was retrieved using the TCGAbiolinks (version 2.30.0) package in R. Tumor classification was determined based on tissue and organ of origin in TCGA metadata, identifying 338 patients as OSCC cases. For survival analysis, “days_to_last_follow_up” (for surviving patients) and “days_to_death” (for deceased patients) were used as survival time metrics. Log rank tests (p < 0.05) were used to determine statistical significance. RNA-seq data from 40 OSCC-derived organoids and OSCC-TCGA patient samples were analyzed using a modified single-sample Gene Set Enrichment Analysis (ssGSEA) algorithm (https://github.com/broadinstitute/ssGSEA2.0). Pathways with statistically significant differences across all morphological subtypes were selected for downstream analysis. Statistical significance between subtypes was assessed using the Kruskal-Wallis test. These pathways are referred to as morphology-specific features (Figure S5). To assign morphological subtypes to TCGA patients, the correlation between morphology-specific GOBP pathway scores in organoids and patient samples was assessed. The following steps were performed.

• (1)Autocorrelation was applied to the morphology-specific feature score matrix of OSCC-TCGA patients.
• (2)Hierarchical clustering was conducted using Ward’s D2 method to generate k number of patient subclusters.
• (3)Pearson correlation analysis was performed between the mean morphology-specific pathway scores of each TCGA subcluster and those of each organoid-derived subtype. Only statistically significant correlations (p < 0.05) were retained.
• (4)The remaining patients were re-clustered into 3 using autocorrelation.
• (5)Each cluster was assigned to a morphological subtype based on the highest positive correlation per subtype.

To determine the optimal number of subclusters generated in step 2, the mean silhouette width was measured with autocorrelation matrix from step 4. The subclusters with the highest mean silhouette width, k = 13, were used for further analysis (Figure S5E).

The resultant subtypes were analyzed for 2-year survival outcomes. The Cox proportional hazard ratio was calculated using the normal-like subtype as a reference. All analyses were performed using R 4.3.1.

Drug response test

Organoids were plated with 20 μL of 10% BME containing 100 cells per well and seeded in 384-well plates. The organoid growth medium was added after incubation for 3 days for organoid formation. Drugs were applied at five doses at 10-fold intervals as follows: 5-fluorouracil (5-FU) (0.01–100 μM), Cisplatin (0.01–100 μM), Paclitaxel (0.01–100 nM), Docetaxel (0.01–100 nM), Tipifarnib (0.01–100 μM), EPZ015666 (0.01–100 μM), AZD4547 (0.01–100 μM), Niraparib (0.01–100 μM), Alpelisib (0.01–100 μM), Nutlin-3a (0.01–100 μM), Sapitinib (0.01–100 μM), AZD7762 (0.01–100 μM), Everolimus (0.01–100 μM), and BAY1895344 (0.001–10 μM). Cell viability was measured after 5 days using a Cell Titer Glo 3D viability assay kit (Promega). Luminescence intensity was detected using SPARK microplate reader (Tecan).

Quantification and statistical analysis

Statistical analyses, including area under the curve (AUC), one-way ANOVA, 2-tailed Student’s t test, and chi-square test, were performed using GraphPad Prism 8.4.3 (GraphPad Software, Inc., CA, USA). Statistical significance: ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, and ∗∗∗∗p < 0.001.

Published: May 12, 2025