Dual-oxygen pancreatic cancer organoids recapitulate basal–classical heterogeneity validated by spatial transcriptomics

Abstract

Background

Pancreatic ductal adenocarcinoma (PDAC) comprises heterogeneous classical and basal-like subtypes. Although pancreatic cancer organoids (PCOs) are widely used, normoxic cultures preferentially maintain the classical subtype, limiting their ability to represent intrinsic intratumoral heterogeneity. Because PDAC develops in profoundly hypoxic microenvironments, the influence of oxygen tension on organoid modelling remains unclear.

Methods

We performed spatially resolved single-cell transcriptomics on 12 primary PDAC tumours to map the distributions of classical and basal-like cells. Paired organoids were generated under normoxic (20% O₂; NORM-PCOs) and hypoxic (1% O₂; HYPO-PCOs) conditions, and morphology, subtype scores, proliferation and drug responses were compared.

Results

Spatial profiling revealed gradients of basal-like–classical heterogeneity across tumours, with basal-like cells enriched in hypoxic regions. HYPO-PCOs tended to retain basal-like traits, including solid or multi-layered morphology, and increased proliferation, along with a tendency toward reduced sensitivity to chemotherapeutic agents. In contrast, NORM-PCOs more often exhibited mono-layered morphology and showed higher classical subtype scores.

Conclusions

Hypoxic culture conditions promote the preservation of basal-like subpopulations in PDAC organoids, more accurately reflecting the basal–classical heterogeneity of primary tumours. Hypoxia-induced basal-like features remained stable after oxygen switching, indicating a fixed phenotype that informs organoid modelling and offers insight into PDAC progression.

Background

Pancreatic ductal adenocarcinoma (PDAC) is a challenging malignancy, with a 5-year survival rate of less than 10%. Clinico-pathologically [1], PDAC is characterised by poor vascularisation and extensive desmoplastic stroma [2], which severely restricts oxygen (O₂) delivery, subjecting the tumour cells to a severely hypoxic microenvironment (under 0.7% O₂) [3, 4]. These hypoxic tensions are thought to be a key driver of PDAC’s aggressive behaviour, contributing to treatment resistance, metastatic potential and poor patient prognosis [5–8]. Breakthroughs in PDAC research are thus currently driven by creating novel approaches that elucidate how PDAC cells adapt to and thrive in such extreme hypoxia.

PDAC progression has long been understood through a multi-step carcinogenesis model [9], which ties morphological changes to the accumulation of genetic mutations [10, 11]. As PDAC progresses, well-differentiated tumour regions gradually give way to poorly differentiated, more aggressive forms, reflecting the genetic shift from the classical subtype to the basal-like subtype [12–14]. Basal-like PDAC, marked by its mesenchymal characteristics and significantly worse prognosis, represents a major therapeutic challenge [15]. Advances in genomic studies, particularly single-cell RNA sequencing, have revealed the mixture of these subtypes within the same tumour [16–18] but have lacked the spatial resolution necessary to fully understand their distribution and functional roles [17, 19].

Hypoxia is central in driving the basal-like subtype [20], as evidenced by the upregulation of hypoxia-related genes in these aggressive cancer cells [21]. Studies have shown that PDAC tumours exhibit O₂ concentrations as low as 0.7%, much lower than other solid tumours (e.g. lung or colorectal cancers) [22]. Hypoxia-adapted cancer cells not only become more resistant to conventional therapies but also gain invasive and metastatic potential, a key factor in the progression of basal-like PDAC [20]. Therefore, investigating the molecular mechanisms of tumour evolution by hypoxic selective pressure is critical for identifying novel therapeutic targets for PDAC.

Patient-derived organoid models have become an essential tool in cancer research, offering an in vitro system that more closely recapitulates the in vivo tumour microenvironment [23–26]. As a result of numerous excellent studies, niche dependence has become the foundation for a highly effective strategy to selectively isolate specific cells [19, 27]. However, recent reports have pointed out a potential selection bias in PDAC organoid cultures, where basal-like subtype cells are often eliminated, causing the classical subtype cells to dominate, even when both subtypes are present in the primary tumour [17, 19]. This phenomenon results in the selective propagation of well-differentiated, classical-oriented cells, while the poorly differentiated, basal-like-oriented cells that drive malignancy are underrepresented. Addressing this limitation is critical for developing organoid models that accurately reflect the intratumoral heterogeneity of PDAC.

In a previous study, we reported a single case that pancreatic cancer organoids (PCOs) established under hypoxic conditions (1% O₂) were enriched in basal-like cells, whereas those under normoxic conditions (20% O₂) preferred classical cells [28]. This finding highlighted the importance of O₂ levels in shaping the cellular composition of PCOs and offered a novel approach to studying the aggressive basal-like subtype in vitro. However, questions remained as to the correspondence between each PCO and the morphological and molecular features of its primary tumour. Recently, spatial transcriptomics technology, such as the Xenium In Situ platform, has been developed to enable analysis of a wide range of genetic expression at single-cell resolution on glass slides while preserving morphological information [29, 30].

In this study, we aimed to overcome selection biases in PCOs modelling by establishing cultures under extreme hypoxia. We hypothesised that hypoxia would preferentially preserve basal-like subpopulations often underrepresented in normoxic conditions. To test this, we analysed phenotypic differences in 12 paired PCOs grown under dual-oxygen conditions, comparing their morphology, transcriptomic profiles, drug responses and phenotypic shifts upon oxygen switching. Spatial transcriptomics (Xenium) was used to evaluate how well these organoids reflected the basal–classical heterogeneity of their primary tumours.

Materials and methods

Clinical samples

A total of 24 consecutive cases of patients who were diagnosed with PDAC and underwent surgery at University of Tsukuba Hospital with informed consent were used in this study. This full study protocol was approved by the Institutional Review Board of the University of Tsukuba Hospital (registration number: H28-090). Patient background items, including age, sex, tumour location, presence or absence of preoperative treatment, and post-operative diagnoses, are shown in Supplementary Table 1.

3D organoids culture and treatment

All 3D organoids were established from clinical specimens. Specimens were washed vigorously and minced into 10 mm³-sized fragments. These fragments were then digested with Liberase TH (#5401151001 Roche Life Science) at 37 °C for 30 min and TrypLE Express Enzyme (#12605020 Thermo Fisher Scientific) at 37 °C for 20 min. Approximately 2.0 × 10⁴ cells per well were seeded in 20 µL Matrigel (#356231 BD Biosciences) domes on a 48-well plate. Basal culture medium was made according to a prior description [27]. The medium for hypoxia was equilibrated in a sealed container with O₂ adsorbent at least 6 h prior to use. The media were changed every 3 or 4 days. A total of three passages were performed in medium without EGF for the purpose of enriching PCOs. All organoid models were routinely tested for Mycoplasma. To confirm the morphological and genetic plasticity, both cultures were eventually switched to opposite O₂ conditions and cultured for a total of three passages (approximately 1 month).

Subtype scores

For the validation study of Moffitt’s classification, gene expression data for basal-like and classical subtypes were extracted based on 25 basal-like genes and 25 classical genes (Supplementary Table 2), and were normalised using Z-scores. Subtype score was then calculated by subtracting the total Z-score of Classical genes from the total Z-score of Basal-like genes [31]; samples with a total score ≥ 0 were classified as Basal-like subtype while samples with total score < 0 were classified as Classical subtype.

Spatial transcriptomic data analysis

Pre-processed data from each sample were further processed and analysed individually using the R package Seurat (v.5.1.0) [32] in RStudio (v.4.0.2). After removing cells expressing over 150 features, we filtered out cells with no feature counts. We then normalised data using the NormalizeData function and extracted highly variable features using the FindVariableFeatures function. Normalised data underwent a linear transformation (scaling) and principal component analysis (PCA) based on variable features using the RunPCA function. Graph-based clustering was then performed according to gene expression profiles using the FindNeighbors and FindClusters functions with default parameters before results were visualised using a nonlinear dimensional reduction UMAP technique running the RunUMAP and DimPlot functions. Cell clusters were annotated based on the expression of canonical markers, including EPCAM, CDH1, KRT19, MUC1 and CEACAM5 for epithelial cells; PPY and INS for islet cells; AMY2A and CPA1 for acinar cells; COL1A1, COL1A2, ACTA2 and VIM for fibroblasts; PECAM1 and CD34 for endothelial cells; SOX2 and CCN3 for nerve cells; and CD3, CD4, CD8, GZMA, ITGAX, GZMB, CD19, FCGR3A, ITGAM and CD68 for immune cells. The FindIntegrationAnchors function was used to identify sources of variation common to multiple datasets, and the data were integrated using the anchor in the IntegrateData function with a canonical correlation dimension of 20. In integrating the data, samples were trimmed to 2.5 ×2.5 mm squares to exclude edge effects and to avoid exceeding the amount of data the computer could handle. The integrated data were scaled, and PCA was performed as was done on the individual datasets. Only the epithelial component was extracted from the integrated data and used for analysis.

Hypoxia scores

To validate hypoxia in each cell, expression data for hypoxia-related genes were extracted and normalised to Z-scores. Hypoxia-related genes were those reported by Buffa et al. [33] and Liberzon et al. (MSigDB) [34] that are included in the Xenium gene panel. The total score was calculated by summing the Z-scores of hypoxia-related genes.

5-group classification by Moffitt’s basal-like and classical genes

To visualise the spatial distribution of subtype gradients (i.e. the transition from Classical to Basal-like) across tissue sections using the Xenium platform, we categorised the samples into five subtype groups as follows. Based on single-cell subtype scores derived from spatially resolved single-cell analysis, cells were stratified into quintiles (20% intervals). Cells in the top 20% were defined as ‘Pure Basal-like’ (orange), those in the 20–40% range as ‘Basal-like’ (mustard), 40–60% as ‘Intermediate’ (yellow-green), 60–80% as ‘Classical’ (jay blue), and 80–100% as ‘Pure Classical’ (royal blue).

Shannon index

The Shannon index was used for evaluation of 5-group subtyping clusters heterogeneity. The formula for the Shannon index is as follows:

$$Shannonindex=-{\sum }_{i=1}^c{p}_i\ln p_i$$

In the Shannon index, p is the proportion (n/N) of individuals of one cluster (n) divided by the total number of clusters (N), ln is the natural log, Σ is the sum of the calculations, and c is the number of clusters.

The theoretical maximum of the Shannon index in this setting is ln (5) = 1.61. We therefore defined low diversity as a Shannon index <1.0, which corresponds to less than approximately 60% of the maximal attainable diversity and typically reflects a community structure in which one subtype dominates while the others are present only as minor components. This operational cutoff provides a biologically interpretable separation between samples with strongly skewed subtype compositions and those with more balanced subtype distributions.

Differentiation score

To quantify the degree of differentiation, we calculated a differentiation score using gene sets reported in previous studies. Markers of high differentiation included HNF4A, MUC1 and CEACAM5, while low differentiation markers included TP63 and LDHA [35–39]. The differentiation score was defined as the sum of Z-score for the high differentiation gene set minus the sum for low differentiation gene set. Thus, a higher score reflects a higher level of differentiation.

Chemotherapeutic response

Chemosensitivity assays were performed using the CellTiter-Glo® luminescent cell viability assay (#G7570, Promega), following the manufacturer’s protocol. Briefly, organoids were dissociated into single cells, and 2.0 × 10³ cells were seeded per well in 96-well plates using 5 μL Matrigel domes (#356231, BD Biosciences), in triplicate. After 96 h of incubation, cells were treated with 5-fluorouracil (#068-01401, FUJIFILM Wako) and gemcitabine (#073-06631, FUJIFILM Wako) for an additional 72 h. Following drug exposure, cells were incubated with CellTiter-Glo reagent for 10 min, and luminescence was measured using a microplate reader (Varioskan LUX, Thermo Fisher). Background luminescence was subtracted using control wells containing medium without cells. IC₅₀ values were calculated and compared between organoid lines to assess differential chemosensitivity.

Statistical analysis

RNA expression levels, tumour volumes of PCO-derived xenografts, Shannon index and subtype score graphs were analysed using the Mann–Whitney U test. Paired two-tailed t-tests were used for organoid-pair experiments, including comparisons of doubling times for cell proliferation and the areas under the curve (AUC) of chemotherapeutic responses. One-way ANOVA and Tukey’s multiple comparison test were used when testing comparisons of three or more groups. Correlation coefficient analysis was calculated using Pearson’s correlation coefficient. Fisher’s exact test was used for comparing the engraftment rate of xenografts. The data were analysed using GraphPad Prism 6.0 and RStudio (v.4.0.2). The P values are expressed as follows: *p < 0.05, **p < 0.01, ***p < 0.001, ****p < 0.0001, ns (not significant).

Results

Establishment of patient-derived organoids under two different oxygen conditions (20% and 1% O₂)

Attempts were made to establish PCOs under 20% O₂ (normoxia) and 1% O₂ (hypoxia) conditions from 24 unique clinical PDAC samples. The tumours were located in the pancreatic head in 5 cases and in the body or tail in 7 cases. Of the 24 total organoid samples, 16 were derived from patients who had had received neoadjuvant chemotherapy. Importantly, none of the patients who underwent preoperative therapy achieved an Evans grade IV (i.e. pathological complete response) postoperatively, indicating that viable tumour cells were present in all cases. Detailed patient characteristics are summarised in Fig. S1a. Successful cultures were obtained for 13 (54.2%) samples under normoxia and 17 (70.8%) samples under hypoxia (Fig. S1b–d). PCOs from 12 cases were then successfully established under both O₂ conditions. Further analyses and experiments were conducted on these 12 paired PCOs.

Spatial transcriptomics maps basal–classical heterogeneity within hypoxia-associated regions of PDAC

To assess the heterogeneity of PDAC, we performed spatially resolved, single-cell transcriptomic analyses of primary PDAC tissue corresponding to successfully established 12 paired PCOs using the Xenium platform. The process for quantifying gene expression at the single-cell level is illustrated in Fig. 1a. Briefly, cells were segmented using the nuclear expansion method based on DAPI fluorescence staining. Gene expression signals within each cell are represented as plots. From the integrated data of these 12 cases (Fig. S2a, b), cancer cells (n = 133,168) were extracted based on cell-type-specific markers (Fig. S2c, d) and analysed. First, to quantify the degree of subtype representation, we calculated a subtype score for every single cell and assessed the distribution range of these scores within each tumour (Fig. 1b). A wide range of cases, from those rich in classical cells to those rich in basal-like cells, were observed among the 12 cases.

Fig. 1. Molecular subtype of each cell and O₂ gradient within PDAC tissue were extremely heterogenous by spatial transcriptomics (Xenium In Situ).

a Spatial Transcriptomics Overview: a schematic analysis shows DAPI and cell segmentation, with basal-like (orange) and classical (blue) gene counts plotted per cell. This method estimates the subtype score for each tumour cell. b In 12 cases with successful organoid establishment in both O₂ conditions, intratumoral subtype scores (mean ± SD) were calculated for each case. c Cells were categorised into five subtypes based on subtype score by dividing the subtype score into quintiles (each representing 20% intervals from the top). The curve represents a kernel density estimate of the observed single-cell subtype score distribution across all analysed cells. d The proportion of the 5-group subtypes in each case is shown above. These 5-group subtypes were then projected onto the H&E-stained primary tissues at different magnifications (scale bar: 200 μm). e For representative primary tumours, a heatmap of hypoxia-related genes was compared with the 5-group subtypes. Well-differentiated and classical regions are indicated by red circles, whereas undifferentiated and basal-like regions are indicated by blue circles (scale bar: 200 μm). f, g Boxplots of hypoxia scores and differentiation scores are shown for each of the 5-group subtypes. Statistical significance was determined by one-way ANOVA and Tukey’s test (mean ± SD, ****p < 0.0001).

Based on this figure, cancer cells were classified into five subtype groups according to 20% intervals of the subtype score (Fig. 1c). Our 5-group subtyping clusters were named ‘Pure Classical’, ‘Classical’, ‘Intermediate’, ‘Basal-like’ and ‘Pure Basal-like’, in order from the lowest to the highest subtype score of each sample. For each PDAC tissue, the percentage of cells classified into each of 5-group subtypes was evaluated (Fig. 1d, upper row). Among the 12 cases, 5 were classified as classical-predominant, with more than 50% of cells falling into the Pure Classical or Classical categories. Another 5 cases were basal-like–predominant, with over 50% of cells categorised as Basal-like or Pure Basal-like. The remaining 2 cases exhibited an intermediate phenotype. The subtype diversity was also assessed using the Shannon diversity index, which reflects the degree of heterogeneity [40], for the 5-group subtype (Fig. S2e). Quantitative analysis of the subtype diversity revealed that 3 samples exhibited a Shannon index below 1.0, indicating low diversity. Among them, P20 was enriched for the classical subtype, whereas P11 and P24 were enriched for the basal-like subtype. In contrast, the other 9 out of 12 cases showed high diversity, with a wide range of coexisting groups from Pure Classical to Pure Basal-like. The distribution of the 5-group subtype was then projected onto the primary PDAC tissues (Fig. 1d). Pure Classical and Classical clusters were distributed in relatively well-differentiated regions with ductal structures (Fig. 1e, red circles) while Basal-like and Pure Basal-like clusters were distributed in relatively poorly differentiated regions with a lack of ductal structures (Fig. 1e, blue circles).

Next, we attempted to visualise the O₂ gradient in PDAC tissue. Currently, there is no retrospective quantification method for assessing hypoxia within tissues. Reports are limited to attempts that preoperatively administered pimonidazole to patients to identify hypoxic regions [41]. In this study, however, preoperative administration of pimonidazole was not performed, making such quantification challenging. Since hypoxia-related genes are useful in assessing tissue hypoxia [42, 43], of the genes previously used in Buffa et al. [33] and the hypoxia gene set of the molecular signatures database (MSigDB) [34], 13 genes (Fig. S2f) from the Xenium panel were used to evaluate the hypoxia scores and heatmaps. Heatmaps of hypoxia-related genes were generated. Comparing the 5-group clusters with the tissue morphology revealed that the classical-oriented cluster, which contained well-differentiated regions, tended to exhibit lower expression of hypoxia-related genes (Fig. 1e, red circles), while basal-like-oriented clusters with poorly differentiated regions tended to have higher expression of hypoxia-related genes (Fig. 1e, blue circles). To assess the correlation between hypoxia and subtype, the hypoxia score of each cell was calculated and compared to the 5-group subtype with the tissue morphology. Although the hypoxia score, treated as a continuous variable, did not exhibit a strong linear correlation with the subtype score, it showed a significant stepwise increase across the five subtype groups, rising from Pure Classical to Pure Basal-like (Figs. 1f and S2g). To further examine the relationship between morphological differentiation and subtype identity, we evaluated differentiation scores for each of the five subtype groups. This analysis revealed a clear trend: more basal-like subtypes were associated with lower differentiation scores, suggesting an inverse relationship between basal-like features and low differentiation (Figs. 1g and S2h).

HYPO-PCOs show multi-layered or solid morphology corresponding to primary tumour structure

The bright-field images of the 12 paired PCOs are shown in Fig. 2a. The morphology of each PCO was classified into three categories: Mono-layer, Multi-layer and Solid [13]. A representative example of these three morphological classifications is shown in Fig. 2b. In NORM-PCOs, 10 cases exhibited a Mono-layer structure, with 0 cases of Multi-layer and 2 cases of Solid. In contrast, in the HYPO-PCOs, 5 cases were Mono-layer, 3 cases were Multi-layer and 4 cases were Solid. Among the 12 pairs, clear morphological differences between NORM-PCOs and HYPO-PCOs were observed in 5 pairs (Fig. 2c, outlined in red). For the 5 pairs, NORM-PCOs were all mono-layered while HYPO-PCOs ranged from multi-layered to solid. All 5 corresponding primary tumours of these pairs exhibited mixed tissue structures, composed of relatively well-differentiated (Fig. 2d, #) and poorly differentiated (Fig. 2d, †) regions. The PCOs’ morphology classified in vitro is retained in the orthotopic xenograft mouse models, where mono-layer PCOs formed well-differentiated tumours while solid PCOs formed poorly differentiated tumours (Fig. 2d, right, and Fig. S3).

Fig. 2. HYPO-PCO tended to exhibit a multi-layered or solid morphology compared to NORM-PCO, plus correspondence between PCO and the primary tissue.

a Bright-field images of 12 NORM-PCO and HYPO-PCO pairs reveal distinct morphologies. NORM-PCOs predominantly exhibited mono-layer forms, while HYPO-PCOs showed multi-layered or solid structures. Scale bars, 100 μm. b Morphological classifications: mono-layer (single wall), multi-layer (multiple walls) and solid (loss of luminal structure), differentiated by encircling lines. c Among NORM-PCOs, 10 were mono-layer and 2 were solid. HYPO-PCOs included 5 mono-layer, 3 multi-layer and 4 solid cases. Five pairs (red boxes) displayed marked differences between NORM and HYPO-PCOs. d Primary PDAC tissues showed heterogeneous histology, ranging from highly differentiated (#) to poorly differentiated (†) areas. Histological differences between NORM- and HYPO-PCOs were evident and recapitulated in orthotopic xenografts, confirmed by H&E and human-HLA staining. Scale bars, 100 μm.

Spatial transcriptomics compares subtype composition between PCOs and primary tumours

We next conducted 5-group Xenium subtyping on 12 paired PCOs. For each organoid line, several hundred individual organoids were analysed, as detailed in Fig. S4a. After integrating the datasets from PCOs and primary tumours, cell types were identified (Fig. S4b–d). In addition, the range of subtype scores within PCOs was then evaluated and compared with those of the corresponding primary tissues (Fig. 3a). The overall subtype score for HYPO-PCOs was significantly higher than that for NORM-PCOs (Fig. S4e). After applying the five‑group subtype classification to both NORM-PCOs and HYPO-PCOs, we found that 7 of 12 NORM-PCOs contained more than 50% classical-oriented cells (i.e. Pure Classical + Classical). By contrast, in HYPO-PCOs, only 2 of 12 cases contained more than half of the classical-oriented cells, while the other 10 cases were predominantly in the Intermediate to Pure Basal-like cells (Fig. 3b). In addition, the Shannon index of our 5-group subtype was evaluated for each PCO (Fig. 3c). Low diversity was observed in 8 samples in NORM-PCOs, whereas only 3 samples showed low diversity in HYPO-PCOs. The index was significantly higher in HYPO-PCOs compared to NORM-PCOs (Fig. 3d).

Fig. 3. Spatial transcriptomics revealed molecular subtypes, O₂ gradient differences and the correspondence between PCOs and the original tissues.

a Subtype score ranges in NORM- and HYPO-PCOs overlapped with those of the primary tissues. b Percentages of 5-group subtypes were compared for NORM- and HYPO-PCOs. c, d Shannon index analysis revealed higher subtype diversity in HYPO-PCOs compared to NORM-PCOs (*p < 0.05, Mann–Whitney U test). A red line was drawn at a Shannon index of 1.0 to indicate the boundary between high and low diversity. e–h Spatial transcriptomics confirmed subtype correspondence between PCOs and primary tissues. For example, in P20 (classical-dominant), both PCOs lacked basal-like components. In heterogeneous P01, NORM-PCO01 was mono-layered and classical-rich, while HYPO-PCO01 was solid and retained basal-like traits. In diverse P07, HYPO-PCOs robustly recapitulated basal-like components, unlike NORM-PCOs.

Earlier, we discussed the diversity in tissue morphology and subtypes between the primary tumours and PCOs. To clarify the correspondence between PCOs and their primary tumours, each of our 5-group subtypes was next spatially compared between the primary tumour tissue and the PCO tissue. Histomorphology and 5-group cluster projections of the primary tissues and the corresponding NORM-/HYPO-PCOs were compared (Figs. 3e–h and S5). In cases with less heterogeneity in the primary tissue morphology and subtype, such as P20 (where the histology was almost homogeneous and skewed in classical subtypes), both NORM-PCO20 and HYPO-PCO20 were of the same PCO morphological classification (mono-layer) and primarily composed of Classical and Pure Classical (Fig. 3e). In contrast, in cases with heterogeneous morphology and subtype status in the primary tissues, remarkable subtype cluster differences were observed between NORM-PCOs and HYPO-PCOs. NORM-PCO01 had a mono-layer morphology and was enriched in Classical and Pure Classical. On the other hand, HYPO-PCO01 was richest in Basal-like and Pure Basal-like, with HYPO-PCO01 having a solid morphology (Fig. 3f). In P07, which exhibited marked subtype heterogeneity within the primary tumour, NORM-PCO07 showed a prominent loss of its basal-like component, whereas HYPO-PCO07 retained this basal-like component to a substantially greater extent (Fig. 3g). Moreover, even in P17 – a case with few basal-like cells in the primary tumour, HYPO-PCO17 successfully recapitulated the basal-like component (Fig. 3h).

Bulk RNA sequencing reveals genomic and transcriptional differences between NORM- and HYPO-PCOs

The Moffitt subtype classification [15] and the putative mutation status inferred from the results of bulk RNA sequencing are shown (Fig. 4a). In NORM-PCOs, none of the organoids were classified as the Basal-like subtype; all were classified as Classical. In contrast, among the HYPO-PCOs, PCO1, PCO12 and PCO24 were classified as Basal-like. The heatmap of subtype scores then showed a uniform blue pattern for NORM-PCOs while some HYPO-PCOs displayed a yellow-green to orange gradient. Gene mutations in the PCOs were also evaluated through RNA sequencing (Fig. 4b). KRAS mutations were observed in 21 cases (87.5%), p53 mutations in 16 cases (66.7%), SMAD4 mutations in 7 cases (29.2%) and CDKN2A mutations in 10 cases (41.7%). Differences in the presence of 4 such gene mutations between NORM-PCO and HYPO-PCO were noted in PCO01, PCO07 and PCO16. The most significant difference was observed with the p53 mutation, where PCO07 and PCO16 had p53 mutations only in the HYPO-PCO. KRAS wild-type was only detected in NORM-PCO01 and both NORM- and HYPO-PCO16, while NORM-PCO01 exhibited a BRAF mutation instead. PCA indicated that NORM-PCOs and HYPO-PCOs tended to separate into distinct clusters (Fig. S6). The results of the GSEA indicated that NORM-PCOs were significantly enriched for Moffitt classical signalling and oxidative phosphorylation, while HYPO-PCOs were significantly enriched for hypoxia, Moffitt basal-like signalling, TNF-α signalling, mTORC1 signalling, EMT and TGF-β signalling (Fig. 4c).

Fig. 4. Bulk RNA sequencing revealed differences in genotype, subtype and specific hallmarks between NORM-PCO and HYPO-PCO.

a Bulk RNA sequencing revealed distinct molecular features. Subtypes were classified by Moffitt’s classification. The samples are arranged in order of their subtype scores. The heatmap of basal-like and classical gene expression (z-score normalised). b Mutational statuses from bulk RNA sequencing. c GSEA identified enriched pathways in HYPO-PCOs, including hypoxia, basal-like, glycolysis, mTORC1, EMT and TGF-beta signalling, while classical and oxidative phosphorylation pathways were enriched in NORM-PCOs. NES and FDR values confirmed these distinctions.

HYPO-PCOs exhibit pharmacological and biological features consistent with basal-like PDAC

The biological and pharmacological characteristics of NORM-PCOs and HYPO-PCOs were confirmed and analysed. The dose-response curves from drug sensitivity assays with 5-fluorouracil (5-FU) are shown (Figs. 5a and S7a). The normalised AUC and Emax values were significantly higher and the IC50 significantly lower in HYPO-PCOs compared to NORM-PCOs (Figs. 5b and S7b). In contrast, dose-response curves for gemcitabine (Figs. 5c and S7c) indicated no significant differences in normalised AUC, IC50, or Emax values between NORM-PCOs and HYPO-PCOs (Figs. 5d and S6d). The cell proliferation curves of the 12 paired PCOs were then evaluated (Fig. S8a). No significant differences were observed between the average proliferation curves of NORM-PCOs and HYPO-PCOs (Fig. S8b). Similarly, there were no significant differences in the doubling times between NORM-PCOs and HYPO-PCOs (Fig. S8c). For 8 pairs, the O₂ concentration was altered to measure the cell proliferation rate. When NORM-PCOs were exposed to 1% O₂, the proliferation rate of NORM-PCOs significantly decreased in 4 out of 8 cases, remained unchanged in 3 out of 8 cases and significantly increased in 1 out of 8 cases. In contrast, when HYPO-PCOs were exposed to 20% O₂, the proliferation rate increased in 1 out of 8 cases, remained unchanged in 7 out of 8 cases, and did not decrease in any case (Fig. S8d). In the overall comparison of NORM-PCOs, there was no statistically significant difference in doubling time between 1% and 20% oxygen conditions. Similarly, for HYPO-PCOs, the difference in doubling time was not statistically significant (Fig. S8e). Subcutaneous and orthotopic xenografts were established from NORM-PCOs and HYPO-PCOs (Fig. S9a). Regarding subcutaneous xenografts, HYPO-PCO07 exhibited a significantly higher engraftment rate compared to NORM-PCO07, while no significant differences were observed for the other PCOs (Fig. 5e). Tumour size was recorded weekly for subcutaneous engraftment tumours (Figs. 5f and S9b). Both subcutaneous and orthotopic xenografts were terminated at 10 weeks and, at that timepoint, the subcutaneous engraftment tumour weights in HYPO-PCOs were significantly higher than NORM-PCOs (Fig. 5g). Regarding orthotopic xenografts, there were no significant differences for engraftment rates between NORM-PCOs and HYPO-PCOs (Fig. S9c). No significant differences in tumour weight were observed in the orthotopic engraftment tumours (Fig. S9d).

Fig. 5. Pharmacological and biological characteristics of HYPO-PCOs in vitro and in vivo were similar to previously reported basal-like characteristics.

a, b HYPO-PCOs exhibited higher resistance to 5-fluorouracil than NORM-PCOs, as shown by normalised AUC values (***p < 0.001, paired t-test). c, d No significant difference was observed in gemcitabine sensitivity between NORM-PCOs and HYPO-PCOs (ns: not significant, paired t-test). e The subcutaneous engraftment efficiency of HYPO-PCO07 was significantly higher than that of NORM-PCO07. No significant differences were observed for the other PCOs (ns, **0.01, Fisher’s exact test). f, g Tumour weight after 10 weeks was significantly greater for HYPO-PCOs compared to NORM-PCOs with biologically independent replicates (n = 4–6/PCO). (*p < 0.05, Mann–Whitney U test).

Oxygen switching reveals morphological stability and stable transcriptional profiles in PCOs

After three passages each for NORM-PCOs and HYPO-PCOs under O₂ conditions opposite to those of initial establishment O₂, morphology and transcriptomics were re-evaluated (Fig. 6a). Although no remarkable morphological changes were observed among 12 paired PCOs (Fig. 6b,c), differentially expressed genes (DEGs) were identified in NORM-PCOs and HYPO-PCOs across initial establishment, O₂ 20%, and O₂ 1% and presented on volcano plots (Fig. 6d). The specific genes of each PCO were then extracted by Venn diagrams (Fig. 6e). Genes noted in bold (SLC2A1, KRT17, KRT6A and KRT5) are included in Basal-like gene sets.

Fig. 6. Switching the O₂ conditions between NORM-PCO and HYPO-PCO revealed morphological rigidity in vitro and stable transcriptome profiles in silico.

a NORM-PCOs and HYPO-PCOs were placed in opposite O₂ conditions for three passages. b, c The 3 morphological classifications and corresponding bright-field images of the 12 pairs of NORM-PCOs and HYPO-PCOs were obtained after three passages under their respective oxygen concentrations. Scale bars, 100 μm. ×200 magnification. No remarkable morphological changes were observed. d Volcano plots highlight stable differential gene expression between NORM-PCOs and HYPO-PCOs under initial and switched O₂ conditions. Red points indicate the consensus genes shown in Fig. 6e. e Venn diagrams of DEGs reveal consistent transcriptomic profiles regardless of O₂ conditions. Genes noted in bold (SLC2A1, KRT17, KRT6A and KRT5) are included in Basal-like gene sets.

Discussion

We first investigated the intratumoral heterogeneity of 12 primary PDAC cases with our 5-group basal–classical classification using Xenium spatial transcriptomics, and revealed that 9 tumours exhibited pronounced subtype diversity. When we established organoids from these tumours under dual-oxygen conditions (normoxia and hypoxia), we observed striking morphological divergence: 7 of 12 HYPO-PCOs exhibited solid or multi-layered morphologies, while 10 of 12 NORM-PCOs showed predominantly mono-layered morphology. When organoids were classified using the same 5-group spatial framework, hypoxic PCOs more faithfully recapitulated the heterogeneous subtype composition of the original tumours, including basal-like populations. Bulk RNA-seq further supported this, showing that hypoxic PCOs had elevated basal-like scores, whereas normoxic PCOs had higher Moffitt classical scores. Notably, the phenotypes of both NORM- and HYPO-PCOs remained stable, even after switching oxygen conditions.

This study demonstrated that PCOs established under hypoxic conditions had higher success rates compared to normoxic conditions. While no prior studies have specifically addressed hypoxic PCO establishment, research on colorectal cancer organoids and induced pluripotent stem cells highlights the advantages of mimicking in vivo O₂ levels [44, 45], such as reducing oxidative stress and maintaining stem cell properties [46, 47]. Considering the highly hypoxic nature of pancreatic tumours, the improved establishment efficiency under hypoxia observed in this study is consistent with these findings.

Spatial analysis revealed significant subtypical and morphological heterogeneity within primary tumours. The Xenium platform enabled single-cell–level assessment of subtype scores across entire PDAC tissues, revealing marked differences in subtype diversity among cases. A 5-group classification system provided greater resolution of heterogeneity compared to binary models, such as Moffitt’s two-group classification. This framework uncovered transitions between subtypes, particularly with decreased differentiation scores from Pure Classical to Pure Basal-like states. This suggests classical cells are well-differentiated, whereas basal-like cells regress to a poorly differentiated state. Transitional zones were observed, where basal-like cells at the edges of classical-rich regions indicated a dynamic shift toward the basal-like subtype.

Hypoxia scores were positively correlated with the shift from Pure Classical to Pure Basal-like clusters, linking hypoxia to aggressive basal-like phenotypes and poor differentiation. However, hypoxia-related gene expression differed even in adjacent regions containing both classical and basal-like cells. This suggests that hypoxia-related genes may not solely reflect hypoxic tensions but also gene enrichment specific to basal-like cells. These findings reinforce the association between hypoxia, basal-like subtypes, and poor differentiation, aligning with previous evidence that basal-like cells adapt to hypoxic environments [20, 35, 48]. The study underscores the importance of hypoxia in PCO establishment and its influence on PDAC heterogeneity and subtype transitions.

Morphological divergence was observed between paired organoids in a subset of cases. Furthermore, orthotopic xenografts derived from these organoids preserved the morphological characteristics observed in vitro. These findings suggest that organoids established under distinct oxygen conditions may selectively preserve pre-existing tumour subclones present within the primary tumour microenvironment.

In contrast, PCOs with no significant morphological differences (e.g. PCO20, PCO21, PCO24) exhibited lower histological diversity within the primary tumour and across subtyping clusters, suggesting that normoxia and hypoxia influence PCO morphology in tumours with more tissue diversity. At the macro level, morphological differences between NORM-PCOs and HYPO-PCOs were observed in 5 out of 12 cases. At the micro level, transcriptomic differences were found in the pairs of all cases. Single-cell analysis revealed that NORM-PCOs had more classical genes, while HYPO-PCOs closely resembled basal-like populations, showing greater phenotypic diversity. Bulk RNA sequencing confirmed these findings, with HYPO-PCOs showing enrichment in genes associated with glycolysis, mTOR, TNF-α, EMT and TGF-β signalling, which are linked to basal-like subtype traits [49, 50].

Xenium projection of the five clusters shows that NORM-PCO and HYPO-PCO reflect distinct morphological and molecular characteristics within primary tissues. These differences may reflect the preferential expansion of distinct tumour cell populations under different oxygen conditions. In P01 and P07, where classical cells were a minority within the primary tumour, NORM-PCOs predominantly reflected the classical component. Similarly, in P17 and P22, where basal-like cells were rare, HYPO-PCO17 and HYPO-PCO22 mirrored the basal-like component. These findings suggest that hypoxic culture conditions may facilitate the establishment of organoids representing basal-like tumour phenotypes that are less frequently maintained using conventional normoxic culture conditions. Interestingly, in P11 and P24, both NORM-PCO and HYPO-PCO included organoids failed to fully recapitulate the characteristics of their corresponding primary tumours. This discrepancy may be attributable to two non-exclusive possibilities. First, a selection bias inherent to the organoid culture process - which cannot be fully explained by oxygen conditions alone. Second, sampling bias, which remains an inevitable limitation when establishing models from clinical specimens [51]. While the latter is difficult to overcome, further studies are warranted to better understand and minimise the former, as it represents a critical challenge in faithfully modelling tumour heterogeneity.

The pharmacological and physiological characteristics of HYPO-PCO reflect the clinical features of the basal-like subtype of PDAC. Although the differences were not striking, HYPO-PCOs displayed somewhat lower sensitivity to 5-FU in our assays, potentially reflecting intrinsic properties of basal-like PDAC cells previously associated with chemoresistance[35, 52]. GSEA revealed upregulation of the mTORC1 signalling pathway in HYPO-PCOs, which is related to chemotherapy resistance [53, 54]. To date, research on the heterogeneity of PDAC has primarily advanced in silico, relying on retrospective studies. In contrast, HYPO-PCO offers a model that better replicates basal-like cell populations, making it more suitable for in vivo and in vitro experiments. Additionally, subcutaneous tumours from HYPO-PCOs grew significantly faster than those from NORM-PCOs, reflecting the aggressive growth of basal-like PDAC in clinical settings [15].

Even after 1 month under switched oxygen conditions, PCOs exhibited no notable morphological alterations, suggesting relatively limited phenotypic plasticity. HYPO-PCOs maintained stable proliferation regardless of oxygen tension and demonstrated preservation of hypoxia-adapted phenotypic features compared with NORM-PCOs. Clinically, pancreatic cancer frequently metastasizes to the liver, an organ characterised by high vascularisation and elevated oxygen tension. Despite this normoxic environment, liver metastases often retain the aggressive, poorly differentiated phenotype of the primary tumour. The observation that hypoxia-induced programmes remain relatively fixed even after returning to normoxic conditions may partly explain why the basal/hypoxic phenotype is retained following liver metastasis. To investigate the molecular basis of this fixed phenotype, we identified genes that were consistently enriched in each model regardless of oxygen exposure, thereby representing model-specific transcriptional programmes. Analysis of these gene sets revealed that NORM-PCO–specific genes, such as GKN1 and GKN2, are associated with early-stage PDAC [55], indicative of a more differentiated, classical-like cellular state. In contrast, HYPO-PCO–specific genes, including SLC2A1, KRT6A, KRT5 and KRT17, are linked to basal-like phenotypes and may reflect a more aggressive malignant population. Notably, the greater number of HYPO-PCO–specific genes compared to those of NORM-PCOs may suggest that HYPO-PCOs harbour a broader and more diverse malignant cell composition. Targeting this fixed phenotype could provide a promising strategy to counter resistance and potentially improve therapeutic outcomes in PDAC.

This study has several limitations. The sample size is relatively modest, and selection bias inherent to organoid establishment cannot be completely excluded. Sampling bias from surgically resected specimens may also influence the representation of molecular subtypes. In addition, although oxygen conditions appear to influence subtype-associated phenotypes, other tumour microenvironmental factors, such as stromal interactions and nutrient gradients, were not modelled in this study.

Despite these limitations, the dual-oxygen organoid system enables the establishment of distinct tumour states from the same patient specimen, providing a platform to explore biological differences between basal-like and classical PDAC. This approach may support the identification of circulating biomarkers associated with aggressive tumour states and may also contribute to studies investigating mechanisms of therapeutic resistance. More broadly, integrating physiologically relevant oxygen conditions into organoid models may help refine experimental systems for studying PDAC heterogeneity and progression.

Conclusion

Our study demonstrates that oxygen conditions during organoid establishment influence the relative preservation of basal-like and classical phenotypes in PDAC models. Rather than supporting the superiority of a single culture condition, our findings suggest that a dual-oxygen strategy may better recapitulate the intrinsic heterogeneity of primary tumours. Spatial transcriptomic profiling revealed that basal-like, poorly differentiated tumour cells are enriched in hypoxic regions of the primary tumour and are preferentially retained under low-oxygen conditions in vitro. By incorporating oxygen gradients into organoid culture, this approach more accurately reflects the spatial and phenotypic complexity of the original tumour morphology. Such hypoxia-based modelling not only enhances the biological relevance of in vitro systems but may also provide new insights into pancreatic cancer progression.

Author contributions

KK conceptualised the study, developed the methodology, curated the data, performed the formal analysis, validation, investigation, visualisation and software development, drafted the original manuscript and participated in manuscript revision. HN contributed to the conceptualisation, methodology, formal analysis, investigation, visualisation, drafting of the original manuscript, manuscript revision and funding acquisition. TO contributed to the conceptualisation, methodology, visualisation, drafting of the original manuscript, manuscript revision, supervision, project administration and funding acquisition. SK contributed to data curation, formal analysis, investigation and software development. AS and YS contributed to data curation. PL contributed to the formal analysis, validation and software development. YA contributed to the formal analysis, software development, manuscript revision and resources. JK contributed to visualisation. YM, OS, NS and DM contributed to the provision of resources, and NS and DM also contributed to the investigation. YK and CRG contributed to manuscript revision. YO contributed to funding acquisition. MS-Y contributed to software development. All authors have read and approved the final manuscript.

Funding

Japan Society for the Promotion of Science (JSPS) KAKENHI (grant numbers JP23K19527, 22K16504 and 22H02911) of Japan.

Data availability

All study data are included in the article and/or Supplementary data. The RNA-seq data used in this study have been deposited in the Gene Expression Omnibus database under accession code GSE283850 and can be found at https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE283850.

Code availability

The codes for key computational analyses are available on GitHub at https://github.com/KumanoKoichiro/PDAC-PCO.

Competing interests

The authors declare no competing interests.

Ethics approval and consent to participate

Informed consent was received from all patients after approval by the Institutional Review Board of the University of Tsukuba Hospital (approval number: H28-090). Animal experiments were performed in accordance with approval from the Institutional Animal Care and Use Committee and the Regulation for Animal Experiments at the University of Tsukuba (approval number: 23-030). The study also conformed to the Fundamental Guidelines for Proper Conduct of Animal Experiments and Related Activities in Academic Research Institutions under the jurisdiction of the Ministry of Education, Culture, Sports, Science and Technology (Japan).

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41416-026-03577-w.