Patient-specific modeling of stroma-mediated chemoresistance of pancreatic cancer using a three-dimensional organoid-fibroblast co-culture system

Background Cancer-associated fibroblasts (CAFs) are considered to play a fundamental role in pancreatic ductal adenocarcinoma (PDAC) progression and chemoresistance. Patient-derived organoids have demonstrated great potential as tumor avatars for drug response prediction in PDAC, yet they disregard the influence of stromal components on chemosensitivity. Methods We established direct three-dimensional (3D) co-cultures of primary PDAC organoids and patient-matched CAFs to investigate the effect of the fibroblastic compartment on sensitivity to gemcitabine, 5-fluorouracil and paclitaxel treatments using an image-based drug assay. Single-cell RNA sequencing was performed for three organoid/CAF pairs in mono- and co-culture to uncover transcriptional changes induced by tumor-stroma interaction. Results Upon co-culture with CAFs, we observed increased proliferation and reduced chemotherapy-induced cell death of PDAC organoids. Single-cell RNA sequencing data evidenced induction of a pro-inflammatory phenotype in CAFs in co-cultures. Organoids showed increased expression of genes associated with epithelial-to-mesenchymal transition (EMT) in co-cultures and several potential receptor-ligand interactions related to EMT were identified, supporting a key role of CAF-driven induction of EMT in PDAC chemoresistance. Conclusions Our results demonstrate the potential of personalized PDAC co-cultures models not only for drug response profiling but also for unraveling the molecular mechanisms involved in the chemoresistance-supporting role of the tumor stroma. Supplementary Information The online version contains supplementary material available at 10.1186/s13046-022-02519-7.


Background
Pancreatic ductal adenocarcinoma (PDAC) has become the third and fourth leading cause of cancer related death in Northern America and Europe, respectively [1]. Chemoresistance represents a major challenge in the treatment of resectable and unresectable PDAC, yet systemic therapies remain largely restricted to conventional cytotoxic drugs administered as single agents or combination regimens [2]. The PDAC tumor microenvironment, consisting mostly of extracellular matrix (ECM), cancer-associated fibroblasts (CAFs), infiltrating immune cells and vasculature, plays a crucial role in driving tumor chemoresistance via diverse, not yet fully elucidated mechanisms [3]. A very prominent desmoplastic reaction driven by CAFs constitutes up to 90% of the total PDAC tumor volume [4]. This vast tumor ECM forms a physical barrier that can lead to decreased drug delivery to tumor cells [5,6]. Several studies suggest that CAFs promote tumor progression, metastasis and therapy resistance [7]. CAF-mediated chemoprotection against gemcitabine in PDAC is thought to involve a variety of mechanisms such as induction of anti-apoptosis/pro-survival pathways in tumor cells [8,9], alteration of tumor gemcitabine metabolism [10,11] and release of exosomes [12].
Suboptimal tumor modeling neglecting tumor-stromal interactions is regarded as an important contributor to the high drug attrition rate of preclinically promising drugs [13]. Incorporation of stromal components into drug screening models is therefore urgently needed. PDAC tumor organoids [14] have emerged in recent years as 3D in vitro disease models able to retain the intrinsic heterogeneity and genetic alterations of the original tumors [15][16][17]. As PDAC organoids are amenable to clinical application and seem to reflect patient drug response [16,18,19], they hold great potential for personalized oncology. Yet, drug screening based on purely epithelial organoid culture models fails to consider the contribution of the patient-specific tumor microenvironment. Hence, incorporation of relevant stromal compartments to organoid cultures is a crucial step for optimization of tumor tissue modeling and drug response prediction.
Recently, heterocellular organoid cultures have aided to grasp CAF heterogeneity and dissect complex tumorstroma interactions. For instance, two distinct yet plastic CAF subtypes, namely myofibroblastic and inflammatory CAFs, were identified using a murine organoid co-culture system [20]. Tsai and colleagues established cultures of PDAC organoids including CAFs and T cells, further demonstrating the suitability of these models for investigating the PDAC tumor microenvironment [21]. However, personalized drug screening assays using matched 3D heterotypic organoid cultures have not been reported yet.
Here, we established direct 3D co-culture models of patient-derived PDAC organoids (PDOs) and patientmatched CAFs. Using the live image-based drug assay DeathPro [22], we investigated the effect of CAFs on PDO chemosensitivity to the first-line chemotherapeutic drugs gemcitabine, 5-fluorouracil (5-FU) and paclitaxel, and elucidated transcriptional changes induced by tumor-stroma interaction at the single-cell level.

PDAC tumor specimens
Tumor tissue biopsies were obtained from surgically resected tumor specimens from patients who underwent primary resection at the Department of General, Visceral and Transplantation Surgery, Heidelberg University Hospital. The study was approved by the ethical committee of University of Heidelberg (ethic votes 301/2001, 159/2002, S-206/2011, S-708/2019 and S-083/2021) and was conducted in accordance with the Helsinki Declaration. All patients provided written informed consent prior to acquisition of tissue. Tumor entity was confirmed by pathological assessment of the resected specimens.

Isolation and culture of cancer-associated fibroblasts
Primary cancer-associated fibroblasts were isolated from the tumor specimens using the outgrowth method [23]. Pieces of minced tissue generated in the process of organoid isolation were separated and incubated with fibroblast medium containing RPMI Medium 1640 (Gibco) supplemented with 10% fetal calf serum (Gibco), 200 mM GlutaMAX (Gibco) and Penicillin/Streptomycin (Gibco). Success rate for CAF culture establishment was 90%. CAFs were passaged approximately every 8-9 days and their identity was confirmed by morphological assessment and immunofluorescence staining for α-SMA. All CAF lines tested negative for mycoplasma contamination and were used for experiments within 4 to 7 passages after isolation.

Image-based drug testing and drug response analysis
Chemosensitivity of organoids in mono-and co-culture was assessed using the live image-based drug assay DeathPro [22]. Mono-and co-cultures of PDOs and PDO/CAFs were seeded into μ-Chamber Angiogenesis 96-well-plates (ibidi) as 10 μl drops in co-culture matrix and cultured with 70 μl co-culture medium. CAFs were stained with Cell Tracker Green CMFDA (Invitrogen) before seeding to distinguish them from PDOs. The drug screen included gemcitabine (Selleckchem), 5-FU (Selleckchem) and paclitaxel (Selleckchem). For drug treatment, serial dilutions (1:4 for gemcitabine and 5-FU, 1:3 for paclitaxel) were prepared by mixing stock solutions with co-culture medium. Drugs were applied 3 days after seeding and washed out after 72 hours. Each PDO line was tested twice independently. Confocal imaging was performed 0 and 120 h after drug application at similar positions. Four hours prior to each imaging cells were stained with 1 μg/ml Hoechst (Invitrogen) and 1 μg/ml propidium iodide (Sigma-Aldrich). Images were taken in a standardized way according to DeathPro image acquisition [22] using the Visual Basic for Applications macro "Autofocus Screen" [24]. Hoechst and propidium signals were acquired simultaneously, cell tracker signals sequentially to avoid spectral overlaps. Two positions were imaged per well. At each imaging position stacks of 17-18 slices with 50 μm slice distance were acquired.
Overall, approximately 90.000 confocal images were acquired, from which around 5.700 maximum intensity projections were created and analyzed. Image processing and drug response analysis was performed using the DeathPro workflow as previously described [22]. Mean values of each independent replicate were used for the analysis.

Single cell RNA-sequencing
Single cell RNA-sequencing (scRNA-seq) was performed for 3 pairs of matched PDOs/CAFs co-cultures and the corresponding PDO and CAF 3D monocultures. For each condition, 3 co-culture matrix drops of 30 μl were seeded into one well of a 12-well plate. Single cell dissociation was performed after 5 days in culture as we previously described [25]. Single-cell sequencing libraries were prepared according to the 10x Genomics Single Cell 3 v2 Reagent Kit User Guide. Libraries were sequenced in one lane per sample with the Illumina NextSeq 500 system in high-output mode (paired-end, 75 bp).

Analysis of scRNA-seq data
Raw sequencing data were processed with CellRanger version 2.1.1 (10x Genomics), using the 10x reference human genome hg19 1.2.0 for alignment. Seurat version 3.0 [26] was used for quality control and downstream analysis. Cells with fewer than 200 genes, as well as genes represented in fewer than 3 cells, were excluded from the analysis. We also determined a maximum number of counts per gene, and a maximum fraction of mitochondrial reads, for each sample based on individual sample quality (Supplementary  Table S1). Count data was lognormalized with a scale factor of 10.000, and the 2.000 most variable genes were identified using the FindVari-ableFeatures function in Seurat. Normalized data were scaled with the ScaleData function.
To differentiate CAF and PDAC cells, expression data from all samples was merged and cells were assigned their identity based on established marker genes [27]; decorin (DCN) and lumican (LUM) expression was used to identify CAFs while expression of keratin 18 (KRT18) and keratin 19 (KRT19) was used to identify tumor organoid cells. For further analysis, PDAC and CAF cells, respectively, from monoculture and co-culture samples were combined using the IntegrateData function in Seurat. Dimensional reduction was performed using the umap-learn package [28]. Cells were clustered by the Louvain algorithm with a resolution of 0.2. Differentially expressed genes were identified using a Wilcoxon rank sum test. To link differentially expressed genes with biological pathways and functions, gene set enrichment analysis was performed using the MSigDB database [29,30], excluding genes with an adjusted p-value > 0.05. In the CAF samples, two clusters that showed enrichment for ribosomal and mitochondrial genes, respectively, were excluded from further analyses.
To distinguish iCAF-like and myCAF-like cells, we performed principal component analysis on the expression of iCAF and myCAF marker genes [27] in CAF cells, followed by Louvain clustering with a resolution of 0.1 to obtain two clusters. Scores for different gene sets were calculated using the AddModuleScore function in Seurat, using published lists of PDAC subtype marker genes [31] and gene sets from the MSigDB database [29,30] for EMT (HALLMARK_EPITHE-LIAL_MESENCHYMAL_TRANSITION) and proliferation (HALLMARK_E2F_TARGETS). Cell cycle scores were calculated with the CellCycleScoring function in Seurat. Potential ligand-receptor interactions between CAF and PDAC cells in co-culture were identified using CellPhoneDB [32]; PDAC clusters 2 and 4, which both comprised cycling cells, were combined for this analysis.

Immunofluorescence staining
Paraffin embedded tumor tissue sections of 4 μm were co-stained with an anti-CD44 polyclonal rabbit antibody (1:150; Sigma-Aldrich, HPA005785) and an anti-HGF monoclonal mouse antibody (1:150; OriGene TA807186) overnight at 4 °C after antigen retrieval at pH 8.5. Donkey anti-mouse Alexa Fluor 555 and donkey anti-rabbit Alexa Fluor 488 conjugated antibodies were incubated at 1:200 for 1 h at room temperature. Slides were mounted using Fluoroshield with DAPI (Sigma-Aldrich). Confocal images were acquired in a Nikon confocal microscope using a 60× oil objective. Image processing was performed in Fiji (ImageJ).

3D co-cultures of matched patient-derived pancreatic tumor organoids and cancer-associated fibroblasts
Five patient-matched pairs of PDAC tumor organoids (PDAC-PDOs) and CAFs were established from surgically resected tumor samples at a success rate of 50%, with individual success rates of 50 and 93% for PDOs and CAFs, respectively. Expansion of PDOs of sufficient purity and biomass for downstream applications was completed approximately between 20 and 90 days, with a median of around 40 days. The clinicopathological characteristics of the patients are described in Table 1.
For all five PDO/CAF pairs, we established direct 3D co-culture models (Fig. 1A). For this, matched organoids and CAFs were combined at a 1:1 ratio and were cultured in a 3D matrix composed of Matrigel and collagen I. Direct physical contact of organoids and CAFs was confirmed by fluorescence staining (Fig. 1B). In line with extensive evidence that CAFs promote proliferation of tumor cells in 3D culture models [34], we observed an enhancing effect of CAFs on PDAC organoid proliferation in 4 out of 5 matched co-cultures (Fig. 1C).

Cancer-associated fibroblasts increase drug resistance of patient-matched pancreatic tumor organoids
To assess the impact of patient-matched CAFs on chemosensitivity of tumor organoids, we used the imagebased drug assay DeathPro [22], which allows to evaluate independently drug-induced cell death and proliferation inhibition (PI, Fig. 1D). Sensitivity of PDOs to three chemotherapeutic drugs frequently used in the clinical treatment of PDAC, i.e., gemcitabine, 5-FU and paclitaxel, was evaluated for PDO mono-and PDO/CAF cocultures. To discriminate CAFs from organoids in the confocal images generated by the DeathPro workflow, CAFs were stained with a cell tracker before setting up the co-cultures (Fig. 1D). By overlaying Hoechst, propidium iodide and the cell tracker signals, individual dose response curves for organoids and CAFs were generated (Fig. 1F).
We observed high reproducibility of the response curves ( Supplementary Fig. S1A) and strong correlations (Pearson R > 0.78) for death (AUC, max. death) and proliferation inhibition (AUCpi, max. PI) parameters between independent replicates ( Supplementary  Fig. S1B). Hoechst and propidium iodide staining had no cytotoxic effects on PDOs or CAFs (Supplementary Fig.  S1C and D). In control conditions without drug treatment, the enhancing effect of CAFs on organoid proliferation resulted in an average of 1.3-fold increase in relative growth of organoids in co-culture compared to monoculture (P < 0.05; Fig. 1C). Basal levels of cell death in organoids under control conditions did not significantly differ between mono-and co-culture ( Supplementary Fig. S1E).
Cell death induced by gemcitabine, 5-FU and paclitaxel treatments was significantly higher in PDO monocultures than in PDO/CAF co-cultures ( Fig. 2A and B). Both AUC and max. Death values were on average significantly lower in co-culture conditions, indicating that the presence of CAFs resulted in increased resistance of organoids to the cytotoxic effects of chemotherapy treatments. The extent of this effect was, however, patient and drug specific, suggestive of heterogeneity in the impact of tumor-stroma interactions on chemoresistance mechanisms. PDOs in co-culture also displayed reduced proliferation inhibition (AUCpi and max. PI) induced by gemcitabine ( Fig. 2C and D). For 5-FU, only a slight but significant decrease in max. PI was observed. No significant difference was found in organoid proliferation inhibition in the presence of CAFs after paclitaxel treatment ( Fig. 2C and D). Worth noting, paclitaxel was less effective at inhibiting organoid proliferation than gemcitabine or 5-FU (Fig. 2C, AUCpi values).
In sum, our results show that CAFs exert a protective effect against the cytotoxic effects of gemcitabine, 5-FU and paclitaxel on PDAC-PDOs, indicating a direct contribution to tumor cell drug resistance. These results further highlight the relevance of heterotypic culture models for personalized in vitro drug testing.

Increased expression of inflammatory pathways in CAFs co-cultured with PDAC organoids
To gain insight into the mechanisms behind the increased chemoresistance of PDAC organoids triggered by CAFs, we performed scRNA-seq to identify transcriptional changes induced in PDO/CAF co-cultures. Three pairs of organoids and patient-matched CAFs (lines 100, 107 and 112) were sequenced after 5 days in 3D mono-and coculture (Fig. 3A). A total of 13.235 PDAC cells and 7.356 CAFs were sequenced and analyzed (Supplementary Fig.  S2A and B). Organoid cells were identified by the expression of the tumor markers KRT18 and KRT19 whereas CAFs were characterized by DCN and LUM expression (Fig. 3B). Patient-specific clustering was observed for organoids and CAF populations, indicating transcriptional heterogeneity between the individual samples (Fig. 3B).
Clustering of all sequenced CAFs distinguished seven transcriptional clusters (Fig. 3C, D and Supplementary Fig. S2C), with distinct functional identities, as indicated by functional enrichment analysis for gene ontology (GO) terms. The largest cluster, cluster 0, was functionally characterized by expression of genes related Fig. 1 Co-culture of PDAC organoids with CAFs and image-based drug testing to de-convolve CAF and organoid responses. A Brightfield images corresponding to the five established direct 3D co-cultures of the patient-matched CAFs and PDAC-PDOs. Scale bar: 250 μm. B Co-culture of matched CAFs and PDAC-PDOs stained for the CAF marker α-SMA (red) and actin cytoskeleton (phalloidin, green). Nuclei were stained with Hoechst (blue). Direct intercellular contact between organoid tumor cells and CAFs can be observed. Scale bar: 50 μm. C Proliferation levels of PDAC organoids measured from culture day 3 to day 8. PDAC organoids in co-culture show significantly higher proliferation levels than in monoculture. Paired t-test, * P < 0.05. D Schematic overview of the established drug test workflow for PDAC-PDO mono-and PDAC-PDO/CAF co-cultures. E Montage of maximum intensity projections as an example of the image data generated from a co-culture model treated with increasing concentration of 5-FU. CAFs (blue) were stained with CellTracker Green CMFDA to be distinguished from PDAC organoids. Hoechst (green) and propidium iodide (red) were used to stain the nuclei of all and dead cells, respectively. to extracellular matrix and epithelial to mesenchymal transition. In co-culture, we observed an increase in the proportion of CAFs from cluster 5 (Fisher's exact test, p < 0.0001; Fig. 3E), which was characterized by a higher expression of genes related to cell cycle, suggesting a potential increase of CAF proliferation in co-culture. This increase was consistent across all three patient samples ( Supplementary Fig. S2C). On the other hand, contribution of cluster 1, related to inflammation, locomotion and TNF-α signaling, decreased in co-culture (Fisher's exact test, p < 0.0001; Fig. 3E), although this was restricted to case 112 at the individual sample level ( Supplementary  Fig. S2C). However, we note that these proportions based on scRNA-seq data may not fully reflect cell numbers in cultures, due to possible differences in dissociation efficiency between CAF subtypes. The two CAF subtypes previously described [20], myCAFs and iCAFs, were identified as CAF subpopulations. Principal component analysis on iCAF and myCAF marker genes [27] for all sequenced CAFs resolved 2 distinct transcriptional clusters (Supplementary Fig. S2D). 88.5% of the CAF population corresponded to myCAFs (6509 cells), whereas 11.5% were classified as iCAFs (847 cells). iCAFs corresponded almost exclusively to the inflammation-related cluster 1 (Fig. 3F). Expression levels of iCAF marker genes were higher in the co-culture setting than in monocultures (Fig. 3G).
We further compared CAFs in mono-and co-culture by gene set enrichment analysis of differentially expressed genes across all cell clusters. Two of the significantly enriched gene sets in co-culture were associated with proliferation (REACTOME_CELL_CYCLE, HALL-MARK_E2F_TARGETS; Fig. 3H), which is in line with the increased proportion of the cell cycle related cluster 5. Remarkably, among the top enriched gene sets in coculture we found hallmark gene sets related to inflammation (TNFα signaling via NFκB, interferon gamma and alpha responses and TGFβ signaling, Fig. 3H and I).
Altogether, these results indicate that co-culture with PDAC organoids induces a pro-inflammatory phenotype in CAFs, which might drive the enhanced chemo-resistance in tumor cells and could be amenable to molecular targeting.

PDAC organoids display increased expression of EMT genes in 3D co-culture with CAFs
Eight transcriptional clusters were identified for all sequenced PDAC-PDO cells (Fig. 4A, B, Supplementary  Fig. S3B and C). The proportion of each cluster differed between organoid lines, reflecting patient-specific transcriptional heterogeneity. For instance, cluster 0, characterized by expression of genes related to adhesion and cell junctions, was the predominant cluster of line 100PO, whereas the immune response related cluster 1  Fig. S3A). No predominant cluster was identified in line 112PO (Supplementary Fig. S3A). Interestingly, we did not observe significant changes in the proliferation score for any of the identified transcriptional clusters of PDAC organoid cells in co-culture compared with monoculture ( Supplementary Fig. S3D).
We further characterized PDO cells in terms of the two major transcriptional subtypes in PDAC, namely classical and basal-like subtypes, by scoring the expression of each subtype gene signatures at the single-cell level. Organoid line 100PO comprised predominantly cells with higher basal-like scores, whereas 107PO and 112PO showed higher classical scores ( Supplementary Fig. S3E). No significant differences in the subtype scores were observed between mono-and co-culture conditions, indicating that the presence of CAFs did not induce a shift in the subtype state of the tumor cells in our model (Supplementary Fig. S3E and F).
To assess the transcriptional changes induced by co-culture with CAFs, we analyzed the differentially expressed genes between PDAC organoids cells in mono-and co-culture conditions (Fig. 4C). Several genes associated with epithelial-to-mesenchymal transition such as COL1A1, COL1A2, FN1, LAMC2 and VIM were upregulated in co-cultured organoid cells. Consistently, Gene Set Enrichment Analysis (GSEA) identified HALLMARK_EPITHELIAL_MESENCHY-MAL_TRANSITION as the most significantly enriched gene set in the co-culture settings (Fig. 4D). Additionally, gene sets associated with inflammation such as HALLMARK_TNFA_SIGNALING_VIA_NFKB, HALL-MARK_IL6_JAK_STAT3_SIGNALING or REACTOME_ INTERFERON_GAMMA_SIGNALING were also found to be enriched in organoid cells co-cultured with CAFs.
Further focusing on the EMT signature of the identified transcriptional clusters of PDAC cells, we found an increased EMT expression score for almost all clusters in co-culture conditions ( Fig. 4E and F), consistent with the result of the GSEA. Expression of several individual EMT-related genes was also upregulated in co-culture (Fig. 4G). Using CellPhoneDB, an open repository of ligand-receptor interactions [32], we were able to detect potential intercellular receptor-ligand interactions between PDO cells (receptor) and CAFs (ligand) (Fig. 5A). In general, most interactions took place between organoid cells in cluster 0 and CAFs. Since cluster 0 was characterized by genes involved in adhesion and cell junctions, this suggests direct intercellular interactions. Additionally, several receptor-ligand interactions related to EMT were identified between PDAC organoid cells and CAFs (Fig. 5B) such as CD44-HGF, CD44-LGALS9 or CD44-FGF2. Most of these EMT-related interactions occurred between CAFs and PDAC organoid cells in cluster 0 as well as clusters 2 and 4 representing cycling PDAC cells (2: S-Phase, 4, G2/M-Phase). Spatial proximity and co-localization of the receptorligand interaction pair CD44-HGF could be confirmed in the original tissues by immunofluorescence staining (Fig. 5C), supporting the in vivo relevance of tumorstroma interactions occurring in our co-culture model.
Taken together, these findings indicate that intercellular interactions between CAFs and epithelial cancer cells are reproduced in our co-culture system, leading to the upregulation of EMT in PDAC organoid cells and increased chemoresistance.

Discussion
Initial studies have demonstrated that PDAC organoids reflect the genetic diversity of the original tumors and that drug sensitivity of organoid lines correlates with clinical response, providing a first indication of their potential for personalized medicine [16,19]. However, the absence of stromal components in these models has not yet allowed to evaluate the influence of stromal activity and signaling on chemosensitivity and precludes screening of drugs targeting stromal elements or tumorstroma interactions.
In this study, we established direct co-cultures of patient-matched PDAC organoids and CAFs and showed that CAFs promoted epithelial cancer cell proliferation as well as resistance of cancer cells to gemcitabine, 5-FU and paclitaxel. The use of the image-based DeathPro assay [22] allowed us to overcome the challenge of differentiating cell type-specific drug responses. Common in vitro viability assays such as MTT or CellTiter Glo are not capable of differentiating drug responses of stromal and tumor cells when directly co-cultured. Hence, in addition to discerning specific tumor cell responses, this system allows monitoring the specific effect of drug treatment on CAFs, providing an excellent tool to study combination therapies targeting tumor and stromal cells.
Previous studies have reported enhanced chemoresistance of primary human PDAC cells in the presence of unmatched stromal cells in 3D tumor spheroid and bioengineered models [35,36]. Similarly, Tsai et al. reported an increase in IC50 after gemcitabine treatment of one patient-derived organoid line in co-culture with CAFs [21]. Our work extends on these observations in a patient-specific manner, highlighting the feasibility and relevance of complex organoid models for ex vivo drug response prediction in the clinical setting. In our scRNAseq data, we observed unsupervised clustering of organoid tumor cells as well as of CAFs by patient of origin; patient-specific heterogeneity is thus preserved in both cell populations. These results accentuate the need and potential to incorporate patient-matched stromal cells into organoids to faithfully model tumor diversity across patients. In our experience, the limiting factor for clinical application of our co-culture model would be the time necessary to reach sufficient purity and biomass of the organoid cultures. Successful establishment and expansion of organoid lines was accomplished in 50% of the cases and downstream applications could be performed after a median of 40 days. Considering that adjuvant therapy is commonly started between 6 and 12 weeks after resection, drug response prediction based on our model  [37] from minimal samples such as fine-needle biopsies have been previously reported, supporting a potential clinical application in the neoadjuvant or palliative settings. Further optimization and patient-specific tailoring of organoid culture conditions to accelerate organoid expansion could be therefore highly beneficial. Recently, Harryvan et al. [37] described a pancreatic mini-tumor model established from patientderived organoids and CAFs, which was able to model the characteristic desmoplastic reaction observed in PDAC tumors. In their study, CAF activation and proliferation was induced by activation of transforming growth factor and platelet-derived growth factor β signaling pathways via modulation of culture media components, supporting the potential of fine-tuning PDO/CAF co-culture conditions to model even more closely tumor drug response. Our scRNA-seq analyses also revealed transcriptional changes towards an EMT phenotype in co-cultured organoid cells, consistent with the results reported by Ligorio and colleagues for 2D co-cultures of patientderived PDAC cell lines derived from metastatic ascites and unmatched CAFs [40]. This observation is also in line with previous reports of EMT induction in PDAC cell lines after indirect co-culture with pancreatic stellate cells (PSCs) cells [41] or exposure to PSC-condition medium [42]. In vivo, orthotopically grafted human organoids surrounded by a stromal microenvironment evidenced upregulation of EMT signaling relative to organoids engrafted intraductally, emphasizing that induction of EMT in tumor cells by the stromal microenvironment is a relevant process in vivo [43]. While the role of EMT in driving PDAC chemoresistance is widely recognized [44,45], our results indicate CAF-mediated induction of EMT in organoids as one of the mechanisms contributing to their chemo-protectant effect in our coculture system.
Despite changes in EMT, we did not observe shifts in the transcriptional subtype signatures of the tumor cells under our culture conditions. Although proliferation of organoids was enhanced in co-cultures, proliferation scores derived from the scRNA-seq analyses were not significantly different between mono-and co-cultures. Similarly, the increment in the proportion of the CAF cluster expressing cell cycle related genes observed in co-cultures was not accompanied by enhanced CAF proliferation as determined by Death-Pro. It is worth noting that in our scRNA-seq data we identified considerably lower numbers of CAFs in coculture compared to CAF monoculture samples. It is thus conceivable that the observed changes in CAF cluster proportions might be due to differential loss of CAF cells in the co-culture setting during sample processing due to, for example, preferential survival of cells with rounded shape or exclusion of CAF-tumor cell doublets. A challenging recovery of CAFs has also been reported for primary pancreatic tumor samples [27], raising awareness that single cell analyses of CAFs might be generally affected by technical biases during dissociation.
Depletion of tumor stroma results in more aggressive tumors and reduced survival in preclinical models [46,47] and stroma-depleting agents have shown no benefit in clinical trials [48]. Targeting of specific tumorstromal cells interactions, in contrast, has become a promising therapeutic approach for PDAC [49]. Results from our scRNA-seq analyses identified several potential interactions involving the cancer stem cell marker CD44 with ligands secreted by CAFs such as HGF, HBEGF, FGF2 and LGALS9. We could also show colocalization of CD44 and HGF in the parental tumor tissues, validating the relevance of predicted interactions from our co-culture model in the in vivo context. CD44 signaling has also been shown to promote stemness via a SPP1-CD44 axis [50] and high expression of CD44 has been associated with poor prognosis in PDAC [51]. Evidence from patient-derived xenografts shows that CD44 + cells are the source of PDAC relapse after gemcitabine treatment, rendering CD44 a promising therapeutic target against recurrent disease [52].

Conclusions
Altogether, our study demonstrates increased chemoresistance of PDAC organoids in co-culture with patientmatched CAFs, emphasizing the relevance of complex co-culture models for personalized medicine applications. This also opens the possibility to investigate efficacy and mode of action for drugs targeting the tumor microenvironment in a patient-specific manner. Hence, our work provides evidence of the feasibility of modeling patient-specific tumor-stroma interactions for target discovery and drug testing in a high-throughput amenable 3D culture setting.