Cell of Origin Influences Pancreatic Cancer Subtype

Pancreatic ductal adenocarcinoma (PDAC) is a deadly cancer that is projected to be the second leading cause of cancer-related deaths in the United States by 2030 (1). The 5-year survival rate for patients with PDAC is a mere 9% (2) and is attributable to late-stage diagnosis—when patients are rarely eligible for surgical resection—and therapeutic strategies being largely ineffective (3). A better understanding of how PDAC arises is vital for improving both early detection and treatment.

The genetics underlying PDAC development are well-described. Human PDAC genome sequencing has identified several recurrent genomic alterations in pancreatic tumors, including activating mutations in KRAS in >90% of PDACs (4, 5) and mutations in the TP53, CDKN2A, and SMAD4 tumor-suppressor genes in 72%, 30%, and 32% of cases, respectively (6). Analysis of human pancreas samples through DNA and IHC analyses has led to a proposed progression model for PDAC, with oncogenic KRAS mutations serving as the initiating event and specific tumor-suppressor gene mutations occurring at defined stages thereafter to promote cancer progression (7). In this model, cancers arise when oncogenic KRAS promotes the formation of preinvasive pancreatic intraepithelial neoplasias (PanIN) that progress through increasingly more dysplastic stages, culminating in frank carcinomas and metastasis. Studies in genetically engineered mouse models in which oncogenic KRAS is expressed and tumor-suppressor genes are inactivated in the pancreas have been instrumental for demonstrating the importance of these combined mutations in driving PDAC development (8–12). Importantly, these models faithfully reproduce the entire progression of human PDAC, from neoplastic lesions to metastatic cancer.

As a complementary approach to genomic analyses, transcriptomic analyses have been utilized to define molecular subtypes of human PDAC, with the goal of identifying distinct classes that could guide clinical decision-making. Three seminal studies established a framework for PDAC classification. Depending on the methodology, transcriptomics analysis of resected PDAC samples from untreated patients revealed between two and four distinct molecular subtypes (13–15). Although differences in sample cellularity and technology used for transcriptomic analysis may underlie the differences in specific subtypes defined by each study, there is a consensus from these initial studies as well as subsequent ones that two broad pancreatic cancer subtypes exist: the classical and the basal-like subtypes, with the latter being associated with poorer prognosis (6, 13–20). The cellular and molecular factors that drive each molecular subtype, however, remain unknown. Unequivocally establishing the factors critical for dictating the subtypes will refine our understanding of the underlying biology of these classes and ultimately improve risk stratification and therapeutic development. Such an understanding, however, requires tractable genetic experiments in model systems to allow interrogation of such factors as specific gene mutations and cell type of origin on the development of different subtypes.

Correlative analyses have suggested particular genetic associations with specific subtypes, such as mutations in the TP53 gene, which encodes the p53 transcription factor, with the basal-like subtype (6, 13, 17, 19, 20). TP53 commonly sustains missense mutations in the DNA-binding domain that not only induce loss of wild-type p53 function, but may also promote gain-of-function properties through the production of a mutant p53 protein (21). Missense mutations comprise two major classes: contact mutations that alter residues required by p53 to contact DNA and structural mutations that affect the three-dimensional structure of the protein (22). The most frequently observed TP53 mutations in PDAC are at codons 175 and 273 (corresponding to mouse codons 172 and 270, respectively), which represent structural and contact residue mutations (22, 23). Although the majority of mouse PDAC studies have relied on combined oncogenic KRAS and p53^R172H expression to drive metastatic PDAC, oncogenic KRAS and Trp53 deletion also promote metastatic PDAC (9, 12, 24). How Trp53 deletion and different classes of p53 point mutants affect the development of different subtypes of PDAC, however, remains unexplored.

It also remains unclear how cell of origin influences pancreatic cancer subtype. Acinar and ductal cells are the major epithelial cell types in the exocrine pancreas, but which of these serves as the cell of origin in PDAC has been debated. Early studies in mouse PDAC models failed to reveal the cell of origin because they relied on strains in which Cre recombinase is expressed in multipotent pancreas progenitor cells during embryogenesis (8). Although PanINs resemble ducts and were originally thought to arise from normal ductal cells, more recent evidence using tamoxifen-regulatable Cre to induce genetic mutations in adult mice has suggested that acinar cells can give rise to PanINs and PDAC through a cellular reprogramming process termed acinar-to-ductal metaplasia (ADM; refs. 25–33). Some evidence from adult mouse PDAC models also supports the notion that PDAC can arise from ductal cells, in the context of oncogenic KRAS expression and homozygous Trp53 or Fbw7 deletion (26, 29, 32). Thus, human PDAC likely originates from pancreatic acinar or ductal cells.

Here, we seek to understand the factors that influence the development of different transcriptional subtypes of PDAC using a comprehensive panel of tractable genetically engineered mouse models. To better recapitulate the process of PDAC development in humans, we have investigated PDAC development in adult mice, from pancreatic acinar or ductal cells, and in the contexts of intact p53, p53 inactivation, or expression of the p53 mutants p53^R172H or p53^R270H. We have performed transcriptomic analysis of tumors arising from acinar and ductal cells and then utilized these profiles to illuminate a connection between cell of origin and human PDAC subtype. Our studies reveal that different cells of origin can influence the ultimate molecular subtype of PDAC, providing critical new insight into the basis for intertumoral PDAC heterogeneity.

To deconstruct the genetic requirements for pancreatic cancer development from adult mouse pancreatic acinar cells, we established a system to activate oncogenic Kras^G12D and modulate Trp53 in mouse pancreatic acinar cells through the use of a tamoxifen-inducible knock-in Ptf1a^CreER allele. Importantly, to induce genetic alterations in cells of the fully developed adult pancreas, mice were aged to adulthood (8–10 weeks) before tamoxifen treatment (Fig. 1A). Tamoxifen treatment induced efficient genomic recombination in the majority of acinar cells, but not other pancreatic cell types, as indicated by the selective expression of the tdTomato reporter allele in amylase-positive acinar cells (Fig. 1B). To initially assess tumor development in the presence and absence of p53, we generated tamoxifen-treated, aged, and examined cohorts of Kras^LSL-G12D/+;Rosa26^{LSL-tdTomato/LSL-tdTomato};Ptf1a^CreER;Trp53^+/+ (Kras^LSL-G12D/+;Rosa26^{LSL-tdTomato/LSL-tdTomato} hereafter denoted as KT), KT;Ptf1a^CreER;Trp53^fl/+, and KT;Ptf1a^CreER;Trp53^fl/fl mice for pancreatic cancer–free survival upon morbidity.

The Kaplan–Meier analysis revealed first that loss of p53 in KT;Ptf1a^CreER;Trp53^fl/fl mice led to rapid, fully penetrant development of PDAC, with a median pancreatic cancer–free and overall survival of 137 days after tamoxifen treatment (Fig. 1C; Supplementary Fig. S1A; Supplementary Table S1). As expected, KT;Ptf1a^CreER;Trp53^fl/+ mice with heterozygous expression of wild-type Trp53 displayed a longer tumor latency with a median pancreatic cancer–free survival of 276.5 days after tamoxifen treatment. Interestingly, KT;Ptf1a^CreER;Trp53^+/+ mice also succumbed to pancreatic cancer, albeit with a significantly longer latency of 670 days. Approximately half of this latter cohort developed pancreatic cancer, with the rest displaying widespread premalignant PanIN lesions at morbidity. Of note, male KT;Ptf1a^CreER;Trp53^+/+ mice displayed significantly shorter pancreatic cancer–free and overall survival than female mice, in contrast to KT;Ptf1a^CreER;Trp53^fl/fl mice, where survival was similar between sexes (Supplementary Fig. S1B and S1C). At morbidity, some mice in all cohorts exhibited common human clinical symptoms of PDAC such as ascites, jaundice, and bowel obstruction (Fig. 1D; Supplementary Fig. S1D).

Histopathologic analysis provided additional insight into PDAC development. We assigned a primary and secondary histologic grade for each tumor (Supplementary Table S1) and found that KT;Ptf1a^CreER;Trp53^fl/+ and KT;Ptf1a^CreER;Trp53^fl/fl mice consistently developed moderately/well-differentiated or poorly differentiated adenocarcinomas characterized by Cytokeratin 19 (CK19) positivity (Fig. 1E and F). Of tumor-bearing KT;Ptf1a^CreER;Trp53^+/+ mice, 44% developed moderately/well-differentiated adenocarcinomas, 22% developed poorly differentiated adenocarcinomas, and 33% developed sarcomatoid carcinomas, a statistically significant shift in spectrum relative to KT;Ptf1a^CreER;Trp53^fl/fl mice. These tumors were either CK19-positive adenocarcinomas or sarcomatoid carcinomas with a notable positivity for vimentin and absence of CK19 immunostaining (Fig. 1F; Supplementary Fig. S1E). The majority of tumors in mice of all genotypes had tumor-adjacent PanIN lesions, as indicated by Alcian blue positivity (Fig. 1F; Supplementary Table S1). Interestingly, one KT;Ptf1a^CreER;Trp53^+/+ mouse had evidence of a rare intraductal papillary mucinous neoplasm (IPMN; Supplementary Fig. S1F). Lineage tracing using the tdTomato allele verified that all PanINs, IPMNs, tumors, and metastases originated from Ptf1a-expressing cells, suggesting an acinar cell origin (Fig. 1F and G; Supplementary Fig. S1F–S1H). In addition, metastatic lesions were observed in all cohorts of mice at both the gross and histologic levels. Liver metastases were the most common, with rare peritoneum, diaphragm, and lung metastases observed upon dissection in some animals (Fig. 1G; Supplementary Fig. S1H). Quantification of liver metastases by histologic analysis revealed similar rates of metastasis, with 33% and 32% of tumors metastasizing in p53-expressing and p53-deficient mice, respectively (Supplementary Table S1). Collectively, these findings demonstrate that oncogenic KRAS expression can drive pancreatic cancer development from adult mouse acinar cells. Although PDAC can develop in mice initially expressing wild-type p53, the latency of tumor development is significantly reduced with targeted p53 inactivation, underscoring the importance of p53 as a barrier to pancreatic cancer development in adults.

To assess how the expression of mutant p53, given its potential gain-of-function effects, compares with complete p53 deficiency in adult acinar cell–derived pancreatic cancer development, we next generated mouse cohorts with oncogenic KRAS^G12D expression and expression of the p53 structural mutant p53^R172H (KT;Ptf1a^CreER;Trp53^fl/LSL-R172H), the p53 contact mutant p53^R270H (KT;Ptf1a^CreER;Trp53^fl/LSL-R270H), or p53 deficiency (KT;Ptf1a^CreER;Trp53^fl/–). We generated hemizygous cohorts in which expression of a Trp53 point mutant allele and a Trp53 null allele in the pancreas allows a clear assessment of potential gain-of-function activity rather than dominant-negative activity observed when the mutants are combined with a wild-type Trp53 allele. Importantly, by using a combination of a Trp53-floxed allele and conditional Lox-stop-lox activatable alleles for the Trp53 mutants (which are Trp53 null until Cre is active), we generated mice with a uniform Trp53 heterozygous genetic background in stromal tissues, for a controlled comparison of pancreatic cancer–free survival across all genotypes (Fig. 2A).

Analysis of these cohorts revealed that KT;Ptf1a^CreER;Trp53^fl/− mice deficient for p53 rapidly developed cancer with a median pancreatic cancer–free survival of 123 days after tamoxifen treatment (Fig. 2B; Supplementary Table S2). This latency is similar to that seen in KT;Ptf1a^CreER;Trp53^fl/fl mice, suggesting that the Trp53 heterozygous stroma does not significantly change tumor latency (Supplementary Fig. S2A). Furthermore, KT;Ptf1a^CreER;Trp53^fl/LSL-R270H mice expressing the contact mutant p53^R270H had a similar pancreatic cancer–free survival of 123.5 days to KT;Ptf1a^CreER;Trp53^fl/− mice (Fig. 2B), indicating that there is no significant gain-of-function effect driven by p53^R270H expression in terms of tumor latency. Interestingly, analysis of the p53^R172H structural mutant revealed a slightly enhanced pancreatic cancer–free survival of 144 days in the KT;Ptf1a^CreER;Trp53^fl/LSL-R172H mice relative to KT;Ptf1a^CreER;Trp53^fl/− mice with Trp53 deficiency, again indicating no clear gain-of-function effects with this mutant in terms of tumor latency (Fig. 2B). As with KT;Ptf1a^CreER;Trp53^fl/fl mice, there was no significant difference in pancreatic cancer–free survival of female and male mice in each cohort (Supplementary Fig. S2B). Some mice in all cohorts exhibited common human clinical symptoms of PDAC such as ascites, jaundice, and bowel obstruction at morbidity at roughly the same frequencies (Fig. 2C; Supplementary Fig. S2C). Histopathologic analysis revealed that p53 deficiency or expression of either p53 mutant consistently led to the development of adenocarcinomas, with 68% to 78% of tumors manifesting a moderately/well-differentiated primary tumor grade (Fig. 2D; Supplementary Table S2). These tumors were characterized by CK19 positivity and frequently had tumor-adjacent PanINs (Fig. 2E; Supplementary Table S2). Histologic analysis of livers, the most common site of metastasis, revealed similar rates of metastasis across cohorts (Fig. 2D and F; Supplementary Fig. S2D and S2E). Each cohort also had evidence of rare metastases to the peritoneum, diaphragm, and lung observed upon dissection (Supplementary Fig. S2F). Lineage tracing using the tdTomato allele confirmed that all tumors and metastases originated from Ptf1a-expressing cells, suggesting an acinar cell origin (Fig. 2E and F; Supplementary Fig. S2D–S2F).

That mutant p53 expression did not obviously change metastatic rates suggests that there is no clear gain-of-function effect of p53 point mutant expression in enhancing metastasis. Accordingly, approximately 25% of KT;Ptf1a^CreER;Trp53^fl/LSL-R172H and KT;Ptf1a^CreER;Trp53^fl/LSL-R270H mice had at least one primary tumor negative for mutant p53 expression, often accompanied by a metastasis negative for p53 expression (Fig. 2G). These observations suggest that there is not an absolute selection for cells expressing mutant p53. Collectively, these findings demonstrate that, like p53 deficiency, mutant p53 promotes PDAC development, but does not exhibit a clear gain-of-function effect in terms of tumor latency and rates of metastasis in acinar cell–derived pancreatic cancer development.

We next sought to define the genetic requirements for PDAC development from adult pancreatic ductal cells. Toward this end, we generated mouse cohorts in which we activated oncogenic Kras^G12D and modulated Trp53 using the tamoxifen-inducible Sox9CreER transgene (Fig. 3A). Tamoxifen treatment of adult mice resulted in efficient genomic recombination in ductal cells, as indicated by expression of the tdTomato reporter allele in CK19-positive ductal cells (Fig. 3B). We first assessed tumor development in mice expressing or lacking p53 by generating, tamoxifen treating, aging, and examining cohorts of KT;Sox9CreER;Trp53^fl/fl and KT;Sox9CreER;Trp53^+/+ mice for pancreatic cancer–free survival upon morbidity.

The Kaplan–Meier analysis revealed that 32% of KT;Sox9CreER;Trp53^fl/fl mice developed pancreatic cancer by approximately 300 days after tamoxifen treatment (Fig. 3C; Supplementary Table S3). In contrast, none of the 18 KT;Sox9CreER;Trp53^+/+ mice developed pancreatic cancer, suggesting that Trp53 loss is necessary for ductal cell–derived pancreatic cancer development. However, the ductal model is complicated by the expression of Sox9 in adult tissues beyond the pancreas, leading to the development of nonpancreatic tumors, such as Harderian gland adenocarcinomas, that necessitate early sacrifice, with a median overall survival of 415 days in KT;Sox9CreER;Trp53^+/+ mice and 253 days in KT;Sox9CreER;Trp53^fl/fl mice (Fig. 3D). Pancreata were therefore analyzed when mice reached morbidity due to either pancreatic or nonpancreatic tumors. At morbidity, some KT;Sox9CreER;Trp53^fl/fl mice succumbing to PDAC exhibited common human clinical symptoms of PDAC such as ascites and bowel obstruction (Fig. 3E). We observed both moderately/well-differentiated and poorly differentiated adenocarcinomas in pancreatic tumor–bearing KT;Sox9CreER;Trp53^fl/fl mice, at similar frequencies to KT;Ptf1a^CreER;Trp53^fl/fl mice (44% vs. 39% poorly differentiated tumors, respectively; Figs. 1E and 3F). The adenocarcinomas were characterized by CK19 positivity, and importantly, lineage tracing with tdTomato suggested a Sox9-expressing pancreatic cell of origin for these tumors (Fig. 3G; Supplementary Fig. S3A). Histologic analysis and Alcian blue staining failed to reveal PanINs associated with tumors in KT;Sox9CreER;Trp53^fl/fl mice (Fig. 3G; Supplementary Table S3). In terms of establishing the incidence of metastasis, neither tumor histology nor tdTomato positivity could be used to distinguish PDAC metastases from primary cholangiocarcinomas resulting from Sox9CreER activity in the liver where it can promote neoplasia. However, visual inspection upon dissection, coupled with histologic analysis, identified small clusters of tdTomato-positive malignant cells in the peritoneum and lungs of mice with clear PDACs, suggesting that metastasis had occurred (Fig. 3H).

Next, to assess the gain-of-function phenotypes of mutant p53 expression in adult ductal cell–derived pancreatic cancer development, we generated mouse cohorts with oncogenic KRAS^G12D expression and expression of the structural mutant p53^R172H (KT;Sox9CreER;Trp53^fl/LSL-R172H), expression of the contact mutant p53^R270H (KT;Sox9CreER;Trp53^fl/LSL-R270H), or conditional Trp53 knockout (KT;Sox9CreER;Trp53^fl/−). As with the acinar model, these mice were generated to create uniform Trp53 heterozygous stroma. Mice in all three cohorts displayed similar pancreatic cancer–free and overall survival (Supplementary Fig. S3B and S3C) and exhibited clinical symptoms of pancreatic cancer development (Supplementary Fig. S3D). However, it was unclear if there was a gain-of-function phenotype, as the incidence of pancreatic cancer development was low. Histopathologic analysis revealed the development of poorly to moderately/well-differentiated adenocarcinomas in all cohorts and a sarcomatoid carcinoma in one KT;Sox9CreER;Trp53^fl/− mouse (Supplementary Fig. S3E and S3F). Again, the adenocarcinomas were characterized by CK19 expression and tdTomato expression, confirming a Sox9-expressing pancreatic cell origin for these tumors (Supplementary Fig. S3F).

To investigate a potential precursor lesion for the ductal cell–derived tumors, we examined the pancreata of mice prior to cancer development. Analysis of KT;Sox9CreER;Trp53^fl/fl mice at approximately 120 days—the midpoint of median overall survival in the ductal model—consistently failed to reveal evidence of potential precursor lesions, including ADMs or PanINs (Fig. 4A). In stark contrast, KT;Ptf1a^CreER;Trp53^fl/fl mice showed early evidence of ADM and PanIN development at approximately 70 days, the midpoint of median overall survival in the acinar model (Fig. 4B). When ductal cell–derived tumors developed, tumors were often surrounded by uninvolved pancreas (Fig. 4C), whereas acinar cell–derived tumors were typically surrounded by ADMs and PanINs (Fig. 4D). Notably, some ducts in the pancreata of KT;Sox9CreER;Trp53^fl/fl mice displayed phospho-ERK1/2 staining, as did ADMs and PanINs of KT;Ptf1a^CreER;Trp53^fl/fl mice, indicative of KRAS activation in these cells (Fig. 4E). This observation, coupled with the lack of other precursor lesions, suggests that ductal cell–derived tumors might originate directly from ducts. Collectively, these findings demonstrate that oncogenic KRAS expression and Trp53 inactivation or mutation can drive pancreatic cancer development directly from adult mouse ductal cells, highlighting a role for p53 in blocking pancreatic cancer development from ductal cells in adult mice.

Our studies have demonstrated that pancreatic cancer can originate from either PTF1A- or SOX9-expressing pancreatic cells, which we posit represent acinar and ductal cells, respectively. To gain insight into whether the cell of origin affects the molecular subtype of PDAC, we performed transcriptomic analysis of bulk acinar cell–derived and ductal cell–derived tumors by RNA sequencing (RNA-seq; Supplementary Fig. S4A). We focused on tumors with Trp53 deletion, rather than Trp53 point mutation, given the robust phenotypes observed with p53 deficiency in both tumor models. Notably, the set of acinar cell–derived and ductal cell–derived tumors we analyzed had similar differentiation profiles (Supplementary Fig. S4A and S4B). Principal component analysis of transcriptomes revealed that acinar cell–derived and ductal cell–derived tumors cluster distinctly, suggestive of different molecular profiles (Fig. 5A). Indeed, analysis of differentially expressed genes by DESeq2 revealed that 1,075 genes are more highly expressed in acinar cell–derived tumors than in ductal cell–derived tumors and that 417 genes are more highly expressed in ductal cell–derived tumors than in acinar cell–derived tumors (log₂ fold change > 1.0, P-adjusted value < 0.05; Fig. 5B; Supplementary Table S4). IHC analysis confirmed the differential expression of proteins encoded by several of these genes, including CLDN18, TFF1, and MUC5AC, all of which showed higher expression in acinar cell–derived tumors than in ductal cell–derived tumors (Fig. 5C).

We sought to understand the significance of the gene-expression profiles specific to acinar cell–derived and ductal cell–derived tumors. We first asked whether each tumor type resembled the cell of origin. Comparison of the acinar cell–derived and ductal cell–derived tumor signatures with single-cell RNA-seq data for normal pancreatic acinar and ductal cells (34) failed to reveal a significant likeness of tumors to the cell of origin (Fig. 5D; Supplementary Fig. S4C). We next used functional annotation to reveal pathways specific to tumors derived from each cell of origin (Fig. 5E and F). Gene Ontology (GO) analysis of the top pathways enriched in acinar cell–derived tumors revealed categories linked to extracellular matrix (ECM) organization, cell adhesion, and digestive system development (Fig. 5E). Consistent with the enhanced ECM organization, Masson's trichrome staining revealed increased collagen staining in acinar cell–derived tumors compared with ductal cell–derived tumors (Fig. 5G). In contrast to the acinar cell–derived tumors, the ductal cell–derived tumors displayed enrichment for GO terms such as ribosome function and glycolysis (Fig. 5F). These findings demonstrate that the molecular profile of pancreatic tumors is dependent on the cell of origin.

As described above, human PDAC has been classified into discrete molecular subtypes by transcriptomic analyses (13–15). We hypothesized that the emergence of these tumor subtypes could be influenced by cell of origin. We therefore sought to determine whether the acinar cell– and ductal cell–derived tumor signatures are enriched in different molecular subtypes of human pancreatic cancer. We defined acinar cell–derived and ductal cell–derived tumor signatures based on the highest-confidence differentially expressed mouse genes, using a log₂ fold change > 1.5 and P-adjusted value < 0.01. We thus identified 640 genes as differentially expressed between acinar cell–derived and ductal cell–derived tumors, 573 of which had human orthologs (Fig. 6A). Of these, 496 genes were more highly expressed in acinar cell–derived tumors than in ductal cell–derived tumors, and 77 genes were more highly expressed in ductal cell–derived tumors than in acinar cell–derived tumors.

To investigate whether these defined signatures were associated with particular human PDAC subtypes, we interrogated whether there is an enrichment of either the acinar cell–derived tumor signature or ductal cell–derived tumor signature in human PDACs classified using published datasets. We began with the four molecular subtype classification scheme—squamous, pancreatic progenitor, immunogenic, and abnormally differentiated endocrine exocrine (ADEX)—defined through analysis of the Australian International Cancer Genome Initiative (ICGC) patient cohort using RNA-seq of untreated bulk primary PDACs of high cellularity (13). Notably, generation of this dataset through RNA-seq analysis of bulk primary tumors mirrors the input material and methodology of our mouse gene-expression analysis. Strikingly, gene set variation analysis (GSVA) of these data revealed a significant enrichment of the ductal cell–derived tumor signature in tumors of the squamous subtype (Fig. 6B and C). In contrast, the acinar cell–derived tumor signature was significantly enriched in the pancreatic progenitor, ADEX, and immunogenic subtypes—which are now thought to be within a single broad subtype known as classical—relative to the squamous subtype (ref. 16; Fig. 6B and D). To confirm subtype enrichment of the cell-of-origin signatures in an independent cohort, we leveraged an RNA-seq dataset of bulk primary PDACs from The Cancer Genome Atlas (TCGA) Research Network (6), classified according to the four-subtype classification scheme. Again, GSVA uncovered a significant enrichment of the ductal cell–derived tumor signature in the squamous subtype relative to the pancreatic progenitor, immunogenic, and ADEX subtypes (Supplementary Fig. S5A and S5B). In contrast, the acinar cell–derived tumor signature was significantly enriched in the pancreatic progenitor, ADEX, and immunogenic subtypes relative to the squamous subtype (Supplementary Fig. S5A and S5C).

We then compared the patterns of enrichment of our signatures in the Australian ICGC PDAC patient cohort processed using the other major classification schemes (14, 15). The ICGC dataset classified using the classical, quasi-mesenchymal, and exocrine-like molecular subtype designations, which were derived from microarray analysis of microdissected epithelium of untreated, primary PDACs (15), revealed a significant enrichment of the ductal cell–derived tumor signature in tumors classified as the quasi-mesenchymal subtype relative to the other subtypes (Fig. 6E and F). In contrast, the acinar cell–derived tumor signature was significantly enriched in the classical and exocrine-like subtypes relative to the quasi-mesenchymal subtype (Fig. 6E and G). The Australian ICGC PDAC patient data was then classified using the basal-like and classical designations originally derived from analysis of bulk, resected untreated primary PDACs and metastases by both microarrays and RNA-seq, followed by the exclusion of transcripts expressed in the normal pancreas (14). The ductal cell–derived tumor signature was significantly enriched in basal-like tumors (Fig. 6H and I), and the acinar cell–derived tumor signature was significantly enriched in classical tumors (Fig. 6H and J). Subsequent studies have led to the consensus that there are two broad subtypes—the basal-like subtype (including the squamous or quasi-mesenchymal subtypes) and classical subtype (including the pancreatic progenitor, immunogenic, ADEX, and exocrine-like subtypes; refs. 6, 16, 18), and our findings suggest that the acinar cell–derived tumor signature is associated with the classical subtype and the ductal cell–derived tumor signature correlates with the basal-like subtype.

We next evaluated the clinical significance of our cell-of-origin signatures by examining associations with overall survival of patients with PDAC in the Australian ICGC cohort. Interestingly, patients whose tumors showed low expression of the acinar cell–derived signature had significantly worse overall survival than those with high expression of the acinar signature (Fig. 7A). In contrast, patients whose tumors showed high expression of the ductal cell–derived signature had significantly worse overall survival than those with low expression of the ductal signature (Fig. 7B). Cox proportional hazard analysis demonstrated that high expression of the ductal cell–derived signature remained significantly associated with decreased overall survival in a model that incorporated key clinical covariates such as age, tumor stage and grade, and subtype classification, whereas high expression of the acinar cell–derived signature was independently associated with improved overall survival (Fig. 7C). These independent survival associations were confirmed in a meta-analysis of three cohorts of patients with PDAC, including the Australian and Canadian ICGC cohorts and the TCGA cohort (Fig. 7D). Similar to the Australian ICGC cohort, the Canadian ICGC cohort again demonstrated significant associations of the acinar cell–derived signature with improved survival and of the ductal cell–derived signature with decreased survival (Supplementary Fig. S6A). It is unclear why the TCGA cohort alone does not reflect the same outcome associations as the ICGC cohorts (Supplementary Fig. S6B), but this could be related to differences in epithelial purity of these samples (6, 13). Collectively, our findings indicate that the acinar cell–derived and ductal cell–derived tumor signatures are independently associated with survival of patients with human PDAC. Moreover, these findings demonstrate that the very earliest events in carcinogenesis can impart on tumors a fundamental program that drives cancer subtype and phenotype.

Here, we use genetically engineered mouse models to better understand the cellular and molecular origins of different subtypes of pancreatic cancer. To best recapitulate human cancer development, we induced mutations in different cellular compartments of the pancreas in adult mice and aged large cohorts until morbidity. Our aging studies revealed first that oncogenic KRAS expression, coupled with Trp53 deletion, in adult acinar cells efficiently drives metastatic PDAC, supporting the idea that p53 imposes an important barrier to PDAC development from acinar cells. Interestingly, by aging our acinar cell–derived tumor cohorts extensively, we observed that oncogenic KRAS expression can also drive metastatic PDAC development in the context of wild-type p53. In addition, KRAS^G12D expression in acinar cells, coupled with Trp53 point mutant alleles, promotes tumorigenesis. Upon interrogating adult ductal cells as a cell of origin, we found that expression of activated KRAS can promote PDAC, but only in the context of Trp53 inactivation or mutation, and that tumor latency is significantly longer than that for acinar cell–derived PDAC development. Transcriptomic analysis of tumors arising from each compartment upon KRAS activation and Trp53 inactivation revealed that the ductal cell–derived tumor signature is enriched in the basal-like subtype, whereas the acinar cell–derived tumor signature is enriched in the classical subtype. Our studies thus illuminate cell of origin as one factor that determines subtypes of PDAC.

The cell of origin of human pancreatic cancer has been controversial. Historically, pancreatic ductal cells were presumed to be the PDAC cell of origin based on the histologic similarity of PDACs to normal pancreatic ducts (35, 36). However, this notion was revised when experiments in mouse models demonstrated that KRAS expression and Trp53 mutation in acinar cells can trigger PDAC, suggesting that PDAC can originate from acinar cells that undergo ADM, which then become PanINs and ultimately PDAC (25–27, 30–32). Subsequent studies of ductal cells—one using 3- to 4-week-old Sox9CreER;Kras^LSL-G12D;Trp53^fl/fl mice and the other using 7- to 9-week-old Hnf1b^CreER;Kras^LSL-G12D;Trp53^{LSL-R172H/LSL-R172H};R26^mTmG mice—suggested that ductal cells can also give rise to PDAC in specific contexts (26, 29). Whereas the former model showed PDAC development accompanied by high-grade PanINs, possibly due to the age of tumor initiation when the pancreas was still rapidly growing (37), our studies in adult mice and the latter model demonstrated PDAC without clear evidence of PanINs, suggesting another trajectory for these tumors. This path contrasts dramatically with that of acinar cell–derived tumors, which are associated with ADM and PanINs, the presumed precursor for PDACs. Alternatively, it is possible that PanINs arise in the ductal model but universally progress into tumors, thus obliterating evidence of the initial lesion. Indeed, earlier studies observed the development of PanIN-like lesions after orthotopic transplantation of genetically modified primary human ductal cells (38). We did not, however, observe PanINs, hyperplasia, or dysplasia in numerous mice, even when pancreata were collected at earlier timepoints. The phospho-ERK1/2–positive ducts seen in these mice, in conjunction with the morphology of early tumors, suggest that tumors may therefore arise from ductal cells. Although the tumors in the Sox9CreER mice most likely originated from ductal cells, it is also formally possible that they originated from other SOX9-expressing cells in the pancreas such as centroacinar cells or rare SOX9-expressing acinar cells (39–42). Regardless of the exact identity of the SOX9-expressing cell that gives rise to PDAC, our results clearly suggest that different cells of origin targeted by Sox9CreER and Ptf1a^CreER give rise to fundamentally different pancreatic cancers.

Interestingly, pancreatic tumor histology was significantly influenced by p53 status. We found that Trp53 status led to the development of primarily poor to moderately/well-differentiated adenocarcinomas, similar to reports in Kras^LSL-G12D/+;Pdx1-Cre;Trp53^LSL-R172H/+ mice (9, 12). In contrast, KT;Ptf1a^CreER; Trp53^+/+ mice developed both adenocarcinomas and sarcomatoid carcinomas tumors. This latter observation is reminiscent of previous reports of sarcomatoid tumors developing in Kras^LSL-G12D/+;Pdx1-Cre and Kras^LSL-G12D/+;Pdx1-Cre;p16^–/− mice, cohorts with intact p53 (12). Notably, an IPMN, which has been associated with Smad4 deficiency (11, 43), was also observed in a KT;Ptf1a^CreER;Trp53^+/+ mouse. Future studies will delineate how loss of different tumor-suppressor genes affects the histologic subtype of pancreatic cancer arising from different cells of origin.

Notably, we did not observe a gain-of-function effect of mutant p53 expression, with either the p53^R270H contact mutant or the p53^R172H structural mutant, in terms of tumor latency and rates of metastasis. Previous characterization of these point mutants in mouse cancer models has similarly shown that tumor latency is not altered by expression of the point mutants, although gain-of-function phenotypes relating to the appearance of new tumor types and enhanced rates of metastasis were reported (44–46). We did not observe enhanced metastasis with expression of the point mutants relative to p53 deficiency, which suggests some context dependency for this phenotype. Indeed, this behavior is not unprecedented, as expression of p53^R172H in a hepatocellular carcinoma model also did not enhance development and progression (47). Interestingly, we noted that acinar cell–derived PDAC was in fact significantly delayed in the presence of p53^R172H, an observation potentially explained by the ability of this protein to fold into the proper conformation in certain settings (48). Overall, at the level of resolution of our studies, it appears that loss of wild-type p53 function, rather than gain-of-function p53 activities, is the predominant driver of PDAC development.

In characterizing PDAC molecular subtypes by transcriptomic and genomic analyses, associations with specific genetic events have been described (13–15, 19). The basal-like subtype has been correlated with certain mutations, such as those in TP53 and KDM6A, along with downregulation of genes of endodermal identity (6, 13–17, 19, 20). The classical subtype is characterized by expression of genes related to an endodermal cell fate such as GATA transcription factors (6, 13–15, 16, 20). Although these molecular subtypes have been linked to certain genetic lesions post facto, our transcriptome analysis directly interrogates the connection between the cell of origin and molecular subtype. We show not only that acinar cells can give rise to tumors of the classical subtype and ductal cells can give rise to tumors of the basal-like subtype, but also that both subtypes can originate from cells deficient for Trp53. In future, single-cell RNA-seq analysis of malignant and adjacent premalignant pancreas tissue in our mouse PDAC models will enhance our understanding of PDAC evolution. In particular, it is now feasible to computationally order cellular trajectories from such data (49), which would provide insight into the dynamics of tumor development from acinar and ductal cells. Single-cell resolution will help not only to identify intermediate states during tumor progression but also to decipher the significance of recent observations of a small set of human PDACs with intratumoral heterogeneity, where tumors are comprised of cells of both the classical and basal-like subtypes (17, 20).

Elucidating the cell of origin of specific cancer types not only provides key insight into the biological basis for tumor evolution, but also illuminates possible routes to improved clinical intervention. Genetically engineered mouse models have been fundamental for identifying the cell of origin of diverse cancers, such as intestinal, breast, and lung cancers, and for defining cell of origin as a critical parameter influencing intertumoral heterogeneity (50, 51). Our study underscores the importance of both acinar and ductal cells as tumor-initiating cells for PDAC, as well as illuminating the distinct molecular subtypes of PDAC arising from these cells. Although different molecular classes of PDAC have been previously defined, the existing subtype signatures are not currently used to guide therapeutic decision-making. Understanding the underlying biological basis for these distinct subtypes may help to refine signatures and thus improve strategies for early diagnosis, prognostication, and therapy. As new PDAC therapies are pursued, it will be instructive to stratify patients using our cell-of-origin signatures, which have a clear biological basis, when assessing the efficiency of such therapies to determine if they preferentially act in tumors of a particular subtype. Indeed, we have identified distinct functional annotation signatures for acinar cell and ductal cell tumors, and these may provide insight into therapeutic vulnerabilities specific to each subtype. For example, our ductal cell–derived signature displays a striking enrichment of genes involved in glycolysis, and, accordingly, some patient-derived cell lines of the squamous subtype are sensitive to glycolysis inhibition (52). In addition, although our study suggests that cell of origin is one factor that influences PDAC molecular subtype, other factors such as genetic and epigenetic changes during tumor progression likely also influence tumor subtype (20). Continued investigation will further reveal which additional factors, such as genetic and epigenetic changes, influence molecular subtype in PDAC and will enable this information to be utilized for improved clinical management of PDAC.

All mouse work was approved and performed in compliance with the Stanford University's Institutional Animal Care and Use Committee, known as the Administrative Panel on Laboratory Animal Care. Mice were maintained on a mixed 129/Sv-C57BL/6 background. Previously described mouse strains carrying the following alleles were used: Ptf1a^CreER (RRID:IMSR_JAX:019378), Sox9CreER (RRID:IMSR_JAX:018829), Trp53^fl (RRID:IMSR_JAX:008462), Trp53^null (RRID:IMSR_JAX:002101), Trp53^LSL-R172H (RRID:IMSR_JAX:008652), Trp53^LSL-R270H (RRID:IMSR_JAX:008651), Kras^LSL-12D (RRID:IMSR_JAX:008179), and Rosa26^LSL-tdTomato (RRID:IMSR_JAX:007909). Mice were genotyped by PCR. Cohorts with similar male and female representation were generated to account for sex as a biological variable. At 8 to 10 weeks of age, mice were administered 5 mg tamoxifen (Sigma; catalog no. T5648) dissolved in 2% ethanol/corn oil via oral gavage once per day for 3 consecutive days. Unless indicated otherwise, mice were aged until morbidity. At morbidity, tumor tissue was fixed for histopathology or flash-frozen in liquid nitrogen for additional analyses.

Hematoxylin and eosin (H&E) staining and Masson's trichrome staining were performed on paraffin-embedded tissues using standard protocols. All tumors were graded under blind review by a pathologist. Masson's trichrome staining was quantified using the FIJI Software (RRID:SCR_002285). Liver metastases were assessed through analysis of one H&E section per liver. Whole mount brightfield and fluorescent images were captured from freshly dissected tissue using a Leica stereo microscope and Nightsea stereo microscope fluorescent adapter.

IHC and immunofluorescence were performed using standard protocols. For unmasking, sections were incubated in 0.01 mol/L citrate buffer (pH 6.0) or Tris-EDTA buffer (10 mmol/L Tris, 1 mmol/L EDTA, and 0.05% Tween 20, pH 9.0) in a pressure cooker for 13 minutes. Primary antibodies directed against the following proteins were used: Amylase (Santa Cruz Biotechnology; catalog no. sc-46657, RRID:AB_626668), CK19 (DSHB; catalog no. TROMA-III, RRID:AB_2133570), CLDN18 (Thermo Fisher Scientific; catalog no. 700178, RRID:AB_2532290), Insulin (Cell Marque; catalog no. 273A-14, RRID:AB_1158520), MUC5AC (Lab Vision; catalog no. MS-145-P1, RRID:AB_62736), phospho-ERK1/2 (Cell Signaling Technology; catalog no. 4370, RRID:AB_2315112), p53 (Leica Biosystems; catalog no. P53-CM5P, RRID:AB_2744683), RFP/tdTomato (Rockland; catalog no. 600-401-379, RRID:AB_2209751), TFF1 (ABclonal; catalog no. A1789, RRID:AB_2763830), and Vimentin (Cell Signaling Technology; catalog no. 5741, RRID:AB_10695459). For immunohistochemistry, the following biotinylated secondary antibodies were used: anti-rabbit (Vector Laboratories; catalog no. BA-1000, RRID:AB_2313606), anti-rat (Vector Laboratories; catalog no. BA-9401, RRID:AB_2336208), and anti-mouse (Vector Laboratories; catalog no. BA-2000, RRID:AB_2313581). Following biotinylated antibody incubation, the VECTASTAIN Elite ABC HRP Kit (Vector Laboratories; catalog no. PK-6100, RRID:AB_2336819) was used. Subsequent chromogenic staining was performed with the DAB peroxidase substrate kit (Vector Laboratories; catalog no. SK-4100, RRID:AB_2336382), and sections were counterstained using Gill's formula hematoxylin (Vector Laboratories; catalog no. H-3401, RRID:AB_2336842) or the NovaUltra Alcian Blue Stain Kit (IHC World; catalog no. IW-3000). For immunofluorescence, the following secondary antibodies were used: Alexa Fluor 488 anti-rat (Molecular Probes; catalog no. A-21208, RRID:AB_141709), Alexa Fluor 546 anti-mouse (Molecular Probes; catalog no. A-11030, RRID:AB_144695), fluorescein anti-rabbit (Vector Laboratories; catalog no. FI-1000, RRID:AB_2336197), and fluorescein anti-mouse (Vector Laboratories; catalog no. FI-2000, RRID:AB_2336176). Sections were mounted with DAPI mounting media (ProLong Fluoromount G Mounting Medium, RRID:SCR_015961). All images were taken using a Leica DM6000B microscope (RRID:SCR_018713), an Aperio AT2 slide scanner (Leica Biosystems), and/or a NanoZoomer 2.0-RS slide scanner (Hamamatsu).

RNA was extracted from bulk tumors using the RNeasy Midi Kit (QIAGEN; catalog no. 75144). RNA quality was assessed using a BioAnalyzer (2100 Bioanalyzer Instrument, RRID:SCR_018043), and only samples with an RNA integrity number greater than 8.0 were included for library preparation. Novogene generated and sequenced RNA-seq libraries using the Illumina TruSeq RNA Sample Preparation Kit (catalog no. RS-122-2001) and Illumina HiSeq4000 (RRID:SCR_016386). Reads were aligned to the GRCm38 mouse genome using HISAT2 (v 2.0.4; RRID:SCR_015530). Sorted BAM files were generated using Samtools (v 1.3.1; RRID:SCR_002105). The number of reads mapping to each gene in the mouse Ensembl database (v 87; RRID:SCR_002344) was counted using HTSeq-count (v 0.6.1; RRID:SCR_011867). Differential gene-expression analysis was performed using DESeq2 (v 1.24.0; RRID:SCR_015687). Pathway analysis was performed using Metascape (v 3.5; RRID:SCR_016620).

The acinar cell– or ductal cell–derived tumor signatures comprise genes with an absolute log₂ fold change >1.5 and an adjusted P value < 0.01 in the differential gene-expression analysis described above. Human orthologs of the derived mouse genes were obtained using R package BioMart (v 2.40.5; RRID:SCR_002987). We used publicly available gene expression and clinical data for the ICGC Australian (13), ICGC Canadian (PACA-CA, downloaded from the ICGC data portal—https://dcc.icgc.org/), and TCGA (6) PDAC cohorts. R package GSVA v 1.32.0 (53) was used to estimate the enrichment of the acinar cell– or ductal cell–derived tumor signatures in human PDAC cohorts.

Human PDAC samples from the ICGC Australian and TCGA cohorts were assigned to molecular subtypes of squamous, pancreatic progenitor, ADEX, and immunogenic based on published subtype designations (6, 13). Subtype designation of quasi-mesenchymal/classical/exocrine-like and basal-like/classical were not published for the ICGC Australian cohort. Therefore, GSVA was used to calculate a signature enrichment score using the published gene sets for each subtype classification (14, 15). Tumors were then clustered using the k-means algorithm for all signatures, and the subtype was assigned to the cluster with the highest score.

GraphPad Prism (v 8.3.1; RRID:SCR_002798) and R (v 3.6.1; RRID:SCR_001905) were used for analysis and graphical representations. The n number is included in the figure legends and represents the number of mice analyzed per genotype for biological replication. GraphPad Prism was used to perform the log-rank test for the Kaplan–Meier analysis. The Survival package (v 2.44-1.1) was used for the cox proportional hazards regression model. The SurvComp package (v 1.32; RRID:SCR_003054) version 1.32 was used to compute meta-analysis statistics. Power analysis was not used in this study because sample sizes depended on available datasets.

Next-generation sequencing data for the RNA-seq experiments have been deposited at the Gene Expression Omnibus (RRID:SCR_005012) under accession number GSE164180.

Outside of the submitted work, C. Curtis reports personal and/or consulting fees from Genentech, NanoString, and GRAIL, Inc. and is a stockholder for GRAIL, Inc. No disclosures were reported by the other authors.

B.M. Flowers: Conceptualization, formal analysis, funding acquisition, investigation, visualization, writing–original draft, writing–review and editing. H. Xu: Formal analysis, visualization. A.S. Mulligan: Investigation. K.J. Hanson: Investigation. J.A. Seoane: Formal analysis, visualization. H. Vogel: Formal analysis. C. Curtis: Supervision. L.D. Wood: Formal analysis. L.D. Attardi: Conceptualization, formal analysis, supervision, funding acquisition, writing–original draft, writing–review and editing.

The authors thank Tyler Jacks and Monte Winslow for mouse strains; the staff at the Stanford Functional Genomics Facility for their technical support; Pauline Chu for the processing and sectioning of all paraffin-embedded mouse tissue samples; Steven Artandi for the insulin antibody; and Monte Winslow, Anthony Boutelle, and Alyssa Kaiser for critical reading of the manuscript. This work was supported by an NCI R35 grant (CA197591, to L.D. Attardi), a Blavatnik Family Fellowship (to B.M. Flowers), and an NIH Director's Pioneer Award (DP1-CA238396, to C. Curtis).

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.