ID1 Mediates Escape from TGFβ Tumor Suppression in Pancreatic Cancer

TGFβ signaling mediates tumor-suppressive or tumor-progressive effects depending on the developmental context of a cancer cell (1). TGFβ causes apoptosis in premalignant cells that harbor RAS oncogenes (2, 3). Some tumors develop with genetic inactivation of this pathway. However, a majority of tumors retain a functionally intact TGFβ pathway, somehow eliminating the tumor-suppressive effect of TGFβ while benefiting from its invasive and immunosuppressive effects (1, 4).

Pancreatic ductal adenocarcinoma (PDA), which has the highest incidence of TGFβ pathway mutations in cancer, is a paradigmatic example of this phenomenon. In normal pancreatic progenitors, the transcription factors KLF5 and SOX4 cooperatively impose an epithelial phenotype. Mutational activation of KRAS, a nearly universal tumor-initiating event in PDA, causes TGFβ to induce expression of the epithelial–mesenchymal transition (EMT) transcription factor SNAIL, which represses KLF5 expression. As cells undergo a SNAIL-driven EMT in this dysregulated, KLF5-depleted context, SOX4 triggers a phenotype checkpoint by activating proapoptotic genes and cell death (3). TGFβ pathway inactivation, most frequently by mutation of SMAD4, occurs in half of PDA cases. How the other half of PDAs retain an intact TGFβ pathway yet avert the proapoptotic effects remains an open question.

Genomic analysis of PDAs has revealed three other frequently mutated genes—KRAS, CDKN2A, and TP53—in addition to SMAD4 (5). Mutations in these three genes are not mutually exclusive with SMAD4 mutations. Beyond these genetic alterations, PDAs present with a long series of low-frequency mutated genes (5). Given that no single high-frequency genetic alteration has emerged as mutually exclusive to TGFβ pathway inactivation, we postulated that multiple alterations may converge on a common regulatory node that is critical to escape from tumor suppression in PDAs with an intact TGFβ pathway. Identifying this regulatory node would provide a potential therapeutic target in PDA.

To investigate this hypothesis, we focused on the analysis of dominant transcriptional networks in PDAs. Transcriptional dysregulation is a common feature of emerging tumors, reflecting adaptation to genetic alterations in cancer cells and inputs from the tumor microenvironment. Using this approach, we found that cancer cells from PDA tumors that develop with an active TGFβ pathway avert apoptosis by transcriptional dysregulation of ID1, an inhibitor of progenitor cell differentiation (6). Transcriptional induction of ID1 uncouples TGFβ-induced EMT from apoptosis. The dysregulation of ID1 expression results from a diverse set of alterations, including PI3K–AKT signaling pathway mutations. ID1 thus emerges as a target of interest in pancreatic cancer.

TGFβ signals through the paired receptor kinases TGFBR1 and TGFBR2 to phosphorylate SMAD2 and SMAD3 transcription factors, which associate with SMAD4 to activate target genes (Fig. 1A; ref. 1). SMAD4 is inactivated in 38% to 43% of human PDAs, and the full set of TGFβ pathway core components collectively is inactivated in approximately 53% of PDAs (Supplementary Fig. S1A). To determine whether PDAs lacking mutations in these components retain a functional TGFβ pathway, we assayed 12 human PDA organoids for responsiveness to TGFβ. Activating KRAS mutations (G12D, G12V, or Q61H) were detected in all of the organoids and deleterious TP53 mutations were identified in 8 of 12, reflecting the mutational spectrum of PDAs (Supplementary Table S1). Using induction of the common TGFβ target gene SMAD7 as a readout, we found that six PDA organoids showed a weak increase (<3-fold; organoids HT22, HT33, and NL5) or no increase in SMAD7 mRNA levels by TGFβ (HT30, HT42, and LMCB3), whereas the other six showed a 5- to 40-fold increase (Fig. 1A). We designate these as “TGFβ-inactive” or “TGFβ-active” organoids. Because the functional transcriptional unit of TGFβ signaling is a trimer of receptor-phosphorylated SMAD2/3 with SMAD4, we determined by immunoblotting whether the organoids expressed SMAD4 and phospho-SMAD2 (pSMAD2) in response to TGFβ. Three of the TGFβ-inactive organoids (HT30, HT33, HT42) exhibited low levels of pSMAD2, consistent with receptor inactivation. HT30 has a TGFBR2 N179Ifs*10 mutation, HT33 a TGFBR2 P154Afs*3 mutation, and HT42 a TGFBR2 R485H mutation. The other TGFβ-inactive organoids showed low levels of SMAD4. All TGFβ-active organoids stained positive for pSMAD2 and SMAD4 (Supplementary Fig. S1B), suggesting that a subset of PDAs retain a functionally intact TGFβ pathway.

To investigate this point in tissue samples, we stained for pSMAD2 in a tissue microarray of 130 human PDAs of which 69% were SMAD4-positive (7). Five percent of samples were negative for pSMAD2 in the tumor cells but were positive in the tumor stroma (Fig. 1B and C), congruent with the frequency of genetic TGFβ receptor inactivation. Sixty-three of the samples were positive for both pSMAD2 and SMAD4, suggesting the presence of TGFβ signaling components in the cancer cells of these tumors (Fig. 1B and C). Mouse PDAs (8) showed pSMAD2 staining regardless of whether the tumors expressed wild-type SMAD4 or not (Supplementary Fig. S1C). These results suggest that a large subset of PDAs retain a functional TGFβ pathway and are exposed to TGFβ in the tumor microenvironment.

Progenitor cells are characterized by the expression of dominant transcription factors that specify the lineage and developmental stage of the cells. We decided to test the hypothesis that PDAs developing with either active or inactive TGFβ pathway achieve a common carcinoma stage endpoint characterized by a particular transcriptional network. We performed principal component analysis (PCA) of gene- expression datasets from 225 cases of PDA, normal pancreas, and pancreatic neuroendocrine tumors (PNET; refs. 9, 10). PCA based on the top five transcription factors expressed within at least one case (67 total; Fig. 1D; Supplementary Fig. S1D) showed separation of normal, PDA, and PNET samples. This separation was similar to that of PCA using a full list of expressed transcription factors (Supplementary Fig. S1E). The same analysis was applied to a second, independent PDA gene-expression dataset (5) with similar results (Supplementary Fig. S1F). Notably, PDA cases with wild-type or mutant TGFβ pathway components clustered together (Fig. 1D), suggesting that PDA evolution selects for similar transcription factor networks regardless of the genetic integrity of the TGFβ pathway.

Of the transcription factors highly expressed in normal tissue, RBPJL, BHLHA15/MIST1, and NR5A2 are crucial in pancreatic specification and development of the acinar lineage, which is the putative cell of origin of PDA (11–13). In PNET, the highly expressed transcription factors correspond to the endocrine lineage: ISL1, MAFB, NKX2.2, PAX6, and ST18 (14–17). In PDA, the highly expressed transcription factors include AHR, HMGB2, ID1, KLF5, and PPARG, which play roles in epithelial progenitor specification and transformation (6, 18–20). The expression level of the top five transcription factors was sufficient to separate normal samples from PDA samples by unsupervised hierarchical clustering (Fig. 1E). In an independent dataset (21), these factors were enriched in PDAs of both the classic and the basal subtypes (Supplementary Fig. S1G), as well as in both primary and metastatic tumors (Supplementary Fig. S1H), suggesting that these five genes define a basic PDA transcriptional network.

To investigate the effect of TGFβ on the expression of the five PDA-associated transcription factors in a tumor-suppressive context, we transduced a SMAD4 vector in cancer cells derived from mouse Kras^G12D;Cdkn2a^−/−;Smad4^−/− PDA tumors (8), which restores the apoptotic effect of TGFβ (3). Notably, treatment of the SMAD4-restored PDA cells with TGFβ for 12 hours, compared with treatment with SB505124 (a TGFBR1 kinase inhibitor used to suppress endogenous TGFβ activity), decreased the expression of the five transcription factors enriched in PDA, as determined by full transcriptome mRNA sequencing (RNA-seq; Fig. 1F). The effect was confirmed in a second mouse PDA cell line and two human PDA cell lines (Fig. 1G). The effect was particularly pronounced with Id1 and Pparg, and was accompanied by induction of Snai1, encoding SNAIL (Fig. 1H; ref. 3), and expression of EMT and apoptosis signatures based on gene set enrichment analysis (GSEA; Fig. 1I). Thus, the PDA core transcription network is repressed in the context of a TGFβ tumor-suppressive response.

ID1 was of particular interest because of the known role of the ID family members (ID1–4) in the negative regulation of cell differentiation (6). ID1–4 expression is high in progenitor cells and low in more differentiated cells. ID proteins lack a DNA-binding domain and function by sequestering basic helix-loop-helix (bHLH) E-proteins to prevent their dimerization with differentiation bHLH transcription factors. TGFβ represses ID1–3 in epithelial progenitors (22) and induces ID1 expression in breast cancer and glioblastoma cells (23–25).

To better understand the role of ID1 in PDA, we stained for ID1 in a panel of human pancreatic tissues including normal pancreas, pancreatic intraepithelial neoplasias, primary PDAs, and PDA metastases. Epithelial cells stained positive for ID1 within PDA samples (Fig. 2A). Of the 30 cases of human PDA that we examined, 24 (80%), including 14 of 16 SMAD4⁺ tumors and 10 of 14 of SMAD4⁻ tumors, exhibited cancer cells with nuclear ID1 staining. In normal human pancreatic tissues from 6 patients who had no signs of pancreatic disease, nuclear ID1 staining was observed in endothelial cells and in fewer than 1% of epithelial cells (Fig. 2A).

We devised a four-color multiplex IHC staining method to examine ID1, KLF5, and SOX4 and cytokeratins together in PDA samples surgically resected from 6 patients of known SMAD4 status (Fig. 2B; Supplementary Table S2). ID1⁺,KLF5⁺,SOX4⁺ cells constituted only a small proportion (mean, 5.5%; range, 0.2%–13.2%) of the cancer cell population in these samples (Fig. 2C; Supplementary Table S1). Of the carcinoma cells, 55.9% expressed none or only one of these progenitor markers. Notably, ID1⁻,KLF5⁻,SOX4⁺ cells were less abundant (mean, 3.8%; range, 0.6%–12.3%) than ID1⁺,KLF5⁻,SOX4⁺ cells (mean, 7%; range, 4.8%–11.8%). Among SMAD4⁺ samples (samples 2, 4, 6; Supplementary Table S1), ID1⁻,KLF5⁻,SOX4⁺ cells were even less abundant (mean, 1.1%; range, 0.6%–2.1%) compared with ID1⁺,KLF5⁻,SOX4⁺ cells (mean, 7.5%; range, 4.9%–10.2%). ID1⁻,KLF5⁻,SOX4⁺ cells were the least abundant population overall in the sample set as well as in individual cases, except for one case of poorly differentiated SMAD4⁻ PDA, which had large ID1⁺ cell populations compared with the other cases (Supplementary Table S1). These results are consistent with our hypothesis that SOX4⁺ cells lacking ID1 and KLF5 expression survive poorly, particularly in SMAD4⁺ PDAs.

As with the human PDAs, 3 of 3 normal mouse pancreata were negative for ID1, whereas 6 of 6 PDAs from different autochthonous and orthotopic mouse models (Kras^G12D;Cdkn2a^−/− or Kras^G12D;Trp53^R172H or Kras^G12D;Cdkn2a^−/−;Smad4^−/−) showed ID1 staining in 46% to 47% of cancer cells regardless of Smad4 status (Supplementary Fig. S2A and S2B). PDX1 is a transcription factor of pancreatic progenitors (26). ID1⁺ cells constitute a subset of PDX1⁺ cells in autochthonous pancreatic tumors (Supplementary Fig. S2C and S2D). Amylase, a marker of mature pancreatic acinar cells, was mutually exclusive with ID1 immunofluorescence (Supplementary Fig. S2D). In freshly dissociated Kras^G12D;Cdkn2a^−/−;Smad4^−/− orthotopic tumors, the ID1⁺ population is also enriched for cell-surface markers (CXCR4, SSEA4, CD44, and CD24) previously associated with pancreatic cancer stem cells (Supplementary Fig. S2E and S2F; refs. 27–29).

To recapitulate the heterogeneity in ID1 expression observed in pancreatic tumors, we created a knock-in GFP reporter at the endogenous Id1 locus in mouse Kras^G12D;Cdkn2a^−/−;Smad4^−/− PDA cells (ID1-GFP cells; Fig. 2D and E; Supplementary Fig. S2G and S2H). GFP^hi cells isolated by FACS from these populations grew better in a surrogate growth assay for stem-like cancer cells in which the cells grow as spheroid colonies (“oncospheres”) in low-attachment plates with limited growth factors (ref. 28; Fig. 2F). The transcriptional signature of pancreatic stem cells is not well defined. However, the GFP^hi cells expressed signatures associated with breast cancer progenitors and intestinal stem cells, as determined by RNA-seq and GSEA (Supplementary Fig. S2I; refs. 30, 31). In contrast, GFP^lo cells expressed signatures of pancreatic epithelial differentiation (Supplementary Fig. S2J). The data indicate that ID1 is expressed in a large population of PDA cells including progenitor cells.

ID genes are frequently coexpressed and have overlapping functions (6). ID2 and ID3 were enriched in human PDAs, albeit less than ID1 was (Supplementary Fig. S2K). Of six human PDA samples in which nuclear ID1 was not detected, three showed nuclear staining for ID2, one for ID3, and two for ID2 and ID3 (Supplementary Fig. S2L and S2M). shRNA-mediated knockdown of Id1 stimulated the expression of Id2 in Kras^G12D;Cdkn2a^−/−;Smad4^−/− PDA cells, and Id2 knockdown stimulated expression of Id1 and Id3 (Supplementary Fig. S2N). Knockdown of Id1 decreased the spheroid-forming ability of these PDA cells, and additional knockdown of Id2 and Id3 further decreased spheroid growth (Fig. 2G). Thus, ID1, ID2, and ID3 are expressed in a compensatory manner and show a partial functional overlap in PDA cells.

We orthotopically implanted Kras^G12D;Cdkn2a^−/−;Smad4^−/− mouse pancreatic cells with doxycycline-inducible expression of Id1–3 shRNAs, and tumor growth was tracked by bioluminescence imaging (BLI) of a transduced firefly luciferase gene. The knockdown of Id1 alone decreased tumorigenic potential by approximately 10-fold (Supplementary Fig. S2O). The knockdown of Id1–3 further decreased this potential (Fig. 2H; Supplementary Fig. S2P) and prolonged survival of the tumor-bearing mice (Fig. 2I), compared with doxycycline treatment of tumors expressing a control shRNA (shRen). ID1 immunostaining was present in the late-emerging Id1–3 shRNA tumors, indicating tumor formation by cells that escaped Id1 depletion (Supplementary Fig. S2Q).

To determine if the requirement for the ID proteins is retained in an established tumor, we orthotopically implanted PDA cells harboring inducible Id1–3 shRNAs, allowed the tumors to grow for 3 weeks, and then randomized the mice into matched bioluminescence groups for treatment with doxycycline. Doxycycline-treated mice had lower tumor burdens and longer survival (Fig. 2J and K). Id1–3 shRNAs also decreased the formation of orthotopic tumors by a cell line derived from a Kras^G12D;Cdkn2a^−/− mouse PDA (Fig. 2L and M). These observations suggest that the ID family supports the tumorigenic potential of PDA cells.

The emergence of ID1 as a top PDA transcriptional network component and its tumorigenic activity in PDA progenitor cells led us to investigate whether ID1 downregulation by TGFβ is deleterious to these cells. To simulate the premalignant, TGFβ-sensitive state, we restored SMAD4 expression in Kras^G12D;Cdkn2a^−/−;Smad4^−/− mouse PDA cells and human SMAD4-null PDAs. In addition, we derived premalignant cells from Kras^G12D;Cdkn2a^−/− mouse pancreata. We then queried the TGFβ response of cells from these various models (Fig. 3A). TGFβ inhibited Id1 expression and induced Snai1 expression (Fig. 3B) and apoptosis (Fig. 3C) in organoids derived from Kras^G12D;Cdkn2a^−/− mouse pancreata. RNA-seq data from the SMAD4-restored mouse PDA cells showed that TGFβ also decreased the expression of Id2 and Id3 (Fig. 3D), in association with the induction of Snai1 (refer to Fig. 1G) and cleaved caspase-3 (Fig. 3E). The decrease in Id1–3 expression was accompanied by an increase in protein binding to an E-box motif oligonucleotide in electrophoretic mobility shift assay (Supplementary Fig. S3A), consistent with a decrease in ID activity in these cells.

In SMAD4-restored mouse PDA cells containing an endogenous ID1-GFP reporter, the sorted GFP^hi cells showed ID1-GFP downregulation by TGFβ (Fig. 3F and G), which was accompanied by induction of Snai1 (Supplementary Fig. S3B) and apoptosis (Fig. 3H; Supplementary Fig. S3C). In contrast, the GFP^lo counterparts did not undergo apoptosis in response to TGFβ (Fig. 3H; Supplementary Fig. S3C), although they showed SMAD2 phosphorylation and Snai1 expression in response to TGFβ (Fig. 3I and J). These ID1-GFP^lo cells express less SOX4 than the ID1-GFP^hi cells (Supplementary Fig. S3D), and high expression of SOX4 enables TGFβ-mediated apoptosis (3). Expression of SOX4 from an inducible promoter increased TGFβ-mediated apoptosis in ID1-GFP^lo cells (Fig. 3K; Supplementary Fig. S3E). These results suggest that TGFβ-induced ID1 downregulation and apoptosis are specific responses of ID1^hi PDA progenitors, whereas the more differentiated progeny is protected from apoptosis by a decrease in SOX4 expression.

In contrast to SMAD4-restored PDA cells, cells derived from PDAs that developed with a functional TGFβ pathway were resistant to TGFβ-induced apoptosis (Fig. 4A–D) and, notably, retained ID1 expression in the presence of TGFβ (Fig. 4E). TGFβ increased ID1 expression in human PDA organoids that had a functional TGFβ pathway (Fig. 4F), and the extent of this induction tracked with that of SMAD7 across the cohort (refer to Fig. 1A).

To recapitulate this phenomenon in cells derived from the same genetic background, we cultured SMAD4-restored mouse PDA cells (S4 cells) for 3 weeks with TGFβ and selected for cells resistant to the proapoptotic effect of TGFβ (Fig. 4G). This selection protocol was done in the presence of the allosteric AKT inhibitor MK2206, which synergizes with TGFβ in the induction of apoptosis in PDA cells (3). The cells that we derived (S4.1 cells) were resistant to the proapoptotic effects of TGFβ (Fig. 4H; Supplementary Fig. S4A) but still induced Smad7 expression (Fig. 4I) and underwent an EMT in response to TGFβ (Supplementary Fig. S4B). Notably, whereas TGFβ treatment downregulated Id1 in the S4 parental population, it increased Id1 expression in S4.1 cells (Fig. 4I–J) and in two other independent populations (S4.2 and S4.3) isolated with the same protocol (Supplementary Fig. S4C). Besides Id1, only four other genes were differentially regulated by TGFβ by more than 4-fold: Id3, Fam167a, Trib3, and Chac1 (Fig. 4I). Id2 was differentially regulated by TGFβ but less than 4-fold. RNA polymerase II (RNAPII) chromatin immunoprecipitation sequencing (ChIP-seq) analysis in S4 and S4.1 cells showed RNAPII already occupied the promoter of Id1 under basal conditions, and TGFβ decreased the occupancy in S4 cells, but increased it in S4.1 cells (Supplementary Fig. S4D). Id1–3 knockdown in S4.1 cells rescued the TGFβ-induced apoptosis (Fig. 4K).

TGFβ had similar effects on SMAD2/3 and SMAD1/5 phosphorylation in S4 and S4.1 cells (Supplementary Fig. S4E), ruling out a switch in SMAD signaling (32) as an explanation for the induction of Id1 in these cells. Functional annotation of the genes differentially expressed between the S4 and S4.1 cells (average expression >10 readcounts, fold change >2, P < 0.05) showed that the top enriched pathways were related to small-molecule metabolism (Supplementary Fig. S4F). The common genes between these pathways were all UDP glucuronosyltransferases (UGT), which are key enzymes for the inactivation of polycyclic compounds, including MK2206 (33). Several UGTs were upregulated in the S4.1 cells (Supplementary Fig. S4G). We generated a signature of differentially expressed genes (average expression > 10 readcounts, fold change > 4, P < 0.05) in S4 cells treated with 2.5 μmol/L MK2206 for 16 hours. The MK2206 gene signature was decreased in the S4.1 cells relative to the S4 cells treated under these conditions (Supplementary Fig. S4H), consistent with a low effectiveness of MK2206 in the S4.1 cells. Collectively, these results indicate that human and mouse PDA cells that retain an active TGFβ pathway do not downregulate ID1 expression in response to TGFβ.

Next, we tested the hypothesis that loss of TGFβ-mediated ID1 repression protects PDA progenitors from proapoptotic TGFβ effects. Enforced expression of Id1 from an inducible promoter in SMAD4-restored mouse PDA cells (Fig. 5A) inhibited the apoptotic effect of TGFβ (Fig. 5B). When these cells were implanted into mice with caerulein-induced acute pancreatitis, which mimics the TGFβ-rich microenvironment of the premalignant pancreas, enforced expression of Id1 conferred resistance to apoptosis (Fig. 5C) and promoted tumor growth (Fig. 5D).

The enforced expression of Id1 caused few changes in basal gene expression (average readcounts >10, fold change >2, P < 0.05, Supplementary Fig. S5A) and in TGFβ gene responses (Fig. 5E), but decreased the ability of TGFβ to induce several genes with described proapoptotic activities: Fbxo32, Rnf152, Bmf, Ndrg1, and Errfi1. Using qRT-PCR, we confirmed that enforced expression of Id1 decreased the induction of Fbxo32 and Rnf152 by TGFβ without affecting the induction of Snail1 or Smad7 (Supplementary Fig. S5B). Id1 expression did not interfere with the enrichment for an EMT gene signature (Fig. 5F), the induction of EMT effectors (Snai1, Zeb1, Zeb2, and Cdh2), or the downregulation of SNAIL-repressed genes Cdh1 (E-cadherin), Klf5, and Krt19 (cytokeratin 19) in response to TGFβ (Fig. 5A and G). Under prolonged (72 hours) treatment with TGFβ, SMAD4-restored cells with enforced Id1 expression survived with a mesenchymal phenotype (Fig. 5H). These results suggest that TGFβ repression of ID1 is necessary for the apoptotic effect, whereas sustained expression of ID1 decouples TGFβ-induced EMT from apoptosis.

SMAD-binding motifs and E-box–binding motifs were enriched within accessible chromatin regions near the ID1-inhibited genes, as detected by the assay for transposase-accessible chromatin sequencing (ATAC-seq; Supplementary Fig. S5C). SMAD2/3 bound to a subset of these chromatin regions in the presence of TGFβ (Supplementary Fig. S5D), including near the ID1-regulated proapoptotic genes (Supplementary Fig. S5E). Focusing on the accessible chromatin regions of Fbxo32 and Rnf152, we confirmed the presence of SMAD- and E-protein–binding sequences in these regions (Supplementary Fig. S5E). TGFβ induced the binding of SMAD2/3 to these regions, as determined by ChIP-PCR (Supplementary Fig. S5F). The E-proteins E12/E47 were also bound to these enhancers (Supplementary Fig. S5G). ID1 expression caused a decrease in E12/E47 binding without significantly affecting the binding of SMAD2/3 (Supplementary Fig. S5F and S5G). These results are consistent with the mode of action of ID proteins, which is by preventing the binding of E-proteins to DNA (6).

The observed switch of the ID1 response to TGFβ from repression to activation provides a mechanism for tumor progression by PDAs that develop with a functional SMAD pathway. Given that the highly recurrent genetic alterations in PDA (i.e., KRAS, TP53, CDKN2A) are not mutually exclusive to TGFβ pathway alterations (refer to Supplementary Fig. S1A), we postulated that low-frequency alterations in multiple factors achieve the same common end of preventing the repression of ID1 by TGFβ. We performed a genome-wide CRISPR/Cas9 screen to identify factors that enable ID1 repression by TGFβ. We transduced SMAD4-restored mouse PDA cells containing the endogenous ID1-GFP reporter with the Gecko_v2 genome-wide library containing 123,411 single guide RNAs (sgRNA; refs. 34, 35) and treated the cells with TGFβ or SB505124 for 36 hours. To select cells that bypassed Id1 repression by TGFβ, we sorted for ID1-GFP^hi cells and repeated this procedure three times to improve the stringency of the screen (Fig. 6A).

sgRNAs targeting TGFβ pathway components were enriched in the samples treated with TGFβ (Fig. 6B; Supplementary Fig. S6A), providing a positive control for the screen. In addition, Biocarta pathway analysis of the data identified PI3K–AKT-related pathways (AKT, EGF, IGF1, IGF1R, INSULIN, PTEN, HER2, MTOR, and NGF) as being depleted in the TGFβ-treated samples (Fig. 6C). Analysis of genetic events in the MSK-IMPACT dataset (36) for alterations occurring with mutual-exclusivity tendencies to TGFβ pathway inactivation demonstrated concordant depletion of PI3K–AKT-related pathways (Fig. 6D).

Several PI3K–AKT-activating events including INSR and AKT2 amplification and PIK3CA and PTEN mutation occur in a subset of PDA with intact TGFβ pathways (Supplementary Fig. S6B). Allowing AKT signaling, either by excluding MK2206 from the media or by expressing the MK2206-resistant mutants AKT1^W80A and AKT2^W80A (37), prevented TGFβ-mediated Id1 repression and apoptosis in SMAD4-restored cells (Fig. 7A and B; Supplementary Fig. S7A), while still allowing TGFβ induction of Snai1 and EMT (Fig. 7C; Supplementary Fig. S7A). These results provided further evidence that PI3K–AKT activation selectively prevents TGFβ-mediated ID1 repression and decouples TGFβ-induced EMT from apoptosis in PDA cells.

Five regions of high accessibility (regions 1–5) were detected in the Id1 locus in SMAD4-restored mouse PDA cells (Fig. 7D). These regions were conserved in the ID1 locus in human Panc1 PDA cells (Supplementary Fig. S7B). SMAD2/3 interacts with regions 1, 3, and 4 in SMAD4-restored mouse PDA cells in a TGFβ-dependent manner (Fig. 7D). Region 1, which is located approximately 1 kb upstream of the transcription start site, is implicated in the downregulation of ID1 by TGFβ in other cell types (22, 38). PI3K–AKT signaling upregulates ID1 expression by excluding the transcription factor FOXO3a from the nucleus (39). Indeed, FOXO1 and FOXO3a interacted with Id1 regions 1 and 2 (Fig. 7E). In pooled CRISPR/Cas9-modified cells, modification of Region 2 blocked the repression of ID1 by TGFβ, as did targeting of Tgfbr2 (Fig. 7F). Mouse PDA tissues demonstrated more extensive FOXO1 nuclear exclusion in Kras^G12D;Cdkn2a^−/− tumors than in Kras^G12D;Cdkn2a^−/−;Smad4^−/− tumors, suggesting higher AKT signaling in the SMAD4⁺ tumors (Supplementary Fig. S7C). Some cells isolated from Kras^G12D;Cdkn2a^−/− PDAs were sensitive to TGFβ-induced apoptosis in the presence of AKT inhibition (Supplementary Fig. S7D). Collectively, these results suggest that PI3K–AKT activation constitutes one mechanism for the dysregulation of ID1 repression to evade TGFβ-mediated tumor suppression in PDA cells.

To survive the apoptotic effect of TGFβ during tumor formation, KRAS-mutant pancreatic progenitors must either genetically alter the TGFβ pathway or decouple it from apoptosis. Mutational inactivation of SMAD4 or TGFβ receptors eliminates TGFβ tumor-suppressive responses in approximately 50% of human PDA cases. Our new findings identify an escape mechanism for the other cases by showing that ID1 decouples TGFβ-induced EMT from apoptosis in PDA cells (Fig. 7G). A key finding of the current study is that transcriptional dysregulation of ID1 is a selected, common feature of PDAs, which imparts both progenitor-like characteristics to PDA cells and protects from TGFβ-induced apoptosis. We show that a combination of signaling inputs and genetic alterations contributes to the maintenance of high ID1 levels in PDAs. Other bHLH factors necessary in pancreatic development, including MIST1 and PTF1A, have tumor-suppressive roles in the pancreas (40, 41), and together with our observations this suggests that a critical balance between inhibitory and activating bHLH factors controls pancreatic homeostasis.

TGFβ-induced apoptosis occurs in PDA progenitor stages that express SOX4 and KLF5, and depend on cooperation between these transcription factors for maintenance of their epithelial progenitor phenotype and survival (3). TGFβ and RAS signaling synergize to potently induce SNAIL expression, which triggers a phenotype checkpoint through a proapoptotic imbalance between SOX4 and KLF5. We now show that ID1 expression averts apoptosis in this context. More differentiated PDA cell progenies that express low levels of ID1 and SOX4 do not die in response to TGFβ, but these ID1/SOX4-low cells are also poorly tumorigenic. By uncoupling TGFβ signaling from apoptosis, ID1 allows PDA progenitors to retain EMT and other effects of TGFβ that provide tumors with selective advantages in immune evasion, invasion, and metastasis (1). The ability to uncouple the proapoptotic and EMT effects of TGFβ provides a framework for targeting specific features of this pathway.

The PI3K–AKT pathway emerged in our work as a signal for ID1 dysregulation. PI3K–AKT-activating mutations are relatively rare in PDA compared with other tumor types. However, insulin signaling is particularly relevant in the pancreas, where blood flows through the insulin-synthesizing islets of Langerhans to the capillaries prior to supplying the pancreatic acini and ducts (42). Furthermore, several studies have connected high levels of circulating insulin with increased incidence of PDA and resistance to the proapoptotic effects of PI3K inhibition (43–45). AKT2 amplification and a microenvironment rich in insulin may potentiate PI3K–AKT signaling in the pancreas and contribute to the observation that receptor tyrosine kinase activation is necessary for PDA formation even in the presence of mutant KRAS (46, 47). The tail of low-frequency genetic alterations that are exclusive of SMAD4 mutations in genome-wide studies of PDA may contain additional mediators of ID1 dysregulation besides AKT pathway components.

Our analysis demonstrates that focusing on highly expressed transcriptional regulators yields commonalities in PDA. Our analysis using a rank-based, transcription factor–centered algorithm allowed us to identify factors that are key to integrating microenvironmental signals and genetic alterations to orchestrate cell fates and states. Transcription factors that are highly and differentially expressed frequently represent lineage dependencies and constitute promising therapeutic targets. Our findings suggest that PDAs of different subtypes possess common transcriptional dependencies, such as on ID1. Small-molecule inhibitors of ID proteins have been reported (48, 49). Based on the dependence of PDA cells on ID1, we predict that PDAs would be particularly sensitive to these inhibitors.

The study was conducted under Memorial Sloan Kettering Cancer Center Institutional Review Board approval (MSKCC IRB 15-149 or 06-107), and all patients provided written informed consent prior to tissue acquisition. Normal pancreas and PDA samples were collected from patients undergoing routine surgical resection and as part of MSKCC's rapid autopsy program for paraffin embedding or organoid generation. Normal pancreatic samples were from 6 patients with tumors at distant sites (lung and brain) with no observed pathology in the pancreas.

Animal experiments were performed as approved by the MSKCC Institutional Animal Care and Use Committee. For orthotopic tumor models, 500 dissociated cells were implanted in 25 μL of Matrigel (Corning Matrigel GFR Membrane Matrix, #356231), injected through 31G syringes into the pancreata of 4-week-old female FVB (JAX, FVB/NJ) or athymic nude (ENVIGO, Hsd: Athymic Nude-Foxn1nu) mice. Mice were started on 2,500 mg/kg doxycycline diet (ENVIGO, TD.07383 2014-2500-B, irradiated) after cell implantation for Tet-On shRNA experiments. Tumor growth was tracked weekly using bioluminescence (Goldbio Firefly d-Luciferin, potassium salt). Tumors were collected from genetic mouse models of PDA (Pdx1-Cre LSL-Kras^G12D;Cdkn2a^fl/fl;Smad4^fl/fl and LSL-Kras^G12D;Cdkn2a^fl/fl) forcell lines, organoid lines, and IHC. Acute pancreatitis was induced by caerulein injection with 8 hourly injections of 50 μg/kg on two consecutive days. For AKT inhibition, mice were dosed by oral gavage with 100 mg/kg MK2206 dissolved in Captisol as a carrier.

Pancreatic organoids were generated as described previously (50) and maintained embedded in Matrigel with Advanced DMEM/F12 (Gibco, 12634-028) supplemented with B-27 (Life Technologies, 12587-010), HEPES (10 mmol/L), 50% WNT/R-spondin/Noggin-conditioned medium (ATCC, CRL-3276), Glutamax (Invitrogen, 2 mmol/L), N-acetyl-cysteine (Sigma, 1 mmol/L), nicotinamide (Sigma, 10 mmol/L), EGF (PeproTech, 50 ng/mL), gastrin (Sigma, 10 nmol/L), fibroblast growth factor 10 (PeproTech, 100 ng/mL), and A83-01 (Tocris, 0.5 μmol/L) as described previously (50). Pancreatic spheroids were grown in Ultra Low Attachment Culture plates (Corning) in DMEM supplemented with Glutamax (2 mmol/L) and heparin (5 μg/mL). All cell lines and organoids were maintained at 37°C and 5% CO₂.

Human pancreatic cancer cell lines BxPC3, MiaPaca2, and Panc1 were obtained from the ATCC in 2012. The human pancreatic cancer cell line A21 was obtained in 2016 (51). The 806 and 906 Kras^G12D; Cdkn2a^−/−;Smad4^−/− mouse PDA cell lines and the NB44 Kras^G12D; Cdkn2a^−/−;Smad^+/+ mouse PDA cell line were provided by N. Bardeesy (Massachusetts General Hospital Cancer Center) in 2011 (8). The 4279 cell line was derived from a Kras^G12D;Cdkn2a^−/−; Smad^+/+ mouse PDA tumor in this lab. Pancreatic cell lines were maintained in high-glucose DMEM supplemented with 10% FBS and 2 mmol/L l-glutamine. Cell lines were authenticated by Western blot verification of expected genetic alterations or transcriptomic evaluation of appropriate gene expression. Cell lines were frozen in FBS with 10% DMSO and stored in liquid nitrogen indefinitely. Organoid lines were frozen in organoid growth media with 50% FBS and 10% DMSO. Cell lines were used within 1 month of culture from cryopreserved stocks and were tested for Mycoplasma contamination at the time of acquisition and cryopreservation. Organoids were used between passages 5 and 25, and were tested for Mycoplasma contamination after use or 1 to 6 passages before use.

Four micron–thick sections obtained from tissue blocks were baked for 3 hours at 62°C with deparaffinization performed on the Leica Bond RX followed by 4 sequential rounds of staining, each round including a combined block with primary antibody (PerkinElmer antibody diluent/block ARD1001) followed by a corresponding secondary horseradish peroxidase (HRP)–conjugated polymer (PerkinElmer Opal polymer HRP Ms + Rb ARH1001). Each HRP-conjugated polymer induces the covalent binding of fluorophores to tissue using tyramide signal amplification. The covalent reaction was followed by heat-induced stripping of the primary antibody in PerkinElmer AR9 buffer (AR900250ML) for 20 minutes at 100°C before the next step in the sequence. The antibodies were sequentially stained in the following order: ID1 (BioCheck, 1:300), KLF5 (R&D Systems, 1:200), SOX4 (Abcam, 1:600), and keratin cocktail [PanCK (Dako, 1:200), CK7 (Abcam, 1:400), and CAM5.2 (Becton Dickinson, 1:150)]. Following incubation of the KLF5 goat polyclonal primary antibody, detection was performed using a rabbit anti-goat secondary (Vector Laboratories, 1:4,000) followed by the aforementioned PerkinElmer Opal polymer. After four sequential rounds of staining, sections were stained with Hoechst 33342 (Invitrogen) to visualize nuclei and mounted with ProLong Gold antifade reagent mounting medium (Invitrogen).

Cells (10⁷) were collected for each ChIP sample. Cells were cross-linked at 37°C for 10 minutes with 1% formaldehyde, quenched with 125 mmol/L glycine, washed with PBS, and sonicated in lysis buffer: 50 mmol/L HEPES/KOH, pH 7.5, 140 mmol/L NaCl, 0.1% Na-deoxycholate, 1% Triton X-100, 1 mmol/L EDTA, complete protease inhibitor cocktail (Roche). Samples were incubated with 5 μg of anti-RNAPII (Abcam), anti-SMAD2/3 (Cell Signaling Technology), or anti-E12 (Santa Cruz Biotechnology) overnight and washed 7 times with 20 mmol/L Tris, pH 7.9, 500 mmol/L NaCl, 2 mmol/L EDTA, 1% Triton X-100, and 0.1% SDS. After one wash with Tris-EDTA (TE), DNA was eluted in TE+1% SDS for 1 hour at 65°C, and reverse cross-linked with RNAse A for 4 hours and Proteinase K for 1 hour at 65°C. DNA was purified using a PCR Purification Kit (Qiagen).

Libraries were prepared using the NEBNext ChIP-seq Library Prep Master Mix Set for Illumina (NEB, E6240L) and quality checked using Agilent Technologies 2200 TapeStation to determine fragment size and PicoGreen (Life Technologies/Invitrogen, P7589) to quantify concentration. Samples were pooled and submitted to New York Genome Center for single-end 50 bp sequencing using a HiSeq 2500.

Fifty thousand cells were collected and washed with 1 mL of cold PBS and then 1 mL of ice-cold ATAC Buffer (10 mmol/L Tris, pH 7.4, 10 mmol/L NaCl, and 3 mmol/L MgCl2). Cells were suspended in 50 μL of ATAC Lysis Buffer (10 mmol/L Tris, pH 7.4, 10 mmol/L NaCl, 3 mmol/L MgCl2, 0.1% NP-40 or IGEPAL-Ca630), incubated on ice for 2 minutes. One milliliter of cold ATAC Buffer was added, and nuclei were pelleted at 1,500 rpm for 10 minutes at 4°C in a bucket centrifuge. Nuclei were resuspended in 22.5 μL of the supernatant and transferred to 2.5 μL Tagmentation Enzyme (transposase) and 25 μL of Tagmentation Buffer (Illumina Nextera DNA Sample Preparation Kit). Reaction was incubated at 37°C for 30 minutes. After tagmentation, SDS (final concentration of 0.2%) was added and sample was incubated at room temperature for 5 minutes before purifying with 2X Agencourt AMPure XP beads (Beckman Coulter A63881). Purified samples were eluted in 50 μL of 0.1X TE.

Libraries were prepared with 50 μL sample + 55 μL of NEBNext Q5 Hot Start HiFi PCR Master Mix (NEB, catalogue M0543L) and 5 μL of primer mix using 25 μmol/L of the Nextera primers (52). PCR amplification was performed as previously described for 12 cycles. Samples were purified with 1.5× AMPure XP beads. Concentration was measured using PicoGreen and median fragment size measured using the Agilent D1000 screentape on an Agilent Technologies 2200 TapeStation. Samples were sequenced (paired-end 50 bp) on a HiSeq 2500.

CRISPR screenings were performed as described (35). For the whole-genome screen, the GeCKO v2 mouse whole genome library (Addgene Pooled Library 1000000049) was introduced into an 806-SMAD4pLVX-ID1GFP-Cas9Blast clone at 10% infection efficiency and 200× representation. Cells were plated at 2 × 10⁶ per plate in 2.5 μmol/L MK2206 for 12 hours and then treated with 2.5 μmol/L SB505124 or 100 pmol/L TGFβ for 36 hours prior to collection for FACS. Between rounds of selection, cells were expanded in 15-cm plates, and representation was kept at >200×.

All RNA-seq, ChIP-seq, and ATAC-seq data were deposited in the Gene Expression Omnibus database under accession number GSE112940.

Genetic knockins, knockouts, and knockdowns and overexpression, immunoblotting, IHC, immunofluorescence, flow cytometry, data analysis methods, oligonucleotide, and antibody reagents are described in the Supplementary Methods.

S.D. Leach is a Scientific Advisory Board member at Nybo Pharmaceuticals, has ownership interest (including patents) in Episteme Prognostics, and has an unpaid consultant/advisory board relationship with Episteme Prognostics. C.A. Iacobuzio-Donahue reports receiving commercial research support from Bristol-Myers Squibb. J. Massagué is a Scientific Advisory Board member at Scholar Rock, Inc., is an advisory board member at the Salk Institute, is a Scientific Review Board member at HHMI, is an External Advisory Board member at Institute Research Biomedicine Barcelona, has received honoraria from the speakers' bureau of Columbia University, has ownership interest in Scholar Rock, Inc., and has an unpaid consultant/advisory board relationship with the editorial boards of Cancer Discovery, Cell, The EMBO Journal, Genes & Development, and Journal of Cell Biology. No potential conflicts of interest were disclosed by the other authors.

Conception and design: Y.-H. Huang, J. Massagué

Development of methodology: Y.-H. Huang, N. Lecomte, C.J. David, P.J. Allen

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): Y.-H. Huang, J. Hu, F. Chen, N. Lecomte, H. Basnet, M.D. Witkin, P.J. Allen, S.D. Leach, T.J. Hollmann, C.A. Iacobuzio-Donahue

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): Y.-H. Huang, F. Chen, T.J. Hollmann, C.A. Iacobuzio-Donahue, J. Massagué

Writing, review, and/or revision of the manuscript: Y.-H. Huang, J. Hu, F. Chen, P.J. Allen, T.J. Hollmann, J. Massagué

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): Y.-H. Huang, J. Massagué

Study supervision: J. Massagué

We thank L. Cantley, A. Ventura, R. Benezra, R. Levine, E. Er, R. Koche, and K. Ganesh for thoughtful discussion of this project. We also thank Y. Li, J. Hampton, S.E. Kim, and M. Overholtzer for assistance with experiments. We gratefully acknowledge the support of the Marie-Josée and Henry R. Kravis Center for Molecular Oncology, the Molecular Diagnostics Service, the Center for Epigenetics Research, the Integrated Genomics Core, and the Flow Cytometry Core of MSKCC. This work was funded by NCI grants R01-CA34610 (J. Massagué), R35-CA220508 (C.A. Iacobuzio-Donahue), and P30-CA008748 (MSKCC). Y.-H. Huang was supported by Medical Scientist Training Program grant T32-GM007739 and Predoctoral Fellowship F30-CA203238 from the NCI.

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.