Epithelial NOTCH Signaling Rewires the Tumor Microenvironment of Colorectal Cancer to Drive Poor-Prognosis Subtypes and Metastasis

1CRC is the second most common cause of cancer death, with metastasis the key contributor to CRC-associated death. Thus, a pressing need exists for therapeutic strategies which target disease subtypes with the poorest prognosis. We have generated a genetically engineered mouse model (GEMM) of metastatic CRC, which represents a fundamental advance in preclinical modeling. In doing so, we identify epithelial NOTCH1 signaling as a critical feature of both the poorest prognosis subtypes and of metastatic seeding at the secondary site. Crucially, targeting NOTCH1-driven neutrophil recruitment and TGF-β signaling with clinically relevant small-molecule inhibitors or therapeutic antibodies has a profound impact on metastatic burden in vivo.

2Patient mortality in CRC is closely associated with metastasis, with an overall 5-year survival rate for late-stage patients of 5%–10%. Resection of both primary and metastatic lesions provides the best prognosis for these patients, but post-intervention recurrence is very common due to disseminated latent or therapy-resistant tumor cells. Consequently, preclinical models which faithfully recapitulate the processes of CRC metastasis are required.

3Stratification of CRC by transcriptional profiling has allowed classification of CRC into four consensus molecular subtypes (CMSs). CMS1 tumors are likely to be microsatellite unstable, hyper-mutated, and characterized by lymphocytic infiltration. CMS2 and CMS3 tumors exhibit high levels of WNT signaling, little immune infiltration and intermediate overall and relapse-free survival. Patients with “mesenchymal” CMS4 tumors have the worst overall and relapse-free survival rate and these tumors are characterized by significant fibroblast and innate immune cell infiltration, and elevated TGF-β signaling. As these transcriptional signatures are generated from whole tumors, the presence of stromal cells contributes significantly, confounding analysis. While the contribution of epithelial cells to stromal infiltration/adaptation is not fully understood, cell-intrinsic transcriptional signatures (CRIS) have been shown to have prognostic implications. In particular, the CRIS-B signature predicts poor prognosis and is enriched for signatures associated with epithelial-mesenchymal transition (EMT) and TGF-β signaling. Notably, while CMS4 and CRIS-B are characterized by the same activated programs, the composition of gene signatures is different and CRIS-B is a composition of mainly CMS1 and CMS4 genes. Crucially, mutational data does not stratify the different CRC subtypes and the mechanisms which drive subtypes are not known.

4There are two postulated routes by which metastatic CRC (mCRC) arises. The classical route is initiated by mutations in the APC tumor suppressor gene which is followed by alterations in mitogen-activated protein kinase (MAPK), TP53, and TGF-β signaling during progression. These tumors develop from adenomas with tubular morphology. Recent efforts to model metastatic disease with compounding mutations in the intestine of APC-deficient mice yielded tumors that readily progressed to adenocarcinoma but showed limited metastasis. Intriguingly, if APC-deficient tumors with compounding mutations are propagated ex vivo as organoids and re-implanted into mice, metastasis occurs. Alternatively, CRC progression can be initiated by KRAS or BRAF mutations, with tumor development from adenomas with a serrated morphology. Importantly, patients with serrated adenoma-associated signatures have a poorer prognosis than those with “classical” tubular adenomas. These adenomas may progress to high-grade carcinoma through p16/CDKN2A promoter hyper-methylation and subsequent gene silencing, or via mutation of TP53. Braf-mutant genetically engineered mouse models (GEMMs) of CRC exhibit activated WNT signaling, indicated by nuclear accumulation of β-catenin, while KRAS-mutant tumors appear to develop independently of WNT pathway activation. Nevertheless, these GEMMs develop few distant metastases and have a long latency.

5Despite some caveats, GEMMs are powerful tools to study tumor biology in an autochthonous setting, and are the gold standard in preclinical CRC research. The major weakness of current CRC GEMMs is the lack of a complete adenoma-carcinoma-metastasis cascade and the absence of highly penetrant metastases, particularly to distant organs such as the liver. For this reason, current models can be seen as excellent tools to study early-stage disease rather than malignant progression, with transplantation of tumor-derived organoids currently being the best alternative for analysis of metastatic spread. Transplantation models have highlighted a key role for LGR5+ stem cells in metastasis and have suggested that TGF-β inhibitors may have efficacy in Apc-mutation-driven mCRC.

6Activated NOTCH1 signaling has been observed in CRC and other cancer types. This activation can occur via NOTCH1 ligands on the surface of tumor cells or by components of the TME such as endothelial or innate immune cells. Tumor cell-autonomous signaling can also occur by NOTCH1 receptor copy-number gain, reported in 22% of CRCs, with negative prognostic value. In addition NOTCH1 signaling can be activated via mutation of FBXW7, found in 11% of human CRCs. Activation of NOTCH1 signaling can contribute to cancer cell stemness, invasion, and metastasis. Moreover, recent combination of activated NOTCH1 signaling and Trp53 deletion in the intestine resulted in metastatic disease, albeit with long latency and relatively low penetrance (10% liver metastases), limiting preclinical relevance. Importantly, the molecular mechanism driving NOTCH1-dependent metastasis and the requirement for additional oncogenic driver mutations remains unclear.

7There is an urgent need for improved therapeutic options for patients with advanced mCRC. Currently, molecular subtyping is the most effective strategy to identify patients with the poorest prognosis. For this reason, subtype-specific preclinical models are vital for development of new therapeutic approaches.

8Given associations between NOTCH signaling and CRC we generated a NOTCH-score, based on expression of pathway components, and applied this to The Cancer Genome Atlas (TCGA) human CRC dataset. We found that a high NOTCH-score is significantly associated with CMS4 and poor prognosis (Figures S1A and S1B). Interestingly, when further stratified, the NOTCH-score remained prognostic when KRAS was mutated (Figure S1C), and segregated the poorest prognosis patients in CMS4 (Figure S1D). In addition, we found a high percentage of human CRC metastasis strongly positive for NOTCH1 intracellular domain (N1ICD), indicative of activated NOTCH1 signaling in human CRC metastasis (Figure S1E).

9In light of these observations, we sought to test the functional role of NOTCH1 signaling in CRC. This was achieved using the inducible enterocyte-specific villinCreER to recombine either one copy of Apcfl/+ or activate KrasG12D/+ in combination with deletion of Trp53fl/fl, and overexpression of the transcriptionally active N1ICD (Figure 1A). Consistent with previous studies, villinCreER Trp53fl/fl Rosa26N1icd/+ (PN) mice developed tumors at long latency (Figure 1B). Importantly, all induced villinCreER KrasG12D/+ Trp53fl/fl Rosa26N1icd/+ (KPN) mice that developed intestinal adenocarcinoma exhibited metastases to lymph nodes, lungs, liver, and/or diaphragm at clinical endpoint (Figure 1C). A total of 83% (24/29) of KPN mice had liver metastases, recapitulating human disease (Figures 1D-1F). In contrast, APC-deficient models such as villinCreER Apcfl/+ Trp53fl/fl Rosa26N1icd/+ (APN) or villinCreER Apcfl/+ Trp53fl/fl (AP) did not develop metastases. PN or villinCreER KrasG12D/+ Trp53fl/fl (KP) mice developed few metastases and very rarely to distant sites: 5% liver and ~20% lymph nodes, respectively (Figure 1C). Expression of two copies of the N1icd allele (Figure S2A) or one copy of mutant Trp53fl/R172H in KPN mice did not change survival and/or metastatic burden (Figures S2B and S2C). Furthermore, mutations had only a mild impact on intestinal homeostasis (Figures S2D–S2G). Given that KP and KPN mice exhibited similar latency (Figure 1B), but only KPN mice displayed significantly increased metastatic burden (Figure 1C), we concluded that epithelial NOTCH1 drives metastasis in a setting where Trp53 is mutated and RAS/MAPK signaling is activated.

10Human serrated CRCs have been associated with KRAS mutations, and these morphological features are reported to be recapitulated in the tumors of KrasG12D-driven intestinal GEMMs. Histological analysis of KPN tumors confirmed a serrated morphology of primary tumors, while tumors driven by Apc deletion exhibited a tubular morphology (Figure 2A). Consistent with the metastatic spread of KPN tumors, primary tumors were highly invasive and poorly differentiated, exhibited a high collagen content, significant infiltration of cancer associated fibroblasts (CAFs) and hypoxia, all features typical of advanced disease (Figures 2B–2D). On average, KPN mice developed two tumors per intestine (Figures 2E, 2F, and S3A). We analyzed the expression of the DNA mismatch repair protein MLH1 in primary tumors of APN and KPN mice. Retained expression of MLH1 indicates that these tumors are microsatellite stable (MSS) (Figure S3B). Therefore, KPN tumors represent models of MSS serrated intestinal cancer in which NOTCH1 signaling drives metastasis without impacting tumor initiation.

11Human serrated polyps show reduced WNT pathway activity compared with tubular adenomas which harbor APC mutations (Figure 3A). Comparison of WNT target gene expression between GEMM primary tumors and human serrated adenoma revealed that KPs and KPNs are closely related to serrated tumors (Figure 3B). APN tumors exhibited significant activation of canonical WNT signaling, indicated by nuclear accumulation of β-catenin, with lower activation observed in KPN tumors (Figure 3C). This was reflected by distinct patterns of WNT target gene expression in each primary tumor type (Figures 3B, 3D, and S3B). Moreover, liver metastases from KPN tumors did not have a marked accumulation of nuclear β-catenin, although elevated expression of some WNT targets, including CD44 and SOX9, was observed (Figures 3C and S3C). This recapitulates activation of WNT seen in human serrated tumors and indicates that hyper-activation of epithelial canonical WNT signaling is not essential for metastasis.

12Given these moderate levels of WNT signaling and reported upregulation of WNT ligands or R-spondins in CRC, we examined ligand deregulation or ligand dependence in our models. RNA-sequencing data exhibited profound expression of WNT ligands (Figure S3D). To test ligand functionality we treated KPN mice with LGK974, a clinically relevant PORCUPINE inhibitor, from 85 days after induction, blocking WNT ligand secretion (Figures 3E and S3E). This treatment had no impact on survival or metastatic rate (Figures 3F and 3G; Table S1). To understand the mechanism of WNT ligand independence, we derived organoid cultures from KPN primary tumors. When these organoids were passaged or seeded as single cells, they grew independently of WNT agonist R-spondin1 and were refractory to LGK974 treatment (Figures 3H–3J). Similarly, KP organoids were refractory to LGK974 treatment indicating a NOTCH1-independent mechanism (Figure S3F). This suggests an epithelial cell-intrinsic mechanism of WNT ligand-independent growth, or independence from WNT signaling altogether. To identify drivers of WNT activation, we applied whole-genome sequencing to ten KPN primary tumor-derived organoid lines. This approach confirmed loss of Trp53 and MSS status (overall 1.59 mutations/Mb; coding mutation rate 1.31 mutations/Mb) (Figures S4A–S4C and Table S2). Strikingly, four out of ten (40%) organoid lines had homozygous mutations in Csnk1a1 (encoding casein kinase 1α), a component of the β-catenin destruction complex (Figure 3K). Intriguingly, Csnk1a1 deletion in intestinal epithelial cells has been shown to trigger tumorigenesis only in combination with loss of Trp53. Furthermore, two lines showed Ephb2 missense mutations (Figure 3K), which has been associated with CRC progression. Importantly, while these mutations drive increased expression of selected WNT targets, they appear to be weaker activators of the pathway than Apc loss and critically mimic levels found in human serrated tumors.

13To better understand how our model, and more broadly, NOTCH1 signaling relates to human CRC, we generated transcriptome-wide expression profiles from tumor tissue (consisting of epithelium and stroma). Comparison of signatures generated from both the serrated (KPN) and tubular (APN) tumors (Table S3) with human data revealed a poorer prognosis for patients resembling the KPN signature (Figure 4A), in line with the poor prognosis associated with human serrated CRC. Interestingly, when we analyzed organoid expression profiles derived from APN or KPN tumors (Table S4) this survival segregation still holds (Figure 4B), demonstrating the predictive value of epithelial KPN signatures. Comparison of mouse intestinal tumors with CMSs revealed a NOTCH1-dependent positive correlation between the KPN transcriptome and CMS4, and a negative correlation with CMS2/3 (Figure 4C). Strikingly, tumor models driven by APC loss correlate with CMS2/3 (Figure 4C), which confers better disease prognosis. Moreover, cross-comparison with CRIS signatures revealed that KPN tumors strongly correlate with CRIS-B (Figure 4D), associated with poor prognosis. Gene set enrichment analysis (GSEA) indicates that KPN tumors are enriched for CMS4/CRIS-B-associated signatures such as vascular endothelial growth factor/vascular endothelial growth factor receptor, EMT, and TGF-β activation (Figures 4E and S4D; Table S5). These data demonstrate that the GEMMs described here exhibit transcriptional overlap with the subset of human CRCs with poorest prognosis and that epithelial NOTCH1 is a key driver of those subtypes.

14The current paradigm suggests that stromal signatures are associated with poor prognosis, which is particularly pertinent to the “mesenchymal” CMS4 tumor signature. In light of this, we have compared our transcriptional profiles with data used to characterize human CMS4 CRC as highly enriched for myeloid, angiogenic, inflammatory, fibroblast, and immunosuppressive cell signatures. While many of these features were recapitulated in metastatic KPN tumors (Figure 5A), most were also present in the non-metastatic KP model (Figures 5A and S4E), implying that they may not be determinants of metastatic spread. Importantly, enrichment of a neutrophil signature was associated with metastatic KPN tumors, but not with non-metastatic KP tumors (Figures 5B and S4E). Similar to human serrated adenoma, we have detected neutrophil accumulation in primary tumors, metastases, and systemically in KPN mice (Figures 5C–5F). Given that metastasis in the KPN model was associated with neutrophil infiltration, we assessed expression of chemokines implicated in neutrophil attraction (Figures 5G and S5A) finding increased expression of Cxcl5 in the epithelium of KPN tumors but not of KP (Figures 5G and 5H and Table S6). Cxcl5 expression was correlated with that of its receptor Cxcr2, which is predominantly expressed on neutrophils (Figure S5B). When we examined the expression of neutrophil-associated genes such as ELANE, MPO, and CXCR2 in human CRC, we found significantly increased expression in human CMS4 (Figure S5C). In addition, the neutrophil infiltration-score was able to predict survival in treatment-naive metastases and is significantly associated with CMS4 and CRIS-B (Figures S5D and S5E). Furthermore, neutrophil infiltration of metastases, analyzed by MPO or CXCR2 expression predicts poor survival in an additional cohort (Figure S5F; Table S6). For these reasons, we hypothesized that neutrophils may be a critical driver of NOTCH1-dependent metastasis in CMS4/CRIS-B CRC.

15To test this, we treated KPN mice with AZD5069, a clinically relevant CXCR2 small-molecule (CXCR2sm) inhibitor, which has been shown to block neutrophil recruitment. Importantly, while CXCR2sm treatment, from day 85 post induction, did not impact survival or primary tumor burden in KPN mice, it profoundly reduced metastasis (Figures 5I, S6A, and S6B; Table S1). Short term treatment of KPN tumor-bearing mice with CXCR2sm resulted in reduced neutrophil counts in both the peripheral blood and primary tumors (Figures 5J and 5K), and an increase in CD8+ T cell numbers, compared with vehicle-treated counterparts (Figures 5K and S6C–S6E). The enhanced CD8+ T cell number, thought to create an anti-metastatic microenvironment at a secondary site, was retained in livers at endpoint (Figure S6F).

16Given that CXCR2 expression may not be restricted to neutrophils, we evaluated the impact of neutrophil depletion with a Ly6G-targeting antibody (1A8). Again, metastasis was suppressed when compared with isotype control (2A3), but survival was unaffected (Figures 5I, S6G, and S6H; Table S1). Circulating neutrophils were reduced at endpoint, indicating a sustained effect of the neutralizing antibody (Figure S6I). We detected an increase in CD4+ and CD8+ T cells in the primary tumors of KPN mice treated with 1A8 in the short term (Figures 5K, S6J, and S6K), and increased CD8+ T cells in livers at endpoint (Figure S6L). Together, this indicates that epithelial NOTCH1 triggers CXCR2-dependent Ly6G+ neutrophil accumulation within the pre-metastatic niche and generates an immunosuppressive environment. Therapeutic targeting of neutrophils results in increased infiltrating CD8+ T cells within the pre-metastatic niche and a reduction in metastasis.

17To determine how epithelial NOTCH1 controls metastasis, we examined the transcriptome of tumor-derived KPN versus KP organoids (Figure 6A; Table S7). Increased expression of canonical NOTCH1 target genes, such as Fjx1 and Dtx1, along with an enriched NOTCH-score was observed in KPN organoids, compared with KP counterparts (Figures 6A and S7A). Strikingly in human CRC, the KPN/KP-score predicts poor prognosis and is associated with CMS4, CRIS-B, and neutrophil infiltration (Figures 6B and S7B–S7E). Interestingly, we found significantly increased expression of the gene encoding the Tgfb2 ligand in KPN organoids (Figure 6A). TGFB2 expression predicts poor survival and is significantly correlated with CRIS-B, CMS4, NOTCH-score, and KPN/KP-score in human CRC datasets (Figures 6B, 6C, and S7F–S7J). Furthermore, TGFB2 expression and the KPN/KP-score are also associated with human serrated adenoma (Figure 6D). Together, these data demonstrate a strong association between epithelial NOTCH1-dependent transcriptional signatures and high TGFB2 expression in serrated tumors which exhibit poor outcome and underscores the human relevance of the KPN model.

18To understand how NOTCH1 controls Tgfb2 expression we confirmed NOTCH1-dependent expression of Tgfb2 in KPN organoids (Figure 6E). Moreover, GSEA showed that KPN primary tumors and organoids are associated with TGF-β activation (Figures 4E and 6F). Promoter analysis of the genomic area around the Tgfb2 transcriptional start site for putative RBPJ binding sites revealed a number of canonical RBPJ-DNA binding motifs (Figure 6G; Table S8). Chromatin immunoprecipitation of RBPJ, the key mediator of NOTCH1-mediated transcription, showed binding to the promoter of Tgfb2 in KPN organoids; however, no binding was detected to an upstream region of the Tgfb2 promoter which lacks RBPJ binding sites or a control region (Chrm1) (Figure 6H).

19Alongside the upregulated neutrophil markers in the primary tumors from our GEMMs (Figures 5A–5D and 6I), Tgfb2 was expressed by epithelial cells in KPN tumors (Figure 6J). Tgfb1 expression was comparable in both primary tumors and organoids derived from KPN or KP mice (Figures 6A, 6E and 6I), with predominantly stromal expression in KPN tumors (Figure S7K).

20We next examined the contribution that the epithelial compartment of KPN tumors makes to the TME in metastatic colonization. Previous studies have reported that intra-splenic transplantation of organoids from mouse intestinal tumors, with combined Apc, KrasG12D, Trp53, and TGF-β signaling mutations, is the most efficient means of generating metastases (Figure 7A). We found no difference in the capacity to colonize the liver between organoids derived from KPN or villinCreER Apcfl/fl KrasG12D/+ Trp53fl/fl TrgfbrIfl/fl (AKPT) tumors (Figure 7B), but found a significant increase in Ly6G+ neutrophils in liver metastases formed by KPN organoids (Figures 7C and 7D), and strong epithelial expression of Tgfb2 only in KPN liver metastases (Figure 7E).

21To investigate the impact of neutrophils on metastatic seeding of KPN organoids in this transplantation model we treated recipient mice with either CXCR2sm or a TGFBR1/ALK5 kinase inhibitor (Alk5i) (Figure 7F). Strikingly, both inhibitors reduced the metastatic number and burden significantly (Figures 7G and S7L), and, as observed in the autochthonous GEMM (Figure S6C), long-term treatment with CXCR2sm led to a significant increase of circulating neutrophils (Figure S7M). Intriguingly, Alk5i treatment had no effect on circulating neutrophils but significantly reduced neutrophils infiltrating metastases (Figures 7H and 7I). In addition, we detected an increase in activated CD69+/CD4+, CD69+/CD8+, and IFNγ+ CD4+ type 1 T helper (Th1) cells upon Alk5i treatment (Figure 7J). Importantly, we detected increased CD8+ GzB+ cytotoxic T lymphocytes in metastases treated with Alk5i or CXCR2sm (Figures 7J, S7N, and S7O). This suggests that the effect of CXCR2sm and Alk5i on metastasis is mediated by alleviation of a neutrophil-dependent immunosuppressive microenvironment. This is supported by the finding that CXCR2sm or Alk5i have no effects on metastasis when KPN organoids were transplanted into immune-deficient nude mice which lack T cells (Figure S7P). Together, these data show that the epithelial programs driven by NOTCH1 in KPN tumor cells rewire the TME and generate an immunosuppressive, pro-metastatic environment.

22Given the NOTCH1-dependent expression of Tgfb2 in the KPN model and the profound effect on metastasis in the transplantation model, we examined the importance of TGF-β signaling to NOTCH1-dependent metastasis. We confirmed TGF-β activity in KPN tumors via nuclear localization of phosphorylated SMAD3 (pSMAD3) and TGF-β signaling targets Smad7, CALD1, and IGFBP7 (Figure 8A). We then applied two independent, clinically relevant therapeutic approaches, either targeting of TGFBR1/ALK5 with Alk5i, or with a ligand-trapping antibody targeting TGF-β1/2/3 (1D11) (Figure 8B). In an early-intervention setting, in which mice are treated from 85 days after induction, inhibition of ALK5 resulted in rapid development of highly invasive (T3) colonic tumors, although with markedly reduced metastasis (Figures 8B and 8C; Table S1). Interestingly, targeting the ligands with 1D11 did not result in accelerated tumorigenesis, but significantly reduced metastasis (Figures 8B and 8C; Table S1). In a late intervention approach, when mice were treated with Alk5i from 130 days after induction, developing tumors with a similar latency and comparable tumor burden as vehicle-treated mice (Figures 8B and S8A), but with significantly reduced metastatic penetrance at endpoint (Figure 8C; Table S1). This was associated with a significant reduction in the number of neutrophils in the liver (Figure 8D), with peripheral blood neutrophils being unaffected (Figure S8B). The reduction in liver neutrophils was accompanied by accumulation of CD3+, CD4+, and CD8+ T cells (Figures 8D, S8C, and S8D). Interestingly, no change in primary tumor fibrosis was detected when KPN mice were treated with Alk5i or 1D11 (Figure S8D). Taken together, these data support a strong role for TGF-β signaling in generating an immunosuppressive pro-metastatic microenvironment in the liver by recruiting neutrophils.

23To address the specific function of TGF-β signaling in the neutrophil population of tumor-bearing mice, we transplanted KPN organoids into the colonic submucosa of syngeneic C57BL/6 mice lacking Alk5 expression specifically in Ly6G+ neutrophils (Ly6GCre Alk5fl/fl) (Figure 8E). Engrafted KPN organoids formed primary tumors that were morphologically indistinguishable from those of the autochthonous GEMM (Figures 8F, S8E, and S8F), with epithelial cells from primary tumors and liver metastases also found to express high levels of Tgfb2 (Figure 8G). As predicted by inhibitor experiments, deletion of Alk5 in the neutrophil population had no beneficial impact on survival (Figure S8G), but led to a striking reduction in metastasis (Figure 8H). This was concomitant with reduced tumor-infiltrating neutrophils, although peripheral blood neutrophil counts were maintained (Figure 8I). Reduced neutrophil infiltration was not associated with an alteration in neutrophil maturity, as expression of CD101 (a marker of mature murine neutrophils) was unaffected by ALK5 deficiency (Figure 8J). Importantly, TGF-β signaling inhibition blocks metastasis by reducing neutrophil attraction, rather than by polarization to an anti-tumor phenotype.

24The genetic progression of CRC has been investigated for many years, yet no robust drivers of metastasis have been identified in GEMMs. Importantly, we demonstrate that alteration of epithelial tumor cell-intrinsic signaling can rewire the TME and, in turn, promote metastasis. Notably, this occurs spontaneously only in concert with specific mutations that drive serrated tumors, and not during the progression of WNT-driven tubular adenoma. These observations are consistent with reported differential responses of tubular and serrated adenomas to TGF-β signaling.

25Whole-genome sequencing of tumors from KPN mice revealed relatively low levels of mutation, although recurrent biallelic mutations in Csnk1a1 were observed. This would suggest a further single sporadic mutation, for example in Csnk1a1, could drive rapid progression to adenocarcinoma and metastasis in the KPN model, possibly by generating an inflammatory TME. One could hypothesize that this recapitulates the “Big Bang” model of human CRC in which the key driver mutations occur early, while later, large tumors exhibit neutral evolution.

26The literature regarding NOTCH and WNT interaction in WNT-driven models reveals different roles of NOTCH1 signaling. Epithelial N1icd expression, in combination with Apc1263 allelic mutation, can drive tumor initiation, or inhibit WNT signaling in the ApcMin/+ model at the transcriptional level. Our results demonstrate that, in the context of Apc and Trp53 loss, N1icd has no significant impact on survival or tumor stage. It is interesting to note, however, that we observed reduced levels of selected WNT target genes in KPN tumors and that a specific WNT pathway mutation was selected in emerging tumors. Importantly, human serrated KRAS-mutant tumors predominantly exhibit low levels of nuclear β-catenin, mirrored in KPN tumors. This demonstrates that the model described here shares key cellular and molecular features with human serrated disease. We speculate that WNT ligand inhibition might be a therapeutic option for this type of CRC as reported for CRC with RNF43 mutations or R-spondin fusions. This said, we found no response of these tumors to PORCUPINE inhibition. From our work, it is interesting to propose that the timing of WNT pathway activation is key. When Apc is lost early, adenoma/polyps are formed that require multiple further mutations to progress to adenocarcinoma. In a serrated model, WNT pathway mutations occur much later (which happens in the KPN model) and drive rapid progression to carcinoma and metastasis.

27Critically, our model suggests therapeutic targets in mCRC. NOTCH inhibitors are currently in non-stratified clinical trials for CRC and have shown some minor benefits; our data suggest benefits for NOTCH signaling inhibition in CMS4/CRIS-B patients. Combinatorial inhibition of MEK/ERK and γ-secretase increases efficacy in melanoma, papillary thyroid cancer, and CRC. Identification of tumors with activated NOTCH1 signaling may be aided with the KPN/KP-score described here, or through use of TGF-β2 as surrogate. To generate the NOTCH1 signature we activated N1icd in intestinal enterocytes, mimicking the activation of NOTCH1 signaling in CRC by NOTCH1 receptor copy gain. Activation of NOTCH1 in the KPN model had no impact upon normal homeostasis, possibly because of targeted activation of NOTCH1 in adult tissue rather than during embryonic development. Interestingly, lateral inhibition, which occurs in the normal small intestinal crypt to maintain the stem cell niche, could influence the dynamics of non-cell-autonomous NOTCH1 activation and may occur in malignant progression. This kind of non-cell-autonomous activation should be considered in the future.

28Our work also elucidates a novel non-cell-autonomous role for NOTCH1 in CRC through control of chemokine expression (TGF-β2 and CXCL5). Interestingly, TGF-β2 shows a 100- to 500-fold higher affinity to betaglycan/TGFBRIII than TGF-β1/3. Previously, NOTCH1 signaling has been shown to impact the cellular secretome of multiple cancers. Interestingly, this cancer-associated secretome includes many inflammatory chemokines, similar to the KPN tumors, such as granulocyte-colony stimulating factor and interleukin-6. The potential release of numerous chemokines may help to explain the dramatic effects on neutrophils both in the blood and at metastatic sites in our model. Non-metastatic KP tumors develop a similarly fibroblast-rich stroma as KPN tumors, but lack TGF-β2 expression and neutrophil infiltration. Therefore, the TME generated by CAFs appears insufficient on its own to prime for metastasis but relies on the infiltration of neutrophils. This suggests that our GEMMs may differ from some of the recently described models in which CAFs were important in defining TGF-β sensitivity. Interestingly, these models were initiated by Apc loss (alongside Kras, Tp53, and Smad4/TgfbrII) and predominantly driven by transplantation or induced colitis. It would be interesting to examine the chemokine profiles of these tumors as they may well be differentially dependent on contributions from components of the TME, as indicated by the difference in TGF-β2 expression.

29Though previously thought to be terminally differentiated cells, neutrophils exhibit phenotypic plasticity and adopt distinct mature phenotypes. Accordingly, neutrophils have recently been shown to express various markers of maturity. In particular, TGF-β signaling is thought to be a key regulator of the pro-tumorigenic state of neutrophils, and TGF-β inhibition has been shown to drive increased tumor infiltration of neutrophils. This is in contrast to our findings, where pharmacological inhibition of TGF-β signaling and genetic deletion of TGF-β signaling activity in neutrophils reduces neutrophil numbers in tumors but has no effect on neutrophil maturation status.

30Future studies should characterize the phenotypic traits of neutrophils in the metastatic niche, when compared with the primary tumor and peripheral blood. Our findings suggest that CXCR2/ALK5-expressing neutrophils are important in CMS4/CRIS-B CRC disease progression and in the genesis of CRC metastases. The unresponsiveness of primary tumors to CXCR2sm treatment pre-clinically in this study and other studies might be explained by active immune checkpoints in the primary site, which are lacking in the metastatic site. It should be noted that our mouse model shows high systemic levels of neutrophils and that clinically high neutrophil-to-lymphocyte ratios (NLR) are often associated with the poorest prognosis. It would be of interest to determine whether these NLRs could be used to stratify patients for CXCR2/TGF-β/NOTCH-targeted therapies. In particular, the sensitivity of metastases to neutrophil inhibition holds promise as a potential treatment option for stage II/III CRC patients with undergoing primary tumor resection before treatment with CXCR2 or ALK5 inhibitors. Critically, we highlight a novel targeted therapeutic approach which may compromise the seeding of metastases in a setting in which major primary and metastatic tumor burden is reduced through surgical resection.

31Requests for further information, reagents, and resources should be directed to and will be fulfilled by the Lead Contact, Owen J. Sansom: (o.sansom@beatson.gla.ac.uk).

32Species used: Mus musculus

33All animal experiments were performed in accordance with a UK Home Office licence (Project License 70/8646), adhered to ARRIVE guidelines and were subject to review by the animal welfare and ethical review board of the University of Glasgow. Both genders were induced with a single injection of 2 mg tamoxifen (Sigma-Aldrich, T5648) by intraperitoneal injection at an age of 6 to 12 weeks, all experiments were performed on a C57BL/6 background (N = 5 or more). Mice were sampled at clinical endpoint, which was defined as weight loss and/or hunching and/or cachexia. Mice were censored ≥ 550 days after tamoxifen administration or if sampled not due to intestinal tumor burden or associated metastasis. The alleles used can be found in the Key Resources Table.

34Alk5 inhibitor (Alk5i) (AstraZeneca, AZ12601011) was administered at 50 mg/kg and CXCR2 small molecule (AstraZeneca, AZD5069) at 100 mg/kg, both in 0.5% Hydroxypropyl Methylcellulose (HPMC) and 0.1% Tween-80 twice daily by oral gavage. As vehicle control for Alk5i and CXCR2sm 0.5% HPMC and 0.1% Tween-80 was given with the same regime. Ly6G-antibody (clone 1A8, BioXcell, BE0075-1) or isotype control (clone 2A3, BioXcell, BE0089) were administered three times a week by intraperitoneal injection at 10 mg/kg. TGF-β ligand-antibody (clone 1D11, BioXcell, BE0057) or isotype control (clone MOPC-21, BioXcell, BE0083) were administered three times a week by intraperitoneal injection at 5 mg/kg. LGK974 (Active Biochem, A-1400) was administered at 5 mg/kg, in 0.5% Methylcellulose (MC) and 0.5% Tween-80 twice daily by oral gavage. As vehicle control 0.5% MC and 0.5% Tween-80 was given with the same regime. Treatments were started 85 or 130 days after initial tamoxifen injection; short term treatments of tumor bearing mice were started when tumors were palpable.

3546 patients who underwent synchronous resection of colorectal primary tumor and liver metastases between April 2002 and June 2010 at Glasgow Royal Infirmary were included in the study, details can be found in Table S6. Patients were identified from a prospectively maintained database and represent a consecutive cohort of resected patients. Application to access patient tissue was approved by the NHS Greater Glasgow and Clyde biorepository and ethical approval granted in biorepository application #357 and informed consent was obtained from all subjects. Patients were followed up at one month, six monthly until two years, and thereafter annually until five years at which point they were discharged. Recurrence data, morbidity, and mortality was prospectively collected. Information on date and cause of death was determined via access to the NHS Greater Glasgow and Clyde clinical portal. Death records were complete until 1st November 2017, which served as the censor date.

36Human liver metastases were anonymised, five micron-thick, formalin fixed and paraffin embedded sections of liver containing metastatic colorectal carcinoma from partial hepatectomy specimens were stained for N1ICD. The use of the human material was approved by the Lothian NRS Human Annotated Bioresource and informed consent was obtained from all subjects (ethical review number 15/ES/0094).

37T staging of tumors was performed by a boarded pathologist according to the following parameters included in the classical TNM classification; T0, no evidence of primary tumor. Tis, Carcinoma in situ: intraepithelial or invasion of the lamina propria (i.e. no extension through the muscularis mucosae and therefore the submucosa). T1, Tumor invades submucosa. T2, Tumor invades muscularis propria. T3, Tumor invades into the subserosa. T4, Tumor invades/perforates the visceral peritoneum and into other adjacent organs/structures.

38Tumor differentiation scoring was performed by a boarded pathologist according to the following parameters; well differentiated tumors exhibit clear glandular differentiation in >95% of the tumor. Moderately differentiated tumors exhibit glandular differentiation in 50-95% of the tumor. Poorly differentiated carcinomas exhibit glandular differentiation in 5-50% of the tumor.

39Blood was collected in EDTA columns after cardiac puncture. Blood samples were analyzed with IDEXX ProCyte Dx.

40Advanced DMEM/F12 was supplemented with penicillin/streptomycin (100 U/ml / 100 μg/ml) (15140122), 2 mM L-Glutamine (25030081), 10 mM HEPES (15630080), N2-supplement (17502001) and B27-supplement (17504044) (all ordered from Gibco, Life Technologies or ThermoFisher-Scientific) and from here on is referred to as ADF. Complete ADF was prepared by supplementing ADF with 50 ng/ml Recombinant Human EGF (Peprotech, AF-100-15), 100 ng/ml Recombinant Murine Noggin (Peprotech, 250-38) and 500 ng/ml Recombinant Mouse R-spondin-1 (R&D systems, 3474-RS). Intestinal epithelium extraction and culture conditions were previously described. These culture conditions were used unless stated differently in figure legends.

41Tumors were cut into small fragments and washed five times in PBS. Tumor fragments were incubated in 5 ml 10x Trypsin (5mg/ml, Gibco), 1x DNase buffer and 200U recombinant DNase I (Roche, 04716728001) at 37°C for 30 minutes. To further dissociate tumor fragments, 5 ml ADF was added and tumor fragments were shaken vigorously. This step was repeated five times. After aspirating the supernatant and re-suspending the pellet in 10 ml ADF, the suspension was passed through a 70 μm cell strainer. The cell pellet was re-suspended in Matrigel (BD Bioscience, 356231) according to pellet volume and seeded. Organoids/spheroids were cultured in complete ADF at 37°C, 5% CO2, 21% O2.

42Organoids were harvested and dissociated by fiercely pipetting. Organoids were washed twice with PBS before being dissociated into single cells by incubating in 2 ml 10x Trypsin (5mg/ml), 1x DNase buffer and 200U recombinant DNase I (Roche, 04716728001) at 37°C for 7 minutes. Cells were passed through a 40 μm cell strainer before 1000 single cells were seeded in 20 μl Matrigel in a 24-wells culture plate. LGK974 10 μM (Active Biochem, A-1400) and appropriate volumes of vehicle (DMSO) were added at the moment of single cell seeding. Organoid forming capacity was assessed after one week by measurement of diameter and counting the number of organoids formed in each culture condition.

43To prepare the cell suspension, KPN (liver metastases derived; C57BL/6, N=10) or villinCreER Apcfl/fl KrasG12D/+ Trp53fl/fl TrgfbrIfl/fl (AKPT; small intestinal derived; C57BL/6, N=7) organoids were cultured in conditions as described above without R-spondin. Tumor cells were harvested and washed with PBS and trypsinized with 0.25% Trypsin in PBS-EDTA for 7 minutes at 37°C. After trypsinization, cells were washed and passed through a 40 μm cell strainer and counted using a haemocytometer.

44C57BL/6 or CD-1/Nude mice (6-12 weeks old males; Charles River, UK) were anesthetized with isoflurane, and a laparotomy was performed to gain access to the spleen. 5x105 single cells in 50 μL PBS were injected into the spleen after which the incision was sutured. The mice were sampled four weeks post transplantation. Organoid donor and recipient mice were sex matched.

45Colonic sub-mucosal injections of organoids were performed as previously described, using a Karl Storz TELE PACK VET X LED endoscopic video unit. KPN liver metastases derived (C57BL/6, N=10) organoids, cultured in conditions as described above without R-spondin, were harvested and dissociated by fiercely pipetting. Organoids were washed twice with PBS before being injected. Approximately 500 organoids in 70 μl PBS were injected in a single injection. At clinical end point tumors and metastasis were quantified.

46Tumor samples were dissected and digested using the Mouse Tumor Dissociation Kit (Miltenyi Biotec, 130-096-730) and the GentleMACS Octo Dissociator with Heaters (Miltenyi Biotec, 130-096-427), using the 37C_m_TDK_1 programme. The cells were passed through a 70 μm cell strainer and then counted. Two million cells were stained with LIVE/DEAD fixable near-IR stain kit (ThermoFisher, L10119) at 1:1000 dilution in 100 μl PBS in the dark for 20 minutes at 4°C, then washed with PBS 1% BSA. TruStain FcX anti-mouse CD16/32 (Biolegend, 101320) was used at 1:200 in 50 μl PBS 1% BSA to block CD16/32 activity, in the dark for 15 minutes at 4°C. After that incubation time, 50 μl of the antibody staining mixes were added: for the neutrophil panel, CD45 (ThermoFisher, 63-0451-82), CD48 (Biolegend, 103424), CD11b (Biolegend, 101243), Ly6G (BD Biosciences, 563978), and CD101 (ThermoFisher, 12-1011-82). For the T cell panel, CD45 (ThermoFisher, 63-0451-82), CD3 (Biolegend, 100228), CD4 (Biolegend, 100510), CD8a (Biolegend, 100730), and CD69 (Biolegend, 104508). The cells were incubated in the dark for 30 minutes at 4°C. The cells were then washed with PBS and re-suspended in 50 μl PBS. To fix the cells, 50 μl of PBS 4% paraformaldehyde were added to the re-suspended cells, followed by incubation at room temperature in the dark for 15 minutes. The cells were washed and re-suspended in PBS. Only for the T cell panel, the cells were re-suspended in Permeabilization Buffer (ThermoFisher, 00-8333-56) instead. The cells were stained intracellularly in 1x Permeabilization Buffer with IFNγ (ThermoFisher, 25-7311-41), and Granzyme B (Biolegend, 515406) in the dark for 30 minutes at 4°C. The cells were washed in Permeabilization Buffer. Finally, the cells were washed in PBS and re-suspended in PBS for flow cytometry acquisition. Both neutrophil and T cell populations were identified with the following initial gating strategy: doublet discrimination by discrepancy between FSC-A and FSC-H signals; live cells: CD45+. Subsequently, for neutrophils: CD48-/loLy6G+, CD11b+Ly6G+ to confirm the neutrophil identity, and CD101 to characterize the neutrophils. Alternatively, for T cells: CD3+, and CD4+ and CD8+ subsets. For both subsets, CD69 was used to analyse for activation. For the CD4 subset, IFNγ was used as an additional activation marker, while Granzyme B was used for the CD8+ subset. The data were acquired with the BD LSR Fortessa flow cytometer (BD Biosciences) and analyzed with FlowJov10.4.2.

47RNA was isolated using the Qiagen RNeasy Mini kit (Qiagen, 74104) according to the manufacturer’s protocol including the optional DNA degradation step using the Qiagen RNase-Free DNase kit (Qiagen, 79254). Cell pellets or tissue were lysed using the Precellys lysing kit (Bertin Instruments, KT03961-1-003-2) in a Precellys Evolution machine (Bertin Instruments).

48Organoids of the respective genotype, at comparable passage ~ 5 were sampled 72 hours post seeding. Organoid pellets were snap-frozen and RNA was isolated as described above.

49RNA of whole tumor samples was isolated at endpoint from genotypes as indicated in figure legends and conserved in RNAlater (Sigma, R0901) at -80°C until further use for RNA isolation as described above. For sequencing tumor fragments were excised from the tumor centre to minimize effects of intra-tumor heterogeneity. Primary tumors from KPN (villinCreER KrasG12D/+ Trp53fl/fl R26N1icd/+; small intestine), KP (villinCreER KrasG12D/+ Trp53fl/fl; small intestine), APN (villinCreER Apcfl/+ Trp53fl/fl Rosa26N1icd/+; small intestine) and AP (villinCreER Apcfl/+ Trp53fl/fl; small intestine) tumors were sampled without exclusion of submucosa or muscularis propria.

50cDNA was generated by reverse transcription of the isolated RNA using the M-MuLV-Reverse Transcriptase kit (ThermoFisher-Scientific, 28025013) according to the manufacturer’s protocol. qPCR was performed using the DyNAmo HS SYBR Green qPCR kit (ThermoFisher-Scientific, F410) according to the manufacturer’s protocol. CT-values were normalized to β-Actin (Actb) CT-values. mRNA expression levels were calculated according to the ΔCT method and expressed as 2ˆ(-ΔCT). Primers sequences can be found in Table S8.

51For chromatin immunoprecipitation (ChIP) the SimpleChIP® Plus Enzymatic Chromatin IP Kit (Cell Signaling Technology, 9005) protocol was used according to the manufacturer’s instructions. In brief, KPN organoids were grown in medium conditions as described above. Cells were cross-linked for 10 minutes at room temperature and chromatin was fragmented by micrococcus nucleases followed by three sonication cycles to generate DNA fragments. Incubation with RBPSUH (D10A4) XP® Rabbit mAb (Cell Signaling Technology, 5313) or recommended concentration of rabbit normal IgG control (Cell Signaling Technology, 2729) for 16 hours at 4°C was performed. The sequences of oligonucleotides used as qChIP primers are listed in Table S8.

52Tissues were fixed in 10% neutral buffered formalin and processed by standard histology processing techniques. The following antibodies were used: CD3 (AbCam Ab16669, pH6 1:50), CD4 (eBioscience 14-9766-82, ER2 Leica, 1:500), CD8 (eBioscience 14-0808-82, ER2 Leica,1:500), S100A9 (Santa Cruz sc-8115, pH6 1:1000), Ly6G (clone 1A8, 2B Scientific BE0075-1, ER2 Leica, 1:60000), CD44 (BD Biosciences 550538, pH6, 1:250), SOX9 (Millipore AB5535, pH6, 1:500), β-catenin (BD Biosciences 610154, pH8, 1:50), MLH11 (Abcam ab92312, pH6, 1:200), CALD1 (Sigma HPA008066, ER2 Leica, 1:400), IGFBP7 (Sigma HPA002196, ER2 Leica, 1:100), pSMAD3 (Abcam Ab52903, pH6, 1:40), αSMA (Sigma-Aldrich A2547, pH6, 1:25000), N1ICD (D3B8 Cell Signaling Technology 4147, Protaqs IX, BioCyc, 401603692, 1:50).

53To stain collagen or fibrin presence within tissue sections Picro Sirius Red staining technique was used. Briefly, de-waxed slides were immersed in Picro Sirius Red solution for 2 hours. Picro Sirius Red Solution: 0.1% Direct red 80 (Sigma, 41496LH) in distilled water and 0.1% Fast green FCF (Raymond Lamb, S142-2) in distilled water were mixed in equal volumes and then diluted 1:9 with Aqueous Picric acid solution. Post staining slides were dehydrated according to standard protocols and mounted for analysis.

54Hypoxia was detected by administration of Hypoxyprobe (Hypoxyprobe HP1-100; 100 μl intraperitoneal) 1 hour before sampling and detected using Hypoxyprobe recognizing antibody (Hypoxyprobe HP1-100, pH6, 1:150).

55In situ hybridisation (ISH) analysis was performed using the RNAscope 2.5 LS (Brown, 322100) detection kit (Advanced Cell Diagnostics, Hayward, CA) on a Leica Bond Rx autostainer strictly according to the manufacturer's instructions. Staining was performed on 4 μm formalin fixed paraffin sections which were cut and then placed in a 60°C oven for 2 hours prior to staining. To ensure the quality and integrity of the available RNA the tissue being investigated was tested with the positive control probe (mm-Ppib, 313918). Only after probe quality control were the results evaluated. To further ensure accuracy and integrity of the staining a negative control probe (mm-DapB, 312038) was used to confirm that the tissue staining seen was accurate due to binding with the target probe and not non-specific. Probes: Cxcl1 (407728), Cxcl2 (437588), Cxcl3 (492758), Cxcl5 (467448), Cxcr2 (487678), Axin2 (400338), Lgr5 (312178), c-Myc (413458), Smad7 (429418), Tgfb1 (407758), Tgfb2 (406188), positive control probe Ppib (313918) and negative control probe DapB (312038).

56The quality of the purified RNA was tested on an Agilent 2200 Tapestation using RNA screen tape. Libraries for cluster generation and DNA sequencing were prepared following an adapted method from the Illumina TruSeq RNA LT Kit. Quality and quantity of the DNA libraries was assessed on a Agilent 2200 Tapestation (D1000 screentape) and Qubit (Thermo Fisher Scientific) respectively. The libraries were run on the Illumina Next Seq 500 using the High Output 75 cycles kit (2x36cycles, paired end reads, single index). Quality checks on the raw RNA-Seq data files were done using fastqc version 0.11.2 and fastq_screen version 0.11.3. RNA-seq paired-end reads were aligned to the GRCh38 version of the mouse genome using tophat2 version 2.0.13 with Bowtie version 2.2.4.0. Expression levels were determined and statistically analyzed by a combination of HTSeq version 0.6.1, the R environment, version 3.2.2, utilizing packages from the Bioconductor data analysis suite and differential gene expression analysis based on the negative binomial distribution using the DESeq2. All RNA-sequencing data have been deposited in the ArrayExpress database under accession number E-MTAB-6363.

57Macroscopic intestinal tumor or metastases were analyzed for size and number and tumor burden or metastases burden was calculated as number of tumors times tumor size. All metastases were confirmed histologically.

58IHC and ISH (RNA-scope) images were digitalized using a SCN400F slide scanner (Leica Microsystems, Milton Keynes, UK) at 20x (IHC) or 40x (ISH) resolution. Scanned images were analyzed using HALO Image analysis software (V2.0.1145, Indica Labs). Tumors were analyzed for the percentage of positive cells for N1ICD, CD3, CD4, CD8a, Ly6G and S100A9. Tumor areas were manually defined using the HALO software and scoring was performed in a blinded manner for all samples. β-catenin staining was analyzed manually and considered as positive when > 10% of the tumor area was strongly positive for nuclear β-catenin.

59For Cxcl5/Cxcr2 co-analysis serial sections (3.5 μm sections) were stained for Cxcl5 and Cxcr2 and scanned at 40x magnification. Slides were then automatically aligned utilising the image registration module within the HALO package. Sequential, non-overlapping, paired fields of view were then individually scored.

60CRC patient data were obtained from different public sources. Expression data and clinical/genetic annotation from the TCGA project were downloaded from the FIREHOSE repository (https://gdac.broadinstitute.org/). This included RNA-seq expression data generated by the Illumina HiSeq (n=326) and Genome Analyzer (n=172) platforms (RSEM normalized data). After log transformation, data from both platforms were combined into a single dataset (n=498), by correcting platform-specific effects with the ComBat algorithm as implemented in the sva R package. From the NCBI GEO repository microarray expression data and clinical/genetic annotations for the following 11 datasets (total n=1981): GSE39582 (n=585), GSE13294 (n=294), GSE14333 (n=157), GSE17536 (n=177), GSE17537 (N=55), GSE20916 (n=81), GSE2109 (n=315), GSE23878 (n=35), GSE33113 (n=90), GSE35896 (n=62) and GSE37892 (n=130). The microarray data were normalized, summarized and log2 transformed using robust multiarray analysis (rma) and batch effects (both between and, where present, within dataset) were removed using Combat. After normalization the probe sets were annotated using the hgu133plus2.db annotation R package. In the case of multiple probe sets interrogating a specific gene, the probe set with the highest mean intensity was selected as representative for that gene. CMS labels for all datasets were obtained from Guinney et al. and CRIS labels were obtained from Isella et al.. A processed gene expression dataset of CRC liver metastases (E-TABM-1112, n=120) was obtained from ArrayExpress.

61Mouse model gene expression signatures were generated by quantifying differential gene expression data using the DESeq2 R package. Patient-based CMS signatures were derived from the TCGA dataset where differential expression was determined using the limma R package. Human-to-mouse orthologue mappings were obtained from Biomart (http://biomart.org) using the interface provided by the biomaRt R package. In case of one-to-many human-to-mouse mappings the mapping with the highest homology percentage was selected. For the mouse model-patient correlation analysis 75 genes were selected that were the up-regulated most significantly (> 0.75 log fold change) in each of the four CMSs (300 genes total) and five CRISs (375 genes total) and calculated the Pearson correlation coefficients of the mouse model signatures with the patient CMS and CRIS signatures.

62Infiltration scores were calculated with the MCPcounter R package and then standardised per cell type.

63KPN (n=9) vs. APN (n=3) tumor and organoid KPN (n=3) and APN (n=4) gene expression signatures were determined using the same procedure described in the previous paragraph. The 500 most significantly regulated genes were selected to construct the KPN vs. APN signature. No formal p value cut-off was used but all Benjamini-Hochberg adjusted p values were below 1x10-6. TCGA and NCBI GEO expression datasets (excluding samples of tumor stage IV, without recurrence data or those without a CMS label n=1018) were combined into a single dataset correcting batch effect using ComBat. Expression values were mean-centred gene-wise and individual samples were scored by calculating the Pearson correlation coefficient with the KPN vs. APN signatures. Samples with a correlation score > 0.1 were assigned to the positive group, those with a correlation score < -0.1 to the negative group. Kaplan-Meier plots were generated using the survival R package and the survival distributions were compared with the log-rank test.

64Geneset enrichment analyses (GSEA) were run using the GSA R package with sample permutation (10,000 permutations) and gene standardization using all genes in the expression dataset, taking the unadjusted p values as output. The input genesets are listed in Table S5. The two TGF-β genesets used to asses TGF-β signaling activity in organoids where obtained from.

65The NOTCH-score was calculated by summing the standardized expression of a panel of NOTCH related genes: JAG1, JAG2, DLL1, DLL3, DLL4, NOTCH1, NOTCH2, NOTCH3, NOTCH4, HES1, HES2, HEY1, HEY2 and DTX1. For the recurrence-free survival analysis of CRC patients from the TCGA dataset, patients were stratified using the within-group median NOTCH-score. KRAS mutants were called when mutations encoding amino acids G12, G13 and A143 were present. Kaplan-Meier plots were generated using the survival R package and the survival distributions were compared with the log-rank test.

66The WNT-score was calculated by summing the standardised expression of a panel of WNT related genes: ASCL2, AXIN2, BMP4, MRTO4, HILPDA, NOP16, KITLG, LGR5, MYC, NOP2, PPIF, SOX4, PAAF1, ZIC2 & ZNRF3.

67A KPN (n=3) vs. KP (n=3) organoid gene signature was generated by quantifying differential gene expression data using the DESeq2 R package. We selected the 100 most significantly up-regulated genes (log fold change > 1). No formal p value cut-off was used but all Benjamini-Hochberg adjusted p values were below 1x10-2. The KPN/KP-score was calculated by summing the standardised expression of the genes in the KPN vs. KP score (Table S7).

68The human serrated signature was derived by comparing WNT target gene expression in serrated and tubular adenomas (GSE45270 (n=13) and GSE79460 (n=16)) using the limma R package. The microarray data were normalized summarized and log2 transformed using robust multi array analysis (rma) and batch effects were removed using Combat. After normalization the probe sets were annotated using the hgu133plus2.db annotation R package. In case of multiple probe sets interrogating a specific gene, the probe set with the highest mean intensity was selected as representative for that gene. The mouse model signatures were generated using the DESeq2 R package.

69Organoid (tumor derived) and tail DNA were extracted using DNAeasy kit (Qiagen, 69504) as per the manufacturer’s instructions. DNA concentration and quality were determined by Nanodrop spectrophotometry and by PicoGreen (Invitrogen, P11496). Whole genome sequencing was performed using 151bp paired-end sequencing on the Illumina HiSeq X platform. Short insert libraries were constructed using prepared flow cells, and clusters generated using standard methods. Samples were sequenced at an average depth of 39x with a minimum coverage of 26x. Data were mapped to the mouse reference genome (GRCm38) using the bwa-mem alignment tool. All whole genome sequencing data have been deposited in the European Nucleotide Archive (ENA) under ENA accession ID: ERP040713.

70Somatic variants were detected using CaVEMan, an expectation maximization–based somatic substitution detection algorithm. Candidate somatic variants were then filtered for quality and to remove known mouse genome variations. Single point mutations overlapping known structural variants in any of the mouse genomes were also removed due to high misalignment rates in these regions. Small insertion and deletion (indel) detection was performed using the cgp-pindel pipeline (v0.2.4w). Detected indels were then filtered for quality, sequence coverage in both tumor and normal, strand bias and for overlap with known simple repeats or indels in the in-house normal panel. Selected mutations were confirmed by Sanger sequencing.

71Tumor specific copy number changes were reported using Control free software.

72The datasets generated during this study are available at the European Nucleotide Archive: ERP040713 and ArrayExpress: E-MTAB-6363.

73The codes supporting the current study have not been deposited in a public repository because of dependencies on in-house software and data infrastructure, but are available from the corresponding author on request.

74Statistical analyses were performed using GraphPad Prism software (v7.03 GraphPad software, La Jolla, CA, USA) and R (version 3.4.3) performing tests as indicated and were considered statistically significant, with ∗ p<0.05, ∗∗ p<0.01, ∗∗∗ p<0.001, ∗∗∗∗ p<0.001.