Stem-like cells from Cancer of Unknown Primary are endowed with distinctive hypermetastatic properties and unveil liability to MEK inhibition


 Cancers of Unknown Primary (CUPs), featuring metastatic dissemination in the absence of a primary tumor, are a biological enigma and a fatal disease. We propose that CUPs are a distinct, yet unrecognized, pathological entity originating from stem-like cells endowed with unique properties. These cells can be isolated in vitro (agnospheres) and propagated in vivo by serial transplantation, displaying exceptionally high tumorigenicity. After subcutaneous engraftment, agnospheres recapitulated the CUP phenotype, by showing the ability to spontaneously and quickly disseminate, and form widespread established metastases. Regardless of different genetic backgrounds, agnospheres invariably displayed cell-autonomous proliferation and self-renewal, mostly relying on unrestrained activation of the MAP kinase/MYC axis, which confers sensitivity to MEK inhibitors in vitro and in vivo. Such sensitivity is associated with a transcriptomic signature predicting that 75% of CUP patients could be eligible to MEK inhibition. These data shed light on CUP biology and unveil an opportunity for therapeutic intervention.

sensitivity to MEK inhibitors in vitro and in vivo. Such sensitivity is associated with a transcriptomic signature predicting that 75% of CUP patients could be eligible to MEK inhibition. These data shed light on CUP biology and unveil an opportunity for therapeutic intervention.

INTRODUCTION
Cancer of Unknown Primary (CUP) is the diagnosis received by patients that present multiple metastases in the absence of a primary tumor anatomically or histologically recognizable through a standardized work-up that includes thorough body imaging and tissue immunohistochemistry (1)(2)(3)(4). Although implying some uncertainties, the definition of CUP applies to 1-2% of all 5 malignancies and features a dismal median survival (< 1 year). The clinical course is however less aggressive in a subset of cases (15-20%) displaying features reminiscent of some specific tissue of origin such as the neuroendocrine carcinoma (2,4).
So far, CUPs were investigated with two main pragmatic clinical aims, such as to uncover the molecular or epigenetic signature of a putative 'tissue of origin', and treat each CUP 10 as a high grade metastatic tumor of that tissue or organ (5)(6)(7), or to identify mutated cancer genes and inform personalized targeted therapies (8, 9). However, apart from a few exceptions, prediction of a putative origin and application of tailored therapies negligibly extended the overall survival of CUP patients, which remain among the poorest in oncology (3,4,10). To make progress, it seems critical to adopt an alternative approach, and investigate CUPs as a group of 15 tumors sharing the common ability to disseminate in a way that is (i) early and rapidly progressing; (ii) unrestrained by those barriers in the tissue of origin that facilitate the growth of a detectable primary mass; and (iii) associated with an early cell differentiation block, which precludes formation of recognizable tissues.
In this study, we isolated a panel of human CUP-initiating, stem-like cells, named 20 'agnospheres', able to reproduce a faithful disease model, featuring early dissemination and formation of established multi-organ metastases. We discovered that agnosphere properties are cell-autonomous and converge on constitutive activation of the proliferative MAP kinase 4 pathway, sustaining expression and activity of the MYC proto-oncogene. MEK inhibition with trametinib caused agnosphere cell death, and necrosis of experimental CUPs, without inducing negative feedback mechanisms. Importantly, we found that response to trametinib is foreseen by an originally set-up gene expression signature, which, applied to patients' tissues, predicts sensitivity in a high percentage of CUPs, suggesting that common disease molecular mechanisms 5 are amply shared.

Make-up of the Original Tumors
In a cohort of 61 early metastatic cancer patients enrolled at our institution, 27 were diagnosed as CUPs through a rigorous ad excludendum protocol that ruled out the presence of a primary tumor 5 or a defined tissue of origin (2,4). In the CUP cohort, fresh human specimen for biological studies could be obtained only in 8 cases (29%), including 5 biopsies and 3 surgeries (Supplementary Tables S1-S2). All 8 samples were transplanted in immunocompromised mice (patient-derived xenografts, PDX), resulting in 6/8 engraftments (75% success rate); interestingly, CUP samples that did not engraft belonged to a patients' subset with prolonged 10 survival: AGN47 (OS >40m) and AGN913 (OS >84m). From PDX, or fresh human samples whenever possible, we tried to obtain spheroid cultures of stem-like cells, named 'agnospheres', with a 75% of successful rate ( Fig. 1A and Supplementary Table S2). Briefly, 4 out of 5 patients that provided agnospheres (AGN901, AGN906, AGN43, AGN67 and AGN914) displayed the typical aggressive CUP presentation, featuring (i) multiple metastases; (ii) the histological aspect 15 of poorly differentiated carcinomas lacking expression of markers associated with specific organs or tissues, and (iii) rapidly lethal clinical outcome (average overall survival: 11.5 months) (Supplementary Table S1). Agnospheres derived from the above patients are hereafter indicated as AS901, AS906, AS43, AS67 and AS914, respectively. As long-survivor patient AGN47 belongs to a CUP subtype with neuroendocrine differentiation, the corresponding agnosphere is 20 henceforth indicated as N-AS47.
We performed the genetic characterization of agnospheres and available human tissues from which they were isolated (hereafter indicated as 'original tissues'), by whole exome sequencing (Supplementary Table S3 and Dataset 1): AS901 and AS67 were hypermutated, consistently with the presence of a POLE mutation in AS901 and POLE, POLQ in AS67; AS906, AS43 and AS914 harbored different combinations of cancer-associated genes, some of which are among the ten most frequently mutated genes in CUPs, being enriched in CUPs compared with metastases from cancers of known origin ( Supplementary Fig. S1). Finally, N-5 AS47 harbored no mutation in known oncogenes or tumor suppressor genes (Supplementary Table S3 and Dataset 1). Copy number variation analysis of the two most commonly altered tumor suppressor genes unveiled heterozygous loss of TP3 accompanied by TP53 hemizygous mutation in AS43, and heterozygous PTEN loss in both AS914 and N-AS47 ( Fig. 1B and Supplementary Table S3). The allele frequency of mutations is consistent with heterozygosity 10 (oncogenes) or hemizygosity (TSG) in agnospheres (Supplementary Table S3), supporting the pathogenic meaning of the genetic alterations and suggesting that agnospheres are monoclonal cell populations, at least relatively to driving genetic lesions. Interestingly, we recently showed that the multiple CUP metastases can be highly genetically related, in particular sharing all the mutations in driver genes (12) thus supporting the conclusion that agnospheres, although derived 15 from a single metastatic lesion can be representative of all patient's metastases.
Conventional and spectral karyotypic analysis showed that AS901, accordingly with its hypermutated state, did not display evident aberrations of chromosome number and structure ( Fig. 1C). AS906, AS43, AS67 (although hypermutated), and AS914 exhibited slightly increased ploidy and multiple numerical and/or structural chromosome aberrations, while N-AS47 showed 20 multiple numerical but few structural aberrations (Fig. 1C-D). Although the sample size is too small to generalize conclusions, the overall genetic analysis indicated mutational and karyotypic heterogeneity among agnospheres, consistently with previous observations in 200 CUP patients (8).
Agnospheres Are Enriched in Stem-like Cells that Self-sustain Their Long-term Propagation 5 Agnospheres were isolated and propagated in suspension, in highly stringent stem culture conditions (13). Surprisingly, all agnospheres were capable of self-renewing and long-term propagating at clonal density in the absence of any exogenous growth factors, such as EGF or FGF, which are usually required for isolation and maintenance of the stem phenotype in cancer stem cells from highly aggressive tumors (11, [13][14][15]. Agnospheres expressed high levels of 10 transcription factors known to be major regulators of self-renewal and stem identity in normal and neoplastic tissues, such as Polycomb repressors (16) and reprogramming transcription factors (17,18). Polycomb repressor EZH2 and BMI1 mRNAs and proteins were highly expressed in all agnospheres and in the majority of agnosphere cells ( Fig Among reprogramming transcription factors, prominent and widespread expression of MYC genes in the absence of gene amplifications ( Supplementary Fig. S2B), together with the related set of targeted genes, was observed in all agnospheres ( Fig. 2A-D). Co-expression of MYC with OCT4, SOX2, and KLF4 was observed only in a fraction of agnosphere cells, likely 20 marking a subpopulation with enhanced stem traits ( Fig. 2B and Supplementary Fig. S2C).
In the transcriptome of agnospheres and original tissues we also evaluated the expression of an embryonic stem cell (ES) expression signature, which was previously shown to be enriched 8 in high-grade, aggressive tumors of various origins (19). The signature is based on 13 gene-sets collected in 4 main groups, including the two above mentioned Polycomb and MYC target groups, a group including targets of NANOG, OCT4 and SOX2, and a group of genes expressed in cultured human ES cells (19). Our analysis showed that this signature is highly enriched in agnospheres, as well as in tumorspheres derived from metastases of a colorectal cancer and a 5 melanoma (Fig. 2B). Interestingly, the signature was highly enriched also in original CUP tissues, with the exception of ES-expressed genes group, according to previous observation in breast cancer (14) (Supplementary Fig. S2D and Dataset 2). This suggests that CUP tissues contain a high proportion of stem-like cells, accordingly to what observed in other poorly differentiated, high-grade tumors (19). 10 Given the interconnection between stem phenotype, metastatic ability and epithelialmesenchymal transition (EMT) (20), in the agnosphere transcriptome we assessed the expression of EMT core transcription factors (ZEB 1 and 2, SNAI 1 and 2 and TWIST1), and other transcription factors alternatively associated with EMT ( Fig. 2E) (21). The global level of EMT core transcription factors was relatively weak in the subgroup of adenocarcinomas and 15 neuroendocrine agnospheres. However, in AS906 and AS914, which derived from carcinomas with sarcomatoid features (Supplementary Table S1), EMT transcription factors such as ZEB1 and SLUG, and mesenchymal markers CD44 and Vimentin were highly expressed, while epithelial cell adhesion markers EpCAM and E-cadherin (CDH1) were barely detectable (Fig.   2C, E and Supplementary Fig. S2E). These features are consistent with the phenotype of AS906 20 and AS914, which formed loose spheroids in culture (Fig. 1A). Interestingly, YAP/TAZ, which has been implied in both EMT and reprogramming of differentiated cells into stem cells (22), was highly expressed in all aggressive and undifferentiated CUPs but almost completely absent in N-AS47 (Fig. 2C, E).
Finally, in the agnosphere transcriptome, we analyzed also the expression of cell surface and functional stem-cell markers previously used for identification and/or prospective isolation of cancer stem cells, finding significant expression of general markers such as CD24, CD47, 5 CD98, CD166, ITGA6, ITGB1, with the notable exception of CD133, and, as expected, weaker or absent expression of tissue-specific stemness markers (Fig. 2F) (23,24). Expression of the same markers was found in the original CUP tissues ( Supplementary Fig. S2E). Interestingly, the MYC target gene and 'don't eat me' immunosuppressive signal CD47 was expressed in all agnospheres (25,26). Nonetheless, the mutually exclusive expression of EpCAM and CD44 ( Concerning functional properties, agnospheres kept in the absence of exogenous growth factors displayed an in vitro estimated fraction of clonogenic cells ranging between 15-44% in 15 AS901, AS906, AS43 and AS67, which decreased to 4% in AS914 and to a mere 0.09% in N-AS47. The clonogenic frequency inversely correlated with the population doubling time, which was ~2,5 days in the highly self-renewing agnospheres, and increased to ~5 and 10 days in AS914 and N-AS47, respectively. In the two latter, constant occurrence of cell death in 5-10% of the overall agnosphere population was observed by trypan blue exclusion (not shown). 20 Overall these data indicate that agnospheres are enriched in cells with transcriptional traits typical of embryonic stem cells, mirrored by functional clonogenic properties.
Interestingly, such transcriptional features are present also in CUP tissues, attesting to the faithful phenotypic correspondence between the original tumors and the agnospheres, and indicating that CUP tissues are highly enriched in stem-like cells.

Agnospheres Generate Phenocopies of the Original Tumors and Contain an Exceedingly
High Percentage of Tumor-initiating Cells 5 Next, to assess the agnosphere tumorigenic potential, we subcutaneously transplanted  Table S1). Importantly, agnospheres regenerated IS-tumors 10 indistinguishable from the original human tumors by histology and expression of immunohistochemical markers ( Fig. 3A and Supplementary Fig. S3A). Interestingly, agnospheres behaved in vitro as tumoroids (27), by retaining the same tissue marker expression In order to evaluate the tumor-initiating potential, we challenged 4 representative 20 agnospheres by stringent in vivo limiting-dilution serial transplantation experiments up to three passages. Remarkably, in AS901, AS906 and AS43 as few as 10 agnosphere cells transplanted subcutis could generate tumors in the majority of mice, leading to an estimated tumor-initiating cell (TIC) frequency of 50% in AS901 and AS906, and 7% in AS43 (Fig. 3D). Such frequencies are exceptionally high after subcutaneous transplantation in NOD/SCID mice, being 1-2 log higher compared with other aggressive tumors such as colorectal or breast cancer (28,29) or glioblastoma (30), assessed in the same conditions. Not surprisingly, N-AS47 displayed a relatively low TIC frequency (0,002%), consistently with its overall in vitro biological properties 5 and modest clinical aggressiveness.

Metastatic Pattern of CUP Patients
We then assessed metastatic ability, by transplanting subcutaneously luciferase-labelled 10 agnosphere cells from the same 4 representative cases (AS901, AS906, AS43, N-AS47). After removing IS-tumors, to prevent both overgrowth and interference with weaker luminescence from metastatic sites, we longitudinally monitored spheropatients by in vivo imaging, detecting increasing signals located in multiple sites, further visualized in explanted organs ( Fig. 4A-C).
The ability of metastases to phenocopy both the original and IS-tumors was confirmed by 15 histopathological and immunohistochemical analysis ( Fig. 4D and Supplementary Fig. S4A-C).
These findings indicate that agnospheres retain specific pseudodifferentiative abilities rather independent from signals by the tissue context, and can adapt to grow in different environments.
In spheropatients transplanted with AS901, AS906 or AS43, the clinical end-point was reached within 3-4 months and the overall post-mortem analysis showed that multiple metastatic sites 20 were present in ~75% of mice (Fig. 4E). In spheropatients transplanted with N-AS47, the disease progressed more slowly, without killing the mice and mirroring the patient's clinical course.
However, at the experimental endpoint (7 months), widespread micrometastases and disseminated single cells were detected in 100% of mice. Interestingly, metastases formed also in mice where the IS-tumor did not grow (54% of mice) (Fig. 4E).
Among the most frequent metastatic sites, we found axillary lymph nodes (Fig. 4F), considered as regional lymph nodes invaded by lymphogenous spread. Lymph nodes are frequently colonized in CUP patients as well (~35%) (4), including those from which 5 agnospheres were derived (Supplementary Table S1). The other organs colonized in spheropatients were likely accessed by hematogenous spreading, with homing occurring not only in the primary capillary district after venous dissemination (lung), but, often, after systemic arterial circulation as well. The frequent colonization of connective tissues, including peripancreatic fibroadipose tissue, gonadal adipose tissue, kidney capsule, and occasional 10 homing between myocardiocytes, suggested a predilection for the connective/mesodermal soil, where micrometastases and intravascular cancerous emboli could be often detected ( Fig. 4D and Supplementary Fig. S4). This, again, is consistent with the metastatic pattern in CUP patients, which can include uncommon sites such as subcutaneous connective tissues and muscles, as observed in patients AGN906, AGN43 and AGN914 (Supplementary Table S1). 15 Strikingly, in CUP spheropatients, metastases efficiently disseminated as early as within 10 days after agnosphere injection, as observed in organs explanted from 4/4 mice transplanted with AS43. Dissemination occurred well before the IS-tumor became palpable, suggesting that cells can initiate the process without prior local expansion and, likely, without induction of a premetastatic niche from an established tumor. Immunostaining of explanted organs showed single 20 disseminated tumor cells, micrometastases and emboli at multiple sites, mostly in the lung and in connective tissues around the organs, i.e. the same sites in which macrometastases develop (Fig.   4G). Interestingly, single disseminated cells were identified by pan-cytokeratin immunostaining, thus retaining their epithelioid phenotype, suggesting that EMT, if ever occurred, rapidly disappeared.
Such a widespread and rapid metastatization is rather uncommon after subcutaneous engraftment of tumor-initiating cells or organoids from other aggressive carcinomas. As an example, mice subcutaneously injected with the metastasis-derived colosphere CRC729 (Fig. 4E   5 and Supplementary Fig. S4D-E) formed IS-tumors in 100% of cases, in some cases locally invasive, but they were always unable to form detectable metastases during 7-months in vivo monitoring and ex-vivo imaging. This is consistent with lack of evidence in similar models reported by the literature. 10

Agnospheres are Sensitive to MEK Inhibition in Vitro
As described above, all agnospheres displayed the distinctive ability to long-term propagate in the absence of any exogenous growth factor. Moreover, agnospheres, including the slowly proliferating N-AS47, were insensitive to an ample panel of exogenously supplied growth factors, with the exception of the rapidly proliferating AS67, which showed a moderate response 15 to some of them (  Table S3 and Dataset 1). However, pan-EGFR family inhibition through lapatinib or afatinib (31) could significantly halt only AS43 proliferation, without inducing apoptosis even after prolonged treatment (Supplementary Fig. S5D-E). This partial or full resistance to EGFR inhibition could be explained by the concomitant presence of other autocrine loops providing bypass mechanisms that kept MAPK and AKT pathways constitutively active ( Supplementary Fig. S5F). 5 As the multiple mechanisms sustaining autonomous proliferation should converge onto the MAPK pathway, we challenged agnospheres with the clinically approved, most potent MEK1/2 inhibitor trametinib (32). Although agnospheres did not harbor a BRAF mutation, a well-established biomarker of trametinib response (33), the drug exerted a dramatic effect at nanomolar concentrations in 4/6 agnospheres, inducing growth arrest and apoptosis, as indicated 10 by the appearance of cleaved PARP protein, within 4 days in AS901 and AS906, and within 10 days in AS43 and AS67 (Fig. 5C-D and Supplementary Fig. S5G). This effect was comparable to that observed in a BRAF-mutated melanosphere (MS321) ( Supplementary Fig. S5G).
Importantly, trametinib blocked phosphorylation of the MEK1/2 target ERK1/2 and caused disappearance of the c-MYC protein, known to be tightly regulated by proliferative signals via 15 MAPK pathway ( Fig. 5D) (34,35). c-MYC disappearance was accompanied by down-regulation of its transcriptional target cyclin D1, responsible for G1-S cell cycle progression via CDK4/6mediated RB hyperphosphorylation, which was consistently impaired (Fig. 5D). Stabilization of p27, known to be prevented by c-MYC, was observed as well ( Fig. 5D) (35). Given the MYC relevance for the stem-like phenotype, and the high expression of MYC targets in agnospheres 20 ( Fig. 2B), we can envisage that trametinib hits the core of agnosphere properties. Interestingly, in agnospheres, trametinib did not cause a rebound, but rather a decreased, EGFR phosphorylation and expression (Fig. 5D). The latter is a frequently observed mechanism of resistance to MEK inhibitors, based on relief of a MAPK-mediated negative feed-back on receptors (36). This suggests that agnospheres may lack this negative feed-back from MAPK to receptors, resulting in a constitutively active, addictive MAPK signal, which would eventually cause hypersensitivity to MEK1/2 inhibitors. Conversely, trametinib was biologically ineffective in N-AS47, consistently with its inability to inhibit ERK1/2 phosphorylation (Fig. 5C-D), suggesting that N-5 AS47 can self-sustain proliferation through pathways independent of MEK1/2. Surprisingly, in AS914 trametinib exerted the paradoxical effect of increasing cell proliferation, likely due to a rebound effect on upstream proliferative signals, as described in several cell lines and tumors (36). 10

The MEK inhibitor trametinib is effective in the CUP preclinical model and its efficacy is reliably predicted in patients by a gene expression signature
To evaluate the therapeutic efficacy of trametinib in preclinical models, CUP tumors were generated by subcutaneous transplantation of 3 representative agnospheres (AS901, AS906 and AS43). After establishment of IS tumors, mice were treated with a dose of trametinib previously 15 shown to be well tolerated and effective in xenograft models harboring BRAF mutation (Fig. 5E) (36,37). We choose to monitor the therapeutic effect in the established IS (subcutaneous) tumor, assuming that (i) the IS-tumor is not representative of a conventional primary tumor, but rather of every metastasis generated by the agnosphere in the mouse, given the homogeneity of IStumors and metastases; (ii) the effect of trametinib is not expected on the early dissemination 20 step of the metastatic process, but on the growth of established metastases; (iii) the IS-tumor can be more accurately and precociously measured than metastases, which are multiple and delayed, and evolve by an unpredictable pattern that prevents reliable longitudinal quantification. The inhibitory effect of trametinib on tumor growth became evident early after the beginning of the treatment (Fig. 5E). Accordingly, statistically significant prolonged survival was observed in all mice treated with trametinib (Fig. 5F). Interestingly, histopathological analysis of tumors collected at the clinical endpoint showed that treated tumors, as compared with controls, contained an ample central necrosis area surrounded by a crown of actively proliferating cells 5 ( Fig. 5G-H). These observations indicate that tumor volume measurement alone underestimates the trametinib biological effect, while in vitro proliferation arrest and cell death seem reliable predictors of the in vivo tumor tissue response.
To evaluate whether sensitivity to trametinib can be predicted in CUP patients, we envisaged the application of a 'trametinib response signature'. To generate such a signature, we  Table   S2), 3 metastatic BRAF-mutated melanomas (including the tumor that originated MS321, 20 MEL321) and an early-metastatic merkeloma (a tumor of neuroectodermal origin) (Fig. 5J).
The signature predicted trametinib sensitivity in all BRAF-mutated melanomas, as expected, and in 9/12 CUPs, while it predicted resistance in the neuroendocrine AGN47 and in merkeloma (MERK44). Agnospheres and their original tumors were predicted consistently, attesting that the signature retains its ability to reliably classify suboptimal samples such as paraffin-embedded tissues. Interestingly, CUP cases associated with trametinib sensitivity independently of the presence of BRAF or RAS mutations, suggesting that constitutive and addictive activation of the MEK/MYC axis can be sustained by alternative mechanisms 5 impinging on the growth factor signaling pathway.

DISCUSSION
Isolation of stem-like cells from CUP patients provided, to our knowledge, the first CUP experimental model, suitable to identify genetic and molecular determinants that explain the disease hypermetastatic phenotype, featuring early widespread dissemination, ability to quickly colonize multiple organs by adapting to different microenvironments, and cell differentiation 5 block.
Irrespective of their different genetic alterations, agnospheres display common properties, among which the first and most striking is proliferative and self-renewal autonomy: agnospheres retain their ability to long-term propagate without differentiating even in the absence of any exogenous growth factor, a property seldom displayed by stem-like cells derived from other 10 metastatic tumors. We could associate this property with the constitutive expression of selfrenewal and reprogramming transcription factors, and the constitutive enrichment of a transcriptomic signature distinctive of embryonic stem (ES) cells (19). Within this signature, genes controlled by Polycomb repressors (41) and by the MYC family (42) were strongly modulated, consistently with widespread expression of their transcriptional drivers (BMI1 and 15 EZH2, and MYC, respectively). Interestingly, this ES signature was enriched not only in agnospheres but also in original CUP tissues, indicating that the entire tumor is mostly formed by cells with stem features, and that, vice-versa, agnospheres well represent the overall CUP cell population, thus behaving like 'tumoroids' (27).
The c-MYC proto-oncogene is well known for its role at the cross-road between stem cell 20 regulation and oncogenesis. Famous for being part of the original Yamanaka cocktail that reprogram differentiated fibroblasts to pluripotency (43), c-MYC has been recognized as a key factor to recruit quiescent stem cells into proliferation, by providing not only direct stimulation of the cell cycle, but also activation of metabolic genes that support the bioenergetics needs of proliferating cells (44). Overexpression of the c-MYC proto-oncogene contributes to tumor onset and/or progression, as indicated by findings in human tumors (that overexpress MYC in ~30% of cases) and mouse models (34,45). Experimental c-MYC overexpression can confer addiction, so that MYC knock-down causes regression of established tumors, attesting its relevance for cell Interestingly, high levels of N-MYC (together with c-MYC) were found in AS914, consistently with observations in ample range of malignancies (48) A second feature shared by all agnospheres, which can be correlated with independency from exogenous signals, is the ability to widely metastasize after subcutaneous transplantation: 15 indeed, agnospheres could quickly disseminate, home, survive and thrive in multiple tissue contexts, where they consistently reproduced the histology of the original tumors, including expression of markers specific to each different patient, together with lack of terminal differentiation. This adaptability is likely conferred by the ability to sustain proliferative and pseudodifferentiative programs in a niche-independent way. Importantly, our data indicate that 20 this ability is passed-on from patients to the experimental CUP model through agnospheres, likely as result of (epi)genetic mechanisms rendering the proliferative pathway constitutive, either through growth factors autocrine loops or altered signal transduction.
In spite of the above common traits, the agnosphere experimental behavior diverged according to the different clinical courses of the respective original patients: not surprisingly, the frequency of tumor-initiating cells was exceptionally high and metastatic dissemination was rapid in agnospheres isolated from the most aggressive cases, while the two parameters were decidedly reduced in the N-AS47 from the long-survivor neuroendocrine CUP. 5 In looking for CUP liabilities, we reasoned that the variegated panel of proliferative signals detected in agnospheres should converge on constitutive activation of the MAPK pathway, known to be a major inducer of MYC expression and activity (34,35). Interestingly, hyperactivation of the MAPK pathway has been recognized as a distinctive feature of metastatic tumors in general (49) and CUP in particular (50), together with MYC overexpression (51). We 10 could demonstrate that 4/6 agnospheres were addicted to the MAPK pathway and highly sensitive to the specific MEK1/2 inhibitor trametinib (36). Sensitivity correlated with complete downregulation of c-MYC expression, proliferation arrest and cell death in vitro, and induction of massive necrosis in experimental tumors. Interestingly, in many cancers displaying MAPK hyperactivation, including those harboring RAS mutations, the administration of MEK1/2 15 inhibitors is poorly effective, as it interrupts negative feedbacks reverberating from the MAPK pathway to tyrosine kinase receptors. These feedbacks may include signals rapidly leading to receptor downregulation, as well as long-term transcriptionally-driven adaptive kinome reprogramming (36). Evidence indicates that these mechanisms are likely and consistently disrupted in responsive agnospheres, as trametinib did not induce a rebound effect of tyrosine 20 kinase receptor reactivation, but it rather induced receptor downregulation. Conversely, in nonresponsive agnospheres, trametinib not only failed to block cell proliferation, but it stimulated MAP kinase activation, suggesting that the above mechanisms of feed-back were intact, or other MEK family kinases might be involved, whose identification is hindered by lack of specific inhibitors. Interestingly, in AS914, trametinib even stimulated proliferation. This is not surprising as, for example, in normal ES cells, MEK inhibitors are used to promote rather than to halt proliferation (18).
To assess whether sensitivity to trametinib, a clinically approved drug, could be applied 5 throughout CUP cases, and provide a tool to stratify patients for trials, we elaborated and validated an original 'trametinib response signature'. This signature correctly anticipated the experimentally-assessed response to trametinib in agnospheres, and was retrieved also in the matched patients' tissues. Thus validated, the signature predicted the response in a retrospective cohort of CUP cases. Despite the absence of BRAF or RAS family mutations, usually associated 10 with trametinib sensitivity, the majority of CUPs were classified as responders. This indicates that constitutive activation of the MAPK pathway, leading to MYC sustained expression and the ensuing stem and proliferative transcriptional programs, can be a CUP widespread pathogenetic mechanism offering opportunities for therapeutic intervention.
Beyond CUP investigation, the integrated experimental platform presented in this study 15 can have far-reaching applications, as it is endowed by unique prerogatives, compared to the metastatic models available so far (52). First, while the latter are mostly based on genetically engineered mice, or the use of conventional cells lines artificially manipulated to become metastatic, human agnospheres are innately endowed with comprehensive metastatic programs that are faithfully inherited from the original tissue and passed on to the mouse model. Moreover, 20 with their properties, agnospheres can help to overcome operational limitations recognized to current metastatic models such as being (i) slow and inefficient, because generated late in progression by rare cell subpopulations; (ii) based on cooperation between cell-intrinsic and environmental signals, that hamper successful colonization.
The model here described thus represent a next-generation tool for functional validation of metastatic determinants, and mechanistically-supported therapeutic interventions in a broad spectrum of aggressive tumors.

Human Subjects
Metastatic patients with diagnosis of CUP, melanoma, colorectal cancer and merkeloma were recruited at the Candiolo Cancer Institute, according to the ethical requirements of the 5 institutional Review Board on human experimentation. Informed consent was obtained from all patients. All patient samples were de-identified before processing. CUP patients were enrolled in an approved prospective observational trial where they received the best standard of care and provided archival and viable biological specimens, and blood samples. The diagnosis of CUP was attained through application of an 'ad excludendum' diagnostic protocol in accordance with 10 ESMO guidelines (2).

Animal Models
All animal procedures were performed according to the institutional guidelines and approved by the Italian Ministry of Health. NOD.CB17-Prkdcscid/NcrCr mice (Charles River Laboratories), 15 5-to 6-week-old male were used for all in vivo studies. Mice were housed at a maximum of 6 per cage with a 14-hour light/10-hour dark cycle with food and water ad libitum. Mice were monitored at a minimum of twice weekly for general performance status and euthanized when volume of xenografts reached 1600 mm 3 , or displayed sign of distress, or weight loss ≥ 20%.

Generation of Patient-derived Xenografts
Human CUP tumor specimens derived either from biopsy or surgery, were subcutaneously transplanted in mice (patient-derived xenograft) and tumors were explanted when reached a maximum of 1600 mm 3 . The tumors were collected for agnosphere derivation and further 15 propagation.

Histopathology
Immunohistochemical staining of formalin-fixed paraffin-embedded tumor tissue sections derived from patients and animal models was performed using Ventana Benchmark ultra System total area for each sample. Necrotic areas were calculated as total area minus vital area.

Genomic DNA was isolated from agnospheres using Relia Prep TM gDNA Tissue Miniprep
System (Promega) according to manufacturer's instructions. DNA was quantified using a Nanodrop ND1000 spectrophotometer (Thermo Fisher Scientific). As normal control, gDNA of 10 PBMCs derived from a pool of 5 healthy individuals was used. PTEN, TP53, MYC and N-MYC copy-number variations were calculated with the 2-ΔΔ Ct methods using GREB1 as normalizer.

Mutational Screening of FFPE Tissue Specimens
Assessment of mutational status for K-RAS, HRAS, N-RAS and BRAF hotspot mutations was 15 performed through OncoCarta™ Panel v1.0 (Agena Bioscience) according manufacturer's instructions.

Chromosomes G Banding
Chromosome analysis by G banding was performed on agnospheres. In order to increase the

RNA-seq and qRT-PCR
Growing agnospheres and melanosphere were harvested and RNA was extracted using RNeasy Micro Kit (Qiagen) following manufacturer' instructions. For qRT-PCR mRNA was converted into first-strand cDNA using superscript II Reverse Transcriptase (Invitrogen) according to manufacturer's instructions. Amplification was performed with ABI PRISM 7900 HT (Applied bases (regions with average quality below 6). Alignment was performed with STAR 2.6.0a (55) on hg38 reference assembly obtained from cellRanger website (Single Cell Gene Expression, 10x Genomics) (Ensembl Assembly 93). The expression levels of genes were determined with htseq-count 0.9.1 by using cellRanger pre-build genes annotations (Single Cell Gene Expression, 10x Genomics) (Ensembl Assembly 93). Genes with an average number of cpm (counts per million) <5 and Perc of duplicated reads >20% were filtered out. Data normalization was performed using edgeR (56). 5

Prediction of Embryonic Stem Cell (ES) Signature in Agnospheres and Patients' Tissues
Transcriptional profiles of each agnosphere or the original patient were first averaged across replicates and then genes with an average CPM less than one across all samples removed. Then averaged transcriptional profiles were converted in z-score and the GSEA (57) was performed using as input the Embryonic Stem cell (ES) signature as published by Ben-Porath et al. (19). 10 GSEA and associated statistics were computed using the fgsea package in R statistical environment version 3.6.

Generation and Validation of Trametinib Response Signature
The basal expression profile of about 1,000 cancer cell lines was obtained using RNA-seq from 15 Cancer Cell Line Encyclopedia (21) (39). Cell lines derived from liquid tumors were discarded, and only cell lines derived from solid tumors were used for the following analysis. The raw counts of each gene across the cancer cell lines were normalized using edgeR (58). Cell lines response to trametinib data used in this study were previously generated by Rees et al. (38) and was expressed in terms of Area Under the Curve (AUC), which reflects the in vitro response to 20 trametinib of each cancer cell line (21) over 72h. In particular, lower values of AUC are associated with a higher sensitivity to trametinib and vice versa. In total 445 CCLs for which trametinib response and expression data were both available were used to identify marker genes associated to trametinib resistance as described below. The gene signature to predict trametinib response from transcriptional data was identified as depicted in Supplementary Fig. S6A-B. First, we computed Pearson Correlation Coefficient (PCC) between the expression of each gene and trametinib potency measured by mean of AUC across the 445 distinct cancer cell lines (CCLs) selected as described above. We then considered the top 1,000 genes with highest PCC 5 as putative marker genes of trametinib resistance, because higher expression of these genes is associated with lower potency of trametinib across the panel of the 445 used CCLs. Finally, a machine-learning approach based on recursive feature elimination (RFE) and support vector machines (SVMs) was used to identify the 500 (out of 1,000) genes whose expression values were best at discriminating resistant from sensitive trametinib CCLs (59,60). In particular, we 10 used linear SVMs that were trained and tested using the kernlab package in the R statistical environment version 3.6 (61). For RFE, trametinib sensitive CCLs were defined as the ones which AUC values were in the 5% quartile of trametinib AUC distribution across the 445 CCLs used in this study. On the other hand, trametinib resistant CCLs were defined as those whose AUC values were in the 95% quartile. We used Gene Set Enrichment Analysis (GSEA) to 15 predict the trametinib response checking whether marker genes of trametinib resistance identified as described above were either down-or up-regulated across a given transcriptional profile ( Supplementary Fig. S6B). Hence, given a transcriptional profile, genes are first sorted form the most to the least expressed and then Gene Set Enrichment Analysis (GSEA) is performed using as input the identified marker genes of trametinib resistance. A positive 20 enrichment score means that genes associated with trametinib resistance are highly expressed in the transcriptional profile, thus predicting resistance of these cells to trametinib treatment.
Conversely, a negative enrichment score indicates a low expression of trametinib marker genes of resistance and thus predicts sensitivity of these cells to trametinib treatment ( Supplementary   Fig. S6B). General accuracy of the method in predicting trametinib response was then evaluated using an independent dataset of 634 CCLs from Garnett et al. (40), for which basal expression from microarray and trametinib response in term of IC50 were available. We used the method described above to predict the trametinib response for each of the 634 CCLs and we then 5 computed percentage of correctly predicted sensitive and resistant CCLs (Supplementary Fig.   S6C). Specifically, CCLs sensitive to trametinib were defined as those with an IC50 <1 M, while all the other were considered as resistant to trametinib treatment. We obtained an average classification accuracy of about 76% across the two classes (sensitive and resistant) of CCLs. In all the analyses described above GSEA and associated statistics were computed using the fgsea 10 package in R statistical environment version 3.6.

Patients' Tissues
Tumorsphere or patient transcriptional profiles were first averaged across replicates when 15 present, then genes with an average CPM less than one across all samples filtered out. Before to apply GSEA averaged transcriptional profiles of each tumorsphere or patient was converted in zscore. GSEA (57) was performed using as input identified marker genes of trametinib resistance with the fgsea package in R statistical environment version 3.6. 20

Immunofluorescence and Cell Immunohistochemistry
Samples undergoing immunofluorescence were either Formalin-Fixed, Paraffin-Embedded tumor specimens or growing agnospheres. The latter were harvested, fixed 10 min with PFA 4% at 4°C, washed in PBS, suspended in bio-agar for cyto-inclusion (Bio-Optica) at 42 °C, and processed for inclusion in paraffin. All staining were performed as previously described (62).
Images were acquired using a LEICA SPEII confocal microscope, equipped with a 40X oil objective and a 1,5X zoom for a final magnification of 600X. Optical single sections were acquired with a scanning mode format of 1024x1024 pixels. Fluorochromes unmixing was 5 performed by acquisition of automated-sequential collection of multi-channel images, to reduce spectral crosstalk between channels. For immunohistochemical staining an additional peroxidase blocking was performed in H2O2 3%/ methanol 50% incubated 20 min in the dark. For primary and secondary antibody concentrations see Reagents. Secondary antibodies were HRPconjugated (Dako), and diaminobenzidine (DAB) substrate chromogen kit (Dako) was used for 10 detection. Nuclei were counterstained with Haematoxylin and images were acquired through LASV4.2 software. Images are representative of at least three independent immunostainings.

Western Blot Analysis
Total protein were extracted using RIPA buffer supplemented with a protease inhibitor cocktail 15 (Roche Life Science), NaVO3 1mM and NaF 1mM, subjected to sonication, quantified using BCA methods (Pierce), and 20 g were separated on SDS-polyacrylamide gradient gel 4-12 % or 4-20 % (Invitrogen) and blotted onto nitrocellulose membrane. After blocking, primary antibodies were incubated at the indicated concentrations (see Reagents). After incubation with HRP-conjugated secondary antibodies (Jackson Lab), enhanced chemiluminescence (Biorad) 20 was used for detection according to manufacturer's instructions and images were acquired with the ChemiDoc Touch TM Imaging System (Biorad) with Image Lab software. Actin or tubulin were used as protein loading control as indicated. The results shown are representative of at least 3 independent experiments.

In vitro Limiting Dilution Assay (LDA)
Agnospheres were dissociated and seeded at limiting dilution concentration (1-100 cells/100l) 5 in ultra-low-attachment 96-well microtiters (Corning). 'Positive tests' were defined as wells with primary spheres with a diameter ≥100m. Stem cell frequency was calculated using the ELDA software (http://bioinf.wehi.edu.au/software/elda/). Means and 95% confidence intervals (CI) are shown (n ≥3 independent experiments). Primary spheres were harvested, dissociated and seeded at the same dilutions for a second and a third passage to assess long-term propagation. was calculated with the ELDA software as above. Mice dead before the experimental endpoint that did not generate a tumor were excluded from the counts. From P0 tumors agnospheres were re-derived in culture by enzymatic digestion with collagenase as above, and, after 1 week of 20 recovery in culture, they were dissociated and transplanted for a second passage in mice (P1) at the same cell dilutions to verify the presence of long-term propagating TIC cells (63). In some experiments, the procedure was further repeated for a third passage (P2),

Assessment of Spontaneous Metastases
Agnospheres were engineered to express luciferase as follows. Single cell suspensions were seeded in ultra-low-attachment 6-wells microtiter at 500.000 cells/well in 1.5 ml of culture medium. Lentiviral particles carrying pCMV-Luciferase or pCMV-Luciferase-IRES-GFP transfer vectors (64) were added at 5 MOI. After 6 h, 1 ml medium was added, and 24 h later the administered to mice by subcutaneous injection. Anesthesia was delivered in the induction chamber with 2.5% isofluorane in 100% oxygen at a flow rate of 1 L/min and maintained during the imaging in the IVIS with a 1.5% mixture as above at 0.5 L/min. To mask the residual signal from the subcutaneous tumor a black tape was applied during the imaging. Luminescent signals 20 in different organs were already visible after surgery and were monitored weekly until mice reached the clinical endpoint. For ex-vivo imaging, luciferin was administered as above, mice were euthanized by carbon dioxide inhalation and all organs were immediately explanted and analyzed with IVIS. Organs were formalin-fixed paraffin-embedded to undergo histopathological evaluation and comparison with original tumors as above. The frequency of metastatic sites detailed in Fig. 4F was assessed by ex-vivo organ luminescence and by immunohistochemistry of organ sections, to detect human tumor cells by expression of specific markers. Such frequency could be underestimated for missed identification of micrometastases in some organs. To assess 5 early dissemination, in additional experiments with AS43, mice were euthanized 10 days after subcutaneous injection and explanted organs were formalin-fixed paraffin-embedded and analyzed as above.
In vivo trametinib treatment 10 Agnosphere were inoculated subcutis in immunocompromised mice. When tumors reached ~150 mm 3 , mice were randomized in two groups (control and trametinib-treated). For treatment, trametinib (GSK1120212) was dissolved in hydroxi-methylcellulose 0.5%-Tween 80 and administered by oral gavage at 1mg/Kg/die (therapeutic range in preclinical models: 0.3-3 mg/kg/die) (37). Tumor volume was measured and calculated as above and expressed as fold 15 changes vs. day 0. For each experiment, n ≥ 5 mice/group were used. Statistical significance was determined by non-parametric test (Mann-Whitney) at the indicated days. Survival curve was generated considering as experimental endpoint a tumor volume of 1600 mm 3 and statistical significance was calculated using log-rank (Mantel-Cox test). 2H. Chi-squared test for limiting dilution experiments was performed by ELDA software (in fig   2G and 3C). A p-value <0.05 was considered significant, * p<0.05, ** p<0.01, *** p<0,001 **** p< 0,0001, ns: not significant. Mantel-Cox test was used for survival curve (in fig. 5F). For 5 RNAseq analysis the fgsea package in R statistical environment version 3.6 was used. Tables 1-3