Loss of Dnmt3a and Dnmt3b does not affect epidermal homeostasis but promotes squamous transformation through PPAR-γ

The DNA methyltransferase Dnmt3a suppresses tumorigenesis in models of leukemia and lung cancer. Conversely, deregulation of Dnmt3b is thought to generally promote tumorigenesis. However, the role of Dnmt3a and Dnmt3b in many types of cancer remains undefined. Here, we show that Dnmt3a and Dnmt3b are dispensable for homeostasis of the murine epidermis. However, loss of Dnmt3a-but not Dnmt3b-increases the number of carcinogen-induced squamous tumors, without affecting tumor progression. Only upon combined deletion of Dnmt3a and Dnmt3b, squamous carcinomas become more aggressive and metastatic. Mechanistically, Dnmt3a promotes the expression of epidermal differentiation genes by interacting with their enhancers and inhibits the expression of lipid metabolism genes, including PPAR-γ, by directly methylating their promoters. Importantly, inhibition of PPAR-γ partially prevents the increase in tumorigenesis upon deletion of Dnmt3a. Altogether, we demonstrate that Dnmt3a and Dnmt3b protect the epidermis from tumorigenesis and that squamous carcinomas are sensitive to inhibition of PPAR-γ. DOI: http://dx.doi.org/10.7554/eLife.21697.001


Introduction
DNA methylation is an epigenetic mechanism that regulates several aspects of gene expression, such as long-term gene silencing, transcriptional elongation, and maintenance of genomic stability (Allis and Jenuwein, 2016;Avgustinova and Benitah, 2016;Rinaldi and Benitah, 2015). It is found throughout the vertebrate genome and is deposited by DNA methyltransferases on the fifth position of cytosine (5-mC), predominantly at CpG dinucleotides. The role of DNA methylation in establishing different cell fates during embryogenesis is fairly well understood. However, if and how DNA methylation is necessary to stably maintain the identity of adult stem cells, and how this process is disrupted during oncogenic transformation, is under intense investigation (Shen and Laird, 2013).
Three DNA methyltransferases are encoded in the vertebrate genome. Dnmt1 is predominantly associated with the maintenance of DNA methylation following cell division due to its high affinity for hemimethylated DNA. Consequently, depletion of Dnmt1 leads to a significant reduction of the global levels of 5-mC (Li et al., 1992;Lei et al., 1996). Dnmt3a and Dnmt3b are de novo DNA methyltransferases that establish genome-wide DNA methylation during mammalian embryogenesis and adult stem cell homeostasis (Okano et al., 1999). In mouse embryonic stem cells, the combined loss of Dnmt3a and Dnmt3b leads to the progressive loss of DNA methylation, suggesting that these enzymes are additionally involved in maintaining 5-mC levels (Chen et al., 2003).
Since Dnmt3a-null mice die perinatally, and ablation of Dnmt1 and Dnmt3b results in embryonic lethality, conditional deletion mouse models have been necessary to study their functions in adulthood (Ueda et al., 2006;Okano et al., 1999). Hematopoietic stem cells (HSCs) lacking Dnmt3a cannot differentiate correctly upon serial transplantation, and end up developing a range of severe myeloid and lymphoid malignancies in aged animals (Challen et al., 2011;Mayle et al., 2015). Conversely, HSCs-depleted of Dnmt3b show no phenotypical differences with respect to wild-type controls, whereas combined ablation of Dnmt3a and Dnmt3b in HSCs result in an enhanced block of hematopoietic differentiation as compared to Dnmt3a loss alone (Challen et al., 2014). Interestingly, the observed phenotype seems specific to stem cells, as the fully differentiated cardiac myocytes carrying a combined deletion of Dnmt3a and Dnmt3b are indistinguishable from wild-type controls (Nührenberg et al., 2015). Similarly, to HSCs, purified murine neural stem cells (SCs) lacking Dnmt3a do not show problems with self-renewal but fail to differentiate properly (Wu et al., 2010). In addition, Dnmt3a acts as a potent tumor suppressor in lung tumorigenesis, to promote adenoma progression but not initiation, downstream of oncogenic K-Ras (Gao et al., 2011). This is in contrast with the pro-tumorigenic activity of Dnmt3b, which at least in the murine intestinal epithelium cooperates with the loss of APC to drive adenoma initiation and growth Lin et al., 2006) Recently, progress has been made in identifying the molecular mechanisms underlying the biological functions of Dnmt3a and Dnmt3b by studying their genome-wide localization. For instance, Dnmt3b associates with and methylates the gene bodies of actively transcribed genes in murine embryonic SCs and human embryonic carcinoma cells (Baubec et al., 2015;Jin et al., 2012;Morselli et al., 2015). Likewise, it has been proposed that gene body methylation is responsible of most of the transcriptional changes underlying the ability of Dnmt3a to promote neural SCs differentiation, and in protecting the lung epithelium from tumor progression (Wu et al., 2010;Gao et al., 2011). eLife digest Most of the cells in our body contain the same DNA. However, our bodies are made of many different types of cell, such as nerve cells or skin cells, which perform very different jobs. In each cell type only certain sets of genes encoded by the DNA are active. Proteins known as epigenetic regulators are responsible for producing the different patterns of gene activity. If epigenetic regulators are switched on or off at the wrong time, they can contribute to ageing and diseases such as cancer.
Enzymes known as DNA methyltransferases are one group of epigenetic regulators. DNA methyltransferases control the activity of genes by adding small chemical groups known as methyl groups to the DNA. Two of these enzymes -known as Dnmt3a and Dnmt3b -are important during development to help cells mature and specialize into different types. Mice that lack both of these enzymes either die as embryos or just after birth. Furthermore, these enzymes are mutated or less active in some skin cancers and various other human cancers.
Here, Rinaldi et al. investigated the role these enzymes play in adult mice. The experiments show that under ordinary laboratory conditions, mutant mice that lacked Dnmt3a and Dnmt3b were as healthy as normal mice. However, when the mice were exposed to chemicals that promote tumor growth, which mimics skin exposure to UV light, the mutant mice developed many more skin tumors than the normal mice. Furthermore, the tumors in the mutant mice were more likely to form secondary tumors in the lung. Rinaldi et al. found that Dnmt3a reduced the production of a protein called PPAR-g, which helps to break down some types of fat molecules. Treating the mutant mice with a drug that inhibits PPAR-g activity slowed the growth of the tumors.
Overall, these experiments show a new way in which DNA methyltransferases act in adult animals. Future research will investigate whether drugs that inhibit the breakdown of fats could help to treat cancers in which the Dnmt3a and Dnmt3b proteins are mutated or less active.
We have recently reported that Dnmt3a and Dnmt3b are required for the self-renewal of human keratinocyte progenitors, whereas Dnmt3a is also required for their proper differentiation (Rinaldi et al., 2016). Mechanistically, Dnmt3a and Dnmt3b bind to and promote the activity of enhancers in both human epidermal progenitors and differentiated keratinocytes (although Dnmt3a having a stronger affinity for enhancers in differentiated keratinocytes). Interestingly, both proteins preferentially associate to super-enhancers rather than typical enhancers. Nonetheless, they differ in their mechanism of action, since Dnmt3a (together with Tet2) is essential to maintain high levels of 5-hydroxymethylcytosine (5-hmC) at the center of its target enhancers, while Dnmt3b promotes 5-mC along the body of the enhancer. These regulatory regions dictate the transcription of essential genes necessary for epidermal stem cell identity and maintenance, such as FOS, ITGA6, TP63, and KRT5. Similar to its role in mouse ES cells, Dnmt3b also binds to and methylates the gene bodies of these genes to reinforce their expression (Rinaldi et al., 2016). Dnmt3a also associates to the enhancers regulating the expression of genes such as IVL, LOR, FLG2, and KRT1 which drive the differentiation of SCs into mature keratinocytes (Rinaldi et al., 2016). Interestingly, DNA methylation at active enhancers has also been recently reported in normal and cancer-derived human colon, mammary and prostate epithelial cells (Charlet et al., 2016).
However, to date, no in vivo studies have investigated the roles of Dnmt3a and Dnmt3b in adult epidermal function and malignant transformation. Using mouse models carrying an epidermis-specific ablation of either Dnmt3a or Dnmt3b, or both, we demonstrate that Dnmt3a and Dnmt3b are dispensable for skin homeostasis. However, Dnmt3a has a critical role in suppressing carcinogeninduced squamous tumor initiation, but not progression, while both Dnmt3a and Dnmt3b concertedly prevent tumor progression.

Results
Epidermal deletion of Dnmt3a does not affect steady-state tissue homeostasis but significantly promotes carcinogen-induced tumorigenesis We first studied the pattern of expression of Dnmt3a and Dnmt3b during epidermal development, and in adulthood. At E14.5, Dnmt3a was expressed in the entire Keratin-14+ compartment comprising the basal layer of the embryonic epidermis and the hair placodes (  (Solanas and Benitah, 2013), and decreased in the interfollicular epidermis, although some basal IFE cells expressed high levels (Figure 1-figure supplement 1A-D). On the other hand, we were not capable of detecting Dnmt3b by immunofluorescence staining in sections of developing or adult mouse epidermis (not shown), suggesting that Dnmt3b is expressed at low levels (Challen et al., 2014). In fact, RNA-seq data confirmed that Dnmt3a was enriched almost fivefold as compared to Dnmt3b in adult basal IFE keratinocytes (Figure 1-figure supplement 1C). However, Dnmt1, the main DNA methyltransferase, was the most abundant DNA methyltransferase, both in interfollicular epidermis and in hair follicle bulge cells (Figure 1-figure supplement 1C).
To gain insight to the roles of Dnmt3a and Dnmt3b in epidermal tissue function, we generated epidermis-specific conditional knockout (cKO) mice by crossing animals containing the Dnmt3a or Dnmt3b gene flanked by loxP sites with animals carrying the Keratin14-CRE-ROSA26-YFP-cassette (hereafter referred to as Dnmt3a/3b-cKO) (Gao et al., 2011). Surprisingly, neither Dnmt3a-nor Dnmt3b-cKO displayed noteworthy epidermal phenotypical differences as compared to their wildtype littermates at different postnatal ages ( Deregulation of DNA methylation can alter gene expression, leading to tumor suppressor silencing or oncogene activation (Witte et al., 2014), and mutation/deregulation of Dnmt3a and Dnmt3b has been observed in several tumor types (Leppert and Matarazzo, 2014;Subramaniam et al., 2014). Recently, Dnmt3a has attracted much attention, as it is one of the most frequently mutated genes in cancer (Kim et al., 2013), especially in acute myeloid leukemia (Garg et al., 2015;Ley et al., 2010). In fact, a loss-of-function mutation of Dnmt3a is one of the earliest mutations that occurs in human acute myeloid leukemia (Shlush et al., 2014). Importantly, these mutations are functional since knock-in mice that model it develop a range of severe myeloid and lymphoid malignancies (Challen et al., 2011;Mayle et al., 2015). In addition, HSCs harboring inactivating mutations of Dnmt3a are clonally selected in ageing humans (Shlush et al., 2014). However, much less is known about how deregulation of Dnmt3a and Dnmt3b affect tumorigenesis in epithelial tissues.
To elucidate the roles of Dnmt3a and Dnm3b in skin tumorigenesis, we first generated tumors from the epidermis using the chemically induced carcinogenesis protocol based on DMBA/TPA (Ewing et al., 1988). The first epidermal squamous malignancies appeared significantly sooner in Dnmt3a-cKO than in their wild-type littermates, after only 2 months from the first DMBA treatment, indicating that Dnmt3a acts as a barrier against tumor initiation ( Figure 1A,B). Dnmt3a-cKO animals also showed a significant increase in tumor burden, with an average of 17 tumors per animal compared to three tumors per wild-type animal after 6 months of initiating the experiment ( Figure 1C and Figure 1-source data 1).
Although Dnmt3a-cKO animals showed a strong increase in tumor initiation and burden, they developed the same percentage of squamous cell carcinomas than wild-type mice ( Figure 1D,E). Indeed, a detailed histological analysis of the tumors collected from Dnmt3a-cKO and wild-type animals indicated that Dnmt3a-cKO mice developed the same percentage of benign tumors, such as keratoacanthomas and papillomas, as well as of malignant invasive papillomas and squamous cell carcinomas (SCCs) ( Figure 1E). Dnmt3a-cKO mice only developed an increase in the percentage of sebaceous adenomas ( Figure 1E). No metastases were scored in any of the animals, as expected using this protocol in mice with a C57/Bl6 genetic background (Sundberg et al., 1997). Altogether, these results indicate that loss of Dnmt3a dramatically increases initiation of epidermal squamous tumors without affecting their malignant progression, and slightly skews the histology of tumors towards the sebaceous lineage.
Deletion of Dnmt3b does not increase tumor initiation, but synergizes with loss of Dnmt3a to promote tumor progression and metastasis Dnmt3a suppresses K-Ras-driven lung tumor progression, whereas Dnmt3b is pro-tumorigenic in APC-deficient colorectal adenomas (Gao et al., 2011;Lin et al., 2006). Hence, we next tested whether Dnmt3a and Dnmt3b also exert opposing effects regarding tumorigenesis in the epidermis. Interestingly, there were no differences between wild-type and Dnmt3b-cKO mice with respect to either the timing of tumor initiation or tumor burden upon treatment with DMBA/TPA ( To assess whether Dnmt3a and Dnmt3b potentially play redundant roles during tumorigenesis, we also induced tumors using the DMBA/TPA protocol in animals carrying an epidermis-specific deletion for both Dnmt3a and Dnmt3b in combination (DcKO). Strikingly, although DcKO animals had a severe depletion of DNA methylation in their epidermises, they formed a morphologically normal skin with all its appendages and did not develop any epidermal abnormality even up to 70 weeks of age (Figure 2-figure supplement 4A-B). This strongly suggests that de novo DNA methylation is dispensable for the long-term homeostasis of undamaged epidermis. However, when subjected to tumorigenesis, DcKO animals displayed a significantly shortened latency and significant higher tumor burden than wild-type mice ( Figure 2B-F). Although these differences were similar to the ones observed in single Dnmt3a-cKO mice ( Figure 2C-D), DcKO mice formed aggressive squamous cell carcinomas at a higher frequency as compared to the single cKOs of Dnmt3a or Dnmt3b ( Figure 2E). In addition, metastatic nodules in the lungs were observed in 30% of DcKO animals, but in none of the wild-type, Dnmt3a-cKO, or Dnmt3b-cKO animals ( Figure 2F). Recent reports show that epidermal squamous cell carcinomas that harbor cells undergoing epithelial to mesenchymal transitions are more metastatic than those that remain predominantly epithelial in nature (Latil et al., 2017;da Silva-Diz et al. 2016). Interestingly, DcKO tumors contained large areas with spindle-shaped cells that expressed lower levels of the epithelial markers E-Cadherin and Keratin14, compared to the wild type and to Dnmt3a-cKO tumors ( Figure 2E, and   Taken together, these results indicate that Dnmt3a and Dnmt3b are dispensable for epidermal homeostasis, and that Dnmt3a, but not Dnmt3b, suppresses skin squamous tumor initiation. However, both Dnmt3a and Dnmt3b repress the malignant transformation of epidermal cells into aggressive squamous cell carcinomas.

Deletion of Dnmt3a results in increased tumor transcriptome heterogeneity and upregulation of genes related to lipid metabolism
We next wanted to characterize the molecular mechanisms that might underlie the tumor-suppressive function of Dnmt3a in the epidermis. To this end, we isolated by FACS-based cell sorting the basal integrin a6 bright tumor cells from four wild-type and eight Dnmt3a-cKO tumors, and performed whole-genome transcriptome profiling by RNA-seq ( Figure 3A). It is important to note that our mouse pathologists scored these tumors as squamous cell carcinomas (SCCs), and that ADFP (Perilipin-2) expressing sebaceous adenomas were not included in this transcriptome study ( Figure 3 The PCA analysis of the RNA-seq samples showed that the four wild-type tumors clustered together, indicating that overall their transcriptomes were defined by common genes ( Figure 3B,C). In contrast, the transcriptomes of Dnmt3a-cKO tumors were substantially more heterogeneous, suggesting that the loss of Dnmt3a could result in the deregulation of numerous different pathways in cancer cells, or that in the context of Dnmt3a loss, different cell of origins (i.e. basal IFE cells, hair follicle stem cells, or Lrig+ stem cells) might be more prone to generate more transcriptionally divergent tumors. Nevertheless, 391 genes were consistently differentially expressed between wild-type and Dnmt3a-cKO tumors, of which 114 were downregulated and 277 were upregulated in the latter (Supplementary file 1). The downregulated genes were mainly associated with apoptosis, suggesting that loss of Dnmt3a promotes cell survival and protects against programmed cell death ( Figure 3D); consequently, TUNEL and active Caspase3 staining confirmed that Dnmt3a-cKO tumors had fewer apoptotic cells as compared to wild-type tumors ( Altogether, this suggests that the loss of Dnmt3a endows pre-cancerous mutant basal cells with a survival and proliferative advantage, which could account for the increased number of tumors these mice develop. However, once tumors are formed, they progress with the same kinetics as wild-type tumors ( Gene ontology (GO) analysis of the 277 genes that were upregulated in Dnmt3a-cKO basal tumor cells highlighted two principal pathways that were over-represented in all eight Dnmt3a-cKO tumors: Wnt signaling (including ligands and receptors), and more predominantly, lipid metabolism ( Figure 3D). Interestingly, recent reports have associated an increase in lipid metabolism with increased tumorigenesis of chronic myeloid leukemia, as well as colorectal, liver, oral, and breast cancer (Beyaz et al., 2016;Ma et al., 2016;Camarda et al., 2016;Ye et al., 2016;Corbet et al., 2016;Schug et al., 2015;Pascual et al., 2017;Wahl et al., 2017;Bensaad et al., 2014;Luo and Puigserver, 2016). A number of genes associated with fatty acid and lipid metabolism were   . Among these, the most upregulated ones encoded the key pro-adipogenic transcription factors PPAR-a and PPAR-g, which promote adipocyte differentiation and the expression of genes involved in fatty acid metabolism, and which are not expressed in homeostatic epidermal cells (Fajas et al., 2001). The role of these transcription factors in cancer is still poorly understood, although they tend to be upregulated in many types of human tumors (Fajas et al., 2001). Importantly, PPAR-g was upregulated at the RNA and protein levels in all the Dnmt3a-cKO sequenced tumors ( Figure 3E and Figure 3-figure supplement 4A-B). Interestingly, the expression of PPAR-g has been extensively reported to be under epigenetic control by repressive mechanisms such as H3K9 methylation and DNA methylation Zhao et al., 2013).
To further dissect the early molecular changes that might result in the tumor-suppressing role of Dnmt3a in the epidermis, we did a short (6-week long) DMBA/TPA carcinogenesis treatment ( Figure 4A). We then FACS-isolated Itga6 bright CD34 neg cells, consisting mostly of epidermal basal cells (IFE), and hair follicle stem cells (Bulge; Itga6 bright CD34 pos ) from pre-cancerous back skin of wild-type or Dnmt3a-cKO animals for RNA-seq analysis (Supplementary file 2). After this short DMBA/TPA treatment, most of the upregulated genes in epidermal cells (IFE) were already predominantly linked to lipid metabolism and cell proliferation, whereas they related mostly to cell proliferation, and Wnt signaling in bulge stem cells (Figure 4-figure supplement 1A-B). Interestingly, we did not observe a diminished expression of genes regulating apoptosis, as we did in tumor cells. Hence, these results suggest that most of the transcriptome changes observed in tumors upon deletion of Dnmt3a occur early, and that the transition from the pre-cancerous epithelium to tumor growth occurs subsequently by bypassing apoptosis.
The active enhancers bound by Dnmt3a contained higher levels of DNA methylation and hydroxymethylation than those not bound by it ( Figure 5A,C). Importantly, loss of Dnmt3a significantly reduced their DNA methylation and hydroxymethylation ( Figure 5A,C). Intriguingly, a significant reduction in DNA methylation also occurred in enhancers not bound by Dnmt3a, albeit to a statistically significantly lesser extent than those directly targeted by Dnmt3a in wild-type cells ( Figure 5A, B). Upon deletion of Dnmt3a, DNA hydroxymethylation was also significantly reduced in its target enhancers, and to a lesser extent in non-Dnmt3a-bound enhancers ( Figure 5C). However, the ratio of 5-hmC levels at enhancers bound by Dnmt3a between wild-type and Dnmt3a-cKO epidermal cells is significantly higher as compared to the ratio of 5-hmC levels between the enhancers that are not normally bound by Dnmt3a ( Figure 5D). This indicates that the presence of Dnmt3a correlates with significantly higher 5-hmC levels, likely because Dnmt3a provides 5-mC as a substrate for generating 5-hmC, as we have previously shown in human keratinocytes (Rinaldi et al., 2016).

Dnmt3a binds to promoters of genes involved in cell proliferation and lipid metabolism to drive their DNA methylation
In addition to active enhancers, a significant proportion (19%) of the enriched regions for Dnmt3a corresponded to promoters/TSSs ( Figure 4B and Supplement file 3). To understand if Dnmt3a was methylating these promoters, we overlaid the Dnmt3a ChIP-seq with the MeDIP-seq data. Notably, the promoters bound by Dnmt3a showed a strong and statistically significant loss of DNA methylation around the corresponding TSS ( Figure 6A). The levels of DNA methylation were not significantly changed at promoters not bound by Dnmt3a ( Figure 6B). Of note, Dnmt3a-target TSSs were not   enriched for 5-hMC (not shown). The loss of DNA methylation at the promoters/TSSs bound by Dnmt3a was also accompanied by a general increase in the transcription of these genes, measured by RNA-seq in the tumors ( Figure 6C). Altogether, these data suggest that Dnmt3a directly represses the expression of a specific subset of genes by methylating their promoters/TSSs. Interestingly, a GO analysis of the promoters bound by Dnmt3a indicated that they regulated the expression of genes predominantly involved in cell proliferation and lipid metabolism, consistent with our RNA-seq results ( Figure 6D and Supplement file 3). Interestingly, Dnmt3a bound the promoters of Ppar-a and Ppar-g in wild-type but not Dnmt3a-cKO epidermis ( Figure 6E and Supplement file 3). Furthermore, 5-mC levels were lower in the TSS of the PPAR-g gene in Dnmt3a-cKO as compared to wild-type tumors, indicating a DNA methylation-dependent mechanism of transcriptional repression ( Figure 6F).
Consistent with a transcriptional derepression of the locus following loss of DNA methylation, PPAR-g mRNA and protein levels were upregulated both in pre-cancerous interfollicular epidermis and in tumors lacking Dnmt3a, suggesting that the upregulation of PPAR-g is acquired at the precancerous stage, even before overt tumors appear ( Figure 6G,H and Figure 6-source data 1).

Inhibition of PPAR-g attenuates the protumorigenic effect elicited by the depletion of Dnmt3a
We next tested whether the increase in the expression of genes involved in lipid metabolism was required for the earlier onset of tumorigenesis and increased tumor burden in Dnmt3a-cKO mice. To this end, wild-type and Dnmt3a-cKO mice were subjected to the DMBA/TPA skin carcinogenesis protocol, but were separated into two cohorts, one treated topically with a PPAR-g chemical inhibitor in combination with DMBA/TPA, and the other with the vehicle ( Figure 7A-diagram) (Sahu et al., 2012;Grabacka et al., 2006). Interestingly, inhibition of PPAR-g significantly delayed the onset of tumor appearance in Dnmt3a-cKO mice, and reduced the number of tumors developed by the Dnmt3a-cKO ( Figure 7B-C and Figure 7-source data 1). However, the average size of the tumors was not affected by the inhibition of PPAR-g ( Figure 7D). Thus, inhibition of PPAR-g could be a potential new therapy for cutaneous squamous cell carcinomas harboring low levels of Dnmt3a.
At last, using public available data from four different studies we determined that the expression of Dnmt3a is significantly reduced in squamous cell carcinomas and actinic keratosis, the premalignant stage of squamous tumors, compared to human healthy epidermis (

Discussion
Dnmt3a modifies cytosine at CpG dinucleotides and is responsible for the proper differentiation of murine hematopoietic stem cells and murine neural stem cells (Challen et al., 2011;Mayle et al., 2015;Shlush et al., 2014). Recently, we and others have shown in human epidermal keratinocytes and murine olfactory sensory neurons, respectively, that Dnmt3a regulates gene expression by cooperating with Tet to maintain high levels of 5-hmC at enhancers (Colquitt et al., 2014;Rinaldi et al., 2016). Importantly, Dnmt3a is not only frequently mutated in human tumors (Kim et al., 2013) but is perhaps one of the first mutations to occur during tumorigenesis (Shlush et al., 2014). Using knockout mouse models, we now have demonstrated that Dnmt3a is also tumor-suppressive toward carcinogen-induced epidermal squamous neoplasia. Its loss not only accelerated the onset of tumors, but also increased tumor burden. However, once formed, Dnmt3a-deficient tumors grew, and progressed to carcinomas with the same kinetics and proportions, respectively, as their wild- 5-mC levels at enhancer "body" bound by Dnmt3a (n=363) 5-mC levels at enhancer "body" not bound by Dnmt3a (n=2734) Bound by Dnmt3a Figure 5. Depletion of Dnmt3a leads to loss of DNA methylation and hydroxymethylation around its target enhancers. (A) Relative methylation score (CpG count) measured around 363 enhancers bound by Dnmt3a (-5 kb, +5 kb) from independent biological replicates of FACS sorted tumor cells from wild type (n = 2) and Dnmt3a-cKO (n = 2) (p<2.2 Â 10 À16 ). (B) Relative methylation score (CpG count) measured around 2734 enhancers not bound by Dnmt3a (-5 kb, +5 kb) from independent biological replicates of FACS-sorted tumor cells from wild-type (n = 2) and Dnmt3a-cKO (n = 2) animals Figure 5 continued on next page type counterparts. Recent works have shown that the absence of Dnmt3a or Tet2 in hematopoietic stem cells predisposes to leukemia formation (Yang et al., 2016;Rasmussen et al., 2015), but that restoring the expression of Dnmt3a after the leukemia had developed did not revert the phenotype (Yang et al., 2016). That said, the role of Dnmt3a in tumorigenesis is tissue specific, since in the lung it does not affect tumor initiation but rather tumor progression (Gao et al., 2011). Interestingly, in the work of Gao et al., most of the changes in gene expression in Dnmt3a-depleted cells were attributed to alterations in gene body methylation, rather than at promoters. Conversely, in our model, we see significant changes at regulatory elements (i.e. promoters and enhancers) that lead to changes in gene expression in Dnmt3a-depleted epidermal tumors.
Our results, together with accumulating evidence from other groups, demonstrates a clear relationship between the levels of Dnmt3a-Tet-5-hmC and tumorigenesis: the inactivation of this axis in adult stem cells predisposes them to tumor initiation. Interestingly, a global reduction of 5-hmC is a hallmark of several cancer types, including squamous cell carcinoma, and is often correlated with poor prognosis Liao et al., 2016;Shi et al., 2016;Ficz and Gribben, 2014;Lian et al., 2012). Our results indicate that Dnmt3a drives the DNA-methylation, and subsequent hydroxymethylation, of a subset of enhancers that regulate the expression of genes involved differentiation. Conversely, Dnmt3a binds to, and DNA methylates, the promoters of cell proliferation and lipid metabolism genes to repress their expression. Interestingly, however, deletion of Dnmt3a does not result in changes in the specification of the different keratinocyte lineages in the skin (epidermis, hair follicles, and sebaceous glands), nor their homeostasis in adulthood. What is more, even the combined deletion of Dnmt3a and Dnmt3b, albeit significantly reducing overall DNA methylation levels, did not result in any skin phenotype even in aged mice. On the other hand, we have recently shown that Dnmt3a and Dnmt3b are necessary for the self-renewal and differentiation of primary human keratinocytes (Rinaldi et al., 2016). This apparent contradiction in the phenotypes observed might be due to the fact that our work with human keratinocytes relied on culturing the cells, which was recently shown in murine skin keratinocytes to induce a wound healing damagedlike reversible state that affects their epigenome and transcriptome (Adam et al., 2015). Thus, in vivo deletion of Dnmt3a might not be sufficient to alter the homeostasis of undamaged skin, but renders the epidermis more susceptible to situations of damage. Accordingly, the epidermis of Dnmt3a-cKO mice responded to the treatment of DMBA/TPA in a much more pronounced manner. This effect might not be specific to Dnmt3a. For instance, Dnmt1 is the DNA methyltransferase most highly expressed in epidermal cells and is responsible for about 70% of DNA methylation levels (Li et al., 1992). Its depletion causes a strong loss of self-renewal of primary human basal keratinocytes (Sen et al., 2010). However, its loss in murine epidermis leads to a mild increase in proliferation, and to a partial alopecia, only in very aged mice (Li et al., 2012). Intriguingly, these results also suggest that Dnmt1 and Dnmt3a/3b exert different functions in epidermal homeostasis, although future work will be required to study these putative differences in depth.
Notwithstanding the differences between in vivo and ex vivo studies, our results show that the genomic localization of Dnmt3a is very similar between intact murine keratinocytes and cultured human keratinocytes. Besides its localization at active enhancers of genes involved in epidermal differentiation, Dnmt3a also bound to, and methylated, promoters of genes that regulate cell proliferation and lipid metabolism to repress their expression. Among these genes, were the master regulators of lipid metabolism and adipogenesis PPAR-a and PPAR-g. Interestingly, a number of recent studies have highlighted the importance of a persistent lipid metabolism in promoting tumor transformation, and tumor metastasis in colorectal, liver, breast and oral squamous carcinomas, as well as for enhancing chemoresistance of leukemia stem cells (Pascual et al., 2017;Ma et al., 2016; Figure 5 continued (p=2.374e À5 ). (C) Global levels of 5-hmC at enhancer center (-2Kb, + 2 Kb) were quantified using HOMER software in independent biological replicates of FACS sorted tumor cells from wild-type (n = 2) and Dnmt3a-cKO (n = 2) mice at enhancers bound or not by Dnmt3a. (D) Ratio between the 5-hmC levels at enhancers bound or not by Dnmt3a in wild-type and Dnmt3a-cKO tumor cells. DOI: 10.7554/eLife.21697.022 The following figure supplement is available for figure 5: Relative methylation score (CpG count) measured around active and silenced promoters bound by Dnmt3a (-5 kb, +5 kb) from independent biological replicates of FACS-sorted tumor cells from wild type (n = 2) and Dnmt3a-cKO (n = 2) animals. (B) Relative methylation score (CpG count) measured around promoters not bound by Dnmt3a (-5 kb, +5 kb) from independent biological replicates of FACS-sorted tumor cells from wild-type (n = 2) and Figure 6 continued on next page Beyaz et al., 2016;Ye et al., 2016). The upregulation of these transcription factors upon deregulation of Dnmt3a might predispose the epidermis to develop more tumors, suggesting that an intriguing mechanistic link between lipid metabolism and the epigenetic regulation of tissue homeostasis through DNA methylation, might exist. A recent large clinical association study has already pointed to this by establishing a correlation between the expression of obesity-related genes and changes in the content of DNA methylation (Wahl et al., 2017). Importantly, our results show that PPAR-g is partially responsible for promoting tumorigenesis in Dnmt3a-deficient epidermis, which considering that human skin tumors express lower levels of Dnmt3a, might provide us with a new therapeutic antitumor avenue against squamous cell carcinomas.

Chemical skin carcinogenesis
The work with mice was approved by the Ethical Committee for Animal Experimentation (CEEA) of the Scientific Park of Barcelona (PCB), and the Government of Catalunya. Inbred male or female Dnmt3a flox/flox (C57/Bl6) backcrossed to Krt14-CRE-YFP (C57/Bl6) for six to nine generations were used for all animal experiments. Chemically-induced skin carcinogenesis was performed as previously described (Nassar et al., 2015;Abel et al., 2009), with a slight modification to yield high-frequency SCCs in the C57/Bl6 genetic background. Briefly, the back skin of 8-week-old mice-at which time hair follicles are in their resting phase (telogen)-was shaved and treated with the mutagen 7,12dimethylbenz[a]anthracene (DMBA; 200 ml of 0.25 mg/ml solution in acetone) and the pro-inflammatory and pro-proliferation agent 12-O-tetradecanoyl phorbol-13-acetate (TPA; 200 ml of 0.02 mg/ml solution in acetone) once weekly for 6 weeks. Specifically, DMBA was given on Monday and TPA always on the Friday of the same week. For short DMBA experiments, animals were sacrificed and back skins were processed 3 days after the sixth TPA application. For tumor formation studies, treatment continued twice weekly with TPA (200 ml of 20 mg/ml solution in acetone) for up to 20 weeks, or until the largest tumor of each mouse reached 1.5 mm diameter, at which point animals were sacrificed. In total, 12 wild type and 15 Dnmt3a-cKO tumors from 6 and 8 mice, respectively, were included for tumor analyses.

PPAR-g inhibitor treatment
The chemically-induced skin carcinogenesis was performed as previously described above (Nassar et al., 2015;Abel et al., 2009). We used the chemical 2-Chloro-5-nitro-N-phenylbenzamide (Sigma Aldrich: GW9662) described as potent PPARg inhibitor. Briefly, the shaved dorsal epidermis of wild-type and Dnmt3a-cKO mice was treated twice a week topically with 200 ml of 100 nmoles of GW9662 solubilized in acetone. We applied the PPAR-g inhibitor together with the first DMBA treatment, and subsequently administered it at every DMBA or TPA treatment. The PPAR-g inhibitor was applied always 2 min before every administration of DMBA or TPA (Sahu et al., 2012;Grabacka et al., 2006).

Single-cell preparation and FACS analysis
To isolate pre-cancerous epidermal cells following short DMBA/TPA treatment, back skins were dissected and processed to single-cell suspensions as previously described (Jensen et al., 2010). For ChIP-seq, single-cell suspensions were cross-linked for 10 min at room temperature with 1% formaldehyde (methanol-free; Thermofisher, 28906) and quenched for 5 min to a final concentration of 0.125M of glycine. Cells were washed 2Â with cold PBS and frozen at -80˚C.
Tumor cells (YFP bright /lineage neg cells), pre-cancerous epidermis of interfollicular epidermis (YFPpos /CD49f high /CD34 neg cells), and bulge hair follicle stem cells (YFP bright /CD49f high /CD34 pos cells) were FACS-sorted using a BD FACSAria Fusion flow cytometer (BD Biosciences). Approximately, 3-20 Â 10 4 cells were sorted and lysed in 1 ml of TRIzol for RNA and DNA isolation. After adding 200 ml chloroform, samples were vortexed for 30 s and then centrifuged at 12,000 g to separate the RNA-containing supernatant from the organic phase. RNA was precipitated with 1Â volume of isopropanol, washed twice with 70% ethanol, and then used for library preparation. The interphase of the TRIzol solution (after removal of the supernatant) was precipitated adding 1Â volume of isopropanol, centrifuged for 1 hr at 4˚C at 13,000 g, washed twice with ethanol, and digested overnight at 55˚C with proteinase K (10 mg/ml) in TE 1Â buffer. The following day, digested material was incubated 1 hr at 37˚C with RNase A and purified using a conventional phenol/chloroform separation. The DNA pellet was quantified, and DNA was used for library preparation for MeDIP-seq and hMe-DIP-seq experiments.

MeDIP and hMeDIP sequencing
Purified genomic DNA (250 ng) from tumor cells was sonicated to obtain fragments of 300-700 bp. Adaptors from the NEBNext Ultra DNA Library Prep Kit for Illumina were added to the fragmented DNA. DNA was denatured for 10 min at 99˚C and cooled to avoid re-annealing. Fragmented DNA was incubated overnight with 1 mg of antibodies (5-methylcytosine, Abcam cat. # ab10805; 5hydroxymethylcytosine, Active Motif, cat. # 39769, RRID: AB_10013602) previously cross-linked with 15 ml of Dynabeads Protein A (Life Technologies). Immunocomplexes were recovered using 8 ml for 2 hr. The following morning, DNA was washed three times for 10 min each, and purified DNA was extracted using QIAquick MinElute (Qiagen). Amplified libraries were prepared using NEBNext Ultra DNA Library Prep Kit for Illumina (E7370L) following the manufacturer's instructions.

RNA library preparation and sequencing
The libraries of total RNA from wild type and Dnmt3a-cKO tumors was prepared using the TruSeq-Stranded Total Sample Preparation kit (Illumina Inc.) according to the manufacturer's protocol. Each library was sequenced using TruSeq SBS Kit v3-HS, in paired end-mode with the read length 2 Â 76 bp. A minimal of 137 million paired-end reads was generated for each sample run in one sequencing lane on HiSeq2000 (Illumina, Inc) following the manufacturer's protocol. Images analysis, base calling, and quality scoring of the run were processed using the manufacturer's software Real-Time Analysis (RTA 1.13.48) and followed by generation of FASTQ sequence files by CASAVA.

RNA-seq data processing
RNA-seq datasets were pre-processed by removing both low-quality bases from the 3 0 -ends of the reads and adapter sequences using Trimmomatic (version 0.33) (Bolger et al., 2014). The trimmed reads were aligned to the mouse genome (UCSC mm10) using TopHat (version 2.0.13) , with default parameters and -g 5. Gene and transcript expression levels were quantified with HTSeq (version 0.6.1p1) (Anders et al., 2015). From the raw counts, counts per million (cpm) and fragments per kilobase of transcript per million mapped reads (fpkm) values were calculated. Differential expression analysis was performed using DESeq2 (Love et al., 2014) using a q-value cutoff of 0.05 and a fold-change cutoff of 1.5 to identify differentially expressed genes.

ChIP-seq data processing
ChIP-seq datasets were aligned to the mouse genome build mm10 using BowTie (version 1.0.1) (Langmead et al., 2009); the parameters used were -k 1, -m 1, and -n 2. UCSC browser tracks (Kent et al., 2002) were created from the mapped bam file after converting it to bedGraph (normalized to 10 million reads) and subsequently bigWig format. Peak calling of Dnmt3A to determine regions of ChIP-seq enrichment over the background was done with the MACS version 1.4.1. Peaks of the methylation and hydroxymethylation datasets were determined similarly. For histone marks, MACS version two was used with parameters -broad, -q 0.01, and -g mm. ChIP-seq peaks were annotated using the annotatePeaks.pl script of the HOMER suite (version 4.6) (Heinz et al., 2010) using the UCSC mm10 annotation. The coverage depths of different ChIP-seq experiments at specified regions were also calculated using the annotatePeaks.pl script. This generated a normalized coverage value of different sequencing experiments at equally spaced bins spanning the region of interest. Bin size was set to 1 bp.
For the differential regulation analysis of MeDIP-seq data with replicates, common peaks were first determined among the replicates of the wild type and KO samples separately. A consensus peakset was then created from the two common peaksets, and the read counts were calculated for all the peaks of the consensus peakset. DESeq2 (Love et al., 2014) was applied to calculate the differentially bound peaks using an adjusted p value of < 0.05.

Statistical analysis
To compare tumor burden between genotypes, we used a T-Test with 95% confidence. To compare free tumor survival differences and anagen entry differences, we used a Chi-Square test. To compare Relative Methylation Score (RMS) levels and to compare normalized 5-hmC levels between wild type and Dnmt3a-cKO sorted tumor cells we used a paired Wilcoxon Test. The same paired-Wilcoxon test was used to measure differences in RNA expression.

KI67 staining and quantification
Skin and tumor sections were stained after deparaffinization with KI67 (Abcam ab15580; RRID: AB_ 443209) for 60 min. After two washes, section were incubated with Power Vision Rabbit (Inmuno-Logic) for 45 min. Positive staining was revealed using a chromogen DAB for 5 min (Dako). Counterstain for hematoxylin was incubated for 3 min (Dako).
Stained sections were scanned using a high-resolution NanoZoomer 2.0 HT (Hamamatsu). KI67positive nuclei in the interfollicular epidermis were measured using the TMarker software (Schüffler et al., 2013). Positive and negative nuclei for the staining were trained using the color deconvolution plugin and quantified using the cancer nucleus classification plugin. The total number of positive nuclei was normalized to the total number of nuclei in the area considered. Unpaired parametric T-test was used to measure statistical difference among groups and genotypes. epidermis. Spreadsheet3 contains the genomic coordinates of the active enhancers not bound by Dnmt3a in precancerous epidermis. Spreadsheet4 contains the genomic coordinates of the promoters not bound by Dnmt3a in precancerous epidermis. Spreadsheet5 contains the genomic coordinates of the active enhancers bound by Dnmt3a in precancerous epidermis. DOI: 10.7554/eLife.21697.032

Major datasets
The following dataset was generated: