Our strategy for the characterization of mouse models of human CRC (Figure 1) relies on gene expression differences and relative patterning across a range of mouse CRC models, normal mouse colon developmental stages, and human CRCs. Achieving this comparison was facilitated by the use of reference RNAs from whole-mouse and normal adult colon reference RNAs for both mouse and human measurements. Mouse tumor samples were profiled on cDNA microarrays using the embryonic day (E)17.5 whole mouse reference RNA identical to that used previously 15 to examine embryonic mouse colon gene expression dynamics from E13.5 to E18.5, during which time the primitive, undifferentiated, pseudo-stratified colonic endoderm becomes a differentiated, single-layered epithelium. This strategy allowed us to construct a gene expression database of mouse colon tumors in which gene expression levels of the tumors could be referenced, ranked, and statistically compared to an average value among the tumors or to embryonic or adult colon gene expression levels on a per-gene basis. First, we compared the four models with each other, then to mouse colon development, and finally to human CRCs using gene ortholog mapping (Figure 1).

To discover gene expression programs underlying differences between etiologically distinct mouse models of CRC, gene expression level values for each transcript in each tumor sample was set to its ratio relative to its median across the series of tumor models. Using non-parametric statistical analyses, 1,798 cDNA transcripts were identified as differentially expressed among the four mouse models of CRC. Five major gene patterns were identified using K-means clustering (clusters C1-C5; Figure 2a, top). Genes belonging to these clusters were strongly associated with annotated gene function categories (see Table 1 for detailed biological descriptions and associations). For example, cluster C1, composed of transcripts that exhibited lower expression in Smad3^-/- tumors and higher expression in AOM, Apc^Min/+ and Tgfb1^-/-; Rag2^-/- tumors, contains 391 transcripts, including Cdk4, Ctnnb1, Myc, Ezh2, Mcm2 and Tcf3. Gene list over-representation analysis using Ingenuity Pathway Analysis applications demonstrated highly significant associations to cell cycle progression, replication, post-transcriptional control and cancer. Similarly, cluster C2, composed of 663 transcripts that exhibited high expression in AOM and Apc^Min/+ tumors, but low in Smad3^-/- and Tgfb1^-/-; Rag2^-/- tumors, included transcripts for contact growth inhibition (Metap1, Pcyox1), mitosis (Mif, Pik1), cell cycle progression and checkpoint control (Id2, Ptp4A2, Tp53).

From the 1,798 transcripts differentially expressed among the four mouse models of CRC, more than 70% (n = 1265) distinguished Apc^Min/+ and AOM tumors versus Smad3^-/- and Tgfb1^-/-; Rag2^-/- tumors (Figure 2a, bottom). If a random or equivalent degree of variance occurred among all classes, there would be far less overlap. The majority of this signature (approximately 75%, n = 904 features) derived from genes over-expressed in Apc^Min/+ and AOM tumors relative to the Smad3^-/- and Tgfb1^-/-; Rag2^-/- tumors (cluster C6). Cluster C6 was functionally enriched for genes linked to canonical WNT signaling (Table 1). These included genes previously identified to be part of this pathway (Cd44, Myc, Stra6, Tcf1, Tcf4 16, Id2, Lef1, Nkd1, Nlk, Twist1 17, Catnb, Csnk1a1, Csnk1d, Csnk1e, Plat, Wif1) as well as genes that appear to be novel canonical WNT signaling targets (for example, Cryl1, Expi, Ifitm3l, Pacsin2, Sox4 16, Ets2, Hnrnpg, Hnrpa1, Id3, Kpnb3, Pais, Pcna, Ranbp11, Rbbp4, Yes 18, Hdac2 19). Moreover, consistent with the over-expression of Myc in tumors from the Apc^Min/+ and AOM models, we detected enrichment of Myc targets, such as Apex, Eef1d, Eif2a, Eif4e, Hsp90, Mif, Mitf, Npm1 20, and the repression of Nibam 20.

To establish a molecular basis for over-expression of canonical WNT target genes in Apc^Min/+ and AOM tumors, we used immunohistochemistry to characterize the relative cellular distribution of β-catenin. Tumors from Apc^Min/+ (Figure 2b, bottom left panel) and AOM (not shown) mice exhibited strong nuclear β-catenin immunoreactivity and reduced membrane staining (see inset), whereas tumors from Smad3^-/- (Figure 2b, bottom right panel) and Tgfb1^-/-; Rag2^-/- (not shown) mice showed strong plasma membrane β-catenin staining with no nuclear accumulation (see inset). Additional tests to confirm the microarray results were also carried out using an independent set of C57BL/6 Apc^Min/+ colon tumor samples analyzed by quantitative real-time PCR (qRT-PCR; Figure 3a) and immunohistochemistry (Figure 3b). All expression patterns identified via microarray analysis were consistent with the qRT-PCR results (n = 9 transcripts, chosen for their demonstration of a range of differential expression characteristics). In situ hybridization analyses using C57BL/6 Apc^Min/+ colon tumor samples also validated that Wif, Tesc, Spock2 and Casp6 were strongly expressed in dysplastic cells of the tumors (data not shown). At the protein level, immunohistochemical analyses confirmed relatively greater expression of the oncoprotein stathmin 1 in Apc^Min/+ mice and tyrosine phosphatase 4a2 in Smad3^-/- mice (Figure 3b).

Overall, cluster C6 genes (that is, genes with greater up-regulation in tumors from Apc^Min/+ and AOM models than in Smad3^-/- and Tgfb1^-/-; Rag2^-/-) were consistent with increased tumor cell proliferation (for example, Myc, Pcna), cytokinesis (for example, Amot, Cxcl5), chromatin remodeling (for example, Ets2, Hdac2, Set) as well as cell cycle progression and mitosis (for example, Cdk1, Cdk4, Cul1, Plk1). It is important to note that Myc is up-regulated in all four mouse tumor models relative to normal colon tissue (see below). Biological processes showing increased transcription in tumors from the Smad3^-/- and Tgfb1^-/-; Rag2^-/- models (cluster C7) included immune and defense responses (for example, Il18, Irf1, Myd88), endocytosis (for example, Lrp1, Ldlr, Rac1), transport (for example, Abca3, Slc22a5, Slc30a4), and oxidoreductase activity (for example, Gcdh, Prdx6, Xdh) (Table 1). Taken together, these transcriptional observations are both consistent with and extend our understanding of the histological features of the CRC models 7. For example, while Apc^Min/+ and AOM tumors are characterized by cytologic atypia (that is, nuclear crowding, hyperchromasia, increased nucleus-to-cytoplasm ratios and minimal inflammation), tumors from Smad3^-/- and Tgfb1^-/-; Rag2^-/- mice show less overt dysplastic changes but exhibit a significant inflammatory component.

We hypothesized that comparisons of genes over-expressed in both colon tumors and embryonic mouse colon could provide valuable insights into tumor programs important for fundamental aspects of tumor growth and regulation of differentiation. To identify genes and observe regulatory patterns that were shared or differed between colon tumors and embryonic development, we applied a global quantitative referencing strategy to both tumor and embryonic samples by calculating the relative expression of each gene as the ratio of its expression in any sample as that relative to its mean level in adult colon. From this adult baseline reference, genes over-expressed in the four mouse tumor models appeared strikingly similar. Moreover, the vast majority of genes over-expressed in tumors were also over-expressed in embryonic colon (Figure 4a). If the fraction of fetal over-expressed genes from the entire microarray (5,796 of 20,393 features; 28.4%) was maintained at a similar occurrence frequency in the tumor over-expressed fraction (8,804 of 20,393), one would expect an overlap of 2,502 transcripts ((8,804/20,393) × 28.4%). Rather, 4,693 out of the 5,796 fetal over-expressed transcripts were observed to be over-expressed in the 8,804 tumor over-expressed genes (Figure 4b). The probability calculated by Fisher's exact test is p < 1^-300, and thus represents highly significant over-representation of fetal genes among the tumor over-expressed genes. Similarly, genes under-expressed in developing colon were disproportionately underexpressed in tumors relative to normal adult colon (3,282 of 3,541; p < 1^-300). Combining these results, approximately 85% of the developmentally regulated transcripts (7,975 out of 9,337 features) were recapitulated in tumor expression patterns relative to adult colon (Figure 4a,b, green and red markers represent the corresponding 7,975 features).

To explore the potential biological significance of genes over-expressed in both embryonic colon development and mouse tumors, we used K-means clustering to generate C8-C10 cluster patterns as shown in a hierarchical tree heatmap (Figure 4c; Table 2). Several sub-patterns were evident, some of which clearly separated Apc^Min/+ and AOM from Smad3^-/- and Tgfb1^-/-; Rag2^-/- tumors. One strong cluster, cluster C8, consisted of genes more strongly expressed in Apc^Min/+ and AOM than Smad3^-/- and Tgfb1^-/-; Rag2^-/- tumors. This group of genes represented a large fraction of all differences found between nuclear β-catenin-positive (Apc^Min/+ and AOM) and negative (Smad3^-/- and Tgfb1^-/-; Rag2^-/-) tumors (approximately 45%; 1,636 out of 3,592 features), as well as differences detected between early (that is, E13.5-E15.5, ED) and late (E.16.5-E18.5, LD) embryonic colon developmental stages. Thus, the fraction of developmentally regulated genes that are more characteristic of the earlier stages of normal colon development (E13.5-E15.5), are clearly expressed at higher levels in nuclear β-catenin-positive tumors. This observation is illustrated by 750 transcripts selected solely for stronger expression in ED versus LD (Figure 4d). Note that most of these transcripts overlap with cluster C6 containing 230 features (Figure 2a, lower panel) and illustrate the tendency of the earlier-expressed developmental genes to be more strongly expressed in Apc^Min/+ and AOM mice. In addition, transcripts associated with increased differentiation and maturation, observed at later stages of colon development E16.5-E18.5 (for example, Klf4 21, Crohn's disease-related Slc22a5/Octn2 22, Slc30a4/Znt4 23, Sst 24), were expressed at higher levels by tumors from Smad3^-/- and Tgfb1^-/-; Rag2^-/- mice.

Since mouse tumors recapitulated developmental signatures irrespective of their etiology, we asked whether a similar commitment to embryonic gene programming was shared by sporadic human CRCs. Tumor classification by microarray profiling is usually accomplished by referencing relative gene expression levels to the median value for each gene across a series of tumor samples. Using this 'between-tumors median normalization' approach, as well as a gene filtering strategy that detects significantly regulated genes in at least 10% of the cases, led to the identification of a set of 3,285 probe sets corresponding to transcripts whose expression was highly varied between independent human tumor cases. As shown in Figure 5, there was striking heterogeneity of gene expression among 100 human CRCs. For example, cluster 15 contained a set of genes (principally metallothionein genes) recently identified to be predictive of microsatellite instability 2526. This analysis indicates that human CRCs have a greater level of complexity than the mouse colon tumors studied here (compare Figures 2 and 5). There was no correlation between these distinguishing clusters and the stage of the tumor (note the broad overlapping distributions of Dukes stages A-D across these different clusters). However, as shown in Table 3, gene ontology and network analysis of the individual gene clusters (clusters C11-C17) that were differentially active in subgroups of the tumors, map to genes highly associated with a diverse set of biological functions, including lipid metabolism, digestive tract development and function, immune response and cancer

To evaluate if similar sets of genes are systematically activated or repressed in human CRC, as in the mouse colon tumors, we undertook two procedures to align the data. First, gene expression values for the mouse and human tumors were separately normalized and referenced relative to their respective normal adult colon controls; second, mouse and human gene identifiers were reduced to a single ortholog gene identifier. The latter is a somewhat complex procedure that requires identifying microarray probes from each platform that can be mapped to a single gene ortholog and undertaking a procedure to aggregate redundant probes within a platform (see Materials and methods). This approach allowed the identification of 8,621 gene transcripts on the HG-U133 plus2 and Vanderbilt NIA 20 K cDNA arrays for which relative expression values could be mapped for nearly all mouse and human samples. A clustering-based assessment of expression across the whole mouse-human ortholog gene set identified a large number of transcripts behaving similarly across colon tumors, many irrespective, but some respective of species. Notably, the great majority of genes over-expressed in all tumors were also over-expressed during colon development (Figure 6a). To evaluate the statistical significance of this pattern, we used a Venn overlap filtering strategy and Fisher's exact test analysis. Approximately 50% of the 2,212 ortholog genes over-expressed in at least 10% of the human cancers relative to adult colon were also over-expressed in developing colon. If there was not a selection for developmental genes among those over-expressed in tumors, the expected overlap would be (2,718/8,621) × 2,212 = 697 transcripts. Using Fisher's exact test for the significance of the increased overlap of 1,080 versus 697 transcripts is p < 1e-300. Similarly, genes under-expressed in mouse colon development and human CRCs also strongly overlapped (Figure 6b; 431 of 737, p < 1e-76). This result is significantly greater than the 8-19% of genes that were estimated to be over-expressed in human colon tumors and fetal gut morphogenesis based upon a computational extrapolation of SAGE data 27. Thus, our findings not only confirm but also significantly expand and experimentally validate the previously suggested recapitulation of embryonic signatures by human CRCs.

All overlaps between tumor expression and development were pooled to form a set of 2,116 ortholog gene transcripts. This was subjected to hierarchical tree and K-means clustering to define six expression clusters, C18-C23 (Figure 6c; Table 4). These clusters provide an impressive partitioning of groups of genes associated with different biological functions critical for colon development, maturation and oncogenesis. Cluster C22 (860 transcripts of genes strongly expressed both developmentally and across all tumors) is highly enriched with genes associated with cell cycle progression, replication, cancer, tumor morphology and cellular movement. Cluster C18 (258 transcripts down-regulated in mouse and human tumors, as well as in development) is highly enriched in genes associated with digestive tract function, biochemical and lipid metabolism. This cluster is clearly composed of genes associated with the mature GI tract. Thus, as opposed to recapitulating developmental gene activation, the cluster C18 pattern indicates a corresponding arrest of differentiation in both mouse and human tumors. Cluster C23 (142 transcripts over-expressed in all mouse models and human CRC, but with low expression in development) maps to genes highly associated with the disruption of basement membranes, invasion and cell cycle progression, as well as altered transcriptional control. Cluster C21 (313 transcripts in which human tumors somewhat variably express a set of genes that are rarely expressed by the mouse tumors) is remarkable for its composition of genes associated with cell cycle proliferation, tissue disruption and angiogenesis. Thus, while categorically quite similar to cluster C23, the genes in cluster C21 represent a separately regulated module that is enriched for genes associated with invasion. Clusters C21 and C23 reveal sets of genes likely involved in tumor progression. Cluster C22 (with genes over-expressed in all mouse and human tumors and strongly expressed in embryonic colon) represents a group of genes highly correlated with transformation. The top-ranked transcription factor present in this cluster, with regulation independent of β-catenin localization, is Myc/MYC (Figure 7b). Although Myc was lower in expression in the Smad3^-/- tumors compared to tumors from the other three models, it was elevated in all four models relative to normal adult colon. Myc/MYC was over-expressed in all mouse and human tumors as well as in development. This contrasts with Sox4, which is unaltered in expression in the Smad3^-/- and Tgfb1^-/-; Rag2^-/- tumors but is up-regulated in AOM and Apc^Min/+ tumors relative to normal adult colon (Figure 7b). Myc/MYC over-expression may be independent of nuclear β-catenin status. Increased Myc/MYC expression may reflect both activation of canonical Wnt signaling, as it is a target of nuclear β-catenin/TCF 28, and deregulation of TGFβ signaling, as TGFβ1 is known to repress Myc/MYC 293031. These observations suggest a fundamental role for Myc/MYC in colonic neoplasia.

Tumors from mouse models of CRC exhibit significant phenotypic diversity 6, and, therefore, were expected to exhibit differential gene expression patterns. Using a combination of inter-model and normal adult gene expression level referencing, our analysis of tumors from mouse models of CRC has revealed a low complexity between models and strains, and has identified common and unique transcriptional patterns associated with a variety of biological processes and pathway-associated activities. Our results demonstrate an imbalance between proliferation and differentiation, with nuclear β-catenin-positive tumors being more proliferative, less differentiated and with lower immunogenic characteristics than tumors from nuclear β-catenin-negative tumors. Mouse tumors characterized by signatures of relative up-regulation of genes associated with cell cycle progression also showed increased canonical WNT signaling activity (Apc^Min/+ and AOM). Tumors from mouse models not showing canonical WNT signaling pathway activation (Smad3^-/- and Tgfb1^-/-; Rag2^-/-) were characterized by up-regulation of genes associated with inflammatory and innate immunological responses, and intestinal epithelial cell differentiation. Recent studies have indicated that chronic inflammation caused either by infection with Helicobacter pylori 32 or Helicobacter hepaticus 13 is a prerequisite for intestinal tumor development in Smad3^-/- and Tgfb1^-/-; Rag2^-/- mice, respectively.

The activation of canonical WNT signaling in AOM tumors was identified using a between-tumor global median normalization to gene expression data. However, when tumor sample expression was referenced to that of normal adult intestinal tissue, many more genes are up-regulated, including developmental genes that are not dependent on nuclear β-catenin. That canonical WNT signaling-related genes are altered similarly in both AOM and Apc^Min/+ tumors suggests biological similarities between the two models. In addition, the relatively consistent programming within the AOM model also emphasizes its value for examining the more complicated genetics that result in strain-specific sensitivity to environmental agents that induce cancer.

Activation of canonical WNT signaling leads to nuclear translocation of β-catenin and, through its interaction with LEF/TCF, the regulation of genes relevant to embryonic development and proliferation 16, as well as stem cell self-renewal 33. Consequently, the activated canonical WNT signaling observed in Apc^Min/+ and AOM models suggests that tumors may arise as a consequence of proliferation of the stem cell or 'transient amplifying' compartment. In the colonic crypt, loss of TCF4 34 or DKK1 over-expression 35 promotes loss of stem cells, suggesting that canonical WNT signaling is required for the maintenance of the intestinal stem cell compartment 343536. Conversely, increased nuclear β-catenin/TCF4 activity imposes a crypt progenitor phenotype on tumor cells 18. In this study, we identified transcriptional activation of the canonical WNT signaling pathway in tumors from Apc^Min/+ and AOM mice. This was confirmed by immunohistochemistry (Figure 2b).

In colon tumors and perhaps intestinal stem cells, activation of canonical WNT signaling promotes a hyperproliferative state. Proliferation-related characteristics of nuclear β-catenin-positive tumors include increased expression of CCND1, MYC, PCNA 18, and Sox4 16. These genes were also identified as a component of our nuclear-β-catenin-positive signatures. In turn, increased MYC decreases intestinal cell differentiation by binding to and repressing the Cdkn1a (coding for p21^CIP1/WAF1) promoter 37, the gene encoding Wnt-inhibitory factor Wif1, the gene encoding the negative regulator of WNT Naked1 38, and the gene encoding the Tak1/Nemo-like kinase, Nlk 39. Wif1 displays a graded expression in colonic tissue, with higher expression in the stem cell compartments and lower expression in the more differentiated cells at the luminal surface, suggesting that Wif1 may contribute to stem cell pool maintenance independent of WNT signaling inhibition. 40.

Canonical WNT signaling not only governs intestinal cell proliferation, but also cell differentiation and cell positioning along the crypt-lumen axis of epithelial differentiation. Increased canonical WNT signaling activity enhances MATH1-mediated amplification of the gut secretory lineages 41. Canonical WNT signaling also influences cell positioning by regulating the gradient of EPHB2/EPHB3 and EPHB1 ligand expression 4243. Together, our data suggest a complex imbalance of crypt homeostasis due to enhanced canonical WNT activity.

Our results indicate that tumors arising in response to abnormal TGFβ1/SMAD signaling 1444 are similar to one another in their specific gene signatures and broadly distinct from those with activated canonical WNT signaling by their absence of nuclear β-catenin. Unique to the dysregulated TGFβ1/SMAD4 signaling models is the strong signature of an immunologically altered state, with up-regulation of genes determining immune and defense responses, such as Il18, Irf1 and mucin pathway-associated genes. Again, these tumors are usually characterized by a strong inflammatory component when evaluated histopathologically, even in the absence of T- and B-cells such as in the Tgfb1^-/-; Rag2^-/- background.

As shown in Figure 2a, the microarray patterns of gene expression for AOM and Apc^Min/+ tumors are mirror images of those for Tgfb1^-/-; Rag2^-/- tumors. It is perhaps not surprising that combining these two transcriptional programs results in increased number and invasiveness of colonic tumors as recently reported for Apc^Min/+ mice crossed to Smad3^-/- mice 45. Moreover, combined activation of canonical WNT signaling and inhibition of TGFβ signaling also results in more advanced intestinal tumors in Apc^delta716/+; Smad4^+/- mice 46, and intestine-specific deletion of the type II TGFβ receptor in Apc^1638N/wt mice 47.

The findings that shared over-expressed signatures are identifiable in all four mouse models of CRC, which are also representative of the majority of embryonic colonic over-expressed signatures, and that these signatures are also present in all human CRCs, suggest that colon tumors may arise independently of canonical WNT signaling status. A likely candidate to impart this oncogenic signaling is Myc, which is an embryonic up-regulated transcript that is also upregulated in all human CRCs and mouse tumor models independently of nuclear β-catenin status.

It has long been suggested that cancer represents a reversion to an embryonic state, partly based upon the observation that several oncofetal antigens are diagnostic for some tumors 4849. To assess the embryology-related aspects of tumorigenesis and tumor progression in CRC, we analyzed and compared the transcriptomes of normal mouse colon development and models of CRC. Our data show that developmentally regulated genes represent approximately 56% of mouse tumor signatures, and that the tumor signatures from the four mouse models recapitulate approximately 85% of developmentally regulated genes.

There are at least two regulatory programs that determine the expression of developmental genes by mouse tumors (Figures 2, 4, and 8). The simpler program is evident by the over-expression of the earliest genes of colon development by the nuclear β-catenin-positive models. The more subtle program could be detected only in reference to adult colon and is highly shared by nuclear β-catenin-negative models. This program, though modified by nuclear β-catenin status, is represented by a large scale over-expression of developmentally expressed genes in tumors that are both positive and negative for canonical WNT signaling. Genes found within this signature have a large overlap with those present in the colon at later developmental stages (E16.5-E18.5).

How do genes tightly regulated during mouse colon development become activated in colon tumors? While activated canonical WNT signaling imparts a strong influence, its absence in Tgfb1^-/-; Rag2^-/- and Smad3^-/- tumors, as determined by the absence of nuclear β-catenin, did not prevent the large scale activation of developmental/embryonic gene expression. One mechanism may be through epigenetic alterations. In human CRCs, these types of alterations in gene expression programs 50 suggest a link between cellular homeostasis and tumorigenesis. The recruitment of histone acetyltransferases and histone deacetylases (HDACs) are key steps in the regulation of cell proliferation and differentiation during normal development and carcinogenesis 51. Induction of Hdac2 expression occurs in 82% of human CRCs as well as in tumors from Apc^Min/+ mice 19. Alternatively, common regulatory controls may operate in parallel growth and differentiation/anti-diifferentiation pathways such that a single or small subset of regulators, such as MYC or one or more micro RNAs, may be responsible for the control of multiple pathways. Indeed, consistent with our observation of nuclear β-catenin-independent activation of Myc in all mouse models and across the board for human CRC, deletion of Myc has recently been demonstrated to completely abrogate nuclear β-catenin-driven small bowel oncogenesis in mouse models 52.

As shown in Figure 5, considerable and intriguing heterogeneity of human CRC is observed among genes highly relevant for differential malignant behavior. However, employing between-tumors normalization and referencing strategies prevents the detection of gene expression patterns that are shared between tumors. Using the adult normal colon as a reference, as shown in Figure 6, a large fraction of differential gene expression relative to adult colon could be demonstrated that recapitulated developmental gene expression by virtue of both activating embryonic colon gene expression and failing to express genes associated with normal colon maturation. Within these developmentally regulated gene sets, our analyses revealed little evidence of CRC subsets, including those suggestive of nuclear β-catenin negative tumors that might approximate the Smad3^-/- and Tgfb1^-/-; Rag2^-/- signature. Our inability to identify distinct subclasses with respect to developmental genes in the human CRCs is perhaps not surprising in that over 80% of microsatellite-unstable (MSI+) CRCs from HNPCC families exhibit nuclear β-catenin 53. In addition, within the developmental genes, little evidence was apparent for signatures related to MSI+ tumors, often associated with HNPCC, although some of this type of signature was perhaps apparent in the median normalized depiction of the tumors as highlighted in Figure 5.

This report constitutes a comprehensive molecular evaluation and comparison of mouse and human colon tumor gene expression profiles. We have greatly improved our ability to compare tumor gene expression profiles between mouse and human tumors by using a referencing strategy in which gene expression levels in the tumor samples are analyzed in relation to gene expression in corresponding normal colon epithelium. This approach has revealed that gene expression patterns are both shared and distinct between mouse models and human CRCs. Although several recent studies have suggested that tumors recapitulate embryonic gene expression 16275455, the present study demonstrates the magnitude of this similarity.

Finally, our results suggest that comparisons made between mouse tumor models, developing embryonic tissues, and human CRCs provides a powerful biological framework from which to observe shared and unique genetic programs associated with human cancer. While ortholog-gene based analyses have been used previously to obtain direct comparison of the molecular features of mouse and human hepatocellular carcinomas 56, our results provide striking support for the hypothesis that cancer represents a subversion of normal embryonic development. By inclusion of detailed mouse embryonic and developmental profile information, our results have revealed critical similarities and differences between the mouse and human tumors that are particularly revealing of oncogenic and tumor suppressor programs, some genes from which should be useful for development of diagnostic biomarkers and identification of therapeutic targets and pathways.