The dog gene expression data set consists of microarray data from 33 mammary samples, comprising 7 healthy and 26 diseased samples. All samples were extracted from a total of 10 animals, each animal bearing one or more tumors.

All dog carcinomas were assigned to grade 1, according to diagnostic criteria proposed by the World Health Organization (WHO) 19. For the purpose of the study, canine samples were grouped in four classes (from 0 to 3), representing increasing levels of aggressiveness (see Additional file 1), as follows. Class 0: 7 normal samples; class 1: 3 'hyperplastic/dysplastic lesions' (representing benign proliferative lesions) and 2 'benign tumors'; class 2: 2 'in situ carcinomas' (malignant tumor with a better prognosis as compared to other malignant lesions) and 2 'tubular carcinomas' (showing well differentiated morphology and no evidence of infiltration); class 3: 6 'simple carcinomas' and 11 'complex carcinomas' (representing aggressive forms of malignant tumors). For simplicity, classes 0, 1, 2 and 3 are respectively, referred to in the text as 'normal', 'benign', 'intermediate' and 'malignant' samples.

The human data set consists of 129 samples, including 68 malignant tumors and 61 matched adjacent normal samples. Human tumors were classified according to TNM classification 20 in stages 1–4, with 3 tumor samples belonging to stage 1, 51 to stage 2, 13 to stage 3 and 1 to stage 4 (see Additional file 2).

Microarray experiments were run on two different custom designed, double channel Agilent microarray platforms (25K probes for human and 44K probes for dog). Both data sets used a pool of normal samples as a common baseline sample, respectively derived from human and dog samples (see Methods). For each sample, gene expression ratio values were calculated by comparing the individual sample (tumor or normal) with the respective common baseline.

To characterize the overall diversity between individual dog samples, we performed an unsupervised hierarchical clustering of all 33 dog samples, arbitrarily assigned to three major classes: 'normal', 'benign and intermediate tumors' and 'malignant tumors' (Figure 1). The first bifurcation of the hierarchical clustering dendrogram identifies two clusters of samples, which represent non-random distributions of normal and malignant samples from the complete population. Normal samples are over-represented in cluster 1 (hypergeometric test; P = 4 × 10^-5), segregating into the same sub-cluster, while malignant tumors are over-represented in cluster 2 (hypergeometric test; P = 1.1 × 10^-3). Interestingly, samples from the 'benign and intermediate tumor' class were evenly distributed across the two clusters, as expected by their intermediate phenotype falling between 'normal' samples and 'malignant tumors'. Genes differentially expressed in the three sample groups were then identified by ANOVA analysis. Using a conservative q-value threshold of 0.001, a list of 1043 genes was selected (see Additional file 3). According to their patterns of expression, six groups of genes could be identified (Figure 2). Each group was analyzed separately by gene set enrichment analysis (see Methods and Additional file 3).

The overall similarity between dog and human samples was analyzed using the unsupervised principal component analysis (PCA) of dog and human samples. The analysis was performed including the expression values of 9.963 genes identified as orthologous between the two species and represented on the respective microarray platforms (see Methods).

All 162 samples from the two species were projected onto the first three principal components (Figure 3). The first component captures most of the variability (31%) and drives the separation of tumor and normal samples of both species. An additional 7% and 6% of the total variability is captured by the second and the third principal components, respectively, which instead drive the separation of groups of normal and tumor samples within each species. As evident from Figure 3, general differences in tumor/normal characteristics of the samples dominates with respect to the dog/human species membership. The species separation is however clearly evident within the group of normal samples while being less pronounced in the group of tumors, most likely because of the higher heterogeneity of these samples as compared to healthy normal tissue. Noticeably, one canine (M6b) tumor clusters in the group of normal samples. In absence of additional information, this could supposedly be explained by a particular cellular composition of this sample, where the epithelial compartment is under-represented with respect to the stromal part.

To identify the most relevant biological 'themes' driving the separation of tumor and normal samples and represented on the first PC, genes most correlated with this axis were analyzed by gene set enrichment analysis (Additional file 4 and Figure 3).

Consistent with the previous analysis run on dog samples only, genes with higher expression in tumors (positive correlation with first PC) were found to be enriched for cell cycle (P = 6e-9) and DNA replication (P = 7e-6) related genes.

Conversely, genes with higher expression in normal samples (negative correlation with first PC) were enriched for genes related to the extracellular matrix (P = 0.0025), cell matrix adhesion (P = 0.0078), cell communication (P = 0.0207) and TGF-β receptor signaling pathway (P = 0.0115). Interestingly, the nine dog tumors classified as 'benign' and 'intermediate' cluster somewhere in between tumor and the normal samples in the PCA space, along the first principal component. This observation is consistent with the biology of low aggressive tumors, which maintain some of the biological features of the normal tissue counterpart such as low proliferation rate and limited vascularization.

By using the Student's t-test we separately identified for human and dog, those genes globally deregulated in tumors as compared to their normal counterparts (Figure 4 and Additional file 5). To allow for a consistent comparison between human and dog, samples classified as 'benign' or 'intermediate' were excluded from the t-test analysis of the dog data set. Out of the set of 9.963 orthologous genes, 1.259 genes were identified as significantly up-regulated in dog tumors (q-value < 0.001) and many of them (717, hypergeometric test p = 9e-62) were also up-regulated in human tumor samples. Similarly, of the 773 genes down-regulated in dog tumors, a significant number (343, hypergeometric p = 1e-16) were also down-regulated in the human tumour samples. Thus, the analysis reveals the presence of a shared core of genes commonly deregulated in both human and dog mammary tumors. Although statistically significant, we believe this number of 'core genes' represents only a conservative estimate. This is most likely due to the limited size of the dog data set, which results in the reduced power of the t-test used for the identification of differentially expressed genes. The large overlap between genes deregulated (either up or down) in dog and human can also be appreciated in a heat map representation including data from both species, after hierarchical clustering (Figure 5) 3738394041.

By gene set enrichment analysis we attempted to identify those gene sets or pathways showing a global deregulation in tumor samples as compared to their normal counterparts (see Methods).

We defined as "up-" or "down-regulated" those gene sets or pathways statistically enriched in the list of genes up- or down-regulated, respectively, in tumors (the complete results are available in Additional file 6). As expected, a large part of the pathways showing up-regulation in tumors from both species are related to increased proliferating activity and to the general reorganization of cells observed in many tumor types (Figure 6). In particular, we observed gene sets related to cell cycle (S phase, M/G1, G1/S), protein and RNA transport (including RAN mediated), nucleotide and nucleic acid metabolism and transport, DNA repair and metabolism.

Conversely, gene sets and pathways related to cell and tissue development, cell matrix adhesion and cell communication emerged as being down-regulated. These findings are consistent with the notion that silencing of these processes favors tumour cell invasion and migration by coordinately impairing cell differentiation, adhesion to the extracellular matrix and cell-cell communication 4243. More specific sets of genes emerging as perturbed in both human and canine mammary tumors are described in more detail below.

To get a comprehensive view of possible relationships existing between various cancer related signatures in the human and dog mammary tumors, we closely analyzed a large and heterogeneous compendium of published transcriptional signatures, all related to breast cancer. The signatures composing the compendium were defined in independent studies using different methodologies and cancer models, including human primary breast tumors, tumors from transgenic mice, and in vitro cell lines (Table 1) 373839404164656667686970717273. Six signatures were developed in the context of prognostic or diagnostic studies ('prognostic' signatures) while eighteen were reported to monitor the activation status of pathways associated to oncogenesis ('oncogenic' signatures). While being derived independently, each of the signatures can be represented as a set of genes whose up-or down-regulation is associated with a given phenotype, such as the activation/engagement of a pathway or a defined clinical prognosis. Because the clinical phenotype of a tumour is closely related to the underlying biology and ultimately depends on the activation status of a complex network of oncogenic genes and pathways, the distinction between 'oncogenic' and 'prognostic' signatures has to be intended as purely indicative. To provide a metric that can be uniformly applied to all signatures, we defined a simple quantitative 'signature score' by using for each sample the weighted average of the gene expression levels of all genes composing a signature (see Methods). Setting this metric allowed us to investigate all pair wise relationships between different prognostic signatures, both in human and dog tumor samples. For each pair wise comparison the genes common to the respective signature pair were excluded from the computation of Pearson's correlation. This way we insured that signature correlation values were not driven by intrinsically similar gene contents. Starting with the analysis of the prognostic signatures, we observed that most pairs are positively correlated with each other in the human tumor data set (Pearson's correlation > 0.48, P value < 0.0001) (Figures 5 and 7). In some cases, these results reproduce what was also observed elsewhere, regarding the relationships between 'Wound' signature and poor prognosis 41, 'Hypoxia' and poor prognosis 71, Proliferation' and 'Van't Veer' signatures 37.

Similar relationships were observed in the dog tumor samples, again underlining the close resemblance of the dog tumor samples to their human counterparts (Figures 5 and 7). To establish how prognostic and oncogenic signatures are related in the two species, the same analysis was extended to the complete set of signatures of the compendium, including now also the 'oncogenic' signatures. Each signature was split into the up- and the down-regulated arm and, after computing all pair wise correlations independently for human and dog samples, signature were clustered accordingly and the results represented in the form of squared heat maps (Figure 8 and Additional files 7 and 8). Coherent clusters of closely correlated signatures emerged, visible in tumours of both species. Looking more closely at the results, it can be noticed that signatures monitoring the same pathway, for example 'Myc' and 'Response to androgens' represented by two independently developed signatures, cluster close to each other, as expected.

Three major clusters of signatures showing close reciprocal correlations are visible in both tumor types (Figure 8). 'Down', or 'Up' arms of the 'SRC',' B-catenin', 'MEK', 'ERBB2' and 'androgen' (both arms) signatures appear in cluster 1 in agreement with previous studies on the respective pathways. For example, ERBB2 and MEK (MAP2K1 and MAP2K2) are part of the larger 'MAPK pathway' 70 related to the response to androgens 74 and it is therefore unsurprising that these three signatures appear to be co-regulated. Likewise, the 'SRC' and 'B-catenin' pathways are known to be co-regulated in breast cancer and associated with poor survival 6775. Most signatures initially defined as 'prognostic' ('Van't Veer', 'Grade1-Grade3', 'Proliferation', 'Wound'), appear in cluster 2, together with additional signatures such as 'Cell Cycle', 'Cyclin D1', 'Resistance to Tamoxifen' and 'PTEN'. This cluster therefore mirrors the reciprocal association between loss of PTEN and cancer progression, cell cycle, poor prognosis and resistance to Tamoxifen 3738394041. Moreover, the presence of both the 'Myc' and the 'Wound' signature in cluster 2 is consistent with the notion that activation of the Wound signature in primary breast cancer is prominently associated with over-expression of MYC due to gene on chromosome arm 8q 76. The 'ERBB2' signature, although present in cluster 1, nevertheless shows strong correlation with other prognostic signatures, as expected for the important role of ERBB2 over-expression in cancer prognosis 77. Cluster 3 contains co-regulations of the 'Response to androgens' and 'MAPK' signatures, whose corresponding pathways are also reported to be correlated 74. In addition, the 'Raf1', 'EGFR' and 'E2F3' signatures also fall into this cluster, confirming their known associations with the MAPK pathway74. Most importantly, the same reciprocal pair wise correlations were maintained in the two species, providing again strong evidence for a conserved relationship between the underlying networks of genes related to the establishment, progression and clinical outcome of breast tumors.

The canine dataset consists of 26 tumors (17 malignant, 5 benign and 4 intermediate) and 7 normal mammary glands from 10 different donors. Tumors derived from the same animal were always extracted from different glands. Detailed pathological information for each canine sample are available in Additional file 1.

The 33 canine mammary tissues samples were classified applying the diagnostic criteria proposed by the World Health Organization (WHO) 19 while histological grading was performed according to 78. Therefore, the overall grade was obtained assessing three morphological features: degree of tubular formation, degree of nuclear and cellular pleomorphism, and mitotic count. Morphological diagnosis allowed identification of normal mammary glands (7 samples), hyperplastic/dysplastic lesions (3 samples), adenomas (2 samples), and malignant tumours (17 samples). All samples were snap-frozen in liquid nitrogen after collection.

Samples were assigned to 4 different classes, as follows (see Additional file 1): Class 0: 'normal' samples; Class 1: 'hyperplastic/dysplastic lesions' (representing benign proliferative lesions) and 'adenomas' (benign tumors); Class 2: 'in situ carcinomas' (malignant tumor with a better prognosis as compared to other malignant lesions) and 'tubular carcinomas' (showing well differentiated morphology and no evidence of infiltration); Class 3: 'simple carcinomas' and 'complex carcinomas' (representing aggressive forms of malignant tumors). Classes 0, 1, 2 and 3 are respectively referred in the text as 'normal', 'benign', 'intermediate' and 'malignant' samples.

The human dataset included 68 infiltrating ductal mammary carcinoma and 61 adjacent non-involved tissues. After surgical removal, only samples with greater than 70% tumor cells were retained. Information regarding tumor stage and percentage of tumor cells are provided in Additional file 2. All samples were snap-frozen in liquid nitrogen after collection. Human samples were purchased from Genomics Collaborative, Inc., which was acquired by SeraCare Life Sciences. The samples were appropriately consented for the purposes of our study, in compliance with the Helsinki Declaration http://www.wma.net/e/policy/b3.htm (see Additional files 9 and 10).

Collection of dog samples was submitted for approval to IRBM IACUC (Institutional Animal Care & Use Committee) and then to the Italian Regulatory Agency for approval (Italian Ministry of Health). IRBM is in compliance with European Legislation for the Protection of Animal Used for Experimental & Scientific Purposes (EEC 86/609) and received FULL AAALAC Accreditation with no. 1242. All samples were obtained from client-owned bitches undergoing routine mastectomy. To obtain the histological samples, no experimental protocols were applied. Nevertheless, a written informed consent was obtained from all dog owners.

RNA was isolated, labelled and hybridized as in 40. Briefly, total RNA was isolated with RNAzolB and finally dissolved in RNase-free water. Then 25 μg of total RNA was treated with DNase using the Qiagen RNase-free DNase kit and RNeasy spin columns. Total RNA was dissolved in RNase-free water to a final concentration of 0.2 μg/μl. cRNA was generated by in vitro transcription using T7 RNA polymerase on 5 μg of total RNA and labelled with Cy5 or Cy3 (Cy Dye, Amersham Pharmacia Biotech). 5 μg of labelled-RNA from each sample were co-hybridized with 5 μg of a normal reference pool, consisting of an equal amount of cRNA extracted from mammary healthy samples of the corresponding species.

Labelled cRNAs were fragmented to an average size of 50–100 nucleotides by heating the samples to 60°C with 10 mM of zinc chloride and then adding an hybridization buffer containing 1 M NaCl, 0.5% sodium sarcosine, 50 mM MES, pH6.5, and formamide to a final concentration of 30%. The final volume was 3 ml at 40°C.

The human samples were hybridized on a Human 25K array containing 23,720 unique probes for ~21,000 human genes. Canine samples were hybridized on a Dog 44k array containing 39,558 unique probes corresponding to ~30,000 canine genes. Based on 3' end distance and cross-hybridization in-silico assessment, potentially unreliable probes were removed from the analysis. This filter yielded a final number of 28,541 and 17,888 probes for dog and human respectively. The selection of reliable probes was done using the genomic assembly hg18 and canFam2 for human and dog respectively. The array designs Human 25k (v 3.0 A1) and Dog 44k 1.0 were submitted to GEO with accessions GPL3991 and GPL7198 respectively. Both human and dog arrays were manufactured by Agilent (Agilent Technologies Inc., Santa Clara, CA, USA). The probes were 60 bp in length and were selected based on the oligonucleotide probe design program 79. Each sample was hybridized in duplicate with fluor reversal to systematically correct for dye bias. After hybridization, slides were washed and scanned using a confocal laser scanner (Agilent Technologies). The intensities obtained after scanning were quantified, background-corrected and normalized. Expression changes between each sample (tumours and normals) and the corresponding normal reference pool were quantified as the log₁₀ of the expression ratio. The Rosetta error model 80 was applied to dye-swapped pairs of hybridizations to compute a weighted average ratio per gene and the corresponding p-value indicating the probability a gene is falsely classified as deregulated. All microarray data were submitted as a unique data series to GEO, with accession GSE14999.

Prior to clustering the dog mammary dataset was filtered for non variable genes by removing those genes that showed significant changes in expression (fold-change > 3 or < 1/3) in less than 3 samples. For each gene, fold change was calculated as the ratio between each sample and the corresponding normal reference pool. Therefore for human samples the fold change was the ratio against the pool of human normal samples while for the dog samples the fold change was the ratio against the pool of dog normal samples.

Using these criteria, 2,008 genes were selected for an average linkage hierarchical clustering based on Pearson correlation coefficients. The overrepresentation of normal, benign-intermediate and malignant samples in each cluster was assessed by hypergeometric test.

Starting from the whole set of 28541 reliable dog probes, we selected the genes differentially expressed among the three groups of samples (M, malignant; I-B, intermediate and benign; N, normal) by applying one-way ANOVA test on gene expression ratio values, calculated against the reference pool of normal samples. The resulting p-values were corrected for multiple testing by converting them into the corresponding q-values using the qvalue package 81 of the R statistical programming language 82. The q-value corresponds to the false discovery rate (FDR). Therefore a q-value threshold of 0.001 implies that 0.1% of the genes identified as differentially expressed are false positive. Using the conservative q-value cutoff of 0.001 we selected 1043 genes (Additional file 3). The selected genes have then been submitted to t-test, in which genes were individually tested for up- or down-regulation in each class of samples against the others. In this way it was determined in which sample class each selected gene was found to be differentially expressed. Genes could therefore be classified in 6 groups, according to their pattern of expression. For display purposes, the groups of samples (M, I-B and N) and genes (A to F) were clustered separately using an average linkage hierarchical clustering and the Pearson correlation as similarity measure.

To compare human and dog datasets, we have identified the set of genes orthologous between the two species. Based on conservative criteria (E-value of homolog relationship < 1E-10; the homologous gene is the best BLAST match in both species; conserved synteny), we defined a set of 9,963 orthologs (the ratio Tumor vs Pool of Normal, the error and the corresponding p-value for the combined dataset are available as Additional files 11, 12 and 13). Prior to combining the dog and human datasets for hierarchical clustering, we selected the subset of most variable genes having a three fold-change (or greater) in at least 3 human and dog samples. This filter yielded 600 orthologs. Genes were then standardized (mean 0 and standard deviation 1) independently in each dataset. Finally, the combined dataset was clustered using an average linkage algorithm based on Pearson correlation coefficients.

Differences in average gene expression between the set of malignant tumours and normal samples were computed independently in the human and canine datasets by t-test. P-values were adjusted for multiple comparisons and genes differentially up- or down- regulated compared to the normal counterpart (q-value < 0.001) were identified. Canine samples classified as intermediate or benign were removed from the analysis.

Groups of genes identified in previous steps were compared to annotated gene sets in order to identify the functional classes that were significantly over-represented. Enrichment p-values were computed according to the Fisher's exact test. A total number of about 12,000 annotated gene sets were obtained from publicly available sources (Gene Ontology 83, KEGG 84, Interpro 85, Panther 86, oPOSSUM 87), gene sets of relevance to cancer taken from published references 37394445688889909192) and commercial sources (GeneGo (GeneGo Inc., St Joseph, MI, USA), Ingenuity (Ingenuity Systems Inc, Mountain View, CA, USA), TRANSFAC 93).

In light of a number of considerations, we decided to use and report the uncorrected p-values and do not correct for multiple testing. First, the effective number of gene sets effectively taken into account to generate the presented results is significantly lower the complete set used to run the automated analysis. Second, there is a very high degree of overlap between different gene sets. Third, as we also demonstrated in the signature correlation analysis, many gene sets are closely related to each other in terms of transcription. As a consequence, single tests performed on the individual gene sets are strongly dependent on each other, violating the assumption of independence required by standard correction methods such as 'Bonferroni', 'Holm' and 'FDR'. Thus in this context, standard correction for multiple testing would have resulted as too conservative. To be noticed also that most of the discussed pathways have a p-value much lower than the standard threshold of 0.05.

We analyzed the expression patterns of 24 published transcriptional signatures of breast cancer collected from different studies (see Table 1 for details). All signatures include genes up- and down-regulated. A signature is then composed of one or two 'arms', containing perturbed genes either down- and up-regulated. In the analyses shown in Figures 5 and 7 a global 'signature score' was assigned to each sample by using the weighted average of the gene expression level of all the genes of the signature. For each gene, the weight was respectively -1 and +1 depending whether the gene was part of the down- or the up-regulated arm of the signature. In the analysis of Figure 8, when signatures were composed of two arms, individual arm scores were computed and analyzed separately for each sample, by averaging the corresponding gene expression levels.

In all correlation analyses pair wise Pearson correlations between signatures were computed independently for the human and dog datasets, after removing the normal samples. For each signature pair, genes in common were excluded from the computation of the scores. In this way we insured that observed correlations could not be ascribed to a similar gene composition of the different signatures.

In Figure 8 all signature arms were clustered using the calculated correlation values and the clustering results were represented as squared heat maps.