Gene expression data

Gene expression data were obtained from microarray experiments, downloaded from the Stanford Microarray Database [13]. Ignoring cell line and time course experiments, we collected a dataset of 1402 human two-color cDNA microarrays corresponding to 58 discrete tissues types and about 15000 individual genes. Screening for tumor relevant information in our dataset resulted in 836 tumor arrays, which could be assigned to 33 tumor tissue types defined as tumor entities in our analysis. Additionally 16 tumor classes were allocated to tumor entities representing the tumor differentiation and developmental origin of the tissue, e.g. tumor entities 'Prostate Tumor, Grade T2b', 'Prostate Tumor, Grade T3a' and 'Prostate Tumor, Grade 3b' were assigned to the tumor class 'Prostate'. Downloaded microarray experiments had been analyzed against diverse human controls [13]. In order to avoid divergence due to diverse control data sets, we filtered all tumor classes for common applied controls used in the respective microarray experiments. We always used those data with the largest number of individual datasets with the same one control per tumor class for further evaluation. All other data were discarded. Finally our dataset was reduced to 649 microarrays belonging to 28 tumor entities respectively 16 tumor classes, which were analyzed with the R Project statistical software [14] to identify differentially expressed genes as follows:

Initially the raw data were loaded into R, invalid cDNA probes removed (e.g. flag 'contaminated' or 'failed') and a background correction was applied. After the calculation of the log2 fold changes relative to the control RNA the quantile normalization was applied to each experiment. In all these initial steps we made use of the 'limma' R package [15]. Next, we calculated the variance as well as mean for each gene within the corresponding tumor entities. Moreover, the pooled variance for each gene was calculated to identify global outliers. Pooled variance was defined as weighted mean of the variances within each tumor entity and weighted by the relative sizes of the respective entity. Genes (426 in total) with a pooled variance of all genes greater than the third interquartile range of the pooled variance (3iQR = 0.038) were defined as outliers and excluded from further analysis, to ensure the comparability and correct technical bias due to heterogeneity of the dataset. Additionally, tumor entity specific outliers were eliminated by considering those genes whose variance were greater than the 3iQR of the corresponding tumor entity; depending on tumor entity between 364 and 1053 genes were excluded from the dataset.

A gene was considered upregulated (respectively downregulated) if it had a M-value > 1.0 (resp. < -1.0) and a p-value smaller than 0.01 in a one sided one-sample t-test. This procedure defined between 391 and 1433 up- and between 424 and 2150 downregulated tumor associated genes (figure 1a). Note: Although the p-values produced this way are merely used as a score and are not interpreted as type-1 error rates, it might be interesting to assess the significance of the individual tests. We therefore performed an FDR correction using "p.adjust" and provided the list of multiple testing corrected p-values in the additional files (Additional file 1).

Functional data

Metabolic reaction and signaling event information were acquired from the PATHWAY database of the Kyoto Encyclopedia of Genes and Genomes (KEGG) [16]. These pathway maps and pathway modules are categorized into metabolism, genetic information processing, environmental information processing, cellular process, organismal systems and human diseases in several organisms. Our analysis was based on the KEGG v47 and comprises 210 human pathway maps in XML file format corresponding to 6 categories and 32 subcategories. The data are publically available and can be accessed from the FTP ftp://ftp.genome.jp/pub/. The KEGG pathway data were parsed by a Perl [17] script using the XML::simple module, resulting in 10001 gene-pathway relations belonging to 208 pathway maps and 4186 unique genes (figure 1b).

Integration of gene expression and functional data

For each tumor entity - KEGG pathway pair, we tested whether the set of genes that are differentially expressed in the given tumor entity is enriched in the set of genes belonging to the given KEGG pathway. Testing was performed using an one-sided exact Fisher test. This resulted in a 28 tumor entities × 208 KEGG pathways matrix containing the p-values of each test.

To get a better overview, we implemented a Python [18] script, using the PIL module, to visualize significant enriched KEGG pathway maps (p < 0.05) within the tumor entities into a clear structured functional cancer map (figure 1c) (figure 2). The same procedure as described here for the set of all differentially expressed genes can be applied to the set of upregulated resp. downregulated genes separately. The corresponding functional cancer maps are given in the additional files (Additional file 2, Additional file 3). Note: A separately generated functional cancer map without prior filtering for outlier genes (pooled variance and tissue specific variance) was added to the additional files (Additional file 4).

Literature Mining

Using the web-based tool PubMatrix [19] for literature mining of PubMed [20], we analyzed our significantly enriched KEGG pathway maps against an assortment of tumor associated terms to identify pathways previously mentioned in context to cancer. Tumor associated keywords used in the PubMatrix searches included "Neoplasm", "Tumors", "Tumor", "Benign Neoplasms", "Neoplasms, Benign", "Benign Neoplasm", "Neoplasm, Benign", "Cancer", "Cancers" and "neoplasms[mesh]". PubMatrix is limited to 10000 results for each search and modifier terms. All search results can be found in the additional files (Additional files 5).

Calculation of tumor phylogeny

In a second step, we used the binarized functional cancer map (1 respectively 0 representing significant resp. non-significant entries) to calculate a similarity score between tumor entities. The similarity between two tumor entities T1 and T2, was quantified by calculating the Cohen's Kappa Coefficient [21] of corresponding columns in the binarized functional cancer map. This coefficient is a statistical measure of agreement for qualitative items.

We calculated the quadratic agreement matrix among the tumor entities. The result was visualized with the R function heatmap [22]. Euclidean distance average linkage hierarchical clustering [23] was chosen for the dendrogram included in the plot (figure 3). Visualizations of the tumor similarity matrices with respect to upregulated and downregulated genes are given in the additional files (Additional file 6, Additional file 7).

Gene expression data

Exposing functional mechanisms

Focusing on deregulated genetic signaling events across multiple tumors as well as metabolic events in tumor progression, we analyzed the differentially expressed genes in context of the functional categories as well as signaling and metabolic pathway maps provided by the Kyoto Encyclopedia of Genes and Genomes (KEGG) [16]. After assigning the KEGG data to gene expression data, a total of 141 KEGG pathway maps were significantly enriched and revealed four well conserved subcategories among the tumor entities: 'Cell Growth and Death', 'Signaling Molecules and Interactions', 'Immune System' and 'Immune System Diseases'. The most highly conserved KEGG pathway was 'Cell adhesion molecules (CAMs)' (p value ranging between 5.58e-13 and 0.01) which is significantly enriched in 26 tumor entities, followed by the KEGG pathway map 'Asthma' (1.37e-12 - 0.01), 'Cell cycle' (1.42e-09 - 0.02) and 'Hematopoietic cell lineage' (1.07e-7 - 0.04) in 25 tumor entities. 'Antigen processing and presentation' (2.59e10 - 0.02), 'ECM-receptor interaction' (7.11e-9 - 0.04), 'Complement and coagulation cascades' (1.66e-15 - 0.03) and 'Graft-versus-host disease' (9.37e-10 - 0.04) KEGG pathway maps were found to be significantly enriched in 24 tumor entities. As a result, we generated a functional cancer map to visualize this large amount of multiple functional tumor data into a clearly structured compendium (figure 2).

Literature Mining

To evaluate the significantly enriched KEGG pathway maps in context of published literature to tumor progression already described and mostly unknown we utilized the web-based tool PubMatrix [19]. The 'Cell cycle' pathway was the most cited pathway with focus on tumor progression displaying more than 1000000 search results per modifier terms in total. Furthermore, 'Melanoma' and 'Acute myeloid leukemia' with 718282 respectively 493765 tumor associated publications were ranked at the top. Taking into count highly enriched KEGG pathway maps, among the KEGG pathway map 'Cell cycle' in KEGG class 'Cellular Processes', the 'Nucleotide excision repair' (174594) in 'Genetic Information Processing' as well as 'Wnt signaling pathway' (21699) in 'Environmental Information Processing' were highly linked to tumor publications. In contrast, highly enriched KEGG pathway maps 'ECM-receptor interaction' (4), the 'Complement and Coagulation cascades' (35) and the 'PPAR signaling pathway' (238) were mostly undescribed. The complete search results can be found in the additional files (Additional file 5).

Exposing tumor phylogeny

Further use of the functional cancer map was made to calculate tumor phylogeny by laying down new starting points for the efficient combination of tumor therapeutics as well as for expanding the application area of well-established tumor therapies. We postulated that tumor entities with comparable significantly enriched KEGG pathway maps demonstrate comparable activation of biochemical processes and, thus, biochemical behavior. As a measure of closer or distant relationship between two tumor entities we applied the Cohen's Kappa Coefficient [21] to the binarized functional cancer map data. Analyzing phylogeny between diverse tumor entities, highest concordance was found between tumor entities within the same tissue class. The tumor tissue 'Conventional Renal Cell Carcinoma, Grade 2' and 'Conventional Renal Cell Carcinoma, Grade 3' showed the highest concordance κ = 0.77, followed by 'Breast Carcinoma' and 'Breast Carcinoma Invasive Ductal Carcinoma, Grade 2' (κ = 0.74) and 'Breast Carcinoma Invasive Ductal Carcinoma, Grade 2' and 'Breast Carcinoma Invasive Ductal Carcinoma, Grade 2' (κ = 0.70). Examining the tumor entities belonging to different tumor classes, 'Breast Carcinoma Invasive Ductal Carcinoma, Grade 2' and 'Adenocarcinoma, Pancreas, Grade II' (κ = 0.69) displayed the most functional similarity. Also, high similarity in biological behavior was found between 'Conventional Renal Cell Carcinoma, Grade 2' and 'Adenocarcinoma, Pancreas, Grade II' (κ = 0.63) as well as 'Acute Lymphoblastic Leukemia Peripheral Blood, Grade Pre-B' and 'Acute Lymphoblastic Leukemia Bone Marrow, Grade Common' (κ = 0.61) (figure 3).

Exposing functional mechanisms

As a proof of principle and supporting our approach, we found well documented biological events such as cell cycle and apoptosis to be highly conserved among the diverse tumor entities. These findings confirmed the reliability of the generated functional cancer map. Deregulated balance between proliferation and cell death represents a pro-tumorigenic mechansim in human carcinogenesis [33]. Both activation of proliferative signals and an inhibition of death process, leading to survival and consecutively to proliferation, may contribute essential steps in the progression to cancer in affected cells. Among highly enriched pathways in multiple tumors were 'Cell cycle', 'Signaling Molecules and Interactions', and 'Cell Growth and Death' summarizing genes involved in cell proliferation and apoptosis, demonstrating that key tumorigenic features of tumor development were identified by our analysis. Furthermore multiple functional pathways within the categories, 'Immune System' and 'Immune System Diseases' summarizing genes involved in immune response and immunological changes were also highly enriched in almost all tumor entities. Changes in these processes have been extensively documented to be involved in cancer development and growth [34–38]. Whenever normal body cells turn into tumor cells, various antigens on their surface change and were constantly secreted into the circulatory system. The so called tumor antigens were recognized by the immune system and cytotoxic T cells, natural killer cells and macrophages were advised to catch and eliminated the cell that undergo malignant transformation. A disturbed immune surveillance may lead to an increased tumor growth [34, 38].

However, besides extensively studied and repeatedly documented signaling pathways highlighted in our analysis several pathways were found to be commonly enriched among differentially regulated genes which have not been clearly attributed to tumorigenesis. Thus, our work identified key roles of additional signaling pathways in cancer biology. Deregulations or mutations in the molecules involved in the 'Cell Cycle' and 'ECM-receptor interaction' pathway, preserving the physical link between extracellular matrix and actin cytoskeleton, are also closely connected to tumor progression [39]. Besides the 'Cell Cycle' pathway, the 'ECM-receptor interaction' pathway interacts with the 'Cell adhesion molecules (CAMs)'. This balance between cell adhesion and extra cellular molecules is essential for normal cell survival, imbalance between those pathways leads to detaching cells from the extra cellular matrix and therefore enhance metastasis [40, 41]. However, the ECM-receptor interaction and mechanisms of focal adhesion were not in the focus of biomedical research on cancer mechanisms. Over the past decades, the majority of cancer related genetic studies were focussed on genes involved in transcription and cell cycle control. This is reflected in a tenfold higher number of abstracts containing the search terms 'cancer and cell cycle' than 'cancer and extracellular matrix or microenvironment' (122 594 vs. 20 287 articles) when entered in a PubMed query. This may be due to the fact that the plethora of biological functions of ECM proteins has been underestimated for a long time. For a long time functional properties of the extracellular microenvironment and matrix was believed to be restricted to form the "glue" between cells, responsible for simply to maintain shape and coherence of tissues and organs as well as a reservoir of body fluids. In the last decade the microenvironment gained increasing attention for many other processes including tumorigenesis [42] e.g. the breast microenvironment plays an important and complex role in hepatocarciongenesis. Therefore, cancers are currently regarded as heterogeneous multicellular entities containing cells of multiple lineages who interact with one another, the extracellular matrix (ECM), and soluble molecules in their vicinity are dynamic. Overall, these interactions favour cancer cell proliferation, movement, differentiation, and ECM metabolism, while simultaneously restricting cell death, stationary polarized growth, and ECM stability [43]. However, the microenvironment is also essential to key processes such as angiogenesis. Another important factor is the appreciation that cell migration, matrix and tissue remodeling are not unique properties of cancer cell growths but instead are tightly regulated programs normally utilized during development and in adult tissues responding to acute tissue injury [43]. Reciprocal interactions between responding "normal" cells, their mediators, structural components of the ECM, and genetically altered neoplastic cells regulate all aspects of tumourigenicity. An additional role was suggested for complex interaction of the extra cellular tumour stroma with (mesenchymal) stem cells involved in cancer growth [44]. Thus, demonstrating that functional pathways involved in ECM formation and loosening of focal adhesions are differentially regulated in almost all cancer entities, our results clearly advocate for enhanced efforts in unravelling molecular mechanisms of cancer development influenced by these pathways.

Almost no studies were published on the functional pathway of 'Complement and Coagulation cascades'. Few studies demonstrated a potential role of individual genes within this pathway in cancer development. Some members of the complement system can be modulated by the antifibrinolytic protein, activated thrombin-activatable fibrinolysis inhibitor (TAFIa), by inactivating the anaphylatoxins C3a, C4a and C5a and dysregulation of those genes has been implicated in tumor growth [45, 46]. In addition a link between the haemostatic system and both tumor stroma and metastasis has been noticed in animal and in vitro models. The relationship between cancer and blood coagulation has been described on a diverse pathophysiological level. Cancer development and growth was shown to go along with hypercoagulable state and predisposes to activation of platelets and thrombosis, blood coagulation and fibrinolysis interfering with tumor cell biology, tumor growth, angiogenesis and metastatic process. However, the detailed molecular mechanisms of anti-coagulants such as warfarin on cancer remain elusive [47, 48]. Given the strong enrichment of differentially regulated genes linked to the functional pathway of complement and coagulation, our data suggested that it may be worth to further elucidate exact underlying molecular mechanisms as they may hold key regulatory mechanisms in cancer development and growth.

Another promising candidate for further evaluation was the highly conserved PPAR signaling pathway. It is known that the peroxisome proliferation-activated receptor (PPAR) signaling pathway regulates a multitude of genes important for diverse cellular functions including cell proliferation, cell differentiation, immune response and apoptosis. Once activated, PPARgamma will preferentially bind with retinoid X receptor alpha and signal antiproliferative, antiangiogenic, and prodifferentiation pathways in several tissue types. Since ligands and other agents influencing this pathway have revealed anticancer effects in a variety of human cancers [49, 50] efforts to target this signaling pathway should certainly be intensified given the broad range of tumors with deregulated PPAR signaling.

Finally, we noticed an enrichment of genes invovlved in the functional pathways "Diabetes type I" and "gluycolysis/gluconeogesis". Besides our data several clinical reports support a synergism between Diabetes mellitus and cancer development. Konishi I et al. reported a correlation between a diabetic pattern of the 75 g oral glucose tolerance test OGTT and hepatocarcinogenesis and considered it a significant risk factor for hepatocellular carcinoma (HCC) in patients with hepatitis C virus [51]. Also Huo et al. found that Diabetes mellitus independently predicts decreased survival in HCC patients [52]. Finally, El-Serag et al. reported that Diabetes mellitus is associated with an increased risk for HCC [53]. Interestingly the PPARgamma activators are utilized for the treatment of diabetes mellitus pointing towards a congruent signaling in of PPAR pathway and diabetes signatures [50].

In summary, there is accumulating evidence for a synergy between diabetes mellitus and an increased risk for and decreased survival of patients with HCC. Our data suggests that important interactions may be identified by analyzing the enriched functional pathways of "Diabetes type I" and "gluycolysis/gluconeogesis" previously not extensively studied in context of cancer.

Exposing tumor phylogeny

We introduced a measure of profile similarity by which tumor entities were structured based on semantic aggregated profiles, which means significant, individual tumor mechanisms calculated out of gene expression data, into a dendrogram (figure 2). In order to get a measurement for genetic distance and thus to estimate the affinity between the diverse tumor entities, we applied the Cohen's Kappa Coefficient [21] to our binarized functional cancer map and clustered the distant relation matrix (figure 3). Predominant clustering of the tumor entities to their corresponding tumor class is indication for the robustness of the tumor phylogeny.

Notably, examining downregulated genes and clustering of the tumor entities by means of KEGG pathways, we found our sample to almost perfectly cluster according to the germ layer of the tissue of tumor origin. However, this finding needs to be further analyzed and applied to robust statistical testing [54, 55].