Targeting changes in cancer: assessing pathway stability by comparing pathway gene expression coherence levels in tumor and normal tissues

Introduction

Cancer is essentially a disease arising from an accumulation of genetic abnormalities (1–3), which are thought to participate in neoplastic development and, in some cases, the development of chemotherapeutic resistance (4–6). Many genes have been implicated in the genesis of various cancers (1, 7). In the process of carcinogenesis, some are found to be mutated whereas others tend to exhibit dysregulated levels of expression (8–11). Both mutation status and RNA or protein expression levels have proved valuable for the development of cancer diagnostic assays, particularly for prediction of prognosis. However, a diagnostic gene expression pattern does not necessarily have a causative role in carcinogenesis (12–17).

A current concept suggests that most genes act as part of one or more pathways. This concept is supported by the frequent observation that qualitative or quantitative changes in the expression of certain genes lead to characteristic cancer phenotypes (18, 19). This observation is supported by the elucidation of distinct biochemical functions for altered cancer genes (20–23). The notion of “pathways” (24) is a convenient abstraction that can be considered in isolation and has been found to be extremely useful in describing and understanding the inner workings of cellular biology (25, 26). The importance of pathways in the context of the entire cellular system is highlighted by the challenges faced in drug discovery today (25, 27), hence the notion of a “systems” approach has gained momentum in identifying pathways related to a disease and suggesting secondary effects of drugs (28–31). The use of pathways also provides a central reference to a more systematic view of biological processes (24, 32, 33), which, when combined with the latest high-throughput experimental and computational methods, has been a driving force for many breakthroughs in systems biology and opportunities to improve the drug discovery process (34–36).

A focused analysis on changes in the expression patterns of specific cellular pathways can reveal biological insights that are not easily apparent from variations in individual genes. Various computational methods have been proposed to analyze gene expression patterns within predefined pathways (37–39). We have previously presented strategies to evaluate the level of coexpression in pathways or functionally related groups of genes using gene expression patterns measured in the National Cancer Institute's 60-cell tumor screen (40) and related pathway gene expressions to differential cytotoxicity of potential anticancer agents screened in the same cell lines and their mechanisms of action (41). The purpose of this study is to extend our analysis to include gene expression data obtained from human tumor and normal tissue samples and evaluate these observations for the purpose of assessing which pathways are deregulated in cancer, and to apply these results toward the development of a rationale for selecting pathway-specific chemo-interventions. The ability to identify and disrupt targets that are characteristic of cancer cells without affecting normal cells is crucial for successful anticancer therapy and is the essence of this investigation. Our analysis will focus on the oligonucleotide microarray samples publicly available at the Whitehead site,³ which encompass gene expression data measured in 190 patient tumor samples spanning 14 common tumor types (18 subtypes) and 90 normal samples including 12 tissue types (13 subtypes; ref. 42). We have previously applied these data to successfully classify tumor tissue samples (43), with the seminal finding that gene expression profiles alone were sufficient to correctly classify most of the tumor tissues according to cancer type.

Our current study examines this data from a new angle (i.e., the pathway perspective) to compare and contrast pathway features using gene expression profiles obtained from normal and tumor tissues. We have organized the tissue gene expression patterns in terms of the widely used, publicly available gene annotations (pathways or functional categories) defined by Kyoto Encyclopedia of Genes and Genomes (KEGG), BioCarta, and Gene Ontology (GO). The gene expression coherence levels in these pathways have been evaluated using our previously developed methods (41) and compared between tumor and normal samples. Pathways that show a significant change in their cohesiveness, which may be an indication of instability in their gene regulation, are identified. We have, in addition, identified pathways that show specific changes in cohesiveness for certain tumor types, evidence supporting their selection as good targets for developing novel anticancer therapies. The pathways that contain known cancer genes have been examined in particular, and their behavior compared with that of other pathways. Finally, we have applied our previously developed methods that connect pathways to small molecules that putatively disrupt their function (41) in an attempt to propose agents that might selectively target cancer-relevant pathways.

Materials and Methods

Gene Expression Data for Tumor and Normal Tissue Samples

Tissue gene expression data were downloaded from the Whitehead website (42).⁴ Data were filtered to exclude all signals labeled as “absent” and all negative signals. Signals were then log transformed by taking the natural logarithms to suppress extreme values, and normalized within each array. For each gene, expression measurements across all samples (either tumor or normal) form a data vector. Gene data vectors of length 90 are built for normal tissue samples and 190 for tumor tissue samples. These data vectors were then filtered such that only genes with measurements in at least one third of the normal tissues or tumor tissues were included for further analysis. This leaves 6,258 gene data vectors for normal tissues and 5,232 gene data vectors for tumor tissues in the final data set.

Pathway Gene Expression Coherence or Cohesiveness Measures

The pathway gene expression coherence or cohesiveness level is evaluated for a total of 962 pathways using the KEGG, BioCarta, and GO annotations (see Supplementary data for details),⁵ following the same procedures as described in an earlier report (40, 41). Briefly, the Kruskal-Wallis H statistic (H-score) is used as a measure of pathway gene expression coherence or cohesiveness to compare intra-pathway and inter-pathway gene-gene correlations. The statistical significance of the H-scores (P values) was established by random permutation tests (41). We consider pathways with H > 3.84 (P < 0.05) as cohesive, and not cohesive otherwise. Cohesiveness of pathways within each particular tissue type (e.g., breast tumor tissue) is calculated using only the gene expression measurements for that particular tissue.

Statistical Significance Evaluation of Changes in Pathway Cohesiveness

Random permutation tests have been conducted to examine the probability of observing “up” and “down” pathways by chance (we define an “up” pathway as a pathway that is not cohesive in normal tissues but has become significantly cohesive in tumor tissues, and a “down” pathway as one that is significantly cohesive in normal tissues but has lost its coherence in tumor tissues). Genes are randomly permutated to generate random pathways. To compare the cohesiveness of each random pathway in tumor versus normal tissues, the total set of 280 tissue samples is also randomly permutated and divided into two sets, containing 190 (number of tumor) and 90 (number of normal) samples each, respectively. The cohesiveness H-score of the pathway within each of the two sets is then calculated and compared. This procedure is repeated 1,000 times and the fractions of up and down pathways were counted. The results in comparison with those observed for the true pathway/tissue systems are summarized in Table 1 . The number of pathways with observed coherence change is significantly higher than random, except for the up pathways.

Gene expression coherence in pathways is also evaluated within each specific tumor or normal tissue following previously described methods (40), except that for each tissue type, only the gene expression measurements for that particular tissue are used in the calculations. The number of samples within a tissue type is typically small (10 per tumor and 5 per normal tissue on average), rendering the cohesiveness measures generated therein inherently volatile (tending to have large variations); therefore, additional randomization tests are done to verify the statistical significance of the observations. A thousand random pathways are built as described above, and the cohesiveness of each pathway is evaluated in tissues with samples selected at random from the total pool of 280 samples. The random permutation of tissue samples is repeated 1,000 times for each random pathway and the cohesiveness H-score of the pathway in the random tumor tissue set is compared with that of the random normal tissue set, and the number of up or down pathways counted. For most tissue types, the number of true up or down pathways observed is significantly higher than random, with the exceptions of the number of up pathways observed for breast, bladder, and renal tissues and the number of down pathways for uterus, pancreas, and ovary tissues (see Supplementary Table S1 for detailed results).⁵ Therefore, the results on these tissue types are not included in our analysis.

Results and Discussion

Coexpression of genes has been observed in certain pathways; however, it is not clear whether any difference exists in the level of pathway gene coexpression between cancer and normal cells. If we assume that coexpression is reflective of coordinated gene regulation, then any change in the level of coexpression in a pathway can be viewed as a change in that pathway regulation, and the propensity of a pathway to changes in regulation may be indicative of instability. The level of gene expression coherence can be evaluated for each pathway annotated by KEGG, BioCarta, or GO (strictly speaking, GO annotates functionally related genes that are not necessarily “pathways” per se; however, we will refer to GO terms as “pathways” in the rest of the text just for convenience), following the same procedures as previously reported (40, 41). Briefly, the Kruskal-Wallis H statistic (H-score) is computed to compare gene-gene expression correlations within a pathway to those between pathways, and used as a measure of pathway gene expression coherence or cohesiveness. We consider pathways with significantly stronger intra-pathway than inter-pathway gene-gene correlations, characterized by a large and positive H-score (P < 0.05), as cohesive, and not cohesive otherwise (see Materials and Methods for details).

Pathway Cohesiveness in Tumor versus Normal Tissues

As an initial assessment of pathway gene expression coherence or cohesiveness in tumor versus normal tissues, all 190 tumor and 90 normal tissue samples are included in computing the H-scores for all three pathway collections (see Supplementary Fig. S1 for pathway H-score distributions).⁵ Genes in vetted pathways are found to be significantly more coherently expressed than a random set of genes regardless of whether the tissue is normal or tumor bearing [normal, P = 2.52 × 10⁻⁴; tumor, P = 4.39 × 10⁻³ (t test)]. However, a more interesting question is whether differences exist in measures of cohesiveness between normal and tumor tissues for individual pathways. Hence, we examine the extent to which each individual pathway has changed its cohesiveness, what types of pathways they are, and whether they have become more or less cohesive in tumor tissues as compared with normal tissues.

The 962 pathways we have analyzed can be divided into four different categories according to their relative pathway gene expression coherence or cohesiveness in tumor compared with normal tissues (see Supplementary Fig. S2):⁵ those that are cohesive only in normal tissues (155; 16.1%), those cohesive only in tumor tissues (89; 9.3%), those cohesive in both (101; 10.5%), or those not cohesive in either (617; 64.1%). This implies that whereas the pathway gene expression coherence or cohesiveness level has remained statistically the same for most pathways (10.5% + 64.1% = 74.6%), there are also pathways that have shown a change in their cohesiveness when comparing normal to tumor tissues (16.1% + 9.3% = 25.4%). Moreover, as shown in Supplementary Fig. S2⁵, significantly more pathways have become less cohesive when compared with the number of pathways that have become more cohesive in tumor tissues (16.1% versus 9.3%; P = 5.78 × 10⁻⁶, Fisher's exact test). In the remainder of the text, we will refer to pathways that have become significantly cohesive in tumor tissues as up pathways and those that have lost their coherence in tumor tissues as down pathways. Random permutation tests to examine the probability of observing up and down pathways by chance (see Materials and Methods for details) show that our observation of 16% down pathways is significantly higher than random (P = 0.002). Taken together, these results indicate that pathways tend to become less cohesive in tumor tissues than in normal tissues.

KEGG pathways are grouped into five functional categories and BioCarta groups its pathways into 12 different categories of biological processes (Table 2 ). These categories are useful for establishing statistics about the type of pathways that are prone to variations in their pattern of gene regulation. For each of those categories, the number of up pathways is compared with the number of down pathways, and a Fisher's exact test is done to determine if any detected difference is statistically significant. The results are listed in Table 2. In most pathway categories, the number of up pathways is not significantly different from the number of down pathways; however, significantly more down pathways are found in the BioCarta pathway categories “Cell Signaling” and “Cytokines/Chemokines” and the KEGG pathway category “Environmental Information Processing,” which consists mostly of signaling pathways as well. This indicates that signaling pathways, when compared with other pathways, are more likely to lose their cohesiveness in tumor-bearing tissues. This is consistent with the results of our previous pathway analysis using the gene expression data measured across National Cancer Institute's 60 tumor cell lines, where signaling pathways were mostly found not to be cohesive (40). In this earlier study, we had no knowledge of the cohesiveness of signaling pathways in normal tissues; the results obtained here use both normal and tumor tissue gene expressions and show that signaling pathways in normal tissues are not inherently incoherent. They are, in fact, significantly coherently expressed in normal tissues, but this coherence is lost for tumor tissues. The KEGG category of “Genetic Information Processing” is the pathway category that shows the least change in their coherence level (37.5% pathways are cohesive in both; see Table 2). Pathways in this KEGG category had been shown in our earlier study to be among the most cohesive pathways, whereas our current analysis additionally reveals that these pathways are inherently cohesive in normal tissues and have maintained their cohesiveness in tumor tissues.

Tissue Type–Specific Pathway Cohesiveness

The pathway gene expression coherence or cohesiveness level was also evaluated within each specific tumor or normal tissue type (see Materials and Methods for details) and was found to be significantly higher than what was expected from random on average (average H-scores are 2.44 for tumor tissues, 1.89 for normal tissues, and 0.18 for random samples; P < 0.05, t test). The number and type of tissues in which a pathway is found to be cohesive can potentially be exploited to postulate connections between tissue-related functions of the pathway and its role in cancer. A number of pathways are found to be cohesive in multiple tissue types, which we consider as “universally cohesive,” whereas others are only cohesive in particular tissue types, which we refer to as “specifically cohesive.” Universally cohesive pathways are probably essential for cell survival in general, whereas pathways that are only cohesive in a tumor or normal tissue might play important roles in tumorigenesis and/or growth, especially for that particular tissue type, and therefore represent potential targets for anticancer therapies. Thirty pathways are found to be cohesive in >50% of the 18 tumor tissue types and 19 pathways found to be cohesive in >50% of the 13 normal tissue types. Fifteen pathways are universally cohesive in both tumor and normal tissues (see Supplementary Table S2)⁵. Most of these pathways are involved in protein synthesis and energy metabolism. Conversely, the cohesiveness of some pathways seems to be tissue specific. There are 194 such pathways, 90 of which are specifically cohesive in one type of normal tissue (not cohesive in any tumor tissues; Table 3A ) and 104 of which are specifically cohesive in one type of tumor tissue (not cohesive in any normal tissues; Table 3B). These pathways are therefore not only tissue specific but also have a change of cohesiveness when comparing normal and tumor tissues, thus representing interesting targets for anticancer therapy.

Gene Mutations and Tissue Type–Specific Pathway Cohesiveness and Stability

Alterations in three types of genes are thought to be responsible for tumorigenesis: oncogenes, tumor-suppressor genes, and stability genes (7). Mutations in these three classes of genes can occur in the germ line, resulting in hereditary predispositions to cancer, or in single somatic cells, resulting in sporadic tumors. Several pathways that contain these cancer genes are found to be specifically cohesive in tissue types that are consistent with the tumor types implicated by the corresponding cancer genes. The GO term “DNA repair” contains cancer predisposition genes ATM and TP53. TP53 is a tumor-suppressor gene, mutations of which have been found in major hereditary cancers such as breast, sarcoma, adrenal, and brain tumors. Mutations in ATM, a stability gene, have been found in major heredity cancers such as leukemias, lymphomas, and brain tumors. DNA repair is found by our analysis to be specifically cohesive in the central nervous system (glioblastoma) tumor tissue (Table 3B). The GO term “negative regulation of cell proliferation,” which contains the tumor-suppressor gene neurofibromatosis type 1 (NF1), is also found to be specifically cohesive in glioblastoma. Mutations in NF1 can result in hereditary predispositions to neurofibroma. The GO term “regulation of cyclin dependent protein kinase activity,” which contains the tumor-suppressor gene PTEN, and the BioCarta pathway “Cell Cycle: G₂-M Checkpoint,” which contains the tumor-suppressor gene BRCA1, are both specifically cohesive in mesothelioma tumor tissue. Mutations in PTEN can result in hereditary predispositions to hamartoma (a focal malformation resembling a neoplasm, composed of an overgrowth of mature cells and tissues that normally occur in the affected area), glioma, and uterine cancers, and mutations in BRCA1 to breast and ovarian cancers. Mutations in tumor-suppressor genes EXT1 and EXT2 and stability genes RECQL4 and WRN have all been implicated in hereditary predispositions to bone cancer. The GO terms in which these genes are categorized, “DNA helicase activity” (RECQL4 and WRN) and “endoplasmic reticulum membrane” (EXT1 and EXT2), are found to be specifically cohesive in normal blood tissues but have lost their cohesiveness in tumor tissues.

Other cancer genes are mutated somatically but not inherited in mutant form, and these mutations can result in sporadic tumors. Amplification of the ERBB2 gene has been found in breast and ovarian tumors, and the BioCarta pathway “Role of ERBB2 in Signal Transduction and Oncology” is found to be specifically cohesive in normal uterine tissues but this cohesiveness is not sustained in tumor tissues. Amplification of MYC has been found in lymphomas, neuroblastomas, and small-cell lung cancers. The BioCarta pathway that contains MYC, “Tumor Suppressor Arf Inhibits Ribosomal Biogenesis,” is specifically cohesive in normal germinal tissues, and two other pathways in which MYC participates, the GO term “cell cycle arrest” and the BioCarta pathway “Inhibition of Cellular Proliferation by Gleevec,” are both found to be specifically cohesive in lung (adeno) tissues. Translocations in ABL1 and MLL are characteristic of certain leukemias. The pathways containing these genes that are also specifically cohesive in leukemia include the BioCarta pathway “Lissencephaly gene (LIS1) in neuronal migration and development” (ABL1) and the GO term “GTPase activity” (MLL), which are found to be specifically cohesive in B- and T-cell acute lymphocytic leukemia, respectively. It is worth noting that the recent selective and specific therapies with Gleevec (44) and Erbitux (45) target the BCR/ABL and the epidermal growth factor receptor signaling pathways, respectively.

Pathway Stability as Reflected by Changes in Gene Expression Coherence

Because cancer is a disease closely tied to genetic instability, the relative stability of pathways or pathway gene expression regulation is of great interest. The tendency for some pathways to change their cohesiveness may be a reflection of instability in the regulation of these pathways, which could be either the cause or consequence of cancer, and thus warrants further investigation. As a measure of tissue-related pathway stability, we have calculated for each pathway category the average number of tissue types in which a pathway is found to have changed its cohesiveness, either up or down. A pathway category is considered “generically unstable” if its pathways have changed their cohesiveness in a significantly (P < 0.05, t test) above average number of tissue types, and specifically unstable if their pathways only change their cohesiveness in certain specific tissue types, as shown in Table 4 . For the five KEGG pathway categories, “Environmental Information Processing” is found to be the most generically unstable. This result is consistent with our earlier findings that this pathway category, which consists of mostly signaling pathways, is also significantly enriched in down pathways (i.e., they tend to lose their cohesiveness in cancer). On the other hand, the pathways categorized as metabolic are mostly specifically unstable when compared with the other KEGG pathway categories. The stability of other KEGG pathway categories is not significantly different from average. “Cell Cycle Regulation,” a BioCarta pathway category, is the most specifically unstable of all, with half of its pathways exhibiting no cohesiveness change and each pathway showing change in only one tissue type on average. On the other hand, “Adhesion” seems to be the most generically unstable BioCarta pathway category and “Cell Signaling” is second. These results are consistent with what we find with the KEGG pathways. Overall, the KEGG pathways seem to be more generically unstable than the BioCarta pathways (P = 1.68 × 10⁻⁹, t test). The three GO categories, “biological process,” “molecular function,” and “cellular component,” did not show any statistically significant difference in their generic instability.

Stability of “Cancer Pathways”

Cancer is a genetic disease and genes operate through pathways. Variations in the gene regulation patterns in a pathway, as reflected by the change of cohesiveness in the pathway, may be indicative of pathway instability, which itself may be a result (or cause) of genetic abnormalities. Approximately 45% of the 962 analyzed pathways contain at least one of the 346 known “cancer genes” according to a recent census of human cancer genes (1). As described earlier in the text, a change in cohesiveness has been found in 25% of all the pathways analyzed, including 16% down pathways and 9% up pathways. The question of interest then is whether the pathways that have shown a cohesiveness change in tumor compared with normal tissues can be considered “cancer pathways” based on the likelihood of them containing cancer genes, or whether cancer pathways in general are more likely to change their cohesiveness. The answer to the first questions is “yes,” as cancer genes are found in 57% of the pathways that have shown a change in their cohesiveness, which is significantly higher than the average probability (45%) of a pathway to contain cancer genes (P = 1.23 × 10⁻³, Fisher's exact test). Moreover, a down pathway is found to be much more likely to contain cancer genes than an average pathway (P = 4.20 × 10⁻⁵, Fisher's exact test). The partitioning of pathways in terms of cancer and noncancer is illustrated in Fig. 1A . Conversely, we have also found that cancer pathways are more likely to change their cohesiveness; i.e., a significantly higher percentage of cancer pathways show a cohesiveness change than the pathways that do not contain cancer genes (32% versus 20%; P = 2.92 × 10⁻⁵, Fisher's exact test). Moreover, the cancer pathways are especially enriched in down pathways (22% versus 11%; P = 1.60 × 10⁻⁶, Fisher's exact test), indicating that these pathways tend to become deregulated or dysfunctional in cancer. The partitioning of cancer and noncancer pathways in terms of cohesiveness change is shown in Fig. 1B. The fact that cancer pathways tend to become less cohesive in tumor-bearing tissues may be exploited to find more cancer pathways or cancer genes; i.e., other pathways that are shown to lose their cohesiveness in cancer, but are not known to contain cancer genes, may be additionally interesting pathways to be considered as therapeutic targets (e.g., the KEGG pathways “Fatty acid metabolism,” “Basal transcription factors,” and “Glycolysis/Gluconeogenesis”).

• Download figure
• Open in new tab
• Download powerpoint

Figure 1.

A comparison between pathways that contain cancer genes (cancer pathways) and those that do not contain known cancer genes (noncancer pathways) in terms of the percentage of pathways that show a coherence change in tumor versus normal tissues. The total pathway count for each pathway category is shown in parentheses next to each pathway type label. A, the pathways labeled as down (cohesive in normal and not cohesive in tumor tissues) contain a significantly larger percentage of cancer pathways than the other types of pathways. B, most pathways show no change in their coherence level comparing tumor and normal tissues; however, cancer pathways contain significantly more down pathways than noncancer pathways.

The instability of cancer pathways is also reflected by the number of tissue types in which they have shown a coherence change. Cancer pathways containing genes with germ-line or somatic mutations (7) both exhibit a coherence change in significantly more tissue types than an average pathway [germ line, P = 1.86 × 10⁻⁷; somatic, P = 4.96 × 10⁻⁵ (t test)], indicating that cancer pathways are generically less stable than other pathways. Interestingly, regardless of the fact that individuals with a germ-line mutation carry that mutation in every cell of their body whereas somatic mutations are found only in an individual's cancer cells (2, 3), the tissue promiscuity/specificity of pathway instability is found to be not significantly different between pathways containing germ-line cancer genes and those containing somatic cancer genes (average unstable tissue counts are 3.64 for pathways containing germ-line and 3.34 for somatic cancer genes; P = 0.31, t test). This may explain the tissue specificity of gene defects; i.e., although the vast majority of inherited cancer genes seem to be expressed in most adult tissues, a germ-line mutation in these genes is manifested in only a limited spectrum of cancers. We have found that most pathways, including the ones containing cancer genes, only show a coherence change in a small fraction of tissue types (15–20%), indicating that probably not all mutations will be translated into pathway instability, and only in the tissue types where the stability of pathways is compromised by a mutation will cancer arise. Just like mutated proteins may render a pathway unstable or deregulated, it is equally probable that pathway instability itself could be the contributing factor to certain gene mutations. Unstable pathways are probably more susceptible to environmental influences, which could trigger or facilitate mutations. The final result would be a malignant cycle that promotes uncontrolled growth. The implication that pathway instability could be the cause of genetic instability makes pathways as a system of interactions, rather than individual genes, an interesting target in itself for therapy considerations.

Strategies to Target Change in Anticancer Therapies: Finding Agents That Can Potentially Perturb Unstable Pathways

Pathways that tend to change their cohesiveness in tumor versus normal tissues represent interesting targets for anticancer therapy, making desirable agents with the potential to specifically disrupt these pathways. We have previously developed a method to relate compound cytotoxic response to pathways and mechanisms of action through gene expression patterns (41). Briefly, we have organized the tumor cell 50% growth inhibition (GI₅₀) data for >40,000 screened compounds that are publicly available into Self-Organizing Maps (SOM; refs. 46, 47). SOM clustering of the GI₅₀ data segregates compounds into nine major response categories: mitosis (M); membrane function (N); nucleic acid metabolism (S); metabolic stress and cell survival (Q); kinases/phosphatases and oxidative stress (P); and four unexplored regions, R, F, J, and V (47–49). Each of these regions is further divided into a total of 80 clades, each of which is a group of clusters (nodes) that share similar cytotoxic responses (M₁–M₈, N₁–N₁₃, P₁–P₈, Q₁–Q₇, R₁–R₇, S₁–S₁₃, F₁–F₈, J₁–J₈, and V₁–V₈). Compounds in SOM clusters that are significantly correlated with a pathway can potentially perturb that pathway. Therefore, for each of the pathways with a change in gene expression coherence, we have found the SOM cluster that is most significantly correlated with the pathway following the previously described methods (41). Figure 2A shows the distribution of these pathways in terms of the number of clusters identified in each SOM clade.

• Download figure
• Open in new tab
• Download powerpoint

Figure 2.

SOM distribution of compounds that are the most significantly correlated with pathways of interest in terms of the number of clusters identified in each SOM clade. A, compound groups that can potentially target pathways with a change in gene expression coherence comparing tumor to normal tissues [gray histograms, up pathways (cohesive in tumor and not cohesive in normal tissues); black histograms, down pathways (cohesive in normal and not cohesive in tumor tissues)]. B, compound groups that can potentially target cancer pathways.

The compounds that are the most significantly correlated with these pathways are mostly clustered in SOM regions F, P, V, and N. This is not surprising because we have previously found that the P-region contains agents with kinase/phosphatase targeting as their putative mechanism of action (47, 50) and many of the pathways that tend to change their cohesiveness are signaling pathways. Some clades contain compounds that can potentially perturb both up and down pathways, such as clades F₂, P₁, and P₂, whereas other clades contain compounds that seem to perturb specifically up or down pathways. Most notably, compounds clustered in clades P₄ and M₃ are associated only with up pathways and those in clades N₉, P₆, F₆, F₈, and V₁ are associated predominantly with down pathways and with very few or no up pathways. The SOM distribution of compounds most significantly associated with pathways containing known cancer genes, and therefore are potential drug candidates for these pathways, is shown in Fig. 2B. The distribution pattern closely resembles those in Fig. 2A, especially those for the down pathways. Most clinically used compounds in anticancer therapies are located in regions M and S, essentially targeting the proliferation stage of a cancer cell. Our analysis, however, points to a much more diverse set of compounds with the potential to disrupt a wide spectrum of cellular processes characteristic of cancer cells, yet they are different from traditional cancer therapy.

Conclusions

In this study, we have developed strategies to analyze pathway stability by comparing pathway gene expression coherence or cohesiveness levels in tumor tissues to their normal counterparts. Changes occurring in cancer as reflected by variations in pathway coherence are considered indicative of pathway instability and possibly genetic instability. We have identified pathways that show a significant change in their coherence level in tumor tissues in general, as well as specific changes in certain tumor types. These pathways may represent good targets for developing novel anticancer therapies. Significantly more pathways are found to lose their coherence in tumor tissues. Signal transduction represents the most unstable pathway category, of which coherence in gene regulation is largely lost in tumor tissues. In contrast, pathways responsible for vital cellular processes are mostly able to maintain their gene expression coherence in tumor tissues and are among the most stable. The combination of homeostatic control over critical pathways to ensure survival and altered regulation of signaling to allow excessive proliferation forms the foundation for the selective growth advantage of cancer cells over normal cells. The instability of metabolic and cell cycle regulating pathways seems to be the most tissue specific. The function of these pathways and their unstable tissue type may provide important clues to finding the molecular mechanisms underlying specific cancer types. We have examined the particular pathways that contain known cancer genes and compared their behavior with other pathways. Cancer pathways are found to be more likely to lose their coherence and thus show a greater level of instability than an average pathway. Finally, we have proposed strategies to target these changes; i.e., to find new agents that can specifically target the unstable pathways that might be relevant in cancer.