Logo of jamiaJAMIA - The Journal of the American Medical Informatics AssociationVisit this articleSubmit a manuscriptReceive email alertsContact usBMJ
J Am Med Inform Assoc. 2012 Mar-Apr; 19(2): 241–248.
Published online Dec 20, 2011. doi:  10.1136/amiajnl-2011-000658
PMCID: PMC3277635

Identifying disease genes and module biomarkers by differential interactions

Abstract

Objective

A complex disease is generally caused by the mutation of multiple genes or by the dysfunction of multiple biological processes. Systematic identification of causal disease genes and module biomarkers can provide insights into the mechanisms underlying complex diseases, and help develop efficient therapies or effective drugs.

Materials and Methods

In this paper, we present a novel approach to predict disease genes and identify dysfunctional networks or modules, based on the analysis of differential interactions between disease and control samples, in contrast to the analysis of differential gene or protein expressions widely adopted in existing methods.

Results and Discussion

As an example, we applied our method to the study of three-stage microarray data for gastric cancer. We identified network modules or module biomarkers that include a set of genes related to gastric cancer, implying the predictive power of our method. The results on holdout validation data sets show that our identified module can serve as an effective module biomarker for accurately detecting or diagnosing gastric cancer, thereby validating the efficiency of our method.

Conclusion

We proposed a new approach to detect module biomarkers for diseases, and the results on gastric cancer demonstrated that the differential interactions are useful to detect dysfunctional modules in the molecular interaction network, which in turn can be used as robust module biomarkers.

Keywords: System biology, bioinformatics

Introduction

Complex diseases are well recognized as generally being caused by the dysregulation of biological systems or molecular networks rather than by the mutations of individual genes. Moreover, it is notoriously difficult and time consuming to identify disease related genes with biological experiments.1 2 On the other hand, many computational methodologies have been presented to predict disease related genes,3 4 most of which focus on the differential expression of genes and static regulation between genes5 6 while ignoring the dynamic regulation or network rewiring between molecules during different disease stages. However, analysis of dynamic regulation between molecules can uncover crucial information that cannot be detected in static conditions.7 Generally, regulations or interactions between molecules vary at different times and in different tissues, which changes are causally related to disease progression. As a result, those molecules found to interact with different partners in the molecular interaction network could be correlated with the relevant disease. Therefore, new potential disease genes can be detected by studying dynamic regulation and network rewiring between molecules, a method ignored by investigations examining differential expressions only7 as proposed in previous reports.

In this paper, a novel approach is presented to identify disease genes and module-based biomarkers in complex diseases by investigating dynamic interactions and network rewiring between molecules related to pathogenesis. We study differential interactions rather than differential genes (or proteins) based on expression data, which is a key difference from existing methods. In particular, we assume that the dysfunction or network rewiring of those interactions (eg, protein–protein interactions, PPIs) that differ between control and disease samples is possibly related to the relevant disease. Using gene or protein expression to quantitatively describe the dynamic interactions between proteins, we identify those interactions (ie, gene pairs rather than individual genes or proteins) that occur in disease but not in control samples and vice versa, which interactions are denoted as differential interactions (eg, differential PPIs for protein interactions) hereafter and are presumed to be highly related to the disease of interest. Taking the analysis of high throughput data on gastric cancer as an example, our method identifies differential interactions involving 34 potential gastric cancer genes from analysis of microarray data on 160 clinic samples (80 disease and 80 control samples). It is found that many of these potential cancer genes have already been reported, for example in the Cancer Gene Census (http://www.sanger.ac.uk/genetics/CGP/Census/, accessed on March 22, 2011)1 and the Genetic Association Database (GAD) (http://geneticassociationdb.nih.gov/, update of August 1, 2011).8 Moreover, results with these 34 genes on two independent validation datasets demonstrate that the identified gastric cancer genes can be used as module biomarkers to diagnose gastric cancers. In addition, analysis of the relationship between subject characteristics and the identified cancer related genes shows that alcohol and smoking increase the risk of gastric cancer, which is consistent with results reported in the literature9–12 and clinic data. These results demonstrate the efficiency and efficiency of our method for identifying module biomarkers, and also provide more evidence on the predicted disease genes.

Materials and methods

Gene expression profiles and protein interactions

The gene expression profiles for gastric cancer were obtained from the GEO database (ID: GSE27342).13 The dataset contains 160 samples from cancer tissues and the adjacent non-cancerous tissues of 80 non-treated gastric cancer patients. All profiles were normalized by the RMA (robust multi-array averaging) method and the probe sets were mapped to their corresponding gene symbols. The expression values of replicated probe sets corresponding to one gene were averaged. Finally, 17 170 genes were identified for the following study. Since there were fewer samples in stage I and stage II than other disease stages, they were combined into one phase (shown in table 1). The human PPIN (protein–protein interactions network) was downloaded from the HPRD (http://www.hprd.org) database and self-interactions and repeat interactions removed, resulting in 37 039 interactions among 9465 proteins.

Table 1
Gastric cancer and control sample characteristics

Identification of differential interactions and candidate disease genes

At each phase, the gene expression profiles were divided into control and disease groups. In each group, the correlation coefficient between a pair of interacting proteins in the human PPIN was calculated based on the samples in the group. Subsequently, only those PPIs with high correlation coefficients (≥0.75) were reserved with the assumption that the PPIs with low correlation coefficients do not occur in corresponding samples. As a result, the rewired new PPIN, which were called PPI in control and PPI in disease, were constructed as shown in figure 1. The PPIs which differed between control and disease samples were then identified and illustrate the dynamic changes in interactions between control and disease status. The PPIs which differed in each phase were identified by combining specific PPIN in control and specific PPIN in disease while removing their common edges, where the common edges are those edges that are present in both control and disease protein-protein interaction networks. as shown in figure 1. Each edge in the differential PPIs represents the dynamic interaction between control and disease samples. The disease related genes are those which are common to specific PPIN in control and specific PPIN in disease. It is very possible that these genes play important roles in progression from control to disease. The common members of the disease related genes in each phase were regarded as potential disease genes and called candidate disease genes (shown in figure 1). Clearly, gene regulations or genetic interactions can also be identified in the same way as the PPIs if a reference network is available.

Figure 1
Identification of disease related genes or module biomarkers based on differential protein–protein interactions (PPIs). The threshold for high correlation was set to 0.75, and ‘Remove common edges’ means removing interactions that ...

Identified disease genes are an effective module biomarker

To determine whether our identified disease genes are related to disease, we used them as a module biomarker to discriminate between control and disease samples. Support vector machine (SVM) regression was applied to the expression values of the predicted disease genes from the module biomarker to distinguish disease from controls. Furthermore, the receiver operating characteristic (ROC) curve and the area under the ROC curve (AUC) were used to evaluate classification efficiency. In addition, two independent human gastric tumor datasets were also used to evaluate the classification performance of the module biomarker. Two R14 packages, kernlab and ROCR, were used to build the SVM classifier and produce the ROC curves with defaulted parameters.

Validation of candidate disease genes

Documented cancer related genes causally implicated in cancer were obtained from the Cancer Gene Census database1 and used to validate the candidate disease genes identified by our approach. The known cancer genes common to the candidate disease genes were used to evaluate the prediction results. A hypergeometric test was used to estimate the enrichment of these candidate disease genes compared to known cancer related genes. The formula of the hypergeometric test is:

p(Xx)=1k=0x1(Mk)(NMnk)(Nn),

where N is the total gene number of the gene expression profiles, M is the number of known cancer genes, n is the number of the candidate disease genes that we identified, and k is the number of common entries between them. P is the enrichment statistical significance of the test. The same enrichment score was then calculated for the candidate disease genes in the cancer pathways from the KEGG database and gastric cancer association genes from GAD.

Random sampling was also used to test probability, where the same number of known cancer genes was randomly selected to estimate whether these known cancer genes included in the previous results were statistically significant. First, the same number of genes as the candidate disease genes was randomly selected from the entire expression profiling gene set. Second, the number of known cancer genes included in the random samples was counted. Third, random sampling was repeated 106 times. Fourth, the p value of the candidate disease genes was defined as the probability that one random sampling might contain a greater or equal number of known cancer genes than in our study samples. The same random sampling method was also used to estimate the statistical significance of the candidate disease genes among the genes in the KEGG cancer pathways and the gastric cancer association genes from GAD.

Clinical outcome for candidate disease genes

The association between cancer genes and the development of gastric cancer is influenced by environmental and lifestyle factors. The significance of the association was calculated by Fisher's exact test to determine if there were non-random associations between two categorical variables based on these candidate disease genes. First, the expression profiles were grouped according to alcohol consumption, smoking status, age, and gender into different subsets. In each subset, the genes with differential expression between controls and cancer were identified using the Wilcoxon test (with p<0.05 considered significant) and the number of differential candidate disease genes counted individually. Since the parametric statistical test requires a normal distribution hypothesis, we utilized the Mann–Whitney Wilcoxon test to measure differential expression between control and cancer subgroups. The statistical significance was then calculated by Fisher's exact test based on the number of differential candidate disease genes according to age, gender, alcohol consumption, and smoking status, respectively.

Results

Identification of differential interactions and gastric cancer genes

To test our approach, we took a gastric cancer study as an example. Table 1 provides information about gastric cancers and their corresponding control samples. All samples were allocated to one of three phases (table 1). The control and cancer specific PPIN were assigned to each phase based on gene expression data and human PPIs, where the common edges between PPIN in control and PPIN in cancer were removed. The differential PPIs were obtained by merging the control specific PPIN and the cancer specific PPIN in the corresponding phase, which include the edges that occur only in the specific control PPIN or only in the specific cancer PPIN. Only those genes that have differential PPIs but are expressed in both control and cancer samples were regarded as potential gastric cancer genes in each phase. Table 2 shows the intermediate results obtained for different phases. Note that there are fewer edges and genes in the middle phase than in the other two phases since the threshold for the middle phase is stricter as it contains fewer samples. Furthermore, the ratio of edge number in control specific PPIN to that in cancer specific PPIN decreases from the early to late phase. This phenomenon may be explained as follows. First, the control samples were obtained from tissue adjacent to the cancer tissue, and cancer tissue can affect adjacent tissue more strongly in the late than early phase. Second, cancer cells may perhaps develop new mechanisms to stabilize cell proliferation and enhance tumor metastasis with worsening disease, resulting in more cancer specific regulations in the late than early phase.

Table 2
Interactions and genes in each phase

Finally, the predicted gastric cancer genes found in all three phases were treated as our final predicted gastric cancer genes or module biomarker. In this work, 34 genes were identified as gastric cancer genes. The interactions of the 34 gastric cancer genes in human PPIN are shown as figure 2. We can see from the figure that 24 out of 34 potential genes are connected to a module component or a subnetwork in the interaction network, where the green and red vertices are separately the genes whose expressions are gradually increasing or decreasing in different phases of gastric cancer. For example, gene SFRS12 shows a gradually increase in expression, while gene RAF1 shows a gradually decrease in expression from the early to late phase. Although some genes show increasing or decreasing levels of gene expression, most of the candidate disease genes do not show statistically significant changes in gene expression between control and cancer status. Therefore, they will not be detected in a conventional general analysis of differentially expressed genes.

Figure 2
The interaction network with the predicted gastric cancer genes. The interactions among the 34 genes were obtained from HPRD, where 24 genes form a giant component and the nodes are marked in white if their corresponding gene expressions are not always ...

Furthermore, functional enrichment analysis of our predicted disease genes for gastric cancer was conducted using DAVID15 (see online supplementary table 1). From gene functions, we can see that some of the disease genes have been reported to play an important role in tumorigenesis. For example, SMAD2, a tumor suppressor,16–18 mediates the signal of the transforming growth factor TGF-β whose growth inhibitory effects are resisted in human cancer,19 and regulates multiple cellular processes, such as cell proliferation, apoptosis, and differentiation. STAT3, a tumor suppressor,20 is a signal transducer and activator of transcription, and can promote oncogenesis by being constitutively active through various pathways. Functional enrichment analysis shows that the identified gastric cancer genes are mainly enriched through an enzyme linked receptor protein signaling pathway, phosphorylation, an intracellular signaling cascade, and protein amino acid phosphorylation, which indicates that gastric cancer may be related to dysregulation of signaling pathways.

In addition, a PPIN containing candidate disease genes and their interaction partners in each phase was constructed (figure 3). NOA (network ontology analysis),21 which annotates biological networks, was used to analyze the enriched functions of the PPIN in each phase. The results of NOA are shown in table 3. In the control samples, the functions are enriched in the signaling pathway, while in the early phase of disease samples, the functions are enriched in nucleic acid related metabolic processes. The functions of the PPIN change to DNA replication and DNA-dependent transcription in the middle and late phases, respectively. The functional annotations of PPIN in each phase demonstrate clearly that the identified cancer genes function reasonably well and their changes reflect dynamic transition between different stages from nucleic acid related metabolic processes to DNA replication and transcription that is an important marker of cancer cell growth.

Figure 3
The protein–protein interaction networks (PPIN) with the predicted cancer genes and their corresponding interaction partners in different cancer phases. All nodes are marked in white except the 34 candidate cancer genes. (A) Specific PPIN in control ...
Table 3
Functional enrichment analysis of PPIN including the identified gastric cancer genes

Confirmation of identified gastric cancer genes

To confirm the predicted gastric cancer genes, two public databases (the Cancer Gene Census database1 and GAD8) were explored to see whether these genes have been previously reported. In the Cancer Gene Census database1 (accessed on March 22, 2011), 457 genes are reported to be related to cancer. There are nine common genes between the identified 34 cancer genes and the 457 cancer genes (figure 4A and table 4). To investigate whether these nine genes could have been obtained randomly, statistical significance was checked using hypergeometric distribution and 106 times random simulation. Two significant p values of 1.799×10−7 from hypergeometric distribution and <1.0×10−6 from random simulation were obtained, which indicate that the identified candidate cancer genes are enriched among known cancer related genes and cannot be obtained randomly. The GAD is an archive of human genetic association studies of complex diseases and disorders.8 The gastric cancer association genes were selected from the latest version (updated on August 1, 2011) of this database. The database lists 146 gastric cancer genes, five of which are also found among the predicted cancer genes (figure 4B and table 4). Similarly, two significant p values of 9.467×10−6 and 8.0×10−6 were obtained with a hypergeometric test and random simulation, respectively. In addition, some pathways were annotated as being related to cancer in the KEGG database.22 We also identified 328 genes involved in cancer pathways, nine of which are also included in our 34 cancer genes (table 4). Significant p values of 1.046×10−8 and <1.0×10−6 were obtained from the hypergeometric test and 106 times random simulation study, respectively. The results show that our predicted gastric cancer related genes are highly related to gastric cancer.

Figure 4
Predicted genes and known cancer related genes. (A) 457 known cancer genes were obtained from the Cancer Gene Census database, nine of which appear in our predictions. (B) 146 known gastric cancer associated genes were obtained from the Genetic Association ...
Table 4
Common genes between the identified cancer genes and documented cancer related genes

In order to further confirm the predicted gastric cancer genes, we used them to distinguish cancer samples from controls. Two independent gene expression datasets (GSE2701 and GSE13911) for gastric cancer were obtained from D'Errico et al23 and Chen et al.24 If our predicted cancer genes can successfully distinguish cancer samples from control samples from the two independent datasets, they can be further shown to be related to gastric cancer. Moreover, we compared our results with those obtained with a published biomarker set for gastric cancer,25 which were derived from differential gene expression. Fivefold cross-validation was used to evaluate performance based on different biomarkers, while SVM regression was used as a classifier. Figure 5A shows the ROC curves obtained with our predicted candidate genes as a module biomarker and known biomarkers individually. It shows that our identified module-based biomarker performs well with respect to the known biomarkers, which proves that the module biomarker is related to gastric cancer.

Figure 5
ROC curves obtained with our module biomarker and published biomarkers. (A) The gene expression dataset used to identify differential protein–protein interactions and two independent datasets ...

Personal characteristics, such as alcohol consumption, smoking status, age, and gender, are important factors affecting tumorigenesis. The relationships between these characteristics and gastric cancer were therefore investigated and significance was evaluated with Fisher's exact test (see online supplementary table 2) using our module biomarker (34 genes). Specifically, the expression data were first separated into differential expression subsets according to alcohol consumption or not, smoking or not, age, and gender. The differentially expressed genes were identified by the Wilcoxon test. Furthermore, the background expression profiles of all samples were also investigated. It was found that alcohol consumption is significantly related to gastric cancer (p=3.51×10−5). Also, only four candidate genes are differentially expressed between control and cancer samples (see online supplementary table 2) as 17 genes have already been dysregulated in those consuming alcohol. Comparison of the alcohol and non-alcohol samples showed that alcohol was significantly related to gastric cancer (p=1.24×10−5). Similarly, smoking was also significantly related to gastric cancer (see online supplementary table 2). Therefore, it is possible that alcohol consumption and smoking are risk factors for gastric cancer. In contrast, other factors, such as gender and age, were not found to be significantly related to gastric cancer. These results that alcohol and smoking are related to gastric cancer are consistent with those reported in literature.9–12

Discussion

The identification of genes related to cancer can help the diagnosis and efficient treatment of cancer and also shed light on other complex diseases. In this work, we have presented a novel network-based approach to predict cancer genes and module biomarkers through differential interactions. In particular, we applied our method to the analysis of gastric cancer based on three-phase microarray data. Many of the predicted 34 gastric cancer genes are known cancer genes and have been reported in the literature, for example, SMAD2 and STAT3 are tumor suppressors, while ABL1 and PLK1 are known oncogenes.

Based on NOA analysis of functional enrichment, the functions of the PPIN including the predicted cancer genes were obviously different between the control and cancer samples. In control samples, the identified subnetwork contains genes that are enriched in signaling pathways, and gene ontology analysis actually gave similar result by employing DAVID. In the cancer samples, the identified subnetwork was regulated by disease genes clearly related to the nucleic acid and DNA regulatory processes. It is possible that those disease genes regulate and participate in signaling pathways in normal cells. However, when these disease genes regulate the process of DNA replication, the factors regulating DNA replication or the nucleic acid metabolic process in normal cells are unable to inhibit the process because they do not identify the disease genes as targets. It is also possible that the identified disease genes mainly promote nucleic acid synthesis and DNA replication in cancer cells. Regulation of transcription in the late phase may be associated with infection and the ability of cancer cells to metastasize. Although we ignored the phase-specific genes and only chose those which overlapped in the three phases, some genes showing phase-specific potential still appeared among the candidate disease genes. Twelve genes (SFRS12, TGFBR2, ABL1, USO1, RAF1, TP53BP2, SMURF2, UBTF, BRCA1, STAT3, SFRS3, and YWHAB) among the predicted disease genes exhibited gene expression which continuously increased or decreased from the early to late phase, as shown in figure 2. These are potential phase-specific genes for gastric cancer.

In particular, USO1 shows high differential expression in the early phase, a sharp decrease in gene expression from the early to middle phase, and then almost normal expression in the late phase (data not shown). It is possible that USO1 may be a trigger or driver gene of gastric cancer which only is upregulated in the early phase, with its expression levels returning to near normal in the late phase. ABL1 and SFRS3 show similar trends to USO1, and also exhibit a sharp decrease in gene expression after the early phase.

On the other hand, some studies have reported that alcohol and tobacco consumption are associated with gastric cancer risk.9–12 The clinical outcome test for the 34 candidate disease genes was based on Fisher's exact test for the number of differential expression candidate disease genes, and we obtained the same result that alcohol and tobacco consumption are risk factors for gastric cancer. This approach also proves that there is an association between the identified disease genes and gastric cancer. Further observation of the results indicated that some of the disease genes, such as KHDRBS1 and SNRPB (see online supplementary table 2), are both differentially expressed between control and cancer status with alcohol consumption and are also differentially expressed in other conditions. It is possible that the two genes are affected by alcohol or respond to the stimulation of alcohol in the stomach. The genes MMP2 and STAT3 (see online supplementary table 2), which are differentially expressed genes except alcohol and smoking habits, can be affected by alcohol or tobacco consumption. In other words, the expression level of the two genes on alcohol or smoking habits in normal cell is similar to cancer cell, and thus there is a high risk to the human on alcohol or smoking habits. The four genes CDC6, PLK1, BRCA1, and MCM3 which were differentially expressed in alcohol consumption, appeared to be differentially expressed in all conditions. They also have higher differential expression than other candidate disease genes (figure 2 and online supplementary table 2). Hence, it is possible that they play important roles in tumorigenesis and cancerization.

Conclusion

In this study, we proposed a differential interaction-based approach for studying complex diseases, in contrast to the conventional methods of examining differential gene or protein expression. In particular, we applied our method to the analysis of gastric cancer progression. Our method utilized the dynamic variety of human PPIN in different phases to predict gastric cancer genes and module biomarkers, resulting in the identification of 34 gastric cancer genes. Gastric cancer information from two public databases was used independently to validate the predicted gastric cancer genes and produced very significant results. These genes can be used as a module biomarker to identify cancer phenotypes and also to evaluate the correlation between lifestyle and carcinogenesis. The functional enrichment of these genes to cancer pathways identified in the KEGG database shows that the predicted gastric cancer genes take part in cancer processes. The identified subnetwork or module shown to be directly connected by the gastric cancer-related genes to all phases was analyzed by NOA, and a functional transition from normal phenotypes to cancer phases was demonstrated. This transition may be the driving factor for gastric cancer progression. Although we mainly used differential PPI interactions in this work, gene–gene interactions or other forms of interactions could also be used for similar analysis of disease. As a future topic, we will further study the dynamic features during disease progression with consideration of both phase-common and phase-specific differential interactions.26

Supplementary Material

Supplementary Data:

Footnotes

Funding: This work was partly supported by the National Natural Science Foundation of China (NSFC) under Grant Nos. 91029301, 61134013, 61103075, 61072149, 91130032, and 31100949; by the Innovation Program of Shanghai Municipal Education Commission (10YZ01); by the Shanghai Rising-Star Program (10QA1402700); by the Chief Scientist Program of Shanghai Institutes for Biological Sciences (SIBS), Chinese Academy of Sciences (CAS) under Grant No. 2009CSP002; by the Knowledge Innovation Program of SIBS, CAS under Grant No. 2011KIP203; by the SA-SIBS Scholarship Program; by the Shanghai Natural Science Foundation under Grant No. 11ZR1443100; by the Shanghai Pujiang Program; by the National Center for Mathematics and Interdisciplinary Sciences, CAS; and by the Aihara Project of the FIRST program initiated by CSTP.

Competing interests: None.

Contributed by

Contributors: XL wrote the paper, performed the experiments and analyzed the data. ZPL analyzed the data and contributed materials and analysis tools. XMZ conceived and designed the experiments. LC conceived and designed the experiments and contributed materials and analysis tools.

Provenance and peer review: Not commissioned; externally peer reviewed.

References

1. Futreal PA, Coin L, Marshall M, et al. A census of human cancer genes. Nat Rev Cancer 2004;4:177–83 [PMC free article] [PubMed]
2. Polager S, Ginsberg D. p53 and E2f: partners in life and death. Nat Rev Cancer 2009;9:738–48 [PubMed]
3. Chuang HY, Lee E, Liu YT, et al. Network-based classification of breast cancer metastasis. Mol Syst Biol 2007;3:140. [PMC free article] [PubMed]
4. Liu ZP, Wang Y, Zhang XS, et al. Detecting and analyzing differentially activated pathways in brain regions of Alzheimer's disease patients. Mol Biosyst 2011;7:1441–52 [PubMed]
5. Liu X, Tang WH, Zhao XM, et al. A network approach to predict pathogenic genes for Fusarium graminearum. PLoS One 2010;5:e13021. [PMC free article] [PubMed]
6. Chen L, Wang RS, Zhang XS. Biomolecular Networks. Hoboken, New Jersey: John Wiley & Sons, Inc, 2009
7. Bandyopadhyay S, Mehta M, Kuo D, et al. Rewiring of genetic networks in response to DNA damage. Science 2010;330:1385–9 [PMC free article] [PubMed]
8. Becker KG, Barnes KC, Bright TJ, et al. The genetic association database. Nat Genet 2004;36:431–2 [PubMed]
9. Tredaniel J, Boffetta P, Buiatti E, et al. Tobacco smoking and gastric cancer: review and meta-analysis. Int J Cancer 1997;72:565–73 [PubMed]
10. Bertuccio P, Chatenoud L, Levi F, et al. Recent patterns in gastric cancer: a global overview. Int J Cancer 2009;125:666–73 [PubMed]
11. Koizumi Y, Tsubono Y, Nakaya N, et al. Cigarette smoking and the risk of gastric cancer: a pooled analysis of two prospective studies in Japan. Int J Cancer 2004;112:1049–55 [PubMed]
12. Sung NY, Choi KS, Park EC, et al. Smoking, alcohol and gastric cancer risk in Korean men: the National Health Insurance Corporation study. Br J Cancer 2007;97:700–4 [PMC free article] [PubMed]
14. R Development Core Team (2011) R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria: ISBN 3-900051-07-0, URL http://www.R-project.org/
15. Huang da W, Sherman BT, Lempicki RA. Systematic and integrative analysis of large gene lists using DAVID bioinformatics resources. Nat Protoc 2009;4:44–57 [PubMed]
16. Kang Y, He W, Tulley S, et al. Breast cancer bone metastasis mediated by the Smad tumor suppressor pathway. Proc Natl Acad Sci U S A 2005;102:13909–14 [PMC free article] [PubMed]
17. Prunier C, Ferrand N, Frottier B, et al. Mechanism for mutational inactivation of the tumor suppressor Smad2. Mol Cell Biol 2001;21:3302–13 [PMC free article] [PubMed]
18. Yang J, Wahdan-Alaswad R, Danielpour D. Critical role of Smad2 in tumor suppression and transforming growth factor-beta-induced apoptosis of prostate epithelial cells. Cancer Res 2009;69:2185–90 [PMC free article] [PubMed]
19. Riggins GJ, Thiagalingam S, Rozenblum E, et al. Mad-related genes in the human. Nat Genet 1996;13:347–9 [PubMed]
20. de la Iglesia N, Konopka G, Puram SV, et al. Identification of a PTEN-regulated STAT3 brain tumor suppressor pathway. Genes Dev 2008;22:449–62 [PMC free article] [PubMed]
21. Wang J, Huang Q, Liu ZP, et al. NOA: a novel network ontology analysis method. Nucleic Acids Res 2011;39:e87. [PMC free article] [PubMed]
22. Kanehisa M, Goto S, Kawashima S, et al. The KEGG resource for deciphering the genome. Nucleic Acids Res 2004;32:D277–80 [PMC free article] [PubMed]
23. D'Errico M, de Rinaldis E, Blasi MF, et al. Genome-wide expression profile of sporadic gastric cancers with microsatellite instability. Eur J Cancer 2009;45:461–9 [PubMed]
24. Chen X, Leung SY, Yuen ST, et al. Variation in gene expression patterns in human gastric cancers. Mol Biol Cell 2003;14:3208–15 [PMC free article] [PubMed]
25. Cui J, Chen Y, Chou WC, et al. An integrated transcriptomic and computational analysis for biomarker identification in gastric cancer. Nucleic Acids Res 2011;39:1197–207 [PMC free article] [PubMed]
26. Chen L, Wang R, Li C, et al. Modelling Biomolecular Networks in cells: Structures and Dynamics. London: Springer-Verlag, 2010

Articles from Journal of the American Medical Informatics Association : JAMIA are provided here courtesy of American Medical Informatics Association
PubReader format: click here to try

Formats:

Related citations in PubMed

See reviews...See all...

Cited by other articles in PMC

See all...

Links

  • MedGen
    MedGen
    Related information in MedGen
  • PubMed
    PubMed
    PubMed citations for these articles

Recent Activity

Your browsing activity is empty.

Activity recording is turned off.

Turn recording back on

See more...