Mining tissue specificity, gene connectivity and disease association to reveal a set of genes that modify the action of disease causing genes
© Reverter et al; licensee BioMed Central Ltd. 2008
Received: 07 January 2008
Accepted: 19 September 2008
Published: 19 September 2008
The tissue specificity of gene expression has been linked to a number of significant outcomes including level of expression, and differential rates of polymorphism, evolution and disease association. Recent studies have also shown the importance of exploring differential gene connectivity and sequence conservation in the identification of disease-associated genes. However, no study relates gene interactions with tissue specificity and disease association.
We adopted an a priori approach making as few assumptions as possible to analyse the interplay among gene-gene interactions with tissue specificity and its subsequent likelihood of association with disease. We mined three large datasets comprising expression data drawn from massively parallel signature sequencing across 32 tissues, describing a set of 55,606 true positive interactions for 7,197 genes, and microarray expression results generated during the profiling of systemic inflammation, from which 126,543 interactions among 7,090 genes were reported.
Amongst the myriad of complex relationships identified between expression, disease, connectivity and tissue specificity, some interesting patterns emerged. These include elevated rates of expression and network connectivity in housekeeping and disease-associated tissue-specific genes. We found that disease-associated genes are more likely to show tissue specific expression and most frequently interact with other disease genes. Using the thresholds defined in these observations, we develop a guilt-by-association algorithm and discover a group of 112 non-disease annotated genes that predominantly interact with disease-associated genes, impacting on disease outcomes.
We conclude that parameters such as tissue specificity and network connectivity can be used in combination to identify a group of genes, not previously confirmed as disease causing, that are involved in interactions with disease causing genes. Our guilt-by-association algorithm should be useful for the discovery of additional modifiers of genetic diseases, and more generally, for the ability to associate genes of unknown function to clusters of genes with defined functions allowing for novel biological inference that can be subsequently validated.
The understanding of the biology underlying phenotype is still a limiting factor in delivering the promise of high throughput genomics. However, as new datasets are available, new data mining methods are developed and the goal appears ever more achievable.
Among the high-throughput technologies, gene expression profiling has led to the identification of genes that perform in a coordinated manner allowing researchers to reasonably predict the role of genes for which no biological function was attributed, based on the known performance of other group members. These predictions rely on the guilt-by-association heuristic, widely invoked in genomics and with proven applicability .
At the same time, a comprehensive atlas of transcribed genes in humans has revealed that genes may be split into two broad categories based on the number of tissues they are expressed in . Genes that are expressed in many tissues are designated as housekeeping (HK) while those that are expressed in few tissues are termed tissue-specific (TS).
Tissue specificity has subsequently been linked to a number of significant outcomes including level of expression , ability to detect cis-acting and trans-acting expression- quantitative trait loci , and differential rates of polymorphism , evolution  and disease-association . In addition, we  and others [9, 10] have demonstrated the importance of exploring differential gene connectivity in the identification of disease-associated genes using microarray gene expression data. More recently, the combination of text mining with gene interaction network analysis has been proposed to infer unknown gene-disease associations .
Furthermore, genes with a high degree of connectivity (network hubs) have been shown to be conserved across species  and their knockout phenotype more likely to be lethal . Finally, based on sequence conservation across species, a computational algorithm has been developed to identify genes associated with disease . However, no study relates gene interactions with tissue specificity and its subsequent likelihood of association with disease.
To address this situation, we mined three large independent datasets and classified transcribed human genes based on transcript abundance, tissue specificity, gene connectivity and disease association. We discuss how these factors relate to each other and, based on this new knowledge, implement a simple yet powerful guilt-by-association algorithm that allows us to identify several candidate genes that, while not previously associated with disease, may impact the development of diseases, including cancers, and hypothesize that many other members of this list will ultimately be confirmed as modifiers of various genetic diseases.
Data resources, edits and nomenclature
We merged three large datasets as follows: Firstly, we accessed expression data drawn from massively parallel signature sequencing (MPSS) covering 182,719 tag signatures across 32 tissues . Tissues represented on the MPSS data included nine different central nervous system (CNS) areas (amygdale, caudate nucleus, cerebellum, corpus callosum, fetal brain, hypothalamus, thalamus, spinal cord, and pituitary gland) and 23 non-CNS organs (adrenal gland, bladder, bone marrow, heart, kidney, lung, mammary gland, pancreas, placenta, prostate, retina, salivary gland, small intestine, spleen, stomach, testis, thymus, thyroid, trachea, uterus, colon, monocytes and peripheral blood lymphocytes). A total of 18,677 unique genes were represented on the MPSS data and the number of expressed genes per tissue averaged 8,943 and ranged from 5,845 in pancreas to 12,267 in testis.
Secondly, we downloaded a set of 55,606 true positive interactions among 7,197 genes that were defined from functional studies . This interactions dataset was built including 2,788 confirmed, direct, physical protein-protein interactions derived from the Biomolecular Interaction Network Database (BIND; http://binddb.org) , 18,176 confirmed human protein interactions from the Human Protein Reference Database (HPRD; http://www.hprd.org/) , 22,012 direct functional interactions from the Kyoto Encyclopedia of Genes and Genomes (KEGG; http://www.genome.jp/kegg) , and 16,295 interactions derived from Reactome http://www.reactome.org.
Finally, we used the microarray expression results generated during the profiling of systemic inflammation across 44,924 probe sets  and from which 126,543 interactions among 7,090 genes were reported . The microarray experiment used 92 Affymetrix GeneChips (Affymetrix, Santa Clara, CA) to examine gene expression profiles in whole blood leukocytes immediately before and at 2, 4, 6, 9 and 24 h after intravenous administration of bacterial lipopolysaccharide (LPS) endotoxin to four healthy human subjects. For the control (placebo) data, four additional subjects were studied under identical conditions but without LPS administration.
For the present study, and to enable the merging of the three datasets, a number of edits were performed as follows: For the MPSS data, tags not expressed at more than 5 transcripts per million (tpm), in at least one tissue, were disregarded. The threshold of 5 tpm corresponds to the sensitivity of MPSS technology as claimed by the manufacturers and independently assessed in our laboratory . Also, when the same gene was represented by more than one MPSS tag, the reading from the most abundant tag, summed across all tissues, was assigned to that gene. Finally, for the true positive interactions and the inflammation datasets, interactions involving genes not surveyed in the MPSS data were also discarded.
These criteria resulted in 15,050 genes [see Additional file 1] of which 5,198 and 4,950 were included in the true positive interactions and the inflammation datasets, respectively, and with 2,499 genes in common. In addition, a total of 6,151 (41%) of the genes were associated with disease according to OMIM database  as of September 19, 2007; and with 1,445 of them defined as disease-causing (i.e., associated with either known disease phenotype or polymorphic sequence known).
Hereafter, we refer to DIS to indicate the 6,151 genes from our resulting dataset that are disease-associated according to OMIM, and to NDIS to indicate the remaining 8,899 non-disease-associated genes also according to OMIM. Similarly, we refer to INT (and NINT) to indicate genes in our dataset for which interactions have (and have not) been reported.
Data mining approaches
In order to further characterize the relationship existing between tissue specificity, gene connectivity and disease association, the 15,050 genes were classified as either TS or HK. To ensure that these two categories together represented the majority of the genes, we searched for category limits from either extreme of the distribution of the number of genes expressed in one, two, and up to 32 tissues, until equivalent categories were defined, cumulatively representing > 50% of the total number of genes. In doing so, there were 4,232 (28%) TS genes expressed in 1 to 4 tissues, and 4,006 (27%) HK genes expressed in more than 25 tissues. The remaining 6,812 (45%) genes were classified as non-specific (NS).
Finally, and in order to identify novel candidate genes impacting disease, we developed a guilt-by-association algorithm. Selection thresholds based on the average number of known interactions combined with the average proportion of DIS genes among their interactors were determined from DIS genes. These thresholds were then applied to genes in the NDIS category. Genes exceeding both thresholds were identified as likely disease-associated candidates.
Results and discussion
Initial gene groupings and unknown biological processes
As expected, the discovery of interactions as well as disease-association for a given gene provides additional biological knowledge, allowing inferences as to its genomic functionality. Nevertheless, the biological process of about 10% of these presumably well-characterized genes remains to be elucidated. On the other extreme, and highlighting the extent to which further research is needed, as many as 85% of NDIS, NINT genes and across the three expression categories (TS, NS and HK) belong to an unknown biological process.
The impact of tissue-specificity
Relationship between the number of tissues in which a gene is expressed and a series of variables.
Non-Interacting (NI) genes only
Interacting genes only
Non-Interacting and Non-Disease
Non-Interacting and Disease
Interacting and Non-Disease
Interacting and Disease
Proportion of interacting genes:
Non-Disease genes only
Disease genes only
Proportion of disease genes:
Non-Interacting genes only
Interacting genes only
Tissue specificity of interactors:
Non-Disease genes only
Disease genes only
Proportion of disease genes among interactors:
Non-Disease genes only
Disease genes only
Gene interactions in the context of tissue-specificity and disease association
Our analyses revealed that interacting HK genes are more likely to interact with genes that are also HK (PCC = 0.89; P < 0.0001) and vice-versa (i.e., TS genes are more likely to interact among themselves). Importantly, this correlation remained strong when conditioning on disease status (Table 1). Also, interactions between two HK genes were 12.8 times more frequent (P < 0.0001) and 3.3 times more cohesive (P < 0.0001) as measured by the clustering coefficient, than interactions between two TS genes. The clustering coefficient is a measure of network cohesiveness and captures how many neighbours of a given gene are connected to each other.
Identification of candidate disease genes via guilt-by-association
Contingency table underlying the guilt-by-association algorithm
Number of Connections
% Disease-associated genes among interactors
To assess the optimality of our approach, we repeated the analyses using only the 1,445 DIS genes (out of the initial 6,151) with known disease phenotype and either sequence mutation or molecular basis known as those declared as truly disease-associated. The new thresholds for connectivity and proportion of DIS genes among interactors were 12 and 35%, respectively. The new list of candidate genes included 127 genes of which 107 were assessed as DIS in the initial list of 6,151. Assuming the remaining 20 genes are indeed false positives, this implies a precision of at least 84%.
It should be noted that precision alone is not enough to assess the goodness of a classifier, as it is only concerned with the ratio of identified genes that are positive, but not with the total number of discovered genes.
Precision analysis of the guilt-by-association algorithm
Threshold for number of connections (TC)
Threshold for % disease genes among interactors (TD)
TC = 1
TC = 4
TC = 13
TC = 12
TD = 12.8
TD = 28.6
TD = 50.0
TD = 35.0
However, the infeasibility of directly computing performance measures associated with a given algorithm in the absence of negative examples should be acknowledged. That is, although one can be relatively sure that certain genes are associated with a disease, it is not possible to ensure that a set of genes is not involved in any disease. In other words: Absence of evidence is not evidence of absence. On the other extreme, some of the genes annotated as disease associated by OMIM could also be false positives. In these situations, partially supervised learning algorithms have been proposed to address this issue and in the context of identifying disease genes .
Nevertheless, a literature survey revealed that 44 of the 112 candidate genes [see Additional file 2] have been previously associated with polymorphisms or differential gene expression leading to a modified risk of disease. A further 10 genes exist within chromosomal regions associated with disease. The remaining 58 genes have no obvious association to disease in any system. The 39% rate of disease association determined here is much higher (hypergeometric P = 7.5 × 10-16) than the 14% predicted by OMIM across the genome, with 2,549 genes defined as the basis of heritable disease out of the 18,091 total.
Clusters of disease among candidate genes
For the case of gastric cancer, another cluster of seven genes (AKT3, KRAS, MAP2K4, PIK3CB, PLCB1, PIK3R5 and PPP3R2) was identified. Four of these genes have been previously associated with gastrointestinal disease while AKT3, PIK3CB and PIK3R5 have not, although the differential expression of AKT3 in gastric cancer is well defined [see Additional file 2]. We suggest these previously non-associated genes are strong candidates for further study into the basis of these diseases and are potential prognostic markers.
Data mining approaches have allowed us to gain an insight into the complex relationships existing between gene expression, disease association, network connectivity and tissue specificity. We have identified elevated rates of expression and network connectivity among broadly expressed genes, and among disease-associated tissue-specific genes.
In particular, when exploring the relationship between tissue specificity and disease association, we found this relationship most interesting. While there is a moderate positive relationship between the number of tissues in which a gene is expressed and the proportion of disease genes, we show that this relationship is reversed when only considering genes for which interactions have been reported. We present this phenomenon as an example of the well-reported Simpson's Paradox. To a great extent, the inclusion of number of interactions as a threshold parameter in our guilt-by-association algorithm obviates the need to also include tissue specificity.
However, it should also be acknowledged that probability values associated with testing the null hypothesis of a given PCC not being statistically different from zero were computed assuming asymptotic normality and as such are prone to inaccuracies. With this in mind, we focussed on combining discrete parameters such as number of connections and the association to disease-associated genes to identify a group of genes, not previously confirmed as disease causing, that are involved in interactions with disease causing genes. The nature of these newly identified interactions could range from epistatic interactions (i.e., the action of one gene is suppressed by another such as the case of RAD51 and BRCA1) to physical gene-gene interactions to correlated co-expression. Based on bibliographical validation and network re-construction we have identified several candidate genes that may impact the development of cancer and hypothesize that many other members of this list will ultimately be confirmed as modifiers of various genetic diseases.
Finally, it should be noted that while new algorithms are being proposed in the literature on a rather frantic pace, the task of comprehensively comparing algorithms could be unattainable if not futile. Instead, we claim that our conservative thresholds for predicting disease association is justified because using thresholds of known disease genes increases our likelihood of success given any estimation process is going to have a degree of false positives. We acknowledge the list does not exhaust all possible disease genes but merely gives researchers the best short list for further study.
massively parallel signature sequencing
Pearson correlation coefficient
genes in our dataset that are disease-associated according to OMIM as of September 19, 2007
genes in our dataset that are non-disease-associated genes also according to OMIM
genes in our dataset for which interactions have been reported
genes in our dataset for which interactions have not been reported.
The authors are grateful to Victor Jongeneel and Christian Haudenschild for providing the gene-centric and tag-centric annotated MPSS data files. The authors would like to acknowledge three reviewers who provided important insights. In particular, comments by Borja Calvo on previous versions of this manuscript greatly improved its final outcome. This work was supported by the CSIRO Centre for Complex Systems Science http://www.csiro.au/science/ComplexSystemsScience.html.
- Wolfe CJ, Kohane IS, Butte AJ: Systematic survey reveals general applicability of "guilt-by-association" within gene coexpression networks. BMC Bioinformatics. 2005, 6: 227-10.1186/1471-2105-6-227.View ArticlePubMedPubMed CentralGoogle Scholar
- Jongeneel CV, Delorenzi M, Iseli C, Zhou D, Haudenschild CD, Khrebtukova I, Kutnetsov D, Stevenson BJ, Strausberg RL, Simpson AJG, Vasicek TJ: An atlas of human gene expression from massively parallel signature sequencing (MPSS). Genome Res. 2005, 15: 1007-1014. 10.1101/gr.4041005.View ArticlePubMedPubMed CentralGoogle Scholar
- Su AI, Wiltshire T, Batalov S, Lapp H, Ching KA, Block D, Zhang J, Soden R, Hayakawa M, Kreiman G, Cooke MP, Walker JR, Hogenesch JB: A gene atlas of the mouse and human protein-encoding transcriptomes. Proc Natl Acad Sci USA. 2004, 101: 6062-6067. 10.1073/pnas.0400782101.View ArticlePubMedPubMed CentralGoogle Scholar
- Pettretto E, Mangion J, Dickens NJ, Cook SA, Kumaran MK, Lu H, Fischer J, Maatz H, Kren V, Pravenec M, Hubner M, Hubner N, Aitman TJ: Heritability and tissue specificity of expression quantitative trait loci. PLoS Genetics. 2006, 2: 1625-1633. 10.1371/journal.pgen.0020172.View ArticleGoogle Scholar
- Zhang L, Li WH: Human SNPs reveal no evidence of frequent positive selection. Mol Biol Evol. 2005, 22: 2504-2507. 10.1093/molbev/msi240.View ArticlePubMedGoogle Scholar
- Yang J, Su AI, Li WH: Gene expression evolves faster in narrowly than in broadly expressed mammalian genes. Mol Biol Evol. 2005, 22: 2113-2118. 10.1093/molbev/msi206.View ArticlePubMedGoogle Scholar
- Winter EE, Goodstadt L, Ponting CP: Elevated rates of protein secretion, evolution and disease among tissue-specific genes. Genome Res. 2004, 14: 54-61. 10.1101/gr.1924004.View ArticlePubMedPubMed CentralGoogle Scholar
- Reverter A, Ingham A, Lehnert SA, Tan SH, Wang YH, Ratnakumar A, Dalrymple BP: Simultaneous identification of differential gene expression and connectivity in inflammation, adipogenesis and cancer. Bioinformatics. 2006, 22: 2396-2404. 10.1093/bioinformatics/btl392.View ArticlePubMedGoogle Scholar
- Choi JK, Yu U, Yoo OJ, Kim S: Differential coexpression analysis using microarray data and its application to human cancer. Bioinformatics. 2005, 21: 4348-4355. 10.1093/bioinformatics/bti722.View ArticlePubMedGoogle Scholar
- Elo LL, Järvenpää H, Oresic M, Lahesmaa R, Aittokallio T: Systematic construction of gene coexpression networks with application to human T helper cell differentiation process. Bioinformatics. 2007, 23: 2096-2103. 10.1093/bioinformatics/btm309.View ArticlePubMedGoogle Scholar
- Özgür A, Vu T, Erkan G, Radv DR: Identifying gene-disease associations using centrality on a literature mined gene-interaction network. Bioinformatics. 2008, 24: i277-i285. 10.1093/bioinformatics/btn182.View ArticlePubMedPubMed CentralGoogle Scholar
- Luscombe NM, Babu MM, Yu H, Snyder M, Telchmann SA, Gerstein M: Genomic analysis of regulatory network dynamics reveals large topological changes. Nature. 2004, 431: 308-312. 10.1038/nature02782.View ArticlePubMedGoogle Scholar
- Liang H, Li WH: Gene essentiality, gene duplicability and protein connectivity in human and mouse. Trends Genet. 2007, 23: 375-378. 10.1016/j.tig.2007.04.005.View ArticlePubMedGoogle Scholar
- Calvo B, López-Bigas N, Furney SJ, Larrañaga P, Lozano JA: A partially supervised classification approach to dominant and recessive human disease gene prediction. Comput Meth Prog Bio. 2007, 85: 229-237. 10.1016/j.cmpb.2006.12.003.View ArticleGoogle Scholar
- Franke L, van Bakel H, Fokkens L, de Jong ED, Egmont-Petersen M, Wijmenga C: Reconstruction of a functional human gene network, with an application for prioritizing positional candidate genes. Am J Hum Genet. 2006, 78: 1011-1025. 10.1086/504300.View ArticlePubMedPubMed CentralGoogle Scholar
- Bader GD, Donaldson I, Wolting C, Ouellette BF, Pawson T, Hoque CW: BIND – The Biomolecular Interaction Network Database. Nucleic Acids Res. 2001, 29: 242-245. 10.1093/nar/29.1.242.View ArticlePubMedPubMed CentralGoogle Scholar
- Mishra GR, Suresh M, Kumaran K, Kannabiran N, Suresh S, Bala P, Shivakumar K, Anuradha N, Reddy R, Raghavan TM, Menon S, Hanumanthu G, Gupta M, Upendran S, Gupta S, Mahesh M, Jacob B, Mathew P, Chatterjee P, Arun KS, Sharma S, Chandrika KN, Deshpande N, Palvankar K, Raghavnath R, Krishnakanth R, Karathia H, Rekha B, Nayak R, Vishnupriya G, Kumar HG, Nagini M, Kumar GS, Jose R, Deepthi P, Mohan SS, Gandhi TK, Harsha HC, Deshpande KS, Sarker M, Prasad TS, Pandey A: Human protein reference database – 2006 update. Nucleic Acids Res. 2006, 34: D411-D414. 10.1093/nar/gkj141.View ArticlePubMedGoogle Scholar
- Kanehisa M, Goto S: KEGG: Kyoto encyclopedia of genes and genomes. Nucleic Acids Res. 2000, 28: 27-30. 10.1093/nar/28.1.27.View ArticlePubMedPubMed CentralGoogle Scholar
- Joshi-Tope G, Gillespie M, Vastrik I, D'Eustachio P, Schmidt E, de Bono B, Jassal B, Gopinath GR, Wu GR, Matthews L, Lewis S, Birney E, Stein L: Reactome: a knowledgebase of biological pathways. Nucleic Acids Res. 2005, 33: D428-D432. 10.1093/nar/gki072.View ArticlePubMedGoogle Scholar
- Calvano SE, Xiao W, Richards DR, Felciano RM, Baker HV, Cho RJ, Chen RO, Brownstein BH, Cobb JP, Tscheke SK, Miller-Graziano C, Moldawer LL, Mindrinos MN, Davis RW, Tompkins RG, Lowry SF, the Inflammation and Host Response to Injury Large Scale Collaborative Research: A network-based analysis of systemic inflammation in human. Nature. 2005, 337: 1032-1037. 10.1038/nature03985.View ArticleGoogle Scholar
- Reverter A, McWilliam SM, Barris W, Dalrymple BP: A rapid method for computationally inferring transcriptome coverage and microarray sensitivity. Bioinformatics. 2005, 21: 80-89. 10.1093/bioinformatics/bth472.View ArticlePubMedGoogle Scholar
- McKusick VA: Online Mendelian Inheritance in Man, OMIM™. [http://www.ncbi.nlm.nih.gov/Omim]
- Simpson EH: The interpretation of interaction in contingency tables. J Royal Stat Soc Ser B. 1951, 13: 238-241.Google Scholar
- Goh K, Cusick ME, Valle D, Childs B, Vidal M, Barabási AL: The human disease network. Proc Natl Acad Sci USA. 2007, 104: 8685-8690. 10.1073/pnas.0701361104.View ArticlePubMedPubMed CentralGoogle Scholar
- Shannon P, Markiel A, Ozier O, Baliga NS, Wang JT, Ramage D, Amin N, Schwikowski B, Ideker T: Cytoscape: a software environment for integrated models of biomolecular interaction networks. Genome Res. 2003, 13: 2498-504. 10.1101/gr.1239303.View ArticlePubMedPubMed CentralGoogle Scholar
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.