Network-based Identification of Novel Cancer Genes

Genes involved in cancer susceptibility and progression can serve as templates for searching protein networks for novel cancer genes. To this end, we introduce a general network searching method, MaxLink, and apply it to find and rank cancer gene candidates by their connectivity to known cancer genes. Using a comprehensive protein interaction network, we searched for genes connected to known cancer genes. First, we compiled a new set of 812 genes involved in cancer, more than twice the number in the Cancer Gene Census. Their network neighbors were then extracted. This candidate list was refined by selecting genes with unexpectedly high levels of connectivity to cancer genes and without previous association to cancer. This produced a list of 1891 new cancer candidates with up to 55 connections to known cancer genes. We validated our method by cross-validation, Gene Ontology term bias, and differential expression in cancer versus normal tissue. An example novel cancer gene candidate is presented with detailed analysis of the local network and neighbor annotation. Our study provides a ranked list of high priority targets for further studies in cancer research. Supplemental material is included. Genes involved in cancer susceptibility and progression can serve as templates for searching protein networks for novel cancer genes. To this end, we introduce a general network searching method, MaxLink, and apply it to find and rank cancer gene candidates by their connectivity to known cancer genes. Using a comprehensive protein interaction network, we searched for genes connected to known cancer genes. First, we compiled a new set of 812 genes involved in cancer, more than twice the number in the Cancer Gene Census. Their network neighbors were then extracted. This candidate list was refined by selecting genes with unexpectedly high levels of connectivity to cancer genes and without previous association to cancer. This produced a list of 1891 new cancer candidates with up to 55 connections to known cancer genes. We validated our method by cross-validation, Gene Ontology term bias, and differential expression in cancer versus normal tissue. An example novel cancer gene candidate is presented with detailed analysis of the local network and neighbor annotation. Our study provides a ranked list of high priority targets for further studies in cancer research. Supplemental material is included. The function of a protein can be expressed in terms of its interactions with other molecules. All interactions between all proteins define the “protein interactome,” i.e. the complete interaction network of the proteins of an organism. These networks form the backbone of molecular pathways and cellular processes. Thus, the construction of interaction networks will shed light on many aspects of the dynamic and interactive function of human proteins. Several efforts in reconstructing the human interactome are ongoing. Interactions may be measured directly with high throughput yeast two-hybrid or pulldown assays (1.Rual J.F. Venkatesan K. Hao T. Hirozane-Kishikawa T. Dricot A. Li N. Berriz G.F. Gibbons F.D. Dreze M. Ayivi-Guedehoussou N. Klitgord N. Simon C. Boxem M. Milstein S. Rosenberg J. Goldberg D.S. Zhang L.V. Wong S.L. Franklin G. Li S. Albala J.S. Lim J. Fraughton C. Llamosas E. Cevik S. Bex C. Lamesch P. Sikorski R.S. Vandenhaute J. Zoghbi H.Y. Smolyar A. Bosak S. Sequerra R. Doucette-Stamm L. Cusick M.E. Hill D.E. Roth F.P. Vidal M. Towards a proteome-scale map of the human protein-protein interaction network.Nature. 2005; 437: 1173-1178Crossref PubMed Scopus (2302) Google Scholar, 2.Stelzl U. Worm U. Lalowski M. Haenig C. Brembeck F.H. Goehler H. Stroedicke M. Zenkner M. Schoenherr A. Koeppen S. Timm J. Mintzlaff S. Abraham C. Bock N. Kietzmann S. Goedde A. Toksöz E. Droege A. Krobitsch S. Korn B. Birchmeier W. Lehrach H. Wanker E.E. A human protein-protein interaction network: a resource for annotating the proteome.Cell. 2005; 122: 957-968Abstract Full Text Full Text PDF PubMed Scopus (1897) Google Scholar). Experimental interactions have been collected from multiple sources to build large interaction networks (3.Bader G.D. Betel D. Hogue C.W. BIND: the biomolecular interaction network database.Nucleic Acids Res. 2003; 31: 248-250Crossref PubMed Scopus (991) Google Scholar, 4.Ramani A.K. Bunescu R.C. Mooney R.J. Marcotte E.M. Consolidating the set of known human protein-protein interactions in preparation for large-scale mapping of the human interactome.Genome Biol. 2005; 6: R40Crossref PubMed Google Scholar, 5.Peri S. Navarro J.D. Amanchy R. Kristiansen T.Z. Jonnalagadda C.K. Surendranath V. Niranjan V. Muthusamy B. Gandhi T.K. Gronborg M. Ibarrola N. Deshpande N. Shanker K. Shivashankar H.N. Rashmi B.P. Ramya M.A. Zhao Z. Chandrika K.N. Padma N. Harsha H.C. Yatish A.J. Kavitha M.P. Menezes M. Choudhury D.R. Suresh S. Ghosh N. Saravana R. Chandran S. Krishna S. Joy M. Anand S.K. Madavan V. Joseph A. Wong G.W. Schiemann W.P. Constantinescu S.N. Huang L. Khosravi-Far R. Steen H. Tewari M. Ghaffari S. Blobe G.C. Dang C.V. Garcia J.G. Pevsner J. Jensen O.N. Roepstorff P. Deshpande K.S. Chinnaiyan A.M. Hamosh A. Chakravarti A. Pandey A. Development of human protein reference database as an initial platform for approaching systems biology in humans.Genome Res. 2003; 13: 2363-2371Crossref PubMed Scopus (890) Google Scholar). The network can be augmented considerably by inferred interactions either in the same or from other species (6.Lehner B. Fraser A.G. A first-draft human protein-interaction map.Genome Biol. 2004; 5: R63Crossref PubMed Google Scholar, 7.Brown K.R. Jurisica I. Online predicted human interactions database.Bioinformatics. 2005; 21: 2076-2082Crossref PubMed Scopus (473) Google Scholar, 8.Persico M. Ceol A. Gavrila C. Hoffmann R. Florio A. Cesareni G. HomoMINT: an inferred human network based on orthology mapping of protein interactions discovered in model organisms.BMC Bioinformatics. 2005; 6: S21Crossref PubMed Google Scholar, 9.Rhodes D.R. Tomlins S.A. Varambally S. Mahavisno V. Barrette T. Kalyana-Sundaram S. Ghosh D. Pandey A. Chinnaiyan A.M. Probabilistic model of the human protein-protein interaction network.Nat. Biotechnol. 2005; 23: 951-959Crossref PubMed Scopus (353) Google Scholar, 10.Jensen L.J. Kuhn M. Stark M. Chaffron S. Creevey C. Muller J. Doerks T. Julien P. Roth A. Simonovic M. Bork P. von Mering C. STRING 8—a global view on proteins and their functional interactions in 630 organisms.Nucleic Acids Res. 2009; 37: D412-D416Crossref PubMed Scopus (1937) Google Scholar). The largest predicted human interactome is currently provided by FunCoup (11.Alexeyenko A. Sonnhammer E.L. Global networks of functional coupling in eukaryotes from comprehensive data integration.Genome Res. 2009; 19: 1107-1116Crossref PubMed Scopus (132) Google Scholar), which uses eight types of evidence and transfers interactions extensively from model organism orthologs. The development of new therapeutics and diagnostics rely on the understanding of disease mechanisms. Therefore, the identification of novel disease-associated genes is of great importance. Disease genes have traditionally been found by genetic linkage analysis or gene association studies, but this is very time-consuming and costly and often fails due to lack of data. Particularly for complex diseases involving many genes, these methods are unreliable (12.Emahazion T. Feuk L. Jobs M. Sawyer S.L. Fredman D. St Clair D. Prince J.A. Brookes A.J. SNP association studies in Alzheimer's disease highlight problems for complex disease analysis.Trends Genet. 2001; 17: 407-413Abstract Full Text Full Text PDF PubMed Scopus (151) Google Scholar). Bioinformatics methods can be used to accelerate disease gene discovery either based on gene annotation and sequence features (13.Perez-Iratxeta C. Bork P. Andrade M.A. Association of genes to genetically inherited diseases using data mining.Nat. Genet. 2002; 31: 316-319Crossref PubMed Scopus (283) Google Scholar, 14.Turner F.S. Clutterbuck D.R. Semple C.A. POCUS: mining genomic sequence annotation to predict disease genes.Genome Biol. 2003; 4: R75Crossref PubMed Scopus (186) Google Scholar) or based on network analysis (15.George R.A. Liu J.Y. Feng L.L. Bryson-Richardson R.J. Fatkin D. Wouters M.A. Analysis of protein sequence and interaction data for candidate disease gene prediction.Nucleic Acids Res. 2006; 34: e130Crossref PubMed Scopus (126) Google Scholar, 16.Lage K. Karlberg E.O. Størling Z.M. Olason P.I. Pedersen A.G. Rigina O. Hinsby A.M. Tümer Z. Pociot F. Tommerup N. Moreau Y. Brunak S. A human phenome-interactome network of protein complexes implicated in genetic disorders.Nat. Biotechnol. 2007; 25: 309-316Crossref PubMed Scopus (751) Google Scholar, 17.Oti M. van Reeuwijk J. Huynen M.A. Brunner H.G. Conserved co-expression for candidate disease gene prioritization.BMC Bioinformatics. 2008; 9: 208Crossref PubMed Scopus (34) Google Scholar, 18.Wu X. Jiang R. Zhang M.Q. Li S. Network-based global inference of human disease genes.Mol. Syst. Biol. 2008; 4: 189Crossref PubMed Scopus (548) Google Scholar, 19.Ideker T. Sharan R. Protein networks in disease.Genome Res. 2008; 18: 644-652Crossref PubMed Scopus (635) Google Scholar). The network-based methods normally connect gene networks with phenotype networks to infer gene-disease relationships. These works, however, are limited to using only direct interaction data and/or were only applied to rank a short list of candidate genes in a genomic interval. Here we describe a new generic network-based approach, MaxLink, for predicting novel candidate members to known biomolecular processes and pathways. A typical application is the identification of new disease genes based on a set of known disease genes. We applied MaxLink to the human interactome generated by FunCoup to screen for new cancer genes. To seed the screen, we compiled a list of 812 known cancer genes, 364 from the Cancer Gene Census (20.Futreal P.A. Coin L. Marshall M. Down T. Hubbard T. Wooster R. Rahman N. Stratton M.R. A census of human cancer genes.Nat. Rev. Cancer. 2004; 4: 177-183Crossref PubMed Scopus (2432) Google Scholar) and 448 genes from text mining. MaxLink assigns a score to every new candidate gene based on the number of links to a seed set. We show that the maxlink score is a useful indicator of candidate reliability by three types of validations: cross-validation, differential cancer expression, and GO 1The abbreviations used are:GOGene OntologyDEdifferential expression scoreHPAHuman Protein AtlasKEGGKyoto Encyclopedia of Genes and Genomes. term analysis. The screen resulted in nearly 2000 candidates of which nearly 200 are connected to over 10 known cancer genes. These genes have, to our knowledge, no clear former evidence supporting association with cancer. However, their network connection to cancer genes makes them worth particular focus when developing biomarkers or studying oncogenesis. As the candidate list is long, it makes sense to explore the top ranking genes first. Gene Ontology differential expression score Human Protein Atlas Kyoto Encyclopedia of Genes and Genomes. The input data set of known cancer genes was collected from Swiss-Prot (21.The UniProt Consortium The Universal Protein Resource (UniProt).Nucleic Acids Res. 2008; 36: D190-D195Crossref PubMed Scopus (858) Google Scholar) and from the Cancer Gene Census (20.Futreal P.A. Coin L. Marshall M. Down T. Hubbard T. Wooster R. Rahman N. Stratton M.R. A census of human cancer genes.Nat. Rev. Cancer. 2004; 4: 177-183Crossref PubMed Scopus (2432) Google Scholar). The Swiss-Prot genes were identified by searching annotations in the CC field, which represents curated annotations and includes a subcategory for annotations indicating disease involvement. The disease annotations of the CC field were matched against cancer-specific terms (see supplemental Table 4), and genes for which a match could be found were added to the set of known cancer genes. Genes and matching keywords are detailed in supplemental Table 1. The Gene Ontology functional term analysis was done using the amiGO web site. Enrichment analysis of terms in the major cluster (348 genes) versus UniProtKB (20,740 genes) resulted in a total of 231 terms with p < 10−2. This list was abbreviated by requiring p < 10−10 and enrichment >5, resulting in 34 GO terms (supplemental Table 2). We used the human FunCoup protein network (11.Alexeyenko A. Sonnhammer E.L. Global networks of functional coupling in eukaryotes from comprehensive data integration.Genome Res. 2009; 19: 1107-1116Crossref PubMed Scopus (132) Google Scholar) to identify network neighbors to the previously retrieved input genes. Only links with a confidence value >0.75 were considered. Each candidate gene was assigned a maxlink score for ranking that equals the number of linked known cancer genes. To identify genes with possible cancer annotations, the complete UniProt (22.Bairoch A. Apweiler R. Wu C.H. Barker W.C. Boeckmann B. Ferro S. Gasteiger E. Huang H. Lopez R. Magrane M. C. N. The protein resource (UniProt).Nucleic Acids Res. 2005; PubMed Scopus Google Scholar) and as as reference were searched for cancer-specific text terms (supplemental Table Genes with a match were from the candidates genes with a gene found in the of P. K. B. M. Y. L. G. F. T. Down T. T. S. J. S. S. M. R. K. N. A. A. D. F. E. D. I. K. P. B. A. B. A. S. D. M. I. G. D. G. S. A.J. J. S. M. E. T. V. R. J. Hubbard A. G. J. A. S. Acids Res. 2008; 36: PubMed Scopus Google Scholar) were the of a connections were to genes, it was of cancer and was this we all genes the known cancer genes all genes with cancer annotation (see and all other genes. The gene of these were and genes over more links to genes with cancer than to the known cancer genes were We a score expression score for differential protein expression levels in cancer types to their normal (see Table from the of the Human Protein was by the expression in a normal from the expression in the cancer for gene and tissue. To for were by based on the expression of tissue. A of represents the the total for gene was by the of all for all cancer and normal in and and and and and and in a new GO terms for all genes were retrieved from P. K. B. M. Y. L. G. F. T. Down T. T. S. J. S. S. M. R. K. N. A. A. D. F. E. D. I. K. P. B. A. B. A. S. D. M. I. G. D. G. S. A.J. J. S. M. E. T. V. R. J. Hubbard A. G. J. A. S. Acids Res. 2008; 36: PubMed Scopus Google Scholar), and the terms were to all All GO terms for the set of known cancer genes were for enrichment with a The set of GO terms was then for < for of the candidates of all genes a number of linked known cancer genes or to a that for were for GO term by the of the by the of the same term for the known cancer genes. We have an analysis to identify and rank candidate cancer genes based on their connectivity to known cancer genes in the FunCoup cancer we gene with clear evidence for cancer involvement. To the and of the known cancer genes, we their network using them as we candidate novel cancer genes and refined this list by to our approach, we used three types of cross-validation, enrichment of cancer GO and differential expression in cancer versus normal tissue. Our with known cancer genes. a previous the Cancer Gene (20.Futreal P.A. Coin L. Marshall M. Down T. Hubbard T. Wooster R. Rahman N. Stratton M.R. A census of human cancer genes.Nat. Rev. Cancer. 2004; 4: 177-183Crossref PubMed Scopus (2432) Google Scholar) identified 364 cancer genes. text mining Swiss-Prot for genes to be involved in cancer, we identified genes. this list with the Cancer Gene Census resulted in 812 cancer genes (supplemental Table To this set of genes in terms of network we cluster This major with with and as in 1. Thus, of the known cancer genes were in a that processes to cancer. To we enrichment of functional annotation terms in the Gene Ontology database M. C.A. J.A. D. H. K. M.A. Hill L. A. S. M. G. Gene for the of Genet. 25: PubMed Scopus Google Scholar) to all human genes. We enrichment p < for terms as and and (supplemental Table 2). This is in with known processes. The FunCoup network was used to candidates connected to known cancer genes by high confidence our was to find genes previously with cancer, the list was further refined by a number of the genes that annotation a association with cancer were from the FunCoup was using data of which were linked to of genes were to that the candidates are in with the This a and genes with only cancer genes be linked to many known cancer genes have many links and are involved in cancer. Thus, genes were in the links to known cancer genes than by their connectivity in the genes may have been found are connected and of a association to the known cancer genes. A of the analysis is in all a list of 1891 candidates with a maxlink score to known cancer genes) between 55 and (see supplemental Table our method it be to the known cancer genes in a We MaxLink of the known cancer genes we were to identify of the genes on However, only of the known cancer genes links to other input the is to the by the As we we enrichment of the known cancer genes the retrieved genes, i.e. their in the retrieved set to their in the The enrichment for all genes was more than < However, this to over for the genes with maxlink score (see The Human Protein Atlas L. E. P. L. A. C.A. A. J. H. P. E. A. S. K. C. S. F. M. A Human Protein Atlas for expression based on 2008; Full Text Full Text PDF PubMed Scopus Google Scholar) protein expression in normal and from a large number of Using these we a normal versus cancer for the candidate cancer genes with expression data in using all cancer To the of a high maxlink we the of genes with i.e. when the differential expression on for all cancer As in this for of the candidates of genes with a maxlink score and was considerably with the known cancer genes and as a This that candidates linked to a high number of known cancer genes are for cancer. To this is by a of normal expression or an in cancer expression, we the expression levels as a function of maxlink score (see The is an of normal and cancer expression but with a in Thus, the MaxLink can genes expressed in a of and the maxlink score is a useful indicator of cancer We that the of the known cancer genes was only than that of This can be by the that was with a cancer focus and is for cancer genes. our candidate cancer genes show the same GO term enrichment as the known cancer genes, this further to their in cancer. To we retrieved all GO terms for the known cancer genes and for a a total of enrichment or to was the for These GO terms were for enrichment in of the candidate genes by of links to known cancer genes. The enrichment to the number of linked known cancer genes that genes more in the cancer network are more with cancer. Our screen resulted in a list of 1891 novel candidate genes (supplemental Table the of the maxlink score as an indicator of cancer it makes sense to focus on candidates with the linked known cancer genes. The list candidates with 10 or more linked known cancer genes, and these be as the targets for cancer To a candidate cancer gene may be we an number which is a of protein A. was to 34 known cancer genes (see we predict cancer is to or be to is expressed in all cancer types and differential expression with normal in and and Thus, may a more in it is to be a cancer The network neighbors of have differential expression supporting the that it is for a of cancer. the of the neighbors is this a very of the of for cancer cancer is it cancer types as and A that in been to in which can to Y. Zhang W. L. R. L. R. W. in in and cancer in Genet. 2005; 37: PubMed Scopus Google Scholar). is in is in which been implicated in and cancer Y. Zhang W. L. R. L. R. W. in in and cancer in Genet. 2005; 37: PubMed Scopus Google Scholar). These by in for and its neighbors as as from annotation of many of the Our analysis based on expression, the gene and the connection between and cancer, for the cancer types with the but that it may cancer in tissue. it was of the top ranking novel cancer gene is no of cancer association in UniProt or However, the presented that it a in We have a general network-based for and ranking candidate novel genes to a or disease and have applied it to find novel cancer genes. The show that the ranked list produced by our method is for cancer genes. Cancer is in this study as This is a but is based on the that in are often by in the same for or is by the that the network of known cancer genes only large cluster which show very Cusick M.E. D. B. Vidal M. The human disease 2007; PubMed Scopus Google Scholar) that are often by the same genes. The of this is that genes found in have a for and of the of the MaxLink other diseases or processes can be A more to disease gene is linkage analysis the gene with a disease is known to be found in a genomic that can in the of a genes. MaxLink could be used to genes for as as a number of genes are known for the disease in The with other methods be the of evidence in the FunCoup short it may be to the with this study to a of of our candidate genes were as cancer genes by the However, a gene have differential expression in this its in cancer. The protein with the progression may be to with the However, the predicted functional coupling to a cancer is and the gene in may to be a useful or The differential used was only to the of the MaxLink and as a indicator of cancer we all cancer a gene expressed in only or a cancer types an total The of this study is the ranked list of novel cancer gene The candidates connected to 10 or more known cancer genes are targets for new that will the to understanding cancer. of the candidates with a maxlink score may cancer biomarkers or but is a for on the high genes. A high maxlink score is an that the candidate as a and a in the and is more to be of in many the other are for and may be as targets for An to this be that in in but only is functional in a a a an cancer-specific We and for and We the Human Resource for to the protein data. with