Locus-Specific Databases in Cancer: What Future in a Post-Genomic Era? The TP53 LSDB paradigm

Introduction

Using protein sequencing, Vernon Ingram was the first to discover that a small change in the hemoglobin protein could lead to human sickle-cell anemia [Ingram, 1956]. Since this pioneering work, it has been largely demonstrated that gene mutations are the basis for most genetic diseases. (Although epigenetic modifications are also an important component in all human diseases including cancer, they will not be discussed here as they are not integrated in LSDBs.) In the late 1970s, the tedious task of protein sequencing was replaced by the revolutionary introduction of DNA sequencing and molecular cloning technologies [Collins, 1995]. Over the years, progress has been made in identifying the genes involved in both monogenic and polygenic disorders, including such complex diseases as cancer. For these genes, numerous and various types of alterations have been described, ranging from point mutations to large deletions or translocations. Reporting, storing, classifying, and analyzing these mutations have been a major challenge [Horaitis and Cotton, 2004]. For many years now, locus-specific databases (LSDBs) have been developed for this purpose. Although LSDBs are independently developed for single genes, they do offer great accuracy as they are curated manually by experts in the field [Claustres et al., 2002; Auerbach et al., 2011]. They provide information that can be used for large-scale analyses and often include structural, functional, or evolutionary data. For constitutional mutations associated with a genetic syndrome, several LSDBs also include phenotypic data useful for the study of genotype–phenotype correlation.

The development of high-throughput methodologies capable of analyzing the pattern of transcription of an entire cell or tissue (expression array or RNA sequencing), defining the structure and organization of the chromatin (chromatin immunoprecipitation) or sequencing an entire genome in a few days (next-generation sequencing, NGS) has radically changed the entire field of biology [Park, 2009; Hawkins et al., 2010; Metzker, 2010; Ozsolak and Milos, 2011]. Furthermore, this methodological revolution led to the discovery of novel layers of complexity in gene organization and how their expression is regulated. These spectacular advances have laid a path to a postgenomic era in which healthcare research and provision will be very different.

In cancer research, genomic studies in the pregenomic era were limited; they were focused either on a small number of genes analyzed in large patient cohorts, or on a more significant number of genes but in only a few tumors (Fig. 1). Indeed, large-scale analysis combining a multitude of genes and tumors was a Herculean and costly task.

Today, in the postgenomic era, these barriers have fallen and whole genome sequencing in a multitude of tumors can be performed in a matter of weeks. The International Cancer Genome Consortium (ICGC, http://dcc.icgc.org/), The Cancer Genome Atlas Project (TCGA, http://cancergenome.nih.gov/), and the Sanger Institute (http://www.sanger.ac.uk/) have undertaken large-scale cancer genome analyses in different types/subtypes of cancer and several reports from these projects have already been published [Hudson et al., 2010; Alexandrov et al., 2013, Garraway and Lander, 2013, Koboldt et al., 2013]. These studies will lead to profound changes in LSDB management.

Whole tumor genome sequencing will result in an enormous increase in the volume of acquired data and lead molecular genetics to a new world of “big data” [Schadt et al., 2010]. Exaoctets of raw data (1 Gigaoctet (Go) = 10⁹ octets; 1 Exaoctet (Eo) = 10¹⁸ octets) will be generated and sophisticated software will be required for mining and interpreting them. Handling, accessing, and analyzing this amount of data will also require novel computing processes such as cloud or heterogeneous computing. Furthermore, the extraction of specific gene information will be a very challenging task for LSDB curators [Metzker, 2010, Schadt et al., 2010].

Recent studies have shown that the organization of the mammalian genome is far more complex than previously thought and that noncoding regions of the genome are also potential targets for alterations [Gerstein et al., 2012]. Evaluation of mutation pathogenicity must therefore evolve and take into consideration a setting larger than the end protein [Sauna and Kimchi-Sarfaty, 2011]. Furthermore, cancer genomes are polluted by thousands of random, “passenger” mutations unrelated to neoplastic progression, necessitating novel bioinformatics tools to identify the few relevant “driver” mutations truly associated with cell transformation [Chanock and Thomas, 2007].

How will the definition of a cancer mutation evolve in the future? What will be the fate of LSDBs in the postgenomic era? Will they even survive these major changes? How will the flow between data, curators, and users be modified by these new technologies, which displace publications as the primary material for data mining? To address these questions, the present review will focus on the TP53 gene (MIM #191170). Indeed, this latter is the most frequently mutated gene in human cancer and the TP53 database is the largest collection of mutations, furthermore compiled from a wide gamut of cancer types.

Thus, the evolution and content of the TP53 LSDB will be used here as a paradigm to explore the current situation and make assumptions for the future of LSDBs. This review will focus exclusively on cancer genes and cancer LSDBs as they have specificities and requirements generally different from those of other genetic diseases, although some issues are also applicable outside the strict setting of cancer. Recent works have demonstrated the presence and predominance of passenger mutations in tumor genomes. These passenger mutations must be differentiated from the rare driver mutations, a challenge very similar to that faced by clinical geneticists to differentiate SNPs from disease-causing mutations. Germline mutations in tumor suppressor genes lead to hereditary cancer syndromes with heterogeneous penetrance managed as complex Mendelian disorders.

Issues related to the storage and management of large amounts of data will not be discussed here as they have already been extensively reviewed [Schadt et al., 2010; Hood and Rowen, 2013].

Mutations in the Pre- and Postgenomic Eras

Cancer is a genetic disease characterized by a wide variety of genomic alterations, ranging from single-nucleotide substitutions to large chromosomal abnormalities such as translocations or duplications [Stratton, 2011; Vogelstein et al., 2013]. Large-scale analyses of several thousand different types of tumors have revealed that tumoral DNA is flooded with single-nucleotide variants (SNVs) scattered across the entire genome, including both the coding and noncoding regions. Changes to DNA methylation, a process involved in the regulation of gene expression, are also key players in cancer.

The classical view prevailing forty years ago divided the genome into two entities: one comprising genes and their regulatory regions responsible for encoding messenger RNA translated in turn into protein; and one comprising “junk DNA” with unknown and consequently nonessential function [Susumo, 1972]. The resulting gene-centric view created an analytical bias toward coding regions. Identification of an SNV as an acquired somatic mutation was checked either by looking at the constitutional DNA or, more often, by comparing the sequence to the dbSNP, which gathers all natural polymorphisms found in the human genome (http://www.ncbi.nlm.nih.gov/SNP/). SNV were dichotomized into synonymous SNV (sSNV) that do not change the protein sequence and nonsynonymous SNV (nsSNV) that lead to an amino-acid substitution. These nsSNVs are the most deleterious for gene function. Nevertheless, a significant number of sSNV detected in exonic sequence should not be considered as silent mutations as they can lead to aberrant RNA splicing or structure and thus a decrease in protein translation [Sauna and Kimchi-Sarfaty, 2011]. The sSNV c.375G>T (p.=T125T) in the TP53 gene had been initially considered as neutral in multiple reports but was later shown to be detrimental for TP53 splicing [Varley et al., 1998].

Recent discoveries have profoundly changed our knowledge of the mammalian transcriptome and genome plasticity and oblige us to reconsider the definition of “cancer associated mutation.” This new knowledge will also have important consequences on mutation analysis and interpretation (Fig. 2).

Plasticity of the Human Genome

Somatic mosaicism denotes the presence of multiple populations of cells with different genotypes in a single organism [Biesecker and Spinner, 2013]. It was long assumed that somatic mosaicism was rare, occurring only occasionally in hereditary diseases (e.g., forms of Turner's syndrome, trisomy). Recent studies have revealed that mosaicism is far more frequent than previously thought and could concern the majority of somatic cells [Stratton et al., 2009; Jacobs et al., 2012]. Starting at fertilization and throughout life, the genome is continuously the target of exogenous or endogenous mutagens, DNA replication errors and various types of recombination (Fig. 3). The pattern and the frequency of these alterations differ according to cell type and to the degree of exposure to various exogenous mutagens, for example, lung epithelial cells exposed to tobacco smoke, skin cells exposed to UV light, or organ tissues exposed to gamma radiation. The vast majority of these somatic alterations are thought to be neutral with no phenotypic contribution. However, they do contribute to the constitution of a heterogeneous somatic mosaicism (Fig. 3).

In human cancer, when a clonal alteration with a selective growth advantage appears, a population with a specific genotype, differing slightly from the primordial genotype, will emerge (Fig. 3). Furthermore, during the transforming process, the mutation rate will continue with the same frequency or even accelerate if repair systems are impaired. In the pregenomic era, this heterogeneity could escape detection by being located in regions of the genome not screened for mutation and/or present in only a few percent of tumors cells, thus not picked up by global genome analyses (Fig. 3). Today, the ability to detect tumor heterogeneity and reconstruct the evolution of the various molecular events in single tumors via NGS and bioinformatics has shed new light on the importance of passenger mutations. However, we currently have no knowledge as to how many of these mutations reflect in reality somatic mosaicism before transformation [Campbell et al., 2008; Yates and Campbell, 2012]. Despite their apparent lack of clinical value, these mutations deserve attention because they exhibit specific mutational signatures highly related to the type of cancer, which is useful information for tracking cancer mutation etiology [Alexandrov et al., 2013, Kandoth et al., 2013, Lawrence et al., 2013]. Furthermore, it remains possible that some passenger mutations randomly selected in primary tumors will come into play later, for example, in the development of resistance to cancer treatments. From a semantic point of view, these mutations are neither driving mutations as they do not participate in transformation, nor passenger mutations as the may be ticking bombs waiting for specific events to occur [McFarland et al., 2013].

Passenger mutations can be found in coding and noncoding regions of the genome. They may also be difficult to distinguish from driving mutations, although this distinction is essential for obtaining an accurate picture of the cancer genome. For coding regions, several statistical approaches have been developed to solve this problem, for example, the comparison of observed and expected ratios of synonymous versus nonsynonymous variants. Alternatively, various bioinformatics methods can be used to indicate whether an amino-acid substitution is likely to damage protein function on the basis of either conservation through species or whether or not the amino-acid change is conservative in combination with other information including structural information obtained from the 3D-structure/homology models [Grantham, 1974; Adzhubei et al., 2010; Ng and Henikoff, 2006; Davydov et al., 2010]. The classification and pathological assessment of SNV in ncRNA genes will be far more complex than it is in coding regions and require better knowledge of ncRNA function and an accurate evaluation of the impact of mutations affecting them [Cooper and Shendure, 2011]. Novel sophisticated bioinformatics tools and carefully curated databases will be necessary to achieve this goal [Khurana et al., 2013; Lawrence et al., 2013]. The complexity of the mutational landscape and the expansion of selected targets to large, unexplored regions of the genome will be important challenges in cancer genetics. The NIH have recently launched several application calls for research projects focused on developing computational approaches for interpreting sequence variants found in the nonprotein-coding regions of the human genome (http://grants.nih.gov/grants/guide/rfa-files/RFA-HG-13--013.html)

Discovery and Publication of Cancer Gene Mutations: The TP53 Paradigm

Upon the discovery of a novel cancer gene, a unique three-phase pattern is observable in publication, that is, a discovery phase, a consolidation phase, and an application phase. The length of the phases, individually and collectively, depends on the popularity of the gene, the type of alteration and its clinical relevance (Fig. 4A and B). During the discovery phase, the publications describe precisely novel mutations and discuss their potential pathogenicity in relation to the disease. A burst of studies then leads to the identification of novel mutants and their diversity rises quickly (Fig. 4A and B). This phase is commonly associated with reports published in journals with a high impact factor and parallels the rate of mutant or clinical novelties. Transition to the consolidation phase occurs quickly when genetic and clinical data become redundant. During this phase, the number of reported new mutants will decrease and the sequencing of multiple new clinical specimens will discover mostly previously described mutants, leading to a plateau in mutant novelty (Fig. 4A and B). This consolidation phase is vital as it adds nuance to and validates data from the discovery phase in a wide variety of clinical or geographical settings. Consequently, mutations are either described in supplementary materials or quoted as unpublished data, leading to a decrease in reported mutations. Except for a few very specific cases, the consolidation phase is accompanied by a decrease in the impact factor of the publishing journals. This decrease in descriptions of mutations does not reflect their frequency in the disease or the incidence of their analysis but rather a lack of interest and/or utility in their publication. If the mutation has no clinical value, the number of studies will drop quickly then stop. It has also been observed that the consolidation phase is associated with an increase in inconsistent studies. An extensive analysis of the various flaws associated with the publication of mutations is provided by Kern and Winter in their 2006 review [Kern and Winter, 2006].

Finally, for (Fig. 4A) mutations with clinical value, a long application phase then begins. However, publications fall off as service laboratories do not consider reporting to be an essential part of their work and descriptions of novel mutations become scarce.

This model is well illustrated by an analysis of the TP53 gene mutation database and the frequency of publications reporting TP53 mutations (Fig. 4C). The discovery phase began in 1989 with the first description of TP53 mutations in lung and colorectal cancer [Baker et al., 1989; Takahashi et al., 1989]. Over the following years, there was a constant increase in publications describing novel TP53 alterations in various types of cancer (http://p53.fr). Thus, culminated in 2001 with more than 2,000 mutations reported in 300 publications [Soussi and Beroud, 2001; Soussi et al., 2006]. More than 85% of the mutant TP53 listed in the database was identified during the discovery phase. The decline in published TP53 mutations began in 2000, corresponding to the beginning of the second, consolidation phase. This decline was because of the difficulty of publishing TP53 mutations in peer-reviewed journals as their novelty wore off. In recent publications, TP53 mutations are not described because of journal space considerations. This trend toward the nonreporting of mutations is not specific to TP53; it applies to most cancer genes and raises important issues for the evolution of cancer LSDBs as they rely predominantly on published materials.

Whole tumor genome sequencing is currently accelerating the discovery of new cancer genes. Since the pioneering discovery of BRAF (MIM #164757) mutations in melanoma by Davies et al. (2002), numerous genes have been shown to carry mutations in various types of cancer [Davies et al., 2002]. For many of these genes, the discovery phase is just getting underway and several questions, such as (1) the frequency of the mutations, (2) the various types of cancer targeted by the alteration, and (3) their clinical relevance, are still awaiting answers.

These studies shed light on genes and pathways that were not previously under extensive analysis. This is best exemplified by the discovery of isocitrate dehydrogenase 1 and 2 (IDH1, MIM #14700 and IDH2, MIM #147650) mutations in gliomas [Parsons et al., 2008]. These enzymes are part of a multienzymatic complex localized in either the cytoplasm (IDH1) or the mitochondria (IDH2) and participate in the Krebs cycle, an essential metabolic pathway. IDH1 mutations are rare in primary glioblastoma multiforme but common in secondary glioblastoma multiforme indicating that they could be useful markers for glioblastoma stratification. A better prognosis for patients with IDH1 mutations has been observed in several studies but its clinical utility has not yet been established [Gupta et al., 2011]. The frequency of IDH1 mutation in other types of cancer must be confirmed but novel IDH1 mutants have not been identified as the majority of mutational events are restricted to a few codons. Analysis of IDH1 and IDH2 is now in the consolidation phase where the utility of IDH mutations as clinical biomarkers will be studied. Furthermore, other genes are currently at the beginning of the discovery phase. For example, the recent discovery of TRAPP and GRIN2 mutations in melanoma or GATA3 (MIM #131320) in breast cancer will open several new fields of investigation as these genes were not previously associated with cancer [Wei et al., 2011; Banerji et al., 2012].

The pace of the discovery phase accelerated rapidly with the release of the sequences of several thousand cancer genomes. These sequence analyses not only confirmed the participation of the usual culprits, such as KRAS (MIM #190070), PIKC3A (MIM #171834), or TP53, but also led to the discovery of a multitude of new suspects, with furthermore enough evidence to identify some of them as true driver genes. The genes ARID1A (associated with chromatin modeling, MIM #603024) and GATA3 (associated with differentiation) are among the few to have been validated to date; others, with lower frequencies of mutation, will need more research.

As discussed by Wood et al. (2007), genes with high frequencies of mutation (gene “mountains” according to those authors) should be easy to identify but will be more so the exception rather than the rule. Inversely, genes with low frequencies of mutation (gene “hills”) will be far more difficult to distinguish from neutral passenger mutations. Furthermore, because of the stochastic nature of mutations, the multiplicity of the various pathways, and the enormous cross-talk between them, we can anticipate that some cancer genes will be mutated very infrequently. As recurrence is a strong criterion for inferring the relevance of a mutation in cancer, a novel mission, named “the 10K project,” is being launched to sequence 10,000 tumors per cancer type with the goal of uncovering very rare cancer genes with sufficient statistical power (http://news.sciencemag.org/biology/2013/03/ready-more-10000-cancer-genomes-projects). With the forthcoming release of the next-generation sequencer, the increasing power and accuracy of computational tools and the completion of the various large-scale cancer genome projects, it is not unrealistic to predict that the majority of cancer genes targeted by small alterations will be identified within the next decade. As these works progress, the accessibility of multiple cancer genome sequences will increase tremendously, and the discovery phase for any newly detected cancer genes will be shortened as their status in other cancers will be quickly determined by data mining. Ultimately, after completion of these sequencing projects, we will reach a plateau where the majority of gene regions involved in cancer will be identified, including a catalog of the SNVs associated with each one. Thus, a full molecular portrait of various types of human cancer will be available. Consequently, medical research will enter a new era where the emphasis is placed on clinical studies and finding the relevance of these data to improve patient care.

The Future of Cancer LSDBs

The explosion of information described in the previous sections will have profound consequences on LSDBs. To date, the majority of cancer gene LSDBs are simple lists of mutations that are very difficult to search. They are also highly heterogeneous in terms of quality, content, and format. Several database management systems, such as the Universal Mutation Database, the Leiden Open Variation Database, and the MUTbase have been developed to standardize the current data via a framework for LSDB curation but these systems are not intercompatible [Auerbach et al., 2011]. The human Genome Variation Society has published several guidelines on database structure and content or on variant nomenclature but with little in-the-field success [Kohonen-Corish et al., 2010]. Several recent surveys noted that less than half of the current LSDBs would meet minimal criteria for ease of use. Furthermore, the international nomenclature for publishing DNA variants (http://www.hgvs.org/mutnomen/recs-DNA.html) is used in less than 20% of publications, leading to data that are often useless for inclusion in LSDBs (Soussi, unpublished observations).

Most of the present LSDBs followed a similar development pathway. The data included in them were generally derived from publications, personal results from the curators or unpublished results from a consortium of specialists (Fig. 5).

LSDBs are maintained by one or several curators who are specialists in the field and thus benefit from strong scientific expertise. Data mining for LSDBs was historically performed manually and was task intensive. To ease the process, several procedures were established. First, algorithms were developed to identify publications that describe variants associated with a specific gene name. Second, tools for extracting mutation data from publications automatically were developed to circumvent error-prone manual entry of mutations in the database. However, these procedures function only with articles that respect the international mutation nomenclature and furthermore they do not distinguish tables that describe mutations in more than one gene. Maintenance is another problem for these LSDBs: more than 50% of LSDBs received regular maintenance and updates for only a few years after their launch. Indeed, maintenance of single LSDBs is time consuming, not rewarding, and poorly funded.

With the release of whole cancer genome sequences, this development pathway will face obsolescence as data are directly included in centralized databases such as those maintained by the Welcome Trust Sanger Institute (COSMIC database, http://cancer. sanger.ac.uk/cancergenome/projects/cosmic/), TCGA, http://cancergenome.nih.gov/), or the ICGC (http://dcc.icgc.org/). Although mutation data currently remain available and extractable from supplementary materials, this source will soon run dry, as all data will be mined from centralized databases in a near future. The aggregation of all cancer gene mutations in individual databases has multiple advantages, but how the data are organized in any one database will dictate the type of studies that can be realized based on it. The TCGA database provides a patient- and tumor type-centric organization that allows multigene analysis in individual patients. Furthermore, multipatient analyses will allow for the determination of statistically significant co-occurrence or co-exclusion of mutations situated in different genes, an important feature for analyzing the specific pathways targeted by these mutations in a specific type or subtype of cancer. The ICGC and TCGA databases only include data obtained via the large sequencing programs of their parent consortiums (Fig. 5). The COSMIC database provides a gene-centric organization providing an accurate description of the various mutational events that target each cancer gene in different types of cancer. Its data are harvested from the literature and are not restricted to specific studies or projects. In contrast, this database is less opportune for tumor-oriented analysis and an absence of literature curation may permit the presence of spurious data.

We are experiencing a transition period where it is difficult for biocuration efforts to keep pace with mutation discovery, leading to large uncertainties in the information included in the various databases [Howe et al., 2008; Meyerson et al., 2010; Sanderson, 2011]. Overlap between the databases is not yet optimal and a single published study can lead to different sets of data in the three databases due to heterogeneous curation procedures. This can be illustrated by an analysis of mutations in several genes such as TTN (MIM #188840) and MUC16 (MIM #606154), which encode the longest proteins of the human body. Mutations in TTN and MUC16 have been reported in multiple tumor types and the COSMIC database reported 1,500 and 1,000 mutations for these two genes, respectively. The large size of these two genes and the lack of direct connections of their functions to cell transformation were highly suggestive of passenger mutations. The development of novel algorithms to infer the significance of cancer gene mutations has led to the discovery of multiple false-positive genes including TTN and MUC16 (Fig. 6) [Lawrence et al., 2013]. Mutational data also fluctuate depending on the algorithm used for variant calling. In several TCGA studies, the status of TP53 has been found to be heterogeneous depending on the database and the procedure used for the analysis (Soussi, unpublished observation, see also Leroy et al. in the same issue).

These uncertainties related to centralized databases, due to global automation and the absence of gene specific expertise, are far more important than those found in LSDBs. However, these issues are innate to the current, unstable transition period, which will surely pass notably as curation algorithms are developed to create accurate mutation databases. The MutSig algorithm of the Broad Institute provides a good example of how quickly algorithms can evolve (Fig. 6).

The most important question to be addressed is whether there is a future for independent cancer LSDBs. In the past, LSDBs have been tremendously useful. They benefit from rigorous expert curation, often coordinated by collaborating researchers with scientific expertise. Impaired gene functions caused by deleterious mutations and associated with specific phenotypes have generated working hypotheses and led to multiple lines of study. The four highly conserved domains of TP53 were identified as early as 1987 and the hot spot mutations within them in 1989. However, the DNA binding activity of this core (Soussi et al., 1987; Baker et al., 1989; Takahashi et al., 1989; Kern et al., 1991) region of the protein was not discovered until 1991.

Notch mutations are localized in various subdomains of the protein in different types of cancer, generating promising leads on the diversity of this pathway (Guruharsha et al., 2012). LSDBs have also been used for the development of diagnostic tools targeting the most significant mutations in one or several genes with clinical potential. Ultimately, they provided lists of potential targets for therapeutic development. Several LSDBs also include expert-curated, gene-specific information. The TP53 mutation database includes functional and structural information that led to the accurate classification of the loss of function of the various TP53 variants, information that is more specific than that provided by global algorithms such as SIFT or MUTAssessor (Reva et al., 2011; Sim et al., 2012). Curated mutations in LSDBs are also used as training or test sets in the development of novel biocomputing tools used for the stratification of driver and passenger mutations. The TP53 database was recently curated using specific statistical tools, resulting in the removal of more than 100 articles reporting artifactual data [Edlund et al., 2012] and Leroy et al. in the same issue. This curated database was used for the development of CHASM, a program for estimating the impact of missense mutations [Cline and Karchin, 2011].

Although cancer gene LSDBs contain mostly somatic mutations, germline mutations in several tumor suppressor genes associated with cancer predisposition are also available. Indeed, mostly germline mutations have been described for several genes, for example, BRCA1 or BRCA2 (MIM #600185) associated with breast and ovarian cancer, or MSH2 (MIM #609309) and MLH1 (MIM #120436) associated with hereditary nonpolyposis colorectal cancer, whereas both germline and somatic mutations are available for other genes, for example, APC (MIM #611731) associated with familial adenomatous polyposis or TP53 associated with Li–Fraumeni syndrome (see the review of Kamihara et al. in this issue for a full discussion on TP53 germline mutations).

Although the pathogenicity of somatic mutations in cancer genes can be easily appraised, assessing the pathogenicity of germline mutations is far more difficult. Providing patient counseling is critical in the presence of a suspected hereditary syndrome. LSDBs have been invaluable as tools for classifying these variants and as a reference for clinical geneticist.

We can nonetheless predict that LSDBs, in their current form, will disappear very quickly. The reasons for this include the lack of mutation descriptions in publications, information redundancy with large databases and an absence of funding. Indeed, since the launch of the various large tumor-sequencing projects, LSDBs on novel cancer genes are no longer being developed. Furthermore, cancer cannot be reduced to the simple presence of an SNV; other genetic and epigenetic events participate actively in the tumorigenesis process. The versatility of novel NGS platforms allows for multiple types of analyses, including sequencing, copy number evaluation, translocation detection, methylation profiling, and expression profiling. By combining these analyses, researchers will be able to draw an integrated picture of each tumor and pinpoint relationships (or absence of relationships) between various types of alterations and specific pathways. The TCGA and ICGC databases are informationally complete and can be browsed via specific Web portals.

Nevertheless, a curated repository for each cancer gene must be available so that the scientific community has access to the same, accurate, bulk information previously supplied by LSDBs. As the number of true cancer genes will most likely be limited, it is not unimaginable that administrators of large cancer databases assign a group of curators/experts to each gene of importance. Specific information relevant to individual genes could be added to the database and made available to the community via specialized but easily accessible gene-specific interfaces. For example, The TP53 database includes more than 100,000 entries concerning functional data (e.g., transactivation, DNA binding, gain of function) for 2,000 TP53 variants (Leroy et al., 2013). In the APC tumor suppressor gene, the distribution of its biallelic nsSNV events is not random and knowledge of this distribution will be important for explaining the mutation spectra observed in colorectal cancer (Fearnhead et al., 2001). Furthermore, the localization of mutations in the APC gene is highly correlated with disease severity and association with extracolonic features.

Ultimately, what should be the most desirable output? For a clinical oncologist dealing with patients or families on a daily basis, reports of genetic modifications must be linked to the full description of their biological consequences so that the information can be used with strong confidence. The number of allelic variants of unknown significance, a nightmare for clinical geneticists, must be reduced to lessen the burden for both clinicians and families. The integrated association of the multiple mutations found in a tumor and the genetic background of the patient will help to define the most effective therapy.

For basic research, as discussed above, disease-associated mutations are of inestimable value. Gene deconstruction via disease-associated mutations can be likened to evolution but it shapes the functionality of the gene product to another purpose, such as hyperactivity, partial or total loss of activity or the gain of a new function. The propensity of a specific nucleotide to be selected during the neoplastic transformation is related to its mutability depending not only on the DNA sequence context, but also and more importantly on the consequences of the alteration at the RNA and/or protein level. Therefore, each residue of a protein can be associated with a decomposition factor in contrast to evolutionary conservation and cancer hot spot mutation residues associated with a high decomposition factor are frequently localized at phylogenetically conserved positions. The value of cold spots for mutation, that is, residues highly conserved phylogenetically and never selected in cancer, have been largely underestimated. In the TP53 gene, only a few conserved residues of the protein are never found to be mutated in human cancer. It appears that these residues control the interaction of the negative regulator mdm2 with TP53, and any alteration would be lethal for the cell (see Leroy et al. in this issue for more information). This information gained via the analysis of the mutation database in association with a phylogenetic analysis, the construction of artificial mutants and structural analysis of the TP53 protein, demonstrates to what extent information dispersed in various databases or publications can raise pertinent questions.

Connecting all the information stored across various databases in such a way as to make it accessible via a single output interface will be one of the most important challenges to meet in the coming decades. Indeed such an orchestrated approach will help calm the sea of big data and thus allow the scientific community to efficiently use this information in its quest to understand and treat cancer.