Authors of initial version

Alvaro N.A.Monteiro¹, Gerhard A. Coetzee², Matthew L. Freedman³, Mariella De Biasi⁴, Graham Casey⁵, Dave Duggan⁶, Angela Risch⁷, Christoph Plass⁷, Pengyuan Liu⁸, Michael James⁸, Haris G. Vikis⁸, Jay W. Tichelaar⁸, Ming You⁸, Simon A. Gayther⁵, Ian G. Mills⁹ on behalf of functional cancer genomics supported by the NIH Post-Genome Wide Association Initiative

1. Risk Assessment, Detection, and Intervention Program, H.. Lee Moffitt Cancer Center and Research Institute, Tampa, Florida, 33612
2. Department of Preventive Medicine and Urology, Norris Cancer Center, University of Southern California, Los Angeles, CA 90033
3. The Eli and Edythe L. Broad Institute of MIT and Harvard, Cambridge MA 02142 and Department of Medical Oncology, Dana-Farber Cancer Institute, Boston MA 02115
4. Department of Neuroscience, Baylor College of Medicine, Houston, TX 77030
5. Department of Preventive Medicine, University of Southern California, Los Angeles, CA, USA
6. Translational Genomics Research Institute (TGen), Phoenix
7. German Cancer Research Center, Division of Epigenomics and Cancer Risk Factors, Heidelberg, Germany
8. Washington University, St. Louis, Missouri
9. Centre for Molecular Medicine (Norway), Nordic EMBL Partnership, University of Oslo, Blindern N-0317 Oslo, Norway

[[ Dear all, thank you for the really great job! We’ve hit now the deadline of December 20th and also a technical limit of 80.000 characters ;-) The ball is now with the original authors and the editor ... We’ll keep you posted on the discussion page   ]]

Introduction

The last few years have seen an explosion of activity in the identification of common, low-penetrance susceptibility alleles for a range of complex diseases and other traits using genome-wide association studies (GWAS) [1]. As of June 2010 , 904 genome-wide associations (at p<5x10^-8) for 165 traits have been published [2]. Of these, 193 loci were associated with modified risk of 11 different cancer types [2]. Even though some risk-associated loci have been successfully identified [3][4][5] , these results are rarely followed by studies  I prefer: frequently these results are not followed by studies ... However, I don't know what sounds better in English  on the molecular basis of risk etiology or on the identification of the causal variant [6] . A recent Nature Genetics editorial began to put this problem in sharper focus and suggested that there should be a more significant investment in functional characterization of identified risk loci [7]. Several factors might discourage the investment of such resources. Firstly the effects are so weak that many GWA SNP discoveries (e.g. in type 2 diabetes) increase the sibling recurrent risk ratio by less than 1 %, and there is little connection between the results of an abundance of earlier family based genetic linkage results and the current GWA results. Systematic comparison between the genome-wide linkage and association results has shown that the extent of genomic convergence between these two lines of hypothesis-free evidence is not beyond what is expected by chance for the majority of complex traits examined to date [8]. For example the strong cluster of positive genetic linkage results around 1q23 for type 2 diabetes has no corresponding signal in multiple GWA studies [9], whereas the replicated association between Crohn's disease and the IL23R gene was not supported by previous linkage data [8]. In the present paper (? - I am referring to as the paper itself . Answer: is it better now?) , we introduce the term post-GWAS to refer to the analysis that follows the identification of a risk loci .  see discussion, this is one of the first time the term Post-GWAS is used in the literature and we should define it properly to avoid future confusion .  Importantly, the development of reliable biomarkers and effective preventive and therapeutic agents made possible by these discoveries is founded on a detailed understanding of biological function. The editorial made a distinction between two types of functional analysis. The first comprises a preliminary investigation that includes re-sequencing, association analysis of all variants within the region of linkage disequilibrium (LD), investigation of risk-associated SNPs as modifiers of monogenic traits, genomic analysis of gene expression in human tissues, screens for somatic mutations on risk haplotypes, and epigenetic analysis of human tissues in the regions of the GWAS SNPs. Similar guidelines have been proposed elsewhere, in the context of loci that underlie complex traits [10]. There is a second stage of functional analysis geared towards understanding the biological mechanism of risk enhancement and causality. Whereas in-depth analysis cannot be the subject of systematic evaluation due to the functional diversity across loci, the identification of general hypotheses that can be tested in a systematic way will improve and accelerate the analysis  is accelerated analysis the problem or rather the lack of functional validation of GWAS outcomes ? . Hence, in this paper we propose principles for the initial functional characterization of cancer  (disease ?; limited to cancer)  risk loci to bridge this information gap. Several challenges lie ahead in assigning functionality to susceptibility SNPs. For example, most effect sizes are small relative to those seen in monogenic diseases, with per allele odds ratios usually ranging from 1.15 to 1.3, despite an occasional outlier such as KITLG in testicular germ cell cancer (OR=3.08) [11][12].  I have trouble with this argument since it gives me the impression that what is sought for through GWAS is diseases with a monogenic-like determination. This kind of diseases are very rare and thus, constitute much more an exception rather than a rule . Answer: The paper assumes that you have already identified an association between a variant and a disease. Then, in this paragraph they say that characterizing the function of a single snp is more difficult than the case of a single gene in a monogenic disease  Thus, the functional effects of SNPs are likely to be subtle. It is unclear whether current molecular biology methods have enough resolution to differentiate such moderate effects. In addition, we anticipate that it will be difficult to address function for biological effects that are non-cell autonomous, are specific to certain developmental stages, or act at a site distant from the tissue-of-origin of the cancer. The study of such effects might benefit from in vivo models. An ultimate goal for the comprehension of a specific disease is to link genetic variation to causation : this is currently beyond the capability of modern molecular and logistical approaches, when it comes to analyzing complex syndromic  (syndrome?)  studies. Our aim here is , therefore, to provide a set of recommendations to optimize the allocation of efforts and resources in order to maximize the chances of elucidating the functional contribution of specific loci to a phenotype associated with a specific disease. Since it has been estimated that 88% of currently identified disease- and trait-associated SNPs are intronic or intergenic [4] , in this paper we will focus on the analysis of non-coding variants and outline a hierarchical approach for post-GWAS functional studies. Connecting a risk allele to a target gene(s) is particularly tractable intermediate point in this work. The unifying hypotheses that are applicable across these studies are:

• There is a transcript , coding or not, that has not yet been annotated, and associates with a risk locus.
• The risk locus is a regulatory element affecting the expression of one or more annotated transcribed regions of the genome.

Defining regional boundaries and the assembly of layered genomic data

The general approach underlying current GWAS is to identify disease association through surrogate SNP markers that capture linkage disequilibrium (LD) relationships across the genome. Taking this approach means that any GWAS data generated using commercial SNP arrays are limited by the depth of the genomic coverage and representation of LD structure on those arrays. Unfortunately, recent data suggest that only 60% or less of the common SNP information (>5 %), and less than 50% of variants with a frequency higher than 2.5% are found on SNP arrays used in the majority of GWAS conducted to date, when compared to the 1000 Genomes Project¹ [13]. While a new generation of SNP arrays have been developed and addresses this issue of genome coverage (at least for European populations), the degree , or lack thereof , of common SNP content in GWAS studies has important implications not only for missing “dark matter” signals, but also for subsequent post GWAS functional characterization studies of known associations. Moreover, in the vast majority of cases any association identified through GWAS would be predicted to be between a surrogate marker (e.g . tagSNP) and the disease trait, rather than a surrogate marker and a causal variant, as SNP arrays were designed using surrogates chosen to capture LD structure on these arrays rather than for any functional reasons. Therefore, evaluation of all common SNPs across associated regions will be desirable to fully characterize the biological implications of disease associations. The clear implication is that it is not the SNP with the low p-value that is of interest but the genomic region surrounding it. Unfortunately, current HapMap information is incomplete with respect to identifying all common variant information and may not capture the genetic diversity of the populations being used in association studies. The ongoing 1000 Genomes Project seeks to address this missing data and will ultimately capture the majority of SNPs with an allele frequency of >1% across the genome [13], including intergenic regions and gene deserts where the majority of associations have been mapped. It is important to note that the recently published pilot data from the 1000 Genomes Project [13] captures the majority of common variants (>5% allele frequency) but also many less common or rare variants , and this resolution will improve strongly with the next data releases. Due to the used resequencing approach and the higher number of individuals per population, the 1000genomes project will provide information of genetic variants with much better resolution than the widely used HapMapII project   [14] . There is a growing interest in the role of multiple rare variants in disease risk, which will require targeted or whole genome sequencing of 1,000s to 10,000s of samples. An examination of the large number of disease causing allelic variants in well known genes (e.g. BRCA1 or CFTR,   ) makes it clear that complex diseases will not only be due to the multiple genes now under scrutiny, but also multiple variants within those genes, each variant with its own genealogy and genetic history. Increasing the sample size to 4000 and subjecting them to genome-wide sequencing, as is being envisaged in the UK10K project  ), will maximize the chances in detection in variation. Moreover, reducing cost, faster CPU processing power and increasing speed of sequencing will allow allele frequency as low as 0. 01 to be detected .

While genome-wide resequencing of thousands of cases and controls currently is not feasible for the vast majority of GWA studies and again, datasets the 1000genomes project will likely prove useful to alleviate this disadvantage. Interestingly, the information from cancer GWA studies can be augmented by genotype imputation based on reference haplotypes [15]. The striking advantages of this bioinformatic approach are (i) the potential identification of additional or more strongly cancer-associated SNPs and (ii) the identification of essentially all potential functional variants that are in strong LD with the originally genotyped SNPs. By using information from GWA studies in order to prioritize SNPs and target genes, such imputation methods are in a good position to respond to some of the challenges in the post-GWA studies. Furthermore, these methods allow for re-analysis of existing GWA datasets and meta-analysis across [16] different genotyping platforms.

Apart from the imperfect interrogation of rare variants, the first wave of GWAS largely discarded other important sources of variation, such as structural variants, noncoding RNAs or epigenetic changes [17]. The most common types of such variants are gains and losses of DNA, called copy number variants (CNVs), which likely exert important phenotypic effects on gene expression and function. Recently developed genotyping chips allow the simultaneous interrogation of SNPs and CNVs across the genome with improved coverage [18]. The genome-wide detection of rare CNVs, as well as the discrimination capacity for variants with wide range of variability in copy number, represent important challenges for genotyping technology research. Since the global inventory of CNVs remains incomplete, and with limited empirical data currently available, the extent to which current arrays capture the structural variome and enable a more comprehensive elucidation of the spectrum of genomic variability remains unclear [19]. Even under optimal design, genome-wide SNP screens will still not be informative about the role of other components of the regulatory architecture of the genome, such as noncoding RNAs or epigenetic modifications (somatically-acquired or trans-generationally inherited); complementary techniques, such as high-throughput methylome analysis, will be ultimately needed [17].

Computational methods for SNP prioritization, Databases and approaches to predict function

 (Note: I am not able to post two sections that we (Khader Shameer, Iftikhar J Kullo Kullo J Iftikhar) would like to add; we are getting "“Text exceeds maximum length " error. Any idea how can I add the text in main article)  

Different methods are available to identify candidates for the causal variant, to predict its function and to infer the biological mechanism behind the association. Computational approaches have the advantage of being quicker and cheaper to run, making them widely available ; moreover, a computational pipeline can be automatized and applied to several identified variants , and makes it easier to compare results from different studies. A successful computational prediction can save time and simplify the design of the subsequent experimental steps, but should be always accompanied by an experimental validation of the result. Samples of the types of methods and tools that may be used to accomplish these investigations are provided in Table 2.

As described in the previous section, the first step after having identified a tagSNP associated with cancer or another congenital disease is to identify the real causal SNP or at least to predict its position or the gene involved in the association. Two computational techniques developed for this task are generally known as SNP prioritization and digital candidate gene approach [23]. SNP prioritization methods aim at predicting the known SNP variants which are the most likely to be involved with the phenotype; generally, these approaches rely on comparison with microarray data [24], or the incorporation of information from different databases [25]. Candidate Gene Approaches also rely on the integration of different sources, such as expression, known association with other phenotypes, biological pathways, etc.. [26] [27] [28] [27] [28] [29]  this paragraph is being written with the help of Emre Guney from the Pompeu Fabra University 

After having identified the association of a SNP or a variant with a disease, the first approach to determine its function is to consult the existing databases to determine whether other information are known about the SNP, or whether the SNP is known to be associated with other diseases or phenotypes. Useful resources for general information on SNPs and variants are listed in Table 2. General genome browsing tools can be employed to explore the context in which a variant is found , which can be used as a first step to delimit the potential functional effect of the SNP (Figure 1). Layers of annotation including known and predicted genes, promoters and other regulatory aspects, and comparative genomics can all be investigated to elucidate features. Identification and comparative analysis of patients suffering from same type of drug reaction could also be a useful tool in identifying SNPs [30]. Cancer-specific resources may offer insights on the presence or absence of variations in normal and/or tumor genomes. Finally, an innovative approach to the analysis of the literature on a region of a SNP is provided by GRAIL [31].

 [note: Next two paragraphs and additions to table2 contributed by Catherine L. Worth, Tammy M Cheng, Sungsam Gong, Tom L Blundell Department of Biochemistry, Tennis Court Road, Cambridge, University of Cambridge, UK; David F Burke Department of Zoology, Downing St, Cambridge, University of Cambridge, UK dfb21@cam.ac.uk] 

After having collected the information already annotated about the SNP identified, further analysis and predictions can be made. Most non-synonymous SNPs (nsSNPs) have little or no effect on the function of the proteins for which they code, and appear to be selectively neutral. However, a few SNP are regulatory in nature and may affect function directly by altering protein-binding sites, or indirectly by altering protein stability, aggregation or post-translational modifications. The sheer volume of SNP data generated by GWAS means that computational methods are useful for prioritizing SNPs for experimental characterization. In the case of variants falling within coding regions, it is possible to predict their effect on the protein structure [32]; a good example of an automated pipeline is described in Burke et al. [33], who presented a framework to predict the effects of known nsSNPs . They analysed over 6000 genes carrying almost 24,000 nsSNPs (2,249 disease-associated nsSNPs from OMIM and 21,471 nsSNPs from dbSNP). In order to increase the coverage of structural information, the authors implemented an automated, large-scale, structure prediction strategy using homology modeling, building models for almost 90% of the genes. These models were then used to generate predictions for 40% of the nsSNPs (using methods that depend on analyzing amino acid substitutions in orthologous families in terms of their local structural environments in the protein structure [34]  . The observed nsSNPs were analysed for affects on protein stability by several methods including SDM [35]   and for affects on protein interactions by Crescendo [36]. Structural databases that identify protein-protein (PICCOLO:   ), protein-nucleic acid ([37] BIPA:   ) and protein-ligand ([38] CREDO:   ) interactions along with a newer analysis tool ( [39],  ) and web interface (SAMUL:  ) in which users can browse structural and functional annotations of proteins at their residue level , have since been used to augment these predictions . Further work by the group demonstrated that combining the predictions from different structure-based methods increases the sensitivity of predictions whilst maintaining good specificity [40]. This methodology has been used to generate new hypotheses regarding the molecular etiology of renal cell carcinoma and pheochromocytoma in the cancer syndrome, von Hippel-Lindau disease ([41].Software that is currently available for testing the effect of nsSNPs in silico is described in Table 2.

Regarding the variants falling within intergenic regions, many approaches are available, depending on the position of the variant identified and the hypothesis of the researcher. A gene prediction tool can be employed to determine whether the region involved is an un-annotated gene or a pseudogene [42] and other Genomic Functional elements, such as splicing regulators, transcription factor binding sites and otehr (other) regulatory elements can be predicted as well [43].  (Methods for predicting regulatory SNP in silico have been developed. Manke, T et al. Hum.Mut. 2010; MacIntyre G et al. Bioinformatics 2010).  Thus, several bioinformatics methods and tools have been developed for the analysis of putative promoters (Eponine [44], McPromoter [45], PromoterScan [46]), transcription factor binding sites (Match/Transfac [47], TFSEARCH  ), and both exonic splicing enhancers (ESEFinder [48], RESCUE-ESE [49]) and silencers [50]. Additionally, suites that include different prediction tools for both coding and regulatory SNP have been developed ([51], [52]), providing the user with an integrated interface that allows an easier and more comprehensive analysis. However, although the prediction of functional effects can help to prioritize SNPs for downstream analysis, it is often found that associations from individual markers show a weak effect because of the complex interactions among genes. It is likely that the combined use of experimental genotyping results and biological information on gene interactions (protein-protein interactions, transcriptional interactions, etc) will improve the power to detect associations of markers to complex traits. Hence, pathway-based approaches constitute a powerful tool to further analyze the results of a GWAS, not only to characterize the function of the variants for which an association is identified but also to possibly discover the effect of other variants with minor effects [53] [54] [55] [56] [57]. In any case, the annotation of biologically relevant pathways in scientific databases like KEGG/Pathways[58], Metacyc [59], Reactome [60] and the Pathway Interaction Database [61] makes it possible to understand the function of the genes in the nearby of the variant identified and to formulate hypothesis on the etiology of the disease .  the following paragraph has been moved here from 'Genetics of Gene Expression '  A recent elegant report  which report? citation needed  demonstrates that network analysis using risk variant and gene expression data is proving to be a powerful and fruitful tool in the dissection of the pathways, driving the pathogenesis of the disease. This ranges from transcriptomic analysis to predict the regulatory influence of transcription factors over gene networks dependency using tools such as ARACNE (Algorithm for the Reconstruction of Accurate Cellular Networks) [62], through Bayesian network approaches to identify predictive relationships between genes from a combination of expression and eQTL data [63]. Whilst these tools are elegant, the ability to translate their outputs into biological significance is heavily dependent on the availability of easy to handle and relevant model systems to test the predicted connectivity. These approaches clearly pose validation challenges for many diseases.

Section on "Integration of tools for enhanced annotation and analysis of variants" and "Visualization of Risk Variants" added to discussion page  Genetics of Gene Expression

 note: include these papers [64] [65]  Having a defined target region, and identified all common variants (frequency > 1-5 %, depending on sequencing depth) within a risk locus by re-sequencing and/or in silico analysis, the next step to achieve functional characterization is to explore associations between statistically significant SNPs and gene expression. Such associations offer a thread from which to build functional validation, although they do not necessarily mean that the genes cause the clinical trait. Gene expression is a heritable trait. Transcript abundance varies in the human population (similar to height and blood pressure) and thus can be considered a trait that is amenable to genetic mapping. A number of landmark studies have unequivocally demonstrated that a substantial fraction of transcripts in the human genome are influenced by inherited variation [66][67][68][69] [70][71]. Genetic variants affecting transcript levels are often referred to as expression quantitative trait loci (eQTLs). eQTLs can be located near the gene they regulate or far away and thus have a local or a more distant effect . The distinction between the two is often arbitrary; however, for most studies local eQTLs have usually been defined as being within 1 mega base from the variant under consideration. Distant can involve interactions between an eQTL and a gene located in different non-homologous chromosomes. As has been previously pointed out , we prefer the terminology of local and distant rather than cis- and trans- which implies a mechanism [72]. Certain principles have emerged from these studies: (i) eQTLs tend to explain a greater proportion of trait variance than is typically seen for risk alleles and clinical traits; this observation translates into eQTLs and gene expression traits that can be discovered with smaller sample sizes than association studies between inherited variation and clinical traits (such as disease risk ); (ii) local eQTLs tend to have larger effects on gene expression than distant eQTLs and are , therefore, easier to discover ; (iii) expression phenotypes are primarily regulated by distant eQTLs [73] (comment: I prefr [74] and it says the opposite: local=cis-eQTL) (iv) the majority of local eQTLs in humans is clustered closely around the transcription start site or or just upstream the transcription end site [75].

Genetic variation on the short arm of chromosome 3 is a good example of a local eQTL which affects the expression of the CCR5 gene significantly. This genetic variation, known as the CCR5-∆32, consists of a 32bp deletion in the single coding exon of the gene causes production of a nonfunctional protein. Consequentially, this gives rise to a phenotype in the form of resistance to HIV infection in individuals who are homozygous and slow progression to AIDS in individuals who are heterozygous for this variation [76] [77] . However, closing the gap between genotype and phenotype in complex diseases is not always straightforward. Unlike Mendelian disorders, a large fraction of the associated loci are located outside known protein-coding regions. Both empirical and computational data support the notion that a considerable proportion of these loci will be eQTLs [78][79][80][81] (and Pomerantz et al., PLoS Genetics, in press). The strategy of applying the genetics of gene expression approach offers an appealing and straightforward way to initiate the complicated task of connecting risk variants to their target genes. Importantly , this strategy does not require knowledge of the actual causal allele. Many of the initial successful eQTL studies relied on the employment of patient-derived lymphoblastoid cell lines largely due to their easier availability [82] [79]. Recently, these anlayses have been performed in primary human tissues , and it has been demonstrated that some of the associations are tissue-specific [83][80] (and Pomerantz et al ., PLoS Genetics, in press). Alternative splicing also adds to the complexity of eQTL associations, allowing for cases where only a particular RNA isoform is affected by the genotype (Burd CE et al .  ).

Add  short paragraph on overlap eQTL and GWA signal here (see discussion @Mark McCarthy and comments herein) .

A complementary and powerful approach to defining local eQTLs is to measure allelic imbalance in individuals that are heterozygous for a risk allele. Any transcript demonstrating a deviation from a 1:1 ratio (as typically measured by a transcribed heterozygous marker) becomes a strong candidate gene [84][85][86]. What if the risk allele is not associated with the expression trait? False negatives can occur because gene expression varies in terms of time and space; therefore, the developmental time point and/or the cellular population(s) that it is being examined may not be entirely appropriate. As a rule of thumb, associations between the risk allele and its expression should initially be tested in the cell lineage that is believed to give rise to the tumor type (and subtypes) under study, and contrasted with the expression patterns and associations in other tissues. Effects on transcript abundance may be subtle and therefore below the sensitivity threshold of a particular platform and/or the sample size may not be adequate . Transcript abundance is usually evaluated under steady-state conditions. Lastly, the impact may only be revealed in certain contexts, such as the activation or the inhibition of a particular pathway. In these cases, alternative assays will be required to involve these genes . Future questions to be addressed in this field include: i) what are the appropriate target tissues to examine? Risk alleles may act in a non-cell or -tissue autonomous fashion and therefore may exert their effect through other cell types that act upon the target tissue under consideration. ii) Should the diseased tissue or the normal tissue or both be evaluated? We will touch some of these points later in the article . Finally, iii) RNA-sequencing (RNA-seq) using next generation sequencing platforms has the potential to reveal the transcriptome in its wholeness [87]. The RNA-seq platforms also possess suitable characteristics such as increased sensitivity for low abundance transcripts and a wider dynamic range than microarrays. The challenge shared by both RNA-seq and microarrays are the limited involvement of time points from the data gained. As cancer development is a temporal and wide process, dynamic monitoring risk allele's expression is necessary to validate its roles. Furthermore, as cancer will develop through a multi-step process, it would be very intriguing to detect the different expression levels of a certain risk allele during the shift between two phases of the disease (e.g. the transition between hyperplasia or dysplasia and carcinoma in situ). Critically, it is now becoming evident that data from functional studies heavily rely on compelling statistical inference. How to design and statistically interpret becomes important for the soundness of the results from the data analysis. Furthermore, the interpretation or biological inference often extrapolates beyond the whole metabolic pathway and even the phenotype (cancer or disease per se) involved. But, frequently, the RNA-Seq projects are only based on samples from one or several specific time points, how reliable for the causality related to the cancer or other disease characteristics is the critical concern for the results. From a physiological point of view, genomic functional data may show different activities of relevant alleles within a polymorphic site at different time points. These differences make challenge for reproducibility, and statistically convincing functional results for the candidate gene(s) being analyzed. Failure to observe a functional difference between alleles is also not a reliable indicator because there is very limited data available for the phenotype development. These concerns make dynamically and systematically necessary to monitor the expression of these selected important alleles .

Genetic variation and epigenetic control of gene expression

Promoter methylation, histone tail modifications and altered expression of non-coding RNAs, such as the large intergenic non-coding RNAs (lincRNAs) [88][89] [89] that associate with chromatin modifying complexes, contribute to gene regulation in normal development as well as to aberrant gene expression in tumorigenesis [90]. Furthermore, epigenetic mechanisms play an important role in mediating the influence of the environment on gene expression [91]. Epigenetic silencing has been shown to be the predominant mechanism of gene inactivation for a subset of genes [92], while for other genes genetic, as well as epigenetic, mechanisms have been shown to jointly contribute to tumor suppressor gene activation [93]. An important research direction in the field is to understand the interplay between environmental, genetic and epigenetic factors, both in healthy tissues as well as in the initiation and progression of malignancies. Here we focus specifically on two questions: (1) Do genetic variants alter the epigenetic landscape and in this way they increase the susceptibility to develop cancer? (2) Do genetic variants increase the risk of a locus to become epigenetically silenced in the tumor? Technologies, in particular for the assessment of DNA methylation, are now sufficiently advanced and affordable enough to screen large tissue collections at a single CpG resolution [94]. Methylation profiling within and around haplotype blocks associated with cancer risk thus represents one appropriate strategy. Mechanisms by which genetic variants have the ability to affect epigenetic marks are known from studies in hereditary non-polyposis colorectal cancer. Here constitutional epimutations exist, that predispose to the early onset of a colorectal cancer [95][96][95] : the MSH2 and MLH1 , present in somatic cells including peripheral blood lymphocytes, are examples of highly penetrant epigenetic predisposing factors. For example, mutation of TACSTD1 leads to transcriptional inactivation through promoter methylation of a single MSH2 allele in normal somatic tissues. It has been proposed that many epimutations are a consequence of cis or trans acting genetic variants (reviewed in [97]). In an elegant experiment Kerkel et al. [98] showed a sequence dependence of allele-specific methylation and demonstrated that cis-regulatory polymorphisms control gene expression and affect chromatin states. Recent ChIP-Seq data provides further compelling evidence that SNPs and structural variants frequently coincide with allele specific differences in transcription factor binding sites and chromatin structure. This can now be investigated using allele-specific sequence analysis approaches [99], although the depth of sequencing required for this approach is still an issue [84]. Genome-wide maps of allelic asymmetries are expected to identify functional regulatory polymorphisms [100]. The fact that SNPs can affect allelic imbalance in a tissue specific manner, as shown for UGT2B15 [101], is an important issue in the experimental design. Further epigenetic mechanisms modulating gene expression on transcriptional and post-transcriptional levels involve long non-coding RNA and small regulatory RNA [102] [103]. For example, piRNA have been implicated in de novo DNA methylation and transposon silencing in mice [104], and several oncogenic miRNAs and their binding sites , were shown to be directly be affected by SNPs and CNVs [105]  <the suggested replacement for the deleted Pelletier & Weidhaas 2010 reference would be more readily available to the readers> . Tandem repeats can also impact gene expression , for instance, by altering transcription factor binding sites and also by affecting chromatin structure (reviewed in [106]). Despite the focus above on enhancers and insulators as genotype-specific, long-distance regulators of gene expression, we are also well aware of that chromatin fibres dynamically explore the nuclear space to establish meta-stable, long-range interactions with other chromatin fibres [107]. Although the functional outcomes of such interactions remain largely unchartered, it has been demonstrated that they are capable of transferring epigenetic marks to modulate transcriptional processes both in cis [108] and in trans [109] [110]. Hence, chromosome crosstalk sets the stage for the spreading and propagation of pleiotropic epigenetic effects in all likelihood in a manner reflecting the topology of the involved network [107]. As sequence polymorphisms can influence communication between different parts of the genome [111] it is probable that some SNPs can generate genotype-specific chromatin networks. This scenario includes the possibility that the genotype promotes or antagonizes the formation of chromatin networks in combinations that for some regions are individual-specific to predispose for various human diseases. It is thus conceivable that the functional annotation of loci identified in GWAS screens will benefit from the incorporation of the chromatin/chromosomal network perspective. Once used the approaches described above to generate a focused list of polymorphisms for functional follow-up, the subsequent challenge is to examine their impact in appropriate in vivo model systems. The principal criterion for taking a polymorphism forward is the association of the eQTL with a disease.So far, epigenetic modifications play important roles in controlling gene expression. But epigenetics is much more diverse than canonical gengetics, related studies are far behind those on the latter. In the future, certain epigenetic modifications might be considered as another type of genetic variations, and could be independently investigated in GWAS studies.

Models for testing function in vitro and in vivo

Gaining a better understanding of the biological mechanisms of cancer development often relies on the analysis of models that reflect the human disease and the application of technologies that facilitate the analysis of these models (Table 1). It is likely that establishing a functional rationale underlying the significance of allelic variation and the identification of candidate genes at common low penetrance susceptibility loci in biologically relevant disease models, will become a major component of following-up the data emerging from GWAS. Disease models can be based on either the in vitro characterization of human tissues (primary tissues or cells in culture) or in vivo models of disease development. However, models are generally limited to studying one variant/ gene at a time. It remains technically challenging to study co-founder effects, the possibility that multiple genes in a locus cooperate in the development of a particular disease . This mirrors challenges of assessing individual risk based on a statistical model in which SNPs interact, and , indeed, the statistical significance of associations between SNPs and disease is based on the idea of independent contributions from each SNP, rather than a complex interplay.  ("One cannot reduce the complex bodies to their simplest parts, nor can one reduce the idea of the body to its simplest parts." ~ Spinoza. See discussion on pathway approach.)  Genome-wide SNP genotyping could be exploited as a 'fingerprinting' tool for finding disease-causing genes in both congenital [112][113] and sporadic diseases, such as cancer [114] [115]. Whereas, copy number variation (CNV) provides gene dosage information which can be used for measuring global genomic instability status and identify novel disease biomarkers during disease progression [116][114][116][115]. The basic hypothesis underpinning these studies is that genetic variation at susceptibility loci influences the initiation of the disease phenotype. Although the expression of several highly penetrant disease genes (e.g. BRCA1, BRCA2, APC) is ubiquitous, the functional effects of genetic variation are reflected in a tissue-specific manner. Therefore, it will be necessary to evaluate the functional impact of disease-associated SNPs in the precursor tissue. Another issue is the size of the functional effects expected for common genetic variants. Whereas the functional effects of several sequence alterations in genes such as BRCA1 and APC are clearly detrimental (they are usually coding sequence changes leading to defective protein products or abrogated expression), the functional effects of SNPs are likely to be subtle, and therefore likely to require large sample sizes in order to achieve sufficient power to detect associations. Thus, large bio-banks of normal tissues may need to be established to evaluate functional differences between the different alleles of a SNP may require establishing large biobanks of normal tissues. Establishing such biobanks will be a significant part of the challenge; whereas extensive efforts within the cancer research community have established tumor tissue bio-repositories, it has been less common to do so for normal tissues from the cells of origin of cancers. This issue is particularly problematic for tumor subtypes in which the cell of origin is still debated. Studies of normal tissues will need to be complemented by in vitro analyses using cell culture models. The location of SNPs with respect to candidate genes will provide multiple testable hypotheses about the functional consequence of genetic variants - for example whether or not SNPs lie in coding sequences, intronic regions, in well-defined gene promoters or stretches of chromatin enhancer sequences that may suggest a role in transcriptional regulation . Progress in establishing good models of normal tissues has been hampered by difficulties in accessing specimens, and the challenges of culturing primary cells. For example prostate epithelial cells are dependent on the presence of a co-cultured stromal component for establishing the secretory cell phenotype and functional differentiation. For the normal colon, most commercially available normal epithelial cell lines are fetal in origin, and differences in fetal and adult cell biology limits the translational potential of work using fetal cells to model adult epithelial cancer genesis. There are exceptions - in breast, well-characterized commercially available cell lines exist that represent good models of normal breast tissue, e.g. MCF10A cells and immortalized HMECs. Three-dimensional (3D) cultures of MCF10As form polarized cystic structures that closely reflect the architecture and molecular features of breast acini in vivo. Using this system, a link between loss of BRCA1 function and impaired luminal differentiation of mammary epithelia was established [117]. The fidelity of this association was later confirmed through analysis of epithelial subpopulations in BRCA1 mutation carriers [118] . Since BRCA1 basal-like breast cancers may originate from luminal epithelial progenitors [119], impaired luminal differentiation may be principal to tumorigenesis associated with BRCA1 mutations. If so, quantitation of multi-cellular phenotypes resulting from BRCA1 depletion in the MCF10A 3D system may serve as an in vitro tool through which genetic associations, such as those identified through GWAS [120], can be validated. For instance, a locus on 19p13 modifies risk of breast cancer in BRCA1 mutation carriers [120]. A smaller 13Kb region, defined by the most strongly associated SNPs, contains three genes, including ABHD8, ANKLE1 and C19orf62 (aka MERIT40 ) [120]. MERIT40 is a component of the BRCA1 A complex and may be the most plausible genetic modifier in this cluster [120], however, variation in other genes may contribute to risk. To test these putative associations, matrix format screens could be used to concurrently deplete each candidate gene with BRCA1 depletion to measure the attenuation or augmentation of impaired luminal differentiation within the MCF10A 3D system. By using such 3D models, it is possible to dissect subtle phenotypes, such as changes associated with gene dosage. For example , a recent study using primary human oral epithelial mucosal keratinocytes on 3D cultures has unraveled an early mechanism of how aberrant expression of FOXM1 transcription factor hijacks adult stem cells to induce cancer initiation [121].

The development of in vitro models of human tumors, which represent the quickest and most accessible way to test the function of candidate genes located at susceptibility loci, are more advanced, but functional effects may be masked by an aberrant genetic background and limited diversity of cell types. Importantly, most SNPs are reported to confer a predisposition to a particular disease and , consequently, we must expect that the greatest functional impact will be achieved in an essential healthy, non-aberrant tissue/background. Such a context in which to study function is perhaps the hardest context to replicate and maintain in a laboratory setting, meaning that there must be a continuous drive for improvements in the models used. Evaluating the functional effects of SNPs using in vivo models represents an even greater challenge. Several significant limitations in the biological validation of candidate genes remain by employing genetically engineered animal models, including:

• Relatively short duration of experimental models compared to human tumorigenesis that typically develops over several decades.
• Important differences in human vs. animal physiology. For example, animal models of HIV, psychiatric illnesses and leprosy hardly exist and no standardized phenotype ontology to map complex human traits in animal model phenotypes exists [122].  can be explained better. Any good reference ?   the example looks like unrelated to cancer risk? 
• Important differences in the structure and sequence of non-coding regions  as above, can be explained better 
• Limited modelling of gene-environment interactions.
• Sensitivity of animal modelling to confirm function of low penetrance alleles when studied in isolation.
• High frequency (35%) of perinatal lethality in animal models when the GWAS-implicated loci are knockedout, thus preventing further elucidation of the responsible molecular mechanisms [122].
• Difference in the environment where the animal grows, as the cage where the animals live can influence their behavior and health. Toys or social interactions can reduce these difficulties. [123]  
• Possible bias toward males , and intrinsic difficulty in designing experiments with females, due to the menstrual cycle. [124]
• Differences in N-Glycosylation and protein modification .
• Different Alternative Splicing regulation.
• Epigenetic differences.
• Differences in the karyotype and chromatin interactions.  this last phrase written on behalf of Rolf Ohlsson 

Exploring the functions of SNPs in regulatory sequences/regions: drawing the themes together.

Functional SNPs found within large, non-coding intergenic intervals highlight the challenges of defining LD structure and pursuing functional validation. One of the hypotheses to explain this is that the risk locus contains regulatory region(s) with element(s) affecting the expression of one or more distantly transcribed region(s) of the genome. Here the starting point is to explore whether they are components of regulatory elements and what their distal effects may be. Although the most abundant of these regulatory sequences are enhancers, they may also include other regulators such as insulators and silencers. Unlike promoters (at transcription start sites of genes), distal regulatory sequences such as enhancers are often cell-type specific [125] and thus may be targets for tissue-specific risk-SNP effects. Annotating such regulatory sequences using chromatin marks or DNase sensitivity has proven to be a powerful method [126][127][127], and is more informative than the alternative approach of using evolutionary conservation, since regulatory elements tend to be unconstrained across mammalian evolution [128][129][130]. Specifically, identifying regulatory sequences containing functional SNPs within response elements by using chromatin annotations has been proposed recently[131]. Additionally, demarcation of such regulatory regions is even more precisely achieved by assessing the association of candidate transcription factors with response elements. Both histone modifications and transcription factor occupied regions are currently identified using ChIP-seq methodologies and signals yield short DNA stretches (~1kb) amenable to detailed analyses. Enhancer activity in such regulatory regions can be assayed using reporter genes in vitro [5] and/or in vivo[126]. Additionally, resequencing the target regulatory regions should then be prioritized to capture all the variation within them. Once activity is demonstrated, identified SNPs within the regions, especially ones within known transcription factor response elements and in LD with a tag-SNP, may be analyzed using biochemical methods for differential transcription factor binding and activity. Resequencing, targeting the regulatory regions, then should be prioritized to capture all the variation within them. Finally, regulatory sequences containing functional SNPs determined in this way can be matched with their physiological target genes to probe functional significance . It is possible that multiple SNPs may function coordinately at a particular locus and each should be taken forward in subsequent mechanistic analyses. Three approaches are conceptually available to identify targets of regulatory sequences: (i) Target genes may be identified in regulatory sequences knockout mouse models (using cre-lox for tissue-specific and timing purposes) by genome-wide gene expression analyses after the knockout; and (ii) Target genes may also be identified by using the regulatory sequences as baits in chromatin conformation capture (3C) based studies [132][133], including genome-wide 3C-seq, to identify pleiotropic chromatin crosstalk; and (iii) Targeted editing using somatic cell knock-in technology, although technically demanding, is another approach. Allelic series in isogenic settings may be created and gene expression differences measured - either in naturally growing cells or in cells that are perturbed in some manner (e.g., radiation, hormones, etc ). Finally, these screening approaches should be followed by matching results from them, creating a list of robust targets that can each specifically be studied further. Finally, these screening approaches should allow the matching of results from and between them, creating a priority list of robust potential targets that might each specifically be studied further. Thus, target genes under control of functional SNP-containing regulatory regions may have important roles in the cancer phenotype, such as proliferation, migration and apoptosis. Endpoints of the cancer phenotype, such as cell division, migration and apoptosis rates and protease secretion may be measured in cultured cancer cells and mouse xenografts after the overexpression of the genes of interest, or their selected siRNA/shRNA knockdown. It will also be important to develop ways to make this fine functional annotation data generated in these analyses freely accessible, because in many cases while they may not constitute publishable data they could be useful for other post-GWAS efforts. In summary the order of investigation is: identify chromatin architecture in tag-SNPS (+LD blocks) -> assay for regulatory activity, such as enhancers -> determine how SNPs in LD with the tag-SNP effect such activity -> identify regulated target genes -> understand biology/cancer risk -> understand causal predisposition.

A role for Digital Life Information? See discussion page  .

Conclusions

The GWAS community has arrived at an important crossroads. As resources are limited, the debate revolves around whether enough progress has been made towards identifying the SNPs that are likely to contribute most to disease causation or to invest in functional post-GWAS analysis. Consensus on a single gene not being the only causative factor in complex diseases like cancer is emerging. Similarly, it is not surprising that functional SNPs affecting a single gene or allele may have a subtle effect on their own. As sequencing technologies become cheaper and more accessible, we argue that this will evolve rapidly as datasets expand and will afford greater confidence in defining both SNPs and LD structure within the region where they lie. This will require detailed mapping and annotation of epigenetic and transcriptomic landscapes within which a major limitation may prove to be the sample collections themselves. Whilst this progresses, which it will be hopefully achieved through consortia assembled from academia and industry, it is vital that proof-of-principle studies take forward the strongest candidate SNPs available so far, not in this case necessarily to test their causative association with disease, but rather to understand their functional impact. What makes for strong candidates are significant associations with transcript expression ( overlap between eQTL and GWA signal ; chromosome conformation capture), tissue specificity and the phenotypic impacts of these transcript associations on model systems in downstream experiments. Successfully making the experimental transitions to progress through this process will require collective working at a consortia/multi-group level and a clear decision tree. It is essential for the field that this overrides the temptation to publish fragmentary work capturing only sub-steps in this sequence. Over time, integration of the re-sequenced, epigenetic and molecular-epidemiological data within different ethnic groups (and thus within different linkage disequilibrium structures) will help localizing causal variants. If we begin to consider how to explore the functional impact of SNPs now, we will, as a community, be in good shape to rise to the challenge of testing causation in the future. The field is still making the first forays into the functional characterization of SNPs and is many steps away from proving causality. Nonetheless, our view is that causality can only be inferred if the eQTL is associated with disease and a SNP leads to expression differences in reliable in vitro and perhaps in vivo assays. Naturally, our ability to get close to this goal must be assessed in the context of current technologies and knowledge on a disease-by-disease basis, but we hope that this article will help to frame the developing debate and the emerging research that seeks to rise to this great challenge.

Acknowledgements

We would like to thank Dr. Fred Bunz and all the members of the NIH Post-Genome Wide Association Initiative for helpful discussions and in particular Dr. Ian Tomlinson (Wellcome Trust Centre for Human Genetics, Oxford). The contributing groups are supported by funding made available through the NIH Post-Genome Wide Association Initiative in response to Call  . This Call sustains research across five cancer organ sites (prostate: 1U19CA148537-01, breast: 1U19CA148065-01, ovarian: 1U19CA148112-01, colorectal: 1U19CA148107-01 and lung: 1U19CA148127-01). For further information on this Initiative please refer to the website:  

Glossary

3C, 4C, 5C, Hi-C, 3C-seq: chromosome conformation capture (3C) is a technique is used to identify interactions between genes and long-range regulatory domains made possible by chromosome loops that bring the two regions to physical proximity. Developments on this method include circularization (4C) of the genomic fragments with the use of inverse PCR primers, carbon copy (5C) technology for multiplexed ligation-mediated amplification, and high throughput analysis by massively parallel sequencing between many baits and targets (Hi-C), or many targets from a single locus (3C-seq).

Supporting references: [134] [135][136][137]

Causal variant: In the context of GWAS it represents the SNP that is mechanistically linked to risk enhancement. This is distinct from SNPs that do not have any functional impact but are statistically associated with the disease phenotype due to its linkage disequilibrium with the causal variant.

Supporting evidence: [138]

ChIP-Seq: Chromatin Immunoprecipitation (ChIP) is a method to study protein-DNA interactions. It identifies genomic regions that are binding sites for a known protein. Analysis of these regions is typically performed by PCR, when there is a hypothesized known binding site, or through the use of genomic microarrays (ChIP-chip). Alternatively, analysis can be done using next-generation sequencing (Seq) technology to analyze DNA fragments.

Supporting references: [139][140]

CNV: Copy Number Variation is a type of structural variation in which a particular segment of the genome, typically larger than 1kb, is found to have a variable copy number from a reference genome.

Supporting references: [141][142][143]

Deep sequencing: a sequencing strategy used to reveal variations present at extremely low levels in a sample. For example, to identify rare somatic mutations found in a small number of cells in a tumor, or low abundance transcripts in transcriptome analysis.

Supporting references: [87]

DNA Methylation: A modification of the DNA that involves predominantly the addition of a methyl group to the C5 position of the pyrimidine ring of a cytosine found in a CpG dinucleotide sequence.

Supporting references: [144]

eQTL: expression Quantitative Trait Locus. A genetic variant underlying the individual differences in quantitative levels of gene (or protein) expression.

Supporting references: [82]

Epigenetic markers: an array of modifications to DNA and histones independent of changes in nucleotide sequence but rather the addition of a methyl group to cytosine and a series of post- translational modifications of histone including methylation, acetylation, and phosphorylation.

Supporting references: [145]

Fine mapping: a strategy to identify other lower frequency variants in a disease-associated region (typically spanning a haplotype block) not represented in the initial genotyping platform with the goal of uncovering candidate causal variants. It can include data mining of publicly available sequencing efforts, such as the 1000 Genomes Project and targeted re-sequencing. Supporting references: [138][146]

Functional variant: a variant that confers a detectable functional impact on the locus. It can represent a change in coding region but also changes in regulatory regions that have an impact on function.

Supporting references: [5]

GWAS: Genome-Wide Association Study is a case-control study design in which most loci in the genome are interrogated for association with a trait (disease) through the use of SNPs, by comparing allele frequencies in cases and controls.

Supporting references: [1]

Haplotype block: linear segments of the genome comprising of co-inherited alleles in the same chromatid.

Supporting references: [147][148]

Homologous recombination: an error-free recombination mechanism that exchanges genetic sequences between homologous loci during meiosis, and utilizes homologous sequences such as the sister-chromatid to promote DNA repair during mitosis.

Supporting references: [148]

Linkage disequilibrium: a nonrandom association between two markers (e.g. SNPs ).

Supporting references: [149][150][147][151].

MicroRNAs: endogenous short (~23 nt) RNAs involved in gene regulation by pairing to mRNAs of protein coding mRNAs.

Supporting references: [152]

Next generation sequencing: a technology to sequence DNA in a massively parallel fashion, therefore sequencing is achieved at a much faster speed and lower cost than traditional methods.

Supporting references: [145]

Non-coding variant: a variant that is located outside of the coding region of a certain locus.

Tagging variant: a variant (SNP) that defines most of the haplotype diversity of a haplotype block.

Supporting references: [147]

Transcriptome: The complete set of transcripts in a cell. In some cases it can also include quantitative data about the amount of individual transcripts.

Supporting references: [87]

r²: square of the correlation coefficient (r) between loci/markers, which is a normalized value for the deviation of the observed frequency of a haplotype from the expected frequency.

RNA-Seq: a method to obtain genome-wide transcription map using deep sequencing technologies to generate short sequence reads (30-400 bp). It reveals a transcriptional profile and levels of expression for each gene.

Supporting references: [87]

SNP: single nucleotide polymorphism

Figure 1

The genomic context in which a variant is found can be used as preliminary functional analysis. Figure shows the genomic context distribution of disease- and trait-associated SNPs annotated in the Catalog of Genome-Wide Association Studies [2] as of December 9th, 2010. Most of the SNPs are located in intergenic and intronic positions, but a small percentage are located upstream and downstream of genes, as well as in regulatory regions and splice sites. SNPs in these locations can be analyzed in more detail using more specific bioinformatics tools.

 

Table 1. Methods for functional validation/enabling technologies

Table 2. Computational methods to characterize risk loci function

References