ANNOTATING INDIVIDUAL HUMAN GENOMES

Abstract

Advances in DNA sequencing technologies have made it possible to rapidly, accurately and affordably sequence entire individual human genomes. As impressive as this ability seems, however, it will not likely to amount to much if one cannot extract meaningful information from individual sequence data. Annotating variations within individual genomes and providing information about their biological or phenotypic impact will thus be crucially important in moving individual sequencing projects forward, especially in the context of the clinical use of sequence information. In this paper we consider the various ways in which one might annotate individual sequence variations and point out limitations in the available methods for doing so. It is arguable that, in the foreseeable future, DNA sequencing of individual genomes will become routine for clinical, research, forensic, and personal purposes. We therefore also consider directions and areas for further research in annotating genomic variants.

INTRODUCTION

The DNA sequencing Era

The introduction and now wide-spread availability of efficient and cost-effective high-throughput DNA sequencing technologies has radically changed the way in which geneticists can approach the study of human diseases. For example, traditional strategies for identifying or ‘mapping’ genes and/or genetic variants that influence a disease typically involved tracing co-segregation or linkage disequilibrium (LD) patterns between a particular genetic variant and a phenotype either among family members, as in classical linkage analysis [1,2], or across unrelated individuals, as in more contemporary LD mapping-based association analysis [1,3]. The goal of these strategies is to identify a unique ancestral chromosomal segment or haplotype that is ‘marked’ by specific variants that track along with a disease from generation-to-generation or between individuals with the phenotype. The basic intuition is that this segment or haplotype contains a variant that causally influences the phenotype and the neighboring material is simply being inherited along with this causal variant. Thus, the identified associations between genetic variants and a particular phenotype resulting from these strategies are often indirect and merely pointed researchers to a genomic region that might harbor variants that are causally-related to the phenotype in question. The subsequent identification of the causal variants in the region implicated in a linkage or LD-mapping study is not always trivial and can require a great deal of effort.

Whole genome sequencing can facilitate tracing segregation and/or LD patterns involving a phenotype of interest or a haplotype that harbors a variant that influences that phenotype. This is the case because one does not have to first identify a region harboring a variant and then seek to determine the trait-influencing variant on the haplotype at a later time, but rather would have all the information needed for identifying an association and a causal variant simultaneously. DNA sequencing a genomic region or entire genomes among individuals, e.g., with and without a particular phenotype, will reveal all the genetic variants the individuals possess, shifting the problem from merely identifying a region harboring a variant to determining which among the many variants observed in the region (or genome as a whole) are likely to influence the phenotype [4]. As with traditional gene mapping methods, the use of DNA sequencing technologies to identify variants associated with a particular phenotype requires sophisticated data analysis methodologies [5]. The search for variants likely to be associated with a phenotype from sequence data can be greatly enhanced by leveraging knowledge about the variants discovered, such as their positions and likely functional effects, since this knowledge can be used to prioritize variants or weight them in relevant statistical analyses [5]. Assigning biological meaning to sequence variants, i.e., ‘annotating’ them, is the subject of this review, as there are a variety of ways variants can be annotated and exploited in various contexts.

There are other reasons why sequence annotations are important in modern human genetics. For example, many clinical diagnoses of congenital conditions require an understanding of variants an individual might possess that could be pathogenic, and this understanding requires the use of methods for classifying and annotating existing and novel variants, as is the case with BRCA1-associated breast cancer and similar conditions [6]. In addition, many contemporary sequencing initiatives focus on the mere identification and cataloguing of naturally-occurring sequence variants, such as the International HapMap Project [7], the 1000 Genomes Project [8], and the Cancer Genome Atlas Project [9], leaving it open to others to determine the biological significance of the variants identified. Finally, as whole genome sequencing (WGS) of individual human genomes becomes more efficient, the need for putting the variation across individual genomes into a biological context will become more pronounced [10–12].

Annotating Sequence Variations

There are many ways in which one can assign a label or biological meaning to a position in the human genome that varies among individuals in the population at large. The simplest is to identify a variant’s position either within the sequence as a whole or in a putative functional genomic element. This is not trivial, as the genomic positions must be based on landmark sites in the genome or with respect to a ‘reference’ genome, such as that used for the human genome browser [13,14] or the HuRef genome [15]. The use of reference genomes can be problematic, however, since human genomes vary in size due to repetitive elements and functional elements are continually being identified and updated. In addition, sequence variants, perhaps despite being in a functional element, do not often perturb the actual functioning of a position or functional element [16]. The evolutionary or population history of a variant might shed light on a variant’s effect, but how one determines and assigns the functional impact of variant given the biological realities of e.g., redundancies and feedback in disease pathways, general buffering effects, and tissue-specific gene effects, is not trivial [16–18], nor is understanding the origins of a variant [19].

One important issue surrounding the annotation of sequence variations involves the nature of the variants themselves [20]. Single nucleotide variants (SNVs) are the most abundant forms of variation and have been given a great deal of attention. Small insertions and deletions (‘indels’), repetitive sequence, copy number variations (CNVs), and inversions have been given less attention, although they are beginning to receive more. Another issue relating to variant annotation involves the potential context-specificity of particular effects. Epistasis, gene x environment interactions, or general multilocus and genetic background effects may mitigate or enhance the effect of any single variant. In addition, phase information (i.e., knowledge of the nucleotide content of the two homologous chromosomes an individual inherits in the vicinity of a variant) is known to influence the functional effects of a variant in many cases [21].

In the remainder of this review we consider the various methods and strategies for annotating DNA sequence variation. These strategies range from merely identifying the genomic elements (e.g., genes, promoters, microRNA binding sites, etc.) a variant reside in, predicting the likely functional effect of an association using sophisticated statistical algorithms, and directly examining the impact of the variant on phenotypic expression using contrived experimental settings involving in vitro or model organism-based assays. By ‘functional’ it is meant that a variant results in a change of phenotype either at the basic molecular level (e.g., change in binding strength of a transcription factor) or at the overt phenotypic level (e.g., the variant is associated with, or known to cause, a disease). Although many of the methods, analyses, and references we comment focus on single nucleotide variants (SNVs), quite a few of these methods are applicable to other forms of variation as well. Table 1 provides a listing of example tools and databases that provide annotations of the type we discuss. We consider the limitations of each of many approaches and provide references to some of the more salient examples of their use. We conclude with a few remarks on the future of genome annotations and the use of such annotations in clinical and non-clinical settings.

Table 1.

Example Tools for Human Variant Annotations

Tool	Website/Reference	Purpose/Theme
UCSC genome browser	http://genome.ucsc.edu/	Position-specific functional organization of the genome
dbSNP	http://www.ncbi.nlm.nih.gov/projects/SNP/	Catalog variants with population-genetic annotations
OMIM	http://www.ncbi.nlm.nih.gov/omim	Catalog known disease-causing mutations
HapMap	http://hapmap.ncbi.nlm.nih.gov/	Catalog variants with population-genetic annotations
COSMIC	http://www.sanger.ac.uk/perl/genetics/CGP/cosmic	Catalog of somatic mutations from tumor sequencing
TAMAL	http://neoref.ils.unc.edu/tamal/	Provides functional and population-genetic annotations
Variant Analyzer	http://www.svaproject.org/	Provides functional annotations
PharmGKB	http://www.pharmgkb.org/	Pharmacogenetics variant annotations
HGDP Selection Browser	http://hgdp.uchicago.edu/cgi-bin/gbrowse/HGDP/	Browser for assessing signs of selection in the human genome
Association Database	www.genome.gov/gwastudies	Results of genome wide association studies (GWAS)
SeattleSeq	http://gvs.gs.washington.edu/SeattleSeqAnnotation/	Variant Annotation
Gene Ontology	http://www.geneontology.org/	Biological, Molecular and Cellular Annotations
KEGG Pathways	http://www.genome.jp/kegg/pathway.html	Pathway Analysis
DAVID	http://david.abcc.ncifcrf.gov/	Multiple Annotations
UniProt	http://www.uniprot.org/	Protein Elements
Transfac	http://www.biobase-international.com	Transcription factor databases
Genenetwork eQTL website	www.genenetwork.org	eQTL database

TYPES OF ANNOTATION

Genomic Position and Functional Elements

The simplest and most obvious form of sequence variant annotation is determining the variant’s position in the genome. Position-based variant annotations are greatly facilitated by the availability of reference genomes and web-based browsers that can be used to graphically display genomic positions on those reference genomes, such as the UCSC browser [13,14]. The position of a variant can also be annotated with respect to known functional elements in a genome, such as introns, exons, promoters, enhancers, silencers, microRNA binding sites, conserved regions, etc. Annotating a variant with respect to its position in functional genomic elements is crucial for putting the likely functional effect of that specific variant into perspective, as discussed in detail below [22]. The UCSC browser provides information on functional genomic elements and their positions in the genome as well as an easy way of representing the position of an identified variant relative to a reference genome [13,14].

There are few caveats with determining the position of a variant in a genome either relative to other variants or with respect to functional elements in a genome. First, not all functional elements in the human genome have been identified, although their existence is constantly being updated through functional studies. Second, humans are a diploid species, so that each human contains two genomes: one maternally-derived and one paternally-derived. The position of genetic variants on a genome relative to other variants should thus be determined with consideration of which chromosome or haplotype they reside on. Characterizing haplotype or ‘phase’ information from sequence data obtained from an individual is not trivial, but may be important for putting the functional effects of those variants into perspective [21, 23]. Third, variant positions can also be determined independently of a reference genome and therefore be assessed merely relative to other variants in an individual genome via de novo assemblies of individual genomes. However, such de novo assemblies are notoriously difficult for human genomes [15].

Population Characteristics

The population characteristics of a variant are an important form of initial annotation information. The frequency of a variant (or the frequencies of alleles at a multiallelic locus) within different populations is a useful indicator of the age of a variant. Frequency information also provides insight into the population-specificity of the variant – which could be a sign of selection effects ([24]; see below) – or suggest that the variant, if it had never been observed before in population studies, is a de novo mutation. However, the best way to determine if a variant is de novo is to genotype the parents of the carrier and determine if the variant was indeed inherited or not. Allele frequency information has also been shown to be associated with likely functional or phenotypic consequences of a variant [25–27].

Other population characteristics that are important for variant annotation include linkage disequilibrium (LD) patterns exhibited by the variant with other variants. Such patterns can provide information about the haplotypic background upon which the variant arose and the potential for use of other variants as surrogates or proxies for the variant in question. Such information can greatly facilitate gene mapping studies [28]. Population characteristic annotations of variants are often recorded in databases such as the hapmap database [7] and dbSNP [29].

Conservation

One of the most widely used methods for estimating the likely functional impact of a variant is to investigate how conserved the nucleotides are encompassing the variant (or amino acids if the variant is in the coding region of a gene). If the wild-type or non-variant nucleotides at the position are conserved across species, then this could suggest that selection acted against alternative nucleotides at that position on an evolutionary time scale, indicating not only the likely functional or phenotypic significance of those nucleotides but also the deleterious nature of the variant nucleotides. Many programs used to estimate the likely functional impact of a variant leverage conservation information (e.g., [30]). However, conservation has been shown to be only a reasonable surrogate, rather than a perfect surrogate, for functionality and phenotypic relevance [31]. In addition, lineage and species-specific functionality of nucleotides raises questions about what species to include in an assessment of conservation (see, e.g., [32]).

Biophysical Modeling

One direct way to assess the likely functional impact of a DNA sequence variant is to determine the influence of the nucleotide substitution, indel, or larger structural anomaly on the structure of DNA sequence encompassing that variant or a protein encoded by the sequence harboring the variant. Although such biophysical modeling has been pursued at the level of DNA to characterize and quantify the structure and properties of unique sequences [33–35, 31] it is far from trivial to perform and its use to assess the effect of nucleotide substitution requires extensive knowledge of not only the basic properties of DNA, but how those basic features impact the functioning of higher-order structures and processes such as proteins, binding events, and interactions involving different DNA sequence-mediated molecular structures.

Knowledge of protein structures associated with a gene whose coding regions harbor variants is on much more solid footing, since the characterization of the effects of particular residues or amino acids on protein function is achievable with appropriate structural assays and computational models [36, 37]. However, not all proteins have had their structures solved, nor do all amino acid substitutions result in an overt protein deformation or admit an easy functional characterization. In addition, relatively few proteins have had both mutant and wild-type structures solved, making it hard to evaluate how well one can predict the influence of specific amino acid substitutions on not only those proteins whose structures are solved, but homologous (or paralogous) protein structures as well [38].

Prediction/Classifier Analyses

In the absence of precise biophysical or functional models of DNA sequence or protein structures, determining the likely effect of a sequence variant on those structures can be pursued by building predictive models based on various features associated with previously characterized variants (e.g., sequence motifs, sequence conservation levels, similarity to other variants, the results of biochemical assays exploring those variants, etc.) that can discriminate known functional from non-functional variants. The information about which features appear to discriminate functional from non-functional variants can then be applied to novel or previously uncharacterized variants in order to estimate their likely functional potential. Prediction and classification analyses of this sort of used widely and form the basis of many computer programs for evaluating the functional significance of coding variants [39, 22, 32, 40–42], variants in regulatory genomic regions [43], and somatic mutations in cancer [44, 45]. Problems inherent in these approaches are the need for specifying the variant ‘features’ that will be used to create predictive models, since the chosen features might not encompass the more important, yet unknown, features [40], and the need for training data sets for the development of the predictor or classifier for which the absolute functionality of variants has been determined.

Literature-based Annotations of Specific Variants

The best way of assigning a functional annotation to an observed sequence variant is to specifically study that variant using, e.g., relevant laboratory functional assays, transgenic mice, families with clinical phenotypes possessing that variant, etc. The literature contains thousands of papers describing the results of various assays investigating the significance of particular sequence variants. Accessing this information to annotate an observed variant in an individual genome is not entirely trivial as the reports in the literature describing any variant will undoubtedly have nuances and caveats, and be very context-specific. Databases such as OMIM [46] can clearly facilitate searches for at least the overt phenotypic influence of a variant, as can phenotype or disease-specific databases such as AlzGene [47].

Many reports describing the functional effects of human variants leverage laboratory-based in vitro or model organism functional assays, the generalizability or ultimate human in vivo significance of which is hard to gauge. Recently, Ioannidis and Kavvoura [48] and Cerulli and Goldstein [49] considered the results of in vitro studies designed to assess the functional effects of specific naturally-occurring human DNA sequence variants on the expression of genes in cis and compared them to the results of association studies involving those same variants with specific diseases. The authors found only a weak correlation, suggesting that the results of contrived laboratory-based functional assays probing the molecular effects of variants do not necessarily predict which variants play a large role in mediating disease susceptibility in the population at large. Despite this, knowledge of the molecular effects of variants, in whatever setting, can be important for making claims about the potential for those variants to influence gross phenotypic expression as well as those variants’ potential for participating in pathogenetic processes and networks that underlie diseases.

There is a growing trend to develop tools for mining the literature on genetic variants by leveraging appropriate keywords, assay types, etc. mentioned in published reports on the internet [50, 51]. The basic idea is to bring together and synthesize information about genetic variants and possibly make predictions about a novel variant based on what is learned from the literature about related variants (see, e.g., [50, 51]). However, there are clearly limits in the available published knowledge that might create biases in developing such predictive schemes [52].

Context-specific perturbation-based annotations

The functional effects of many sequence variants require appropriate environmental conditions to be present, and hence annotations assigned to variants that exhibit an effect only in certain contexts are important. For example, many variants have been found to exhibit interactions with environments of one sort or another in the context of disease susceptibility [53]. However, more specific interactions involving drugs or chemicals have received a great deal of recent attention and have led to databases cataloging evidence for such interactions, such as the ToxPO database [54] and, importantly, the Pharmacogenetics Knowledge Base or ‘PharmGKB’ [55]. Many studies and databases providing context-specific functional annotations of genetic variants are not necessarily population-based, but may leverage specific laboratory or model-organism based functional assays [56].

Molecular phenotype-based association study results

Many researchers have pursued tests of associations between genetic variants and the expression levels of genes measured on a genome-wide scale via microarray technologies [57,58]. Such studies typically involve measuring gene expression levels from tissues obtained on many individuals, genotyping those individuals, and then testing the statistical association between the genotyped variants and the expression levels of the genes. These studies have revealed thousands of putative statistical associations between genetic variations, working either in cis or in trans, and the expression levels of many genes [59]. Association studies of this type have been expanded to consider associations between genetic variants and protein levels [60] as well as the methylation levels of different genomic sites [61]. In fact, some investigators have considered integrating information about genetic variants influencing both, e.g., expression and methylation levels of certain genes [62].

Annotating variants with respect to their statistical association with changes in expression levels (expression Quantitative Trait Loci or ‘eQTLs’), protein levels (‘pQTLs’), and methylation levels (‘mQTLs’) can be very revealing about the likely functional effects of the variant, especially since it has been shown that variants that influence expression levels of genes from association studies are more likely to be associated with overt phenotypes and/or disease [63]. Two important issues in annotating variants in the context of their influence on expression level, protein level, and methylation level are that these phenomena are tissue-specific, making it important to consider the tissue or cell type that the assays were performed with. In addition, the statistical criteria for declaring a significant association between a variant and, e.g., the expression level of a gene must be considered since genome-wide association studies with small sample sizes – which is the rule for most eQTL, pQTL, and mQTL studies – can result in many false positive and false negative associations.

Clinical phenotype-based association study results

By far the most important and relevant variant annotation concerns knowledge of what overt, clinical phenotypes and diseases a variant may either cause or be associated with. The assessment of which overt or clinical phenotypes a variant might cause or be associated with is in contrast to variants exhibiting associations with molecular physiologic phenotypes such as gene expression or protein levels, protein structural deformations, or environmentally-induced phenotypes, since the molecular effects of a variant may be mitigated or offset by the effects of other genes or environmental conditions, such as pharmacotherapies, and hence not result in an overt phenotype. Databases such as OMIM [64] catalog disease-associated variants and are extremely useful for annotating previously observed variants likely to be causally associated with rare Mendelian diseases. In addition, the results of recent genome wide association studies (GWAS) have provided over 1000 very compelling associations between particular genetic variants, mostly SNVs, and common, complex diseases and phenotypes of all sorts ([3] and http://www.genome.gov/GWAstudies/).

Two important Issues with the use of association study results like those recorded in OMIM and GWAS results databases [65] for genome annotation include the fact that they rely on previous knowledge of the association and hence cannot be used for novel variants, and that one must consider the statistical significance of the association (as with eQTL, pQTL, and mQTL analysis) [66].

Multilocus phenotypic predictions/disease risk assessment

Once associations between multiple variants and a phenotype have been described, it is important to consider the collective effects of these variants on phenotypic expression. This is not trivial since it requires knowledge or assumptions about how the variants interact, but is of great interest since such models could be used for predicting or classifying individuals harboring specific variants into, e.g., diagnostic or prognostic categories with respect to disease. There are two ways relevant modeling and analyses can be pursued. The first is to directly assess the collective effects of variants in an appropriate data set with individuals studied before and after the onset of the phenotype so that predictions can be made. The second is to model the impact of the collection of variants by making assumptions about, e.g., their effect sizes, frequencies, combined influence (such as additive or multiplicative effects), the incidence and prevalence of the phenotype in question and how incidence and prevalence are impacted by the collective influence of the variants, and the influence of non-genetic factors. The second method is used routinely, as many association studies and candidate gene analyses have been pursued in isolation and not in one all-encompassing longitudinal study, creating a need to synthesize the information across such studies. Obviously, such studies are highly problematic since the assumptions they make may not be correct [67].

Despite this fact, phenotype prediction from genetic (and non-genetic) data is seen as the holy grail of genomic medicine and disease prevention, even though there are great concerns about how such information will be perceived and utilized [68]. Recent studies in diabetes and cardiovascular disease suggest that the addition of previously associated genetic risk loci in clinical risk models of those diseases increases both discriminative and predictive accuracy, albeit only marginally. Typically, the strongest predictors of disease onset are known clinical risk factors such as body mass index, age, or gender. While genetic risk variants can have unequivocal associations with disease, the addition of multilocus genetic risk to clinical models rarely provides large gains in predictive accuracy, and models built solely on genetic factors tend to perform comparably to non-genetic risk models. Table 2 provides a list of studies exploring multilocus risk or classification models for diabetes and a summary of the outcomes of these studies.

Table 2.

Studies Investigating Multilocus Genetic Models for Predicting or Classifying Diabetics.

Reference	# Variants	Study Design	Outcome
Lyssenko, V. et al. (2005)	6	Predictive	Effect of multilocus genetic risk is roughly equivalent to estimated relative risk
Weedon, M. et al. (2006)	3	Classification	Low predictive power for multilocus genetic model only
Lango, H. et al. (2008)	18	Classification	Adding genetic risk variants only marginally increases discriminative accuracy above BMI, Age and sex
Lu and Elston (2008)	12	Classification	Addition of 9 variants and 4 environmental variables to model by Weedon et al., (2006) marginally increases AUC estimates
Lyssenko, V. et al. (2008)	16	Predictive	Addition of genetics to clinical models improves discriminatory power and reclassification of patients in a small but statistically significant way
Meigs, J. et al. (2008)	18	Predictive	Addition of genetics only slightly increases risk prediction accuracy above common clinical risk factors alone
Cauchi, S. et al. (2008)	22	Classification	15-loci model sufficient to attain high discriminative accuracy within French population
vanHoek, M. et al. (2008)	18	Predictive	Adding genetic risk variants only marginally increases discriminative accuracy above BMI, Age and sex
Lu, Q. et al., (2009)	18	Classification	Addition of 15 genetic loci to model from Lu and Elston (2008) provided only limited improvement in discriminative accuracy
Miyake, K. et al., (2009)	11	Classification	Adding genetic risk variants only marginally increases discriminative accuracy above BMI, Age and sex
Talmud, P. et al., (2010)	20	Predictive	Addition of genetics to clinical models only minimally improves predictive accuracy and reclassification

Relevant analyses that lead to genomic variation annotations of this sort are problematic for many reasons that go beyond mathematical assumptions about how independently characterized multiple variants might work together to influence phenotypic expression. For example, most studies that have identified variants to be used in a multilocus model exploit extant or prevalent cases and controls. Hence the resulting models provide classifications of individuals based on phenotype rather than true predictions of who will develop a phenotype over time. In order to make true predictions one would need either to pursue a longitudinal study with individuals without the phenotype at the start or have very reliable incidence data, and getting appropriate incidence data is not trivial [67]. Other issues involve the need to accommodate environmental effects, interrelationships with other diseases, and a need to recognize the limitations of classifications and predictions based solely on what is likely to be incomplete available knowledge of the genetic architecture of the phenotypes in question [69, 70].

Pathway-based annotations

Genes do not work in isolation to influence the expression of a phenotype, but rather work collectively to influence a particular biological process, biochemical pathway or metabolic network that mediates that phenotype’s expression. These processes and networks are replete with feedback and compensatory mechanisms, making it difficult to predict the phenotypic impact of isolated genetic variants, especially in the case of common chronic diseases. Similarly, gene regulatory elements tend to contain redundant and/or tissue-specific elements making it difficult to interpret the ultimate phenotypic impact of variants that disrupt, e.g. transcription factor binding sites. Thus, one needs to consider the influence of a variant that has not been characterized before in the context of the other genes and genomic elements it interacts with [71–73].

Annotations that involve sets of variants, rather than an individual variant, are limited by the available knowledge as to how genes and genomic elements interact and mediate basic molecular processes. Despite this fact, many attempts have been made to annotate the functioning of entire pathways or processes on the basis of variants that disrupt the function of relevant genes [74–78]. Conversely, the phenotypic impact of variants can be predicted on the basis of the biological processes the impacted genes participate in, allowing the inference of gene-phenotype relationships without prior phenotypic information. Moreover, methods such as co-expression or other network inference techniques can be utilized to infer biological process and the potential phenotypic impact of variants in genes of unknown function [71–78]. Thus, pathway and network based annotation approaches can be powerful approaches to inferring phenotypic information where direct links to phenotype do not exist.

De Novo Association Analyses Involving Multiple Genomes

In the absence of prior information one might leverage to annotate the likely effect of a variant, one could conduct a de novo study to determine the phenotypic relevance of the variant. Such studies could include eQTL mapping, in vitro laboratory or model organism-based functional assays, association analyses, or any one of the study types that reflect the annotation types discussed in this review. Of these, the study types receiving the most attention are, as noted, association studies involving medically-relevant phenotypes. Designing and conducting such studies are not trivial as they are costly, laborious and time-consuming, especially if they involve the notion that an identified variant is one of many different rare variants that perturb a genomic element in a similar way [5]. Association studies can be made more efficient through the use of extant control data or pooling strategies [79], although such strategies are not without their problems.

The creation and use of genomic databases, such as dbGAP [80] or other medical genomics databases [81], which harbor DNA sequence variation information on individual genomes with some phenotypic information available from the subjects whose genomes were assayed, will facilitate the annotation of variations via de novo association studies. In fact, it is not hard to imagine a time when whole genome sequencing (WGS) will become routine and databases harboring the WGS information will be created, such that if a variant or variant type is observed, phenotypic associations involving that variant can be pursued in short order. Obviously, and justifiably, informed consent, anonymity, liability, and other issues concerning the use and distribution of individuals’ genome information will loom large in these contexts.

Ancestry

Variation in the frequency of a sequence variant across different populations can point to the origins (or even signs of selection) of that variant. With enough genetic markers across the genome that exhibit frequency differences across populations (‘Ancestry Informative Markers’) estimates of an individual’s ancestry can be made [82]. These same principles can be applied to specific genomic regions implicating a variant [83,84].

Within-Species Evidence for Selection

Many approaches to assessing evidence that a genomic region shows signs of selection – and hence is very likely to be functional or phenotypically-relevant – leverage within-species, as opposed to cross-species, information [24]. It is therefore possible to annotate variants and genomics regions harboring variants on the basis of evidence that they exhibit signs of selection. There are many different methods for assessing signs selection, and these methods do not always agree [24]. However, a recent paper describes an aggregate or consensus method for assessing selection that shows promise in differentiating variants that are truly reflective of selection vs. those that are not [85]. In addition, Lohmuller et al. [86] have considered how admixed populations can be studied for evidence of selection in particular genomic regions, thus providing an additional level of annotation for variants identified in individuals with mixed ancestry.

Somatic Cell and Tumor Genome Annotations

There is growing interest in identifying somatic mutations that contribute to cancer [87]. Annotating somatic mutations for their properties involves determining if, in fact, a variant is somatically acquired. This involves comparing an individual’s tumor genome with his or her germline genomic DNA [88]. Databases that catalog somatic mutations in cancer exist, such as the COSMIC database [89]. These databases build off individual sequencing studies of tumors, such as those recently pursued by Link et al. [90] which identified a number of inherited and somatic mutations involved in a leukemia, and Ding et al. [88] which identified somatic variants contributing to a metastastic tumor which were verified in functional analyses of those variants via xenograft models. Annotating novel somatic mutations for their likely tumorigenic effect is not trivial but can be pursued through the assessment of patterns among genes and pathways observed to be mutated in large consortium studies of tumors such as the The Cancer Genome Atlas (TCGA) and the International Cancer Genome Consortium [91].

EXAMPLE STUDIES LEVERAGING ANNOTATION

There have been many very recent precedent-setting individual whole genome variant annotation studies (Table 3). Most of these studies have leveraged disease and trait association study-based annotations along with an assessment of overall ancestry of the individuals whose genome were sequenced. For example, J. Craig Venter’s genome and variations of relevance of to many phenotypes are discussed in his autobiography [92] as well as a paper describing the diploid assembly of his genome [15]. The genome of scientist Steven Quake was the subject of a recent report that estimated his lifetime risk of many common polygenic diseases as a function of variants identified in his genome that had been shown previously to be associated with these diseases [93]. Similarly, a recent study of an individual Charcot-Marie-Tooth neuropathy leveraged WGS to identify variants likely to contribute to that disease [94]. A similar study used WGS in a 4-member family and sophisticated variant annotation ‘filters’ to identify variants very likely to contribute to the Miller syndrome phenotype in the 2 offspring of the family [95].

Table 3.

Recent Individual Whole Genome Sequencing Studies with Variant Annotations

Individual	Reference	Platform	Annotations
JC Venter	Venter (2007) [92]; Levy et al. (2007) [15]	Sanger Sequencing	Disease, traits, ancestry
S. Quake	Ashley et al. (2010) [93]	Helicos	Disease, traits, ancestry
Family with Miller Syndrome	Roach et al. (2010) [95]	Complete Genomics, Inc.	Specific disease mutations
J. Lupski	Lupski et al. (2010) [94]	SOLiD	Specific disease mutations
NA19240	Moore et al. (2011) [11]	SOLiD	Disease, traits, ancestry
NA18507	Moore et al. (2011) [11]	SOLiD; Illumina	Disease, traits, ancestry
Anonymous Chinese Asian	Moore et al. (2011) [11]	Illumina	Disease, traits, ancestry
Anonymous Korean Asian	Moore et al. (2011) [11]	Illumina	Disease, traits, ancestry
J. Watson	Moore et al. (2011) [11]	Roche 454	Disease, traits, ancestry
NA07022	Moore et al. (2011) [11]	Complete Genomics	Disease, traits, ancestry
NA12878	Moore et al. (2011) [11]	SOLiD	Disease, traits, ancestry

Two additional studies have considered annotating the genomes of multiple individuals. Moore et al. [11] considered disease association-based annotation and ancestry of 10 recently published individual genomes. Pelak et al. [4] explored the frequency and likely functional impact of variants identified in 20 individual whole genomes and, although they did not consider detailed annotations of each individual, they did characterize the number of likely deleterious and novel mutations per genome and found that, on average, each human genome may carry as many as 144,000 unique variants.

The use of DNA sequencing and variant annotations in clinical decision-making has also recently witnessed some sensational successes. For example, in a 5 year old boy who had undergone more than 100 surgical operations, was almost constantly hospitalized and intermittently septic, whole exome sequencing proved to be life-saving. The idiopathic disease that this boy was suffering from proved to be from a heretofore not seen mutation in the gene XIAP, and led his team of physicians to have him undergo a bone marrow transplant. This patient is now thriving and represents the first case of an idiopathic disease diagnosed and treatment-guided by sequencing [96]. Similarly, there are examples in tumor sequencing studies in which the causative mutation, as determined by sequencing and contrast of the tumor and germline genomes, led to appropriate tailored therapy and an excellent therapeutic response [97, 98]. In aggregate, such cases are the first to document the remarkable potential of not only sequencing, but appropriate annotation of the data, to improve the prognosis of individuals with serious medical conditions.

FUTURE DIRECTIONS FOR ANNOTATING INDIVIDUAL GENOMES

It is quite clear that advances in DNA sequencing technology, as well as assembly and variant calling algorithms, will lead to a situation in which the generation of DNA sequence is not a major impediment to identifying variants that may contribute to disease. Rather, future genetics research will focus on interpreting and making sense of the variations identified from the use of such sequencing technologies [10, 99], for example via the strategies outlined in this review.

However, annotating variants is an activity that also has some inherent challenges. For example, humans are diploid, so that annotating variants may require understanding the sequence context within which variants reside. Thus, phase information from DNA sequencing studies may be necessary for appropriate functional annotations, but is not trivial to obtain currently [21, 23]. Importantly, any one of the annotation methods described here has limitations, as we have tried to make clear. In addition, depending on the application, integrating different types of annotation make sense and the best way to do this is an open question. For example, in clinical decision making settings involving novel variants, the knowledge that other variants in a gene in question have been shown to influence disease, the particular variant in question has been shown to likely damage the gene or genomic element it resides in, and the variant has been shown to be associated with the expression level of that gene in tissues of relevance to the disease, would make the case for the role of that variant in mediating a patient’s disease easier to make than if only evidence from a single annotation existed. Also, in the context of tumor variant annotations, it makes sense to ask what other variations may exist in the tumor beyond the one of focus, how those variants may impact gene function, and, further how the genes impacted by those variants influence a fundamental process or pathway that includes the gene with the primary variant of interest.

Facilitating the integration of variant annotations will thus be an important future activity. The creation of databases, graphical methods displaying integrated annotation results, and statistical analysis and query tools will all help. In the context of disease risk predictions, longitudinal and clinical studies exploring the role of multiple variants must be pursued. Furthermore, the annotation needs to be developed must be adapted for different uses. For example, the interpretative depth and needs for the research community are quite distinct from that of individual physicians or patients. As more individuals access their whole genomic sequence to determine their unique biologic features including disease susceptibility, pharmacogenomic interactions, and ancestry, a different set of annotation tools to represent this information will be required. For all such applications it will be vital for the dissemination of the information to be done in intuitively interpretable, reliable, and ethically-aware ways.