Functional single nucleotide polymorphism-based association studies

Abstract

Association studies hold great promise for the elucidation of the genetic basis of diseases. Studies based on functional single nucleotide polymorphisms (SNPs) or on linkage disequilibrium (LD) represent two main types of designs. LD-based association studies can be comprehensive for common causative variants, but they perform poorly for rare alleles. Conversely, functional SNP-based studies are efficient because they focus on the SNPs with the highest a priori chance of being associated. Our poor ability to predict the functional effect of SNPs, however, hampers attempts to make these studies comprehensive. Recent progress in comparative genomics, and evidence that functional elements tend to lie in conserved regions, promises to change the landscape, permitting functional SNP association studies to be carried out that comprehensively assess common and rare alleles. SNP genotyping technologies are already sufficient for such studies, but studies will require continued genomic sequencing of multiple species, research on the functional role of conserved sequences and additional SNP discovery and validation efforts (including targeted SNP discovery to identify the rare alleles in functional regions). With these resources, we expect that comprehensive functional SNP association studies will soon be possible.

Introduction

Association studies of common, complexly inherited human diseases have the potential to provide us with insights into causes of enormous human suffering [1]. While thousands of such studies have been published (typically using single nucleotide polymorphisms [SNPs]), only a handful of these finding have been clearly and consistently replicated. While some findings are doubtless real, [2] debate continues over most. There are only a small number of genetic variants that have been clearly and consistently associated with a common disease, many of which are listed in Table 1.

Table 1 Some clear, consistent common disease associations.

Types of association studies

Researchers, typically, carefully weigh comprehensiveness and efficiency in designing an association study. A highly comprehensive study would assess every variant in the region(s) under study, regardless of type, location and allele frequency. A highly efficient study would be designed to reduce costs, including genotyping and/or multiple testing costs. Genotyping costs can be saved by determining which SNPs are in linkage disequilibrium (LD). For example, if you knew that two SNPs were in complete LD in the specific population of interest, you would only need to genotype one to assess them both. Multiple testing costs can be reduced by only looking at SNPs with a high a priori chance of being associated. Note that as multiple testing correction should account for the effective number of independent tests performed, genotyping only one of two SNPs in complete LD does not reduce multiple testing costs; if the SNPs are in complete LD, only one effective independent test is being performed, regardless of whether one or two SNPs are genotyped (Bonferroni correction is overly conservative). As 'per SNP' genotyping costs continue to fall, it seems likely that multiple testing costs will become the predominant concern in efficiency. Therefore, we discuss efficiency in terms of the a priori likelihood for an SNP to be associated with the phenotype studied.

Different types of large-scale association studies and the balance they strike are shown in Figure 1, although, obviously, many studies are hybrids of these types. These approaches, which have been applied to candidate genes, regions and recently to the whole genome, [21, 76] are discussed in detail below, along with another technique (re-sequencing), which can currently only be applied on a small scale. Additional techniques that may be useful in 'special' populations, such as isolated founder and admixed populations, are discussed elsewhere [77–79].

Figure 1

Association study approaches: Efficiency versus comprehensiveness. Studies vary in their efficiency (the a priori likelihood of a tested single nucleotide polymorphism [SNP] being associated with a disease), which has an impact on genotyping and multiple testing costs. Highly efficient designs (as defined by multiple testing costs) are shown on the right, with less efficient designs on the left. Studies also vary in comprehensiveness, both in terms of the allele frequency spectrum assessed (A) and the extent the region under study is assessed (B). Highly comprehensive studies extend from top to bottom. The efficiency (or comprehensiveness) for a specific study type relative to another in this figure is certainly not meant to be quantitative but merely indicative of the direction (bigger or smaller). This figure is applicable to large-scale studies of candidate genes, regions or the whole genome. Different functional SNP approaches are represented in blue, while non-functional approaches are represented in green. Re-sequencing is currently only feasible for examining one or a few candidate genes and is therefore not depicted. (A) Using linkage disequilibrium (LD) approaches, rare alleles are less likely to be tagged and hence the rare allele region is not covered. Since non-synonymous SNPs (nsSNPs) are assessed directly, association with rare alleles can be readily detected; however, this is limited by the availability of these SNPs. The light colour in the rare allele region is to indicate that coverage is dependent on SNP discovery. In this figure, we consider the most obvious functional SNPs, the nsSNPs. We presume the efficiency of the other functional categories may be significantly lower. (B) Typically, there is a trade-off between efficiency and comprehensiveness. One may limit the study to nsSNPs in order to have high efficiency at the cost of comprehensiveness. Further increase in the efficiency (and decrease in comprehensiveness) can be achieved by focusing only on nsSNPs predicted or known to have a functional consequence. Similarly, it has been proposed that a study utilising SNPs that tag the highest number of other SNPs (ie SNPs in high LD regions) would be more efficient (but less comprehensive) than a study aiming at LD coverage of the full genome [75].

Re-sequencing

When there is strong a priori evidence that a gene may be involved in a disease, it is possible to sequence that gene in cases and controls [43, 80, 81]. This requires no prior knowledge of variants in the region and allows researchers comprehensively to evaluate all variants in a gene, regardless of their allele frequency. Usually, it is necessary to group the very rare variants (< 1 per cent) for power considerations [43, 80, 81]. While this approach is now possible for one or a few candidate genes, it is by no means comprehensive across the genome and dramatic reductions in sequencing costs are necessary for its implementation on a large scale [82–84].

LD

Given the high rate of LD in the genome, many variants do not need to be directly genotyped in order to be assessed. They may instead be assessed by genotyping another SNP in high LD. The goal of LD-based ('tagging') approaches is to test a sufficient number of common SNPs so that SNPs that are not directly tested are assessed through their high correlation with the genotyped SNPs. This can create efficiency in genotyping but does not reduce multiple testing costs (as discussed previously, multiple testing corrections should account for the effective number of independent tests, rather than the number of SNPs genotyped). Additionally, the efficiency of the approach is modest, since there is a low a priori chance that a specific assessed SNP is associated with disease. By focusing only on regions with high LD (in which a single SNP is likely to tag several other SNPs), one improves the efficiency because there is an increased likelihood for any assessed SNP (ie for one test) to be tagging a functional SNP that is associated with the phenotype of interest [75].

Tagging allows most common SNPs to be comprehensively assessed in linkage regions, [85] candidate genes [86] or the whole genome [87]. Tagging, however, is not comprehensive in terms of allele frequencies because it tends to work poorly on rare polymorphisms [88–92]. Given the clear importance of rare polymorphisms (Box 1), this presents a substantial drawback. While some analytical work suggests that long haplotypes may be used to achieve a degree of 'tagging' of the rare allele, this comes with a dramatic multiple testing cost [106]. The adequate assessment of rare alleles requires direct interrogation.

Functional SNPs

Functional variants are the most likely to be associated with diseases (in fact, non-functional variants should only be associated secondary to LD); therefore, genotyping studies using only functional SNPs are relatively efficient. Since these variants are directly assessed, these studies are comprehensive in terms of allele frequency, covering rare and common variants present in the databases or discovered during focused SNP discovery. Our poor ability to predict functional SNPs, however, means that this approach is generally far from comprehensive in terms of coverage of the region under study. Nevertheless, by focusing on the most obvious classes of potentially functional SNPs, such as those causing non-synonymous changes in proteins, researchers have had notable successes with association studies in candidate genes [107] or linkage regions [3, 22]. It is now possible to apply this method on a genome-wide scale, [75, 108] which increases comprehensiveness with some reduction in efficiency.

Extending the (potentially) functional SNP approach

There are many attractive features of the functional SNP approach, including its efficiency and ability to assess rare and common alleles. Additionally, a positive association automatically provides a candidate causative polymorphism.

A major criticism of the functional approach is its lack of comprehensiveness, [96] and extending the coverage has been difficult, given our poor ability to predict functional SNPs. We can, however, broadly define functional SNPs as SNPs in any class predicted to have an above-average chance of having a functional effect. Recent progress in comparative genomics is likely to dramatically increase the comprehensiveness of this approach.

Below, we address some traditional functional elements (non-synonymous, splicing and promoter SNPs), as well as functional sequences emerging from the study of genome conservation.

Non-synonymous

The most obvious class of potentially functional SNPs is those causing non-synonymous changes in proteins (nsSNPs). Over 60 per cent of known Mendelian disease mutations and almost all the consistent, common disease mutations in Table 1 involve nsSNPs [109]. While there is a clear ascertainment bias for studying and confirming associations with nsSNPs, they are inarguably important in disease.

Additional evidence that many nsSNPs are functional and subject to selection comes from candidate gene sequencing studies, which find that 60 per cent of the expected number of nsSNPs are missing [110, 111]. Furthermore, nsSNPs have lower minor allele frequencies than do synonymous SNPs [110, 111]. When we examined all coding SNPs currently in the SNP database (dbSNP), we also found a dearth of nsSNPs; these are expected to comprise two-thirds of coding SNPs [111] but instead comprised less than one-half (20,463 nsSNP out of 42,387 coding SNPs). The deficiency of nsSNPs was even more notable when the analysis was limited to conserved coding regions in which only one-third of SNPs were non-synonymous (8,828 of 23,397). (SNP definitions were derived from the Ensembl database, and conserved regions were as defined previously [112].)

Large-scale studies of nsSNPs maintain high efficiency while allowing reasonable coverage [75]. One could choose to further increase efficiency (and decrease comprehensiveness) by limiting a study only to nsSNPs with a high predicted likelihood of being damaging. A substantial proportion of such SNPs have already been implicated in human disease [103, 113].

Splicing

Perhaps the next most obvious class of potentially functional variants is SNPs around splice junctions. Mutations that affect splicing underlie 15 per cent of mutations in Mendelian diseases and hence are likely to play some role in common diseases [114].

Splicing is catalysed by weakly conserved 5' and 3' splice sites and a branch site, as well as exonic and intronic enhancers and silencers. Sites far from splice junctions can affect splicing, and a few mutations in these distant sites have been shown to cause human disease [115–120]. It appears, however, that most control of splicing lies in the 20 base pairs (bp) flanking each side of exon - intron boundaries [120]. These regions contain a high density of splicing enhancers (SEs), [120] have fewer SNPs than sequences further from splice junctions [120] and contain most of the known splicing mutations [114]. We find that these sequences are significantly conserved and have a relative dearth of SNPs (Table 2).

Table 2 Conservation and relative single nucleotide polymorphism (SNP) density in different types of functional regions.

Rather than testing all SNPs within the vicinity of a splice junction, one could increase efficiency by limiting the analysis to SNPs specifically predicted by computational models to affect splicing [121, 122]. Conversely, one can increase comprehensiveness by assessing SEs beyond 20 bp of splice junctions. SEs are most prevalent in exons [123, 124]. Some synonymous SNPs have also been shown to alter splicing [122]. Several programs are now available to predict SEs [125, 126]. In addition to SNPs within 20 bp of the junction, the interrogation of synonymous SNPs predicted to disrupt SE activity [126] increases study comprehensiveness.

Promoters

Promoters are cis-elements that lie upstream of transcription start sites and are responsible for transcription initiation [127]. The existence of regulatory variants affecting transcription has long been established [128, 129] and that have been shown to play a role in human disease [130, 131].

Even though the exact promoter sequence may not be easily discerned, recent work has shown that the 500 bp upstream from the transcription start site is almost always able to function as a promoter [132]. Defining the promoter, however, requires determining the 5' end of transcripts, which is typically done experimentally and hence is laborious [133–135]. As shown in Table 2, conservation in the promoter sequences is threefold higher than expected.

In addition to promoters, numerous other cis-acting elements (for example enhancers) contribute to gene regulation. These elements have been more difficult to identify because they can lie within coding sequences, introns or as far as 1 megabase away [120, 136, 137]. Defining these elements is a main goal of the ENCODE project [138]. Genomic work aimed at identifying transcription factor binding sites and other regulatory sequences experimentally and informatically is ongoing, [87, 139, 140] and study of conserved sequences holds promise for the identification of these regions.

Conserved sequences

Computational efforts have consistently found that approximately 5 per cent of the human genome shows conservation with other species [112, 141–148]. Although some regions may be conserved due to low mutation rates, clearly many, and perhaps most, of these regions are functionally important [149]. Indeed, most coding exons and many untranslated regions show interspecies conservation, although these only account for a minority of conserved regions. Conserved elements have been show to affect gene transcription levels, [150–156] RNA editing [112] and genome stability [157]. Additionally, conserved regions are enriched in intronic stretches surrounding alternatively spliced exons and have an excess of predicted secondary structure [112, 143, 158] and matrix-scaffold attachment regions [159]. Furthermore, they are enriched in stable gene deserts, which have been postulated to contain long range cis-regulatory regions [112]. Two lines of evidence suggest that many SNPs in conserved regions are subject to selection and, hence, are presumably functional: these regions contain a relative dearth of SNPs (Table 2), and the SNPs present there show a shift in allele frequency distribution towards rarer alleles [160, 161].

The identification of conserved non-coding elements has generated a paradigm shift for the definition of functional elements. Without knowing the exact function of each element, sequences conserved across species define a map of likely functional regions in the genome and SNPs in the regions are candidates for functional SNP association studies.

The study of conserved regions is a vibrant field, with diverse methods of defining conservation and views on the correct number and types of species to compare. Some groups have focused on very large regions while others have examined conservation of regions as small as 4 bp [112, 143, 144]. Analyses can be performed using very closely related species (such as primates) or very distant species (such as a range of eukaryotes) [112, 143, 144]. The study of species that are moderately distant (< 75 million years) has yielded many of the conserved elements, [162] while study of primates has provided insight on primate-specific regulatory elements [146]. In addition to identifying conserved elements subject to purifying selection, comparative genomics has identified genes with evidence of positive selection [163, 164]. Similar analyses may eventually be able to identify non-coding elements subjected to positive selection.

The proportion of functional elements that can be identified by comparative genomics is not yet clear. In a study using sequences from multiple yeast species, essentially all the known non-coding regulatory regions were identified as conserved [157]. Another study in yeast could identify conserved elements at the resolution of 6 bp transcription factor binding sites [165]. In mammals, using the currently available genomic sequences, most of the coding sequences and known regulatory sequences are conserved [166]. The analysis of more mammalian genome sequences will undoubtedly refine the current picture of conserved elements, although it is not clear that it will reach the same resolution achieved in yeast [162]. Nevertheless, it is likely that some functional sequences may not be identified through comparative genomics. If these SNPs do not fall into another obvious class of functional elements (like promoter regions), they may be missed by function-based association studies.

Generating a whole genome set of functional SNPs

The current feasibility of genome-wide function association studies depends upon the total number of functional SNPs and the extent to which such SNPs are represented in the databases. In the following discussion, we define functional SNPs as SNPs that fall into any of the above classes (ie non-synonymous, splicing, promoter, conserved [112]). Ongoing improvements in the definition of conserved regions may slightly change these estimates.

To estimate the total number of functional SNPs, we have utilised publicly available data from ENCODE regions. Ten regions (500 kilobases each) were re-sequenced in 48 unrelated individuals (16 Yoruba, 16 Centre D'Etude Du Polymorphisme Humain [CEPH], eight Han Chinese and eight Japanese). The SNPs in these regions, including those already present in the dbSNP and those newly discovered in sequencing, were then genotyped in the full 270 HapMap samples.

We first determined the total number of functional SNPs currently in dbSNP (using the above definitions). We then used the ENCODE regions to determine the allele frequency distribution (ie percentage rare and common) of conserved-region SNPs already in the dbSNP (ignoring those newly discovered by the ENCODE re-sequencing effort). We subsequently used information on the newly discovered ENCODE SNPs and our internal SNP discovery efforts to infer the percentage of SNPs missing from the dbSNP. This allowed us finally to estimate the total number of such SNPs. Implicit in this estimation is that the distribution of the allele frequency of functional SNPs is the same as the distribution of the subset of these SNPs that are in conserved elements (which account for over 75 per cent of the functional SNPs).

There are approximately 380,000 functional SNPs in dbSNP build 124. We infer from the ENCODE data that approximately 190,000 of these are common and 85,000 are rare (the remaining SNPs are very rare or database errors). Results were similar using data from both the CEPH and Yoruban samples. These results differ markedly from the expectations under the standard neutral model that there should be similar numbers of rare and common SNPs, suggesting that rare SNPs are missing in the dbSNP database [167]. Of the conserved region SNPs detected in the ENCODE Yoruban samples, the dbSNP database contained 23 per cent of the rare and 55 per cent of the common SNPs. Coverage was higher for conserved-region SNPs detected in the ENCODE CEPH samples, as the dbSNP database contained 35 per cent of the rare as well as 71 per cent of the common SNPs. Given that limited numbers of chromosomes typically are used for SNP discovery, both the dbSNP database and ENCODE are biased to miss rare SNPs^a. The extent of this bias estimated using our internal SNP discovery efforts suggests that dbSNP coverage of rare SNPs is between approximately 25 per cent (in Caucasian) and approximately 15 per cent (in African).

From the above data, we estimate that there are approximately 350,000 common and 570,000 rare functional SNPs in the Yoruban samples and 270,000 common and 340,000 rare functional SNPs in the CEPH samples. Hence, a study that assayed only common functional SNPs would require a similar number of SNPs as an LD tagging study [161, 168]. Even greater genotyping efficiency could be found by combining the approaches. Additionally, the number of rare functional SNPs is within the ability of new genotyping technologies [98, 99, 169].

Discussion

Association studies based on functional SNPs are highly efficient as they study the set of SNPs most likely to cause disease. In the past, these studies have been criticised as not being comprehensive due to our incomplete knowledge of the functional elements of the human genome. Research into conserved sequences and the continuing influx of genomic sequences into the public domain promises to delineate many of these elements and increase the comprehensiveness of functional SNP association studies. The use of functionalbased association studies can, in principle, adequately assess rare alleles, poor coverage of which is a major drawback for LD-based association studies.

It may be possible to improve the balance between the comprehensiveness and efficiency (defined in terms of multiple testing costs) of a functional SNP-based study by incorporating the a priori probability that an SNP is functional into the statistical tests used for analysis. For instance, one might set a less stringent p-value threshold for a nonsense SNP than for one in a putative promoter. Additionally, one might set a lower p-value threshold for an SNP that was in two functional categories rather than in a single functional category. For example, Table 3 indicates that SNP density (which over the whole genome probably reflects selection and, hence, functionality) is particularly low in coding regions that are also conserved or flank splice junctions.

Table 3 SNP density per kilobase (kb) and counts in different types of functional regions.

For comprehensive functional-based association studies to become practical, several goals need to be accomplished. First, the definition of functional elements needs to be refined through the availability of more genomic sequences. Secondly, SNP discovery efforts must be continued and expanded. Targeted re-sequencing in the functional regions may be necessary in order to compensate for bias against rare alleles in the databases, especially those that are population-specific and hence more likely to be functional [105]. The availability of extra sequencing capacity and efficient SNP discovery technologies can help to achieve this goal [170]. Thirdly, SNPs must be genotyped in the major ethnic populations to determine allele frequencies. HapMap now includes millions of SNPs, although these are biased to common SNPs [161]. Given the high-throughput genotyping technologies available, testing additional candidate functional SNPs to identify the common and rare SNPs can be readily performed. Indeed, we have recently undertaken the task of genotyping approximately 30,000 nsSNPs from the public databases to identify a set of approximately 20,000 that are polymorphic in at least one population [105].

With the availability of the functional elements and the SNPs, only approximately 270,000 - 350,000 SNPs must be genotyped to assess common functional SNPs in the genome. Furthermore, the genotyping of 300,000 - 500,000 additional SNPs will allow assessment of rare functional SNPs which have been implicated in many common diseases and are inadequately assessed by other approaches.

Box 1. Common variant/common disease versus rare variant/common disease

For the purposes of this review, we use the standard definition of a polymorphism as a variant whose minor allele frequency (MAF) is above 1 per cent, and define common alleles/polymorphisms as those with MAF > 10 per cent, rare alleles/polymorphisms as those with MAF 1 - 10 per cent and very rare alleles/variants as those with MAF < 1 per cent. In the past decade, there has been substantial debate over the importance of common alleles versus rare alleles (or even very rare variants) in common, complex human diseases. Theoretical work has been used to argue all points of view: that causative common disease alleles are most likely common alleles, or rare alleles, or very rare alleles [93–95].

One key argument for common alleles relies on the perceived greater practical difficulties in studying rare alleles rather than common alleles. First, analysis methods are particularly sensitive to genotyping errors of rare alleles and rare alleles have been particularly prone to genotyping errors [96, 97]. Recent improvements in genotyping technologies, however, dramatically lessen these concerns [98, 99]. Secondly, rare alleles are more likely to be population specific and therefore are more likely to generate spurious associations due to population substructure. Again, improvements, this time to analytical methods, allow us to detect and adjust for these artifacts [100, 101]. Thirdly, it has been argued that the power to detect associations with rare alleles appears low when compared with that to detect common alleles. While this is certainly true if one assumes the same genotypic relative risk, this assumption is arbitrary, and if one instead uses another arbitrary assumption of equal population attributable risk, then the power to detect rare alleles would be significantly better than that for common alleles. Probably, a more reasonable approach is to consider a specific genetic effect size (eg defined by likelihood of the odds (LOD) score in sibling-pair analysis) of a locus and assume that causative alleles generate this specific effect size [102]. Given this assumption, the power to detect common and rare alleles is fairly similar (data not shown). Finally, rare alleles are difficult to 'tag' and therefore need to be assessed directly, creating two problems: alleles must be in databases in order to be assessed and genotyping all of the rare alleles in the genome would be at least an order of magnitude larger than contemplated for the linkage disequilibrium (LD)-based approach for common alleles. These concerns, while substantial, may be addressed by single nucleotide polymorphism (SNP) discovery and focusing genotyping efforts on rare SNPs that are also potentially functional.

One theoretical argument for rare alleles is that purifying selection should keep the frequency of deleterious functional alleles low. Indeed, in a study of approximately 30,000 non-synonymous SNPs, we confirmed previous observations that SNPs predicted by PolyPhen [103, 104] to be damaging have significantly lower allele frequencies than SNPs predicted to be benign. This effect is largely due to an enrichment of damaging SNPs in the MAF < 10 per cent category [105].

Perhaps the strongest argument comes from an examination of Table 1, which indicates that both common and rare alleles are important. In light of these data, it is clearly essential for common disease association studies to investigate rare, as well as common, alleles.

End notes

^a SNP discovery efforts interrogate a limited number of individuals and hence are more likely to find a common minor allele than a rare minor allele. For example, a study using only one individual (two chromosomes) has a 50 per cent chance of including both alleles of a 50 per cent allele frequency SNP, but only a 2 per cent chance of finding both alleles of a 1 per cent frequency SNP. Hence 1 per cent alleles are more likely to be missed in both dbSNP and the targeted re-sequencing than 10 per cent alleles. In addition, SNPs in dbSNP and those identified in this targeted re-sequencing effort are more biased to be more common in a different ethnic population where they may have been discovered. Indeed when studying alleles that are rare in the Caucasian population, we found the frequency in other populations to be higher for SNPs already in dbSNP than for SNPs identified through SNP discovery in the Caucasian population (MF unpublished results).