Reducing GWAS Complexity

ABSTRACT

Genome-wide association studies (GWAS) have revealed numerous genomic 'hits' associated with complex phenotypes. In most cases these hits, along with surrogate genetic variation as measure by numerous single nucleotide polymorphisms (SNPs) that are in linkage disequilibrium, are not in coding genes making assignment of functionality or causality intractable. Here we propose that fine-mapping along with the matching of risk SNPs at chromatin biofeatures lessen this complexity by reducing the number of candidate functional/causal SNPs. For example, we show here that only on average 2 SNPs per prostate cancer risk locus are likely candidates for functionality/causality; we further propose that this manageable number should be taken forward in mechanistic studies. The candidate SNPs can be looked up for each prostate cancer risk region in 2 recent publications in 2015^1,2 from our groups.

Genome-wide association studies (GWAS) of complex phenotypes became more and more powerful as sample sizes of cases and controls increased and meta-analyses were employed. Also, as next generation sequencing techniques became more feasible and increasingly affordable, more and more single nucleotide polymorphisms (SNPs) with lower and lower minor allele frequencies (MAFs) have been identified. Thus, association signals at any given locus have become increasingly complex in large part due to the many candidate risk SNPs, correlated with each other due to linkage disequilibrium (LD). Consequently, it is virtually impossible to assign functionality, let alone causality, to any given SNP at a risk locus. This dispiriting situation is only made more daunting by the unexpected finding that 90% or more of these risk SNPs are located in non-coding DNA.

To address these issues, we and others have used chromatin biofeatures to inform potential functionality on the original discovery SNPs, known to the field as index SNPs, and their many surrogate SNPs, the former revealed by GWAS and the latter defined by r² of population-specific LD. Thus, software tools such as FunciSNP,³ RegulomeDB,⁴ Haploreg,⁵ Annovar,⁶ IGV,⁷ and more recently FunSeq⁸ and motifbreakR⁹ were developed to utilize correlated risk SNPs from the 1000 Genomes Project, co-locating with chromatin annotations (such as obtained from ChIP-seq and nucleosome occupancy data); this significantly reduces the number of candidate functional SNPs. This became necessary since the SNP surrogates were plentiful (for example, there are on average ∼500 per prostate cancer risk region at r² ≥ 0.5 to the index SNP - Table 1). Over the last couple of years, we have successfully used this annotation approach for prostate¹⁰ and breast¹¹ cancer risk regions. Despite the significant reduction in candidate functional SNPs using this approach, the on average ∼10 (median = 5) candidate SNPs per prostate cancer risk locus still make a detailed and comprehensive wet-lab analysis of functionality intractable (Table 1). However, more recently a new approach emerged, known as fine-mapping.

Table 1.

SNPs per PCa risk region.

Selection Criteria	Average (Median) # of SNPs/region
All SNPs (1K) at 1MB/region	∼50 thousand (MAF ≥ 1%)10
SNPs r2 ≥ 0.5 with index SNP	∼50010
SNPs r2 ≥ 0.5 in biofeatures	9.8(5)10
SNPs significantly fine mapped	12(6)1 – 22(13)2
SNPs significantly fine mapped AND in biofeatures	3.5(2)1 – 5.4 (3)2 ∼2 common in the 2 independent studies

Grant support: R01 CA 136924, R01 CA 190182 & UO1 CA 184826

With the advent of fine-mapping strategies and corresponding analysis methods, the complexity at any given locus or region can be reduced further. Many new analysis aproaches have been developed to go beyond simple rankings of marginal p-values to statistically identify a putative set of candidate SNPs for further functional analysis. These methods include Bayesian and re-sampling approaches that formally incorporate the uncertainty in estimation and ranking,^1,12,13 model selection approaches to condition on multiple SNPs in the region using either individual-level data^2,14 or marginal test statistics,^15,16 and approaches that formally incorporate prior information, such as SNP annotation, into the final inference.^17,18

Recently, we and others have fine-mapped prostate cancer risk regions using a multi-ethnic¹ and a single large European population.² This reduced the candidate risk SNPs per region (Table 1). Coupled with FunciSNP³ annotated functionality, the 2 fine mapping studies further reduced the number of common candidates to on average only about 2 SNPs per region (Table 1). This clearly means that without functional annotation or fine-mapping a significant number of false positives at each locus may well lead to non-productive functional analyses that have little to do with risk. Ultimately SNPs must be functional to be causal, but not all functional SNPs are inevitably involved in risk.

However, even with fine-mapping data at hand, downstream analyses will be complicated when multiple biologically functional SNPs with measurable allelic effects are present at a single locus, of which only one is the major driver of cancer. Also plausible are haplotypes with 2 or more variants partially complementing each other. Finally, an additional complication is that the etiological target tissue of a given risk locus may lie outside the tissue of origin for a particular disease. To give an example, risk for prostate cancers (and others) are likely affected by immune cells in addition to dysregulation in prostate epithelia. This may explain the roughly 1/3 of risk loci where no functionality may be assigned based upon extant epigenomics data in prostate epithelial cell types.¹⁰ To give another example, there may be several plausible tissues of origin, which may be at least partially addressed by assessing candidate loci for enrichment in the epigenomic features of those various tissues. We showed for serous ovarian cancers that finemapped risk SNPs were enriched in regulatory sites of immortalized primary cell lines derived from fallopian tube serous epithelium (FTSE) over those of similar cell lines derived from ovarian serous epithelia, supporting FTSE as the tissue of origin.¹⁹ While this by no means settles the debate on such questions, it may provide an important clue as to the tissue of origin and suggest therapeutic targets for prevention and early intervention measures. Going forward, we anticipate that comparison among multiple candidate tissues (or all known cell types) will become standard in the field. Thus, in the midst of so much uncertainty, it is essential that biological assays give repeatable and reliable measures of these complex interactions. It must also be stressed that functional annotations alone are clearly not comprehensive and that other, as yet unknown, chromatin-related functions likely are to be considered in the future. The ultimate goal would be to identify the true functional/causal SNP (or allele with more than 1 SNP or even multiple causal alleles) at every risk locus. Attainment of this goal may require development of specific assays designed to measure the allelic affect on cancer processes once the variant functionality vis-a-vis the epigenome (or other) has been verified. The eventual utility from this will be the compilation of more informative nomograms for risk assessments and the identification of risk mechanisms.

Disclosure of potential conflicts of interest

No potential conflicts of interest were disclosed.