Introduction
Functional annotations provide valuable information for prioritizing potential causal variants within complex-trait loci identified through genome-wide association studies (GWASs).1–12 Profiles of such functional genomic annotations, including transcription factor binding sites and open chromatin regions from hundreds of cell types, are rapidly becoming available.13–15 But, the most informative annotation is not always known. The most informative genomic annotations for fine mapping a particular trait are most likely related to mechanisms important for that trait. For example, binding sites for transcription factors that regulate key pathogenic pathways might prioritize variants for diseases,5,7 and promoters active in a specific cell type might be able to prioritize expression quantitative trait locus (eQTL) variants from that cell type. Informative annotations such as these can be used for identifying likely causal variants, and these variants can then in turn be functionally interrogated for elucidating mechanisms underlying the trait.
Identifying the most informative annotations requires a robust statistical strategy that controls for two important types of potentially confounding genomic features: (1) local structure of genetic variation near SNP associations and (2) colocalization of multiple functional genomic annotations. First, trait-associated SNPs often map to regions with greater gene density, genetic variation, and linkage disequilibrium (LD) than the rest of the genome. Second, functional annotations that colocalize are often enriched within trait-associated loci. For example, DNase-I hypersensitive sites (DHSs) colocalize with exons,16,17 and regulatory elements cluster together near and within gene transcripts. Therefore, an observed enrichment of one annotation might be the consequence of unaccounted colocalization with other annotation, thus confounding inferences of causality.
We developed Genomic Annotation Shifter (GoShifter), an enrichment test that controls for local genomic structure. GoShifter employs an intuitive method that locally shifts sites of tested features within each locus to generate a null distribution of annotations overlapping associated variants by chance. Other methods, such as Genome Structure Correction (GSC), assess the relationships between genomic features by shifting them.1,18–20 Although GSC can assess the significance of overlap between two genomic features, it does not provide a clear application to individual GWAS loci and their local LD structure. Here, we apply the shifting approach to identify informative annotations for fine mapping GWAS loci. We benchmark the performance of GoShifter against that of commonly employed matching-based methods. These methods rely on inferring the enrichment of the SNP-annotation overlap in the observed data by contrasting it with the overlap in the null set of SNPs derived by random sampling of variants from the genome. In order to control for plausible genomic confounders, these methods sample SNPs by matching for a selection of defined genomic parameters. The selection of these parameters is based on the assumptions about possible analytical confounders. In contrast, GoShifter does not require prior knowledge because the null distribution is derived within the tested loci, ensuring that the density of SNPs, annotations, and the spatial distribution of genomic features are preserved.
We show that compared with commonly used SNP-matching-based methods, GoShifter is able to robustly identify informative annotations under a range of different scenarios. We show that matching-based approaches are prone to inflating observed enrichment values: we highlight that the lack of matching on SNPs in LD can lead to misleading results. Furthermore, we implemented a stratified test to distinguish contributions from two colocalized annotations. Using the local-shifting approach, GoShifter allows prioritization of loci by determining the most informative functional variants driving the observed enrichment.