SNPsea: an algorithm to identify cell types, tissues and pathways affected by risk loci

Abstract

Summary: We created a fast, robust and general C++ implementation of a single-nucleotide polymorphism (SNP) set enrichment algorithm to identify cell types, tissues and pathways affected by risk loci. It tests trait-associated genomic loci for enrichment of specificity to conditions (cell types, tissues and pathways). We use a non-parametric statistical approach to compute empirical P-values by comparison with null SNP sets. As a proof of concept, we present novel applications of our method to four sets of genome-wide significant SNPs associated with red blood cell count, multiple sclerosis, celiac disease and HDL cholesterol.

Availability and implementation: http://broadinstitute.org/mpg/snpsea

Contact: soumya@broadinstitute.org

Supplementary information: Supplementary data are available at Bioinformatics online.

1 INTRODUCTION

As genome-wide association studies (GWAS) continue to find disease alleles, investigators seek to identify the set of pathways and tissue types affected by these alleles, and the physiological conditions under which they act (Elbers et al., 2009; Lango Allen et al., 2010; Raychaudhuri, 2011; Wang et al., 2013; Yaspan and Veatch, 2011). For example, we have previously presented statistical methods to identify immune cell types for further functional investigation by finding cell type-specific expression of genes in linkage disequilibrium (LD) with autoimmune disease-associated single-nucleotide polymorphisms (SNPs) (Hu et al., 2011). Presumably, alleles influence disease risk through pathways specific to these cell types.

We sought a general implementation of these methods to leverage data from high-throughput functional assays that assess genome-wide transcription, protein binding, epigenetic modifications and other functional parameters across diverse cellular conditions and tissue types. Each of these diverse data types can be represented as a continuous matrix of genes and conditions (e.g. cell types, tissues, pathways, experimental conditions). Databases such as Gene Ontology (GO) (Botstein et al., 2000) offer expert-defined pathways and complementary gene annotations that can be represented as binary values.

Investigators have already described strategies to assess enrichment of GWA results for pathways or gene sets but not for condition specificity (Holden et al., 2008; Weng et al., 2011). In contrast to these methods, we do not require genotypes, P-values, a priori gene sets or pathways or a priori definitions of gene–SNP associations. We require only a list of SNP identifiers, use LD structures to identify plausibly influential genes and use a simple sampling approach to identify the conditions they influence.

SNPsea is a general algorithm to identify the conditions relevant to a trait by assessing the genes within associated loci for enrichment of condition specificity.

2 METHODS

For a given set of SNPs, SNPsea tests genes implicated by LD, in aggregate, for enrichment of specificity to a condition in a given matrix of genes and conditions. The matrix must be normalized so that conditions are comparable.

First, we identify genes implicated by each SNP using LD from reference genomes. Second, we calculate a specificity score for each condition with these genes. Finally, we compare these scores with scores obtained with null sets of matched SNP sets to calculate an empirical P-value for each condition (see Supplementary Notes for algorithm details).

We empirically calculate P-values because we previously found that analytical distributions can result in inaccurate P-values (Hu et al., 2011). SNP linkage intervals, gene densities, gene sizes and gene functions are correlated across the genome and are challenging to model analytically.

We used C++ for fast computation of P-values because Python was prohibitively slow. The online reference manual details compilation and installation procedures; we also provide executable files for immediate use on select platforms.

2.1 Multiple genes implicated by LD

Accurate analyses must address the critical issue that SNPs from GWA studies frequently implicate more than one gene (50% of GWAS Catalog SNPs, Supplementary Fig. S2).

We defined LD intervals with SNPs from the 1000 Genomes Project (EUR) (Genomes Project Consortium, 2010) and a previously described strategy (Supplementary Fig. S1) (Rossin et al., 2011). A SNP implicates genes overlapping its LD interval, defined by the furthest SNPs in a 1 Mb window with r² > 0.5. To ensure the associated genes are included, we extend each interval to the nearest recombination hotspots with recombination rate >3 cM/Mb (HapMap3) (Myers et al., 2005). We merge SNPs with shared genes into a single locus.

By default, we assume that each associated locus harbours a single influential gene rather than multiple genes. We provide an alternative scoring method to account for multiple genes (Supplementary Notes) that produces similar results in four traits we tested (Supplementary Fig. S4).

Because interval lengths depend on the choice of r² threshold, we looked for an effect of this choice (Supplementary Fig. S3). The significant result for the Gene Atlas and blood cell count SNPs is robust to different thresholds. Similarly, the choice of r² threshold has little effect on the significant GO enrichment result for these SNPs.

2.2 Type I error estimates

We tested 10 000 sets of 100 randomly selected LD-pruned SNPs. For each condition (tissue or GO term), we observed appropriate proportions of P-values <0.5, 0.1, 0.05, 0.01 and 0.005 (Supplementary Fig. S5).

3 EXAMPLES

We used SNPsea to identify tissues relevant to blood cell count by testing 45 genome-wide significant SNPs (van der Harst et al., 2012) with expression data (Gene Atlas) for 17 581 genes across 79 human tissues (Su et al., 2004). Bone marrow CD71+ early erythroid cells are significantly enriched for cell type-specific expression of the genes within the trait-associated loci (P = 2 × 10⁻⁷) (Fig. 1).

Fig. 1.

Empirical P-values for specificity to each condition. 25 of 79 tissues (Gene Atlas) are shown. Adjacent: Pearson correlation coefficients for pairs of expression profiles ordered by hierarchical clustering with UPGMA

The genes in these loci are enriched for the term hemopoiesis (GO:0030097) (P = 2 × 10⁻⁵) (Supplementary Fig. S6), suggesting that blood cell count may be influenced by the genes expressed specifically in early erythroid cells and involved in forming blood cellular components.

We provide additional examples for SNPs associated with multiple sclerosis, celiac disease and HDL cholesterol. Each includes Gene Atlas and GO enrichments, r² comparisons and comparisons of results assuming a single or multiple causal genes (Supplementary Figs S7–9).

Funding: The National Institutes of Health (5K08AR055688, 1R01AR062886-01, 1U01HG0070033, T32 HG002295/HG/NHGRI, 7T32HG002295-10), the Arthritis Foundation and the Doris Duke Foundation.

Conflict of Interest: none declared.