Low-Frequency and Rare-Coding Variation Contributes to Multiple Sclerosis Risk

Abstract

Multiple sclerosis is a complex neurological disease, with ∼20% of risk heritability attributable to common genetic variants, including >230 identified by genome-wide association studies. Multiple strands of evidence suggest that much of the remaining heritability is also due to additive effects of common variants rather than epistasis between these variants or mutations exclusive to individual families. Here, we show in 68,379 cases and controls that up to 5% of this heritability is explained by low-frequency variation in gene coding sequence. We identify four novel genes driving MS risk independently of common-variant signals, highlighting key pathogenic roles for regulatory T cell homeostasis and regulation, IFNγ biology, and NFκB signaling. As low-frequency variants do not show substantial linkage disequilibrium with other variants, and as coding variants are more interpretable and experimentally tractable than non-coding variation, our discoveries constitute a rich resource for dissecting the pathobiology of MS.

Graphical abstract

Figure S1

Data Quality Overview, Related to STAR Methods (A) QC process. We assembled 42 cohorts of data (either entire country-level collections or groups of samples processed as a batch; Table S1). We called common variant genotypes with the standard algorithm provided by Illumina (GenCall), and low-frequency variants with zCall, an algorithm specifically developed to call these variants on the exome chip (Goldstein et al., 2012). We performed initial quality control on each cohort separately to account for variation between batches and cohorts (upper gray region), then merged cohorts into 13 country-level strata. To ensure that these strata were uniform we then performed stringent quality control on each stratum (lower gray region) to produce our final dataset. (B) the exome chip captures a large fraction of ExAC (release version 1) low-frequency miss-sense variants. The exome chip captures the majority of variants present in ExAC (Lek et al., 2016) down to a minor allele frequency ∼0.0005, below which a large number of variants is observed (left). Thus, the overall coverage at very rare alleles (5 × 10⁻⁴ > MAF > 1.5 × 10⁻⁵, corresponding to a single allele seen in 33,370 non-Finnish European individuals in ExAC) is low (right).

Figure 1

Rare-Coding Variants Are Associated to Multiple Sclerosis Risk in a Multi-cohort Study (A–C) We analyzed 120,991 low-frequency non-synonymous coding variants across all autosomal exons in 32,367 MS cases and 36,012 controls drawn across the International Multiple Sclerosis Genetics Consortium centers. We find evidence for association with both common variants with combined MAF > 5% (A) and with rare variants across the autosomes (B). We sourced samples from Australia, 10 European countries, and the United States (C). See also Figures S2 and S3.

Figure S2

Low-Frequency Variant Association Statistic Characteristics, Related to Figure 1 (A) effect sizes increase at low minor allele frequency. We conducted a meta-analysis of 120,991 low-frequency coding variants across all autosomal exons, concentrating on non-synonymous variants which are more likely to have a phenotypic effect. We analyzed a total of 32,367 MS cases and 36,012 controls in thirteen strata. Here, we show that estimates of effect size (β or log odds ratio, y axis) increase at low allele frequency (number of minor alleles present in control samples, x axis). Because many low-frequency variants are not present in all cohorts, we stratify these data by number of cohorts in which a variant is polymorphic (subplots). Rarer variants have larger estimated effect sizes and are present in fewer cohorts. (B) forest plots for genome-wide significant low-frequency variants. Seven variants in six genes are significant in our analysis (p < 3.5 × 10⁻⁷, Bonferroni correction for the total number of variants genotyped). Two of these (TYK2 p.Pro1104Ala and GALC p.Asp84Asp), are in linkage disequilibrium with known GWAS hits. Studies are ordered by increasing sample size.

Figure S3

Patterns of Association for Common and Rare Variants in Seven Genome-wide Significant Loci, Related to Figure 1 Plots are centered on the seven variants reported in Figure 1 and Table 1. Each show LD and association signal of low frequency variants (circles, this study), and common variants from our most recent GWAS (squares, 14,802 MS cases and 26,703 controls; International Multiple Sclerosis Genetics Consortium et al., 2017) and the ImmunoChip meta-analysis (diamonds; Beecham et al., 2013). For GALC and TYK2, our most associated variants, rs11552556 and rs34536443 respectively, capture the common variant signals we have previously reported (panels A and G). For the remaining loci, our most associated variants show no LD to other variants, with no evidence of association in our common variant studies (panels B-F).

Figure 2

Rare Variants Explain a Substantial Portion of Multiple Sclerosis Heritability We estimated the MS risk heritability explained by common variants (MAF > 5%) and low-frequency non-synonymous coding variation (MAF < 5%) in each of 13 cohorts genotyped on the exome chip using genome-wide complex trait analysis (GCTA; top). By meta-analyzing these estimates across cohorts, we found that low-frequency variants explain 11.34% of heritability on the observed scale, which corresponds to 4.1% on the liability scale (right top). After dividing the low-frequency variants into intermediate (5% > MAF > 1%) and rare (MAF < 1%; bottom), we found that the latter alone explains 9.0% heritability on the observed scale (3.2% on the liability scale; bottom right). Meta-analysis confidence intervals are small and visually occluded by the mean estimate plot characters. Cohorts (abbreviations as in Table S1) are ordered by sample size, with the percentage of the overall sample size shown in each subplot title. We could not obtain estimates for either model for our Finnish cohort (see STAR Methods; not shown), or for the three-component model for our Belgian cohort (bottom, top row, fourth from left). Both cohorts are small, which may explain the failure to converge.