Negligible impact of rare autoimmune-locus coding-region variants on missing heritability

Genome wide association studies (GWAS) have identified common variants of modest effect size at hundreds of loci for common autoimmune diseases - however a substantial fraction of heritability remains unexplained, and to which rare variants may contribute^1,2. To discover rare variants and test them for association with a phenotype, the majority of studies re-sequence a small initial sample size and then genotype the discovered variants in a larger sample set^3-5. This approach will fail to analyse a large fraction of the rare variants present in the entire sample set. Here we perform simultaneous amplicon sequencing-based variant discovery and genotyping for coding exons of 25 GWAS risk genes in 41,911 white-European origin UK subjects comprising 24,892 subjects with six autoimmune disease phenotypes and 17,019 controls, and show that rare coding-region variants at known loci play a negligible role in common autoimmune disease susceptibility. These results do not support the rare variant - synthetic genome-wide association hypothesis⁶. Many known autoimmune disease risk loci contain multiple independent common and low frequency variant association signals, and so genes in these loci are a priori stronger candidates to harbor rare coding-region variants than other genes. Our data suggest that the missing heritability for common autoimmune diseases may not lie in the rare coding-region variant portion of the allelic spectrum, but perhaps as others have proposed in very many common variant signals of weak effect^7-10.

Recent large scale human sequencing studies have revealed an abundance of rare variants (which we define as minor allele frequency (MAF)<0.5%), shown that these are geographically localised, and are more likely to have deleterious functional consequences^11,12. Nelson et al, in the largest sample size studied to date¹², resequenced 202 genes in 14,002 people and found ~95% of exonic variants identified to be rare, with 74% only observed in one or two subjects. More broadly, across ~15,000 genes, similar findings were observed in recent exome sequencing studies of 2,440 and 6,515 subjects^13,14. Importantly, these studies demonstrate that even if we had reference variation databases from a million subjects, most of the rare variant allelic spectrum of any given sample set (e.g. a case-control cohort) will be unique and only identifiable by direct resequencing of the entire sample set.

There are only a handful of published examples of rare coding-region variants associated with common autoimmune diseases (although many examples in familial/Mendelian immune-mediated diseases). Coding-region variants in IFIH1 associated with type 1 diabetes (MAF in controls 0.67 - 2.2%)³, TYK2 with multiple autoimmune diseases¹⁵, and IL23R and inflammatory bowel disease⁵, for example, are low frequency (which we define as MAF 0.5 - 5%) rather than particularly rare. In other examples, the existing evidence for association, and/or the effect sizes, are relatively weak (e.g. CARD14 & psoriasis¹⁶, IL2RA/IL2RB and rheumatoid arthritis¹⁷). The association of rare coding-region variants in CARD15/NOD2 with Crohn’s disease probably provides the best example, albeit three low frequency variants comprise over 80% of all the disease causing mutations¹⁸. The majority of studies also lose power (especially for pooled across gene variant tests) by initially sequencing only a small sample subset rather than testing the entire rare variant content of a large case-control sample set. We sought to improve on these methods by performing highly multiplexed sequencing of sufficiently high quality to enable direct genotyping in the entirety of a large autoimmune disease case-control collection.

We selected subjects from a single population, individuals of white Northern-European ethnicity living in the UK (Supplementary Methods), to minimise any effects of population stratification. We selected to re-sequence all RefSeq exons for 25 genes from 20 GWAS-identified risk loci showing overlap between six common autoimmune disease phenotypes (autoimmune thyroid disease, coeliac disease, Crohn’s disease, psoriasis, multiple sclerosis and type 1 diabetes). All genes studied were from risk loci for at least two phenotypes; all genes had known immune system function; 18 of 20 loci had either a single candidate immune gene or all immune genes at a locus were selected (the remaining 2 loci had partial transcripts of another immune gene within the 0.1cM linkage disequilibrium block); and all genes & loci were densely genotyped on the Illumina Immunochip (Supplementary Table 1)¹⁹. We attempted high throughput sequencing of 52,224 samples (including positive and negative controls, and repeats). We performed extensive quality control on both samples and variant calls (Supplementary Methods). The final dataset comprised 41,911 phenotyped individuals (immune disease cases and controls), with Immunochip array genotypes available for 32,806 of these individuals (Supplementary Table 2). We discovered 4,377 variant sites across all amplicons, the genotype call rate was 99.9989% (reference homozygote as well as non-reference genotypes) across 41,911 individuals. Of these, 2990 variants were in protein coding regions (including exon splice sites) of the 25 genes (Table 1, Supplementary Table 3); 97.1% of which are rare (MAF in 17,019 controls <0.5%); 73.6% being novel when compared with current published datasets (dbSNP137, 1000 Genomes Project, NHLBI) containing >6000 individuals, and 67.3% novel compared to an unpublished dataset of 25,994 exome-sequenced individuals (MacArthur DG, personal communication); and 68.9% were only seen in one or two individuals. These proportions of novel, and rare, variants are similar to recent data from other large re-sequencing studies¹².

Table 1. Variant types in protein coding regions of 25 genes in 41,911 phenotyped individuals.

Numbers shown are after quality control steps. Annotation performed with GENCODE V14 gene definitions. Triallelic (n=124) and quadrallelic (n=3) sites (combined single nucleotide variants and indels) are shown as multiple separate variants with the appropriate annotation for each non-reference allele.

Variant type	All variants	Rare (MAF<0.5%)a	Novelb
Nonsynonymous SNV	1,792	1,758	1,379
Splicing SNV	86	85	65
Stopgain SNV	47	47	42
Synonymous SNV	1,024	972	674
Frameshift indels	31	31	31
Nonframeshift indels	10	10	10
Total variants	2,990	2,903	2,201
Singleton	1,602	1,598	1,411
Doubleton	470	468	378

^aMAF in 17,019 sequenced controls.

^bNot seen in dbSNP137; nor 1000 Genomes Project (April 2012 release); nor NHLBI (data release ESP6500SI, with 6,503 individuals)

Our very high coverage data (99.8% of 183.4m site x sample genotype calls had read depth≥40, 96.6% read depth>100, Supplementary Figure 1) enabled stringent data filtering on call rate per sample, per variant site, and other criteria (Supplementary Methods). To confirm data quality, we performed further experiments and analyses: (a) we genotyped one control sample 296 times (on different 48-sample microfluidic chips), the genotype call error rate was 2 non-consensus genotype calls of 1,295,581 called genotypes (0.00015%); (b) 32,806 of 41,911 subjects also had dense Immunochip genotyping data at the 25 genes, genotype concordance at 91 variant sites genotyped on both platforms was 99.994%; (c) transition/transversion (Ti/Tv) rates, a quality control measure based on expected human mutation types, were 2.434 at coding-region variants (2.427 at singletons), 2.44 at rare (MAF<0.5%) variants (2.437 at singletons), and 2.275 at novel variants (2.273 at singletons)(definitions as Table 1); (d) we selected all (35) nonsense single nucleotide variants (SNV) and all (39) frameshift indels in the Immunochip genotyped samples for Sanger sequencing, 2 variants failed assay/PCR design, and there was 1 false positive SNV and 1 false positive indel (overall false positive rate 2.8%). All 70 validated SNVs and indels had the same alleles in high-throughput and Sanger sequencing assays; (e) proportions of rare, and of known, variants were similar to those found by other large sequencing studies, and we identified no common or low frequency novel variant sites.

We first attempted to identify any low frequency or rare variants of larger effect. We performed for each coding-region variant and each of 7 phenotypes (including all autoimmune disease cases combined) a single-variant association analysis. Only previously reported loci were observed with common variants (MAF>5%), as expected. We identified three low frequency (MAF 0.5 - 5%) and rare exonic variants (MAF in 17,019 controls <0.5%) with single SNP association P<10⁻⁴ (chosen as a partial Bonferroni multiple testing correction for 25 genes and 7 phenotypes, but not correcting for all variants per gene)(Supplementary Table 4, Supplementary Data). We next analysed low frequency and rare exonic variants, conditioning on common variant non-coding signals at each locus, and observed no additional association signals (Supplementary Data). An association between type 1 diabetes and the low frequency UBASH3A SNP rs17114930 was observed, but conditional regression analysis showed this signal to be secondary to a stronger common frequency variant/haplotype previously identified by GWAS²⁰. We identified novel low frequency (nearly “common” as MAF in 17,019 controls = 4.97%) NCF2 coding-region variant associations with coeliac disease at two SNPs (rs17849502, non-synonymous and rs17849501, synonymous; in almost complete linkage disequilibrium r²=0.992). Both variants were present on the Illumina Immunochip, but just failed quality control criteria in our previous coeliac disease study due to missing data¹⁹. We replicated the UK findings in 4,313 coeliac cases and 3,954 controls (European samples, Supplementary Methods, rs17849502 P_CMH=4.46×10⁻⁵, OR 1.35 (95% CI 1.17-1.55)). Logistic regression analysis conditioning on rs17849502 in the UK resequencing dataset revealed no further single-variant coeliac disease association signals below P<10⁻⁴. NCF2 is a component of the neutrophil NADPH oxidase respiratory burst complex, different disease-causing mutations cause the recessive Mendelian phenotype chronic granulomatous disease. The rs17849502 variant is associated with the autoimmune disease systemic lupus erythematosus, and rs17849502/H389Q reduces the binding efficiency of NCF2 with the guanine nucleotide exchange factor Vav1²¹. These data now implicate a disease mechanism of impaired neutrophil function in coeliac disease, a condition previously thought to be of predominantly B and T cell mediated immuno-pathogenesis, and where neutrophils may play a role in regulating adaptive immunity²².

We noted that even with ~7,000 cases and ~17,000 controls the power to detect association signals using single-variant tests for variants (MAF< 0.5%) of modest effect (e.g. odds ratio <3) is limited (Supplementary Figure 2) and therefore performed gene-based pooled variant association tests to better detect the combined effect of multiple variants. We defined coding-region variants as functional candidates if the variants were rare (MAF in 17,019 controls <0.5%) and predicted to be of potential functional impact (non-synonymous, premature-stop, splice site altering, see Supplementary Methods). We pooled variants (by gene) in analyses to detect different scenarios (Figure 1, Supplementary Data), including: C-alpha test which can detect a combination of risk and protective variants; burden tests to detect either an excess of risk variants in cases or protective variants in controls; a modified version of the burden test using conditional regression and common variant non-coding signals at a locus as covariates; a test to detect an excess of rare variants seen uniquely in cases (the case or control unique tests being particularly suitable for the study of the large numbers of singletons and doubleton variants we observe); and a test to detect an excess of rare variants seen uniquely in controls. The distribution of association statistics for all five pooled gene tests across each of the 6 or 7 phenotypes tested was consistent with the global null of no association.

Figure 1. Association analyses of discovered rare functional variants in autoimmune diseases.

We define rare functional variants as MAF<0.5% in 17,019 controls and predicted non-synonymous, premature stop or splice site annotation. Quantile-quantile plots compare observed versus expected test-statistic distributions, with shading indicating 99% confidence intervals. Full results are available in Supplementary Data. Each of six individual diseases, and all autoimmune diseases combined, were tested as phenotypes.

a. Gene based C-alpha test (25 genes by 7 phenotypes, n=41,911 subjects) allowing for both risk and protective effects for rare functional variants (n = 41,911 subjects). Singleton variants pooled into a single binomial count per phenotype.

b. Gene based burden tests (25 genes by 7 phenotypes, n=41,911 subjects) comparing summed allele counts for rare functional variants in cases versus controls with Fisher’s exact test.

c. Conditional gene based burden test (25 genes by 6 phenotypes, n=32,806 subjects): rare functional variant allele counts are summed for each individual per gene and introduced in a logistic regression, including Immunochip covariates for multiple independent top (common) variant signals selected on the basis of a stepwise regression (down to P>10⁻⁴). The psoriasis phenotype was not tested as most samples do not have Immunochip data.

d. UNIQ-cases tests (25 genes by 7 phenotypes, n=41,911 subjects) that compares the number of rare functional variants only observed in cases with the distribution of this value upon random permutation (10,000 times) of the phenotypes.

e. UNIQ-controls, same as e but for rare functional variants uniquely observed in controls.

On the basis of these results, in (to the best of our knowledge) the largest human disease sample sequencing study to date, we find little support for a significant impact of rare coding-region variants in known risk genes for the autoimmune disease phenotypes tested. Our data provide little stimulus in support of large scale whole exome sequencing projects in common autoimmune diseases. Using average genetic effect estimates from our data (Supplementary Information), over all loci and phenotypes we have tested, we estimate that rare variants contribute to less than 3% of the heritability explained by common variants at these known risk loci²³.

METHODS

Gene selection

All genes studied (listed in Supplementary Table 3) were risk loci for at least two phenotypes, had a known immune system function, were from loci with only a single strong candidate immune gene (or all immune genes were selected at four loci - IL18R1/IL18RAP; CTLA4/CD28/ICOS; IL2/IL21; PTPRK/THEMIS), and all genes & loci were densely genotyped with all 1000 Genomes pilot project variants on the Illumina Immunochip (for design of this chip, see ref ¹⁹). Additional criteria favouring locus selection were known multiple independent association signals; risk (not necessarily same variants/haplotype or signal direction) for many autoimmune diseases; fine-mapping or other data strongly suggesting a single candidate gene; and smaller cDNA size,

Samples

UK samples for the six component immune disease phenotypes have been described in previous publications (which also contain full details of Ethics Committee approvals)^19,20,24-27, as have the 3 control populations^19,28. Informed consent was obtained from all subjects. Individuals with self-reported autoimmune disease were excluded from the UK Blood Services - Common Controls and NIHR Cambridge Biomedical Research Centre Cambridge BioResource controls. Samples with self-stated non-white European ethnicity were excluded (later further confirmed by Immunochip-based principal component ethnicity analysis for 32,806 samples). Samples with gross discordance with Immunochip genotypes and/or with known gender or genotype mismatch issues from previous GWAS were excluded. Samples with known duplicates or relatedness (as distant as first cousins) were excluded, relatedness was later confirmed by Immunochip genome wide identity-by-state analysis and by analysis of multiple rare variant sharing in Fluidigm sequencing data. Additional independent European samples genotyped for rs17849502 (4,313 coeliac cases and 3,954 controls) were previously described².

Wet-lab

PCR primers were designed for all RefSeq exons of 26 genes, and amplicons selected to be 150-200bp in size. There was minor primer design dropout at IL18R1, STAT4, THEMIS, ZMIZ1, although >94% of exon sequence was still covered at these genes. Variant calls at the gene YDJC later proved unreliable with highly biased allele depths at heterozygote sites, likely due to the very high exon GC content (~70%), and this gene was not further analysed nor is it discussed elsewhere in this study. The total length of (overlapping) amplicons was 95,927bp, with primers removed (still overlapping) 72,612bp, and with primers removed and unique sequence 58,550 bp. PCR amplification was performed using 50ng gDNA per sample on the 48 sample/plate Fluidigm microfluidic Access Array system. PCR primers for 511 PCR reactions were pooled up to 13-plex per well in 48 pools. Individual per sample per pool PCR reactions took place in ~35nl reaction chambers with ~300 DNA haplocopies/reaction. All pools per sample were combined. Each sample’s pool was then individually barcoded in a second PCR reaction with one of 1536 10bp Fluidigm-designed unique barcodes (Fluidigm unidirectional sequencing protocol).

Sequencing

34 libraries (each of 1536 barcoded samples) were generated. Libraries were first sequenced on an Illumina MiSeq for rapid quality control of the barcoding step, and to optimise loading concentrations/cluster density. Libraries were then sequenced one per lane using 101bp paired-end reads and an 11bp index read (the last base of each read being only used for chemistry cycle phasing purposes) on Illumina HiSeq sequencers. Lanes were repeated if target cluster density or target clusters passing filter were not achieved. Individual samples were demultiplexed by Illumina CASAVA software, allowing zero mismatches per 10bp barcode. Sanger sequencing was performed on PCR products using an ABI 3730×l DNA analyser and ABI big dye terminator 3.1 cycle chemistry. We sequenced all samples with rare variant allele genotypes, and a control sample, for the 74 sites selected.

Bioinformatics

PCR primers were trimmed from the 5′ end of individual reads using a modified version of btrim²⁹. Trimmed sequences were aligned to the GRCh37 human reference genome using gapped quality-aware alignment, and base call quality recalibration implemented in novoalign V2.07.18 with settings ‘-t 100 -H -g 65 -x 7 -o FullNW’. Data was realigned against known (1000 genomes and Mills-Devine 2-hit) indels and per-sample called indels. SNPs were called using GATK 1.6-5 and settings ‘-- min_base_quality_score 15 -stand_call_conf 30 --baq CALCULATE_AS_NECESSARY -glm SNP --baqGapOpenPenalty 65 -- downsampling_type BY_SAMPLE --downsample_to_coverage 250’ and then hard filtered using GATK settings ‘QUAL<80.0 DP<20 MQ<40.0 QD<2.0 MQRankSum<-12.5 HRun>5’ (several other recommended best practice GATK settings were not appropriate for PCR amplicon data), and around indels. Small indels (up to 15bp gaps from novoalign) were called using GATK and settings ‘--min_base_quality_score 15 -stand_call_conf 30 --baq CALCULATE_AS_NECESSARY -glm INDEL -- baqGapOpenPenalty 65 --downsampling_type BY_SAMPLE --downsample_to_coverage 250’ and then hard filtered using GATK settings ‘QUAL<80.0 DP<20 QD<2.0’ (several other recommended best practice GATK settings were not appropriate for PCR amplicon data). The most important of these settings were likely to be calling genotypes as missing with sequencing depth <20 high quality bases and the minimum Phred 15 recalibrated base call quality score to define high quality bases. Both samtools and vcftools software were also used to process data. SNP genotypes (including non-reference genotypes) were called at all 58,550 bases of amplicon sequence. Samples with <57,600 SNP genotype calls (98.4%, a threshold determined by inspection of the call rate plot) were removed and scheduled for repeat processing. Clusters of very close non-reference genotypes in an individual sample were removed. Non-reference genotype sites were then identified across all samples, and vcf-level data reduced to variants at polymorphic sites (in one or more samples). A combined vcf file of all polymorphic sites and samples was then loaded into PLINK/SEQ v0.09. Multiple step filtering based on call rate per-sample and call rate per-variant site was applied, with final requirements >99.95% call rate per-sample and per-variant site. Lower call rate samples at this stage were also scheduled for repeat processing. We removed variants if the sum of heterozygote genotype allele depths was <25% or >75%. The final filtered data was then exported to a vcf file containing all variants and samples for analysis in R. Immunochip data was loaded into Illumina GenomeStudio software from .idat files, and all samples called together in GenomeStudio using the cluster settings as previously described¹⁹. Data was merged with HapMap Phase 3 genotypes, principal component analysis performed, and the first two principal components used to validate ethnicity (Supplementary Figure 3).

Barcode and sequencing amplicon performance

Barcode evenness was excellent, with typically 99.0% of the 1536 barcodes producing pass-filter read numbers which were between 0.033% and 0.13% of the total pass filter reads per lane (0.065% expected), with most of the failing barcodes tagging known water negative control samples or (based on repeat amplification with a different barcode) due to poor DNA quality. Amplicon evenness was good, and for many genotype calls we required to downsample data to 250 bases per site per sample (Supplementary Figure 1). However 10 of 511 amplicons effectively failed PCR. In a typical analysis of 100 high quality samples, 2% of the 58,550 unique amplicon bases had minimum mean read-depth <20, nearly all accounted for by the 10 failing amplicons.

Variant annotation

Annotation of all variants was first performed using annovar (Feb 2013) and the GENCODE V14 dataset. Coding variants were identified. Rare functional variants were identified based on stop, frameshift indel, non-synonymous (SNP or 3n indel) or splice predictions. We performed an additional layer of annotation for high confidence loss of function mutations, using the methods described by MacArthur et al³⁰. The Variant Effect Predictor (VEP v2.5) tool from Ensembl was modified to produce custom annotation tags and additional loss of function (LoF) annotations. The additional LoF annotation was applied to variants which were annotated as STOP_GAINED, SPLICE_DONOR_VARIANT, SPLICE_ACCEPTOR_VARIANT, and FRAME_SHIFT and flagged if any filters failed. Filters included LoF is the ancestral allele; Exon is surrounded by non-canonical splice site (i.e. not AG/GT); LoF removes less than 5% of remaining protein; LoF is rescued by nearby Start Codon which results in less than 5% of protein truncated; Transcript only has one coding exon; splice site mutation within intron smaller than 15 bp; Splice site is non-canonical OR other splice site within same intron is non-canonical; Unable to determine exon/intron boundaries surrounding variant. A LoF variant is predicted as high confidence (HC) if there is one transcript that passes all filters, otherwise it is predicted as low confidence (LC). We noted that LoF mutations were seen in 21 of 25 genes, all were heterozygous genotypes, and mainly (87 of 97) as singletons or doubletons in the 41,911 samples (Supplementary Table 3).

Statistical analysis

Most analysis was performed in R using custom code (available on request). For tests using permutations (C-alpha, UNIQ-cases and UNIQ-controls in Figure 1), we randomly permuted in R the case control status 10,000 times. The unconditional burden test (Figure 1B) used a Fisher exact text. Conditional burden tests used the glm function in R, including selected Immunochip common variants as covariates (selection based on a stepwise regression analysis up to 10⁻⁴). For the C-alpha statistic computation (Figure 1A), the expected proportion of rare alleles in the case/control cohorts was set to the proportion of cases and controls. Figure 1 was generated using the fact that under the null of no association −2log(p) is distributed as chi-squared with 2 degrees of freedom. PLINK/SEQ v0.09 (http://atgu.mgh.harvard.edu/plinkseq/index.shtml) was used for Ti/Tv statistics, and to confirm findings of R analyses (not shown). We used PLINK/SEQ for the genotype concordance analysis between Immunochip and Fluidigm-sequencing data. Discordant calls were observed at 169 of 2985255 (0.0056%) genotypes, occurring at 36 of 91 polymorphic variant sites present in both datasets. We inspected Illumina Immunochip R theta intensity plots for the discordant genotypes, and observed 8 discordant genotypes to be likely due to Immunochip data mis-clustering, and 11 discordant genotypes to be due to a 3^rd or 4^th observed allele in the high-throughput sequencing data. At the sites with 3^rd and 4^th alleles, we note the Immunochip array assays can only call two alleles, therefore is not possible to determine whether these sequence genotype calls are real or errors. R code used for analysis is available from V.P.

Estimation of average genetic effect contributed by rare variants

For each combination of locus by disease, we combined all rare functional variants (frequency < 0.5% in 1,000 Genomes/NHLBI datasets and non-synonymous, LOF or splicing) in a burden statistic X and computed the combined frequency of X in the sample. Using a logistic regression model with the disease phenotype as outcome, we estimated the odds ratio associated with the burden variable X. This knowledge of frequency and odds ratio for the burden variable X enables the estimation of the average genetic effect (AGE, as defined by Liu and Leal²³) version of the variance explained. We then compared this variance at each combination of locus/gene with the variance explained by what we consider to be a typical common variant association (odds ratio 1.2, MAF 20%, assuming a single common variant per locus). To deal with the uncertainty in estimated odds ratio and obtain a confidence interval for this value, we randomly sampled the odds ratio from their estimated distribution for each pair of disease/locus. Averaging over the 150 combinations of 6 diseases by 25 loci, we estimate the ratio of heritability explained for all rare variants by all common variants to have a mean value of 1.6%, with a confidence interval of [1.2%-2.3%]. Liu and Leal point out that the AGE estimate can underestimate the true explained variance by rare variants. Nevertheless, assuming that rare variants are generally all risk or all protective at a given gene, their simulations also show that the under-estimation is limited, in the range of a 25% decrease. Taking this conservative estimate of the under-estimation level, we find the upper bound of the 95% of the confidence interval to be 3.05%. Hence, our data indicate that the aggregate contribution of rare variants to the heritability (<0.5% MAF, and averaged over these loci/diseases) is unlikely to exceed approximately 3% of the heritability assigned to common variants. We acknowledge that a much larger under-estimation (and therefore a much larger heritability explained for rare variants) is possible in the presence of a combination of high risk and highly protective rare variants at the same locus. While we cannot exclude such scenario, it is unlikely to be widespread. We also assumed in our estimates that rare variants act additively at the log scale. While this assumption is standard, we cannot exclude that a combination of rare variants results in a much stronger predictive outcome than rare variants individually, hence under-estimating the heritability associated with rare variants.