Analysis of immune-related loci identifies 48 new susceptibility variants for multiple sclerosis

Abstract

Using the ImmunoChip custom genotyping array, we analysed 14,498 multiple sclerosis subjects and 24,091 healthy controls for 161,311 autosomal variants and identified 135 potentially associated regions (p-value < 1.0 × 10^-4). In a replication phase, we combined these data with previous genome-wide association study (GWAS) data from an independent 14,802 multiple sclerosis subjects and 26,703 healthy controls. In these 80,094 individuals of European ancestry we identified 48 new susceptibility variants (p-value < 5.0 × 10^-8); three found after conditioning on previously identified variants. Thus, there are now 110 established multiple sclerosis risk variants in 103 discrete loci outside of the Major Histocompatibility Complex. With high resolution Bayesian fine-mapping, we identified five regions where one variant accounted for more than 50% of the posterior probability of association. This study enhances the catalogue of multiple sclerosis risk variants and illustrates the value of fine-mapping in the resolution of GWAS signals.

Multiple sclerosis (OMIM 126200) is an inflammatory demyelinating disorder of the central nervous system that is a common cause of chronic neurological disability.^1,2 It has its greatest prevalence amongst individuals of Northern European ancestry³ and is moderately heritable,⁴ with a sibling relative recurrence risk (λ_s) of ~ 6.3.⁵ Aside from the early success in demonstrating the important effects exerted by variants in the Human Leukocyte Antigen (HLA) genes from the Major Histocompatibility Complex (MHC),⁶ there was little progress in unravelling the genetic architecture underlying multiple sclerosis susceptibility prior to the advent of genome-wide association studies (GWAS). Over the last decade, our Consortium has performed several GWAS and meta-analyses in large cohorts, ^7-10 cumulatively identifying more than 50 non-MHC susceptibility alleles. As in other complex diseases, available data suggest that many additional susceptibility alleles remain to be identified.¹¹

The striking overlap in the genetic architecture underlying susceptibility to autoimmune diseases^9,10,12,13 prompted the collaborative construction of the “ImmunoChip” (see Supplementary Note and Supplementary Figs. 1 and 2 for details of IMSGC nominated content), an efficient genotyping platform designed to deeply interrogate 184 non-MHC loci with genome-wide significant associations to at least one autoimmune disease and provide lighter coverage of other genomic regions with suggestive evidence of association.¹⁴ Here, we report a large-scale effort that leverages the ImmunoChip to detect association with multiple sclerosis susceptibility and refine these associations via Bayesian fine-mapping.

After stringent quality control (QC), we report genotypes on 28,487 individuals of European ancestry (14,498 multiple sclerosis subjects, 13,989 healthy controls) that are independent of previous GWAS efforts. We supplemented these data with 10,102 independent control subjects provided by the International Inflammatory Bowel Disease Genetics Consortium (IIBDGC)¹⁵ bringing the total to 38,589 individuals (14,498 multiple sclerosis subjects and 24,091 healthy controls). We performed variant level QC, population outlier identification, and subsequent case-control analysis in 11 country-organized strata. To account for within-stratum population stratification we used the first five principal components as covariates in the association analysis. Per stratum odds ratios (OR) and respective standard errors (SE) were then combined with an inverse variance meta-analysis under a fixed effects model. In total we tested 161,311 autosomal variants that passed QC in at least two of the 11 strata (Online Methods). A circos plot¹⁶ summarising the results from this discovery phase analysis is shown in Figure 1.

Figure 1. Discovery phase results.

Primary association analysis of 161,311 autosomal variants in the discovery phase (based on 14,498 cases and 24,091 healthy controls). The outer most track shows the numbered autosomal chromosomes. The second track indicates the gene closest to the most associated SNP meeting all replication criteria. Previously identified associations are indicated in grey. The third track indicates the physical position of the 184 fine-mapping intervals (green). The inner most track indicates −log(p) (two-sided) for each SNP (scaled from 0-12 which truncates the signal in several regions, see Supplementary Table 1). Additionally, contour lines are given at the a priori discovery(−log(p) = 4) and genome-wide significance (-log(p) = 7.3) thresholds. Orange indicates -log(p) ≥ 4 and < 7.3, while red indicates −log(p) ≥ 7.3. Details of the full discovery phase results can be found in ImmunoBase.

We defined an a priori discovery threshold of p-value <1 × 10^-4 and identified 135 primary statistically independent association signals; 67 in the designated fine-mapping regions and 68 in less densely covered regions selected for deep replication of earlier GWAS. Another 13 secondary and 2 tertiary statistically independent signals were identified by forward stepwise logistic regression. A total of 48 of the 150 statistically independent association signals (Supplementary Table 1) reached a genome-wide significance p-value <5 × 10^-8 at the discovery phase alone. Next, we replicated our findings in 14,802 multiple sclerosis subjects and 26,703 healthy controls with available GWAS data imputed to the 1000 Genomes European phase I (a) panel (Online Methods). Finally, we performed a joint analysis of the discovery and replication phases.

We identified 97 statistically independent SNPs meeting replication criteria (p_replication < 0.05, p_joint < 5 × 10^-8, and p_joint < p_discovery); 93 primary signals (Supplementary Figs. 3-95) and four secondary signals. Of these, 48 are novel to multiple sclerosis (Table 1) and 49 correspond to previously identified multiple sclerosis effects (Table 2). An additional 11 independent SNPs showed suggestive evidence of association (p_joint < 1 × 10^-6) (Supplementary Table 2).

Table 1.

48 Novel non-MHC susceptibility loci associated with multiple sclerosis at a genome-wide significance level

				Discovery			Replication			Joint
SNP	Chr	Positiona	RA	RAF	P-value	OR	RAF	P-value	OR	P-value	OR	Geneb	Function
rs3007421	1	6530189	A	0.12	9.6 × 10-7	1.12	0.13	8.8 × 10-5	1.10	4.7 × 10-10	1.11	PLEKHG5	intronic
rs12087340	1	85746993	A	0.09	5.1 × 10-12	1.22	0.09	2.9 × 10-10	1.20	1.1 × 10-20	1.21	BCL10	intergenic
rs11587876	1	85915183	A	0.79	8.4 × 10-8	1.12	0.81	2.9 × 10-3	1.06	4.4 × 10-9	1.09	DDAH1	intronic
rs666930	1	120258970	G	0.53	7.5 × 10-8	1.09	0.53	1.3 × 10-5	1.07	6.0 × 10-12	1.08	PHGDH	intronic
rs2050568	1	157770241	G	0.53	1.3 × 10-6	1.08	0.54	2.3 × 10-5	1.07	1.5 × 10-10	1.08	FCRL1	intronic
rs35967351	1	160711804	A	0.67	1.7 × 10-6	1.09	0.68	5.9 × 10-6	1.09	4.4 × 10-11	1.09	SLAMF7	intronic
rs4665719	2	25017860	G	0.25	6.8 × 10-6	1.09	0.25	1.1 × 10-4	1.08	3.1 × 10-9	1.08	CENPO	intronic
rs842639	2	61095245	A	0.65	1.7 × 10-9	1.11	0.67	1.4 × 10-6	1.09	2.0 × 10-14	1.10	FLJ16341	ncRNA
rs9967792	2	191974435	G	0.62	1.8 × 10-9	1.11	0.64	1.2 × 10-4	1.07	3.5 × 10-12	1.09	STAT4	intronic
rs11719975	3	18785585	C	0.27	5.4 × 10-6	1.09	0.28	4.1 × 10-4	1.07	1.1 × 10-8	1.08		intergenic
rs4679081	3	33013483	G	0.52	1.2 × 10-5	1.08	0.55	3.7 × 10-4	1.07	2.2 × 10-9	1.07	CCR4	intergenic
rs9828629	3	71530346	G	0.62	5.5 × 10-6	1.08	0.64	8.5 × 10-6	1.08	1.9 × 10-10	1.08	FOXP1	intronic
rs2726518	4	106173199	C	0.55	1.2 × 10-5	1.09	0.58	4.7 × 10-4	1.06	3.9 × 10-8	1.07	TET2	intronic
rs756699	5	133446575	A	0.87	3.0 × 10-6	1.12	0.88	6.5 × 10-6	1.11	8.8 × 10-11	1.12	TCF7	intergenic
nonec	5	141506564	C	0.61	6.0 × 10-5	1.07	0.62	1.5 × 10-5	1.08	3.6 × 10-9	1.07	NDFIP1	intronic
rs4976646	5	176788570	G	0.34	1.0 × 10-12	1.13	0.36	5.0 × 10-7	1.10	4.4 × 10-18	1.12	RGS14	intronic
rs17119	6	14719496	A	0.81	1.9 × 10-6	1.11	0.80	1.2 × 10-5	1.10	1.0 × 10-10	1.10		intergenic
rs941816	6	36375304	G	0.18	4.5 × 10-9	1.13	0.20	8.3 × 10-5	1.08	3.9 × 10-12	1.11	PXT1	intronic
rs1843938	7	3113034	A	0.44	2.2 × 10-6	1.08	0.44	1.1 × 10-5	1.08	1.2 × 10-10	1.08	CARD11	intergenic
rs706015	7	27014988	C	0.18	1.3 × 10-9	1.14	0.18	9.9 × 10-3	1.06	1.1 × 10-9	1.10		intergenic
rs917116	7	28172739	C	0.20	2.1 × 10-8	1.12	0.21	5.8 × 10-3	1.06	3.3 × 10-9	1.09	JAZF1	intronic
rs60600003	7	37382465	C	0.10	2.5 × 10-8	1.16	0.10	4.2 × 10-7	1.14	6.0 × 10-14	1.15	ELMO1	intronic
rs201847125d	7	50325567	G	0.70	2.9 × 10-8	1.11	0.70	6.7 × 10-5	1.09	1.2 × 10-11	1.10	IKZF1	intergenic
rs2456449	8	128192981	G	0.36	2.2 × 10-8	1.10	0.37	3.8 × 10-3	1.05	1.8 × 10-9	1.08		intergenic
rs793108	10	31415106	A	0.50	5.6 × 10-8	1.09	0.51	1.8 × 10-5	1.07	6.1 × 10-12	1.08		intergenic
rs2688608	10	75658349	A	0.55	6.4 × 10-5	1.07	0.56	2.0 × 10-4	1.06	4.6 × 10-8	1.07	C10orf55	intergenic
rs7120737	11	47702395	G	0.15	7.6 × 10-8	1.13	0.15	1.0 × 10-3	1.08	1.0 × 10-9	1.10	AGBL2	intronic
rs694739	11	64097233	A	0.62	1.3 × 10-5	1.08	0.62	3.8 × 10-5	1.07	2.0 × 10-9	1.07	PRDX5	intergenic
rs9736016	11	118724894	T	0.63	2.2 × 10-8	1.10	0.63	2.6 × 10-8	1.10	3.0 × 10-15	1.10	CXCR5	intergenic
rs12296430	12	6503500	C	0.19	3.6 × 10-10	1.14	0.21	1.7 × 10-5	1.09	7.2 × 10-14	1.12	LTBR	intergenic
rs4772201	13	100086259	A	0.82	1.7 × 10-7	1.12	0.83	1.1 × 10-4	1.09	1.3 × 10-10	1.10	MIR548AN	intergenic
rs12148050	14	103263788	A	0.35	1.5 × 10-5	1.08	0.36	4.3 × 10-9	1.10	5.1 × 10-13	1.09	TRAF3	intronic
rs59772922	15	79207466	A	0.83	4.0 × 10-6	1.11	0.83	5.4 × 10-4	1.08	1.2 × 10-8	1.09	CTSH	intergenic
rs8042861	15	90977333	A	0.44	9.8 × 10-7	1.08	0.45	3.4 × 10-4	1.06	2.2 × 10-9	1.07	IQGAP1	intronic
rs6498184	16	11435990	G	0.81	2.1 × 10-10	1.15	0.82	6.5 × 10-9	1.14	7.4 × 10-18	1.15	RMI2	intergenic
rs7204270*	16	30156963	G	0.50	9.3 × 10-8	1.09	0.49	3.7 × 10-5	1.08	1.6 × 10-11	1.09	MAPK3	intergenic
rs1886700	16	68685905	A	0.14	8.8 × 10-6	1.11	0.14	3.2 × 10-4	1.08	1.3 × 10-8	1.10	CDH3	intronic
rs12149527	16	79110596	A	0.47	1.7 × 10-6	1.08	0.47	4.3 × 10-6	1.08	3.3 × 10-11	1.08	WWOX	intronic
rs7196953	16	79649394	A	0.29	2.6 × 10-5	1.08	0.30	7.9 × 10-7	1.09	1.0 × 10-10	1.09	MAF	intergenic
rs12946510	17	37912377	A	0.47	8.5 × 10-6	1.08	0.48	8.0 × 10-5	1.07	2.9 × 10-9	1.07	IKZF3	intergenic
rs4794058	17	45597098	A	0.50	1.6 × 10-5	1.07	0.52	3.5 × 10-10	1.11	1.0 × 10-13	1.09	NPEPPS	intergenic
rs2288904	19	10742170	G	0.77	9.6 × 10-10	1.14	0.78	5.4 × 10-4	1.07	1.6 × 10-11	1.10	SLC44A2	exonic
rs1870071	19	16505106	G	0.29	5.7 × 10-10	1.12	0.30	4.6 × 10-7	1.09	2.0 × 10-15	1.10	EPS15L1	intronic
rs17785991	20	48438761	A	0.35	6.4 × 10-7	1.09	0.34	5.9 × 10-3	1.05	4.2 × 10-8	1.07	SLC9A8	intronic
rs2256814	20	62373983	A	0.19	8.3 × 10-7	1.11	0.21	6.4 × 10-4	1.08	3.5 × 10-9	1.09	SLC2A4RG	intronic
Secondary
rs7769192e	6	137962655	G	0.55	1.3 × 10-5	1.08	0.54	5.1 × 10-5	1.07	3.3 × 10-9	1.08		intergenic
rs533646f	11	118566746	G	0.68	3.6 × 10-7	1.10	0.68	3.9 × 10-5	1.08	7.6 × 10-11	1.09	TREH	intergenic
rs4780346g	16	11288806	A	0.23	6.8 × 10-6	1.09	0.25	1.5 × 10-5	1.09	4.4 × 10-10	1.09	CLEC16A	intergenic

All listed signals had a discovery P-value ≤ 1.0 × 10^-4, a replication P-value ≤ 5.0 × 10^-2, and a joint P-value ≤ 5.0 × 10^-8

All P-values are two-sided

RA= Risk Allele, RAF = Risk Allele Frequency

^aPosition is based on human genome 19 and dbSNP 137.

^bNearest gene listed if within 50Kb. Bold indicates Gene Ontology Immune System Process.

^cA proxy SNP (rs1036207, r² = 0.99) and

^d(rs716719, r²=1.00) was used in replication.

^eThe P-value and OR values provided are after conditioning on rs67297943 (Known – see Table 2),

^frs9736016, and

^grs12927355 (Known – see Table 2).

^*Note primary was rs11865086 (P-value = 1.77 × 10^-8) in Discovery but not available in Replication so the best proxy was used.

Table 2.

49 Known non-MHC susceptibility loci associated with multiple sclerosis at a genome-wide significance level

				Discovery			Replication			Joint
SNP	Chr	Positiona	RA	RAF	P-value	OR	RAF	P-value	OR	P-value	OR	Geneb	Function
rs3748817	1	2525665	A	0.64	1.3 × 10-12	1.14	0.65	1.2 × 10-15	1.15	1.3 × 10-26	1.14	MMEL1	intronic
rs41286801	1	92975464	A	0.14	7.9 × 10-16	1.20	0.16	2.1 × 10-12	1.17	1.4 × 10-26	1.19	EVI5	UTR3
rs7552544*	1	101240893	A	0.56	3.7 × 10-6	1.08	0.43	3.3 × 10-12	1.12	1.9 × 10-16	1.10	VCAM1	intergenic
rs6677309	1	117080166	A	0.88	1.5 × 10-28	1.34	0.88	4.1 × 10-16	1.24	5.4 × 10-42	1.29	CD58	intronic
rs1359062	1	192541472	C	0.82	1.8 × 10-13	1.18	0.83	2.1 × 10-8	1.13	4.8 × 10-20	1.15	RGS1	intergenic
rs55838263	1	200874728	A	0.71	1.4 × 10-9	1.12	0.71	3.9 × 10-11	1.13	4.0 × 10-19	1.13	C1orf106	intronic
rs2163226	2	43361256	A	0.71	7.0 × 10-8	1.10	0.73	3.8 × 10-10	1.14	2.1 × 10-16	1.12		intergenic
rs7595717	2	68587477	A	0.26	3.3 × 10-7	1.10	0.27	6.8 × 10-8	1.10	1.2 × 10-13	1.10	PLEK	intergenic
rs9989735	2	231115454	C	0.18	7.8 × 10-14	1.17	0.19	6.8 × 10-11	1.14	4.2 × 10-23	1.16	SP140	intronic
rs2371108	3	27757018	A	0.38	2.1 × 10-6	1.08	0.39	5.8 × 10-11	1.12	1.5 × 10-15	1.10	EOMES	downstream
rs1813375	3	28078571	A	0.47	5.7 × 10-18	1.15	0.49	4.4 × 10-16	1.15	1.9 × 10-32	1.15		intergenic
rs1131265	3	119222456	C	0.80	2.0 × 10-15	1.19	0.81	4.8 × 10-10	1.14	1.4 × 10-23	1.17	TIMMDC1	exonic
rs1920296*	3	121543577	C	0.64	6.8 × 10-15	1.14	0.64	5.5 × 10-9	1.10	6.5 × 10-22	1.12	IQCB1	intronic
rs2255214*	3	121770539	C	0.52	5.3 × 10-13	1.13	0.52	3.3 × 10-13	1.13	1.2 × 10-24	1.13	CD86	intergenic
rs1014486	3	159691112	G	0.43	1.2 × 10-9	1.11	0.44	1.4 × 10-10	1.11	1.1 × 10-18	1.11	IL12A	intergenic
rs7665090	4	103551603	G	0.52	2.4 × 10-6	1.08	0.53	5.0 × 10-4	1.13	1.0 × 10-8	1.09	MANBA	intergenic
rs6881706	5	35879156	C	0.72	4.9 × 10-9	1.12	0.73	1.7 × 10-9	1.12	4.3 × 10-17	1.12	IL7R	intergenic
rs6880778	5	40399096	G	0.60	1.7 × 10-8	1.10	0.61	3.9 × 10-13	1.13	8.1 × 10-20	1.12		intergenic
rs71624119	5	55440730	G	0.76	2.7 × 10-9	1.12	0.76	1.9 × 10-5	1.09	3.4 × 10-13	1.11	ANKRD55	intronic
rs72928038	6	90976768	A	0.17	7.6 × 10-7	1.11	0.19	9.0 × 10-11	1.17	1.5 × 10-15	1.14	BACH2	intronic
rs11154801	6	135739355	A	0.37	2.3 × 10-9	1.11	0.37	1.0 × 10-12	1.13	1.8 × 10-20	1.12	AHI1	intronic
rs17066096	6	137452908	G	0.23	5.9 × 10-12	1.14	0.25	4.1 × 10-13	1.15	1.6 × 10-23	1.14	IL22RA2	intergenic
rs67297943	6	138244816	A	0.78	4.8 × 10-8	1.12	0.80	2.5 × 10-6	1.11	5.5 × 10-13	1.11	TNFAIP3	intergenic
rs212405	6	159470559	T	0.62	1.4 × 10-15	1.15	0.64	1.8 × 10-7	1.10	8.0 × 10-21	1.12	TAGAP	intergenic
rs1021156	8	79575804	A	0.24	5.6 × 10-10	1.12	0.26	2.1 × 10-8	1.11	8.5 × 10-17	1.11	ZC2HC1A	intergenic
rs4410871	8	128815029	G	0.72	2.0 × 10-9	1.12	0.72	3.4 × 10-8	1.11	4.3 × 10-16	1.11	MIR1204	intergenic
rs759648	8	129158945	C	0.31	2.8 × 10-6	1.09	0.31	3.7 × 10-5	1.08	5.0 × 10-10	1.08	MIR1208	intergenic
rs2104286	10	6099045	A	0.72	7.6 × 10-23	1.21	0.73	3.6 × 10-26	1.23	2.3 × 10-47	1.22	IL2RA	intronic
rs1782645	10	81048611	A	0.43	4.3 × 10-7	1.09	0.41	6.2 × 10-10	1.11	2.5 × 10-15	1.10	ZMIZ1	intronic
rs7923837	10	94481917	G	0.61	4.6 × 10-9	1.11	0.62	2.0 × 10-9	1.11	4.3 × 10-17	1.11	HHEX	intergenic
rs34383631	11	60793330	A	0.40	5.7 × 10-10	1.11	0.39	4.5 × 10-15	1.15	3.7 × 10-23	1.13	CD6	intergenic
rs1800693	12	6440009	G	0.40	6.9 × 10-16	1.14	0.41	1.0 × 10-13	1.14	6.7 × 10-28	1.14	TNFRSF1A	intronic
rs11052877	12	9905690	G	0.36	5.4 × 10-9	1.10	0.38	1.2 × 10-5	1.08	5.6 × 10-13	1.09	CD69	UTR3
rs201202118c	12	58182062	A	0.67	7.4 × 10-13	1.14	0.67	1.6 × 10-10	1.12	9.0 × 10-22	1.13	TSFM	intronic
rs7132277	12	123593382	A	0.19	1.9 × 10-6	1.10	0.19	1.4 × 10-8	1.13	1.9 × 10-13	1.12	PITPNM2	intronic
rs2236262	14	69261472	A	0.50	1.2 × 10-5	1.08	0.50	3.8 × 10-8	1.09	2.5 × 10-12	1.08	ZFP36L1	intronic
rs74796499	14	88432328	C	0.95	8.5 × 10-11	1.31	0.95	4.5 × 10-11	1.33	2.4 × 10-20	1.32	GALC	intronic
rs12927355	16	11194771	G	0.68	8.2 × 10-27	1.21	0.69	4.3 × 10-21	1.18	6.4 × 10-46	1.20	CLEC16A	intronic
rs35929052	16	85994484	G	0.89	3.3 × 10-7	1.14	0.88	3.6 × 10-6	1.15	5.9 × 10-12	1.15	IRF8	intergenic
rs4796791	17	40530763	A	0.36	1.8 × 10-8	1.10	0.36	1.2 × 10-13	1.14	3.7 × 10-20	1.12	STAT3	intronic
rs8070345	17	57816757	A	0.45	5.4 × 10-16	1.14	0.46	1.9 × 10-9	1.10	2.2 × 10-23	1.12	VMP1	intronic
rs1077667	19	6668972	G	0.79	3.5 × 10-13	1.16	0.79	8.4 × 10-13	1.16	1.7 × 10-24	1.16	TNFSF14	intronic
rs34536443	19	10463118	C	0.95	1.2 × 10-8	1.28	0.96	2.9 × 10-7	1.30	1.8 × 10-14	1.29	TYK2	exonic
rs11554159	19	18285944	G	0.73	2.6 × 10-13	1.15	0.74	1.4 × 10-12	1.15	1.9 × 10-24	1.15	IFI30	exonic
rs8107548	19	49870643	G	0.25	2.0 × 10-6	1.09	0.26	2.5 × 10-10	1.13	5.7 × 10-15	1.11	DKKL1	intronic
rs4810485	20	44747947	A	0.25	1.8 × 10-5	1.08	0.25	1.4 × 10-12	1.14	7.7 × 10-16	1.11	CD40	intronic
rs2248359	20	52791518	G	0.60	9.8 × 10-5	1.07	0.62	8.2 × 10-11	1.12	2.0 × 10-13	1.09	CYP24A1	intergenic
rs2283792	22	22131125	C	0.51	1.1 × 10-6	1.08	0.53	5.4 × 10-11	1.11	5.5 × 10-16	1.10	MAPK1	intronic
Secondary
rs523604d	11	118755738	A	0.53	2.5 × 10-7	1.09	0.54	4.0 × 10-9	1.11	6.2 × 10-15	1.10	CXCR5	intronic

All listed signals had a discovery P-value ≤ 1.0 × 10^-4, a replication P-value ≤ 5.0 × 10^-2, and a joint P-value ≤ 5.0 × 10^-8

All P-values are two-sided

RA = Risk Allele, RAF = Risk Allele Frequency

^aPosition is based on human genome 19 and dbSNP 137.

^bNearest gene listed if within 50Kb. Bold indicates Gene Ontology Immune System Process.

^cA proxy SNP (rs10431552, r² = 0.99) was used in replication.

^dThe P-value and OR values provided are after conditioning on rs9736016 and rs533646 (both Novel – see Table 1).

^*These three SNPs were not primary in the 2011 GWAS, two were secondary and the third was tertiary in that study.

The strongest novel association, rs12087340 (p_joint = 1.1 × 10^-20, OR = 1.21), lies between BCL10 (B-cell CLL / lymphoma 10) and DDAH1 (dimethylarginine dimethylaminohylaminohydrolase 1). The protein encoded by BCL10 contains a caspase recruitment domain (CARD) and has been shown to activate NF-kappaB.¹⁷ The latter is a signalling molecule that plays an important role in controlling gene expression in inflammation, immunity, cell proliferation, and apoptosis. It has been pursued as a potential therapeutic target for multiple sclerosis.¹⁸ BCL10 is also reported to interact with other CARD domain containing proteins including CARD11.¹⁹ We have also identified a novel association of rs1843938 (p_joint = 1.2 × 10^-10, OR = 1.08), which is only 30 kb from CARD11.

One novel SNP was found within an exon, rs2288904 (p_joint = 1.6 × 10^-11, OR= 1.10); a missense variant in SLC44A2 (solute carrier family 44, member 2). Notably, this variant is also reported as a monocyte-specifccis-acting eQTL for the antisense transcript of the nearby ILF3 (interleukin enhancer binding factor 3).²⁰ This protein was first discovered to be a subunit of a nuclear factor found in activated T-cells, which is required for T-cell expression of IL2, an important molecule regulating many aspects of inflammation.

Of the 49 previously identified effects,^9,10,21 37 are in designated fine-mapping regions, and 23 of these 37 signals were localized to a single gene based on genomic position (Supplementary Table 3). Recognizing that proximity does not necessarily indicate functional importance, this emphasizes the utility of dense mapping in localizing signals from a genome-wide screen. The ImmunoChip analysis furthered the understanding of previously proposed secondary signals at three loci (Supplementary Note and Supplementary Tables 4-6); in particular we showed that the effects of two previously proposed independent associations at the IL2RA locus are driven by a single variant, rs2104286.^7,22.

In an effort to define the functionally relevant variants underlying these associations, we further studied the regions surrounding the 97 associated SNPs using both a Bayesian and frequentist approach in 6,356 multiple sclerosis subjects and 9,617 healthy controls from the UK (Online Methods).²³ Based on imputation quality, fine-mapping was possible in 68 regions (Supplementary Table 7): 66 of 93 primary (Fig. 2A) and two of four secondary. Eight of the 68 regions were fine-mapped to high resolution (Table 3, Fig. 2B and Supplementary Fig. 96). One third of the variants identified in these eight regions were imputed, indicating reliance on imputation even with dense genotyping coverage.

Figure 2. Bayesian fine-mapping within primary regions of association.

a) Summary of the extent of fine-mapping across 66 regions in 9,617 healthy controls from the UK, showing the the physical extent of, the number of variants, and the number of genes spanned by the posterior 90% and 50% credible sets. b) Detail of fine-mapping in region of TNFSF14. Above the x-axis indicates the Bayes Factor summarizing evidence for association for the SNPs prior to conditioning (blue markers) while below the x-axis indicates the Bayes Factor after conditioning on the lead SNP (rs1077667). Mb=Megabases.

Table 3.

The 18 variants from the 8 regions with consistent high resolution fine-mapping

Gene	SNP	Chr	Positiona	Posterior	GERP	Functional Annotationb
TNFSF14	rs1077667	19	6668972	0.74	-3.89	intronic, TFBS / DNase1 peak, correlates with serum levels of TNFSF14
IL2RA	rs2104286	10	6099045	0.93	-0.47	intronic, correlates with soluble IL-2RA levels
TNFRSF1A	rs1800693	12	6440009	0.69	2.53	intronic, causes splicing defect and truncated soluble TNFRSF1A
	rs4149580c	12	6446990	0.10	2.06	intronic
IL12A	rs1014486	3	159691112	0.67	0.24	-
CD6	rs34383631	11	60793330	0.20	1.66	-
	rs4939490c	11	60793651	0.14	-0.53	-
	rs4939491c	11	60793722	0.14	-0.37	-
	rs4939489	11	60793648	0.10	3.25	-
TNFAIP3	rs632574	6	137959118	0.27	-1.15	-
	rs498549c	6	137984935	0.20	0.52	-
	rs651973	6	137996134	0.17	2.41	downstream of RP11-95M15.1 lincRNA gene
	rs536331	6	137993049	0.15	0.19	upstream of RP11-95M15.1 lincRNA gene
CD58	rs6677309	1	117080166	0.21	-1.18	intronic, TFBS / DNase1 peak
	rs35275493c	1	117095502	0.24	0.75	intronic (insertion)
	rs10754324c	1	117093035	0.22	0.32	intronic
	rs1335532	1	117100957	0.17	-1.32	intronic
STAT4	rs78712823	2	191958581	0.59	-3.98	intronic

All listed variants have posterior ≥ 0.1 in regions where ≤ 5 variants explain the top 50% of the posterior and the top SNP from the frequentist analysis lives in the 90% confidence interval, ordered by maximum posterior.

Posterior denotes the posterior probability of any variant driving association. GERP denotes Genomic Evolutionary Rate Profiling.

^aPosition is based on human genome 19 and dbSNP 137.

^bFunctional data from VEP, eQTL browser, Fairfax et al. (2012), pubmed searches, 1000G. Dash indicates intergenic with no additional annotation. Variants without annotation are intergenic and have no reported regulatory consequence.

^cImputed variant.

To assess whether functional annotation²⁴ provides clues about the molecular mechanisms associated with genetic risk, we considered the relationship of variants to described coding and regulatory features in these eight regions. Although we found no variants with missense or nonsense effects, there was a notable enrichment for variants with functional effects: one known to affect splicing,²⁵ three known to correlate with RNA or serum protein levels^22,26,27 and several in transcription-factor binding and DNase I hypersensitive sites.^{28, 29} Four of the 18 variants in the fine-mapped regions are within conserved regions (GERP > 2).³⁰ This lack of functional annotation likely reflects the limited repertoire of reference expression and epigenomic profiles and suggests that the function of the variants may be cell-type or cell-state specific, as has been reported for many eQTLs in immune cell types.²⁰

To determine the Gene Ontology (GO) processes of the 97 associated variants, we used MetaCore from Thomson Reuters (Online Methods). We found the majority of the 97 variants lie within 50 kb of genes having immunological function. Of the 86 unique genes represented, 35 are linked to the GO immune system process (Table 1 and Table 2). We do not see a substantial over- or under- representation of certain GO processes when comparing the novel and previously identified loci, but this may be a limitation of ImmunoChip targeting genomic loci enriched for immunologically active genes, with more subtle distinctions between them not adequately captured by broad annotations such as GO.

Finally, we explored the overlap between our findings and those in autoimmune diseases with reported ImmunoChip analyses. We calculated the percentage of multiple sclerosis signals (110 non-MHC, Supplementary Table 8) overlapping those of other autoimmune diseases by requiring an r² ≥ 0.8 between the best variants reported in each study using SNAP.³¹ In total we find that ~22% of our signals overlap at least one other autoimmune disease. More specificially, ~9.1% overlap with inflammatory bowel disease (IBD) - ~7.3% with ulcerative colitis (UC), ~9.1% with Crohn’s disease (CD) -¹⁵, ~9.1% with primary biliary cirrhosis (PBC),^{32, 33} ~4.5% with celiac disease (CeD),³⁴ ~4.5% with rheumatoid arthritis (RhA),³⁵ ~0.9% with psoriasis (PS),³⁶ and ~2.7% with autoimmune thyroid disease (AITD).³⁷ We report the same top variant seen in PBC for 7 loci. We also note that our best TYK2 variant (rs34536443)³⁸ is also the most associated variant for PBC, PS and RhA. Lastly, AITD, CeD, PBC, and RhA report variants with pairwise r² ≥ 0.8 to the multiple sclerosis variant near MMEL1³⁹ (Supplementary Table 8).

In summary, we have identified 48 new multiple sclerosis susceptibility variants. These novel loci expand our understanding of the immune system processes implicated in multiple sclerosis. We estimate that the 110 non-MHC established risk variants explain 20% of the sibling recurrence risk; 28% including the already identified MHC effects⁹ (Supplementary Note). Additionally, we have identified five regions where consistent high resolution fine-mapping implicated one variant which accounted for more than 50% of the posterior in previously identified regions of TNFSF14, IL2RA, TNFRSF1A, IL12A, and STAT4. Our study further implicates NF-kappaB in multiple sclerosis pathobiology¹⁸, emphasizes the value of dense fine-mapping in large follow-up data sets, and exposes the urgent need for functional annotation in relevant tissues. Understanding the implicated networks and their relation to environmental risk factors will promote the development of rational therapies and may enable the development of preventive strategies.

Online Methods

ImmunoChip data (discovery set)

Details of case ascertainment, processing and genotyping for the discovery phase are provided in the Supplementary Note (Supplementary Table 9). Genotype calling for all samples was performed using Opticall.⁴⁰ Samples that performed poorly or were determined to be related were removed (Supplementary Table 10). The data were organized in 11 country level strata: ANZ (Australia + New Zealand), Belgium, Denmark, Finland, France, Germany, Italy, Norway, Sweden, United Kingdom (UK), and the United States of America (USA). SNP level quality control (Supplementary Table 11) and population outlier identification using principal components analysis (Supplementary Fig. 97) were done in each stratum separately.

Discovery set analysis

We applied logistic regression, assuming a per-allelic genetic model per data set, including the first five principal components as covariates to correct for population stratification (Supplementary Table 12 lists the per data set genomic inflation factors, λ). We then performed an inverse-variance meta-analysis of the 11 strata, under a fixed effects model, as implemented in PLINK.⁴¹ To be more conservative and account for any residual inflation in the test statistic, we applied the genomic control equivalent to the per-SNP standard error in each stratum. Specifically, we corrected the SNP standard errors by multiplying them with the square root of the raw genomic inflation factor λ, per data set, if the λ was >1.

Within the designated fine-mapping intervals, we applied a forward stepwise logistic regression to identify statistically independent effects. The primary SNP in each interval was included as a covariate, and the association analysis was repeated for the remaining SNPs. This process was repeated until no SNPs reached the minimum level of significance (p-value <1 × 10^-4). Outside of the designated fine-mapping intervals, all SNPs having a p-value <1 × 10^-4 were identified and grouped into sets based on a physical distance of less than 2Mb and a similar stepwise regression model was applied. Any SNPs to enter the model with p-value <1 × 10^-4 after conditioning were considered statistically independent primary signals.

In addition, because of the close physical proximity between some fine-mapping intervals and SNP sets, independence was tested for all identified signals within 2Mb of one another. The and cluster plots (Supplementary Fig. 98) of all independent SNPs were examined, and the SNP was excluded if unsatisfactory. If any SNP was excluded, the forward stepwise logistic regression within that fine-mapping interval or SNP set was repeated after removal of the SNP. During this process, 17 additional SNPs were excluded based on cluster or forest plot review.

Replication Set

The replication phase included GWAS data organized into 15 strata. Within each stratum, poorly performing samples (call rate < 95%, gender discordance, excess heterozygosity) and poorly performing SNPs (Hardy-Weinberg equilibrium (HWE) p-value <1 × 10^-6, minor allele frequency (MAF) < 1%, call rate < 95%) were removed. Principal components analysis was performed to identify population outliers per stratum, and the genomic control inflation factor was < 1.1 for each. The data included in the final discovery and replication analyses are summarized in Supplementary Table 13 and Supplementary Table 14. All the samples used in the replication set were unrelated to those in the discovery set; verified by identity-by-descent analysis.

We attempted replication of all non-MHC independent signals that reached a discovery p-value of <1 × 10^-4 in a meta-analysis set of GWAS. Each data set was imputed to the 1000 Genomes European phase I (a) panel using BEAGLE⁴² to maximize the overlap between the Immunochip SNP content and the GWAS data. Post-imputation genotypic probabilities were used in a logistic regression model, per stratum, to estimate SNP effect sizes and p-values. By using the post-imputation genotypic probabilities, we penalized SNPs that didn’t have good imputation quality, thus ensuring a conservative analysis. Furthermore, we accounted for population stratification in each data set by including the first five principal components in the logistic model. We then meta-analysed the effect size and respective standard errors of the 15 strata using a fixed effects model inverse-variance method. We applied the genomic control equivalent to the per-SNP standard error in each stratum, controlling for the respective genomic inflation factor λ (Supplementary Table 14).

To replicate the primary SNPs per identified signal in the discovery phase, we used the replication effect size and respective standard error. For the secondary and tertiary SNPs, we fitted the same exact models as in the discovery phase, per data set. We then performed fixed effects meta-analysis to estimate an effect size that corresponds to the same logistic model. In the case that a SNP was not present in the replication set, we replaced it with a perfectly tagging SNP, i.e. a SNP that had r² and D’ equal to 1. If a perfectly tagging SNP was not available, we selected a SNP that had equivalent MAF and the highest possible r² and D’. Estimation of r² and D’ for this objective were based on the ImmunoChip control samples.

Joint analysis (discovery and replication sets)

The discovery and replication phase effect sizes and respective standards errors were meta-analysed under a fixed effects model. A SNP was considered replicated when all three of the following criteria were met: 1) replication p-value <5.0 × 10^-2, 2) joint p-value <5 × 10^-8, and 3) the joint p-value was more statistically significant than the discovery p-value. SNPs that reached a p-value of <1 × 10^-6 but did not pass the genome-wide threshold, were coined suggested if the above criteria 1) and 3) were met.

Fine-mapping of association signals

To fine-map signals of association we used a combination of imputation and Bayesian methodology.²³ Around each of the 97 associated SNPs, 2Mb were isolated in the discovery and replication phase UK data as well as the European samples from the Phase 1 1000G.²⁸ Forming the single largest cohort, only UK samples were considered to minimize the effects of differential imputation quality between populations of different ancestry. In addition to the previous quality control, SNPs with failed alignment or a difference in MAF > 10% between the typed cohorts and the 1000G samples, MAF < 1%, or HWE p-value <1.0 × 10^-4 were removed.

Imputation was performed separately for the UK discovery and replication cohorts on each 2Mb region using the default settings of IMPUTEv2.^43,44 Missing genotypes in the genotyped SNPs were not imputed, and any imputed SNP that failed the HWE and MAF threshold was subsequently removed. We carried out frequentist and Bayesian association tests on all SNPs in each cohort separately, assuming additivity, using the default settings of SNPTESTv2.⁴⁵ Frequentist fixed-effect meta-analysis was carried out using the software META.⁴⁶ Bayesian meta-analysis was carried out using an independence prior (near-identical results were obtained using a fixed-effect Bayesian meta-analysis).

To identify regions where reliable fine-mapping could be achieved, we used the information score (INFO, obtained from IMPUTEv2) as identified from the 1000G samples. Specifically, we measured the fraction of variants with both r² > 0.5 and r² > 0.8 to the primary associated variant, having greater than 50% and 80% INFO scores respectively. Regions where any SNP with r² > 0.5 had INFO < 50% were excluded. We also excluded regions where the top hit from imputation had an INFO score less than 80%. Regions were considered to be fine-mapped with high quality when all variants with r² > 0.8 had at least 80% INFO. Within these regions, we excluded variants where the inferred direction of association was opposite in the UK discovery and replication cohorts.

To measure the posterior probability that any single variant drives association, we calculated the Bayes Factor. Under the assumption that there is a single causal variant in the region, this is proportional to the probability that the variant drives the association.²³ We identified the smallest set of variants that contained 90% and 50% of the posterior probability. We called a region successfully and consistently fine-mapped if there were at most five variants in the 50% confidence interval and the top SNP from the frequentist analysis lived in the 90% confidence interval. For these regions, we annotated variants with information about evolutionary conservation, predicted coding consequence, regulation, published associations to expression or DNase I hypersensitive sites using ANNOVAR,⁴⁷ VEP,²⁴ and the eQTL browser, a recent immune cell expression study²⁰, and other literature.

Gene Ontology

To determine the GO processes for which our associated variants were involved, we used MetaCore from Thomson Reuters. We annotated the processes for the unique genes within 50Kb of the variants.

Cross disease comparison

In order to explore the potential overlap with variants identified across other autoimmune diseases, we calculated the percentage overlap of reported variants found in other ImmunoChip reports to our ImmunoChip results. The top variants reported as either novel or previously known in other ImmunoChip reports were compared with the 110 variants representing both our novel and previous discoveries in multiple sclerosis. In order for a signal to be considered as overlapping, we required an r² ≥ 0.8 using the Pairwise LD function of the SNAP tool in European samples.³¹

Secondary analyses

We performed a severity based analysis of MSSS in cases only from the discovery phase (Supplementary Fig. 99). In addition, a transmission disequilibrium test was done in 633 trios to test for transmission of the 97 identified risk alleles (Supplementary Fig. 100). Details are given in the Supplementary Note.