Empirical Characteristics of Family-Based Linkage to a Complex Trait: the ADIPOQ Region and Adiponectin Levels

Abstract

We previously identified a low frequency (1.1%) coding variant (G45R; rs200573126) in the adiponectin gene (ADIPOQ) which was the basis for a multipoint microsatellite linkage signal (LOD=8.2) for plasma adiponectin levels in Hispanic families. We have empirically evaluated the ability of data from targeted common variants, exome chip genotyping, and genome-wide association study (GWAS) data to detect linkage and association to adiponectin protein levels at this locus. Simple two-point linkage and association analyses were performed in 88 Hispanic families (1150 individuals) using 10,958 SNPs on chromosome 3. Approaches were compared for their ability to map the functional variant, G45R, which was strongly linked (two-point LOD=20.98) and powerfully associated (p-value=8.1×10⁻⁵⁰). Over 450 SNPs within a broad 61 Mb interval around rs200573126 showed nominal evidence of linkage (LOD>3) but only four other SNPs in this region were associated with p-values<1.0×10⁻⁴. When G45R was accounted for, the maximum LOD score across the interval dropped to 4.39 and the best p-value was 1.1×10⁻⁵. Linked and/or associated variants ranged in frequency (0.0018 to 0.50) and type (coding, non-coding) and had little detectable linkage disequilibrium with rs200573126 (r²<0.20). In addition, the two-point linkage approach empirically outperformed multipoint microsatellite and multipoint SNP analysis. In the absence of data for rs200573126, family-based linkage analysis using a moderately dense SNP dataset, including both common and low frequency variants, resulted in stronger evidence for an adiponectin locus than association data alone. Thus, linkage analysis can be a useful tool to facilitate identification of high impact genetic variants.

Introduction

Family-based linkage analysis has been highly successful in identifying genetic loci underlying Mendelian disorders. In contrast, linkage analysis of complex traits and diseases in the general population has resulted in little success. While many complex trait and disease variants have been identified through genome-wide association studies (GWAS), the majority of these loci have small effect sizes and cumulatively explain relatively little of the overall risk(Kiezun et al. 2012; Manolio et al. 2009). These observations have led investigators to assess new approaches and to reassess methods such as family-based linkage analysis.

Large-scale exome- and genome-wide sequencing have facilitated creation of extensive resources for analysis of common, primarily non-coding variants and more recently common, low frequency, and rare coding variants through exome chip genotyping. This provides an ability to re-address the failures of family-based linkage approaches to identify complex trait loci. Family-based approaches remain a potentially powerful methodology for identification of complex trait loci. Patterns of segregation of uncommon or rare variants amplify power for detection compared to conventional population-wide association studies. With these new data resources the question arises as to how linkage analysis will perform in this framework and especially what can be expected when translated to empirical studies.

In prior reports we have described one of the few contemporary examples of family-based linkage (and association) with a complex trait(An et al. 2013; Bowden et al. 2010). In a microsatellite-based multipoint linkage analysis, a linkage peak for plasma adiponectin protein levels (LOD=8.2) overlying the ADIPOQ locus on chromosome 3 was identified in Hispanic families in the Insulin Resistance Atherosclerosis Family Study (IRASFS)(Guo et al. 2006). Common, non-coding variations did not account for this linkage, but a combination of conventional and exome sequencing revealed a novel coding variant (G45R; rs200573126) that segregated with low adiponectin levels (average 80% reduction) and was highly associated with plasma adiponectin (p=5.03×10⁻⁴⁰)(Bowden et al. 2010). This G45R variant was present at 1.1% frequency in the sample, contributed significantly to the variance in adiponectin levels (20%), and accounted for the previously observed linkage signal.

This low frequency coding variant which was the source of linkage and association with a complex trait was identified using targeted methods. An agnostic search for novel variants contributing to complex traits would likely include a genome-wide approach searching for linkage (and association) to complex traits in families. Here we have evaluated the performance of a combined linkage and association analysis approach in a locus-wide re-analysis of the ADIPOQ region. This analysis provides insight into the empirical signature of a low frequency, high impact causal variant in a background of genotype data from GWAS and exome chip sources. The characteristics of this linkage are relevant especially to the scenario in which a novel trait-defining variant has not been directly genotyped.

Materials and Methods

Samples

The samples used in this study were from the Hispanic cohort of the Insulin Resistance Atherosclerosis Family Study (IRASFS)(Henkin et al. 2003). Briefly, subjects were ascertained on the basis of large family size in San Luis Valley, Colorado, and San Antonio, Texas. The sample consisted of a maximum of 1414 individuals from 88 families with available genotype data. Detailed relationship information about these samples is included in Supplemental Table 1. A subset of these individuals (n=1150) had plasma adiponectin levels measured by radioimmunoassay (RIA; Linco Research, St. Charles, MO, USA) as previously reported(Bowden et al. 2010; Guo et al. 2006). IRB approval was obtained at all clinical and analysis sites and all participants provided informed consent.

Genotype Data

Single nucleotide polymorphism (SNP) genotype data from a variety of sources were collected. These included individually genotyped SNPs at the ADIPOQ locus (n=33)(An et al. 2012; An et al. 2013) which were combined with data from 7497 SNPs in the ADIPOQ region (chr3:161,560,463-197,838,262, hg19, 25 MB proximal and to the distal end of the chromosome) derived from Illumina OmniExpress (GWAS chip) genome-wide genotyping as part of the GUARDIAN Consortium(Goodarzi et al. 2013) and data from all chromosome 3 SNPs (n=3428) on the Illumina HumanExome Beadchip(Hellwege et al. 2014) for a total of 10,958 non-redundant, polymorphic SNPs. Quality control for each group of SNP data has previously been described in detail(An et al. 2012; Hellwege et al. 2014). All datasets underwent Mendelian error checking using PedCheck(O'Connell and Weeks 1998).

Statistical Analyses

All SNPs were evaluated for both two-point family-based linkage and single SNP association using Sequential Oligogenic Linkage Analysis Routines (SOLAR)(Almasy and Blangero 1998). Both analyses used age, sex, body mass index (BMI), and study center as covariates. Association analysis additionally included three admixture variables as covariates. Admixture was estimated using ADMIXTURE (Alexander et al. 2009) assuming five ancestral populations from exome chip-wide SNP data pruned for linkage disequilibrium (LD) to produce admixture estimates for the greatest number of samples. HapMap samples from CEU, YRI and CHN-JPT were included in the analysis. Three variables were selected as representing the variation in these Hispanic samples as larger values for K began to isolate individual pedigrees. Plasma adiponectin levels were transformed (natural log) to best approximate the distributional assumptions of the test. To test the influence of kurtosis, empirical LOD scores were calculated for each series of analyses using the lodadj procedure in SOLAR as appropriate. This procedure simulates the data to generate a distribution of LOD scores under the null hypothesis of no linkage at a given number of repetitions. For the purposes of this study, 100 replicates were used to generate the empirical LOD adjustment. Additional linkage and association analyses were performed adjusting for G45R as a covariate in the respective models.

To evaluate the relative performance of multipoint versus two-point linkage analysis, the previously published(Guo et al. 2006) microsatellite multipoint linkage analysis encompassing 25 microsatellite markers from chromosome 3 was contrasted with the two-point linkage analyses. The analyses were computed with and without the G45R polymorphism to observe the impact of functional variant on the evidence for linkage. A SNP-based linkage analysis panel was created from the GWAS data for multipoint analysis to reflect a more contemporary approach to linkage analysis. This SNP panel used the Illumina Human Linkage Panel IVb genotyping panel as its backbone. Of the 409 SNPs present from the linkage panel on chromosome 3, 160 (39%) of these were directly available from the GWAS. The remaining SNPs were selected using proxies for linkage panel SNPs (D'=1.00; r² > 0.98) while maintaining low linkage disequilibrium (D' < 0.40; r² < 0.20) between other markers in the panel. Unanticipated linkage disequilibrium between markers required pruning these 409 markers to an uncorrelated subset of 156 SNPs for multipoint analysis. Multipoint identity by descent (IBD) matrices were calculated using Loki(Heath 1997; Heath et al. 1997) as specified previously(Guo et al. 2006) and multipoint linkage analysis was performed using SOLAR.

Results

Two-point Linkage Analysis

Two-point linkage analysis was chosen as the primary linkage tool to avoid signal inflation due to LD between closely linked markers. Linkage analysis was carried out with a total of 10,958 SNPs on chromosome 3 comprising common, low frequency, and rare non-coding and coding variants in 1414 DNAs from 88 families. The highest density of SNPs was in a 36 Mb interval surrounding the ADIPOQ locus. Family sizes ranged from 2 to 33 individuals. A breakdown of the demographic characteristics by SNP variant class is summarized in Table 1. Comprehensive sample characteristics have been described previously(Bowden et al. 2010). SNP minor allele frequencies (MAF) ranged from 0.0018 to 0.5. The previously documented causal variant G45R (rs200573126) was the most strongly linked variant with a two-point LOD=20.98 (Table 2, Fig. 1, Supplemental Table 2). Additionally, many variants across chromosome 3 (n=453) also showed evidence of linkage, with LOD scores ranging from 3 to 14.85 (Table 2, Fig. 1). These SNPs spanned 56 MB proximal and 5.2 MB distal to the ADIPOQ locus (n=7,203 SNPs in the region) and included both common non-coding and lower frequency coding variants. In contrast, a very large number of SNPs (n=8,910) showed little evidence of linkage (LOD<1.0), including 73% of SNPs (n=5,293) within the 61MB region specified above. Minor allele frequencies for the nominally linked variants ranged from 0.0044 to 0.5. Adjusting for G45R as a covariate, the maximum observed two-point LOD score across the chromosome was reduced to 4.39 at rs10937349 (Fig. 2). This SNP rs10937349 is an intronic variant in the LPP gene (encoding LIM domain containing preferred translocation partner in lipoma), which is 1.65 Mb distal of ADIPOQ. This LPP SNP was not associated with adiponectin levels with or without adjustment for the G45R variant (p=0.53 and 0.76, respectively) (Figs. 1 and 2). The G45R adjusted analysis identified only four variants with LOD scores greater than 3 (Fig.2) which were widely dispersed across the greater ADIPOQ region, some being greater than 15 MB away from the ADIPOQ locus.

Table 1.

Demographic characteristics of the IRASFS Hispanic samples with adiponectin levels.

Characteristic	Directly Genotyped (33 variants)	Exome Chip (3428 variantsa)	GWAS (7497 variantsb)
Samplesc	1150	1144	955
Age (years)	41.1 (18-81)	41.1 (18-81)	39.3 (18-81)
% Female	58.4	58.4	58.4
BMI (kg/m2)	28.8 (16.5-58.1)	28.8 (16.5-58.1)	28.2 (16.5-58.1)
Adiponectin (μg/ml)	13.6 (1.2-99.8)	13.6 (1.2-99.8)	13.4 (1.3-39.7)
% T2Dd	15.3	15.3	4.9

Data presented as mean (range) or percent.

^aEntire chromosome 3.

^bchr3:161,560,463-197,838,262, hg19.

^cFrom 88 pedigrees.

^dType 2 diabetes.

Table 2.

Linkage and association of variants with LOD>8.0 and ADIPOQ variants.

SNP	Source	Position	Gene	Annotation	MAFa	LOD	G45R adjusted LOD	Association p-value	Proportion of Variance	D’	R2
rs4148556	GWAS	183734990	ABCC5	intron	0.12	8.75	1.90	0.79	0.000	0.406	0
rs12493550	GWAS	183752773	HTR3D	intron	0.15	10.16	2.93	0.54	0.001	0.77	0.001
rs9877051	GWAS	184326485	intergenic		0.38	9.81	2.17	0.22	0.002	1	0.006
rs10937185	GWAS	184409778	intergenic		0.15	14.85	1.89	0.0073	0.009	0.659	0.025
rs4687478	GWAS	184437480	intergenic		0.46	8.30	0.46	0.075	0.004	0.501	0.003
rs6770129	GWAS	184814444	C3orf70	intron	0.38	8.13	1.25	0.16	0.001	0.758	0.003
rs7647305	EC	185834290	intergenic		0.17	8.10	0.00	0.0046	0.016	0.527	0.016
rs11924144	EC	185845583	intergenic		0.36	9.81	0.58	0.081	0.006	0.878	0.014
rs4583642	GWAS	186126310	intergenic		0.43	9.65	1.46	0.059	0.006	0.81	0.005
rs1900618	GWAS	186338135	AHSG	intron	0.46	10.02	0.95	0.18	0.003	1	0.008
rs13073106	GWAS	186342060	intergenic		0.46	8.54	0.55	0.19	0.003	1	0.008
rs10937273	DG	186549695	intergenic		0.33	10.75	0.08	0.058	0.003	0.798	0.015
rs710445	DG	186561518	ADIPOQ	intron	0.44	6.09	0.78	0.64	0.000	0.435	0.001
rs16861205	DG	186561634	ADIPOQ	intron	0.13	0.14	0.13	0.74	0.000	1	0.001
rs16861209	DG	186563114	ADIPOQ	intron	0.064	0.04	0.07	0.032	0.006	1	0.001
rs822391	DG	186563803	ADIPOQ	intron	0.19	8.65	0.22	0.015	0.005	0.799	0.001
rs822394	DG	186566728	ADIPOQ	intron	0.19	9.25	0.20	0.028	0.004	1	0.003
rs822396	DG	186566877	ADIPOQ	intron	0.20	9.15	0.09	0.037	0.003	1	0.003
rs12495941	DG	186568180	ADIPOQ	intron	0.32	10.64	0.11	0.044	0.005	1	0.023
rs7649121	DG	186568785	ADIPOQ	intron	0.24	0.21	0.18	0.25	0.000	1	0.003
rs7627128	GWAS	186568799	ADIPOQ	intron	0.25	0.85	0.24	0.34	0.000	1	0.003
rs9877202	DG	186569607	ADIPOQ	intron	0.006	0.00	0.00	0.79	0.000	1	0
rs2036373	DG	186570191	ADIPOQ	intron	0.044	0.94	0.58	0.97	0.000	1	0
rs17366568	EC	186570453	ADIPOQ	intron	0.066	0.00	0.19	0.32	0.002	0.038	0
rs2241766	DG	186570892	ADIPOQ	G15G	0.18	10.71	0.03	0.11	0.002	1	0.036
rs16861210	DG	186570943	ADIPOQ	intron	0.058	0.01	0.03	0.16	0.003	1	0
rs200573126	EC	186570980	ADIPOQ	G45R	0.009	20.98	0.00	8.1×10−50	0.201	#N/A	#N/A
rs1501299	GWAS	186571123	ADIPOQ	intron	0.24	0.57	0.00	0.55	0.001	1	0.003
rs2241767	DG	186571196	ADIPOQ	intron	0.18	10.71	0.07	0.12	0.002	1	0.036
rs3821799	DG	186571486	ADIPOQ	intron	0.49	4.58	0.01	0.35	0.000	1	0.012
rs3774261	DG	186571559	ADIPOQ	intron	0.43	4.23	0.11	0.71	0.000	1	0.01
rs3774262	DG	186571814	ADIPOQ	intron	0.18	10.62	0.02	0.17	0.002	1	0.044
rs62625753	EC	186572026	ADIPOQ	G90S	0.004	0.62	0.70	0.011	0.004	1	0
rs17366743	EC	186572089	ADIPOQ	Y111H	0.016	0.50	0.64	0.86	0.000	1	0
rs6444174	DG	186573189	ADIPOQ	3′ UTR	0.029	0.16	0.20	0.94	0.000	1	0
rs6773957	EC	186573705	ADIPOQ	3′ UTR	0.43	4.09	0.07	0.84	0.000	1	0.013
rs1063537	GWAS	186574075	ADIPOQ	3′ UTR	0.18	3.78	0.00	0.30	0.001	1	0.044
rs2082940	DG	186574164	ADIPOQ	3′ UTR	0.19	10.69	0.01	0.10	0.003	1	0.035
rs1063538	GWAS	186574183	ADIPOQ	3′ UTR	0.43	2.33	0.10	0.59	0.000	1	0.013
rs9842733	DG	186575482	ADIPOQ	3′ UTR	0.004	0.00	0.00	0.81	0.000	1	0
rs1403697	DG	186576693	ADIPOQ	3′ UTR	0.004	0.00	0.00	0.81	0.000	1	0
rs7628649	DG	186585381	intergenic		0.21	10.95	0.00	0.17	0.001	1	0.029
rs9865432	GWAS	186597161	intergenic		0.47	8.59	1.08	0.014	0.006	1	0.009
rs2117986	GWAS	186600420	intergenic		0.47	8.65	1.10	0.018	0.005	1	0.009
rs4686443	GWAS	186619943	intergenic		0.36	10.56	1.17	7.0×10−4	0.016	1	0.006
rs7645316	GWAS	186635984	intergenic		0.38	8.38	1.05	0.0033	0.011	1	0.006
rs12054151	GWAS	186784230	ST6GAL1	intron	0.054	8.00	0.00	1.0×10−7	0.031	1	0.167

DG= directly genotyped. EC= exome chip. GWAS= GWAS chip.

^aMinor Allele Frequency

Fig. 1.

Adiponectin two-point linkage and association on chromosome 3

Fig. 2.

Adiponectin two-point linkage and association on chromosome 3, adjusted for the presence of G45R

Multipoint Linkage Analysis: Microsatellite and SNP Performance

We have also revisited multipoint approaches such as the previously observed microsatellite multipoint linkage (maximum LOD = 8.2; Table 3)(Guo et al. 2006) as a complementary analysis to the two-point linkage results summarized above. When not adjusting for G45R (LOD=20.98), the largest two-point LOD score is observed with rs10937185 (LOD=14.85). By comparison, adding G45R as a marker to the microsatellite multipoint increased the LOD score marginally from 8.2 to 8.4 (Table 3). Review of the microsatellite data revealed that one allele of a neighboring microsatellite marker TTTA040 (allele 160 bp) was in linkage disequilibrium with G45R (D'=0.96). Two other microsatellite markers were also in LD (D3S2427; D'=0.62 and D3S2398; D'=0.48) with G45R. Thus, how much this modest increase in the LOD score is due increased linkage content versus bias due to linkage disequilibrium is unclear (Table 3). To evaluate the performance of SNPs in the region in a multipoint framework, GWAS genotypes were pruned to replicate the Illumina HumanLinkage IVb genotyping panel (N=8000 SNPs), and further pruned to eliminate linkage disequilibrium between SNPs. Multipoint linkage analysis results with these SNP markers on chromosome 3 (N=156) yielded a LOD score of 6.4. When G45R was included, the LOD score rose to 10.9. Thus comparison of the approaches yielded similar overall results, albeit with substantially different maximum LODs.

Table 3.

Comparison of linkage analysis methods with and without G45R.

Variant Source	Analysis Type	Maximum LOD score
		Without G45R	Including G45R
GWAS+Exome Chip	Two-point Linkage	14.9	20.9
Microsatellite Panel	Multipoint Linkage	8.0	8.4
SNP Linkage Panel	Multipoint Linkage	6.4	10.9

Association Analysis

In parallel with the two-point linkage analysis, the SNP dataset was analyzed for association with adiponectin. The SNP association analysis across the greater ADIPOQ region confirmed the G45R association with adiponectin (p-value=8.1×10⁻⁵⁰) but provided little evidence of association at other loci (Table 4). The lack of association of other SNPs across the region (Fig. 3) is consistent with the existence of only a single functional variant in the region and a lack of linkage disequilibrium between G45R and any other SNP (r²≤0.20). This observation was reinforced when focusing on association results for SNPs with LOD scores over 8 (Table 2) where association p-values ranged from nominal 2.5×10⁻⁴ to non-significant. Overall a total of 866 variants demonstrated at least nominal evidence of association (p<0.05), including only three with p-values less than 1.0x10^-5. In addition to G45R, among these highly associated SNPs were an intronic variant (rs12054151; p-value=1.0×10⁻⁷) in the ST6GAL1 gene over 200kb distal of G45R and a coding variant (rs200567888; K207N; p-value=4.3×10⁻⁶) in the EPHA3 gene which is located nearly 100Mb distal to ADIPOQ. Results of association in the immediate region of 400 Kb surrounding ADIPOQ are presented in Supplementary Figure 1.

Table 4.

Top Association results

SNP	Position	Source	Gene	Variant	MAFa	N	Associationb			G45R Adjusted Associatione		Linkage	G45R Adjusted	LD with G45R
							P-value	β c	PoVd	P-value	β	LOD	LOD	D’	r2
rs200573126	186570980	EC	ADIPOQ	G45R	0.009	1136	8.1E-50	−1.56	0.201	1.00	0.00	20.98	0.00	N/A	N/A
rs12054151	186784230	GWAS	ST6GAL1	intron	0.055	947	1.0E-07	−0.26	0.031	0.88	−0.01	7.99	0.00	1	0.167
rs200567888	89259477	EC	EPHA3	K207N	0.002	1136	4.3E-06	−1.32	0.025	1.00	0.00	1.23	0.31	0.491	0.048
rs7635994	181938870	GWAS	intergenic		0.12	947	1.3E-05	−0.16	0.013	0.0089	−0.09	5.64	0.00	0.42	0.013
rs17540772	181498753	GWAS	intergenic		0.080	947	5.0E-05	−0.18	0.007	0.029	−0.09	4.20	0.00	0.467	0.025
rs3907825	186812036	EC	intergenic		0.37	1136	5.7E-05	−0.10	0.017	0.040	−0.08	2.39	0.69	0.747	0.009
rs192746498	170811629	EC	TNIK	E899A	0.006	1136	8.7E-05	−0.67	0.022	1.00	0.00	0.95	0.01	0.619	0.152
rs188780512	69244498	EC	FRMD4B	V418I	0.002	1136	1.3E-04	−1.09	0.010	1.00	0.00	1.29	1.48	1	0
rs3752904	183996068	GWAS	ECE2	A210A	0.40	947	1.5E-04	0.09	0.019	0.016	0.05	6.77	0.00	0.85	0.01
rs1382269	137108482	EC	intergenic		0.45	1136	1.9E-04	0.09	0.019	1.1E-05	0.16	0.00	0.06	1	0.008
rs7639661	181955695	GWAS	intergenic		0.076	947	2.0E-04	−0.17	0.010	0.062	−0.08	6.35	0.04	0.413	0.02
rs55872908	167512569	EC	SERPINI1	A280T	0.014	1136	2.1E-04	−0.37	0.006	0.94	−0.02	3.09	0.29	0.538	0.18
rs76947105	42448455	EC	LYZL4	P59S	0.005	1136	2.3E-04	−0.63	0.016	1.00	0.00	1.98	1.17	0.208	0.019
rs2228243	186395113	EC	HRG	H340R	0.173	1136	2.5E-04	−0.11	0.012	0.61	0.04	7.70	0.61	0.863	0.036
rs16860974	186387299	GWAS	HRG	intron	0.149	947	2.6E-04	−0.12	0.013	0.32	−0.03	3.95	0.86	0.868	0.042
rs9840545	181484270	GWAS	intergenic		0.107	947	2.6E-04	−0.14	0.012	0.056	−0.07	3.15	0.00	0.506	0.021
rs902954	171098694	GWAS	TNIK	intron	0.189	947	2.7E-04	0.11	0.009	0.0039	0.08	1.53	0.00	0.039	0
rs9856368	187756856	GWAS	intergenic		0.039	947	3.6E-04	−0.20	0.020	0.0050	−0.14	0.71	0.46	0.202	0.01
rs17601879	172258131	GWAS	intergenic		0.072	947	3.9E-04	−0.16	0.007	0.086	−0.07	2.93	0.05	0.477	0.029
rs1709679	183943141	GWAS	intergenic		0.22	947	4.2E-04	−0.10	0.015	0.037	−0.05	4.19	0.00	0.558	0.011
rs16832893	181855149	GWAS	intergenic		0.17	947	4.3E-04	−0.11	0.015	0.040	−0.06	3.60	0.00	0.467	0.01
rs2378358	187754948	GWAS	intergenic		0.12	947	4.7E-04	−0.12	0.014	0.012	−0.08	1.06	0.01	0.438	0.014
rs61736574	63968083	EC	ATXN7	N325S	0.005	1136	5.4E-04	−0.55	0.013	1.00	0.00	1.21	0.22	0.272	0.026
rs16825471	175057188	GWAS	NAALADL2	intron	0.038	947	5.8E-04	−0.20	0.013	0.30	−0.06	3.83	0.16	0.516	0.066
rs10513798	184488916	GWAS	intergenic		0.13	947	5.8E-04	−0.12	0.005	0.082	−0.05	3.90	0.00	0.426	0.012
rs4686599	191845112	GWAS	intergenic		0.23	947	6.2E-04	−0.09	0.006	0.043	−0.05	1.11	0.00	0.468	0.007
rs4686443	186619943	GWAS	intergenic		0.37	947	7.0E-04	0.08	0.016	0.040	0.05	10.56	1.17	1	0.006
rs9816189	162287573	GWAS	intergenic		0.16	947	7.0E-04	−0.11	0.000	0.15	−0.04	4.31	0.00	0.725	0.027
rs10804838	169857007	GWAS	PHC3	intron	0.064	947	7.3E-04	−0.16	0.013	0.21	−0.06	1.17	0.10	0.544	0.042
rs12632166	184485263	GWAS	intergenic		0.084	947	7.9E-04	−0.14	0.014	0.27	−0.04	3.34	0.00	0.477	0.024
rs56181422	39306947	EC	CX3CR1	L352L	0.001	1136	8.0E-04	−0.77	0.017	1.00	0.00	0.00	0.13	0.66	0.065
rs9821439	186791795	GWAS	ST6GAL1	intron	0.46	947	8.1E-04	−0.08	0.014	0.14	−0.03	4.33	0.20	1	0.012
rs13059427	173514581	GWAS	NLGN1	intron	0.049	947	8.1E-04	−0.17	0.019	0.079	−0.08	2.81	0.00	0.448	0.038
rs6783938	184906201	GWAS	LOC339926	intron	0.17	947	8.9E-04	−0.10	0.017	0.22	−0.03	5.35	0.00	0.794	0.029
rs11716002	186613364	GWAS	intergenic		0.30	947	9.0E-04	0.08	0.023	0.034	0.05	5.79	0.06	1	0.004
rs35713889	49162583	EC	LAMB2	G914R	0.028	1136	9.2E-04	−0.24	0.005	1.00	0.00	0.86	1.38	0.131	0
rs149492706	119219579	EC	TIMMDC1	C78R	0.010	1136	9.4E-04	−0.37	0.010	1.00	0.00	0.08	0.09	0.182	0.028
rs13067039	173563692	GWAS	NLGN1	intron	0.052	947	9.4E-04	−0.17	0.019	0.34	−0.05	2.28	0.00	0.612	0.068
rs6796724	172374119	GWAS	NCEH1	intron	0.29	947	9.5E-04	0.08	0.002	0.0065	0.06	0.00	0.16	1	0.004
rs9290844	186807282	GWAS	intergenic		0.35	947	9.5E-04	−0.08	0.017	0.041	−0.05	1.36	0.46	0.764	0.01

DG= directly genotyped. EC= exome chip. GWAS= GWAS chip.

^aMinor Allele Frequency.

^bModel 1 included age, gender, BMI, study center, and admixture estimates as covariates.

^cNatural log units per copy of the minor allele.

^dProportion of variance explained.

^eModel 1 + G45R carrier status.

Fig. 3.

Linkage disequilibrium surrounding G45R (indicated by the red star) in a subset of 225 unrelated individuals from IRAS-FS

Color scheme represents pairwise r2 values.

Association analysis was also computed adjusting for the presence of G45R (Fig. 2, Table 3). It is of note that the associations above (rs12054151 and rs200567888) were reduced to p-values of 0.88 and 1.0, respectively, upon adjustment for G45R. A total of 477 variants also showed evidence of association (p<0.05), with the most strongly associated SNPs having p-values of 1.1×10⁻⁵ and 4.1×10⁻⁵. These two variants, however, were located 49 MB distal from G45R and separated by 22 kb. Both variants are intergenic and were genotyped as part of the exome chip. Of the 477 associated variants, 240 of them were nominally associated (P<0.05) in the original model (i.e. without adjustment for G45R), suggesting that adjusting for G45R identifies additional variants which were masked when the strongly associated causal variant was present.

Discussion

In this study we explored the characteristics of a genetic linkage signal due to a low frequency, high impact mutation in the ADIPOQ gene. These characteristics are especially relevant to ongoing searches for high impact variants contributing to the variance of complex traits. The targeted coding variant, ADIPOQ G45R, is a definitive example of a low frequency, high impact variant: it was present in 7 of 88 families in IRASFS with a 1.1% frequency, yet accounts for a high multipoint LOD score (8.2) and contributes substantially to variance in adiponectin levels in this Hispanic sample (17%)(Bowden et al. 2010) is thus a high impact variant. We examined the characteristics of linkage using a contemporary genetic dataset of Illumina Omni Express GWAS SNPs and exome chip data from chromosome 3. Due to the characteristics of this variant (low frequency and only observed in Hispanics), we sought to determine whether surrounding variants would have been sufficient to draw the attention of an investigator to this region, i.e. if we did not know of or have data for the functional variant, would we be able to detect its presence? The two-point linkage analysis across this chromosome implicated many SNPs, aside from G45R, with the most strikingly linked being the intergenic SNP rs10937185 (LOD=14.85; MAF=0.15) which is nearly 2 MB away from G45R (Figure 1). Thus, linkage results without G45R would present a broad region effectively similar to linkage peaks observed in past multipoint linkage analyses of Mendelian traits. Therefore, evidence of linkage can be observed with both common and low frequency surrogates for a low frequency coding and functional variant in the absence of data for the functional variant. To emphasize, this result is obtained with a high density contemporary dataset consisting primarily of GWAS and exome chip data. After adjustment for the functional G45R, the LOD scores dropped substantially with just one SNP with a LOD score greater than 3. For example, the LOD score for rs10937185 dropped to 1.89 (Fig. 2, Table 2). There was very little linkage evidence implicating additional variants nearby. In the opposite case, adjusting for rs10937185 as a covariate reduced the LOD score at G45R only slightly, from 20.98 to 17.17 (data not shown).

In parallel with the linkage analysis we performed single SNP association analysis. The two-point linkage analysis results are an easily interpretable complement to the association analysis which readily facilitates alignment of linkage and association signals. The strongest association signal other than G45R was rs12054151 (p-value=1.0×10⁻⁷), located in the ST6GAL1 gene which is 213 kb from ADIPOQ and has a LOD score of 7.99 (Table 4). This variant nominally explained an estimated 3.1% of the variance in adiponectin levels. In the context of gene discovery, this combination of linkage and association would be fairly promising; however, there was no residual association after adjusting for the presence of G45R (p-value=0.88; Table 4). Overall, there was limited evidence of association across chromosome 3. When viewed from the perspective of a GWAS, there were some notable SNPs, but none with compelling evidence of association that would draw the eye of an investigator and few had any evidence of support by nearby SNPs.

Examining the intersection of linkage and association identified 119 variants which had both p-values less than 0.05 and a LOD score greater than 3. Looking at the results more stringently (p<0.001 and LOD >3) left just 20 variants (Supplemental Figures 2 and 3), though the G45R mutation with association p-value of 8.1×10⁻⁵⁰ and LOD score of over 20 shows that the functional variant substantially outperforms surrogate markers. Conversely, when examining the G45R adjusted results, there were just four variants with a LOD score greater than 3 and only 10 with a p-value less than 0.001. These two groups were, however, mutually exclusive: there were no nominally associated SNPs with a LOD score greater than 3.

This analysis of the characteristics of linkage and association across a region when an uncommon causal variant is known leads to several conclusions. Perhaps most obvious is that the evidence of two-point linkage for the functional G45R variant was substantial, yielding a LOD score of 20. Parenthetically, the adiponectin trait demonstrates high kurtosis (lack of normality after adjustment for covariates) in the analyses presented here. When accounting for kurtosis, the maximum LOD score with G45R is reduced (LOD≈10) but overall inferences are the same (Supplemental Table 2). Large numbers of variants surrounding the causal variant over an extended region show substantial evidence of linkage (LODs of up to 14.85). This linkage evidence is driven by the G45R variant, however (Fig. 2), indicating the power to detect a strikingly causal variant even if it were not typed and analyzed. Thus, linkage in this scenario of a low frequency, high impact variant is perhaps surprisingly consistent with conventional linkage patterns even though the 1.1% MAF variant appeared in only 7 of 88 pedigrees.

The previously observed striking microsatellite linkage may be due to the chance LD between G45R and a microsatellite allele (D'=0.96). In comparing multipoint with two-point methods using SNP data, this was not as strongly detected when using a SNP panel constructed from 409 common SNPs (pruned to n=156) which apparently do a poorer job of capturing LD with the functional variant. It is unclear whether this is due to the primarily European-derived information used in developing the original linkage chip being different from that of our Hispanic samples, or if there is some other confounding aspect to the SNPs chosen. It is unclear if information content is the reason for differences in performance, but it seems implicit that a two point analysis of a high density dataset of GWAS and exome chip based SNPs would perform at least at the equal with “traditional” approaches.

The ability of multiple markers in the simple two-point analysis to detect a linkage signal has important implications that suggest extension of this approach to other traits and other study samples. The two point linkage approach has clear advantages especially for a seamless alignment with complementary association data (e.g. Fig. 1). First, this documents the power of linkage analysis to facilitate identification of high impact genetic variants even in a sample of families which would be considered modest in size by current standards for genetic association studies. Importantly, this could encompass more sophisticated phenotypes rather than the generic phenotypes addressable in most GWAS studies (e.g. acute insulin response rather than type 2 diabetes). Second, and most importantly, significant evidence of linkage can be observed even with both common and low frequency surrogates for a low frequency coding and functional variant. Notably, in this example, the low frequency causal variant had a much higher LOD score than non-causal neighboring SNPs, providing further support to its relevance. Consequently, linkage analysis has the potential for discovery of novel variants from genome-wide resources of common (GWAS) or coding (exome chip) data sets even when the functional variant was not directly genotyped such as a private variant in a single family.

Further implications are that common, non-coding variants identify the linkage signal. Thus GWAS datasets that have been generated in most large ongoing studies are readily applicable for this approach. In addition, while this manuscript has focused on the scenario of low frequency high impact variant in our prior work (Hellwege 2014) we showed that application of this approach is also relevant to common high impact variant discovery. It is implicit that such variants (APOE and CETP) should be identified in GWAS, it is noteworthy that APOE was strongly linked to APOB and much more weakly to LDL (for example). Thus the approach may be useful for common variants with more sophisticated phenotypes. Finally, it is increasingly clear that high impact non-coding variants exist and are potential contributors to variance in traits. This linkage method should be directly applicable to discovery of such variants also.