Genome-wide linkage and admixture mapping of type 2 diabetes in African American families from the American Diabetes Association GENNID (Genetics of NIDDM) Study Cohort.

OBJECTIVE: We used a single nucleotide polymorphism (SNP) map in a large cohort of 580 African American families to identify regions linked to type 2 diabetes, age of type 2 diabetes diagnosis, and BMI. RESEARCH DESIGN AND METHODS: After removing outliers and problematic samples, we conducted linkage analysis using 5,914 SNPs in 1,344 individuals from 530 families. Linkage analysis was conducted using variance components for type 2 diabetes, age of type 2 diabetes diagnosis, and BMI and nonparametric linkage analyses. Ordered subset analyses were conducted ranking on age of type 2 diabetes diagnosis, BMI, waist circumference, waist-to-hip ratio, and amount of European admixture. Admixture mapping was conducted using 4,486 markers not in linkage disequilibrium.
RESULTS: The strongest signal for type 2 diabetes (logarithm of odds [LOD] 4.53) was a broad peak on chromosome 2, with weaker linkage to age of type 2 diabetes diagnosis (LOD 1.82). Type 2 diabetes and age of type 2 diabetes diagnosis were linked to chromosome 13p (3-22 cM; LOD 2.42 and 2.46, respectively). Age of type 2 diabetes diagnosis was linked to 18p (66 cM; LOD 2.96). We replicated previous reports on chromosome 7p (79 cM; LOD 2.93). Ordered subset analysis did not overlap with linkage of unselected families. The best admixture score was on chromosome 12 (90 cM; P = 0.0003).
CONCLUSIONS: The linkage regions on chromosomes 7 (27-78 cM) and 18p overlap prior reports, whereas regions on 2p and 13p linkage are novel. Among potential candidate genes implicated are TCF7L1, VAMP5, VAMP8, CDK8, INSIG2, IPF1, PAX8, IL18R1, members of the IL1 and IL1 receptor families, and MAP4K4. These studies provide a complementary approach to genome-wide association scans to identify causative genes for African American diabetes.

Type 2 diabetes is marked by a clear genetic propensity, a high concordance in identical twins, tendencies for both diabetes and age of onset to be familial (1), and marked differences in prevalence among ethnic groups (2). Despite considerable evidence for a genetic predisposition, unraveling the genetic etiology has been daunting, with few confirmed genes identified from genome-wide linkage scans. Recent successes with genome-wide association scans (3) have greatly increased the number of confirmed genetic loci, but these successes have been limited primarily to Caucasian populations. With few exceptions (TCF7L2), these regions from genome-wide association scans overlap neither with published linkage scans nor with known candidate genes or pathways, and in many cases, the variants are of unknown functional significance. These confirmed diabetes genes are uniformly of small effect size, and together, they explain only a small portion of diabetes risk even in Caucasians (4). Only variants at TCF7L2 show evidence for an association in African American populations (5), but the association of TCF7L2 with diabetes and reduced insulin secretion has not been easily reproducible (6). Thus, the genetic predisposition for type 2 diabetes in minority populations remains largely unknown. Particularly for African-based populations, low levels of linkage disequilibrium increase the challenge of mapping genes by genome-wide association and argue for the importance of complementary methods.

Previous genome-wide linkage scans for type 2 diabetes in African American or African families have been limited. A small African American sample of 229 individuals (124 sibpairs) was included in the original microsatellite linkage scan of GENNID (Genetics of NIDDM) Study families (7) and showed linkage on chromosome 10 near microsatellite marker D10S1412. Sale et al. (8) examined 392 microsatellite markers in 675 individuals (638 affected sibpairs) and reported suggestive linkage on chromosome 6 near D6S1035 (163.5 cM); other regions on 7p showed an association with earliest onset diabetes. Rotimi et al. (9) examined 390 microsatellite markers in 691 individuals (343 affected sibpairs) from West Africa and found evidence for linkage on chromosomes 20q13.3 and 12q24. Studies of single nucleotide polymorphism (SNP)-based linkage scans have not been published. In contrast to Caucasian linkage scans, the few available studies of African-derived populations have not replicated regions of linkage.

From 1993 to 2003, the American Diabetes Association sponsored the collection of families and sibpairs with at least two type 2 diabetic siblings, known as the GENNID Study (10). The families were collected in three phases. Phase 1 (1993–1995) included multiplex families, but only one site (University of Chicago) contributed African American families. In 1995, criteria were changed to nuclear families, and subsequently, ascertainment was further expanded with four new sites for African American families. Hence, the sample set was expanded considerably beyond the earlier analysis of Ehm et al. (7). In the present study, we examined 1,496 individuals from 580 pedigrees ascertained at 10 centers and providing 1,021 sibpairs with genotypes. This study with 5,619 SNP markers is thus the largest study of type 2 diabetes genetics in families of African descent to date.

RESEARCH DESIGN AND METHODS

African American subjects were ascertained and tested using the protocols and procedures of the GENNID project, as described previously (10), from 10 sites: Birmingham, AL; Little Rock, AR; Chicago; Denver, CO; Los Angeles (two sites); Baltimore, MD; Raleigh-Durham, NC; Charleston, SC; and St. Louis, MO (Table 1). Pedigrees were ascertained between 1993 and 2003 in three phases, all requiring two affected siblings. Phase 1 included multigenerational families ascertained in Chicago. Phase 2 included nuclear families with affected siblings, parents if available, or unaffected siblings where both parents were not available. Phase 3 included affected sibpairs. The distribution of individuals and sibpairs by site is listed in Table 1. Age of diagnosis was obtained by history using a standardized questionnaire (10). Physical examination data and DNA were available on 1,496 individuals from 580 pedigrees (Table 1). After removal of apparent sample discrepancies, the sample was reduced to 1,450 individuals, of whom 73 were the only members of the families and 9 were unrelated to any other pedigree member. Additional samples were removed subsequently for missing diabetes status (n = 2), missing current age (n = 1), age of diagnosis reported <15 years (n = 36), or BMI <14 or >55 kg/m2 (n = 65). Relationships included 1,021 full sibpairs and are shown in supplementary Table S1, available in an online appendix at http://dx.doi.org/10.2337/db08-0931. For the analyses reported, we included 1,082 affected individuals for sibpair analyses and variance component analysis with age of type 2 diabetes diagnosis and 1,344 individuals for analysis of BMI and type 2 diabetes.

TABLE 1

Distribution of study participants by site

Site	Individuals	Sibling pairs
AL	50	33
AR	238	274
CH	411	432
CO	55	65
LA	126	113
LX	70	33
MD	101	45
NC	155	98
SC	137	54
SL	153	146
Totals	1,496	1,293

Data are n. Distribution of individuals and sibpairs by site, as listed in research design and methods. AL, Birmingham, AL; AR, Little Rock, AR; CH, Chicago; CO, Denver, CO; LA, Cedars Sinai, Los Angeles, CA; LX, University of California, Los Angeles, CA; MD, Baltimore, MD; NC, Raleigh-Durham, NC; SC, Charleston, SC; SL, St. Louis, MO.

DNA preparation and genotyping.

DNA was prepared from lymphoblastoid cell lines by Coriell Cell Repository (Camden, NJ) by salting out; DNA quality was checked by pico-green, and concentration was adjusted for submission to the Center for Inherited Disease Research (CIDR). Genotyping was completed using the Illumina Golden Gate assay to type Illumina Linkage Panel IVb, comprising 6,008 SNPs, of which 5,958 were released. Data were analyzed using the Illumina BeadStudio Genotyping Module, v2.3.41.

Quality control and error corrections.

CIDR reported 0.22% missing genotypes and 99.926% Mendelian consistency. Concordance between 66 blind duplicates was 99.9985%. Pedigree structures were examined for 1,413 SNPs selected for spacing at ≥2 cM using the Eclipse2 program (11). Likelihoods of eight relationships (monozygotic twins, full siblings, half siblings, parent-offspring, avuncular, grandparent-grandchild, and unrelated) were calculated. Two pairs of monozygotic twins were identified, 11 individuals were removed due to apparent sample mix-ups, and 89 families with nonpaternity or nonmaternity were identified and corrected. An additional eight pedigrees were merged based on likelihoods supporting first- or second-degree relationships between pairs of pedigrees from the same site. Resulting corrections reduced the sample size to 1,450 members of 534 pedigrees, for which 4,004 genotypes (0.047%) were missing. Genotyping errors were identified using Pedcheck (12) to remove family members with Mendelian inconsistencies and MERLIN (13) to identify markers with excess recombination. These steps removed 6,435 genotypes (0.075% of the total). For seven markers, 5–10% of genotypes were missing. An additional 37 SNPs showed deviations from Hardy-Weinberg equilibrium at P < 0.005, and 31 showed deviations with 0.005 < P < 0.01. Removal of these possibly poor-quality SNPs had minimal or no effect on logarithm of odds (LOD) scores (data not shown).

Statistical analysis.

We conducted linkage analysis using variance component likelihood analysis for the traits type 2 diabetes, type 2 diabetes age of diagnosis/onset, and BMI and nonparametric linkage (NPL) analysis implemented in MERLIN (13) for type 2 diabetes. We report variance component linkage analysis using 5,914 markers and NPL analyses using 5,870 markers. For variance component analysis, we used a likelihood analysis implemented in jPAP (14). Sample size was 1,344 for type 2 diabetes and BMI and 1,082 for age of type 2 diabetes diagnosis. Detailed methods are provided in supplemental materials in the online appendix. Briefly, we modeled type 2 diabetes risk to account for age of type 2 diabetes diagnosis in affected pedigree members using a modification of the age at diagnosis regressive model described elsewhere (15). We modeled the logit probability of type 2 diabetes on age, BMI, and sex. Both age of type 2 diabetes diagnosis and BMI were modeled as a normal density with mean ± SD. Linkage was tested using multipoint identity by descent probabilities computed in MERLIN (13) and by comparing the maximized likelihood with a quantitative trait locus (QTL) to that without the QTL. The LOD score was calculated as the common logarithm of the likelihood ratio. NPL was calculated using MERLIN with NPL and S-Pairs statistics. We present best findings from all three methods; the NPL score gave more conservative results.

Linkage disequilibrium between SNPs could inflate LOD scores for linkage studies. To address this issue, SNPs were clustered using r2 thresholds of 0.25, 0.50, and 0.70, as implemented in MERLIN, assuming no recombination within clusters and no linkage disequilibrium between clusters. Using the three thresholds, we identified 338 SNPs in 147 clusters (r2 > 0.7), 462 SNPs in 202 clusters (r2 > 0.5), and 770 SNPs in 332 clusters (r2 > 0.25). The different thresholds had little effect on LOD scores (data not shown); we present results when correcting for clusters of SNPs with r2 > 0.7.

Ordered subset analysis was conducted using family-specific LOD scores from the MERLIN output for both NPL and S-Pairs statistics. Families were ranked from low to high and from high to low based on average-within-family values for age of type 2 diabetes diagnosis, BMI, waist circumference, waist-to-hip ratio (WHR), diabetes duration, and admixture. The maximum LOD score was calculated using the program FLOSS (16). Significant differences from the unranked LOD score were calculated based on a minimum of 500 permutations with ranks randomly assigned.

Association of each marker was calculated using the regression models described above and comparing the model with the SNP to one without the SNP. Genotypes were scored as 0/1/2 for homozygotes, heterozygotes, and rare homozygotes, thus assuming an additive model. An approximate χ2 statistic was calculated as twice the natural logarithm of the likelihood ratio with the SNP to the likelihood without the SNP, and a 1 degree of freedom (d.f.) P value was calculated.

Admixture mapping was conducted on 1,450 individuals using the program ADMIXMAP (17) to estimate the proportion of European admixture for each subject. To avoid effects of linkage disequilibrium, markers were limited to 0.05-cM spacing, resulting in 4,486 markers. Markers were selected to optimize the frequency difference between Caucasian and African (Yoruban) HapMap samples. X chromosome markers were excluded.

RESULTS

We first examined three traits, type 2 diabetes, age of type 2 diabetes diagnosis, and BMI, using variance components linkage in 5,914 markers for 1,344 members of 530 families. Given the close association of BMI with type 2 diabetes, the likely role in modifying age of type 2 diabetes diagnosis, and as a surrogate for insulin resistance, our primary goal was to determine whether BMI mapped to type 2 diabetes or age of type 2 diabetes diagnosis (“diabesity” loci). The study population was drawn from 10 sites (Table 1). Characteristics of the study population are shown in Table 2, and linkage results are shown in Fig. 1. The LOD scores >1.5 for all three traits are summarized in Table 3. Type 2 diabetes was most strongly linked to chromosome 2 (124–132 Mb), with maxima of 4.53 at 127 cM and 2.98 at 108 cM, and this region showed suggestive linkage to age of type 2 diabetes diagnosis (LOD 1.82 at 84 cM and 1.66 at 115 cM). Both type 2 diabetes and age of type 2 diabetes diagnosis showed suggestive linkage also to chromosome 13p (3–22 cM; LOD 2.42 for type 2 diabetes and 2.46 for age of type 2 diabetes diagnosis). The strongest linkage for age of type 2 diabetes diagnosis was on chromosome 18p (LOD 2.96; 54–96 cM). Type 2 diabetes also showed suggestive linkage to chromosome 7p (79 cM, 2.93), a region implicated previously (18). Other suggestive type 2 diabetes peaks were on chromosome 4 (135 cM; LOD 2.26), 11 (123 cM, 2.36), and 16 (56 cM, 2.43). The chromosome 18 linkage for age of type 2 diabetes diagnosis corresponded with previous reports in Caucasians (19,20). Best linkages for BMI were on chromosomes 1 (LOD 2.30, 69 cM), 18 (116 cM, at 2.45; telomeric to type 2 diabetes and age of type 2 diabetes diagnosis), and 21 (14 cM) but did not overlap with type 2 diabetes or age of type 2 diabetes diagnosis linkages except on chromosome 1 (70–90 cM and 123–130 cM; Table 3; Fig. 1).

TABLE 2

Characteristics of study population

Variable	Men	Women	Total
n (%)	521 (34.8)	975 (65.2)	1,496
Age (years)	53.3 ± 13.1	53.8 ± 12.8	53.6 ± 12.9
Diabetes	406 (77.9)	802 (82.3)	1,208 (80.8)
Onset age of diabetes diagnosis (years)	45.2 ± 11.7	44.8 ± 12.5	44.9 ± 12.3
BMI (kg/m2)	30.1 ± 6.2	34.4 ± 8.6	32.9 ± 8.1
WHR	0.96 ± 0.10	0.93 ± 0.13	0.94 ± 0.12
Fasting glucose (mmol/l)	8.77 ± 4.53	8.59 ± 4.05	8.65 ± 4.23
Fasting insulin (pmol/l)	136 ± 235	148 ± 224	144 ± 228

Data are n (%) or means ± SD.

FIG. 1.

Linkage plots for variance component and NPL methods. Fig. 1 shows all linkage analyses conducted in one figure. As noted in the text, numbers of individuals and numbers of markers varied across analyses. Variance component linkage for type 2 diabetes is black, for age of type 2 diabetes diagnosis is blue, and for BMI is green. LOD scores using MERLIN NPL scores are shown in red.

TABLE 3

Summary of linkage from variance component analysis

Chromosome	Location	Trait	Maximum LOD	SNP
1	32	Type 2 diabetes	2.27	Rs3766306
	69	BMI	2.30	Rs1707302
	114	AOD	1.98	Rs1215512
	129	BMI	1.77	Rs338491
2	84	Type 2 diabetes	2.28	Rs1919481
	108	Type 2 diabetes	2.98	Rs981525
	115	AOD	1.66	Rs1567804
	127	Type 2 diabetes	4.53	Rs925848
3	27	BMI	1.92	Rs2251166
	150	BMI	1.54	Rs1479137
4	135	Type 2 diabetes	2.26	Rs426029
6	118	Type 2 diabetes	1.82	Rs2032558
7	79	Type 2 diabetes	2.93	Rs956523
	135	BMI	1.84	Rs1862083
11	123	Type 2 diabetes	2.36	Rs1073636
12	27	Type 2 diabetes	2.14	Rs732868
13	4	Type 2 diabetes	2.33	Rs765244
	16	Type 2 diabetes	2.42	Rs306395
	22	AOD	2.45	Rs1572881/rs2491222
16	35	AOD	1.77	Rs30222
	56	Type 2 diabetes	2.43	Rs11901
18	66	AOD	2.96	Rs920783
	116	BMI	2.45	Rs1944566
20	62	Type 2 diabetes	1.77	Rs1406966
21	14	BMI	2.14	Rs6517799
22	31/41	Type 2 diabetes	1.68	Rs737622/rs715550

As expected, NPL analysis of type 2 diabetes gave overall smaller peaks and incomplete overlap with variance component scores. On chromosome 7 (27 cM; 1.73), type 2 diabetes linkage corresponded to a region identified in a preliminary report in African American Gullah-speaking families (18). Other regions on chromosome 13 (20 cM; LOD 1.63), 16 (51 cM; 1.84), and 22 (rs815550; 1.19) were near variance component peaks (Fig. 1; supplementary Table S2, available in the online appendix).

We performed ordered subset analysis on type 2 diabetes NPL scores by ranking families (both from low to high and high to low) by mean age of type 2 diabetes diagnosis, BMI, waist, WHR, diabetes duration, and admixture. Locations with scores that increased significantly are shown in Table 4, with the most prominent regions in a subset of at least 100 families on chromosomes 1 (30 cM; admixture), 3 (11 cM; age of type 2 diabetes diagnosis), 7 (131 cM; WHR), and 14 (57 cM; duration of diabetes). Ranking families by mean BMI produced no significant result. Only the region on chromosome 1 overlapped with a variance component linkage signal (Table 4).

TABLE 4

OSA of type 2 diabetes

Chromosome	Covariate	Peak	Maximum OSA score	Maximum unconditional score	Difference	P value	Families and ranking*
1	Admix.	30.00	4.11	2.33	1.79	0.0353	−117/414
3	AOD	10.96	3.88	2.17	1.71	0.0279	+175/404
4	AOD	109.31	3.93	1.80	2.13	0.0419	−20/404
7	WHR	131.47	4.29	2.72	1.57	0.0326	+211/371
11	AOD	104.00	4.10	2.58	1.53	0.0334	+18/404
14	Duration	56.59	4.05	2.64	1.41	0.0297	+222/351
17	Waist	56.10	3.95	1.43	2.52	0.0101	−58/373
20	WHR	2.28	3.75	2.48	1.27	0.0372	−33/371

Data are Z scores; P values are based on 500 permutations. Best signals from ordered subset analyses are shown. Admix, Caucasian admixture; duration, duration of diabetes; waist, waist circumference.

Families ranked from lowest to highest mean covariate value, −; families ranked from highest to lowest mean value, +.

Admixture mapping provides an alternative to family-based linkage to identify genes in recently admixed populations (21,22). Although our SNP-based linkage map was limited for admixture mapping by the relatively small proportion of ancestrally informative markers, the large number of markers nonetheless made mapping with the ADMIXMAP program feasible. We tested autosomal markers at 0.05-cM intervals to avoid confounding by linkage disequilibrium. Among the resulting 4,486 markers, the strongest association was for SNP rs1565728 (chromosome 12, 90 cM; P = 0.0003). No other marker approached Bonferroni-corrected experiment-wide significance (supplemental data and supplementary Table S3, available in the online appendix), and no marker that achieved nominal uncorrected significance (P < 0.05) mapped to a linkage peak. Individual admixture estimates varied from 0.010 to 0.614 but with mean estimates by study site ranging from 0.069 (South Carolina Gullah population) to 0.201 (Colorado) (supplemental data and supplementary Table S4, available in the online appendix). Heterogeneity of admixture proportions across sites was highly significant (χ2 = 154.76, P < 5 × 10−11; 10 df). Admixture proportion significantly affected both type 2 diabetes risk (P = 0.01) and age of type 2 diabetes diagnosis (P = 0.003) but had no effect on BMI (P = 0.95).

DISCUSSION

Considerable progress has been made in the identification of type 2 diabetes susceptibility genes for individuals of European ancestry. Nonetheless, only a small portion of the genetic risk has been explained, and of the 18 reported susceptibility genes (4), only TCF7L2 shows any evidence for a role in African-derived populations (5). Thus, most of the susceptibility genes in high-risk ethnic groups remain to be identified. Genome-wide association studies in populations of African descent may identify a subset of these genes, but a much larger number of SNPs will be required to overcome lower levels of linkage disequilibrium, with accompanying difficulties in separating real signals from noise in populations that are readily available. Family-based linkage studies, particularly SNP-based methods with higher information content, thus remain an important alternative. Based on European data, linkage peaks are unlikely to represent single common variants of large effect sizes. Instead, these peaks likely represent multiple rare variants of large effect, copy number polymorphisms, or multiple common variants that together provide a linkage signal. Only the latter model is amenable to genome-wide association studies, but even for such variants, the additional information of linkage studies may be essential to select significant associations from the large number of nominally associated SNPs.

Previous studies of populations of African descent have been limited. As of this writing, no genome-wide association studies of African-based populations have been reported, and most European susceptibility loci do not appear to be important in African American populations (5). Among previous linkage studies, Sale et al. (8) used a microsatellite map to analyze 675 individuals in 247 families ascertained primarily for diabetic renal disease. The strongest signal from that study (chromosome 6q, 163 cM) is considerably telomeric to our strongest chromosome 6 peak (LOD 1.82 at 118 cM), whereas the second best signal from that study (chromosome 22, 32 cM, LOD 1.33) is in the same location as a minor peak in our current study (33 cM, LOD 1.68). Our strongest signals do not overlap with those found by Sale et al., however. A microsatellite scan in West African families (9) reported signals on chromosome 20 (LOD 1.8; 60–78 cM), which overlapped with a minor peak of similar size in our study at rs1406966 (LOD 1.77; 62 cM). A second minor peak on chromosome 4 from the West Africa study (LOD 1.37, 125 cM) also overlapped with a peak on chromosome 4 (135 cM; LOD 2.26) in our study. The African American families in the report of Ehm et al. (7) are entirely contained in the present multicenter study, and thus, any overlap in linkage signals would not be independent. More recently, Sale et al. (18) analyzed 471 individuals from the Gullah-speaking African American population of South Carolina for the same SNP markers as the present study. Chromosome 7 linkage to type 2 diabetes from that study overlaps with both NPL and variance component peaks near 27, 40, and 78 cM reported here (18) (M. Sale, personal communication). Because some of the GENNID samples included Gullah-speaking families from South Carolina, we repeated our chromosome 7 analysis excluding all samples from South Carolina. The maximum LOD score dropped from 2.93 to 2.03, consistent with a contribution to the total score from the South Carolina samples. Whereas Sale et al. (18) found chromosome 7 to be most strongly implicated in early-onset type 2 diabetes, ordered subset analyses in our sample with families ranked by age of type 2 diabetes diagnosis implicated no peak in this region.

In a recently published meta-analysis of 23 microsatellite linkage scans across ethnic groups that included both Chicago and Arkansas African American GENNID families (23), five regions of linkage overlapped with the current study: chromosomes 4 (135 cM), 6 (118 cM), 16 (56 cM), 20 (62 cM), and 22 (40 cM). Despite the partial overlap in samples, the only possible overlap in linkage signals between our study and the African American subgroup from the meta-analysis was on chromosome 16 (33–66 cM). Other African/African American peaks from the meta-analysis (chromosomes 4, 6, 10, and 14) were not observed in the present study.

Simulation studies and comparisons of microsatellite and SNP-based linkage scans have suggested that scans based on 10,000 SNPs provide superior information to the standard 10-cM microsatellite scans for diseases including prostate cancer (24), rheumatoid arthritis (25,26), and bipolar disorder (27). Each of the earlier studies used the 10,000-SNP Affymetrix platform, whereas we used the 5,914-SNP Illumina platform. We compared the information content between the Illumina SNP mapping panel and the earlier CIDR microsatellite panel (8) using Arkansas families that had been typed for both marker sets. The information content for all chromosomes was increased from 50 to 60% for the microsatellite scans to >80% for the SNP scan (supplemental data and supplementary Fig. S1). This increased information content would be expected to generate a larger number of significant or suggestive linkage peaks, narrower and better defined peaks, and perhaps could explain the lack of overlap with earlier, generally smaller linkage scans in African American families.

The typing of 5,914 SNPs at relatively even intervals allowed us to also map by admixture. In contrast to linkage, admixture mapping depends on the presence of extended haplotypes deriving from parental populations to map a disease such as type 2 diabetes that is more prevalent in the African than European populations. This approach has more power than linkage or genome scan approaches to identify genes that derive primarily or entirely from either the African or European populations when the markers also show large frequency differences between populations. Admixture mapping would not be a powerful method to identify susceptibility loci that are common and of similar frequency in both ethnic groups (22), and power would be reduced if marker frequencies are similar between parental populations. Admixture mapping has recently been used successfully in mapping hypertension susceptibility genes (28,29). The marker set in the current study was not specifically designed for admixture mapping, but the amount of data generated was sufficient to easily identify admixture. We found a strong negative correlation between European admixture and both type 2 diabetes risk and age of type 2 diabetes diagnosis. Furthermore, our data easily demonstrated significant heterogeneity in the amount of admixture across centers. Nonetheless, <10% of the markers showed a minor allele frequency difference between Yoruban and Caucasian HapMap samples exceeding 0.5. Thus, a specifically designed admixture map might generate different and/or more significant signals.

Our strongest signal was at rs1565728 on chromosome 12 (90 cM). When we examined different combinations of markers in this region, we obtained P values from 0.00003 to 0.03, suggesting that no single marker was responsible for the signal. The only gene in the immediate vicinity of this marker is the transcription factor, E2F7, which is involved in cell cycle regulation but is expressed at low levels in most tissues (30). However, admixture and linkage mapping peaks did not overlap, and no admixture signal achieved genome-wide significance (supplementary Table S2). Chromosome 1 marker rs1007460, the second strongest admixture signal (supplementary Table S2), was close to a well-replicated linkage signal on chromosome 1q21-q24 (31). These findings will require follow-up with markers specifically selected for large differences in allele frequency between ancestral populations.

We tested for family-based association with type 2 diabetes, BMI, and age of type 2 diabetes diagnosis, although recognizing the limitations of the relatively low-density SNP linkage panel. No SNP attained Bonferroni-corrected significance (supplementary Table S5, available in the online appendix); but rs741923 on chromosome 19 (P = 0.0001) was suggestive for association with type 2 diabetes; and SNPs rs1459085 (chromosome 5; P < 0.0002), rs729958 (chromosome 9, P = 0.00005), and rs1941487 (chromosome 18, P = 0.0002) were suggestive for association with diabetes age of type 2 diabetes diagnosis. SNPs on chromosomes 2 (rs838715, P = 0.0001) and 8 (rs1483457, P = 0.0001) showed suggestive associations with BMI. SNPs rs1459085 (chromosome 5) and rs729958 (chromosome 9) were associated with both type 2 diabetes and age of type 2 diabetes diagnosis. Although association and linkage signals generally did not overlap, rs1051783 (type 2 diabetes, P < 0.002) was within the chromosome 2q linkage peak and near gene MFSD9, which is not an obvious candidate. Several SNPs on chromosome 6q21 (113–118 cM) were modestly associated with type 2 diabetes (P < 0.002) in a region of modest linkage (LOD 1.54). Among the known genes in this region are peptidylprolyl isomerase like 6 and zinc finger protein ZBTB24; neither is an obvious diabetes candidate gene. SNP rs2491222 (P < 0.0005 with type 2 diabetes) was under the chromosome 13 linkage peak for type 2 diabetes and age of type 2 diabetes diagnosis and within FLT3, which encodes a tyrosine kinase with known roles in hematopoiesis and leukemias. SNP rs11901 (chromosome 16) showed linkage to type 2 diabetes (LOD 1.76) and association with type 2 diabetes and age of type 2 diabetes diagnosis (P < 0.002) and is upstream of the serine/threonine kinase, TAOK2, which acts in the Jun NH2-terminal kinase mitogen-activated protein (MAP) kinase pathway by activating upstream MAP kinases (MKK) MKK3 and MKK6.

Three peaks were most prominent in the linkage analysis (Fig. 1): chromosome 2 extending from 41 to 121 Mb and encompassing 354 genes (498 unique transcripts), chromosome 13 from 18 to 31 Mb and encompassing 73 genes (89 unique transcripts), and chromosome 18 from 27 to 64 Mb and encompassing 130 genes (178 unique transcripts). Within these 557 genes are many strong candidates for type 2 diabetes. On chromosome 2, candidates include hexokinase 2, transcription factor TCF7L1, vesicular transport proteins VAMP5 and VAMP8, liver fatty acid-binding protein FABP1, dual-specificity phosphatase DUSP2, diabetes-related ankyrin repeat protein ANKRD23, and potential insulin signaling molecules MAP4K4 and INPP4A, as well as INSIG2 and STEAP3. On chromosome 13, candidates include CDK8, IPF1, and TNF superfamily member TNFRSF19. The chromosome 18 peak included another TNF superfamily member, TNFRSF11a, as well as MC4R, TCF4, phosphoinositide-3-kinase, class 3, and phosphatidylinositol glycan anchor biosynthesis. Only IPF1 was previously evaluated in African Americans (32,33) but was not found to be associated with type 2 diabetes.

In addition to the limitations noted above, our analysis of age of diagnosis was based on historical data. Even when obtained by standardized questionnaire, historical age of type 2 diabetes diagnosis is inexact and likely not an accurate reflection of the true age of onset. Nonetheless, this trait showed 60% heritability with similar values across centers and good correspondence with diabetes as a dichotomous trait.

In summary, using a large multicenter collection of nuclear families, a SNP-based linkage panel, and analytical methods based on variance component, admixture mapping, and ordered subsets, we have identified multiple regions of interest. Phenotypes of type 2 diabetes and diabetes age of diagnosis frequently mapped to the same locations, lending additional confidence to these results. Based on current understandings of complex disease pathogenesis, the majority of these regions probably represent collections of causative variants, rather than single SNPs or genes that can explain the linkage signal. Dense maps suitable for narrowing linkage signals, further exploring admixture mapping, and testing family-based association using maps that tag all variants in known genes will be logical steps. With most genetic risk factors yet to be discovered, these studies will undoubtedly complement other genome-wide association efforts in identifying those genes that contribute to diabetes in African Americans and other populations.