Linkage Analysis of a Composite Factor for the Multiple Metabolic Syndrome: The National Heart, Lung, and Blood Institute Family Heart Study

The NHLBI FHS is a multicenter population-based family study aimed at investigating the genetic and nongenetic determinants of CHD, preclinical atherosclerosis, and cardiovascular disease risk factors (18). Probands for the NHLBI FHS included a random sample of 2,000 participants and a second sample of 2,000 with a family history of CHD. Both samples were recruited through the Framingham Heart Study (Framingham, MA), the Utah Health Family Tree Study (Salt Lake City, UT), and the Atherosclerosis Risk in Communities Study (Minneapolis, MN, and Forsyth County, NC). Family history of CHD was based on information collected in the parent studies, and family risk scores for each individual were represented by the ratio of the observed numbers of CHD events within a family to the numbers expected according to age- and sex-specific incidence rates derived from the Framingham Study. Probands’ family members, including siblings, parents, spouses, and children >25 years of age, were invited to participate. A total of 5,975 individuals from 588 randomly recruited families and 656 families with elevated family risk scores participated.

The clinical examination of each study participant was performed according to a standardized protocol to measure anthropometry, blood pressure, lipids, lipoproteins, fibrinolytic activities, insulin, glucose, and blood chemistries. Participants were asked to fast for 12 h before the clinical examination. Blood was drawn in all participants for the laboratory tests and stored for genotyping and additional phenotyping.

All blood assays were performed in a laboratory at Fairview-University Medical Center at the University of Minnesota. Total cholesterol and triglyceride concentrations were measured with a Roche Cobas Fara high-speed centrifugal analyzer (Roche Diagnostic System) (19). HDL cholesterol was measured after precipitation of other lipoprotein fractions by dextran sulfate (20). PAI-1 antigen was measured using an enzyme-linked immunosorbent assay (21), and fibrinogen was measured by the Clauss method (22). Serum uric acid was measured by using the Ortho Clinical Diagnostics (Rochester, NY) Vitros thinfilm clinical analyzer method (23). An enzymatic (glucose oxidase) method (Kodak Ektachem 700 Analyzer; Kodak, Rochester, NY) was used to measure serum glucose, and a radioimmunoassay technique (Coat-A-Count; Diagnostic Products, Los Angeles, CA) was used to measure serum insulin. Insulin resistance, as evaluated by the HOMA index, was defined as (fasting insulin [μU/ml] × fasting glucose [mmol/l])/22.5 (24).

Anthropometric data, including BMI (kg/m²), WHR, and subscapular skinfold, were collected with subjects wearing scrub suits. Subscapular skinfold was measured twice with the “Lange” caliper, and the mean of the two measurements was used in our analysis. Hypertension was defined as SBP ≥140 mmHg, diastolic blood pressure (DBP) ≥90 mmHg, or treatment for hypertension. Diabetes was defined by the American Diabetes Association criteria (fasting glucose ≥126 mg/dl or use of hypoglycemic medication).

Detailed information on NHLBI FHS genotype data and quality control has been previously described (18,25). In brief, genotyping was conducted on two complementary samples of NHLBI FHS with different sets of markers. Only those individuals who had complete phenotype data in the final factor analysis model were included in our genetic analyses. The first sample (Marshfield sample) consisted of 2,467 white individuals distributed across 387 of the largest three-generation pedigrees. In this sample, approximately half of the pedigrees were recruited at random from the parent cohort populations, and the other half were ascertained because of large pedigree size and high familial risk and occurrence of CHD. These families were genotyped by the NHLBI Mammalian Genotyping Service, located in Marshfield, Wisconsin. The CHLC (Cooperative Human Linkage Center) Screening Set 10 was used for the genome scan and consists of 402 microsatellite (autosomal) markers spaced about every 9 cM. The average marker heterozygosity was 0.76. The second sample consisted of 1,082 white participants from 256 sibships (1,365 sibpairs). This sample contained all validated sibpairs with CHD and sibships with at least one sibling who had a high individual risk score for CHD (18). A total of 243 markers with an average marker distance of ∼18 cM was developed and typed by the Utah Molecular Genetics Laboratory (Salt Lake City, UT). A combined marker sample was formed by pooling the two samples. The combined group consisted of 2,831 individuals from 486 pedigrees, including 718 subjects genotyped for both marker sets. Family size of the combined group ranged from 2 to 14, with a median of 4 per family. For the combined marker set, the genetic map locations were based on the Marshfield map, with novel markers from the Utah collection incorporated by interpolation of sites between anchor markers present on both maps (25). Marker consistency between family members was tested, and a single individual or the entire family was set to “missing” for inconsistent markers. Due to overlapping of 718 individuals in the two individual samples, the subsequent factor and linkage analyses were conducted in the combined group unless otherwise mentioned.

A factor analysis was applied to the following MMS-related variables: BMI, WHR, subscapular skinfold, triglycerides, HDL, HOMA, PAI-1, uric acid, DBP, SBP, fibrinogen, and LDL and total cholesterol. Three variables, triglycerides, HOMA, and PAI-1, were log transformed to remove skewness of the distributions. In the factor model (26), each individual phenotype is expressed as a linear function of a set of latent factors and an error term. The relationship of the latent factors to the phenotypes is reflected by factor loadings, the regression coefficients of the phenotypes on the latent factors. A factor loading of 0.40 with a variable, corresponding to 15% of variance explained by a factor, was commonly used to judge meaningful correlation between the factor and the variable (26). The contribution of each factor to the set of variables is evaluated by eigenvalues (i.e., eigenvalues ≥1.0), defined by the sums of the squared factor loadings (26). Using PROC FACTOR procedure implemented in SAS (SAS version 6.12; SAS, Cary, NC), factor patterns were obtained by maximum-likelihood methods.

For linkage analysis, the major factor that accounted for the largest amount of the covariation among the variables formed the basis of the phenotype. To increase the relative contribution of the major factor and reduce noise from variables with small loadings on this factor, variables with factor loadings <0.4 were removed from the model in a stepwise manner until all variables correlated with the factor ≥0.4. Finally, a factor score, the estimated value of the underlying factor, was calculated for each individual based on the final factor pattern and the values of observed phenotypes. The factor score was adjusted for the effects of age, age squared, and field center in sex-specific models, and residuals from the regression models were standardized to zero mean and unit variance for linkage analysis. Individual variables contributing to the factor underwent the same adjustment procedure. Twelve subjects whose BMI, WHR, or HOMA residuals were outliers (5 SDs from the sex-specific means) were removed from the entire analysis. The final sex-specific residual phenotypes were approximately normally distributed (both skewness and kurtosis <0.6).

A multipoint variance components approach implemented in GENEHUNTER version 2.1 (27) was used to examine evidence for linkage of the composite factor. The full probability distribution of allele sharing at genotyped loci and at five evenly spaced points between adjacent markers was estimated with an exact multipoint algorithm (27). The mean trait value, variance components due to additive genetic contributions of a major quantitative trait locus (QTL), residual additive genetic effects, and environmental effects were estimated by a maximum likelihood method. Significance of a genetic contribution of the QTL was tested by comparing the maximum likelihood of the full model with that of a reduced model that constrained the QTL component to be zero. Twice the log_e likelihood ratio of the two models is asymptotically distributed as a one-half/one-half mixture of a χ² variable and a point mass at 0 (27). The difference between the two log₁₀ likelihoods produced an LOD score. Allele frequencies for the Marshfield marker set were estimated based on the marker allele frequencies of founders, and those for the Utah marker set were based on 200 randomly selected, unrelated subjects. Linkage plots for the MMS factor were compared with those of individual phenotypes at chromosomal regions where strong evidence for linkage (LOD >3.0) was detected for the composite factor.

We simulated Mendelian transmission of a completely informative marker in every pedigree under the null hypothesis of no linkage with the quantitative traits under investigation (17). The data for the MMS factor, individual traits included in the final factor analysis model, and the simulated marker were then submitted to a variance components QTL linkage analysis in MERLIN (28), and LOD scores were retained. (We find that MERLIN gives the same LOD scores as GENEHUNTER but that MERLIN is faster.) This process of simulation and linkage analysis was repeated 9,999 times. We then computed empirical single-point P values using the method recommended by North et al. (29).

Table 1 shows phenotypic characteristics of participants stratified by genotyping subset and sex. Table 2 presents the factor-loading pattern from the initial factor modeling of all 13 variables. Four factors were retained with the eigenvalues ≥1.0 criterion. The first factor explained ∼47% of the common variance and 25% of the total variance. This factor correlated at least moderately (factor loading ≥0.4) with anthropometric (BMI, WHR, and subscapular skinfold) and biochemical (triglycerides, HDL, HOMA, PAI-1, and uric acid) measures and captured many features of the metabolic syndrome. Therefore, this factor was labeled an “MMS” factor. Table 3 presents factor loadings on the MMS factor (factor 1) in the initial model and five reduced models in which minor variables were dropped in a stepwise fashion according to factor loadings <0.4. Exclusion of total cholesterol resulted in a two-factor model, with the proportion of common variance explained by the MMS factor substantially increased (model 2). Further exclusion of LDL, fibrinogen, DBP, and SBP did not result in appreciable changes in factor loading or the proportion of common variance explained by the MMS factor, although the relative contribution of the factor to the total variance increased steadily. In the final model, the MMS factor comprised BMI, WHR, subscapular skinfold, triglycerides, HDL, HOMA, PAI-1, and uric acid, accounting for 81% of the common variance and 41% of the total variance.

Heritability of the MMS factor was 0.41 ± 0.03 (P < 0.0001) after accounting for the effects of age, age squared, sex, and field center. Heritabilities for the individual phenotypes were as follows: h² = 0.44 ± 0.04 for BMI, 0.34 ± 0.03 for WHR, 0.43 ± 0.03 for subscapular skinfold, 0.45 ± 0.03 for triglycerides, 0.50 ± 0.03 for HDL, 0.35 ± 0.03 for HOMA, 0.28 ± 0.03 for PAI-1, and 0.38 ± 0.04 for uric acid.

Table 4 presents locations from the multipoint genome scan for the MMS factor with at least suggestive signals (LOD scores >1.9) in the combined sample. The strongest evidence for linkage was detected on chromosome 2. The maximum LOD score was 3.34 (empirical P = 0.0004) at 240 cM between markers D2S427 (LOD 2.84, empirical P = 0.0007) and D2S1279 (LOD 3.29, empirical P = 0.0004). The 1-LOD unit support interval extends from 223 to 246 cM (Fig. 1). The second highest LOD (2.86, empirical P = 0.0006) was observed at marker D12S1052 (83 cM) on chromosome 12. In addition, suggestive evidence for linkage (LOD >1.9) was observed at a region at 135 cM, marker D14S606, and a region at 26 cM on chromosomes 7, 14, and 15, respectively (Table 4).

Multipoint linkage results for the MMS factor on chromosome 2 were compared with those of the eight individual traits at their peak LOD locations (Fig. 2). The peak LODs of the eight traits, which were within the 1-LOD unit support interval of the MMS factor, are as follows: BMI (LOD 2.40 at 228 cM, empirical P = 0.0017), WHR (LOD 1.72 at 241 cM, empirical P = 0.0037), subscapular skinfold (LOD 2.55 at 240 cM, empirical P = 0.0012), triglycerides (LOD 1.35 at 227 cM, empirical P = 0.0089), HDL (LOD 1.68 at 241 cM, empirical P = 0.0041), HOMA (LOD 0.41 at 225 cM, empirical P = 0.0881), PAI-1 (LOD 0.95 at 241 cM, empirical P = 0.0211), and uric acid (LOD 1.88 at 228 cM, empirical P = 0.0021). Except for HOMA index, the peak LOD scores for the other seven traits were significant at P < 0.05 (without regard to multiple tests, since we were only testing the specific hypotheses at the QTL location on chromosome 2).

We eliminated 174 diabetic subjects and repeated the factor modeling and linkage analysis of the MMS factor on chromosomes 2, 7, 12, 14, and 15 in the combined sample. The factor-loading pattern and the peak LOD score on chromosome 7 did not change appreciably (data not presented). The peak LOD scores on chromosomes 2, 12, 14, and 15 were attenuated: LOD 2.07 at 240 cM on chromosome 2, LOD 1.30 at 83 cM on chromosome 12, LOD 1.93 at 92 cM on chromosome 14, and LOD 1.74 at 26 cM on chromosome 15.

Additional data can be found in two online appendixes available at http://diabetes.diabetesjournals.org.

We incorporated a factor analysis approach with a multipoint genome-wide scan to localize genetic loci influencing the underlying factor structure for MMS-related phenotypes in the NHLBI FHS. We identified a region on chromosome 2q36 that shows strong evidence for a major gene contributing to the MMS factor. The fact that each of the eight individual phenotypes mapped to the same region and that peak LOD scores for seven traits were significant at P < 0.05 further strengthens our findings.

The factor analysis result obtained in our study agrees with other reports using similar methods to investigate MMS-related traits. The major factor accounting for the most variation in the data was strongly loaded by insulin and obesity variables (3,5,7–12), as well as by glucose (5,10,12) and dyslipidemia (3,8,10–12). For example, using principal component factor analysis, Meigs et al. (8) reported a central metabolic syndrome factor representing covariation among fasting and postchallenge insulin, BMI, WHR, HDL, and triglycerides in the Framingham Offspring Study. Importantly, a recently published study showed that using directly measured insulin sensitivity versus HOMA index yielded similar factor-loading patterns (12). Also consistent with our observations are studies revealing that uric acid (3) and PAI-1 (5), but not fibrinogen (5), clustered with the factor exhibiting strong loading for obesity and insulin variables.

Our highest LOD score, located on chromosome 2, is in close proximity to regions linked to various obesity traits, HDL, and type 2 diabetes across studies, as summarized in Table 5. For example, Hager et al. (30) and Deng et al. (31) observed modest linkage for a QTL influencing obesity (BMI >27 kg/m²) and percent fat mass at marker D2S206, which is located at the same chromosome position as our peak LOD for the MMS factor. In a study of 27 Mexican-American families, a region about 4 cM centromeric to our peak was reportedly linked to HDL with an LOD of 1.3 (32). As for type 2 diabetes, Elbein et al. (33) obtained an LOD of 2.18 at maker D2S336. This marker maps to 245 cM on the Marshfield map. Additionally notable are findings from the mouse that indicate that regions syntenic to human 2q31-q37 contain QTLs for various obesity phenotypes (34).

Nevertheless, two studies have reported different pleiotropic loci for insulin resistance syndrome–related traits than our study (35,36). Kissebah et al. (35) reported a QTL on chromosome 3q27 that was linked to BMI, waist and hip circumferences, weight, insulin, and insulin/glucose in 507 nuclear Caucasian families. The second study (36) found strong linkage between two loci on chromosome 6 and measures of insulin, obesity, and lipids in nondiabetic subjects from 27 Mexican-American families. This discrepancy may be due to inclusion of “nontraditional” insulin resistance traits such as PAI-1 and uric acid in our study, which may pinpoint different pathways. In addition, other factors that differ among the studies may also explain the discrepancy, including population characteristics, genetic and environmental modifiers, sample size, and analytic approach.

As for BMI as an individual phenotype, the location of our peak on chromosome 2 (LOD 2.4 at 228 cM) is slightly different from that reported by Feitosa et al. (25) for a QTL influencing BMI in the NHLBI FHS population (peak LOD 1.5 at 241 cM). The discrepancy may be caused by differences in sample size and the adjustment procedure.

It is of concern that factor scores obtained from analysis of the combined sample might be inaccurate if the factor structure differed substantially between the Marshfield and Utah samples. Misspecified factor loadings introduce errors in the estimates of factor scores and may inflate type I and/or type II errors in linkage analysis. To evaluate heterogeneity of the factor-loading pattern and its effect on linkage analysis, we conducted separate factor analysis for the Marshfield and Utah samples and performed linkage analysis on chromosome 2 for each sample separately. The factor-loading pattern of the two individual samples was comparable to that of the combined sample (data not presented). In particular, for the MMS factor, factor loadings differed by <10% for four variables (BMI, subscapular skinfold, HOMA, and PAI-1), 10–20% for two variables (WHR and HDL), and 20–25% for two variables (triglycerides and uric acid). Furthermore, when the MMS factor was derived based on sample-specific factor analysis models, it peaked with LODs of 3.29 and 2.29 at 233 cM and 241 cM for the Marshfield and Utah samples, respectively, compared with LODs of 3.46 and 2.26 at the corresponding similar locations when the samples were combined to derive factor scores. Findings from these analyses suggest the linkage signal obtained on chromosome 2 was not an artifact that could have occurred from using a unifying factor analysis model.

To evaluate heterogeneity of linkage signals by sampling scheme, we conducted separate linkage analysis in the Marshfield and Utah samples for chromosome 2. The peak LOD was 3.46 at 233 cM for the Marshfield sample and 2.26 at 241 cM for the Utah sample, as compared with the LOD of 3.34 at 240 cM from the combined analysis. It is important to note that there were 718 individuals present in both samples. Therefore, similar linkage findings from the two samples could be due to the presence of overlapping individuals and, therefore, should not be judged as an independent replication.

The approach of deriving composite factors for genome scans has been used by Arya et al. (17) to map genetic loci contributing to the underlying factors for MMS-related traits in nondiabetic Mexican-American families. In this study, strong evidence for linkage was detected at two regions on chromosome 6 for a factor loaded by BMI, leptin, and fasting specific insulin and a region on chromosome 7 for a lipid factor (HDL and triglycerides). In our study, the location on chromosome 7 with suggestive evidence for linkage (LOD 2.42 at 135 cM) is ∼12 cM telomeric to the locus linked to the lipid factor by Arya et al. This locus has been previously reported for linkage to BMI in the NHLBI FHS population (25) and contained the leptin gene, an obvious candidate gene for obesity.

Our search for genes in the peak LOD region on chromosome 2 flanked by markers D2S360 (223 cM) and D2S336 (245 cM) revealed several possible candidates that may be influential in determining the multivariate correlation among MMS-related traits. These genes include thyroid hormone receptor interactor 12 (TRIP12, 2q36.1), 5-hydroxytryptamine (5-HT, serotonin) receptor 2B (HTR2B, 2q36.3-q37.1), and insulin receptor substrate 1 (IRS1, 2q36). Thyroid hormones, mediating via thyroid hormone receptors, exert pleiotropic effects on many physiological aspects, including metabolism of lipids, carbohydrates, and proteins (37). TRIP12 is one of thyroid hormone receptor–interacting proteins and has been shown to be highly expressed in human skeletal muscle and testis (38). The second candidate, HTR2B, is one of several mediators of the neurotransmitter, serotonin. Serotonin plays an important role in the regulation of appetite. In rats, stimulation of the HTR2B receptors leads to hyperphagia or hypophagia depending on the level of basal feeding (39). Polymorphisms in the genes encoding 5-HT 2A and 2C receptors have been associated with obesity and type 2 diabetes (40,41).

Our findings suggest the existence of a common QTL contributing to the variation and covariation of the MMS-related traits. It is possible that pleiotropic genes regulate the traits simultaneously through independent and parallel pathways or act through primary phenotype(s), such as obesity variables (2), insulin resistance (3), or diabetes, to mediate the correlation in these traits. The obesity hypothesis fits our data because the obesity phenotypes (BMI and subscapular skinfold) show stronger genetic signals than other traits and because all the traits are associated with obesity. The weak linkage for insulin resistance may reflect inaccuracy in measuring insulin resistance with the HOMA index. Diabetic subjects were included in the study because diabetes represents an advanced stage of insulin resistance. To evaluate the influence of diabetes on our linkage findings, 174 individuals with diabetes were excluded and the factor modeling and linkage analysis were repeated on the five chromosomes with at least suggestive linkage. Although the linkage for chromosome 2, 12, 14, and 15 were attenuated, there still was modest evidence for linkage among nondiabetic subjects. The relatively large reduction in the LOD score after removing diabetic subjects suggests that the locus that was identified in the study for metabolic syndrome or insulin resistance is related to diabetes susceptibility or that the diabetic individuals contribute a substantial amount to the linkage signal for the MMS factor. This later interpretation is supported by the observation that the diabetic individuals had higher values for the MMS factor, BMI, WHR, subscapular skinfold, triglycerides, HOMA index, and PAI-1, as well as lower HDL, compared with nondiabetic subjects (data not presented). It is of concern that the degree of glycemic control by medication or other types of intervention affects insulin resistance traits and, therefore, may decrease the power to detect insulin resistance loci compared with an untreated population. Among the 174 diabetic subjects in our study, 122 (70.1%) were taking diabetes medication at the time of study. Information on other types of intervention for glycemic control was not collected in this study. The diabetic subjects who were on medication had higher values for the MMS factor, WHR, and HOMA index than those who were not on medication (data not presented). Other traits did not differ significantly between the two groups. Taken together, these data suggest that the degree of glycemic control in the 174 diabetic subjects was not substantial enough to bring mean levels of these traits close to those of the nondiabetic subjects.

In summary, we have found strong evidence for the presence of a genetic locus on chromosome 2 linked to the MMS factor comprising BMI, WHR, subscapular skinfold, triglycerides, HDL, HOMA, PAI-1, and uric acid. In addition, several other regions were detected with suggestive evidence that may also contribute to the underlying correlation structure of these traits.

This study was supported in part by NHLBI cooperative agreement grants U01 HL56563, U01 HL56564, U01 HL56565, U01HL56566, U01 HL56567, U01 HL56568, and U01 HL56569. W.T. is supported by NHLBI training grant 1T32-HL07972-01.

We wish to thank the University of Minnesota Supercomputing Institute for use of the IBM SP supercomputer.