Both QTL and covariate-based linkage analyses produced substantial linkage signals from the BN cohort. QTL linkage analysis produced four suggestive signals, one when MENAR was the outcome and the remainder when ANX was the outcome. Covariate linkage analyses revealed a number of substantial linkage scores from the BN cohort: for minimum BMI, one significant and three suggestive; for CM, two significant and three suggestive; and for OBF, one significant and five suggestive. Analysis of the AN cohort was less fruitful. QTL linkage analysis provided two suggestive signals (for OBS and ANX) and five suggestive signals for covariate-based linkage analysis (1 for BMI, 2 for CM and 2 for OBF). The AN cohort is roughly one-half the size of the BN cohort, and this difference might account for the limited linkage signals.
Under the null hypothesis of no linkage and a single scan, we expect only one linkage score to exceed the suggestive threshold. The distribution of the number of regions in which the scores exceed the threshold can be modeled as a Poisson with mean one, and it is simple under this assumed distribution to calculate the probability of regional scores exceeding the threshold at least X times for a single scan. The probability of exceeding the threshold at least X = 3 times is ˜ 0.06, so the results for BMI, CM and OBF are striking for the BN cohort (–), even after accounting for multiple tests. Likewise multiple suggestive signals for both the AN and BN cohorts for CM and OBF are highly significant (p < 0.001 for either).
Disappointingly, however, all significant
signals from the BN cohort were dampened slightly or substantially when the BN and AN cohorts were combined. Diminished linkage signals could result from either statistical or biological causes, or both. In terms of statistical causes for the disparity, the simplest explanation is that the substantial linkage signals in the BN sample are false positives. A more subtle explanation appeals to the observation that large signals from linkage scans are expected to be highly biased, greatly exaggerating the true effect of the locus (or loci) generating the signal (Goring et al., 2001
). This phenomenon might account for some of what we observe. Nonetheless, based on statistical theory, for any liability locus in common in the two cohorts, one expects linkage signals in both data sets to produce positive evidence for linkage, even if the magnitudes of the signals are quite different. This expectation is not realized; instead the complementary sample tends to produce weak evidence against linkage when the other is strongly positive.
We suspect a better explanation for the disparity lies in biology: AN and BN cohorts differ at a fundamental biological and genetic level, and that these differences are not easily resolved by the quantitative traits/covariates we used in the analyses. Biologically, the trait might not have the same salience for AN and BN cohorts. BMI is an example. Individuals with consistent AN, almost by definition, display lower lifetime BMI than individuals with consistent BN. The ability to attain and maintain extremely low body weights is apparently a fundamental biologically- and genetically-based difference between individuals with AN and BN. Our genetic explanation hypothesizes that a locus that generates substantial liability to BN will often – but not always – fail to generate liability to AN. For example, by using QTL linkage analysis, significant linkage is obtained on 10p for MENAR in the BN cohort, but LOD = 0.01 for the AN cohort. As described in Bulik et al. (2003b)
, the same region of 10p, which produced significant linkage based on diagnosis, overlaps substantially with regions showing linkage for obesity (Hager et al., 1998
; Hinney et al., 2000
). Rates of obesity higher than that expected by chance are observed in families accessed through BN probands, but not in families accessed through AN probands. Thus, while speculative, the putative QTL on 10p could be genetically correlated with the disinhibition
characteristic of BN and obesity, but have no impact on the inhibition
characteristic of AN. It is interesting to note that early age-of-menarche has been associated on a population level with disinhibition (Johannson and Ritzen, 2005
) and with the development of binge-eating in the absence of compensatory behaviors independent of BMI (Reichborn-Kjennerud et al., 2004
). It might also be important to note that OBF produces a suggestive signal in the region from the BN cohort, but OBS does not.
The other two traits producing significant linkage in the BN sample were CM and OBF, both detected by covariate-based linkage analysis. Again the putative loci underlying these traits could pleiotropically impact liability to BN. It is also possible that different values of CM and OBF are associated with liability to AN versus BN (e.g., low OBF for AN and high OBF for BN). If so, this might be detectable by QTL linkage analysis. To investigate the latter possibility, we tested chromosomes 14 and 16 for QTL linkage to CM and OBF. We find supporting evidence for QTL at 14q22.2 for the BN cohort (CM: LOD = 2.70; OBF, LOD = 1.20), but not the AN cohort (CM, LOD = −.46; OBF, LOD = −.14). We also find some weak support for a QTL at 16p13.3 region for the BN cohort (CM: LOD = 1.39; OBF, LOD = 0.50), but not the AN cohort (CM, LOD = −.43; OBF, LOD = −.23). Our data, therefore, support the idea that loci near 14q21.1 and 16p13.3 affect liability to BN, but have little impact on liability to AN.
On the other hand, consistent with our genetic hypothesis that some loci do confer liability to both AN and BN, some linkage signals were amplified by combining the data sets (). For OBS, the LOD score at 1q31.1climbs from 1.55 for the AN cohort to 1.98 for the combined sample and, at 7p21.2, from 1.56 for the BN cohort to 1.79. Another notable increase occurs at 4q23 for MENAR, which achieves LOD = 2.01 for the combined sample from scores for the individual BN and AN cohorts of roughly 1.0. Other regions/traits showing similar changes include 5p15.33/BMI (LOD = 1.71), 10q21.3/OBF (LOD = 2.14) and 3q13.32/OBF (LOD = 1.84). See for more details.
Certain regions of the genome repeatedly show positive linkage signals for multiple traits with different samples (web-Table 1
). An obvious example is the 14q21 region, which showed significant linkages in the BN cohort for BMI, CM and OBF (–). Because of the pattern of linkage signals, we were curious if these covariates up-weighted/down-weighted the same families for linkage analysis (the probability of membership of the families in the linked group), even though the covariates were largely uncorrelated (see Figs. 3 and 4 in Bulik et al. in review
). Thus we evaluated the correlation of the weights, and found they were not highly correlated across these three traits. Maximum pairwise correlation was 0.21 and minimum was 0.10. This is the region that showed suggestive linkage in a previous analysis of diagnosis and IBD-sharing (Bulik et al., 2003b
). Clustering families by these covariates gives greater weight to families with higher IBD-sharing, without
information about IBD-sharing per se
A complicated region is 4p16.1 - 4p15.33. Within 4p16.1, suggestive and close to significant linkage occurs for OBF in the BN cohort; two other traits/analyses yield LOD scores greater than 1.0, for BMI (1.14; BN) and CM (1.49; BN/AN). At the adjacent chromosome band, 4p15.33, suggestive linkage occurs for CM in the BN/AN cohorts and LOD = 1.21 occurs for BMI in the AN cohort. Thus, ignoring the possibility of no liability locus in 4p16.1 - 4p15.33, the same locus could be generating all the signals. An alternative interpretation puts a locus in 4p16.1 affecting liability to BN and/or a locus in 4p15.33 affecting liability more generally. See web-Table 1
for more regions of overlap.
A few regions of overlapping linkage signals coincide with our previous research. In addition to 10p13 and 14q21.1, the 1q31 region again comes to the fore (, web-Table1
). In our first exploratory analysis using covariates, 1q31 showed significant linkage (Devlin et al., 2002a
; Bacanu, 2005
) when both OBS and drive-for-thinness were used as covariates. (Unfortunately, drive-for-thinness was not measured in the BN cohort, and no combination of traits measured in both samples accurately predicts drive-for-thinness in the AN cohort.) Our current analyses, using OBS in a QTL linkage analysis, produces a strong suggestive signal for linkage at 1q31.1 in the BN/AN cohorts. Complementing this linkage is another suggestive linkage at 1q31.1 for ANX in the BN cohort, which shifts to 1q25.1 (a difference of only 12 cM) and increases to suggestive (LOD = 2.0) when this sample is combined with the AN cohort.
Intriguingly, 11q22 shows some overlap of linkage signals for the BN and AN cohorts for OBS and BMI (, web-Table 1
). This region contains DRD2, polymorphisms which demonstrate linkage and association in our analyses (Bergen et al., in press
). The −141 C/- insertion/deletion (−141 Indel) polymorphism, which affects DRD2 transcription efficiency, shows significant association with diagnosis at the level of alleles, genotypes and haplotypes; the insertion C allele, which appears to increase expression of DRD2, is transmitted from parents to their affected offspring at rates significantly greater than that expected by chance; and haplotypes containing the insertion C allele and other SNP variants show even greater transmission distortion. Therefore the linkage results (, web-Table 1
) could be attributable to the impact of DRD2 polymorphisms.
After measuring 100 features relevant for eating disorders in multiplex families for eating disorders, we used a multilayer decision process to select six traits for linkage analysis and team the traits with an appropriate analytic method (Bulik et al., in review
). Insofar as we are aware, this is the first study to explore the phenotypic space in this way, and it could prove a useful blueprint for other studies of its kind, such as ongoing studies of the genetic basis of Type II diabetes and hypertension. When the results of the phenotypic analyses were applied to that genetic data from two cohorts of multiplex families, a number of linkage signals worthy of follow-up study arise. It is tempting to conclude that our approach to these complex data has been successful because we have identified a greater number of significant and suggestive linkages than that expected by chance. Nonetheless we are cognizant that proof of success will only come when alleles generating liability to eating disorders – or affecting the traits underlying liability to eating disorders – are convincingly identified under our linkage peaks. We have pursued two approaches to achieve this goal: bolstering our linkage results by linkage studies on new samples of multiplex families; and by direct identification of genetic variation generating liability to disease in these linkage regions. Looking at the results from both cohorts, two promising features stand out: regions of the genome which repeatedly show positive linkage signals for different traits (14q21.1, 4p16.1 - 4p15.33, 1q31, 11q22, 10p13) and different samples; and suggestive linkage signals that emerge by combination of the two samples (1q31.1, 7p21.2, 4q25, 5p15.33, 10q21.3, 3q13.32).