This meta-analysis of all 23 type 2 diabetes autosomal genome scans provides modest evidence for linkage on several chromosomes, notably 14 (~58 cM in the deCODE map), 10 (~138 cM), 4 (~76 cM and ~176 cM), and 16 (~44 cM). When we restricted our attention to geographically/racially defined subgroups in an effort to decrease genetic heterogeneity, we found evidence for several additional chromosomal regions, notably chromosomes 1 (~164 cM in Europeans/European Americans, and ~144 cM in Africans/African Americans), 2 (~246–258 cM in Mexican Americans), and 6 (~116 cM in Europeans/European Americans, and ~152 cM in Africans/African Americans). Many of these regions are only identified in one subgroup and so may reflect true etiologic differences by group or a combination of false positives and false negatives.
Two regions which have received considerable attention in the type 2 diabetes genetic linkage literature owing to repeated reports of genetic linkage are 1q [beginning with [5
]] and 20q [beginning with [4
]]. While neither of these regions was among the most interesting in this meta-analysis, both yielded some evidence for linkage: for chromosome 1 at ~166 cM, pTP
= 0.013, and for chromosome 20 at ~62 cM, pTP
= 0.071 and pLOD
= 0.039. In subgroup analyses on European/European American samples both regions are significant at the 0.01 level, but are not interesting in African/African American or Mexican American samples (all p > 0.10).
Several genome-wide association studies [46
] have recently been published on susceptibility loci for type 2 diabetes. Except the TCF7L2
gene which is identified by all the six studies and is one of the strongest signals in our meta-analysis, the other main signals found in these association studies are not close to the signals in meta-analysis. But there are intriguing overlaps between some modestly associated single nucleotide polymorphisms (SNPs) and our meta-analysis peaks. For example, our signal on chromosome 4 (~76 cM) is ~1 Mb from SNP rs282705 reported by Sladek et al. [46
] (p = 9.0 × 10–6
), and 5 Mb from SNP SNP_A-4299379 reported by Saxena et al. [48
] (p = 7.2 × 10–5
). Our signal on chromosome 14 (~58 cM) is ~8 Mb from SNPs rs1256517 and rs1256526 in Sladek et al. [46
] (p = 4.7 × 10–6
). Our signal on chromosome 16 (~44 cM) is ~1 Mb from a modestly associated SNP rs724466 in Steinthorsdottir et al. [50
] (p = 2.7 × 10–5
), and ~8 Mb from one of a modestly associated SNP rs10521095 in Scott et al. [47
] (p = 3.8 × 10–5
Demenais et al. [52
] conducted a meta-analysis of four type 2 diabetes linkage scans of European samples using the GSMA method with unweighted ranks. These four studies: University of Lund Botnia, University of Lund Sibs, United Kingdom, and Pasteur Institute Lille, are included in our analyses, although with slight differences in the samples. Demenais et al. found six regions on the genome to be nominally significant at 0.05 level. Their region on chromosome 16 (29–58 cM) directly overlaps one of our top four signals (chromosome 16 ~44 cM). Four more of their regions on chromosomes 1 (145–174 cM), 2 (58–87 cM), 6 (116–145 cM), and 17 (29–58 cM) are also at least nominally significant at the 0.05 level in our analyses (table ), and the chromosome 1, 16, and 17 regions also are significant at the 0.01 level in our subgroup analyses of 11 European/European American samples (table ).
While these results are of interest, it should be noted that the evidence for linkage at individual locations and overall is at best modest. For the ~30 cM bins, the smallest p values for the truncated p value, LOD score, and weighted rank methods were 0.0020, 0.0051, and 0.0138, respectively. None of these p values met the threshold for genome-wide significance using the Bonferroni correction. Indeed, among 115 tests based on the ~30 cM bins, 2 were significant at the 0.01 level and 3 at the 0.05 level for the truncated p value method; 1 and 6 for the analyses based on LOD scores, and 0 and 9 for those based on weighted ranks. 1.15 and 5.75 such results would be expected at the 0.01 and 0.05 levels in the absence of linkage information. Similarly, the minimum p value observed for 2 cM bins for the truncated p value based meta-analysis was 0.0017, while those for the LOD score and weighted rank based analyses were 0.0011 and 0.0020. Among 1,758 tests based on the 2 cM bins, 18 were significant at the 0.01 level and 94 at the 0.05 level for the truncated p value method; these numbers were 18 and 83 for the analyses based on LOD scores and 8 and 90 for those based on weighted ranks. 17.58 and 87.90 such results would be expected at the 0.01 and 0.05 levels even in the absence of linkage information. Again, no excess of small p values was observed.
The GSMA method proposed by Wise et al. [29
] provides an approach to combine the results of multiple linkage studies to identify susceptibility loci for the disease of interest. It suggests the use of ranks to unify various measures of linkage results and the use of bins (~30 cM) to obtain independence of test statistics. Our simulations have shown that GSMA results in appropriate type I error rates under the conditions of homogeneous marker sets and constant marker density (simulation 3), but results in modestly inflated type I error rates on average and some variability in these rates by location under more realistic settings (simulations 1 and 2). The alternative scoring schemes of maximum LOD scores and truncated p values provided similar but slightly better type I error control. The truncated p value method not only gives the best average type I error rate compared to the nominal value, but also the smallest standard deviations among the methods using the three scores (table ).
Ranks allow the advantage of easy conversion of different types of linkage scores (such as LOD scores, mean IBD sharing, or p values) into a uniform scale, but do not directly reflect the strength of the corresponding linkage signals. Even under the null hypothesis of no linkage anywhere in the genome, high ranks still must be assigned and often will be assigned to regions with little evidence for linkage. Since most linkage studies report statistics that are of (approximately) known asymptotic distribution, conversion to a common scale such as LOD scores or p values should usually be possible, as was the case for the 23 samples in the Consortium data.
Analysis based on LOD scores preserves the magnitude of the linkage results from each study. Analysis based on truncated p values emphasizes the gap between ‘significant’ (smaller than a cut-off value) and ‘insignificant’ (greater than a cut-off value) results by replacing the latter by the extreme value of 1. Both methods make use of more information than the ranks. In addition, since the bins falling into the same linkage peak are obviously correlated, the truncated p value based analysis narrows the width of the peak by truncating ‘insignificant’ results into the baseline and hence reduces the extent of correlation among the test statistics. This may explain the better control of the type I error rates using this method. Forabosco et al. [53
] showed that only using strongly significant results, such as LOD score >1, 2, or 3, will result in loss of power in GSMA. In our analyses, we use the more modest threshold value of 0.05 for the p values, which corresponds to a LOD score of 0.59. We also conducted the meta-analyses using a cut-off p value of 0.10 (results not shown), and the results are similar to those presented here with no additional interesting regions identified. Our analyses results from using the LOD scores (and so no p value truncation) and truncated p values are also highly correlated, with Pearson correlation of 0.88 between the logarithm of the p values.
We also carried out meta-analyses and simulations using ranks without weighting by sample size and minimum p values without truncation as measures of linkage signals (results not presented). Unweighted ranks are expected to perform less well than the other methods since they ignore differences in the amount of information contained in individual studies. (Non-truncated) minimum p values produced results highly correlated with those using the maximum LOD scores, with the Pearson correlations 0.96 and 0.92 in ~30 cM and 2 cM bin methods respectively, and identified three out of the four best regions; the region ~76 cM on chromosome 4 was not identified at 0.05 significance level by this method.
Besides considering alternative scoring schemes, we also considered narrower bin widths. Although ~30 cM bins can contain some linkage peaks and potentially reduce the correlation between adjacent bins, it is still difficult to develop a good binning approach to keep all linkage peaks from being split into two or more bins. Further, wide bins decrease the resolution of the linkage results, and may attenuate the distinction between weaker and stronger linkage signals. We therefore decided to use 2 cM bins as an approximation of continuous evaluation of the linkage statistic. An advantage of the narrow bins is that most of the bins contain zero or one markers in the linkage studies we collected. The effect of marker densities is likely minimized, consistent with the better control of type I errors using narrower bins in simulations 1 and 2.
Since LOD scores were not calculated at every 2 cM in every study, we interpolated the LOD scores linearly (see ‘Materials and Methods’). To evaluate the impact of interpolation, we calculated the LOD scores at every 10 cM instead, using the same simulated pedigrees and genotypes as those in our simulations, and interpolated the scores at every 2 cM. The interpolated scores are very close to the actual values and should not have changed the power of GSMA noticeably.
Because the test statistics (ranks, LOD scores, or truncated p values) are permuted within study to assess the significances, splitting a linkage peak into multiple bins, which we shall call correlated bins, causes dependence of bins and may bias the estimated p values from permutation. Using 2 cM bins obviously does increase the number of correlated bins. However, and more important, it does not change the proportion of correlated bins compared to that using ~30 cM bins. For example, we simulated 100 chromosomes of 150 cM under the null hypothesis, and counted the number of bins belonging to the same linkage peak; here a peak is defined as the contiguous bins with LOD >0. The proportion of correlated bins was 0.429 for 2 cM bins, and 0.508 for 30 cM bins. For the truncated p values with cut-off at 0.05, that is, LOD >0.59, the proportions dropped to 0.043 and 0.058 for 2 cM and 30 cM bins, respectively.
A more important issue with 2 cM bins is the problem of multiple testing. With more bins, a substantially more stringent threshold for p values will be required to declare genome-wide significance of any bin. On the other hand, the 2 cM bins do provide higher resolution for the interesting regions, which can be more easily compared with other linkage studies and genome-wide association studies. The 2 cM bin-based methods also control the type I error rates better than the ~30 cM-based methods, especially when there is substantial variability in marker density. Therefore, we suggest an initial screening using ~30 cM bins followed by using 2 cM bins to localize linkage signals and control for false positives.
To obtain more accurate evaluation of significance, we may need to know the multi-dimensional distributions of the test statistics under the same settings of our meta-analyses of the Consortium data. Ideally, if we could use the complete pedigree information in simulation 1 without the pedigree reduction described in ‘Materials and Methods’, it would provide an estimate for the desired distribution, given that the number of simulation replicates is sufficiently large. A similar approach was suggested by Wise [54
] to combine candidate region studies with genome-wide studies using the GSMA method. If we are willing to ignore the differences between the Consortium data and the reduced data for simulation, we can estimate the p value of each bin as the proportion of simulation replicates in which the score of the bin in simulation was greater than that in our meta-analyses. The same regions were then identified on chromosome 10 and 14 using either ~30 cM bins or 2 cM bins, and on chromosome 4 using 2 cM bins at the 0.01 significance level, with p values similar to those obtained by permutations.
When data are available, a joint linkage analysis may also be performed by pooling pedigree information across the same 23 individual studies. Its results should be similar to ours from the meta-analyses described here, but some differences can also be expected. The meta-analysis considers only positive linkage signals, while in the joint analysis, deficiencies of sharing in some studies may cancel excess sharing in other studies. In the meta-analysis, individual studies can apply various methods of analysis, including parametric/non-parametric and variance component linkage analyses, and might consider auxiliary phenotype data. Although this flexibility is appreciated when original genotype data cannot be accessed, it may add heterogeneity among the individual studies’ results to be summarized by the meta-analysis. On the other hand, while the joint analyses can make a better use of information, they are sometimes limited by computational powers needed for the large sample size of the pooled sample.
In summary, our meta-analyses of the International Type 2 Diabetes Linkage Analysis Consortium data suggested evidence for linkage on regions of chromosomes 4, 10, 14, and 16 for type 2 diabetes, with no signal reaching genome-wide significance, but those on chromosomes 10 and 14 reaching a level expected to occur by chance once per genome. Subgroup analyses are consistent with the possibility that genetic heterogeneity of the collected samples may be a cause for the modest linkage signals. Computer simulations showed that variations of the marker density within and between studies could result in modestly inflated or deflated type I errors in current analyses. Smaller bins (2 cM) and alternative test statistics more directly based on the linkage evidence, such as LOD scores or truncated p values, may help to draw the type I errors towards the nominal value, but more sophisticated approaches need to be considered to correct this problem fully.