Search

Family-based study designs are again becoming popular as new next-generation sequencing technologies make whole-exome and whole-genome sequencing projects economically and temporally feasible. Here we evaluate the statistical properties of linkage analyses and family-based tests of association for the Genetic Analysis Workshop 17 mini-exome sequence data. Based on our results, the linkage methods using relative pairs or nuclear families had low power, with the best results coming from variance components linkage analysis in nuclear families and Elston- Stewart model-based linkage analysis in extended pedigrees. For family-based tests of association, both ASSOC and ROMP performed well for genes with large effects, but ROMP had the advantage of not requiring parental genotypes in the analysis. For the linkage analyses we conclude that genome-wide significance levels appear to control type I error well but that "suggestive" significance levels do not. Methods that make use of the extended pedigrees are well powered to detect major loci segregating in the families even when there is substantial genetic heterogeneity and the trait is mainly polygenic. However, large numbers of such pedigrees will be necessary to detect all major loci. The family-based tests of association found the same major loci as the linkage analyses and detected low-frequency loci with moderate effect sizes, but control of type I error was not as stringent. [ABSTRACT FROM AUTHOR]

Random forests (RF) is one of a broad class of machine learning methods that are able to deal with large-scale data without model specification, which makes it an attractive method for genome-wide association studies (GWAS). The performance of RF and other association methods in the presence of interactions was evaluated using the simulated data from Genetic Analysis Workshop 16 Problem 3, with knowledge of the major causative markers, risk factors, and their interactions in the simulated traits. There was good power to detect the environmental risk factors using RF, trend tests, or regression analyses but the power to detect the effects of the causal markers was poor for all methods. The causal marker that had an interactive effect with smoking did show moderate evidence of association in the RF and regression analyses, suggesting that RF may perform well at detecting such interactions in larger, more highly powered datasets. [ABSTRACT FROM AUTHOR]

Random forests (RF) is one of a broad class of machine learning methods that are able to deal with large-scale data without model specification, which makes it an attractive method for genome-wide association studies (GWAS). The performance of RF and other association methods in the presence of interactions was evaluated using the simulated data from Genetic Analysis Workshop 16 Problem 3, with knowledge of the major causative markers, risk factors, and their interactions in the simulated traits. There was good power to detect the environmental risk factors using RF, trend tests, or regression analyses but the power to detect the effects of the causal markers was poor for all methods. The causal marker that had an interactive effect with smoking did show moderate evidence of association in the RF and regression analyses, suggesting that RF may perform well at detecting such interactions in larger, more highly powered datasets.

Genome-wide linkage analysis using microsatellite markers has been successful in the identification of numerous Mendelian and complex disease loci. The recent availability of high-density single-nucleotide polymorphism (SNP) maps provides a potentially more powerful option. Using the simulated and Collaborative Study on the Genetics of Alcoholism (COGA) datasets from the Genetics Analysis Workshop 14 (GAW14), we examined how altering the density of SNP marker sets impacted the overall information content, the power to detect trait loci, and the number of false positive results. For the simulated data we used SNP maps with density of 0.3 cM, 1 cM, 2 cM, and 3 cM. For the COGA data we combined the marker sets from Illumina and Affymetrix to create a map with average density of 0.25 cM and then, using a sub-sample of these markers, created maps with density of 0.3 cM, 0.6 cM, 1 cM, 2 cM, and 3 cM. For each marker set, multipoint linkage analysis using MERLIN was performed for both dominant and recessive traits derived from marker loci. Our results showed that information content increased with increased map density. For the homogeneous, completely penetrant traits we created, there was only a modest difference in ability to detect trait loci. Additionally, as map density increased there was only a slight increase in the number of false positive results when there was linkage disequilibrium (LD) between markers. The presence of LD between markers may have led to an increased number of false positive regions but no clear relationship between regions of high LD and locations of false positive linkage signals was observed.