Gene expression in large pedigrees: analytic approaches

Rita M. Cantor, Heather J. Cordell

We currently have the ability to quantify transcript abundance of messenger RNA (mRNA), genome-wide, using microarray technologies. Analyzing genotype, phenotype and expression data from 20 pedigrees, the members of our Genetic Analysis Workshop (GAW) 19 gene expression group published 9 papers, tackling some timely and important problems and questions. To study the complexity and interrelationships of genetics and gene expression, we used established statistical tools, developed newer statistical tools, and developed and applied extensions to these tools. To study gene expression correlations in the pedigree members (without incorporating genotype or trait data into the analysis), 2 papers used principal components analysis, weighted gene coexpression network analysis, meta-analyses, gene enrichment analyses, and linear mixed models. To explore the relationship between genetics and gene expression, 2 papers studied expression quantitative trait locus allelic heterogeneity through conditional association analyses, and epistasis through interaction analyses. A third paper assessed the feasibility of applying allele-specific binding to filter potential regulatory single-nucleotide polymorphisms (SNPs). Analytic approaches included linear mixed models based on measured genotypes in pedigrees, permutation tests, and covariance kernels. To incorporate both genotype and phenotype data with gene expression, 4 groups employed linear mixed models, nonparametric weighted U statistics, structural equation modeling, Bayesian unified frameworks, and multiple regression. Regarding the analysis of pedigree data, we found that gene expression is familial, indicating that at least 1 factor for pedigree membership or multiple factors for the degree of relationship should be included in analyses, and we developed a method to adjust for familiality prior to conducting weighted co-expression gene network analysis. For SNP association and conditional analyses, we found FaST-LMM (Factored Spectrally Transformed Linear Mixed Model) and SOLAR-MGA (Sequential Oligogenic Linkage Analysis Routines -Major Gene Analysis) have similar type 1 and type 2 errors and can be used almost interchangeably. To improve the power and precision of association tests, prior knowledge of DNase-I hypersensitivity sites or other relevant biological annotations can be incorporated into the analyses. On a biological level, eQTL (expression quantitative trait loci) are genetically complex, exhibiting both allelic heterogeneity and epistasis. Including both genotype and phenotype data together with measurements of gene expression was found to be generally advantageous in terms of generating improved levels of significance and in providing more interpretable biological models. Pedigrees can be used to conduct analyses of and enhance gene expression studies.