Genetical genomics [1] integrates genome-wide expression profile data of microarray experiments and marker-based measures of genetic variation. It has newly become a central methodology in quantitative trait studies in order to determine loci involved in regulatory expression of quantitative variation in RNA level. As a result of the reduced cost of microarray expression arrays and marker genotyping, these studies have been extended to study quantitative traits in linkage and association studies [2, 3]. In microarray experiments, research on determining which normalization method best adjusts for systematic experimental variation to produce unbiased data is necessary. Among the normalization methods used for the common Affymetrix GeneChip, the robust multiarray average (RMA) method and the statistical algorithm implemented in Affymetrix's Microarray Suite (MAS5) program are the gold standard to control for systematic variation in samples of unrelated individuals [4, 5]. Investigation of the effects of various normalization methods in family data are needed [6].

RMA adjusts for systematic variation by performing a quantile normalization procedure, which assumes that the data for the variable considered (such as study sample) all are sampled from the same or similar distributions and the values for that variable are then normalized to a standard distribution. However, it is not yet known which standard distribution is the best to use for family data in genetical genomics studies. Linkage analysis utilizes data from pedigrees, which usually have more homogeneity within pedigrees for the trait under study than between pedigrees, both biologically and environmentally. Thus, the increased background sharing within pedigrees can result in increased correlation between related individuals for expression levels. Consequently, the familial correlation within each pedigree due to biological similarity is of interest in linkage studies. Therefore, we hypothesized that linkage analysis that uses expression trait data normalized by ignoring pedigree membership could improperly 'correct' the trait values by imposing equal means and variances across pedigrees. A conceptual graph (Fig. 1) plots the values of one trait for each individual. It shows how an outlier in a family, whose members have a different distribution of expression values than other families, will be systematically adjusted (solid arrow) toward the familial mean value if its value is normalized by within-pedigree normalization methods. However, the normalized trait values will be over-corrected and biased toward the overall mean if normalized using all individuals because this method falsely assumes that they are unrelated individuals sampled from a single distribution. A second member of the same family would be adjusted in a different direction (dashed arrow) away from the overall mean and toward the family mean if its value is normalized by within-pedigree normalization.

Our aim is to maintain the individual familial distributions with normalization. We address this problem of normalization of expression data within pedigrees by comparing two possible standard distributions for normalization: 1) applying RMA across all arrays as a pool, which assumes all individuals are independent and share the same distribution of trait values, and 2) applying RMA to arrays within pedigrees assuming family members within a pedigree share the same distribution.

Study subjects consisted of 194 individuals from 14 Centre d'Etude du Polymorphisme Humain (CEPH) Utah pedigrees, and 2882 autosomal and X-linked single-nucleotide polymorphism (SNP) genotypes were available from The SNP Consortium [2]. Quantitative phenotype data were generated from immortalized B cells, and 8793 gene expression values were available from the microarray raw CEL (cell intensity file) data files from the Affymetrix Genechips and Hgfocus CDF (chip description file) files. In cases in which a subject had more than one expression file, the first replicate was chosen for this analysis so that we used one array per subject. To evaluate the performance of different scenarios, two data sets were used: 1) RMA normalized using all individuals (arrays) as the normalization pool, and 2) RMA normalization applied to a within-pedigree pool to be able to allow for as many distributions as the number of pedigrees. The 'affy' package v1.8 in R v2.3.1 was used to perform the normalizations of the expression data [7].

In order to examine the effect of using different normalization pools consisting of different types of individuals, we performed paired t-tests using 8793 gene expression values of four founders of one family in two ways: comparing the values of the four founders after normalizing using themselves as the pool to paired gene expression values after normalizing 1) using all family members including themselves as the pool, and 2) using independent individuals as the pool (other grandparents from other families) [8]. The purpose of microarray normalization methods is to remove systematic variation while preserving biological variation. Therefore, comparison of the normalized gene expression data using the four founders as their own normalization pool to the normalized data using either of the other two pools (family members or independent individuals) should not show significant differences if the normalization pools are all only removing random error variation. Thus, we should observe non-significant p-values for these paired t-tests if only random variation is removed by each method. We assumed that these t-tests indicated a significant difference in the normalization methods if the p-value was less than a conservative p-value of 0.001 (since we were performing over 8,000 tests). However, we also evaluated this using a p-value of 0.05 as the significance threshold.

To examine the effect of normalization methods on linkage results, we selected 18 cis-acting transcriptional regulator phenotypes with previous evidence of cis-acting linkage to the known location of each corresponding structural gene [2]. Nonparametric quantitative linkage analysis was performed using Merlin v1.0 with the qtl option. We plotted both the negative p-values of the nonparametric linkage score [8] and the allele-sharing LOD score of Kong and Cox [9]. Based on the change of mean and variance of the trait values in each array, some individuals may have different trait values when using different normalization methods. FCOR in S.A.G.E v5.1.0 was used to calculate familial correlations (e.g., parent-offspring, sibling, and grandparent), which were compared for each trait across the different normalization methods.

Because of the presence of systematic variation in generating gene expression data, proper normalization of raw values is required to separate true signals from background noise. Previous gene expression studies have typically focused on unrelated individuals. With the emergence of genetic linkage studies of gene expression traits, the use of pedigrees poses concerns for proper normalization. We investigated the need to account for pedigrees by normalizing within families. Our results for the 18 linked traits show strong familial correlations, which generally increased when traits were normalized within pedigrees. However, in general, the strength of the maximum LOD score decreased (13/18) when traits were normalized within pedigrees. This is not surprising because the increased correlation of a quantitative trait within a family may decrease evidence for linkage. It is also possible that linkage signals are inflated when normalizing using all individuals as the pool. Because we did not evaluate linkage of these 18 traits to all markers on other chromosomes in this data set, we cannot evaluate possible inflation of linkage signals. However, the pattern of the chromosomal linkage plot for these genes (15/18) did not differ across the two normalization strategies. This lack of difference suggests that normalization within pedigrees may not be necessary, at least in studies with small pedigree sizes. However future linkage studies on expression data with larger pedigree sizes and/or large sample sizes may benefit from normalization by pedigree.