Linkage analysis using principal components of gene expression data

There is great interest in understanding genetic factors related to variable expression of genes. Recently, several studies have shown the first evidence of heritability of mRNA between individuals [1–6]. By treating the expression phenotypes for each transcript (or probe) as a quantitative phenotype, a variance-components linkage analysis could be used [1–6]. The expectation is to detect linkage signals between the gene expression values and genomic regions. As pointed out by William et al. [6], several issues plague these studies, including the selection of informative expression values. Principal components (PC) is a dimension-reduction approach [7] and it has been shown to be a valuable tool in linkage analysis of correlated phenotypes [8]. Multivariate linkage analysis has been shown to be useful in identifying genomic regions with pleiotropic effect [9]. Given that the PC approach is another way to combine information from multiple phenotypes, it can be hypothesized that PC analysis might also be helpful in the identification of pleiotropic effects. Because of the large number of expression phenotypes in the Genetic Analysis Workshop 15 (GAW15) Problem 1 data set, we first selected the most genetically informative phenotypes based on those with high heritability estimates [10]. In this paper, we examine whether combinations of correlated expression phenotypes improved the linkage signals using PC and whether the PC results suggest new pleiotropic effects.

Figure 1 shows the relative weighting for the first five PCs. For each component, the bar height represents the relative influence of a particular probe. Also included is the gene and gene location associated with each expression phenotype. The first PC is dominated by probe 209480.at, which has a relative weighting of 0.979. This first probe was the only component to show any increased linkage signal, changing from a LOD of 9.0 to 10.7. One possible reason for this increase is that the first PC is composed of two phenotypes (209480.at, 204769.s.at) that are associated with the HLA region on chromosome 6. Figure 2 shows the linkage analysis for chromosome 6 using these two phenotypes. Separate lines are drawn for the two univariate analyses, the first PC, and the bivariate analysis using these probes.

Table 1 compares our results with those of Morley et al. [2]. They used the Affymetrix normalization method and their multipoint genome-wide linkage analysis was done using SIBPAL in S.A.G.E. [2]. Two of our six best linkage signals for the single phenotype analysis agreed with Morley's; four, including the HLA region identified using the PC approach, were not found by Morley. For the remaining nine top phenotypes identified by Morley et al., we were unable to compare the results because the specific phenotypes were not part of our final 45.

Additional review of all the bivariate estimates using the de Andrade screening approach showed only two new areas to investigate that would not have been previously flagged using a criteria of p < 10^-9. The DDX17 gene signal, identified by Morley using probe 208151.x.at (22q13.1), was increased from p < 10^-6 to p < 10^-8 when used with probe 207598.x.at (XRCC2, 7q36) in a bivariate analysis. One area in chromosome 6 had a screening p-value of 10^-8 using probes 220386.s.at and 320.at but actual bivariate analysis yielded a p-value similar to the stronger of the phenotypes (10^-6). Neither of these potential associations were strongly grouped in the first 10 PCs. The Dead/H Box 17 (DDX17) is a member of the DEAD box (asp-glu-ala-asp/his) protein family of RNA helicases that are involved in diverse cellular functions including mRNA splicing, ribosome assembly, translation initiation, mRNA stability, and cell growth and division, and XRCC2 is essential for the efficient repair of DNA double-strand breaks by homologous recombination between sister chromatids.

We compared the total number of LOD scores greater than three across the genome for the first 10 PCs with the regions identified by each of the 45 phenotypes and found that in general, the PC analyses under-performed the single probe set analysis for known signals. The component that best reproduced the single probe set analysis (Component 1) was primarily composed of that probe set and also included another probe set that focused on that region. No new signals were detected using PCs despite the strong correlation between the probes. However, we observed a strong linkage signal on chromosome 6 in the HLA region when two probe sets from the HLA region were analyzed as bivariate traits. The two other increases in linkage signals from bivariate analysis that increased our list of interesting areas were not picked up using PCs.

We observed that although PC has been suggested as a potentially useful screening tool for identifying genes linked to a cluster of highly correlated variables/phenotypes, it was not helpful in identifying linkage signals in this data set. Based on this analysis, other data reduction techniques should be investigated.