The Genom of Homo sapiens.pdf

HAPLOTYPE EXPRESSION ASSOCIATIONS 91nificance of the result in the face of multiple testing ofmany markers. For this purpose, the omnibus statistic isrecalculated for up to 6,250 randomly permuted data setsrepresenting the null hypothesis of no true association.The permutation was done in such a way that the expressionvalues were randomly redistributed among the subjects,whereas the genetic assignments were left alone.The fraction of these instances of the null hypothesis thatresult in a more extreme omnibus statistic is a direct measurefor the probability of the observed statistic havingarisen by chance, and is called the p value. More specifically,if N random permutations are performed, and M ofthe permutations yield an omnibus statistic more extremethan that of the original data, the probability of the nullhypothesis is p = M/N. This procedure is nonparametric,meaning it makes no assumptions about the distributionof the expression values. In addition, because of the useof the omnibus statistic, the testing of multiple markers isfully accounted for. In fact, most other methods used toadjust for multiple comparisons would result in a largeloss of sensitivity, since there is no way to accurately accountfor the high degree of correlation between themarkers, many of which differ only slightly from eachother. For efficiency, the permutation testing was done onsuccessively larger sets of permutations N = (50, 250,1250, 6250), and the next higher set of permutations wasskipped when the number of false positives M was 10 ormore in the previous set.After the full association analysis was run, we furtherfiltered the results based on the quality of the expressiondata. Each expression level for each sample is associatedwith both a quantitative expression value and binary present/absentlabel. This label is calculated by a statisticalalgorithm in the MAS 5.0 software, and a present labelmeans that there is a probability of less than 5% that thereis no signal and what is measured is only noise. For mostanalyses, we excluded fragments for which there werefewer than 10 present calls out of 89 total. Additionally,for some analyses, we excluded fragments for which themaximum fold change was less than two. This is calculatedby taking the maximum and minimum expressionvalues for the expression level in the subset of samplewith “present” calls and taking the ratio. In addition, weexcluded in the analysis the genes that showed more thantwofold of mRNA expression levels within the three samplesfrom the same individual.RESULTSAssociation analysis was carried out examining 6,184gene loci by 22,242 expression levels. At a significancelevel of p ≤ 0.001, we found 66,141 significant associationsbetween gene loci and expression levels. Becausethis is such a large and rich data set, we have concentratedour analysis on two interesting subsets of associations.The first of these analyses concentrates on a set of cis-actingSNPs; i.e., SNPs that affect the expression levels ofthe genes they reside in. The second example examines agene coding for a largely uncharacterized zinc finger protein(ZNF295), which likely forms part of a transcriptionfactor complex. Two amino acid-changing SNPs in thislocus are strongly associated with expression levels in alarge number of other genes. Finally, we examine theglobal statistics of the data set and attempt to compare ourresults to some reported previously.cis-Acting Genetic VariationsNot all of the gene loci for which we have haplotypedata are represented by expression probes, and, viceversa, not all of the genes represented by expressionprobes have known haplotypes associated with them. Inour experiment, there were 4,762 gene loci and 7,553 expressionprobes where each of the loci belonged to thesame gene as one or more of the probes. We identified 22genes from our initial screen that have candidate cis-associationssignificant at a level of p