Figure 3

AUCs for gene expression levels. Comparison between AUC for 67 gene expression levels considering the best 50 predictive variables chosen among all markers and cDNA levels (red solid squares), the best 50 variables chosen among all markers (green solid triangles) and considering the best 50 variables chosen among all transcript levels (blue open circles). All three AUCs for each expression level are in the same abscissa's position, genes were ranked according to AUC using all variables. It can be seen that using only markers results in consistently lower AUC, whereas there are no large differences between using all variables or only transcript levels. For some genes (23 out of 67), AUCs using only cDNAs were slightly better than using all variables, this occurred because the RFE algorithm [10] may not completely remove redundant information from all variables and thus does not always guarantee the absolute maximum. The 67 genes shown were chosen within those with most significant QTLs in Chesler et al. (2005). Thus, one should expect that markers are better predictors, and consequently higher AUC, for these genes than for a random gene. Note that an AUC of 50% means than the criterion is no better than a random ordering.