Response-related filtering of the NCI-60 data - continued. (a) The projection score as a function of log10(α), where α is the p-value threshold used for inclusion when contrasting each of four cancer types with all the other eight types in the NCI60 data set. For the Melanoma, Leukemia and Renal types, small groups of variables form the most informative subsets. For the NSCLC type, the entire variable collection is the most informative variable subset. (b) The p-value distribution for all variables when contrasting NSCLC with all other groups, indicating that there are essentially no truly significantly differentially expressed genes for this contrast. (c) The p-value distribution for all variables when contrasting Melanoma with all other groups.