What genes are useful for classification?
	This is the classical machine-learning (statistical) problem of feature (variable) selection. Rank the genes according to their separation between classes, using some kind of t-statistic, or recursively choose the next best feature from the list of unused features. Consider a series of classifiers incorporating the first p genes on the list, p = 1,2,… Estimate the performance of the p-feature classifier with leave-out-one analysis. Pick the best performing p.