Bringing Out the Best Features of Expression Data

Scientists are constantly classifying objects based on observation: A Drosophila geneticist sexes flies; a taxonomist sorts butterflies according to genus and species; a physician interviews patients, observing symptoms and rapidly classifying patients according to their disease, and predicting how they will respond to various therapies. The most successful patient interview is accompanied by prior knowledge of what questions to ask, and which observable variables (“features”) are most salient to the classification problem at hand. Although the process of learning salient features in biology and medicine has traditionally been based on a combination of experience, intuition, and anecdotal evidence, it has increasingly been approached from a statistical perspective. However, choosing patient features salient to diagnosis from among tens of thousands of potential features, after drawing on experience from only tens of patients, is a whole new ball game. This is the challenge facing those who seek to derive diagnostic markers from gene expression array data.