Motivation: Genome-wide microarray data are often used in challenging classification problems of clinically relevant subtypes of human diseases. However, the identification of a parsimonious robust prediction model that performs consistently well on future independent data has not been successful due to the biased model selection from an extremely large number of candidate models during the classification model search and construction. Furthermore, common criteria of prediction model performance, such as classification error rates, do not provide a sensitive measure for evaluating performance of such astronomic competing models. Also, even though several different classification approaches have been utilized to tackle such classification problems, no direct comparison on these methods have been made. Results: We introduce a novel measure for assessing the performance of a prediction model, the misclassificationpenalized posterior (MiPP), the sum of the posterior classification probabilities penalized by the number of incorrectly classified samples. Using MiPP, we implement a forward stepwise cross-validated procedure to find our optimal prediction models with different numbers of features on a training set. Our final robust classification model and its dimension are determined based on a completely independent test dataset. This MiPP-based classification modeling approach enables us to identify the most parsimonious robust prediction models only with two or three features on well-known microarray datasets. These models show superior performance to other models in the literature that often have more than 40–100 features in their model construction.

Availability: Our MiPP software program is available at the Bioconductor website (http://www. bioconductor. org). Contact: jaeklee@ virginia. edu