Current cancer classifications using morphological criteria produce heterogeneous classes with variable prognosis and clinical course. By measuring gene expression for thousands of genes in a single hybridization experiment, microarrays have the potential to contribute to more effective classifications based on molecular information. This gives hope to improve both prognosis and treatment. Statistical methods for molecular classification have focused on using high dimensional representations of molecular profiles to identify subclasses. These can be noisy, unstable, and highly platform-specific. In this article, we emphasize the notion of molecular profiles based on latent categories signifying under-, over-, and baseline expression. Following this approach, we can generate results that are more easily interpretable, more easily translated into clinical tools, more robust to noise, and less platform-dependent. We illustrate both the methods and the associated software for molecular class discovery on a data set of 244 microarrays comprising six known leukemia classes.