MOTIVATION: Microarray data has been shown recently to be efficacious in distinguishing closely related cell types that often appear in different forms of cancer, but is not yet practical clinically. However, the data might be used to construct a minimal set of marker genes that could then be used clinically by making antibody assays to diagnose a specific type of cancer. Here a replication algorithm is used for this purpose. It evolves an ensemble of predictors, all using different combinations of genes to generate a set of optimal predictors. RESULTS: We apply this method to the leukemia data of the Whitehead/MIT group that attempts to differentially diagnose two kinds of leukemia, and also to data of Khan et al. to distinguish four different kinds of childhood cancers. In the latter case we were able to reduce the number of genes needed from 96 to less than 15, while at the same time being able to classify all of their test data perfectly. We also apply this method to two other cases, Diffuse large B-cell lymphoma data (Shipp et al., 2002), and data of Ramaswamy et al. on multiclass diagnosis of 14 common tumor types. AVAILABILITY: http://stravinsky.ucsc.edu/josh/gesses/.