Personalized medicine aims at using molecular signatures of individual patients for better matching of disease with specific therapies. The success of personalized medicine depends on having accurate models that can classify patients into distinct subgroups; each subgroup corresponds to an optimal treatment for the patient in the subgroup. This talk presents a method to identify sets of markers sharing compatible patterns across a subset of patients via bicluster analysis. A bicluster analysis generates a collection of binary classifiers from the collection of biclusters. A composite prediction model is then developed to classify patients into several disjoint subgroups. Each subgroup can represent different tumor subtypes or different genotypes of patients related to treatment response. The approach is illustrated by applications to synthetic data and real datasets.