As recently discovered, a comprehensive profiling of antibodies in a patient's blood can be obtained using random-sequence peptides on microarrays and analyzed for medical diagnosis. In this paper, we propose a novel adaptive learning methodology for biothreat detection and classification, which extracts and models appropriate stochastic features from such immunosignatures. The technique is based on the use of Dirichlet process mixture models to adaptively cluster the microarray measurements in feature space. This learning-while- classifying strategy provides the capability of adaptively detecting new biothreat agents on the fly. We demonstrate the utility of the proposed method by classifying diseases using real experimental peptide microarray data.