The recent advent of the microarray has enabled simultaneous monitoring of the expression of thousands of genes, and with it have come vexing questions of how best to interpret the new wealth of data being gathered. Although many important uses have been made of this technology, such as for diagnosis, prediction of treatment response and clinical outcome, and prediction of metastatic status, we will focus attention on only these specific applications. It is clear that there is no consensus at present on the most effective methods for analysis of such gene expression data, and here we consider only some approaches that have forged a link between nonlinear system identification and construction of effective gene‐expression‐based class predictors. As we will discuss, one hallmark of these approaches is that they can produce effective predictors with considerably less training data than some statistically based prediction methods.