Some initial experiments using Genetic Programming to evolve comprehensible robust predictive models based on gene expression as measured by DNA chips will be described. DNA chip data is notoriously difficult to data mine. It is noisy, has thousands of features, but often few training examples. The approach is based on selecting genes via a small number of iterations of many Genetic Programming runs in parallel. In a final iteration a non-linear model is evolved using a few tens of genes. Simple models are preferred over predictive accuracy.