Recently, researchers have made some progress in using microarrays to validate predicted exons in genome sequence and find new gene structures. However, current methods rely on separately making threshold-based decisions on intensity of expression, similarity of expression profiles, and arrangements of exons in the genome. We have taken a Bayesian approach and developed GenRate, a generative model that accounts for both genome-wide expression data taken from multiple conditions (e.g. tissues) and co-location and density of probes in DNA sequence data. GenRate balances probabilistic evidence derived from different sources and outputs scores (log-likelihoods) for each gene model, enabling the estimation of false-positive and false-negative rates. The model has a number of local minima that is exponential in the length of the DNA sequence data, so direct application of the EM learning algorithm produces poor results. We describe a novel way of parameterizing the model using examples from the data set, so that good solutions are found using an efficient algorithm. We apply GenRate to a subset of mouse genome-wide expression data that we have created, and discuss the statistical significance of the genes found by GenRate. Three of the highest-ranking gene structures found by GenRate, each containing thousands of bases from the genome, are confirmed using RT-PCR experiments.