We describe CONTRAST, a gene predictor which directly incorporates information from multiple alignments rather than employing phylogenetic models. This is accomplished through the use of discriminative machine learning techniques, including a novel training algorithm. We use a two-stage approach, in which a set of binary classifiers designed to recognize coding region boundaries is combined with a global model of gene structure. CONTRAST predicts exact coding region structures for 65% more human genes than the previous state-of-the-art method, misses 46% fewer exons and displays comparable gains in specificity.