Analysis of gene expression data in terms of a priori-defined gene sets typically yields more compact and interpretable results than those produced by traditional methods that rely on individual genes. The set-level strategy can also be adopted in predictive classification tasks accomplished with machine learning algorithms. Here, sample features originally corresponding to genes are replaced by a much smaller number of features, each corresponding to a gene set and aggregating expressions of its members into a single real value. Classifiers learned from such transformed features promise better interpretability in that they derive class predictions from overall expressions of selected gene sets (e.g. corresponding to pathways) rather than expressions of specific genes. In a large collection of experiments we test how accurate such classifiers are compared to traditional classifiers based on genes. Furthermore, we translate some recently published gene set analysis techniques to the above proposed machine learning setting and assess their contributions to the classification accuracies.