Schematic overview of COMBINER. COMBINER first infers candidate modules as activity vectors from each pathway in an inference dataset. It then validates these modules in validation datasets by regenerating activity vectors and performing supervised classification. Finally, the modules present in at least half of the validation sets are considered to be core modules. The resulting core module markers are then projected onto a known protein-protein interaction network. We generated 250 groups of 500 classifiers in parallel using LDA with recursive feature elimination. Both the classifier AUC and weight vectors were computed, and each feature was then ranked by its average normalized weight. The most consistently low-ranking feature was then removed recursively until the average AUC threshold was achieved. At this point, the remaining markers were considered to comprise the final modules.