Mining functionally relevant gene sets for analyzing physiologically novel clinical expression data

Gene set analyses have become a standard approach for increasing the sensitivity of transcriptomic studies. However, analytical methods incorporating gene sets require the availability of pre-defined gene sets relevant to the underlying physiology being studied. For novel physiological problems, relevant gene sets may be unavailable or existing gene set databases may bias the results towards only the best-studied of the relevant biological processes. We describe a successful attempt to mine novel functional gene sets for translational projects where the underlying physiology is not necessarily well characterized in existing annotation databases. We choose targeted training data from public expression data repositories and define new criteria for selecting biclusters to serve as candidate gene sets. Many of the discovered gene sets show little or no enrichment for informative Gene Ontology terms or other functional annotation. However, we observe that such gene sets show coherent differential expression in new clinical test data sets, even if derived from different species, tissues, and disease states. We demonstrate the efficacy of this method on a human metabolic data set, where we discover novel, uncharacterized gene sets that are diagnostic of diabetes, and on additional data sets related to neuronal processes and human development. Our results suggest that our approach may be an efficient way to generate a collection of gene sets relevant to the analysis of data for novel clinical applications where existing functional annotation is relatively incomplete.