Effects of functional bias on supervised learning of a gene network model.

Gene networks have proven to be an effective approach for modeling cellular systems, capable of capturing some of the extreme complexity of cells in a formal theoretical framework. Not surprisingly, this complexity, combined with our still-limited amount of experimental data measuring the genes and their interactions, makes the reconstruction of gene networks difficult. One powerful strategy has been to analyze functional genomics data using supervised learning of network relationships based upon reference examples from our current knowledge. However, this reliance on the set of reference examples for the supervised learning can introduce major pitfalls, with misleading reference sets resulting in suboptimal learning. There are three requirements for an effective reference set: comprehensiveness, reliability, and freedom from bias. Perhaps not too surprisingly, our current knowledge about gene function is highly biased toward several specific biological functions, such as protein synthesis. This functional bias in the reference set, especially combined with the corresponding functional bias in data sets, induces biased learning that can, in turn, lead to false positive biological discoveries, as we show here for the yeast Saccharomyces cerevisiae. This suggests that careful use of current knowledge and genomics data is required for successful gene network modeling using the supervised learning approach. We provide guidance for better use of these data in learning gene networks.