Progress and challenges in the computational prediction of gene function using networks

In this opinion piece, we attempt to unify recent arguments we have made that serious confounds affect the use of network data to predict and characterize gene function. The development of computational approaches to determine gene function is a major strand of computational genomics research. However, progress beyond using BLAST to transfer annotations has been surprisingly slow. We have previously argued that a large part of the reported success in using "guilt by association" in network data is due to the tendency of methods to simply assign new functions to already well-annotated genes. While such predictions will tend to be correct, they are generic; it is true, but not very helpful, that a gene with many functions is more likely to have any function. We have also presented evidence that much of the remaining performance in cross-validation cannot be usefully generalized to new predictions, making progressive improvement in analysis difficult to engineer. Here we summarize our findings about how these problems will affect network analysis, discuss some ongoing responses within the field to these issues, and consolidate some recommendations and speculation, which we hope will modestly increase the reliability and specificity of gene function prediction.