Analyzing protein function on a genomic scale: the importance of gold-standard positives and negatives for network prediction.

The concept of 'protein function' is rather 'fuzzy' because it is often based on whimsical terms or contradictory nomenclature. This currently presents a challenge for functional genomics because precise definitions are essential for most computational approaches. Addressing this challenge, the notion of networks between biological entities (including molecular and genetic interaction networks as well as transcriptional regulatory relationships) potentially provides a unifying language suitable for the systematic description of protein function. Predicting the edges in protein networks requires reference sets of examples with known outcome (that is, 'gold standards'). Such reference sets should ideally include positive examples - as is now widely appreciated - but also, equally importantly, negative ones. Moreover, it is necessary to consider the expected relative occurrence of positives and negatives because this affects the misclassification rates of experiments and computational predictions. For instance, a reason why genome-wide, experimental protein-protein interaction networks have high inaccuracies is that the prior probability of finding interactions (positives) rather than non-interacting protein pairs (negatives) in unbiased screens is very small. These problems can be addressed by constructing well-defined sets of non-interacting proteins from subcellular localization data, which allows computing the probability of interactions based on evidence from multiple datasets.