We consider the problem of predicting the functions of individual proteins in protein-protein interaction (PPI) networks. Existing techniques assume that proteins that are topologically close in the network tend to have similar functions. We hypothesize that better predictive accuracy can be obtained by generalizing this assumption. We call two functions collaborative if proteins with one function often interact with proteins performing the other function. Our hypothesis is that techniques that extract such function collaboration information from networks, and exploit it, can yield better predictions.We propose and evaluate two such techniques. A comparative evaluation on three S. cerevisiae interaction networks, at different levels of detail, shows that the new techniques consistently improve over state of the art function prediction techniques, with improvements in F-measure ranging from 3% to 17%.