As with any new method, it is important when interpreting the results (our large set of predicted complex structures) to keep in mind the limitations of the approach. First, our study is not comprehensive, so conclusions should not be drawn about absences; in particular we eliminated proteins that arose from recent duplication due to difficulty in identifying orthologs in other organisms, and thus only surveyed 2/3 of the entire yeast proteome. Second, the approach likely misses interactions restricted to a small set of organisms, or which vary rapidly during evolution, due to weaker co-evolutionary signals. Third, the approach likely works less well for transient interactions which generally involve smaller and weaker interfaces which may be under lower selective pressure, in particular those involving intrinsically disordered regions which are poorly represented in the PDB. The majority of known interactions identified by our approach are likely obligate assemblies and involve ordered structural elements. Fourth, interactions between single hydrophobic or amphipathic helices, such as single transmembrane helices or coiled coils, may be overpredicted (in initial studies of human complexes, interactions solely between single-pass transmembrane regions appear to be over represented). Fifth, and perhaps most importantly, for proteins that form high-order obligate protein complexes, binary complex models may be quite inaccurate, as illustrated by the SNARE example.