Knowing how proteins interact in a cell gives important clues to their function, and in recent months the debate has intensified over how best to measure protein-protein interactions. Earlier this year, Cellzome and MDS Proteomics published rival protein-protein interaction studies in yeast — and both sides claimed to have used the superior approach. However, a paper published online May 8 in Nature attempts to prove that these and other approaches each have their own bias, and that no single method is sufficient.
To compare high-throughput interaction methods, Peer Bork of the European Molecular Biology Laboratory, the corresponding author of the Nature paper, and his colleagues at Cellzome, the University of Manchester, and the University of Washington chose datasets from studies of protein interactions in yeast obtained by five different methods. The group analyzed purified protein complexes analyzed by mass spectrometry, yeast two-hybrid interactions, correlated mRNA expression (synexpression) studies, genetic interactions (synthetic lethality), and in silico predictions from the genome.
Although all these methods are aimed at finding interactions, they differed vastly in their results: Of the about 80,000 interactions from the combined five approaches, only about 2,500 were supported by more than one method.
To explain this, Bork and his colleagues describe how some methods are better than others at discovering interactions within certain functional protein classes. Also, some of the approaches were biased towards highly abundant proteins, and almost every method favored particular subcellular compartments.
However, not all the measured protein-protein interactions are likely to be “real”. To compare the accuracy of the various approaches, the authors compared each dataset against a “trusted reference set” of manually-curated protein complexes obtained from the Munich Information Center for Protein Sequences (MIPS) and the Yeast Proteome Database (YPD). These were derived by a variety of methods and added up to about 11,000 interactions.
The authors found that the tandem affinity protein complex purification method employed by Cellzome ranked best in its coverage of the reference set, better than the equivalent method used by MDS Proteomics, which relies on overexpression and a single purification step.
Another measure of quality the authors used was the fraction of interactions between proteins belonging to the same functional group — these are more likely to be “true,” according to theory. By this measure, the MDS Proteomics approach seemed to contain a high number of false positives.
However, some scientists, particularly those from the MDS camp, are concerned that the “benchmark” set used in the analysis might overrepresent abundant complexes, and still contain spurious interactions. In addition, the paper may have found a high number of false positives in the MDS data because the authors failed to use appropriate filters, said Mike Tyers, a professor at the University of Toronto and co-founder of MDS Proteomics.
Tyers also claimed that the discrepancies could be accounted for by the different baits used by the MDS and Cellzome teams in their earlier analysis. In the paper, however, Bork and his co-authors claim that accounting for the different baits would not change their overall results.
Whichever method might be more accurate, overall the authors did not rank any of them as especially reliable, estimating that more than half of all current high-throughput interactions are essentially false-positives. “To improve confidence in detected or predicted protein interactions, as many complementary methods as possible should be used.
For companies concentrating on a single platform technology, this might mean joining forces to get reliable data, said Tyers. “There are going to be some very interesting alliances made between companies,” he said.