CAMBRIDGE, Mass.—For years, scientists have used computational methods to model protein surfaces and virtually dock druglike molecules, searching for new therapeutics. But according to Dr. Emanuele Perola of Vertex Pharmaceuticals, for all the improvements in hardware, software and the number of protein structures, the hit rates of most studies don't go beyond 10 percent.
Perola used Vertex's vast compound libraries and the structures of six kinases to look for features that might allow one to differentiate true from false positives, reporting the work in Proteins.
Perola used Glide-based virtual screening to test 5000 compounds and as was expected, found that under normal scoring conditions, most true positives ranked highest while true negatives ranked lowest. False negatives, however, were often found to be docked incorrectly when compared to crystal structures, indicating that the error was not a function of scoring, as had been suspected.
Furthermore, Perola noted that key active-site hydrogen bonds between ligand and protein found in the true positives were absent from the false positives. When these critical interactions were added as a scoring rule, the screening success rate generally rose, suggesting that in some cases, computational screening still needs a human touch.