Modeling three-dimensional protein structures for CASP5 using the 3D-SHOTGUN meta-predictors

Full-atom models were generated for all CASP5 targets by using the fully automated 3D-SHOTGUN fold recognition meta-predictors (Fischer D, Proteins 2003;51:434-441). The 3D-SHOTGUN meta-predictors assemble hybrid 3D models by combining structural information of a number of independently generated, fold recognition models. At the time CASP5 took place, the 3D-SHOTGUN servers generated unrefined C(alpha)-only models. Fischer's participation in CASP had three main goals. The first was to test the value of using 3D-SHOTGUN models as input to a refinement procedure. The second goal was to test whether human intervention could result in a better performance than that of the automated servers. The third goal was to evaluate which human procedures, not yet implemented within the 3D-SHOTGUN servers, can be implemented in the future. For CASP5, our group's predictions applied a very simple approach using the multiple parent option of the Modeller program (Sali and Blundell, J Mol Biol 1993;234:779-815). The input to Modeller was different combinations of the unrefined 3D-SHOTGUN models and the sequence-template alignments used by 3D-SHOTGUN's assembly step. Our evaluation of the accuracies of the refined versus the SHOTGUN models shows that the refined models were consistently slightly more accurate than SHOTGUN's. For a few targets, the manual use of the information from the CAFASP servers resulted in better human models. This manual intervention was particularly valuable in the identification of domains, still a difficult feature for automated servers. The CASP5 results indicate that 3D-SHOTGUN's hybrid models can be a valuable starting point for full-atom refinement and that the resulting refined models are, on average, more accurate than those produced by the servers. Thus, we conclude that our three goals were achieved. A preliminary automated version of the refinement procedure, named SHGUM, is now available.