Cooperative approach for the protein fold recognition

We, four independent predictors, organized a team and tackled blind protein structure predictions using fold recognition methods. We tried to assign the homologous or analogous folds in the protein structure database for a number of target sequences that showed no apparent sequence homology to the proteins of known folds. After primary analyses by conventional softwares, these sequences were threaded through the structural library using three different programs developed by ourselves, which employed different compatibility functions. Collecting the results of our individual analyses, and the available biological knowledge about the target, we held meetings and discussed all plausible structures for the target. For 25 target sequences, we submitted 56 models including NONE: This was the first time the fold was determined. At the time of the meeting (CASP3), 19 protein structures (21 domains) categorized as the threading targets were available. We succeeded in predicting eight out of 18 targets (20 domains) that we submitted; however, alignment accuracies were not satisfactory for some of the models. We often obtained correct answers even if some of us missed the right prediction; therefore it would appear that our threaders compensated each other. When all the information is managed effectively, the prediction gains more accuracy. Proteins Suppl 1999;3:126–132. © 1999 Wiley‐Liss, Inc.