CASP provides a unique opportunity to compare the performance of automatic fold recognition methods with the performance of manual experts who might use these methods. Here, we show that a novel automatic fold recognition server, Pmodeller, is getting close to the performance of manual experts. Although a small group of experts still perform better, most of the experts participating in CASP5 actually performed worse even though they had full access to all automatic predictions. Pmodeller is based on Pcons (Lundström et al., Protein Sci 2001; 10(11):2354-2365) the first "consensus" predictor that uses predictions from many other servers. Therefore, the success of Pmodeller and other consensus servers should be seen as a tribute to the collective of all developers of fold recognition servers. Furthermore we show that the inclusion of another novel method, ProQ2, to evaluate the quality of the protein models improves the predictions.
Copyright 2003 Wiley-Liss, Inc.