On the transferability of folding and threading potentials and sequence-independent filters for protein folding simulations

Abstract Significant progress has recently been made in de novo protein structure prediction. The Rosetta method by Baker and colleagues, which is based on the idea of assembling putative models from a library of k-mer fragments derived from known three-dimensional protein structures, proved to be particularly successful. Critical components of the Rosetta approach are various sequence-dependent as well as sequence-independent measures that are used to rank alternative models and to enhance sampling of native-like conformations. In the present work we revisit several sequence-independent filters that have been used to enhance the discrimination of native and native-like structures from misfolded structures, such as the overall compactness of the structure and its contact order. We also propose a novel sequence-independent filter, based on the shape of the mean inter-residue radial distribution function. Using the Rosetta, Park–Levitt and CASP4 sets of decoys it is shown that sequence-independent filters are in fact more successful in distinguishing native structures in Rosetta and CASP4 tests than commonly used knowledge-based pairwise potentials. The latter are typically designed to distinguish native structures in a population of well-folded alternatives, and they fail to discriminate between native-like and non-physically packed misfolded structures from Rosetta simulations. Moreover, a rigorous attempt to optimize pairwise potentials for recognition of homologous structures in threading by using a linear programming approach leads to further deterioration of performance in terms of recognition of native structures from the Rosetta set. These findings shed light onto the success of tailored scoring functions used in the Rosetta protocol and provide support for explicit inclusion of both sequence dependent and sequence independent measures in the design of scoring functions. A Web server that enables ranking of decoy structures according to sequence independent filters considered here is available at http://sift.chmcc.org.