LoCo: a novel main chain scoring function for protein structure prediction based on local coordinates

摘要

Background Successful protein structure prediction requires accurate low-resolution scoring functions so that protein main chain conformations that are close to the native can be identified. Once that is accomplished, a more detailed and time-consuming treatment to produce all-atom models can be undertaken. The earliest low-resolution scoring used simple distance-based "contact potentials," but more recently, the relative orientations of interacting amino acids have been taken into account to improve performance. Results We developed a new knowledge-based scoring function, LoCo, that locates the interaction partners of each individual residue within a local coordinate system based only on the position of its main chain N, C_α and C atoms. LoCo was trained on a large set of experimentally determined structures and optimized using standard sets of modeled structures, or "decoys." No structure used to train or optimize the function was included among those used to test it. When tested against 29 other published main chain functions on a group of 77 commonly used decoy sets, our function outperformed all others in C_α RMSD rank of the best-scoring decoy, with statistically significant p-values Conclusions Our function demonstrates an unmatched combination of accuracy, speed, and simplicity and shows excellent promise for protein structure prediction. Broader applications may include protein-protein interactions and protein design.