Decoy Detection of Computational Protein Designs

Decoy detection is one way to reframe protein folding, not in terms of folding a protein, but in terms of differentiating a well-folded protein from a poorly folded one. An effective decoy scoring metric would both improve prediction methods and indicate how prediction methods fail. The O’Hern group has been making progress in understanding protein structure by focusing on the core regions of proteins, which are inaccessible to solvent. Core structure is uniquely specified by purely repulsive atomic interactions, as hard-sphere interactions are able to predict core structure. In this work, we apply this framework to the decoy detection problem and find that state-of-the-art protein predictions in the CASP11, 12 and 13 competitions often have core regions that are overpacked, due to overlapping residues. Additionally, cores in the predicted protein structures are often too small and too solvent-exposed, suggesting that the prediction methods do not properly capture hydrophobic collapse. Finally, by scoring CASP predictions based on its core structure, we can effectively distinguish between high- and low-quality computational protein designs.