Contact area difference (CAD): a robust measure to evaluate accuracy of protein models

Abstract

A simple unified measure to evaluate the accuracy of three-dimensional atomic protein models is proposed. This measure is a normalized sum of absolute differences of residue-residue contact surface areas calculated for a reference structure and a model. It employs more rigorous quantitative evaluation of a contact than previously used contact measures. We argue that the contact area difference (CAD) number is a robust single measure to evaluate protein structure predictions in a wide range of model accuracies, from ab initio and threading models to models by homology, since it reflects both backbone topology and side-chain packing, is smooth, continuous and threshold-free, is not sensitive to typical crystallographic errors and ambiguities, adequately penalizes domain and/or secondary structure rearrangements and protein plasticity, and has consistent linear and matrix representations for more detailed analysis. The CAD quality of crystallographic structures, NMR structures, models by homology, and unfolded and misfolded structures is evaluated. It is shown that the CAD number discriminates between models better than Cartesian root-mean-square deviation (cRMSD). Structural variability of the NMR structures was found to be three times larger than deformations of crystallographic structures in different packing environments.