A new method is presented for the redesign of protein-protein interfaces, resulting in specificity of the designed pair while maintaining high affinity. The design is based on modular interface architecture and was carried out on the interaction between TEM1 beta-lactamase and its inhibitor protein, beta-lactamase inhibitor protein. The interface between these two proteins is composed of several mostly independent modules. We previously showed that it is possible to delete a complete module without affecting the overall structure of the interface. Here, we replace a complete module with structure fragments taken from nonrelated proteins. Nature-optimized fragments were chosen from 10(7) starting templates found in the Protein Data Bank. A procedure was then developed to identify sets of interacting template residues with a backbone arrangement mimicking the original module. This generated a final list of 361 putative replacement modules that were ranked using a novel scoring function based on grouped atom-atom contact surface areas. The top-ranked designed complex exhibited an affinity of at least the wild-type level and a mode of binding that was remarkably specific despite the absence of negative design in the procedure. In retrospect, the combined application of three factors led to the success of the design approach: utilizing the modular construction of the interface, capitalizing on native rather than artificial templates, and ranking with an accurate atom-atom contact surface scoring function.

Methods for protein modeling and design advanced rapidly in recent years. At the heart of these computational methods is an energy function that calculates the free energy of the system. Many of these functions were also developed to estimate the consequence of mutation on protein stability or binding affinity. In the current study, we chose six different methods that were previously reported as being able to predict the change in protein stability (ddG) upon mutation: CC/PBSA, EGAD, FoldX, I-Mutant2.0, Rosetta and Hunter. We evaluated their performance on a large set of 2156 single mutations, avoiding for each program the mutations used for training. The correlation coefficients between experimental and predicted DeltaDeltaG values were in the range of 0.59 for the best and 0.26 for the worst performing method. All the tested computational methods showed a correct trend in their predictions, but failed in providing the precise values. This is not due to lack in precision of the experimental data, which showed a correlation coefficient of 0.86 between different measurements. Combining the methods did not significantly improve prediction accuracy compared to a single method. These results suggest that there is still room for improvement, which is crucial if we want forcefields to perform better in their various tasks.