(681b) Developing Similarity Matrices for Protein-Protein Interactions

The last decade has seen unprecedented advances in the state-of-the-art in computational protein engineering and design. Today, it is becoming increasingly possible to accurately predict the structures of naturally occurring proteins or design new proteins computationally. With increasing successes in solving protein structures, more and more research focus is turning towards predicting protein-protein interactions. The prevalence of very large-scale datasets from –omics technologies provides an abundance of data for addressing these problems.

Similarity matrices provide scores on the relative effects of mutating one amino acid to another. They are statistical tools that have a long history of successful use for predicting protein structures. As such, they are also being utilized for predicting protein-protein interaction properties. However, existing similarity matrices were developed based on rates of amino acid mutations in homologous protein sequences. The relative importance of amino acids for stabilizing protein-protein interactions is likely to be different than for stabilizing protein structures.

We previously identified a non-redundant database of 492 antibody-protein complexes and demonstrated that most of the binding energy comes from a subset of 5-6 amino acids. Using three protein forcefields, CHARMM, Amber and Rosetta, we have conducted a systematic mutation of these important amino acids to all alternatives and calculated the corresponding changes in predicted binding energies. From this data, we have constructed three similarity matrices for protein-protein interactions, one for each forcefield. We will present on the development of the matrices, the similarities and differences between the results for the forcefields, and the similarities and differences with popular BLOSUM and PAM similarity matrices.