(471f) Using Pre- and Post-Binding Interaction Stabilization Features to Identify Non-Realistic Computationally Predicted Protein Interfaces

The last few years have seen the rapid proliferation of machine learning- (ML) based protein design methods. Although these methods have resulted in large increases in experimental success rate compared to prior efforts, the majority of their predictions still fail when experimentally tested. This stands in stark contrast to ML methods for protein structure prediction, which are accurate within experimental uncertainty for nearly all proteins.

To understand the types of interactions that should occur in protein binding interfaces, we conducted short, 5 ns molecular dynamics (MD) simulations of 20 single chain variable fragment (scFv) antibody-protein complexes. Hydrogen bonds, hydrophobic interactions, and salt bridges that appear in the experimental structures and spontaneously during the MD simulations were evaluated for their persistences and energies based on their pre- and post-binding stabilization features. The interactions were all determined using geometric criteria, rather than being based on energy values.

The results demonstrate that there are large and statistically significant requirements for the interactions that persist in a bound complex and contribute meaningful energy to it. Specifically, the interactions must be present in the initial complex; both residues involved in a hydrogen bond must be stabilized in the bound complex; at least one of the residues in a salt bridge must be stabilized in the bound complex; and pre- and post-binding stabilization is unimportant for hydrophobic interactions. These findings have allowed us to rapidly scan computationally predicted protein complexes and identify ones that are experimentally wrong but appear excellent by previous computational methods.