Comparative Evaluation of Computational Methods for Modeling SARS-CoV-2 Spike Protein RBD - Antibody Complexes

Computational modeling of protein-protein interactions (PPIs) is a high priority because it is essential for understanding nearly all cellular functions and for addressing the limitations of time-consuming experimental methods. This approach is crucial for drug discovery, as it enables rapid identification of disease targets, the design of therapeutics that modulate PPIs, and the simulation of interactions under realistic, crowded cellular conditions. The SARS-CoV-2 Spike protein's Receptor-Binding Domain (RBD) is the crucial region that directly engages the human ACE2 receptor to initiate infection. Therefore, studying the RBD and its interactions with neutralizing antibodies is paramount for understanding immune protection, assessing the threat posed by viral variants, and guiding the design of effective vaccines and antibody-based therapeutics. Here, we evaluated the performance of several computational tools for modeling complexes between the SARS-CoV-2 Spike protein receptor-binding domain (RBD) and various neutralizing antibodies. Recent breakthroughs in computational chemistry have moved beyond traditional protein-protein docking towards various generative methods and co-folding tools that predict a protein's 3D structure and binding poses from its sequence. Therefore, this motivates us to compare the performance of traditional docking methods, such as pyDockWEB and ClusPro, with that of novel, promising AI-driven tools, such as AlphaFold 3, Boltz-2, Protenix, and Chai-1. All these methods were systematically evaluated for their ability to reproduce the 3D structures of known Spike RBD-antibody complexes. We found that traditional docking tools, such as ClusPro and pyDockWeb, perform well at capturing correct protein-protein interactions for relatively small antibodies with well-defined interaction interfaces, but fall short at reproducing more complex protein-protein assemblies. AlphaFold 3 performs best at reproducing the 3D structures of the five studied RBD-antibody complexes among the four AI-driven prediction tools considered. Our study sheds light on understanding protein-protein interactions and provides a practical guide for accurate modeling of viral Spike protein RBD-antibody interactions.