Integrating artificial intelligence-based epitope prediction in a SARS-CoV-2 antibody discovery pipeline: caution is warranted

Abstract

SARS-CoV-2-neutralizing antibodies (nABs) showed great promise in the early phases of the COVID-19 pandemic. The emergence of resistant strains, however, quickly rendered the majority of clinically approved nABs ineffective. This underscored the imperative to develop nAB cocktails targeting non-overlapping epitopes. Undertaking a nAB discovery program, we employed a classical workflow, while integrating artificial intelligence (AI)-based prediction to select non-competing nABs very early in the pipeline. We identified and invivo validated (in female Syrian hamsters) two highly potent nABs. Despite the promising results, in depth cryo-EM structural analysis demonstrated that the AI-based prediction employed with the intention to ensure non-overlapping epitopes was inaccurate. The two nABs in fact bound to the same receptor-binding epitope in a remarkably similar manner. Our findings indicate that, even in the Alphafold era, AI-based predictions of paratope-epitope interactions are rough and experimental validation of epitopes remains an essential cornerstone of a successful nAB lead selection. Full list of funders is provided at the end of the manuscript.