Exploring the impact of structure prediction algorithms on structural convergence of antibody repertoires

Abstract

The antibody repertoire is essential in the defense against pathogens. Exploring the overlap in functionally similar antibodies between individuals can improve our understanding of the antibody response. Previous sequence-based convergence analyses only detected a limited amount of shared functional antibody clusters. It has been proposed that structure-based analyses can discover more shared clusters since antibodies can have similar structures despite sequence dissimilarity. For these analyses, high-throughput structure prediction algorithms are essential. It remains unknown how consistent convergence analysis results are across different prediction algorithms. In this study, we aim to clarify the effect of this algorithm choice on convergence results. To examine this, repertoirewide sequencing data is used as input to IgFold, ESMFold and ABB2 to model antibody structures. Similarity across these algorithms correlated poorly based on biophysical properties, while a strong correlation was observed based on RMSD. Therefore, only RMSD results were used for convergence analyses. These analyses detect distinct sets of most similar antibodies, and therefore distinct public structures, depending on the choice of prediction algorithm. Furthermore, results depended on choices in the RMSD threshold, clustering method and included antibody regions. Standardization of these choices and reported metrics is essential for reliable implementation of structure-based convergence analyses. The dependence of reliable thresholds on sequence data and structure-based algorithms complicates this. A wider investigation into standardization should be performed. Still, structure-based analyses have the potential to yield insights that can not be discovered based on sequences alone. This could help to unravel the underlying mechanics behind epitope binding and improve future efforts in pharmaceutical antibody development.