AbDesign: database of point mutants of antibodies with associated structures reveals poor generalization of binding predictions from machine learning models.

Antibodies are naturally evolved molecular recognition scaffolds that can bind a variety of surfaces. Their designability is crucial to the development of biologics, with computational methods holding promise in accelerating the delivery of medicines to the clinic. Modeling antibody-antigen recognition is prohibitively difficult, with data paucity being one of the biggest hurdles. Current affinity datasets comprise a small number of experimental measurements, which are often not standardized between molecules. Here, we address these issues by creating a dataset of seven antigens with two antibodies each, for which we introduce a heterogeneous set of mutations to the CDR-H3 measured by ELISA. Each of the parental complexes has a known crystal structure. We perform benchmarking of state-of-the-art affinity prediction algorithms to gauge their effectiveness. Current computational methods exhibit substantial limitations in accurately predicting the effects of single-point mutations. In contrast, the older empirical, physics-based method FoldX performs well in identifying mutants that retain binding. These findings highlight the need for more resources like the one presented here, i.e. large, molecularly diverse, and experimentally consistent datasets.