Learned features of antibody-antigen binding affinity

Defining predictors of antigen-binding affinity of antibodies is valuable for engineering therapeutic antibodies with high binding affinity to their targets. However, this task is challenging owing to the huge diversity in the conformations of the complementarity determining regions of antibodies and the mode of engagement between antibody and antigen. In this study, we used the structural antibody database (SAbDab) to identify features that can discriminate high- and low-binding affinity across a 5-log scale. First, we abstracted features based on previously learned representations of protein-protein interactions to derive 'complex' feature sets, which include energetic, statistical, network-based, and machine-learned features. Second, we contrasted these complex feature sets with additional 'simple' feature sets based on counts of contacts between antibody and antigen. By investigating the predictive potential of 700 features contained in the eight complex and simple feature sets, we observed that simple feature sets perform comparably to complex feature sets in classification of binding affinity. Moreover, combining features from all eight feature-sets provided the best classification performance (median cross-validation AUROC and F1-score of 0.72). Of note, classification performance is substantially improved when several sources of data leakage (e.g., homologous antibodies) are not removed from the dataset, emphasizing a potential pitfall in this task. We additionally observe a classification performance plateau across diverse featurization approaches, highlighting the need for additional affinity-labeled antibody-antigen structural data. The findings from our present study set the stage for future studies aimed at multiple-log enhancement of antibody affinity through feature-guided engineering.