A compact vocabulary of paratope-epitope interactions enables predictability of antibody-antigen binding

Abstract

Antibody recognition of antigen relies on the specific interaction of amino acids at the paratopeepitope interface. A long-standing question in the fields of immunology and structural biology is whether paratope-epitope interaction is predictable. A fundamental premise for the predictability of paratope-epitope binding is the existence of structural units that are universally shared among antibody-antigen binding complexes. Here, we identified structural interaction motifs, which together compose a vocabulary of paratope-epitope binding that is shared among investigated antibody-antigen complexes. The vocabulary (i) is finite with less than 10 4 motifs, (ii) mediates specific and non-redundant interactions between paratope-epitope pairs, (iii) is immunity-specific (distinct from the motif vocabulary used by non-immune protein-protein interactions), and (iv) enables the machine learning prediction of paratope or epitope. The discovery of a vocabulary of paratope-epitope interaction demonstrates the learnability and predictability of paratope-epitope interaction.

Current challenges for epitope-agnostic TCR interaction prediction and a new perspective derived from image classification

“…This necessitates relevant feature encoding. Indeed, studies have shown that feature engineering is essential for the prediction of paratope-epitope binding (22) and that it can also improve the performance of models for more general bioinformatics problems (23). Moreover, when complex machine learning methods are applied to small datasets, the dangers of overfitting and memorization of examples pose an even bigger threat.…”

Using B cell receptor lineage structures to predict affinity

“…from sequence data alone, is much more difficult. It would involve de novo structure and binding prediction, which are not currently practical, although much recent work focuses on these problems [56][57][58][59][60].…”

Modern Hopfield Networks and Attention for Immune Repertoire Classification

“…Each simulated repertoire was then randomly assigned to either the positive or negative class, with 2, 500 repertoires per class. In the repertoires assigned to the positive class, we implanted motifs with an average length of 4 AAs, following the results of the experimental analysis of antigenbinding motifs in antibodies and T-cell receptor sequences by (Akbar et al, 2019). We varied the characteristics of the implanted motifs for each of the 18 datasets with respect to the following parameters: (a) ρ, the probability of a motif being implanted in a sequence of a positive repertoire, i.e.…”