| Summary: | Predicting how antibodies bind to their targets is a fundamental problem of immunology, and a critical step in accelerating the development of vaccines and therapeutics against foreign pathogens. In particular, the task of predicting the 3D structure of an antibody-target complex, otherwise known as docking, is an important tool in drug design, providing valuable insights such as ways to increase antibody potency or methods to limit the likelihood of a mutational escape. State of the art models of antibody docking treat the task as a regression problem, outputting a single prediction. We hypothesized that while the performance after a single try might be poor, the likelihood of producing a good docking pose in 𝐾 tries could be significantly higher. To achieve this without having to alter existing docking models, we propose to first train a B-Cell epitope predictor and to subsequently use it to produce a diverse set of candidate binding sites. Our epitope predictor achieves state of the art performance, with an ROC-AUC score of 76. We then show that, by properly post-processing the epitope model’s predictions to select 𝐾 promising candidate docking sites, the success rate of a docking model on an independent test set can be increased by a factor of almost 10, with as little as 10 tries. Our approach is compatible with any docking model and offers an alternative to pure generative modeling, while being able to guarantee a diverse set of solutions, without the need to leverage complex sampling strategies.
|
|---|