NanoBEP - A Machine Learning Based Tool for Nanobody Binding Energy Prediction

Nanobody is a special class of antibodies comprising only one variable heavy chain. Its small size and high stability over a wide range of temperature and pH, makes it an ideal candidate for biomedical applications. Designing a nanobody that can bind to a specific target protein, either for therapeutic or diagnostic purposes, requires a quick estimation of binding affinity of nanobody-protein complex. Many predictive models for protein-protein interactions have been developed leveraging the capability of machine learning techniques. The popular protein-protein interaction models, however, could not accurately predict the binding affinity of available nanobody-protein complexes. We, therefore, have developed a random forest based model that can predict the value of dissociation constant (log10 Kd) at high accuracy with a Pearsons correlation coefficient value of 0.95 and a mean absolute error of 0.44. Our cherry-picked model identifies the best protein features for the prediction through two stages of selection strategy that includes elimination of highly correlated features through graph network analysis, followed by the recursive feature elimination through random forest. Despite being a class of antibodies, a model trained only on antigen-antibody complexes couldnt accurately predict the binding affinity of the nanobody-protein complexes. The predictability improved only when we included the data on monomeric protein complexes and some nanobody-protein complexes during training.