DoRIAT: A Bayesian framework for interpreting and annotating docking runs

The advent of sequence-to-structure deeplearning models have transformed protein engineering landscape by providing an accurate and cost effective way to determine crystal structures. Despite their accuracy, deep-learning predictions tend to give limited insights around protein dynamics. To improve conformation exploration we have developed a machine learning pipeline that combines deep-learning predictions with molecular docking. In this report, we propose Do cking R un I ntepretation and A nnotation T ool (DoRIAT). In contrast to frameworks that score models based on interface interactions, DoRIAT uses a set of parameters that summarize binding conformation. We use DoRIAT to score output from docking runs, identify complexes close to the native structure and create ensembles of models with similar binding conformations. Our results demonstrate that the single structural model DoRIAT selects to be the closest representation of the crystal structure lies within the top 10 of docked models, ranked by Root Mean Squared Distance(RMSD), in around 80% of cases.