Bayesian optimization for small molecule hits generation
Pavel Yakovlev · 2018
In this talk we present the hit generation platform, that is based on direct evolution of ligands set with docking-based survival scoring. Starting with a fully random-generated set of small molecules we score them by docing score, ADME(T) parameters and poses to estimate specified objective function. Then we use Bayessian approach to select a number of support points and recreate a new generation using them. We use generative models based on RNN-autoencoders to create an internal representation of molecules by their SMILES strings and then use this representation to create novel molecules with target properties similar or increased.