Optimizing Drug Activity Using Docking-Informed Machine Learning

Bayesian optimization (BO) has been widely used to optimize drug potency towards a protein target by searching through fixed libraries of molecules and selecting compounds guided by machine learning (ML) predicted activities. We find that by pre-computing docking scores for the full library, and by using estimated binding energies and 3D descriptors from docking as features in the ML model, BO can be significantly accelerated relative to using traditional 2D features such as molecular fingerprints. Additionally, we have observed that a docking-based initialization scheme is often superior to the commonly-used diversity-based or random initialization. Using docking-based features and docking-based initialization required on average 24% (up to 77%) fewer data points to find the most active compound and gave on average 32% (up to 159%) improvement in enrichment factors, relative to a typical BO approach. We applied our method to 14 ChEMBL data sets, and 4 more highly challenging LIT-PCBA data sets with low hit rates and high molecular diversity. Our approach combines the generality of structure-based virtual screening (SBVS) with the inference power of ML ligand-based virtual screening (LBVS) to offer a more data-efficient hybrid approach.