Combining physics-based and machine-learning methods for de-novo drug design

Abstract

This thesis presents a proof-of-concept for a novel de-novo drug design algorithm that uses forcefield parameters to generate molecules in 3D space directly in the active site of a target. The algorithm efficiently samples possible molecules and their bound conformations using an approach inspired by Configurational-Bias Monte Carlo (CBMC). It is wholly atomistic and strings together atoms to construct the final molecule and uses forcefield interaction parameters to find the optimal binding partner for the target. The atom types used are parameterized in CHARMM-27 and are well-established. We have previously validated the algorithm's accuracy in predicting strong binders through rigorous free-energy calculations. Adding to this physics-based approach, we use reinforcement learning to bias the atom type selection towards making molecules synthesizable using SYBA, an established classifier for predicting whether a molecule is synthesizable. The program shows good results by generating a diverse set of synthesizable molecules for streptavidin and HSP90, which are our test systems. The algorithm can also suggest modifications to existing ligands, thus allowing it to inspire ligand affinity improvement through minor modifications.