Optimization of Generator Reward Functions Targeting Non-covalent KRAS Inhibitors

While machine learning (ML) and artificial intelligence (AI) technologies for molecular generation have seen rapid progress in the last five years, there is still a need for careful set up and optimization of these tools. This study documents the optimization of reward functions for a reinforecement learning (RL)-based generator targeting non-covalent inhibitors of KRAS. In particular, we study the relative effect of additive versus multiplicative reward components in a multiobjective function using docking environments and pharmaocphore modeling. Our findings illustrate how to maximize the outcome of a target criterion, using the number of molecules that meet or surpass a target docking score. In the process, we find an apparent compounding effect of docking and pharmacophore modeling on the generators’ results. From these results, we identify and assess a promising non-covalent inhibitor candidate molecule based on docking score and likely interactions across a variety of KRAS mutants.