Ligand-based drug design (LBDD) utilizes similarity (in 2D-, 3D- or descriptor space) to existing ligands on a project to determine the fitness of new candidate ligands. While 3D similarity-based metrics have been extensively utilized in virtual screening, 3D similarity measures have not been integrated as extensively into the de novo design via generative modeling. Here, we describe how de novo compound generation can be guided by ROCS shape- and chemical feature-based similarity as the metric. As proof of concept, we incorporate this metric into two published generative chemistry models, the reinforcement-learning method REINVENT and the evolutionary algorithm-based method Argenomic. We (1) demonstrate the feasibility of generating compounds close to the bioactive conformers of several reference compounds and (2) apply these generative models for scaffold hopping by generating compounds while replacing user-defined segments of those bioactive conformers. We also discuss the critical role of reward shaping in the context of both reinforcement learning and the fitness functions for evolutionary algorithms in guiding the generation of compounds with desirable scaffold diversity.

Reinforcement learning with 3D reward fnctions

ACS Fall 2021

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.