A major application of deep generative models in drug discovery aims at designing chemical structures exploring the space, and optimizing where possible, a promising scaffold. Traditional deep generative models, optimizing without any scaffold constraints, typically fail to generate a large ratio of molecules with the desired scaffold, thus hindering the practical application of these models for lead-optimization. To address this limitation, we designed a de novo design method that implements a scaffold-constrained, fragment-based molecular generator featuring single- and multiple- R-group optimization. Our method uses the well-established SMILES-based Recurrent Neural Network (RNN) generative model to sample and combine different types of fragments (R-groups or linkers) to generate molecules that meet the scaffold-constraint. The resulting generative method is coupled with multi-objective optimization available via a composite or pareto scoring function within a reinforcement learning framework. During training, learning is achieved through fragment-scaffold and fragment-fragment correlations to improve physicochemical and biological properties of interest in a multi-objective setting. In this presentation we describe results from the application of the method in various scaffold-constrained scenarios. Results obtained are evaluated using different metrics and compared with other state-of-the-art deep generative models.

Scaffold-constrained Molecular Generation with Multi-objective Optimization

Fall 2022 abstracts, posters, and presentations will launch in mid-September.

ACS Fall 2022

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