Multi-objective molecule optimization with graph history and transfer refinement
Elton L. Cao · 2023
Deep generative de novo drug discovery models have shown promise as a novel path for molecular development. Many objectives, including property and protein optimization, have been effectively implemented for discovering drugs with multiple objectives. De novo models attempt to generate molecules completely from scratch; however, oftentimes chemists may need to optimize a given lead drug by generating analogs. Therefore, in this work, we propose a molecule optimization pipeline consisting of analog-based (ASB) scaffolds and a graph generative variational autoencoder (VAE). In our approach, molecules can be optimized through multiple objectives, including molecular property refinement through conditions and protein target optimization through a transfer refinement cycle. Additionally, we also implement a graph history network as well as a cyclic annealing schedule to control the degree of uniqueness within the molecule.