P07-31
Efficient Single Step Synthesizable Molecular Design using Wasserstein Autoencoder
Jinzhe ZHANG *1, Jiawen LI1, 3, Mizuki TAKEMOTO1, Ryuichiro ISHITANI1, 2, 4
1Drug Discovery, Preferred Networks Inc
2Department of Computational Biology and Medical Sciences, The University of Tokyo
3Division of Computational Drug Discovery and Design, Medical Research Institute,, Tokyo Medical and Dental University
4Department of Biological Sciences, The University of Tokyo
( * E-mail: jzhang@preferred.jp )
De novo molecular design algorithms address the inverse design problem by creating chemical structures that optimize a given set of desired properties. However, these generative models often overlook the synthetic accessibility of the generated candidates, leading to increased time and cost at the synthesis stage. Generative models that do account for synthetic accessibility frequently underperform in molecular generation tasks.
We propose a Wasserstein autoencoder-based generative model that designs chemical compounds with desired properties while ensuring all generated compounds can be synthesized using single step chemical reaction from a predefined set of possible reaction types, based on a given set of building blocks. By crafting a smoother latent space, we demonstrate that our method outperforms non-synthesizability-aware models in sampling efficiency while maintaining synthesizability. Additionally, our model surpasses existing synthesizability-aware models in terms of optimizing target properties.
This approach facilitates the screening of Bespoke library[1] and the rapid synthesis and testing of designed candidates during the quick prototyping stage of molecular research.
[1] Kaplan, Anat Levit, et al. "Bespoke library docking for 5-HT2A receptor agonists with antidepressant activity." Nature 610.7932 (2022): 582-591.