One of the limitations to the success of virtual screening is the breadth and quality of the compound library in available databases. We propose to address this problem using a family of neural network architectures known as variational autoencoders (VAEs), which has recently achieved significant results in the discovery of novel chemical structures with favorable chemical and pharmaceutical properties. VAEs can yield new active molecules and identify leads that may not have been in the original library at a fraction of the computational costs of virtual screening and physics-based simulations.<br/><br/>Using a syntax-aware grammar variational autoencoder (GVAE), we characterize known active molecules in the Pfizer compound and encode the process adopted by medicinal chemists in optimizing lead activity as a learnable sequence of geometric transformations in latent space. We demonstrate that applying these learned transformations, in combination with Bayesian optimization techniques, to new leads enables the directed expansion of our compound collection and the systematic discovery of new molecular structures with favorable properties.

Leveraging generative models for the directed expansion of compound libraries

256th National Meeting (2018)

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