For de novo molecule generation, what specific deep learning architecture is used to map discrete molecular graphs into a continuous latent space for optimization?
The specific deep learning architecture used to map discrete molecular graphs into a continuous latent space for optimization is the Variational Autoencoder, often referred to as a VAE. A molecule is discrete because it consists of specific atoms and fixed bonds that do not exist in intermediate states, while a continuous latent space is a mathematical representation where data points are mapped to coordinates in a multi-dimensional vector spa....
Community Answers
Sign in to open profiles and full community answers.
Lara Emad Aljudaibi
“The specific architecture is a Junction Tree Variationsl Autoencoder (JT-VAE), or more broadly a Graph Variational Autoenconder (Graph VAE) These architectures combine Graph Neural Network encoders with a Variational Autoencoder framework. The GNN maps non-Euclidean discrete molecular graphs into a continious low-dimensional latent space using probabilistic latent variables. Once embedded in this smooth manifold, gradient-based optimization algorithms can easily navigate the latent space to perform property optimization before a graoh decoder reconstructs novel chemical structures.”
71.0%