GTC Silicon Valley-2019: Deep Generative Models for Computational Drug Discovery
GTC Silicon Valley-2019 ID:S9699:Deep Generative Models for Computational Drug Discovery
David Koes(University of Pittsburgh)
Generative modeling is a new paradigm for molecular design that translates the growing amount of biomolecular data into an efficient method of predicting novel drug candidate molecules called ligands. We've applied convolutional neural networks to a space-filling atomic density representation of molecular structures to perform protein-ligand scoring and molecular docking. We'll describe our recent work with 3D generative models of molecular structure that generate ligand atom densities conditioned on a receptor-binding site. Our dual-encoding network architecture allows us to interpolate in the protein space and the ligand space. We'll show that our models successfully generate ligand densities conditional on a given receptor, discuss the challenges in mapping a continuous latent space to discrete chemical space, and explain our approaches to latent space regularization.