Pedro O. Pinheiro
Affiliation
Prescient Design / Genentech
Talk Title
All-atom structure-conditioned molecule generation with voxels and neural fields
Abstract
In this presentation, I will talk about how we can perform structure-conditioned generation by representing molecules as atomic density fields, either by discretizing the space (voxel grids) around molecules or modeling it implicitly (neural fields). These representations allow us to use very expressive architectures from the computer vision community, known to work well in practice. They are also flexible, allowing us to model different molecular systems with the same model. We show the effectiveness and versatility of these representations with (in silico) experiments on three generation tasks: small molecules or macro-cyclic peptides conditioned on protein pockets and antibody CDR H3 redesign.
Bio
Pedro is a principal research scientist in the Frontiers Research team at Prescient Design/Genentech. His current interests include representation learning, generative models and applications to drug discovery. Prior to Genentech, Pedro got a PhD in computer vision at EPFL in Switzerland and worked in two Canadian start-ups: element AI and deep genomics.