Design in Voxel Space Decode in SMILES Space: Plixer Generates Drug-Like Molecules for Protein Pockets

Abstract

We introduce Plixer, a two-stage generative model for de novo drug design that generates small-molecule ligand binders conditioned on an empty protein pocket. Plixer combines a conditional voxel inpainting network to generate 3D ligand hypotheses with an independently trained voxel-to-SMILES decoder that translates these voxel representations into valid chemical structures. By de-coupling the learning of spatial protein-ligand interactions from the learning of chemical grammar, our approach leverages large libraries of 3D ligand conformers to augment the limited data available for protein-ligand complexes. We show that this approach generates molecules with higher predicted binding affinity than recent methods.