Speaker: Trond Linjordet
Title: Deep Generative Modeling of Ligands and Metal Complexes
Abstract: Transition metal complexes (TMCs) comprise compounds used in industrial chemical catalysis, in the energy transition, and as pharmaceutical products. The task of discovering of novel, promising TMCs poses an inherently more complex challenge than typical drug discovery, which considers only organic molecules. This talk presents work where deep neural models are trained with TMC data from the tmQM dataset series, which in turn is based on the Cambridge Structural Database. The focus is primarily on work now submitted for review, using the junction-tree variational autoencoder (JT-VAE) deep neural architecture to learn to generate transition metal ligands. This project investigated both unconditional generation via sampling the learned latent space, as well as conditional generation via optimization of initial SMILES-based prompts towards desired target properties. The target properties included both TMC-level properties and ligand-level properties. Thousands of novel and diverse monodentate and bidentate ligands were generated and assessed. The talk will also briefly overview the early stages of a subsequent project using denoising diffusion probabilistic models (DDPMs) to generate whole-TMC geometries.