Sampling from molecular unnormalized distributions with Deep Generative Models

Licentiatavhandling, 2024

This thesis investigates how Deep Generative Models (DGMs) can address important drug discovery problems involving sampling from unnormalized
distributions. It consists of two papers focusing on this challenge’s aspects: molecular design and conformational sampling. The first paper proposes a new training scheme to fine-tune graph-based DGMs for de novo molecular design. Our method can produce molecules with specific properties even when they are scarce or missing in the training data and outperforms previously reported graph-based methods on predicted dopamine receptor type D2 activity while maintaining diversity. The second paper develops Surrogate Model-Assisted Molecular Dynamics (SMA-MD), which combines a DGM with statistical reweighting and short Molecular Dynamics simulations to generate equilibrium ensembles of molecules. SMA-MD can produce more diverse and lower energy ensembles than conventional molecular dynamics simulations. These contributions constitute important stepping stones towards the automation of the drug discovery process.