Deep generative models for molecular dynamics and design

Abstract

This thesis explores how Deep Generative Models (DGMs) can accelerate molecular modeling tasks central to drug discovery by addressing conditional sampling problems. It consists of four studies, the three first focusing on molecular dynamics (MD), and the last on molecular design. The first paper introduces Surrogate Model-Assisted Molecular Dynamics (SMA-MD), which combines a DGM with statistical reweighting and short MD simulations to efficiently sample Boltzmann ensembles of small molecules, producing more diverse and lower-energy configurations than conventional simulations. The second paper presents Transferable Implicit Transfer Operators (TITO), a transferable generative surrogate that learns time-integrated molecular dynamics directly from data, enabling propagation at arbitrarily large time steps with up to four orders of magnitude acceleration while maintaining thermodynamic and kinetic fidelity. The third paper, Boltzmann Priors for Implicit Transfer Operator learning (BoPITO), introduces equilibrium-aware priors to surrogate models of MD, improving data efficiency and long-term dynamical accuracy. Finally, the fourth paper develops a reinforcement learning scheme to fine-tune graph-based DGMs for \textit{de novo} molecular design, guiding models toward molecules with desired properties even when such examples are rare or absent in the training data. These contributions constitute important stepping stones towards the automation of the drug discovery process.