Published December 2026 | Version v1
Dissertation Open

Representation and Reinforcement Learning for Molecular Dynamics

  • 1. ROR icon University of Chicago

Description

Molecular dynamics simulations can describe biomolecular motion and kinetics at atomic resolution, but the resulting trajectories are difficult to analyze and may contain few transitions between long-lived metastable states. This dissertation studies two coupled problems: how molecular conformations should be represented to allow automatic analysis and how this representation enables efficient sampling of important conformational space.

Denoising pretraining on small-molecule structures provides all-atom geometric graph features that can be reused for molecular and protein learning. The scaling experiments show that greater model size does not reliably predict better transfer; architecture, conformational coverage, and the evaluation split also matter. SubFormer addresses computational cost and long-range communication by applying attention to chemical fragments after local atomic message passing. For trajectory analysis, geom2vec freezes the pretrained encoder in SubFormer, stores its features, and trains token mixers with time-lagged objectives. Analyses of fast-folding proteins test whether these features retain side-chain and folding information. A token-merging extension further reduces computational cost by combining sequence, graph, and attention updates, extending the analysis to large biomolecular complexes. Together, these methods separate the cost of geometric feature extraction from repeated downstream training while retaining structural detail for kinetic analysis.

Time-lagged objectives connect these representations to dynamical questions. Slow-coordinate and predictive-state models describe motion recorded in trajectories, while committor analysis addresses the probability of reaching one endpoint before another. These quantities require validation on temporally separated data or independent trajectories: a structural embedding alone does not establish a transition pathway or determine a rate.

GESPO uses group relative policy optimization and a geom2vec model to allocate a simulation budget toward informative conformational space. A generative policy proposes starting conformations for short simulations, and trajectory-level rewards guide updates between groups of simulations. Peptide and protein–ligand applications evaluate state discovery and the dynamical information recovered from the collected segments. Because the starts are chosen adaptively, pooled frame counts do not directly measure equilibrium populations. The dissertation therefore distinguishes representation quality, observed transition information, and the statistical assumptions needed to infer kinetic observables. It presents a closed-loop approach in which learned representations support kinetic inference, and kinetic feedback guides the collection of new trajectories.

Files

Dissertation_revised_2026-10-07.pdf

Files (75.2 MB)

Name Size Download all
md5:29d4874c77a222b72f535992e16db3e4
75.2 MB Preview Download

Additional details

UChicago Information

Division(s)
Physical Sciences Division
Department(s)
Chemistry