Physics-Inspired Generative Deep Learning with Application to RNA Structure Prediction

Abstract

Biomolecular structure is often treated as a single native conformation, but many molecular systems are more accurately described as thermodynamic ensembles of interconverting structures. This thesis develops a probabilistic generative modeling framework for learning such ensembles from limited molecular data. The central premise is that generative models are most effective for molecular science when physical inductive biases are included. The thesis begins by benchmarking existing classes of probabilistic generative models on molecular data and distributions with tunable dimensionality. These comparisons identify denoising diffusion probabilistic models as especially capable for the moderate-dimensional, multimodal landscapes common in molecular systems. Building on that result, the next study shows that diffusion models trained on replica exchange molecular dynamics data can extrapolate to temperatures not present in the training set. The core methodological contribution, the Thermodynamic Maps framework, represents physical temperature in a diffusion model. This adjustment enables precise thermodynamic extrapolation from sparse data and is validated on both the Ising model and RNA conformational sampling. The framework is then extended to \textit{ab initio} RNA ensemble prediction in RNAnneal, where hierarchical structure generation, thermodynamic scoring, and physics-grounded ranking are combined to recover three-dimensional RNA conformational ensembles from sequence alone. A complementary application, DiffDock-Glide, demonstrates that a similar physics- and data-driven pipeline improves molecular docking and virtual screening outcomes. Together, these studies demonstrate that generative models can extrapolate from highly limited data when provided physical and chemical inductive biases.