(389j) Latent Space-Driven Molecular Simulations and Markov State Models for Uncovering RNA Folding/Unfolding Thermodynamics and Kinetics

RNA stem-loop folding and unfolding are fundamental processes that underlie the relationships between RNA structure and its function. However, simulating the folding and unfolding dynamics remains a computational challenge due to RNA’s rugged folding landscape. To address this, we adapted DeepDriveMD (DDMD), a deep learning framework, to accelerate the sampling of RNA folding/unfolding via molecular dynamics (MD) simulations1. DDMD employs an autoencoder to adaptively generate a compact latent representation from ensembles of ongoing MD simulations. Using this latent space, DDMD directs simulations toward undersampled regions of the conformational landscape, ensuring computational resources are focused on exploring the most relevant areas of phase space, thereby drastically reducing the computational cost to generate converged results. We used this approach to determine the folding landscape and pathways of RNA tetraloops having GNRA motifs, which are one of the most frequently occurring sequence patterns in rRNA hairpins2. The method generates reasonable free energy estimates at room temperature1. The latent space embeddings generated on the fly provide low-dimensional yet information-rich features that serve as the basis for constructing Markov State Models (MSMs), enabling a detailed quantitative investigation of RNA folding kinetics. Unlike traditional local order parameters, latent features retain essential structural information while dramatically reducing dimensionality, accelerating sampling, and improving the quality of kinetic rates. We identified metastable states, estimated transition rates, and applied transition path theory to reveal that the folding and unfolding pathways of RNA stem-loops are symmetric. This integration of latent embeddings with MSMs allows us to rapidly uncover transition rates with high accuracy at room temperature. Overall, our workflow provides a powerful new paradigm for dissecting RNA folding landscapes and kinetics, offering both methodological advances in kinetic modeling and novel insights into RNA structural dynamics.

1. Gupta, A., Ma, H., Ramanathan, A., & Zerze, G. H. (2024). A Deep Learning-Driven Sampling Technique to Explore the Phase Space of an RNA Stem-Loop. Journal of Chemical Theory and Computation, 20(20), 9178-9189.

2. Woese, C. R., Winker, S., & Gutell, R. R. (1990). Architecture of ribosomal RNA: constraints on the sequence of" tetra-loops". Proceedings of the National Academy of Sciences, 87(21), 8467-8471