(622b) Multi-Multimolecular Latent Space Simulations of DNA Hairpin-Duplex Competition

As computational resources continue to expand and force fields become more exact, molecular dynamics simulations will inevitably be applied to larger and more complex systems. Increasingly, these systems involve multiple components whose dynamics are coupled when in bound or interacting states and are otherwise entirely independent. Traditional kinetic modeling methods such as Markov state models (MSMs) have been applied to these multi-molecular systems but can fail to capture an exponentially increasing number of often degenerate states. We previously developed latent space simulators (LSS) as a continuous approach to low-dimensional kinetic modeling that (i) encode a molecular system into a slow latent space, (ii) propagate dynamics in this latent space, and (iii) generatively decode a synthetic molecular trajectory. Here we expand on the LSS framework to develop a model, specialized toward multi-multimolecular systems, which learns the dynamics of each molecule independently and varies the reconstruction approach as a function of intermolecular distance and orientation. We apply our model to a coarse-grained DNA system consisting of oligomers that can both hybridize into a duplex and fold into unimolecular hairpins. Our model generates ultra-long synthetic molecular trajectories that reproduce the thermodynamics and kinetics of molecular dynamics data and lends insights into competition between duplex and hairpin states. This work lays the foundations of the approach for applications to other multi-molecular systems such as protein dimerization, DNA-protein binding, and protein-ligand interactions.