(657a) Deep Learning of Low-Dimensional Latent Space Molecular Simulators

The long-time microscopic evolution of molecular systems is governed by the leading eigenfunctions of the transfer operator that propagates the system through time. The low-dimensional latent space defined by these eigenfunctions parameterize the slow manifold to which the system is constrained to evolve. We have developed deep neural networks of various architectures to (i) learn the leading transfer operator eigenfunctions and project molecular configurations into this latent space, (ii) propagate the molecular dynamics within the encoded latent space, and (iii) decode the latent space back to the all-atom coordinate space. We train the networks over short simulation trajectories to establish low-dimensional latent space simulators capable of performing molecular dynamics simulations at orders-of-magnitude lower cost than traditional physics-based simulation. In an application to Trp-cage protein, we train latent space simulators over 200 µs of traditional physics-based simulation trajectories costing millions of CPU-h to generate. Our trained model can generate simulations of equal length in just 2.3 s of wall time on a single CPU core, and the synthetic trajectories are in excellent accord with the structure, thermodynamics, and kinetics of the original simulations: our all-atom configurations are lie within <5 Å RMSD and our free energy landscapes and kinetic rates are in quantitative agreement. The trained latent space simulators can generate thousands of folding and unfolding events in a matter of minutes, enabling us to quantify long-time kinetic phenomena at unprecedented statistical accuracy while maintaining a continuous-time, continuous-space, all-atom representation of the molecular dynamics.