Molecular Latent Space Simulators for Distributed and Multimolecular Trajectories

Abstract

All atom molecular dynamics (MD) simulations offer a powerful tool for molecular modeling, but the short time steps required for numerical stability of the integrator place many interesting molecular events out of reach of unbiased simulations. The popular and powerful Markov state modeling (MSM) approach can extend these time scales by stitching together multiple short discontinuous trajectories into a single long-time kinetic model but necessitates a configurational coarse-graining of the phase space that entails a loss of spatial and temporal resolution and an exponential increase in complexity for multimolecular systems. Latent space simulators (LSS) present an alternative formalism that employs a dynamical, as opposed to configurational, coarse graining comprising three back-to-back learning problems to (i) identify the molecular system’s slowest dynamical processes, (ii) propagate the microscopic system dynamics within this slow subspace, and (iii) generatively reconstruct the trajectory of the system within the molecular phase space. A trained LSS model can generate temporally and spatially continuous synthetic molecular trajectories at orders of magnitude lower cost than MD to improve sampling of rare transition events and metastable states to reduce statistical uncertainties in thermodynamic and kinetic observables. In this work, we extend the LSS formalism to short discontinuous training trajectories generated by distributed computing and to multimolecular systems without incurring exponential scaling in computational cost. First, we develop a distributed LSS model over thousands of short simulations of a 264-residue proteolysis-targeting chimera (PROTAC) complex to generate ultralong continuous trajectories that identify metastable states and collective variables to inform PROTAC therapeutic design and optimization. Second, we develop a multimolecular LSS architecture to generate physically realistic ultralong trajectories of DNA oligomers that can undergo both duplex hybridization and hairpin folding. These trajectories retain thermodynamic and kinetic characteristics of the training data while providing increased precision of folding populations and time scales across simulation temperature and ion concentration.

Machine Learning in Soft Matter: From Simulations to Experiments

“…[ 110,111 ] Therefore, some alternative approaches to MSM have been proposed, such as methods based on latent space simulators. [ 112,113 ] This method uses three different deep learning architectures, which involve encoding molecular trajectories into a latent space, propagating low‐dimensional trajectories within this latent space, and decoding the latent space back to configuration space, to train on short, discontinuous trajectories and multiple‐molecule systems for trajectory prediction. Additionally, some studies have attempted to learn dynamic processes based on recurrent neural network.…”

Artificial Intelligence for Conjugated Polymers

“…In addition, the size and complexity of molecular dynamics or DFT simulation trajectories often pose a barrier to their analysis and are often overlooked or neglected. However, recent advances like the Grisanov Reweighting Enhanced Sampling Technique (GREST), Latent Space Simulators (LSS), and contributions by Vacher et al through Bayesian neural networks have shown potential to convert simulation trajectories into machine friendly and physically meaningful inputsenhancing collective variable discovery, configuration sampling, and property estimation . Although these methods have been applied to macromolecules as large as pentapeptides, their application to larger polymeric materials has not been explored.…”

A Deep Learning-Driven Sampling Technique to Explore the Phase Space of an RNA Stem-Loop

“…[31][32][33][34] DDMD is a deep learning-driven MD simulation technique that avoids the problems outlined above by finding undersampled states from an ensemble of running simulations without requiring the user to define a set of physical CV(s). 35,36 It does so by learning a lowdimensional latent representation [37][38][39] of the simulation data on the fly and treating it as a proxy for the CV. DDMD has recently been shown to enhance the conformational sampling of proteins and protein-ligand complexation.…”