Generalizable coordination of large multiscale workflows

Abstract

The advancement of machine learning techniques and the heterogeneous architectures of most current supercomputers are propelling the demand for large multiscale simulations that can automatically and autonomously couple diverse components and map them to relevant resources to solve complex problems at multiple scales. Nevertheless, despite the recent progress in workflow technologies, current capabilities are limited to coupling two scales. In the first-ever demonstration of using three scales of resolution, we present a scalable and generalizable framework that couples pairs of models using machine learning and in situ feedback. We expand upon the massively parallel Multiscale Machine-Learned Modeling Infrastructure (MuMMI), a recent,

Asynchronous Reciprocal Coupling of Martini 2.2 Coarse-Grained and CHARMM36 All-Atom Simulations in an Automated Multiscale Framework

“…Subsequent translation of the GROMACS input file to AMBER format takes an additional ∼0.6 h on a single core (Figure D), resulting in an average time of ∼1.6 h for converting and preparing one CG snapshot to run at the AA scale using the Backmapping module. By design, this module runs on CPUs only, allowing the MuMMI workflow to maintain a pool of ready-to-run AA simulations using otherwise unused CPU resources while occupying nearly all GPUs for concurrently running AA and CG simulations …”

“…To incorporate an AA scale, we introduced new MuMMI modules (highlighted in magenta): The Backmapping module, which converts a CG simulation frame to AA simulation utilizing sinceCG, the AA Simulations and Analyses module, which executes and monitors equilibration and production AA simulations and performs on-the-fly analysis, and the AA-to-CG Feedback module, which aggregates analysis from all running simulations to update CG parameters. Part (B) adapted from ref .…”

“…After initial processing, a modular and extensible series of analyses are performed, including calculation of the root-mean-square deviation (RMSD) of C α atoms from their initial positions in the RAS HVR and the RAF CRD (independently and combined), the end-to-end distance across the HVR, the orientation of the RAS G-domain, and the value of d z . Analysis results are stored per frame in NumPy array format using MuMMI’s pytaridx interface. , In addition, when CRD loops are sufficiently close to the membrane ( d z < 3 nm), protein coordinates are passed to the AA-to-CG Feedback module in standard PDB file format.…”

“…The sampling approach used here extends the DynIm sampling framework 55 to balance randomness and novelty and to work directly in this three-dimensional space. All selected CG frames are scheduled by the MuMMI workflow 51 for CG-to-AA transformation by the Backmapping module (Figure 1B). This module retrieves the selected snapshot from the Martini CG ddcMD 77 trajectory via a customized MDAnalysis tool with ddcMD trajectory support, 83 calls sinceCG to perform the CG-to-AA transformation (see below), converts input files from GROMACS format to AMBER format using ParmEd, 84 and logs the status of the transformation.…”

“…This manuscript describes in detail the addition of a third (AA) scale to the Multiscale Machine-Learned Modeling Infrastructure (MuMMI). − In MuMMI, a quasi-two-dimensional (quasi-2D) continuum model uses dynamic density functional theory (DDFT) to rapidly sample large-scale fluctuations in the local concentrations of lipid mixtures around diffusing point definitions of membrane proteins. Diverse local arrangements of lipids and proteins that emerge in DDFT simulation are automatically resampled to construct an ensemble of smaller systems that can be simulated with enhanced precision, and later Boltzmann-weighted .…”