Massively Parallel Machine Learning Driven Multiscale Simulations to Explore RAS-RAF Biology

Abstract

Presented by H.I. Ingólfsson for the Joint Design of Advanced Computing Solutions for Cancer, Pilot 2 Team. Molecular dynamics simulations can provide detailed insights into biological mechanisms. However, even on the largest supercomputers available today micro-scale simulations cannot reach experimentally relevant time- and length-scales which significantly lessens their potential impact. Yet, macro-scale simulations that can cover experimental conditions normally do not provide the necessary detail to directly observe phenomena such as protein bonding or signaling. The multi-resolution framework uses machine learning, to couple a macro-scale model at experimentally relevant time- and length-scales with a vast ensemble of successively higher resolution micro-scale simulations. Given sufficient resources, the system will converge to macro-scale results at micro-scale resolution and provide fundamentally new insights. Here we present an extension to the massively parallel Multiscale Machine-Learned Modeling Infrastructure (MuMMI) incorporating the third (all-atom) resolution scale, multilevel feedback, higher fidelity macro model and support for different protein types. We are currently running a large-scale simulation campaign of RAS and RAF proteins on a complex plasma membrane using Summit, the second fastest supercomputer in the world, and will present results on the performance of the infrastructure and interactions of RAS, RAF and the lipid bilayer. This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory under Contract DE-AC52-07NA27344. Release number: LLNL-ABS-815242.