Adaptively Iterative Multiscale Switching Simulation Strategy and Applications to Protein Folding and Structure Prediction

Abstract

Structure prediction is an important means to quickly understand new protein functions. However, the prediction of effects of proteins that have no detectable templates is still to be improved. Molecular dynamics simulation is supposed to be the primary research tool for structure predictions, but it still has limitations of huge computational cost in all-atom (AA) models and rough accuracy in coarse-grained (CG) models. We propose a universal multiscale simulation strategy named AIMS in which simulations can iteratively switch among multiple resolutions in order to adaptively trade off AA accuracy and CG high-efficiency. AIMS follows the idea of CG-guided enhanced sampling so that final results always keep AA accuracy. We successfully achieve four ab initio and four data-assisted protein structure predictions using AIMS. The prediction result is an ensemble rather than a structure and provides special insights on folding metastable states. AIMS is estimated to achieve a computational speed about 40 times faster than that of conventional AA simulations.

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

“…This class of computational methods, in use since at least 1972, − generally requires (a) a scale that allows faster, larger, or less computationally demanding simulations; (b) a scale that provides enhanced accuracy, flexibility, or relevant details; (c) a mechanism to couple the scales; and (d) multiscale consistency, without which conflicts may drive the joint multiscale behavior to poorly represent constituent models . Multiscale CG/AA simulations have been used to study lipid membranes, protein dynamics, , and the effects of protein crowding, among many other applications. , Approaches for CG-to-AA scale conversion have been recently reviewed and include both rule-based procedures like Backward and CG2AT2, and a topology-free machine learning (ML) method called GLIMPS . Alternative approaches to accelerate simulations while retaining atomic resolution for molecules of interest include hybrid simulations whose Hamiltonians directly mix CG and AA components − (which is conceptually similar to quantum mechanical/molecular mechanics approaches , ) and adaptive resolution simulations, in which Hamiltonians can change on the fly. − …”

“…This class of computational methods, in use since at least 1972, 21−25 generally requires (a) a scale that allows faster, larger, or less computationally demanding simulations; (b) a scale that provides enhanced accuracy, flexibility, or relevant details; (c) a mechanism to couple the scales; and (d) multiscale consistency, without which conflicts may drive the joint multiscale behavior to poorly represent constituent models. 26 Multiscale CG/AA simulations have been used to study lipid membranes, 27 protein dynamics, 28,29 and the effects of protein crowding, 30 among many other applications. 31,32 Approaches for CG-to-AA scale conversion have been recently reviewed 33 and include both rule-based procedures like Backward 34 and CG2AT2, 28 and a topology-free machine learning (ML) method called GLIMPS.…”

Ribonucleic Acid Folding Prediction Based on Iterative Multiscale Simulation

“…We ran the targeted MD simulation to generate a local double-helix stem. Previously, sampling small protein conformations with the AIMS model was performed directly by running AAMD simulations; however, RNA secondary structure is harder to generate, and the simulation time is likely to need >1 ms. Dynamic simulations for such a long time are almost impossible for current computers, so we started the molecular dynamics simulation from the possible conformations generated with targeted MD simulation, which also provided a solid foundation for the subsequent CG force field fitting.…”

“…47 However, the resulting conformation may be unreasonable, even with high local energies. On the basis of the AIMS 48 multiscale simulation strategy, taking into account the properties of the three-bead model and RNA base pairing, the back-mapping was performed in two steps. In the first step, we ran ABMD combined with AA/virtual site (AA/VS) technology.…”

“…The base pairing of other parts is mainly generated and repaired by molecular dynamics simulation, including noncanonical base pairs. On the basis of the multiscale simulation strategy AIMS, 48 combined with the properties of RNA, we extend it to RNA folding. The flowchart is shown in Figure 3.…”