2024 AIChE Annual Meeting

(169x) Generalizability of Machine Learning Derived Interatomic Potentials of Peptides and Isomeric Structures

Authors

Joshi, N., University of Washington
Pfaendtner, J., University of Washington
Biomolecular simulations are typically performed with classical force fields or using ab initio molecular dynamics (AIMD). Accuracy at the electronic level is lost in the former, while computational time limits the scalability of the latter.1,2 Machine learning potentials (MLPs) trained on AIMD data offer higher accuracy results while overcoming the computational limitations of AIMD.3 In this study we explore the generalizability of MLPs trained on AIMD data obtained from the density functional theory (DFT) version of Car-Parrinello Molecular Dynamics for alanine dipeptide and its peptoid isomer disarcosine. We demonstrate the computational speed up and ability of these MLPs to capture DFT level accurate simulations with a limited training set. We illustrate the versatility of MLPs to simulate peptide and peptoid systems. These results demonstrate how with a limited dataset a sufficient MLP can be developed to capture DFT level accuracy and generalizability across different simulation parameters of interest like temperature. Furthermore we demonstrate the capability of MLPs to capture rare events of the systems and obtain their free energy profile. This work contributes to the production of faster simulations with increased accuracy and generalizability and can be used to extend the types of biomolecules studied.

References:

  1. D. Marx and J. Hutter, Ab initio Molecular Dynamics: Basic Theory and Advanced Methods (Cambridge University Press, 2009).
  2. Jörg Behler; Perspective: Machine learning potentials for atomistic simulations. J. Chem. Phys. 7 November 2016; 145 (17): 170901.
  3. Chmiela, S.; Sauceda, H. E.; Müller, K.-R.; Tkatchenko, A.Towards Exact Molecular Dynamics Simulations With Machine-learned Force Fields.Nat. Commun.2018,9, 3887.