Modeling of Peptides with Classical and Novel Machine Learning Force Fields: A Comparison.

Area of Science:

• Computational Chemistry
• Materials Science
• Biophysics

Background:

• Classical force fields (FFs) are established for molecular dynamics (MD) simulations of soft matter.
• Neural network (NN) based machine learning (ML) potentials offer a novel alternative for MD simulations.
• ML potentials learn atomic energies from their chemical environment, promising for novel compound exploration.

Purpose of the Study:

• To evaluate the performance of the ANI-2x ML potential against classical FFs (CHARMM27, GROMOS96 43a1).
• To assess the applicability of ML potentials for simulating soft matter systems, specifically water and peptides.
• To identify limitations and areas for improvement in current ML potentials for MD simulations.

Main Methods:

• Comparative molecular dynamics (MD) simulations.
• Testing of ANI-2x, CHARMM27, and GROMOS96 43a1 force fields.
• Simulation systems included bulk water, trialanine, and an α-aminoisobutyric acid 9-mer in vacuum and water.

Main Results:

• ANI-2x accurately reproduced the ordered water structure, comparable to *ab initio* MD.
• Peptide simulations showed similar secondary structure basins but differed in basin positioning and stability compared to classical FFs.
• Divergent sampled structures suggest limitations related to ML potential range or training data.

Conclusions:

• The ANI-2x ML potential demonstrates potential for MD simulations of soft matter, particularly for water.
• Discrepancies in peptide simulations highlight the need for further refinement of ML potentials.
• Findings provide insights for enhancing ML potentials for improved accuracy and broader applicability in soft matter simulations.