Tripeptide dynamics from empirical and machine-learned energy functions.

Area of Science:

• Computational Chemistry
• Spectroscopy
• Biophysics

Background:

• Molecular dynamics (MD) simulations are crucial for understanding molecular behavior.
• Accurate energy functions are essential for reliable MD simulations.
• Machine learning (ML) offers a promising approach to develop accurate energy functions.

Purpose of the Study:

• To present molecular dynamics simulations for tripeptides using machine-learned energy functions.
• To validate ML potentials against experimental spectroscopic data.
• To explore the conformational space and vibrational properties of tripeptides in different environments.

Main Methods:

• Development and application of machine-learned potential energy surfaces (ML-PESs).
• Molecular dynamics simulations in both gas and solution phases.
• Hybrid ML/Molecular Mechanics (ML/MM) simulations in explicit solvent.
• Comparison of simulation results with experimental spectroscopic data (e.g., amide-I vibrations, vibrational circular dichroism).

Main Results:

• ML-PESs trained on high-level quantum chemical data achieved quantitative agreement with experimental vibrational splittings for AAA tripeptides.
• ML/MM-MD simulations accurately reproduced experimental amide-I splitting in solution for AAA.
• ML-PESs were developed for both zwitterionic and neutral forms of AMA, revealing red-shifted NH- and OH-stretch spectra due to cyclization and H-bonding.
• Simulations provided a consistent interpretation for experimental vibrational circular dichroism data of AMA.

Conclusions:

• Machine-learned potentials enable stable, quantitative, and meaningful MD simulations for hydrated tripeptides on nanosecond timescales.
• ML-PESs provide valuable insights and an interpretive framework for experimental spectroscopic studies of peptides.
• This approach demonstrates the feasibility of using ML for complex molecular simulations in biophysical research.