Learning reaction coordinates via cross-entropy minimization: Application to alanine dipeptide.

Area of Science:

• Computational Chemistry
• Molecular Dynamics
• Machine Learning

Background:

• Identifying reaction coordinates is crucial for understanding complex molecular systems.
• Existing methods may struggle with high-dimensional collective variables.

Purpose of the Study:

• To propose a novel cross-entropy minimization method for reaction coordinate identification.
• To extend likelihood maximization by fitting the committor function with a sigmoid.
• To incorporate L2-norm regularization for improved model stability with numerous variables.

Main Methods:

• Cross-entropy minimization applied to collective variables.
• Sigmoidal fitting of the committor function (pB*).
• L2-norm regularization and cross-validation for parameter selection.
• Application to alanine dipeptide isomerization using 45 dihedral angles.

Main Results:

• The method successfully identified a reaction coordinate involving key dihedral angles.
• Results are consistent with prior studies on alanine dipeptide isomerization.
• Identified points with pB* ~ 0.5 act as a clear separatrix between reactant and product states on the potential of mean force.

Conclusions:

• The proposed cross-entropy method effectively determines reaction coordinates from high-dimensional data.
• The method provides a robust way to analyze molecular dynamics simulations and understand reaction pathways.
• This approach enhances the analysis of complex molecular transformations.