Committors without Descriptors.

Area of Science:

• Computational chemistry and molecular dynamics.
• Statistical mechanics and enhanced sampling techniques.

Background:

• Rare events pose significant challenges in atomistic simulations.
• The committor function offers a formal description for rare events.
• Previous methods utilized committor-based approaches with neural networks and physical descriptors.

Purpose of the Study:

• To further automate and enhance committor-based methods for rare event simulations.
• To leverage graph neural networks for direct processing of atomic coordinates.
• To improve the sampling of transitions between metastable states and transition state ensembles.

Main Methods:

• Development of a committor-based enhanced sampling method.
• Iterative optimization of a neural network-based committor parametrization using a variational criterion.
• Integration of graph neural networks to process atomic coordinates directly, bypassing the need for physical descriptors.
• Application to benchmark systems and complex molecular processes.

Main Results:

• Successful automation of the committor-based rare event simulation procedure.
• Demonstration of graph neural networks' ability to directly use atomic coordinates.
• Improved sampling of transitions and transition state ensembles.
• Highlighting the advantages of graph-based methods in systems involving solvent molecules.

Conclusions:

• The proposed graph neural network-enhanced committor method offers a more automated and powerful approach to studying rare events in atomistic simulations.
• This method effectively captures the role of solvent molecules in processes like ion pair dissociation and ligand binding.
• The approach significantly advances the capabilities for extensive sampling in complex molecular systems.