Learning Collective Variables with Synthetic Data Augmentation through Physics-Inspired Geodesic Interpolation.

Area of Science:

• Computational chemistry
• Biophysics
• Molecular dynamics simulations

Background:

• Enhanced sampling techniques in molecular dynamics simulations often rely on collective variables (CVs) to study rare events like protein folding.
• Identifying effective CVs is challenging due to limited prior knowledge of the event's pathway.

Purpose of the Study:

• To develop a simulation-free data augmentation strategy for improving the efficiency of enhanced sampling techniques.
• To generate realistic protein folding transition data without requiring true transition state samples.

Main Methods:

• A data augmentation strategy using physics-inspired metrics to generate geodesic interpolations.
• Creating data that mimics protein folding transitions.
• Leveraging interpolation progress parameter for regression-based CV model learning.

Main Results:

• Successfully generated simulation-free data resembling protein folding transitions.
• Demonstrated potential for improving classifier-based methods with augmented data.
• Showcased the utility of regression-based learning for CV models.

Conclusions:

• The proposed data augmentation strategy enhances sampling efficiency in molecular dynamics simulations.
• This method offers a viable alternative for generating crucial transition data when real samples are scarce.
• Improves the accuracy and applicability of collective variable models for studying complex biomolecular events.