MLCV: Bridging Machine-Learning-Based Dimensionality Reduction and Free-Energy Calculation.

Area of Science:

• Computational chemistry and biophysics
• Molecular dynamics simulations
• Enhanced sampling techniques

Background:

• Importance-sampling algorithms use reaction coordinates (RCs) to study slow molecular processes.
• Manually defining RCs from collective variables (CVs) is challenging due to complex conformational changes.
• Machine learning (ML) methods can extract low-dimensional RCs, but integrating them into simulations is difficult.

Purpose of the Study:

• To present MLCV, a tool simplifying the use of ML-derived CVs in enhanced-sampling simulations.
• To enable seamless integration of ML-based CVs into molecular dynamics (MD) workflows.
• To facilitate the study of complex biochemical processes using ML-guided enhanced sampling.

Main Methods:

• Developed the MLCV tool for integrating ML-CVs with the Colvars module in MD simulations.
• Implemented hard-coded neural networks within Colvars for direct use of ML-CVs.
• Tested the approach using three case studies involving small peptides.

Main Results:

• Demonstrated that MLCV effectively bridges deep learning and enhanced-sampling in MD simulations.
• Showcased the versatility of MLCV across various CVs and neural network architectures.
• Successfully applied the method to analyze complex molecular transformations in peptides.

Conclusions:

• MLCV offers a powerful and user-friendly platform for ML-guided CV discovery and enhanced-sampling.
• The tool is accessible to both experts and non-experts in computational chemistry.
• MLCV aids in unveiling molecular mechanisms of complex biochemical processes through advanced simulations.