EncoderMap: Dimensionality Reduction and Generation of Molecule Conformations

Area of Science:

• Computational chemistry
• Biophysics
• Machine learning

Background:

• Molecular simulations generate large, high-dimensional datasets.
• Extracting meaningful information requires dimensionality reduction.
• Existing methods vary in efficiency and linkage between high- and low-dimensional spaces.

Purpose of the Study:

• To introduce EncoderMap, a new dimensionality reduction algorithm.
• To establish a functional link between high- and low-dimensional representations.
• To demonstrate EncoderMap's utility in analyzing molecular simulation data.

Main Methods:

• Utilized a neural network autoencoder combined with a nonlinear distance metric.
• Applied EncoderMap to molecular simulation data of a flexible peptide and protein folding.
• Evaluated the algorithm's ability to project data and generate high-dimensional structures.

Main Results:

• Successfully projected high-dimensional structures to an informative low-dimensional map.
• Identified major conformational states and transitions in molecular simulations.
• Demonstrated the generation of high-dimensional conformations from the low-dimensional map (inverse mapping).

Conclusions:

• EncoderMap efficiently reduces dimensionality and preserves essential information from molecular simulations.
• The inverse mapping capability is crucial for enhancing conformational space exploration and transition sampling.
• This method offers a powerful tool for analyzing complex molecular dynamics data.