EncoderMap III: A Dimensionality Reduction Package for Feature Exploration in Molecular Simulations.

Area of Science:

• Computational chemistry
• Machine learning for scientific data analysis
• Molecular dynamics simulations

Background:

• Dimensionality reduction is crucial for analyzing complex molecular simulation data.
• Existing methods like standard autoencoders may not fully capture the intricate relationships within high-dimensional molecular data.
• The sketch-map algorithm's multidimensional scaling (MDS)-like loss offers improved correlation between high- and low-dimensional similarities.

Purpose of the Study:

• To introduce a new, enhanced version of the EncoderMap package.
• To incorporate new features and customization options for analyzing molecular simulation data.
• To improve the usability and applicability of EncoderMap for researchers.

Main Methods:

• Utilizing a neural network autoencoder architecture.
• Augmenting the autoencoder with a multidimensional scaling (MDS)-like loss term.
• Porting the package to TensorFlow 2 for modern compatibility and new features, including sparse input capabilities.

Main Results:

• The updated EncoderMap (version 2) offers improved low-dimensional projections with better correlation between high- and low-dimensional similarities.
• New features include enhanced visualization, modularity for better understanding of the training process, and user-defined custom loss functions.
• The package demonstrates successful application to diverse molecular dynamics datasets, including topologically different proteins and the ubiquitin system.

Conclusions:

• The enhanced EncoderMap provides a powerful and flexible tool for dimensionality reduction in molecular simulations and general high-dimensional data analysis.
• The new features and improved architecture facilitate deeper insights into molecular dynamics and complex datasets.
• EncoderMap version 2 is readily usable on modern hardware and offers significant advantages over previous methods.