ivis Dimensionality Reduction Framework for Biomacromolecular Simulations.

Area of Science:

• Computational Biology
• Biophysics
• Machine Learning

Background:

• Molecular dynamics (MD) simulations generate high-dimensional data, complicating protein analysis.
• Existing dimensionality reduction methods (PCA, t-ICA, t-SNE) have limitations in preserving structural information and handling noise.

Purpose of the Study:

• To evaluate ivis, a deep learning-based dimensionality reduction technique, for analyzing protein MD simulations.
• To compare ivis with traditional methods for its effectiveness in capturing protein dynamics and structure-function relationships.

Main Methods:

• Application of the ivis deep learning framework to MD simulation data of diatom aureochrome 1a (PtAu1a) LOV domains.
• Comparative analysis of ivis against linear (PCA, t-ICA) and nonlinear (t-SNE) dimensionality reduction methods.

Main Results:

• Ivis demonstrated superior performance in constructing Markov state models (MSMs).
• Ivis preserved both local and global distance information with minimal loss.
• The method successfully identified residue-level allosteric mechanisms via neural network feature weights.

Conclusions:

• Ivis is a powerful tool for analyzing complex protein dynamics from MD simulations.
• This deep learning approach offers enhanced insights into protein structure-function relationships and allostery.
• Ivis represents a promising advancement in the computational biology analysis toolbox.