Random Functions as Data Compressors for Machine Learning of Molecular Processes.

Area of Science:

• Computational chemistry
• Biophysics
• Materials science

Background:

• Machine learning (ML) is revolutionizing molecular dynamics (MD) simulations.
• ML algorithms often require dimensionality reduction for analyzing complex conformational landscapes.
• Feature selection is crucial but challenging due to high dimensionality and computational costs.

Purpose of the Study:

• To introduce random nonlinear projections as an efficient feature compression technique for ML in MD.
• To demonstrate the method's ability to reduce computational cost without substantial information loss.
• To validate the approach for protein folding and trajectory analysis.

Main Methods:

• Development of an efficient random projection method for feature space compression.
• Application of the method to MD trajectory data from protein folding simulations (NTL9 and villin headpiece).
• Analysis of static and dynamic information retention after compression.

Main Results:

• Random nonlinear projections effectively compress high-dimensional feature spaces.
• Compressed feature spaces lead to faster computations in ML analyses.
• Core static and dynamic information relevant to protein folding was retained.
• Trajectory analysis using compressed features proved more robust.

Conclusions:

• Random nonlinear projections offer a powerful and efficient strategy for dimensionality reduction in ML for MD.
• This technique enhances the feasibility and robustness of analyzing complex molecular dynamics data.
• The method has broad applicability in materials modeling and biological simulations.