Evaluation of Dimensionality-reduction Methods from Peptide Folding-unfolding Simulations

Area of Science:

• Computational chemistry
• Molecular dynamics
• Biophysics

Background:

• Dimensionality reduction is key for studying molecular free energy landscapes.
• Non-linear methods often outperform linear methods like principal component analysis (PCA) in simple systems.
• The B1 domain of streptococcal protein G's second beta-hairpin is a relevant model system.

Purpose of the Study:

• To evaluate and compare the performance of various non-linear dimensionality reduction techniques (locally linear embedding, Isomap, diffusion maps) against principal component analysis (PCA).
• To assess these methods using the equilibrium folding/unfolding trajectory of a protein beta-hairpin.
• To determine the most effective dimensionality reduction strategy for complex molecular systems.

Main Methods:

• Utilized CHARMM parm19 polar hydrogen potential function for molecular simulations.
• Generated equilibrium folding/unfolding trajectories for the second beta-hairpin of streptococcal protein G.
• Applied and evaluated principal component analysis, locally linear embedding, Isomap, and diffusion maps.

Main Results:

• Principal component analysis (PCA) demonstrated comparable performance to non-linear methods on this complex protein system.
• No single dimensionality reduction method excelled across all evaluation criteria.
• Each method exhibited specific limitations when applied to the protein folding data.

Conclusions:

• PCA is a viable and effective method for analyzing complex molecular systems, not necessarily inferior to non-linear techniques.
• A comprehensive assessment of dimensionality reduction requires multiple evaluation criteria, not just one.
• Careful consideration and multiple criteria are essential when using dimensionality reduction, especially for interpreting free energy landscapes from limited dimensions.