Correlation-Based Feature Selection to Identify Functional Dynamics in Proteins.

Area of Science:

• Computational Biology
• Biophysics
• Molecular Dynamics

Background:

• Interpreting molecular dynamics (MD) simulations requires dimensionality reduction techniques like Principal Component Analysis (PCA) and Time-Lagged Independent Component Analysis (TICA).
• Selecting relevant input coordinates (features) is critical, as noise or uncorrelated motions can be mistaken for important dynamics.
• PCA and TICA can be misled by high-amplitude or long-timescale noise, obscuring true functional motions.

Purpose of the Study:

• To develop a robust method for identifying functionally relevant collective motions in biomolecular systems from MD simulations.
• To overcome limitations of traditional PCA and TICA by effectively discriminating collective dynamics from noise.

Main Methods:

• Utilized a block-diagonalization approach on the correlation matrix of input coordinates.
• Employed a combination of linear correlation measures and the Leiden community detection algorithm for clustering.
• Validated the method on model systems, including T4 lysozyme functional motion and villin headpiece folding.

Main Results:

• The combination of linear correlation and Leiden community detection successfully identified collective motions underlying functional dynamics.
• This approach effectively discriminated relevant motions from uncorrelated noise, avoiding biases associated with PCA and TICA.
• Demonstrated accurate identification of collective motion in T4 lysozyme and provided physical interpretation for villin headpiece folding.

Conclusions:

• The proposed method provides a reliable strategy for feature selection in MD simulations.
• This approach enhances the interpretation of biomolecular dynamics by accurately identifying functionally relevant collective motions.
• The findings offer a more precise understanding of protein dynamics and mechanisms.