Nearl: extracting dynamic features from molecular dynamics trajectories for machine learning tasks.

Area of Science:

• Computational Biology
• Biophysics
• Machine Learning

Background:

• Machine learning is rapidly advancing in biomolecular applications.
• Information regarding protein dynamics remains underutilized in current predictive models.
• Extracting dynamic features from molecular dynamics (MD) trajectories is crucial for understanding protein function.

Purpose of the Study:

• To introduce Nearl, an automated pipeline for extracting dynamic features from large ensembles of molecular dynamics trajectories.
• To identify intrinsic patterns of molecular motion and generate informative features for predictive modeling.
• To bridge the gap between raw MD data and actionable insights for machine learning applications.

Main Methods:

• Nearl implements two novel classes of dynamic features: marching observers and property-density flow.
• These features capture local atomic motions and global conformational changes.
• Protein substructures are transformed into 3D grids using voxelization, compatible with 3D convolutional neural networks (3D-CNNs).

Main Results:

• Nearl successfully extracts dynamic features from molecular dynamics trajectories.
• The generated features are suitable for input into 3D-CNNs for predictive modeling tasks.
• The pipeline demonstrates flexibility in handling various input formats and customizable feature extraction.

Conclusions:

• Nearl addresses the underutilization of protein dynamics data in machine learning.
• The pipeline provides a robust method for converting complex MD data into valuable features.
• Nearl facilitates the development of more accurate predictive models by incorporating dynamic information.