A random forest learning assisted "divide and conquer" approach for peptide conformation search

Area of Science:

• Computational chemistry
• Biophysics
• Machine learning

Background:

• Determining peptide conformations computationally is complex due to high-dimensional search spaces.
• The "divide and conquer" strategy offers a promising method for reducing search space size.
• Existing methods require significant human effort for optimization.

Purpose of the Study:

• To enhance the "divide and conquer" approach for peptide conformation prediction using machine learning.
• To develop a method that can automatically identify and utilize patterns in peptide structures.
• To improve the efficiency and scalability of computational peptide conformation searches.

Main Methods:

• Utilized a random forest classification algorithm to identify equivalent amino acid residue groups ("words").
• Developed a random forest supervised learning model to learn the conformational rules ("grammar") of backbone dihedral angles (φ-ψ units).
• Integrated learned "words" and "grammar" into the "divide and conquer" framework for efficient conformational searching.

Main Results:

• Amino acid residues were successfully grouped into equivalent "words", enabling fragment substitution.
• A distinct "grammar" governing low-energy peptide conformations was identified and learned.
• The machine learning-assisted method efficiently searched conformations for peptides of varying lengths, including GGG/AAA/GGGG/AAAA/GGGGG.
• Computational cost demonstrated a slow increase with peptide length.

Conclusions:

• Machine learning, specifically random forests, can significantly improve the "divide and conquer" method for peptide conformation prediction.
• The identification of equivalent amino acid "words" and conformational "grammar" provides a powerful, automated approach to search space reduction.
• This novel method offers a scalable and efficient solution for computational peptide structure determination, with minimal human intervention.