Resolving Protein Conformational Plasticity and Substrate Binding via Machine Learning.

Area of Science:

• Biophysics
• Computational Biology
• Molecular Dynamics

Background:

• Identifying specific protein conformations crucial for ligand binding is challenging due to protein flexibility and the random nature of molecular recognition.
• Existing methods struggle to link distinct protein structures with specific ligand-bound states.

Purpose of the Study:

• To develop a computational framework that accurately resolves protein conformational plasticity and substrate binding processes.
• To identify key protein residues and allosteric sites involved in ligand recognition pathways.

Main Methods:

• Integration of Random Forest (RF) machine learning with time-structured independent component analysis (TICA) for molecular dynamics (MD) simulations.
• Application of RF-TICA-MD to analyze multi-microsecond MD trajectories of ligand recognition in T4 lysozyme and cytochrome P450cam.
• Utilizing supervised classification to identify ligand-sensitive amino acid residues followed by unsupervised dimension reduction.

Main Results:

• The RF-TICA-MD framework successfully delineated protein conformational landscapes, distinguishing ligand-bound from unbound states.
• Identified critical amino acid residue pairs essential for ligand binding, even when unsupervised methods failed.
• Demonstrated robustness with different machine learning classifiers (e.g., XGBoost) and data types.

Conclusions:

• RF-TICA-MD offers an efficient and unambiguous solution for studying protein conformational dynamics in ligand binding.
• The framework can identify allosterically important regions, providing insights into molecular recognition mechanisms.
• A Python implementation is available, facilitating broader application in computational biophysics.