Predicting protein conformational motions using energetic frustration analysis and AlphaFold2

Area of Science:

• Structural Biology
• Computational Biology
• Biophysics

Background:

• Protein function is intrinsically linked to dynamic conformational changes.
• Predicting static protein structures using deep learning (e.g., AlphaFold2) is advanced, but predicting dynamic motions remains a significant challenge due to limited experimental data.
• Current purely data-driven machine learning approaches struggle with the complexity of protein dynamics.

Purpose of the Study:

• To develop a novel method for generating protein allosteric motions by integrating physical energy landscape information into deep learning frameworks.
• To demonstrate that local energetic frustration can be used to enhance deep learning models like AlphaFold2 for predicting protein conformational dynamics.
• To provide a strategy for predicting dynamic structures of allosteric proteins.

Main Methods:

• Developed an integrative method combining deep learning with physical energy landscape information.
• Utilized local energetic frustration, a quantifiable feature of the protein energy landscape, to empower AlphaFold2 (AF2).
• Input multiple sequence alignment sequences with progressively enhanced energetic frustration features to generate alternative protein structures and motion pathways.

Main Results:

• The method successfully generated protein conformational motions, starting from static ground state structures.
• Generated motions for adenylate kinase were consistent with experimental and molecular dynamics simulation data.
• Successfully predicted alternative conformations for KaiB and ribose-binding proteins, which exhibit large-amplitude conformational changes.
• Demonstrated a method to extract features from the AlphaFold2 energy landscape, addressing its 'black box' nature.

Conclusions:

• Integrating physical principles, specifically energy landscape information, into deep learning models is a viable strategy for predicting protein conformational motions.
• The developed method effectively generates realistic protein allosteric motions and alternative conformations.
• This approach offers a promising solution for the long-standing challenge of predicting dynamic protein structures.