Quantified Dynamics-Property Relationships: Data-Efficient Protein Engineering with Machine Learning of Protein...

Area of Science:

• Computational Biology
• Protein Engineering
• Machine Learning

Background:

• Machine learning excels at predicting protein mutation effects but typically requires extensive training data.
• Experimental data collection for training is often costly and time-consuming, limiting its scalability.
• High-throughput molecular dynamics simulations offer a potential data source but are underutilized in this context.

Purpose of the Study:

• To develop a novel method for selecting optimal protein variants using limited experimental data and molecular dynamics simulations.
• To demonstrate the effectiveness of this approach in protein engineering and directed evolution.
• To establish a practical framework for integrating protein dynamics information into engineering workflows.

Main Methods:

• Utilized deep neural networks trained on molecular dynamics simulation data to generate dynamic property descriptors.
• Quantified relationships between a small set of experimentally determined labels and these dynamic descriptors.
• Applied this method to select desirable protein variants for optimization.

Main Results:

• Achieved highly optimized protein variants using minimal experimental data, surpassing alternative supervised machine learning methods.
• Accurately predicted key residues influencing protein properties based on limited experimental labels and dynamics-property relationships.
• Demonstrated that crucial residue information can be uncovered even when not predictable from simulations or experimental data alone.

Conclusions:

• The developed method provides a practical and efficient framework for protein engineering guided by simulation-derived dynamics.
• This approach effectively leverages small experimental datasets, making protein engineering more accessible and cost-effective.
• The study highlights the power of integrating computational dynamics with experimental validation for predicting and optimizing protein function.