mGPfusion: predicting protein stability changes with Gaussian process kernel learning and data fusion.

Area of Science:

• Biochemistry
• Computational Biology
• Protein Engineering

Background:

• Proteins are vital in the biochemical industry, but predicting how mutations affect their stability is challenging due to limited experimental data.
• Accurate computational methods are needed for efficient protein design and property refinement through mutations.

Purpose of the Study:

• To develop a novel computational method for predicting protein stability changes upon single and multiple mutations.
• To improve the accuracy of predictive models by integrating limited experimental data with large-scale molecular simulation data.

Main Methods:

• Developed mGPfusion, a Gaussian process (GP) based method utilizing Bayesian data fusion to combine experimental and in silico data.
• Modeled proteins using contact maps and employed a mixture of graph kernels to infer mutation-induced stability effects.
• Created protein-specific models requiring minimal experimental data for the protein of interest.

Main Results:

• mGPfusion demonstrated superior performance over state-of-the-art methods in predicting protein stability across 15 different proteins.
• Integrating molecular simulation data significantly enhanced model learning and prediction accuracy.
• The method showed robust performance even with limited experimental measurements.

Conclusions:

• mGPfusion offers an accurate and efficient computational approach for predicting protein stability changes.
• The integration of diverse data sources, including molecular simulations, is crucial for advancing protein design.
• The developed software and datasets are publicly available to facilitate further research.