Convolution Neural Network-Based Prediction of Protein Thermostability.

Area of Science:

• Biochemistry
• Computational Biology
• Protein Engineering

Background:

• Protein thermostability is crucial for industrial applications but often limited in natural proteins.
• Computer-aided rational design and machine learning methods aim to improve protein thermostability.
• Existing prediction methods have limitations due to overlooking protein sequence features.

Purpose of the Study:

• To develop an accurate method for predicting protein thermostability changes induced by mutations.
• To address limitations of current methods by incorporating protein sequence features.
• To enhance the design of thermostable proteins, especially enzymes.

Main Methods:

• Utilized convolutional neural networks (CNNs) for protein thermostability prediction.
• Developed a three-dimensional coding algorithm incorporating protein mutation information.
• Employed multiscale convolution to extract neighboring features at mutation sites.

Main Results:

• Achieved high prediction accuracies of 86.4% (S1615) and 87% (S388).
• Demonstrated a significantly improved Matthews correlation coefficient compared to other methods.
• Outperformed the RIF strategy (Rosetta ddg monomer, I Mutant 3.0, FoldX) in predicting lipase mutant thermostability (75.0% vs 66.7%).

Conclusions:

• The proposed CNN-based method effectively predicts protein thermostability and distinguishes key features.
• This approach offers a powerful tool for devising mutations to enhance protein thermostability.
• The method shows significant potential for advancing enzyme engineering and industrial biotechnology.