SESNet: sequence-structure feature-integrated deep learning method for data-efficient protein engineering.

Area of Science:

• Computational Biology
• Protein Engineering
• Machine Learning

Background:

• Deep learning models for protein engineering require extensive experimental data for accurate fitness prediction.
• Predicting functional fitness of high-order protein mutants remains challenging due to data scarcity.

Purpose of the Study:

• To develop SESNet, a supervised deep learning model for predicting protein mutant fitness.
• To leverage sequence and structure information with an attention mechanism for improved prediction accuracy.
• To introduce a data augmentation strategy for enhancing model performance with limited experimental data.

Main Methods:

• SESNet integrates local and global evolutionary sequence contexts with protein structure information.
• An attention mechanism is employed to weigh the importance of different input features.
• A data augmentation strategy uses unsupervised learning data to pre-train the model before fine-tuning with experimental data.

Main Results:

• SESNet outperforms state-of-the-art models on 26 deep mutational scanning datasets for sequence-function relationship prediction.
• The proposed data augmentation strategy enables high prediction accuracy for higher-order mutants (>4 sites) with minimal experimental data (<50).
• The model demonstrates practical value for biochemical groups with limited experimental resources.

Conclusions:

• SESNet provides an accurate and efficient method for predicting protein mutant fitness.
• The data augmentation strategy significantly improves the prediction of complex mutations.
• This approach lowers the experimental burden for protein engineering studies.