Data-efficient protein mutational effect prediction with weak supervision by molecular simulation and protein...

Area of Science:

• Computational Biology
• Biophysics
• Machine Learning

Background:

• Protein mutational effect prediction is vital for protein engineering and pathogenicity assessment.
• Current methods face challenges due to limited experimental data and high costs.
• Previous data augmentation relied on molecular simulations, limited to thermostability.

Purpose of the Study:

• To develop a new data augmentation technique for protein mutational effect prediction.
• To extend the applicability of computational predictions to diverse protein properties.
• To improve prediction accuracy in low-data regimes.

Main Methods:

• Combined molecular simulation with zero-shot prediction from protein language models for data augmentation.
• Utilized computational estimates as 'weak' training data.
• Dynamically adjusted the weight and inclusion of weak data based on experimental data availability.

Main Results:

• The new method enhances prediction accuracy, particularly when experimental data is scarce.
• Successfully extended applicability to protein binding affinity and enzymatic activity prediction.
• Demonstrated improved performance in benchmark tests for small data scenarios.

Conclusions:

• The proposed method effectively supplements experimental data for machine learning models.
• Offers a powerful approach for protein engineering and pathogenicity prediction with limited data.
• Advances the field of protein mutational effect prediction in data-scarce environments.