Evaluation of machine learning-assisted directed evolution across diverse combinatorial landscapes.

Area of Science:

• Biochemistry
• Computational Biology
• Protein Engineering

Background:

• Machine learning-assisted directed evolution (MLDE) enhances protein variant identification compared to traditional methods.
• Optimal MLDE strategy selection is challenging due to limited understanding of performance factors across diverse proteins.

Purpose of the Study:

• To systematically analyze MLDE strategy performance across various protein fitness landscapes.
• To provide practical guidelines for selecting effective MLDE strategies in protein engineering.

Main Methods:

• Evaluated multiple MLDE strategies, including active learning and focused training with six zero-shot predictors.
• Analyzed 16 diverse protein fitness landscapes, quantifying landscape navigability with six attributes.

Main Results:

• MLDE shows greater advantage on more challenging directed evolution landscapes, particularly when focused training is combined with active learning.
• Focused training with zero-shot predictors, utilizing evolutionary, structural, and stability data, consistently outperformed random sampling for binding and enzyme activity.
• Performance advantage varied across different landscapes.

Conclusions:

• Findings offer practical guidelines for selecting MLDE strategies in protein engineering.
• The study highlights the benefit of combining active learning with focused training for challenging protein landscapes.