Learning protein fitness landscapes with deep mutational scanning data from multiple sources

Area of Science:

• Biochemistry
• Computational Biology
• Machine Learning

Background:

• Accurate fitness landscape mapping is crucial for machine learning-assisted directed evolution (MLDE).
• Existing MLDE methods often struggle with generalizing to new protein targets.
• Deep mutational scanning (DMS) provides valuable data for understanding protein sequence-function relationships.

Purpose of the Study:

• To develop and validate a multi-protein training scheme for improving fitness landscape prediction in MLDE.
• To assess the scheme's ability to generalize to new proteins and predict higher-order variant effects.
• To identify potential limitations and pitfalls in current MLDE approaches.

Main Methods:

• A novel multi-protein training scheme was designed, utilizing existing DMS data from diverse proteins.
• Proof-of-concept trials were conducted to validate the scheme through random and positional extrapolation, zero-shot predictions, and higher-order effect extrapolation.
• Performance was benchmarked against strong baseline models.

Main Results:

• The multi-protein training scheme demonstrated effectiveness in aiding the understanding of new protein fitness landscapes.
• The study validated the scheme's performance in random/positional extrapolation, zero-shot predictions, and higher-order effect prediction.
• Unexpectedly strong performance from baseline models revealed significant pitfalls in current MLDE practices.

Conclusions:

• The proposed multi-protein training scheme offers a promising approach to enhance MLDE by improving fitness landscape learning.
• The findings underscore the importance of robust baselines and highlight areas for improvement in MLDE methodologies.
• This work contributes to a better understanding of protein fitness profiles and advances the development of more effective protein engineering tools.