Navigating high-order protein fitness landscapes via deep learning on directed evolution trajectories

Area of Science:

• Protein Engineering
• Computational Biology
• Biophysics

Background:

• Predicting the functional impact of multiple mutations (high-order mutations) in proteins is a significant challenge.
• Current models, including protein language models, often fail to capture complex interactions between multiple residues.
• Understanding these interactions is crucial for protein engineering and deciphering disease mechanisms.

Purpose of the Study:

• To introduce DENet, a deep learning framework for predicting high-order mutation effects.
• To reconstruct high-resolution protein fitness landscapes using comutation information from directed evolution.
• To enhance the engineering of complex protein variants and elucidate their mechanisms.

Main Methods:

• Developed DENet, a deep learning framework leveraging directed evolution trajectories for comutation information.
• Applied DENet to reconstruct fitness landscapes for protein targets like KRAS and MEK1.
• Created an in silico strategy to simulate directed evolution for inferring comutation information from single-mutant data.

Main Results:

• DENet successfully identified potent high-order KRAS mutants and uncovered allosteric mechanisms.
• For MEK1, DENet nominated variants with over 1,000-fold increased drug resistance and identified >75% of known clinical mutations.
• DENet outperformed existing models in predicting mutation effects and identifying key variants.

Conclusions:

• DENet offers a quantitative framework for navigating complex protein fitness landscapes.
• The framework facilitates rational engineering of multi-mutation proteins.
• DENet aids in elucidating protein mechanisms and their clinical implications.