Rapid in silico directed evolution by a protein language model with EVOLVEpro.

Area of Science:

• Biotechnology
• Computational Biology
• Protein Engineering

Background:

• Directed protein evolution is crucial for biomedical applications but faces challenges like complexity and inefficient optimization.
• Existing in silico methods using protein language models (PLMs) have limited generalization and mapping to protein activity.

Purpose of the Study:

• To develop a novel framework for rapid protein activity improvement using minimal experimental data.
• To overcome the limitations of current protein engineering and in silico guidance methods.

Main Methods:

• Introduced EVOLVEpro, a few-shot active learning framework.
• Integrated protein language models (PLMs) with regression models for fitness landscape guidance.
• Applied the framework to diverse protein engineering tasks with minimal data.

Main Results:

• Achieved up to 100-fold improvements in desired protein properties.
• Demonstrated effectiveness across six proteins in RNA production, genome editing, and antibody binding.
• Showcased the superiority of few-shot active learning over zero-shot predictions.

Conclusions:

• EVOLVEpro significantly accelerates AI-guided protein engineering.
• The framework offers a powerful approach for optimizing protein activity in biology and medicine.
• Highlights the advantage of active learning with limited experimental data for protein design.