Rapid protein evolution by few-shot learning with a protein language model

Area of Science:

• Protein engineering
• Computational biology
• Artificial intelligence in life sciences

Background:

• Traditional directed evolution is labor-intensive and struggles with multi-property optimization.
• Current in silico methods using protein language models (PLMs) lack generalizability across protein families.
• Efficient protein engineering is crucial for advancements in research, medicine, and biotechnology.

Purpose of the Study:

• To introduce EVOLVEpro, a novel few-shot active learning framework for rapid protein engineering.
• To enhance the efficiency and effectiveness of in silico protein evolution.
• To demonstrate the broad applicability of AI-guided protein engineering.

Main Methods:

• Integration of protein language models (PLMs) with protein activity predictors.
• Implementation of a few-shot active learning strategy.
• Validation across diverse protein engineering applications.

Main Results:

• Achieved significant protein activity improvements within as few as four rounds of evolution.
• Demonstrated up to 100-fold improvement in desired protein properties.
• Showcased successful application in engineering RNA polymerase, CRISPR nucleases, prime editors, integrases, and antibodies.

Conclusions:

• EVOLVEpro significantly outperforms current state-of-the-art in silico protein evolution methods.
• Few-shot active learning with minimal experimental data is superior to zero-shot predictions for protein engineering.
• EVOLVEpro enables broader applications of AI-guided protein engineering in biology and medicine.