Efficient, Few-shot Directed Evolution with Energy Rank Alignment

Area of Science:

• Biochemistry
• Computational Biology
• Protein Engineering

Background:

• Directed evolution is crucial for protein engineering but is limited by the high cost of experimental data.
• Current machine learning methods for sequence selection are constrained by sparse data, restricting model complexity.

Purpose of the Study:

• To develop a more efficient method for protein engineering by adapting large-scale protein language models.
• To leverage natural protein sequence distributions for navigating complex fitness landscapes.

Main Methods:

• Adapted large-scale pre-trained protein language models using experimental data.
• Employed a post-training algorithm grounded in statistical physics.
• Utilized quantitative experimental rankings to generate diverse, high-fitness sequences.

Main Results:

• Achieved remarkable efficiency improvements in protein engineering.
• Required fewer experimental data points compared to competing methods.
• The adapted models provided insights into the biophysical characteristics of high-fitness sequences.

Conclusions:

• Adapting protein language models offers a powerful alternative for efficient protein engineering.
• This method effectively navigates high-dimensional fitness landscapes.
• The adapted models are interpretable and enhance understanding of protein properties.