Large language models generate functional protein sequences across diverse families

Area of Science:

• Biotechnology
• Computational Biology
• Protein Engineering

Background:

• Deep-learning language models show potential in protein design and engineering.
• Generating functional protein sequences across large families remains a challenge.

Purpose of the Study:

• Introduce ProGen, a language model for generating protein sequences with predictable functions.
• Demonstrate ProGen's adaptability and performance in protein family generation.

Main Methods:

• Trained ProGen on 280 million protein sequences from over 19,000 families.
• Augmented the model with control tags for specifying protein properties.
• Fine-tuned ProGen on curated sequences and tags for improved controllable generation.

Main Results:

• ProGen successfully generated functional protein sequences across diverse families, including lysozymes, chorismate mutase, and malate dehydrogenase.
• Artificial lysozymes engineered by ProGen exhibited catalytic efficiencies comparable to natural lysozymes.
• Achieved significant sequence identity reduction (as low as 31.4%) while maintaining protein function.

Conclusions:

• ProGen offers a powerful tool for de novo protein design and engineering.
• The model's ability to generate functional proteins with high control advances biotechnological applications.
• ProGen demonstrates broad applicability across various protein families.