Identifying promising sequences for protein engineering using a deep transformer protein language model.

Area of Science:

• Computational Biology
• Protein Engineering
• Artificial Intelligence in Biochemistry

Background:

• Discovering novel protein sequences with desired properties is challenging due to the vast sequence space.
• Identifying promising sequences for applications like nanobody discovery and protein optimization is often costly and time-consuming.

Purpose of the Study:

• To develop and validate a deep learning-based method for efficiently identifying high-potential protein sequences.
• To introduce a 'Promise Score' to guide protein engineering efforts in sequence discovery and optimization.

Main Methods:

• Utilized a deep transformer protein language model and its self-attention map to calculate a Promise Score.
• Applied the Promise Score to nanobody (Nb) discovery and protein optimization workflows.
• Analyzed self-attention maps to identify key protein regions involved in intermolecular interactions.
• Explored fine-tuning strategies for the protein language model for predictive property modeling.

Main Results:

• The Promise Score effectively selects promising lead sequences from nanobody repertoires.
• The Promise Score guides site-specific mutagenesis experiments, identifying a high percentage of improved protein sequences.
• Self-attention maps provide insights into protein regions driving specific interactions.
• Demonstrated the utility of fine-tuning protein language models for targeted protein engineering tasks.

Conclusions:

• The Promise Score offers a computationally efficient approach to accelerate protein engineering.
• Deep learning models, particularly transformer-based language models, are powerful tools for predicting and designing functional proteins.
• Understanding protein interactions through self-attention maps enhances the design process.