Unsupervised evolution of protein and antibody complexes with a structure-informed language model

Area of Science:

• Computational biology
• Protein engineering
• Structural biology

Background:

• Large language models (LLMs) excel at learning protein design principles from sequence data alone.
• Protein function, activity, and evolvability are critically determined by their three-dimensional structures, not just sequences.
• Existing LLMs often lack the structural context necessary for comprehensive protein design.

Purpose of the Study:

• To develop a general protein language model augmented with structural information to guide protein evolution.
• To demonstrate the model's capability in engineering protein complexes and improving therapeutic antibodies.
• To validate the effectiveness of structure-informed protein language models in enhancing protein function without task-specific training.

Main Methods:

• Augmenting a general protein language model with protein structure backbone coordinates.
• Extending the ESM-IF1 model, initially trained on single-chain structures, to engineer protein complexes.
• Screening approximately 30 variants of two therapeutic antibodies against SARS-CoV-2.

Main Results:

• The structure-augmented language model successfully guided protein evolution for diverse proteins.
• The extended ESM-IF1 model enabled the engineering of protein complexes.
• Significant improvements were observed in antibody neutralization (up to 25-fold) and affinity (up to 37-fold) against SARS-CoV-2 variants BQ.1.1 and XBB.1.5.

Conclusions:

• Integrating structural information into protein language models is advantageous for identifying efficient protein evolution pathways.
• This approach allows for protein engineering without the need for task-specific training data.
• The findings pave the way for improved protein design and therapeutic development, particularly for antibodies targeting viral escape variants.