Generative Artificial Intelligence-Empowered Virtual Evolution of Enzyme with the VERnet Model

Area of Science:

• Biochemistry
• Computational Biology
• Protein Engineering

Background:

• Large language models (LLMs) struggle with designing active proteins requiring complex intramolecular interactions, limiting biocatalyst engineering.
• Current artificial intelligence (AI) research in biology often relies on laboratory assays, which are time-consuming for complex tasks.

Purpose of the Study:

• To develop a standardized strategy using function-targeted AI models to interpret the impact of protein sequence variations on function.
• To guide enzyme evolution by focusing on specific functional interpretations rather than just binding affinity.

Main Methods:

• Established the VERnet model using deep mutation scanning data and self-distillation.
• Applied generative AI and directed evolution to identify functional CYP2C9 variants.
• Utilized AlphaFold3 for fine-tuning models to predict double amino acid substitution variants.
• Performed molecular dynamics simulations to analyze structural and dynamic changes in evolved variants.

Main Results:

• The VERnet model achieved 93.5% accuracy in interpreting Cytochrome P450 2C9 (CYP2C9) variants.
• Directed evolution identified multiple CYP2C9 variants with diverse functional alterations.
• AI model predictions were strongly corroborated by in vitro validation of metabolic activity.

Conclusions:

• Function-targeted AI models offer substantial potential for predicting and guiding protein functional evolution.
• The developed strategy effectively decodes the subtle effects of sequence variations on protein function.
• This approach advances biocatalyst design and protein engineering through AI-driven insights.