An enzyme-specific protein language model for catalytic property prediction

Area of Science:

• Biochemistry
• Computational Biology
• Enzyme Engineering

Background:

• Enzymes are crucial for cellular metabolism, but predicting their catalytic functions from amino acid sequences is difficult.
• Current protein language models (PLMs) are general-purpose and lack enzyme-specific adaptations, limiting their efficiency and accuracy for enzyme-related tasks.

Purpose of the Study:

• To develop an enzyme-specific PLM, EnzGFM, that accurately predicts enzyme properties and accelerates screening.
• To create an agentic pipeline, EnzGFM-Agent, for practical enzyme engineering applications.

Main Methods:

• EnzGFM utilizes a Mamba-Transformer hybrid architecture with hierarchical pre-training.
• The model was evaluated on enzyme property prediction benchmarks, including kinetic parameters, enzyme-reaction mapping, EC number classification, and mutation effect assessment.
• EnzGFM-Agent was developed as an enzyme-focused agentic pipeline for experimental validation.

Main Results:

• EnzGFM outperforms existing Transformer-based PLMs by 2-5 fold in acceleration.
• Significant relative improvements were observed: 16.67% in kinetic parameter prediction, 15.69% in enzyme-reaction mapping, 13.19% in EC number classification, and 20.04% in mutation effect assessment.
• EnzGFM-Agent demonstrated the ability to enrich beneficial variants in small candidate pools, reducing wet-lab screening burden.

Conclusions:

• EnzGFM effectively captures enzyme-specific sequence-function patterns.
• EnzGFM-Agent translates predictions into actionable candidates for enzyme engineering.
• The developed tools offer a more efficient and accurate approach to enzyme discovery and engineering.