Robust enzyme discovery and engineering with deep learning using CataPro.

Area of Science:

• Biochemistry
• Computational Biology
• Enzyme Engineering

Background:

• Accurate prediction of enzyme kinetic parameters (kcat, Km, kcat/Km) is vital for enzyme exploration and modification.
• Existing predictive models often suffer from low accuracy or poor generalization due to overfitting issues.

Purpose of the Study:

• To develop an accurate and generalizable deep learning model for predicting enzyme kinetic parameters.
• To evaluate the performance of the proposed model on unbiased datasets and demonstrate its utility in enzyme mining and engineering.

Main Methods:

• Developed unbiased datasets for rigorous evaluation of enzyme kinetic parameter prediction methods.
• Proposed CataPro, a deep learning model utilizing pre-trained models and molecular fingerprints.
• Predicted key enzyme kinetic parameters: turnover number (kcat), Michaelis constant (Km), and catalytic efficiency (kcat/Km).

Main Results:

• CataPro demonstrated significantly enhanced accuracy and generalization ability compared to baseline models on unbiased datasets.
• In an enzyme mining project, CataPro combined with traditional methods identified an enzyme (SsCSO) with a 19.53-fold increase in activity.
• The identified enzyme was successfully engineered, further improving its activity by 3.34 times.

Conclusions:

• CataPro is a highly effective deep learning tool for accurate prediction of enzyme kinetic parameters.
• The model shows significant potential for accelerating enzyme discovery and facilitating enzyme modification efforts.
• This work highlights the practical application of advanced computational methods in biocatalysis and enzyme engineering.