TCNeKP: A Novel Deep Learning Architecture for Enzyme Catalytic Activity Prediction.

Area of Science:

• Biochemistry
• Computational Biology
• Enzyme Kinetics

Background:

• Accurate prediction of enzyme kinetic parameters, including Kcat (catalytic turnover rate) and Km (Michaelis constant), is vital for enzyme rational design and engineering.
• Existing computational models face challenges in effectively integrating diverse data types like enzyme sequences, substrate information, and reaction conditions.

Purpose of the Study:

• To develop novel computational models, TCNeKP, for predicting enzyme kinetic parameters (Kcat and Km).
• To enhance the accuracy and robustness of enzyme kinetic parameter prediction for both wild-type and mutant enzymes across various enzyme classes and substrates.

Main Methods:

• Enzyme sequences were autoembedded and processed using a temporal convolutional network (TCN) for feature extraction.
• Substrates were encoded using a pretrained SMILES-Transformer language model, and catalytic conditions (pH, temperature) were encoded via radial basis function (RBF).
• A fully connected network integrated these features for single-task prediction, and a multitask TCNeKP model with a cross-task dynamic parameter-sharing module and attention mechanism was developed.

Main Results:

• The TCNeKP models demonstrated robust performance in predicting Kcat and Km across 7 EC classes, outperforming state-of-the-art models like MPEK, UniKP, and DLKcat.
• The multitask TCNeKP model achieved superior R-squared values (0.677 for Kcat, 0.657 for Km) compared to benchmark models.
• Collaborative learning between Kcat and Km prediction tasks significantly improved feature extraction for enzyme-substrate interactions and catalysis.

Conclusions:

• The developed TCNeKP models provide a powerful tool for accurate prediction of enzyme kinetic parameters.
• Multitask learning enhances predictive performance by leveraging shared information between Kcat and Km prediction.
• These findings contribute to advancing enzyme engineering and rational enzyme design through improved computational prediction.