MPEK: a multitask deep learning framework based on pretrained language models for enzymatic reaction kinetic...

Area of Science:

• Biochemistry and Molecular Biology
• Computational Biology and Bioinformatics
• Enzyme Engineering

Background:

• Enzymatic reaction kinetics, including turnover number (kcat) and Michaelis constant (Km), are vital for understanding enzyme mechanisms and optimizing enzymes for biomanufacturing.
• Experimental determination of kcat and Km is resource-intensive, necessitating advanced computational approaches.

Purpose of the Study:

• To develop a universal pretrained multitask deep learning model (MPEK) for simultaneous prediction of kcat and Km.
• To improve prediction accuracy by integrating pH, temperature, and organismal information and considering the intrinsic relationship between kcat and Km.

Main Methods:

• A multitask deep learning architecture was employed, pretrained on extensive datasets.
• The model, MPEK, was designed to predict kcat and Km concurrently, incorporating environmental and biological context.
• Performance was evaluated against existing models using established kcat and Km test datasets.

Main Results:

• MPEK significantly outperformed previous models, achieving a Pearson coefficient of 0.808 for kcat prediction (14.6% and 7.6% improvement) and 0.777 for Km prediction (34.9% and 53.3% improvement).
• The model demonstrated sensitivity to enzyme sequence variations and revealed enzyme promiscuity.
• Case studies highlighted MPEK's potential for assisting in enzyme mining and directed evolution.

Conclusions:

• MPEK offers a powerful and efficient in silico tool for predicting key enzyme kinetic parameters.
• The model facilitates enzyme discovery and engineering by providing accurate predictions and insights into enzyme behavior.
• A web server has been developed to make MPEK accessible for broader research applications.