Enzyme Commission Number Prediction and Benchmarking with Hierarchical Dual-core Multitask Learning Framework

Area of Science:

• Bioinformatics
• Computational Biology
• Enzymology

Background:

• Enzyme commission (EC) numbers are crucial for understanding enzyme function and cellular metabolism.
• Existing computational methods for EC number prediction struggle with novel proteins, showing decreased performance, usability, and efficiency.
• There is a need for improved methods to accurately predict EC numbers for a wider range of protein sequences.

Purpose of the Study:

• To develop a novel deep learning framework, HDMLF, for accurate EC number prediction.
• To enhance the prediction performance, usability, and efficiency of EC number prediction methods, especially for newly discovered proteins.
• To demonstrate the capability of the model in uncovering enzyme promiscuity.

Main Methods:

• HDMLF utilizes a hierarchical dual-core multitask learning framework.
• The framework incorporates a protein language model for sequence embedding and a gated recurrent unit with an attention layer for EC prediction.
• A greedy strategy was employed for model integration and fine-tuning.

Main Results:

• HDMLF achieved significantly higher accuracy and F1 scores compared to four representative methods, improving them by 60% and 40%, respectively.
• The model demonstrated stable and superior performance across various metrics.
• A case study successfully predicted enzyme promiscuity, identifying tyrB's potential role in compensating for aspC loss.

Conclusions:

• HDMLF offers a substantial advancement in EC number prediction accuracy and reliability.
• The developed framework addresses limitations of existing methods, particularly for novel proteins.
• A user-friendly web platform (ECRECer) and offline bundle were created to enhance accessibility and usability.