CACLENS: A Multitask Deep Learning System for Enzyme Discovery.

Area of Science:

• Biotechnology and Bioinformatics
• Computational Biology
• Enzyme Engineering

Background:

• Deep learning models excel at predicting enzyme properties but struggle with high-performance screening due to limitations in multimodal and multitask learning.
• Existing methods lack the integrated capabilities needed for efficient enzyme discovery in complex biological processes.

Purpose of the Study:

• To introduce CACLENS (Cross-Attention & Contrastive Learning-enabled Enzyme Selection), a novel multitask deep learning framework designed to overcome the limitations of current enzyme screening methods.
• To enhance the prediction accuracy and efficiency of identifying functional enzymes for applications in biosynthesis and biodegradation.

Main Methods:

• Developed CACLENS, a multitask deep learning framework utilizing Customized Gate Control, contrastive learning, and cross-attention mechanisms.
• Integrated reaction type classification, EC number prediction, and reaction feasibility assessment into a unified enzyme screening pipeline.
• Applied CACLENS to identify Zearalenone (ZEN) degrading enzymes.

Main Results:

• CACLENS demonstrated robust performance in multitask prediction with reduced computational resources.
• Successfully predicted 10 potential ZEN-degrading enzymes, with one achieving over 90% degradation efficiency for ZEN and its analogue α-ZOL.
• Established a user-friendly web server for CACLENS (https://ai.caclens.com/) to facilitate enzyme discovery.

Conclusions:

• CACLENS significantly advances the high-performance screening of functional enzymes by enabling multimodal and multitask predictions.
• The framework accelerates the discovery of industrial enzymes, as evidenced by the identification of highly efficient Zearalenone degrading enzymes.
• The accessible web server empowers researchers to discover novel catalytic elements for various applications.