CACLENS: A Multitask Deep Learning System for Enzyme Discovery

Abstract Deep learning greatly advances large‐scale predictions of enzymatic structure, function, and properties. However, existing deep learning models remain limited in high‐performance screen of functional enzymes, due to a lack of multimodal learning and multitask prediction capabilities. To address these challenges, CACLENS (Cross‐Attention & Contrastive Learning‐enabled Enzyme Selection) is introduced, a multitask deep learning framework incorporating Customized Gate Control, contrastive learning, and cross‐attention mechanisms. CACLENS demonstrates robust performance across three key functions–reaction type classification, EC number prediction, and reaction feasibility assessment with fewer computational resources. These three functions are seamlessly incorporated into the enzyme screening pipeline for efficient screening of desired enzymes in biosynthesis and biodegradation processes, thereby significantly expediting the discovery of industrial enzymes. Using CACLENS, 10 potential degrading enzymes against Zearalenone (ZEN) are predicted and expressed, and one of them achieves a degradation efficiency of over 90% for ZEN and its analogue α‐ZOL. In addition, a user‐friendly web server for CACLENS is established and is accessible at https://ai.caclens.com/ for researchers to discover catalytic elements.