EnzHier: Accurate Enzyme Function Prediction Through Multi-scale Feature Integration and Hierarchical Contrastive Learning

Hongyu Duan, Ziyan Li, Yixuan Wu, Bozhen Ren, Wen Chen, Fanghua Wang, Dongming Lan, Yonghua Wang, Li C. Xia

Abstract: Accurate enzyme function prediction is essential for enzyme design and discovery. Existing methods face challenges with understudied enzyme families and multifunctional enzymes. We present EnzHier, a machine learning model that combines multi-scale feature integration with hierarchical triplet loss to predict Enzyme Commission (EC) numbers. By leveraging the hierarchical structure of EC classifications and multi-scale sequence similarity, EnzHier captures both local motifs and global sequence patterns. EnzHier outperforms state-of-the-art methods by achieving a 23% higher F1-score on benchmark cross-validation, and exhibits superior generalizability in external validations. The model also has high performance with challenging cases, correctly classifying difficult halogenases and identifying multifunctional enzymes like farnesyl pyrophosphate synthase–areas where other models often fail. Overall, EnzHier provides a robust and interpretable tool for enzyme function prediction, particularly for solving challenging cases where previous methods have shown limitations. EnzHier is publicly available at https://github.com/labxscut/EnzHier.

External IDs:dblp:conf/isbra/DuanLWRCWLWX25