A novel benchmark dataset for enzyme function prediction reveals the limitations of state-of-the-art models

Accurate computational prediction of enzyme function, standardized by Enzyme Commission (EC) numbers, is essential for large-scale genome annotation and generative enzyme design. However, it remains unclear whether state-of-the-art predictors learn the intrinsic structural determinants of catalytic activity or merely rely on global sequence similarity to annotated homologues. To address this gap, we introduce EnzymARC, a novel benchmark dataset of putative non-functional decoy sequences generated via structure-guided, systematic disruption of active sites (targeting catalytic residues and surrounding 5 A, 10 A, and 15 A radii) from experimentally annotated enzymes. We evaluated three distinct prediction paradigms against this dataset: homology-based annotation (DIAMOND), contrastive learning with protein language models (CLEAN), and a deep learning model incorporating non-enzyme discrimination (DeepEC). Our findings reveal that current models are highly vulnerable to phylogenetic shortcuts. Both DIAMOND and CLEAN exhibited false positive rates exceeding 90\% for low-perturbation decoys, confidently assigning the original EC numbers despite the destruction of the catalytic machinery. While DeepEC demonstrated improved sensitivity at higher perturbation levels, highlighting the benefit of negative training examples, all models struggled to identify targeted active-site disruptions. We demonstrate that modern EC predictors largely fail to distinguish catalytically incompetent variants from functional enzymes, and we propose that integrating structure-aware negative examples into both training and benchmarking is critical for developing functionally robust models in computational enzymology.