UniZyme: a unified enzyme property prediction framework via geometric deep learning and organism-optimum pretraining

Enzymes are central to industrial biocatalysis. However, most natural enzymes evolved under mild physiological conditions and are not well adapted to industrial environments. Recent AI-based predictors of enzyme properties facilitate the discovery and engineering of enzymes with desirable properties under industrial conditions. Nevertheless, existing methods usually model these properties as isolated tasks and rely solely on protein sequences, limiting their ability to exploit shared physicochemical determinants and capture structural mechanisms. Here, we present UniZyme, which integrates protein language model embeddings, surface features, and geometric structural information through a geometry-aware graph neural network, and jointly predicts optimal pH, optimal temperature, melting temperature, and solubility. We further incorporates organism-optimum pretraining to introduce coarse-grained ecological adaptation priors. Experiments on four public benchmarks demonstrate that our model consistently outperforms state-of-the-art methods and maintains strong generalization on low-homology test subsets. Additional analyses show that UniZyme learns a unified enzyme representation space while focusing on task-relevant structural regions. UniZyme provides an accurate and interpretable framework for enzyme multi-property prediction, supporting enzyme screening and engineering for industrial conditions.