Multi-purpose enzyme-substrate interaction prediction with progressive conditional deep learning

Understanding and modeling enzyme-substrate interactions is crucial for designing enzymes with tailored functions, thereby advancing the field of enzyme engineering. The diversity of downstream tasks related to enzyme catalysis calls for a computational architecture that actively perceives enzyme-substrate interaction patterns to make unified predictions for multiple objectives. Here, we introduce MESI, a progressive conditional deep learning framework for multi-purpose enzyme-substrate interaction prediction. By decomposing the modeling of enzyme-substrate interactions into a two-stage process, MESI incorporates two conditional networks that respectively emphasize enzymatic reaction specificity and crucial catalytic interactions, facilitating a gradual shift in the feature latent space from the general domain to the catalysis-aware domain. Across various downstream tasks, MESI consistently outperforms state-of-the-art methods on top of a unified architecture. Furthermore, the proposed conditional networks implicitly capture the fundamental patterns of enzyme catalysis with negligible additional computational overhead, as evidenced by extensive ablation experiments. With the support of this conditional perception mechanism, MESI enables cost-effective and accurate identification of active sites without requiring any structural information, highlighting enzyme residues and substrate functional groups involved in diverse and critical catalytic interactions. Overall, MESI represents a unified prediction paradigm for downstream tasks related to enzyme catalysis, paving the way for deep-learning-based catalytic mechanism cracking and enzyme engineering with strong generalization and interpretability.