Abstract
Nanozymes-artificial nanomaterials with enzyme-like catalytic properties-offer superior durability compared to natural enzymes, enabling broad applications in biosensing and therapeutics. However, their rational design remains challenging due to immense structural heterogeneity and a reliance on trial-and-error. While machine learning (ML) accelerates materials discovery, conventional models often lack physicochemical grounding. We present a multi-output, physics-informed ML framework for predicting the core catalytic kinetics (Km, kcat, Vmax) of nanozymes. Using a curated dataset from NanozymeDB, we integrated the Arrhenius equation (activity-specific activation energies), pH-dependent Gaussian zones, and material properties (electronegativity, redox potential). Our model achieves robust predictive performance: R² = 0.555 for Km, 0.319 for kcat, and 0.524 for Vmax, with 55.6 of Vmax predictions falling within a factor of two of experimental values. Feature importance analysis confirmed that the Arrhenius factor (importance = 0.228) and pH zones (0.30 combined) dominate predictions, validating the physics-informed approach. This demonstrates that embedding thermodynamic constraints into ML enhances both predictive reliability and interpretability for rational nanozyme design.