An abundant number of nanomaterials have been discovered to possess enzyme-like catalytic activity, termed nanozymes. It is identified that a variety of internal and external factors influence the catalytic activity of nanozymes. However, there is a lack of essential methodologies to uncover the hidden mechanisms between nanozyme features and enzyme-like activity. Here, a data-driven approach is demonstrated that utilizes machine-learning algorithms to understand particle-property relationships, allowing for classification and quantitative predictions of enzyme-like activity exhibited by nanozymes. High consistency between predicted outputs and the observations is confirmed by accuracy (90.6%) and R (up to 0.80). Furthermore, sensitive analysis of the models reveals the central roles of transition metals in determining nanozyme activity. As an example, the models are successfully applied to predict or design desirable nanozymes by uncovering the hidden relationship between different periods of transition metals and their enzyme-like performance. This study offers a promising strategy to develop nanozymes with desirable catalytic activity and demonstrates the potential of machine learning within the field of material science.