Recent advances have enabled high-quality computationally generated structures for proteins with no solved crystal structures. However, protein function data remains largely limited to experimental methods and homology mapping. Since structure determines function, it is natural that methods capable of using computationally generated structures for functional annotations need to be advanced. Our laboratory recently developed a method to distinguish between metalloenzyme and nonenzyme sites. Here we report improvements to this method by upgrading our physicochemical features to alleviate the need for structures with sub-angstrom precision and using machine learning to reduce training data labeling error. Our improved classifier identifies protein bound metal sites as enzymatic or nonenzymatic with 94% precision and 92% recall. We demonstrate that both adjustments increased predictive performance and reliability on sites with sub-angstrom variations. We constructed a set of predicted metalloprotein structures with no solved crystal structures and no detectable homology to our training data. Our model had an accuracy of 90%-97.5% depending on the quality of the predicted structures included in our test. Finally, we found the physicochemical trends that drove this model's successful performance were local protein density, second shell ionizable residue burial, and the pocket's accessibility to the site. We anticipate that our model's ability to correctly identify catalytic metal sites could enable identification of new enzymatic mechanisms and improve de novo metalloenzyme design success rates.