Developing computational methods for assigning protein function from tertiary structure is a very important problem, predicting a catalytic mechanism based only on structural information being a particularly challenging task. This work focuses on helping to understand the molecular basis of catalysis by exploring the nature of catalytic residues, their environment and characteristic properties in a large data set of enzyme structures and using this information to predict enzyme structures' active sites. A machine learning approach that performs feature extraction, clustering and classification on a protein structure data set is proposed. The 6,376 residues directly involved in enzyme catalysis, present in more than 800 proteins structures in the PDB were analyzed. Feature extraction provided a description of critical features for each catalytic residue, which were consistent with prior knowledge about them. Results from k-fold-cross-validation for classification showed more than 80% accuracy. Complete enzymes were scanned using these classifiers to locate catalytic residues.