Since there is a strong need for computational methods to predict and characterize functional sites for initial anno- tations of protein structures, a new methodology that relies on descriptions of the functional sites based on local prop- erties is proposed in this paper. This new approach is in- dependent of conserved residues and conserved residue ge- ometry and takes advantage of the large number of protein structures available to construct models using a machine learning approach. Particularly, the proposed method per- formed feature extraction, clustering and classification on a protein structure data set, and it was validated on metal- binding sites (Ca2+, Zn2+, Na+,K+, Mg2+, Mn2+, Cu2+, Fe3+, Hg2+, Cl-) present in a non-redundant PDB (a total of 11,959 metal-binding sites in 3,609 proteins). Feature extraction provided a description of critical fea- tures for each metal-binding site, which were consistent with prior knowledge about them. Furthermore, new in- sights about metal-binding site microenvironments could be provided by the descriptors thus obtained. Results using k-fold cross-validation for classification showed accuracy above 90%. Complete proteins were scanned using these classifiers to locate metal-binding sites. Keywords: Functional Genomics, Protein functional sites, Feature Extraction, Clustering, Classification, Metal- binding sites. Java source code available upon request. Supplementary Website: http://dis.unal.edu.co/~biocomp/metals/