The last decade has witnessed an unprecedented accumulation of proteins in large online databases which has led to the need for automatic prediction of protein function essential for massive and timely annotations of the proteins in these datasets. Protein databases, combined with functional annotations and machine learning (ML) techniques, offer many potential benefits, including significantly facilitating rapid pharmacological target identification. The main objective of this study is to identify, for the problem of enzyme classification, the most powerful combinations of descriptors taken from different protein representations. To achieve this objective, four approaches for representing the Position-Specific Scoring Matrix (PSSM) combined with three methods for representing the Amino Acid Sequence (AAS) are evaluated with the aim of experimentally producing a powerful ensemble of descriptors for enzyme function prediction. Each protein descriptor is classified by a Support Vector Machine (SVM), with the set of SVMs finally combined by sum rule. Cross-validation experiments using these descriptors on single-functional enzymes (n=44,661) extracted from the PDB database demonstrate that the ensemble proposed here achieves superior classification rates compared to state-of-the-art ML techniques reported in the literature on the same dataset. Although the proposed ensemble strongly outperforms these other techniques, it is computationally much heavier, mainly because the PSSM extraction process is time consuming. However, there is a growing repository of proteins where PSSM has already been extracted, making the proposed method more practical and attractive. The MATLAB code and the dataset used in the experiments reported here are available at https://github.com/LorisNanni.