Automated structural elucidation of carbohydrates can significantly speedup biomedical research globally. Such a tool would have an impact on a broad range of basic and applied research ranging from understanding disease to drug development and even to homeland security. In this paper we present a machine learning approach using a mixture of probabilistic neural networks (PNNs) and a clustering that is able to elucidate the structure of PMAA from their gas chromatography-electron impact mass spectra (GC-EIMS). In this three-level hierarchical approach, we first identify the number of carbons in the molecule, then whether the sugar has a five or six membered ring structure, and finally the exact and complete structure of the PMAA (including the linkage information). In this paper we report an overall accuracy of 95.58 percentage in structural elucidation of D-Arabinose (ara), D-Fucose (fuc), D-Galactose (gal), D-N-Acetylgalactosamine (galNac), D-Glucose (glc), D-N-Acetylglucosamine (glcNac), D-Mannose (man), D-N-Acetylmannosamine (manNac), D-Rhamnose (rha), D-Ribose (rib), and D-Xylose (xyl), many of which are diastereomers. In this paper we also report 97.51 percentage grouping accuracy for sugars that were not present in the training set. Using this approach we have been able to demonstrate that our approach is able to fully identify the structure of PMAAs present in the training set and partially identify the structure of PMAAs that were not included in the training set
