Fragment-based analysis of protein three-dimensional (3D) structures has received increased attention in recent years. Here, we used a set of pentamer local structure alphabets (LSAs) recently derived in our laboratory to represent protein structures, i.e. we transformed the 3D structures into one-dimensional (1D) sequences of LSAs. We then applied Hidden Markov Model training to these LSA sequences to assess their ability to capture features characteristic of 43 populated protein folds. In the size range of LSAs examined (5 to 41 alphabets), the performance was optimal using 20 alphabets, giving an accuracy of fold classification of 82% in a 5-fold cross-validation on training-set structures sharing < 40% pairwise sequence identity at the amino acid level. For test-set structures, the accuracy was as high as for the training set, but fell to 65% for those sharing no more than 25% amino acid sequence identity with the training-set structures. These results suggest that sufficient 3D information can be retained during the drastic 3D->1D transformation for use as a framework for developing efficient and useful structural bioinformatics tools.