Accuracy of secondary structure and solvent accessibility predictions for a clostridial neurotoxin C-fragment

Earlier studies used Rest and Sander's artificial neural network [(1993a), J. Mol. Biol. 232, 584-599] to predict the secondary structures [Lebeda and Olson (1994), Proteins 20, 293-300] and residue solvent accessibilities [Lebeda and Olson (1997), J. Protein Chem, 16, 607-618] of the clostridial neurotoxins. Because the X-ray crystal structure of the 50-kDa C-terminal half of the heavy chain of tetanus toxin was recently determined, this report evaluates the accuracy of these network-derived predictions. For this predominantly P-strand-containing fragment, predictions, on a per-residue basis, for both secondary structure and solvent accessibility were about 70% accurate. A more flexible and realistic analysis based on overlapping segments yielded accuracies of over 80% for the three-state secondary structure and for the two-state accessibility predictions. Because the accuracies of these predictions are comparable to those made by Rest and Sander using a dataset of 126 nonhomologous globular proteins, our predictions provide a quantitative foundation for gauging the results when building by homology the structures of related proteins.