A novel computational approach was examined for predicting epitopes from primary structures of the seven immunologically distinct botulinum neurotoxins (BoNT/A-G) and tetanus toxin (TeTX). An artificial neural network [Rost and Sander (1994), Proteins 20, 216] was used to estimate residue solvent accessibilities in multiple aligned sequences. A similar network trained to predict secondary structures was also used to examine this protein family, whose tertiary fold is presently unknown. The algorithm was validated by showing that it was 80% accurate in determining the secondary structure of avian egg-white lysozyme and that it correctly identified highly solvent-exposed residues that correspond to the major contact regions of lysozyme-antibody cocrystals. When sequences of the heavy (H) chains of TeTX and BoNT/A-G were analyzed, this algorithm predicted that the most highly exposed regions were clustered at the sequentially nonconserved N- and C-termini [Lebeda and Olson (1994), Proteins 20, 293]. The secondary structures and the remaining highly solvent-accessible regions were, in contrast, predicted to be conserved. In experiments reported by others, H-chain fragments that induced immunological protection against BoNT/A overlap with these predicted most highly exposed regions. It is also known that the C-terminal halves of the TeTX and BoNT/A H-chains interfere with holotoxin binding to ectoacceptors on nerve endings. Thus, the present results provide a theoretical framework for predicting the sites that could assist in the development of genetically engineered vaccines and that could interact with neurally located toxin ectoacceptors. Finally, because the most highly solvent-exposed regions were not well conserved, it is hypothesized that nonconserved, potential contact sites partially account for the existence of different dominant binding regions for type-specific neutralizing antibodies.