We present results of structural modeling of the variable fragment of M alpha 2,3, an antibody capable of neutralizing all short snake toxins. Three different methods were used to model the hypervariable loops: the conformational search algorithm CONGEN (Bruccoleri and Karplus, Biopolymers 26:137-168, 1987), high-temperature molecular dynamics (Bruccoleri and Karplus, Biopolymers 29:1847-1862, 1990), and a combined knowledge-based and energy-based algorithm (Martin et al., Proc. Natl. Acad. Sci. USA 86:9268-9272, 1989). Ninety plausible conformations were generated and were clustered into 13 classes. The clustering results indicate that there was little overlap of the conformational space explored by the different methods. Canonical loop structures were found by all methods for two of the loops, in agreement with previously established empirical modeling criteria. Nine of the 13 classes of structure were rejected on the ground of their lacking common features of antibody combining-site structure. The remaining four models were refined using restrained molecular dynamics. It was found that interconversion between the four resulting structures is possible with no significant energy barriers, suggesting that they are in thermodynamic equilibrium at 300 K. Features of the combining-site structure likely to be particularly important for antigen binding are discussed.