Threading algorithms attempt to solve the inverse protein folding problem: given a group of structures and a sequence, identify the structure that is most compatible with this sequence. A recent study of this class of algorithms by S. J. Wodak and colleagues suggests that while threading algorithms are capable of recognizing many folding motifs, their performance in truly blind predictions is disappointing, and the underlying alignments upon which the selections are based are frequently errant. To help overcome this problem we have developed a Test of Optimal Mutagenesis algorithm (TOM) that exploits information inherent in the variation between several homologues in a multiple sequence alignment. This information is used to help select the correct structural motif for the sequence from a database of known structures. A total of 305 high-resolution structures were selected to represent the set of known folds; 56 proteins were chosen that had at least one close structural match in this set. To test TOM, we attempted to determine which of the 305 folds was a match to each of the 56 protein sequences. TOM correctly predicts a close structural match for 45% of these proteins. THREADER, an algorithm chosen as a literature standard, correctly matched 20% of the test set. By comparing the performance of TOM, THREADER, and TOM NOVAR (a version of TOM without variability information), we conclude that the tendency of an amino acid to be buried or exposed is the dominant determinant of the success of threading algorithms. In addition, the structural alignments produced by TOM suggest that the exact alignment of just 30 to 50% of the residues in a sequence with the correct fold is necessary to select it as the highest scoring match in a set of folds.