Fold assignments for amino acid sequences of the CASP2 experiment

New and newly extended methods for fold assignment were tested for their abilities to assign folds to amino acid target sequences of unknown three-dimensional structure. These target sequences, released through the CASP2 experiment, are not obviously related to any sequence of known three-dimensional (3D) structure. We assigned 3D folds to target sequences and filed these predictions with CASP2 before their 3D structures were released. The methods tested were (1) Environmental 3D profiles of Bowie and colleagues [Bowie, J.U., Luthy, R., Eisenberg, D. Science 253:164-170, 1991]; (2) A variation of this is termed Directional Profiles; (3) The H3P2 five-dimensional sequence-structure substitution matrix of Rice and Eisenberg [Rice, D., Eisenberg, D.J. Mol. Biol. 267:1026-1037, 1997]; and (4) The Sequence Derived Property methods of Fischer and Eisenberg [Fischer, D., Eisenberg, D. Prot. Sci. 5:947-955, 1996]. When the 3D structures of the sequences were released, 17 of our predictions were evaluated. Of these 17, we assigned high probabilities to 11, of which 9 were correct. Five of these correct predictions were of known 3D structures similar to the targets and four of these were of new folds. The evaluation demonstrated that our methods were effective in assigning the proper fold to more than half of the CASP2 targets with known folds (5/9) and also were able to detect half of the sequences that corresponded to no known folds (4/8). Even when the correct fold is assigned to a sequence, proper alignment of the sequence to the structure remains a challenge. Our methods were able to produce accurate alignments (< 1.2 mean residue shift error from the structural alignment) for four of the targets, including the particularly difficult alignment (only 7% residue identity in the structurally aligned regions) of the ferrochelatase sequence to the fold of a periplasmic binding protein.