Measures of threading specificity and accuracy

Threading predictions for CASP2 target proteins were compared to their true structures using a series of precisely defined measures of agreement, calculated in a fully automatic way. Fold recognition specificity was calculated as the proportion of a predictor's "bet" that was placed on previously-known structures similar to the prediction target, as identified by a "jury" of well-tested structure-structure comparison methods. Values approaching 100% indicate that a prediction correctly identified the structural and/or evolutionary family to which a target belongs. Alignment specificity was calculated as the proportion of aligned residue paris in the predicted target-to-known-structure alignment that also occur in the structure-structure alignments produced by the "jury" methods. Contact specificity was calculated as the proportion of nonlocal residue contacts in the molecular model implied by threading alignment, that also occur in the experimental structure of the target. Alignment specificity and contact specificity measure the accuracy of a predicted 3-dimensional model. Values approaching 100% indicate that target residues have been assigned to the correct spatial locations and that the model is as accurate as possible for a threading prediction.