Evaluation of disorder predictions in CASP5

Abstract

This paper reports an analysis of the accuracy of predictions of structural disorder received as part of the CASP5 experiment. Six groups made predictions of disorder. The predictions of the four most active groups have been compared with the experimental results, in terms of the sensitivity and specificity of the methods. All four methods succeed in detecting over half the disordered residues in the targets, with a generally low rate of over-prediction. Two of the methods perform significantly better when the structure of a related protein is available. There is a trade-off between the fraction of disordered residues detected and the extent of over-prediction, and groups have adopted different compromises in this respect. Comparison of performance at the same over-prediction rates highlights the role of related structures in some methods rather than others, with different groups achieving the highest sensitivity for different target sets. Over-all, the methods are clearly of considerable use in identifying potential disorder.