Evaluating a set of protein structure predictions is difficult as each pred
iction may omit different residues and different parts of the structure may
have different accuracies. A method is described that captures the best re
sults from a large number of alternative sequence-dependent structural supe
rpositions between a prediction and the experimental structure and represen
ts them as a single line on a graph. Applied to GASP2 and CASP3 data the be
st predictions stand out visually in most cases, as judged by manual inspec
tion. The results from this method applied to GASP data are available from
the URLs http://PredictionCenter. Ilnl.gov/casp3/results/th/ and http://www
.sanger. ac.uk/similar to th/casp/. (C) 1999 Wiley-Liss, Inc.