Fig. 6. : Agreement between NMR models and close homologues. Using one NMR model to predict the default DSSP assignment of another highlights the problem of discrete assignments. The influence of small structural variations between NMR models from the same protein was substantial, e.g., only 67% of all models had more than 85% of the residues assigned to the same DSSP state (b: 8 classes). When grouping the eight states into three classes, 95% of all models had more than 85% identical three-class assignments. Translated to the 20 NMR models usually deposited in PDB, this implies that 1-2 of them would have a three-class prediction below 85%. (d) The correlation was markedly higher for the continuous DSSP assignment: 11% of the protein pairs had an average correlation of 0.8 or worse using the DSSP assignment, while only 5% for DSSPcont (3 classes). (a, c) The agreement was considerably lower for X-ray structures of homologues proteins. This variation was most likely due to 3D-misalignments and sequence-induced structural changes. These results suggest that the assignment problem will dominate more strongly as predictors increase in performance.