The results of a blind test of the new Jnet and other prediction methods that apply multiple sequence alignments for prediction are shown in Table 5, and Figure 4. Predictions were made for 406 proteins not used to develop the methods (see Methods). On this set of 406 proteins, Jnet gave an average accuracy of 76.4%. This is 3.1% better than the best previous method (PHD, 73.3%) and 1.8% better than the Jpred [34] consensus method.
Table 7 summarises a closer investigation of the
differences between PHD and the Jnet method. From these figures it is
clear that while the Jnet method is more accurate than PHD, the
-strand state is not predicted any more accurately by Jnet than
PHD (0.1%). Most of the improvement is coming from the clearer
delineation of the helix and coil states, (1.6 and 4.1%
respectively).