Table 6 shows the differences between the RS126 and CB396 set of proteins. For all methods, the average accuracy drops by between 1.3% (NNSSP) and 2.7% (DSC) for the CB396 protein set. The NNSSP and PREDATOR programs used in this analysis were trained on larger numbers of proteins than RS126, and so should be less degraded by evaluation on a different test set. However, these methods still show a decrease in accuracy with the CB396 set.
Table 6 illustrates that the percentages for the segment overlap correlate well with the Q3 values. The SOV score for the consensus method is somewhat higher (74.5%) than the previous published value for PHD of 72%[64]. Table 6 also shows that the PHD method does exceedingly well at predicting segments. As measured by the SOV method, it is on average 2% better than any of the other methods tested. The difference between the consensus and PHD segment overlap scores is smaller than the corresponding Q3 value. This may be due to segment overlap being a more sensitive method to assess secondary structure predictions, or that the PHD method is the only one that has been optimised to predict segments scored by Equation 2.
Table 6 summarises the differences between SOV and Q3 accuracies for each method on the RS126 and CB396 sets. Although Q3 shows a consistent reduction on moving from RS126 to CB396, SOV shows a general improvement. Of the individual methods, NNSSP increases in accuracy the most (0.7%) while the consensus method increases by 0.9% to 75.4% SOV accuracy.