All alignments discussed so far had gaps in the query sequence and residues in the corresponding column below each gap removed. Although this appears severe, it was standard practice for earlier prediction methods such as NNSSP . As shown in Table 2, including gaps in the query sequence reduced the accuracy of prediction by 1.1%.
This behaviour may be explained, because regions where there are gaps in the query sequence, are most likely to be in the coil state, and as such have no effect on the prediction. The gapped alignments have another significant draw back in that the profiles are considerably larger than the profiles with the gaps removed, and as such take much longer to train.