The same neural network architecture was trained with PSIBLAST profiles. PSIBLAST generates two types of profile, a simple frequency, and a position-based scoring matrix. Both were examined, and the results are shown in Table 4.
The predictions based on the PSIBLAST alignments were 0.5% more accurate on average, than the predictions from the CLUSTALW alignment method (see Table 3). When both predictions were combined in an arithmetic sum, the average accuracy rose to 73.4%, Table 3. However, when the position specific scoring matrix of PSIBLAST was applied, the accuracy improved from 72.1% to 75.2%. The HMMER2 [45] package was then used to re-score the CLUSTALW alignments. This raised the accuracy to 74.4% over 71.6%. The scoring schemes used in both PSIBLAST PSSM profile and the HMMER2 profiles are more sensitive than using simple frequency counts, as both apply scoring methods that use prior knowledge of amino acid relationships, derived from the BLOSUM62 matrix. In addition, the sequences are weighted by the amount of information they carry [40,45,47]. This has the effect of removing redundancy in the alignments which has previously been shown to improve prediction accuracy [34]. Position specific scoring schemes have also previously been shown to be more successful in sequence searching [41].
The alignments from PSIBLAST gave more accurate predictions (75.2%) than those derived by the CLUSTALW alignments, re-scored as HMM profiles (74.4%). The alignment approach described in the Methods section (Figure 2), was applied to retain the divergent sequences found by PSIBLAST. The predictions from these `progressive' alignments were compared to the original predictions derived from the CLUSTALW alignments. No significant improvement was found. The alignment method applied internally by PSIBLAST could not be improved upon in this work, as assessed by secondary structure prediction accuracy.