The effect of two different alignment methods, CLUSTALW and AMPS [42] is summarised in Table 2. AMPS may be run in several different modes but in this study we used normalised alignment scores to generate the tree order, and the corresponding alignment. For the same sequences, predictions from alignments for CLUSTALW were 1.1% more accurate for the same cross-validated test than those for AMPS. This test shows CLUSTALW to be more favorable for generating multiple sequence alignments for secondary structure prediction. However, AMPS alignments are improved [42] when randomisation is performed to generate SD scores to define the sequence similarities, which can then be used to order the tree for the subsequent alignment. As this process is particularly computationally intensive it was not investigated here.
Table 2 also summarises the effect of two different methods for scoring the alignment profiles; simple frequency counting of residues in the alignment, and scoring positions by their BLOSUM62 matrix scores. The frequency matrix was 1% more accurate than the matrix scored by BLOSUM62 mutation scores regardless of the alignment method.