| advertisement: compare things at compare-stuff.com! |
The results for fold recognition using structure-mapped hydrophobicity were not particularly encouraging. Correct folds ranked consistently lower than in the trials using unmapped sequences discussed above. At one stage it seemed that alignment quality was improved using mapped information, but this difference disappeared when gap penalties were optimised. It is not clear why some pairs of domains appear to have more stable and better alignments using mapped sequences whilst others perform better using unmapped sequences. Since the numbers each way are similar (Figure 5.14), the behaviour is probably random, although further investigation will be necessary. One explanation could be that the mappings are more stable in these examples.
It has already been hinted at why structure-mapped library sequences are less suitable for fold recognition. In Chapter 4, we proposed that direct numerical comparison of sequence properties was more sensitive than indirect comparison using log-odds substitution matrices which encode the same information, but with loss. The mapping process smoothes and quantises the sequence information, also leading to a loss of information. It appears that the exact representation of the hydrophobicity information is more important than the extraction of essential features by the mapping process. In the next section we discuss what might be done to use the structurally relevant sequence features extracted by the self-organising maps whilst avoiding this degradation in alignment specificity.