advertisement: compare things at compare-stuff.com!

Can we recognise all the recognisable folds?

So far, this study has paid little attention to the structural relatedness of the folds in our dataset. Inspection of hit lists shows that some queries yield a succession of correct top hits, whilst other queries of the same fold topology yield none. Are similarities being missed for these queries because they are too weak, or because the method fails to discriminate signal from noise for these examples? In Figure 4.4 the relationships between structural similarity and fold recognition performance are explored. Figure 4.4(a) shows that a greater number of similar folds (with lower RMSd) are more likely to be aligned with a higher Z-score. However, the majority of similar pairs, with RMSd$<5$Å for example, have been aligned unreliably ($z<1.0$). These are pairs of folds which the method has missed. Interestingly, the inverse situation also exists: reasonably high scoring alignments ($z>1.0$) are made for quite dissimilar pairs (RMSd$>10$Å), indicating that close structural similarity is not necessarily required in order for fold recognition to be successful. Figure 4.4(b) shows basically the same trend, this time using the fraction of aligned structure rather than RMSd, as a semi-independent measure of similarity. We cannot yet explain why so many pairs are missed, and will undertake an analysis of their alignments and look in more detail at the structural comparisons.

Figure 4.4: Plots of fold recognition results vs. structural similarity measures for all 562 alignable pairs (same topology) from the set of 78 query and 197 library folds, using hydrophobicity and secondary structure prediction information. (a) Alignment Z-score vs. RMS deviation of all structurally aligned C-$\alpha$ coordinates by SSAP. (b) Alignment Z-score vs. fraction of shortest domain aligned by SSAP. (c) Mean alignment shift error vs. fraction of shortest domain aligned by SSAP (see Section 4.4.5). (d) Mean alignment shift error vs. Alignment Z-score.