In the misclassification tests we test each similarity level alone against
the set of negatives (members of other folds). So we can measure of the
performance of the methods on the different similarity levels. As we
perform these tests on the consensus hierarchy between CATH and SCOP
the levels are well defined.
In the superfamily misclassification test we use only the similarities of the
superfamily level (i.e. no family members - no very closely related proteins,
but still closely related proteins) and test a given query against the template library.
So we have only two possibilites for the best hit:
- it may come from the same superfamily (but different family) - this would be the good case
- or it may come from a different fold - the method reports that a domain in
an other fold is more similar than the all members of the superfamily