QS_global is the "original" QS-score, whereas QS_best only considers
residue that are actually mapped, i.e. does not punish if one structure
has an additional chain or more residues in one of the chains.

All structures, be it target or model structures undergo additional
processing so they only contain residues with complete backbone.
For peptides thats: N, CA, C, CB (no CB for GLY), for nucleotides
thats O5', C5', C4', C3' and O3'. Reason for that is to give some
guarantees to downstream algorithms.

One downstream algorithm that could be adapted because of that is
the transformation in the ReprResult class. If number of residues
is smaller than 3, all those backbone atoms are used to derive
a transformation in order to not crash. Behaviour for ReprResult
objects with >= 3 residues does not change. Only CA/C3' atoms
are used.

just keeping a list of ResidueView objects seems not enough to guard
the associated chain/entity views from being garbage collected

The reason for that is that one might start with an initial seed
somewhere in Nirvana that leads to no possible increase in QS-score.
Stoping there might hinder a non-zero QS-score when continuing with
other seeds. And even if the QS-score is 0.0 in the end, the mapping
should make sense from an lDDT perspective more or less...

Reason is that these modules add a dependency on scipy which should be
optional.

Besides raw mapping, this also contains alignments etc for further processing

Searches for representations of a target stubstructure in model. Main intent
for this function was to find most similar interfaces but in principle it can
also be used to implement the CAMEO lDDT-BS

see previous commits...

see previous commits...

see previous commit

HHblits unit test on Gitlab CI runner seems to give slightly different results
than on any other machine. This is a "debug commit" to trigger the CI pipeline
and see whether its the only difference or if there are more.