One training data element. We can store sequence and known structure
(primary, secondary) as well as a predicted structure (stored, if
weight is how strongly this element should influence a
extendedKnowledge is True, if the data element comes from
a source which knows about extended secondary structures, like the PDB.
Otherwise it is False.
NOTE During training, one should not penalize non-canonical predictions in interior loops and multibranch loops, unless they hinder formation of true pairs.
TODO at some point we will move toward pseudoknots and other fun