This is try6-opt2.  It did not score well with our cost functions, but
we include it as the only model we came up with that included much
beta structure (our neural nets predicted helices, so we don't really
believe in this structure).