As is often the case with TIM barrels, we had a lot of difficulty
aligning the strands to orient the polar residues the right way.
We had no trouble forming very pretty barrels, but they often had
polars buried under the helices.

After fussing with alignments and sheet constraints, we finally got a
barrel model that seemed adequate.  At that point we noticed that the
remainder of the protein was still disordered.  We chopped it off as a
separate domain (something we should have done much sooner), and
optimized it separately.  We first tried splitting at H236, but
decided that the helix N253-L271 belongs to the barrel domain, so
split again at G274.  

We did further optimization of the barrel domain by itself, though the
automatic prediction from the domain as target was pretty good.  We
ended up cutting-and-pasting in a strand from a model we had
previously generated, though, since K50 and Q52 faced the wrong way in
the moels generated automatically from the domain alone.

The second domain did have a fold-recognition match that optimized
fairly easily, so we pasted the two domains back together and did an
optimization run with undertaker to fix up the breaks and clashes
between the domains.

This is try12-opt2, the best-scoring model with an unconstrained
undertaker costfcn.  The E22 residue has somehow worked its way to the
outside of the barrel, though it is in the right phase of the strand
to be on the inside.