Question:
We have a bit of a problem with our refinement 
Some of the classsums look a bit off centered and it gets worse as I do more rounds of refinement. Our structure looks more funny as I refine further. We wonder if it's worth removing the bad classsums in the initial rounds of refinements to get the rough shape of the molecule (if yes, how?).
I started with a Gaussian sphere as initial model and used these commands:
refine 2 mask=25 pad=128 hard=25 ang=10 classkeep=1 classiter=8 median phasecls refine 10 mask=25 pad=128 hard=25 ang=10 classkeep=1 classiter=5 median phasecls refine 14 mask=25 pad=128 hard=25 ang=10 classkeep=1 classiter=3 median phasecls refine 16 mask=25 pad=128 hard=25 ang=8 classkeep=1 classiter=8 median phasecls refine 24 mask=25 pad=128 hard=25 ang=8 classkeep=1 classiter=5 median phasecls refine 32 mask=25 pad=128 hard=25 ang=8 classkeep=1 classiter=3 shrink=8 median phasecls refine 42 mask=25 pad=128 hard=25 ang=8 classkeep=1 classiter=3 shrink=8 median phasecls refine 44 mask=25 pad=128 hard=25 ang=6 classkeep=1 classiter=8 shrink=8 median phasecls
Answer: (dwoolford, Nov 20th 2007) - This has happened to me before - try generating a better starting model - use refine2d.py to generate some unbiased class sums, pick a bunch of good classes from the result, and then use startcsym (or is it startAny) to generate a starting model. You should be able to tell if you're getting the off-center problem very early in the refinement - i.e. just do several (or even 1) refinments, and check that the class sums look normal.
Answer: (dwoolford, Nov 20th 2007) - If you want to remove bad class sums from the classes.?.img, you can do it in v2 - just remove the class sums and their associated projection using the delete option, then save (to a different filename). Then run make3d with the pruned image - all that make3d requires is the Euler angles, which are stored in the .hed file. (Technically, you do not have to remove the associated projection image, but it is good practice to do so).
Answer:Another possible problem is contrast. Are the particles dark (negative) on a white background ? If so, you need to invert the density. There are places in EMAN where white objects on a dark background are assumed. With symmetry this problem doesn't happen much , since the symmetry 'anchors' things around the symmetry axis. Oh, one other thing to try. If you use xfiles=, the 3rd number is the number of an iteration to align the 3D model to. ie if you do 'refine 8 ...' after the first couple of iterations you can include xfiles=X,Y,2 which will align subsequent results to the results of iteration 2.