EMAN Wiki

This page is read only. You can view the source, but not change it. Ask your administrator if you think this is wrong.
**Q:** I have an asymmetric particle, how do I get a good starting model and refine it ?

**A:**
//Note: The EMAN2 workflow includes a method for automating the process of generating many initial models, automatically refining each with downsampled data, and assessing the results. This may be a much easier solution than the EMAN1 methods below. You can use EMAN1 class-averages with the e2initialmodel.py program in EMAN2, then take the resulting model back into EMAN1 (if you really want to move back to EMAN1).//

This is among the most difficult classes of single particle reconstruction problems. If you have both complete asymmetry as well as a particle with significant heterogeneity (dynamics or multiple states), the problem may be impossible using the standard approaches, and methods like single particle tomography might be required. If your particle is fairly homogeneous, you may be fine. The way to assess this is using refine2d.py as documented in several other FAQs.

There are two apporaches for tackling asymmetric problems. One, as documented in the 4 step EMAN tutorial interface (run the 'eman' program) is to try to come up with a good initial structure using common-lines, then refine it. The problem with this approach is that common-lines can produce an incorrect structure which is, at low resolution, still consistent with the data. If you happen to get one of these 'apparently correct' structures, it is very difficult for the refinement algorithm to work its way out of the local minimum. However, if you start with a really bad structure, in most cases you will end up converging to the right structure. So, while following the tutorial instructions and using refine2d.py combined with startAny is a good thing to try first, you also need to re-refine with some randomized starting models to insure that you are finding the best (and a self-consistent) answer. To make a randomized starting model, use makeinitialmodel.py. It takes no options and will prompt you through the process of generating a starting model.

For initial refinements intended to find the correct overall quaternary structure, I would suggest using heavily downsampled data for speed. That way you can run several different refinements with different starting models, and not spend a lot of time. Once you have the correct overall shape, you can scale the model back to its original sampling, and process with the full data. Typical goals for starting models would be 20-30 A resolution so a pixel sampling of 4-8 A/pix would be appropriate for this initial step. 

If you have a directory with a start.hed/img and threed.0a.mrc file in it, and want first to do some initial refinements, do something like this:
<code>
mkdir 1
proc2d start.hed 1/start.hed shrink=3    (3 is just an example)
proc3d threed.0a.mrc 1/threed.0a.mrc shrink=3    (or make a starting model in 1)
cd 1
refine ...

etc. for 2, 3, 4 or however many initial refinements you want to try. Say you are happy with the model you get from itertion 8 in directory 3:

cd 3
proc3d threed.8a.mrc ../threed.0a.mrc scale=3 clip=n,n,n   (where n is the original box size of your data)
cd ..

refine ...   (real refinement)
</code>

There are many ways to decide if you have a good or bad model. One method is to compare the projections to the class-averages (in classes.?.img). If ALL of the class averages are a very good match to the projections, then you probably have a good model (taking noise into account, of course. Another test is the 'convergence' plot run from the 'eman' browser. There are a number of additional tips in the tutorial.