$0
In this step we analyze the quality of the model generated in step 3. In the next version of EMAN, you can expect a lot of this process to be automated, but for now, you'll have to do the analysis yourself. There are several steps to this process. First, you should have a look at the class-averages generated after each step of the reconstruction. Start by running eman from the same directory you ran refine from.
Press the 'Browse Files/History' button. This will cause the main browser window to appear. In the lower left will be a list of all of the files in the current directory. In the lower right is a list of the command history for this directory. In the top of the window are the various inspectors. Start by selecting the classes file for the first iteration, classes.1.img in the file list. Then select the 'File' inspector tab. You'll see the first image in the file displayed in the upper left. The classes files contain projections and the corresponding class averages. The first image is a projection, the second is a class average, the third is a projection, etc. First, just to get an overview, hit the Big View button. This will let you look at all of the images in the file at once. In the first iteration, the class averages may not match the projections very well. Look through the other classes.?.img files and see if things seem to be improving.
For a more qualitative view, select the Compare Images tab. When this inspector is in 'classes' mode (selected to the right of the images), a projection will be displayed on the right, and the corresponding class average will be displayed on the right. The difference between these 2 images is displayed in the middle. When the refinement has converged, the projections and class averages should be nearly identical except for noise. That is, the center image should look like flat noise without any recognizable structure.
If the class averages look ok, you can take a look at the 3d reconstructions. There are 3 volume images generated in each iteration: threed.?.mrc, threed.?a.mrc and x.?.mrc. The third image will be generated only if you specified the xfiles option in the refine command. The first file is the raw reconstruction, the second is the reconstruction with iterative real-space corrections, and an optional mild low-pass filter, and the third has been cropped to twice the mask radius, optionally scaled to a particlular mass, and optionally aligned to an earlier iteration. If the third set is present, they are the best ones to examine. Otherwise look at the second set. If you have some other program (like Iris Explorer) you like to use to examine MRC volume data, go ahead and look at the reconstructions after each iteration. The EMAN file browser will allow you to examine the 3D model in projection. Select the 3D model in the file list, then use the right mouse button to rotate it. v4 allows you to compare several models side by side. For isosurface views, and better comparisons, we suggest vis5d (http://www.ssec.wisc.edu/~billh/vis5d.html). A copy of this program is included in the EMAN distribution for your convenience, although it is a completely separate product, and credit should be given to the authors of that package if it is used for publications, etc. EMAN includes mrc2v5dt to convert MRC data to a format compatible with vis5d. To use this program, just do:
mrc2v5dt <mrc file 1> <mrc file 2> ... <output .v5d file>
If you recall, in step 1 it was suggested that you apply a high-pass filter to the raw particle data. If you followed this advice, we now need to remove this filter to really get a proper view of the final reconstructions. If you generated x-files, they should already have been corrected for this factor. Otherwise, this can be accomplished with the following command:
proc3d threed.1a.mrc x.1.mrc unhp=1
If the 3D reconstructions seem to be converging towards something, you should
take a look at the convergence more qualitatively. Run eman and select
the directory the refinement was run in. Then select Convergence on the
Analysis menu. This will calculate the the Fourier Shell Correlation
between the 3D models generated after each iteration. It will also generate
one for the even/odd models generated by the ttest command (if present).
All of these curves will then be plotted. What you will generally see is a curve
starting high at low resolution, gradually falling off at higher resolution,
and eventually levelling off near zero. Occasionally you will see this curve
rising again at high resolution. This is not indicative of improving resolution
in your model, but rather undersampling at high resolution in Fourier space. If
you see this effect occuring at a lower resolution than your goal in the
reconstruction, this means you need to use a smaller angular increment (ang=
in the refine command. The point at which this curve begins to fall off
rapidly is a good indicator of your final resolution.
For a more reliable resolution measurement, however, a more robust test
is required. EMAN contains an automatic routine for performing a 2-way t-test.
EMAN does this by generating 2 class averages for each cls???? file. Each
of these averages will use only 1/2 of the images in the clsfile. Two
3D models are then generated, threed.te.mrc and threed.to.mrc.
Comparing these models will give an accurate estimate of the reconstruction
resolution. Note that if you have a barely sufficient number of images in each
class before the t-test, the t-test's results will give an inaccurate
resolution (it will be lower than it should be). The class alignment procedure
really requires a minimum of 8-10 particles to work reliably. If you have
less than twice this many particles in many of your classes, the t test is
likely to underestimate your resolution. The options in the ttest
command are taken from the options used in the refine command. We currently
suggest using a fixed .5 FSC threshold for determining resolution in the
convergence plot. At present there is no option for plotting a 3-sigma curve,
due to the large uncertainties in resolution estimates made with this
criteria. Keep in mind, however, that real-space resolvability is actually
~2.2 times better than the Fourier resolution cutoff. Qualitatively this means if
you use a fairly strong criteria, say, .7-.8, and determine a resolution of, for
example, 18 Å, you will actually be able to (barely) resolve 8-9 Å
features in the map. Coincidentally, if you use the 3-sigma criteria, you will
generally find a measured resolution somewhere near this value. The community
needs to get together and settle on a consistent and reliable resolution
criteria. For now, understand the issues involved, publish your FSC curve, and
be self-consistent.
If everything has gone well, you now have a self-consistent model at some resolution. However, this does not say that you have the correct model, and, of course, you may not have the optimal resolution yet. If your particle has at least 1 C3 symmetry and you've determined this much properly, your model is likely to be correct. Asymmetric particles are more difficult. There are cases where several different models can be constructed from the same set of data and all will appear to be consistent with the data. This generally occurs when the particles have a strongly preferred orientation. In any case, you must be very careful about making conclusions about asymmetric models. Additional tests that should be performed in this case is beyond the scope of this document.