Building a preliminary model. This step varies considerably depending on the symmetry of your model.
First, use startnrclasses. This program will group your boxed out particles into self similar classes (reference free classification), then align and average the particles in each class. This will give you a set of low-noise images representative of the various projections present in your data set. If you didn't do it earlier, process your particles with cenalignint before running:
startnrclasses start.hed #-Ptcl-per-class [mask=radius]
Generally the number of particles in each class should be at least 20 or so. Usually 40 or 50 classes will be enough to analyze, so the total number of particles divided by 50 is usually pretty good. The optional mask allows you to exclude information outside a given radius (in pixels). This number should be slightly larger than the longest axis of your particle. If you don't provide a mask, the box radius will be used.
Next, you should manually examine the class averages in classes.hed. Select some class averages that look good, and put them in a file called good.hed. The precise number of classes to select depends on the particle. Generally at least 7 or 8 should be selected. You can take as many more as you like, but try to avoid particularly noisy ones, or duplicates of the same view. Once you have good.hed prepared, run:
startAny good.hed [sym=symetry]
If your particle has some symmetry, but was not amenable to the other methods, then specify it above. For asymmetric particles, don't specify sym=.
You should now have start.hed/img and threed.0a.mrc. The hardest part is over. This is what you need to start a refinement. Go on to step 3.