Hi Klim,
In order not to spend too much time on the angular accuracy calculations 
(which are only a metric,  not an essential part of the optimisation 
itself), RELION only calculates those accuracies for the most highly 
populated classes. For all classes that contain less than 1 % of the 
particles this calculation is skipped and the values are set to a 
meaningless 999. So:
1) if you have many particles, classes with less than 1% may still look 
good and hence have good accuracies (as high-frequency components in the 
classes result in better accuracies)
2) no
3) no

HTH,
S


On 05/12/2014 11:01 PM, Kliment Verba wrote:
> Hi,
>
> as part of my 3D reconstruction in Relion I always "filter" my particle stack by running 2D classification and getting rid of the particles in low signal to noise looking classes. Looking through the resulting classes however, I have noticed a curious occurrence.
>
> Visually, I get roughly 3 groups of classes(starting with 120000 particles going into 150 classes): excellent,sharp looking classes with secondary structure, ok looking classes with hints of secondary structure (but even if without secondary structure, clearly my particles) and junk looking classes. However, when I sort the classes based on the Accuracy of Rotations, there are really only two groups of classes. The excellent, sharp looking classes with secondary structure have quite high accuracies (<3 degrees) and all the other classes show 999 in that field, even the ones which have hints of secondary structure.
>
> Now, if I take a class which visually looks good, but has 999 as the accuracy of rotations, and reclassify the particles(about 1000 particles) in that class using all same parameters into 5 classes, the accuracy of rotations for each resulting class end up being better than 3 degrees, and I get one class with secondary structure showing, couple reasonable looking classes and a junk class.
>
> So, I guess I have a couple of questions:
> 1) How is the accuracy of rotations metric generated? How can I have a class which looks good but has 999 as accuracy of rotations?
> 2) Does the above scenario simply mean that I need to use much more classes to start with?
> 3) Or does it mean that the classification have not yet converged? (I ran it for 50 cycles, resolution stopped improving and the particles per class stopped changing in the couple of classes I checked)
>
> Thanks,
>
> Klim

-- 
Sjors Scheres
MRC Laboratory of Molecular Biology
Francis Crick Avenue, Cambridge Biomedical Campus
Cambridge CB2 0QH, U.K.
tel: +44 (0)1223 267061
http://www2.mrc-lmb.cam.ac.uk/groups/scheres