Hi, I’ve been having this problem from the very start. All classes seem to have a similar number of particles in ab initio and heterogeneous refinement. I don’t think the particles are very similar as seen from 2D classes. Is there a way to get a better classification for the particles in ab inito and hetero. refinement? Or is it because my dataset is too heterogeneous and doesn’t have any preference for a given class and thus gets assigned with equal probability?
Would be great to get some insights/tips
Thanks in advance!
Sunanda
Hi again, since I did not get reply yet, I assume that its a problem that has not been faced by others or I havent provided the relevant information. Would be happy to do so. I’m attaching the output of my recent heterogeneous classification, just to highlight how the particles sort by around 25% in each of the four classes
Hi Sue,
When you say “I don’t think the particles are very similar as seen from 2D classes”, what do you mean?
Do you mean that the 2D classes look different from one another (which is expected even in a homogeneous dataset as they represent different views of your 3D particle), or do you mean that you can see clear differences between equivalent views (e.g. dissociation of a subunit or movement of a domain).
It may be that your sample is genuinely homogenous (and hence particles are evenly splitting between identical classes), or it may be that your initial model is of poor quality - it is difficult to tell without more details. E.g. what do your 2D classes look like? Do you see clear secondary structural features? Does your ab initio model look sensible and protein-like? etc.
Cheers
Oli
Hi Sue,
I was wondering if you find out a solution for this problem. I’m having the same situation. Ab initio or heterogeneous refinement is splitting the particles equally in the number of classes. When I do 2D classification on the output of those classes I see in all of them similar good 2D classes as well as bad classes. Thanks.
Hi Juliana,
That time I was working on a very heterogeneous dataset, with an additional problem of flexibility. After improving my sample I did have a better particle distribution in ab initio. For this I did one ab initio run (3-4 classes) and used the particles from the best class to do a further ab initio (2-3 classes) which gave me a better ‘classification’. The first ab initio had equal distribution of particle problem, but it got better with the second round and I could see a clear improvement in the map. But I think some ‘good’ particles also get lost with the junk in this process, but not a problem for me as I had a lot of particles.
Once I get a consensus refinement I do 3D classification without alignment in relion and that gave me better results than heterogeneous refinement in csparc.
Do you have 2D classes with nice features?
Hope this helps.
Best,
Sue