Hello! I am currently struggling with an issue where for my given particle set, homogeneous reconstruction seems to produce a map with better resolution, a better GSFSC curve, and a better cFAR score than Homogeneous or Non-Uniform Refinement.
Does anyone have an explanation as to why this might occur, and what I could do about it?
You have duplicate particles in the homogeneous reconstruction - look at the FSC curve, it does not fall to zero. The estimated resolution and cFAR score for that reconstruction are meaningless (and dangerously misleading) because of this.
Look at the maps - the FSC tells only a small part of the tale.
Also check whether you imposed the correct symmetry during refinement.
Welcome to the forum. 
Lots of people happy to help with any questions. 
Once you’ll have removed the duplicates, try setting maximum alignment resolution to 6 or 4 angstroms (I mean… you can try both) to see how much the noise of high resolution is messing things up.
The first homogenous refinement looks good! But it does have particle duplicates in the half-sets as the FSC does not go below zero. I would recommend using the remove duplicates job and testing different cutoffs until you see the FSC drop below zero in a homogenous refinement job.
The other homogenous and NU refinements do look worse and could be caused by the initial volume not being aligned on a symmetry axis if symmetry is being applied.
I’m sure it is doable to get a good NU refinement with a good cFAR out of this.
Does homogeneous refinement and NU refinement check for duplicate particles and ignore them (as 2D classification does)?
It depends on what you mean by duplicate particles. If you connect multiple stacks of particles to the input of any job, any duplicate particles with the same UID will be filtered out. Check out this post for more information:
It’s important to note that these jobs only detect duplicate particles by checking for duplicate IDs, and they don’t actually compare the particle locations on the micrograph. For example, if the same particle is picked twice, it will get assigned two unique UIDs, and the duplicates won’t get filtered out. You will have to filter out the duplicates using a Remove Duplicate Particles job and specifying the minimum separation distance that must exist for two particle picks to be considered unique particles. As you mentioned, jobs like 2D classification also give you the option to specify a minimum distance between two particles.
If the duplicate particles are the result of a symmetry expansion, these particles cannot be used for any global refinement (heterogeneous, homogeneous, non-uniform). The only jobs that properly handle symmetry expanded particles are local refinement, 3D classification 3DVA, and 3D Flex.
You can find some more helpful information in these threads:
Hello all! Thank you for your help. After making sure to Remove Duplicate particles correctly, aligning my volume and particles with the correct symmetry axis, and further classifying in 3D, I have a homogeneous reconstruction whose FSC curve falls to zero:
However, this reconstruction still has a much better cFAR score than my homogeneous or NU refinements. Currently, I’m getting cFAR=0.61 on my reconstruction and no more than 0.1 in my refinements.
Nice, the FSC looks much better now! Can you clarify what the relationship is between the homogeneous reconstruction (which produced the top FSC plot) and the refinement (which produced the bottom cFSC plot)? I assume that the reconstruction is for a class from 3D classification, but was the refinement performed on the reconstruction?
Additionally, how many particles went into the reconstruction? What does the orientation distribution plot look like, and how does it compare to the consensus used for 3D classification?
No problem! I’ll go through your questions one by one so you have all the information you’re looking for. I appreciate your help!
The reconstruction and the refinements use the same particle set. The original reconstruction was the result of reconstructing a class from Heterogeneous refinement (generated initial volumes via Ab-Initio reconstruction of a particle subset and used the volumes as classes with the full particle set), and the subsequent homogeneous refinement of that class produced the bottom cFSC plot.
The reconstruction contains around 176,000 particles, and the orientation plot for the reconstruction is:
The refinement orientation plot is:
Given that I used Heterogeneous Refinement, I don’t think there is a consensus volume.
Hi rpat,
I can’t explain the difference in the cFSC score between the reconstruction and the refinement, but it’s important to visually inspect the actual maps. Does the refinement show more evidence of preferred orientation than the reconstruction? How does it compare to the actual map output by heterogeneous refinement? If you do a 2D classification on these particles, do you see an assortment of views? I wonder if some of the other heterogeneous refinement classes are soaking up rarer views.
In my experience, an orientation distribution plot that looks that homogeneous usually means that there’s still a lot of junk particles in the particle stack, because junk particles don’t typically show a preferred orientation. You may want to do further iterations of hetero refinement to remove more junk and get a cleaner particle stack, which can sometimes help with getting a better-quality reconstruction.
A “decoy classification” strategy for removing junk particles seems to be the preferred approach for several users on the forum. You can find more information about it here:
Best,
cbeck
