This may just be more brainstorming than feasible but who knows. More often with proteins I am working with due to odd shape or being localized on a nanodisc/tube or liposome, centering the pick is an issue. While Topaz and cryolo do a good job, it is really challenging. What happen usually is the membrane bound or odd shaped molecule defines a center based on some high detail/bright spot. If I increase box size to be big enough and some other strategies, I can eventually center it and train better models.
Would it be possible to have an interactive 2D classification job in which the circular mask can be shifted across the 2D square. The output of that jobs would be particle stack with coordinates shifted to reflect the back calculated new pick center.
I would also be very interested in such a feature. Often I need to recenter particles in 2D (in the absence of a 3D volume) and the automated recentering is failing for some of my cases. What I typically end up doing is migrating to relion where an arbitrary extraction center can be manually defined in the extraction job. The interactive feature suggested above would be really cool, but the possibility to define a new shifted extraction center would already be much appreciated.
I’ve also encountered the problem “What happen usually is the membrane bound or odd shaped molecule defines a center based on some high detail/bright spot” when I was doing the 2D classification with the proteins I am working with. May I ask what strategies you’ve used to approach this other than Topaz and Cryolo?
Please find below some suggestion made by @kstachowski related to that topic.
"As a possible solution and work around, you should be able to perform a 3D reconstruction (ab-initio follows by non-uniform refinement) and following that, provide a shift to all particles (not just individual classes) by obtaining a new voxel centering location and shifting with the particles and associated volume to a new location. This is based on the assumption that you will be able to get at least a medium resolution reconstruction of the head and portion of the tale. I would choose a location about halfway between the head and the end of the tail.
Items needed - volume from reconstruction, particle stack from reconstruction, and then the new center for the volume/particles.
After this you would need to reextract the particles with the new 'centers". I would then try to perform a homogeneous reconstruction followed up by local refinement. I should mention this is guaranteed to work, because the head has good SNR and appears to be rather rigid. 2D and 3D classification or refinement methods that use particles shifts in their job types will likely shift the particles back to put the head at the center of the box.