Exactly based on “Case Study: Yeast U4/U6.U5 tri-snRNP” in the cryosparc tutorial, I have run a 4-class ab-initial reconstruction. Using the output particle and output volume from each class from the ab-initial reconstruction step, I run a homogeneous refinement. For class 0 from the ab-initial reconstruction step, the resolution of the homogeneous refinement step was 6.52. For class 1 from the ab-initial reconstruction step, the resolution of the homogeneous refinement step was 5.77.
Using the output volume from each class from the ab-initial reconstruction step and all particles input for the above mentioned ab-initial reconstruction step, I run a homogeneous refinement. For class 0 from the ab-initial reconstruction step, the resolution of the homogeneous refinement step was 3.71. For class 1 from the ab-initial reconstruction step, the resolution of the homogeneous refinement step was 3.82.
We can see, with the same class volume from the ab-initial reconstruction step, if we used the full stack particles as input for the ab-initial step, the map resolution output from the homogeneous refinement step was much higher.
Then can we conclude that, the model volume used for the homogeneous refinement step, can heavily lead to the output model volume bias?
What happens if you use the ab initio volumes, with all the particles, to do heterogeneous refinement? Try using 2 copies of each ab initio volume.
I predict classes from heterogeneous refinement will do a bit better when refined separately. Ab initio is can use a random initialization, but it sometimes separates different views and doesn’t classify as well as reference based 3D classification.
This can certainly be the case, especially since different dynamic masks are generated and used for each refinement, so you can even see variance in the final resolution run-to-run with the same particles and initial volume because FSC calculation is mask sensitive. If each class showed a relatively minor conformational change, this could likely explain the 0.1 Å resolution different you saw. For larger differences between classes, it’s certainly true that the reference bias can exist. This can be reduced by tricks such as using a stronger initial low-pass filter.
Let us continue to discuss the issue. Exactly based on “Case Study: Yeast U4/U6.U5 tri-snRNP”, I got the 4 class-partcles and the corresponding 4 class-volumes with the ab-initial reconstruction. For class 0 volume, I run homogeneous refinement with the class 0 particle and I got the refined A.mrc. For class 0 volume, I run homogeneous refinement with all particles and I got the refined B.mrc. For class 1 volume, I run homogeneous refinement with the class 1 particle and I got the refined C.mrc. For class 1 volume, I run homogeneous refinement with all particles and I got the refined D.mrc.
Chimera checking showed all conformations from A.mrc, B.mrc, C.mrc, D. mrc were strongly different with each other.
Take the refined results from class 0 volume as example, do you think both A.mrc and B.mrc are reliable (practically speaking, the reliable model can be used to explain biological function without bias and error, and the reliable model was theoretically correct), or only one of them was reliable? If only one reliable, which one was reliable, and why?
You mentioned “This can be reduced by tricks such as using a stronger initial low-pass filter”. Here I would ask, to what extent of “low-pass filter” it can be regarded as “low-pass filter”, to what level of strongness it can be regarded as “a stronger initial low-pass filter”, and how to process step by step to get “This can be reduced by tricks such as using a stronger initial low-pass filter”?