I am wondering if there is a way to disable reprojections in heterogeneous refinement, or if a similar task can be done using a different job type. I am working with a dataset that has a mixture of conformations, some of which appear to be the same at low resolution for certain projections. I suspect that the low resolution is preventing efficient separation of these conformations, which is stalling my refinements at 8A resolution.
I was thinking I would try adding some Fabs to my particle to help with alignments, but it would be nice if I could address this in silico. I have high resolution maps of these conformations from different data sets, so I would like to try sorting particles using those high-res maps as the reference. It would be nice if particle alignment was done as well so I don’t need to guess the correct orientation for the reference maps. I assume if I set the initial low-pass resolution to 4 A that after the first reprojection at iteration 1, the resolution would go back down to ~8 A. I understand that this will introduce a lot of reference bias, but I think a subsequent ab-initio would address that.
In my hands, hetero refinement is not very efficient for separating structures that are very close to each other. I first try to understand the movement with 3DFlex and 3DVA (which might already separate the classes), then do masked 3D Classification - which might or not be reference-based. Adding a smart solvent mask at this point - one that envelops all the possible conformations that you can see - seems to help a lot. No matter which method, I keep resolution low (8 - 20 angstroms?) to avoid too much interference by noise and detailed movements such as side chains or others, which will distract the algorithm from the large scale movements that make the real difference between distinct conformations.
In the reciprocal space, resolution is about distances, not angles. So if a 5° movement of a helix makes a 20 angstroms difference at the opposite end, it should be detectable at very low resolution.
I’m still curious to read other people’s strategies, especially those who believe to have benefited from the high resolution at this point.
I would second the suggestion of 3D classification without alignments, although if the different conformations really look similar at low resolution, this may cause issues when using a low target resolution in 3D classification - so some experimentation may be required there.
Alternatively, you might try multi-class ab initio, starting at a higher resolution (say starting at 9Å rather than 35). If the relevant differences are in this resolution range, this might help sort things out.
Concerning heterogeneous refinement, what have you tried so far? Have you tried using the maps of both conformations as initial models for heterogeneous refinement, even with the default initial lowpass resolution?
Another way to separate conformations which we have found effective is to use identical starting maps in heterogeneous refinement, but with a more aggressive initial lowpass resolution (e.g. 40 or 60 Å), sometimes coupled with increasing the initial learning rate.
If I understand what you’re looking for, you have particles which are only refining to about 8 Å, and you believe this may be preventing them from separating into classes. You have high-quality maps of these classes from a prior investigation of the same/similar target, so you want to classify them using those maps.
How to do this
If I have all this right, you could try a 3D Classification job, and
Provide your high-quality volumes to the job as inputs (I recommend low-pass filtering them first, to at least 6-ish Å)
Set Initialization mode to Input
Set O-EM learning rate init to 0. This will prevent updating the volumes between iterations, meaning they stay the same as your input volumes. This learning rate will slowly anneal to 1 according to the O-EM learning rate halflife (%), which you could also set to 0 to prevent the annealing
This would still use your initial 8 Å pose alignments, but I think this is the best way to accomplish what you’re looking for.
I do not recommend doing this
However, I would caution against this approach, especially if you use high-resolution maps to classify your particles. If you use 3D Classification (or any job) to classify particles, you run a high risk of template bias, even if you later perform Ab Initio as you say.
For example, consider a particle stack of images which only goes to 8 Å. If you use a high-res 3D Classification to classify into a class that looks like a 3 Å map of your target and some number of other classes, you’ll keep particle images with noise that looks like the high-resolution template. When you later do Ab initio, these noise images will align well based on their similarity to your reference, even though you are no longer providing the reference. You could end up introducing spurious high-resolution noise which looks like signal.
Provide your high-quality volumes to the job as inputs (I recommend low-pass filtering them first, to at least 6-ish Å)
Just to add - this will only work at all if the orientation of these volumes matches with the orientations of the particle set you are classifying - as 3D classification does not modify the alignments (this may be obvious but just in case)