If I use Volume Alignment to align symmetry, and use that aligned model with symmetry alignment disabled in homogeneous refinement, it works perfectly (first iteration looks like this)(1):
This doesn’t happen all the time (some datasets seem more prone to it than others(2), even for apoferritin) but I’m curious whether others have seen the same thing?
Anyway, since aligning to symmetry with Volume Alignment Tools solves the problem, this isn’t a major problem. That said, if others encounter it in the future I hope it serves as a guide to work around the issue.
…
(1) When you can tell whether something is good or bad on iteration zero is when you’ve run way too much of that sample…
(2) Thinking about it, I think it might be related to particle proximity as it seems more likely to happen on datasets where particles are very concentrated and the initial model has weak “ghosts” around it. As such, @olibclarke’s previous request for masking in ab initio would be nice to have (even if just to test whether this hypothesis is correct…)
Thanks for this detailed post, this is a great example of how strongly the initial model can affect results (esp. in the case of a symmetric structure).
To dig in a bit more, it would help us if you could answer the following questions:
Could you show the output slices or projections of the Volume Alignment Tools job that you used to align the ab-initio structure?
Did you set any parameters of the Volume Alignment Tools job aside from specifying octahedral symmetry alignment (e.g. lowpass filtering parameters)?
Could you try repeating the homogeneous refinement (with the ab-initio model directly inputted to it) but changing the initial lowpass resolution to be something in the range of 8-12 Å? Does this help the job converge to a correct solution?
We have noticed that with high symmetry structures, starting from a slightly higher initial lowpass resolution generally performs better. This appears to be because when solving high symmetry structures from a low resolution model, there can be a lot of run-to-run variance and sometimes refinement converges to a locally optimal solution that is clearly wrong. Since the underlying volume alignment code is identical in both jobs, I wonder if some of the discrepancy shown here is due to run-to-run variance (i.e. the volume is aligned in a satisfactory way both times, but due to particle ordering and other random-seed-dependent changes, one job converges and the other doesn’t).
No adjustments to VATs other than specifying octahedral symmetry. The ab initio model had a target res of 6 Å (I find that 6-8 Å gives more reliable initial models for smaller higher symmetry targets).
Interestingly, 15, 12, and 8 Å starts all align badly and give bad starts (and resulting maps). I tried 15 and 12 Å previously so was expecting that, but wasn’t expecting 8 Å to give a bad start alignment. Also tried 6; same result. But VATs gives a good alignment. Does it search over a larger range of possibilities?
Possibly; maybe related to particle input order? But 99% of the time if I get a good alignment the final model turns out well, even if it takes a while to converge.
Another question – in VAT, are you connecting a mask to the job at all, or not?
One more question – if you call VAT with the Initial shift grid sampling parameter set to 7, do you notice the alignment results deteriorate in any way?
I took a look at the difference in the code between these two call sites, and really the only thing different between these two is that:
Refinement uses Initial shift grid sampling of 7 (unmodifiable), VAT uses 9 by default. This just parameterizes the number of shifts (along one dimension) to search over initially.
VAT applies a mask to the volume prior to alignment. Refinement doesn’t apply the mask.
Internally we are also trying to recreate this issue.
