I may be missing the mark here, but my “canonical volume” output from the training looks quite good, with features which look like they are reaching the extent of the cropped training nyquist. When I go to run the Flex Reconstruct job, which has the un-cropped particles as input, I get a map which does not recapitulate these features.
Has anyone had these issues before? Any suggestions on how to remedy? I’m cautiously optimistic that 3D Flexible Refinement is learning something, I would just love to have it apply this at the slightly higher resolution.
Hi @NathanaelCaveney! I’d love to take a swing at helping you out here.
If possible, could you share some images of the features you’re noticing disappear between training and reconstruction? It may help me figure out what’s going on.
Additionally, do you observe realistic motions in the 3D Flex generate maps? Are the motions supported by 3DVA jobs (if you’ve run those)?
In any case, I have two main ideas as to what could be happening:
High-frequency noise
Is it possible that the features are in fact still present in the reconstruction volume, but are being swamped out by high-frequency noise? If this is the case, we’d expect to see them re-appear if you apply a low-pass filter (or just use volume gaussian {vol id} sd 2 in chimerax)
FSC filtering
It could be that the structure is still present in the reconstruct job, but is being FSC filtered away. Recall that the train job does not use half-sets, and so does not perform FSC filtering, whereas the reconstruct job does.
If your mask is especially large it may be that the FSC is too poor and the features are filtered away. A smaller mask may reduce the noise and therefore improve the FSC (with the usual caveats about mask tightness introducing FSC artifacts).
Please let me know!
2 Likes
Thanks @rposert - I think you are probably correct re FSC filtering. I managed to get the crucial info out of this via some local refinements, so this is a bit back burner for now. I will make sure to update this thread if/when I return to it and have anything useful to interject into the convo.
1 Like