3D classification without alignment

Hi @Lan, I refer to the parameters used for mask creation in relion. ini_threshold is the initial threshold applied to the input map (which is already lowpass filtered) used for binarization. extend_inimask x will make dilate the mask x A beyond this ini_threshold. I will set the threshold in chimera such that the lowpass filtered map encompasses the region of interest, then take note of the contour level, which is used as the ini_threshold parameter. Hope this helps!

1 Like

Hi User123,
Thank you very much. Sorry for the late. How would you judge which lowpass filter to be used in molmap pdb is the right parameter? L

Hi L, the mask should not have high resolution features, so molmap from 8-30A is probably fine. You want to generate a mask that encompasses the entirety of your protein density plus regions that may have movement, heterogeneity, low occupancy, etc. molmap 8A will be tighter than molmap 20A (the ini_threshold and extend_inimask also dictates this). Generally, probably better to go with a looser mask.

3D classification can take a lot of optimization to get optimal particle separation and tightness of the mask can sometimes be a factor. Every dataset is different and parameters can be tested iteratively to learn what works best. Small movements (like rotation of an alpha helix about its axis) can be detected with a relatively tighter mask than larger movements (like a shift of many angstroms). A tighter mask might be helpful in excluding areas of high heterogeneity, like a micelle or lipids, if that is an area you are not interested in and find its signal is dominating 3D classification results.

thank you user123 for your fast reply. At which stage, the 3D classification starts to work? Lan

Hi Lan,

The answer to this depends on both the data/sample and what you are trying to accomplish or sort out. Very generally speaking, 3D classification with a circular mask and alignments works well at early stages if you are trying to split classes with large differences in composition or conformation. 3D classification with fixed alignments and a tight or local mask works well for obtaining the best set of particles or the best set of particles within the specified mask once you already have a pretty good reconstruction and typically after polishing. Again, these are very general statements that aren’t always true and every dataset will require trial and error to figure out the best strategy.

Best,

Ryan

edit: when I say tight mask I mean a mask made from 15A lowpass filtered with a soft edge of ~6 pixels. For local masks, I like to segment the lowpass filtered map in chimera/segger and group based on smoothing to obtain the region of interest

2 Likes

I’m bringing this thread back from the dead for a minute, as I’m more seriously thinking about implementing cryoDRGN in my super heterogeneous dataset that cryoSPARC can’t seem to handle. Guillaume, could you comment on that floating density that shows up in the movies from the e-Life paper? Did you do some sort of masking that resulted in an artifact? Or is this a known physiological part of the complex? Also, does cryoDRGN let you output a random walk through the variability? Maybe this is what’s meant by the graph traversal you mentioned above.

Congrats on getting the paper out btw, the dynamics of the ATPase lobes next to the nucleosome look really cool and it’s really exciting that you got it out of less than 50k particles. It gives me hope!

I think the floating density on the other side of the nucleosome is a second ALC1 remodeler. It could be either one that is really there but more weakly bound to the other side (a reasonable hypothesis, given how heterogeneous the post-translational modification that allows it to bind is), so a nucleosome with two remodelers bound. Or it could be an image processing artifact due to the nucleosome driving the alignment: if you have a projection of a nucleosome in disc view, you can only distinguish “heads” versus “tails” if something big enough to see is bound on only one side, so it’s possible that we really only have 1:1 complexes, and a few particles with a more dynamic ALC1 bound on the same side were actually aligned wrongly because there wasn’t enough rigid density at the remodeler location to overcome the alignment being driven by the nucleosome. Does it make sense?

A paper on the CHD4-nucleosome complex (Nucleosome-CHD4 chromatin remodeler structure maps human disease mutations | eLife) reports something similar, and found that it was likely an image processing artifact and they really only have a 1:1 complex.

1 Like

Also, does cryoDRGN let you output a random walk through the variability? Maybe this is what’s meant by the graph traversal you mentioned above.

I forgot to answer this. cryoDRGN’s graph traversal is not a random walk. You give it a starting and an end coordinate in the latent space (and optionally intermediate coordinates that you want it to visit), and it finds the path between the two through nearest neighbors (and always with coordinates that correspond to observed particles).

I think cryoSPARC works very well with heterogenous datasets, and I’m not sure you will see the miracle with cryoDRGN.