Clustering in 3D variability display

Hi CryoEM/cryosparc team,

I am curious to know more about the ‘Clustering’ programme in 3D Var job display programme.
What I find is that the clustering is always following a ‘histogram’ like approach (where there is a ‘peak’ for a cluster with most particles, regardless of if it is dependent on the PCA component selected and not a discrete (black and white) separation. Essentially after selecting a PCA component to cluster my particles, is there a way to drive the clustering such that I can cluster a group of particles more in a discrete manner? Or I have to run multiple rounds of 3DVAR to pull out some states? Any extra info. around how the clustering function works would be helpful.

Many thanks,
Abid.

Welcome to the forum @javeda . Would you like to post an image of your clustering result?

Sure. For instance:

The two peripheral segments relate to the principal component selected: Whereas the middle peak (where majority of the particles get grouped) 3D takes me back to the 3D state where I started before 3D variability. I tried multiple rounds of 3D var on this state and hence was curious if I can run a clustering approach such that I don’t have to run multiple rounds of 3D variability.

The heterogeneity in this instance is the ‘occupancy’ + ‘flexibility’ of the sub-complex I am trying to resolve - hence trying the PCA approach.

Hi @javeda,

In 3DVA Display of a single mode the data are represented in a histogram (as you show) that plots number of particles assigned to each bin across the component. Generally, volumes from the central bins (containing the most particles) resemble mostly closely the input volume, and the less populated bins nearer the upper and lower bounds of the plot resemble the first and last frames that are obtained from simple mode 3D Var Display for that mode. Clustering in 3D Var Display is performed based on gaussian fitting.

You might like to try 3D Var Display with intermediates mode to output different particle sets across the component, and if you don’t want the particles to overlap between classes, you can set Intermediate: window (frames) to 0. Once you are happy with the range of volumes that you have you could also try using those volumes as references for 3D classification in input mode to cross-validate the 3DVA result of discrete separation.

Would this approach give you the sort of result you are looking for?

Hello,

This is helpful. Thank you. I am going to try your suggestion shortly and will update accordingly :slight_smile: