After doing 3DVA and displaying in cluster modes ( cluster =5 ), cluster 0-4.
Cluster 0 & 1 looks similar and cluster 2 & 3 also seems to be the same.
So I run a heterogeneous refinement to further classify these clusters.
After re-classification by heterogeneous refinement, the particle numbers change significantly in each cluster.
Cluster 0: 142974 -> 216415
Cluster 1: 142583 -> 41047
Cluster 2: 141894 -> 279346
Cluster 3: 170592 -> 37916
Cluster 4: 133705 -> 157024
The particles composed of cluster 1 & 3 were classified to other classes.
Will the 3DVA generate repeated clusters?
Should I trust heterogeneous refinement that some of the clusters are exactly the same?
And, can I combine those particles (ex: cluster 0 & 1) to do a further refinement?
The 3DVA Display job uses a Gaussian Mixture Model to cluster particles in the reaction coordinate space. Clusters can ‘overlap’ in this space, so particles from two different clusters may actually be quite close in this space.
Should I trust heterogeneous refinement that some of the clusters are exactly the same?
If pairs of clusters (0&1 and 2&3) are similar, the heterogeneous refinement job can indeed ‘collapse’ those two classes into a single class – this would explain why the number of particles in clusters 1 and 3 drop significantly.
In general, two clusters will never be exactly the same and identifying similar classes will probably require some domain knowledge and extensive processing through different classification jobs (3DVA, hetero refinement, 3D class). Note that heterogeneous refinement and 3D classification can also be sensitive to initial pseudo-random seeds, so that even between two different runs of the same job you may see different outputs.
And, can I combine those particles (ex: cluster 0 & 1) to do a further refinement?
You can try combining the particles from the pairs of clusters, running a homogeneous refinement or reconstruction, and then using those volumes as initializations in 3D class / heterogeneous refinement.
