Multiple rounds of 3DVA


Do I need to run a homogeneous refinement after selecting particles from cluster 3DVA display mode for a second round of 3DVA or can I use the output particles and volume directly?

Best regards,

Depends on what you want to achieve but generally speaking, yes, you can use the particles and the volume directly for refinement. Do make sure that your particles used for refinement are not downsampled. Generally 3DVA runs much faster using adequately downsampled particles.

1 Like

Hi Nuno,

I don’t think you need to run a refinement (for alignement) in-between two rounds of 3DVA, but I usually do it, believing that this new round of refinement will be more efficient than the first one.

1 Like

Thanks for the replies!


Hi all! Just a quick addendum to the excellent replies:

3DVA assumes that all particle alignments are fixed. Depending on the type of heterogeneity present and cluster size, it can be useful to do a refinement for a particular cluster to improve the alignment quality. However, this may not always be the case (e.g., the cluster may have too few particles which may lead to worse alignments than the original consensus refinement).

Out of curiosity, @nbustorff, can you say a bit more about your multiple-round 3DVA approach? Are you running 3DVA again on each one of the clusters? Have you found this to be a useful strategy to separate classes further?



Dear @vperetroukhin,

I have been using exactly this strategy. It is very useful to further separate classes. Although, I recently became unsure about the true meaning of the clusters. What I mean by this is the following.

Although the job seems to work. For some masks, the 3D map from the reconstructions look artificial in the mask region. They look like they have mask induced artifacts. I am a little puzzled.

I am working with translation of ribosomes. One of the many things I tried was to cluster particles that have the tRNA occupied differently in the PTC. Either they have A/P/E sites occupied.

I am working with downsampled 256px particles because our computers can’t handle full particles of 512px.

The way I create the mask is:


  • Open volume of Homogeneous Refinement job
  • Load PDB of 70S ribosome with tRNAs
  • Fit PDB to volume
  • Delete all chains except tRNA
  • molmap of pdb with 15A onGrid volume
  • Save map


  • Import volume
  • Volume tools to create mask

    input: map
    output: maks
    threshold: depending on the molnmap volume (normally: 0.1)
    fill holes: true
    dilation: 3
    padding: 6

Mostly I run 3DVA with 3 modes. Depending on the result components graphs I choose the number of clusters for the 3DVA display job.

I am super happy with the 3DVA! I would like to understand it a bit better to make sure everything is working as it should and that I am not introducing artifacts.

Thanks a lot.


Thanks for the description!

Although the job seems to work. For some masks, the 3D map from the reconstructions look artificial in the mask region. They look like they have mask induced artifacts. I am a little puzzled.

Ah, strange. Could you share a screenshot of these artifacts?

Dear @vperetroukhin,

I hope these figures can help. I wish that the density is real, but as you can see in figure 10, it is not very convincing.


3DVA Job:

Figure 1: 3D refinement of input particles plus masking region

Figure 2: 3D refinement of input particles plus masking region (90° turn)

Figure 3: 3DVA inputs

3DVA display Job:

Figure 4: 3DVA inputs

Figure 5: 3DVA Consensus map input slices

Figure 6: Results mode 0

Figure 7: Result PCA distribution

Figure 8: 3DVA Cluster 0 reconstruction slices

Figure 9: 3DVA Cluster 1 reconstruction slices

Figure 10: 3D Homogeneous Refinement Cluster 0

Figure 11: 3D Homogeneous Refinement Cluster 0 + mask from 3DVA job

Ah interesting, thanks for the figures (and the numbering :slight_smile:!). It is quite difficult to tell if that’s real density, especially because of the relatively large negative values you see in Figure 9 . A couple suggestions:

  1. perhaps try expanding the mask to include more of the ribosome and then rerunning 3DVA, and/or
  2. try running a (masked) 3D classification and see if you observe similar behaviour.

Thanks for the suggestions. I will definitely try.