Hi,
I have been working with a dataset of around 79000 particles which gives good 2D classes.
I would like to mention that using the default parameters in 2d classification module didn’t give good results but as mentioned in the blog, I tried changing following parameters:
class2D_sigma_use_white to false
class2D_force_max to false
class2D_clamp to true . This produced very good 2D classes. I can see a lot more features in 2D classes than I was able to see using other softwares.
The size of the protein is small ~ 110kDa.
Doing ab-initio (single class) produces very good model.
But if I do ab-initio (2 classes), the program produces models which separates the top views from the side views ( I checked it by doing 2D classification again on the output of ab-initio).
Further, when I do refinement from ab-initio models, the FSC plot shows resolution of 7-8A but the structure looks worse than the ab-initio with spikes and protrusions.
Based on the previous discussion on the forum I have tried following:
=> lowering the initial low-pass filter to 60A.
=> I have tried lowering the dynamic_mask_start_res to 16A
=> increasing the refine_num_final_iterations to 5
=> also tried applying the symmetry: C2
I am looking for any suggestions to help improve the results, or any comments if I am doing something wrong.
Have you tried using heterogenous refinement starting from your good initial model? That can be useful as a means of separating different conformations, and as a way to eliminate junk (for the latter, initialize with one good model and some unrelated models).
How does the orientation distribution look, is there a severe preferred orientation issue?
Thanks for the reply. I have tried doing heterogenous refinement into 3 classes. It produced classes with ~34% particles in each class and shows ~8.6A resolution in each. In chimera, as I increase the threshold, it shows lots of spikes.
About orientation distribution: ab-initio splits the data into 80% and 20%. It has preferred orientation issue.
For this, I tried to take the equal number of top views as of the side views. It produces similar kind of ab-initio model, which looked good but again the refinement shows lots of spikes and protrusions.
Hmmm - maybe give heterogeneous refinement a go as a means of removing junk (as opposed to just separating conformations)? I have had luck with this for small particles, see here for description: https://twitter.com/OliBClarke/status/880227665797672960
I’m interested in using heterogeneous refinement to remove junk. So when you say you initialize with one good model and some unrelated models, do you mean that one should start with multiple initial models in heterogeneous refinement (1 good model and say 3 junk/unrelated models) and then classify the datset into the same number of classes (4 in this case if you start with 4 initial models)?
Hi Sameer - yes, if I understand you correctly. I often start with one good model (from ab initio), or two if there are two very distinct entities in the dataset (e.g. dimer v trimer), and 4-6 junk classes. These can be either taken from a previous run (e.g. failed ab initio runs), or just different structures that are unrelated to the target - it doesn’t really seem to matter. In all the cases I’ve tried so far, this gives a very efficient cleaning of the dataset. Sometimes it requires a couple of rounds - just keep going until it converges (should end up with 99%+ in the good class, essentially nothing in the junk classes once the set is clean).