Top-viewed and side-viewd particles seperated into 2 classes during Ab initial reconstruction

Hi, everyone. Recently I’m dealing witth a large membrane protein complex with distinct soluble features in 2D classes. And more importantly, densities of top views and side views seem to be the identical soluble features. However, when ab initial reconstruction was performed, top views and side views were classified seperatedly into different classes. And this happened even only 2 classes were assigned.

To avoid this problem, different parameters were tested for reconstruction, including:

  1. number of ab initial classes, we have tested 2,3 and 6;
  2. SGD momentum and sparsity prior, ranging from 0 to 0.8
  3. initial learning rate, ranging from 0.05 to 0.4

Currently, all parameters tested fail to generate reasonable models. I also tried to exclude all side views for reconstruction. In this way, densities of transmembrane region seemed to be better, but inclusion of top views to facilitate heterogenous refinement failed again.

Is there any other thing I can try to get better results?


Have you tried just running ab initio with a single class?

Other than that, I would suggest trying varying the initial and final resolution. For large, flexible membrane protein complexes, we often find setting the range to lower resolution works best (e.g. start 30, final 25).


1 Like

Hi, Oli.

Thank you for your suggestions. Actually I haven’t tried to set final resolution to values like 25, but a range of initial resolution is now under test. I shall continue testing different final resolution in combination with different initial resolution to see whether things can be better.

As for single class reconstruction, previous attemps proved that it didn’t work out well, perhaps ill particles intefered transmembrane signals of good particles when only one class was designated.

Thank you again for your kind advice!

If you know what you’re expecting to see, preselect only quality 2D classes which aren’t ill and comprise at least both if not more views. 1 class an initio. Then het refine 5x of the 1 volume with all particles and hopefully they don’t separate on pose but on quality and structure

Wondering if there was any resolution to this issue from the OP? I am having the same issue and have tried everything. My protein is not a membrane protein but an elongated beta helical protein where the top views looking “down-the-barrel” are much smaller than the side views which are quite long and thin. As stated in the OP, ab-initio is separating classes by views. Ive tried everything I can think of but would be curious to know if you ever found a solution.

Can’t speak for the OP, but in our experience switching off “Enforce Non-negativity” and “recenter in real space” often helps in such cases (switch off both parameters, which are enabled by default).

Still unclear why this helps, but it often (not always) helps.

1 Like