"side view prior" for ab initio?


In cases where we have all side views, and no top views, of a disk-shaped sample, ab initio frequently fails in this mode, where we end up with a cylinder, and all the different side views are assigned to views orthogonal to the long axis of the cylinder.

Is there any way to introduce some kind of “shape” prior to prevent this? E.g. ab-initio, but instead of starting from random density, start from a disk, or similar?

In a case like this the underlying data is fine, and can give a high resolution reconstruction if refined against the correct model, it is just ab initio that is failing.

There is also another, related failure mode we have observed, where we have side views and “oblique” views, but no top views, whereupon the oblique views are assigned as top views. That is a little more dangerous, as it almost looks sensible (just a little squashed), and will give decent resolution (internally self consistent) but incorrect reconstructions upon refinement.



(If I recall right, there was an option in v0 to initialize ab initio from a starting model - am I misremembering that? If not that could maybe be brought back as an option…?)

@olibclarke – yep, you’re correct, an initial model for ab-initio was an option in v0. We’ve been doing some internal testing to see if initial models can improve ab-initio performance in different edge cases (including the type you mention here and the ‘high symmetry’ capsid-type volumes). So far the results are inconclusive…we still observe collapsing to certain preferred configurations even with appropriate initializations. We’re looking into adjusting the way we handle symmetry in the loss function. We’ll keep you posted!


Thanks @vperetroukhin! Just to clarify, this particular case (and others where I have observed this situation) do not exhibit symmetry, this was in C1 (and the object is intrinsically asymmetric)

Hey Oli,

Bit of a thread hijack, but I’ve also encountered the second failure mode you’ve seen. Any pointers for getting ab initio to properly align views in this scenario?

Nothing 100% reliable, but a couple of points:

  • Failures of this type tend to pop up more often in multi-class jobs than single class, with the same set of particles.
  • Experimenting with resolution limits is worthwhile. For small membrane proteins, we often get better results starting at higher res (e.g. initial res 9, final res 7), while for larger, dynamic complexes or cases like this where there is a degree of preferred orientation, lower resolution seems to work better. For example with the ankyrin complex, initial res of 30 and final res of 25 works best, while the defaults or higher resolution limits gave poor results.
  • If the lack of top views is the problem, it may be worth looking harder for them - try splitting top (or near top) and side views and 2D-classifying separately, to get a more balanced set of views prior to ab initio.

Hope that helps!


1 Like