Heterogeneous refinement all classes turned to spheres

Very odd result in heterogeneous refinement. I’ve never seen this before. Same thing happened when I repeated a new job with the same inputs. These are ribosomes. Everything is OK until checkpoint 5, then checkpoint 6 preview images all look blank. The image is from the final checkpoint 7.


At checkpoint 5, the FSC curves go to Nyquist (~8Å). The final ones are noisy curves decaying around ~15Å.

More details:

Normal iterations up to 80.
Iteration 80 - 2 normal good classes, 1 normal bad class, 1 almost empty class.
Iteration 81 - All references are filled with 0, but FSC and particle counts are OK.
Iteration 82 - Random angle assignments based on empty references lead to spherical reconstructions. (Actually, they are ellipsoidal, there must be some kind of bias).

The zeroed references are otherwise normal MRC files with correct headers. There are no error messages in the log within the web UI.

HI @DanielAsarnow this is really bizarre - never seen this before.
When you say iteration 80 has one almost empty class, how empty are we talking?
Have you done any other processing with this same set of particles (2D class, abinitio, etc)?

They seem to have at least a few thousand particles in each class. Some of the reconstructions seem blank at earlier iterations, as well. I exported the job, perhaps I could upload it for you somewhere?

Other jobs seem to be running normally for a fairly high resolution ribosome dataset.