Encountered exception when trying to mark failed 3D class job as completed

3D Classification
1

I’ve been running 3D classification jobs on large datasets with the following parameters:

  1. Number of classes: 5
  2. Filter resolution (Å): 3
  3. Convergence criterion (%): 1
  4. RMS density change convergence check: false
  5. Per-particle scale: input
  6. Initialization mode: PCA

With a wall clock limit of 48 hours these jobs don’t run long enough to converge but are able to be marked as completed and the resulting classes used with success. We set up a separate lane with a wall clock of 96 hours and these jobs get much closer to finishing where I’m happy with the sorting result, but when trying to mark the jobs as completed I get the following errors:

[CPU: 87.12 GB]

Finalizing Job…

[CPU: 87.12 GB]

Passing through outputs for output group particles_all_classes from input group particles

[CPU: 91.43 GB]

Encountered exception: Could not add rows to dataset

OR

[CPU: 91.65 GB]

Finalizing Job…

[CPU: 91.65 GB]

Passing through outputs for output group particles_all_classes from input group particles

[CPU: 91.65 GB]

Encountered exception: ‘Unknown dataset field uid or field type 0’

2

Hi @rabdella,

What are you attempting to classify and what led you set that combination of parameters? Additionally, what is the box size of your target?

Thanks,
Kye

3

Hi Kye,
I’m attempting to classify the rocking of a 40 kDa subunit on a 90 kDa protein. The resolution of the preceding local refinement is quite high, 1.9 Å but the distal portion of the 40 kDa subunit is averaged out due to flexibility. My best estimate is that the distal portion moves by ~5 Å. There is a loop that is part of the 40 kDa subunit that exists in two conformations depending on if it is engaged with itself or the 90 kDa protein. The input particles are 768 x 768 but the 3 Å filter resolution adjusts the box size to 288 x 288.

4

Hi @rabdella,

Thanks for that info.

A couple other Q’s.

  1. Is this a single protein or a complex?
  2. Would you be able to post an image of the your FSC plot and an image of your volume with with transparent masks covering the 90 kDa subunit and 40 kDa subunit?
  3. Have you tried 3DVA? Using 3DVA with a filter res of 6Å might show what your looking for. Volumes of 3DVA can be used as inputs to 3D-classification or you can refine those particles separately.

Here are some recommendations around parameterization for 3D-classification with your target and goal in mind:

  • Use a filter resolution of 6-10 Å (especially since you’re looking for domain motion as opposed to very tiny structural changes of a region). This will prevent CS from just trying to classify high frequency noise over a larger region.
  • Keep Convergence criterion and RMS density change at defaults - these are likely leading to your very, very long runtimes.
  • Other params are fine
  • Changing class similarity to a lower value might yield good results too - this will help to diversify the volumes as much as possible.

Additionally, you could run a 3D-classification with a filter res of 15Å or 20Å with a mask around your 40 kDa subunit to try to tease our a particle set where it is fully present and not dissociated from the 90 kDa component.

Best,
Kye