Homogeneous Refinement leads to 2D Density

Hi everyone,

I have been trying to do a single particle reconstruction of an RNA-protein complex. It appears that the sample is heterogeneous so I did ab initio reconstruction followed by heterogeneous refinement then 3D classification (BETA). The 3D classification appears to give a consistent density for my sample, if I perform heterogeneous refinement on the sample I can get ~ 14 Å resolution reconstructions.

However, if I try to do a homogeneous refinement from the sample it leads to a 2D density at ~ 22 Å resolution. Any idea what is going on?


Thanks,
Dylan

I’m a little confused about the workflow so let me know if I’m interpreting it wrong. You did a single-class ab initio followed by some number of classes of a heterogeneous refinement and then took one of those classes for 3D classification? Did you do additional heterogeneous refinement after 3D classification? Have you tried 3DVA and tried selecting a subset of particles from there for a more homogenous reconstruction?

I’m also confused about the 22A structure output from homogenous refinement. You said it’s 2D density that you’re seeing? I don’t know how that could come from a 3D refinement. Is it possible that you’ve just opened the wrong file in your favorite visualization software?

Another parameter that could help troubleshoot the low resolution is knowing the number of particles you have to start with, and the number of particles that went into or was output from each subsequent job.

Hi Jenchem,

So the workflow goes as follows:

  1. I performed ab initio with 5 classes (I assumed there was heterogeneity in the sample). (82, 597 particles)
  2. Heterogeneous refinement on the 5 classes from the ab initio. (82, 597 particles)
  3. 3D classification on the particles from heterogeneous classification. (82, 597 particles)
  4. Homogenous refinement on a similar classes from 3D classification. (14, 234 particles or 20, 631 particles), performed heterogenous refinement too.

I have also done additional heterogeneous refinement after 3D classification which does not give the problem that homogenous refinement is giving me.
I can try the 3DVA to select the subset of particles from there.

For the 2D density I have opened from the homogologous reconstruction I have opened all of the map outputs and they are all 2D except for precision.mrc.

I have added the number of particles at each step.

Thanks,
Dylan

It sounds like you did a 5-class ab initio and then heterogeneous refinement on one of those classes? Perhaps you could try a 1-class ab initio and then try to parse the heterogeneity after getting a solid ab initio. With only 400k particles total it might not be able to output good ab initio classes if you split them 5 ways.

What is it that made you initially think the complex is heterogeneous?

82k particles isn’t that many/ almost certainly isn’t enough to divide into 5 classes.

Also, do the 14A volumes look like they could be your complex when you download them? Or do they look like noise? At that kind of resolution the FSC number can often be off by a fair bit, so the “22A” number you see in your homogenous reconstruction job might not be surprising. You may simply just not have enough particles under your classification scheme (splitting things 5 ways) to see much. I’d try fewer classes.

The sample is suspected to have a few confomers from SHAPE probing data that we have. I will try with 1-class ab initio and if that doesn’t work I will collect more data.

Thanks

Processing RNA is challenging. Except for some very particular molecules, it seems to be a struggle to get good particle picks, good ab initio results, good 2D & 3D classifications, and good (half-map controlled) refinements. You should be very skeptical of any 3D reconstructions, checking carefully that they are robust.

At the beginning I recommend you use the EMAN2 neural net picker (it only requires < hundreds of examples, and classifies on radial averages), downsample to a pixel size ~3-4Å, and perform several rounds of 2D classfication. The approach worked the best for me until I understood the data very well. You can see also that e.g. Wah’s lab structures use several rounds of 2D classification.