Good 2D classes but bad ab-initio reconstruction

Hello everyone,

I am currently trying to process cryo-EM data, and I struggled with particle picking, which I was able to drastically improve using Topaz training and extraction. After this, I got good 2D classes from a subset of images that I use for finding good parameters (I trained two models for picking top views and side views separately) :

Then I tried to use Ab-Initio reconstruction, because I thought that this subset was quite homogeneous, still using 4 classes to separate bad particles. The results are shown here:

As you can see, the Ab-Initio job was not able to build at least one clear 3D class from my particles. I also tried 3D Classification but it is not better. I also extended this analysis to the whole dataset (more than 150 000 particles used for Ab-Initio), tried to extend the box size for extraction, but still the same issue: good 2D but nothing with 3D reconstruction. Do you have any suggestion that I can try to obtain a better 3D reconstruction ? Any idea about the reason why it is so difficult despite good 2D classes ?

Thank you very much for your time and advices.

your 2D classes are decent but not good - though they have the desired ~shape suggesting the sample of interest has been found and picked, they don’t have any 8Å features. Even with amazing 2D classes of 1 million particles, the ab initio can still be quite difficult to achieve. “3D classification” requires already having good model and good assignment of particles to the model, so not expected to help. I would suggest using your best picks to train another iteration of Topaz, and at the same time you can try decoy classification in heterogeneous refinement - where you give all particles, one barely decent 3D volume from ab initio or elsewhere, and a varying number of complete junk volumes - maybe 4-10. This gives the program lots of “room” to move particles around, and in the process often guides any decent/correct particle to classify into one class and build a better model. There are plenty of other tricks if there is a much larger particle set, but 150k particles where many are not homogeneous is an uphill battle.

1 Like

Thank you very much for your answer, I will try your suggestions

Your boxes are way too tight. Double them at least. Looks like PSI/PSII or other large photosynthetic complex, which frequently show less-than-perfect 2D, but you’ll probably get much better results for 2D with a larger box.

3 Likes