Hello, I have a difficult Glacios dataset - a very dense mixture of 40S particles and a sphere-resembling protein which is approx. 10x smaller then 40S.
Picking the smaller particle isn’t easy. I am trying to optimize the Topaz Train, using 100 micrograms and ~1000 manually picked objects. Custom parameters in the following job were:
Estimated Particle Diam. = 100 A
Expected Number of Particles = 200 (I am definitely not picking all of them per micrograph)
Number of Epochs = 20
While Training Precision seems to be OK, Test Avarage-Precision is very low and it does not rise towards the plateau. Any suggestions on how to improve on that?
run the manual picks through 2D 6 classes and pick only those which represent something believable (probably 1 class). can also take that one template to template pick to repeat the exercise to prepare quality particles for Topaz.
Old school particle pickers like DoGpick (appion) work amazingly well for this (ignoring large particles and getting all smalls).
Could you also get the blob picker to work?
I’m guessing “picking” all particles then curating will work fine, either by inspect picks or 2D class
Thank you for the feedback. While the blob-picker, followed by rigorous curating in 2D classification, worked decently in the end, I managed as well to get the Topaz Train working. The picks generated by Topaz were much better.
Interestingly, Topaz Train required non-denoised micrographs to perform well - comparison below.
Other customized parameters were: par_diam = 100, num_particles = 200, num_epochs = 30
I was using cryoSPARC 4.6.0 and micrograph denoising was done with the Micrograph Denoiser BETA.
