Dear All,
I am trying to create a model for a topaz picking job. My particle is small (130A) and heterogenous.
I have tried autopicking the full dataset and after several attempts and cycles of class2D I have some classes that I show below. However when I train Topaz with this, the Average precision curve is going up and down with the epochs… What is causing such behaviour ? What can I try to improve this ?
With many thanks,
GIa
how many particles are we talking about? The 2D class all the way to the right seems to be the only one with reasonable features. If it has >10,000 particles, consider selecting only that class and running 2D again. Try using ~500-5,000 total particles for Topaz training, but make sure they are AMAZING particles. Unfortunately given you have only one good class shown, the best result from this will be better picking of that one orientation (particle pose). I would try to trim down to a tiny set of amazing particles for a couple other views also, if they exist. These 2D look a little strange, likely white noise model or force max poses or something. for above suggestions also try ~standard settings. These tiny jobs will run very fast.
In those classes there are 13000 particles (all of them). standard 2DCLass settings.
why do you think they look strange ? well actually, they are… the protein is very flexible.
and i probably don’t have amazing particles… or maybe yes, but they are hidden in a gigaton of junky ones.
I will try what you suggest.
But if I select only 1 class, would be representative of my particle ?
Thanks,
GIA
they just look a little too “white on black” for standard settings which I would guess would have more shades of grey in the particle and a dark or medium dark background.
Unfortunately no, if you select 1 class it will not be representative of the whole particle. I am trying to help get Topaz to work, and work well. however, all the views shown are ~the same view, and if there are only 13k then you will need more data, Topaz will likely not be enough of a boost.
I’m hopeful in the suggested experiment that Topaz is trained well to pick that one view, and maybe also helps get better particles of other views which can be used to train new topaz models.
