I am working with a C15 particle that tends to “stick” together in a weird helical-like structures and I of course don’t want to pick these…I want to pick just the single ones. My micrographs contain these but of course in small numbers. From a previous dataset I managed to pick 25 k particles by first using manual picker, picking some 400-500 good particles, extract, 2d classify, and picking one top and one side view as templates in a template picker job. I have a 2.5 Angstrom construction now, and I thought I could use it to create templates to make my life easier as now I got a huge new dataset with 16000 micrographs. However, when I inspected the results after 2D classification of the picks, I could see that it failed miserably and picked the worst, polymerized helically arranged particles. I went over my new micrographs using manual picker, and I could pick almost 1000 single particles…so my micrographs should have many more…so why is the template picker failing to help with that? Is there any suggestions for improving my difficult particle picking process??? I can of course keep picking manually by eye, but its a tiresome process and takes a huge amount of time…
Train a topaz model - if you have 1000 manual picks that should be sufficient
I did not use Topaz, but I used the deep picker train with my previous dataset of about 5000 micrographs. I had 25 k particles, and I used the defaults. The first model I got was not good as I could see the training and validation curves diverge in later iterations. So I trained to use the generated model as an input in another deep picker train job, but for some reason I just got a flatout line (no curve)…so I guess that failed…then I just trained using a slightly higher number of micrographs (500 instead of 300)…I got a slightly better validation/training curve (I attached these curves). When I used this for subsequent inference and picking, I got worse particles than I did with the template picker with manually picked templates.
I would recommend using Topaz with your manual picks.