Hi all,
Are all input particles fully used in a homogeneous refinement, or are some downweighted or not used? Is there a way to take the particles cryoSPARC is most certain about from a given homogeneous refinement and use only those to do another? Similar to how when you select a 2D class you can say take only the ones you’re 99% certain about.
I ask because for fun I ran a homogeneous refinement on my completely raw dataset (500k particles) and it yields a 3.2 Å structure that looks pretty much correct. Of course, ResLog reminds me that I absolutely do not need that many particles. I was curious to see if this would help screen decent particles from nonsense vs my usual methods of classification.
Cheers,
Sara
Hi Sara,
Good question.
In a homogeneous refinement (with default settings) there is no analog to the classification certainty score from 2D or 3D classification, since that certainty score is related to the relative similarity of a particle image to different 2D/3D references, but in homogeneous refinement there is only one reference to compare against,
In fact it’s a relatively open problem to be able to directly distinguish good from bad particles using only a single reference. Typically people use multiple rounds of 2D/3D classification to sort particles, but even there the best you can hope for is that there are multiple bad particles that all have common features and therefore form their own “junk” class that can be discarded. In reality, it seems that most junk particles are very different from each other and so they don’t aggregate into junk classes, instead they simply spread out over good classes.
We’re working on new methods to detect bad particles while doing a refinement, but those are still in research stages.
Hope that helps!
Ali