Hi,
Multiple new reconstruction approaches (e.g. see AR-Decon: https://www.biorxiv.org/content/10.1101/2025.02.23.639707v1.full; GitHub - yifancheng-ucsf/AR-Decon: A deconvolution pipeline designed to enhance the quality of 3D cryo-EM maps that suffer from anisotropic resolutions.) used to mitigate map anisotropy benefit from being used in an iterative manner, ideally used to reconstruct the half maps used to initialize the next round of refinement/classification.
Would it be possible in the future to add support for a plug and play external reconstruction module in refinement and/or classification, to facilitate such approaches in difficult cases?
Cheers
Oli
4 Likes
Hi @olibclarke thanks for sharing! Just as a note for others who may read this: we don’t currently have an option for plug-and-play external reconstruction modules during refinement. One of the tricky parts of this is that with any external method in play, we can’t provide any guarantees about the validity of the results since reconstruction methods can easily hallucinate incorrect structure and also cause correlation between half-maps that can no longer be diagnosed with FSC and related tools.
Thanks!
2 Likes
Totally fair, it does add complications. Thanks!
If it were possible to like this multiple times, I would. I could not agree more.
EMDB has several examples now where I did a double take at the unmasked FSC and had to ask, “How did this get through validation?” and “Did the PI even check?” and that’s without “AI” routines (at least according to the methods of the paper in question) which can introduce all sorts of oddball effects (from various talks/presentations).
Agreed - I wasn’t really referring to “AI” routines or anything involving machine learning - just to things like AR-decon, sidesplitter etc, where there is no neural-network funny business involved. But once you open the door to one external reconstruction approach, you let the others in as well, and I understand that it could be a bit of a Pandora’s box situation…