spIsoNet

Overcoming the preferred-orientation problem in cryo-EM with self-supervised deep learning

What does it do?

spIsoNet trains and implements a generative artificial intelligence algorithm how to minimize the effects of preferred orientations on a given 3D reconstruction. In single particle analysis, missing views can lead to streakiness in 3D reconstructions because viewing directions are missing (and these missing views affect the alignment of other particles). spIsoNet tries to mitigate this by training an algorithm to learn to dampen the missing view artifacts.

Here is an example provided by the authors on EMDB entry 8731 (HA trimer). On the left is the original reconstruction and on the right is the spIsoNet Anisotropy Correction applied to the half map 3D reconstructions.

Note: COSMIC2 only offers "Anisotropy Correction" from spIsoNet as of June 2025. "Misalignment correction" will added in the future.

Learn more:

How is it different than other tools?

spIsoNet trains a deep learning algorithm to mitigate the effects of preferred orientation. A comparable tool on COSMIC2 is AR-Decon, which uses deconvolution to minimize preferred orientation problems. Noteably, AR-Decon is 5-8X faster than spIsoNet.

Running this tool on COSMIC2

Required inputs

Half maps - you will need each half map from a given 3D refinement.
Mask - a mask file is required.
Resolution limit - provide the highest resolution to use during the anisotropy correction

Outputs

3D FSC - This will be the estimated coverage of Fourier space as shown in a 3D volume that can be opened in ChimeraX.