FAQ & Troubleshooting

(Work in progress)

More resources can be found by searching the github issues page.

1. The expected heterogeneity is not present in the cryoDRGN reconstruction.

   • There may be too many junk particles or other imaging outliers in your dataset.

2. How do I update cryoDRGN versions?

   • See this section of the installation instructions here.

3. How can I check on intermediate results during training?

   • See the run.log that is in the output directory. The cryodrgn analyze command may be run at any time on the intermediate epochs.

4. Can I use the loss function or learning curve to tell when my training has converged?

   • In general, no. The training curve is typically used to diagnose any training instabilities (e.g. spikes in the training curve). The loss function on the training set does not generally indicate convergence of the model.

5. Do I need to use backproject_voxel? (Step 4 in the GitHub)

   • We recommend running this step as a brief sanity check that the poses/CTF parameters have been parsed correctly. cryodrgn backproject_voxel can also be used to validate any particle selections.

6. I am seeing weird artifacts (e.g. spherical ringing) in the reconstructed volumes.

   • This is most likely an issue with the input CTF parameters, or some type of aliasing from signal pre-processing, e.g. in signal-subtracted particle stacks. Please reach out if you run into these artifacts or file a github issue.

7. Does cryoDRGN work with membrane protein complexes?

   • It does! However, we find that cryoDRGN can capture the heterogeneity of the micelle instead of the protein complex. You can try training on signal-subtracted images.

8. Does particle crowdedness affect the training result?

   • Particle crowdedness means signals from adjacent particles are included in cryoDRGN training. Particle crowding does affect the results (e.g. learning heterogeneity of neighboring particles). You can reduce the real-space windowing applied to the input images by setting --window .6 when using cryodrgn train_vae.

9. Is cryoDRGN sensitive to orientation bias?

   • As long as sufficient views are covered, cryoDRGN should be able to reconstruct 3D density maps.

10. How does one generally recognize a cluster as a junk cluster? Are there any visible features in the UMAP that can be used?

    • This requires some digging into the results and images. In obvious cases, you might be able to tell by looking at the maps or images. Sometimes, you'll need to do follow-up analysis to validate, e.g., 2D classification or follow-up refinements.

11. Do better-separated clusters more likely mean distinct conformational states? If a UMAP shows clusters that are close to one another (but with clear boundaries), can one interpret it as a set of conformational states that are very close to one another?

    • No, not in general. Sometimes there is "repetition" in the latent space where homogeneous states are split into multiple clusters (maybe there is subtle variation in the background). You may want to try out the new landscape analysis tool if you want to more quantitatively define conformational states.