DeepEMhancer – non-linear 3D reconstruction post-processing

DeepEMhancer is a deep-learning-based algorithm to perform automatic sharpening of unmasked, unfiltered reconstructions.

Inputs:

  • Required: 3D reconstruction
    • Recommended running information:
      • Use half maps not full reconstructions
      • Do NOT use post-processed or masked maps
  • Required: Select deep learning model for post-processing
    • wideTarget – trained on reconstructions with a large mask
    • tightTarget (default) – trained on reconstructions with a tight mask
    • highRes – trained on reconstructions for reconstructions <4Å
  • Optional: Mask
    • If you have a map that was masked or you are running into errors with deepEMhancer running with wideTarget, tightTarget, or highRes, providing a mask will help deepEMhancer calculate the noise statistics more accurately.
  • Optional: Select batch size for processing
    • IF your job is crashing with a CUDA error message in the STDERR output file, your volume size is too large for the default batch size of 6.
    • Try decreasing the batch size until the job runs successfully. For example, we have found that 512 requires a batch of 5.

Outputs

  • Post-processed ‘sharpened’ map that can be downloaded from the task output page
  • Example: