IsoNet

IsoNet is a deep-learning-based algorithm that tries to restore the missing wedge in cryo-ET tomograms through an iterative process.

Usage: To guide segmentation, particle picking, and other tomographic features that will be extracted from the unmodified reconstruction.

Disclaimer: IsoNet is a generative neural network. This means it generates information based on how you train the neural network. At best, it is trying to use other views of objects to try to fill in the missing wedge. At worst, it is generating artifacts.

For general information on COSMIC2 job submission, please see here.

Data preparation for IsoNet on COSMIC2

At the moment, we expect that you will perform data preprocessing on your local machines. This means that you will run these steps locally:

isonet.py prepare_star

isonet.py deconv (if you choose to use this option)

isonet.py make_mask

isonet.py extract

After these steps, you will have randomly generated extracted areas from your sub-tomograms in a directory (e.g., ‘subtomo’).

  • Parameter tuning: 
    • CTF deconvolve: snrfalloff
    • Make mask: density_percentage and std_percentage
    • Extract: cube_size
      • Recommended: increase to 80 or 96 when running on COSMIC2. (Default = 64)

Place STAR file into directory containing extracted sub-tomograms

Place the STAR associated with your subtomo data into the subtomo folder. COSMIC2 expects the STAR file to be within a given directory during upload. In the example, the STAR file after extract is: subtomo/subtomo.star.

Upload data using Globus

When you are ready, use the “Globus Upload / Download” button to upload your data to COSMIC2.  Follow this video tutorial on how to use Globus for uploading. 

For IsoNet, use the Globus navigator on COSMIC2 to click the subtomo folder and then click upload.

After upload, you will see the STAR file displayed in the data page.

Running IsoNet Refine on COSMIC2

Select IsoNet as the tool in addition to your subtomo STAR file.

Approach for iterations and denoising – indicates how many iterations and noise levels to use:

    • Regular: (Works for nearly all use cases)
      • Iterations = 30
      • Iteration that start to add noise of corresponding noise level = 11,16,21,26
      • Level of noise after the iteration defined previously = 0.05,0.1,0.15,0.20
    • Exhaustive:
      • Iterations = 40
      • Iteration that start to add noise of corresponding noise level = 11,16,21,26,31,36
      • Level of noise after the iteration defined previously = 0.05,0.1,0.15,0.2,0.25,0.3
    • Disabled:
      • Iterations = 30
      • Iteration that start to add noise of corresponding noise level = 0
      • Level of noise after the iteration defined previously = 0

Click Submit! Typical run times are 4 – 10 hours using four GPUs.

Monitoring job status: 

Be sure to check out this page to see how to check the running job status and the job’s current state.

Output

After completion, the .h5 model weight files will be found in the Globus upload folder named isonetOutputXXXXX with a time stamp (large numbers are at later timepoints). You can download .h5 files using Globus.

Predicting on tomogram data

Using the .h5 weight file from IsoNet refine, you can run predictions on your remaining tomograms.

  • Parameter tuning: To improve resulting tomograms, you should consider modifying the cube_size

Software citation:

“Isotropic Reconstruction of Electron Tomograms with Deep Learning.” Yun-Tao Liu, Heng Zhang, Hui Wang, Chang-Lu Tao, Guo-Qiang Bi, Z. Hong Zhou. bioRxiv 2021.07.17.452128; doi: https://doi.org/10.1101/2021.07.17.452128. Link.