T20S Proteasome: Deep Particle Picking Tutorial

Step 1 - Preprocess Data

  • Preprocess the T20S subset by completing steps 1-12 in the tutorial found here:
  • Ensure that the Inspect Picks, and Select 2D jobs from the linked tutorial are completed as they will be required for training the deep picker. The ouputs from these two jobs will be used as inputs for the deep picker jobs.

Step 2 - Create Training Job

  • Select Deep Picker Train (BETA) from the Job Builder. Drag and drop the micrographs output from the completed Inspect Picks job and the particles_selected output from the completed Select 2D job into the micrographs and particles inputs respectively.
Passing inputs to the Deep Picker Train job
  • Change the Number of epochs parameter to 50.
  • Queue the job.
  • The job is training a deep convolutional neural network model on the subset of 20 micrographs from the T20S tutorial. It is highly recommended to train deep picker models on subsets of micrographs as acquiring training picks for all micrographs takes time and is not required. Once the model learns on a sufficient subset of the micrographs, it can pick particles from the entire dataset.

Step 3 - Create Inference Job

  • Select Deep Picker Inference (BETA) from the Job Builder. Drag and drop both the deep_picker_model and micrographs outputs from the Deep Picker Train job into the corresponding inputs on the Job Builder.
Passing inputs to the Deep Picker Inference job
  • Queue the job.
  • The job is using the trained model to infer picks from the input micrographs. Even though this tutorial job is picking from the same micrographs used to train the model, a properly trained model will pick particles that were not used as training picks from the micrographs.

Step 4 - Acquire Particles from Inference

  • Select Extract from Micrographs from the Job Builder. Drag and drop both the micrographs and the particles outputs from the Deep Picker Inference job into the corresponding inputs on Job Builder.
  • Queue the job.
  • This job will update the particle picks with information required for further processing.
  • Select 2D Classification from the Job Builder. Drag and drop the particles output from the Extract from Micrographs job into the particles input of the 2D Classification job.
  • Queue the job.
  • Select Select 2D classes from the Job Builder. Drag and drop both outputs of the 2D Classification job into their corresponding inputs in the Select 2D classes job.
  • Queue the job.
  • Wait for the job status to change to "Waiting" and then select the particle templates that should be kept for further processing.
  • The 2D Classification and Select 2D classes jobs serve to filter out unwanted particles from the particle picking. Once the Select 2D classes job is complete, the particles output from the job can be used as particle picks to process further into the pipeline.

Next Steps

Now that a basic deep picking pipeline has been completed, the more advanced aspects of particle picking with the deep picker can be explored. The following are some of these aspects:
  • Ideally, deep picking models are trained on a subset of micrographs and then perform inference on an entire dataset, as mentioned before. The deep picking model trained in this tutorial can be applied to the entire T20S dataset rather than the subset used in this tutorial. Potential refinement results will improve with the resultant increased number of picks.
  • The Deep Picker Train (BETA) job has many training parameters that can be fine tuned to affect the quality of the model. See the job documentation for more details on the training parameters.
Last modified 1yr ago