3D particle picking¶

nextPYP provides a suite of methods for 3D particle picking that includes size-based, geometry-based, template-search, and neural network-based methods. It also features an interactive interface to pick particles manually and import or export coordinates to use with external programs.

Main particle picking block¶

The Particle-Picking block takes Tomograms as input and implements four particle picking methods.

Import coordinates¶

nextPYP can import particle coordinates saved as IMOD models (*.spk) or xyz coordinates saved in plain text format (*.box):

1. Click on Tomograms (output of the Pre-processing block) and select Particle-Picking

2. Select “import” as the Detectiom method

3. Select the location to import the coordinates from

4. Set the Particle radius (A) (for visualization purposes)

5. Click Save, Run, and Start Run for 1 block

6. Once the run completes, navigate to the Particle-Picking block and confirm that the coordinates were imported correctly by checking the position of the green markers

Manual picking¶

nextPYP also provides a user-friendly UI to quickly pick particles from datasets with many tomograms:

1. Click on Tomograms (output of the Pre-processing block) and select Particle-Picking

2. Select “manual” as the Detection method

3. Set the Particle radius (A) (for visualization purposes)

4. Click Save, Run, and Start Run for 1 block

5. Once the run completes, navigate to the Particle-Picking block, inspect different tomogram slices using the slider and select particles by clicking on them. Go to other tomograms using the navigation bar at the top of the page and add particles as needed

Size-based picking¶

This method, described in Jin et al. (2024), can be used to detect particles based on their size. It works both on purified and in-situ samples:

1. Click on Tomograms (output of the Pre-processing block) and select Particle-Picking

2. Select “size-based” as the Detection method

3. Set the Particle radius (A) and other parameters as needed (see the classification tutorial for an example)

4. Click Save, Run, and Start Run for 1 block

5. Once the run completes, navigate to the Particle-Picking block to inspect the results

Template search¶

This method uses the package pytom-match-pick to detect particles using template search (GPU required):

1. Click on Tomograms (output of the Pre-processing block) and select Particle-Picking

2. Select “template-search (pytom)” as the Detection method

3. Set the Particle radius (A), specify a Template, and set other parameters as needed

4. Click Save, Run, and Start Run for 1 block

5. Once the run completes, navigate to the Particle-Picking block to inspect the results

Geometry-based picking blocks¶

nextPYP’s geometry-based picking is used to detect membrane proteins attached to the surface of virions or vesicles, as described in Liu et al. (2023). The method is composed of three stages:

Detection of virion centers¶

The first step is to estimate the position and the approximate radius of each virion or vesicle:

1. Click on Tomograms (output of the Pre-processing block) and select Particle-Picking

2. Select “virions” as the Detection method

3. Set the expected Virion radius (A) and other parameters as needed

4. Click Save, Run, and Start Run for 1 block

5. Once the run completes, navigate to the Particle-Picking block to inspect the results

Virion segmentation¶

The next step is to segment virions in 3D using methods described in Bartesaghi et al. (2005):

1. Click on Particles (output of the Particle-Picking block) and select Segmentation (closed surfaces)

2. Adjust the segmentation parameters as needed (defaults should work fine for EMPIAR-10164, for example)

3. Click Save, Run, and Start Run for 1 block

4. Once the run completes, navigate to the Particle-Segmentation (closed surfaces) block and go to the Segmentation tab to inspect the results

nextPYP calculates segmentations using implicit surface representations that require specifying a threshold value to uniquely define the detected surface. To facilitate the threshold selection process for each virion, the user can choose from 8 different threshold values (represented as columns in a table). If none of the columns look reasonable (or if a virion should be ignored), the last column, labeled as “-”, should be selected. This process must be repeated for all virions in a tilt-series and for all tilt-series in the dataset

Here is a screenshot of the user interface for virion segmentation:

Particle picking from virions¶

The last step is to pick particles from the surface of virions:

1. Click on Segmentation (closed) (output of the Segmentation (closed surfaces) block) and select Particle-Picking (closed surfaces)

2. Select a particle detection Method and corresponding parameters: “uniform” can be used to select uniformly spaced positions on the virion surfaces, while “template search” is used to search for positions on the surface that have high-correlation with an external template (provided as an *.mrc file with the correct pixel size saved in the header)

3. Click Save, Run, and Start Run for 1 block

4. Once the run completes, navigate to the Particle-Picking (closed surfaces) block to inspect the results

Neural-network picking blocks¶

nextPYP uses two blocks to implement neural network-based particle picking, as described in Huang et al. (2024). This method uses consistency regularization to minimize the number of annotations and speedup training:

Model training¶

The first step is to obtain a set of particles using any of the methods implemented in the Pre-processing or MiloPYP (eval) blocks to train the neural network:

1. Click on Particles (output of the Particle-Picking or Particle-Picking (closed surfaces) blocks), or MiLoPYP Particles (output of the MiLoPYP (eval) block) and select Particle-Picking (train)

2. Adjust the parameters for training as needed. If using MiLoPYP particles, see instructions on how to set parameters here

3. Click Save, Run, and Start Run for 1 block

4. Once the run completes, navigate to the Particle-Picking (train) block to inspect the training loss

Model evaluation¶

Once the model has been trained, evaluation can be run on the entire dataset:

1. Click on Particles Model (output of the Particle-Picking (train) block) and select Particle-Picking (eval)

2. Select the location of the Trained model (*.pth) using the file browser, and adjust the evaluation parameters as needed (the file browser’s default location will be the train/ folder from the parent block)

3. Click Save, Run, and Start Run for 1 block

4. Once the run completes, navigate to the Particle-Picking (eval) block to inspect the particle picking results