This code relates to

A systems-level study reveals regulators of membrane-less organelles in human cells

by Doris Berchtold, Nico Battich, and Lucas Pelkmans.

Published in Molecular Cell: November 29, 2018. DOI: https://doi.org/10.1016/j.molcel.2018.10.036

---

Dependencies

ClassifyPixels.m and supporting functions have been tested on Windows and MATLAB 2014a.

Example data set

The example data set comprises 10 images of unperturbed cells of the splicing speckle screen that were stained with antibodies against SRRM2 (in TIFF), a measurement file to correct these images for uneven illumination, and the segmentations of cells and nuclei (in SEGMENTATION). For comparison, the folder segmented_splicing_speckles contains the segmentation of splicing speckles obtained with the model that was trained on the screen.

How to train a segmentation model with the ClassifyPixels GUI and apply it on the example data set

1. Launch MATLAB 2014a on a Windows machine
2. Add the folder PixelClassification to your MATLAB path
3. Open CreateSettingsFile.m and run
4. Launch the GUI by typing ClassifyPixels in the MATLAB command line
5. Read through the detailed GUI description below. In brief:
6. Load the settings file, click next image
7. Train the classifier by clicking on single pixels of multiple images and mark them as signal or background
8. Save the classification model (default: PCProject_01.mat)
9. Open run_pixel_classification_example.m, change the name of the classification model and run the script. This will apply the trained model on all images of the example data set. The segmentation files will be saved into the folder PixelClassification\ExampleDataSet\TEST_seg.

Detailed description of the ClassifyPixels Graphical User Interface

Top section

• In /path/to/settings insert the full path to the settings file
• In PCProject01.mat insert the name of the output file, which will contain the classification model
• Click Load to load the settings file
• Save Classification saves the classification model after training

Bottom section

• Click Next image to display the first or next image
• Set the Low Quantile and High Quantile and click refresh to rescale the intensities of the displayed image
• Go to section allows to move to another field of view of the same image

Drawing tools

• Select Signal (green) before clicking on pixels of the object to segment; select Background (red) before clicking on background pixels; select eraser to remove and correct wrongly classified pixels
• The Brush scale bar allows to change the number of pixels selected upon clicking on the image

Modeling

All segmentations of MLOs in Berchtold et al. were achieved using the Gaussian option of SVMs and the Smooth Probability checked. Other implementations have not been tested extensively.

• Select the type of SVM model that will be trained on the selected pixels
• Click Train after marking sufficient pixels as signal or background
• Click Classify to classify all pixels of the current image; to visualize the result click Show Prob, see below
• The Smooth Prob and More options allow smoothing of the estimated probabilities of being an object
• Click Optimize (SVM only) to run a feature selection to optimize the misclassification error; this option only works for Linear, Gaussian, or Polynomial modeling

Visualization

• Define the probability threshold for segmentation in the Seg box; default = 0.5
• Click Show Image to see the current image
• Click Show Prob to see the current probability of pixels of being an object; click Classify before
• Click Show Seg to see the current segmentation of objects according to the Seg threshold
• Click Original Image to generate a new figure of the full original image
• Check Show cell seg to also plot the cell segmentations
• Click Plot Features to generate figures of all pixel features used for classification; looking through these features helps to select the most informative pixel features for object segmentation

---

Pixel features used for classification

Feature name : Description
log_5_2: Laplacian of Gaussian filter, size = 5 pixels, sigma = 2
log_6_3: Laplacian of Gaussian filter, size = 6 pixels, sigma = 3
log_8_4: Laplacian of Gaussian filter, size = 8 pixels, sigma = 4
log_10_5: Laplacian of Gaussian filter, size = 10 pixels, sigma = 5
log_20_10: Laplacian of Gaussian filter, size = 20 pixels, sigma = 10
gau_20_10: Gaussian filter, size = 20 pixels, sigma = 10
gau_9_4: Gaussian filter, size =9 pixels, sigma = 4
gau_5_2: Gaussian filter, size = 5 pixels, sigma = 2
diffgau_1: difference of Gaussians, 1st size = 15 pixels, sigma = 8, 2nd size = 15 pixels, sigma = 3
diffgau_2: difference of Gaussians, 1st size = 15 pixels, sigma = 9, 2nd size = 15 pixels, sigma = 5
average_3: averaging filter, size = 3 pixels
average_5: averaging filter, size = 5 pixels
average_10: averaging filter, size = 10 pixels
average_20: averaging filter, size = 20 pixels
edges: Sobel and Prewitt filters, and edges on Laplacian transformations
simple_edges: Sobel and Prewitt filters
edge_log_s8_4: Prewitt filter on Laplacian-of-Gaussian transformation
hole_log_8_4: A hollow filter of 5 pixel size applied on the Laplacian-of-Gaussian transformation
entropy_log_8_4: Entropy filter applied on the Laplacian-of-Gaussian transformation size = 8 pixels, sigma = 4
entropy_log_5_2: Entropy filter applied on the Laplacian-of-Gaussian transformation size = 5 pixels, sigma = 2
morph_opening: Morphological opening with filter sized of 5, 8, and 10 pixels
morph_close: Morphological opening with filter sized of 3, 5, 7, 10,15, and 20 pixels
morph_close_small: Morphological opening with filter sized of 3, and 5 pixels
loc_back_small and loc_back_large: local background estimation
deconv_10_7: Lucy-Richardson deblur with Gaussian filter size = 10 pixels, sigma = 7
deconv_10_3: Lucy-Richardson deblur with Gaussian filter size = 10 pixels, sigma = 3
deconv_5_3: Lucy-Richardson deblur with Gaussian filter size = 5 pixels, sigma = 3
std_dif_1: standard deviation differences of deblur image
zscore_1: z-scoring filters
watershed_lines, watershed_lines_smooth, watershed_lines_smooth_low, and watershed_lines_smooth_very_low: Filters based on watershed lines
tophat_12: Tophat filter of size 12 applied on Gaussian filtered image
orig_tophat_30: Tophat filter of size 12 applied on original image

---

Pixel features used for classification of MLOs in Berchtold et al.

Cajal bodies:
log_5_2, log_6_3, log_8_4, log_10_5, gau_9_4, gau_5_2, diffgau_1,diffgau_2, average_3, average_5, average_10, morph_opening, morph_close, edges, deconv_10_7, deconv_10_3, deconv_5_3, std_dif_1, zscore_1, watershed_lines_smoothed
Nucleoli:
gau_20_10, gau_9_4, gau_5_2, average_3, average_5, average_10, morph_opening, morph_close_small, orig_tophat_30
P-bodies:
log_5_2, log_5_3, log_6_3, log_8_4, gau_9_4, gau_5_2, diffgau_1, diffgau_2, average_3, average_5, average_10, edges, edge_log_8_4, hole_log_8_4, deconv_5_3, std_dif_1, zscore_1, watershed_lines
PML nuclear bodies:
log_5_2, log_5_3, log_6_3, log_8_4, gau_9_4, gau_5_2, diffgau_1, diffgau_2, average_3, average_5, average_10, edges, edge_log_8_4, hole_log_8_4, deconv_5_3, std_dif_1, zscore_1, watershed_lines
Stress granules:
log_8_4, log_10_5, log_20_10, average_20, simple_edges, entropy_log_8_4, entropy_log_5_2, morph_close, tophat_12, orig_tophat_30
Slicing speckles:
log_5_2, log_8_4, log_20_10, gau_20_10, gau_9_4, gau_5_2, diffgau_1, diffgau_2, average_5, average_10, average_20, edges, hole_log_8_4, morph_opening, morph_close, deconv_10_7, deconv_10_3, std_dif_1, zscore_1