Reproduce the statistics-paper compartments

Build the supplied CellProfiler nuclei/cell/cytoplasm analysis in VIPP, then compare measurements on identified image fields. The workflow follows Marcotti and colleagues' statistics paper and its executable author pipeline. Its scientific target is a per-cell relative nuclear-localisation measurement.

New in 0.16.0a1

This workflow requires the CellProfiler profile nodes introduced in 0.16.0a1. Earlier versions do not contain them.

Keep the comparison target explicit

Agreement with the original CellProfiler 4.2.6 runtime on the same inputs checks VIPP's implementation. Reproducing the paper's exported measurements also requires matching the authors' original image pixels and cohort. Current acquired-image reruns do not yet establish full published-result equality. Preserve discrepancies instead of tuning settings to conceal them.

Prepare the three stains

Use a nuclear stain, Actin and the protein being measured. Keep the protein image unsmoothed for measurement. Select a single YX plane from each input and verify matching physical grids.

Dataset	Nuclear stain	Cell guidance	Measured protein
IDR0139	DNA / DAPI, acquisition A01	Actin, A03	Fascin, A02
IDR0028	Hoechst, TIFF frame 0	Actin, frame 2	YAP/TAZ, frame 3

IDR0139's filename suffix C04 belongs to Actin A03; C03 belongs to a separate NuclearActin A04 stain. Use the acquisition mapping above. IDR0028's four TIFF frames are stain channels, even when a reader describes them as four time frames. Review axes explicitly before selecting planes.

Add Convert Dtype, choose float32, and preserve the input intensity values. Follow it with Rescale Intensity: choose explicit Values, input limits 0 and 65535, output limits 0 and 1, with inversion disabled. This conversion matches CellProfiler's float32 division across the full uint16 range. You can apply it to a multichannel stack before extracting stains. Do not stretch each image's observed range or use display contrast as the scientific normalization. Retain the raw input files.

Connect the compartment stages

1. Add Smooth — CellProfiler Gaussian to the nuclear image and a second instance to Actin. Set Artifact diameter (pixels) to 2 for both.
2. Connect smoothed nuclear intensity to Segment Nuclei — CellProfiler Shape. Keep Minimum diameter (pixels): 15, Maximum diameter (pixels): 50 and Threshold smoothing scale: 1.3488. Click Calculate. Inspect both Retained nuclei and Before filtering.
3. Connect smoothed Actin to Threshold — CellProfiler Minimum Cross-Entropy. Keep Threshold smoothing scale: 0. Inspect the Boolean foreground mask.
4. Add Prepare Seeds — CellProfiler Propagation. Connect the nuclear Before filtering output to its matching port and Retained nuclei to its matching port. The output intentionally includes excluded border nuclei while neighbouring cells compete for pixels.
5. Add Grow Regions from Seeds — CellProfiler Propagation. Connect smoothed Actin to Guidance image, prepared seeds to Seed labels and the Actin threshold mask to Foreground mask. Keep Distance regularization: 0.05, then click Calculate.
6. Connect those grown labels and Retained nuclei to Finish Cell Regions — CellProfiler. Keep Fill labelled holes enabled. This removes regions belonging solely to excluded border nuclei and restores accepted nucleus IDs.
7. Connect the finished cells and retained nuclei to Extract Cytoplasm — CellProfiler. Keep Shrink nuclei before subtraction enabled. The nuclear outline remains in the cytoplasm by design.
8. Overlay nuclei, cells and cytoplasm on their source images. Inspect touching nuclei, image borders, unassigned regions and small objects. Save the frozen workflow before processing the rest of the cohort.

Generic Gaussian, threshold, watershed and subtraction nodes have different edge or object-handling rules. For this reproduction, use the named profile stages; see the numerical reference.

Measure and preserve cell identity

Use Measure Objects + Intensity to obtain the unsmoothed protein's mean intensity in Retained nuclei and in the Cytoplasm output. Export the tables with the object label and complete source identity, keeping dataset, plate, well and field. Label 1 in one image is unrelated to label 1 in another image. See measurement tables for table and CSV handling.

The reproduction companion joins the exported means and calculates nuclear_mean / (nuclear_mean + cytoplasm_mean) for each paired cell. Image processing runs through the VIPP graph; ratios and statistics use companion Python postprocessing, as in the original paper. Use the same policy for missing values and zero denominators in both implementations; do not silently replace them with zero. The authors' selected reference tables contain finite ratios without such exclusions.

For descriptive exploration inside VIPP, connect the compartment measurement tables to Statistics or Plot Results, or open Results Workspace from a table node. Collect saved batch tables through Collect measurement results before comparing images. Choose the observation level deliberately: objects, image averages and declared-sample averages answer different questions. These tools do not reproduce the authors' downstream subsampling or inferential analysis; the companion remains the reference for that part of the workflow.

The main IDR0139 comparison uses wells J05, O02, E22 and L08 on plate 1093711385, with four fields per well. The larger selected cohort contains ten wells. IDR0028 uses twelve selected wells across plates 1A, 2A and 2B, with 28 fields per well. The authors' selected-well tables and notebooks define the cohort; a few example fields cannot reproduce a whole-well result.

Compare in stages

First compare image identity, scaling, threshold masks, object counts, paired object positions and compartment means. Then compare per-cell ratios and well-level summaries. Preserve cell-to-well and well-to-plate relationships when reproducing the authors' subsampling and replicate plots.

Keep three conclusions separate: agreement with the reference implementation, agreement with the published measurements, and biological validity. A successful segmentation comparison does not establish an independent experimental replicate or validate the statistical inference.

The authors' pinned code and data instructions provide the downstream statistical notebooks. See validation status for the current evidence and unresolved source differences.

For an interface-only example, open Exhaustive Inspector Showcase and inspect its CellProfiler compartment profile lane. It uses an existing synthetic image for both nuclear and cell guidance to exercise the controls; it is not a paper-data reproduction.