Cell Painting Pipelines — From Images to Phenotypic Profiles

The problem — A Cell Painting "profile" sounds like a measurement, but it is the output of a pipeline — six dyes across five channels yielding ~1,500 morphological features per cell (Bray et al.) — and the number anyone reasons over arrives only at the very end. The well image is not the profile; the segmentation mask is not the profile; the raw feature table is not the profile. Each stage transforms the data, and a failure at any one of them propagates silently into a final vector that still looks perfectly well-formed. Treating the assay as "stain and read" hides where it actually breaks.

What it is / how it works — The pipeline is a fixed chain, and each link gates the next:

• Illumination correction (preprocess) — fluorescence fields have vignetting and uneven excitation; uncorrected, position-on-plate leaks into every feature. Per-channel illumination functions (BaSiCPy / CellProfiler) flatten it before anything is measured (Bray et al.).
• Segmentation (segment) — Cellpose or CellProfiler objects define which pixels are which cell; merged or split nuclei corrupt every per-cell feature downstream.
• Feature extraction (features) — classical CellProfiler intensity/texture/shape features, or learned embeddings — the decade review reports deep-learning features beating CellProfiler by up to +29% mAP.
• Batch correction (correct) — the central nemesis. Plate, well, and source effects sit inside the profile; Arevalo et al. benchmark ten methods (Harmony and Seurat RPCA best) and find none removes batch without some loss of biology.
• Readout (readout) — only here does a number exist: percent-replicating or mAP via copairs, measuring whether replicates of the same perturbation retrieve each other above a null. JUMP reports a phenotype for only ~68% of genetic perturbations (genetic map), and on CPJUMP1 cross-modality compound↔gene matching sits barely above chance.

Where it breaks — The general failure is reading the endpoint without instrumenting the chain that produced it — and it is the same failure whatever the assay. Oversmoothed illumination correction flattens real intensity signal; a segmentation drift silently rebases every shape feature; un-modelled batch makes two replicates of one compound land far apart, so the profile is non-replicating for a technical reason no pixel metric can see. The readout is the only honest verdict, which is why it belongs with task-conditioned quality — and why a strong feature extractor does not collapse the eight stages around it. The discipline: gate on percent-replicating / mAP, but trace a bad readout back through correction, features, and segmentation — never report the end without auditing the chain.

References

• Bray et al. (2016). Cell Painting, a High-Content Image-Based Assay for Morphological Profiling Using Multiplexed Fluorescent Dyes. Nature Protocols.
• Chandrasekaran et al. (2024). Three Million Images and Morphological Profiles of Cells Treated with Matched Chemical and Genetic Perturbations (CPJUMP1). Nature Methods.
• Arevalo et al. (2024). Evaluating Batch Correction Methods for Image-Based Cell Profiling. Nature Communications.
• Chandrasekaran et al. (2025). Morphological Map of Under- and Over-Expression of Genes in Human Cells. Nature Methods.
• Seal et al. (2025). Cell Painting: A Decade of Discovery and Innovation in Cellular Imaging. Nature Methods.