Cell Painting is a high-content imaging assay that stains cells with six fluorescent dyes, captures images across five channels, and reveals eight major cellular structures in a single experiment. Automated software then measures roughly 1,500 features per cell, from size and shape to texture and intensity, producing a rich morphological fingerprint that can detect even subtle changes caused by drugs, gene knockouts, or environmental chemicals.1PubMed Central. Cell Painting, a high-content image-based assay for morphological profiling using multiplexed fluorescent dyes The technique has become one of the most widely adopted tools in image-based profiling, and the science around it is evolving fast.
Why “Paint” a Cell
Most biological assays are designed to answer a specific question: does this compound lower a particular enzyme’s activity, or does that gene knockout affect a known signaling pathway? Cell Painting takes the opposite approach. Rather than looking for one predetermined outcome, it captures a broad snapshot of cell morphology and lets the data reveal what changed. The idea is to maximize the biological information you can extract from a single, relatively cheap experiment.
The assay accomplishes this by staining cells for eight organelles and sub-compartments simultaneously, using six well-characterized dyes chosen for their compatibility in high-throughput settings.2GigaScience. A dataset of images and morphological profiles of 30 000 small-molecule treatments using the Cell Painting assay Because the assay is unbiased, it can pick up effects that a researcher might never have thought to look for. A compound designed to target mitochondria might also alter nuclear shape or endoplasmic reticulum texture, and Cell Painting catches all of it in one pass.
The Six Dyes and What They Reveal
The standard Cell Painting protocol labels the nucleus, nucleoli, endoplasmic reticulum, actin cytoskeleton, Golgi apparatus, plasma membrane, and mitochondria.3PubMed Central. Application of Cell Painting for chemical hazard evaluation in support of screening-level chemical assessments Six dyes cover these eight targets, with some dyes pulling double duty because their fluorescence overlaps enough to be captured in a shared imaging channel but still illuminates distinct structures. For instance, DNA-binding dyes mark the nucleus and nucleoli together, while a mitochondrial stain highlights that organelle’s distribution and membrane potential.
The dyes were selected not just for biological relevance but for practical reasons: they are inexpensive, commercially available, and tolerate the fixation and permeabilization steps used in high-throughput screening. Cells are typically grown in multiwell plates, treated with whatever perturbation is being tested (a small molecule, a genetic edit, an environmental chemical), and then fixed, stained, and imaged. The entire wet-lab portion, from seeding cells to finished plates, takes a few days in most labs, and the protocol is straightforward enough that it has been adopted across dozens of academic and pharmaceutical settings worldwide.
Imaging and Feature Extraction
Once stained, the plates are loaded onto an automated high-content microscope that captures images in five fluorescence channels. Each channel corresponds to a different excitation/emission wavelength, so the dyes are separated optically rather than physically. The result is a set of multichannel images for every well, with hundreds to thousands of individual cells visible per image.
The images then go into an automated analysis pipeline. Software such as CellProfiler identifies each cell in the image, draws boundaries around it, and measures approximately 1,500 morphological features.4Nature Protocols. Cell Painting, a high-content image-based assay for morphological profiling using multiplexed fluorescent dyes These features span a wide range: overall cell area, nuclear roundness, the graininess of mitochondrial staining, the intensity distribution of actin fibers, and many more. Each cell ends up represented as a long numerical vector, and those vectors can be averaged across a well or kept at single-cell resolution depending on the analysis goal. One study using the assay for gene-function profiling extracted 1,384 features per cell following feature selection.5eLife. Systematic morphological profiling of human gene and allele function via Cell Painting The exact count varies slightly depending on the software version and the segmentation settings, but the order of magnitude stays the same.
Cleaning Up the Data
Raw morphological profiles are noisy. Plates processed on different days, imaged on different microscopes, or stained with slightly different dye batches can introduce systematic variation that has nothing to do with biology. Batch effects are one of the biggest practical headaches in Cell Painting, and a dedicated preprocessing pipeline is needed before the data can be meaningfully compared.
A typical pipeline strips out features that barely vary across the dataset, normalizes the remaining features plate by plate so that systematic shifts are dampened, transforms each feature’s distribution to something closer to a bell curve, and removes redundant features that are highly correlated with one another.6Nature Communications. Evaluating batch correction methods for image-based cell profiling After these steps, you are left with a smaller, cleaner set of features that more faithfully represents the actual biological signal. Getting batch correction right matters enormously: if systematic noise is not adequately removed, two identical treatments processed on different days can look different, and two genuinely distinct treatments processed together can look the same.
Drug Discovery and Mechanism of Action
One of Cell Painting’s most prominent applications is in drug discovery, where it can help researchers figure out how a compound works. Because the assay captures broad morphological changes rather than a single readout, compounds that act through the same biological pathway tend to produce similar profiles. You can compare the fingerprint of a new, poorly understood compound against a library of “landmark” compounds whose mechanisms are already known. If the profiles match closely, that is a strong hint that the two compounds share a mechanism of action.7PubMed Central. Cell Painting: A Decade of Discovery and Innovation in Cellular Imaging
This approach is sometimes combined with other data sources. Molecular structure information, transcriptomic readouts, and metabolomic data can all be layered on top of morphological profiles to improve mechanism-of-action predictions.8Artificial Intelligence in the Life Sciences. Combining molecular and cell painting image data for mechanism of action prediction The morphological data adds something that purely molecular descriptors miss: a readout of what the compound actually does to a living cell, not just what its structure suggests it might do. For pharmaceutical companies screening thousands of molecules in early discovery, this can help prioritize which compounds move forward and flag potential off-target effects before expensive animal studies.
Genetic Screening Without Target-Specific Markers
Cell Painting is not limited to chemical compounds. It works equally well with genetic perturbations, and this has opened up a powerful approach to understanding gene function. In a CRISPR-based screen, cells are engineered to have individual genes knocked out, and then they are stained and imaged with the standard Cell Painting protocol. The resulting morphological profiles reveal which genes, when lost, produce visible changes in cell shape, organelle structure, or overall organization.
A pooled screening platform described recently combines Cell Painting with optical pooled screening, enabling researchers to knock out many genes in the same experiment and then read out both the genetic identity of each cell and its morphological profile. By feeding these images into self-supervised deep-learning models, gene networks emerge without any need for predefined biomarkers.9PubMed Central. A pooled Cell Painting CRISPR screening platform enables de novo inference of gene function by self-supervised deep learning In that study, a vision transformer model trained specifically on Cell Painting images outperformed both classical feature-extraction methods and a general-purpose model pre-trained on everyday photographs, recovering more known biological associations between genes.10Nature Communications. A pooled Cell Painting CRISPR screening platform enables de novo inference of gene function by self-supervised deep learning The practical upshot is that researchers can discover what a gene does without first guessing which pathway it belongs to.
Toxicology and Chemical Safety Screening
Another area where Cell Painting has gained traction is toxicology. Regulatory agencies and chemical manufacturers need to evaluate whether thousands of industrial, agricultural, and consumer chemicals pose health risks, and traditional animal-based testing is too slow and expensive to keep pace. Cell Painting offers a fast, cell-based alternative that can flag biologically active chemicals and estimate the concentrations at which those effects appear.
In one large study, over 1,200 chemicals from the EPA’s ToxCast library were screened at multiple concentrations in human bone-cancer-derived cells (U-2 OS, the standard Cell Painting cell line). For each active chemical, researchers estimated a “phenotype-altering concentration,” a threshold at which the compound begins to change cell morphology in detectable ways.3PubMed Central. Application of Cell Painting for chemical hazard evaluation in support of screening-level chemical assessments An earlier study adapting Cell Painting for environmental chemicals similarly concluded that the assay is an efficient and reproducible screening method for characterizing biological activity and potency, with potential use in cell-based safety assessments.11PubMed Central. Bioactivity screening of environmental chemicals using imaging-based high-throughput phenotypic profiling
The appeal here is the assay’s untargeted nature. A traditional toxicity screen might test whether a chemical kills cells or disrupts a specific receptor. Cell Painting catches a much wider range of effects, from mitochondrial stress to cytoskeletal disruption to changes in nuclear morphology, all without needing to design a separate assay for each endpoint. That makes it a useful first pass that can identify chemicals warranting more detailed follow-up.
Deep Learning Versus Classical Feature Extraction
The original Cell Painting analysis workflow relies on CellProfiler, which extracts handcrafted features like area, shape, and texture measurements that scientists have defined in advance. This works well for many applications, but there has been growing interest in whether deep-learning models can squeeze more information out of the same images by learning features that humans would not think to measure.
The evidence is mixed but increasingly favors deep learning for certain tasks. One comparative study found that CellProfiler lacked the ability to differentiate morphological changes between certain compounds, whereas convolutional-network-based tools produced features that were better at distinguishing them, though at the cost of interpretability.12Scientific Reports. Attention-based deep learning for accurate cell image analysis Self-supervised models, which learn image representations without being told what to look for, have shown particular promise. When trained directly on Cell Painting images, they capture richer embeddings than models trained on photographs of everyday objects and can recover more known gene-gene relationships.9PubMed Central. A pooled Cell Painting CRISPR screening platform enables de novo inference of gene function by self-supervised deep learning
The field has not abandoned CellProfiler, though. Its features are transparent and well understood, which matters when you need to explain results to regulators or clinicians. Many labs now run both approaches in parallel, using classical features for interpretability and deep-learning embeddings for discovery, then comparing where they agree and diverge.
Moving Into Three Dimensions
Standard Cell Painting is performed on flat layers of cells growing on the bottom of a well plate. That is convenient for imaging, but it is a poor approximation of how cells behave in a living tissue, where they are surrounded by other cells in three dimensions and receive signals from all directions. Recognizing this limitation, researchers have begun adapting Cell Painting for three-dimensional spheroid cultures, small balls of cells that self-organize and more closely mimic tissue architecture.
A recent method combines the standard Cell Painting dyes with tissue-clearing techniques that make the spheroid transparent enough for confocal microscopy to image through. Cells inside the spheroid can be individually segmented, and morphological features extracted in a way analogous to the traditional two-dimensional workflow.13bioRxiv. High-content morphological profiling by Cell Painting in 3D spheroids This is still early-stage work, and the throughput is lower than flat-culture Cell Painting because confocal imaging is slower than widefield microscopy. But it addresses a real gap: some drug effects and disease phenotypes only manifest when cells are in a tissue-like context, and capturing those complex phenotypes requires moving beyond monolayers.
Virtual Cell Painting
One of the practical downsides of Cell Painting is that it requires chemical fixation. Once cells are fixed and stained, they are dead. You cannot follow the same cells over time to watch how they respond to a treatment, and the staining process itself introduces labor and reagent costs. A growing body of work asks whether deep-learning models can predict what the fluorescent channels would look like using only a standard brightfield image, no dyes needed.
The idea is called virtual or label-free Cell Painting. In one study, researchers trained deep-learning models on paired brightfield and fluorescent images, then used the models to predict the five Cell Painting channels from brightfield alone. The predicted images matched the real ones with a mean correlation of 0.84 across all channels, strong enough to preserve biologically meaningful signals in downstream analyses.14PubMed Central. Label-free prediction of cell painting from brightfield images Another effort, called MONET, used a diffusion model trained on a large dataset to predict Cell Painting channels from brightfield, explicitly aiming to enable studies of cell dynamics that fixation-based staining makes impossible.15arXiv.org. MONET – Virtual Cell Painting of Brightfield Images and Time Lapses Using Reference Consistent Diffusion
Virtual Cell Painting would also free up fluorescence channels for more targeted stains. Instead of devoting all five channels to the generic Cell Painting dyes, a lab could use the brightfield-based prediction for the generic organelle channels and reserve the fluorescence for a specialized marker of particular interest, getting the best of both worlds.
The JUMP Consortium and Open Data
Cell Painting generates enormous datasets, millions of images and billions of feature measurements across a screening campaign. Recognizing that the value of morphological data grows when different labs can compare their profiles against a shared reference, a consortium called JUMP (Joint Undertaking for Morphological Profiling) set out to build the largest publicly available Cell Painting dataset. The project brings together ten pharmaceutical companies, six technology firms, and two nonprofit partners. The planned dataset covers images and profiles for over 116,750 unique compounds, overexpression of more than 12,600 genes, and CRISPR knockout of nearly 8,000 genes, all in U-2 OS cells.16bioRxiv. JUMP Cell Painting dataset: morphological impact of 136,000 chemical and genetic perturbations
Having chemical and genetic perturbations profiled side by side in the same cell type under the same conditions makes it possible to connect drugs to their gene targets purely through morphological similarity. If a compound’s profile closely resembles the profile of cells where a particular gene has been knocked out, that gene is a candidate target. The open nature of the dataset also means that academic groups without the resources to run large-scale screens can still develop and benchmark new analysis methods on real, high-quality data.
Generative Models for Predicting Cell Responses
An emerging frontier goes beyond analyzing existing images and asks whether AI can generate realistic predictions of how cells will look under a treatment that has never been tested experimentally. Two recent approaches illustrate the concept.
One model, called IMPA, uses the chemical structure of a drug to predict what the treated cells would look like. It works by interpolating through a learned space of morphological phenotypes, generating intermediate images that show gradual transitions from control cells to treated cells. When tested on actin-disrupting and tubulin-destabilizing compounds, the model produced smooth visual progressions, with features like increased actin contrast and reduced nuclear area appearing progressively as the predicted “dose” increased.17Nature Communications. Predicting cell morphological responses to perturbations using generative modeling
A second framework, MorphoDiff, takes a different route by using a diffusion model to produce high-resolution predicted images of cell morphology under both chemical and genetic perturbations. The developers describe it as the first system capable of generating guided, high-resolution morphology predictions that generalize across both types of intervention, and they validated it against three publicly available Cell Painting datasets.18PubMed Central. MorphoDiff: Cellular Morphology Painting with Diffusion Models
These generative tools are not replacing experiments, but they could reshape how experiments are prioritized. If a model can predict that a candidate drug will produce a toxic-looking morphological profile, the lab can deprioritize it before spending resources on a physical screen. Conversely, compounds predicted to produce desirable phenotypes could be fast-tracked. The reliability of such predictions still needs extensive validation, but the direction is clear: Cell Painting data is becoming both the training ground and the testing ground for AI-driven biology.