CELL PAINTING ASSAY: High-Content Phenotypic Profiling with the Araceli Endeavor®

U2OS cells (Cat#300364, Cell Line Service, Germany, RRID:CVCL_0042, female) were seeded at a density of 1,600 cells per well in 384-well plates. After a 4-hour incubation at 37 °C with 5% CO₂ to allow for cell attachment, selected compounds from our in-house chemical library were added using the Echo 520 acoustic dispenser (Labcyte). A final concentration of 0.1% DMSO served as the negative control for compounds tested at 10 µM. The compound set was curated based on their

known mechanisms of action and previously observed phenotypic outcomes. Selection was further supported by complementary biochemical and molecular biology assays to confirm phenotype-driven hypotheses. Following compound addition, U2OS cells were incubated for 20 hours. Subsequently, the Cell Painting Assay was performed in accordance with the protocol described by (Pahl et al., 2023). This assay utilizes a panel of fluorescent dyes to label eight distinct cellular components: DNA was stained with Hoechst 33342; the endoplasmic reticulum with concanavalin A; nucleoli and cytoplasmic RNA with SYTO 14; F-actin filaments with phalloidin; Golgi apparatus and plasma membrane with wheat germ agglutinin (WGA); and active mitochondria with MitoTracker Deep Red, following the framework established (Bray et al., 2016).

Images were acquired from the same plates using the Endeavor imaging system and a prior imaging solution, in order to enable a direct performance comparison. We acquired four sites with the Endeavor, corresponding to a total imaged area of approximately 5.7 mm2 for each well of the 384-well plate. Acquisition included five imaging channels, one of which was transmitted light, and was completed in only 7 minutes per plate. In contrast, the prior imaging solution required acquisition of 9 sites per well, covering a total area of 4.6 mm2, also across 5 channels. However, this process took approximately 3.5 hours per plate. This demonstrates a significant improvement in imaging speed and area coverage with the Endeavor, achieving greater spatial sampling in a drastically reduced acquisition time, while maintaining equivalent channel depth.

Standard pre-processing steps, including illumination correction, were applied using CellProfiler 4.2.8. The illumination correction pipeline focused on generating illumination correction functions for multiple imaging channels. The process begins by loading image data from a CSV file, with grouping based on metadata such as well, site, and plate. Images from each channel are first downsampled by a factor of 0.125 in X and Y dimensions to reduce computational complexity. For each downsampled image, an illumination correction function is calculated using the “Regular” method, with Gaussian smoothing and spline fitting based on all images across cycles. These correction functions aim to address uneven illumination and improve downstream image analysis. After computation, each illumination function is rescaled back to its original dimensions by applying a resizing factor of 8 in X and Y. The pipeline subsequently saves both the full-sized illumination-corrected images and their averaged correction functions in .npy format for each imaging channel. This pipeline ensures consistent illumination across datasets, which is critical for robust quantitative image analysis in high-content screening experiments.

The Endeavor’s large field of view was efficiently managed through an optimized pipeline, ensuring high-quality input for analysis.

Feature extraction was conducted using CellProfiler version 4.2.8, chosen for its enhanced scalability in processing high-resolution, large-format images, as well as its expanded repertoire of morphological and intensity-based measurements. While our established image analysis pipeline is based on CellProfiler 3 (Pahl et al., 2023), the high-content imaging data generated by the Endeavor system, characterized by increased spatial resolution and larger image dimensions, exceeded the performance constraints of the older version. Although the underlying architecture and segmentation algorithms between CellProfiler versions 3 and 4 remain largely conserved, version 4 introduces additional feature extraction modules and performance optimizations. These enhancements enabled more efficient processing and quantification of complex phenotypic profiles, and were systematically integrated into our updated workflow to maximize the extraction of biologically relevant descriptors from the acquired image data (Stirling et al., 2021).

Images were processed on a high-performance Linux computing cluster using parallelization, where each image was independently analyzed. The analysis was conducted using CellProfiler 4.2.8, with 20 modules for processing high-resolution, multi-channel images from the Endeavor imaging system, and also images from the prior imaging solution. The LoadData module imported images and metadata, rescaling intensities as needed. CorrectIlluminationApply corrected illumination artifacts across all channels, and MeasureImageQuality calculated key metrics, such as blur and intensity, using global thresholding. Nuclei were identified using IdentifyPrimaryObjects, followed by merging of nearby objects with SplitOrMergeObjects to account for multinucleated cells. Cells were segmented using the IdentifySecondaryObjects module, with cytoplasm further defined by IdentifyTertiaryObjects. The RelateObjects module quantified nuclei within cells, and MeasureColocalization assessed co-localization of cellular components. Additional measurements of granularity, intensity, and texture were obtained with the MeasureGranularity, MeasureObjectIntensity, and MeasureTexture modules. Outlines of cells and nuclei were visualized using OverlayOutlines, and results were saved as PNG images with SaveImages.

Finally, ExportToSpreadsheet generated a data file for further analysis. This pipeline enables efficient, high- throughput analysis of cellular morphology and co- localization. Extracted features were saved in per-image .csv files for modular downstream processing.

Extracted features were first consolidated into a unified plate-level dataset using the ingest function from the pycytominer package. Median aggregation at the well level was then performed using the aggregate function to produce well-level morphological profiles.

To correct for potential inter-plate variability, we applied a custom layout correction method based on plate-shift normalization using internal Python scripts. Biological replicates (n=3) were then merged to generate consensus profiles for each treatment condition. Feature selection was performed using a reproducibility-based filter to retain robust features across biological replicates (Woehrmann et al., 2013). Two biological repeats were conducted on separate plates, and feature values from each replicate were compared. Features with a similarity score of 0.8 or higher between repeats were retained in the final set. This process, applied to our current dataset, resulted in a robust feature set of 579 features after a single round of filtering. Additionally, an alternative feature selection approach using pycytominer (Serrano et al., 2025) was also applied. This method performs correlation analysis between features and ranks them based on how often they are highly correlated with others. Features exhibiting high correlation with multiple others were dropped. Pycytominer also filters out features with more than 5% missing values, low variance, or outlier values. Comparison of the features selected by CellProfiler 4.2.8 and the reproducibility-based method revealed significant differences; only about 20% of features overlapped between the two methods. For the purposes of this study, we focused on the reproducibility-based method to ensure consistency with our current dataset and methodology.

To visualize and assess phenotypic clustering, we applied Uniform Manifold Approximation and Projection (UMAP) for dimensionality reduction. Clustering was performed using k-means, and cluster quality was evaluated using the Adjusted Rand Index (ARI). ARI values near 1.0 indicate strong concordance between predicted and reference cluster labels, while values near 0 or negative indicate random or discordant clustering, respectively.

Annotations for biological clusters in the Endeavor dataset were transferred from reference profiles previously generated using the prior imaging solution, which serves as our benchmark for CPA-based analysis.

To evaluate the profiling capabilities of the Endeavor imaging system, we compared its performance to our previous imaging platform, which serves as an internal benchmark for clustering and phenotypic separation. Although the Endeavor captures one fewer fluorescent channel (four instead of five), it achieved similarly high clustering performance, as assessed by the Adjusted Rand Index (ARI), which quantifies the concordance between known biological classes and unsupervised clusters. Prior studies have shown that dropping any single channel has minimal impact on overall profile strength (Cimini et al., 2023), and specifically, excluding the RNA channel affects replicate-level reproducibility more than mechanism of action retrieval (Tromans-Coia et al., 2023). In this study, the same plate was imaged on both systems, meaning RNA was present in the assay even though it was not directly imaged in the Endeavor green channel. Notably, we observed signal leakage from the RNA stain into the ER channel, evidenced by visible nucleoli within the nuclei (Figure 8, ER panel). While this unintended crosstalk complicates the assessment of the RNA channel’s necessity, it suggests that acquiring the green channel with either stain may preserve most informative features. We therefore cannot definitively recommend which channel to exclude without further targeted experiments, but our results indicate that dropping the RNA channel does not substantially compromise profiling performance in this context.

We tested two feature sets extracted from the Endeavor data: 251 features obtained using a stringent reproducibility threshold (correlation ≥ 0.9), and 565 features using a more lenient threshold (correlation ≥ 0.85). Both sets performed well, achieving ARIs of 0.66 and 0.59, respectively (Figure 3A,B). For the prior imaging system, the standard 579-feature set originally optimized for CellProfiler 3 was used. After removing RNA-associated features to match the Endeavor’s channel configuration, 438 features remained, achieving an ARI of 0.52 (Figure 3C). When including the RNA features, the full 579-feature set yielded an ARI of 0.59 (Figure 3D). These results demonstrate that the Endeavor can match or even exceed the performance of the prior imaging system, depending on the selected feature set. Specifically, the most stringent Endeavor-derived feature set (correlation ≥ 0.9) achieved a higher ARI (0.66) than the full 579-feature set from the prior system (ARI 0.59), despite the Endeavor operating with only four imaging channels compared to five. This suggests that the reduced channel configuration in the Endeavor does not compromise clustering performance and may still support high-quality morphological profiling.

The feature selection process for the Endeavor resulted in substantial dimensionality reduction—from an initial ~4,000 features to as few as 251—without loss of clustering performance. This reduction reflects the application of stringent reproducibility filters, which resulted in a best performance regarding clustering of known phenotypes. Notably, the RNA stain was still present in the assay and some signal leakage into the ER channel was observed, suggesting partial retention of RNA-related morphological information. Thus, the dimensionality reduction does not imply degraded image quality, but rather a more focused and reproducible feature set tailored to the Endeavor’s imaging configuration.

To understand whether specific feature types contributed to the differences in performance, we analyzed the distribution of feature classes. When comparing the 251 Endeavor features to the 579 features used in the prior system, we observed distinct patterns (Figure 4). The Endeavor profiles were enriched in AreaShape features, which accounted for nearly 50% of the selected features, suggesting a strong morphological signal. In contrast, Texture features dominated the prior system’s profiles, reflecting differences in image acquisition and resolution.

When expanding the Endeavor’s feature set to 565 (Figure 5), we found that all 251 features were retained, with additional features primarily from the Texture class. This shift resulted in a feature type distribution more similar to that of the prior system, highlighting how selection thresholds influence the balance between morphology- and texture-derived features.

To evaluate the potential for faster acquisition through binning, we simulated 2×2 binning using nearest-neighbor pixel aggregation via the CellProfiler “Resize” module. While this post-hoc approach does not replicate the full benefits of on-sensor binning—such as improved signal- to-noise ratio due to charge-level pixel integration—it provides a conservative estimate of the impact on feature quality.

Clustering performance remained consistent across both feature sets, indicating that even this simplified binning approach does not significantly degrade the quality of extracted features (Figure 6). In practice, true on-sensor 2×2 binning would yield a fourfold reduction in image storage requirements, faster downstream image analysis, and potentially shorter exposure times, all while improving signal- to-noise. Specifically, on-sensor binning combines charge from adjacent pixels during acquisition, leading to an approximate twofold increase in signal-to-noise ratio. In contrast, our simulation used a post-hoc nearest-neighbor resizing method, which does not replicate this signal enhancement and therefore likely underestimates the benefits of hardware-level binning. These findings suggest that true on-sensor binning could substantially enhance throughput and efficiency without compromising the quality of biological insight.

Representative images acquired with both the Endeavor and the prior imaging system revealed clear cellular morphology across all channels (Figures 7-8). While the Endeavor lacks the RNA channel, key subcellular structures—DNA, ER, Actin/Golgi/Membrane, and Mitochondria—remain well resolved. Images were normalized for contrast to highlight structural details. Despite differences in acquisition configuration, the Endeavor consistently captured phenotypic differences across treatments, supporting its use in high-content profiling workflows.

The Araceli Endeavor® imaging system delivers robust and scalable CPA profiles suitable for high-content screening, even with fewer imaging channels and a reduced feature set. Despite using only four channels compared to five in the prior system, the Endeavor maintained or exceeded clustering performance, demonstrating that high-resolution morphological profiling is achievable with streamlined configurations.

The platform supports faster image acquisition, lower data storage demands, and shows strong compatibility with future enhancements like 2×2 binning—without compromising the quality of biological insight. Notably, the Endeavor imaged the same assay plates in just 7 minutes per 384-well plate, acquiring four sites per well (5.7 mm² total area) across five channels, including transmitted light. In comparison, the prior imaging solution required approximately 3.5 hours per plate to acquire nine sites per well (4.6 mm²). This represents a substantial gain in acquisition speed and spatial sampling efficiency. The greater area coverage achieved by the Endeavor has the potential to increase the number of detectable phenotypes by capturing more cells per well, thereby enhancing the statistical power and sensitivity of phenotypic profiling (Tromans-Coia et al., 2023). These results position the Endeavor as a compelling solution for large-scale profiling applications, offering both operational efficiency and scientific depth.

Image Analysis Software: CellProfiler 4 was selected over version 3 due to improved performance with large images and more availability of feature extraction modules, ensuring efficient and accurate processing of high- throughput data.

Feature Selection Strategy: Feature sets were filtered using a reproducibility-based method to reduce redundancy and focus on biologically informative features. This approach yielded smaller, faster-to-analyze profiles that preserved clustering performance.

Feature Set Adaptability: Optimal feature selection may vary depending on the specific compound library or phenotypic space being explored. The availability of both stringent and broader feature sets (e.g., 251 vs. 565 features) provides flexibility to adapt the pipeline based on experimental goals.

Bray, M.-A., Singh, S., Han, H., Davis, C.T., Borgeson, B., Hartland, C., Kost- Alimova, M., Gustafsdottir, S.M., Gibson, C.C., Carpenter, A.E., 2016. Cell Painting, a high-content image-based assay for morphological profiling using multiplexed fluorescent dyes. Nat. Protoc. 11, 1757–1774. https://doi.org/10.1038/nprot.2016.105

Cimini, B.A., Chandrasekaran, S.N., Kost-Alimova, M., Miller, L., Goodale, A., Fritchman, B., Byrne, P., Garg, S., Jamali, N., Logan, D.J., Concannon, J.B., Lardeau, C.-H., Mouchet, E., Singh, S., Shafqat Abbasi, H., Aspesi, P., Boyd, J.D., Gilbert, T., Gnutt, D., Hariharan, S., Hernandez, D., Hormel, G., Juhani, K., Melanson, M., Mervin, L.H., Monteverde, T., Pilling, J.E., Skepner, A., Swalley, S.E., Vrcic, A., Weisbart, E., Williams, G., Yu, S., Zapiec, B., Carpenter, A.E., 2023. Optimizing the Cell Painting assay for image-based profiling. Nat. Protoc. 18, 1981–2013. https://doi.org/10.1038/ s41596-023-00840-9

Pahl, A., Schölermann, B., Lampe, P., Rusch, M., Dow, M., Hedberg, C., Nelson, A., Sievers, S., Waldmann, H., Ziegler, S., 2023. Morphological subprofile analysis for bioactivity annotation of small molecules. Cell Chem. Biol. 30, 839-853.e7. https://doi.org/10.1016/j.chembiol.2023.06.003

Serrano, E., Chandrasekaran, S.N., Bunten, D., Brewer, K.I., Tomkinson, J., Kern, R., Bornholdt, M., Fleming, S.J., Pei, R., Arevalo, J., Tsang, H., Rubinetti, V., Tromans- Coia, C., Becker, T., Weisbart, E., Bunne, C., Kalinin, A.A., Senft, R., Taylor, S.J., Jamali, N., Adeboye, A., Abbasi, H.S., Goodman, A., Caicedo, J.C., Carpenter, A.E., Cimini, B.A., Singh, S., Way, G.P., 2025. Reproducible image-based profiling with Pycytominer. Nat. Methods 22, 677–680. https://doi.org/10.1038/s41592-025-02611-8

Stirling, D.R., Swain-Bowden, M.J., Lucas, A.M., Carpenter, A.E., Cimini, B.A., Goodman, A., 2021. CellProfiler 4: improvements in speed, utility and usability. BMC Bioinformatics 22, 433. https://doi.org/10.1186/s12859-021-04344-9

Tromans-Coia, C., Jamali, N., Abbasi, H.S., Giuliano, K.A., Hagimoto, M., Jan, K., Kaneko, E., Letzsch, S., Schreiner, A., Sexton, J.Z., Suzuki, M., Trask, O.J., Yamaguchi, M., Yanagawa, F., Yang, M., Carpenter, A.E., Cimini, B.A., 2023. Assessing the performance of the Cell Painting assay across different imaging systems. Cytom. Part J. Int. Soc. Anal. Cytol. 103, 915–926. https://doi.org/10.1002/cyto.a.24786

Woehrmann, M.H., Bray, W.M., Durbin, J.K., Nisam, S.C., Michael, A.K., Glassey, E., Stuart, J.M., Lokey, R.S., 2013. Large-scale cytological profiling for functional analysis of bioactive compounds. Mol. Biosyst. 9, 2604–2617. https://doi. org/10.1039/C3MB70245F