Student or educator? Sign up with your school email and get 6 months of Pro free.

The DOI matches its Crossref record; 56 of the 57 references it cites were found in a registry or an index; 16 of the 29 sources its claims cite were checked for support.

A statistical framework for high-content phenotypic profiling using cellular feature distributions

Open access

Abstract

High-content screening (HCS) uses microscopy images to generate phenotypic profiles of cell morphological data in high-dimensional feature space. While HCS provides detailed cytological information at single-cell resolution, these complex datasets are usually aggregated into summary statistics that do not leverage patterns of biological variability within cell populations. Here we present a broad-spectrum HCS analysis system that measures image-based cell features from 10 cellular compartments across multiple assay panels. We introduce quality control measures and statistical strategies to streamline and harmonize the data analysis workflow, including positional and plate effect detection, biological replicates analysis and feature reduction. We also demonstrate that the Wasserstein distance metric is superior over other measures to detect differences between cell feature distributions. With this workflow, we define per-dose phenotypic fingerprints for 65 mechanistically diverse compounds, provide phenotypic path visualizations for each compound and classify compounds into different activity groups.

Overview

CoChat Summary

TL;DR

The paper presents a broad-spectrum high-content screening (HCS) assay and statistical workflow that uses cell-level feature distributions rather than well-averaged summaries to profile phenotypic responses. It reports that the Wasserstein/earth mover’s distance (EMD) is more sensitive than Z-score or Kolmogorov–Smirnov metrics for detecting differences between cell feature distributions and uses the workflow to classify 65 compounds into activity groups.

Key findings

  • The assay labels ten cellular compartments across four panels and extracts 174 texture, shape, count, and intensity features; feature reduction removes 60% of features, leaving 69 active features.
  • Positional effects were common for intensity features: 45% of intensity-related features showed significant row or column dependency (P < 0.0001), versus only 6% of morphological features and cell counts.
  • EMD outperformed Z-score and KS metrics for detecting replicate dissimilarity and distribution differences; it was used to profile 65 compounds at seven concentrations with three technical replicates, using 330 control wells and 455 treatment wells.
  • Phenotypic profiling separated treatments into four broad groups—control, “low stress,” active phenotypes, and broad cytotoxic responses—and showed dose-dependent trajectories; for example, methotrexate reduced cell counts by ~25% and induced G1 arrest, while tanespimycin caused G2/M arrest at 2.5 μM.

Methods

  • U2OS cells were seeded in 384-well plates at 1,800 cells/well; 65 diverse compounds were tested at seven concentrations with three technical replicates, alongside 55 DMSO control wells placed in a diagonal pattern. Four staining panels labeled DNA, RNA, mitochondria, plasma membrane/Golgi, lysosomes, peroxisomes, lipid droplets, ER, actin, and tubulin.
  • Images were acquired on a Cellomics ArrayScan XTI with a ×20 objective, nine fields per well; 174 features were extracted using Cellomics Compartmental Analysis.
  • Analysis included two-way ANOVA for row/column positional effects, median polish adjustment, B-score and per-cell BZ standardization to per-plate controls, comparison of Z-score, KS, and EMD metrics, replicate merging, EMD profiling against a global control, feature reduction, hierarchical clustering, and UMAP; analyses used R and ggplot2.

Limitations

  • The paper states that EMD does not take directional changes into account, and that extending it to account for direction could improve downstream profile similarity analysis.
  • The paper states that the dataset is not ideal for compound similarity or MOA classification because the 65 compounds were chosen to be diverse with low structural similarity; it also notes that no community-wide consensus or standard HCS workflow exists, limiting cross-study comparison and data integration.
  • Evident from the text: the study uses a single cell line (U2OS) and a 24 h treatment period, and some features—especially lysosomal, lipid, and RNA puncta features—were irreproducible or removed due to weak signal, bleed-through, or low biological relevance, which may limit generalizability.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1038/s42003-022-04343-3

Cites

56 of 56 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Fig. 1 High-content screening (HCS) assay panels and data analysis workflow. a The high-content screening (HCS) assay system comprises fluorescent stains and genetically encoded markers for ten cellular components, split across four panels. PMG: plasma membrane and Golgi, ER: endoplasmic reticulum. b Sample images of U2OS-labeled cellular components. Scale bar: 20 µm. c-h Overview of analysis workflow. c Assays are performed in 384-well plates. Cytological features capturing information on texture, shape, count, and intensity are measured for each panel using high-throughput microscopy. d Well- and cell-level data for each plate are adjusted and standardized for any positional (row or column) effects detected. e Assay reproducibility is assessed according to consistency among replicates of the same treatment. Three statistical metrics are compared by their sensitivity in detecting differences between replicates. f Consistent replicate data are aggregated and differences between treatments and controls are scored by comparing feature distributions. g Redundant, noisy, and uninformative features are removed. h Fully processed data are used to generate normalized phenotypic fingerprints, which are analyzed and visualized using hierarchical clustering and dimensional reduction.
Fig. 1 High-content screening (HCS) assay panels and data analysis workflow. a The high-content screening (HCS) assay system comprises fluorescent stains and genetically encoded markers for ten cellular components, split across four panels. PMG: plasma membrane and Golgi, ER: endoplasmic reticulum. b Sample images of U2OS-labeled cellular components. Scale bar: 20 µm. c-h Overview of analysis workflow. c Assays are performed in 384-well plates. Cytological features capturing information on texture, shape, count, and intensity are measured for each panel using high-throughput microscopy. d Well- and cell-level data for each plate are adjusted and standardized for any positional (row or column) effects detected. e Assay reproducibility is assessed according to consistency among replicates of the same treatment. Three statistical metrics are compared by their sensitivity in detecting differences between replicates. f Consistent replicate data are aggregated and differences between treatments and controls are scored by comparing feature distributions. g Redundant, noisy, and uninformative features are removed. h Fully processed data are used to generate normalized phenotypic fingerprints, which are analyzed and visualized using hierarchical clustering and dimensional reduction.CC BY · PubMed Central
Fig. 2 HCS experimental design and inspection of cell counts and cell cycle distributions. a A set of 65 compounds with diverse chemical structures and a wide range of ChEMBL target classes were selected to examine the performance of the marker panels. Hierarchical clustering using the Tanimoto distance metric shows pairwise chemical structure similarity. b The 384-well plate layout for chemical screening: control wells (DMSO treatment) are placed in a diagonal pattern, and a dilution series of compounds arranged horizontally with high to low concentration (left to right). The outermost rows and columns are excluded from scanning and analysis to mitigate edge effects due to evaporation. c Dilution series of 65 compounds and control wells were spread over two plates, with three replicates for each condition. The heatmap shows well-level cell counts in different plates. d Scatter plot showing cell counts in control (left panel) and treatment (right panel) wells, ordered by plate and replicate number (grey dashed lines correspond to min, max and median of control cell count). e Hoechst 33342 (nuclear DNA) stain distribution showing density curves for individual control wells (red) and aggregated control wells (global control, black, sample size: 265,638 cells). f Distribution of total nucleus intensity of cells after the treatment with increasing doses of Mitoxantrone (sample sizes from low to high dosage (n0.078125 – n5): 711, 669, 559, 372, 325, 363, 363 cells). g Cells treated with different compounds and concentrations (A = irinotecan 5 μM, B = monensin 0.3125 μM, C = rapamycin 10 μM, D = vincristine 10 μM) show a diverse distribution of total nucleus intensity (sample size: Control = 265,638, nA = 1544, nB = 1817, nC = 1849, nD = 1173 cells).
Fig. 2 HCS experimental design and inspection of cell counts and cell cycle distributions. a A set of 65 compounds with diverse chemical structures and a wide range of ChEMBL target classes were selected to examine the performance of the marker panels. Hierarchical clustering using the Tanimoto distance metric shows pairwise chemical structure similarity. b The 384-well plate layout for chemical screening: control wells (DMSO treatment) are placed in a diagonal pattern, and a dilution series of compounds arranged horizontally with high to low concentration (left to right). The outermost rows and columns are excluded from scanning and analysis to mitigate edge effects due to evaporation. c Dilution series of 65 compounds and control wells were spread over two plates, with three replicates for each condition. The heatmap shows well-level cell counts in different plates. d Scatter plot showing cell counts in control (left panel) and treatment (right panel) wells, ordered by plate and replicate number (grey dashed lines correspond to min, max and median of control cell count). e Hoechst 33342 (nuclear DNA) stain distribution showing density curves for individual control wells (red) and aggregated control wells (global control, black, sample size: 265,638 cells). f Distribution of total nucleus intensity of cells after the treatment with increasing doses of Mitoxantrone (sample sizes from low to high dosage (n0.078125 – n5): 711, 669, 559, 372, 325, 363, 363 cells). g Cells treated with different compounds and concentrations (A = irinotecan 5 μM, B = monensin 0.3125 μM, C = rapamycin 10 μM, D = vincristine 10 μM) show a diverse distribution of total nucleus intensity (sample size: Control = 265,638, nA = 1544, nB = 1817, nC = 1849, nD = 1173 cells).CC BY · PubMed Central
Fig. 3 Adjustment of plate positional effects and data standardization across different plates. a For each feature, a two-way ANOVA model is applied to detect non-uniformity among control wells due to row and column effects. Here total nucleus intensity is shown as an example, with the first replicate for plate 1 (“plate1 rep1”) showing positional effects among control wells. All statistical output from two-way ANOVA analysis, including F-statistic and corresponding P values for both row and column effects detection, is provided in Supplementary Data 1. b Summary of row effects, shown as negative log of P values (y axis) for each feature across different plates (x axis), grouped by individual markers and cell counts. For example, the RNA marker shows a clear separation of intensity (light green curves) from non-intensity features (dark-green curves). c Well-level positional effects adjustment and standardization: when positional effects are detected (−log(P) >10) for a particular feature, median polish is applied. The B score standardizes well medians to per-plate controls to account for plate-to-plate variation. d Cell-level adjustment and standardization: cell populations in each well are adjusted for positional effects based on the adjustment amount calculated at well level. The adjusted cell-level data are further standardized to per-plate controls.
Fig. 3 Adjustment of plate positional effects and data standardization across different plates. a For each feature, a two-way ANOVA model is applied to detect non-uniformity among control wells due to row and column effects. Here total nucleus intensity is shown as an example, with the first replicate for plate 1 (“plate1 rep1”) showing positional effects among control wells. All statistical output from two-way ANOVA analysis, including F-statistic and corresponding P values for both row and column effects detection, is provided in Supplementary Data 1. b Summary of row effects, shown as negative log of P values (y axis) for each feature across different plates (x axis), grouped by individual markers and cell counts. For example, the RNA marker shows a clear separation of intensity (light green curves) from non-intensity features (dark-green curves). c Well-level positional effects adjustment and standardization: when positional effects are detected (−log(P) >10) for a particular feature, median polish is applied. The B score standardizes well medians to per-plate controls to account for plate-to-plate variation. d Cell-level adjustment and standardization: cell populations in each well are adjusted for positional effects based on the adjustment amount calculated at well level. The adjusted cell-level data are further standardized to per-plate controls.CC BY · PubMed Central
Fig. 4 Statistical metric comparison and feature reproducibility. a Feature reproducibility is assessed by estimating statistical distance among all pairwise replicates in both control samples (left) and treatment samples (right). b Hypothetical probability density (PDF) and cumulative density (CDF) curves for two random samples of the same feature are illustrated to show how the Kolmogorov–Smirnov (KS) distance and Wasserstein metric (EMD) are estimated. c Distributions of statistical scores measured by all pairwise differences between replicates are consistent among both treatments and controls, with EMD score showing higher sensitivity in detecting discrepancies. A full summary of replicate pairwise differences is provided in Supplementary Data 2 which lists feature, treatment (compound_concentration), plate, well id, sample size (as cell count), KS score, EMD score, and Z-score. d Features are sorted by their average EMD score between all replicates as an indicator of reproducibility. A high average EMD score indicates higher variation of a feature among replicates (low reproducibility). Features with poor reproducibility (outliers) falling above the upper interquartile (IQR) threshold value of 0.65 (upper threshold value = 1.5 × IQR + upper quartile) are highlighted in red.
Fig. 4 Statistical metric comparison and feature reproducibility. a Feature reproducibility is assessed by estimating statistical distance among all pairwise replicates in both control samples (left) and treatment samples (right). b Hypothetical probability density (PDF) and cumulative density (CDF) curves for two random samples of the same feature are illustrated to show how the Kolmogorov–Smirnov (KS) distance and Wasserstein metric (EMD) are estimated. c Distributions of statistical scores measured by all pairwise differences between replicates are consistent among both treatments and controls, with EMD score showing higher sensitivity in detecting discrepancies. A full summary of replicate pairwise differences is provided in Supplementary Data 2 which lists feature, treatment (compound_concentration), plate, well id, sample size (as cell count), KS score, EMD score, and Z-score. d Features are sorted by their average EMD score between all replicates as an indicator of reproducibility. A high average EMD score indicates higher variation of a feature among replicates (low reproducibility). Features with poor reproducibility (outliers) falling above the upper interquartile (IQR) threshold value of 0.65 (upper threshold value = 1.5 × IQR + upper quartile) are highlighted in red.CC BY · PubMed Central
Fig. 5 Replicate aggregation and EMD profiling using global controls. a Replicates of treatment samples with sufficient reproducibility are merged to form larger populations (orange curve) for subsequent EMD profiling relative to the global control (black curve). Shown is an example using the area of the nucleus feature with a strong phenotypic response to treatment with 20 µM vincristine (sample sizes: nr1 = 348, nr2 = 386, nr3 = 366, ncontrol = 265,638 cells). The empirical CDF (ECDF) curve shows a global right shift in the treatment condition, indicating an increase in global nuclear area. Inset: area between distributions measured by EMD (gray). b, c PDFs and (insets) CDFs for normalized control (per well) cell populations (b) and replicate-merged treatment populations (c). Differences between individual distributions relative to the global control are measured using the EMD metric and sample sizes corresponding to each statistical difference are listed in Supplementary Data 3 for 330 control (nc-min = 419, nc-max = 990 cells) and 455 treatment samples (nt-min = 52, nt-max = 2763 cells). d Radial plot of scaled EMD scores for 69 measured features among individual controls (gray lines). The EMD profile is log-transformed and min-max scaled to [0,1]. The median score of all controls fluctuates between 0.16 and 0.46 (black dashed line). e Radial plot of residual EMD scores of individual controls (gray lines) relative to the median score (black dashed line, zero). Residual score is defined as the difference between the score of the individual control and the median of all controls. The residuals naturally fluctuate around median zero with values between (−0.29, 0.45). The values have been offset by 0.5 to expand the plot for better visualization. f Radial plot of residual scores for Vincristine treatment at multiple concentrations, relative to control median. d–f Feature labels are color-coded by their corresponding cellular components.
Fig. 5 Replicate aggregation and EMD profiling using global controls. a Replicates of treatment samples with sufficient reproducibility are merged to form larger populations (orange curve) for subsequent EMD profiling relative to the global control (black curve). Shown is an example using the area of the nucleus feature with a strong phenotypic response to treatment with 20 µM vincristine (sample sizes: nr1 = 348, nr2 = 386, nr3 = 366, ncontrol = 265,638 cells). The empirical CDF (ECDF) curve shows a global right shift in the treatment condition, indicating an increase in global nuclear area. Inset: area between distributions measured by EMD (gray). b, c PDFs and (insets) CDFs for normalized control (per well) cell populations (b) and replicate-merged treatment populations (c). Differences between individual distributions relative to the global control are measured using the EMD metric and sample sizes corresponding to each statistical difference are listed in Supplementary Data 3 for 330 control (nc-min = 419, nc-max = 990 cells) and 455 treatment samples (nt-min = 52, nt-max = 2763 cells). d Radial plot of scaled EMD scores for 69 measured features among individual controls (gray lines). The EMD profile is log-transformed and min-max scaled to [0,1]. The median score of all controls fluctuates between 0.16 and 0.46 (black dashed line). e Radial plot of residual EMD scores of individual controls (gray lines) relative to the median score (black dashed line, zero). Residual score is defined as the difference between the score of the individual control and the median of all controls. The residuals naturally fluctuate around median zero with values between (−0.29, 0.45). The values have been offset by 0.5 to expand the plot for better visualization. f Radial plot of residual scores for Vincristine treatment at multiple concentrations, relative to control median. d–f Feature labels are color-coded by their corresponding cellular components.CC BY · PubMed Central
Fig. 6 Hierarchical clustering and dimension reduction. a Hierarchical clustering of EMD scores for control and treatment profiles shows four main clusters: (1) control cluster, (2) “low stress” cluster, (3) active phenotypes, (4) broad phenotypic responses with cytotoxicity. b UMAP dimensional reduction of phenotypic profiles based on EMD scores for all samples. Dimensions 1 and 2 separate most controls (triangles) from treatment samples; dimension 3 additionally separates treatments by cell count. Light gray curves represent per-compound phenotypic trajectories along a concentration gradient.
Fig. 6 Hierarchical clustering and dimension reduction. a Hierarchical clustering of EMD scores for control and treatment profiles shows four main clusters: (1) control cluster, (2) “low stress” cluster, (3) active phenotypes, (4) broad phenotypic responses with cytotoxicity. b UMAP dimensional reduction of phenotypic profiles based on EMD scores for all samples. Dimensions 1 and 2 separate most controls (triangles) from treatment samples; dimension 3 additionally separates treatments by cell count. Light gray curves represent per-compound phenotypic trajectories along a concentration gradient.CC BY · PubMed Central
Fig. 7 Phenotypic characterization of selected compounds. a Phenotypic trajectories of selected compounds from different activity groups are highlighted in the UMAP. Green: low stress; purple: active (dose-insensitive); blue: active (dose-responsive); red: active and toxic. b Raw cell counts across the dilution series of representative compounds in each activity group. Each curve indicates a replicate; the dashed line represents the mean, and dotted lines the maximum and minimum of control cell counts. c Cell cycle distributions (total nucleus intensity) across the compound dilution series for four representative compounds from each activity group, color-coded as in (a). The full set of compounds as described in (a–c) are provided in Supplementary Fig. 5. d Radial plot of fingerprints for representative compounds from (c) at each of the seven concentrations measured. Feature labels are color-coded by their cytoplasmic component.
Fig. 7 Phenotypic characterization of selected compounds. a Phenotypic trajectories of selected compounds from different activity groups are highlighted in the UMAP. Green: low stress; purple: active (dose-insensitive); blue: active (dose-responsive); red: active and toxic. b Raw cell counts across the dilution series of representative compounds in each activity group. Each curve indicates a replicate; the dashed line represents the mean, and dotted lines the maximum and minimum of control cell counts. c Cell cycle distributions (total nucleus intensity) across the compound dilution series for four representative compounds from each activity group, color-coded as in (a). The full set of compounds as described in (a–c) are provided in Supplementary Fig. 5. d Radial plot of fingerprints for representative compounds from (c) at each of the seven concentrations measured. Feature labels are color-coded by their cytoplasmic component.CC BY · PubMed Central

Verification

Claims the paper makes

10 Valid · 3 Not valid · 7 Not checked · 110 Own results and interpretations

65 more claims

Create a free account to see all 130 claims and every source verdict.

See all claims — free account
  • It may state two separate claims as one. 48
  • Probably not a substantial claim. 22

Create a free account to see these claims.

Citations

Cites

Show all 57 references

Research works printed but not linked: 1

  1. Daniel, W. W. BiostatisticsBasic Concepts and Methodology for the Health Sciences (J. Wiley, 2010).

Cited by

See all 19

Metadata

Authors
Pearson, Y. E., Kremb, S., Butterfoss, G. L., Xie, X., Fahs, H., & Gunsalus, K. C.
Affiliations
New York University Abu Dhabi (AE)
Corresponding author
Gunsalus, Kristin C.
Venue
Communications Biology
Published
2022-12-22
Volume
5
Issue
1
Pages
1409-1409
ISSN
2399-3642 (ISSN-L)
OpenAlex
W4312192705
Cited by
19
References
57
Citations by year
2026: 8 · 2025: 8 · 2024: 2 · 2023: 1
FWCI
1.15
Type
journal article
Language
en
Open access
Licence
Locations
  • Communications Biology (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Funders
This work was supported by Tamkeen by an NYUAD Research Institute grant to the NYUAD Center for Genomics and Systems Biology
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Biophysics › Cell Image Analysis Techniques
  • Life Sciences › Pharmacology, Toxicology and Pharmaceutics › Drug Discovery › Computational Drug Discovery Methods
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › Single-cell and spatial transcriptomics
Keywords
high-content screening, phenotypic profiling, cell morphology, Wasserstein distance, biological replicates, feature reduction, quality control, compound classification
MeSH terms
Microscopy, Quality Control, High-Throughput Screening Assays, Workflow

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-07, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-07 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: A statistical framework for high-content phenotypic profiling using cellular feature distributions. https://app.cochat.ai/papers/10.1038/s42003-022-04343-3
The paper
Pearson, Y. E., Kremb, S., Butterfoss, G. L., Xie, X., Fahs, H., & Gunsalus, K. C. (2022). A statistical framework for high-content phenotypic profiling using cellular feature distributions. Communications Biology. https://doi.org/10.1038/s42003-022-04343-3

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.