henhaobio
Sign in

Cell SystemsPharmacology & ToxicologyPeer-reviewed journal article

Cell Painting for cytotoxicity and mode-of-action analysis in primary human hepatocytes

Jessica EwaldiD, Katherine TittertoniD, Alex BäuerleiD, Alex Beatson, Daniil A. BoikoiD, Ángel Alexander CabreraiD, Jaime H. CheahiD, Beth A. CiminiiD, Bram L. GorisseniD, Joshua HarrilliD, Thouis R. JonesiD, Konrad J. KarczewskiiD, Christine CruteiD, David RouquiéiD, Srijit SealiD, Erin WeisbartiD, Brandon WhiteiD, Anne Elizabeth CarpenteriD, Shantanu SinghiD

PDFDOIPreprintCodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

High-throughput, human-relevant approaches for predicting chemical toxicity are urgently needed for better decision-making in human health. Here, we apply image-based profiling (the Cell Painting assay) and two cytotoxicity assays (metabolic and membrane damage readouts) to primary human hepatocytes after exposure to eight concentrations of 1085 compounds that include pharmaceuticals, pesticides, and industrial chemicals with known liver toxicity-related outcomes. Three computational methods (CellProfiler, a Cell Painting-specific convolutional neural network, and a pretrained vision transformer) were compared to extract morphology features from single cells or entire images. We used these morphology features to predict activity in the measured cytotoxicity assays, as well as in 412 curated ToxCast assays that span cytotoxicity, cell-based, and cell-free categories. We found that the morphological profiles detect compound bioactivity at lower concentrations than standard cytotoxicity assays. In supervised analyses, they predict cytotoxicity and targeted cell-based assay readouts, but not cell-free assay readouts. We also found that the various feature extraction methods performed relatively similarly and that filtering out non-bioactive or cytotoxic concentrations did not boost supervised assay prediction performance for any assay endpoint category, although it did have a large influence on unsupervised cluster analysis. We envision that image-based profiling could serve as a key component of modern safety assessment.

In-depth summary

Claude Opus 5.5 · xhigh effort · 5 Nov 2026 · from the full text of version 1 (24 Jan 2025) · CC BY 4.0

Cell Painting in primary human hepatocytes exposed to 1,085 compounds detects bioactivity at lower concentrations than cytotoxicity assays and predicts cell-based, but not cell-free, ToxCast endpoints across feature extractors.

High-throughput in vitro screens are needed to assess chemical hepatotoxicity at scale, as running hundreds of targeted ToxCast assays is slow and unsupervised morphological clustering can be confounded by general cell stress. Applying supervised machine learning to 8-concentration Cell Painting profiles in primary human hepatocytes benchmarks three classical and deep-learning feature extractors and concentration-filtering strategies across 412 ToxCast endpoints.

Key results

  1. ShownCell Painting morphological profiling in primary human hepatocytes detects compound bioactivity for up to 59% of tested chemicals and at lower concentrations than biochemical cytotoxicity assays.

    Across 1,085 OASIS compounds (8 concentrations, 0.01-100 uM), Mahalanobis distance curve-fitting detected bioactivity for 16-59% of compounds (~50-60% for CP-CNN, DINO, and CellProfiler vs 40% in RealTime-Glo MT, 20% in cell count, and 13% in LDH); across 131 co-active compounds, morphological PODs were lower than MT (2.5-fold), cell counts (7.9-fold), and LDH (15.8-fold, all p < 1.0e-14). · 1.2.1 40% of compounds induce activity as…; 1.2.2 Morphological profiling detects…; 1.3 Morphological profiles are more sensitive…

  2. ShownMorphological profiles predict continuous metabolic viability readouts better than cell-count and plate-position baselines, with prediction outliers pinpointing specific signaling and metabolic targets.

    XGBRegressor on Cell Painting profiles beat technical baselines for MT; wells better predicted by morphology (n = 138, 1.0%) enriched in 480 protein targets including cell cycle (29/97, FDR = 1.46e-12), PI3K-Akt (53/256, FDR < 1.0e-14), MAPK (47/249, FDR = 1.90e-12), and p53 (21/64, FDR = 3.43e-10) pathways, whereas overpredicted MT wells (n = 261, 2.0%) enriched in 147 targets including CYP450s and transporters. · 2.1 Morphological profiles predict cytotoxicity…; 2.2 Prediction outliers reveal biologically…

  3. ShownSupervised classifiers trained on hepatocyte Cell Painting profiles predict compound activity in ToxCast cytotoxicity and targeted cell-based assays, but not cell-free assays.

    Across 412 curated human ToxCast endpoints (48 consensus cytotoxicity, 292 cytotoxicity-filtered cell-based, and 72 cell-free endpoints covering 963 overlapping compounds), binary XGBoost classifiers trained on median-aggregated CellProfiler features exceeded random and cell-count baselines for cytotoxicity and cell-based categories, while cell count alone only predicted cytotoxicity. · 3.1 Preparing ToxCast assay endpoints; 3.2 Cell Painting profiles predict activity in…; ToxCast curation

  4. ShownClassical CellProfiler features, pretrained CP-CNN, and DINOv2 vision transformers achieve comparable predictive accuracy, and aggregating all concentrations performs nearly as well as POD-filtered consensus profiles.

    DINOv2 gave a 0.02 higher mean AUROC on cell-based ToxCast assays than CellProfiler (p = 2.7e-5) and CP-CNN (p = 7.4e-6) with no other AUROC/PRAUC differences; filtering to concentrations above the morphological POD raised mean AUROC by only 0.04 (FDR < 1.0e-14) for cytotoxicity and 0.02 (FDR = 0.002) for cell-based assays, and excluding cytotoxic concentrations gave no improvement. · 3.2.1 Consensus profiles that contain all…; 3.2.2 Alternative cell representations perform…

Methods, data and reproducibility

Exposed 5-donor pooled primary human hepatocytes in 384-well plates for 44 h to 1,085 compounds at 8 concentrations (0.01-100 uM, 2 replicates) and assayed LDH release, RealTime-Glo MT, and 5-channel Cell Painting. Extracted features via CellProfiler (5,640/cell), DeepProfiler CP-CNN, and DINOv2 ViT-L/14 for Mahalanobis curve-fitting (fastbmdR) and 5-fold XGBoost modeling against MT, LDH, and 412 ToxCast endpoints.

  • 5-channel + brightfield 40x Cell Painting, cell counts, LDH, and RealTime-Glo MT in 5-donor primary human hepatocytes across 1,085 OASIS compounds · human · n = 1,085 compounds x 8 concentrations x 2 replicates (~800 cells, 15 FOV/well) · cpg0037-oasis
  • Curated human ToxCast/Tox21 bioactivity hit calls and AC50 values from EPA invitrodb v4.1 matched by DTXSID · human · n = 963 compounds across 412 assays (48 cytotoxicity, 292 cell-based, 72 cell-free)
Methods & tools
CellProfiler (v4.2.4), Cellpose (v2.3.2), DeepProfiler (CP-CNN), DINOv2 (ViT-L/14), Pycytominer (v1.2.0), fastbmdR (v0.0.0.9), XGBoost (XGBRegressor / XGBClassifier), invitrodb (v4.1), SQLalchemy (v1.4.54), Distributed-CellProfiler
Software & environment
CellProfiler 4.2.4, Cellpose 2.3.2, DeepProfiler (commit bed9d6a with PR 358), DINOv2 ViT-L/14, Pycytominer 1.2.0, fastbmdR 0.0.0.9, invitrodb 4.1, SQLalchemy 1.4.54, XGBoost
Limitations (3)
  • Hepatocyte Cell Painting profiles cannot predict compound bioactivity in cell-free ToxCast assays measuring isolated protein binding or enzymatic activity. · stated by the authors
  • The LDH membrane-damage assay suffered from strong plate and well-position effects, leading to low concordance between technical replicates. · stated by the authors
  • High-content imaging features are less directly interpretable for specific molecular events than transcriptomics or targeted assays, and ToxCast endpoints were measured across heterogeneous cell types, concentrations, and exposure durations. · stated by the authors

Read this if you use Cell Painting or high-content imaging for chemical toxicity screening, mode-of-action prediction, or benchmarking deep-learning versus engineered morphological representations.

Citations
16
FWCI
29.50
Field percentile
top 1%

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISOllitrault '26Seal '24Dahlin '22Cross-Zamir '22Lambert '25Wedler '26
74% semantic matchPharmacology & Toxicology
Guillaume Ollitrault, Constance A. Mitchell, Stephen Ferguson+4 more·bioRxiv·Oct 12, 2026