henhaobio
Sign in

HeliyonPathologyPeer-reviewed journal article

Advanced Tissue Technologies of Blood-Brain Barrier Organoids as High Throughput Toxicity Readouts in Drug Development

Luisa Bell, Claire Simonneau, Chiara Zanini, Elena KassianidouiD, Christelle Zundel, Rachel Neff, Bernd Steinhuber, Marco Tecilla, Alex OdermattiD, Roberto VillaseñoriD, Nadine Stokar‐RegenscheitiD

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

Abstract

Recent advancements in engineering Complex in vitro models (CIVMs) such as Blood-brain barrier (BBB) organoids offer promising platforms for preclinical drug testing. However, their application in drug development, and especially for the regulatory purposes of toxicity assessment, requires robust and reproducible techniques. Here, we developed an adapted set of orthogonal image-based tissue methods including hematoxylin and eosin staining (HE), immunohistochemistry (IHC), multiplex immunofluorescence (mIF), and Matrix Assisted Laser Desorption/Ionization Mass Spectrometry Imaging (MALDI-MSI) to validate CIVMs for drug toxicity assessments. We developed an artificial intelligence (AI) algorithm to increase the throughput and the reliability of histomorphologic evaluations of apoptosis for in vitro toxicity studies. Our data highlight the potential to integrate advanced morphology-based readouts such as histological techniques and digital pathology algorithms for use on CIVMs, as part of a standard preclinical drug development assessment. Graphical abstract O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=121 SRC="FIGDIR/small/611987v1_ufig1.gif" ALT="Figure 1"> View larger version (40K): org.highwire.dtl.DTLVardef@190530eorg.highwire.dtl.DTLVardef@3cce5forg.highwire.dtl.DTLVardef@10ec3borg.highwire.dtl.DTLVardef@1051585_HPS_FORMAT_FIGEXP M_FIG C_FIG The graphical abstract was partially created with biorender.com. HighlightsO_LIAdvanced Complex in vitro models (CIVMs) like Blood-brain barrier (BBB) organoids show promise for preclinical drug testing. However, robust and reproducible techniques are crucial for the acceptance of CIVMs in drug development processes, especially for toxicity assessments which are highly regulated by health authorities. C_LIO_LIWe developed orthogonal image-based readouts on histological sections to enable the use of BBB organoids for future compound toxicity assessment. C_LIO_LIA newly established artificial intelligence (AI) algorithm provides automated and label-free detection of apoptotic cells in drug screening using BBB organoids and provides an alternative killing assay on single cell resolution (40x) to current standards. C_LI

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (13 Sept 2024) · CC BY 4.0

Paraffin histology, multiplex immunofluorescence and MALDI imaging work on blood-brain barrier organoids; an AI apoptosis detector on hematoxylin-eosin slides (F1 0.817 vs pathologists) tracked staurosporine dose responses.

Complex in vitro models such as blood-brain barrier (BBB) organoids need standardized, higher-throughput readouts to be accepted for regulated toxicity testing, where histopathology is already a main in vivo endpoint. Extending earlier hydrogel-patterned BBB organoid arrays, it moves paraffin histology, automated immunohistochemistry (IHC) and multiplex immunofluorescence onto organoid sections and tissue microarrays and adds automated single-cell apoptosis scoring on hematoxylin and eosin (HE) slides.

Key results

  1. SuggestedOn 40× HE whole-slide images the AI segmented nuclei almost perfectly. Its apoptotic-nucleus calls scored higher against three pathologists (object-based) than against Caspase-3 IHC training labels (pixel-based), but the metrics differ. Scores come from the 44 organoids also used for ground truth and training.

    Nuclei segmentation, pixel-based on training organoids: P 0.978, R 0.982, F1 0.98. Apoptotic nuclei vs Caspase-3 IHC labels, pixel-based: P 0.341, R 0.770, F1 0.460. Object-based vs three blinded veterinary pathologists: P 0.767, R 0.884, F1 0.817. Pathologist inter-rater accuracy M = 0.93, kappa 0.85. · High object-based accuracy for the detection of…; Workflow; AI algorithm design and optimization; AI algorithm validation

  2. SuggestedOn a separate set of 16 naive organoids (staurosporine-treated or untreated) not used in training, recall fell moderately and precision fell sharply. Many apoptotic nuclei called by the AI were not called by the pathologists.

    Object-based comparison with three blinded pathologists on a tissue microarray HE whole-slide image of 16 organoids: recall 0.709, precision 0.478, F1 0.571. On the development organoids the figures were P 0.767, R 0.884, F1 0.817. The authors describe this as good performance on unseen data. · High object-based accuracy for the detection of…; AI algorithm validation

  3. ShownTwo readouts both rose with staurosporine dose: the AI readout (percent apoptotic nuclei per section on 40× HE) and a live CellEvent Caspase-3/7 fluorescence assay imaged at 5×. The AI showed a higher baseline in untreated organoids, which the authors read as better detection of rare apoptotic events.

    Staurosporine at 250, 500, 1000 and 2000 nM for 8 h, plus an untreated control. Each condition had five wells of approx. 55 organoids, split between HE processing (approx. 165 organoids) and the Caspase-3/7 assay (two wells per condition). Agreement is shown only as per-organoid plots (mean ± SD); no correlation coefficient or statistical test is reported. · The newly developed automated image analysis AI…; Staurosporine treatment of BBB organoids; AI algorithm validation

  4. ShownFive-plex immunofluorescence on paraffin sections showed organoids with an outer P-glycoprotein (P-gp)+ endothelial layer around an inner meshwork of GFAP+ astrocytes and NG-2+ pericytes. Ki67 appeared in all three cell types, with baseline Caspase-3+ apoptosis. The same markers stained human frontal cortex.

    Representative images only, with no quantification. Organoid panel: P-gp, NG2, Ki67, Caspase-3, GFAP (Opal dyes, Vectra Polaris at 40×). Cortex panel: P-gp, GFAP, NG2 on one FFPE frontal-cortex block from a single 83-year-old male donor. IHC also detected P-gp, NG-2 and GFAP in monoculture cell pellets, at variable fractions. · BBB organoids recapitulate IHC expression…; Human brain tissue sample source; Automated multiplex immunofluorescence on PFA-PE…

  5. SuggestedMatrix-assisted laser desorption/ionization (MALDI) mass spectrometry imaging on a Fourier transform ion cyclotron resonance (FTICR) instrument located the P-gp substrate rhodamine 123 in organoid cryosections. The P-gp inhibitor zosuquidar only tended to raise rhodamine 123 signal in whole-mount imaging.

    Rhodamine 123 (100 µM, 1 h) was identified at m/z 345.12337 within a 2 ppm window at 50 µm lateral resolution; widefield fluorescence was read at 508/528 nm. Zosuquidar (1 µM, 1 h pre-incubation) gave a trend toward higher organoid-core intensity per area. Neither n per group nor a statistical test is reported. · Detection of rhodamine 123 as a proof of concept…; Rhodamine 123 treatment for MALDI-MS detection; MALDI-MS imaging

Methods, data and reproducibility

Human BBB organoids were profiled on paraffin sections by automated IHC and five-plex immunofluorescence, and on gelatin cryosections by MALDI imaging. Tissue microarrays of organoids were used for HE-based AI analysis. A two-step Visiopharm U-Net, trained on 40× HE whole-slide images labelled via Caspase-3 restaining, counts apoptotic nuclei; it was checked against pathologists and a live Caspase-3/7 assay.

  • BBB organoids self-assembled from hCMEC/D3 cerebral microvascular endothelial cells, primary astrocytes and brain microvascular pericytes in GRI3D 96-well hydrogel arrays · human
  • Paraffin-embedded cell pellets of each monocultured cell type (astrocytes, pericytes, hCMEC/D3) · human · n = three cell pellets
  • FFPE frontal cortex block from a healthy donor (Asterand Bioscience / BioIVT) for marker validation · human · n = one block, one 83-year-old male donor
  • AI development set: HE 40× whole-slide images of organoids with Caspase-3 destain/restain labels and pathologist annotations · human (organoids from human cells) · n = 44 organoids, 14 WSIs (5422 nuclei and 2250 apoptotic nuclei labelled)
  • Held-out naive organoids (staurosporine-treated or untreated) annotated by three blinded pathologists · human (organoids from human cells) · n = 16 organoids
  • Staurosporine dose series (250 to 2000 nM, 8 h) plus untreated control, for AI vs Caspase-3/7 assay comparison · human (organoids from human cells) · n = 4 doses plus control; five wells of approx. 55 organoids each
  • Organoids treated with rhodamine 123, with or without zosuquidar, for whole-mount imaging and MALDI imaging · human (organoids from human cells)
Methods & tools
Visiopharm VIS 2021.09, Visiopharm pre-trained U-Net nuclei detection app (#10167), Ventana Discovery Ultra, Opal Polaris dyes (Akoya Biosciences), Vectra Polaris multispectral imaging, inForm, HALO, Hamamatsu NanoZoomer S360, Bruker solariX XR 7 T FTICR MS, SCiLS Lab, CellEvent Caspase-3/7 (Invitrogen), Fiji
Reagents & environment
Visiopharm VIS 2021.09; NDP.view2 v2.9.22; PhenoChart v1.0.12; inForm v2.4; HALO v3.2.1851.328; SCiLS Lab 2023a Core; Fiji ImageJ 1.54. No containers or requirement files.
Reproducibility gaps
No code released: the algorithm was built in proprietary Visiopharm software on a commercial pre-trained U-Net, and the custom Fiji quantification script is not shared. Ground-truth data are available from the corresponding author on reasonable request.
Limitations (6)
  • Markers seen in monoculture cell pellets may not match assembled organoids (passage number, media, co-culture, extracellular matrix). IHC is also not sensitive to low expression and is not quantitative. · stated by the authors
  • MALDI imaging is acquired slide by slide and trades 50 µm resolution against sensitivity, so throughput is limited and the output is often qualitative rather than quantitative. · stated by the authors
  • Pathologist ground truth requires training on organoid histology, including baseline apoptosis in static culture, and is partly subjective; this is why Caspase-3 IHC was added as an objective reference. · stated by the authors
  • The model learns general histomorphology and may not separate apoptosis from necrosis or other forms of cell death. · stated by the authors
  • The headline accuracy (F1 0.817) comes from the 44 organoids also used for ground truth and training. The only held-out test is 16 organoids from the same protocol and one tool compound, and there precision fell to 0.478. · noted in this summary
  • The claim that the HE-based AI beats Caspase-3 IHC compares a pixel-based F1 (0.460) with an object-based F1 (0.817). The AI vs Caspase-3/7 assay agreement has no statistics, and the 2D vs 3D equivalence rests on data not shown. · noted in this summary

Read this if you run organoids or other complex in vitro models for safety assessment and want paraffin histology, multiplex immunofluorescence or an AI apoptosis count on hematoxylin-eosin slides as standardized readouts.

Citations
21
FWCI
3.45
Field percentile
top 6%

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
THISKoh '23Acharya '24Deben '21Wang '26Jaume '24Kim '26
59% semantic matchBioengineering
I. S. Y. Koh, M. Hagiwara·bioRxiv·Feb 25, 2023