Uncovering the mode of action of engineered T cells in patient cancer organoids

Human material

All human BC and head and neck PDO samples were retrieved from a biobank through the Hubrecht Organoid Technology (HUB; www.hub4organoids.nl). Authorizations were obtained by the medical ethical committee and biobank research ethics committee of UMC Utrecht (UMCU) at the request of HUB, to ensure compliance with the Dutch Medical Research Involving Human Subjects Act. Normal breast organoids were generated from milk obtained via the Moedermelkbank Amsterdam (Amsterdam UMC). Primary patient-derived DMG cultures (no. DMG-VI/SU-DIPG-VI) were kindly provided by M. Monje (Stanford University), M. Vinci (Ospedale Pediatrico Bambino Gecù, nos. DMG-002/OPBG-DIPG-002 and DMG-004/OPBG-DIPG-004-aa) and A. M. Carcaboso (Hospital San Juan de Dios, no. DMG-007/HSJD-DIPG-007). For TEG and WT1 T cell generation, peripheral blood of anonymous healthy donors was purchased from the Dutch blood bank (Sanquin). For CAR T cell generation, cord blood was collected with approval from the Ethical Committee of UMCU. Informed consent was obtained from all donors.

Animal material

NOD.Cg-PrkdcscidIl2rgtm1Wjl/SzJ (NSG) mice were purchased from Charles River Laboratories. Experiments were conducted with permission from the Animal Welfare Body Utrecht (nos. 4288-1-08 and 4288-1-09) as per current Dutch laws on animal experimentation. Mice were housed under 45–65% humidity and a daily 12/12-h light/dark regime, in sterile conditions using an individually ventilated cage system and fed with sterile food and water. Irradiated mice were given sterile water with antibiotic ciproxin for the duration of the experiment. Mice were randomized with equal distribution by age and initial weight measured on day 0 and divided into groups of ten (13T) or 15 (169M).

Organoid culture

Breast cancer and normal breast organoids were seeded in basement membrane extract (BME, Cultrex) in uncoated 12-well plates (Greiner Bio-one) and cultured as described previously^29,43. Briefly, Advanced DMEM/F12 was supplemented with penicillin/streptomycin (pen/strep), 10 mM HEPES, GlutaMAX (adDMEM/F12+++), 1× B27 (all Thermo Fisher), 1.25 mM N-acetyl-l-cysteine (Sigma-Aldrich), 10 mM nicotinamide (Sigma-Aldrich), 5 μM Y-27632 (Abmole), 5 nM Heregulin β-1 (Peprotech), 500 nM A83-01 (Tocris), 5 ng ml^–1 epidermal growth factor (Peprotech), 20 ng ml^–1 human fibroblast growth factor (FGF)-10 (Peprotech), 10% Noggin-conditioned medium²⁰, 10% Rspo1-conditioned medium⁴⁴ and 0.1 mg ml^–1 primocin (Thermo Fisher); and, in addition, with 1 μM SB202190 (Sigma-Aldrich) and 5 ng ml^–1 FGF-7 (Peprotech) for PDO propagation (type 1 culture medium⁴³), or with 20% Wnt3a-conditioned medium⁴⁴, 0.5 μg ml^–1 hydrocortisone (Sigma-Aldrich), 100 μM β-estradiol (Sigma-Aldrich) and 10 mM forskolin (Sigma-Aldrich) for normal organoid propagation (type 2 culture medium⁴³). Organoids from passages 5–30 after cell isolation were used for T cell coculture.

For T cell coculture, organoids were recovered from the BME by resuspension in TrypLE Express and collected in adDMEM/F12+++ (BC and head and neck cancer PDOs) or resuspended and collected in adDMEM/F12+++ (DMG PDOs). Organoid suspensions were filtered through a 70-μm strainer (Greiner) to remove large organoids and pelleted before coculture.

T cells engineered to express a γδ TCR (TEGs and LM1s)

TEG001 (T cells engineered to express a highly tumor-reactive Vγ9Vδ2 TCR)^6,45,46, LM1s (mock T cells engineered to express a mutant Vγ9/Vδ2 TCR with abrogated function)⁸ and TEG011 (mock T cells engineered to express HLA-A*24:02-restricted Vγ5/Vδ1 TCR, used as control for in vivo studies)^47,48 were produced as previously described⁸. Briefly, packaging cells (Phoenix-Ampho) were transfected with helper constructs gag-pol (pHIT60), env (pCOLT-GALV) and pMP71 retroviral vectors containing both Vγ9/Vδ2 TCR chains separated by a ribosomal-skipping T2A sequence, using FugeneHD reagent (Promega). Human peripheral blood mononuclear cells (PBMCs) from healthy donors were preactivated with anti-CD3 (30 ng ml^–1; Orthoclone OKT3, Janssen-Cilag) and IL-2 (50 IU ml^–1; Proleukin, Novartis) and subsequently transduced twice with viral supernatant within 48 h in the presence of 50 IU ml^–1 IL-2 and 6 mg ml^–1 polybrene (Sigma-Aldrich). TCR-transduced T cells were expanded by stimulation with anti-CD3/CD28 Dynabeads (500,000 beads 10^–6 cells; Life Technologies) and IL-2 (50 IU ml^–1). Thereafter, TCR-transduced T cells were depleted of nonengineered T cells by magnetic-activated cell sorting (MACS) as previously described⁸. This depletion protocol establishes a predominantly αβ TCR^– population (Extended Data Fig. 4a), which has been shown to result in complete loss of alloreactivity (Extended Data Fig. 1e)⁴⁵. To separate CD4⁺ and CD8⁺ TEGs and LM1s, we performed positive selection using either CD4 or CD8 Microbeads (Miltenyi Biotech) following the manufacturer’s instructions. After incubation with magnetic microbeads, cells were applied to LS columns and CD4⁺ or CD8⁺ TEGs or LM1s were selected by MACS. After the MACS selection procedure, Vγ9/Vδ2 TCR⁺ CD4⁺ or Vγ9/Vδ2 TCR⁺ CD8⁺ subsets of TEGs were stimulated every 2 weeks using a rapid expansion protocol⁸ where TEGs were cultured in ‘T cell culture medium’ (RPMI-GlutaMAX supplemented with 2.5–10% human serum (Sanquin), 1% pen/strep and 0.5 M beta-2-mercaptoethanol) on a feeder cell mixture comprising sublethally irradiated allogenic PBMCs, Daudi and LCL-TM in the presence of IL-2 (50 U ml^–1), IL-15 (5 ng ml^–1; both R&D Systems) and PHA-L (1 μg ml^–1; Sigma-Aldrich). To monitor the purity of CD4⁺ and CD8⁺ TEGs, as well as the absence of allogenic irradiated feeder PBMCs, cells were analyzed weekly by flow cytometry before functional assays using the antibodies anti-pan γδTCR-PE (Beckman Coulter), anti-αβTCR-FITC (eBioscience) anti-CD8-PerCP-Cy5.5 (Biolegend) and anti-CD4-APC (Biolegend). TEGs of purity <90% were reselected as described above. TEGs were used for coculture assays 4–5 days after the last IL2/IL15/PHA-L stimulation.

Live-cell imaging of T cells and organoid cocultures

Engineered T cells (20,000) were cocultured with normal organoids, PDOs or control cell lines (Daudi or HL-60) in an effector/tumor cell (E:T) ratio of 1:30 or 1:25 (only for Fig. 4d,e and Extended Data Fig. 5a). CD4⁺ and CD8⁺ TEGs were mixed in a 1:1 ratio immediately before plating. Cells were incubated in 96-well, glass-bottom SensoPlates (Greiner) in 200 µl of ‘coculture medium’: 50% type 1 organoid culture medium, 50% ‘TEG assay medium’ (RPMI-GlutaMAX supplemented with 10% fetal calf serum and 1% pen/strep), 2.5% BME and pamidronate for the accumulation of the phosphoantigen IPP to stimulate tumor cell recognition⁸ (1:2,000). Coculture medium was supplemented with both NucRed Dead 647 (two drops ml^–1; Thermo Fisher) and TO-PRO-3 (1:3,000; Thermo Fisher) for fluorescent labelling of living and dead cells (‘Imaging medium’). The combination of NucRed Dead 647 and TO-PRO-3 labels dead cells when excited with a 633-nm laser and living cells with a 561-nm laser (Extended Data Fig. 1a,b). Both were combined to achieve the optimal fluorescent intensity ratio between dead and living cells for live-cell imaging. Before coculture, TEGs were incubated with eBioscience Cell Proliferation Dye eFluor 450 (referred to as eFluor-450; 1:4,000; Thermo Fisher) in PBS for 10 min at 37 °C to fluorescently label all T cells. When CD4⁺ and CD8⁺TEGs were simultaneously imaged, both eFluor-450 and Calcein AM (1:4,000; Thermo Fisher) were used to label the different TEG subsets in PBS for 10 min at 37 °C. For NCAM1 prelabelling experiments, a combination of eFluor-450 (1:4,000; Thermo Fisher) and Hilyte-488-conjugated NCAM1 nanobodies (1:400; QVQ) was used to label CD8⁺ TEGs in PBS for 20 min at 37 °C before coculture. The plate was placed in a LSM880 (Zeiss Zen Black Edition v.2.3) microscope containing an incubation chamber (37 °C, 5% CO₂) and incubated for 30 min to ensure settling of TEGs and organoids at the bottom of the well. The plate was imaged for up to 24 h with a Plan-Apochromat ×20/0.8 numerical aperture dry objective with the following settings: online fingerprinting mode, bidirectional scanning, optimal Z-stack step size, Z-stack of 60 μm in total and time series with either a 30-min interval (up to 60 conditions simultaneously; resolution 512 × 512) or a 2-min interval (up to four or ten conditions simultaneously; resolution 512 × 512 and 200 × 200, respectively). To minimize photobleaching of NCAM1-prelabelled TEGs, the 488-nm laser was activated during only one Z-stack each hour within the first few hours of imaging. Directly after imaging, production of IFN-γ in the supernatant was quantitated using an ELISA-ready-go! Kit (eBioscience) and cell pellets were used to measure organoid viability with the CellTiter-Glo Luminescent Cell Viability Assay (Promega).

IFN-β stimulations

PDOs were harvested as described above and incubated in 96-well, round-bottom culture plates (Thermo Fisher) in 100 µl of type 1 organoid culture medium, supplemented with 2.5% BME and with or without the presence of 100 pg ml^–1 recombinant human IFN-β (Peprotech). After 24 h of incubation (37 °C, 5% CO₂), TEGs or LM1s were added to either IFN-β-preincubated or unstimulated organoids (E:T ratio 1:30) in 100 µl of TEG assay medium, supplemented with 2.5% BME and pamidronate (1:1,000) and with or without the presence of 100 pg ml^–1 recombinant human IFN-β (Peprotech). Medium without T cells was added for ‘organoid only’ controls. After 16 h of incubation (37 °C, 5% CO₂), plates were used to measure organoid viability using the CellTiter-Glo Luminescent Cell Viability Assay.

In vivo targeting by TEGs

Adult female NSG mice (15–16 weeks old) received sublethal total body irradiation (1.75 Gy) and subcutaneous implantation of a β-estradiol pellet (Innovative Research of America) on day –1. On day 0, PDOs (1 × 10⁶ 13T or 0.5 × 10⁶ 169M organoid cells in 100 μl of BME per mouse) were prepared as described previously⁴³ for subcutaneous injection in the right flank on day 0, and mice received two injections of 10⁷ TEGs or TEG011 mock cells on days 1 and 6 in pamidronate (10 mg kg^–1 body weight) as previously reported⁷. On day 1, together with the first T cell injection, all mice also received 0.6 × 10⁶ IU of IL-2 (Proleukin, Novartis) in incomplete Freund’s adjuvant (IFA; MD Bioproducts) subcutaneously. Tumor volume was measured once per week using a digital caliper and calculated by the following formula: 0.4 × (length x width²). Mice were monitored at least twice per week for weight loss and clinical appearance scoring (scoring parameters included hunched appearance, activity, fur texture, piloerection and respiratory/breathing problem). Humane endpoint was reached either when mice experienced 20% weight loss from initial weight, tumor volume reached 2 cm³ or when a clinical appearance score of 2 was reached for an individual parameter or an overall score of 4. In no case was the tumor burden exceeded.

Image processing

For 3D visualization, cell segmentation, extraction of statistics and time-lapse videos were processed with Imaris (Oxford Instruments) v.9.2–9.5. The Channel Arithmetics Xtension was used to create new channels for specific identification of organoids (live and dead) and eFluor-450-labelled or calcein AM-labelled T cells (live and dead) and to exclude cell debris. The Surface and ImarisTrack modules were used for object detection and automated tracking of both T cells (autoregressive motion) and organoids (‘connected components’ or no tracking). The Distance Transformation Xtension was used to measure the distance between TEGs and organoids, with thresholds for defining organoid–T cell interactions visually determined. For tracked TEGs, time-lapse data containing the coordinates of each cell, the values of cell speed, mean square displacement, distance to organoids and dead cell dye channel intensity were exported. For experiments with NCAM1 prelabelling, the mean intensities of the NCAM1 channel per T cell were exported. For tracked organoids, time-lapse data containing the coordinates of each organoid, the surface area, volume and mean dead cell dye channel intensity were exported.

PDO killing dynamics

To quantify the cell death dynamics of PDO cultures, >5,000 single organoids were analyzed at each time point (48 in total). The mean dead cell dye intensity within single organoid surfaces was quantified and rescaled to a range between 0 and 100 per experiment to normalize for variation in absolute dead cell dye intensity. To analyze whether organoid sensitivity to TEGs was dependent on initial organoid size, we compared the initial area (0 h) of organoids killed by TEGs at 10 h compared with the area of TEGs remaining alive at 10 h.

T cell dynamics analysis and multivariate time series clustering

For the analysis of TEG behavior over time, the following parameters were used: T cell death, contact with organoids, speed, square displacement and interaction with other T cells. For each T cell time series, linear interpolation was used to estimate the values in several cases of missing time points. To compare time series independently of their length, cell tracks were cut to a length of 3.3 h. Similarity between distinct cell tracks was measured using a strategy that allows for best alignment between time series, previously applied for mitotic kinetics⁴⁹ or temporal module dynamics comparisons⁵⁰. A cross-distance matrix based on multivariate time series data was computed using the dynamic time-warping algorithm. To visualize distinct cell behaviors in two dimensions, dimensionality reduction on the multidimensional feature count table was performed by the UMAP method^51,52. Clustering was performed using the k-means clustering algorithm with outlier detection. To confirm the identity of each cluster, T cell cluster assignments were back-projected to visualize the surfaces and tracks of particular T cell populations in the imaging dataset (Fig. 2a and Extended Data Figs. 3a,b and 4b).

Cell behavior classification using a random forest classifier

For standardized integration of new experiments, we used a random forest classification approach⁵³ to relate cell behavior to the nine behavioral signatures that we found in our global TEG behavior atlas (Fig. 2b). To allow for inclusion of experiments with a low E:T ratio of 1:25, where the parameter of T cell interaction would be influenced as compared with the standard E:T ratio of 1:30, the following parameters were used: T cell death, organoid contact, speed and square displacement. The reference dataset used to build the global TEG behavior atlas was split into cell tracks for use as either a training dataset (95%) or a test dataset (5%). To reduce dimensionality, for each cell track four time series descriptive statistics were quantified and used to train the classifier. For numeric variables, the following measures were computed for each cell track: mean, median, the top 90% of the distribution and standard deviation. For binary values, such as contact with organoids, the mean was calculated as well as the mean and maximum of cumulative interaction. The random forest classifier was trained using 100 trees on the above-mentioned variables using the nine behavioral signatures as labels (Extended Data Fig. 3c,d). The test dataset was used to assess accuracy of the classifier and to determine in which behavioral signatures the errors occurred (Extended Data Fig. 3e). A slightly updated version of the classifier was used in Fig. 3.

Correlation between TEG behavior and organoid killing dynamics

To estimate the correlation between onset of death in individual organoids and engagement with T cells belonging to the engaging clusters (CL7–9), we implemented a technique of sliding window correlation analysis previously used for functional brain connectivity⁵⁴ and genome analysis⁵⁵. We calculated the Pearson correlation coefficient between the cumulative number of organoid contacts with TEGs from each cluster and the increase in dead cell dye intensity in each over a sliding window of 3 h (Fig. 2f and Extended Data Fig. 3k).

NCAM1 prelabelling quantification using 3D imaging data

Behavioral classification of NCAM1-prelabelled TEGs was performed as described above, by prediction of behavioral signatures with the random forest classifier. NCAM1^+/– TEGs were identified based on an NCAM1 intensity threshold in individual TEGs, visually defined at the time points where the 488-nm laser was turned on. To ensure inclusion of true NCAM1^– or NCAM1⁺ TEGs, two intensity thresholds were defined.

Pseudotime trajectory inference

Two experimental SORT–seq replicates of TEGs cocultured with 13T PDOs, generated as described above, were used for trajectory interference (Extended Data Fig. 6b). Proliferating T cells were excluded from the analysis because they did not show any dynamic inflammatory genes during analysis. Afterwards, the gene expression table was log normalized with a 10,000 scaling factor. Shared nearest-neighbor, graph-based clustering was done as described above at a resolution of 2. Based on marker gene expression of CD8, CD4 and IL17RB⁵⁶, TEGs were subclustered into three subtypes: IL17RB⁻CD8^+eff, IL17RB⁻CD4^+eff and IL17RB⁺CD4^+mem. Downstream analyses were performed on each subset separately and compared with each other where mentioned.The RunFastMNN function from the SeuratWrappers package was utilized to correct for batch effects between the two SORT–seq replicates. We used the package Monocle3 (ref.⁵⁷) to infer the pseudotime trajectory and significantly dynamic genes for each T cell subtype. For each cell subtype, either no-target control or nonengaged^Enriched TEGs were designated as the root of the trajectory. To acquire comparable results from both Seurat and Monocle3 packages, the FastMNN batch-corrected UMAP coordinates were imported and used throughout the trajectory analysis in Monocle3. In IL17RB⁻CD4^+eff and IL17RB⁺CD4^+mem subtypes, Monocle identified no-target control cells as a separate partition. To have all cells along with a single pseudotime spectrum, we added maximum pseudotime values of no-target control T cells to pseudotime values of remaining cells in that subtype. For all TEG subtypes, significant dynamic genes along with the pseudotime trajectory were calculated and identified using Monocle3’s graph_test function, with 1 × 10^–20 q-value as the significance cutoff. Afterwards, using both k-means clustering and visual inspection of gene behavior over the pseudotime, TEGs were clustered into subclusters of similar pattern (CL1–8; Fig. 5g). The expression profile of the genes, along with the pseudotime trajectory, was plotted using the package pheatmap⁵⁸ using row-scaled (z-score) expression values. Smoothed gene behavior was calculated and visualized recruiting the gam smoothing function in the ggplot2 package⁵⁹.

Behavior signature inference over pseudotime

To align pseudotime inference with the different behavioral signatures that we identified with BEHAV3D, we built a probability map distribution for different behavioral signatures over the pseudotime based on the fundamental principle of transitivity of probabilistic distribution (Fig. 5f). We defined three states of cells quantified by different methods:

• Behavioral_signatures (B_sig): (Static, Lazy, Medium scanner, Scanner, Super scanner, Tickler, Engager, Super engager). Behavioral signatures of cells identified by imaging (Fig. 5b).

• Experimental_engagement_state (Exp_eng): (No-target control, Nonengaged, Nonengaged^enriched, Engaged, Super engaged). Cell distribution among different experimental conditions (Fig. 5a).

• UMAP_cluster (U_cl): (1…X). Cell assignment to distinct clusters grouping cells of similar gene expression. Shared nearest-neighbor, graph-based clustering was repeated several times using the Seurat package FindNeighbors and FindClusters functions with resolution in the range 1–7.

From these three different cell states, the following information was quantified:

• p(B_sig|Exp_eng): for each Experimental_engagement_state we quantified the probability distribution of each Behavioral_signature (Fig. 5f). This was achieved by reproducing the Experimental_engagement_states in silico on our imaging data. These values were calculated separately for CD4⁺ and CD8⁺ TEGs.

• p(Exp_eng|U_cl): for each UMAP_cluster, we quantified the probability of each Experimental_engagement_state belonging to this cluster.

Given these probabilities, we then quantified for each T cell the probability distribution of each unique Behavioral_signature in each UMAP_cluster using the equation:

As a result, each cell was assigned a certain probability distribution for different behavioral signatures. To refine the probability map, the same process was repeated for seven runs with different cluster sizes and final probability distributions were averaged per cell. Note that, for cells belonging to the No-target control Experimental_engagement_state, a Behavioral_signature called No-target control was assumed. Given that the nonengaged behavioral signatures (Static, Lazy, Slow scanner, Medium scanner, Super scanner) exhibited an identical probability map, their values were plotted together. For visualization purpose, extreme outlier values of skewed distributions were transformed to a maximal cutoff value. Based on the probability distribution of different behavioral signatures, pseudotime was divided into four stages—Baseline (no organoids), Environmental stimuli, Short engagement and Prolonged engagement—for each TEG subtype (CD8^+eff, CD4^+eff and CD4^+mem).

Serial killer gene signature analysis

Genes of CL7 (Fig. 5g and Supplementary Tables 4 and 5) were analyzed to identify a unique signature for killer TEGs. Sixty-one of 83 genes comprising this cluster were common to TEGs incubated with 13T and 10T organoids and underwent extensive literature curation to identify those with a known role in T cell cytotoxicity, T cell biology (not related to cytotoxicity), morphological plasticity or other processes such as GTPase signaling, ribogenesis and transcriptional regulation.

Cytotoxic in vivo T cell signature definition and projection on TEGs

To generate a signature gene set for cytotoxic CD8⁺ T cells in samples from patients with BC, we downloaded two publicly available datasets from GEO (accession nos. GSE114724 (ref.⁴⁰) and GSE110686 (ref. ⁴¹)). Raw data were downloaded and analyzed with Seurat, using the same procedure utilized for TEG data processing. Clusters were identified and named using the marker genes defined in the study of Savas et al.⁴¹. From the study of Azizi et al.⁴⁰, only TILs were used for analysis. Clusters were generated with a resolution of 0.9. For the Azizi and Savas studies, two marker gene lists were identified for cytotoxic CD8⁺ T cells (based on the 2,000 variable features and an average log(fold change) cut off of 0.3; Supplementary Table 6). The overall enrichment of the identified gene sets for each study was calculated using VISION⁶⁰ and visualized on top of UMAP cell embeddings for each study. In addition, the overall enrichment of in vivo identified gene sets was projected on the UMAP of TEGs.

For the following methods we refer to Supplementary Protocols: primary DMG patient-derived lines and head and neck cancer PDO cultures, cell lines, WT1 T cells, ROR1 CAR T cells, flow cytometry analysis of NCAM1 and ROR1 expression, sorting of NCAM1^–/+ TEGs, T cell serial killing capacity analysis, PDO bulk RNA-seq, SORT–seq sample preparation, SORT–seq library preparation and sequencing, mapping and quantification of SORT–seq data, SORT–seq and 10X Genomics data integration and TEG subpopulation analysis, differential gene expression analysis of TEGs cocultured with distinct PDO cultures and gene set enrichment analysis.

Statistics and reproducibility

Statistical analysis was performed using either R or Prism v.7 software (GraphPad), and results are represented as mean ± s.e.m. unless indicated otherwise; n represents independent biological replicates. Two-tailed unpaired t-tests were performed between two groups unless indicated otherwise. Pearson correlation was used for paired comparison among three different readouts (IFN-γ production, cell viability and live imaging). For live-cell imaging, the increase in dead cell dye between the first and last time points was used as a measure. To compare tumor volumes in mice treated with TEGs or TEG001 mock cells, two-way analysis of variance (ANOVA) with repeated measures was performed. To compare frequencies of different behavioral signatures among PDOs, a Pearson’s chi-squared test was applied. To compare the percentage of dead organoids when TEGs were cocultured with different PDOs, one-way ANOVA followed by Bonferroni correction was performed. To estimate the change in correlation between PDO death dynamics and cumulative contact with TEGs for different behavioral signatures, data were fitted to a linear mixed model with experimental replicate as the random effect to account for variation between them. For cell type enrichment analysis of TEG first and second action after engagement, a hypergeometric test was used (Fisher’s exact test). For comparisons of percentages of distinct TEG subtypes in the same well (CD4⁺ versus CD8⁺ or NCAM⁺ versus NCAM), for each behavioral signature data were fitted to a linear regression model with each individual replicate set as the random effect to account for variation between them. For comparisons of percentages between different T cell lines (different wells), the standard deviation of the difference between mean cluster percentages for pairs of T cell lines was calculated by taking the square root of the sum of the variances of both separate distributions (Fig. 3j). For each fitted model, ANOVA was computed with an F-test. For comparison of IFN-β treatment, paired t-tests were performed. To ensure global TEG behavior atlas (Fig. 2a,b) reproducibility, we pooled 22 different imaging datasets comprising TEGs and LM1 cells cocultured with 13T or 100T organoids. Supplementary Table 8 summarizes the value of n per condition for Figs. 2b, 3f–j and 6e–g and includes statistical test details from Fig. 2f.

Reporting summary

Further information on research design is available in the Nature Research Reporting Summary linked to this article.