ZFP36L2 orchestrates stress-adaptive plasticity in regeneration and cancer

Patient biospecimens and data

All patient material was obtained through Memorial Sloan Kettering (MSK) Institutional Review Board protocols 06-107, 12-245, 14-244 and 22-404. No statistical method was used to predetermine sample sizes. scRNA-seq datasets from matched normal colon, primary tumour and metastasis samples were obtained from a previous study¹⁰. Patient-derived organoids from two primary tumours (MSK125P and OKG146P) and two metastases (MSK125Li and OKG146Li) were generated and validated as previously described^7,10. Tumour whole-exome sequencing (WES) was performed by recapturing DNA processed originally for targeted exon sequencing using MSK-IMPACT⁶⁷. The median WES target coverage for tumour and normal samples was 129× and 101×, respectively. WES data were analysed using the TEMPO pipeline (available from GitHub: https://github.com/mskcc/tempo). The OncoKB precision oncology knowledgebase, a US Food and Drug Administration (FDA)-recognized human genetic variant database curated by experts at MSK⁶⁸, was used to distinguish between oncogenic alterations (presumed drivers) and variants of unknown significance (presumed passengers). Only somatic alterations labelled as oncogenic, likely oncogenic or predicted oncogenic by OncoKB were included for analyses. Archival formalin-fixed, paraffin-embedded (FFPE) ZFP36L2 WT and ZFP36L2 mutant clinical tissue blocks for immunostaining were identified through WES of corresponding tumour DNA originally collected for MSK-IMPACT. Tissue processing, FFPE section selection and histopathological data interpretation were overseen by an expert gastrointestinal pathologist (J.S.).

scRNA-seq data analysis of CRC samples from patients

Normalized gene expression matrices were obtained from the Human Tumour Atlas Network (HTAN) Data Portal (http://humantumouratlas.org/publications/hta8_crc_moorman_2024) and processed as previously described¹⁰. Downstream analyses were restricted to patient-matched pairs of primary tumour and liver metastasis samples (n = 25 patients). After sample filtering, a total of 38,272 cells remained: primary tumour (13,102 cells), metastatic (12,215 cells) ISC (1,145 cells), absorptive precursor (2,805 cells), enterocyte (1,214 cells), BEST4⁺ enterocyte (1,339 cells), secretory precursor (3,818 cells), goblet (1,720 cells), tuft (751 cells) and enteroendocrine (163 cells). All subsequent differential expression, pathway analyses and correlation-based analyses were restricted to this filtered epithelial compartment.

Data visualization

Two-dimensional embeddings were generated using Scanpy (v.1.9.1). A k-nearest neighbours (KNN) graph was constructed on the principal components using Euclidean distance (k = 20), followed by visualization with a force-directed layout (ForceAtlas2). Subsequent plots were produced using Matplotlib (v.3.6.0)

Gene signature scores

Gene signature scores were calculated using the score_genes function from Scanpy (v.1.9.1), which calculates the average expression of a given gene set relative to matched reference genes. To reduce bias from differences in gene expression levels in each signature, we used z-scored expression data as input.

Generation of ISC-like cell annotations and gene signature

To identify tumour cells exhibiting an ISC-like transcriptional program, we first filtered the dataset to retain only malignant epithelial cells. Raw counts from these primary tumour and metastatic cells were normalized using total-count normalization followed by log-transformation. Highly variable genes (HVGs; n = 3,000) were selected for principal component analysis (PCA), neighbourhood graph construction (n_neighbours=20) and Leiden clustering.

To assess stemness expression across cells, we computed the median expression per cluster of ISC-specific marker genes from Hotspot-derived tumour ISC-like gene module 29 (Supplementary Table 1a), as previously identified¹⁰. The distribution of cluster-level ISC module scores was bimodal, which informed the selection of a threshold in identifying clusters that were ISC-like or not ISC-like. This tumour-derived ISC-like threshold value was then kept constant in labelling ISC-like normal cell types. We calculated a gene signature score on the z-normalized expression of these genes using the score_genes function in Scanpy (v.1.9.1) (Fig. 1b and Extended Data Fig. 1a).

DEG analysis between normal and tumour ISC-like cells

Differential gene expression (DEG) analysis was performed using the Wilcoxon rank-sum test as implemented in scanpy.tl_rank_genes_groups, comparing normal and tumour cells stratified by their ISC-like annotations. The analysis was conducted on the log-normalized gene expression matrix. We filtered for previously defined¹⁰ tumour ISC-like Hotspot gene module 29 genes and selected the most significant ones for downstream visualization (Fig. 1b). The log₂ fold changes were computed relative to each gene across the specified cell-type groups.

To further characterize activity of the tumour ISC-like module, we computed the mean imputed expression of each gene in ISC module 29 across canonical ISC cells. Genes were then ranked by mean imputed expression to highlight ISC-enriched transcripts.

Computation of gene autocorrelation signatures in normal and tumour cells

We used Hotspot⁶⁹ (v.0.9.1) to calculate pairwise local correlations between genes on the basis of the KNN graph, which captures local cell–cell similarity and genes with high local autocorrelation. This local strategy is also inherently more robust to gene dropouts and technical noise, which enables sensitive detection of context-specific gene co-expression patterns.

We first partitioned the data to only include tumour cells (25,317 cells) or normal cells (12,955 cells). HVGs (n = 2,000) were selected on the untransformed expression layer, with the manual inclusion of ZFP36L2. Genes prone to technical or non-informative variation (for example, ribosomal, mitochondrial and long non-coding RNAs) were excluded from the analyses, as were genes with non-zero counts across the dataset. A 50-nearest neighbour graph was constructed using the precomputed PCA embedding, and raw UMI counts were re-added to a desiccated counts layer. The resulting dataset was passed to Hotspot, using a depth-adjusted negative binomial model with the neighbourhood structure defined by a 15-nearest neighbourhood graph. Spatial autocorrelation scores (Moran’s I) were computed for all genes. Genes with a FDR < 0.05 were considered significant and retained for downstream analyses (Fig. 1i and Extended Data Fig. 1g).

GSEA

GSEA using GO biomolecular process, all Hallmark, PID and KEGG gene sets and Hotspot modules have been previously defined¹⁰ (Supplementary Table 1b). GSEA was performed using the prerank function of the Python package GSEApy (v.0.14.0) with 10,000 permutations and the default parameters.

Statistical analyses

Pairwise comparisons of PID–AP-1 pathway⁶⁵ scores (Fig. 1j), Hotspot tumour ISC-like module 29 scores¹⁰ and ZFP36L2 expression (Extended Data Fig. 1) were calculated across all cell-type groups by Mann–Whitney U-tests, and P values were adjusted for multiple testing using Bonferroni correction (q < 0.05).

Mouse colon epithelial cell scRNA-seq data analysis

Data acquisition and preprocessing

scRNA-seq data (GSE168448)⁶⁶ were downloaded from the Gene Expression Omnibus (GEO). Cells were first filtered on the basis of mitochondrial content, excluding those with mitochondrial transcript percentages exceeding 0.2. Outliers were identified using the median absolute deviation (MAD) method, following criteria adapted from previous studies^70,71. A cell was classified as an outlier if it met at least two of the following five conditions: (1) log-transformed library size below 3 MADs under the median or above 5 MADs over the median; (2) log-transformed number of expressed genes below 3 MADs under the median or above 5 MADs over the median; (3) percentage of counts from the top 20 most highly expressed genes exceeding 5 MADs below or above the median; (4) feature-count distance (as previously defined⁷⁰) exceeding 5 MADs below or above the median; and (5) percentage of mitochondrial counts exceeding 2.5 MADs above the median and greater than 8%. Same-sample doublets were identified and removed using scDblFinder (v.1.16.0)⁷². Outliers and doublets were excluded from downstream analyses.

Normalization, HVG selection, PCA and denoising

Count data were median-normalized and log-transformed with a pseudo-count of 1. The top 2,000 HVGs were selected using Scanpy⁷³ (v.1.9.8) with the Seurat (v.3) method⁷⁴. PCA was performed on these HVGs, retaining 280 principal components that accounted for 75% of the variance. Denoising was conducted using MAGIC^74,75,76 via the Scanpy external API, using default parameters except for the number of principal components, for which all 280 retained components were used for neighbourhood calculations.

Cell typing

Cell-type annotation was performed by scoring the denoised expression of cell-type marker gene sets using Scanpy’s score_genes function, which reimplements a previously described method⁷⁷. Each cell was assigned to the cell type for which it exhibited the highest score.

Mouse experimental methods

Mouse strains

All animal procedures were approved by the Institutional Animal Care and Use Committee of the Memorial Sloan Kettering Cancer Center (MSKCC). The following mouse strains were obtained from The Jackson Laboratory: C57BL/6J (strain 000664); Vil1^cre mice³⁹ (B6.Cg-Tg(Vil1-cre)1000Gum/J; stock 021504); Lgr5^{eGFP-IRES-creERT2} mice¹⁶ (B6.129P2-Lgr5^{tm1(cre/ERT2)Cle}/J; stock 008875); and NSG mice (NOD.Cg-Prkdc^scidIl2rg^tm1Wjl/SzJ; stock 005557). Zfp36l2^fl/fl mice⁴⁰ were provided by M. Turner and Lgr5^DTR-eGFP mice⁵ by F. de Sauvage. Primers and relevant information used for genotyping are listed in Supplementary Table 4b.

Xenograft experiments used only female NSG mice. All other experiments used male and female mice in equal proportions, with littermates allocated across groups to minimize variability. Sample sizes were not predetermined by statistical methods; group sizes (n = 5–15 per condition) were based on previous experience with similar xenograft and colitis models, keeping animal numbers to the minimum required under the 3Rs principles. Mice were randomly assigned to treatment groups, with blinding applied wherever feasible.

DSS-induced colitis

Experimental cohorts were generated by inbreeding on a C57BL/6J background, and all experiments included littermate controls. To evaluate the role of ZFP36L2 in colonic regeneration, Zfp36l2^fl/fl mice were crossed with Vil1^cre mice to generate intestinal epithelial knockout mice (Zfp36l2^IEC). Mice (6–8 weeks old) received 3.5% (w/v) DSS (molecular weight 36–50 kDa; MP Biomedicals) in drinking water ad libitum for 7 days, followed by 7 days of regular water before euthanasia. Body weight was recorded daily. Caecal and colon tissues were collected at specified time points, processed as Swiss rolls, fixed in 4% paraformaldehyde and embedded in paraffin. Tissue sections were subjected to haematoxylin and eosin staining, immunostaining or FISH (see below).

Quantification of LGR5–eGFP by crypt flow cytometry

Lgr5^{eGFP-IRES-creERT2} mice (Jackson Laboratory, stock 008875) were crossed with Vil1^cre (stock 004586) and Zfp36l2^fl/fl mice to generate Lgr5^{eGFP-IRES-creERT2+};Vil1^cre+;Zfp36l2^IEC mice. Littermate controls included Lgr5^{eGFP-IRES-creERT2+};Vil1^cre+;Zfp36l2^WT and Lgr5^{eGFP-IRES-creERT2–};Vil1^cre+;Zfp36l2^WT mice. For each experimental batch, three littermate pairs of Lgr5^{eGFP-IRES-creERT2+};Vil1^cre+;Zfp36l2^WT and Lgr5^{eGFP-IRES-creERT2+};Vil1^cre+;Zfp36l2^IEC mice (aged 6–8 weeks) were collected. Colons were dissected, flushed with PBS to remove faecal content, opened longitudinally and cut into around 1-cm segments. Tissue samples were incubated in dissociation buffer (PBS with 8 mM EDTA, 0.5 mM DTT and 10 U ml^–1 DNase I (Roche 04716728001)) at 4 °C with gentle shaking for 60 min. Crypts were detached by vigorous shaking, pelleted and further dissociated with TrypLE (Thermo Fisher Scientific). The resulting cell suspension was treated with DNase at room temperature for 5 min and passed through a 40 μm strainer to generate single-cell suspensions for flow cytometry. Viable epithelial single cells were gated by forward scatter, side scatter and negative staining for 4′,6-diamidino-2-phenylindole (DAPI). Lgr5^{eGFP-IRES-creERT2–};Vil1^cre+;Zfp36l2^WT mice were processed in parallel as gating controls. Two independent experimental batches were performed, and data were pooled for analyses.

In vivo Lgr5
^DTR ablation

Mice were bred to generate Lgr5^DTR-eGFP+ and littermate Lgr5^DTR-eGFP− controls. Lgr5^DTR-eGFP+ZFP36L2^IEC mice were generated by crossing Lgr5^DTR-eGFP+;Vil1^cre+ and ZFP36L2^fl/fl together with Lgr5^DTR-eGFP+;Vil1^cre− and ZFP36L2^fl/fl to generate littermate controls. For in vivo depletion of Lgr5^DTR-expressing cells, DT (0.005 μg μl^–1 in PBS) was administered intraperitoneally at a dose of 0.05 μg per g of body weight every 48 h, for a total of 4 doses. Mice aged 8–10 weeks were used for all experiments. Colons were collected at the time points indicated in the figures or figure legends, Swiss-rolled and fixed in 4% paraformaldehyde for 24 h at room temperature. Tissues were then transferred to 70% ethanol and stored at 4 °C before paraffin embedding.

In vitro Lgr5
^DTR ablation and dedifferentiation assay

Colon organoids were established from Lgr5^DTR-eGFP+;Vil1^cre+;ZFP36L2^fl/fl (Lgr5^DTR-eGFP+ZFP36L2^IEC) and Lgr5^DTR-eGFP+;Vil1^cre−;ZFP36L2^fl/fl (Lgr5^DTR-eGFP+ZFP36L2^WT) mice (6–8 weeks old; n = 3 mice per genotype). Approximately 1,000 crypts, isolated as described above, were suspended in 40 μl Matrigel and cultured in mouse WRENAFI (mWRENAFI) medium containing advanced DMEM/F12 (AdDF12; Thermo Fisher Scientific), 2 mM GlutaMAX (Thermo Fisher Scientific), 10 mM HEPES (Thermo Fisher Scientific), 1 mM N-acetyl-l-cysteine (Sigma-Aldrich), 1:50 B27 supplement (Thermo Fisher Scientific), 1:100 N2 supplement (Thermo Fisher Scientific) and 100 μg ml^–1 Pimocin (InvivoGen), supplemented with growth factors and the following inhibitors: 50 ng ml^–1 EGF (Peprotech), 100 ng ml^–1 murine Noggin (Peprotech), 500 nM A8301 (Sigma-Aldrich), 50 ng ml^–1 FGF2 (Peprotech), 100 ng ml^–1 IGF-I (Peprotech), 1 nM NGS-WNT (ImmunePrecise N0001) and 1 μg ml^–1 murine R-spondin1 (Peprotech). After stable organoid lines were established, organoids were dissociated into single cells and seeded at 20,000 cells per 40 μl Matrigel. Cells were cultured in mWRENAFI medium for 4 days, with medium changed every 2 days. On day 4, cultures were switched to fresh mWRENAFI medium with or without 0.6 μg ml^–1 DT for 24 h. Organoids were then dissociated into single cells, and depletion of LGR5–eGFP⁺ cells in the DT-treated group was confirmed by flow cytometry using organoids derived from C57BL/6J mice as a gating control. Live DAPI-negative cells were sorted, embedded in Matrigel at 2,000 cells per 40 μl and cultured in mWRENAFI medium for 5 days to induce dedifferentiation. Regenerated organoid numbers were quantified using BioTek bright-field imaging with whole-dome z-projection, and eGFP⁺ cells from each group were quantified by flow cytometry.

Orthotopic xenograft experiments

NSG mice (Jackson Laboratory, stock 005557) were used for transplantation experiments at 6 weeks of age. Mice were housed in a specific pathogen-free facility under controlled temperature and humidity, with a 12-h light–dark cycle and ad libitum access to either standard chow or an irradiated diet supplemented with 2,500 ppm doxycycline (Modified LabDiet 5053), along with water.

For orthotopic caecal and intrasplenic injections, organoid lines expressing thymidine kinase–eGFP–luciferase and transduced with pTRIPZ lentivirus expressing doxycycline-inducible shRNAs targeting ZFP36L2 or a Renilla control were transduced with lentivirally expressed pLenti-PGK-Akaluc. Transplantation procedures were performed as previously described^7,10. For generation of the OKG146Li-MS2 line, OKG146Li organoids were transduced with pLenti-PGK-Akaluc, and around 500,000 cells were injected orthotopically into the mouse liver in 50% Matrigel as previously described¹⁰ to generate liver tumours. At the end point, tumours were collected, minced and plated in HISC medium containing antibiotics to select for CRC organoids. This in vivo selection procedure was repeated twice to generate the liver-metastasis-selected OKG146Li-MS2 organoid line. For caecal injections, 5 × 10⁵ cells from canonical primary tumour organoids (MSK125P and OKG146P) or liver metastasis (OKG146Li and MSK107Li) organoids, representing canonical and non-canonical cell states, respectively, transduced with lentivirus expressing eGFP–luciferase and either shZFP36L2 or shCtrl were resuspended in 10 μl (50% Matrigel, 50% PBS) and injected into the caecal submucosa of 6-week-old NSG mice. Mice were monitored by BLI 1 week after injection to confirm primary tumour engraftment. Any BLI-signal-negative animals were excluded from subsequent analyses. Mice were then started on a doxycycline diet with weekly BLI monitoring for 15 weeks before euthanasia and organ collection. For splenic vein injections, 6-week-old NSG mice were placed on a doxycycline diet 24 h before surgery. CRC liver-metastasis-derived organoids MSK107Li and OKG146Li-MS2 transduced with lentivirus expressing eGFP-luciferase and either shZFP36L or shCtrl were treated with doxycycline ex vivo for 7 days before surgery and injected into the spleens of mice (5 × 10⁵ cells per mouse). Around 500,000 cells were injected into the splenic vein in 50 μl PBS supplemented with 10% Matrigel, immediately followed by splenectomy. BLI was performed immediately after surgery to establish a post-injection baseline (week 0). Mice were maintained on doxycycline diet for 10–15 weeks until they reached the humane end point.

BLI was performed immediately after surgery and weekly thereafter using an IVIS Spectrum Xenogen system (Caliper Life Sciences). Data were analysed using Living Image (v.2.50) software. Experimental group sizes (n ≥ 5 mice per group) were determined based on practical constraints (five mice per cage) and were age-matched and sex-matched.

Mice with caecal or intrasplenic tumours were monitored daily. Humane end points were defined as the onset of distress, including hunched posture, impaired grooming or weight loss of >20%, together with a route-specific criterion: a maximum experimental duration of 6 months for caecal tumours, or tumour burden exceeding 10% of body mass and/or a body condition score <2 for intrasplenic tumours (with ascites as an additional distress sign specific to intrasplenic tumours). These criteria were not exceeded in any experiment. Animals were euthanized at the end point, and tissues were collected for further analyses. Where indicated, tissues were fixed in 4% paraformaldehyde for 24 h. Histopathological scoring was performed in a blinded manner.

Mouse organoid dedifferentiation assay

Three mice from each genotype (aged 6–8 weeks) were euthanized, and colons were dissected and processed to collect crypts as described above. Approximately 1,000 crypts were suspended in 40 μl Matrigel and cultured in mWRENAFI medium, consisting of advanced DMEM/F12 (AdDF12; Thermo Fisher Scientific), 2 mM GlutaMAX (Thermo Fisher Scientific), 10 mM HEPES (Thermo Fisher Scientific), 1 mM N-acetyl-l-cysteine (Sigma-Aldrich), 1:50 B27 supplement (Thermo Fisher Scientific), 1:100 N2 supplement (Thermo Fisher Scientific) and 100 μg ml^–1 Primocin (InvivoGen), supplemented with growth factors and the following inhibitors: 50 ng ml^–1 EGF (Peprotech), 100 ng ml^–1 murine Noggin (Peprotech), 500 nM A8301 (Sigma-Aldrich), 50 ng ml^–1 FGF2 (Peprotech), 100 ng ml^–1 IGF-I (Peprotech), 1 nM NGS-WNT (ImmunePrecise N0001) and 1 μg ml^–1 murine R-spondin1 (Peprotech). Once stable organoid lines were established, 2,000 single cells from the organoids were plated in 40 μl Matrigel and cultured in mWRENAFI medium supplemented with 10 μM Y-27632 for 3 days. The medium was then switched to mENAFI (removing NGS-WNT and R-spondin1) to induce organoid differentiation for 2 days. Efficient loss of LGR5⁺ ISCs under these differentiation conditions was confirmed by flow cytometry analysis of GFP signal in LGR5–eGFP⁺ organoids cultured under the same conditions. The differentiated mature organoids were dissociated into single cells using TrypLE, and viable cells were sorted and plated at 2,000 cells per 40 μl Matrigel domes to evaluate dedifferentiation via organoid formation. Five to six technical replicates from three biological replicates were pooled for subsequent analysis and plotting.

Multiplexed staining and imaging

Organoid collagen embedding

Day 7 patient-derived organoids (PDOs) from patients with CRC were recovered from Matrigel as described above. Intact organoids were resuspended in neutralized collagen I (250 μl collagen I solution mixed with 28.4 μl of 10× PBS and 5.7 μl of 1 M NaOH) and deposited into a tissue capsule in a 6-well plate to form small domes. Once the collagen I droplets solidified in the incubator for 30 min, the domes were covered with tissue culture medium and cultured overnight.

The medium was aspirated the following day, and the tissue capsule was washed two to three times with PBS until the solution appeared clear, free of residual medium. The tissue capsule was then transferred to freshly prepared 4% paraformaldehyde for 24 h with gentle agitation at room temperature. After fixation, the capsule was washed with PBS and stored in 70% ethanol at 4 °C until paraffin embedding.

FISH

Paraffin-embedding and tissue sectioning were performed by the MSKCC Molecular Cytology Core. Paraffin-embedded tissue sections (5 μm) were stored at 4 °C until use. Sections were processed on a Leica Bond RX stainer, baked for 30 min at 60 °C, dewaxed using Bond Dewax solution (Leica, AR9222) and subjected to EDTA-based epitope retrieval (Leica, AR9640) for 15 min at 95 °C.

For single FISH, RNA probes (Advanced Cell Diagnostics, ready-to-use) were hybridized to the tissue sections for 2 h at 42 °C. Probe detection was performed using a RNAscope 2.5 LS Reagent kit–Brown (ACD, 322100) following the manufacturer’s instructions, with the modification of substituting DAB application with Fluorescent CF594/Tyramide (Biotium, 92174) for 20 min at room temperature. After washing in PBS, slides were incubated with 5 μg ml^–1 DAPI (Sigma-Aldrich) in PBS for 5 min, rinsed and mounted using Mowiol 4–88 (Calbiochem).

For double FISH, a mixture of two probes (hLGR5-C1 and hZFP-C2) was hybridized to 5-μm FFPE tissue sections at 42 °C for 2 h. The hybridized probes were detected using a RNAscope LS Multiplex Reagent kit (ACD, 322800) according to the manufacturer’s protocol. Fluorescence detection was achieved by incubating the slides with Alexa Fluor 488 Tyramide (Life Technologies, B40953) and CF594 Tyramide (Biotium, 92174) for 30 min at room temperature. After PBS washing and DAPI incubation (5 μg ml^–1 for 5 min), slides were mounted in Mowiol 4–88.

The following RNAscope 2.5 LS probes were used in this study: Hs-ZFP36L2 (581268), Mm-Zfp36l2 (518908), Mm-Lgr5 (312178), Hs-LGR5 (311028), Hs-GDF15 (600308), Hs-HILPDA (501958), Hs-GADD45A (477518), positive control probe_Hs-PPIB (313908) and negative control probe_dapB (312038).

Multiplex FISH with IF

After FISH staining, slides were scanned, and coverslips were carefully removed. The slides were then processed following the standard IF protocol (as detailed below).

Single IF staining

For single IF, 5-μm FFPE tissue sections were loaded onto a Leica Bond RX system and dewaxed as the first step. Epitope retrieval was performed using EDTA-based ER2 solution (Leica, AR9640) for 20 min at 100 °C. Primary antibodies against each target were incubated for 1 h at room temperature. Following primary antibody incubation, slides were incubated with Leica Bond Polymer anti-rabbit HRP (included in the Polymer Refine Detection kit, Leica, DS9800) for 8 min. For pan-cytokeratin staining, after primary antibody incubation, slides were incubated with Leica Bond Post-Primary reagent (rabbit anti-mouse linker, included in the Polymer Refine Detection kit, Leica, DS9800) for 8 min, followed by an additional 8 min of incubation with Leica Bond Polymer anti-rabbit HRP (included in the Polymer Refine Detection kit, Leica, DS9800). Signal detection was performed using either Alexa Fluor Tyramide signal amplification reagents (Life Technologies, B40953, B40958) or CF dye Tyramide conjugates (Biotium, 92172, 96053, 92174). After IF staining, slides were washed in PBS and incubated with 5 μg ml^–1 DAPI (Sigma Aldrich) in PBS for 5 min, rinsed in PBS and mounted using Mowiol 4–88 (Calbiochem). Slides were stored overnight at –20 °C before imaging.

Multiplex IF staining

For multiplex IF, antibody staining and detection were performed sequentially according to the single IF staining procedures. Secondary antibody detection was performed using Leica Bond Post-Primary rabbit anti-mouse linker (Leica, DS9800) for human samples stained with mouse-derived primary antibodies, or with rabbit anti-mouse secondary antibody linker (Abcam, ab133469) for mouse samples stained with mouse-derived primary antibodies, followed by Leica Bond Polymer anti-rabbit HRP (Leica, DS9800). After each round of IF staining, epitope retrieval was carried out to denature the primary and secondary antibodies before applying the next primary antibody. After the final staining round, slides were washed in PBS, incubated with DAPI, rinsed, and mounted in Mowiol 4–88 (Calbiochem). Slides were stored at –20 °C overnight before imaging.

The following primary antibodies were used for IF: ZFP36L2 (mouse, 200 µg ml^–1, Santa Cruz, sc-365908); MUC2 (rabbit, 0.0625 μg ml^–1, Abcam, ab97386); pan-cytokeratin (mouse, 0.125 μg ml^–1, Abcam, ab8068); KI67 (rabbit, 0.25 μg ml^–1, Abcam, ab15580); CK20 (rabbit, 1 μg ml^–1, Novus Biologicals, NBP3-03605); CK5 (rabbit, 0.016 μg ml^–1, Cell Signaling Technology, 81817); CHGB (rabbit, 0.1 μg ml^–1, Invitrogen, PA5-52605); DDX6 (rabbit, 0.517 mg ml^–1, Abcam, ab307418); G3BP1 (mouse, 1 mg ml^–1, Abcam, ab56574) and PABPC1 (rabbit, 1 mg ml^–1, Abcam, ab21060). The following secondary antibodies were used for IF: Leica Bond Post-Primary (rabbit anti-mouse linker) (Leica, DS9800) for staining human samples with mouse-derived antibodies, rabbit anti-mouse secondary antibody linker (Abcam, ab133469) for staining mouse samples with mouse-derived antibodies, Leica Bond Polymer anti-rabbit HRP (Leica, DS9800).

FISH and IF imaging and quantification

FISH or IF slides were scanned on a Pannoramic scanner (3DHistech) using a ×40/0.95NA objective.

Mouse crypt FISH and IF multiplex quantification

A total of 31 crypts were drawn and exported as TIFF files from these scans using Slide Viewer (3DHistech). These images were then analysed using ImageJ/FIJI (NIH). Thresholding and Watershedding were used to segment the cells using the DAPI channel. The Lgr5 channel was thresholded and the mask was used to create a distance map. Then, the segmented cells were overlaid onto the distance map and thresholded images to determine their positivity for each marker and their distance from Lgr5.

Tumour marker gene FISH and IF quantification

ROIs around tissues were drawn and exported as TIFF files from scans using Slide Viewer (3DHistech). Images were then analysed using ImageJ/FIJI (NIH). Thresholding and Watershedding were used to segment cells using the DAPI channel. Each marker was thresholded to measure the area and intensity of the fluorescent signal and to classify positivity for each marker per cell.

Patient-derived CRC organoid experimental methods

CRC organoid culture

Organoids were cultured and expanded in HISC medium comprising advanced DMEM/F12 (ADF12; Thermo Fisher Scientific), 2 mM GlutaMAX (Thermo Fisher Scientific), 10 mM HEPES (Thermo Fisher Scientific), 1 mM N-acetyl-l-cysteine (Sigma-Aldrich), 1:50 B27 supplement (Thermo Fisher Scientific), 1:100 N2 supplement (Thermo Fisher Scientific) and 100 μg ml^–1 Primocin (InvivoGen), supplemented with growth factors and inhibitors, including 50 ng ml^–1 EGF (Peprotech), 100 ng ml^–1 Noggin (Peprotech), 500 nM A8301 (Sigma-Aldrich), 50 ng ml^–1 FGF2 (Peprotech) and 100 ng ml^–1 IGF-I (Peprotech). Where indicated, organoids were cultured in IGFF containing all base components of the medium but lacking EGF, Noggin, A8301, FGF2 and IGF. Medium was supplemented with 2 μg ml^–1 doxycycline where indicated. Medium was refreshed every 2 days. Human organoid lines used in this study were verified by short tandem repeat analysis at the time of establishment and before each experiment. All organoid lines used in this study were routinely tested for mycoplasma contamination using a MycoAlert PLUS Detection kit (Lonza), and all tested negative.

For downstream applications, organoids were collected on day 7 by dissolving Matrigel with 2 mM EDTA in DPBS, followed by orbital agitation at 4 °C for 30–60 min. Organoids were then dissociated into single cells using TrypLE (Thermo Fisher Scientific) for 5–10 min at 37 °C. The reaction was quenched by adding 5 volumes of wash buffer (DMEM/F12 supplemented with 2 mM GlutaMAX, 10 mM HEPES, 100 U ml^–1 penicillin–streptomycin (Thermo Fisher Scientific) and 5% FBS (Sigma-Aldrich)), and the cell suspension was filtered through a 40 μm cell strainer. For passaging, cells were embedded at 375 cells per μl Matrigel. For organoid formation and IC₅₀ assays, cells were seeded at 50 cells per μl Matrigel. Matrigel domes were incubated at 37 °C for 30 min to solidify before adding culture medium. Lentiviral transduction was performed as previously described⁴², and antibiotic selection was initiated after the first passage post-transduction using 2 μg ml^–1 puromycin, 15 μg ml^–1 blasticidin and 400 μg ml^–1 geneticin (Thermo Fisher Scientific).

Plasmids

For doxycycline-inducible knockdown of ZFP36L2, miR-30a-based shRNA sequences (RHS4430-200267269 and RHS4430-200276644, pGIPZ) were subcloned into the pTRIPZ lentiviral vector using XhoI and MluI restriction sites. The resulting constructs encoded the following inducible mature antisense sequences: 5′-TGCACAAGAAGTCGACATC-3′ and 5′-TGTTGAGCAGGGCTGTGCC-3′. Lentiviral particles were produced as previously described⁷⁸ in HEK293T cells (American Type Culture Collection) using a non-targeting control vector (Addgene, 127696). For ZFP36L2 knockout, the doxycycline-inducible Cas9 vector TLCV2 (Addgene, 87360) was cloned with either sgRNA1 (5′-ACCCTTAAGGAGCCGTCGGG-3′) or sgRNA2 (5′-TGCTGGCCGAGTGCCGTCGG-3′) targeting ZFP36L2 to generate indel-mediated gene disruption. These constructs were delivered to organoids by lentiviral transduction. Transduced organoids were selected using puromycin and treated with 2 μg ml^–1 doxycycline for 1 week before sorting eGFP⁺ cells. Sorted cells were seeded at 2,000 cells per 40 μl Matrigel dome and cultured in the continued presence of doxycycline. On day 10, individual organoid clones were picked, dissociated and expanded as single clones. ZFP36L2 knockout was confirmed by western blot analysis. For inducible protein expression, the coding sequence of WT human ZFP36L2 was PCR-amplified from organoid cDNA, in-frame fused to an eGFP or meGFP tag and cloned into AgeI-digested and NheI-digested TLCV2 vector (Addgene, 87360) via Gibson assembly using 20-bp homologous overlaps. A frameshift mutant (fsZFP36L2) was generated using a similar approach by dividing the ZFP36L2 coding sequence into two fragments (nucleotides 1–429 and 433–1359) for Gibson assembly. A single-nucleotide deletion (guanine deletion at codon Gly143) was introduced on the primer through site-directed mutagenesis. Lentiviral particles carrying doxycycline-inducible WT ZFP36L2–eGFP, ZFP36L2–meGFP, fsZFP36L2–eGFP or eGFP-only control were produced for downstream experiments. To generate ADAR-fusion constructs for HyperTRIBE, the hyperactive catalytic domain of Drosophila melanogaster ADAR was PCR-amplified from a plasmid template (Addgene, 166969) and inserted in-frame into the doxycycline-inducible WT and fsZFP36L2 expression constructs. The resulting fusion constructs encoded ADAR-tagged ZFP36L2 or fsZFP36L2, followed by a self-cleaving T2A peptide linked to eGFP (Extended Data Fig. 11a). Lentiviral particles were generated and used to transduce organoids. For nuclear labelling during live-cell imaging, organoids were stably transduced with mCherry-tagged histone H2B in a pLenti6/V5-DEST backbone (Addgene, 89766). All the plasmid sequences are listed in Supplementary Table 4a.

Western blotting

Approximately 2 million organoid cells were recovered from Matrigel by disrupting domes in 2 mM EDTA in DPBS with orbital shaking at 4 °C for 30 min. Cells were washed, centrifuged (500g, 5 min, 4 °C) and lysed in 1× RIPA buffer (50 mM Tris-HCl, pH 8.0, 150 mM NaCl, 0.1% SDS, 0.5% sodium deoxycholate and 1% NP-40) supplemented with protease inhibitors (Roche, complete EDTA-free protease inhibitor cocktail) and phosphatase inhibitors (Sigma-Aldrich, 524636-1SET) and benzonase nuclease (2% v/v; Thermo Fisher Scientific, 70-664-3) for 30 min on ice. Lysates were sonicated using a Bioruptor (10 s on–10 s off for 10 cycles at 4 °C). The protein concentration was determined using a Pierce BCA Protein Assay kit (Thermo Fisher Scientific, 23227).

For immunoblotting, 50 μg of total protein per sample was separated by SDS–PAGE on Bis–Tris polyacrylamide gels (Thermo Fisher Scientific), transferred to methanol-activated 0.45 μm PVDF membranes (Millipore, IPFL00010) and blocked with 5% non-fat milk in TBST for 30 min at room temperature. Membranes were incubated overnight at 4 °C with the following primary antibodies: mouse anti-β-actin (1:2,000; Abcam, ab8226), anti-GAPDH (1:2,000; Abcam, ab181602) and anti-ZFP36L2 (1:200; Santa Cruz, clone A-3, sc-365908).

After washing, membranes were incubated with HRP-conjugated anti-mouse or anti-rabbit secondary antibodies (1:10,000; Cell Signaling Technology, 7076 for mouse and 7074 for rabbit). Signals were detected using an iBright CL1000 imaging system (Invitrogen) with SuperSignal West Femto Maximum Sensitivity substrate (Thermo Fisher Scientific).

Reverse transcription and qPCR

Total RNA was extracted from organoids using a RNeasy Mini kit (Qiagen) according to the manufacturer’s instructions. cDNA was synthesized from 2 μg of total RNA using a Transcriptor First-Strand cDNA Synthesis kit (Roche). qPCR was performed using TaqMan gene expression assay primers (ZFP36L2: Hs00272828_m1; GAPDH: Hs02758991_g1, Mm99999915_g1; ACTB: Hs01060665_g1; Thermo Fisher Scientific) on an ABI QuantStudio 7 Pro Real-Time PCR system (Applied Biosystems). Relative expression was computed using Design & Analysis software (v.2.6.0) using the \({2}^{-\Delta \Delta {C}_{t}}\) method and normalized to GAPDH or ACTB expression.

CRC organoid dedifferentiation and scRNA-seq

OKG146P organoids expressing shZFP36L2 or shCtrl were cultured initially in HISC (with 2 μg ml^–1 doxycycline) for 7 days, and live cells were collected for scRNA-seq. A portion of the same batch of organoids was then passaged into IGFF medium (with 2 μg ml^–1 doxycycline) for 7 days to induce differentiation as previously described¹⁰. Organoids were then passaged into complete HISC medium (with 2 μg ml^–1 doxycycline) for an additional 7 days to induce dedifferentiation into an ISC-like state, followed by scRNA-seq. Organoids continuously maintained in complete HISC medium (with 2 μg ml^–1 doxycycline) for 7 days served as controls.

Organoids were dissociated into single cells as described above. For sample multiplexing, single-cell suspensions were incubated with TotalSeq hashtag antibodies (BioLegend, B0251, B0252, B0253) for 30 min on ice. Viable (DAPI-negative) cells were sorted using a 130 μm nozzle on a SH800S (Sony) cell sorter and collected into DPBS containing 0.04% BSA. Equal numbers of hashtag-labelled shZFP36L2 and control cells from each condition were pooled for scRNA-seq using a Chromium Single Cell 3′ (v.3.1) platform (10x Genomics) according to the manufacturer’s protocol. FACS-sorted cells were washed with DPBS + 0.04% BSA and resuspended at 700–1,300 cells per μl. Cell viability was above 90%, as determined by 0.2% Trypan Blue exclusion. Up to 10,000 cells per sample were targeted for droplet-based encapsulation and barcoding. After reverse transcription, emulsions were broken and cDNA was purified using Dynabeads MyOne SILANE, followed by PCR amplification. Libraries were sequenced on an Illumina NovaSeq S4 platform using the following read configuration: read 1, 28 cycles; i7 index, 8 cycles; read 2, 90 cycles.

Data processing, quality control and batch correction

Raw scRNA-seq data were pre-processed to remove ambient RNA contamination using CellBender (v.1.0.0). Doublets were identified and excluded using DoubletDetection (v.3.0). Cells with log-transformed library sizes deviating by more than 3 MADs below or 5 MADs above the median were considered outliers and removed. Additional quality control metrics, including total UMI counts, number of detected genes, mitochondrial gene fraction and ribosomal gene fraction, were computed per cell. Cells with fewer than 500 total counts, fewer than 400 detected genes, greater than 20% mitochondrial content or otherwise poor-quality metrics were excluded. Hashtag-related genes were removed to eliminate multiplexing artefacts.

Normalization and batch correction

After quality filtering, normalization was performed using Scran (v.3.20), and log-transformed expression values were generated. The dataset was aggregated for downstream analyses. Batch correction was performed using single-cell variational inference (scVI)⁷⁹. A raw count matrix was input, and Seurat (v.3) was used to identify 5,000 HVGs. The scVI model was initialized with 100 latent dimensions, incorporating mitochondrial content as a covariate.

Dimensionality reduction and clustering

PCA was performed, retaining 280 principal components that explained 75% of the variance. Cells were clustered using the Leiden algorithm (resolution of 1.8), and UMAP embeddings were generated for visualization. Differential gene expression analysis was performed using the Wilcoxon rank-sum test.

Local variability analysis

To assess local variability in cell states, KNN entropy was computed by calculating the Shannon entropy of ZFP36L2 expression in each cell’s 15-nearest neighbour graph. Entropy distributions across experimental groups (for example, shZFP36L2 versus shCtrl) were visualized using boxplots. Kernel density estimation plots were also used to visualize expression variability in low-dimensional space.

Differential abundance analysis

To evaluate shifts in cell population composition, MELD⁸⁰ was used to partition the UMAP embedding into transcriptionally similar neighbourhoods. The method quantified relative differences in the abundance of these regions between conditions (for example, shZFP36L2 versus shCtrl).

Cell-state classification

Cell-state classification was performed using PhenoGraph⁸¹, which integrates graph-based clustering with phenotypic similarity. An affinity matrix was constructed based on nearest-neighbour relationships, and batch effects were corrected using Harmony⁸². Cell-state labels were assigned to organoid cells by propagating annotations from reference patient-derived cell states.

Bulk RNA-seq

Total RNA was extracted from organoids with conditions specified in the figure legends using a RNeasy Mini kit (Qiagen) according to the manufacturer’s instructions. Three technical replates for each sample were prepared for all the experiments. After RiboGreen quantification and quality control by Agilent BioAnalyzer, 300 ng of total RNA with RIN values of 9.8–10 underwent polyA selection and TruSeq library preparation according to instructions provided by Illumina (TruSeq Stranded mRNA LT kit, RS-122-2102), with 8 cycles of PCR. Samples were barcoded and run on NovaSeq 6000 in a PE100 flowcell, using a NovaSeq 6000 S4 NovaSeq Reagent kit (200 Cycles) (Illumina). An average of 109 million paired reads was generated per sample. Ribosomal reads represented 0.4–0.7% of the total reads generated and the per cent of mRNA bases averaged 92%.

Bulk RNA-seq data analysis

Adapter sequences were trimmed using Cutadapt (v.4.8), and processed reads were aligned to the human reference genome (hg38) using STAR (v.2.7.11b). Gene-level read counts were quantified using HTSeq (v.2.0.5), and differential expression analysis was performed using DESeq2 (v.1.42.0).

HyperTRIBE data analysis

WT ZFP36L2 or fsZFP36L2 was expressed as a fusion with the hyperactive E488Q mutant catalytic domain of ADAR, which enables the detection of ZFP36L2-interacting mRNAs through capture of adenosine (A) to inosine (I, read as guanosine (G)) edits through RNA-seq⁴⁸. The construct design is shown in Extended Data Fig. 11a. Following stable lentiviral integration, organoids were seeded as single cells, cultured in HISC medium for 3 days and then treated with doxycycline for 4 days before flow sorting for eGFP⁺ (ADAR fusion protein expressing) cells and RNA-seq. Day 5 organoids stably transduced with doxycycline-inducible ZFP36L2–ADAR or fsZFP36L2–ADAR constructs were treated with 2 μg ml^–1 doxycycline for 2 days to induce transgene expression. eGFP⁺ live cells were isolated by flow cytometry, and total RNA was extracted using a RNeasy Mini kit (Qiagen). mRNA libraries were prepared and sequenced as described above. ADAR-seq data were processed as previously described⁴⁹. In brief, adapter sequences were trimmed using Cutadapt (v.4.8), and reads were aligned to the hg38 genome and construct sequences using STAR (v.2.7.11b). RNA editing sites were identified using GATK HaplotypeCaller (v.3.8.1.0), with stringent filtering to retain only A-to-G substitutions on the forward strand and U-to-C substitutions on the reverse strand. Known variants annotated in dbSNP (human_9606_b144_GRCh38p2) were excluded. The editing frequency at each site was assessed using a Beta-binomial test, with multiple testing correction performed using the Benjamini–Hochberg method.

Gene decay rate comparison

PDO MSK107Li and OKG146Li organoids expressing shCtrl or shZFP36L2 were cultured as described above, supplemented with 2 μg ml^–1 doxycycline. On day 5, organoids were treated with or without 5 μg ml^–1 ActD for 5 h. Organoids were then collected, and total RNA was extracted using a RNeasy Mini kit (Qiagen). mRNA libraries were prepared and sequenced as described above.

Adapter sequences were trimmed using Cutadapt (v.4.8), and processed reads were aligned to the human reference genome (hg38) using STAR (v.2.7.11b). Gene-level read counts were quantified using HTSeq (v.2.0.5), and differential expression analysis was performed using DESeq2 (v.1.42.0) to compare the 5-h-treated versus untreated samples for shCtrl or shZFP36L2 separately to capture gene decay rates in each condition. To compare decay rates between shCtrl and shZFP36L2 experiments, gene-level read counts were normalized globally for both shCtrl and shZFP36L2 cells, with and without treatment. The decay rate dependency on ZFP36L2 for each gene was calculated using the following formula:

Thiol-linked alkylation for the metabolic sequencing of RNA

To measure RNA decay kinetics, we performed SLAM-seq⁵² in CRC organoids using a SLAM-seq Kinetics kit–Catabolic Kinetics Module (Lexogen) followed by library preparation with a QuantSeq 3′ mRNA-seq V2 Library Prep kit FWD with Unique Dual Indices (Lexogen). The SLAM-seq workflow is based on metabolic labelling of newly synthesized RNA with 4sU, followed by alkylation of 4sU-containing RNA and detection of T > C conversions after sequencing. For all steps from 4sU handling up until library preparation, samples and reagents were protected from white light and handled in the dark or wrapped in foil whenever possible, as recommended by the manufacturer because 4sU is light sensitive and can crosslink.

We first performed an organoid viability titration to determine a non-toxic working concentration of 4sU. shCtrl-expressing MSK107L organoids were dissociated and seeded in 96-well plates as 40 μl Matrigel domes containing 2,000 cells per well in HISC medium. On day 6, organoids were treated with a range of 4sU concentrations, with medium replaced every 2 h, and cell viability was measured after 13 h using CellTiter-Glo assays. The concentration causing 10% growth inhibition over the assay window was selected for the final experiment, as per the manufacturer’s recommendations. On the basis of this titration, 500 μM 4sU was used for all subsequent labelling experiments.

For the catabolic labelling experiment, organoids were labelled with 500 μM 4sU for 6 h, with fresh 4sU-containing medium supplied every 2 h to maintain labelling efficiency. At the end of labelling, one set of samples was collected immediately as the t₀ time point. To measure mRNA decay kinetics, 4sU-containing medium was removed and organoids were washed by adding 3 ml warm PBS and incubating for 10 min to allow residual labelling medium to equilibrate out. The PBS was then replaced with HISC medium containing 50 mM uridine and samples were collected at 2 h, 5 h and 12 h. At each time point, the medium was removed completely and organoids were lysed directly in the culture well by adding 1 ml TRIzol and pipetting 15 times to fully dissolve the Matrigel domes. Lysates were transferred to pre-labelled 1.5 ml tubes, snap-frozen and stored at −80 °C until RNA extraction. Total RNA was subsequently isolated using a standard TRIzol-based extraction protocol.

Purified RNA was then subjected to chemical derivatization according to the SLAM-seq Kinetics kit–Catabolic Kinetics Module manufacturer’s instructions. In brief, 4sU-labelled RNA was alkylated with iodoacetamide under the recommended reaction conditions to modify incorporated 4sU residues before library preparation. Alkylated RNA was purified and used as input for sequencing library construction. SLAM-seq chemistry enables reverse transcriptase to incorporate G opposite alkylated 4sU during cDNA synthesis, which results in diagnostic T > C conversions in sequencing reads derived from newly labelled transcripts. Sequencing libraries were generated from total RNA using a QuantSeq 3′ mRNA-seq V2 Library Prep kit FWD with Unique Dual Indices according to the manufacturer’s instructions. This protocol uses oligo(dT)-primed first-strand synthesis, followed by RNA removal, random-primed second-strand synthesis, bead purification and PCR amplification to generate strand-specific Illumina-compatible libraries enriched for the 3′ ends of polyadenylated transcripts. Unique dual indices were introduced during PCR amplification. Libraries were prepared from total RNA without previous poly(A) enrichment or rRNA depletion, consistent with the QuantSeq workflow.

SLAM-seq data processing and half-life analysis

Raw SLAM-seq reads were processed using the nf-core/slamseq pipeline (v.1.0.0) implemented in Nextflow. Sequencing reads were trimmed for adapters and low-quality bases using Trim Galore, and processed reads were aligned using Slamdunk, which performs conversion-aware mapping, alignment filtering, multimapper recovery, SNP calling to mask genomic T > C variants and quantification of both total and T > C-converted reads. Read counts were first generated at the UTR level and then collapsed to the gene level. The pipeline also generated quality-control metrics summarizing nucleotide-conversion frequencies across reads, read positions and gene–UTR positions, together with trimming and alignment statistics compiled in a MultiQC report. Human organoid datasets were processed against the GRCh38 reference genome. Generation of 3′ UTR annotations and Nextflow pipeline configurations followed the settings described in a previous study⁸³. For RNA half-life estimation, gene-level T > C conversion rates were background-corrected using matched unlabelled samples, with control samples corrected using unlabelled control samples and knockdown samples corrected using unlabelled knockdown samples. In organoid samples, residual 4sU retention in Matrigel resulted in delayed peak labelling, with many transcripts reaching maximal T > C signal at 2 h after chase initiation. Therefore, for the 5,888 genes that peaked at t = 2 h in control samples, t₂ was considered the first effective decay time point, and half-lives were calculated by fitting a single-exponential decay model to the t₂, t₅ and t₁₂ time points. RNA half-lives were calculated as t_1/2 = ln(2)/k, where k is the fitted first-order decay constant. For control versus knockdown comparisons, only these 5,888 genes were considered, and genes were further filtered for good decay-model fit in the control condition (R² exp > 0.6, where R² exp is the coefficient of determination of the single-exponential decay model fit for each gene).

Live-cell confocal microscopy

Confocal imaging was performed on stably transduced, doxycycline-inducible eGFP, fsZFP36L2–eGFP, ZFP36L2–eGFP and meGFP organoids seeded in a 1:1 mixture of culture medium and Matrigel on Ibidi glass-bottom chamber slides (80821). Imaging was carried out using a Nikon CSU-W1 SoRa spinning disk confocal microscope equipped with a ×60 oil-immersion objective, 1.49 NA, 300 ms exposure time, 5% laser power and 1 μm z-stack intervals. Variable settings, treatments or imaging time windows specific to each experiment are detailed in the figure legends. For nuclear–cytoplasmic shuttling experiments, organoids expressing eGFP-tagged constructs were co-transduced with a H2B–mCherry nuclear marker. Live-cell imaging was performed using the same microscope setup at ×60 magnification, with 300 ms exposure time, 5% laser power and continuous acquisition of 0.2 μm z stacks.

Condensate tracking analysis

Time-lapse videos were first motion-corrected for sample drift using the Correct 3D Drift plugin in Fiji, with drift correction restricted to the x–y plane and based on the condensate channel to avoid instability in z axis estimation. Background subtraction was then performed on the condensate channel using a rolling-ball radius of 50 μm to reduce variability arising from elevated background fluorescence in some cells. Condensates were subsequently detected and tracked in 3D using the TrackMate plugin in Fiji (TrackMate 7)⁸⁴. Spot detection was performed with the difference-of-Gaussians detector using a sigma of 1.25 μm. The quality threshold was manually adjusted for each video on the first frame to optimize condensate detection while minimizing false-positive assignments. Tracking was performed using the Simple LAP tracker, with a maximum linking distance of 4 μm between consecutive frames and a maximum gap-closing distance of 2 μm. Tracklets spanning fewer than two frames were excluded from further analyses. The remaining trajectories were overlaid on maximum-intensity projections of the drift-corrected videos. For visualization, tracks were displayed retrospectively for up to five preceding frames across all z planes.

FRAP

FRAP experiments were performed using a Stellaris confocal imaging system (Leica) equipped with a ×63 1.4 NA oil-immersion objective. Stably transduced, doxycycline-inducible eGFP, fsZFP36L2–eGFP and WT ZFP36L2–eGFP organoid lines were seeded on Lab-Tek chamber slides (155409) 3 days before imaging, with doxycycline (2 μg ml⁻¹) added 24 h before the experiment to induce construct expression. A ROI measuring 1.5 × 1.5 μm was photobleached at 35% laser power for 15 frames (0.6 s per frame), followed by acquisition of 30 frames (1 s per frame) to monitor fluorescence recovery. Fluorescence intensity in the ROI was quantified using Fiji (ImageJ), and the mobile fraction was calculated as:

where F(∞) is the fluorescence intensity at steady state after recovery, F(0) is the intensity immediately after bleaching and F(pre) is the pre-bleach intensity.

Statistics and reproducibility

No statistical method was used to predetermine sample sizes. The number of samples (n) is indicated in each figure panel, and statistical analyses are described in the corresponding figure legends. In vitro experiments were repeated a minimum of three times independently with similar results unless otherwise noted. All staining experiments were performed with a minimum of three biological replicates. Mice that died within 24 h of surgery were excluded from analysis as procedure-related complications; no other data were excluded.

Reporting summary

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