Large-scale CRISPR screening in primary human 3D gastric organoids enables comprehensive dissection of gene-drug interactions

Abstract

Understanding how genes influence drug responses is critical for advancing personalized cancer treatments. However, identifying these gene-drug interactions in a physiologically relevant human system remains a challenge, as it requires a model that reflects the complexity and heterogeneity among individuals. Here we show that large-scale CRISPR-based genetic screens, including knockout, interference (CRISPRi), activation (CRISPRa), and single-cell approaches, can be applied in primary human 3D gastric organoids to systematically identify genes that affect sensitivity to cisplatin. Our screens uncover genes that modulate cisplatin response. By combining CRISPR perturbations with single-cell transcriptomics, we resolve how genetic alterations interact with cisplatin at the level of individual cells and uncover an unexpected link between fucosylation and cisplatin sensitivity. We identify TAF6L as a regulator of cell recovery from cisplatin-induced cytotoxicity. These results highlight the utility of human organoid models for dissecting gene-drug interactions and offer insights into therapeutic vulnerabilities in gastric cancer.

CRISPR-Cas9-based screens have allowed the study of gene-drug interactions. Here, the authors develop CRISPR-Cas9 knock-out, activation and repression screens in human gastric 3D organoids, also integrating single-cell CRISPR screens, to identify genes involved in the response to cisplatin in gastric cancer.

Introduction

Recent years have witnessed an emergence in the use of human 3D organoids as a methodological complement to conventional 2D cell lines. Organoids represent robust in vitro cultures that preserve tissue architecture, stem cell activity, multilineage differentiation, genomic alterations, histology, and pathology of primary tissues^1,2. As a result, primary human organoids have enabled the exploration of biological processes in both normal physiology and various pathological states that were previously difficult to study^3,4. Additionally, organoids offer widespread applications in cancer biology, such as oncogene modeling, therapeutic evaluation, and precision medicine, due to their ability to be derived from both normal and tumor tissues of cancer patients. By replicating the therapeutic vulnerabilities observed in clinical settings in response to genomic alterations, organoids provide a powerful model for uncovering the genetic mechanisms behind synthetic lethal and buffering gene-drug interactions. However, even with the availability of well-defined therapeutic agents that facilitate the identification of genetic background-specific drug responses, a comprehensive understanding of gene-drug interactions by using human 3D organoids has been limited by the lack of genome editing tools for high-throughput functional analysis.

The development of CRISPR technology has revolutionized genome editing by providing a precise, efficient, and versatile tool for targeted DNA modification in various organisms^5,6. In human immortalized 2D cell lines, large-scale CRISPR-based genetic screens have yielded significant insights into both normal and disease biology, serving as an unbiased method for investigating disease mechanisms and gene-drug interactions^7,8. While CRISPR applications in 3D organoids are more challenging than in 2D cell lines, they have been successfully used to disrupt individual genes. Several preclinical human transformation organoid models have been developed through sequential CRISPR modifications of single or multiple oncogenic loci^9–13. Even more exciting is the implementation of CRISPR cutting screens in 3D organoids to explore factors such as growth factor dependency for cell growth^14–16, identifying potential cancer-driving forces^17–19, and uncovering key genetic events in cell differentiation^20–22. However, the feasibility of using CRISPR screens in 3D human organoids to investigate gene-drug interactions remains to be fully established.

In addition to using CRISPR to generate DNA double-strand breaks for indel formation, various other CRISPR-based approaches have been developed to investigate biological mechanisms. For example, CRISPRi (interference)^23–25 and CRISPRa (activation)^26–30 technologies utilize catalytically inactive Cas9 (dCas9) fused to either a transcriptional repressor (e.g., KRAB, Krüppel associated box) or a transcriptional activator (e.g., VPR, VP64-p65-Rta). CRISPRi and CRISPRa offer the advantages of precise, adjustable gene expression without causing genomic indels or the nonspecific toxicity associated with Cas9-induced DNA double-strand breaks³¹. Additionally, single-cell RNA-sequencing pooled CRISPR screens can simultaneously sequence transcriptomes and sgRNAs from individual cells, enabling a comprehensive analysis of sgRNA-specific effects on genetic regulatory networks at single-cell resolution^32–34. While some of these methods have been tested on a small scale^35,36, they have not yet been fully adopted for large-scale screens in tissue-derived organoids.

Here, we report a systematic approach that enables a full suite of high-representation CRISPR-based genetic screens, including CRISPR cutting, CRISPRi, CRISPRa, and single-cell CRISPR screens, to study gene-drug interactions in organoids, using oncogene-engineered human gastric tumor models. Specifically, we identify previously unappreciated genes that contribute to cell growth and sensitivity to the chemotherapy drug cisplatin in gastric cancers. Coupled with single-cell CRISPR screens, our data reveal DNA repair pathway-specific transcriptomic convergence in cisplatin-treated organoids, manifested by distinct high-dimensional gene expression profiles and growth phenotypes. We uncovered an unexpected functional connection between protein fucosylation and cisplatin sensitivity. We identified TAF6L as a key gene involved in cell proliferation during the recovery phase following cisplatin-induced DNA damage. Our study highlights the power of robust CRISPR-based genetic screens in primary human organoids to explore fundamental and translational biological questions.

Results

Establishment of CRISPR screening in 3D oncogene-engineered human gastric tumor organoids

To test the feasibility of CRISPR-based genetic screening in 3D organoids, we utilized an oncogene-engineered human gastric tumor organoid model. The TP53/APC double knockout (DKO) organoid line was established in our previous studies by sequentially disrupting TP53 and APC³⁷, two common oncogenic loci in gastric adenocarcinoma^38,39, from non-neoplastic human gastric organoids (Fig. S1A). This engineered organoid model provides a relatively homogeneous genetic background, minimizing variability and enabling precise identification of gene-function relationships in CRISPR-based screens. Building on our previous study that Cas9 integration enhances CRISPR knockout efficiency in gastric organoids⁹, we first generated stable Cas9-expressing TP53/APC DKO organoids using lentiviral transduction (Fig. 1A). To demonstrate highly efficient CRISPR/Cas9 cleavage in organoids, we delivered a second lentiviral construct containing a GFP reporter and a GFP-targeting single guide RNA (sgRNA) (Fig. 1B). In contrast to the nearly universal GFP expression observed in parental cells, over 95% of Cas9-expressing TP53/APC DKO cells became GFP-negative, indicating robust Cas9 activity (Fig. 1B).

Fig. 1. CRISPR KO screen in TP53/APC DKO oncogene-engineered human gastric organoids.

A Establishment of a stable Cas9-expressing engineered TP53/APC DKO human gastric organoid line. Cas9 expression was confirmed by immunoblot analysis. B Highly efficient CRISPR/Cas9 cleavage in Cas9-expressing organoids. A lentiviral construct that contains a GFP-targeting sgRNA was delivered into organoids to deplete the GFP reporter in the same construct. GFP-positive cells were quantified by flow cytometry. C Timeline of CRISPR KO screens. CRISPR KO screens were conducted as two completely independent experiments, involving separate lentiviral transductions performed at different times. D Volcano plot summarizing knockout phenotypes and statistical significance (determined by two-sided Mann-Whitney U test) for sgRNA targets in the pooled CRISPR KO screen. Each dot represents an sgRNA target. Significant hits are labeled in red, with a more negative phenotype score indicating a stronger growth defect. E Gene ontology (GO) analysis identified top terms significantly associated with identified growth defect hits. F Independent validation of significant hits. Stable KO organoid lines were generated individually by lentiviral transduction of single sgRNA with a BFP reporter. 20,000 BFP-positive single cells were sorted and plated in 40 μL Matrigel domes. Brightfield images were taken for organoid growth on day 12.

In a pilot experiment to test the feasibility of CRISPR screens in organoids, we transduced a validated pooled lentiviral library of 12,461 sgRNAs targeting 1093 membrane proteins, each with ~10 sgRNAs/gene, alongside 750 negative control non-targeting sgRNAs (Supplementary Data 1) into Cas9-expressing TP53/APC DKO organoids. Following lentiviral transduction, we ensured that the number of infected cells provided a cellular coverage of >1000 cells per sgRNA from the outset. After harvesting a subpopulation 2 days post-puromycin selection (time point 0, T0), we continued to culture the remaining organoids at this same cellular coverage (>1000 cells per sgRNA) throughout the screening until day 28 (time point 1, T1). We then measured the relative abundance of each sgRNA by next-generation sequencing to reveal how each sgRNA affected cellular growth, with increasing or decreasing sgRNA abundance inferring growth advantages or disadvantages, respectively (Figs. 1C and S1B). Based on sgRNA counts, we observed a near-complete library representation at T0 (99.9%, 1092 target genes) with overall consistency between independent experimental replicates (Fig. S1C). The gene-level phenotype score (see “Methods”) indicated that control sgRNAs were clustered around zero (no phenotype) (Fig. 1D). Upon sequencing at T1 versus T0, a group of 68 significant drop-out genes whose corresponding sgRNAs were under-represented (compared to the control sgRNA distribution) was identified (Fig. 1D; Supplementary Data 2), representing a growth defect phenotype. These genes were enriched in pathways related to essential biological processes such as transcription, RNA processing, and nucleic acid metabolic processes (Fig. 1E). In contrast, only a few genes showed increased sgRNA abundance, indicating a growth advantage phenotype upon knockout (Supplementary Data 2). Notably, the depletion of tumor suppressor LRIG1, a negative regulator of ERBB receptor tyrosine kinases^40–42, was identified as the top hit contributing to increased cell proliferation (Fig. 1D).

We independently validated the significant hits in the array using individual sgRNAs rather than a pooled library. Four significant growth defect sgRNA hits, including CD151, KIAA1524, TEX10, and RPRD1B, were randomly selected from the primary screen, which contained a broad range of p-values. Compared to control organoids harboring a negative control sgRNA, the depletion of all four selected genes reproduced the growth defect phenotype (Figs. 1F and S1D). Overall, these results demonstrated the establishment of a large-scale CRISPR-based genetic screening platform in human 3D organoids, thereby increasing confidence in the identified target genes.

Regulation of endogenous gene expression in 3D organoids by inducible CRISPRi and CRISPRa systems

To establish an inducible system that allows for controlled, temporal regulation of endogenous gene expression, we engineered TP53/APC DKO gastric organoid lines with doxycycline-inducible dCas9-KRAB (iCRISPRi) or dCas9-VPR (iCRISPRa) systems using a sequential two-vector lentiviral approach (Fig. 2A). First, we generated organoid lines expressing rtTA, followed by the introduction of a doxycycline-inducible cassette containing a dCas9 fusion protein along with an mCherry reporter. To establish stable iCRISPRi and iCRISPRa organoids, we sorted mCherry-positive cells after induction. We observed no obvious growth defects in mCherry-positive dCas9-expressing organoids, indicating the low toxicity of dCas9 fusion proteins (Fig. 2B). The expression of dCas9-KRAB or dCas9-VPR was confirmed by Western blotting (Fig. 2C). Doxycycline withdrawal induced degradation of dCas9 fusion proteins, whose expression was quickly restored by re-induction, suggesting tight control of iCRISPRi and iCRISPRa cassettes (Fig. 2C).

Fig. 2. Establishment of inducible CRISPR activation (iCRISPRa) and CRISPR inhibition (iCRISPRi) in human organoids.

A Schematic of the generation of stable iCRISPRi and iCRISPRa organoid lines. A rtTA-expressing TP53/APC DKO human gastric organoid line was first generated by lentiviral transduction, followed by secondary lentiviral transduction of iCRISPRi or iCRISPRa cassettes. After doxycycline induction, the rtTA-doxycycline complex can bind the TRE3G promoter and thus drive the expression of dCas9 fusion proteins and mCherry reporters. The mCherry-positive cell population was sorted to establish stable iCRISPRi and iCRISPRa human gastric organoid lines. B Brightfield and Immunofluorescence images suggested that dCas9-expressing iCRISPRi TP53/APC DKO human gastric organoids (mCherry-positive) were morphologically well-organized after doxycycline induction. C Western immunoblotting time course analysis of doxycycline induction of dCas9-KARB or dCas9-VPR fusion proteins. D iCRISPRi-sgCXCR4 and iCRISPRa-sgCXCR4 stable organoid lines were established by lentiviral transduction of an sgRNA-BFP construct that targets endogenous CXCR4 promoter regions. CXCR4-positive cells were quantified by antibody staining and flow cytometry 7 days post-doxycycline induction. E Western immunoblotting analysis demonstrated genetic knockdown (iCRISPRi) or overexpression (iCRISPRa) of SOX2 protein using two different sgRNAs, sgSOX2#1 and sgSOX2#2.

To test the function of iCRISPRi and iCRISPRa, we designed sgRNAs targeting the promoter of CXCR4, a cell surface receptor, and then analyzed CXCR4-positive cell populations by antibody staining and flow cytometry. Five days post-induction, iCRISPRi-sgCXCR4 organoids exhibited a decreased CXCR4-positive population (3.3%) compared to the parental iCRISPRi organoids (13.1%) (Fig. 2D). In contrast, the CXCR4-positive population increased in iCRISPRa-sgCXCR4 organoids (57.6%) (Fig. 2D). Moreover, individual sgRNAs targeting the promoter of the transcription factor SOX2 were sufficient to inhibit (iCRISPRi-sgSOX2) or enhance (iCRISPRa-sgSOX2) SOX2 expression (Fig. 2E). Overall, the inducible genetic knockdown (iCRISPRi) and overexpression (iCRISPRa) systems enabled effective modulation of individual target gene expression in organoids, demonstrating their utility for precise gene regulation in a 3D culture context.

Functional iCRISPRi and iCRISPRa screens reveal cisplatin sensitization pathways

Next, we developed iCRISPRi and iCRISPRa screening platforms in organoids to investigate the mechanism of cisplatin action, a chemotherapeutic drug that induces DNA crosslinking, while also providing a phenotypic trait for robust positive selection (Fig. 3A). Since many genes directly involved in DNA repair and cisplatin response encode DNA and RNA-binding proteins, we designed two customized CRISPRi and CRISPRa sgRNA libraries specifically targeting an identical group of 1952 genes encoding DNA and RNA-binding proteins (Supplementary Data 3). This targeted gene set includes key GO terms associated with cisplatin responses, such as RNA polymerase II complex, mediator complex, mRNA spliceosome, nucleotide-excision repair, damaged DNA binding, histone modification, and chromatin remodeling complex (Supplementary Data 3). Each sgRNA construct in the library expressed a GFP reporter and puromycin-resistant gene for selection, with a total of ~12,500 sgRNAs with 5 independent sgRNAs targeting each gene, as well as ~5% non-targeting negative control sgRNAs. These CRISPRi and CRISPRa sgRNA libraries were then evaluated in the TP53/APC DKO tumor organoid lines with stable iCRISPRi and iCRISPRa cassette integration, first in the absence of cisplatin to define baseline cisplatin-independent growth dependencies, and then with cisplatin to identify drug sensitization interactions.

Fig. 3. Large-scale iCRISPRi and iCRISPRa screens in human organoids reveal gene-drug dependencies.

A Schematic of iCRISPRi and iCRISPRa screening strategies. Two new customized CRISPRi and CRISPRa sgRNA libraries that targeted an identical group of 1952 genes (~13,000 guides) were delivered into iCRISPRi and iCRISPRa stable organoids, respectively. Organoids were assigned into two groups - vehicle and cisplatin treatment - after sgRNA library delivery. The frequencies of cells expressing a given sgRNA were determined by next-generation sequencing, and the phenotype scores were quantified with the depicted formula. All screens were conducted as two completely independent experiments, involving separate lentiviral transductions performed at different times. B Summary of cell growth phenotypes (determined by screen score, see details in “Methods”) for sgRNA targets in the pooled iCRISPRi (x-axis) and iCRISPRa screens (y-axis). Each dot represents an sgRNA target. Significant hits were labeled in red, with either a cell defect or a growth advantage phenotype. C Cisplatin-associated gene-drug interactions were identified. Each dot represents a targeted gene. Gene hits were labeled in red, with either a cisplatin sensitivity or a cisplatin persistence phenotype. D A total of 41 significant top-scoring sgRNAs of the CRISPR cisplatin sensitivity screen were listed. Several top hits function as well-established DNA repair genes, and potential DNA damage sensors and mediators were identified as indicated. The dot color indicates the P-value (-log₁₀), and the dot size indicates the screen score (log₂). Two-sided Mann-Whitney test. E Independent validation of significant cisplatin sensitization hits, including ERCC6, ERCC4, and GTF2H5 in TP53/APC DKO iCRISPRi organoids. Stable knockdown organoid lines and control organoids that contain non-targeting sgRNAs were generated individually by lentiviral transduction of a single sgRNA. Fully titrated cisplatin treatment was performed to determine cisplatin sensitivity. Error bars represent the SEM of three independent experiments. F Identification of new cisplatin sensitization hits, including ELOF1, LEO1, and ZNF677. Stable knockdown organoid lines were established, followed by fully titrated cisplatin treatment. Error bars represent the SEM of three independent experiments.

We performed iCRISPRi and iCRISPRa screens in the presence of cisplatin in parallel with the cisplatin-naïve studies (Fig. S2A). After doxycycline induction, most organoids co-expressed dCas9 fusion proteins (mCherry+) and sgRNAs (GFP+) (Fig. S2B). In the treatment group, in each cycle, organoids were treated with cisplatin (IC50 treatment, 1.6 μg/mL) for 24 h, followed by 5 days of recovery. A total of 4 rounds of cisplatin treatment cycles were repeatedly conducted to achieve ~6–7 doubling differences between the treated and untreated groups, which provided sufficient statistical power to measure gene-drug interactions—either buffering or sensitization effects (Fig. S2C). The genomic DNA was sequenced at day 2 (time point 0, T0) versus day 26 (time point 1, T1) after puromycin selection to enumerate the frequency of individual sgRNA lentiviral insertions. We quantified the relative abundance of each sgRNA at T1 versus T0 to determine the degree of clonal expansion or contraction over the culture period, representing growth advantage or disadvantage, respectively. Our data for each screen, which included two independent experiments as replicates, showed high consistency in the iCRISPRi screen, while the iCRISPRa screen exhibited lower replicate correlation, consistent with known technical challenges associated with CRISPRa-based perturbations⁴³ (Fig. S2D).

The iCRISPRa screen in cisplatin-naïve organoids identified 42 cell growth-associated sgRNA activation loci, with 28 causing growth defects and 14 promoting increased growth (Supplementary Data. 4). Among these genes, the well-known oncogenic driver MYC was ranked as the top hit for growth advantage, along with additional loci such as SAMD4B (Fig. 3B). The cisplatin-naïve iCRISPRi screen yielded 394 hits (384 growth defects and 10 growth advantages) that exhibited a cell growth phenotype (Fig. 3B; Supplementary Data. 5). Among the growth defect hits, gene sets controlling essential biological processes, such as RNA catabolism, telomere maintenance, DNA biosynthesis, and cell cycle, were enriched (Fig. S2E).

To identify cisplatin sensitization genes, the relative abundance of individual sgRNA was measured in each population to reveal how genetic perturbations modulate cisplatin sensitivity. In particular, the drug sensitivity phenotype represents the normalized difference in abundance between the treated and untreated populations for each sgRNA⁴⁴, where positive scores indicate that gene knockdown or activation confers protection against drug treatment (enrichment of sgRNA abundance of the treated population versus untreated) and negative scores indicate drug sensitization (further depletion/dropout of sgRNA abundance of the treated population versus untreated) (Fig. S3A). After this normalization, we identified 9 cisplatin sensitization genes whose overexpression in the iCRISPRa screen conferred drug sensitivity (Figs. 3C and S3B; Supplementary Data 6). Of note, the complementary iCRISPRi screen with cisplatin was particularly interesting due to the potential therapeutic relevance of genes whose knockdown could enhance cisplatin sensitivity. The iCRISPRi screen yielded 41 cisplatin sensitization hits, and 9 persistence (buffering) hits whose knockdown correlated with cisplatin-specific sgRNA dropout and enrichment, respectively (Figs. 3C and S3C; Supplementary Data 7). As expected, the top gene ontology (GO) terms of the cisplatin-sensitive genes included the response to DNA damage-induced cellular stress, single- and double-strand break DNA repairs, and RNA splicing (Fig. S3D).

The significant cisplatin sensitization hits contained several gene candidates that both confirmed known pathways and included hits representing uncharacterized mechanisms (Fig. 3D). 14 genes were key components of DNA repair pathways, including nucleotide-excision repair (NER genes, ERCC6, ERCC4, XPA, ELOF1, GTF2H5, and ERCC8), homologous recombination repair (HRR genes, BRCA1, HELQ, and MCM9), and Fanconi anemia (FA genes, SETD1A, BRCA2, FAN1, and BOD1L1) pathways. Moreover, consistent with the function of RNA polymerases as sensors of DNA damage-induced lesions⁴⁵, inhibition of RNA polymerase subunits, such as POLR1D, POLR2B, and BRF2, may block sensing of DNA damage, resulting in cisplatin hypersensitivity. In addition, in agreement with the importance of pre-mRNA processing factors in the DNA damage response⁴⁶, several mRNA 3′ end processing factors and RNA-binding proteins were identified, such as CPSF2, ZFP36L2, and SRSF7. Notably, ELOF1, a recently identified transcription elongation factor that plays an important role in transcription-coupled nucleotide-excision repair⁴⁷, and LEO1, a subunit of the polymerase-associated factor 1 (Paf1) complex RNA polymerase that is required for RNA polymerase II removal in response to DNA damage, were identified^48,49. These results suggested that the CRISPRi screens provide a strong capability to uncover genes associated with cisplatin sensitivity.

To validate the findings of the bulk cisplatin iCRISPRi screen, a total of 14 hits were randomly selected, and representative individual knockdown organoid lines were established by lentiviral transduction of a single sgRNA in iCRISPRi TP53/APC DKO organoids. Importantly, consistent with the pooled screen, 9 of these newly established iCRISPRi knockdown lines confirmed cisplatin sensitivity compared to control organoids carrying non-targeting sgRNA, indicating the high reliability of the original screen (Fig. S4A). As expected, inhibition of DNA repair genes, such as ERCC6, ERCC4, and GTF2H5, conferred cisplatin sensitivity (Fig. 3E). Moreover, ELOF1, LEO1, and a previously underexplored gene, ZNF677, from the primary screen were successfully reconfirmed as cisplatin sensitization hits (Fig. 3F). On the other hand, the remaining hits, including POLR1D, CPSF2, BOD1L1, ZKSCAN2, and HMGN2, showed a rapid general growth defect phenotype after knockdown in the absence of cisplatin (Fig. S4B) and were, therefore, not used for the cisplatin sensitivity assay. These findings support the known function of cisplatin in DNA damage-induced cell death, demonstrating the feasibility of our iCRISPRi and iCRISPRa screening platforms and the ability to map mechanistic pathways and targeted loci.

A multiplexed single-cell CRISPR screening platform in human organoids

We aimed to assess the individual single-cell phenotypes of the diverse cisplatin sensitization loci. A high-dimensional phenotype, capable of bridging the gap between individual sgRNA-associated transcriptional profiles and drug perturbations, can provide deeper mechanistic insights into the known and previously uncharacterized cisplatin sensitization hits (Fig. S5A). Perturb-seq is a multiplexed lentiviral platform combining CRISPR-based genetic perturbations and single-cell RNA-seq (scRNA-seq) to accurately identify transcriptional profiles and single-cell states affected by individual genetic perturbations, extending beyond simple bulk cellular growth phenotypes^32–34 (Fig. 4A). Accordingly, our Perturb-seq plasmids encode a BFP reporter (see “Methods”), allowing the selection of lentivirus-transduced cells via fluorescence-activated cell sorting (FACS). Importantly, in addition to the functional sgRNA driven by a mouse U6 promoter, the sgRNA barcode is also included in the BFP mRNA transcribed by RNA polymerase II and polyadenylated, which can be detected by the 10X Genomics scRNA-seq platform via polyA enrichment. We reasoned that Perturb-seq, by simultaneously identifying the sgRNA perturbation and transcriptome of single cells, would more robustly illuminate the individual mechanisms underlying each gene-drug sensitization and buffering interaction identified in our iCRISPRi and iCRISPRa screens.

Fig. 4. Multiplexed single-cell iCRISPRi and iCRISPRa screens in human organoids.

A Schematic of the Perturb-seq platform in iCRISPRi and iCRISPRa organoids. The droplet-based single-cell RNA-sequencing (scRNA-seq) profiles each individual cell transcriptome with the expressed sgRNAs. B Timeline of Perturb-seq screens. dCas9 fusion proteins were induced by doxycycline 2 days before lentiviral transduction of the Perturb-seq sgRNA library. Six days after sgRNA delivery, organoids were treated by cisplatin or vehicle for 24 h, followed by 10X Genomics scRNA-seq. C The heatmap depicted the clustering of sgRNAs in the multiplexed iCRISPRi and iCRISPRa single-cell CRISPR screening experiments. Gene expression was derived by averaging all single cells that express the same sgRNA for transcript signatures of individual sgRNAs. The color bar represents signed log10(FDR). FDR false discovery rate. D UMAP plot of perturbation subpopulations with sgRNA targets MYC, SAMD4B, HHEX, MXI1, and ZBTB7A from the iCRISPRa experiment in (C). Control cells contain non-targeting control sgRNA. Each dot represents a single cell. E Cell cycle composition of subpopulations from (D). The heatmap indicated changes in cell numbers (percentage, %) in different cell cycle stages. F Bubble chart of gene set enrichment analysis (GSEA) indicated significant changes in gene sets of individual perturbation subpopulations, compared with control cells. NES normalized enrichment score.

We thus created targeted Perturb-seq sgRNA libraries containing manually selected hits from our primary iCRISPRi and iCRISPRa screens (69 and 22 loci, respectively) and non-targeting negative control sgRNAs (Supplementary Data 8). After doxycycline induction, sgRNA library-transduced organoids were treated with either vehicle or cisplatin for 24 h (Fig. 4B), followed by FACS isolation of mCherry+/BFP+ single cells for scRNA-seq. The mRNA expression of a total of 20,414 cells from iCRISPRi and 9026 cells from iCRISPRa, each carrying a single sgRNA per cell, was captured, with an average of 142 cells/sgRNA and 198 cells/sgRNA, respectively (Fig. S5B). Because iCRISPRi and iCRISPRa regulated endogenous transcription, the efficiency of sgRNA effects can be validated by mRNA expression of the individual target genes (Fig. S5C). Moreover, high sgRNA and transcript capture rates indicated a comparable quality in organoids to previous datasets in various cell lines^32,50 (Fig. S5D, E). Finally, the heatmap hierarchical clustering of transcriptional profiles revealed correlations of individual genetic perturbations, uncovering functional gene clusters (Fig. 4C).

As a proof-of-principle to determine functional connections between (1) gene-specific phenotypes (growth), (2) single-cell transcriptional profiles (scRNA-seq), and (3) sgRNA perturbations, all within individual cells, we initially focused on the top iCRISPRa hits that elicited cell proliferation (MYC and SAMD4B) or growth defects (HHEX, ZBTB7A, and MXI) in the absence of cisplatin (Fig. 3D). The UMAP analysis indicated distinct clusters, grouped by individual targeting sgRNAs (Fig. 4D). By decomposing the cell cycle states of individual cells based on single-cell transcriptomes, we observed increased mitotic M-phase populations in cells transduced with sgRNAs targeting MYC, consistent with its known oncogenic properties (Fig. 4E). Conversely, overexpression of HHEX and MXI1 caused a dramatic increase of cells in S and G2-M phases respectively, suggesting potential molecular mechanisms of growth arrest (Fig. 4E). In addition, consistent with the growth advantage phenotype of MYC-overexpressing single cells, the gene set enrichment analysis (GSEA) revealed enrichment of translation, DNA replication, cell cycle progression, gene expression, and telomere maintenance (Fig. 4F). As expected, a significant reduction of the same gene sets was often observed in individual CRISPRa cells exhibiting growth defects, such as those with ZBTB7A and HHEX sgRNAs (Fig. 4F). Notably, a strong defective signature of purine biosynthetic pathways was identified in iCRISPRa-sgHHEX cells, suggesting a potential metabolic component underlying HHEX overexpression-induced S phase arrest (Fig. 4F). Taken together, these studies demonstrated the power of multiplexed single-cell CRISPR screening in human organoids, bridging the gap between genetic screens and molecular traits and illuminating how specific genotypes contribute to unique phenotypes.

Mechanistic interrogation of cisplatin sensitization partners using single-cell CRISPR screens

The iCRISPRi gene-drug interaction screens highlighted numerous gene candidates that confirmed known cisplatin sensitivity and suggested previously uncharacterized mechanisms (Fig. 3D). In particular, many genes identified from our iCRISPRi pooled screens in 3D human gastric organoids were not determined as cisplatin sensitization partners in previous genome-scale CRISPR KO screens in immortalized 2D human cell lines⁵¹, suggesting potential 3D specificity (Fig. S6A). With some exceptions, this intriguing list of gene hits could be generally grouped into known biological processes related to cisplatin-associated DNA repair signaling based on published literature (Fig. S6B). However, we were particularly intrigued by the potential application of single-cell CRISPR screening to reveal, in an unsupervised and unbiased fashion, the mechanistic complexity among known cisplatin sensitization hits and in contrast to previously uncharacterized loci. To this end, we analyzed single-cell transcriptional profiles of 29 cisplatin sensitization loci identified by the iCRISPRi single-cell CRISPR screen to explore the mechanistic complexity of their drug sensitivity phenotypes (Fig. S6C).

The cisplatin-naïve organoids in the iCRISPRi single-cell CRISPR screen revealed substantial heterogeneity upon individual genetic knockdowns (Fig. 5A). For example, divergent single-cell transcriptomes were observed in cells transduced with sgRNAs targeting ERCC4, ERCC6, ERCC8, XPA, and GTF2H5, even though these cisplatin sensitization genes are conventionally grouped into the nucleotide-excision repair (NER) pathway. Similar divergence was also evident in transcriptomes of cells transduced with sgRNAs targeting BRCA1, HELQ, and MCM9, which function in the homologous recombination repair (HRR) pathway. In fact, upon hierarchical clustering, the transcriptomes of NER and HRR locus knockdown cells were substantially intermixed rather than segregating to NER- or HRR-specific clusters (Fig. 5A).

Fig. 5. High-dimensional single-cell transcriptomes reveal synergistic effects of gene-drug interactions.

A The cisplatin-naïve organoids in the single-cell iCRISPRi screen revealed heterogeneity upon individual genetic knockdown. For transcript signatures of individual sgRNAs, gene expression was derived by averaging all single cells that express the same sgRNA. B In cisplatin-treated organoids, inhibition of individual genes in the NER or HRR pathways resulted in highly convergent signatures governed by biological pathways. Such clustering was not observed in (A). C Example single-cell profiling of ERCC4 sgRNA-cisplatin gene-drug interactions. The columns represent heatmap expression of the 50 most synergistically regulated mRNAs from the ERCC4 sgRNA-cisplatin combination. Depicted are the average transcriptional profiles for the single perturbations (condition 1—cisplatin alone, condition 2—sgERCC4 alone) and the theoretically predicted gene expressions (condition 3—additive model). The actual measured expression is shown in condition 4—actual model. D Synergistic repression of fucosylated proteins by combining ERCC4 knockdown and 2 µg/ml cisplatin treatment for 48 h. Total fucosylated proteins were determined by AAL Western blotting. E Cisplatin treatment inhibited GMDS expression. GMDS mRNA levels were measured by qPCR. Relative fold change values are shown. Data points represent three technical replicates from one representative experiment out of two independently performed experiments. GMDS expression was detected by Western immunoblotting. Organoids were treated with 2 μg/mL cisplatin for 72 h. The samples derive from the same experiment, but different gels for the tested antibodies and GAPDH were processed in parallel. F Stable GMDS-overexpressing organoid lines were established by lentiviral transduction. GMDS mRNA levels were measured by qPCR. Relative fold change values are shown. Data points represent three technical replicates from one representative experiment out of two independently performed experiments. GMDS overexpression and total fucosylated proteins were detected by Western immunoblotting. The samples derive from the same experiment, but different gels for the tested antibodies and GAPDH were processed in parallel. G Constitutive expression of GMDS sensitized the cisplatin sensitivity of organoids. A fully titrated cisplatin treatment was performed in three independent experiments (N = 3) to determine cisplatin sensitivity. Error bars represent the SEM. H Quantification of γH2AX immunofluorescence staining in nuclei. Each dot represents a single nucleus. Data points represent one representative experiment out of two independently performed experiments. The red horizontal bar represents the mean value. Organoids were treated with 2 μg/mL cisplatin for 48 h.

By contrast, cisplatin-treated organoids exhibited highly convergent transcriptional signatures of NER or HRR pathways upon inhibition of cisplatin sensitization genes. Indeed, cisplatin-treated transcriptomes of the ERCC4, ERCC6, ERCC8, GTF2H5, and ELOF1 iCRISPRi knockdown cells strongly clustered (Fig. 5B), consistent with the common involvement of these genes in NER. Similarly, knockdown of the HRR pathway iCRISPRi hits BRCA1, HELQ, and MCM9 also resulted in similar single-cell transcriptomes, which further did not intermix with the NER knockdown cluster, with the exception of the NER gene XPA (Fig. 5B). This cisplatin sensitization-induced transcriptomic harmonization faithfully reflected the known distinct biology of the NER versus HRR DNA repair pathways versus transcriptomic divergence observed upon iCRISPRi inhibition in cisplatin-naïve organoids.

A significant advantage of single-cell CRISPR screens is the identification of high-dimensional synergistic effects between individual sgRNAs and drug perturbations on single-cell transcriptomes. Indeed, genes that undergo synergistic and concordant regulation by both an individual guide RNA and cisplatin treatment would represent candidate loci for mediating the sensitization interaction. We mined the single-cell transcriptomes for loci that were synergistically repressed or activated by the iCRISPRi sensitization hits and cisplatin, similar to the previously described genetic interaction analysis⁵². In particular, for each individual sgRNA, we quantified effects on distinct mRNAs after treatment with (1) cisplatin alone (i.e., negative control sgRNA plus cisplatin) or (2) sgRNA alone (i.e., no cisplatin). (3) The predicted theoretical sum effect of conditions (1) and (2) on a given mRNA expression was depicted (“additive model”). (4) Finally, the actual measured level of mRNA upon combined treatment with sgRNA and cisplatin constituted a fourth metric, sgRNA + cisplatin (“actual measurement”) (Fig. 5C). We also conducted a similar analysis, implementing a “linear model” of epistasis as previously published⁵², which reached the same conclusions.

This analysis identified numerous gene-drug interactions in which the sgRNA and cisplatin combination synergistically repressed or activated downstream loci much more in the actual scRNA-seq data (condition 4) than would be predicted by the additive model (condition 3). The top 50 such regulated genes were depicted (Fig. 5C). Interestingly, the combination of ERCC4/cisplatin, ERCC6/cisplatin, and ERCC8/cisplatin all synergistically and profoundly repressed GMDS upon single-cell CRISPR analysis to a much greater extent than predicted by the additive model (Figs. 5C and S7A; Supplementary Data 9). GMDS encodes GDP-mannose 4,6-dehydratase, which catalyzes the first step in the GDP-fucose de novo synthesis pathway⁵³ (Fig. S7B). GDP-fucose is the sole donor for protein fucosylation, a critical glycan modification for numerous membrane and secreted proteins. Accordingly, we next tested if the ERCC4 sgRNA + cisplatin combination-induced synergistic repression of GMDS could functionally reduce total cellular protein fucosylation in TP53/APC DKO organoids. Consistent with the single-cell CRISPR screening result, the combination of ERCC4 knockdown and cisplatin treatment strongly inhibited reactivity to fucose-recognizing Aleuria aurantia lectin (AAL) (Fig. 5D) and Ulex europaeus agglutinin I (UEA-I) (Fig. S7C), revealing a decrease of total cellular fucosylated proteins. More broadly, in addition to ERCC4, ERCC6, and ERCC8 (Fig. S6G), single-cell iCRISPRi knockdown of additional NER loci (XPA, ELOF1, GTF2H5) plus cisplatin elicited similar synergistic repression of GMDS (standard deviation > 4), which was not observed upon single-cell iCRISPRi knockdown of HRR genes (Fig. S7D). As GMDS was also synergistically repressed by the combination of cisplatin with either ZKSCAN2 or LEO1 sgRNA, it is conceivable that these loci are unsuspected participants in NER.

To gain deeper insight into the relationship between cisplatin sensitivity and GMDS expression, we first validated the effect of isolated cisplatin treatment on GMDS levels. Consistent with the Perturb-seq data (Fig. 5C), cisplatin treatment decreased GMDS expression (Fig. 5E). To determine whether GMDS expression functionally impacts cisplatin-induced cytotoxicity, we ectopically delivered a full-length GMDS into organoids (GMDS OE) using lentiviral transduction (Fig. 5F). As expected, GMDS expression was sufficient to induce total cellular protein fucosylation (Fig. 5F). Furthermore, GMDS expression effectively rescued the suppressed GMDS levels in cisplatin-treated organoids (Fig. S7E) and resulted in GMDS OE organoids exhibiting an increased sensitivity to cisplatin treatment (Fig. 5G), accompanied by an increase in nuclear phospho-H2AX (γH2AX) (Fig. 5H). Notably, GMDS OE organoids did not compromise the baseline viability of parental organoids in the absence of cisplatin (Fig. S7F), further ruling out nonspecific toxicity associated with elevated GMDS levels. This discrepancy thus dissociated GMDS downregulation from cisplatin-induced cytotoxicity and suggested that the synergistic suppression of GMDS observed with combined cisplatin treatment and NER gene knockdown was not inherently cytotoxic but rather reflected a positive feedback mechanism promoting cell survival following DNA damage.

Next, we tested the hypothesis that the downregulation of fucosylation could act as a pro-survival signal to protect cells from cisplatin-induced cytotoxicity. Inactive fucose analogs, such as 2F-PerAc-Fuc, are taken up by cells through the salvage pathway and inhibit fucosylation by competitively saturating GDP-fucose biosynthetic enzymes, potently depleting endogenous GDP-fucose (Fig. S7B). Accordingly, 2F-PerAc-Fuc inhibited total protein fucosylation (Fig. S7G) but reversed cisplatin cytotoxicity, evidenced by increased cisplatin IC₅₀ (Fig. S7H). These results supported a potential working model in which the downregulation of GMDS-associated fucosylation functions as a DNA damage-induced pro-survival signal to reduce cisplatin sensitivity, independent of additional NER gene knockdown (Fig. S7I). Taken together, these findings reveal a previously uncharacterized crosstalk between GMDS expression and cisplatin sensitivity, highlighting a potential mechanism by which cancer cells modulate their response to cisplatin.

TAF6L is essential for cell recovery following cisplatin treatment

To identify potential previously uncharacterized genes involved in cisplatin sensitivity, we focused on TAF6L, as it is the highest-ranking unreported cisplatin sensitization hit identified from our CRISPR screens (Fig. 3D), aside from ZKSCAN2, whose knockdown causes a growth defect phenotype (Fig. S4B). TAF6L encodes a TATA-box binding protein-associated factor 6-like protein and is a component of the PCAF histone acetylase complex. We first established a stable TAF6L iCRISPRi knockdown TP53/APC DKO organoid line. After doxycycline induction, TAF6L expression was reduced by 85% (Fig. 6A). Consistent with the primary iCRISPRi screen, TAF6L knockdown increased sensitivity to cisplatin compared to the control organoids carrying a non-targeting sgRNA (Fig. 6B). Next, to test whether TAF6L expression influences the cellular response to cisplatin treatment, a full-length TAF6L fused with a GFP reporter (TAF6L-GFP) was transduced into TP53/APC DKO organoids using lentivirus. The expression of endogenous TAF6L and TAF6L-GFP was further confirmed by Western blot analyses (Fig. 6C). As expected, ectopic expression of TAF6L-GFP protected the organoids from cisplatin-induced cell death (Fig. 6D). These results were independently validated in an additional TP53 KO organoid line (Fig. S8A). Previous studies suggested that TAF6L plays a role in maintaining the self-renewal and proliferation of embryonic stem cells⁵⁴. To explore the molecular mechanism of TAF6L-associated cisplatin sensitivity, we first evaluated the role of TAF6L in DNA repair signaling. We examined whether TAF6L expression directly influences HRR and FA pathway activation. However, we did not observe significant changes in RAD51, FANCD2, or ubiquitinated FANCD2 (Ub-FANCD2) expression in TAF6L knockdown or TAG6L-GFP organoids (Fig. S8B, C), suggesting that TAF6L does not directly regulate DNA repair signaling activity. Next, we tested the impact of TAF6L on cell proliferation using an EdU assay before and after cisplatin treatment. Viable EdU+ cells were quantified by flow cytometry (Figs. 6E and S8D). After 48 h of cisplatin treatment, the EdU+ cell population significantly decreased from an average of 18.1% to 3.6% in control organoid lines (Figs. 6E and S8D). However, TAF6L knockdown reduced the EdU+ cell population to an average of 8.0% compared to control organoids (18.1% EdU+ on average), indicating that TAF6L depletion inhibited cell proliferation. Under TAF6L knockdown conditions, only about 1% of cells remained EdU+ following cisplatin treatment. In contrast, TAF6L-GFP expression enabled cells to maintain their ability to proliferate after cisplatin treatment, with an average of 4.8% EdU+ post-treatment. Notably, the re-expression of TAF6L-GFP was sufficient to rescue cell proliferation in TAF6L knockdown organoids (Figs. 6E and S8D). These results suggested that TAF6L is critical for the proliferation that occurs during recovery following cisplatin treatment.

Fig. 6. TAF6L expression affects cisplatin sensitivity by regulating cell proliferation.

A Real-time PCR analysis of complementary DNAs synthesized from mRNA isolated from iCRISPRi TP53/APC DKO organoids carrying either sgControl (Ctrl) or sgTAF6L (KD) 5 days post-induction. Relative fold change is shown from three technical replicates. Data points represent one representative experiment out of two independently performed experiments. B Cell viability of Control (Ctrl) or TAF6L knockdown (KD) iCRISPRi TP53/APC DKO organoids after 3 days of cisplatin treatment. TAF6L (KD) showed a trend of increased sensitivity to cisplatin. Relative cell viability is shown from three technical replicates. Data points represent one representative experiment out of two independently performed experiments. C Western immunoblotting analysis demonstrated TAF6L expression in control (Ctrl), TAF6L knockdown (KD), and TAF6L-GFP overexpression (OE) organoids. D Cell viability of Control (Ctrl) or TAF6L-GFP overexpression (OE) TP53/APC DKO organoids after 3 days of cisplatin treatment. TAF6L-GFP OE showed a trend of decreased sensitivity to cisplatin. Relative cell viability is shown from three technical replicates per condition. Data points represent one representative experiment out of two independently performed experiments. E EdU assay indicated the EdU+ cell population (percentage, %) in parental control, TAF6L KD, TAF6L-GFP OE, and TAF6L-GFP re-expression in TAF6L KD organoids, with or without cisplatin treatment. EdU+ cells were quantified using flow cytometry. Cells were treated with 2 µg/ml cisplatin for 24 h. Quantification of three independent experiments was shown (N = 3). Two-sided T-test. *p < 0.01. F ATAC-seq analysis indicated significant differences in ATAC peaks between control and TAF6L KD organoids at the promoter regions. The pie chart shows the distribution of total differential ATAC peaks. G The pie chart shows the distribution of total differential ATAC peaks in control versus TAF6L-GFP OE organoids. H The lollipop plot shows the fold enrichment (obs/exp) of ATAC-seq peaks across annotated genomic regions in TAF6L KD and TAF6L OE conditions relative to the human reference genome (hg38). Differentially accessible peaks were classified into upregulated and downregulated groups for each condition.

To determine potential biological processes associated with TAF6L, we examined TAF6L-associated transcripts using bulk RNA-sequencing (RNA-seq) in TAF6L knockdown, TAF6L-GFP, and their control organoid lines. Compared to control organoids, the TAF6L knockdown organoids showed 452 differentially expressed genes (fold change > 2 and adjusted p-value < 0.05), with 321 genes consistently upregulated and 131 downregulated (Supplementary Data 10). GO enrichment analysis of the upregulated genes in TAF6L knockdown organoids revealed several key biological processes, such as triglyceride metabolic process, cholesterol transport, and endoplasmic reticulum lumen (Fig. S9A). Moreover, the top GO terms of the downregulated genes in TAF6L knockdown cells included maintenance of gastrointestinal epithelium and epithelium development and differentiation (Fig. S9A). On the other hand, in the TAF6L-GFP-expressing organoids, we identified a total of 1969 differentially expressed genes (fold change > 2 and adjusted p-value < 0.05), including 1079 genes consistently upregulated and 890 downregulated (Supplementary Data 11). Numerous significant GO terms were identified, including the upregulation of brush border membrane, vascular transport, and basolateral plasma membrane (Fig. S9B), as well as the downregulation of extracellular matrix organization, regulation of cell migration, and collagen-containing extracellular matrix (Fig. S9B). We did not identify any significant GO terms related to the DNA damage response, suggesting that TAF6L does not directly regulate the expression of DNA repair genes or other related genes. This was consistent with the single-cell CRISPR results (Fig. 5), as TAF6L did not cluster with NER- or HRR-specific groups, indicating that other mechanisms may be at play. However, in agreement with our conclusion that TAF6L regulates cell proliferation, several GO terms related to cell proliferation were identified (Fig. S9C). Finally, we performed ATAC-seq to explore TAF6L-associated transcriptional regulation and chromatin accessibility. Surprisingly, TAF6L knockdown resulted in a dramatic decrease in ATAC-seq peaks compared with control iCRISPRi organoids carrying non-targeting sgRNA. A total of 36,658 ATAC peaks significantly decreased after TAF6L loss, while only 28 peaks increased (Fig. 6F; Supplementary Data 12). Of note, among the significantly reduced ATAC peaks in promoter regions, the peak located at the TAF6L promoter ranked as the 1st peak (Fig. S9D), consistent with a strong autoregulatory effect at the TAF6L promoter. In contrast, TAF6L-GFP expression did not cause notable changes in chromatin accessibility, with only 7 ATAC peaks increasing and 28 decreasing (Fig. 6G; Supplementary Data 13). To assess whether differential ATAC-seq peaks are preferentially enriched in specific genomic regions, we compared their distribution to genome-wide expectations and visualized the resulting enrichment patterns (Fig. 6H). Taken together, these results suggest that TAF6L may act as a cisplatin sensitization gene by maintaining global chromatin accessibility of loci required for proliferation recovery after chemotherapy exposure.

Discussion

In this study, we employed functional CRISPR screens to uncover the mechanisms underlying gene-drug interactions in primary human 3D organoid cultures. These large-scale genetic screens facilitated both negative selection (cell growth phenotype) and positive selection (drug sensitivity phenotype) in primary human gastric organoids. Our results demonstrated the feasibility of conducting large-scale CRISPR screens with high representation (>1000× coverage) in human 3D organoids across three different CRISPR platforms: CRISPR cutting, CRISPRi, and CRISPRa. This approach yielded highly reliable hits, minimizing the high background noise typically caused by sparse sgRNA coverage and ensuring strong discovery power. Notably, several key findings from our primary screens have been independently validated, further reinforcing the credibility of the identified target genes. Similar unsupervised organoid-based screening strategies could be adapted to a variety of functional assays beyond drug sensitivity, enabling the systematic exploration of multiple pathways and processes in parallel and accelerating discovery in a human model system.

Recent studies have shown that phenotypes associated with CRISPR-mediated knockout of putative oncogenic loci are more pronounced when cancer cells are grown as 3D spheroids than 2D monolayers, suggesting that 3D culture systems may be particularly useful for uncovering cancer vulnerabilities⁵⁵. This advantage likely extends to more physiologically relevant 3D organoid models. In line with this, using iCRISPRi screens, we identified numerous hits already known to participate in DNA damage response to cisplatin treatment, such as genes involved in nucleotide-excision repair (NER), homologous recombination repair (HRR), and Fanconi anemia (FA) pathways. Interestingly, most genes with strong effects on cisplatin sensitivity were identified exclusively in the iCRISPRi screens, but not in the iCRISPRa screen. This is consistent with the idea that genetic knockdown and overexpression of a given target may not produce reciprocal phenotypes due to saturation effects from homeostatic gene expression levels. Importantly, our 3D CRISPR screens in human organoids revealed gene-specific cisplatin vulnerabilities that have not been detected in 2D genetic screens. We focused on TAF6L, a gene not previously reported to be associated with cisplatin sensitivity. Our experiments across different organoid lines demonstrated that TAF6L expression can influence cellular sensitivity to cisplatin. By studying the mechanism directly in human 3D organoids, we showed that TAF6L impacts cell proliferation, with its knockdown preventing cells from recovering after cisplatin-induced cell death, while re-expression rescued this effect. Although TAF6L does not appear to directly regulate DNA damage-related genes, it still emerged as a significant hit, once again underscoring the robustness of our CRISPR screen in 3D organoids. Our data suggested that TAF6L may have additional, intriguing functions beyond its role as a TATA-box binding protein-associated factor in regulating transcription machinery. This was supported by its known role in embryonic stem cell self-renewal⁵⁴ and our findings on its importance in cisplatin sensitivity and maintenance of chromatin accessibility. While our findings clearly demonstrate that TAF6L influences cellular viability and proliferation following cisplatin treatment, the precise molecular mechanism underlying this effect remains to be elucidated and warrants further investigation. From a clinical perspective, targeting TAF6L may offer therapeutic strategies to pharmacologically enhance sensitivity to cisplatin.

We combined our organoid iCRISPRi and iCRISPRa systems with single-cell RNA-sequencing to deeply explore individual sgRNA resistance phenotypes. Interestingly, single cells with cisplatin-sensitizing sgRNAs displayed divergent transcriptomes in the absence of cisplatin, but when exposed to cisplatin, they clustered according to known functions in NER and HRR DNA repair pathways. This type of unbiased, high-dimensional gene expression clustering could be used in the future to classify the functional role of specific genes within particular pathways. Beyond DNA repair, our single-cell CRISPR screens in 3D organoids further uncovered a previously unrecognized link between cisplatin sensitivity and fucosylation. We found that cisplatin, when combined with the knockdown of several NER genes, synergistically suppresses GMDS, which encodes a rate-limiting enzyme in GDP-fucose biosynthesis, leading to a marked reduction in total cellular fucosylated proteins and suggesting a functional connection between DNA repair and fucosylation. Interestingly, re-expressing GMDS did not rescue cells from cisplatin-induced cytotoxicity; instead, it enhanced cell death, indicating that GMDS/fucosylation suppression might serve a protective role in the context of cisplatin treatment. Supporting this hypothesis, pharmacological inhibition of fucosylation further sensitized cells to cisplatin, validating the functional relevance of this pathway in modulating drug response. These unexpected results suggest a potential crosstalk between DNA repair and fucosylation, highlighting the need for further investigation into its mechanistic basis. Future studies should explore the implications of GMDS/fucosylation modulation in cancer progression, chemoresistance to various genotoxic and non-genotoxic agents, and immune surveillance, potentially uncovering therapeutic opportunities.

Our study has certain limitations. First, while 2D cell line screens are more economical, 3D screens incur higher costs, and implementing complex CRISPR/Cas9 genome editing technologies in primary human organoids is technically more demanding across multiple biological replicates. Additionally, organoids naturally retain a heterogeneous cell population, and shifts in cell proportions can occur due to variations in culture conditions. Variability may occur between organoid lines due to genetic and cellular heterogeneity, as well as differences in the regional origins of the organoids⁵⁶. Moreover, direct comparisons between CRISPR screens conducted in different systems (e.g., 2D vs. 3D, or across distinct gene libraries and analysis pipelines) may not be appropriate, as variations in cell type, experimental conditions, and statistical thresholds can all influence the identification of true hits. These factors necessitate careful design of functional assays and interpretation of data. The application of single-cell CRISPR screens may help mitigate some of these issues. In addition, CRISPRa-based screens are inherently less reproducible than CRISPRi or knockout screens⁴³. These known technical limitations likely contributed to the lower correlation observed in our CRISPRa replicates. Nonetheless, the identification of biologically relevant hits, such as MYC, supports the potential utility of CRISPRa when cautiously interpreted. Finally, our oncogene-engineered human tumor organoids represent a pre-cancerous stage^9,37, meaning the gene-specific dependencies and biological processes identified may be distinct to early cancer development as opposed to established tumors.

Overall, our studies underscore the power of high-resolution approaches that link genotypes and phenotypes in human 3D organoids, enabling the mechanistic and unbiased dissection of gene-drug interactions. CRISPR screening platforms in CRISPR/Cas9-engineered human organoids could thus serve as a template for the future exploration of multiplexed genetic interactions in primary 3D cultures. Although our current study employed oncogene-engineered tumor organoids, similar studies could be performed using patient-derived organoids and induced pluripotent stem cell (iPSC)-derived organoids or extended to compact genome-scale sgRNA libraries. Furthermore, the same research approach could also be applied to explore the effects of other anticancer drugs beyond cisplatin. Our study provides a valuable reference for future applications of large-scale chemical profiling or genome-wide CRISPR screens in primary human organoids. The application of organoid-based functional genomics to systematically evaluate mechanistic pathways should not only enhance our understanding of fundamental biological processes but also provide an avenue for discovering drug-target interactions.

Methods

Generation of engineered human gastric tumor organoids

Wild-type human gastric organoids were derived from surgical discards under an approved IRB protocol at Stanford University and cultured in standard submerged Matrigel conditions as previously described⁹ and as elaborated below. The TP53 and APC genes were sequentially disrupted by CRISPR/Cas9 by lipofectamine transient transfection (Thermo Fisher Scientific, #11668019) of an all-in-one construct (Addgene, #42230) expressing both Cas9 and sgRNA. TP54 exon 4 and APC exon 15 were targeted by the following sgRNA sequences: sgTP53 5′-GGGCAGCTACGGTTTCCGTC-3′ and sgAPC 5′-GTTTGAGCTGTTTGAGGAGG-3′. A nutlin-3 functional selection was applied for selecting TP53 KO organoids. APC KO organoids were selected by growing organoids in a medium without Wnt/R-spondin. After clonal expansion, genomic DNA was extracted by DNeasy Blood & Tissue Kits (QIAGEN, #69506). Sanger sequencing was utilized to determine Cas9-induced indels. The TAF6L-GFP lentiviral construct was purchased from Origene (Origene #RC220716L4). The Tet-3G inducible CRISPRi and CRISPRa constructs were gifts from Dr. Stanley Qi (Stanford University). We purchased the Cas9 (Addgene, #52962) and the rtTA3 lentiviral (Addgene, #26429) constructs from Addgene.

Organoid culture media

The organoid culture medium was prepared using Advanced DMEM/F-12 (Thermo Fisher Scientific, #12634028) supplemented with 0.5% Penicillin/Streptomycin/Glutamine (Thermo Fisher Scientific, #10378016), 5% fetal bovine serum (FBS), 10 mM HEPES (Thermo Fisher Scientific, #15630080), 1 mM N-acetylcysteine (Sigma, A9165), 1X N2-Max Supplement (R&D Systems, #AR008)and 1X N21-Max Supplement (R&D Systems, #AR009). Additional components included 1 μM A83-01 (Tocris Bioscience, #2939), 1X GlutaMAX (Thermo Fisher Scientific, #35050061), 10 μM SB-202190 (Biogems, #1523072), 50 ng/mL EGF (PeproTech, AF-100-15), 100 μg/mL Normocin (InvivoGen, ant-nr-1), 10 nM gastrin (Sigma, G9145), and 50% Wnt-3A/R-spondin/Noggin conditioned medium. For iCRISPRi and iCRISPRa induction, 500 ng/mL doxycycline (Sigma, #D3072) was included in the organoid culture media.

Generation of lentivirus

Lentiviral particles were produced by co-transfecting 293T cells with the lentiviral plasmids, the packaging plasmid psPAX2 (Addgene, #12260), and the envelope plasmid pCMV-VSV-G (Addgene, #8454) using Lipofectamine 2000 (Invitrogen, #11668-019) according to the manufacturer’s protocol. Viral supernatants were harvested at 48 and 72 h after transfection and concentrated with PEG-it Virus Precipitation Solution (System Biosciences, LV825A-1). The resulting lentiviral particles were collected by centrifugation at 1500 × g for 30 min at 4 °C and resuspended in organoid culture medium supplemented with 10 μM Y-27632 (PeproTech, #1293823).

Lentiviral transduction of organoids

Organoids were washed with PBS and dissociated into small clusters using TrypLE™ (Invitrogen, #12604-012) at 37 °C for 15–20 min. The resulting clusters were resuspended in a transduction mixture containing organoid culture medium supplemented with 10 μM Y-27632 (PeproTech, #1293823), 8 μg/mL polybrene (Sigma, #107689), and concentrated lentivirus. Spinoculation was carried out by centrifuging the suspension at 600 × g for 1 h at 32 °C. Following spinoculation, organoids were incubated at 37 °C for 12–14 h, then embedded in Matrigel and replated in a new 24-well plate. For CRISPR screens, the multiplicity of infection (MOI) was adjusted to ~25–30% to minimize the likelihood of multiple viral integrations per cell.

Immunoblotting

Western blot analyses were performed using standard protocols. Briefly, cell pellets were lysed in RIPA buffer (150 mM NaCl, 1% Nonidet P-40, 0.5% deoxycholate, 0.1% SDS, 50 mM Tris-HCl, pH 7.5) supplemented with protease inhibitor cocktail (Roche, #04-693-124-001) and phosphatase inhibitor cocktail (Sigma, P5726). Protein concentrations were determined with the BCA assay kit (Thermo Fisher Scientific, #23227). Lysates were resolved by SDS–PAGE (Invitrogen, #NP0323) alongside the PageRuler Plus Prestained Protein Ladder (Thermo Fisher Scientific, #26619) and transferred onto PVDF membranes (Millipore, IPVH00010). Membranes were blocked with 5% non-fat dry milk in 1 × TBS (pH 7.4, Quality Biological, #351-086-151) containing 0.05% Tween-20, then incubated overnight at 4 °C with primary antibodies. Bound antibodies were detected by chemiluminescence (Thermo Fisher Scientific, #34580) after incubation with horseradish peroxidase-conjugated secondary antibodies and exposure to AccuRay Blue X-ray film (E&K Scientific, #EK5129). Antibodies were purchased from the following vendors: SOX2 (Abcam, #ab97959, dilution 1:1000), TAF6L (Proteintech, # 15745-1-AP, dilution 1:1000), CAS9 (Santa Cruz, #sc-517386, dilution 1:1000, dilution 1:1000), β-actin (Thermo Fisher Scientific, #MA1-140, dilution 1:5000), RAD51 (Santa Cruz, sc-398587, dilution 1:500) and FANCD2 (Abcam, ab108928, dilution 1:500), and GAPDH (Cell Signaling, #5174, dilution 1:5000).

Independent validation of cisplatin sensitivity

For 12-point full titration treatment of cisplatin, a total of 5000 cells were resuspended into 10 μL Matrigel and cultured in a well of a 96-well plate for 5 days before drug treatment. For each drug dose, 4 technical replicates (4 wells) were included. The expression of dCas9 fusion proteins was induced by adding 500 ng/mL doxycycline (Sigma, #D3072) in the organoid culture media described above. Cell viability was quantified 3 days after cisplatin treatment. For the cell viability assay, AlamarBlue^TM Cell Viability Reagent (Invitrogen, #DAL1100) in organoid culture media was added into the plate and incubated with organoids for 4 h before being quantified using a Synergy H1 Hybrid Multi-mode Plate Reader (BioTek). Each data point in the IC50 curve has been normalized to the no-drug condition of control sgRNA for each sgRNA target. The IC50 was determined by GraphPad (https://www.graphpad.com/guides/prism/latest/curve-fitting/reg_dr_inhibit_variable.htm).

Total RNA extraction and real-time quantitative PCR

Total RNA from organoids was isolated using the RNeasy Kit (Qiagen, #74106). On-column DNase digestion (Qiagen, #79254) was performed to remove genomic DNA. A total of 0.5–1 μg of RNA was used to synthesize complementary DNA (cDNA) with the iScript Reverse Transcription Supermix (Bio-Rad, #1708841). Quantitative PCR was then performed using Power SYBR Green PCR Master Mix (Thermo Fisher Scientific, #4368708). The forward primer, 5′-CGGCGGTTTGTGGAGATCC-3′, and the reverse primer, 5′-CCTCTCTCAGACGATAGCACAC-3′, were used for TAF6L quantitative PCR.

CRISPR libraries and screens

All screens (CRISPR KO, iCRISPRi, and iCRISPRa) were conducted as two completely independent experiments, involving separate lentiviral transductions performed at different times. For the pilot CRISPR KO screen, the sgRNA library targeting membrane proteins was a gift from Dr. Michael Bassik (Stanford University). For the iCRISPRi and iCRISPRa screens, a sgRNA library was manually curated and designed to target DNA and/or RNA-binding proteins. To select the optimal sgRNAs targeting each gene, we relied on our previously published hCRISPRi and hCRISPRa v2 libraries and screens⁵⁷. Oligonucleotide pools were designed with flanking PCR and restriction sites (BstXI and BlpI), synthesized by Twist Biosciences, and cloned into the sgRNA expression vector pCRISPRia-v2 (Addgene, #84832), as described previously⁵⁷. The expression vector contains a U6 promoter driving the sgRNA expression, as well as an EF1α promoter driving puromycin-T2A-BFP.

For CRISPR screens, (d)Cas9-expressing organoids were washed by PBS and mechanically dissociated into smaller pieces by pipetting and resuspension in TrypLE^TM (Invitrogen, #12604-012) plus 10 μM Y-27632 (Peprotech, #1293823) at 37 °C for 15–20 min. After incubation, fetal bovine serum (FBS) was added to quench TrypLE^TM activity. Organoids were then centrifuged at 600 × g for 5 min and washed once using organoid culture media before cell counting. Approximately 4 × 10⁷ suspension (d)Cas9-expressing cells were used for lentiviral transduction with the sgRNA library (~12,500 sgRNAs), corresponding to ~3000× coverage. Lentivirus was added at an infection rate of ~30%, resulting in an effective post-transduction coverage of ≥1000×. Lentiviral transduction of organoids was performed as previously described above. Following transduction, organoids were centrifuged at 600 × g for 5 min and resuspended in Matrigel at a density of 25,000–30,000 cells per 40 μL Matrigel. Transduced organoids/Matrigel domes (40 μL each dome) were plated onto new pre-warmed cell culture plates. After Matrigel polymerization, organoid culture media (described above) were added to each well. Three days after sgRNA library transduction, sgRNA-expressing cells were selected by 2 μg/mL puromycin (Thermo Fisher Scientific, #A111380-03) for 3 days. More than 90% of sgRNA library-transduced cells became sgRNA positive after puromycin selection. Organoids were recovered from the selection for another 2 days. The genomic DNA was harvested from each sample at the initial time point (T0) and the end time point (T1) with 1000× coverage of the sgRNA library for deep DNA sequencing. For passaging, organoids were pooled and dissociated in bulk into single cells. After cell counting, cells were resuspended into fresh Matrigel with 1000× coverage of the library (cell density: 25,000–30,000 cells/40 μL Matrigel) and replated.

DNA libraries of T0 and T1 were prepared for deep sequencing essentially as previously described^57,58. Briefly, genomic DNA was isolated using a NucleoSpin Blood XL kit (Macherey–Nagel). Then, isolated gDNA was directly amplified by 23 cycles of PCR using NEBNext Ultra II Q5 PCR Master Mix (NEB), appending Illumina adapters and unique sample indices. The libraries were then sequenced at high coverage on a HiSeq 4000 (Illumina) using single-end 50 base pair reads, as previously described⁵⁷.

Sequencing reads were aligned, counted, and quantified using the Python-based ScreenProcessing pipeline (https://github.com/mhorlbeck/ScreenProcessing)⁵⁷. The generation of negative control genes and calculation of phenotypes and Mann-Whitney p-values were performed as described previously^28,57. Drug sensitivity phenotypes were calculated by calculating the log2 change in the enrichment of a sgRNA in the treated and untreated samples, subtracting the equivalent median value for all non-targeting sgRNAs, and dividing by the number of population doubling differences between the treated and untreated populations²⁸. Similarly, untreated growth phenotypes (γ) were calculated from the untreated and t0 samples, divided by the total number of doublings of the untreated population. Phenotypes from sgRNAs targeting the same gene were collapsed into a single sensitivity phenotype for each gene using the average of the top three scoring sgRNAs (by absolute value, labeled as “ave phenotype of top 3 sgRNAs” in the Supplementary Data) and assigned a p-value using the Mann-Whitney test of all sgRNAs targeting the same gene compared to the negative controls. To call gene hits from the screens, a “screen score” was defined as | γ z-score from negative control gene distribution | × –log₁₀ p-value.

All additional CRISPR screen data analyses and plots were generated in Python 2.7 with the libraries NumPy (v1.12.1), Pandas (v0.17.1), SciPy (v0.17.0), and scikit-learn (v0.19.1). Gene Ontology analysis was performed using DAVID (v6.8).

Perturb-seq library and screens

Perturb-seq sgRNAs were individually cloned by annealing complementary synthetic oligonucleotides (Integrated DNA Technologies) containing flanking BstXI and BlpI restriction sites, and ligating the resulting duplex into a modified CROP-seq vector backbone³⁴. The CROP-seq backbone, a gift from Christoph Bock (Addgene #86708), was modified to match the pCRISPRia-v2 vector used in the CRISPR screens by introducing the mouse U6 promoter, BstXI and BlpI restriction sites, and an optimized sgRNA constant region. Additionally, the EF1α-driven puromycin resistance cassette was replaced with BFP.

The Perturb-seq sgRNA library was pooled and packaged into a lentiviral library by transfecting HEK293T cells with TransIT-LT1 (Mirus). Organoids were infected with the lentiviral library at an initial infection rate of ~10%, ensuring 1000× coverage of the library. Following infection, organoids were treated with 500 ng/mL doxycycline (Sigma, #D3072) and cultured for 2 days before FACS sorting (BD FACS Aria2) to isolate the mCherry⁺/BFP⁺ population. At 7 days post-infection, cells were harvested and processed for single-cell RNA-seq using the 10X Chromium Controller and Chromium Single Cell 3′ Library & Gel Bead Kit v2 (10x Genomics). sgRNA sequences from the final library were specifically amplified by PCR, as previously described⁵⁸, using the following primers (synthesized by Integrated DNA Technologies, IDT):

5′-AATGATACGGCGACCACCGAGATCTACAC-3′

5′-CAAGCAGAAGACGGCATACGAGATCAGCCTCGGTCTCGTGGGCTCGGAGATGTGTATAAGAGACAGgtgttttgagactataaGtatcccttggagaaCCAcctTGTTG-3′

The sequence in bold represents the i7 index. PCR amplification was performed using the following cycling conditions: (1) 95 °C for 3 min; (2) 15 cycles of 98 °C for 15 s and 70 °C for 10 s; (3) 72 °C for 1 min. The resulting sgRNA barcode library was purified using a 0.8× SPRI bead selection, followed by additional size selection with BluePippin (Sage Science). The sgRNA barcode libraries were then sequenced as 5% spike-ins alongside the parent RNA-seq libraries on a NovaSeq6000, according to the manufacturer’s instructions.

Perturb-seq analysis

Cell Ranger (version 2.1.1, 10X Genomics) with default parameters was used to align reads and generate digital expression matrices from single-cell sequencing data. Cell sgRNA identity assignments were processed using custom Python scripts described previously (https://github.com/thomasmaxwellnorman/perturbseq_demo)^32,52. The specifically amplified sgRNA reads were aligned to a library of expected sgRNA sequences using Bowtie (flags:-v2-q-m1). Reads sharing the same cell barcode (CBC), UMI, and read identity (including some unmapped by Bowtie due to low quality) were collapsed to generate a table listing possible sgRNA identities for each cell, along with the number of supporting reads and UMIs. The coverage of each proposed sgRNA identity was defined as the number of reads divided by the number of UMIs. The coverage distribution was bimodal. A proposed identity was considered to have good coverage if it (1) fell within the upper mode of the distribution (above a defined threshold), (2) was supported by at least 50 raw reads, and (3) was supported by at least 3 UMIs. Cells with a single identity meeting these criteria were assigned that sgRNA. Cells with two or more such identities were classified as multiple (likely due to multiple infection, PCR artifact, or multiple encapsulation during emulsion generation). Cells with no identities meeting the criteria were classified as unidentifiable.

Custom Python scripts (https://github.com/thomasmaxwellnorman/perturbseq_demo), as previously described, were used to analyze the digital expression matrices, including normalization, quality control, and filtering. Normalized expression data for each cell were generated by (1) equalizing UMI counts across cells and (2) z-normalizing gene expression relative to the mean and standard deviation observed in untreated control cells carrying non-targeting sgRNAs. sgRNAs represented by fewer than 10 cells were excluded from the analysis.

Mean normalized gene expression profiles were calculated by averaging across all cells harboring the same gene perturbation. For hierarchical clustering, gene–gene Spearman’s rank correlation matrices were computed, and clusters were optimally ordered to minimize the distance between successive leaves. Heatmaps of the correlation matrices were generated for visualization, with genes ordered using Ward’s method of hierarchical clustering.

Cell cycle positions for individual cells were inferred using scores derived from panel markers specific to each cell cycle stage, as previously described³². The relative occupancy of cells in each stage was then computed for each perturbation. Heatmaps of deviations display the percentage change in each stage relative to control cells bearing non-targeting sgRNAs.

To assess differences in gene expression between perturbed cell populations, we applied a random forest classifier. This approach is based on the principle that genes important for a given perturbation can accurately predict its identity. Differentially expressed genes were identified by training a random forest classifier to distinguish each gene perturbation from negative control sgRNAs, using each cell as a training data point and its gene expression profile as input. Specifically, we used the randomized trees implementation in scikit-learn, typically with 1000 trees per forest. A two-stage fitting process was performed: first, 20% of the cells were set aside as a test set, and the remaining 80% were used to train a random forest classifier (with 1000 estimators) to predict the perturbation identity based on the normalized expression profiles. All genes with mean expression >0.25 UMI per cell were included as candidate features. During training, the random forest assigned feature importances reflecting their predictive value. The top 100 genes, ranked by importance, were selected as the most informative features. To assess their predictive power, the classifier was retrained using only these top 100 genes and evaluated on the 20% test set. The advantage of this approach is that it scales trivially to comparisons among more than two perturbations, which is critical in our case of analyzing gene-drug interactions.

Gene ontology analysis of the gene expression was performed using Gene Set Enrichment Analysis (GSEA) with GSEAPY in Python 2.7. All additional analyses and plotting were performed in Python 2.7 using a combination of Numpy (v1.12.1), Pandas (v0.17.1), Scipy (v0.17.0), scikit-learn (v0.19.1), and Seaborn.

ATAC-seq assay

ATAC-seq assay and libraries were prepared at the MD Anderson Epigenomics Profiling Core following the protocol as previously described^59,60 with minor modifications. Briefly, an equal number of 50,000 viable single cells per technical replicate from TAF6L knockdown, TAF6L-GFP, and their control organoid lines were sorted and incubated with Tagment DNA TDE1 enzyme (Illumina), followed by DNA purification and library preparation. The resulting libraries were sequenced 2 × 75 bp on an Illumina NextSeq500. For ATAC-seq data processing, we followed the guidelines from the Galaxy Training Materials ATAC-seq tutorial (https://training.galaxyproject.org/training-material/topics/epigenetics/tutorials/atac-seq/tutorial.html). Briefly, adapter sequences were removed using Cutadapt, and filtered reads were aligned to the hg38 reference genome using Bowtie2. Picard was used to remove duplicated reads (https://broadinstitute.github.io/picard/) and reads with a quality score below 30 were filtered out. To verify data quality, the insert size distribution was examined to confirm peaks around 50 bp and 200 bp. Peak calling was performed using MACS2 callpeak (Galaxy Version 2.2.9.1+galaxy0) with the following parameters: –shift −100, –extsize 200 (to account for Tn5 transposase-specific 9 bp insertions), –gsize 2.7e9, –nomodel, and –qvalue 0.05. For identifying differentially accessible peaks, two BED files (treatment and control groups) were analyzed using MACS2 with the same cutoff (–qvalue 0.05). Finally, peaks were converted to BigWig format for visualization using IGV and downstream analyses. For the enrichment analysis of ATAC-seq peak annotations, the expected genomic distribution was calculated based on the human reference genome (hg38), representing the baseline proportions of various genomic regions. The observed distribution of ATAC-seq peaks in the TAF6L knockdown (TAF6L KD) and overexpression (TAF6L OE) groups was obtained from the annotation results described in the ATAC-seq analysis section. To assess enrichment, the fold change was calculated for each genomic region by dividing the observed distribution by the expected distribution (Obs/Exp).

Immunofluorescent staining

Immunofluorescent staining of whole-mount organoids was performed as described. In brief, remove the medium and wash twice with 1X PBS. Fixed with 4% PFA (Electron Microscopy Sciences, #15714-S) for 30 min at RT. Quench with 1X PBS:Glycine buffer with gentle shaking for 30 min. Block the sample for 1 h at RT with 10% Normal Donkey Serum (Jackson ImmunoResearch, #AB_2337258) diluted in IF buffer (1.3 M NaCl, 132 M Na₂HPO₄, 34.5 mM NaH₂PO₄, 77 mM NaN3, 1% BSA, 2% Triton X-100, 1% Tween-20). Incubate primary antibodies, pH2A.X (Cell Signaling, #9718T, dilution 1:500) and Histone H3 (Cell Signaling, #14269S, dilution 1:1000), overnight at 4 °C in IF buffer. Wash 3 times with PBS and incubate with DAPI (Invitrogen, #D3571) and correlated secondary antibodies Alexa Fluor® 594 AffiniPure Donkey Anti-Mouse (Jackson ImmunoResearch, #715-585-150) and Alexa Fluor® 647 AffiniPure Donkey Anti-Rabbit (Jackson ImmunoResearch, #711-605-152) for 4 h at RT. Wash 3 times with PBS and mount the samples with mounting buffer (VECTASHIELD antifade mounting medium, H-1000). Images were acquired with confocal microscopy (ZEISS, #LSM 710). Quantification of the mean intensity of pH2A.X fluorescent signals is proceeded through QuPath software. The nucleus parameters were set as 25 µm background radius, 0 µm median filter radius, 1 µm sigma, 1 µm² minimum area, and 1000 µm² maximum area. Intensity threshold was set as 25.

EdU assay by flow cytometry

CRISPR/Cas9-engineered human gastric organoids expressing BFP, GFP, or mCherry reporters were treated with 1 μM EdU (Thermo Fisher Scientific, #C10634) for 2 h. Organoids were dissociated into single cells with TrypLE Express (Gibco, #12604013) at 37 °C for 15 min and washed with 1% BSA (Sigma-Aldrich, #A7906) in PBS. EdU incorporation was detected using the Click-iT EdU Alexa Fluor 647 Kit (Thermo Fisher Scientific, #C10634) per manufacturer’s instructions, via a copper-catalyzed click reaction. Cells were resuspended in PBS with 2% FBS (Gibco, #26140079) and analyzed on a CytoFLEX flow cytometer (Beckman Coulter). Reporter fluorescence was gated (BFP: 450/45 nm; GFP: 525/40 nm; mCherry: 610/20 nm), followed by EdU-AF647 detection (660/20 nm). Proliferation was quantified in FlowJo v10 (BD Biosciences) as the percentage of EdU-AF647-positive cells relative to parental controls, averaged across three biological replicates.

Statistics and reproducibility

All statistical details, including the definition of n and the type of replicates, were provided in the figure legends. Representative images were shown in Figs. 1A, B, F, 2B–E and 5D–F. Experiments shown in Figs. 1A, B, 2B, 5F, 6E, S2B, S7F and S8B, C were independently repeated at least three times with similar results. Experiments shown in Figs. 1F, 2C–E, 5D, E, H, 6A, B, D, S4A, B, S7C, G and S8A were independently repeated at least twice with similar results. Immunoblots shown in Figs. 1A, 5E, F, 6C and S1A were independently repeated at least twice with similar results. No statistical method was used to predetermine sample size. No data were excluded from the analyses. The experiments performed in this study were randomized. The Investigators were not blinded to allocation during experiments and outcome assessment.

Reporting summary

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

Author contributions

Y.-H.L. conceived this study. Y.-H.L. and J.C. designed experiments. Y.-H.L. established and validated engineered organoid lines and performed CRISPR screens. J.C. generated sgRNA libraries for iCRISPRi and iCRISPRa screens and analyzed CRISPR screen data. Y.-H.L. and H.T.H. performed drug sensitivity validation. M.-F.H analyzed ATAC-seq and RNA-seq data. Y.-H.L. and Martina Towers generated lentivirus. Madeline Tomaske, Martina Towers and W.-C.Y. generated organoid culture media. Y.-H.L., W.-C.Y., C.-M.Y. and Q.L. performed functional assays related to TAF6L and GMDS. A.D. and L.S.Q. provided inducible CRISPRi and inducible CRISPRa constructs. M.C.B. provided sgRNA libraries for CRISPR KO screens. Y.-H.L., J.C., D.-F.L., C.J.K. and J.S.W. supervised this study and acquired funding. Y.-H.L., J.C., and C.J.K. wrote and edited the manuscript.

Peer review

Peer review information

Nature Communications thanks the anonymous reviewers for their contribution to the peer review of this work. A peer review file is available.

Data availability

The datasets generated in this study are publicly accessible. The RNA-seq data generated in this study have been deposited in the Gene Expression Omnibus (GEO) under accession code GSE280256. The ATAC-seq data generated in this study have been deposited in the GEO under accession code GSE280257. The scRNA-seq data generated in this study have been deposited in the GEO under accession code GSE280506. Source data are provided with this paper.

Code availability

The code used to perform the analyses reported in this study can be found on GitHub (https://github.com/jinpjchen/organoid_GI) and archived with a DOI at Zenodo (DOI:10.5281/zenodo.15767754).

Competing interests

C.J.K. declares outside interest in Surrozen, Inc., Mozart Therapeutics, and NextVivo, Inc. J.S.W. declares outside interest in KSQ Therapeutics, Maze Therapeutics, Chroma Medicine, Amgen, Tessera Therapeutics, 5AM Ventures, and Third Rock Ventures. J.S.W. has filed patent applications related to CRISPRi and CRISPRa screening. J.C. is an employee and shareholder of Altos Labs. The remaining authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41467-025-62818-3.