CRISPR-StAR enables high-resolution genetic screening in complex in vivo models

Abstract

Pooled genetic screening with CRISPR–Cas9 has enabled genome-wide, high-resolution mapping of genes to phenotypes, but assessing the effect of a given genetic perturbation requires evaluation of each single guide RNA (sgRNA) in hundreds of cells to counter stochastic genetic drift and obtain robust results. However, resolution is limited in complex, heterogeneous models, such as organoids or tumors transplanted into mice, because achieving sufficient representation requires impractical scaling. This is due to bottleneck effects and biological heterogeneity of cell populations. Here we introduce CRISPR-StAR, a screening method that uses internal controls generated by activating sgRNAs in only half the progeny of each cell subsequent to re-expansion of the cell clone. Our method overcomes both intrinsic and extrinsic heterogeneity as well as genetic drift in bottlenecks by generating clonal, single-cell-derived intrinsic controls. We use CRISPR-StAR to identify in-vivo-specific genetic dependencies in a genome-wide screen in mouse melanoma. Benchmarking against conventional screening demonstrates the improved data quality provided by this technology.

Pooled genetic CRISPR screening is improved in complex models through leveraging internal controls.

Main

In vitro functional genomic screening, in particular with CRISPR–Cas9, has uncovered many essential genes in cell types such as cancer cells¹. However, this approach has yielded few useful therapeutic targets. One reason is the limitation of two-dimensional (2D) cell culture as a model for disease. For example, cancer cell lines readily proliferate by a factor of 10³ per week in vitro, whereas a tumor volume doubling typically occurs within timescales of months^2,3, resulting in different sets of essential or growth-limiting gene sets. Moreover, cancer driver genes often do not fully exhibit cancer-causing effects in vitro^4–6, likely due to the lack of certain tumor physiological conditions, such as limited space, nutrients and oxygen, an acidic milieu as well as various interactions with the extracellular matrix, immune system and endothelial and stromal cells that together shape the development of a tumor. In contrast, the effects of cancer driver genes are more easily revealed in vivo, especially in the presence of an intact immune system⁵. Thus, therapeutic targets often become manifest in vivo but not in vitro^7,8, underlining the importance of studying gene function in vivo.

To date, high-throughput genetic screening in vivo has revealed physiologically relevant novel targets only in a few easily transplantable and highly transformed models or has been limited to small libraries of perturbants^4,7,9–12. One major caveat has been excessive experimental noise, because engraftment of cancer cells typically results in low survival rates and heterogeneous growth. Although genetic experiments would ideally embrace this diversity of tumors, in practice it impedes genetic screening due to introduction of excessive noise. To overcome this problem, we developed a screening paradigm called CRISPR-StAR (Stochastic Activation by Recombination), which is based on Cre-inducible single guide RNA (sgRNA) expression¹³ and single-cell barcoding¹⁴. This method introduces an internal control on a single-cell level to overcome the noise concomitant to complexity bottlenecks and clonal diversity in heterogeneous screening scenarios. Benchmarking CRISPR-StAR to conventional CRISPR screening in vivo resulted in greatly improved accuracy in hit calling. We also discovered in-vivo-specific genetic dependencies in Braf-targeted therapy-resistant melanoma, thereby highlighting the relevance of in vivo functional genetics for identification of potential novel drug targets.

Results

CRISPR-StAR internally controls for growth heterogeneity

Target discovery using CRISPR–Cas9 screening typically requires a coverage of 500–1,000 cells per sgRNA or approximately 0.5–1 × 10⁸ cells per genome-wide library to fend off stochastic noises^15–17. Although this can be easily achieved in vitro, it is almost impossible in vivo due to significant bottlenecks in cell survival during engraftment, resulting in random sampling of the sgRNA library^18–20. Screening in tissues or in organoids²¹ is likewise confounded by transduction efficiency as well as heterogeneity in clonal cell outgrowth kinetics (Fig. 1a). Together, this results in introduction of noise, preventing meaningful comparison of sgRNA population in the plasmid library relative to the established tumor. In the present study, we measured engraftment and growth heterogeneity of various cell lines using genome-wide CRISPR libraries carrying unique molecular identifiers (UMIs)¹⁴ as single-cell barcodes marking clonal progenitor populations. Depending on the cell line, we detected approximately 4,800–20,500 barcodes upon injecting up to 1 million tumor cells, indicating that only a few thousand tumor cells engrafted and contributed to the resulting tumor (Fig. 1b). This is 5–30-fold less than the number of sgRNAs in a typical genome-wide library, thus rendering screening results stochastic even when multiple animals are used. Although injection of more cells had little effect on total engraftment (Extended Data Fig. 1a), immune suppression by genetic means or through depletion of CD8 T cells enhanced barcode recovery (Extended Data Fig. 1b), especially when cells harbored neoepitopes such as Cas9 (Extended Data Fig. 1c–e). Moreover, our findings reveal that engrafted cells display marked variability in their contribution to the total tumor mass as measured by read counts per barcode. Most barcodes recovered exhibited low representation, accounting for 1–5% of the total reads (Fig. 1b, orange, green and light blue), whereas a select few barcodes demonstrated remarkably high read abundance. Notably, 50% of tumor mass comprised only 22–536 barcodes (Fig. 1b, black). This observation of highly skewed clonal expansion dynamics within the engrafted cell population, with very few clones exhibiting exceptional proliferative capacity, is in alignment with previous lineage-tracing results²². Such distribution of sgRNA abundance across 3–5 log scales reflects the intricacies of tumor biology, the heterogeneity of engraftment and clonal outgrowth and represents experimental noises. In contrast, typical gene effect sizes observed in in vitro pooled genetic screens range within 20-fold enrichment/depletion for tumor suppressor (TS) genes or essential genes, respectively. Therefore, the experimental noise significantly surpasses anticipated signals from gene depletion/enrichment, posing a substantial challenge for high-resolution genetic screening in vivo. For that reason, in vivo studies are typically limited to small gene sets and/or particularly well-engrafting cell line models^9,23–27.

Fig. 1. CRISPR-StAR overcomes experimental noise.

a, Schematic of allograft in vivo CRISPR screening workflow, whereby cells with Cas9 receive the viral sgRNA library. This population can be screened in vitro or in vivo. At the point of in vivo injection, a small random sample of cells engrafts to form a tumor. Most cells and sgRNAs drop out of the screening population. Engrafted cells grow out very heterogeneously. The graph depicts this drop in sgRNA representation and growth heterogeneity to illustrate the challenges when comparing the sgRNA population in established tumors relative to the starting population. b, Average number of barcoded mural cancer cells that were engrafted in immunodeficient (Rag2^−/−) mice. Injected cells range between approximately 1 × 10⁵ and 1 × 10⁶ cells per mouse, two mice per cell line. Colors indicate the percentage of NGS reads making up for the engrafted barcodes. c, Conceptional outline of screening with active (purple) and inactive (magenta) sgRNAs that serve as internal control. d, Schematic illustration of the CRISPR-StAR construct. CreERT2 recombination leads to either an active or an inactive sgRNA construct. e, Correlation of the log₂FC of two biological replicates in high (1,024 cells per sgRNA) and low (four cells per sgRNA) complexity using conventional analysis (active sgRNA versus sgRNA before limiting dilution). Each sgRNA is represented by a dot. In red are sgRNAs that target essential genes; in black are sgRNAs that target non-essential genes; and in gray are sgRNAs targeting genes that are not specified in the previous groups (‘other’). R is given for the regression correlation line. f, Correlation of the log₂FC of two biological replicates in high complexities using CRISPR-StAR analysis (active sgRNA versus inactive sgRNA). g, Bar plot of R for each limited dilution condition and analysis method. h, Values for AUROC for each limited dilution condition and analysis method in duplicate. Random bar represents a random distribution of the CRISPR-StAR data for each limited dilution.

Source data

Extended Data Fig. 1. Engraftment efficiency and heterogeneity of various injected cancer cell lines.

a-e) Bar plots for the average number of barcoded mural cancer cells that were engrafted in the mouse. Injected cells range between ~1*10⁵ and 1*10⁶ cells, per cell line and condition two mice were injected. Colors indicate the percentage of NGS reads making up for the engrafted barcodes. a) For each cell line two conditions with different numbers of injected cells are shown. b) Number of engrafted 4T1 or Yumm450R cells in wildtype mice (BALB/c or C57BL/6), immunodeficient (Rag2^-/-) mice or wildtype mice with a CD8 T-cells depletion. c) 4T1 Cas9-CreERT2 clonal cell line (Clone 1) engraftment rate in wildtype mice (BALB/c), immunodeficient (Rag2^-/-) mice or wildtype mice with a CD8 T-cells depletion. d) 4T1 Cas9-CreERT2 clonal cell line (Clone 2) engraftment rate in wildtype mice (BALB/c), immunodeficient (Rag2^-/-) mice or wildtype mice with a CD8 T-cells depletion. e) Yumm450R Cas9-CreERT2 clonal cell line engraftment rate in wildtype mice (C57BL/6), immunodeficient (Rag2^-/-) mice or wildtype mice with a CD8 T-cells depletion.

To overcome these limitations, we devised a method in which (1) the stochasticity of sgRNA loss during engraftment is overcome by initiation of the screen in cells that have survived the engraftment bottleneck; (2) each single-cell-derived clone is tracked and evaluated independently by means of a UMI¹⁴; and (3) the intrinsic (cell type) and extrinsic (microenvironment) heterogeneity of each single-cell-derived clone is controlled for by the generation of internal control (Fig. 1c). In such a scenario, each UMI-marked clone will comprise experimental cells carrying an active sgRNA intermingled with a corresponding wild-type (WT) population harboring the identical sgRNA and UMI in an inactive state, serving as internal control. Proximity to nutrients and oxygen would be equally controlled for as well as exposure to acidification or immune pressures. We previously developed CRISPR-Switch, a method to trigger activation of sgRNAs by excision of a floxed stop cassette placed at the apex of the repeat:antirepeat hairpin where spacer and tracr RNAs had been fused to generate sgRNAs^13,28. By placing an additional pair of intercalated lox5171 sites, which are incompatible with loxP²⁹, into the tandem repeat, we generated an inducible sgRNA construct with two distinct recombination outcomes. Cre activity would either excise the stop cassette to generate an active sgRNA or excise the tracr RNA and maintain an inactive state. Both recombination events occur in an irreversible and mutually exclusive manner (Fig. 1d). We hypothesized that induction of Cre::ERT2 recombinase by tamoxifen upon establishment of single-cell-derived clones—tracked by UMIs¹⁴—would fulfill all conceptual requirements for internally controlled screening, and we termed this method CRISPR-StAR.

To test this concept, we cloned a library containing 5,870 sgRNAs, targeting 1,245 genes into the first-generation CRISPR-StAR backbone and transduced mouse embryonic stem cells (mESCs) expressing Cas9 and Cre::ERT2 at a representation of more than 1,000 cells per sgRNA. Upon selection, cell complexity, or sgRNA coverage, was taken through artificial bottlenecks of ~1, ~4, ~16, ~64, ~256, and ~1,024 cells per sgRNA via limiting dilution in two replicates, re-expanded to more than 1,000 cells per sgRNA and induced by 4-OH tamoxifen (day 0) for recombination. Cells were harvested after 14 d, and the abundance of active and inactive sgRNAs within each clonal UMI population was quantified. Representation of active sgRNAs at day 14 was compared either to the representation before limiting dilution akin to a conventional analysis in a typical CRISPR–Cas9 screen or to the inactive internal UMI controls on day 14 within the CRISPR-StAR paradigm. Representation of sgRNAs targeting core essential genes separated well from neutral controls at high coverages of more than 256 cells per sgRNA (Fig. 1e,f and Extended Data Fig. 2). As anticipated, sgRNAs targeting core essential genes also depleted at low coverage. However, representation of neutral sgRNAs ranged more than 1,000-fold (log₂ fold change (FC) = 10), including complete depletion. The introduction of such noise confounded the separation of essential and non-essential genes by means of pooled screening as quantified using correlation analysis, differential area under the curve (dAUC) and area under the receiver operating characteristic curve (AUROC) (Fig. 1g,h and Extended Data Figs. 3 and 4). Consequently, conventional analysis dropped to very low reproducibility between experiments (Pearson correlation coefficient (R) of 0.07 for one cell per sgRNA). In stark contrast, reproducibility remained higher than 0.68 for all conditions in the CRISPR-StAR paradigm, clearly indicating superiority of CRISPR-StAR over conventional analysis, especially at low coverages.

Extended Data Fig. 2. CRISPR-StAR screening in complex system mimicked by artificial bottlenecks.

a) Schematic illustration of Screening setup, whereby an artificial bottleneck is introduced by diluting the sgRNA library containing cells to ~1, ~4, ~16, ~64, ~256, and ~1024 cells/sgRNA. CreERT2 recombination is induced at Day 0 and cells are maintained for 14 days after which sgRNA were read by NGS. Gene effect was calculated with the conventional analysis by comparison of end population versus starting population (before limiting dilution) or CRISPR-StAR analysis by comparison of active and inactive sgRNAs. b) Correlation of the log₂ fold change of two biological replicates from high to low complexity using conventional analysis (active sgRNA vs sgRNA prior to limiting dilution). Each sgRNA is represented by a dot, in red are sgRNAs that target essential genes, in black sgRNAs that target non-essential genes and in gray are sgRNAs targeting genes that are not specified in the previous groups (other). Pearson correlation coefficient (R) given for the regression correlation line and p-value by a two-sided testing for non-correlation. c) Correlation of the log₂ fold change of two biological replicates from high to low complexity using CRISPR-StAR analysis (active sgRNA vs inactive sgRNA).

Extended Data Fig. 3. Differential area under the curve (dAUC) analysis of CRISPR-StAR screening with artificial bottleneck.

a) Assessment of the recall of the sgRNAs targeting essential genes and non-essential genes, when sgRNAs were ranked according LFC as calculated with conventional analysis. Red line for sgRNAs targeting essential genes, black line for sgRNAs targeting non-essential genes. b) dAUC analysis when LFC was calculated with CRISPR-StAR.

Extended Data Fig. 4. Performance assessment of CRISPR-StAR screening with artificial bottleneck using Receiver Operating Characteristic (ROC) curves.

ROC curves for each limited dilution of ~1, ~4, ~16, ~64, ~256, and ~1024 cells/sgRNA and each replicate when using conventional analysis (blue), CRISPR-StAR (magenta) and random (grey) datasets. Comparing the true positive rate and the false positive rates of essential and non-essential genes.

Optimizing the CRISPR-StAR vector system

To ensure the robust readout of CRISPR-StAR, it is crucial to achieve a balanced ratio between the control populations, which harbor the sgRNAs in an inactive state, and the experimental populations harboring the active sgRNAs. Upon Cre-mediated induction, our initial vector design gave rise to 35–41% of cells with active sgRNAs within a UMI clone (Fig. 2a, vector 1; see Methods for details), limiting the dynamic range for depletion. Additionally, the polymerase chain reaction (PCR) strategy should exhibit minimal bias toward amplifying one conformation over the other. Through adjustments of the relative sequence context and distance between loxP or lox5171 sites, we refined the design to achieve a final ratio of 55–45% active to inactive sgRNAs for the StAR 4GN vector (GFP–neomycin) in vitro (Fig. 2a). This ratio was reproducible across three different cell lines in vitro and across different viral integration sites within each cell line and displayed minimal confirmation-specific PCR bias within 28 amplification cycles (Extended Data Fig. 5a). We also mitigated the risk of premature vector recombination by Cre::ERT2, which could occur during reverse transcription of the viral backbone in the cytoplasm, by placing the selection cassette in a location excised in both active and inactive conformations (Fig. 1d). This design also prevents amplification of unrecombined constructs by PCR due to the length of the amplicon. Collectively, these findings demonstrate that CRISPR-StAR consistently activates at a stable ratio across various cell lines, species and integration sites, underscoring its robustness for internally controlled screening applications.

Fig. 2. Optimizing the pooled screening setup in vivo with CRISPR-StAR.

a, In three independent cell lines transduced with a barcoded library, the percentages of inactive (magenta) and active (purple) UMIs for various CRISPR-StAR vectors were calculated after CreERT2 was induced and the cells went through a bottleneck of approximately 5,000 cells. StAR vectors 1–4 have a retroviral backbone, and 4B (blasticidin resistance), 4BN (blasticidin–neomycin resistance) and 4GN (GFP–neomycin resistance) have a lentiviral backbone. b, Experimental outline of the genome-wide in vitro and in vivo screen in two batches (green and blue/purple). c, Representative images showing YummR engrafted cells stained with H&E. Engrafted cells or tumors were harvested 1 h after subcutaneous injection (p.i.), 5 d p.i. or 10 d p.i., illustrating fully established tumors. Scale bars, 200 µm; for magnified image, 60 µm. Images are representative of two sections of a tumor and two tumors per condition. H&E, hematoxylin and eosin.

Source data

Extended Data Fig. 5. Setup and quality of genome wide mouse and human library.

a) Violin plots depict the distribution of the percentage of active UMIs after cells were put through a bottleneck and subsequent CreERT2 induction, showing the reproducibility of the ratio between active and inactive UMIs independent on genomic integration site. b) Setup of the genome-wide sgRNA library in 9 subpools. Green column contains genes that are drugged or druggable, the blue column is defined by biological categories and in the last purple group fall all genes that could not be assigned to functional groups. c-d) Genome wide library is divided in 9 subpools, A1-A4 containing drugged or druggable genes, B1-B4 containing genes that are defined by biological categories and a rest group. Number of genes per subpool and overlap between subpools are displayed. c) Library targeting mouse genome. d) Library targeting human genome. e-h) Library representation of each genome wide library. On the x-axis are the guides ranked by its Log₁₀ NGS reads on the y-axis. The 10th and 90th percentile are marked by red dashed lines and fold-change difference between the two is indicated. Cloning coverage for each sgRNA in the subpools of each library are depicted in the tables beside each plot. e) Mouse CRISPR-StAR library. f) Human CRISPR-StAR library. g) Mouse pLenti-UMI library. h) Human pLenti-UMI library.

Generating a versatile genome-wide human and mouse CRISPR-StAR library

We cloned a genome-wide library targeting the human or mouse genome with five sgRNAs per gene into the optimized viral backbones. The library was subdivided into nine subpools that are organized in three groups. In one group, column A, we included genes that are drugged, druggable or relevant in the context of druggability. Column B includes biological categories, such as genes involved in transcription, RNA binding and others. The third subgroup of the library contains genes that could not be assigned to the functional groups mentioned above (Extended Data Fig. 5b). Genes in columns A and B are assigned such that no overlap between subpools exists within the same column, and genes assigned to both groups are present only in the higher-order pool. Gene overlap between columns was tolerated, resulting in some genes being present in both sublibraries (Extended Data Fig. 5c,d; see Methods for more details). Each sgRNA was combined with more than 10³ UMIs⁷ to enable single-cell tracing within pooled screens. To assess the library quality, we performed next-generation sequencing (NGS) and found a very even representation of the sgRNAs within the library (Extended Data Fig. 5e–h).

Benchmarking CRISPR-StAR in in a genome-wide in vivo screen

Our study aimed to assess the efficacy and power of CRISPR-StAR in vivo using the Yumm1.7 450R mouse melanoma cell line, chosen for its notably low engraftment rates and considerable growth heterogeneity (Fig. 1b). These cells are driven by the Braf^V600E mutation, Cdkn2a loss and Pten loss yet are selected to be resistant to therapy with Braf inhibitors, such as dabrafenib^30,31. Initial targeted therapy for melanoma often involves Braf inhibition, which results in significant tumor shrinkage. Most patients experience rapid relapse, as reactivation of the MAPK pathway is a frequently acquired resistance mechanism^32–34. Next to targeted therapy, immune checkpoint blockade has greatly enhanced treatment; however, pre-existing or acquired resistance mechanisms limit clinical use for many patients, underscoring the need for alternative therapeutic strategies^35,36. Identification of specific genetic vulnerabilities in therapy-resistant cells in vivo could, thus, uncover options for novel targeted therapy in Braf inhibitor-resistant melanoma.

An advantage of CRISPR-StAR is that the internal controls overcome batch effects and, thus, allow for additive data acquisition. The genome-wide screen could thus be performed in two batches (sublibraries A and B) both in vitro and in vivo (Fig. 2b and Extended Data Fig. 6a,b). Notably, our screening setup allowed us to assay gene function exclusively in fully established tumors by activation of sgRNAs 10 d after cell engraftment (Fig. 2c), thereby excluding in vitro viability and engraftment efficiency from the screen readout.

Extended Data Fig. 6. Genome wide screen median ratio, library representation in vivo and conventional analysis in vitro.

a) Median ratio percentage of active and inactive UMIs in the two batches of the genome wide screen. b) Bar plot for the average number of cancer cells engrafted in the mouse, defined by the number of UMIs. Engrafted numbers are averages per tumor, as 5-7 tumors were combined during genomic DNA extraction. Of note, sgRNAs were initially paired with > 10³ UMIs each during the cloning process and the library is thus far more complex. Thus, the number of UMIs detected serves as proxy for the number of engrafted cells. c) Bar plot of the frequency of each UMI in the genome wide screen in vivo. d) Bar plot in which each bar stands for how many times an x number of UMIs per guide was found. On average the library was covered with 2.3 UMIs per guide. e) Here the number of UMIs per gene were assessed. On average the library was cover with 8.7 UMIs per gene. Furthermore, in vivo we detected 81110 sgRNAs of the 104715 sgRNAs in the plasmid library. f) Volcano plots and density plots of in vitro conventional analysis. Dots represent genes, axis display LFC and -Log₁₀RRA score for each screen. In vitro depleting genes (IDGs) in red, non-Essential genes in black and grey represents genes that are not defined by either of the groups, termed other. g) LFC correlation of genes overlapping in two batches of the genome wide screen for the in vitro conventional analysis. Colors as in e, additionally yellow dots are other genes that in vitro showed minimal enrichment or depletion (LFC in vitro < 1 and > −1), therefore called neutral other. Pearson correlation coefficient (R) between the two batches, when taken into account all depicted gene groups are displayed. R values when neutral other genes are excluded are R = 0.93.

As anticipated, in vitro screening followed by NGS and comparison of reads relative to the sgRNA library demonstrated a clear separation of sgRNAs targeting non-essential (Fig. 3, black) and essential (Fig. 3a, red) genes. In vivo, the screen was conducted across 143 animals (one tumor per animal), yet the complexity of sgRNA representation was, on average, only 2.3 cells per sgRNA or 8.7 cells per gene as quantified by enumerating UMIs within tumors, highlighting the low engraftment rate (~1,300 cells per tumor) of this cell line (Extended Data Fig. 6b–e). Comparing sequencing reads for in vivo active sgRNAs against the plasmid library, akin to a conventional CRISPR–Cas9 analysis, highlighted how even-represented the sgRNAs are within our library. However, active reads were distributed across 5 orders of magnitude in the in vivo dataset irrespective of the genes targeted (Fig. 3b). Of note, sgRNAs targeting in vitro depleting genes (IDGs; see below) did not separate from non-essential genes. In total, 12.8% of sgRNAs did not obtain any active read in the in vivo data, because of either strong essentiality or a failure of cells harboring these sgRNAs to engraft—two possibilities that conventional screening analysis is unable to distinguish. When plotting the same active reads in vivo (that is, maintaining the y axis) against the inactive reads in vivo, sgRNAs targeting IDGs showed a good separation from controls in vivo across all scales of sgRNA representation (Fig. 3b, right panel). Moreover, of the sgRNAs without active reads, 45.9% also had no inactive reads in vivo, strongly indicating that they dropped out during engraftment before sgRNA activation. These sgRNAs would, therefore, have been falsely assessed as strongly essential by a conventional analysis (false positives), but they can be easily weeded out in the CRISPR-StAR paradigm. For the remaining 54.1% of sgRNAs without active reads, inactive reads were also detected and could, therefore, be classified as true depleted sgRNAs. Taken together, the CRISPR-StAR paradigm separated sgRNAs targeting essential from non-essential genes in sparse and heterogeneous datasets, where conventional CRISPR screening methods failed.

Fig. 3. Benchmarking conventional analysis versus CRISPR-StAR.

a, log₁₀ active reads per sgRNA (y axis) versus log₁₀ inactive reads per sgRNA of the in vitro genome-wide screen. Dots represent sgRNAs: in red, sgRNAs that are in vitro defined to target essential genes (LFC < −3 and −log₁₀RRA score ≥ 10); in black, sgRNAs targeting non-essential genes (mouse orthologs of never-depleting genes in DepMap¹). b, Reads per sgRNA of the in vivo genome-wide screen using conventional analysis (active sgRNAs versus plasmid library sgRNAs) or CRISPR-StAR (active versus inactive sgRNAs); axes and colors are as in a. c,d, Volcano plots and density plots. CRISPR-StAR analysis in vitro (c) and conventional and CRISPR-StAR analysis (d). Dots represent genes, with LFC and −log₁₀RRA score on the axes. IDGs are in red; non-essential genes are in black; and ‘others’ (genes not defined by either of the groups) are in gray. e, LFC correlation of genes overlapping in two genome-wide screen batches. Colors are as in c and d; additionally, yellow dots are ‘neutral other’ genes showing minimal enrichment or depletion in vitro (LFC in vitro < 1 and LFC in vitro > −1). R is shown (when all gene groups are considered). R values when excluding ‘neutral other’ genes: in vitro R = 0.94; in vivo conventional analysis R = 0.24; and in vivo CRISPR-StAR R = 0.6. f, dAUC analyses, assessing recall of the sgRNAs targeting essential genes and non-essential genes, when ranked according to LFC. Gene groups are defined by the in vitro dataset, depleting (solid line) and non-essential (dotted line). Colors indicate the analysis: in vitro (black), conventional (blue) and StAR (magenta). g,h, Screen performance was evaluated by calculating ROC curves and PR curves for the in vitro (black), conventional (blue), StAR (magenta) and random (gray) datasets. g, Comparing the true-positive rates and the false-positive rates of in vitro defined depleting and non-essential genes. h, PR curves render the tradeoff between the positive predictive value (precision) and the true-positive rate (recall).

Source data

We combined the results of individual UMIs to genes and calculated gene effects using MAGeCK³⁷ (Methods). Visualization of in vitro log₂ fold change (LFC) against robust rank aggregation (RRA) score yielded anticipated outcomes: non-essential genes exhibited no change (Fig. 3c and Extended Data Fig. 6f). On these gene-level data, we defined the IDGs by a negative change of at least eight-fold and an RRA score of −log₁₀ > 10 as used already in Fig. 3a,b. Upon evaluating gene-level effects in vivo, the performance of CRISPR-StAR was notably superior, with a better discrimination between IDGs and controls compared to conventional analysis (Fig. 3d), accompanied by higher statistical significance. As before (Fig. 1e,f), this separation was driven primarily by the dramatic reduction of noise in the data obtained from neutral genes rather than a failure to obtain depletion for essential genes.

Moreover, leveraging the partial overlap of genes across our two screening batches, we evaluated reproducibility of these two datasets cloned independently and generated 2 years apart of each other. Although conventional analysis achieves an R coefficient of 0.14, CRISPR-StAR demonstrated substantially improved reproducibility with R = 0.54 (Fig. 3e and Extended Data Fig. 6g). As additional benchmarking of CRISPR-StAR, we computed separation of guides targeting IDG from non-essential genes by dAUC, ROC as well as precision-recall (PR) curve (Fig. 3f–h). Collectively, all data corroborate the results obtained in the in vitro pilot screen, affirming that CRISPR-StAR effectively mitigates noise stemming from data sparseness and heterogeneity, thereby empowering genetic screening in vivo.

Context-specific genetic dependencies in melanoma

To elucidate genetic dependencies specific to the in vivo versus in vitro context, we plotted results relative to each other and observed a clear separation of core essential genes from controls under both conditions (Fig. 4a). Cellular component analysis of the genes depleting more than 1.75-fold as much in vivo compared to in vitro revealed a strong overrepresentation (two-fold, P = 7.5 × 10⁻⁵) of genes associated with the mitochondrial inner membrane (blue dots; Gene Ontology (GO) term 0005743). This included genes encoding for proteins of the tricarboxylic acid (TCA) cycle, the electron transport chain, oxidative phosphorylation and the mitochondrial ribosome. Therefore, a genetic dependency on oxidative phosphorylation is specifically observed in vivo. Genetic dependency of these gene products is also not observed across publicly available in vitro datasets (Fig. 4b).

Fig. 4. Context-specific gene essentiality in therapy-resistant melanoma.

a, Correlation of LFC in vitro (x axis) and in vivo (y axis). Each dot represents a gene: in red, core essential genes (mouse orthologs of depleting genes in DepMap¹); in black, non-essential genes; in blue, genes that are part of the cellular component GO term ‘mitochondrial inner membrane’ (GO: 0005743); and in gray, the genes that are not defined by these aforementioned groups, called ‘other’. Size indicates the −log₁₀RRA score in vivo. LFCs in vivo and in vitro lower than −1 (below or left of the dashed line) are genes shown to have a depleting effect. From this group, all genes that are right of the dotted-dashed line have an in vivo depleting tendency. In this group, mitochondrial inner membrane genes are enriched. b, Violin plot of DepMap gene effects of mitochondrial inner membrane–associated genes (GO: 0005743) from the in vivo depleting tendency group in a. Additionally, the LFC (gene effect of the CRISPR-StAR in vitro and in vivo screen) is represented by a blue triangle or circle, respectively. c, Volcano plot of in vivo validation screen with an in vivo depleting library (orange) and an in vivo enriching library (green); other colors are as described in a. LFC is on the x axis, and −log₁₀RRA score is on the y axis. d,e, Correlation of LFC gene effects in vitro versus in vivo. Colors are as defined in a and c. In vitro screen was performed in the presence of the Braf inhibitor dabrafenib (d) and in the absence of dabrafenib (e). i.membrane, inner membrane.

Source data

We aimed to validate individual genes with a clear bias for stronger dependency in vivo (here referred to as ‘in vivo specific’) with putatively therapeutic relevance. Based on screening results of column A with a cutoff of in vivo LFC < −2 and in vitro LFC > −1.5 or in vivo LFC > 2.5 and at least three UMIs, a validation library comprising 80 putative in-vivo-specific essential (orange) and 23 putative in-vivo-specific TS genes (green) was generated (Fig. 4c–e). Controls, as well as approximately 276 additional genes based primarily on lower cutoff criteria and 112 genes belonging to the mitochondrial respiratory chain, were included. The library was screened in vitro as well as in vivo (Fig. 4c and Supplementary Fig. 1a). Because our therapy-resistant melanoma cells are routinely grown in the presence of dabrafenib in vitro, yet in the absence thereof in vivo, we ran the in vitro arm with and without dabrafenib (Fig. 4d,e). Although TSs showed a tendency to also enrich in vitro in the absence of dabrafenib, essential genes depleted even less in that arm overall, presumably due to reduced cell cycles in this suboptimal culturing condition. Genes associated with the Erk pathway stood out, displaying a strong dabrafenib-dependent selection in vitro, namely Mapk1 (also known as Erk2) and Hsp90ab³⁸, whereas the negative regulator Stk40 (ref. ³⁹) showed mild depletion in this condition. We could validate the context-specific dependency for 12 of 23 (52.2%, LFC > 1) and 23 of 80 (28.8%, LFC in vivo < −1.5 and LFC in vitro > 1.5) of the nominated genes that showed putative enrichment or in-vivo-specific depletion, respectively. Of note, genes included under a lower stringency cutoff displayed a markedly lower frequency of validation as anticipated (45/267, 16.3%) (Supplementary Fig. 1b). Similarly, we confirmed the strong in-vivo-specific genetic dependency on genes of the mitochondrial inner membrane. As expected, genes that we included in the validation set based on lower stringency cutoffs did not validate at a similar frequency. In summary, the validation experiment confirmed the experimental power of CRISPR-StAR and our implemented inclusion criteria in the validation set and uncovered context-specific genetic dependencies, emphasizing the relevance for genetic experiments in vivo.

Cross-species validation of in-vivo-specific hits

To corroborate our findings in Yumm450R murine melanoma cells and to assess whether human therapy-resistant melanoma cells also rely on these gene products, we sought to evaluate the set of validation genes in both mouse and human cell lines concurrently. Given that CRISPR-StAR assays established tumors in an inducible manner, and given that genes often function in a highly context-specific manner, we also tested whether gene essentiality extends tumor initiation and growth as assayed with conventional screening paradigms. To this end, we cloned libraries targeting the mouse genes and their human orthologs into a CRISPR-StAR backbone as well as into a constitutive sgRNA backbone and screened sensitive and Braf inhibitor-resistant mouse Yumm cells^30,31 and human melanoma (A375)^40–42. The constitutive screen was performed in duplicate, and we added 4T1, a mouse breast cancer cell line, as an outside comparison. The CRISPR-StAR screen in YummR and A375 was performed as before (Fig. 5a).

Fig. 5. Cross-species validation of genetic dependencies.

a, Heatmap of CRISPR-StAR screening in vivo with three different melanoma cell lines: Yumm450R, Braf inhibitor-resistant murine melanoma; A375R, Braf inhibitor-resistant human melanoma; and A375; Braf inhibitor-sensitive human melanoma. Color range depicts the LFC gene effect: red for depleting, white for no effect and blue for enriching. b, Heatmap of conventional CRISPR screening in vivo with Yumm450R, A375R, A375 and, additionally, Yumm (Braf inhibitor-sensitive murine melanoma) and 4T1 (murine mammary carcinoma). LFC was calculated by comparison of the number of UMIs found in the tumor by NGS to the NGS reads in the cells at the day of engraftment.

Source data

The constitutive screen was enabled by the small size of the library, targeting only 38 genes that resulted in a median of 116 engrafted cells per sgRNA and 575 engrafted cells per gene. In the constitutive arm, we took advantage of the high library coverage and the presence of UMIs in the library, and we analyzed the conventional arm, both the dropout of UMIs relative to the day of engraftment as a proxy for cell death (Fig. 5b) as well as the dropout of reads relative to engraftment as a proxy for cell fitness (Supplementary Fig. 2). As expected for CRISPR-StAR, all assayed genes showed dropout in YummR. Essentiality was mostly confirmed for both A375 cell lines tested as well, indicating that the in vivo essential genes are conserved across species. For the conventional analysis, essentiality by UMI dropout and read dropout resulted in very reproducible results, which were also consistent across replicates and sgRNAs targeting the same gene (Supplementary Fig. 3). However, the conventional results did not correlate with the CRISPR-StAR results in all cases. In particular, we noticed an at least much more pronounced genetic dependency for Stk40 and Zbtb10 in mature tumors assayed by CRISPR-StAR.

Given the discrepancy between constitutive and inducible sgRNAs, we next validated genes identified in the CRISPR-StAR individually. We had noticed that tamoxifen induction fails to result in complete recombination of StAR vectors, presumably due to poor vascularization of the tumors. In our NGS preparation, unrecombined events are excluded by recombination-specific PCR amplification (Methods). Thus, we built a vector that would report recombination state by three fluorophores (Fig. 6a and Extended Data Fig. 7a,b). TripleColour-StAR is miRFP positive before recombination, whereas recombined active constructs express dTomato, and inactive constructs are marked with TagBFP. We confirmed that the TripleColour-StAR successfully links the fluorescent markers to genotypes in in vitro and in vivo setups using sgRNAs targeting EGFP and the essential gene Tuba1b (Extended Data Fig. 7c–h). We decided to validate six genes, including Stk40 and Zbtb10, for which we detected no essentiality in a constitutive validation setup and ran the TripleColour-StAR competition assays in vitro and in vivo. TripleColour-StAR was induced in three independent tumors per gene at a size of approximately 100 mm³ and harvested 14 d later at a size of approximately 500–1,000 mm³, thus accounting for approximately 2–3 volume doublings. Cells in vitro were analyzed 7 d and 14 d after induction to match both a situation of few cell doublings as well as time.

Fig. 6. Single gene validations of genes with context-specific genetic dependency.

a, Schematic description of TripleColour-StAR construct. b,c, Ratio between the number of dTomato⁺ and TagBFP⁺ cells in each gene KO, in vivo, 14 d after tamoxifen injection. The cells were quantified by FACS. Statistical analysis was conducted using an unpaired t-test (two-sided). Each bar represents the mean ratio of three biological replicates ± s.d (Supplementary Table 1). ‘NS’ indicates non-significant; *, **, *** and **** represent P values of less than 0.05, 0.01, 0.001 and 0.0001, respectively. Exact P values in order of in vivo D14 to in vitro D7 and in vivo D14 to in vitro D14: sgEmpty 0.1111 and 0.0585, sgEGFP 0.1147 and 0.0653, sgTuba1b 0.0174 and 0.2539, sgAip 0.0011 and 0.0326, sgBirc6 both <0.0001, sgUba6 <0.0001 and 0.0001, sgStk40 <0.0001 and 0.0003, sgWdr48 both <0.0001 and sgZbtb10 0.0003 and 0.0001. D, day; KO, knockout.

Source data

Extended Data Fig. 7. TripleColour-StAR links its genotypes with designated fluorophores.

a) Schematic representation of TripleColour-StAR construct before and after Cre recombination. b) Table showing the expected fluorescent protein expression based on the three possible outcomes of TripleColour-StAR. c) Representative immunofluorescence images demonstrating fluorescent protein expression with TripleColour-StAR with sgEGFP in YummC2 cells. In vitro, 7 days after 4-OH Tamoxifen treatment. Scale bars represent 50 µm. Representative for two independent experiments. d) FACS histogram of miRFP expression in negative control and TripleColour-StAR with sgEGFP after Neomycin selection. In vitro, without 4-OH tamoxifen treatment. e-h) FACS analysis of TripleColour-StAR with control sgRNAs. The miRFP- population underwent further analysis. In the case of sgEGFP, TagBFP+ and dTomato+ populations were further analyzed for EGFP expression. e) sgEGFP, in vitro, 7 days after 4-OH tamoxifen treatment. f) sgEGFP, in vivo, 14 days after tamoxifen injection. g) sgTuba1b, in vitro, 7 days after 4-OH tamoxifen treatment. h) sgTuba1b, in vivo, 14 days after tamoxifen injection.

Although the ratio of active and inactive sgRNAs remained largely unchanged for neutral controls, and active sgRNAs depleted completely for the essential control gene Tuba1b in vivo and in vitro (Fig. 6b), all other genes displayed significantly stronger dependency in vivo than in vitro (Fig. 6c). Of note, this was true despite a likely lower number of cell cycles in vivo. In particular, Zbtb10 displayed a strong context dependency, mirroring the results obtained in the validation screens (Fig. 4c–e). Zbtb10 and Stk40 confirmed to be in vivo essential as seen before in the CRISPR-StAR setup, despite the fact that no depletion was observed in a conventional screening setup (Fig. 5b), pointing toward a strong context specificity. Thus, all single gene validations strongly support the context-specific function identified by CRISPR-StAR.

Discussion

CRISPR–Cas9 screens in survival and growth assays have enabled the mapping of gene essentiality across more than 1,000 cancer cell lines in culture¹. However, recent years have seen a strong push toward more complex systems, including three-dimensional (3D) cell cultures, such as spheroids^43–45, organoids^46–48, assembloids^49,50 and others. However, the complexity and heterogeneity of these models interfere with the reproducible cellular behavior necessary to translate phenotypes into NGS reads in pooled screening. Although the heterogeneity introduced with editing outcomes can be overcome with the implementation of UMIs¹⁴, very pronounced cell state heterogeneity in populations can be controlled for only by upscaling experimental setups. Similarly, in the allograft setting in vivo, we found 50% of cells making up a tumor stem from just 0.2–4% of engrafted cells. For many cell lines, more than 99% of the cells transferred into animals do not engraft at all (Fig. 1b). Technical solutions must embrace this heterogeneity, as it is a critical part of tumor biology, rather than attempting to select for conditions that blunt it. We performed all screens in this study in the Yumm1.7 450R cell line, which showed to be a particularly poorly engrafting cell line (Fig. 1b and Extended Data Fig. 1e) with approximately 1,300 cells per tumor.

Twin spot analysis in Drosophila melanogaster enabled by mitotic recombination⁵¹, as well as clonal analysis of neurons in mice^52,53, has shown the power of comparative phenotypic analysis of cells with different genotypes within the same model system or tissue. We aimed to set up a similar internally controlled system for an NGS readout but with the following important difference: cell populations of WT and loss-of-function (LOF) cells are born at the same time and place, not one individual cell each. This setup makes comparative analysis more robust and is ideally suited also to study cell–cell competition^54–56. To enable an internally controlled genetic screening paradigm that would allow for an NGS readout in high throughput, we built upon an inducible sgRNA technology that we developed previously¹³ and, in essence, combined the Switch-ON and Switch-OFF design to generate CRISPR-StAR. Upon clonal expansion of a single cell harboring the CRISPR-StAR construct, cells get induced, and WT allelles and LOF alleles are generated in a salt-and-pepper manner to control not only for the cell identities within the clone but also for any extrinsic influence on viability and growth.

By benchmarking to conventional analysis using multiple parameters, we show, in an allograft setting, where screen complexity is low and heterogeneity is high, that CRISPR-StAR overcomes the high degree of systemic noise and generates robust results even at representations of two cells per sgRNA and approximately nine cells per gene. This is made possible because each cell-derived tumor clone is an independent and self-contained experiment. In comparison, a conventional CRISPR screen typically requires a library coverage of more than 500 cells per sgRNA. The greatly reduced library coverage required by CRISPR-StAR will lead to a markedly reduced number of animals needed and can empower screening in many more cell line models that typically have low engraftment rates. This greatly improved scalability of experiments enables anchored screens overlaying genetic or pharmacologic perturbation in further screening arms. Thus, CRISPR-StAR empowers high-resolution genome-wide screening in vivo.

We performed a genome-wide screen in Braf inhibitor-resistant mouse melanoma cells and validated identified genes in the screening cell line but also in human A375 melanoma cells in vivo. We found that aerobic metabolism is an in-vivo-specific genetic dependency, indicating that a pronounced Warburg effect of aerobic glycolysis at least in this cell line is only observed in vitro. This finding aligns with previous studies^57–60 and is plausible given the unphysiological nutrient conditions and the small volume fraction of cells in culture. Similarly, we identified several known TSs, such as Dyrk1a^61,62, Cdkn1a^63,64, Nf2 (ref. ⁶⁵), Ambra1 (ref. ⁶⁶) and Crebbp^67,68, to positively select almost exclusively in vivo. In addition, we identified loss of Kctd5, Trip12 and others to result in robust positive selection and, thus, putatively representing TSs in the context of melanoma. Indeed, mutations in these genes are also found in human melanoma (Supplementary Fig. 4 (ref. ⁶⁹)). In 31% of 854 patients with melanoma, mutations in one or more of the genes identified in this study are reported. Taken together, the physiologic and genetic context of tumor biology, with respect to both metabolism and tumor drivers, is more faithfully modeled in vivo, even though many genes are essential in vitro and in vivo (Figs. 3 and 4a). This highlights the importance of conducting genetic analyses to unveil genetic dependencies in vivo, a task that can most faithfully be achieved through screening in established tumors.

As anticipated for an Erk pathway-driven tumor, we observed cell depletion upon Mapk1 loss in vitro in the presence of dabrafenib. In vivo or in the absence of dabrafenib, however, loss of Mapk1 resulted in positive selection, which is somewhat counterintuitive. Similarly, the Braf chaperone Hsp90 (refs. ^38,70) positively selected upon dabrafenib withdrawal in vitro. In addition, we found loss of Inppl1 encoding for Ship2, which is typically necessary for tumor growth and a drug target^71,72, to result in positive selection. Ship2 is required for EGFR recycling⁷³, pointing toward the possibility that loss of Inppl1 also dampens the ERK signaling pathway. In contrast, loss of Stk40 was shown to result in reactivation of Erk signaling in six of 14 cell lines tested upon inhibition of the pathway by a Ship2 inhibitor³⁹. We found Stk40 to be specifically essential in vivo and displaying a mild essentiality in vitro in the absence of dabrafenib (Fig. 4e). In response to Braf inhibition, the ERK pathway typically gets upregulated by various mechanisms⁷⁴. Our findings thus argue that, in the absence of dabrafenib, in vivo tuning down the excessive Erk pathway signaling in Braf-resistant cells by loss of Erk2 (in the presence of Erk1/Mapk3) or Hsp90 is beneficial^75–77. In turn, this excessive Erk signaling state should, thus, represent a vulnerability especially in vivo and might open a therapy option of pharmacologically modulating Erk throughput in opposite directions in an alternating regimen^78,79. Stk40 inhibition might indeed achieve such untolerated high output of the Erk pathway.

Several other genes were identified to display a genetic dependency that is more pronounced in vivo than in vitro. Some of them also did not display an equally pronounced essentiality when screening them in a constitutive in vivo setting. We speculate that this might be due to the different physiological context of cells engrafted in Matrigel during tumor establishment as opposed to the physiology of a fully established tumor. By implementing a fluorescent readout through the TripleColour-StAR construct, we were able to validate the genetic dependency in the context of our screening setup. We found that Uba6 and Birc6 displayed very similar effects across all experiments as previously observed in vitro⁸⁰. It was recently shown that Uba6 is the only E2 ligase ubiquitylating the antiapoptotic gene Birc6 (refs. ^80–82). Indeed, inhibitors against several Birc6 paralogs are in clinical development^83,84. The zinc finger protein Zbtb10 binds an atypical telomeric repeat and, thus, is thought to be involved in alternative telomere lengthening⁸⁵ as well as mRNA processing⁸⁶. However, in agreement with our results, in vitro gene loss is not resulting in major RNA expression changes⁸⁶. We do not know the mechanism for its in-vivo-specific essentiality.

Activation of the Aryl hydrocarbon Receptor (AhR) pathway by environmental toxins such as dioxin results in a melanocyte overproliferation and hyperpigmentation phenotype⁸⁷. Likewise, activation of AhR was shown to induce resistance to Braf inhibition, and the AhR pathway is frequently upregulated in Braf-resistant melanoma⁸⁸. In contrast, loss of the AhR nuclear translocator Arnt and, thus, loss of AhR signaling leads to increased mitochondrial activity and reactive oxygen species (ROS) production, resulting in faster growth and metastasis formation in late-stage melanoma⁸⁹. In line with the in vivo model being representative of later-stage tumor and dependent on oxidative phosphorylation, we found loss of Arnt to result in increased proliferation in an exclusively in-vivo-specific manner. In support of that, we found the AhR-interacting protein Aip as an in-vivo-specific dependency. Indeed, loss of Aip will result in translocation of AhR into the nucleus and, thus, increased AhR signaling⁹⁰, explaining the opposing phenotype to loss of Arnt. It is, thus, tempting to speculate that, similar to enhanced Erk signaling, AhR signaling might also pose a vulnerability generated by Braf therapy resistance, which can be leveraged, for example, by Aip inhibition. Future studies will be needed to reveal the precise context-specific biology of these genes.

Our experiments also highlighted a limitation of CRISPR-StAR: the need for establishment of cell lines with robust expression of Cas9 and CreERT2 that must not silence upon engraftment. Incomplete recombination of CreERT2 might further be affected by poor vascularization of allografts that could result in suboptimal exposure to 4-OH tamoxifen. A low fraction of unrecombined StAR constructs does not harm the screening setup due to the recombination-specific PCR amplification of sgRNAs for NGS. However, poor editing efficiency due to silencing of Cas9 will, as in vitro, result in poor depletion and limits the achievable signal-to-noise ratio. Thus, care must be taken to select for cell clones with robust expression of Cas9 and CreERT2 transgenes, and the use of genetic safe harbors may be an efficient way to ensure this. We developed a multiplexed setup to validate clones expressing required transgenes in vivo (Supplementary Fig. 5).

Overall, we anticipate that the CRISPR-StAR paradigm of internally controlled screening will facilitate the study of challenging model systems, such as in xenograft screens shown here, as well as 3D cell culture and directly in in vivo tissues.

Methods

StAR vector design and generation

Different CRISPR-StAR vectors were created to obtain the vector with an equal ratio of recombination products. For the initial proof-of-concept screen in mESCs, CRISPR-StAR vector 1 was generated using the original Switch-ON vector (p206)¹³, which is a retroviral backbone, and insertion of a lox5171 site at a XbaI restriction site and lox5171 at EcoRI restriction side (Supplementary Table 2). Additionally, an extra antibiotic selection cassette between the sets of lox sites was ligated at the SbfI restriction site, using oligos with an SbfI recognition sequence overhang to amplify PGK-blasticidin. In this way, cells that could induce recombination before induction with 4-OH could be removed from the population by blasticidin selection. Cells transduced with the CRISPR-StAR vector can be selected for with neomycin.

Vector 1 was used for further adaptations, such as an extra loxP site after the first lox5171 site (vector 2) and, additionally, sequence removal between the lox5171 sites to shorten the active sgRNA product (vector 3). Vector 4 contained, again, two sets of lox sites but kept the reduced sequence for the shorter active sgRNA product. Hereafter, vector 4 was used to amplify the full CRISPR-StAR cassette with oligos containing NheI recognition sites, which were used to ligate the purified PCR product into a lentiviral backbone. The lentiviral vectors are 5–7 and differ only in the antibiotic-resistant or GFP cassettes. Vector 4B has only the blasticidin selection cassette. Vector 4BN contains the blasticidin and a GFP–neomycin selection cassette, and vector 4GN has between the lox sites the GFP–neomycin cassette and no blasticidin selection method anymore. All vectors were verified by Sanger sequencing.

Library cloning

The human and mouse genome-wide sgRNA libraries were synthesized by Twist Bioscience. Each single-stranded oligonucleotide consists of an sgRNA of 18–20 bases, flanked by the sequences 5′-CGTCTCACACC-3′ and 5′-GTTTAGAGACGT-3′, which contain BsmBI restriction sites and a fitting overhang sequence. Additionally, primer binding sites for PCR amplification of each subpool were added. Each oligo subpool was independently amplified using NEBNext High-Fidelity 2× PCR Master Mix (for primers, see Supplementary Table 3). The purified PCR product was cloned into the CRISPR-StAR (4GN lentiviral vector) or pLenti-UMI by Golden Gate cloning. Each of the vectors contained random barcodes, which were integrated as previously described in Michlits et al.¹⁴. For the Golden Gate assembly, five parallel reactions with 245 ng of vector, 5.8 ng of dsDNA, 15 U of BsmBI v2 (New England Biolabs (NEB)), 1,000 U of T4 DNA ligase (highly concentrated; NEB) and 2.5 µl of T4 ligase buffer in 25 µl were cycled at 37 °C for 5 min, followed by 5 min at 16 °C for 75 cycles. The assembled plasmids were purified and transformed into Endura electrocompetent cells (Lucigen) with a Bio-Rad Pulser II according to the manufacturer’s instructions. Bacteria were recovered for 1 h at 37 °C and 200 r.p.m., and, subsequently, the bacteria were expanded over approximately 16 h at 37 °C and 200 r.p.m. in LB medium with ampicillin. Transformation efficiency was confirmed to be at least 1,000-fold over the size of the pool. Plasmid DNA was harvested using a Macherey-Nagel NucleoBond Midi Prep Kit. Subpools were mixed according to the number of sgRNAs in each pool to generate the two batches in which the genome-wide screen was performed. All plasmid sequences are available as a supplementary file; plasmids may be further obtained from Addgene. sgRNAs and sublibrary information are available in Supplementary Tables 5–9.

Cell culture

mESCs (feeder free) were maintained in DMEM supplemented with 15% FCS, 2 mM L-glutamine (Gibco), 1× non-essential amino acid (NEAA; Gibco), 1 mM sodium pyruvate (Sigma-Aldrich), 1× penicillin–streptomycin (Sigma-Aldrich), 100 µM β-mercaptoethanol and 100 U ml⁻¹ leukemia inhibitory factor (LIF).

Yumm1.7, Yumm1.7 450R, A375 and A375R cells were cultured in DMEM/F12 medium supplemented with 10% FCS, 2 mM L-glutamine (Gibco) and penicillin–streptomycin (Sigma-Aldrich). Additionally, Yumm1.7 450R medium contained 100 nM B-Raf inhibitor dabrafenib (Selleck Chemicals), and A375R medium contained 100 nM vemurafenib (Selleck Chemicals).

4T1, 4T07, B16F0, B16F1, B16F10 and Hepa1.6 cells were grown in DMEM high-glucose medium supplemented with 10% FCS, 2 mM L-glutamine (Gibco), 1× NEAA (Gibco), 1 mM sodium pyruvate (Sigma-Aldrich) and penicillin–streptomycin (Sigma-Aldrich).

Lastly, EMT6, EO771 and CT26 cells were grown in RPMI 1640 medium supplemented with 10% FCS, 2 mM L-glutamine (Gibco), 1× NEAA (Gibco), 1 mM sodium pyruvate (Sigma-Aldrich) and penicillin–streptomycin (Sigma-Aldrich).

Generation of Cas9-CreERT2-expressing monoclonal cells

All cell lines for CRISPR screening were generated by transducing them with a lentivirus containing Cas9-p2A-Puro and subsequent selection with puromycin. Cells that were also used for CRISPR-StAR screening, Yumm1.7 450R, A375R and A375, consecutively received retrovirus with CreERT2 and GFP or mCherry. Cells with the fluorescent marker were single cell sorted and expanded. Using CRISPR Switch-ON¹³ and an sgRNA against GFP or mCherry, the monoclonal cells were in vitro tested for Cas9 and CreERT2 induction, functionality and tightness. Seven days after CreERT2 recombinase induction, clones were subjected to single-cell fluorescence-activated cell sorting (FACS) analysis, and GFP reduction was measured. We aimed for a GFP reduction of at least 80–90% and, without 4-OH tamoxifen treatment, little to no reduction of GFP. Clones that matched these criteria were used for in vivo testing. For in vivo testing, we transduced the cells with a sublibrary and assessed by NGS the dropout of essential genes. For maximal Cas9 functionality (and minimizing silencing), we transduced the Yumm1.7 450R cells with an additional Cas9-p2A-blasticidin. Another option to select for functional clones is a PCR-based strategy (Supplementary Fig. 5a). This assay uses a common sgRNA against an essential gene but different UMIs that get transduced to individual clones to be tested. Upon pooling of all cellular clones in vitro, engraftment, tamoxifen induction and tumor harvest, each clone can be addressed individually if (1) it engrafted well (that is, the UMI is detected); (2) it recombined well (the active and inactive band of the CRISPR-StAR construct appears); and (3) Cas9 is active in all cells (that is, the active sgRNA band disappears due to gene essentiality) (Supplementary Fig. 5b).

Lentivirus production

For the production of lentivirus, Lenti-X cells (Clontech, 632180) were transiently transfected with the DNA library pool. Cells were seeded at 70% confluency in 15-cm dishes and transfected 6 h later. Then, 25 µg of library DNA, 12 µg of Gag-pol and 6 μg of VSV-G were mixed together with 129 µl of PEI in 3 ml of DMEM high-glucose medium for the transfection of one 15-cm dish of Lenti-X cells. The transfection mixture was incubated for 20 min at room temperature and added dropwise to the cells. The next morning, media were exchanged, and viral supernatant was harvested 48 h after transfection. The virus-containing supernatant was pooled, filtered through a 0.45-μm PES filter (VWR) and frozen at −70 °C. With a virus titration assay, the dilution ratio was determined for each viral sgRNA library pool to achieve a multiplicity of infection (MOI) of 0.25.

In vitro screening

For in vitro screening, at least 1,000 cells per sgRNA were infected at an MOI of 0.25 to ensure integration of one sgRNA per cell. The determined amount of viral supernatant was added to the cells, together with the appropriate medium for each cell line and 4 μg ml⁻¹ polybrene (Merck Millipore). Cells were maintained as three technical replicates, whereby the sgRNA coverage was preserved at more than 500 cells per sgRNA. Over the course of 5–7 d, cells were selected with 1 mg ml⁻¹ neomycin (500 µg ml⁻¹, G418; Life Technologies, 11811-031). A culture dish with a control population (cells without library) was taken along for assessment of neomycin selection. After elimination of all control cells, the library containing cells was used for in vivo screening or treated with 4-OH tamoxifen to continue the in vitro screen for 14 d.

Mice

For all screening experiments, 5–16-week-old male/female immunocompromised B6(Cg)-Rag2^tm1.1Cgn/J or B6.SJL-Rag2^tm1FwaPtprc^a mice (Rag2^−/−) were used. Additionally, for the experiments to assess the engraftment rate of cell lines, WT mice C57BL/6J and BALB/cJ were used. Mice were bred in-house, and all animal experiments were carried out according to project licenses approved by Austrian veterinary authorities. All mice were housed in groups of 3–6 in individually ventilated cages (IVCs), type II (Tecniplast), in a specific pathogen-free (SPF) animal facility. The animal housing rooms were temperature and humidity controlled (20 ± 3°, 50 ± 15%) with a 14/10-h light/dark cycle. All animals had ad libitum access to food (Ssniff) and water throughout the entire study.

In vivo CRISPR screening

Mice were anesthetized with isoflurane, followed by the subcutaneous injection of 5 × 10⁵–5 × 10⁶ cells harboring Cas9-CreERT2 and the StAR sgRNA library into the flank of Rag2^−/− mice in 50 µl of Matrigel:PBS (1:1; Corning). The cells of the three technical replicates were injected in equal groups of mice. Tumor size was measured 1–2 times per week using a calliper and calculated with the following formula: volume (in mm³) = (long side × short side²) / 2. At day 10 (Yumm1.7 450R) or day 14 (A375R) after cell injection, tamoxifen (dissolved in corn oil; Sigma-Aldrich) was intraperitoneally administrated for CreERT2 recombinase induction, followed by 14 d screening time. Tumors were harvested, cut into pieces and subsequently lysed with lysis buffer (10 mM Tris pH 8, 10 mM EDTA pH 8, 100 nM NaCl, 1% SDS) and 1 mg ml⁻¹ Proteinase K (VWR, 1.24568.0500) for 48–72 h.

Genomic DNA extraction and NGS library preparation

Lysed tumor samples were combined per 5–7 mice, after which genomic DNA was extracted from lysed cell or tumor samples by phenol-chloroform extraction and precipitated with isopropanol. PacI restriction digestion facilitated the accessibility of the DNA for PCR amplification. Each sample was individually amplified with barcoded primers (Supplementary Table 4) in 48 reactions of 50 µl, with an input of 4 µg of DNA per reaction, using KAPA HiFi DNA Polymerase HotStart ReadyMix (Roche). PCR reactions were pooled per sample, purified and quantified by a fragment analyzer. Finally, samples were pooled equally by concentration and sequenced on an Illumina NextSeq2000 P2 in a paired-end sequencing run, using custom primers (Supplementary Table 4). To identify afterwards active or inactive sgRNAs, we needed to read at least 75 bp in read 1; to read the dual index barcodes index 1 and index 2, we obtained 9 bp; and for the UMI barcode, we obtained 11 bp.

Data analysis

For data analysis, BAM files were retrieved from Illumina, and different samples were demultiplexed using the distinct experimental indices. Furthermore, sgRNAs and UMI sequences were identified using Bowtie, SAMtools and FASTX-Toolkit. To confirm an active or inactive sgRNA, we assessed, for each read 1 sequence, the last six bases. Reads that contained ‘TTTT’ in the end were assigned as inactive sgRNA, and ‘CAGC’-containing reads were marked as active sgRNAs.

Subsequently, data were processed in RStudio using a wide variety of packages, such as tidyverse, data.table, dplyr, ggExtra, ggpubr, ggrepel, stringr, patchwork and pROC.

For the in vitro CRISPR-StAR data, reads per UMI were pooled on the sgRNA level, and a pseudocount of 0.5 was added to each active and inactive sgRNA, to avoid 0 values. Because the screen was performed in triplicate, we could do paired analysis. MAGeCK version 0.5.9-foss-2018b³⁷ was used to calculate gene effects, by defining the treatment sample (-t) as the active reads and the control sample (-c) as the inactive reads. The MAGeCK software provided output files with the median log₂FC per gene (neg.lfc in output files) and RRA score (neg.score in output files).

In vivo CRISPR-StAR NGS data were, after separation of the active and inactive sgRNAs in R, processed at a UMI level. Before log₂FC gene effects were calculated, we implemented several filtering steps to clean the data in a standardized manner.

UMI hopping

The first filtering step that we applied was to address the issue of UMI hopping, which refers to instances where multiple distinct sgRNA are erroneously assigned the same UMI due to technical limitations in the NGS process. Because the UMI barcodes and the sgRNA library are cloned at a very high complexity, it is not common to have multiple sgRNAs combined with the same UMI. However, many sgRNAs were found to have the same barcode but with a much lower read count. To this end, we calculated the sum of the active and inactive reads of each unique sgRNA–UMI combination (sumReads), after which we sorted the dataset in descending order based on the total number of reads associated with each UMI. The next step was an iteration through the UMIs with the highest number of reads until the total number of reads associated with a UMI drops below the predefined threshold of 100,000, whereby, in each iteration, the UMI with the highest number of reads was identified from the top of the sorted list. A ratio was calculated for each occurrence by dividing its number of reads by the maximum number of reads among all occurrences of the same UMI. These ratios were then used to filter out instances where the ratio was less than or equal to 0.001, indicating a low proportion of reads compared to the maximum observed for that UMI. This process ensures that UMIs with a disproportionately small number of reads, compared to the highest observed count for that UMI, are filtered out, thereby mitigating the effects of UMI hopping and improving the accuracy of downstream analyses.

Polymeric UMIs

The second filtering step addresses the matter of polymeric sequences within the UMIs. Polymeric sequences are repetitive sequences of nucleotides that may arise due to technical artifacts or biases during library preparation or sequencing. Again, due to the high complexity of the library, it is not expected to have an extremely high number of reads for polymeric UMIs. For this reason, we excluded UMIs containing seven or more consecutive occurrences of cytosine, seven or more consecutive occurrences of adenine, seven or more consecutive occurrences of thymine or five or more consecutive occurrences of guanine. For guanine, the filter is less strong, as Illumina’s NextSeq 2000 assigns the nucleotides with a two-color/channel sequencing by synthesis technology, whereby no color is interpreted as guanine⁹¹, and this is more likely to happen than misdetection of a color.

Active and inactive read filter

After we combined the filtered data of the two screened batches, we applied a last filter. The sumRead of every sgRNA–UMI of each replicate in the two screening batches was calculated. Subsequently, we omitted the sgRNA–UMI replicate combinations with 20 or fewer sumReads.

Lastly, a pseudocount of 0.5 was added to each active and inactive sgRNA–UMI replicate, and gene effects were calculated with MAGeCK in the same way as for the in vitro CRISPR-StAR data. For the in vivo CRISPR-StAR gene effect, only genes with three or more UMIs were used for further analysis.

To pool the data of batch 1 and batch 2 of the genome-wide screen, we calculated the median ratio in each batch for the non-essential genes. Hereafter, the normalization factor between batch 1 and batch 2 for the active and inactive reads could be defined for both the in vitro and the in vivo screen. Before pooling the data of the two batches, reads of batch 2 were multiplied by these normalization factors.

The sgRNAs in the plasmid library subpools were sequenced with individual experimental indices to assess the representation of the library (Extended Data Fig. 5c–f). For the conventional analysis, the reads for each subpool were normalized to the median sgRNA reads of all subpools. Subsequently, we performed MAGeCK analysis, in which active reads were assigned as -t and the sgRNA reads in the library as -c, resulting in log₂FC gene effect values. Note, UMI analysis is not possible in the case of the conventional genome-wide screen in vivo. As in this context, the comparison of the very complex plasmid library with the very few number of UMIs in the tumors would not be fair.

Various gene groups were defined in this paper. IDGs are genes that were, in the genome-wide screen, found to be essential, as these had an effect of LFC lower than −3 and an RRA score of at least −log₁₀ 10 or higher. Non-essential genes were, in DepMap¹, found in no screen to have a gene effect below −0.5. Lastly, core essential genes were, in DepMap, the genes with a gene effect of −1 and lower in at least 400 screens.

TripleColour-StAR assay

The TripleColour-StAR vector was generated from the lentiviral CRISPR-StAR vector by introducing EF1a-TagBFP after the stop cassette, NLS-P2A-miRFP703-T2A-NeoR replacing EGFP-P2A-NeoR and PGK-dTomato after the tracr. These cassettes were synthesized and PCR amplified for subcloning using restriction enzymes. Each sgRNA used for the TripleColour-StAR assay was synthesized, annealed and subsequently inserted into the vector digested with Esp3I restriction enzyme. TripleColour-StAR vectors containing each sgRNA were transfected into Lenti-X cells as previously described. Yumm1.7 450R cells with Cas9 and CreERT2 were treated with the lentivirus and selected with neomycin. After selection, cells were expanded and either subcutaneously injected into Rag2^−/− mice or cultured in vitro with and without 4-OH tamoxifen. In vitro samples were analyzed on day 7 and day 14 after 4-OH tamoxifen treatment via FACS. The Cre recombination in mice was induced at day 7 after cell injection with intraperitoneal tamoxifen. Tumors were harvested on day 14 after tamoxifen injection, dissected into small pieces and incubated with Collagenase IV (2 mg ml⁻¹; Worthington, LS004188) and DNaseI (0.2 mg ml⁻¹; Sigma-Aldrich, 4536282001) for 1 h at 37 °C. Cells were then filtered (70 µm), cultured overnight to isolate tumor cells and subsequently analyzed via FACS. FACS data were analyzed using FlowJo version 10.9.0 software (Supplementary Fig. 6 and Supplementary Table 1). Cells carrying TripleColour-StAR with sgEGFP were re-plated on eight-well chamber slides (ibidi, 80826), incubated overnight, fixed with 4% paraformaldehyde solution (Sigma-Aldrich, 28908) and washed with PBS.

Fixed cells were permeabilized with 0.1% Triton X-100 (Sigma-Aldrich, T8787) and blocked with 10% donkey serum (Sigma-Aldrich, D9663). Subsequently, cells were stained with primary antibodies (TagBFP: NanoTag Biotechnologies, N0502-CUSTOM-FluoTag-X2 anti-TagBFP Dylight 405, 1:500; EGFP: Abcam, A13970, 1:1,000; dTomato: Biorbyt, ORB11618, 1:500) and secondary antibodies (Invitrogen, A78948 and A11057; 1:500). Nuclear staining was performed with To-Pro-3 (Invitrogen, T3605; 1:1000). Imaging was conducted using an LSM 800 Axio Observer microscope, and ZEN Black (version 2.3) and ZEN Blue (version 2.3 Sp1) were used for image processing.

Reporting summary

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

Online content

Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at 10.1038/s41587-024-02512-9.

Author contributions

The concept of this study was conceived by U.E. Proof-of-concept experiments were performed by A.S., N.B., E.U. and A.E. with support from G.V. E.U. generated the genome-wide libraries and performed the genome-wide screen and validation screens, with support from J.B., N.B., J.Y. and S.L., as well as bioinformatic analysis, with support from M.N. and G.V. Validation experiments were performed by E.U. with help from J.S. and C.L. J.L. generated the validation vector with support from T.P. and performed validations with support from E.U. J.L., C.L. and D.S. supported experimental design, execution and analysis. M.G. developed the method for in vivo clone selection. E.U. and U.E. generated the figures and wrote the paper with support from J.L., A.O., D.S. and all other co-authors.

Peer review

Peer review information

Nature Biotechnology thanks Michael McManus and the other, anonymous, reviewer(s) for their contribution to the peer review of this work.

Data availability

All primary and processed data of the genome-wide CRISPR–Cas9 screen are uploaded to the Gene Expression Omnibus database under series GSE262309 (ref. ⁹²). As the screen was performed in two batches, batch 1 is available under GSE262307, and batch 2 is available under GSE262308. Mutation profiles from all available human melanoma studies are accessible via the cBio Cancer Genomics Portal (https://www.cbioportal.org/). Source data are provided with this paper.

Code availability

Code used for analyzing the genome-wide screens in this study are available at GitHub⁹³.

Competing interests

E.U., A.S. and U.E. are listed as inventors of EP3889259A1, and U.E. is listed as inventor of EP3219799A1. D.S. and U.E. are co-founders of Viverita Therapeutics and consultants to Tango Therapeutics. The other authors declare no competing interests.

Extended data

is available for this paper at 10.1038/s41587-024-02512-9.

Supplementary information

The online version contains supplementary material available at 10.1038/s41587-024-02512-9.