Base editing screens map mutations affecting interferon-γ signaling in cancer

Summary Interferon-γ (IFN-γ) signaling mediates host responses to infection, inflammation and anti-tumor immunity. Mutations in the IFN-γ signaling pathway cause immunological disorders, hematological malignancies, and resistance to immune checkpoint blockade (ICB) in cancer; however, the function of most clinically observed variants remains unknown. Here, we systematically investigate the genetic determinants of IFN-γ response in colorectal cancer cells using CRISPR-Cas9 screens and base editing mutagenesis. Deep mutagenesis of JAK1 with cytidine and adenine base editors, combined with pathway-wide screens, reveal loss-of-function and gain-of-function mutations, including causal variants in hematological malignancies and mutations detected in patients refractory to ICB. We functionally validate variants of uncertain significance in primary tumor organoids, where engineering missense mutations in JAK1 enhanced or reduced sensitivity to autologous tumor-reactive T cells. We identify more than 300 predicted missense mutations altering IFN-γ pathway activity, generating a valuable resource for interpreting gene variant function.

Correspondence mathew.garnett@sanger.ac.uk In brief Coelho et al. use functional genomics to systematically analyze the genetic determinants of IFN-g response in colorectal cancer cells. Base editing mutagenesis of key signaling components facilitates the parallel assessment of gene variant function at scale, relevant for understanding IFN-g signaling in cancer immune surveillance and immune disorders.

INTRODUCTION
Cellular responses to the cytokine interferon-g (IFN-g) are essential for normal inflammatory responses, but pathway dysfunction and disease can occur through mutation, leading to hematological malignancies and immunological disorders. 1,2 JAK inhibitors are used to treat myeloproliferative disorders such as polycythemia vera and inflammatory disorders such as rheumatoid arthritis and ulcerative colitis, 2 reflecting the central role of JAK-STAT signaling in these diseases. Furthermore, IFN-g signaling in cancer cells is a critical aspect of anti-tumor immunity. 3,4 Clinical resistance to immune checkpoint blockade (ICB), such as antibody therapies targeting programmed cell death 1 (PD-1) and CTLA-4, has been associated with somatic mutation and homozygous inactivation of IFN-g pathway components in tumor cells, [5][6][7][8] or inactivation of genes involved in antigen processing and presentation (e.g., B2M) 9,10 that are expressed in response to IFN-g. For example, mutations in JAK1 and JAK2 can confer resistance to ICB 5,6 and chimeric antigen receptor T cells. 11 Since somatic mutations in cancer are predominantly single nucleotide changes, which often result in missense mutations with unknown consequence 12,13 (i.e., variants of uncertain significance [VUS]), interpreting their functional relevance remains challenging, representing an impediment to diagnosis, patient stratification, and the management of drug-resistant disease.
Experimental approaches are important to assess the functional effects of VUS. This is due to the ability to establish causality between VUS and disease-related phenotypes, as well as a scarcity of clinical datasets (e.g., from sequencing ICB-resistant tumors) and the infrequent occurrence of some variants in patient cohorts. One approach to prospectively assess endogenous gene variant function at scale is base editing [14][15][16][17][18][19][20] ; a cluster regularly interspaced short palindromic repeats (CRISPR)-based gene editing technology that uses cytidine 21 or adenine 22 deaminases to install C->T or A->G transitions, respectively. Base editors achieve high editing efficiencies within the activity window, which is typically focused around positions 4-8 of the protospacer (where the PAM spans position 21-23), with minimal generation of DNA insertions and deletions. 23 In this study, we use CRISPR-Cas9 screening to identify mediators of sensitivity and resistance to IFN-g in colorectal adenocarcinoma (CRC), and use cytidine base editors (CBEs) and adenine base editors (ABEs) to perform mutagenesis of the top-scoring genes, thereby systematically mapping loss-offunction (LOF) and gain-of-function (GOF) variants modulating IFN-g pathway activity ( Figure 1A), including VUS associated with diseases such as cancer.

RESULTS
CRISPR-Cas9 screens identify mediators of sensitivity and resistance to IFN-g Mechanisms of immune evasion can be cancer-cell intrinsic, 24,25 and thus systematically explored in vitro using CRISPR-Cas9 screens in cancer cell models. 26 This approach has identified tumor IFN-g signaling as essential for sensitivity to anti-tumor (A) Schematic of the integrated CRISPR-Cas9 and base editing screening approaches to identify genetic mediators of sensitivity and resistance to IFN-g. Cas9 screens identify pathways and genes regulating IFN-g response in colorectal cancer cell lines and base editing mutagenesis screens assess the functional consequence of VUS in key regulators. (B) Gene-level volcano plots of CRISPR-Cas9 screens comparing IFN-g-treated and control arms. (C) gRNA-level analysis of top resistance or sensitizing genes, representing essential components of the IFN-g pathway.
(D) Common and private genes conferring sensitivity and resistance to IFN-g in HT-29 and LS-411N CRC cell lines identified from CRISPR-Cas9 screens. Hits were selected using MAGeCK; p < 0.05 and a false discovery rate of less than 0.05. All data are the average from two independent screens performed on separate days. See also Figure S1 and Table S1. T cells, but has focused on melanoma 10,27,28 or a limited number of mouse syngeneic cell lines, 29-32 leaving other indications for ICB, such as CRC, 33 relatively unexplored. To systematically evaluate genetic, cell-intrinsic determinants of IFN-g signaling, and nominate genes for further investigation, we performed CRISPR-Cas9 screens in two BRAF-mutant CRC cell lines, HT-29 (microsatellite stable) and LS-411N (microsatellite unstable [MSI]) ( Figure 1A). Cas9-expressing derivative cell lines 34 were transduced with an immuno-oncology focused guide RNA (gRNA) gene knock-out (KO) library, containing 10,595 gRNAs targeting 2,089 genes with a median of five gRNAs per gene (Table S1) and selected with cytotoxic doses of IFN-g. Screen quality was verified by efficient depletion of gRNAs targeting essential genes 35,36 ( Figure S1A) and a high correlation between independent biological screening replicates ( Figure S1B). MAGeCK 37 ( Figure 1B) and Drug-Z 38 ( Figure S1C) analyses indicated that the KO of genes involved in the regulation of IFN-g signaling, JAK-STAT signaling, and the downstream transcriptional response, caused the strongest resistance, including IFNGR1, IFNGR2, JAK1, JAK2, STAT1, and IRF1 (Figure 1B), each of which had multiple gRNAs with significant enrichment specifically in the presence of IFN-g (Figures 1C and S1D). Changes in gRNA abundance were generally greater for HT-29, reflecting higher sensitivity to IFN-g and a faster growth rate than LS-411N ( Figure S1E). Identification of hits common to both cell lines ( Figure 1D) and STRING network analysis 39 revealed genes centered around IFN-g signaling, protein ubiquitination, RNA processing, and mammalian target of rapamycin (mTOR) signaling ( Figure S1F).
KO of mTOR, AKT1, and WDR24 were significantly associated with resistance to IFN-g, whereas negative regulators of mTOR, TSC1, and STK11 were sensitizing hits, consistent with the pleiotropic, immunosuppressive effects of rapamycin, and mTOR signaling potentiating IFN-g signaling. 40 The inactivation of genes involved in protein degradation such as tumor suppressor genes KEAP1 and FBXW7 has been previously implicated in sensitivity and resistance to cancer immunotherapy, respectively. 25,32 Interestingly, FBXW7 was a significant resistant hit in HT-29, but not LS-411N, where FBXW7 is already mutated. 41 Moreover, sensitizing hits included KO of SOCS1 and STUB1, 30 which are negative regulators of IFN-g signaling that function through inhibition and proteasomal degradation of JAK1 42 and IFNGR1. 43 Top-scoring regulators of apoptosis, CASP8, BAX, and MCL1, indicated the mode of cell death induced by IFN-g, and support the association of CASP8 mutations with immune evasion in TCGA pan-cancer analyses. 9 Finally, KO of autophagy-related genes enhanced cell death in the presence of IFN-g ( Figure 1D) (ATG2A, ATG5, ATF3, and ATF16L1), consistent with autophagy mediating cancer cell resistance to anti-tumor T cells. 29 Collectively, our CRISPR-Cas9 screens identified key nodes of resistance and sensitivity to IFN-g in colorectal cell lines for further study, with considerable overlap with clinical reports of ICB resistance in patients [5][6][7] and genetic screens interrogating cancer immune evasion in vitro 10,28,29,32 and in vivo [29][30][31] (Table S1).
Base editing mutagenesis screening of JAK1 with BE3-NGG In an attempt to analyze spontaneously acquired resistance to IFNg, we grew HT-29 cells in the presence of IFN-g for 2 months, but failed to derive resistant clones, necessitating the use of orthogonal approaches. JAK1 KO caused robust resistance to IFN-g in CRISPR-Cas9 screens, and mutation causes acquired resistance to ICB. 5,6 JAK1 somatic mutations in cancer are most frequently missense mutations (58.2%), with C->T or G->A transition mutations predominating (52.7%), which can be installed using CBEs (Figure 2A). Therefore, we set out to use base editing mutagenesis screens to assign functional scores to VUS in JAK1. To obviate potential toxicity associated with constitutive expression of deaminases, we generated doxycycline-inducible base editor 3 (iBE3) 21 HT-29 and LS-411N cell lines. The base editing activity reporter (BE-FLARE) 44 estimated base editing efficiencies of approximately 40% in HT-29 iBE3 ( Figure 2B). Despite both cell lines having similar ploidy (approximately 3n), base editing efficiency was considerably lower in LS-411N (approximately 15%) ( Figures S2A and S2B), and associated with apparent silencing of base editor expression ( Figure S2C). Since LS-411N is MSI with an inactivating mutation in MLH1, we also tested whether mismatch repair may affect base editing by KO of MLH1 in HT-29 iBE3 cells ( Figure S2D), but found that MLH1 was dispensable for base editing in this context ( Figure S2E). We deemed high editing efficiency important because clinical resistance to ICB is associated with homozygous mutations in IFN-g pathway components, often occurring with loss of heterozygosity. 5,6 Using a pooled library of 2,000 gRNAs, we tiled JAK1 in HT-29 iBE3 cells with exon-targeting gRNAs and gRNAs targeting JAK1 promoter regions, non-targeting (NT), intergenic targeting, and control gRNAs designed to introduce stop codons in 72 essential and 28 non-essential genes (Table S2). We adopted two screening approaches: a long-term proliferation screen and a short-term flow cytometry-based assay based on major histocompatibility complex (MHC-I) and programmed death-ligand 1 (PD-L1) induction with IFN-g ( Figure 2C). gRNAs predicted to install stop codons within essential genes were significantly depleted ( Figure 2D), achieving recovery of known essential genes in both screens ( Figure S2F). There was no relationship between JAK1 gRNA functional scores and the number of offtarget sites ( Figure S2G); however, the gRNA Rule Set 2 score 45 ( Figure S2H), or considering the immediate sequence context of the target cytidine ( Figure S2I), was somewhat predictive of gRNA performance. 17,18 Correlation between independent replicates ( Figure 2E) and the proliferation and fluorescenceactivated cell sorting (FACS) screens ( Figure 2F), was driven by highly enriched gRNAs after positive selection with IFN-g, representing candidate JAK1 LOF variants. As GOF variants were rare, we could only practically sort for JAK1 LOF cells by FACS, and so only recovered LOF gRNAs in the FACS screen ( Figure S3A). We selected 24 gRNAs for validation studies, representing 15 LOF and 5 GOF unique variants, mostly predicted to generate missense variants with clinical precedence in cancer ( Figure 2G; validation cohort). In addition, we included JAK1 Glu890 gRNA, which was unusual as it scored in the proliferation screens but not the FACS screens ( Figure 2G), and the Trp690* gRNA as a control; predicted to generate a nonsense mutation observed in a CRC patient that failed to respond to ICB. 6 Base editing mutagenesis of the IFN-g pathway To achieve a more comprehensive overview of the effect of missense mutations in the IFN-g pathway, we expanded our (E) Correlation between screening replicates. z-scores for gRNAs targeting JAK1 were compared between replicates and alternative screening assays. The shaded line represents the 95% confidence interval. (F) Correlation between different base editor screening assays for JAK1 variants in HT-29 iBE3 cells. (G) Identification of LOF and GOF alleles in JAK1 protein affecting sensitivity to IFN-g. z-scores for the base editing screens using FACS vs proliferation were plotted to select potential LOF (blue) and GOF (red) JAK1 variants. Labeling illustrates amino acid positions that were selected for further validation. All data are representative of two independent experiments or screens performed on separate days. See also Figure S2 and Table S2. base editor mutagenesis screens to include top hits of our CRISPR-Cas9 screens in HT-29 iBE3-NGG cells ( Figure 1B). We tiled JAK1, JAK2, IFNGR1, IFNGR2, STAT1, IRF1, B2M, and SOCS1 with 4,608 gRNAs, including the previous JAK1 gRNAs to serve as internal controls ( Figure 3A and Table S2). All of these genes had 3n ploidy, except B2M, which was 4n. 41 Although not a component of the IFN-g pathway, B2M was included because of its role in MHC-I presentation and anti-tumor immunity, but it was not a hit in our initial IFN-g survival screens as B2M variants should not have an effect on cell proliferation in vitro.
Proliferation and FACS screens were significantly correlated ( Figure 3B), as were independent replicate screens ( Figure S3B), each displaying a high level of enrichment of gRNAs predicted to introduce splice variants, stop codons, and start-lost mutations ( Figure 3B). Once again, JAK1 Glu890 gRNA was enriched in the proliferation screen, but not in the FACS screen. Such behavior was rare for most proteins, except for the transcription factor STAT1, where a cluster of predicted LOF missense mutations was enriched only in the proliferation screen ( Figure 3B), possibly indicating separation-of-function mutants. Encouragingly, we recovered validated gRNAs targeting JAK1 in this larger screen (Table S2 and later sections). In addition to protein truncating mutations, we used JAK1 LOF and GOF gRNAs from our validation cohort as a benchmark for setting the thresholds to call missense variants altering IFN-g pathway activity with high confidence ( Figure 3B).
Because of its short gene length and thus relatively few gRNAs predicted to install missense mutations, we only recovered highly enriched gRNAs predicted to install splice site or stop codon variants in B2M, and these only scored in the FACS screen, as expected. For the negative regulator of IFN-g signaling, SOCS1, LOF mutations were significantly depleted. Editing of JAK1, JAK2, IFNGR1, IFNGR2, and IRF1 predominantly gave rise to LOF missense mutations, but STAT1 was a notable outlier; it displayed a high proportion of GOF mutations (Figure 3B). Of STAT1 LOF missense variants, 66.7% were clustered around the SH2 and transactivation domains, compared with 6.7% of nonsense and splice LOF mutations ( Figure 3C). Conversely, 55.6% of STAT1 GOF mutations were within coiled-coil and DNA-binding domains, consistent with previous reports of GOF mutations within these domains in patients with chronic mucocutaneous candidiasis. 46 LOF missense mutations in IRF1 were enriched in the DNA-binding domain (88.9%), whereas SOCS1 LOF missense variants were enriched in the SOCS box and SH2 domains (84.2%) or within the JAK inhibitory region 42 (SOCS1 His61Tyr), demonstrating that base editing can highlight functional protein domains.
Comparison of base editing technologies for mutagenesis screening An analysis of amino acid mutations predicted from gRNA sequences suggested the BE3-NGG library targeted approximately 21.4% of the amino acids in JAK1. To improve the saturation of mutagenesis achievable with base editing, we used a Cas9 variant with a relaxed NGN PAM requirement, 47 generating BE3.9max-NGN. 18,48 Second, we sought to increase product purity by using a YE1-BE4max-NGN architecture that decreases non-C->T outcomes, 48,49 decreases Cas9-independent off-target editing, and improves editing precision by using an engineered deaminase (YE1) with a narrower editing window. 50,51 Finally, we used an ABE 22 (ABE8e-NGN) 52 to incorporate a wider variety of amino acid substitutions than can be achieved by C->T transitions alone. Using our panel of HT-29 base editor cell lines ( Figure S3C), we re-screened JAK1 with a library of 3,953 gRNAs (Table S2) targeting JAK1 exons ( Figure 4A). For NGN base editors, we detected significantly enriched gRNAs using all four PAMs ( Figure S4A). ABE cannot introduce stop codons, but predicted splice variants in JAK1, which could be introduced with both CBE and ABE, were significantly enriched over NT control gRNAs in all screens ( Figure S4B). Given the PAM utility and editing windows of each base editor, we predicted non-synonymous amino acid mutation coverage of JAK1 was improved to approximately 39.6% for BE4max-YE1-NGN, 50.8% for BE3.9max-NGN, 64.9% for ABE8e-NGN, and 85.1% when combining cytidine and adenine NGN mutagenesis. However, we cannot guarantee the editing efficiency of all gRNAs, so the absence of a significant score cannot be used as evidence for the lack of function of an amino acid position.
When combined, CBE and ABE editors can achieve substitutions of all 20 amino acids to at least two alternative amino acids. Substitution of amino acids with disparate chemical properties achieved larger average effect sizes, especially Leu->Pro missense mutations introduced with ABE, presumably because of the uniquely restricted 4 and c peptide bond angles available to proline ( Figures 4A and S4C). A comparison of functional scores with in silico predictions of variant effect (SIFT, PolyPhen, and BLOSUM62) demonstrated imperfect predictions in each case ( Figure S4D), implying that high-throughput experimentation is often required to complement bioinformatic prediction of variant effect. 53 Functional comparisons of BE3 and BE4max-YE1 editing of JAK1 confirmed the narrower editing profile of the YE1 engineered deaminase ( Figure S5A), with approximately 40.5% of JAK1 gRNAs predicted to edit only one cytosine (Table S3), but we observed a lower editing efficiency for BE4max-YE1-NGN compared with the wild-type (WT) deaminase ( Figure S5B), consistent with a decreased number of significant missense, splice, and stop codon variants compared with alternative NGN base editor architectures ( Figure S4B). Functional gRNAs present in both BE3 and BE4max-YE1 screens had target cytosines within the YE1 5-7 activity window (e.g., Asp775Asn gRNA 908510028), whereas out-of-window targeting gRNAs were not enriched in the BE4max-YE1 screens (e.g., Trp690* gRNA 908510274) ( Figures S5A and S5C).

Deep mutagenesis of JAK1 reveals LOF and GOF variants with clinical precedence
To aid in the interpretation of our mutagenesis screens, we compiled a database of clinical mutations and aligned this with predicted base edited JAK1 variants. We defined clinical precedence as a non-synonymous mutation of the residue in COSMIC, 12 TCGA, ClinVar, 54 literature on JAK1 mutations with known effect, 1 and data from patients receiving ICB, where cancer exome sequencing data are publicly available, 6,8,55-60 but absence from gnomADv3.1 61 (Table S3). LOF and GOF variants made with CBE were more likely to have clinical precedence than ABE variants, with 88% of significant CBE variants  Figure S3 and Table S2. identified occurring at residues with precedence of mutation in cancer genomes, 32% of which were predicted to recapitulate the amino acid substitution with CBE (vs. 6% for ABE), perhaps reflecting the APOBEC deamination signature in cancer. 13 Our analysis revealed candidate GOF variants in the JAK1 pseudokinase domain with clinical precedence in cancer. gRNAs targeting position Arg724 were significantly depleted with IFN-g ( Figure 4A). The predicted base edited variant, Arg724His, has been implicated in activating JAK1 signaling in acute lymphoblastic leukemia through dysregulating intramolecular inhibition of the kinase domain. 1 Another GOF position, JAK1 Val658, is mutated in acute myeloid leukemia (AML); this residue is structurally analogous to JAK2 Val617, which is commonly mutated in polycythemia vera. 1,2 CBE and ABE screens converged on a cluster of GOF variants in the C-terminus of the kinase domain (Met1099, Arg1103) in a known protein-protein interaction motif for SOCS1 42 ( Figure 4A), a significant negative regulator in our CRISPR-Cas9 screens. These variants presumably disrupt this interaction, increasing JAK1 protein abundance and activity ( Figure 4B). Indeed, the amplification of SOCS1 has been found in patients that failed to respond to ICB, 7 implying that this regulatory mechanism may be of clinical relevance.
LOF positions included Gly887 ( Figure 4A), which is within the kinase active site, with the crystal structure, 42 suggesting that mutation of this residue would negatively affect Mg 2+ and adenosine triphosphate(ATP)/adenosine diphosphate coordination ( Figure 4B). Other LOF mutations involving kinase catalytic residues included Asp1003 (proton acceptor), and Asp1021 (within the DFG motif), which were detected with increased (NGN) saturation ( Figure 4A). ABE screens were more likely to detect sites of post-translational modification because of their ability to modify tyrosine, threonine, serine (phosphorylated), and lysine (ubiquitinated), revealing Tyr993, and the known activating Tyr1034 phosphosite as candidate LOF positions ( Figure 4A) and Lys267 as a putative GOF site.
Of the candidate LOF variants found in cancer biopsies (Table S3), Gly655Asp, Gly182Glu, and Gly590Glu 56 ( Figure 4A) were all VUS detected in patients who failed to respond to ICB. In addition, JAK1 Asp775Asn has been independently verified as a LOF variant in melanoma. 6 The overexpression of FLAGtagged WT JAK1 in HEK293T cells resulted in a pSTAT1 signal, even in the absence of IFN-g, and supraphysiologic stimulation with IFN-g, whereas the JAK1 Gly590Arg mutant failed to induce STAT1 phosphorylation to the same extent in either context (Figure 4C), verifying Gly590 as a bona fide LOF mutation.
Functional validation of variants conferring altered sensitivity to IFN-g We set out to functionally validate 24 gRNAs comprising our JAK1 validation cohort ( Figure 2G) in an arrayed format, with multiple assays assessing cell proliferation, signaling, protein expression, RNA expression, and flow cytometry (Figures 5A-5C). This analysis was germane to screening results from multiple base editing modalities, because of their convergence on JAK1 residues within the validation cohort (e.g., Arg108, Gly590, Asp775, Gly887, and Met1099) ( Figure 4A and Table S3). The growth of HT-29 iBE3 cells with engineered JAK1 variants in the presence of IFN-g tracked with screen results, with GOF variants having no survival benefit and LOF variants having robust resistance to IFN-g, relative to controls ( Figures 5A and S5D).
Many of the candidate LOF variants had decreased levels of pSTAT1 and IRF1 induction ( Figures 5B and S5D). The Met1099 and Arg1103 GOF variants had increased levels of JAK1 protein and JAK-STAT signaling, consistent with disruption of the SOCS1 binding interface and decreased E3 ubiquitin ligase-mediated destruction. 42 Surprisingly, the Gly590 LOF variants also had increased levels of JAK1 protein, despite decreased sensitivity to IFN-g. We speculated that increased JAK1 Gly590Arg protein could also be attributable to altered binding to SOCS1; however, we did not observe any change in binding in co-immunoprecipitation experiments ( Figure S5E). JAK1 706/707 gRNA targets a splice region and had severely decreased JAK1 protein expression, similar to the clinical Trp690* nonsense control ( Figure 5B). The Glu1123 splice variant decreased JAK1 RNA abundance to levels comparable with Trp690*, which we presumed was targeted for nonsense-mediated decay. However, basal JAK1 variant RNA expression levels were generally only modestly affected, and RNA expression was not entirely indicative of JAK1 protein levels, consistent with the complex post-translational control of JAK1. 42 Signaling assays for LOF variants were performed with pre-selection with IFN-g to decrease contributions from remaining WT, unedited cells. However, similar results were obtained using an HT-29 iBE3 single cell clone with superior editing efficiency, which did not require prior selection with IFN-g ( Figure S5F).
Next, we generated 40 additional knock-in lines using base editing to install mutations in other IFN-g pathway genes in HT-29, and a primary MSI colorectal cancer tumor organoid, CRC-9 (harboring FBXW7 and TP53 driver mutations), and used flow cytometry to assess the induction of MHC-I and PD-L1 expression upon stimulation with IFN-g ( Figure 5C). LOF mutations decreased responses to IFN-g, except for LOF mutations in the negative regulator SOCS1, which increased the induction of MHC-I and PD-L1 ( Figure S6A). Notably, we confirmed separation of function variants specific to STAT1 ( Figure 3B), which had minimal effects on induction of MHC-I and PD-L1 ( Figures 5C and S6A), but conferred a significant proliferation advantage in the presence of IFN-g ( Figure S6B), highlighting the value of using two screening assays and base editing to gain new, as yet poorly understood, insights into STAT1 function. In contrast, the Asp257 STAT1 GOF mutation significantly increased sensitivity to IFN-g in both cell models ( Figure S6B); this variant effect was stronger than the putative JAK1 GOF variants tested, of which only the JAK1 Met1099 variant displayed significantly increased IFN-g sensitivity in HT-29.
Verification of base editing genotypes with nextgeneration sequencing We performed amplicon sequencing of the endogenous JAK1 loci to unambiguously assign base edited genotypes ( Figure 6A). This analysis confirmed accurate predictions of base editing outcomes, detecting C->T editing focused within the BE3 activity window (approximately 4-9 relative to the PAM at position 21-23), with a minority of gRNAs (22.7%) exhibiting lower frequency edits upstream or downstream (Figures 6B and S6C). Collectively, this resulted in two unanticipated coding mutations from the validation cohort (JAK1 Asp1122Asn and Gly590Glu) caused by editing at protospacer positions 2, 3, and 11. LOF variants were enriched in the presence of IFN-g without exception (JAK1 Glu890 was modestly enriched), verifying the associated resistance phenotype and LOF classification ( Figure 6A). Co-enrichment of LOF variants with synonymous mutations (63.6% of gRNAs) implied selection for edited cells, with co-occurring neutral edits.
To more comprehensively assign the genotypes of base edits en masse, we used single-cell DNA-sequencing (scDNA-seq) ( Figure 6C). Of the 87 gRNAs assigned JAK1 edits, a comparison of scDNA-seq data to amplicon sequencing showed strong concordance between genotypes (Table S4). Combining amplicon and scDNA-seq datasets facilitated genotyping of edits associated with 98/665 JAK1 targeting gRNAs. For gRNAs where we detected JAK1 editing, predictions of amino acid changes from gRNA sequences overlapped with observed protein changes from amplicon sequencing or scDNA-seq in all cases, with 81% of predictions capturing all observed amino acid changes. An advantage of scDNA-seq over bulk amplicon sequencing is the assessment of the penetrance of editing (zygosity). Based on clinical data, we expect that homozygous editing of all JAK1 alleles is required for a LOF phenotype, 5,6 and this was the most frequently observed editing outcome (59% of edits). Of homozygous editing, 83% was focused within the 4-9 base editing window, compared with 76% for homozygous and heterozygous edits combined ( Figure 6C). Moreover, edits outside of this activity window were private to a smaller proportion of cells with the same gRNA assignment.

Article
For 37 gRNAs, we did not detect cells with JAK1 edits; 4 of these gRNAs did not have target cytosines within the base editing window, 6 had exclusively GC targets in the editing window, and 2 had gRNAs with poly-T tracts that could act as U6 transcription termination signals. The remaining 25 gRNAs were expected to have a higher propensity to install base edits, giving an estimation of the false-negative rate of iBE3 cytidine base editing screens (28.7%) using our strict criteria for a calling edited genotypes (STAR Methods), although this rate varies based on the editing efficiency of each cell system. Nonetheless, the falsenegative rate reinforces that we cannot interpret the lack of a significant score as evidence for a residue not being important for protein function. 18 Taken together, these data represent a comprehensive profile of base editing outcomes at endogenous DNA loci and in single cells and indicate the predictability and precision with which variants can be installed using transient expression of base editors from a doxycycline-inducible system. For variants where the edited genotype was not assigned through sequencing, variants are predictions from gRNA sequences and based on the profiling of base editing outcomes at endogenous loci.

Classified JAK1 missense mutations alter sensitivity to autologous anti-tumor T cells in primary human tumor organoids
To understand the broader functional implications of base editing variants, we mined an extensive collection of cancer cell models (n = 1,357) with associated exome sequencing data 41 for pre-existing JAK1 LOF and GOF variants discovered here. The AML cell line OCI-M1 harbored the JAK1 Val658Phe GOF mutation, and 10 cell lines had homozygous inactivating frameshift or nonsense JAK1 mutations. HT55 (CRC) and K2 (melanoma) cell lines harbored homozygous Glu1051Gln and Ala760Val putative JAK1 LOF missense mutations, respectively ( Figure S7A). As predicted, HT55 and K2 failed to respond to IFN-g compared with JAK1 WT cancer cell lines, as measured by failure to induce MHC-I and PD-L1 expression ( Figure 7A). OCI-M1 had relatively high basal expression of MHC-I and PD-L1, consistent with increased levels of JAK-STAT signaling, with further induction of PD-L1 upon IFN-g stimulation. The endogenous C->T mutation in K2 cells was amenable to correction by adenine base editing. ABE8e-NGN-mediated reversion of this JAK1 mutation led to restoration of IFN-g sensitivity ( Figure 7B), verifying that this variant is  Figure 6A in the absence of IFN-g. (C) Single-cell DNA sequencing of base editing in HT-29 iBE3 cells across 50 gRNAs reveals C->T (or G->A) editing focused in the gRNA activity window. Penetrance (zygosity) 0/1/1 is heterozygous (het), and 1/1/1 is homozygous (homo). The proportion of cells with the same gRNA assignment harboring that edit is indicated. See also Figure S6 and Table S4. responsible for resistance to IFN-g. Interestingly, all of these cancer cell lines were derived before ICB was widely available, which suggests these variants arose from in vivo immunoediting 3,9 rather than resistance to therapy.
To further assess the relevance of our findings in other cell models and in a more translational setting, we applied base editing to CRC-9 primary tumor organoids, derived from an MSI colorectal cancer patient where autologous, tumor-reactive T cells have been derived from the patient's peripheral blood mononuclear cells (PBMCs). 62,63 First, we reperformed the BE3-NGG base editing screen of the IFN-g pathway components in the CRC-9 tumor organoid. There was a significant correlation between independent screening replicates (Figure S7C) and, crucially, with screening results from HT-29 iBE3 NGG ( Figure 7C). These data indicate that our base editing variant data are broadly applicable and not private to a particular cell model. Furthermore, we validated clinically observed JAK1 missense variants in CRC-9 tumor organoids using individual gRNAs, as shown by altered sensitivity to IFN-g in three-dimensional growth assays, with LOF missense mutations at JAK1 residues 108, 590, and 775 conferring resistance ( Figure S7D).
Next, we used a co-culture of matched tumor-reactive T cells with genetically engineered tumor organoids ( Figure 7D) to assess T cell-mediated killing by flow cytometry (Figure 7E). After (F) Quantification of T cell-mediated killing of autologous tumor organoids from flow cytometry analysis. Data represent the average ± standard deviation of three biological replicates and were compared against parental co-culture controls using an unpaired, two-tailed Student's t-test (**p < 0.01, *p < 0.05). NT, nontargeting gRNA; ø par., parental tumor organoid. All data are representative of two independent experiments performed on separate days. See also Figure S7. enrichment for tumor-reactive populations and expansion, cocultured PBMCs were exclusively CD3 + , implying a high proportion of T cells 62,63 (Figure S7E). Strikingly, all JAK1 LOF mutant tumor organoids had significant resistance to anti-tumor T cellmediated killing relative to WT controls, with some mutants achieving survival comparable with antibody blockade of MHC-I or growing tumor organoids in the absence of T cells (Figure 7F). Conversely, the GOF mutant Met1099Ile increased sensitivity to T cell-mediated attack. Antibody-mediated neutralization of IFN-g in the co-culture medium significantly alleviated cytotoxicity in the WT and GOF tumor organoids, but had no effect in JAK1 LOF cells (Figures S7F and S7G), consistent with a high level of IFN-g release from autologous anti-tumor T cells upon exposure to tumor cells, 63 which was modestly increased with nivolumab ( Figure S7H). Taken together, these data illustrate that IFN-g pathway-variant maps from base editing screens may be predictive of anti-tumor immunity.

DISCUSSION
In this report, we perform 20 screens with CRISPR-Cas9 and base editors to systematically catalogue the genetic dependencies of IFN-g response in CRC cells and map more than 300 predicted missense mutations affecting IFN-g pathway activity. Through the use of multiple cytidine and adenine base editors, this study systematically probes protein structure and function throughout an entire signaling pathway. We provide BE-view as an online resource for exploration of these data: www.sanger.ac.uk/tool/be-view.
Tumor cell sensitivity to IFN-g is an important determinant of ICB response in multiple tumor types. 5-8 JAK1 is mutated in approximately 10% of CRC and 6% of skin cutaneous melanoma, with a decrease in survival for melanoma patients with deleterious JAK1 alterations. 6 We detected known LOF variants (JAK1 Asp775Asn, Trp690*) 6 and assigned LOF to VUS in JAK1 that may have contributed to primary or acquired resistance to ICB resistance in the clinic (e.g., JAK1 Gly590Arg, Gly182Glu, Gly655Asp, and Pro674Ser). 56,58 We also discovered a splice mutation in JAK1 as a LOF variant (Arg110 splice variant); however, this tumor mutation was recorded in a patient with a partial response to anti-CTLA-4. 56 This highlights that the presence or absence of LOF variants in the IFN-g pathway in a tumor biopsy is not an absolute determinant of ICB response 64 ; rather, the outcome depends on multiple factors, including the penetrance of the mutation itself (i.e., zygosity), tumor clonal architecture, co-occurring mutations, tumor mutational burden, oncogenic signaling, tumor microenvironment, antigen presentation, and immune checkpoint engagement. 4,65 Many of the variants we discovered with functional effects on IFN-g signaling had clinical precedence, implying that immunoediting in cancer, particularly for immune-hot tumors, may be more prevalent than previously thought. 3 CRISPR-Cas9 screening identified druggable targets that sensitized tumor cells to IFN-g when inactivated, such as MCL1 and TBK1, highlighting potential ICB combination therapies in CRC. In line with this, TBK1 inhibition has been reported to increase immune reactivity to tumor organoids ex vivo. 66 Interestingly, inactivation of KEAP1, FBXW7, NF2, and STK11, modulated sensitivity to IFN-g, emphasizing important non-cell autonomous roles for these tumor suppressor genes. Although we found KO of STK11 sensitized to IFN-g, STK11 mutation is associated with resistance to immunotherapy in lung adenocarcinoma, 67 implying potential tissue-type or genotype-specific differences, and highlighting that our reductionist in vitro approach does not consider the potential effects on immune cells. KO of NF2 resulted in an increased resistance to IFN-g and has also been linked to BRAF inhibitor resistance, 45,68 consistent with an overlap between ICB resistance and mitogen-activated protein kinase inhibitor resistance pathways, 69 with possible implications for the efficacy of ICB in melanoma patients pre-treated with BRAF inhibitors.
IFN-g signaling through the JAK-STAT pathway is not only relevant for cancer immunotherapy, but also underpins pathology in myeloproliferative neoplasms, chronic mucocutaneous candidiasis, primary immunodeficiency, and several inflammatory diseases. 1,2 The molecular understanding of JAK-STAT signaling has been hindered by the lack of a full-length crystal structure of JAK1 and the complex intra-molecular regulation by the JAK1 pseudokinase domain. 1 We report base editing screens mapping LOF and GOF variants in key regulatory regions across JAK1, including catalytic residues (ATP coordination), post-translational modifications, the pseudokinase-kinase domain interface, and inter-molecular protein-protein interactions with SOCS1, demonstrating that base editing may be used to understand complex protein biology without prior detailed structural information. 70 Our study provides a resource for improving the interpretation of IFN-g pathway variants in diseases such as cancer, and highlights the potential of semi-saturating base editing mutagenesis, which we envisage will complement SGE, 71 in silico, 72 and prime editing 73 approaches in establishing the functional consequence of genetic variation.

STAR+METHODS
Detailed methods are provided in the online version of this paper and include the following:

INCLUSION AND DIVERSITY
We support inclusive, diverse, and equitable conduct of research.  Primary cell cultures PBMCs and CRC-9 tumor organoids were from the Netherlands Cancer Institute (NKI). CRC-9 is genetically female. Derivation of tumor organoids, enrichment of tumor reactive T cell populations from patient PBMCs were performed as described. 62 Cells were maintained in a 5% CO 2 , 95% air, humidified incubator at 37 C.

Cell culture
Where indicated, CellTiter-Glo proliferation assays (Promega) were performed to assess drug response following manufacturer's instructions. For the long-term culture of HT-29 in IFNg to derive resistant cells, we treated cells with a pre-optimized dose that killed 80% of parental cells (1,000 U/mL; Thermo Fisher Scientific), and refreshed the media with the addition of IFNg twice a week for two months.
Tumor organoid culture Growth and maintenance of CRC-9 tumor organoids in 3D was achieved by growth in 80% basement membrane extract (BME; R&D Systems). Co-culture killing assays were performed as described. 62 Briefly, PBMCs were cultured in anti-CD28 coated plates for 24 h with IL-2 (150 U/mL; Thermo Fisher Scientific). CRC-9 cells were pre-stimulated with IFNg (Thermo Fisher Scientific, 400 U/mL) overnight to increase MHC-I expression, then seeded in suspension in non-tissue culture treated 96 well plates at a 3:1 E:T ratio for 72 h, with or without anti-CD28 coating, nivolumab (20 mg/mL; Selleckchem), MHC-I blocking antibody (W6/32; 50 mg/mL) and IFNg neutralizing antibody (NIB42; 60 mg/mL) in RPMI supplemented with human serum and primocin (Invivogen). T cell mediated killing of CRC-9 tumor organoids is dependent on MHC-I, pre-exposure of organoids to IFNg to increase MHC-I expression and antigen presentation, but not PD-1 inhibition with nivolumab, or CD28 co-stimulation ( Figure S8). Cells were harvested and stained with anti-CD3 FITC antibody (UCHT1; Thermo Fisher Scientific, 1:100), washed in FACS buffer before the addition of DAPI and flow cytometry analysis. 123count eBead counting beads (Thermo Fisher Scientific) allowed for quantification of absolute cell counts based on volumetric measurements from bead counts. The IFNg ELISA assay was performed on neat cell culture medium from the co-culture according to manufacturer's instructions (Thermo Fisher Scientific; #EHIFNG). Base editing screens with tumor organoid CRC-9 was performed in suspension.
base editing window and used VEP 84 to assign amino acid changes. For BE3 and ABE8e we assumed a lenient window of 4-9 and for BE4max-YE1 NGN we used a window of 5-7, where 20-23 is the PAM. We focused our analysis on VEP output of MAINE selected canonical protein coding transcripts. For annotation of edit consequence, we consolidated multiple predicted consequences by giving priority to the most deleterious as follows: stop gain > start loss > splice variant > missense > UTR > synonymous variant. For base editing screens, we filtered out samples with <100 gRNA read counts for any sample in either replicate, and one gRNA that was overrepresented (>50,000 reads) in the library. For CRC-9 tumor organoid screens, we also excluded seven gRNAs from downstream analysis that had > 3-fold read count difference in the control samples between the two experiments. For annotation of post-translational modifications, we used the PhosphoSitePlus database. 59 qPCR 72 h after base editing (induced by the addition of doxycycline), RNA was extracted and genomic DNA was removed (RNeasy columns and DNase I; Qiagen), followed by cDNA synthesis with SuperScript IV and random hexamers, and analysis using SYBR Green reagents on the Step One Plus (Thermo Fisher Scientific), with the following primers: Human JAK1 5 0 -GAGACAGGTCTCCCACAAA CAC-3 0 , 5 0 -GTGGTAAGGACATCGCTTTTCCG-3 0 , Human GAPDH 5 0 -GTCTCCTCTGACTTCAACAGCG-3 0 , 5 0 -ACCACCCTGTTGC TGTAGCCAA-3 0 .

Giemsa staining
After six days of selection with IFNg (1500 U/mL; Thermo Fisher Scientific), cells were washed with PBS, fixed with 4% PFA for 20 min and then stained with Giemsa working solution (1X in water; Sigma-Aldrich) for 2 h at room temperature with gentle rocking. Wells were rinsed with deionized water three times and then allowed to dry before images were taken by scanning.

QUANTIFICATION AND STATISTICAL ANALYSIS
Statistical tests, exact value and description of n, definition of center, dispersion and precision measures are described in the figure legends. No randomization was performed and no statistical methods were used for sample size determination. For CRISPR-Cas9 screening analysis with MAGeCK, p < 0.05 and a false discovery of <5% were used as significance thresholds. For Student's t-test, significance was defined as p < 0.05.

ADDITIONAL RESOURCES
BE-view is an R Shiny app that facilitates exploration of our data: www.sanger.ac.uk/tool/be-view.