Unleashing Cell-Intrinsic Inflammation as a Strategy to Kill AML Blasts

Immunotherapy has revolutionized the treatment of some malignancies. CD19-directed chimeric antigen receptor (CAR) T-cell therapies in B-cell acute lymphoblastic leukemia (B-ALL; ref. 1) and immune-checkpoint inhibitors in adult malignancies, such as melanoma (2) and non–small cell lung cancer (3), highlight transformative advances in cancer treatment. Nevertheless, some forms of cancer, such as acute myeloid leukemia (AML), remain poorly treated despite improvements in targeted therapy. AML has yet to be successfully addressed by newer immunotherapy approaches beyond hematopoietic stem cell transplantation. In AML, however, an alternative therapeutic approach might be to exploit innate inflammatory pathways intrinsic to the leukemia cells.

Genetic events perturbing hematopoietic stem or myeloid progenitor cells can give rise to AML blasts and prevent maturation to terminally differentiated immune cells, which normally possess a limited life span (4). We hypothesized that dysregulation of inflammatory pathways could contribute to the maintenance of the leukemic state in these immune cells gone awry. Moreover, we sought to decipher whether the rewiring of cell-inherent inflammation can be exploited as cell-intrinsic, self-directed immunotherapy. Myeloid leukocytes represent the dominant cellular components of the innate immune system (5). Their mission is to protect the organism, sometimes at the expense of their own survival. Leukemic blasts, however, exhibit rapid clonal expansion at an immature cell differentiation state (6).

The innate immune response constitutes the first line of defense against foreign pathogens. Activation of innate immune cells relies on the discrimination between self and foreign antigens, and inflammation is the immune system's response to the detection of harmful stimuli. Although significant discoveries have broadened our understanding of inflammatory signaling in mature myeloid immune cells, the effects of enhanced or repressed intrinsic inflammation within an AML cell have yet to be fully elucidated. Several studies in B-cell malignancies, however, hint at potential therapeutic opportunities for perturbing inflammatory signaling. For example, the activating L265P mutation in the common Toll-like receptor (TLR) adaptor MyD88, prevalent in multiple lymphoid malignancies, leads to the activation of NFκB via Bruton tyrosine kinase (BTK) and the interleukin-1 receptor-associated kinases (IRAK1 and IRAK4; refs. 7, 8). Targeting BTK with ibrutinib has shown high response rates in patients with Waldenström macroglobulinemia with MyD88 mutations (9). Intriguingly, direct activation of members of the TLR family has also been reported to induce apoptosis in subsets of multiple myeloma and ALL (10, 11).

A strong interconnection between inflammatory signaling and leukemia also comes in the discoveries of a role for TET2 mutations in association with both AML and increased risk for atherosclerosis. Loss-of-function mutations in TET2, an epigenetic regulator, are present in approximately 15% of myeloid malignancies, causing DNA hypermethylation and enhanced self-renewal in hematopoietic stem cells. In addition, TET2 is one of the most commonly mutated genes in clonal hematopoiesis of indeterminate potential (CHIP; refs. 12, 13). In the context of CHIP, TET2 mutations increase the risk of subsequent development of a hematologic malignancy but are also strongly associated with atherosclerotic cardiovascular disease and inflammation (14). For instance, Tet2-deficient macrophages exhibit an increase in NLRP3 inflammasome-mediated interleukin-1β secretion (15). Similarly, plasma interleukin-8, a CXC chemokine not present in mice, was increased in humans with TET2 mutations (14).

IRF2BP2 provides another example of a gene involved in both cancer and inflammatory conditions. IRF2BP2 was initially identified as a corepressor molecule for IRF2 in a yeast two-hybrid library screen (16). Subsequently, it was found to be involved in fusion oncoproteins in a number of malignancies, including AML, and implicated in inflammatory diseases. Studies on IRF2BP2 and leukemia are restricted to a handful of case reports (17–21) identifying IRF2BP2 as a fusion partner to RARA, leading to the phenotype of acute promyelocytic leukemia. In addition, IRF2BP2 was reported to be fused to CDX1 in mesenchymal chondrosarcoma (22). Interestingly, a small number of studies also link IRF2BP2 to inflammation in cardiovascular disease and sepsis (23, 24).

To test the hypothesis that AML cells depend on the proper regulation of inflammatory networks, we analyzed large data collections from primary AML samples and functional genomic screens. Applying CRISPR/Cas9 genome editing technology, the Broad Institute Cancer Dependency Map has screened close to 800 cancer cell lines with a genome-scale small guide RNA (sgRNA) library (25, 26). We integrated data from the Cancer Dependency Map with two independent genome-wide screens, one using RNAi (27, 28), and a second CRISPR/Cas9 screen with a unique library (29). An advantage to performing large-scale screens across hundreds of cancer cell line models is the ability to identify disease-selective versus pan-essential targets, which should increase the therapeutic index of identified targets. We thus focused on novel dependencies that are stronger in AML than in other cancer models screened. Herein, we identified and validated IRF2BP2 as a strong immune-related dependency in AML and elucidated its mechanism of action as a gatekeeper to prevent suicidal inflammatory signaling in AML cells.

To test the hypothesis that inflammatory pathways are inherent to AML blasts, we performed single-sample gene set enrichment analysis (ssGSEA) across large collections of primary AML samples versus the hallmark gene set collection of the Molecular Signature Database (MSigDB). ssGSEA is an extension of GSEA that calculates separate enrichment scores for each pairing of a sample and gene set. Each ssGSEA enrichment score represents the degree to which the genes in a particular gene set are coordinately upregulated or downregulated within a sample. Unsupervised clustering applied to the ssGSEA projection of the 179 primary AML samples from The Cancer Genome Atlas (TCGA) in the space of hallmark gene sets identified an AML subgroup of approximately 40% of all samples enriched for immune and inflammatory pathways (Fig. 1A–C). Quantitatively, the ssGSEA z-score enrichment for the canonical pathway HALLMARK_INFLAMMATORY_RESPONSE was significantly higher in these immune/inflammatory samples when comparing them with all other AML samples within the TCGA collection (Fig. 1D). Moreover, we were able to fully recapitulate these findings in three additional primary AML collections, namely from the European Hovon study group (GSE14468), the Beat AML trial, and the Target AML data collection from children with AML (Supplementary Fig. S1A–S1C). We estimated the lineage enrichment within AML blasts in cohorts of primary AML samples using a gene-expression deconvolution method implemented in CIBERSORT (30). We identified a positive correlation between the enrichment of inflammatory response and enrichment of a monocytic lineage signature in these primary AML samples (Supplementary Fig. S1D). Consistently, the AML group with a high immune/inflammatory response displayed a significant enrichment for FAB M4/M5 samples as compared with the noninflammatory counterpart (Supplementary Fig. S1E).

As part of the Broad Institute Cancer Dependency Map v20q3, 23 AML cell lines were screened with a genome-wide CRISPR/Cas9 library and compared with 766 non-AML cancer cell lines. Twenty-one days after transduction, we sequenced the sgRNAs and quantified their relative abundance as CERES dependency scores (31). This dropout screen allowed us to identify sgRNAs depleted relative to the input, suggesting that the genes targeted by those sgRNAs are genetic dependencies. We selected as top hits those sgRNAs that were more highly depleted in AML cell lines relative to other cancer cell lines and identified 60 genes on which AML cells were significantly more dependent than other cancer cell lines (based on the empirical Bayes statistics for differential dependency with significance cutoffs: effect size ≤ −0.3, adjusted P ≤ 0.10; Fig. 1E and Supplementary Table S1A). GSEA performed for the genome-wide differential dependency scores in AML versus non-AML cancer cell lines identified several immune and inflammatory Gene Ontology gene sets associated with depletion in AML (Supplementary Fig. S1F). To enhance the robustness of our AML dependency findings, we interrogated two additional genome-wide screens for AML-enriched dependencies and intersected these results: Only five genes were significantly enriched in all three screens using these strict criteria (Fig. 1E; Supplementary Tables S1A–S1C). Within this core group of five genes, IRF2BP2 stood out as the strongest when comparing the differential CERES scores in the CRISPR (Avana) AML cell lines with high enrichment for monocytic lineage compared with all other AML cell lines (Fig. 1F). A dependency on IRF2BP2 correlated strongly with enrichment for the monocytic lineage in both CRISPR (Avana) and RNAi (DEMETER2) screens (Supplementary Fig. S1G). Of note, IRF2BP2 was also the single unexplored dependency among the other well-characterized candidates, namely MYB, CBFB, SPI1, and LMO2, scoring in all three data sets (Fig. 1E and G).

Comparing the dependency scores of all AML cell lines with either other hematopoietic or solid tumor cell lines within the CRISPR (Avana) screen, AML cells were significantly more dependent on IRF2BP2 than any other group, representing the only cancer type with a mean dependency score of less than −0.5 (Fig. 1H). Gene expression is one of the strongest predictive biomarkers of dependencies (25). Indeed, when profiling IRF2BP2 expression across cancer cell types, we found that IRF2BP2 expression was highest in AML cell lines (Supplementary Fig. S1H) and AML patient samples (Supplementary Fig. S1I). However, we did not observe a significant correlation between IRF2BP2 expression and dependency on IRF2BP2 in AML cell lines across independent dependency screens (Supplementary Fig. S1J), likely due to high IRF2BP2 expression with a low variation in AML. Moreover, we discovered that IRF2BP2 was marked by a superenhancer in all 71 AML samples in a collection published by McKeown and colleagues, consistent with its high level of expression in AML (ref. 32; Supplementary Fig. S1K and S1L).

We next sought to validate IRF2BP2 as a dependency in vitro and in vivo. We transduced the AML cell lines MV4-11, U937, MOLM13, THP1, and NB4 with IRF2BP2-directed CRISPR guides used in the genome-wide screen and confirmed guide efficacy by Western blot (Fig. 2A; Supplementary Fig. S2A). We observed a marked decrease in the growth of cells infected with IRF2BP2 targeting guides in comparison with nontargeting CRISPR guides (Fig. 2A; Supplementary Fig. S2A). In addition, using an orthogonal shRNA approach, we validated the dependency on IRF2BP2 using the shRNAs from the screen (Fig. 2B) and by targeting IRF2BP2 with unique doxycycline-inducible hairpins (Supplementary Fig. S2B and S2C). To account for potential off-target effects of these inducible hairpins, we performed site-directed mutagenesis of an open reading frame (ORF) for IRF2BP2, introducing a silent point mutation at the binding site of the shRNAs. We then expressed this ORF in MV4-11 cells, followed by knockdown of IRF2BP2 (Fig. 2C). When these cells were exposed to doxycycline, we no longer observed any effect on viability (Fig. 2C), confirming that the observed effect of the shRNA is on target for IRF2BP2. We next transduced luciferase-expressing MV4-11 cells (MV4-11-luc) with these inducible hairpins and the resistant, mutated IRF2BP2 ORF. These cells were then injected into NOD/SCID/IL2rγ^null (NSG) mice. After leukemia engraftment was documented by bioluminescence, the mice were fed a doxycycline-containing diet for 2 weeks. We then assessed for leukemia burden by bioluminescence imaging (Fig. 2D and E) and by determining the chimerism of the human isoform of the pan-leukocyte marker CD45 in peripheral blood (PB) and bone marrow (BM) by flow cytometry (Supplementary Fig. S2D and 2F). Both approaches resulted in a significant reduction in leukemia burden in mice with leukemia cells containing IRF2BP2 targeting hairpins compared with nontargeting shRNAs or mice injected with leukemia cells containing the rescue ORF and the IRF2BP2 targeting shRNAs (Fig. 2E and F).

To study the acute loss of IRF2BP2, we leveraged an IRF2BP2 degrader system applying a previously reported technology known as the degradation tag (dTAG) system to selectively degrade FKBP12^F36V-tagged fusion proteins (33). This degradation approach relies on the small-molecule dTAG-13 that binds to the FKBP12^F36V tag and recruits the E3 ubiquitin ligase cereblon, thereby inducing proteasome-mediated degradation of the FKBP12^F36V-tagged target. We generated AML cells with degradable IRF2BP2 by coinfection with FKBP12^F36V-tagged IRF2BP2 and an endogenous IRF2BP2-targeting sgRNA. Exogenous tagged IRF2BP2 was expressed at comparable levels to the endogenous IRF2BP2 expression level with effective depletion of endogenous IRF2BP2 by sgIRF2BP2 (Fig. 3A; Supplementary Fig. S3A). We found IRF2BP2 degradation to be maximal at 100 to 200 nmol/L dTAG-13 with maximal loss after 12 hours of treatment with dTAG-13 (Fig. 3A and B). Additionally, fractionation of cellular compartments revealed degradation of IRF2BP2 in the cytoplasmic and nuclear extracts, including the chromatin-bound and soluble fractions (Supplementary Fig. S3B). Degradation of IRF2BP2 translated into a reduction of AML cell viability and decreased colony formation capacity (Fig. 3C and D). AML cells tolerated expression of exogenous IRF2BP2 without depletion of endogenous IRF2BP2 and showed a comparable growth rate to cells expressing an ORF for GFP (Fig. 3C). Mechanistically, cells treated with dTAG-13 underwent cell death with hallmarks of apoptosis, as evidenced by an increase in annexin V/PI-positive cells (Fig. 3E) and induction of cleaved caspase-3 (Fig. 3F) and cleaved caspase-8 (Fig. 3G). Similarly, we also saw evidence of apoptosis using CRISPR-based knockout of IRF2BP2 (Supplementary Fig. S3C and S3D). We did not observe markers of pyroptosis [cleaved Gasdermin D (GSDMD)] nor necroptosis (phospho-MLKL) over 24 hours after degradation of IRF2BP2 (Supplementary Fig. S3E and S3F).

We next determined whether there is a therapeutic window between targeting IRF2BP2 in AML cells versus normal hematopoietic progenitor cells. Because of the low efficiency of lentiviral transduction of Cas9 into CD34⁺ hematopoietic stem and progenitor cells, we nucleofected human CD34⁺ healthy BM cells with Cas9–sgRNA ribonucleoprotein complexes (RNP) to enable genome editing as previously reported (34) and subsequently performed colony-forming assays and serial cell counting to evaluate their progenitor activity. We achieved a complete loss of IRF2BP2 protein (Fig. 4A). Colony numbers on day 10 were not significantly different in human CD34⁺ BM cells with IRF2BP2 targeting or a control guide, whereas knockout of the essential gene RPA3 led to a significant reduction in colony number (Fig. 4B). Consistently, serial cell counting in human CD34⁺ BM cells showed a stable number of cells with IRF2BP2-targeting guides (Supplementary Fig. S4A).

To assess response to IRF2BP2 loss in primary patient models of AML, we selected six patient-derived xenograft (PDX) samples (Table 1; refs. 35, 36), confirmed their colony-forming capacity, and performed nucleofection as previously reported (34). We then assessed colony numbers 10 days after seeding nucleofected PDX cells. We found a significant reduction in colony-forming capacity in PDX cells that had received IRF2BP2-targeting guides (Fig. 4C). Interestingly, the effect was least pronounced in PDX cells with a CBFA2T3–GLIS2 fusion (PDX CPCT-0027; Fig. 4C). ssGSEA for monocytic lineage markers (37) in these PDX samples found significant enrichment for monocytic lineage markers in all PDX models but PDX CPCT-0027 (Supplementary Fig. S4B). In addition, to study the acute effects of IRF2BP2 loss in PDX cells, we generated a PDX model with degradable IRF2BP2 and validated IRF2BP2 degradation (Supplementary Fig. S4C), and impaired cell growth with IRF2BP2 degradation (Supplementary Fig. S4D).

To validate the relevance of the IRF2BP2 dependency in primary patient AML cells in vivo, we next assessed the effects of knockout of IRF2BP2 in PDX blasts on leukemia burden and survival. To first test our experimental system, we selected two monocytic PDX models that tolerated stable Cas9 expression and can be passaged for a limited time in vitro (PDX 16-01 with a CALM–AF10 fusion and PDX 17-14 with an MLL–AF10 fusion; see also Fig. 4D and extended information in Table 1). We infected these Cas9-expressing PDX cells with sgIRF2BP2_1 and sgIRF2BP2_3 cloned into a vector with mAmetrine as a selection marker. When we assessed the percentage of mAmetrine-positive cells over time, we found a rapid depletion of cells infected with guides targeting IRF2BP2, whereas the percentage of cells infected with nontargeting guides was stable for 2 weeks (Fig. 4E). We then infected PDX 16-01 and PDX 17-14 cells with a control guide or CRISPR guides against IRF2BP2 under a doxycycline-inducible promoter. Flow cytometry for GFP-positive cells expressing the CRISPR guides revealed a rapid decline in GFP-positive cells in PDX cells infected with sgIRF2BP2 over time. In contrast, the relative number of GFP-positive cells remained stable in cells infected with the nontargeting control (Supplementary Fig. S4E). PDX 16-01 and PDX 17-14 cells infected with IRF2BP2-directed CRISPR guides or a control guide were flow-sorted for GFP to increase purity and then injected by tail vein into irradiated NSG mice. Western blot analysis on day 4 after doxycycline induction showed decreased IRF2BP2 protein in these sorted cells in vitro (Supplementary Fig. S4F and S4G). Upon detection of at least 1% GFP-positive cells in the PB on day 9 (PDX 16-01) and day 11 (PDX 17-14) after transplantation, all mice received a doxycycline-containing diet to induce knockout of IRF2BP2. When we performed BM aspirations after 2 weeks on doxycycline, we detected a significant reduction of leukemia burden in mice transplanted with PDX cells with IRF2BP2-targeting guides compared with those with a control guide (Fig. 4F and G). When we then followed these mice for survival under a continuous doxycycline diet and knockout of IRF2BP2, we observed a significantly prolonged survival in mice transplanted with PDX cells with the targeting guides as compared with control guides: the median survival was 49 (PDX 16-01; sgIRF2BP2_1), 63.5 (PDX 17-14; sgIRF2BP2_1), 52 (PDX 16-01; sgIRF2BP2_2), and 60 (PDX 17-14; sgIRF2BP2_2) days, respectively, in mice transplanted with PDX cells with IRF2BP2 targeting guides as compared with 34 (PDX 16-01; sgNT) and 38 (PDX 17-14; sgNT) days in mice receiving PDX cells with a control guide (Fig. 4H and I). All mice showed a high leukemia burden with %hCD45⁺/live cells ≥95% at the time when they met endpoint criteria. When we assessed IRF2BP2 protein levels in BM cells of mice transplanted with PDX cells transduced with IRF2BP2-targeting guides, these cells had regained IRF2BP2 expression (Supplementary Fig. S4H).

Having validated IRF2BP2 as a strong dependency in AML, we next evaluated the underlying mechanisms explaining this dependency. We were particularly eager to understand how IRF2BP2 was implicated in inflammatory pathways. IRF2BP2 has been reported to be involved in transcriptional control (38), prompting us to study the localization of IRF2BP2 on chromatin. We performed chromatin immunoprecipitation (ChIP) for IRF2BP2 in MV4-11 cells using three different antibodies against IRF2BP2 followed by sequencing. We identified 6,269 binding sites that were identified in at least two out of three experiments, defining a set of IRF2BP2-binding targets (Fig. 5A; Supplementary Fig. S5A). Approximately 2,300 of these peaks were found in promoter regions and 2,900 marked enhancers defined by H3K27ac binding (Fig. 5A). H3K4me1 marks coincided with binding sites defined as gene enhancers, and the histone mark H3K4me3 was found in H3K27ac-marked promoter regions. Additionally, we identified H3K27ac-marked promoter regions to coincide with RNA polymerase II binding. We assigned genes to genome-wide IRF2BP2 chromatin binding peaks using the model-based analysis of the ChIP-sequencing (ChIP-seq; macs2) algorithm for the merged IRF2BP2 ChIP peaks. GSEA for all genes bound by IRF2BP2 identified immune response pathways to be the most significantly enriched (Fig. 5B); the top gene set was TNFα signaling via NFκB (Fig. 5B). To validate these findings, we performed ChIP for HA-tagged IRF2BP2 and H3K27ac in two PDX models and one additional AML cell line. We consistently found genome-wide IRF2BP2 binding in H3K27ac-marked promoter and enhancer regions (Fig. 5A, heat maps 6 and 7; Supplementary Fig. S5B). We identified core promoter and enhancer IRF2BP2-binding regions across all four models (Supplementary Fig. S5C), with 2,097 genes bound in all four experiments (Supplementary Fig. S5D). When we performed GSEA for all genes bound by IRF2BP2 in PDX 16-01, THP1, and PDX17-14, we identified TNFα signaling via NFκB as the top enriched gene set (Fig. 5B). TNFα signaling via NFκB is also a top enriched pathway in the IRF2BP2 “core” bound gene set (Fig. 5B; Supplementary Fig. S5D).

We next explored whether genes bound by IRF2BP2 were also differentially expressed following the degradation of IRF2BP2. We performed RNA sequencing (RNA-seq) in MV4-11 and PDX 16-01 cells with degradable IRF2BP2 after 6 hours, and in MV4-11 cells after 24 hours of dTAG-13 treatment. Significantly more genes were increased than decreased in expression at both time points (Fig. 5C; Supplementary Fig. S5E and S5F), in line with IRF2BP2's proposed function as a transcriptional repressor. Strikingly, IRF2BP2 bound most of the upregulated genes at 6 hours (Fig. 5C and D) and 24 hours (Supplementary Fig. S5F and S5G). Virtually none of the genes with decreased expression at 6 hours (Fig. 5C) and only a minor percentage at 24 hours (Supplementary Fig. S5F) were bound by IRF2BP2. In contrast, there was a significant overlap in genes upregulated and bound by IRF2BP2 at 6 and 24 hours (Supplementary Fig. S5H). In addition, differentially expressed genes following degradation of IRF2BP2 at 6 hours were enriched within bound and upregulated genes at 24 hours after degradation of IRF2BP2 and vice versa (Supplementary Fig. S5I, top). When we analyzed enriched gene sets within the bound and differentially expressed genes at 6 hours in MV4-11 and PDX 16-01 cells, we again found immune response signatures to be the top enriched gene sets at both time points (Fig. 5E), with genes regulated by NFκB in response to TNFα being most significantly enriched (Fig. 5E; Supplementary Fig. S5I, bottom). Of note, we also found a strong enrichment for the immune/inflammatory high AML cluster defined by the HALLMAK_INFLAMMATORY RESPONSE gene set within bound and upregulated genes following treatment with dTAG-13 for 6 hours (Fig. 5F). Moreover, when we performed ChIP-seq for H3K27ac after degradation of IRF2BP2 for 24 hours in MV4-11 cells, we observed a significant increase in genome-wide peaks assigned to genes that had increased expression upon IRF2BP2 degradation for 6 hours (Fig. 5G). These results again support a role for IRF2BP2 as a transcriptional repressor in AML cells. When we performed GSEA for the 41 genes with increased expression and increased H3K27ac binding after dTAG-13 treatment (Fig. 5F, right upper quadrant), we again identified TNFα signaling via NFκB as the top gene set (Supplementary Fig. S5J).

Results from the gene expression and ChIP-seq profiling led us to hypothesize that IRF2BP2 is a repressor of NFκB-mediated TNFα signaling that, when acutely perturbed, leads to leukemia cell death. To test this hypothesis, we first established an NFκB luciferase reporter assay in MV4-11 cells with degradable IRF2BP2. When we degraded IRF2BP2 with dTAG-13 treatment, we observed an acute upregulation of NFκB sig­naling at 6 and 24 hours followed by a decrease in signaling at 72 hours (Fig. 6A; Supplementary Fig. S6A). Cells treated with human TNFα as a positive control showed a similar pattern over time. Consistent with the reporter assay results, nuclear NFκB (RELA) protein levels increased at 24 hours with dTAG-13–induced IRF2BP2 degradation in MV4-11 cells (Fig. 6B) and PDX 16-01 cells (Supplementary Fig. S6B). Moreover, we observed an increased enrichment for NFκB (RELA) chromatin binding upon degradation of IRF2BP2 as compared with the DMSO control when performing ChIP PCR for the canonical RELA binding site at the NFKBA promoter (Fig. 6C). If cell death induced by perturbation of IRF2BP2 was mediated by derepressing NFκB signaling, then super-repression of NFκB should rescue this phenotype. To test this hypothesis, we overexpressed the wild-type (WT) form of the inhibitor of NFκB (IκBα), a mutant “super-repressor” allele of IκBα (IκBα MUT) harboring two amino acid substitutions (S32A/S36A) that render the mutated IκBα resistant to phosphorylation and degradation by IκB kinase (IKK; ref. 39), or a GFP control ORF, in AML cells (Supplementary Fig. S6C). When we treated these MV4-11 cells with TNFα to activate NFκB signaling, we detected an increase in IκB phosphorylation in cells expressing the GFP control or WT IκBα ORF, whereas there was no phosphorylation in cells expressing IκBα MUT (Supplementary Fig. S6D). We next used CRISPR/Cas9 to knock out IRF2BP2 in MV4-11 cells with the control plasmid, IκBα MUT, and IκBα WT (Fig. 6D). Mutated IκBα expression led to a complete rescue of the impaired cell growth with IRF2BP2 deletion in MV4-11 cells, and a greater tolerance of IRF2BP2 knockout was also evidenced by Western blot (Fig. 6D and E). The overexpression of IκBα WT led to the partial rescue of cell death (Fig. 6E), consistent with its repressive effect and high levels of expression. Consistently, doublings over time were significantly increased in cells expressing mutated IκBα compared with cells with a GFP control vector upon IRF2BP2 knockout, the latter of which underwent cell death (Supplementary Fig. S6E). Similarly, the mutant “super-repressor” allele of IκBα rescued the effects of IRF2BP2 knockout in PDX 17-14 cells (Supplementary Fig. S6F and S6G). Of note, when we treated MV4-11 cells overexpressing IκBα or a GFP control with JQ1 or THZ1, two compounds that have been shown to perturb oncogenic transcription, we did not observe any significant difference in the IC₅₀ for these drugs when comparing IκBα WT, mutant, or GFP-expressing cells (Supplementary Fig. S6H). All told, these data are consistent with the cell death associated with IRF2BP2 loss being mediated through activation of NFκB signaling.

To identify mediators of NFκB activation upon perturbation of IRF2BP2, we further explored genes bound by IRF2BP2 that were differentially expressed at 6 hours after degradation of IRF2BP2 in the RNA-seq data. Evaluation of the top differentially expressed IRF2BP2 targets revealed IL1β bound by IRF2BP2 (Fig. 6F and G; Supplementary Fig. S6I) and increased in expression with IRF2BP2 degradation at 6 hours, also scoring as one out of two genes in three different gene sets for immune response and TNFα signaling (Fig. 6F). These results were validated by quantitative PCR; IL1β was upregulated at 6 hours, returning to near baseline by 24 hours after IRF2BP2 degradation (Supplementary Fig. S6J).

If IL1β is a key positive regulator of NFκB signaling upon perturbation of IRF2BP2, then loss of IL1β function should attenuate NFκB signaling and the negative effects of IRF2BP2 degradation on cell growth. When we knocked down IL1β in MV4-11 cells (Supplementary Fig. S6K) expressing degradable IRF2BP2, we indeed observed attenuation of NFκB nuclear translocation (Supplementary Fig. S6L) and rescue of the viability reduction following degradation of IRF2BP2 (Fig. 6H). Consistently, we observed a diminished and delayed induction of cleaved caspase-3 (Supplementary Fig. S6M), supporting the hypothesis that IL1β enhances an inflammatory response upon perturbation of IRF2BP2 that results in the death of AML blasts. Moreover, to test whether constitutive abrogation of IL1β activity would influence cell growth in the context of IRF2BP2 degradation, we pretreated MV4-11 cells with degradable IRF2BP2 for 12 hours with a neutralizing anti–h-IL1β antibody followed by dTAG-13. Pretreatment with an anti–h-IL1β antibody abrogated the activation of NFκB signaling in our reporter cells (Supplementary Fig. S6N). When we assessed cell viability, anti–h-IL1β antibody–treated cells showed superior viability compared with cells pretreated with the antibody solvent water when degrading IRF2BP2, resulting in a rescue of the IRF2BP2 dTAG phenotype (Supplementary Fig. S6O).

We next sought to assess if we could rescue the enhancing effect of IL1β upon IRF2BP2 perturbation in PDX cells in vivo. We transduced Cas9-expressing PDX 16-01 cells with an inducible IRF2BP2-targeting CRISPR guide, and, in a second step, knocked out IL1β with two different sgRNAs in these cells (Supplementary Fig. S6P). We observed a partial rescue of the viability effect in vitro, whereas knockout of IL1β alone did not increase proliferation (Supplementary Fig. S6Q). We next transplanted sublethally irradiated NSG mice with the above engineered PDX 16-01 cells with IRF2BP2-targeting guides. Upon confirmation of engraftment, the mice received a doxycycline-containing diet to induce knockout of IRF2BP2. We performed BM aspirations on all mice after 2 weeks on doxycycline and observed significantly less reduction of leukemia burden in mice transplanted with PDX 16-01 cells with knockout of IL1β (Fig. 6I). This difference in disease burden translated into prolonged median survival in mice transplanted with IRF2BP2 knockout PDX 16-01 cells as compared with mice with concurrent IRF2BP2 and IL1β knockout (Fig. 6J). All told, our data support a role for IRF2BP2 in controlling IL1β-mediated NFκB signaling. Its derepression leads to acute inflammatory signaling, prompting AML cell death.

This study was designed to identify key players at the crossroads of immune and inflammatory signaling and selective dependencies in AML. To this end, we applied genome-wide CRISPR/Cas9-dropout screens for the unbiased identification of novel AML dependencies. By intersecting hits from three independent screens using complementary RNAi and CRISPR/Cas9 approaches, we increased the confidence in prioritized hits and reduced the likelihood of off-target effects confounding hit selection. We identified IRF2BP2 as a strong dependency in AML among a small group of well-established genes, including MYB and CBFB, providing additional confidence in this unexplored target. AML cell lines were more dependent on IRF2BP2 than any other cancer type, nominating IRF2BP2 as a selectively enriched dependency in AML. Pathway enrichment analysis identified a subset of AML with strong enrichment for immune/inflammatory pathways. We found the strongest enrichment in samples from patients with AML that expressed a monocytic signature. IRF2BP2 scored as the top dependency within monocytic AML cell lines when we assessed AML selective, core dependencies.

Our study explored the mechanistic underpinnings of IRF2BP2 in its dual role as a strong dependency and transcriptional regulator. Although most other studies focus on the role of transcriptional activators as cancer dependencies (e.g., BRD4, MYB, WT1, and MYC; refs. 40–43), our data provide evidence that IRF2BP2 primarily acts as a transcriptional repressor in AML. Genome-wide chromatin studies revealed that IRF2BP2 binds in both enhancer and promoter regions of AML cells. Assessing global gene-expression changes following degradation of IRF2BP2, however, the majority of differentially expressed genes upregulated at 6 and 24 hours were bound by IRF2BP2 at baseline. Moreover, we observed a significant gain in H3K27ac in upregulated IRF2BP2 targets following degradation of IRF2BP2, consistent with its role as a transcriptional repressor.

Our study focusing on the molecular function of IRF2BP2 suggests the repression of inflammation as a key strategy for AML cell survival. Although tumor evasion from immune surveillance is an established hallmark of cancer, to our knowledge, cell-intrinsic transcriptional repression of inflammatory signaling within cancer cells has not been extensively studied. Specifically, by acting as a transcriptional repressor, IRF2BP2 puts a brake on NFκB-mediated signaling. Previous studies suggest that cells require a precise set point of NFκB signaling (44), and our results indicate a similar requirement in AML cells. Our data suggest that derepression of NFκB signaling by perturbation of IRF2BP2 in leukemia cells induces an overwhelming inflammatory response. In this model, IRF2BP2 suppresses NFκB signaling as a cell-intrinsic innate immune response that would otherwise be detrimental to the leukemia cell. Thus, perturbation of IRF2BP2 serves to release the safety brake, resulting in an inflammatory burst and driving the AML cell toward cell death.

Overexpression of an IκBα super-repressor mutant (Supplementary Fig. S6C), which suppresses NFκB signaling, rescued AML cells with IRF2BP2 knockout from death. Importantly, overexpression of an IκBα super-repressor mutant in MV4-11 cells did not influence the IC₅₀ of transcription-perturbing compounds such as JQ1 and THZ1 (Supplementary Fig. S6H), suggesting that the hyperactivation of NFκB signaling is not a nonspecific mechanism leading to cell death with transcriptional perturbation.

TNFα acts as a pleiotropic cytokine inducing various cellular responses, ranging from inflammatory cytokine production to cell death (45). By activating the strongly proinflammatory transcription factor NFκB, TNFα plays an essential role in the acute-phase reaction, modulating fever and inflammation (46). Our understanding of its functions in leukemic blasts, however, remains limited. It is well established that in resting cells, NFκB resides in the cytoplasm, where it is controlled by IκB proteins. Upon stimulation (e.g., by IL1β), IκB is phosphorylated by the IKK complex, which leads to the rapid degradation of IκB and the release of NFκB. NFκB subsequently translocates into the nucleus, where it controls gene expression (47). Accordingly, with IRF2BP2 degradation, we observed an upregulation of NFκB signaling based on an NFκB reporter, global gene-expression changes, and an increase in nuclear NFκB protein. Although at first glance a paradoxical pairing, activation of TNFα-mediated NFκB signaling can also lead to apoptotic cell death. Specifically, it has been reported that TNFα-induced caspase-8 activation promotes apoptosis in cells treated with either cycloheximide or a SMAC mimetic (48). With IRF2BP2 repression in AML, we observed an increase in annexin V/PI–positive cells, induction of cleaved caspase-3 (Fig. 3E and F; Supplementary Fig. S3D), and an increase in cleaved caspase-8 (Fig. 3G), all consistent with apoptotic cell death.

In exploring key mediators of TNFα signaling, we found IL1β to be a molecular protagonist in the early phase following the degradation of IRF2BP2. When we reduced IL1β expression in cells and subsequently degraded IRF2BP2, we observed less profound and more delayed cell death (Fig. 6H; Supplementary Fig. S6M). Moreover, we observed a partial rescue of the effect on leukemia burden and survival in PDX cells when we compared knockout of IRF2BP2 with or without additional loss of IL1β in vivo (Fig. 6I and J).

Several studies have examined the effects of an inflamed microenvironment on hematopoietic stem cells (HSC) and leukemogenesis (49–52). For example, some groups have focused on transcription factors that activate inflammatory pathways; in this context, GATA2 has been reported to positively regulate both IL1β and CXCL2 expression in AML (53). Others have explored the concept of an inflamed milieu providing an advantage for selective clones. For example, a comprehensive study by Hormaechea-Agulla and colleagues showed the clonal advantage of DNMT3A loss-of-function clonal hematopoiesis under chronic mycobacterial infection and IFNγ treatment (54). Avagyan and colleagues showed in a zebrafish model that clonal fitness of a mutant CHIP clone can be driven by enhanced resistance to extrinsic inflammatory signals (55). Although our study focused on a role for IRF2BP2 as a repressor of cell-intrinsic inflammation, cell-extrinsic, environmental factors might also contribute to the inflammatory phenotype of leukemic cells and are the subject of ongoing investigations. One hypothesis in light of these recent findings is that in a steady-state situation, IRF2BP2 controls cell-intrinsic inflammatory signaling, allowing leukemic blasts to tolerate more environmentally mediated inflammatory signals, conferring a selective advantage.

The role of IL1β, as one major mediator of inflammatory signaling in hematopoietic cells, is complex. Collectively, published studies suggest a dose- and time-dependent (acute versus chronic exposure) response to IL1β. Importantly, some studies assess cell-intrinsic effects, whereas others are focused on cell-extrinsic mechanisms; some focus on impact on normal hematopoietic progenitors and others on leukemic cells. For example, Yang and colleagues observed lower IL1β expression in CD34⁺/CD38⁻ cells from patients with AML compared with CD34⁺ cells from healthy volunteers. Moreover, they were able to induce apoptosis by higher concentrations of IL1β treatment, and forced expression of IL1β, which also significantly diminished the self-renewal capacity of IL1β-transduced CD34⁺/CD38⁻ cells. Furthermore, when they transplanted IL1β overexpressing blasts into immunocompromised mice, the engraftment and AML reconstitution were impaired (56). Carey and colleagues (57), however, described an expansion of myeloid progenitors in longer-term assays in a subgroup of samples from patients with AML when culturing them with one tenth the IL1β concentration (10 ng/mL) used in the study led by Yang and colleagues (56). Work by Pietras and colleagues on the effects of IL1β in normal hCD34⁺ cells further supports a time-dependent function of IL1β; although IL1β supports inflammation as an emergency signal in an acute setting, chronic IL1β exposure restricts the lineage output of HSCs (58).

Although our data suggest that cell-intrinsic IL1β acts as a key enhancer of an inflammatory burst following the depletion of IRF2BP2, we do not think that IL1β is the sole mediator of the observed hyperinflammation. First, the knockdown of IL1β does not fully rescue the phenotype of IRF2BP2 perturbation. Second, our gene expression and chromatin studies suggest that IRF2BP2 is a direct transcriptional repressor of other members of TNFα-mediated NFκB signaling, such as TNFR2, suggesting NFκB signaling activation on multiple levels. Third, it is well established that canonical NFκB sig­naling can be activated by numerous and even multiple parallel stimuli, feeding into and maintaining a highly dynamic circuit, with NFκB at the crossroads as the master regulator of an innate immune response (59). Our mechanistic model suggests IL1β as a direct target of IRF2BP2, which is derepressed upon IRF2BP2 perturbation to activate NFκB signaling. IL1β is also an established target gene of NFκB, which will then limit its expression in a negative feedback loop, consistent with our observations of an early increase in IL1β expression that trends down by 24 hours.

TNF, first described by O'Malley and colleagues (60), was initially reported to induce programmed cell death or apoptosis. Following depletion of IRF2BP2, we consistently observed activation of caspase-8 and caspase-3, which were previously thought to trigger apoptosis exclusively. With an increasing understanding of distinct forms of regulated cell death (61), caspase-8 and caspase-3 were more recently found to mediate both apoptotic and pyroptotic cell death. Pyroptosis is a rapid, lytic form of cell death, induced by caspase-8 or caspase-3 cleavage of GSDMD leading to pore-forming fragments of GSDMD (62, 63), resulting in the release of inflammatory cytokines. Necroptosis, yet another form of regulated cell death, is associated with an increase in cytokine levels, promoting inflammation by the release of cytokines from dying cells, including IL1β, executed by the oligomerization of MLKL and a subsequent plasma membrane rupture (64–66). Within 24 hours following degradation of IRF2BP2, we did not observe induction of cleaved GSDMD nor phospho-MLKL as hallmarks of pyroptosis or necroptosis (Supplementary Fig. S3E and S3F). Whether pyroptotic or necroptotic cell death might be contributing to the phenotypic consequences of IRF2BP2 loss as secondary effects at later time points remains to be determined but will be the focus of future studies.

The kinetics of CRISPR- or RNAi-based approaches to target repression are slow, limiting the assessment of target perturbation on acute gene-expression changes. In contrast, our study leveraged a degradation technology that allowed us to study immediate, direct changes in global gene expression (33). Moreover, chemical degradation more closely approximates treatment with a small molecule and can better enable preclinical target validation in the absence of a small- molecule inhibitor.

Using PDX models, we demonstrated that depletion of IRF2BP2 is a viable and highly efficacious therapeutic strategy for AML. In an in vivo study approximating a clinical scenario where we induced knockout of IRF2BP2 in patient-derived leukemia cells after the mice had developed leukemia, we observed a significant reduction in leukemia burden and an increase in median survival in mice that had received AML cells with IRF2BP2-targeting guides. When mice transplanted with AML cells with IRF2BP2-targeting guides did eventually die, they died of leukemia with regained IRF2BP2 expression (Supplementary Fig. S4H). The therapeutic efficacy of single- target perturbation of IRF2BP2 positions this transcriptional repressor as a promising target for future clinical translation. Although there are no small-molecule inhibitors of IRF2BP2, one might envision a natural glue-like degrader or proteolysis targeting chimera strategy for the repression of IRF2BP2 therapeutically. Importantly, we confirmed in hCD34⁺ BM cells that targeting IRF2BP2 will not indiscriminately kill normal hematopoietic progenitor cells (Fig. 4B), suggesting potential for a high therapeutic window.

Myeloid blasts were described as abnormal white blood cells more than 200 years ago, inspiring Rudolf Virchow to name this condition of an excessive number of white blood cells as leukemia (“white blood”) in 1845 (67), yet there are many unexplored aspects of the blast intrinsic immune programs. Our study revealed a subgroup of AML strongly enriched for immune-inflammatory pathways. Within this “immune/inflammatory response high” subgroup, we observed more samples with a monocytic signature than in the control group. We exploited this observation to uncover immune dependencies that led to the identification of IRF2BP2. Intriguingly, recent studies highlight a monocytic AML state associated with resistance to venetoclax-based treatment (68)—not only supporting the need for alternative or combination treatment strategies for this AML subgroup but also suggesting the immune differentiation status of AML as a biomarker for clinical decision-making.

In summary, we identified IRF2BP2 as a dependency in AML and determined that IRF2BP2 acts as a transcriptional repressor, controlling NFκB-mediated TNFα signaling to keep AML cells alive. Perturbation of IRF2BP2 induces an acute inflammatory response, enhanced by IL1β, which promotes leukemic blast cell death. More broadly, the functional characterization of IRF2BP2 reveals transcriptional repression as a strategy for leukemia cell survival. This discovery provides a basis for leveraging acute inflammation as self-directed immunotherapy and a therapeutic approach to control leukemic blasts.

The AML cell lines MV4-11, U937, MOLM13, NB4, and THP1 were cultured in RPMI 1640 (Cellgro) supplemented with 10% fetal bovine serum (FBS; Sigma-Aldrich) and 1% penicillin–streptomycin (PS; Cellgro). HEK293T cells were cultured in DMEM with 10% FBS and 1% PS. All cell lines were short tandem repeat profiled. Primary patient samples were acquired following written informed consent in accordance with the Declaration of Helsinki, and PDX models were established under protocols approved by Dana-Farber Cancer Institute and Cincinnati Children's Hospital Medical Center institutional review boards. Characteristics of the PDX models are provided in Table 1. For short-term in vitro culture, PDX cells were maintained in Iscove's Modified Dulbecco's Medium containing 20% FBS and 1% PS, and supplemented with 10 ng/mL human SCF, thrombopoietin (TPO), FLT3L, IL3, and IL6 (PeproTech 300-07, 300-18, 300-19, 200-03 and 200-06).

GSEA v4.2.0 (details below), GraphPad PRISM 9, R 4.0.1, and Python 3.7.4 software packages were used to perform the statistical analyses. Statistical tests used are specified in the figure legends. Errors bars represent standard deviation unless otherwise stated. The threshold for statistical significance is P ≤ 0.05 unless otherwise specified.

The CRISPR screen was performed on 789 cancer cell lines, including 21 AML cell lines at the Broad Institute using the Avana sgRNA library, containing 73,372 guides with an average of 4 guides per gene (69, 70). The combined RNAi (Broad, Novartis, Marcotte) screen included 712 cancer cell lines with 23 AML cell lines as previously described (27, 28). The CRISPR Sanger screen was performed at the Sanger Institute (29, 71) on 317 cancer cell lines including 2 AML cell lines, using a 90,709 sgRNA library targeting 18,009 genes (∼5 sgRNAs/gene) as previously described (72).

To knock out/knock down IRF2BP2, the sgRNAs/shRNAs from the CRISPR (Avana)/Broad RNAi library were used. Inducible hairpins targeting IRF2BP2 were purchased from Dharmacon (TRIPZ lentiviral inducible shRNA collection). To rescue these hairpins, we inserted silent point mutations (A to G at bp position 1,056, A to G at 1,059 bp, and A to C at 1,062 bp) at the hairpin binding sites in the IRF2BP2 ORF (Precision LentiORF, Dharmacon; OHS5897-202624069). The pLX317-GFP ORF was obtained from the Broad Institute. MV4-11 cells were transduced with pMMP-luc-neo retrovirus (the pMMP-luc-neo vector was a kind gift from Andrew Kung) and selected with neomycin at 1 mg/mL. MV4-11-luc cells were then stably infected with either a nontargeting or an IRF2BP2-directed doxycycline-inducible hairpin (RHS4740-EG359938, Dharmacon).

The lentiviral Cas9 vectors coexpressing a blasticidin resistance gene or GFP (pXPR_BRD101 and pXPR_BRD104) and the lentiviral sgRNA expression vectors coexpressing a puromycin resistance gene, mCherry or mAmetrine (pXPR_BRD003, pXPR_BRD043 and pXPR_BRD052) were provided by the Genetic Perturbation Platform at the Broad Institute. The Cas9-T2A-mCherry and doxycycline-inducible sgRNA vectors were obtained from Addgene (70182 and 70183). sgRNAs were cloned into BsmBI-digested vectors.

IRF2BP2 was cloned into pLEX_305-N-dTAG using gateway recombination cloning strategies (Invitrogen) as previously described (33), and dTAG-13 was synthesized as previously described (73). To knock out endogenous IRF2BP2, sgIRF2BP2_1 (sgRNA IRF2BP2_1: CCTCGTAGTTGACGCAGCCG) and sgIRF2BP2_2 (sgRNA IRF2BP2_2: TCTCGATGACGAACTCGACG) from the Avana library were used. To rescue the knockout of endogenous IRF2BP2 by sgIRF2BP2_1 and 2, we introduced the following point mutation to the respective PAM sites: PAM_1: C to T at bp position 111 and PAM_2: C to T at bp position 141.

PGreenFire1-NFκB (EF1a-puro) Lentivector was purchased from SBI System Biosciences (TR012PA-P). Luminescence was detected by the Dual-Luciferase Reporter Assay System and Dual-Glo Luciferase Assay System (E1919 and E2920, Promega) following the manufacturer's protocol.

IκBα WT (pDONR223_NFKBIA_WT) was purchased from Addgene (81833) and cloned into pLEX_307 (Addgene, 41392) using gateway cloning following the manufacturer's protocol. To generate the super-repressor mutant (IκBα MUT), two amino acids were substituted (S32A/S36A).

Lentiviral human shRNA targeting IL1β (RHS4531-EG355) and a nontargeting control (empty vector, RHS4349) were purchased from Dharmacon. To knock out IL1β, the sgRNAs from the CRISPR (Avana) library were used.

An anti–hIL1β-IgG monoclonal antibody was purchased from Invivogen (mabg-hil1b-3).

For subcellular fractionations, a subcellular protein fractionation kit for cultured cells (Thermo Fisher Scientific, #PI78840) was used following the manufacturer's instructions.

Nucleofection was prepared using the P3 Primary Cell 4D-Nucleofector X Kit S (Lonza, #V4XP-3032) for CD34⁺ hematopoietic stem/progenitor cells (HSPC) and PDX cells. Per transfection, 6 μg purified Cas9 nuclease (Integrated DNA Technologies, #1081058) was mixed with 100 pmol/L of chemically modified synthetic sgRNA (Synthego) in P3 buffer and incubated at room temperature for 10 minutes to form RNP. Thawed CD34⁺ HSPCs were cultured in StemSpan SFEM II medium (STEMCELL Technologies, #09605) supplemented with 100 ng/mL human SCF, TPO, and FLT3L and 10 ng/mL IL3 and IL6, for two days before nucleofection. For each nucleofection, 2–3 × 10⁵ CD34⁺ cells or PDX cells were harvested, washed once with PBS, then mixed with the RNP and nucleofected using the Lonza 4D-nucleofector X Unit (program DZ-100). Two days after nucleofection, cells were seeded to assess colony-forming capacity.

Nucleofected cells were resuspended in MethoCult Express methylcellulose medium (STEMCELL Technologies, #04437) and plated at the indicated numbers; colonies were scored on day 10 after seeding.

Apoptosis was assessed with an APC Annexin V Apoptosis Detection Kit (BioLegend) following the manufacturer's protocol. Mouse samples from BM and PB were stained with human CD45 (Invitrogen, #MHCD4528) and murine CD45 (BioLegend, #103113) antibodies, unless the cells had been previously infected with a fluorescence marker (GFP, mCherry, or mAmetrine). Flow cytometry was performed on an LSRFortessa or FACSCelesta flow cytometer. Cells were sorted on a FACSAria II flow cytometer (BD Biosciences). The data were analyzed with FlowJo software (TreeStar).

Proteins were extracted using lysis buffer (Cell Signaling Technology) supplemented with complete, EDTA-free protease inhibitor cocktail (Roche Diagnostics) and phosphatase inhibitor cocktail (Roche Diagnostics). Protein samples were separated by SDS-PAGE and subsequently transferred to polyvinylidene difluoride (PVDF) membranes, which were blocked in 5% BSA and incubated with primary antibodies against IRF2BP2 (Proteintech, #18847-1-AP), vinculin (Cell Signaling Technology, #4650S), HA (Cell Sig­naling Technology, #3724S), caspase-3 (Cell Signaling Technology, #9662S), cleaved caspase-3 (Cell Signaling Technology, #9664S), caspase-8 (Cell Signaling Technology, #9746S), cleaved caspase-8 (Cell Signaling Technology, #9496S), actin (Cell Signaling Technology, #3700S), lamin B1 (Abcam, #ab16048), H3K4me3 (Cell Signaling Technology, #9733S), NFκB p65 (Cell Signaling Technology, #8242S), IL1β (Novus Biologicals, #NB600-633), phospho-IκBα (Cell Signaling Technology, #9246L), IκBα (Cell Signaling Technology, #9242S), GSDMD (Cell Sig­naling Technology, #96458), cleaved GSDMD (Cell Signaling Technology, #36425), MLKL (Abcam, #ab184718), phospho-MLKL (FabGennix, #PMLKL-140AP). Phospho-MLKL-positive control from FabGennix (#PC-MLKL) was used as a positive control. For immunoblots showing phosphorylation and total protein content, the same protein lysates were run on separate blots to present one experiment and were normalized to its loading control. Membranes were washed in TBS-T and incubated with the appropriate horseradish peroxidase–conjugated secondary antibodies. Signal was detected by enhanced chemiluminescence (Thermo Fisher Scientific).

Cells were suspended at a concentration of 15,000 cells/mL and seeded at 40 μL/well into 384-well plates. Cells were assessed for cell viability on day 0 and subsequent time points using the Cell-TiterGlo luminescent assay kit (Promega) according to the manufacturer's protocol. Luminescence was read on a Fluostar Omega Reader (BMG Labtech), and the viability was normalized to day 0.

Virus was produced using HEK293T cells transfected with lentiviral expression vectors, together with envelope VSVG and the gag-pol psPAX2 constructs. For transduction, AML cells were mixed with viral supernatant and 4 to 8 μg/mL polybrene. In some experiments, cells were centrifuged in viral supernatant at 1,000 × g for 1 hour at 33°C to enhance the transduction efficiency.

All in vivo studies were conducted under the auspices of protocols approved by the Dana-Farber Cancer Institute Animal Care and Use Committee. Six- to 8-week-old NSG mice (The Jackson Laboratory) were injected with 0.5 to 1 × 10⁶ cells. Prior to the transplantation of patient-derived cells, mice were conditioned with sublethal radiation. For bioluminescence imaging, mice were injected with 75 mg/kg intraperitoneal d-Luciferin (Promega), anesthetized with 2% to 3% isoflurane, and imaged on an IVIS Spectrum (Caliper Life Sciences). A standardized region of interest encompassing the entire mouse was used to determine total bioluminescence flux. To assess leukemia burden, BM was harvested from femur, tibia, and spine, and red blood cells were lysed (BD PharmLyse) prior to analysis by flow cytometry.

Twenty-five million (for histone marks) to 100 million (for IRF2BP2) AML cells per condition were cross-linked with 1% formaldehyde in cell culture media for 10 minutes and then quenched with 0.125 mmol/L glycine. Cells were lysed in Nuclei EZ Isolation Buffer (Sigma EZ Prep Nuclear Isolation Kit, #NUC-101) supplemented with protease inhibitor. Chromatin was sheared using a Covaris instrument. Subsequently, 0.5% of the lysate was removed as the input control. The remaining lysate was incubated with 30 μL (for histone marks) or 100 μL (for IRF2BP2) of protein G dynabeads (Thermo Fisher) and 5 μg of H3K27ac (ab4729 Abcam; H3K4me1: ab8895, H3K4me3: ab8580) or 10 μg of IRF2BP2 antibody (r1: abcam_ab114997, r2: Bethyl_A303-189A-T, r3: Bethyl_A303-190A), HA (Cell Signaling Technology_3724S), or NFκB (RELA) antibody (Sigma-Aldrich_17-10060) at 4°C overnight (antibodies and beads were incubated together for 30 minutes at 4°C prior to the addition of cell lysate). The precipitated lysate was then washed sequentially in two washes of each of the following buffers: ice-cold low salt buffer (20 mmol/L Tris-HCl pH 8.1, 150 mmol/L NaCl, 2 nmol/L EDTA, 1% Triton-X100, 0.1% SDS), high salt buffer (20 mmol/L Tris-HCl pH 8.1, 2 mmol/L EDTA, 500 mmol/L NaCl, 1% Triton-X100, 0.1% SDS), LiCl buffer (10 mmol/L Tris-HCl pH 8.1, 0.25 M LiCl, 1 mmol/L EDTA, 1% deoxycholic acid, 1% IGEPAL CA-630), and TE buffer (10 mmol/L Tris-HCl ph8.1, 1 mmol/L EDTA). Samples were eluted in elution buffer (1× TE pH 7.4, 1% SDS, 150 mmol/L NaCl, 5 mmol/L DTT). RNase A, proteinase K, and 0.2 M NaCl were added to the elution buffer and samples were incubated at 65°C for at least 4 hours to reverse cross-linking. DNA was purified using AMPure XP beads (Agencourt). ChIP-seq libraries were prepared using Swift S2 Acel reagents on a Beckman Coulter Biomek i7 liquid handling platform from approximately 1 ng of DNA according to the manufacturer's protocol and 14 cycles of PCR amplification. Finished sequencing libraries were quantified by Qubit fluorometer and Agilent TapeStation 2200. Library pooling and indexing were evaluated with shallow sequencing on an Illumina MiSeq. Subsequently, libraries were sequenced on a NovaSeq targeting 40 million 100 bp read pairs for H3K27ac, and 20 to 40 million 50 bp reads for H3K4me1, H3K4me3, and IRF2BP2 by the Molecular Biology Core facilities at Dana-Farber Cancer Institute.

All experiments were performed in technical replicates for each time point and condition. ERCC RNA spike-in mix was added to all samples following the manufacturer's instructions (ThermoFischer, 4456740). RNA was extracted from cells with the RNeasy Kit and on-column DNA digestion (Qiagen). For RNA-seq of MV4-11 and PDX 16-01 cells, polyA mRNA was isolated, and libraries were prepared using the TruSeq Stranded mRNA Kit (Illumina) according to the manufacturer's protocol. All samples were sequenced on an Illumina NovaSeq 6000 instrument with paired-end 150 bp reads to a depth of 40 to 60 million reads per sample.

RNA-seq for all parental PDX samples was performed in one replicate following the above-described protocol except for PDX 2017-63 and 2017-94. For these two models, RNA was extracted with the mirVana miRNA Isolation Kit (Thermo Fisher) following the manufacturer's protocol. The library for RNA-seq was prepared using NEBNext Ultra II Directional RNA Library Prep Kit (New England BioLabs). Sequencing was performed with single-end 51 bp reads at the University of Cincinnati.

The CRISPR Avana 20Q3 public data were downloaded from the depmap.org portal at https://depmap.org/portal/download/. The gene effect scores summarizing the guide depletion were determined based on the CERES algorithm (26). The combined RNAi 20Q3 public data were downloaded from the depmap.org portal at https://depmap.org/portal/download/. The effect of gene knockdown on cell line viability was inferred based on the DEMETER2 model (28). The Sanger CRISPR public data were downloaded from the Sanger Score Project site at https://score.depmap.sanger.ac.uk/. The genetic differential dependencies enriched in the AML cell lines were identified separately for each of the three screens based on the eBayes empirical Bayes linear model implemented in the limma v3.42.2 R package available in Bioconductor v3.9 (74) by performing a two-tailed t test for the difference in the distribution of gene dependency scores in AML compared with all other screened cell lines. Statistical significance for differential gene dependency was calculated as a q-value derived from the P value corrected for multiple hypothesis testing using the Benjamini and Hochberg method (ref. 75; limma eBayes |effect size| ≥ 0.3, q-value ≤ 0.1). Genes that scored as AML-enriched dependencies in all three screens were defined as core AML dependencies. Statistical significance of overlapping AML-enriched dependency hits identified in pairs of screens was calculated by a two-tailed Fisher exact test (P ≤ 0.05).

GSEA v4.2.0 software was used to identify functional associations of genome-wide molecular profiles (76, 77) within the MSigDB v7.1 database collection of gene sets (http://www.gsea-msigdb.org/gsea/downloads.jsp) and http://www.bu.edu/nf-kb/gene-resources/target-genes/. Significance cutoffs were assessed based on the GSEA standard recommendations: absolute normalized enrichment score (NES) ≥1.3, P ≤ 0.05, Benjamini–Hochberg false discovery rate ≤0.25.

ssGSEA is an extension of GSEA that calculates separate enrichment scores for each pairing of a sample and gene set. Each ssGSEA enrichment score represents the degree to which the genes in a particular gene set are coordinately upregulated or downregulated within a sample.

Quality control tests for unmapped sequences were performed based on the FastQC v.0.11.5 software (http://www.bioinformatics.babraham.ac.uk/projects/fastqc/). The ChIP sequences were aligned to the GRCh37/hg19 genome using bowtie2-2.3.5 (78). PCR duplicates were removed with the Picard v2.18.2 Mark Duplicates tool (https://broadinstitute.github.io/picard/). The mapped reads were normalized in units of reads per kilobase per million (RPKM or rpm/bp) and coverage tracks for the RPKM signal were created as bigwig files for bins of size 20 base pairs by using the bam coverage tool available in deepTools v2.5.3 (79). Peak calling was performed against the input control using the model-based MACS2 v2.1.1.20160309 software (80), FDR ≤ 0.01. The area under the curve (AUC) RPKM-normalized signal across genomic regions was computed with the bwtool software (81). Peaks with a low AUC coverage (<300 RPKM) were disregarded, and the ENCODE blacklisted regions for hg19 (available at https://www.encodeproject.org/annotations/ENCSR636HFF/) were removed from each set of peak regions. Quality control tests for the mapped reads were performed using the ChIPQC library available from Bioconductor v3.9 (82). The pairwise correlations between IRF2BP2 replicates were estimated based on the multiBamSummary function available from the deepTools v2.5.3 (79) and visualized on correlation heat maps and PCA plots. The IRF2BP2 binding sites were identified for the merged data (macs2, FDR ≤ 0.01). The peaks were annotated with the closest hg19 genes by using the annotatePeaks function implemented in the Homer v4.11 platform (83) and the GREAT annotation platform (84). Binding peaks and normalized binding signals were visualized on the Integrative Genomics Viewer (IGV) v2.4.0 platform (85). Gene promoter regions were defined as the ± 2.5 kb intervals around the hg19 gene transcription start site (TSS). Enhancer regions were defined as the H3K27ac binding regions outside TSS ± 2.5 kb gene promoter regions. The BEDTools v2.27 suite (86) was used to perform various genomic region analyses (sorting, intersection, and merging). Heat maps of AUC ChIP-seq–normalized signal occupancy on genomic regions were created using the computeMatrix and the plotHeatmap tools available in deepTools v2.5.3. The plotProfile tool from deepTools v2.5.3 was used to create metaplots based on the average normalized scores across genomic regions. Differential genome-wide mark binding in between treatment conditions was quantified based on the unpaired t test with Welch correction (cutoff P ≤ 0.05) for the AUC RPKM-normalized signal across the genome-wide regions in the two conditions. For a specific genomic region, the changes in signal were classified as increase and decrease (or unchanged) based on the absolute cutoff 1.5 for the delta AUC scores.

The spiked-in human reads were mapped to the GRCh37/hg19 human genome using STAR v2.7.3 (87). The ERCC92 mix1 spike-in sequences were mapped using STAR v2.7.3 according to the manufacturer's protocol. Gene-level reads were summarized by using the featureCounts v1.6.3 method implemented in the Subread v2.0.0 package (http://subread.sourceforge.net; ref. 88). The technical quality metrics of the spiked-in mapped reads were assessed using the standard methods implemented in the erccdashboard package (89). Additional quality control tests for the mapped reads and for replicate reproducibility were performed using SARTools v1.7.3 (90). Gene counts were normalized and used to quantify differentially expressed genes between the experimental and control conditions using the DESeq2 v1.24.0 method implemented in Bioconductor v3.9 (91). Genes with ≥10 reads across at least three samples were annotated as expressed. Differentiability for expressed genes was assessed with DESeq2 based on the robust shrunken log₂ fold change scores and the approximate posterior estimation for GLM coefficients (apeglm v1.6; ref. 92; method for effect size). The cutoffs for differentially expressed genes were |fold change expression|≥1.5 and adjusted P ≤ 0.10. Heat maps for transcriptional data visualization were created by using the Morpheus software platform (https://software.broadinstitute.org/morpheus/) based on the log₂(TPM+1) DESeq2-normalized count data.

Plasmids generated in this study are being deposited to Addgene.

ChIP-seq and RNA-seq data generated during this study are available at the Gene Expression Omnibus (GEO; GSE168649). The ChIP-seq Pol2 data were downloaded from the GEO repository (GSE80779). We analyzed the following publicly available AML expression data sets: TCGA LAML (179 AML samples; ref. 93), GSE14468 (526 AML samples; ref. 94), Beat AML (451 samples; ref. 95), Cancer Cell Line Encyclopedia v20Q3 (96), and TARGET AML (232 samples; ref. 97).

S. Lin reports grants from the Leukemia & Lymphoma Society and the NCI during the conduct of the study. N.V. Dharia reports grants from Julia's Legacy of Hope St. Baldrick's Foundation Fellowship during the conduct of the study, as well as other support from Genentech, Inc., a member of the Roche Group, outside the submitted work. M. Wunderlich reports grants from the NIH/NCI outside the submitted work. L. Benajiba reports grants from Gilead and Pfizer outside the submitted work. B. Nabet reports a patent for WO/2017/024318 pending, a patent for WO/2017/024319 pending, a patent for WO/2018/148440 pending, a patent for WO/2018/148443 pending, and a patent for WO/2020/146250 pending. N.S. Gray reports grants from Deerfield Ventures, Arbella Therapeutics, Springworks, and Interline Therapeutics during the conduct of the study, as well as personal fees from Syros, C4 Therapeutics, Allorion, Lighthorse, Matchpoint, Inception, Larkspur, Soltego, GSK, Cobroventures, and B2S outside the submitted work. K. Stegmaier reports grants from the NCI and St. Baldrick's Foundation during the conduct of the study, as well as personal fees from AstraZeneca and Bristol Meyers Squibb, grants and personal fees from KronosBio, grants from Novartis, and personal fees and other support from Auron Therapeutics outside the submitted work. No disclosures were reported by the other authors.

J.M. Ellegast: Conceptualization, data curation, formal analysis, funding acquisition, validation, methodology, writing–original draft, writing–review and editing. G. Alexe: Data curation, formal analysis, writing–review and editing. A. Hamze: Validation. S. Lin: Conceptualization, validation, methodology. H.J. Uckelmann: Methodology. P.J. Rauch: Conceptualization, validation, methodology. M. Pimkin: Data curation, methodology. L.S. Ross: Validation, methodology. N.V. Dharia: Conceptualization, data curation. A.L. Robichaud: Methodology. A. Saur Conway: Conceptualization, validation. D. Khalid: Validation. J.A. Perry: Resources, data curation. M. Wunderlich: Resources, data curation. L. Benajiba: Conceptualization, data curation, validation. Y. Pikman: Resources, data curation. B. Nabet: Conceptualization, validation, methodology. N.S. Gray: Conceptualization. S.H. Orkin: Conceptualization, resources, methodology. K. Stegmaier: Conceptualization, resources, supervision, methodology, writing–review and editing.

This work was supported by the Swiss National Science Foundation (J.M. Ellegast), the Lady Tata Memorial Trust (J.M. Ellegast), the Pediatric Cancer Research Foundation (J.M. Ellegast), The Helen Gurley Brown Presidential Initiative (The Pussycat Foundation; J.M. Ellegast), the NCI K99 CA263161 (S. Lin), the NCI R50 CA211404 (M. Wunderlich), the NCI K08 CA222684 (Y. Pikman), the NCI K22 CA258805 (B. Nabet), the NCI R35 CA210030 (K. Stegmaier) and P50 CA206963 (K. Stegmaier), a St. Baldrick's Foundation Robert J Arceci Innovation Award (K. Stegmaier), St. Baldrick's Foundation Consortium Grant (K. Stegmaier), the Four C's Fund (K. Stegmaier), and Team Crank (K. Stegmaier). S. Lin is a Fellow of the Leukemia and Lymphoma Society. S.H. Orkin is an investigator of the Howard Hughes Medical Institute. We thank Zach Herbert for contributing to the methods of this article. We thank Peter Libby and Amélie Vromman for sharing their expertise on IL1β targeting therapies. We thank all members of the Stegmaier Lab, Charles Hatton, and Haihua Chu for important discussions. We express our deep gratitude to all patients and families who contributed to the research in this study.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.