Drug addiction unveils a repressive methylation ceiling in EZH2-mutant lymphoma

Drug addiction, a phenomenon where cancer cells paradoxically depend on continuous drug treatment for survival, has uncovered cell signaling mechanisms and cancer codependencies. Here we discover mutations that confer drug addiction to inhibitors of the transcriptional repressor polycomb repressive complex 2 (PRC2) in diffuse large B-cell lymphoma. Drug addiction is mediated by hypermorphic mutations in the CXC domain of the catalytic subunit EZH2, which maintain H3K27me3 levels even in the presence of PRC2 inhibitors. Discontinuation of inhibitor treatment leads to overspreading of H3K27me3, surpassing a repressive methylation ceiling compatible with lymphoma cell survival. Exploiting this vulnerability, we show that inhibition of SETD2 similarly induces the spread of H3K27me3 and blocks lymphoma growth. Collectively, our findings demonstrate that constraints on chromatin landscapes can yield biphasic dependencies in epigenetic signaling in cancer cells. More broadly, we highlight how approaches to identify drug addiction mutations can be leveraged to discover cancer vulnerabilities. Profiling the resistance landscape to PRC2 inhibitors in EZH2-mutant lymphoma with CRISPR-suppressor scanning reveals drug addiction mutations and a repressive methylation ceiling. Surpassing the ceiling with SETD2 inhibition halts lymphoma growth.

Drug addiction, a phenomenon where cancer cells paradoxically depend on continuous drug treatment for survival, has uncovered cell signaling mechanisms and cancer codependencies. Here we discover mutations that confer drug addiction to inhibitors of the transcriptional repressor polycomb repressive complex 2 (PRC2) in diffuse large B-cell lymphoma. Drug addiction is mediated by hypermorphic mutations in the CXC domain of the catalytic subunit EZH2, which maintain H3K27me3 levels even in the presence of PRC2 inhibitors. Discontinuation of inhibitor treatment leads to overspreading of H3K27me3, surpassing a repressive methylation ceiling compatible with lymphoma cell survival. Exploiting this vulnerability, we show that inhibition of SETD2 similarly induces the spread of H3K27me3 and blocks lymphoma growth. Collectively, our findings demonstrate that constraints on chromatin landscapes can yield biphasic dependencies in epigenetic signaling in cancer cells. More broadly, we highlight how approaches to identify drug addiction mutations can be leveraged to discover cancer vulnerabilities. Drug resistance represents a major challenge for anticancer therapeutics. In some cases, drug resistance leads to drug addiction, where cancer cells depend on sustained drug treatment for survival. The mechanisms of drug addiction have been most widely reported for kinase inhibitors and typically occur through activation of compensatory signaling pathways [1][2][3] . Due to the cumulative effects of bypass signaling and restored kinase signaling, withdrawal of the kinase inhibitor leads to overactivation of signaling and subsequent arrest of cell growth, revealing an upper limit of mitogenic signaling in cancer cells 3,4 . This mitogenic signaling ceiling likely reflects checkpoint mechanisms such as oncogene-induced senescence, which can be triggered by overactivation of Ras, BRAF and Myc signaling [5][6][7] . Drug addiction remains largely unexplored for other classes of drug targets.
Drug addiction may result from fine-tuned cell signaling thresholds inherent to the tissue of origin. As a key example, B cells are selected for an optimized level of signaling through the B-cell receptor (BCR), which promotes survival and positive selection during B-cell development. However, hyperactivation of BCR downstream kinases, for example in autoreactive B cells, results in excessive signaling and triggers negative selection 8 . Properly tuned signaling levels therefore equip B cells with an appropriately functional receptor that recognizes foreign antigens, but not self-antigens. This requirement for precise levels of BCR signaling can be preserved in malignant B cells, with both the suppression and activation of B-cell kinase signaling leading to lymphoma cell death 2,9 . Whether similar thresholds exist in other pathways beyond kinase signaling in malignant B cells remains to be explored.
We next sought to generate clonal Karpas-422 CXCdel cells for further study. However, despite the strong enrichment of EZH2 CXCdel in the presence of PRC2 inhibitors, clonal cell lines containing CXCdel could not be expanded from single cells when sorted into fresh media lacking inhibitor. We therefore considered whether the CXCdel-mutant cells have impaired growth in the absence of inhibitors. To test this possibility, the PRC2 CRISPR-suppressor scan was repeated and further extended by 3 weeks either in the continued presence of PRC2 inhibitors or after their withdrawal (Fig. 2a). Through this adapted method, which we term 'CRISPR-addiction scanning,' we observed that the previously enriched CXC-targeting sgRNAs were highly depleted upon inhibitor removal, while the sgRNAs targeting the respective inhibitor-binding sites remained enriched ( Fig. 2b-d). Accordingly, genotyping of bulk cells revealed that EZH2 CXCdel was highly depleted upon inhibitor withdrawal ( Fig. 2e and Extended Data Fig. 2a), supporting the notion that cells containing this mutation are drug addicted (that is, have reduced intrinsic fitness upon inhibitor removal) 1,3 .
An alternative hypothesis for the depletion of CXC-targeting sgR-NAs is that cells containing these sgRNAs were out-competed upon inhibitor removal by other cells containing previously suppressed sgRNAs. To test this, we fit a system of ordinary differential equations describing competitive logistic growth (the competitive Lotka-Volterra equations; see Supplementary Note for detailed description) to the sgRNA prevalence data at weeks 0, 5 and 8 (ref. 25). By comparing the estimated intrinsic growth rate r for each sgRNA-containing population on-drug (r on ) versus off-drug (r off ) relative to a drug-resistant mutant that disrupts inhibitor binding, with intrinsic fitness that is assumed to be unaffected by drug (that is, EZH2 sgI109 for GSK343 and EED sgY365 for EED226), we derived an addiction score, A = r on -r off . This score is positive when the sgRNA's effect on intrinsic fitness is greater on-drug than off-drug (addiction), negative when the effect is greater off-drug than on-drug (expected wild-type behavior), and zero when the effect is equal on-drug and off-drug (resistance without addiction). We further restricted the inferred set of addicted sgRNAs (A > 0) to those that comprised at least 5% of the pool by week 5, indicating robust growth while on-drug ( Fig. 2f-g and Extended Data Fig. 2b-h). By this metric, the fitness of the CXC-targeting sgRNAs as well as others declined upon inhibitor removal and thus these sgRNAs were identified as addicted sgRNAs (Supplementary Dataset 3). Altogether, our CRISPR-addiction scan suggests that multiple resistance mutations enriched in the presence of PRC2 inhibitors may act as drug addiction mutations.
In particular, chromatin pathways are critical for B-cell maturation and are frequently dysregulated in lymphoma. Recurrent activating mutations in EZH2, the catalytic subunit of the transcriptional repressor polycomb repressive complex 2 (PRC2), promote neoplastic growth in a variety of lymphomas, including 20% of follicular lymphomas (FL) and 10% of diffuse large B-cell lymphomas (DLBCL) 10,11 . Within the germinal center, PRC2 represses cell cycle checkpoint genes, allowing for somatic hypermutation and B-cell expansion upon presentation of a foreign antigen 12 . In lymphoma, activating mutations in EZH2 repress these same cell cycle checkpoint genes to induce unrestrained, autonomous proliferation of lymphoma cells and expansion of germinal centers, predisposing cells to transformation 13 . The discovery of these activating mutations prompted the development of inhibitors targeting PRC2, leading to the approval of the EZH2 active site inhibitor tazemetostat for FL and the clinical development of allosteric inhibitors targeting the PRC2 subunit, EED. Both active and allosteric site PRC2 inhibitors reduce H3K27me3 levels, derepress tumor suppressor genes and promote lymphoma differentiation and ultimately cell death 14,15 .
Through the discovery of drug addiction mutations that hyperactivate PRC2, here we show that not only lower but also upper thresholds exist for the repressive PRC2 mark, H3K27me3, in EZH2-mutant DLBCL cells. We find that lymphoma cells survive in a 'Goldilocks' epigenetic state, which is bounded by a previously unknown repressive methylation ceiling. Pharmacological inhibition of SETD2, a PRC2 antagonist, can cause cells to surpass this upper methylation limit and consequently block EZH2-mutant DLBCL growth, suggesting that PRC2 activation may represent a new therapeutic vulnerability in this context. Overall, our study demonstrates that there are upper limits to epigenetic modifications in cells, highlighting the importance of drug addiction mutation identification methods for the discovery of new signaling thresholds.

EZH2 CXC domain is a drug resistance hotspot
We surveyed the drug resistance landscape to PRC2 inhibitors in EZH2-mutant lymphoma using CRISPR-suppressor scanning, a method to identify drug resistance hotspots within a protein target 16 . Specifically, we performed CRISPR-suppressor scanning in the DLBCL cell line Karpas-422, which is heterozygous for the most common activating EZH2 mutation, EZH2 Y646N 17 . We used a pool of 650 single-guide RNAs (sgRNAs) spanning the coding sequence of the core members of PRC2 (EZH2, EED and SUZ12) and selected for resistance to two PRC2 inhibitors: GSK343, an EZH2 active site inhibitor 18 ; and EED226, an EED allosteric inhibitor 14 that blocks PRC2 allosteric activation by H3K27me3 ( Fig. 1a and Supplementary Dataset 1). Following 5 weeks of inhibitor treatment, we isolated genomic DNA from surviving cells and determined the enrichment of sgRNAs under each treatment condition. The top-enriched sgRNAs unique to each inhibitor treatment (EZH2 sgI109, EZH2 sgN699 and EED sgY365) targeted positions proximal to the inhibitor-binding sites of the respective inhibitors ( Fig. 1b- [19][20][21] . The identification of mutants at the inhibitor-binding site confirms that CRISPR-suppressor scanning of PRC2 robustly identified sgRNAs leading to drug resistance. Unexpectedly, several sgRNAs targeting positions within the CXC domain of EZH2 were also highly enriched in both inhibitor treatment conditions (sgA596/D597, sgD597 and sgD597/H598; Fig. 1b,c,f, Extended Data Fig. 1a and Supplementary Dataset 2). These sgRNAs target a loop in the CXC domain that lies 17 Å from the substrate nucleosome and over 40 Å from either inhibitor-binding site (Fig. 1c)   Addiction score  Fig. 3e), suggesting that resistance is mediated by maintaining levels of H3K27me3 despite the presence of PRC2 inhibitors. At higher concentrations of PRC2 inhibitors, H3K27me3 was depleted in both wild-type and CXCdel cells, consistent with decreased cell viability observed in the growth assays ( Fig. 3b and Extended Data Fig. 3f).
We considered if the CXCdel mutation requires or cooperates with the activating EZH2 SET domain mutations to maintain H3K27me3 levels and confer resistance. To assess EZH2 genotypes in a nonclonal setting, we performed targeted amplicon sequencing of   Pfei er sgD597 GSK343

Deletions Insertions
Number of amino acids changed:   Fig. 4a-i). CXCdel-mutation frequencies were higher in the wild-type EZH2 allele in Karpas-422 ( Fig. 3g and Extended Data Fig. 4j-k). Notably, the most abundant deletion in Karpas-422 transduced with EZH2 sgD597 under GSK343 treatment was the 10-amino acid CXCdel mutation in the wild-type EZH2 allele, whereas a smaller 3-amino acid deletion was more frequently observed in the EZH2 Y646N allele, suggesting a bias against larger deletions in the Y646N allele ( Fig. 3g-h and Extended Data Fig. 4b-d,f-h,j,k). However, little allele preference for CXCdel was observed with respect to A682G mutation in Pfeiffer, with a significant fraction of alleles containing CXCdel and A682G in cis ( Fig. 3i and Extended Data Fig. 4e,i,l). Besides the CXCdel mutation, other large in-frame deletions were enriched in Pfeiffer cells treated with EED226, with little allele preference (Extended Data Fig. 4l). Smaller in-frame deletions were also enriched in Pfeiffer under vehicle treatment ( Fig. 3f and Extended Data Fig. 4a,e,l). The contrasting genotypes between Karpas-422 and Pfeiffer suggest that mechanistic differences between their distinct SET domain activating mutations may influence allelic preference. Altogether, these findings support the notion that the EZH2 CXCdel mutation is predominantly heterozygous in DLBCL cell lines and acts synergistically with the SET domain activating mutations. Given that CXCdel-mutant DLBCL cells can harbor both CXCdel mutation and activating SET domain mutation either in trans or in cis, we sought to characterize the impact of CXCdel on PRC2 activity in an otherwise wild-type EZH2 setting. We knocked in CXCdel into the endogenous wild-type EZH2 locus in K562 cells (Extended Data Fig. 5a,d). In contrast to Karpas-422 CXCdel cells, no significant changes in the fitness of K562 CXCdel cells were observed in the absence of PRC2 inhibitors (Extended Data Fig. 5b). H3K27me3 levels were increased in homozygous K562 CXCdel +/+/+ compared to K562 wild-type cells ( Fig. 4a and Extended Data Fig. 5c), indicating that CXCdel alone is sufficient to increase H3K27me3. We did not detect global changes in H3K36me3 (ref. 27). Next, we biochemically assessed EZH2 CXCdel activity by purifying wild-type and CXCdel recombinant human PRC2 (EZH2 wild-type or CXCdel, EED, SUZ12, RBAP48 and AEBP2; Extended Data Fig. 5g) 28 . PRC2 CXCdel exhibited a tenfold increase in methyltransferase activity on unmodified mononucleosomes, demonstrating that the hypermorphic effect of the mutation is intrinsic to the PRC2 complex (Fig. 4b).
The EZH2 CXC domain mediates interactions between PRC2 and the substrate nucleosome 22,23 , leading us to consider if enhanced PRC2 activity results from increased substrate affinity. In agreement, electrophoretic mobility shift assays (EMSA) demonstrated that PRC2 CXCdel has increased affinity toward mononucleosomes ( Fig. 4c and Extended Data Fig. 5h). Taken together, these results support the notion that the CXCdel mutation activates PRC2 activity by increasing the affinity of PRC2 nucleosome. While most cancer mutations in the EZH2 CXC domain abrogate PRC2 activity 29 , the hyperactivity of CXCdel suggests that some CXC domain mutations might also enhance PRC2 activity by increasing nucleosome binding. In support, EZH1 Q571R, a mutation present in 27% of thyroid adenomas, has been shown to activate PRC2 activity 30 . EZH1 is a homolog of EZH2 that is predominantly expressed in nondividing cells 31 . The homologous residue in EZH2, Q575, inserts into the major groove of DNA wrapping the substrate nucleosome, suggesting that its mutation to a positively charged residue might increase nucleosome binding (Fig. 4d) 22,23 . Consistent with this notion, purified PRC2 containing EZH2 Q575R exhibited elevated methyltransferase activity and mononucleosome binding relative to wild-type PRC2 (Fig. 4b,c and Extended Data Fig. 5h). Notably, PRC2 CXCdel and Q575R exhibited increased activity across all three steps of methylation, from me0 to me3 (Fig. 4e). Moreover, K562 cells containing knock-in of EZH2 Q575R also exhibited elevated H3K27 methylation (Fig. 4a and Extended Data Fig. 5a,e). As such, we next considered if EZH2 Q575R confers drug addiction. We knocked in the EZH2 Q575R mutation into Karpas-422 cells in the presence of 1 μM EED226 (Extended Data Figs. 3c and 5f). The EZH2 Q575R mutant cells displayed biphasic growth responses to increasing concentrations of GSK343 and EED226, demonstrating drug addiction (Fig. 4f). Thus, both deletions and point mutations in the EZH2 CXC domain can augment PRC2 nucleosome binding and methyltransferase activity to confer drug addiction in EZH2-mutant lymphoma cells.

Pervasive H3K27me3 spreading is a vulnerability of lymphoma
To investigate how CXCdel hyperactivity impacts the chromatin landscape, we profiled H3K27me3 using quantitative spike-in ChIPseq (ChIP-Rx) in wild-type and CXCdel +/− Karpas-422 cells continuously grown in GSK343 and after 72 h of inhibitor withdrawal. Upon inhibitor withdrawal, H3K27me3 substantially increased and broadly spread across the genome in CXCdel +/− cells (Fig. 5a-d), suggesting that H3K27me3 might encroach into neighboring regions typically blocked by active histone modifications 27,32,33 . To determine which regions are most affected by H3K27me3 spreading upon inhibitor removal, we assessed the correlation between the levels of active histone modifications and the change in H3K27me3 signal at the borders of H3K27me3 domains upon inhibitor withdrawal. H3K36me3 was most associated with increases in H3K27me3 (Fig. 5e). In agreement, we observed a greater increase in H3K27me3 within gene bodies compared to promoters of expressed genes for CXCdel +/− cells, consistent with the enrichment of H3K36me3 at gene bodies ( Fig. 5f and Extended Data Fig. 6a,b) 34 . H3K27me3 ChIP-Rx also revealed the spreading of H3K27me3 genome wide and within gene bodies in CXCdel +/+/+ versus K562 wild-type cells (Fig. 5g-h and Extended Data Fig. 6c-e), showing that PRC2 CXCdel is sufficient for H3K27me3 domain spreading. Despite these global changes in H3K27me3 induced by CXCdel, genome-wide binding of EZH2 was largely unchanged between CXCdel and wild-type cells for both Karpas-422 and K562 (Extended Data Fig. 6f-g). Together, our data show that inhibitor withdrawal in CXCdel-mutant cells induced the pervasive spreading of H3K27me3, which is pronounced at gene bodies marked with H3K36me3.
RNA-seq before and after inhibitor withdrawal in Karpas-422 CXCdel +/− cells revealed 338 differentially expressed genes (|log 2 (fold-change)| > ± 0.5, p adj < 0.05), of which 218 genes (64.5%) were downregulated after 72 h of inhibitor removal (Fig. 5i and Supplementary Dataset 4). For the downregulated genes, an even greater increase in H3K27me3 within gene bodies was observed upon inhibitor withdrawal-with genes gaining the most H3K27me3 decreasing in expression to the greatest extent ( Fig. 5j and Extended Data Fig. 7a-b). Of note, several key genes related to B-cell proliferation and survival, including MYB, BCL9 and CDK6, were downregulated and gained H3K27me3 within their gene bodies (Fig. 5a and Extended Data Fig. 7c) 35,36 . In addition, we performed RNA-seq for both wild-type and CXCdel +/+/+ K562 cells and similarly observed the largest gains in H3K27me3 for downregulated genes, along with a correlation between gene repression and increases in H3K27me3 signal (Extended Data Fig. 7d-f and Supplementary Dataset 5). Collectively, our observations support the notion that H3K27me3 overspreading upon inhibitor withdrawal contributes to the silencing of genes critical for lymphoma growth.
Our findings demonstrate a biphasic dependency on H3K27me3 levels to maintain EZH2-mutant lymphoma proliferation, revealing that activating PRC2 is a potential avenue to block lymphoma growth. Active histone modifications, such as H3K36me3, can antagonize PRC2 activity 32 . Given the observation that H3K27me3 aberrantly spreads into H3K36me3-marked regions, we tested whether inhibition of SETD2, the only methyltransferase known to deposit H3K36me3, could facilitate H3K27me3 spreading and halt DLBCL proliferation. We treated a panel of EZH2-wild-type and -mutant lymphoma cell lines with SETD2-IN-1, a recently disclosed, selective SETD2 inhibitor 37,38 . Of the seven EZH2-mutant cell lines tested, six were sensitive to SETD2-IN-1 in comparison to only one of four wild-type EZH2 cell lines (Fig. 6a). SETD2-IN-1 treatment substantially reduced H3K36me3 levels across all cell lines, suggesting that the EZH2-mutant sensitivity for the inhibitor is not correlated with a decrease in global H3K36me3 levels (Extended Data Fig. 8a). To independently validate the effects of SETD2 inhibition, we depleted SETD2 in EZH2-wild-type and -mutant cell lines using short-hairpin RNAs (shRNAs) (Extended Data Fig. 8b,c). Pfeiffer and Karpas-422 (EZH2-mutant) showed the highest level of growth inhibition upon SETD2 knockdown, consistent with the observation that they  are more sensitive to SETD2 inhibition than SUDHL4 and SUDHL5. As expected, SUDHL5, an EZH2 wild-type DLBCL cell line, did not respond to SETD2 knockdown (Extended Data Fig. 8c). Besides their antagonism with PRC2 activity, SETD2 and H3K36 trimethylation are also implicated in many other biological processes [39][40][41][42] , including DNA damage repair, and loss of SETD2 can lead to genome instability 40,42 . However, treatment with SETD2-IN-1 had little effect on levels of phosphorylated histone variant H2AX (γH2AX), a marker for the presence of DNA damage, suggesting that DNA damage is unlikely to be a major cause of sensitivity to SETD2-IN-1 in Karpas-422 cells (Extended Data Fig. 8d).
Profiling H3K36me3 and H3K27me3 in Karpas-422 using ChIP-Rx revealed that SETD2 inhibition substantially decreased levels of H3K36me3 and led to concomitant H3K27me3 spreading (Fig. 6b-e and Extended Data Fig. 9a,b). This spreading was pronounced in gene bodies of expressed genes, particularly those downregulated in Karpas-422 CXCdel +/− cells upon GSK343 withdrawal (Extended Data Fig. 9b). RNA-seq upon SETD2 inhibition revealed 225 differentially expressed genes, of which 137 genes (60.8%) were downregulated (Extended Data Fig. 9c Fig. 9d,e). Consistent with the observation for K562 and Karpas-422 CXCdel cells, gene repression and increases in H3K27me3 were modestly correlated (Extended Data Fig. 9e). More broadly, within gene bodies, changes in H3K27me3 observed upon SETD2 inhibition were correlated with changes in H3K27me3 observed for Karpas-422 CXCdel +/− cells upon inhibitor withdrawal, and this trend was even more evident for SETD2-IN-1 downregulated genes (Extended Data Fig. 9f). Loss of H3K36me3 was correlated with increases in H3K27me3 (Extended Data Fig. 9g). Gene ontology analysis of downregulated genes upon SETD2 inhibition revealed enrichment in pathways associated with cell division and apoptosis (Extended Data Fig. 9h). Altogether, these data suggest that SETD2 inhibition can induce epigenomic changes that partially phenocopy those induced by CXCdel.
Lastly, we reasoned that if CXCdel and SETD2 inhibition both lead to H3K27me3 spreading and growth inhibition, then CXCdel +/− cells should be more sensitive to SETD2-IN-1. Indeed, CXCdel +/− cells were most sensitive to SETD2-IN-1 in the absence of GSK343, while wild-type cells were the least sensitive (Fig. 6f,g). Collectively, these data suggest that SETD2 inhibition selectively blocks the growth of EZH2-mutant DLBCL cell lines, in part by inducing H3K27me3 spreading, nominating

Discussion
PRC2 is not only an important drug target for anticancer therapeutics but also serves as an instructive paradigm for the study of chromatin regulation 15,43 . By adapting CRISPR-scanning approaches, here we discover mutations in PRC2 that unexpectedly confer drug addiction in EZH2-mutant lymphoma and explore their broader implications on chromatin regulation. To the best of our knowledge, our study presents the first example of drug addiction to epigenetic inhibitors, revealing chromatin thresholds in specific cancer-mutant contexts. Through genomic and biochemical approaches, we show that drug addiction mutations in the EZH2 CXC domain cooperate with activating cancer mutations in the EZH2 SET domain to maintain H3K27me3 levels in the presence of mechanistically distinct PRC2 inhibitors. The CXC domain mutations increase PRC2 binding to mononucleosomes, which allosterically activates PRC2 by enhancing substrate affinity. Consistent with this idea, the incorporation of cofactors like AEBP2 and PHF1 into the PRC2 complex has also been found to stimulate its activity by increasing substrate binding affinity [44][45][46] . While it remains to be determined whether drug resistance or addiction mutations within the EZH2 CXC domain may arise in a clinical setting, our work nominates this region as a potential hotspot for PRC2-activating mutations in cancer. Indeed, we show that the homologous Q571R mutation found in EZH1 enhances EZH2 activity and confers drug addiction in EZH2-mutant lymphoma 30 . The selective occurrence of CXC-and SET-domain activating mutations in EZH1 and EZH2, respectively, in different cancer contexts suggests that their distinct mechanisms of PRC2 activation are functionally distinct for tumorigenesis and their possible differences remain to be explored.
While the mechanism underlying the activating CXC point mutation Q575R is intuitive, how deletion of the CXCdel loop enhances EZH2 nucleosome binding remains unclear. Rather than CXCdel being neomorphic, we speculate that it instead disables a yet-to-be-identified allosteric mechanism that inhibits PRC2 activity. Notably, while the CXC domain is highly evolutionarily conserved, the CXCdel loop is considerably less so (Extended Data Fig. 1j-k), suggesting that it is not necessary for protein folding but is instead involved in regulation. Furthermore, the loop directly contacts SUZ12. We speculate that deletion of the CXCdel loop reduces SUZ12 contact to confer greater conformational flexibility to the CXC domain 23,47 , which may subsequently enhance nucleosome binding. Although the precise mechanism of EZH2 activation remains to be determined, our studies highlight how drug resistance mutations can be employed to uncover distal protein sites involved in potential allostery 48 .
Beyond clinical implications, drug resistance mutations can reveal new target biology 48 . Notably, drug addiction mutations in PRC2 enabled the discovery of an H3K27me3 ceiling in EZH2-mutant lymphoma, informing our understanding of constraints on chromatin landscapes. This H3K27me3 ceiling forms the upper limit of a 'repressive methylation index'-a window of H3K27me3 levels permissible for lymphoma cell proliferation (Fig. 6h). Activating EZH2 SET domain mutations in lymphoma elevates H3K27me3 levels to block B-cell differentiation but in doing so creates a 'Goldilocks' epigenetic state near this repressive methylation ceiling. Whether H3K27me3 ceiling exists in non-EZH2-mutant tumors or regulates normal B-cell development remains to be further explored. These findings raise the possibility that other epigenetic signaling pathways may exhibit biphasic dependencies in cancer contexts.
Our study demonstrates how PRC2 activation represents an approach to target EZH2-mutant lymphoma by driving the cells beyond this H3K27me3 repressive ceiling. While direct pharmacological activation PRC2 has not been achieved, our studies show that indirect PRC2 activation via SETD2 inhibition could be used to surpass the H3K27me3 ceiling. Indeed, our results show that EZH2-mutant lymphoma cell lines are more sensitive to SETD2 inhibition. Overspreading of H3K27me3 upon inhibitor withdrawal and SETD2 inhibition is pronounced at the borders of H3K27me3 domains and within gene bodies, suggesting that subsequent down-regulation of cancer growth genes may mediate the antiproliferative effects. However, H3K27me3 overspreading may also alter physicochemical properties of chromatin and lead to broader biophysical consequences on chromatin organization (for example, hypercompaction) 49 . Consequently, this antiproliferative mechanism may starkly differ from cell growth inhibition caused by kinase inhibitor drug addiction, which activates pre-existing negative feedback circuits. Regardless of the underlying mechanisms, our findings suggest that strategies to activate PRC2, either directly or indirectly, may be promising therapeutic opportunities for EZH2-mutant lymphoma.
The paradoxical phenomenon of drug addiction has uncovered new cell signaling thresholds and therapeutic vulnerabilities, yet remains largely unexplored with most drug target classes beyond kinases 1,2 . To explore these possibilities, we developed CRISPR-addiction scanning, which leverages a drug withdrawal arm with accompanying analytical frameworks to identify drug addiction mutations across protein targets. While simple to implement, CRISPR-addiction scanning offers a conceptual advance to study underappreciated dimensions of drug-target relationships. More broadly, our work provides an instructive case study of how the identification of drug addiction mutations may be exploited to reveal new cancer vulnerabilities. Consequently, intentionally searching for drug addiction mutations using methods like CRISPR-addiction scanning may be warranted to interrogate these possibilities across broader drug target classes.

Online content
Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/s41589-023-01299-1.

Lentiviral infection
For lentivirus production, plasmids were cotransfected with GAG/POL and VSVG plasmids into 293FT using FuGENE HD (Promega). Media was exchanged after 7 h, and the viral supernatant was collected 48-72 h after transfection and filtered (0.45 μm). Cells were transduced by spinfection at 1,070g for 90 min at 37 °C with 8 μg ml −1 polybrene (Santa Cruz Biotechnology). After 48 h post-transduction, media was changed and puromycin (Thermo Fisher Scientific) selection was carried out for 7 d at 1-2 μg ml −1 .

sgRNA pooled cloning and CRISPR-suppressor scanning experiments
The PRC2 tiling library included all sgRNA with an NGG PAM cleavage site and off-target score (MIT Specificity Score) greater than 20 within the coding sequence of the three members of the core PRC2 complex as follows: EZH2 (NP_004447.2), EED (NP_001294936.1) and SUZ12 (NP_056170.2) 50 . These sgRNA sequences are listed in Supplementary Dataset 1. Oligos containing the sgRNA sequences were cloned into pLentiCRISPR.v2, a gift from F. Zhang (Addgene no. 52961), in a pooled fashion 50,51 . Lentiviral particles carrying the resultant PRC2 tiling library were generated as described above and titered 51,52 . Karpas-422 cells (4.8 × 10 7 ) were transduced at a multiplicity of infection (MOI) <0.3. After puromycin selection, cells were split into pools and treated with 1 μM GSK343, 1 μM EED226 or vehicle. Genomic DNA was isolated at specified time points. To measure the composition of the population, sgRNA sequences for all replicates were amplified with PCR primers and sequenced 16 .

Validation of enriched sgRNAs from CRISPR-suppressor scan
sgRNAs were cloned into pLentiCRISPR.v2; Karpas-422 and Pfeiffer cells were transduced with the resultant plasmids and selected with puromycin. Transduced Karpas-422 cells were treated with 1 μM of GSK343 or 1 μM of EED226 for 6 weeks. Transduced Pfeiffer cells were treated with incremental concentrations of GSK343 or EED226 (for GSK343-50 nM for 2.5 weeks, 100 nM for 1.5 weeks and 200 nM for 2 weeks; for EED226-200 nM EED226 for 2.5 weeks and 500 nM EED226 for 3.5 weeks). Viability was monitored by flow cytometry with Helix NP NIR viability dye (BioLegend).

Generation of clonal drug-resistant EZH2 mutant and knock-in cell lines
To generate drug-resistant mutant cells, Karpas-422 cells were transduced with pLentiCRISPR.v2 EZH2 sgD597 and treated with 1 μM GSK343 for 6 weeks to enrich for drug-resistant mutant cells. Surviving cells were single-cell-sorted into media containing 1 μM GSK343.
To generate CXCdel, Q575R Karpas-422 or K562 knock-in cell lines, electroporation was performed using NEON system with ribonucleoprotein (RNP) complex. gRNA complex was formed by reconstituting gene-specific crRNA and tracrRNA (IDT) at 1:1 ratio, hybridized by incubation at 95 °C for 5 min, followed by cooling to room temperature. RNP complex was formed by mixing gRNA complex with Alt-R Cas9 enzyme (IDT) and incubation at room temperature for 15 min. 2 × 10 5 cells were washed twice with PBS and resuspended in buffer R. RNP complex, single-stranded oligodeoxynucleotides and Alt-R Cas9 electroporation enhancer (IDT) were added to the cell suspension and electroporated at 1,000 V (50 ms pulse width for 1 pulse) for Karpas-422, 1,450 V (10 ms pulse width for 3 pulses) for Karpas-422. After electroporation, cells were immediately transferred to pre-warmed media supplemented with HDR Enhancer (IDT). To generate single-cell clones, cells were sorted on a MoFlo Astrios EQ cell sorter.

Generation of inducible SETD2 knockdown cell lines
shRNAs targeting SETD2 were cloned into Tet-regulated pRRL.RTREVIR miR-E shRNA expression vector. DLBCL cell lines Karpas-422, Pfeiffer, SUDHL4 and SUDHL5 were transduced with shRen.713 or shSETD2 (ref. 53). Transduced cell populations were sorted based on Venus fluorescence by flow cytometry. Expression of shRNA was induced with 1 μg ml −1 doxycycline. Knockdown efficiency was determined by quantification of SETD2 transcript levels. Total RNA was isolated from cells using RNeasy Mini Kit (Qiagen). cDNA synthesis was carried out using SuperScript III First-Strand Synthesis Supermix (Thermo Fisher Scientific). Quantitative PCR analysis was performed with SYBR Select Master Mix (Thermo Fisher Scientific) on a Bio-Rad CFX96 real-time PCR system. All signals were quantified using the delta Cq method and normalized to levels of GAPDH. The shRNA sequences and primers for the amplification of SETD2 transcripts are listed in Supplementary Table 2.

Genotype determination
Genomic DNA was isolated using the QIAamp Blood Mini Kit (Qiagen). For library preparation, genomic PCR primers with Illumina adapter sequences were used to amplify specified regions of EZH2 as previously described 16 . Samples were sequenced on a MiSeq instrument (Illumina). The sequencing reads were analyzed using CRISPResso2 (v.2.0.40) 54 .

Cell growth assays
Cells were seeded in 96-well plates with 45,000 cells per well in triplicate with drug or vehicle treatments. The number of viable cells was quantified at days 3, 7, 10 and 14 with CellTiter-Glo (Promega) on a SpectraMax i3x plate reader (Molecular Devices). ATP standard curve was prepared using known concentrations of ATP and used to calculate the ATP content of cells. On the days of measurement, cells were replated with fresh media containing drugs or vehicles.

Cell cycle analysis
Wild-type and CXCdel +/− Karpas-422 cells were treated with 1 μM GSK343 or vehicle for 7 d. Cells were washed with PBS and then fixed with cold 70% ethanol and incubated for a minimum of 2 h at −20 °C. After fixing, cells were washed twice with PBS and then resuspended in propidium iodide (PI) staining solution (1× PBS supplemented with 50 μg ml −1 PI, Invitrogen; 100 μg ml −1 RNase A, Qiagen; and 2 mM MgCl 2 ) and incubated for 30 min at room temperature in the dark. Data acquisition was performed on ACEA Novocyte flow cytometer (Agilent) using NovoExpress software (version 1.5.6). Data were analyzed with FlowJo (version 10.7.1).

Purification of recombinant PRC2
Human recombinant PRC2 used contained five members: EZH2, SUZ12, EED, RBAP48 and AEBP2 (UniProtDB entry isoform sequences Q15910-2, Q921E6-3, O75530-1, Q09028-1, Q6ZN18-1, respectively). Recombinant CXCdel PRC2 contained the same five complex members, with a 10-amino acid deletion in EZH2 from T592-S601. Recombinant Q575R PRC2 contained the same five complex members. All members were tagged with an N-terminal MBP tag cleavable by PreScission protease. pFastBac plasmids for the five wild-type PRC2 members were gifts from T. Cech. The proteins were co-expressed in insect cells according to the Bac-to-Bac baculovirus expression system. Detection of gp64 was used to determine baculovirus titer (Expression Systems). For expression, SF9 cells were grown to a density of 2 × 10 6 cells per ml and infected with PRC2 baculovirus at an MOI of 1.2 for AEBP2 and 0.7 for all other subunits. The cells were incubated for 72 h (27 °C, 120 rpm), collected, and then frozen with liquid nitrogen for future purification.
PRC2 complexes were purified according to a published procedure 28,55 . All protein purification steps were performed at 4 °C. Insect cells were first lysed by incubating for 1 h with lysis buffer (10 mM Tris-HCl, pH 7.9, 250 mM NaCl, 0.5% Nonidet P-40 Substitute, 1 mM TCEP) at a ratio of 8 ml of buffer per gram of cell pellet. The lysate was clarified by centrifugation at 29,000g for 40 min. The clarified lysate was added to equilibrated amylose resin (NEB, E8021S; 0.3 ml of resin per gram of cell pellet) and bound in batch for 2 h. The resin was then washed with >10 column volumes (CV) of lysis buffer, followed by >16 CV of Wash buffer 1 (10 mM Tris-HCl, pH 7.9, 500 mM NaCl, 1 mM TCEP) and lastly >16 CV wash buffer 2 (10 mM Tris-HCl, pH 7.9, 150 mM NaCl, 1 mM TCEP). PRC2 was eluted by the addition of 3 CV of wash buffer 2 + 10 mM maltose. Protein was concentrated to 15 mg ml −1 using Amicon Ultra-15 Centrifugal Filter Unit, 30 kDa molecular weight cut off (MWCO). In-house-made PreScission protease was added at a ratio of 1:100 (protease: PRC2) and the NaCl concentration was adjusted to 250 mM by the addition of the appropriate amount of 5 M NaCl. After overnight incubation, cleavage of the MBP tags was confirmed by SDS-PAGE. Next, the cleaved protein was subjected to a 5 ml HiTrap Heparin column at a flow rate of 1.5 ml min −1 with gradient elution from Buffer A (10 mM Tris-HCl, pH 7.9, 150 mM NaCl, 1 mM TCEP) to Buffer B (10 mM Tris-HCl, pH 7.9, 2 M NaCl, 1 mM TCEP) over 35 CV. The salt concentration was diluted to 350 mM, and the protein was concentrated and subjected to size exclusion over a Superose 6 Increase 10/300 GL column with running buffer (20 mM HEPES, pH 7.9, 150 mM NaCl, 1 mM TCEP) at a flow rate of 0.5 ml min −1 . PRC2-peak fractions were verified with SDS-PAGE. The correct fractions were pooled and concentrated as indicated above. The final concentration was determined by UV absorbance at 280 nm measured by Nanodrop. The ratio of absorbance at 260 nm/280 nm ratio was less than 0.7 indicating minimal nucleic acid contamination.

Purification of mononucleosomes
Xenopus histones were expressed in Escherichia coli and purified as previously described with minor modifications 56,57 . Briefly, for the histone purification, inclusion bodies were prepared and extracted as described previously 56,57 . These pellets were subsequently dissolved into an unfolding buffer (7 M guanidinium HCl, 20 mM Tris-HCl, pH 7.5, 10 mM DTT), dialyzed into SAU-200 buffer (7 M urea, 20 mM sodium acetate pH 5.2, 200 mM NaCl, 5 mM 2-mercaptoethanol, 1 mM EDTA) and loaded onto a 5 ml HiTrap SP HP column, eluted by gradient elution over 30 CV from SAU-200 buffer to SAU-600 buffer (7 M urea, 20 mM sodium acetate pH 5.2, 600 mM NaCl, 5 mM 2-mercaptoethanol, 1 mM EDTA). Histone-containing fractions were pooled, dialyzed into distilled water containing 2 mM 2-mercaptoethanol (at least three changes) and lyophilized for storage at −20 °C 58 . For histone octamer assembly, the histones were first dissolved in an unfolding buffer (20 mM Tris-HCl pH 7.5, 6 M guanidinium HCl, 5 mM DTT), combined in equimolar amounts and dialyzed into refolding buffer (10 mM Tris-HCl pH 7.5, 2 M NaCl, 1 mM EDTA, 5 mM b-ME) and purified over a Superdex 200 column (GE Healthcare). Octamer-containing fractions were pooled according to purity and histone stoichiometry, and then flash-frozen for storage at −80 °C after the addition of glycerol to a final concentration of 10% (vol/vol).
The 185-base pair (bp) nucleosome DNA used with Xenopus nucleosomes was atcgctgttcaatacatgcacaggatgtatatatctgacacgtgc ctggagactagggagtaatccccttggcggttaaaacgcgggggacagcgcgtacgt gcgtttaagcggtgctagagctgtctacgaccaattgagcggcctcggcaccgggattctc cagggcggccgcgtatagggat. The 601-nucleosome positioning sequence is underlined. The plasmid containing the 185-bp nucleosomal DNA sequence was a gift from Philip Cole (Harvard Medical School). The 185-bp nucleosomal DNA sequence was prepared via large-scale plasmid preparation followed by cleavage with the EcoRV-HF restriction enzyme. The cleavage product was purified with Q Sepharose HP resin with a stepwise elution from Buffer A (20 mM Tris pH 8.0, 1 mM EDTA) to 40% Buffer B (20 mM Tris pH 8.0, 1 mM EDTA and 1 M NaCl) and elution with 70% Buffer B. Eluted DNA was subsequently purified by size exclusion on a Superdex 200 column in Buffer B. The resulting product was buffer-exchanged into Buffer A and stored at −20 °C. Nucleosomes were reconstituted by mixing the DNA and the octamer in a 1:1.1 ratio and performing gradient dialysis as described previously 57 . Reconstituted nucleosomes were purified by anion exchange with a TSKgel DEAE-5PW (7.5 mm × 7.5 cm) column; their composition was confirmed by native gel electrophoresis and was subsequently concentrated and stored as described previously 56 . All dialysis steps employed 6-8 kDa MWCO dialysis tubing (Spectra-Por), and 10 kDa MWCO Amicon centrifugal filter units were used for concentration and desalting.

EMSA with mononucleosomes
45 nM of mononucleosomes purified as above were incubated with increasing amounts of protein in 20 μl of assay buffer (50 mM Tris-HCl, pH 7.5, at 25 °C, 25 mM KCl, 2.5 mM MgCl 2 , 0.1 mM ZnCl 2 , 2 mM 2-mercaptoethanol, 0.1 mg ml −1 BSA, 10% vol/vol glycerol). After incubation at 30 °C for 30 min, 15 μl of the reaction was loaded onto a 3.5% TBE polyacrylamide gel. Gel electrophoresis was performed at 140 V for 1 h in cold 1× TBE at 4 °C. The gel was stained with SYBR Green I (Thermo Fisher Scientific), and fluorescence images were acquired by an Azure c200 imager. Densitometry of the bound PRC2-mononucleosome band was performed by ImageJ.

MTase-Glo enzyme assay
Endpoint methyltransferase measurements were performed in a 20 μl reaction volume using 80 nM of PRC2 enzyme with 1 μM of mononucleosomes and 40 μM SAM in reaction buffer (50 mM Tris-HCl, pH 8 at 30 °C, 10 mM KCl, 2.5 mM MgCl 2 , 0.1 mM ZnCl 2, 2 mM 2-mercaptoethanol, 0.1 mg ml −1 BSA, 5% v/v glycerol). Reactions were incubated at 30 °C for 2 h and quenched with 0.1% TFA. Enzyme activity was assessed by S-adenosyl homocysteine (SAH) production as determined by luminescence through the MTase-Glo assay (Promega). Briefly, SAH production for each reaction was calculated by correlating the raw luminescence to an SAH standard curve correcting for basal luminescence of mononucleosome and buffer to determine SAH production.

RNA-seq
Total RNA was isolated from cells in triplicate separate cultures using RNeasy Mini Kit (Qiagen). Library preparation was performed using the QuantSeq 3′ mRNA-Seq Library Prep Kit FWD (Lexogen). Libraries were sequenced using a NovaSeq SP kit (Illumina) on a Novaseq instrument (Illumina).

ChIP-seq data analysis
For quantitative ChIP-seq analysis, reads were aligned to a composite reference genome for dm6 and hg38 utilizing bowtie2 (version 2.3.2) according to the Spiker analysis framework (version 1.03). Aligned reads were sorted using samtools (version 0. 1.19), and the Spiker split_bam.py script was run to calculate scaling factors and to split reads into an alignment file containing reads specifically aligning to the hg38 genome (Spiker, version 1.0.3) 60 . This alignment file was then converted into spike-in normalized bigwig file for further analysis and visualization utilizing the deepTools bamCoverage command with the following parameters: -binSize 50-ignoreDuplicates-normalizeUsing RPKM-ignoreForNormalization chrM-extendReads -scaleFactor. Peaks for H3K27me3 and H3K36me3 were called using Sicer with the parameters: -e 8000 -s hg38 -g 12000 -w 4000 (version 2.0). Individual replicate peaks were subsequently combined utilizing Bedtools2 merge -d 12000 (version 2.26.0). Border regions of peaks were called by using the Bedtools2 'flank' command with the parameter -b 50000 (version 2.26.0). For bin-level analysis, the H3K27me3 signal value was averaged across 50 kb bins utilizing the multiBigwigSummary function (deepTools, version 3.4.3). To generate metaprofile plots comparing between treatment conditions, the log2 option in the bigwigCompare function was used before plotting with the plotHeatmap function (deepTools, version 3.4.3). To generate fingerprint plots investigating H3K27me3 spreading, the plotFingerprint function was used on the hg38 split alignment files for individual replicates with the parameters: --ignoreDuplicates --extendReads --skipZeros (deepTools, version 3.4.3). Subsequent data processing was performed in Python (version 3.9.6) using standard plotting functions.
For EZH2 ChIP-seq analysis, reads were aligned to the hg38 genome utilizing bowtie2 (version 2.3.2). Aligned reads were sorted using samtools (version 0. 1.19). This alignment file was then converted into bigwig format for further analysis and visualization utilizing the deepTools bamCoverage command with the following parameters: -binSize 50 --ignoreDuplicates --normalizeUsing RPKM --ignoreForNormalization chrM -extendReads --blackListFileName. Specifically, the file 'ENCFF356LFX.bed' downloaded from Encode was used to define blacklisted regions for hg38. For bin-level analysis, the EZH2 signal value was averaged across 50 kb bins utilizing the multiBigwigSummary function (deepTools, version 3.4.3). Subsequent data processing was performed in Python (version 3.9.6) using standard plotting functions.

RNA-seq data processing and differential gene expression analysis
RNA-seq data were processed according to the QuantSeq 3′ mRNA-Seq Library Prep recommended analysis pipeline through alignment to the Ensembl transcriptome (GRCh38). DESeq2 with R (v.3.5.1) was employed for the differential expression analysis. All genes with adjusted P < 0.05, calculated using the Benjamini-Hochberg correction, and |log 2 (fold-change)| > 0.5 were considered to be differentially expressed. The location of the transcription start site (TSS) and transcription end sites (TES) of genes was determined using the biomaRt package (version 2.41.1), and only genes with a defined hgnc symbol name were included in the annotation list. When multiple isoforms were present, the TSS and TES were defined to be at the minimum and maximum positions, respectively. A table of substantially differentially expressed genes is provided in Supplementary Datasets 4-6.

Quantification and statistical analysis
Microsoft Excel (version 16.69.1), PRISM (version 9.3.1) (GraphPad) and R (version 3.6.2) software packages are used to perform statistical analyses. Statistical parameters including the exact value and definition of n, the definition of center, dispersion, precision measures (mean ± s.d. or s.e.m.), statistical significance and statistical tests used are reported in figures and figure legends. Protein structures were visualized using PyMOL (version 2.5.2) (Schrödinger).

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

Data availability
ChIP-seq and RNA-seq data have been deposited to NCBI GEO (GSE199889). Transformed CRISPR-suppressor scanning reads (log 2 + 1) used for Fig. 1 and Extended Data Fig. 1 are supplied in Supplementary Dataset 2. Addiction scores and growth rates used for Fig. 2 and Extended Data Fig. 2 are supplied in Supplementary Dataset 3. Oligonucleotide sequences are provided as Supplementary Information. Unprocessed gel and immunoblot images, as well as additional data generated in this study, are provided as Source Data with this paper. The following publicly available datasets were used Protein Data Bank (PDB) accession codes 6WKR, 5K0M, 5LS6, EMDB-7306; and ENCODE datasets ENCFF932BQZ, ENCFF435PXJ and ENCFF603BHK. Source data are provided with this paper.

Code availability
Computer code employed in Fig. 2 for the calculation of the addiction score is publicly available at https://github.com/skissler/EZH2 under the GNU General Public License, Version 3.0 (details included in code repository). Custom code used to analyze CRISPR-suppressor scanning data is available at https://github.com/liaulab/CRISPR-suppressor_ scanning. All other computer codes employed are available upon request.