Targetable vulnerabilities in T- and NK-cell lymphomas identified through preclinical models

T- and NK-cell lymphomas (TCL) are a heterogenous group of lymphoid malignancies with poor prognosis. In contrast to B-cell and myeloid malignancies, there are few preclinical models of TCLs, which has hampered the development of effective therapeutics. Here we establish and characterize preclinical models of TCL. We identify multiple vulnerabilities that are targetable with currently available agents (e.g., inhibitors of JAK2 or IKZF1) and demonstrate proof-of-principle for biomarker-driven therapies using patient-derived xenografts (PDXs). We show that MDM2 and MDMX are targetable vulnerabilities within TP53-wild-type TCLs. ALRN-6924, a stapled peptide that blocks interactions between p53 and both MDM2 and MDMX has potent in vitro activity and superior in vivo activity across 8 different PDX models compared to the standard-of-care agent romidepsin. ALRN-6924 induced a complete remission in a patient with TP53-wild-type angioimmunoblastic T-cell lymphoma, demonstrating the potential for rapid translation of discoveries from subtype-specific preclinical models.

T -cell lymphomas (TCLs) comprise approximately 10% of all non-Hodgkin lymphomas (NHLs) in Western countries. Two-thirds of these lymphomas are peripheral TCLs (PTCLs) and the remainder are cutaneous TCLs (CTCLs). Many patients with CTCL have an indolent course that can span decades. In contrast, nearly all PTCLs are aggressive lymphomas and include both αβ and γδ T-cell lymphomas 1 . Emblematic of our poor understanding of PTCL biology, the most common subtype of PTCL is PTCL-NOS (not otherwise specified).
Treatment of PTCLs to date has been largely derivative or empiric. CHOP-based regimens, which were adopted based on their activity in B-cell lymphomas, remain the standard for firstline therapy. Except for ALK-rearranged anaplastic large-cell lymphoma (ALK+ ALCL), over 75% of PTCLs fail to respond or relapse after first-line therapy in <2 years. PTCL-NOS, angioimmunoblastic T-cell lymphoma (AITL), ALK-negative (ALK−) ALCL, and NK-TCL, the most common subtypes of PTCL, are each associated with <35% 5-year overall survival 2 . Long-term survivors of some like hepatosplenic T-cell lymphoma (HS-TCL) are extremely uncommon 2 .
Current approaches to treat PTCL are simply inadequate. The National Comprehensive Cancer Network guidelines list clinical trial as the top choice for all patients with PTCL (other than ALK + ALCL) in both the first-line and relapsed/refractory settings. Improved understanding of the pathobiology of TCLs could highlight therapeutic targets, but there is a severe lack of wellcharacterized cellular and in vivo models of these lymphomas. To address this, we here report a comprehensive characterization of a newly established panel of patient-derived xenografts (PDXs) from multiple subtypes of TCL and available TCL cell lines. We highlight multiple potentially targetable genetic alterations and functional vulnerabilities. In TP53-wild-type TCL, MDM2, and MDMX are prominent therapeutic targets. ALRN-6924 shows promising pre-clinical activity in TP53-wt cell lines and PDX models and induced a complete remission in a patient with AITL. Thus, we demonstrate the potential of pre-clinical target validation and drug testing in models of rare cancers with direct clinical translation.

Results
Comprehensive characterization of TCL models. We collected a panel of previously established cell lines of T-and NK-cell lymphomas and validated their immunophenotype and STR-profiles (Supplementary Table 1, Supplementary Table 2). Furthermore, we defined the immunophenotypes of 17 TCL PDX models by flow cytometry and immunohistochemistry (IHC) and confirmed the expected histologic and phenotypic characteristics (Supplementary Table 2, Supplementary Fig. 1). To mitigate genetic drift that may occur over serial passaging 3 , we utilized PDXs in passage ≤2 and made these passages and affiliated data available through the Public Repository of Xenografts (www.PRoXe.org) 4 .
To characterize genomic alterations across TCL models, we performed whole exome and transcriptome sequencing (RNAseq) of 20 TCL lines using the algorithm established for the Cancer Cell Line Encyclopedia (https://portals.broadinstitute.org/ ccle). We also sequenced 17 TCL PDX models (15 disseminated/ orthotopic, 2 subcutaneous) and 13 additional cell lines using RNA-seq and targeted sequencing panels for recurrently mutated genes (Fig. 1a, Supplementary Tables 3, 4 and 5). Mutations identified in models derived from individual PTCL subtypes largely recapitulated those observed in patients (Fig. 1a, Supplementary Fig. 2). For example, ALK-ALCL frequently harbor JAK1 and/or STAT3 mutations 5 ; these were also observed in TCL models (Fig. 1a), including 3 of 4 breast implant-associated ALCL models.
AITL is the second most-common subtype of PTCL but there were previously no cell lines or faithful in vivo models available. We generated four AITL PDXs, including models that harbor mutations in the most commonly mutated genes in this disease (TET2, DNMT3A, and/or RHOA) (Fig. 1a) 6,7 . Similarly, no in vivo models of T-cell prolymphocytic leukemia (T-PLL) were previously available. We created two T-PLL PDXs, and confirmed ATM mutations and TCL1 rearrangement that characterize this disease (Fig. 1a, Supplementary Fig. 1f).
To identify copy number alterations, we applied the GISTIC (Genomic Identification of Significant Targets in Cancer) algorithm to whole exome sequencing data from 20 cell lines. We identified recurrent amplifications and deletions, including deletions of 9p21.3 that include CDKN2A and 17p13.1 that include TP53 (Fig. 1b, Supplementary Data 1). In fact, TP53 was deleted and/or mutated in 15 of 21 (71.4%) TCL lines, whereas TP53 alterations occur in <20% of patients with untreated PTCL or CTCL [6][7][8] . Of note, none of the 17 PDX models had TP53 mutations (p < 0.001 compared to cell lines, Fisher's exact test).
We utilized RNA-Seq to identify structural variants in TCL cell lines and PDXs. Recent studies have reported fusions involving DUSP22 or TP63 as recurrent events of prognostic significance in ALCL lacking ALK rearrangements 9 . RNA-Seq identified a DUSP22 fusion in the FEPD cell line and fusions of TP63 in 4 lines (Fig. 1c). The latter included TBL1XR1-TP63 and FOXK2-TP63 fusions, which were confirmed by sequencing and fluorescence in situ hybridization ( Supplementary Fig. 3a-c). In all TP63 fusions, the N-terminus of TP63 is deleted, creating a TP63 element similar to the ΔNTP63 (p40) variant (Supplementary Fig. 3a) 10 .
To characterize the transcriptomes of TCL cell lines and PDXs, we applied a sample-sample correlation matrix to cluster gene expression profiles across models. Both ALK+ ALCL and AITL formed distinct clusters, consistent with their unique cells-oforigin, while the remaining models formed a heterogeneous cluster ( Supplementary Fig. 4a). As expected, gene set enrichment analysis (GSEA) identified enrichment of STAT3 target genes in ALCL models and enrichment of follicular helper T-cell (T FH ) genes in AITL models ( Supplementary Fig. 4b, c and Supplementary Data 3 and 4).
CRISPR-Cas9 screen identifies vulnerabilities in TCL. To begin defining additional vulnerabilities across TCL subtypes, we utilized a genome-scale CRISPR-Cas9 loss-of-function screen. Of 21 lines, 8 (representing 4 subtypes) were adequately transducible with sufficient Cas9 activity for screening. We ranked T-cell lymphoma-specific dependencies by comparing to 383 additional lines screened using the same platform 14 (Supplementary Data 5 and 6). Integration of gene expression, recurrence and statistical significance (FDR q-value <0.05) resulted in a list of 30 genes as top vulnerabilities (Fig. 2a-d). IRF4, STAT3, and BCL6 were previously implicated as vulnerabilities in ALCL 5, 15 but the remaining genes are largely novel discoveries for TCL.
We validated IRF4 dependence across multiple cell lines with shRNA ( Supplementary Fig. 5a). IRF4 is known to be regulated by the IKAROS Family Zinc Finger 1 (IKZF1) transcription factor, which was also a vulnerability across multiple lines ( Fig. 2a-d). IRF4 was previously identified as a vulnerability in ALCL, where it is positively regulated by CD30 16 and STAT3 17 . Our data demonstrate that cell lines from both PTCL-NOS and CTCL subtypes are also dependent on IRF4. In B-cell malignancies, lenalidomide is known to induce IKZF1 and IKZF3 degradation by acting as a proteolysis-targeting chimera (PRO-TAC) 18 . Indeed, lenalidomide treatment reduced proliferation in 4 of 7 tested cell lines ( Supplementary Fig. 5b). A preliminary analysis demonstrated lower IRF4 protein levels in lines that were sensitive to lenalidomide (p = 0.0025, unpaired two-sided t-test;     Supplementary Fig. 5c, d). A larger analysis of primary specimens will be needed to determine whether low IRF4 protein level could serve as a predictive biomarker for lenalidomide response in patients with TCL. We noted that the phosphatase PTPN2 was the 2nd ranked vulnerability across lines. PTPN2 is an important regulator of Tcell tolerance and development 19 and PTPN2 loss in tumor cells can enhance sensitivity to immune checkpoint blockade 20 . Confirming the results of the CRISPR-CAS9 screen, we validated PTPN2 as a specific vulnerability in KI-JK cells but not in SMZ-1 cells, which were not sensitive to PTPN2 knockout in the screen and in an independent validation ( Supplementary Fig. 5e, f).
All 4 ALK+ ALCL lines tested in the CRISPR-CAS9 screen were sensitive to the ALK inhibitor crizotinib but the ALK-ALCL line FEPD was not. We noted that ALK did not score as a vulnerability in any of the ALK+ ALCL lines. This was expected as all 4 sgRNAs targeting ALK in the CRISPR-CAS9 screen were directed against the 5′ portion of the gene that is not included in the NPM-ALK fusion ( Supplementary Fig. 5g, h).
MDM2 and MDMX are targetable vulnerabilities in TP53wild-type TCL. We noted that both MDM2 and MDMX, which negatively regulate wild-type p53 function, scored as vulnerabilities in the CRISPR-Cas9 screen exclusively in the 4 TP53wild-type TCL lines ( Fig. 2a-e, Supplementary Fig. 6a). Notably, the SUP-M2 ALCL line harbors an MDMX gene amplification with high protein expression (Fig. 3a, b) and had the highest dependency score for MDMX among the 391 cell lines screened (Fig. 2f). We validated functional dependences on both MDM2 and MDMX using doxycycline-inducible shRNA (Fig. 2g).
To inhibit MDM2 and MDMX, we utilized the stapled peptide ALRN-6924, which is currently in phase I/II trials 21 . ALRN-6924 was potently active against all 9 TP53-wild-type cell lines (IC50 100 nM-4 μM) and in MTA cells, which harbor a heterozygous S215G mutation that is believed to confer loss-of-function. In contrast, all 11 lines with homozygous, hemizygous or dominant negative (p.R273H 22 in SUDHL1) TP53 mutations had IC50 values >10 μM (Fig. 3c, d). In TP53-wild-type lines, ALRN-6924 induced a dose-dependent increase in levels of both p53 protein and the p53-target p21, which was associated with G0/G1 cell cycle arrest and induction of apoptosis (Supplementary Fig. 6c-e). ALRN-6924 had superior potency to the small molecule MDM2 inhibitor RG-7112 in a subset of lines, including 2 lines with MDMX amplification (Fig. 3a,d, Supplementary Fig. 6f-h).
Because of the broad activity of ALRN-6924, we assessed its efficacy in eight TP53-wild-type PDX models representing six different TCL subtypes. We established a maximum tolerated dose (MTD) for the standard-of-care agent romidepsin of 1 mg per kg given intraperitoneally on days 1, 4, and 7 ( Supplementary  Fig. 7a), which was the same dose previously used in lymphoma cell line xenografts 23 . We compared romidepsin at this dose to ALRN-6924 dosed at 20 mg per kg given intravenously on days 1, 4, and 7. This dose of ALRN-6924 results in a similar area under the plasma concentration curve (AUC, 686-759 µg x h x mL −1 ) in mice to the AUC in patients (831 µg x h x mL −1 ) after a single dose of 3.1 mg per kg 21 , which is the recommended phase II dose.
Mice were xenografted with tumor, allowed to develop significant disease burden and then randomized to treatment with vehicle, romidepsin or ALRN-6924 on days 1, 4 and 7. All mice were sacrificed on day 8 to assess disease burden across compartments. ALRN-6924 treatment induced p21 expression by IHC and apoptosis in treated tumors ( Supplementary Fig. 6i, j). ALRN-6924 was broadly active across bone marrow, spleen and other involved compartments in all 8 models. ALRN-6924 was superior or equivalent to romidepsin across models, with the exception of bone marrow disease in 2 models (Fig. 3e).
In the phase 1/2a trial of ALRN-6924 (NCT02264613), the first patient treated for TCL experienced a complete remission (Fig. 3f). She was diagnosed with AITL and relapsed almost immediately after first-line CHOEP chemotherapy. A repeat biopsy confirmed AITL with wild-type TP53 and no copy number alterations at TP53, MDM2, or MDMX. She was started on ALRN-6924 given weekly for 3 weeks out of 4. She remains in a complete remission after two years of therapy.

Discussion
We established and comprehensively characterized preclinical models of TCL to facilitate both mechanistic discovery and biomarker-driven therapy. Immunophenotyping, DNA sequencing and RNA-based clustering all suggest that these models recapitulate important aspects of TCL biology. Importantly, the PDX and cell line models provide a pre-clinical toolbox for interrogating genetic, epigenetic, metabolic and other alterations of potential clinical significance.
Our CRISPR-CAS9 screen identified multiple new targets but was limited by the transducibility of available cell lines. Clearly, additional efforts are needed to extend genome-wide vulnerability screening to additional lines from other subtypes. In the 8 lines we screened, we rediscovered multiple known targets, most notably IRF4. It remains unclear whether clinical responses to lenalidomide, which occur in approximately 22% of patients with relapsed/refractory PTCL 24 , involve the degradation of IKZF1, as shown in B-cell malignancies 18 . Further studies are needed to clarify whether responses to lenalidomide among patients with PTCL result from tumor cell-autonomous responses and, if so, to identify the relevant targets that are degraded in the presence of lenalidomide.
We were surprised that PTPN2 scored as a vulnerability in the CRISPR-CAS9 screen in 7 of 8 cell lines. PTPN2 was previously described as a tumor suppressor in T-ALL 25,26 . PTPN2 can negatively regulate JAK/STAT signaling in CD8+ T-cells 27 . In fact, deletion of PTPN2 in mature T cells results in autoimmunity in aged mice, possibly due to a decreased threshold for T-cell receptor (TCR) signaling 28 . A previous analysis identified biallelic frameshift mutations in one of 69 PTCLs 29 , suggesting that PTPN2 could also function as a tumor suppressor in mature T cell malignancies. An alternate possibility is that hyperactivation of TCR or JAK/STAT signaling resulting from PTPN2 loss (or inhibition) is toxic to some or all mature T cells. This concept was previously demonstrated for multiple phosphatases that buffer Bcell receptor signaling in B-cell malignancies 30 . Further studies are needed to determine the role of PTPN2 in PTCLs and whether loss of this phosphatase can induce a central tolerance checkpoint in lymphoma cells 31 .
We validated JAK2 fusions and wild-type p53 as targets for small molecule inhibitors using in vivo PDX models. The presence of a JAK2 fusion in a T-PLL PDX was unexpected, but aberrant activation of this pathway is consistent with the activating JAK/STAT pathway mutations that have been described in T-PLL 32 . The in vivo response that we observed suggests that identifying JAK2 fusions in this disease could offer significant clinical benefit, especially in a disease like T-PLL with no standard cytotoxic or targeted regimens for second-line therapy. We showed that ALRN-6924 has broad activity against multiple TCL lines and PDXs. It is extremely complicated to compare the activity of multiple drugs across multiple models, especially because the pharmacokinetics of each drug differ between mouse and human. Nonetheless, we used a maximum tolerated dose of romidepsin as a comparator to provide some relevance to current clinical treatment. The data across all 8 lines suggest that romidepsin has impressive activity in the bone marrow but ALRN-6924 was active across nearly all involved sites. Further studies are needed to understand whether these differences in compartmentspecific activity are related to drug distribution or other microenvironmental determinants of efficacy.
Our study demonstrates that data from genome-wide vulnerability screens and in vivo PDX trials are highly complementary and can form the basis for clinical testing in rare cancers. We continue to generate models, including two additional AITL    PDXs and one additional PTCL-NOS PDX that engrafted during the review of this manuscript. All models are available to both academic and industry investigators and will hopefully facilitate the development of more effective treatment strategies for patients with TCL.
Proliferation, cell cycle, and apoptosis assays. Proliferation assays with ruxolitinib and ALRN-6924 were performed in a 384-well format, using a JANUS Automated Workstation (PerkinElmer) and CellTiter-Glo Luminescent Cell Viability reagent (Promega) at 0 h and 72 h according to the manufacturer's protocol. Each data point represents quadruplicates and experiments were repeated at least twice. IC50 values were calculated, using the GR calculator at www.grcalculator.org. Proliferation assays with lenalidomide were performed in triplicate and performed twice, with a concentration of 1 μM vs DMSO, cells were stained with PI, counted and medium was replaced every 48 h. Statistical significance was assessed by two way ANOVA and Bonferroni post-test. Detection of apoptosis was performed by Annexin V staining using the biolegend® FITC Annexin V Apoptosis Detection Kit with 7-AAD. Cell cycle analysis were performed with Hoechst33342 from Sig-maAldrich® according to the manufacturer's protocol.
shRNA toxicity assay. Tet-pLKO-puro was a gift from Dmitri Wiederschain (Addgene plasmid # 21915). We exchanged the sequence of IRES-PuroR in the plasmid with that of P2A-EGFP using gBlocks® from Integrated DNA Technologies. Each shRNA was cloned into the Tet-pLKO-EGFP vector according to the protocol provided on the webpage (https://media.addgene.org/data/plasmids/21/ 21915/21915-attachment_Jws3xzJOO5Cu.pdf). We induced each plasmid into the target cells by lentivirus. psPAX2 and pCMV-VSV-G were used for lentiviral transduction. psPAX2 and pCMV-VSV-G were gifts from Didier Trono and Bob Weinberg (Addgene plasmids # 12260 and # 8454, respectively). Four days after the transduction, shRNA was induced by administration of doxycycline (final 1 μg permL, Clontech, Cat no. 631311). Fractions of GFP were measured over time by flow cytometry (BD FACSCanto II HTS, BD Biosciences).
RT-PCR. SuperScript III Reverse Transcriptase (Thermo Fisher Scientific) was used to prepare cDNA per manufacturer's protocol. Primers for TBL1XR1-TP63 were F, ttgcaagagtcaggattttctca, and R, gaggagccgttctgaatctg. Primers for FOXK2-TP63 were F, tcttcagggtacaaggtggg, and R, gaggagccgttctgaatctg. Primers for ACTB were F, gctcgtcgtcgacaacggctc and R, caaacatgatctgggtcatcttctc. . For the remaining cell lines, Whole exome sequencing was performed by the Cancer Cell Line Encyclopedia Platform at the Broad Institute according to standard protocols of the Cancer Cell Line Encyclopedia. Briefly, DNA was extracted using the QIAGEN® QIAamp DNA Mini Kit according to the manufacturer's protocol. Prior to library preparation, DNA was fragmented (Covaris sonication) to 250 bp and further purified using Agentcourt AMPure XP beads. Size-selected DNA was then ligated to specific adapters during library preparation (Kapa Library Prep, Kapa Biosystems). Each library was made with sample-specific barcodes. Libraries were pooled and sequenced over a HiSeq3000.
All the raw sequencing data was processed using CCLE variant calling pipeline. Mutation analysis for single nucleotide variants (SNVs) was performed using MuTect v1.1.6 in single sample mode with default parameters. Short indels were detected using Indelocator (http://archive.broadinstitute.org/cancer/cga/ indelocator) in single sample mode with default parameters. To ensure high quality variant calls, we required minimum coverage of 4 reads with minimum two reads supporting the alternate allele. Moreover, variants with low allelic fraction (AF < 0.1) and variants outside protein coding region were excluded. To remove germline-like variants, any variant with allelic frequency greater than 1E-5 in Exome Aggregation Consortium (ExAC) project 63 were excluded with the exception of any cancer recurrent variant defined by minimum TCGA frequency of 3 or COSMIC frequency of 10.
For targeted exon capture and next-generation sequencing of all coding exons of 265 genes previously implicated in PTCL (Supplementary Table 4), genomic DNA was extracted from banked xenografted tumor cells harvested from ≥ P1 mice following immunomagnetic depletion of murine cells (EasySep Mouse/Human Chimera Isolation Kit, #19849, StemCell Technologies, Vancouver, BC, Canada). DNA underwent customized hybrid-capture target enrichment (SureSelect, Agilent, Santa Clara, CA, USA) and Illumina sequencing. Genomic DNA from the splenocytes of a normal NSG mouse was sequenced in order to enhance speciesspecific filtering of human reads. Known germline polymorphisms from the Exome Sequencing Project and the dbSNP (build 142) databases were excluded. We applied the MutSigCV algorithm to identify genes that were altered more often than expected by chance given the background mutation rate. We only considered variants predicted to be non-silent (i.e., missense, nonsense, translational start site, or splice site alterations, or in-frame or frame-shift insertions/deletions) and with variant allele frequencies of 5% or more. For every alteration meeting these criteria, sequencing reads were individually visualized to evaluate mapping quality as well as the phase and spatial distribution of alterations within reads. Mutations were then called by cross-referencing candidate variants to ClinVar, COSMIC, and a review of published literature.
To validate mutations, cell lines and PDX tumors were sequenced using the Rapid Lymphoma Panel, an amplicon-based resequencing platform levying a custom panel of oligonucleotide probes which target a bevy of genes (1420 positive and negative strand probes for 98 gene targets, Supplementary  Figure 3) we used the expression matrix from Cufflinks 67 . All snoRNA and miRNA genes were first filtered out as were all genes that were only significantly expressed in 3 or fewer samples based on a FPKM threshold of 5. A custom R script then calculates the Pearson correlation between all of the samples on a pairwise basis and generates the heatmap.
For GSEA, differential expression between experimental criteria was determined using raw counts and normalization procedures within the DESeq2 R package, using a negative binomial distribution. Genes were ordered by the Wald statistics and these ordered lists were used in GSEA (Broad Institute). The false discovery rate (FDR-Benjamini and Hochberg) method was used to adjust for multiple comparisons.
To permit analysis of the aggregated cell line and PDX data, we first performed quantile normalization to adjust for library depth and platform-specific differences. Principal components analysis revealed that the primary source of postnormalization variation derived from batch effects. These batch effects within PDX and Cell Line were successfully removed using the ComBat approach from SVA 3.18.00 68 .
Archer fusion panel. An Anchored Multiplex PCR (AMP) assay was used for targeted fusion transcript detection using next generation sequencing (NGS). Briefly, total nucleic acid was isolated from cell pellets, blood, bone marrow aspirate or smears. The total nucleic acid was reverse transcribed with random hexamers, followed by second strand synthesis to create double-stranded complementary DNA (cDNA). The double-stranded cDNA was end-repaired, adenylated, and ligated with a half-functional adapter. Two hemi-nested PCR reactions using the ArcherDx Heme Fusion kit primers were performed to create a fully functional sequencing library that targets specific genes (exons, Supplementary Table 6). Illumina NextSeq 2 × 150 base paired-end sequencing results were aligned to the hg19 human genome reference using bwa-mem 69 . A laboratory-developed algorithm was used for fusion transcript detection and annotation. The integrity of the input nucleic acid and the technical performance of the assay were assessed with a qualitative reverse transcription qPCR assay and assessing the DNA/RNA content in the sequencing results. Although this assay may detect several potential fusion variants, only the most prevalent one is reported. The assay is validated for samples showing 6% or higher tumor cellularity.
CRISPR-Cas9-screen. Genome-scale CRISPR-Cas9 loss-of-function screening was performed on 391 cancer cell lines including 8 T-cell lymphoma cell lines using the Avana library 70,71 . Briefly, cancer cell lines were transduced with Cas9 using a lentiviral system 71 . Cell lines that met criteria, including acceptable Cas9 activity measuring ability to knock out transduced GFP, appropriate growth properties and other parameters, were then screened with the Avana library. The Avana library 72 designed at the Broad Institute contains >70,000 sgRNAs (single guide RNAs) and an average of 4 guides per gene with approximately 1000 guides that do not target any location in the reference genome as negative controls. A pool of guides was transduced into a population of cells. The cells were cultured for approximately 21 days in vitro, and at the end of the assay, barcodes for each guide were sequenced for each cell line in replicate. Reads per kilobase were calculated for each replicate and then the log 2 fold change compared to the initial plasmid pool was calculated for each guide. Samples with poor replicate reproducibility (<0.7 Pearson coefficient), total read counts less than 15 million, or those that failed to match the fingerprint of parent lines were removed from the analysis. Additionally, guides that have low representation in the initial plasmid pool and those with suspected off-target activity were removed from analysis. Dependency scores were generated for each gene in each cell line using the CERES algorithm 70 . Briefly, CERES estimates gene dependency levels from CRISPR-Cas9 essentiality screens while accounting for the copy-number-specific effect, as well as variable sgRNA activity. The gene dependency scores generated by CERES are absolute dependency scores with each cell line scaled to have an average dependency score of 0 for negative control sgRNAs and an average dependency score of −1 for a set of positive control core essential genes as defined by Hart et al.
To calculate the probability and false-discovery rate (FDR) that a gene dependency score represents a true dependency in a given cell line, we fit a twocomponent mixture model in each cell line. The two components were 1 the empirically determined distribution of true dependent scores, identified using the pan-essential gene scores in that cell line, and 2 the empirically determined distribution of true non-dependent scores, identified as genes that were not expressed in that line. We defined as pan-essential genes 1871 genes whose dependency scores falls in the bottom 24% of gene scores in at least 90% of the cell lines. The probability of dependency for each gene score is the probability that it was generated from the distribution of true dependent gene scores. To correct for noise in the tails of the distributions, all gene scores below −1.5 were assigned probability 1 of being dependencies and all gene scores above 0.25 were assigned probability 0. A Gaussian smoothing kernel with width 0.15 was applied to the final probability scores to further reduce noise.
In order to identify outlier dependencies in T-cell lymphoma lines compared to the other cancer cell lines screened, Z-scores were calculated for each gene across all cancer cell lines using the median dependency score for each gene and the median average deviation. Genes with Z-scores <−4 in at least 2 T-cell lymphoma lines with an FDR < 0.05 in at least 1 T-cell lymphoma line with expression of the gene >1 RPKM were considered candidate T-cell lymphoma dependencies.
The sequences used for sgRNAs were: PTPN2 #1:CCATGACTATCCTCATAGAG PTPN2 #2:CCGCGACTCACCAAGTACAG Control:GCACTACCAGAGCTAACTCA In vivo studies. Viably frozen patient-derived xenograft cells were thawed and washed in 1 × PBS before tail-vein injection at 0.7-1.7 × 10 6 cells per mouse. For subcutaneous models, tumor fragments were implanted in the right flank under isoflurane anesthesia. Tumor burden was monitored periodically based on engraftment kinetics of each model by flow cytometry of peripheral blood, analysis of target organ involvement from sentinel animals for non-circulating models, or tumor size by caliper measurement. Blood was processed with Red Blood Cell Lysis Buffer (Qiagen) before staining with antibodies against human CD45 (Pacific Blue V450-conjugated, BD Bioscience 560367) and human CD2 (APC-conjugated, BioLegend 300213) or human CD56 (APC-conjugated, BD Bioscience 555518) in 1 × PBS with 1 mM EDTA plus 1% fetal bovine serum. All antibodies were used at a dilution of 1:20. Flow cytometry data was analyzed with FlowJo. Upon engraftment, mice were randomized to vehicle or treatment arms. Ruxolitinib was given at 90 mg per kg in PBS + 0.1% Tween 20, orally by gavage twice per day; romidepsin at 1 mg per kg diluted in PBS, given intraperitoneally on days 1, 4, and 7; ALRN-6924 at 20 mg per kg diluted in 20 mM sodium phosphate, 240 mM trehalose and 0.03% (w per v) polysorbate 20, pH 7.5 given intravenously on days 1, 4, and 7. The vehicle for romidepsin and ALRN-6924 was PBS intraperitoneally and ALRN6924 vehicle intravenously on the same schedule as the drugs.
All drugs were dosed in 10 mL per kg volume based on body weight. Mice were humanely sacrificed 16 h after the final dose; blood and target organs were collected for complete blood count, immunohistochemistry, and flow cytometry to determine tumor burden using the same antibodies as above. All animal work was performed in Nod.Cg-Prkdc scid IL2rg tm1Wjl /SzJ (NSG) mice purchased from Jackson Laboratories and handled according to Dana-Farber Cancer Institute's Institutional Animal Care and Use Committee approved protocol #13-034.
Contribution of human participants. For PDX generation, de-identified patient samples were obtained with informed consent and xenografted under Dana-Farber/Harvard Cancer Center Institutional Review Board (IRB)-approved protocol #13-351. The patient reported in Fig. 3 participated in the phase 1/2a trial of ALRN-6924 (NCT02264613). Informed consent was obtained according to the study protocol and was approved by the University of Alabama-Birmingham Cancer Center Institutional Review Board.
Data availability. All primary data from RNAseq, exome sequencing, and CRISPR screening are available in Supplementary Data 7-10. Raw data files are available at https://portal.gdc.cancer.gov/legacy-archive/search/f and have been deposited in NCBI's Gene Expression Omnibus and are accessible through GEO Series accession number GSE114085 73 . Mutation and RNASeq data for PDXs are also available at