A KDM4A-PAF1-mediated epigenomic network is essential for acute myeloid leukemia cell self-renewal and survival

Epigenomic dysregulation is a common pathological feature in human hematological malignancies. H3K9me3 emerges as an important epigenomic marker in acute myeloid leukemia (AML). Its associated methyltransferases, such as SETDB1, suppress AML leukemogenesis, whilst H3K9me3 demethylases KDM4C is required for mixed-lineage leukemia rearranged AML. However, the specific role and molecular mechanism of action of another member of the KDM4 family, KDM4A has not previously been clearly defined. In this study, we delineated and functionally validated the epigenomic network regulated by KDM4A. We show that selective loss of KDM4A is sufficient to induce apoptosis in a broad spectrum of human AML cells. This detrimental phenotype results from a global accumulation of H3K9me3 and H3K27me3 at KDM4A targeted genomic loci thereby causing downregulation of a KDM4A-PAF1 controlled transcriptional program essential for leukemogenesis, distinct from that of KDM4C. From this regulatory network, we further extracted a KDM4A-9 gene signature enriched with leukemia stem cell activity; the KDM4A-9 score alone or in combination with the known LSC17 score, effectively stratifies high-risk AML patients. Together, these results establish the essential and unique role of KDM4A for AML self-renewal and survival, supporting further investigation of KDM4A and its targets as a potential therapeutic vulnerability in AML.


Introduction
Epigenetic regulators are frequently mutated in acute myeloid leukemia (AML), leading to epigenomic alterations 1 . Inhibitors that target epigenetic modulators to rectify epigenomic abnormalities represent valid therapeutic strategies. Further understanding of how epigenetic dysregulation in AML contributes to leukemogenesis may uncover tractable therapeutic targets and biomarkers for AML patient treatment and/or prognostic evaluation.
Primary AML blasts from patients with poor prognosis feature global H3K9me3 hypomethylation 7 positing an oncogenic role for H3K9me3 demethylases in AML. Cheung et al. identified an H3K9me3 demethylase, KDM4C as a cofactor of the PRMT1 complex in MLL rearranged (MLLr) and MOZ-TIF2 AML 8 . Simultaneous knockout (KO) of all three members of the Kdm4 family (kdm4a/b/c) in mice attenuates MLL-AF9 AML 9 , indicating roles for the Kdm4 family in murine myeloid leukemia. However, the therapeutic benefit of targeting the KDM4 family in human AML is not well understood.
Our previous lentiviral knockdown (KD) screen targeting epigenetic regulators in 12 human AML cell lines representing several AML subgroups found that KMD4A KD leads to significant suppression of leukemia cell proliferation 10 . KDM4A has different roles in normal tissue development compared to other members of the KDM4 family; it is amplified/overexpressed in various malignancies including AML (Fig. S1A) and correlates with poor outcome in ovarian cancer 11,12 . Herein we demonstrate that KDM4A KD induces AML apoptosis by a unique mechanism to KDM4C in myeloid leukemia. Apoptosis results from a global accumulation of H3K9me3 and H3K27me3 at KDM4A genomic loci thereby causing downregulation of a KDM4A-PAF1mediated oncogenic program, including a 9-gene signature enriched with leukemia stem cell (LSC) activity, which can stratify high-risk patients. These findings support an essential and unique role of KDM4A for AML cell self-renewal and survival.

Reagents, plasmids, and virus manufacture
Puromycin, IOX1 dev , and lentiviral constructs for KD experiments (Supplemental Table) were purchased from Sigma-Aldrich (St. Louis, MO, USA). IOX1 was from Tocris (4464). pLenti-HA-KDM4A wt and mut (H188A/ E190A) were a gift from Dr. Gary Spencer (CRUK Manchester Institute). Lentiviral and retroviral supernatants were prepared, and leukemic human and murine cells transduced with viral particles as previously described 10 . A list of antibodies used in flow cytometry and Western blot/immunoprecipitation and immunofluorescence staining is in the Supplemental Table. Culture of cell lines and primary cells Leukemia cell lines were from DMSZ (Braunschweig, Germany) and grown in RPMI (10% FBS and 2 mM L-Glutamine) at 37°C in 5% CO 2 . They were recently authenticated and tested for mycoplasma contamination. Murine and human primary AML and normal BM samples were cultured in serum-free media (SFM) (H5100, Stem Cell Technologies, UK) supplemented with appropriate cytokines as described 13 . Murine MLL-AF9 AML cells were leukemic BM cells extracted from a mice cohort with MLL-AF9 AML established by Somervaille et al. 14 , and cultured in the conditional medium with mIL-3 (100 ng/ml). Cytokines purchased from PeproTech (London, UK).

Cell proliferation, apoptosis, and cell cycle analysis
Cell viability was measured by cell count using trypan blue dye (Sigma, T8154) /hemocytometer or resazurin (Alamarblue dye, Sigma) with the Envision Fluorescent Reader (Perkin Elmer). Apoptosis was assessed using Annexin-V/Dead Cell Apoptosis Kit (V13241, Thermo-Fisher) and cell cycle analysis with PI/RNase Staining solution (F10797, ThermoFisher) as per manufactures' instructions. Data were acquired using an LSRII flow cytometer and an Aria III flow cytometer (BD Biosciences, UK) and analyzed using FlowJo software (Tree Star Inc., USA).
Immunofluorescence (IF) staining 6 × 10 4 cells per condition were incubated on poly-Llysine coated Hendley-Essex 12 well glass microscope slides for 1 h before being fixed in 4% formaldehyde in PBS. The cells were permeabilized in 0.5% Triton-X-100 PBS followed by 2 h of blocking in 5% BSA, 0.2% Triton-X-100 TBS. Primary antibody was applied overnight in a humidified chamber at 4°C. Appropriate secondary antibody (1:500 dilution) was applied for further 1 h incubation at room temperature after removal of a primary antibody using PBS 0.1% Tween 20. Antifade mountant with DAPI reagent (Thermo Fisher #P36962) was used to seal each sample and images were captured on the Zeiss Axioimager M1 Epifluorescence and Brightfield Microscope. CellProfiler v2.2.0 image analysis software (Cell-Profiler) was used to quantify IF signals.

Murine transplantation experiments
Mice experiments were approved by the local animal ethics board and performed under a project license issued by the United Kingdom Home Office, in keeping with the Home Office Animal Scientific Procedures Act, 1986. Non-obese diabetic. Cg-Prkdc scid Il2rgtm1Wjl/SzJ (NSG) mice were purchased from Jackson Laboratories (Bar Harbor, ME, USA) for transplantation as previously described 10 . Primary AML patient samples for xenograft transplantation were unfractioned primary blasts from our and Manchester biobank collections. Control or KDM4A KD human AML THP1 cells or primary AML patient blasts were FACS sorted 40 h following lentiviral infection and immediately transplanted into sub-lethally irradiated (450 cGy) NSG mice of 6-8-week-old, mixed-sex (10,000 THP1 cells or 10^6 primary AML cells) via tail vein injection.
RNA isolation, quantitative PCR, RNA-seq, and ChIP-seq RNA was extracted using QIAshredder™ columns and RNeasy Plus Microkit™ (Qiagen). RNA-seq libraries were produced using the TruSeq® stranded mRNA kit (Illumina) and sequenced on the Illumina NextSeq™ 500 platform. For ChIP-seq, DNA was purified using Diagenode's iPure kitv2 and libraries made using the TruSeq ChIP Library Preparation Kit according to the manufacturer's instructions. For QPCR, reverse transcription was carried out using Invitrogen SuperScript III First-Strand Synthesis kit. A SYBR ® green-based fluorescent system was used to quantify dsDNA using the Applied Biosystems 7900 HT Fast Real-Time PCR system. Each qPCR plate included technical triplicates of each specific target alongside two housekeeping genes (GAPDH and ß-Actin). Delta-Delta CT method was used for analysis of gene expression against control. RNA-seq and ChIP-seq reads were mapped to the hg19 human genome or mm9. Transcript abundances were calculated in transcripts per million (TPM) using Kallisto 15 . SICER was used for peak calling on default settings. Both data files are available in the Gene Expression Omnibus (GEO): GSE125376.

Gene signature construction
The KDM4A-9 signature score is calculated as shown below using the least absolute shrinkage and selection operator (LASSO) linear regression 16,17 .
x i r i Where x i and r i denote the gene expression and coefficient of the ith gene (out of the total 9 genes) in the signature, respectively. Specifically, KDM4A-9 score = (TPM2 × 0.097292) Correlating the KDM4A-9 signature and LSC activity LSC enrichment classification and raw expression data from AML samples (GSE76008) 17 were analyzed using the lumi 2.36.0 package. Intensity values were normalized by Robust spline normalization. The diagnostic capability of each gene signature to predict LSC activity across AML samples (GSE76008) was assessed by ROC (Receiver operating characteristic) curve analysis. The Youden index was used to identify the optimal cut-off value.
Construction of KDM4A-9 and LSC17 combined score Min-Max scaling of the KDM4A-9 and LSC17 scores was performed prior to the linear summation of KDM4A-9 and LSC17 scores for each patient to generate a combined KDM4A-9/LSC17 score.

Network construction and visualization
To visualize the connections between the LSC17 and KDM4A-9 signature genes, as well as KDM4A and PAF1 in AML patients, a filtered edge list (TOM ≥ 0.05) 18 was constructed. An undirected network graph was generated using the graph.edgelist function from the igraph 1.2.5 package.

Statistical analysis
Normally distributed groups were compared using a two-tailed student t-test unless stated otherwise. Survival probabilities estimated by Kaplan-Meier method. For RNA-seq counts, a pseudo-count of 1 was added prior to log 2 -transformation. Statistical significance of differential gene expression was assessed by Welch's t-test unless otherwise stated. For RNA-seq, differential expression analysis was performed using the DESeq2 1.26.0 R package. Statistics were calculated using R-3.6.1.

KDM4A is required for the survival of human and murine AML cells
KDM4A expression is unique (Fig. S1B) being highly enriched in AML-LSC + populations (Fig. S1C), suggesting that KDM4A is important for LSC, which are negatively correlated with AML patient survival. We performed lentivirus shRNA KD of KDM4A, KDM4B, and KDM4C in human AML MLL-AF9-driven THP1 cells to confirm its essential role. KDM4A KD THP1 cells exhibited the greatest decrease in cell proliferation compared to nontargeting cells (NTC) (Fig. 1A-C). Consistent with previous work, lentiviral KD of KDM4C had an inhibitory effect on cell proliferation 8 (Fig. 1A). CFC potential was positively correlated with the KDM4A in a dosedependent manner when five KDM4A KD shRNA targeting constructs were compared (Fig. 1D). KDM4A KD induced apoptosis ( Fig. 1E and S1D) rather than cell cycle arrest (Fig. S1E). These results were further confirmed in primary patient blasts ( Fig. 1F and G) and murine AML cells (Fig. 1H). Importantly, we determined the impact of KDM4A KD on AML initiation in vivo by transplanting KDM4A KD THP1 cells (Fig. 1I) or primary AML cells (Fig. 1J-K and S1F-S1H) into recipient NSG mice. Control cells induced short-latency disease with splenomegaly (Fig. 1K). Loss of KDM4A significantly prolonged overall survival (OS) of mice with only one mouse succumbing to leukemia over the follow-up period by either KDM4A#1 KD or KDM4A#2 KD (Fig. 1J-K and S1F-S1H). Together, these data demonstrate a specific and essential role for KDM4A in AML cell survival.

Targeting KDM4A's demethylase activity inhibits AML cell proliferation
Functional rescue experiments determined that the demethylase activity of KDM4A is required for AML. Forced-expression of wild-type human KDM4A rescued the clonogenic activity of AML cells transduced with kdm4a KD virus ( Fig. 2A and B). This rescue phenotype was not observed by a catalytically inactive mutant of KDM4A (KDM4A H188A/E190A ) 19,20 ( Fig. 2A and B) in murine MLL-AF9 cells. Next, we assessed KDM4A substrates H3K9me3 and H3K36me3 in KDM4A KD THP1 cells. There was an increase in H3K9me3 shown by immunoblotting ( Fig. 2C  We next investigated whether KDM4A is dispensable for normal hematopoiesis. Our data showed no significant loss of colonies in kdm4a KD normal murine BM c-kit + cells in CFC assays ( Fig. 2E and F; S2E). Similarly, reduced levels of KDM4A in human CD34 + HSPCs from healthy donors are tolerated ( Fig. 2G and H; S2F) with fewer colonies due to a reduction of CFU-GM in KDM4A KD #1 cells. Although a KDM4A specific inhibitor is lacking, there are several pan-KDM4 inhibitors available including IOX1 21 and IOX derivatives (IOX1 dev ) 22,23 . These displayed significant inhibitions of cell proliferation in THP1 cells and primary AML patient blasts, inducing differentiation and apoptosis (Figs. S3A-S3G) with minimum effect on normal human CD34 + BM HSPCs (Fig. S3E). These phenotypes were accompanied by an increased level of H3K9me3 and H3K27me3 (Fig. S3H), suggesting the anti-leukemic effect was due to KDM4A inhibition.

PAF1 identified as a KDM4A co-regulator is required for human AML cell survival
To determine the impact of KDM4A on global gene expression, we compared the transcriptome of KDM4A KD THP1 cells with NTC control by RNA-seq. 3375 differentially expressed (DE) genes were significantly deregulated following depletion of KDM4A (Log 2 fold change (FC) ≥0.5 or ≤−0.5; adj. p ≤ 0.05; Fig. 3A; Supplemental file). 67% (2274 out of 3375) were direct targets of KDM4A; ChIP-seq revealed that KDM4A bound at their TSS (Supplemental file). 61% (1387 out of 2274) of putative KDM4A direct target genes were downregulated associated with the enrichment of transcriptional repressive marks at H3K9me3 24 and H3K27me3 25 (Fig. 3A).   To provide insights into the survival pathways regulated by KDM4A, we performed gene-set enrichment analysis (GSEA) and revealed significant enrichment of genes regulated by PAFc [26][27][28] (Fig. 3B). This is consistent with downregulation of PAF1, following KDM4A KD at the transcript ( Fig. 3C and D) and protein (Fig. S4A) level in human MLLr-AML cell lines, and primary patient blasts. PAF1 KD phenocopied KDM4A KD in MLLr-AML cells, inducing significant apoptosis (Fig. 3E-G; S4B) and loss of CFU potential (Fig. 3H). Together, these data suggest loss of KDM4A impairs PAF1 function to maintain leukemic cell survival, supporting PAF1 as an important cofactor of KDM4A in human AML.

KDM4A-PAF1 maintains appropriate expression of the MLLr-fusion oncogenic program in MLLr-AML
Our ChIP-seq data reveal substantial overlap amongst PAFc 26 , MLL-AF9 29 , and KDM4A binding sites ( Fig. 4A-C). Specifically, KDM4A bound the PAF1 promoter region (Supplemental file), suggesting a direct regulatory mechanism. There is no enrichment of either histone methylation mark at non-KDM4A binding genomic loci ( Fig. 4D and E), indicative of a human KDM4A-specific epigenomic profile. In marked contrast, there is a global gain of both H3K9me3 and H3K27me3 upon KDM4A KD in THP1 cells at KDM4A binding sites ( Fig. 4F and G; 5A), including the genomic loci of PAF1 and its targets (Fig. 5B). These were validated by ChIP-QPCR in cell lines and primary patient blasts ( Fig. 5C and D).
Furthermore, genes with significant expression changes following KDM4A silencing were also enriched in direct PAF1 target genes 26 (Fig. 5E), suggesting a transcriptional network co-regulated by both KDM4A and PAF1. This notion is supported by the fact that KDM4A bound promoters share almost identical enrichment of transcription factor (TF) binding motifs as the ones bound by PAF1, including homeobox (HOX) TFs, such as TLX2 and DBX (Fig. 5F). Further GSEA analysis on the overlapped DE genes between KDM4A KD and PAF1 KD revealed a significant downregulation of MLLr fusion target genes 30 , as well as HOX family target genes 31 including notably the pro-survival gene, BCL2, and marked upregulation of a mature hematopoiesis program 32 (Fig. 5G) and pro-apoptotic gene, BCL2L11 (BIM). Although the expression of HOXA9 itself was not affected by either KD, our data suggest KDM4A and PAF1 coregulate their downstream targets in a parallel manner.
A core 9-gene signature downstream of KDM4A strongly associated with LSC activity and clinical outcome Supporting the collaborative role of KDM4A-PAF1 in AML, KDM4A expression is highly associated with PAF1 expression in patient datasets ( Fig. 6A and B); KDM4A-PAF1 expression can identify patients with inferior OS (Fig. 6C). KDM4A KD induced a significant reduction of cell proliferation in human AML cell lines representative of different subtypes (Fig. S5A), coupled with an increase in apoptosis and loss of CFC potential (Figs. S5B and S5C). This evidence suggests that KDM4A is required across AML.
These data led to our hypothesis that a core gene expression signature (GES) downstream of the KDM4A-PAF1 regulatory axis, is associated with AML patient outcomes compared with the known LSC score, LSC17 17 . For this, we utilized LASSO regression analysis 16,17,33 on KDM4A regulated genes (KDM4A KD Log 2 FC ≥1 or ≤−1; adj. p ≤ 0.05; Supplemental file) and defined a KDM4A-9 score (Fig. 6D). High KDM4A-9 was highly associated with poor OS in AML cohorts (Fig. 6E-H) independent of age, cytogenetic risk score, and frequent mutation status. The robust prognostic value of KDM4A-9 indicates that the score may be related to the important biological activities of AML-LSCs. Indeed, KDM4A-9 correlates with the LSC17 score of AML samples and over 75% of KDM4A-9 high score fractions are LSC+ (Fig. 7A), therefore KDM4A-9 is a strong predictive indicator of  AML-LSC activity (Fig. 7B). Interestingly, there is no overlap between these two gene signatures. A combined signature score (KDM4A-9/LSC17) achieves an optimal balance between specificity and sensitivity (Fig. 7C) overcoming the limitations of either score alone with an improved ability to predict survival over LSC17 (Fig. 7D and E).
KDM4A binds at the promoter regions of KDM4A-9 genes, whereupon H3K9me3 and H3K27me3 are enriched after KDM4A KD ( Fig. 8A; Supplemental file), suggesting direct regulation. In addition to the QPCR validation in human AML cell lines and primary AML blasts following KDM4A KD or PAF1 KD (Fig. 8B), we observed that the majority of KDM4A-9 genes show correlation with KDM4A and PAF1 expression in patient AML cohorts ( Fig. 8C and D). Furthermore, weighted gene correlation network analysis (WCGNA) 34 demonstrated a strong relationship between KDM4A, PAF1, and the KDM4A-9 and LSC17 GESs across AML. This network possessed high topological overlap (topological overlap matrix (TOM) ≥0.05)) with KDM4A as a highly connected node (Fig. 8E) suggesting that KDM4A-PAF1 regulates the KDM4A-9/LSC17 network in AML.
KDM4A has a distinct function to another KDM4 family member, KDM4C in AML Previously, Cheung et al. showed that KDM4C is required for MLLr-AML cell survival 8 , indicating an overlapping role of KDM4A and KDM4C in AML. However, forced-expression of wild-type human KDM4C failed to rescue the clonogenic activity of murine MLL-AF9 AML cells transduced with kdm4a KD virus (Figs. S6A and S6B), suggesting KDM4A has a distinct role from that of KDM4C. This is in line with previously reported data showing no increase of global H3K27me3 level upon pharmacological inhibition of KDM4C in MLLr-AML cells 8 . Consistently, at the molecular level, KDM4A KD led to transcriptional changes distinct from KDM4C KD via GSEA comparison (Fig. S6C), further supporting a unique role for KDM4A compared to KDM4C in human AML. In particular, KDM4A KD has no significant impact on gene expression of two established targets of KDM4C, HOXA9, and MEIS1 in human MLLr-AML cells. These results are also validated by Q-PCR using shRNAs targeting HOXA9 as control (Fig. S6D). More importantly, kdm4c KD had no impact on PAF1 expression, nor its associated genes targeted by KDM4A including KDM4A-9 and LSC17 GESs (Fig. S6E). Together, these data demonstrate an essential role of KDM4A in human AML.

Discussion
Previous reports indicate that the KDM4 family is required for normal hematopoiesis 9,35 , whilst loss of individual members is tolerated in normal cells 35 highlighting the importance of identifying KDM4 family members that are essential for the survival of AML cells. Our data demonstrate KDM4A is unique; it is selectively required for AML cell survival, with no negative effect on normal hematopoiesis offering a therapeutic window. Lack of tractable enzymatic activities limits the potential of PAF1 or other subunits of the PAFc as therapeutic targets in cancer. Herein, we identify a novel KDM4A-PAF1 signaling axis co-regulating oncogenic transcriptional networks in human AML, providing a way to eliminate leukemic cells with broad therapeutic applications. Our study provides a strong rationale for the further development of KDM4A inhibitors, presenting a promising strategy for novel epigenetic-based therapy in AML.
KDM4A-9 shows strong therapeutic implications comparable with LSC17 17 . A high KDM4A-9 score may reflect an important biological property of KDM4A in leukemogenesis. The function of the KDM4A-9 genes in leukemogenesis is unknown; except Tetraspanin (CD82) 36,37 which plays an important role in AML. Corroborating recent findings 5,6 , our data suggest KDM4A regulates H3K9me3 to direct leukemogenesis. H3K9me3 has emerged as a key player in repressing lineage-   H) showing that the KDM4A-9 score can predict survival across AML patients of varying subtypes. Patients were dichotomized into high and low groups based on whether they possessed a score above or below the median signature score; p by log-rank test. Fig. 7 KDM4A-9 enriched with LSC activity, is a poor prognosis marker for AML. A Scatterplot showing a moderate correlation between the KDM4A-9 score and LSC17 score in primary AML patient samples (GSE76008). LSC enriched (LSC+, n = 138) cell fractions from 78 patient samples are colored blue whilst those that lack LSC enrichment (LSC−, n = 89) are colored red. Over 75% of KDM4A-9 high score (above median value) fractions are LSC+. Pearson correlation used to assess correlation. Significance determined by t-test. B Box plot showing KDM4A-9 or LSC17 signature scores in two comparative groups: LSC+ and LSC− from (A); unpaired t-test, *p < 0.0001. C ROC curves of KDM4A-9 (blue), LSC17 (yellow), and KDM4A-9/LSC17 (green) show the diagnostic capability of each signature to predict LSC enrichment in AML samples. The black bars in each plot are the 95% confidence intervals for the optimal cut-off. The Youden index was used to determine the optimal cut-off for each signature. D-E Patients in the Beat AML/Vizome dataset were dichotomized into high and low groups based on whether they possessed a score above or below the median signature score. Kaplan-Meier survival analysis conducted showing that the combined KDM4A-9/LSC17 score (E) is effective in the prediction of AML patient survival over the LSC17 score alone (D). inappropriate genes, impeding the reprogramming of cell identity during development and cell fate determination 24,38 . It would be interesting to determine further the clinical diagnostic relevance of H3K9me3 in relation to KDM4A in AML patients.