Loss of Kat2A Enhances Transcriptional Noise and Depletes Acute Myeloid Leukemia Stem-Like Cells

Acute Myeloid Leukemia (AML) is an aggressive hematological malignancy with abnormal progenitor self-renewal and defective myelo-monocytic differentiation. Its pathogenesis comprises subversion of transcriptional regulation, through mutation and by hijacking normal chromatin regulation. Kat2a is a histone acetyltransferase central to promoter activity that we recently associated with stability of pluripotency networks, and identified as a genetic vulnerability in AML. Through combined chromatin profiling and single-cell transcriptomics, we demonstrate that Kat2a contributes to leukemia propagation through homogeneity of transcriptional programs and preservation of leukemia stem-like cells. Kat2a loss reduces transcriptional bursting frequency in a subset of gene promoters, generating enhanced variability of transcript levels but minimal effects on mean gene expression. Destabilization of target programs shifts cellular equilibrium out of self-renewal towards differentiation. We propose that control of transcriptional variability is central to leukemia stem-like cell propagation, and establish a paradigm exploitable in different tumors and at distinct stages of cancer evolution.


INTRODUCTION
Acute Myeloid Leukemia (AML) is the most prevalent leukemia in adults with a dismal prognosis of less than 30% 5-year survival (Dohner et al., 2017). It is a heterogeneous disease, clinically and pathologically, with common cellular themes of myeloid differentiation block, and recurrent molecular targeting of chromatin and transcriptional regulation. Effects on transcription are reflected in the AML mutational spectrum (Cancer Genome Atlas Research et al., 2013), as well as through the implication of general transcriptional co-regulators in AML pathogenesis, in the absence of specific mutation events (Roe and Vakoc, 2013). Examples of these are specific AML dependencies on BRD4 (Dawson et al., 2011;Zuber et al., 2011), LSD1 (Harris et al., 2012) or DOT1L (Bernt et al., 2011;Daigle et al., 2011). Moreover, chemical inhibitors exist to target these regulators and have progressed to clinical trials (Gallipoli et al., 2015). More recently, TFIID and SAGA subunit TAF12 was shown to be critical for MYB protein stability and transcriptional activity in AML cells through its participation in the TFIID complex (Xu et al., 2018).
In a recent CRISPR drop-out screen of genetic dependencies in AML, we have identified several members of the SAGA complex, including histone acetyl-transferase KAT2A, as being required for AML maintenance (Tzelepis et al., 2016). KAT2A was suggested to impact cell survival and differentiation status, but its precise molecular mechanisms of action remain to be elucidated, and it is unclear whether it is required in AML initiation, as well as maintenance. Kat2a is a mammalian orthologue of yeast histone acetyl-transferase Gcn5, and is required for H3K9 acetylation (H3K9ac) (Jin et al., 2014), a modification that fine-tunes, rather than initiates, locus-specific transcriptional activity. Kat2a is required for specification of mesodermal derivatives during early embryonic development (Lin et al., 2007) (Wang et bursting nature of gene expression (Chubb and Liverpool, 2010): for most if not all loci, transcriptional activity is not continuous, but burst-like or episodic, with locus-specific rates of locus activation (κON), inactivation (κOFF), and RNA production (κRNA), as well as RNA degradation (Raj et al., 2006) contributing to the net effect. Frequency of bursting depends on the κ ON rate, whilst κ RNA impacts the burst size (Cai et al., 2006). Both parameters contribute to mean gene expression, whilst transcriptional noise is more strictly dependent and anti-correlated with burst frequency . In yeast, size and frequency of bursts are influenced by histone acetylation in gene bodies and promoters, respectively (Weinberger et al., 2012).
In functional terms, transcriptional noise has been directly implicated as a mechanism of cell fate choice in yeast (Blake et al., 2006) and bacteria (Suel et al., 2006), and recurrently associated, albeit correlatively, with cell fate transitions in mammalian systems (Moris et al., 2016). We had previously shown that normal transitions into hematopoietic lineage specification associate with cell-to-cell heterogeneity in gene expression (Pina et al., 2012;Teles et al., 2013). More recently, we have inhibited the activity of Kat2a in mouse embryonic stem cells, and observed an increase in transcriptional heterogeneity that impacted the stability of pluripotency with reconfiguration of correlation gene regulatory networks (GRNs) (Moris et al., 2018). Whilst we have not mechanistically linked enhanced heterogeneity with the loss of pluripotency, we observed propagation of variability of transcriptional levels through the GRNs downstream of nodes with differential H3K9ac.
Cancer, and in particular leukemia, can be perceived as an imbalance between self-renewal and differentiation in favor of self-renewal. We postulated that enhancing transcriptional variability in AML cells would enhance the probability of cell fate transitions out of selfrenewal into differentiation, with loss of leukemia stem-like cells (LSC

Conditional loss of Kat2a does not affect normal hematopoiesis and allows MLL-AF9driven transformation
We sought to investigate Kat2a requirements in vivo during early leukemia initiation and  1A). We obtained locus excision by treatment of experimental and control mice with a course of intra-peritoneal polyinosylic-polycytidylic (pIpC) acid injections, as described (Chan et al., 2011). Excision was tested 4-6 weeks after injection and consistently achieved values greater than 80% in stem and progenitor cell compartments (Fig. 1B), reflected in a profound loss of gene expression, including amongst myeloid-biased (LMPP) and committed (GMP) progenitors critical for AML initiation (Goardon et al., 2011) (Fig. 1C). Of note, locus excision generates an in-frame product that joins the first 2 and the last exons ( Supplementary   Fig. 1A); this product is transcribed ( Supplementary Fig. 1B), but should not code for catalytic or acetyl-binding activity ( Supplementary Fig. 1A). In agreement with a previous report (Bararia et al., 2016), Kat2a was dispensable for HSC maintenance and function, as assessed by BM composition acutely after excision and throughout aging (Supplementary Transformation of progenitor-enriched, lineage-depleted (Lin-) cells with an MLL-AF9 fusion transcript was initially assessed in vitro through serial re-plating of WT and KO cells in in semi-solid medium-based colony-forming assays. Transformation was observed for cells of both genotypes, with similar efficiency (Fig. 1D). Locus excision (Fig. 1E) and gene expression loss (Fig. 1F) were maintained or even increased during transformation, suggesting that loss of Kat2a does not impede the initial selection of a leukemia-transformed clone.

Kat2a depletion impairs establishment of MLL-AF9-initiated leukemia
We investigated the longer-term effects of Kat2a loss in transformation progression in vitro by continued serial re-plating. Whilst no differences were seen in re-plating ability ( Fig. 2A), it was noted that the colonies obtained had a clear component of differentiated cells ('dubbed' mixed or type II colonies (Johnson et al., 2003)) ( Fig. 2B), which could also be observed in colonies initiated from primary BM transformed colonies (Fig. 2C). Accordingly, Kat2a KO colonies showed increased levels of the differentiation marker CD11b (Supplementary Fig.   2A). Interestingly, establishment of clonal liquid cultures from in vitro transformed cells revealed a relative advantage in culture initiation from WT cells (Fig. 2D), suggesting that an imbalance between self-renewal and differentiation in the KO setting.
We probed the effect of Kat2a in MLL-AF9-driven transformation in vivo by injecting lethally-irradiated recipients with WT and KO Lin-BM cells transduced for 2 days with retrovirus encoding the MLL-AF9 oncogenic fusion. Animals developed leukemia 3 months after transplantation, as previously described, with a modest survival advantage for recipients of KO cells (Fig. 2E). At the point of culling, no differences in leukemia burden inspected the pattern of distribution of H3K9ac in promoter and enhancer elements in MLL-AF9 primary leukemia initiated by Kat2a KO or WT cells. Although global H3K9ac was minimally changed between genotypes, there was a specific depletion of H3K9ac peaks at promoters in regions devoid of concomitant H3K27ac activation mark (Fig. 3A). Conversely, H3K9ac was mildly increased at candidate active enhancer regions marked by the presence of H3K27ac (Fig. 3B), suggesting a possible pattern of imbalance of H3K9ac regulation between promoters and enhancers.
We focused on those promoter peaks with unique loss of H3K9ac upon Kat2a depletion, and used the ENCODE database (Auerbach et al., 2013) to confirm enriched experimental binding of KAT2A (aka GCN5) in other model systems ( Fig. 3C and Supplementary File 1).
Similar to a previous study of the effects of Kat2a and H3K9ac loss in embryoid bodies (Wang et al., 2018), we also found evidence for increased representation of MYC targets, which is a known Kat2a interacting protein (Hirsch et al., 2015). Genes associated with differentially-acetylated promoter peaks fell into 3 main categories (

Differential H3K9ac subsequent to Kat2a loss results in transcriptional variability
The role of yeast Gcn5 as a regulator of locus-specific intrinsic transcriptional noise  (Field et al., 2008;Tirosh and Barkai, 2008).

Kat2a regulates transcriptional bursting activity in cells with stem-like characteristics
Having established that loss of Kat2a associates with increased cell-to-cell variability in expression levels of a subset of directly-targeted genes, we asked whether the variability reflected differential regulation of locus transcriptional bursting, and hence modulation of transcriptional noise. We made use of the D3E code developed by the Hemberg lab (Delmans Kat2a targets using the cells in cluster 7 revealed significantly lower frequency of bursting and associated high CV in KO cells ( Fig. 5C and Supplementary Fig. 5C). Again, we observed a mild gain in burst size (Fig. 5C), which associates with unchanged mean expression levels ( Supplementary Fig. 5C). In contrast, modelling of cells in cluster 6, with the lowest STEM-ID score, revealed no differences in transcriptional parameters between WT and KO cells (Fig. 5D). Of note, Kat2a targets had lower average gene expression in cluster 6 ( Supplementary Fig. 5D). Overall, the data suggest that Kat2a target genes associate with candidate stem-like clusters and that Kat2a regulates their expression through buffering of transcriptional variability.

Kat2a regulates the activity of translation-associated genes
Having established that Kat2a loss results in deregulation of transcriptional activity with decrease of bursting frequency, we asked if this effect was biased towards particular classes of genes. Indeed, we found that translation-associated genes, including ribosomal protein genes and translation initiation factors, were significantly overrepresented (

Kat2a loss depletes functional MLL-AF9 leukemia stem-like cells
Finally, we asked whether the enhanced transcriptional variability observed in STEM-ID high Interestingly, we note a dissociation between surface phenotype and stem-like function, suggesting that identification of the classical L-GMP surface antigen phenotype (Krivtsov et al., 2006) may not absolutely associate with function. Also, importantly, we, like others, did not observe that loss of Kat2a introduced changes to normal hematopoiesis (Bararia et al., 2016), particularly in HSC, LMPP or GMP compartments that directly contribute to MLL-AF9 transformation. Repeated 5-FU treatment or secondary transplantation (data not shown) also failed to reveal a stem cell function defect, indicating specific dependency of leukemia stem-like cells on expression of Kat2a. 1 We used single-cell transcriptional analysis to capture cell-to-cell heterogeneities within seemingly phenotypic equivalent primary AML from both genotypes. Whilst differences between genotypes were minimal in terms of average gene expression, we identified a clear distinction in cell-to-cell variability in transcript levels that was specific to a subset of promoters characterized by H3K9, but not H3K27, acetylation, and which were dependent on bursting activity, which correlates with transcriptional noise. The number of mRNA molecules produced by each burst, or burst size, on the other hand, was not changed or was even mildly increased upon loss of Kat2a, suggesting a mechanistic link between H3K9ac and bursting frequency. This recapitulates findings in yeast linking H3K9ac at gene promoters with noise, but not level of gene expression (Weinberger et al., 2012), and provides mechanistic insight into the cell-to-cell heterogeneity elicited by Kat2a loss. In a recent study, the Naëf lab has shown that locus-specific manipulation of promoter, but not distal or enhancer, H3K27 acetylation can change transcriptional bursting frequency (Nicolas et al., 2018). Whilst the association with H3K27ac is unclear in our study, there is a clear contribution of H3K9 acetylation to bursting frequency, which matches an association of promoter H3K9ac, in addition to H3K27ac, and frequency of locus activation in the Naëf study (Nicolas et al., 2018). The mild gain in burst size, although unproductive in terms of transcriptional level, could reflect the differential reconfiguration of H3K9ac at promoters and enhancers upon Kat2a loss, and will be interesting to follow-up in subsequent studies.
Indeed, our lab has recently developed a KAT2A-Cas9 fusion capable of catalyzing targeted acetylation events (data not shown), that will be instrumental in answering these questions.
In linking the promoter-specific effects of Kat2a on H3K9ac and frequency of transcriptional bursting to the observed depletion of leukemia stem-like cells we found that general metabolic categories, in particular related to RNA processing, rather than known leukemiaassociated programs, were affected in their chromatin signature and frequency of bursting.
While we cannot exclude that our focus on loss, rather than reduction, of H3K9ac, combined with current limitations of single-cell RNA sequencing in capturing low-expressed genes, may have missed individual candidates, the agreement between the two levels of analysis, and indeed the ontology overlap with published studies of Kat2a-depleted or inhibited ES cells (Hirsch et al., 2015;Wang et al., 2018), including ours (Moris et al., 2018), suggest that Kat2a may regulate pervasive, rather than cell specific programs. The identification of a candidate nucleosome displacement motif in Kat2a target promoters also indicates specific regulation of highly and widely expressed genes. Amongst these, we found that translation as a category was targeted by Kat2a depletion, and demonstrated that not only is the assembly of polysomes perturbed by Kat2a inhibition, but that perturbation of the translational machinery can re-capture defects in in vitro propagation of leukemia-initiating cells akin to those imposed by Kat2a depletion. In agreement, Morrison and collaborators (Signer et al., 2014) have reported that impaired protein synthesis upon genetic depletion of the ribosomal protein machinery impedes leukemia self-renewal, whilst having non-linear dose-dependent effects on normal hematopoiesis, mimicking our own observations in the Kat2a KO setting.
Future studies directing Kat2a histone acetylation activity to single or multiple loci will illuminate individual vs. global target gene contributions to the leukemia phenotype.
However, it is tempting to speculate that the generic nature of the programs impacted by

Isolation of mouse BM stem and progenitor cells
BM was isolated from mouse long bones as described before (Pina et al., 2015). Following red blood cell lysis, total BM suspension was depleted of differentiated cells using a cocktail of biotinylated lineage antibodies (Table B) and streptavidin-labeled magnetic nanobeads (Biolegend), according to manufacturers' instructions. Cells were directly used in transplants, colony-forming assays or flow cytometry for analysis of normal hematopoiesis. For leukemia studies, cells were cultured overnight at 37°C 5% CO2 in RPMI supplemented with 20% Hi-FBS (R20), 2mg/mL L-Glutamine, 1% PSA, 10 ng/mL of murine Interleukin 3 (mIL3), 10 ng/mL of murine Interleukin 6 (mIL6), and 20 ng/mL of murine Stem Cell Factor (mSCF) (cytokines from Peprotech) (supplemented R20), followed by retroviral transduction.

Colony forming cell (CFC) assays
For analysis of normal progenitors, sorted mouse BM cells were plated at a density of 200-

cells/plate in duplicates, in MethoCult GF M3434 (STEMCELL Technologies). Colonies
were scored at 7-9 days. For analysis of MLL-AF9 leukemia, retroviral-transduced BM cells were plated in M3434 at an initial density of 10000 cells/condition and scored and re-plated every 6-7 days. Re-plating was performed up to passage 9, with 4000 cells/condition used from plate 3. CFC assays from mouse MLL-AF9 transformed lines were seeded in M3434 and scored 6-7 days later. RPS6K inhibition studies were set by adding 3.3uL DMSO, either as vehicle or with a final concentration of 3.5uM of PF4708671 (Tocris), directly to the methylcellulose medium, with mixing prior to cell addition.

In vivo analysis of leukemia initiation and engraftment
For analysis of normal hematopoiesis, 10 6 Kat2a WT or Kat2a KO cKit+ cells were intravenously injected via tail vein into lethally irradiated (2*5.5Gy) CD45.1 recipient mice.

Retroviral transduction
Retroviral construct MSCV-MLL-AF9-IRES-YFP was previously described (Fong et al., 2015). For viral particle production, Human Embryonic Kidney (HEK) 293T cells were seeded at 2.5x10 6 cells/10cm dish in DMEM supplemented with 10% Hi-FBS, 2mg/mL L-Glutamine, 1% PSA and cultured overnight at 37°C 5% CO2. The following day, a transfection mix [per plate: 47.5 uL of TranSIT (Miros), 5ug of packaging plasmid psi Eco vector (5ug), retroviral vector (5ug) and 600uL of Optimem Medium (Gibco)] was prepared according to manufacturer's instructions and added dropwise to cells followed by plate swirling and overnight culture at 37°C 5% CO2. Medium was replaced with R20 the next day. At 24 and 48 hours after R20 replacement, medium was collected and filtered through a 1 8 from 6-10 week-old Kat2a WT and Kat2a KO mice were collected and Lineage-depleted as described above (Isolation of mouse BM stem and progenitor cells), and cultured overnight at 37°C 5% CO2 in supplemented R20. For viral transduction, BM cells were briefly centrifuged at 400G, 5 minutes, and viral particle suspension medium supplemented with 10 ng/mL mIL3, 10 ng/mL mIL6, and 20 ng/mL mSCF added to a final density of 10 6 cells/mL. Cells were plated in 6-multiwell plates and centrifuged for 1 hour at 2000rpm, 32°C. After, cells were incubated for 4 hours at 37°C 5% CO2. A second round of viral transduction was performed, with post-centrifugation incubation performed overnight. Next day, cells were collected, pelleted and washed three times with PBS (2x) and R20 (1x). YFP level was accessed by Flow Cytometry in a Gallios Analyser (Beckman Coulter).

Establishment of MLL-AF9 transformed cell lines
MLL-AF9 clonal liquid cultures were set up using MLL-AF9 retrovirus-transduced primary

BM cells (see Retroviral Transduction section). Transformed cells enriched in vitro by 3
rounds of serial plating (CFC assays) were maintained in R20 supplemented on alternate days with mSCF, mIL3 and mIL6, all at 20ng/mL. Cells were cultured at 2*10 5 cells/ml and passaged when they reached a density of 1*10 6 /ml.

Flow Cytometry
Cell surface analysis of BM and Sp was performed using a panel of antibodies described on Table B according to sorting strategies detailed on Table C. Data was acquired using a Gallios Analyser (Beckman Coulter) and analysis performed in Kaluza software (Beckman Coulter). For sorting, an Influx or an AriaII BD sorter were used.

Quantitative Real time PCR (Q-RT-PCR)
Total RNA was extracted using Trizol Reagent (Invitrogen). RNA from equal numbers of cells was reverse-transcribed using Superscript II (Invitrogen), following manufactures' instructions. Complementary (c)DNA was analyzed in duplicate or triplicate by qPCR using Taqman gene expression assays (Table D) and Taqman Gene Expression Mastermix (Applied Biosystems). Gene expression levels were determined by the Pfaffl method following normalization to Reference gene, as stated. For exon 2-18 in-frame products, qPCR using Sybr Green Master Mix (Applied Biosystems) was performed in triplicates.
After centrifugation (Beckman SW40Ti rotor) at 260 900g for 3 hours at 4°C, gradients were fractionated at 4°C using a Gilson Minipulse 3 peristaltic pump with continuous monitoring (A254nm) and polysome profiles recorded using a Gilson N2 data recorder.

Chromatin Immunoprecipitation sequencing (ChIP-seq)
Total BM cells from duplicate pools of MLL-AF9 Kat2a WT and Kat2a KO primary leukemia samples were crosslinked with 1% Formaldehyde Solution (Sigma Aldrich) for 10 min at room temperature (RT), with gentle rotation (50rpm). Fixation was stopped with Glycine, and cells incubated for 5 min, RT, with gentle rotation (50rpm), followed by two washing steps in ice-cold PBS. Cell pellets were resuspended in Lysis buffer (Table E) followed by Nuclei preparation. Chromatin pellets were sheared in a Bioruptor Pico Plus (Diagenode) in TPX tubes, using 3 runs of 11 cycles (Cycle: 30sec ON 30sec OFF) on high setting. A short spin was performed between runs and samples were transferred to new TPX tubes. 1:10 of total sheared chromatin was kept for input reference. Immunoprecipitation was set up using Dilution Buffer, Protease cocktail Inhibitor, and the respective antibody (Table   F) and the sheared chromatin incubated overnight at 4°C with rotation. On the following day, protein A/G magnetics beads were pre-cleared with Dilution Buffer supplemented with 0.15% of SDS and 0.1%BSA, then mixed with immunoprecipitation mix and incubated for at least 4hours at 4°C with rotation. Chromatin-Antibody-Beads mixes were sequentially washed with ChIP Wash1, ChIP Wash2, ChIP Wash3 (Table E)  Institute.
Raw ChIPseq reads were analyzed on the Cancer Genomics Cloud (CGC) platform (Lau et al., 2017). Reads were aligned to the mouse mm10 Genome obtained from UCSC genome browser using the Burrows-Wheeler Aligner (BWA). Peaks from the aligned reads were obtained using the MACS2 peak calling algorithm with a significance q-value of 0.05. The deepTools bamCoverage command (Ramirez et al., 2016) was used to compare the enrichment of reads in the ChIPseq samples relative to corresponding controls. ChIP-seq samples with distinct separation between control and sample pair for a given marker were retained with exclusion of one H3K4me1 and one H3K27ac replicate. To analyze the changes in acetylation patterns at promoter and enhancer elements, H3K4me3 and H3K4me1 peaks from WT and KO were crossed with H3K9ac-only peaks, H3K27ac-only peaks and dual H3K9ac and H3K27ac peaks from the corresponding genotypes. The H3K9ac-only peaks associated with me3 (promoter elements) were used for further analysis. Genomic peaks were obtained for Kat2a WT and Kat2a KO genotypes separately using Bedtools intersect (Quinlan and Hall, 2010) and H3K4me3 K9ac peaks exclusive to WT genotypes retained as putative Kat2a peaks. Peak locations were converted to fastq sequences using UCSC table browser tool (Karolchik et al., 2004). Genomic Regions Enrichment of Annotations Tool (McLean et al., 2010) was used to assign gene identities to the fastq sequences associated with putative Kat2a peaks. Using the GREAT tool, the genomic region for gene identification was restricted to 1kb upstream and 500bp downstream of the transcription start site (TSS) to infer genes regulated at the promoter level. We used ENCODE ChIP-Seq Significance Tool (Auerbach et al., 2013) to obtain putative transcription factors regulating these targets, as well as lists of genes experimentally bound by GCN5/KAT2A, to confirm the identity of putative Kat2a targets. MEME-chip tool version 4.12.0 (Bailey et al., 2009) (Butler et al., 2018) was used for pre-processing the count-matrix data and obtaining differential gene expression between the two genotypes. RaceID/StemID (Grun et al., 2016) algorithms were used for clustering using t-SNE and obtaining pseudo-temporal arrangement of clusters based on entropy information and cluster stem scores. Parameters for the stochastic gene expression were fitted to the two-state promoter model using the D3E algorithm (Delmans and Hemberg, 2016). R scripts were written for plotting the results as boxplots and for bootstrapping of distance-to-median measure between Kat2a WT and KO.

Statistical analysis
Statistical tests performed are specified in the figure legends. Differences were obtained at pvalue significant at 0.05. All analyses were performed in statistical language R (version 3.4.4).

Data deposition
All single-cell RNAseq data and ChIPseq data were deposited in GEO (SuperSeries GSE118769).

DECLARATION OF INTERESTS
S.P. is a co-founder on Noncodomics, a data analysis company.

ACKNOWLEDGEMENTS
The Kat2a      parameters for genes in the Robust gene set in Kat2a WT and KO primary leukemic cells.
Parameters derived by applying D3E algorithm to single cell RNAseq data. (E) Estimated burst frequency (top) and burst size (bottom) parameters for Kat2a target genes. In (D) and (E), *p<0.05, ***p<0.001, computed with Wilcoxon rank sum test with continuity correction.