Epigenetic mechanisms controlling human leukemia stem cells and therapy resistance

Abstract

Cancer stem cells are essential for initiation and therapy resistance of many cancers, including acute myeloid leukemias (AML). Here, we apply functional genomic profiling to diverse human leukemias, including high-risk MLL- and NUP98-rearranged specimens, using label tracing in vivo. Human leukemia propagation is mediated by a rare quiescent label-retaining cell (LRC) population undetectable by current immunophenotypic markers. AML quiescence is reversible, preserving genetic clonal competition and epigenetic inheritance. LRC quiescence is defined by distinct promoter-centered chromatin and gene expression dynamics controlled by an AP-1/ETS transcription factor network, where JUN is necessary and sufficient for LRC quiescence and associated with persistence and chemotherapy resistance in diverse patients. This enables prospective isolation and manipulation of immunophenotypically-varied leukemia stem cells, establishing the functions of epigenetic plasticity in leukemia development and therapy resistance. These findings offer insights into leukemia stem cell quiescence and the design of therapeutic strategies for their clinical identification and control.

Fig. 1. Quiescent human AML patient cells maintain leukemia initiation, propagation and chemotherapy resistance.

a Experimental design to prospectively isolate quiescent AML cells from human patients using optimized CFSE labeling and orthotopic transplantation in immunodeficient mice (left panel). Representative flow cytometry analysis (for three independent experiments) of quiescent label-retaining human CD45-positive AML cells (LRCs, red box) with high CFSE fluorescence, as compared to their non-label-retaining cells (non-LRCs, black box) that have lost CFSE fluorescence through cell division and proteome turnover, corresponding to Fig. 1b–d (right panel). b–d Leukemia-free survival of mice secondarily transplanted with equal numbers of patient AML LRCs (red) and non-LRCs (black; 5 mice in each group; 900, 1000, and 360 cells/mouse for MSK011, MSK162 and MSK165, respectively), where MSK011 (b) and MSK162 (c) LRCs initiate fully penetrant leukemia, whereas non-LRCs do not (log-rank p = 0.021 and 0.013 for MSK011 and MSK162, respectively). MSK165 (d) LRCs initiate leukemia in 60% of mice, whereas non-LRCs do not (limiting dilution analysis, chi-square test p = 0.019). e Experimental design to analyze mice transplanted with CFSE-labeled human patient AMLs and treated with cytarabine (AraC) and doxorubicin (DXR). f Representative flow cytometry plots (for six biological replicates for each condition) to analyze LRC frequencies in bone marrow human leukemia cells isolated from mice transplanted with CFSE-labeled MSK011 patient AML cells and treated with AraC/DXR chemotherapy or vehicle, corresponding to Fig. 1g, h MSK011 (Supplementary Fig. 7a, b for MSK162 and MSK165). g Combined AraC/DXR chemotherapy treatment reduces bone marrow disease burden of human CD45-positive MSK011 (left), MSK162 (center) and MSK165 (right) AML cell numbers in mice (two-tailed Welch’s t test p = 7.6 × 10⁻⁵, 2.2 × 10⁻⁴, and 3.5 × 10⁻³, respectively). Bars represent mean values of six biological replicates. h LRC frequencies of MSK011 (left), MSK162 (center) and MSK165 (right) patient AML cells are significantly increased upon AraC/DXR chemotherapy (red) as compared to vehicle-treated controls (black), exhibiting the chemotherapy resistance of LRCs (two-tailed Welch’s t test p = 8.7 × 10⁻⁶, 4.9 × 10⁻⁶, and 3.1 × 10⁻⁶, respectively), in contrast to non-LRCs that are largely eradicated by chemotherapy treatment (Supplementary Fig. 7e). Bars represent mean values of measurement of six biological replicates. a, e Free illustration materials from Kenq Net (https://www.wdb.com/kenq/illust/mouse) and SciDraw (10.5281/zenodo.4152947 and 10.5281/zenodo.5204473) are used. Source data are provided as a Source Data file.

Fig. 2. Human patient AML quiescence is reversible evading genetic clonal evolution that maintains disease propagation in serial orthografts in vivo.

a–c Comparison of variant allele frequencies (VAF) of genomic DNA sequencing of LRCs (red) versus non-LRCs (gray) in MSK011 (a), MSK162 (b), and MSK165 (c) human patient AMLs upon serial transplantation of isolated LRCs and non-LRCs, demonstrating several genetic subclones that are enriched in non-LRCs as compared to LRCs. Dashed lines mark 95% confidence intervals (top panels). Bar graphs exhibit VAFs of essential oncogenic mutations in individual patient leukemias, confirming that both LRCs and non-LRCs are all leukemic (bottom panels). Detailed data are provided in Supplementary Table 3−5 and deposited in Zenodo, which include raw read counts of reference and variant alleles. d Experimental design to examine the reversibility of a quiescent state in patient AML cells. Bulk leukemia cells and purified non-LRCs (blue) isolated from the primary recipient mice are re-labeled and sequentially transplanted into secondary recipient mice. Free illustration materials from Kenq Net (https://www.wdb.com/kenq/illust/mouse) and SciDraw (10.5281/zenodo.4152947 and 10.5281/zenodo.5204473) are used. e Representative flow cytometry plots (for three biological replicates for each condition) of human leukemia cells isolated from bone marrow of secondary recipient mice transplanted with bulk leukemia cells (upper panels) or non-LRCs (lower panels) demonstrate gating strategies to measure LRC frequencies (corresponding to Fig. 2f). f Initially proliferating, non-LRC exhibit significant ability to acquire LRC quiescence (blue), and LRC frequencies generated from non-LRCs (blue) are approximately half as compared to parental bulk leukemia cells (black; two-tailed Welch’s t test p = 0.0016). Data show LRC frequencies of human leukemia cells (left y axis) and efficiencies of generated LRCs from non-LRCs relative to bulk cells (right y axis). Bars represent mean values of measurement of biological triplicates. Source data are provided as a Source Data file.

Fig. 3. Human patient AML quiescence is associated with promoter-centered chromatin accessibility dynamics.

a–c Histograms of differentially accessible chromatin regions in quiescent LRCs (red) as compared to non-LRCs (black) as a function of their distance from transcription start sites (ΔTSS) in MSK011 (a), MSK162 (b), and MSK165 (c) human patient leukemias. Data represent measurement triplicates. d Human quiescent LRCs (red) exhibit increased chromatin accessibility of transcriptional start promoter regions as compared to non-LRCs (gray) in MSK011, MSK162, and MSK165 leukemias (two-tailed Fisher’s exact test p = 6.9 × 10⁻², 2.2 × 10⁻¹⁶, and 2.2 × 10⁻¹⁶, respectively). Data represent measurement triplicates. e–j Transcription factor binding sequence motifs enriched in differentially accessible chromatin regions in LRCs (red) as compared to non-LRCs (blue) in MSK011 (e), MSK162 (f), and MSK165 (g), as a function of their statistical significance of enrichment, with specific motif sequences shown (h, i and j, respectively). Motif enrichment statistics are calculated using cumulative binomial distributions implemented in Homer findMotifsGenome.pl. Dashed lines mark p values of 1.0 × 10⁻¹¹. k Venn diagrams represent overlapping transcription factor binding sequence motifs enriched in differentially accessible chromatin regions in LRCs (left) and non-LRCs (right) across three patient AMLs. Source data are provided as a Source Data file.

Fig. 4. Coherent gene expression dysregulation in quiescent patient AML LRC cells.

a Representative gene expression of MSK165 human patient AML cells with statistical significance of measurement triplicates as a function of differential gene expression of LRCs versus non-LRCs. Notable upregulated and downregulated genes are labeled in red and blue, respectively. Results of MSK011 and MSK162 LRC gene expression analysis are described in Supplementary Fig. 10d, e. Differential gene expression statistics are calculated using Wald test implemented in DESeq2 (Fig. 4a–d and Supplementary Fig. 10d, e). Data represent measurement triplicates. b Heatmaps showing commonly differentially regulated genes between LRCs and non-LRCs which are significantly upregulated (left panel) and downregulated (right panel) in at least 2 human patient leukemias (adjusted p < 0.1, fold change >1.5), with blue to red color gradient representing relative decrease and increase of fold change of gene expression, respectively. c Heatmap showing differential gene expression of specific AP-1, KLF and ETS transcription factors (TFs) between LRCs and non-LRCs in MSK011, MSK162, and MSK165 human patient leukemias with blue to red color gradient representing relative decrease and increase of fold change of gene expression, respectively (* denotes adjusted p < 0.1, fold change >1.5). d Heatmap showing differential gene expression of cell surface proteins between LRCs and non-LRCs in MSK011, MSK162, and MSK165 human patient leukemias with blue to red color gradient representing relative decrease and increase of fold change of gene expression, respectively (* denotes adjusted p < 0.1, fold change >1.5). e GO transcription factor target analysis exhibits known AP-1 (JUN or ATF3) target genes that are enriched in LRCs. Enrichment scores and their statistical significance are estimated using the GSEA algorithm (GSEAPreranked v4.4.0). f Gene set enrichment analysis (GSEA) for differentially expressed genes in MSK011, MSK162, and MSK165 LRCs versus non-LRCs, showing commonly dysregulated expression of genes regulating specific cellular signaling pathways. Venn diagrams show the numbers of significantly positively correlated (left) and negatively correlated (right) gene sets in LRCs (p < 0.01, false discovery rate <0.25). Selected significantly enriched gene sets are listed. Enrichment scores and their statistical significance are estimated using the GSEA algorithm (GSEAPreranked v4.4.0). Source data are provided as a Source Data file.

Fig. 5. Shared gene expression dynamics associated with chemotherapy resistance and quiescence in diverse human AML patient specimens.

a Schematic of comparative gene expression analyses of AML cells isolated from the bone marrow (BM) of patients before and after treatment with induction chemotherapy (left) on indicated days, and LRC quiescence of CFSE label retention in mouse orthografts (right). Free illustration materials from Kenq Net (https://www.wdb.com/kenq/illust/mouse) and SciDraw (10.5281/zenodo.4152947 and 10.5281/zenodo.5204473) are used. b Heatmaps showing specific enrichment of distinct gene sets between LRCs in MSK162, MSK011, and MSK165 specimens and chemotherapy-resistant cells in AML329 and AML707B specimens analyzed after induction chemotherapy treatment. Red to blue color gradient represents positive and negative normalized enrichment scores (NES), respectively. c Heatmaps of significance of similarity in gene expression between MSK162, MSK011, and MSK165 LRCs and chemotherapy-resistant cells in AML329 and AML707B patient specimens (numbers indicate two-tailed hypergeometric test p values). White-to-red and white-to-blue colour gradients represent positive and negative odds ratios, respectively. d Commonly dysregulated gene sets shared by both LRCs and residual leukemia cells after chemotherapy were listed. For example, TNF-alpha and NRG1 signaling pathways were positively correlated, whereas MYC, RNA processing and mitochondrial biogenesis pathways were negatively correlated in both LRCs and residual leukemia cells after chemotherapy. Source data are provided as a Source Data file.

Fig. 6. AML stem cell quiescence is controlled by a distinct transcription factor network.

a Experimental design to identify regulatory factors controlling AML stem cell quiescence using lentiviral transduction of doxycycline-inducible cDNA library in human patient AML cells, followed by CFSE labeling to isolate quiescent LRCs (red) and proliferating non-LRCs (black; a representative flow cytometry plot for biological triplicates, corresponding to Fig. 6b–d). Days of induction of cDNA expression using doxycycline treatment are indicated. Each lentiviral doxycycline-inducible cDNA vector includes a specific barcode sequence, enabling quantitative identification by DNA sequencing (bottom). Free illustration materials from Kenq Net (https://www.wdb.com/kenq/illust/mouse) and SciDraw (10.5281/zenodo.4152947 and 10.5281/zenodo.5204473) are used. b Volcano plot showing genes whose enforced expression induces (red) or depletes (blue) quiescent LRCs in MSK162 patient AML specimen. TagBFP serves as a negative control. Statistical significance values are determined using two-tailed Welch’s t test in biological triplicates. c Enforced expression of JUN (center panel) or ETS1 (right panel) in MSK162 patient AML specimen induces or depletes quiescent LRCs (red) as compared to non-LRCs (black), respectively (two-tailed Welch’s t test p = 1.4 × 10⁻³ and 2.8 × 10⁻³, respectively). Enforced TagBFP expression serves as negative control (left panel, two-tailed Welch’s t test p = 0.54). Bars represent mean values of normalized read counts of biological triplicates. d Schematic of the LRC regulatory interaction network. Red and blue circles denote LRC activators and repressors, respectively, with circle size proportional to effect size. Solid lines indicate physical interactions with line thickness corresponding to STRING confidence scores. Source data are provided as a Source Data file.

Fig. 7. Enforced JUN expression promotes LRC quiescence in multiple diverse patient leukemias.

a Experimental design to investigate the function of JUN as a regulator for LRC quiescence in diverse patient AMLs. Patient AML cells are transduced with mCherry-expressing doxycycline-inducible JUN or TagBFP lentivirus vectors, and after labeled with CFSE, transplanted into NSG mice with or without doxycycline diet in vivo (left panel). Representative flow cytometry plots (for 5 biological replicates) show gating strategies to analyze LRC frequencies in JUN or TagBFP-transduced patient leukemia cells by gating on human CD45-positive, mCherry-expressing cells (right panel), corresponding to Fig. 7b, c. Free illustration materials from Kenq Net (https://www.wdb.com/kenq/illust/mouse) and SciDraw (10.5281/zenodo.4152947 and 10.5281/zenodo.5204473) are used. b Representative flow cytometry plots (for 5 biological replicates) exhibit LRC distribution of CD45-positive, mCherry-expressing MSK011 patient leukemia cells, JUN (lower panels) versus TagBFP (upper panels) with (right panels) or without doxycycline (left panels), corresponding to Fig. 7c. c LRC frequencies of human CD45-positive, mCherry-expressing cells are measured in five different patient AMLs (refer to Supplementary Fig. 11g). Dot plots demonstrate fold changes of LRC frequencies in cells with versus without doxycycline induction in JUN (red) or TagBFP (black) transduced cells. LRC frequencies are significantly increased in JUN-transduced cells upon doxycycline induction compared to TagBFP-transduced cells in 4 of 5 patient AML cells, exhibiting that enforced JUN expression induces LRC quiescence. (two-tailed Welch’s t test p = 9.2 × 10⁻⁴, 2.1 × 10⁻², 4.2 × 10⁻³, 2.8 × 10⁻² and 0.10 for MSK011, MSK162, MSK165, MSK136 and MSK183, respectively). Bars represent mean values of measurement of 5 biological replicates, except for MSK183 in which four and three mice are analyzed for JUN and TagBFP, respectively. Source data are provided as a Source Data file.

Fig. 8. Fine-tuned expression of LRC regulators is required for leukemia progression.

a Experimental design for competitive transplantation of JUN-overexpressing-patient leukemia cells. Doxycycline-inducible JUN or TagBFP-transduced, mCherry-expressing patient leukemia cells are separately engineered and isolated by fluorescence-activated cell sorting. Equal cell numbers of JUN and TagBFP-transduced cells are transplanted into NSG mice with or without doxycycline diet in vivo. b Leukemia-free survival of mice transplanted with a mixture of JUN and TagBFP-transduced MSK165 patient leukemia cells with or without doxycycline diet is shown (25,000 cells/mouse, 11 mice for each group). (Supplementary Fig. 12a, b for MSK011 and MSK162). c–e Genomic DNA of engrafted patient leukemia cells is analyzed to determine the dominant clones of propagated leukemia cells in each mouse (refer to Supplementary Fig. 12c–f). Stacked bar charts represent the proportion of dominant clones in each group with (bottom bar) or without (upper bar) doxycycline induction in MSK165 (left), MSK011 (center) and MSK162 (right). JUN-transduced clones are relatively reduced upon doxycycline induction, exhibiting enforced JUN expression impairs leukemia progression (28.6% to 0%%, 33.3% to 0%, and 62.5% to 33.3% for MSK165, MSK011, and MSK162 leukemias, respectively). f Experimental design to investigate leukemia progressing property of JUN-knockout patient leukemia cells, which are generated using Cas9 crRNA;ATTO550-labeled-tracrRNA RNP electroporation, isolated by fluorescence-activated cell sorting, and transplanted into NSG mice. g Leukemia-free survival of mice transplanted with JUN-knockout or control AAVS1-targeted MSK165 patient leukemia cells (30,000 cells/mouse, 15 mice for each group), where JUN-knockout cells propagate leukemia with similar kinetics as control AAVS1-targeted cells (log-rank p = 0.74 and 0.85 for JUN gRNA-1 and gRNA-7 versus AAVS1, respectively). (Supplementary Fig. 14b–e for MSK011 and MSK162). h Mutant allele frequencies are analyzed using the Tracing of Indels by Decomposition (TIDE) method before and after transplantation of JUN-knockout or control AAVS1-targeted MSK165 patient leukemia cells. JUN-knockout clones are enriched upon leukemia progression in vivo, whereas AAVS1-targeted cells are not (n = 10, 9 and 8 mice; two-tailed Welch’s t test p = 0.031, 0.064 and 0.21 for JUN gRNA-1, gRNA-7 and AAVS1, respectively). a, f Free illustration materials from Kenq Net (https://www.wdb.com/kenq/illust/mouse) and SciDraw (10.5281/zenodo.4152947 and 10.5281/zenodo.5204473) are used. Source data are provided as a Source Data file.

Fig. 9. JUN/AP-1 activity is required for chemotherapy resistance.

a Experimental design to investigate whether enforced JUN expression confers chemotherapy resistance on patient leukemia cells. mCherry-expressing, doxycycline-inducible JUN or TagBFP-transduced MSK011 patient leukemia cells are labeled with CFSE and transplanted into NSG mice under doxycycline diet in vivo, followed by combined AraC and DXR chemotherapy in vivo. Free illustration materials from Kenq Net (https://www.wdb.com/kenq/illust/mouse) and SciDraw (10.5281/zenodo.4152947 and 10.5281/zenodo.5204473) are used. b Representative flow cytometry plots (for 5 biological replicates) to analyze bone marrow human leukemia cells isolated from mice transplanted with MSK011 patient leukemia cells containing mCherry-expressing, JUN or TagBFP-transduced cells, corresponding to Fig. 9c–f. c, d Combined AraC and DXR chemotherapy treatment reduces total human CD45-positive (c) and mCherry expressing (d) human leukemia cell numbers in mouse bone marrow, regardless of JUN or TagBFP transduction (two-tailed Welch’s t test p = 7.9 × 10⁻⁴ and 1.3 × 10⁻² for CD45-positive cell numbers of TagBFP and JUN; 1.5 × 10⁻³ and 1.4 × 10⁻² for mCherry-positive cell numbers of TagBFP and JUN, respectively). Bars represent mean values of 5 biological replicates. e, f There is no significant difference in fold reduction of total human leukemia cell numbers in AraC/DXR-treated mice relative to vehicle-treated mice between JUN versus TagBFP transduction group (two-tailed Welch’s t test p = 0.074; l), whereas mCherry-positive JUN-expressing cells exhibit increased resistance to AraC/DXR treatment compared to mCherry-positive TagBFP-expressing cells (two-tailed Welch’s t test p = 0.030). Bars represent the mean values of 5 biological replicates. Source data are provided as a Source Data file.

Fig. 10. Single-cell analysis of LRC quiescence reveals shared and distinct gene expression programs.

a Experimental design for single-cell RNA sequencing of patient LRCs versus non-LRCs performed using three different patient AMLs. Human leukemia cells isolated from mice transplanted with CFSE-labeled patient leukemia cells are separately collected based on CFSE-label retention levels of high, middle and low using fluorescence-activated cell sorting. Gene expression profiles are mapped to a reference atlas of healthy human bone marrow hematopoiesis. Free illustration materials from Kenq Net (https://www.wdb.com/kenq/illust/mouse) and SciDraw (10.5281/zenodo.4152947 and 10.5281/zenodo.5204473) are used (left panel). Created in BioRender. Takao, S. (2025) https://BioRender.com/k45s869 (right panel). b Representative flow cytometry plots (for 3 independent experiments) to isolate human patient leukemia cells based on the levels of CFSE-label retention in MSK011, corresponding to Fig. 10c–f (Supplementary Fig. 17c, e for MSK162 and MSK165, respectively). c Uniform manifold approximation and projection (uMAP) for high, middle-high, middle-low and low-label-retaining cells of MSK011 exhibit that higher label-retaining cell fractions are comprised of more diverse cell types, including HSC- and MPP-like cells. (uMAPs for MSK162 and MSK165 are described in Supplementary Fig. 17d, f). d Stacked bar charts represent projected cell type composition of indicated label-retaining cell fractions of MSK011 (left panel), MSK162 (center panel) and MSK165 (right panel) patient leukemia cells. High label-retaining cells exhibit shared and distinct gene expression programs among three different patient AMLs. e, f uMAP depicts unsupervised clustering of G1 MSK162 cells, which identifies 12 distinct clusters (e). Stacked bar charts represent cell status of label retention in each 12 cluster (f), exhibiting cells with high label retention are enriched in cluster 9 and 11. g Schematic of the mechanisms controlling leukemia cell quiescence and progression. Created in BioRender. Takao, S. (2025) https://BioRender.com/t25j847. Source data are provided as a Source Data file.