Patients with acute myeloid leukemia (AML) frequently relapse after chemotherapy, yet the mechanism by which AML reemerges is not fully understood. Herein, we show that primary AML cells enter a senescence-like phenotype following chemotherapy in vitro and in vivo. This is accompanied by induction of senescence/inflammatory and embryonic diapause transcriptional programs, with downregulation of MYC and leukemia stem cell genes. Single-cell RNA sequencing suggested depletion of leukemia stem cells in vitro and in vivo, and enrichment for subpopulations with distinct senescence-like cells. This senescence effect was transient and conferred superior colony-forming and engraftment potential. Entry into this senescence-like phenotype was dependent on ATR, and persistence of AML cells was severely impaired by ATR inhibitors. Altogether, we propose that AML relapse is facilitated by a senescence-like resilience phenotype that occurs regardless of their stem cell status. Upon recovery, these post-senescence AML cells give rise to relapsed AMLs with increased stem cell potential.
Despite entering complete remission after chemotherapy, relapse occurs in many patients with AML. Thus, there is an urgent need to understand the relapse mechanism in AML and the development of targeted treatments to improve outcome. Here, we identified a senescence-like resilience phenotype through which AML cells can survive and repopulate leukemia.
This article is highlighted in the In This Issue feature, p. 1307
Acute myeloid leukemia (AML) is the most lethal of all blood cancers, killing around three of four patients within five years (https://seer.cancer.gov/statfacts/). Even though 70% to 80% of patients with AML (<60 years) achieve clinical complete remission, the majority of patients will relapse (1) through a largely cryptic mechanism. One prevailing idea is that relapse of AML emerges from an immature drug-resistant subpopulation, the so-called leukemia stem cells (LSC; ref. 2). The LSC concept derived from the observation that only a subpopulation of AML cells with an immature immunophenotype (CD34+, CD38−) can reconstitute leukemia in engraftment experiments using NOD/SCID mice (3). Yet, recent analysis of AML cells surviving chemotherapy in mice and patients at nadir (remaining fraction of chemotherapy-persistent cancer cells after treatment) did not reveal an enrichment of LSCs (4, 5), suggesting additional nuances are involved in relapse. Thus, there is an urgent need to better understand the emergence of relapse in AML.
Genotoxic stress by chemotherapy induces, besides apoptosis, senescence (6, 7), a cellular stress response characterized by aberrant metabolic activity in the absence of proliferation (8). Senescent cells display phenotypic alterations including enlarged cell morphology with increased cytoplasmic granularity and are most commonly characterized by the induction of senescence-associated beta-galactosidase (SA-β-gal; ref. 9). Senescent cells undergo epigenetic changes (10) and secrete extracellular matrix (ECM) modulators as well as proinflammatory cytokines, generally termed the senescence-messaging secretome (11), senescence-associated inflammatory response (SIR; ref. 12), and senescence-associated secretory phenotype (SASP; ref. 13). SASP factors include ECM proteins (e.g., fibronectin and collagen), proteases (e.g., cathepsins and matrix metalloproteinases), interleukins (e.g., IL1 and IL6), chemokines (e.g., IL8/CXCL8 and MCP1), and growth factors (e.g., FGF and PDGF; refs. 11, 14, 15). The SASP factors mediate wound healing (16, 17), developmental senescence (18, 19), chronic inflammation (20), and cellular plasticity (21, 22). Reconciling these contradictory features of senescent cells from an evolutionary standpoint has been challenging.
The initial studies on cellular senescence were performed on nonmalignant cells describing a natural barrier to unlimited proliferation, termed “replicative senescence” or the so-called Hayflick limit (23). Yet, cells that have not reached their replicative life span can also undergo cell arrest in response to sublethal stress, often referred to as (stress-induced) premature senescence (24). A landmark study in the late 1990s demonstrated that expression of oncogenes in primary cells induced premature senescence, defined as oncogene-induced senescence (OIS), that led to the concept that senescence might function as a tumor-suppressive mechanism (25). However, more recent studies have shown that expression of oncogenes in even fully transformed cancer cell lines activates OIS, and that both malignant and nonmalignant cells can escape OIS after a prolonged period in the senescence-arrested state (26, 27). Thus, the premature senescence phenotype is not as static and irreversible as initially thought, particularly in malignant cells (27–30).
Overall, cellular senescence is a complex phenomenon, initiated by a wide range of stimuli including DNA damage, oncogene activation, telomere erosion, protein misfolding, oxidative damage, extracellular signaling by mitogens or cytokines, and high-fat diet (8). One interpretation that potentially reflects the common effect of these various senescence inducers could be that senescence acts as a conserved stress-response mechanism. Given the frequency with which cells encounter environmental stress, it is not unreasonable to expect cells to possess evolutionarily conserved mechanisms to tolerate harmful conditions. For example, the embryonic diapause is an evolutionary strategy in many mammals that describes a reversible growth arrest (up to many months) of blastocysts (31, 32), presumably enabling survival during stressful environmental conditions. We hypothesized that AML relapse is dependent on these cells undergoing a senescence-like resilient state, potentially mimicking a diapause state, to survive genotoxic stress caused by chemotherapy and eventually repopulate the disease.
Chemotherapy Induces Cellular Senescence in Patient-Derived AMLs
To determine if chemotherapy induces a senescence-like state, we analyzed published gene expression microarray profiles of 15 primary patient-derived AML specimens exposed to cytarabine (Ara-C), a standard component of AML therapy, for three days in a coculture system (ref. 33; Fig. 1A). Gene set enrichment analysis (GSEA; ref. 34) showed significant enrichment of senescence and SASP/inflammatory signatures in >90% of primary patient-derived cases (Fig. 1A). We also observed significant enrichment for diapause-associated genes (Supplementary Table S1; ref. 35) after chemotherapy (Fig. 1A). Reciprocally, we observed significant negative enrichment for genes that are downregulated during diapause, as well as genes associated with cell division, consistent with the growth arrest linked to senescence.
To validate whether chemotherapy induces a senescence-like phenotype in myeloid leukemia cells, we examined induction of X-gal–based β-galactosidase staining as a canonical biomarker for this state (9). Using an ex vivo coculture system that allows reliable expansion of primary AML cells collected from patients (36), we observed key features of senescence including increased cell size, granularity, and SA-β-gal activity after three days of exposure to Ara-C or anthracyclines, the two chemotherapy drugs used for standard treatment of this disease, among cell lines and patient specimens (Fig. 1B and C; Supplementary Fig. S1A; ref. 6). The X-gal–based assay has two main limitations: the requirement for chemical fixation, precluding viable cells for further analysis, and limited ability to quantify populations of cells through microscopic assessment. In order to quickly and quantitatively assess SA-β-gal activity within viable cells, we used the fluorogenic substrate 5-dodecanoylaminofluorescein di-β-d-galactopyranoside (C12-FDG; ref. 37; Supplementary Fig. S1B and S1C). Using this assay, we confirmed induction of this senescence marker as well as increased granularity in additional patient-derived AML specimens (Fig. 1D; Supplementary Fig. S1D). Because of the inherent fluorescence property of anthracyclines (38) that would interfere with the C12-FDG signal in our subsequent analyses, we focused largely on chemotherapy treatment with Ara-C. Chemotherapy-induced senescence (CIS) occurred in myeloid leukemia cells regardless of p16 or p53 mutation or expression (Supplementary Fig. S2A and S2B), consistent with previous reports (12, 39, 40). Senescence was not simply a marker of differentiation, because an LSD1 inhibitor (36, 41, 42) induced strong expression of CD11b without an increase of β-gal activity (Supplementary Fig. S2C and S2D).
To explore the gene signatures induced by chemotherapy exposure at an earlier time point that might further indicate the role of senescence as a primary effect on these cells, we performed RNA sequencing (RNA-seq) on patient-derived AML cells exposed to Ara-C and harvested 24 hours later. We identified 1,553 differentially expressed genes [DEG; at least 2-fold change (FC) and adjusted P < 0.01] with approximately 88% of genes being upregulated (Fig. 1E; Supplementary Table S2). Among the upregulated genes were many that encoded typical SASP factors including several members of the cathepsin family and cytokines like IL8. GSEA demonstrated significant enrichment for gene signatures associated with senescence, inflammation, and diapause (Fig. 1F; refs. 12, 13, 35, 43, 44). Reciprocally to the senescence phenotype, cell-cycle genes were significantly downregulated (Fig. 1G). Among the downregulated genes was MYC (Supplementary Table S2), depletion of which was shown to induce a dormant state mimicking diapause (45). Additional pathway analysis of genes upregulated by Ara-C exposure showed enrichment for gene signatures associated with ECM modulation and transcriptional response to genotoxic stress (Fig. 1H). We observed significant enrichment of our early upregulated signature in 14/15 of the patient specimens profiled at day 3 post–chemotherapy exposure (33), suggesting a conserved response mechanism in AML (Fig. 1I).
Senescence-Like AML Cells Can Repopulate AMLs after a Latent State
Relapse of AML frequently occurs following a variable nadir phase. To mimic this, we first seeded primary human AML cells in our coculture organoid system and exposed them to increasing Ara-C concentrations in the range of clinically relevant concentrations (ranging between the induction and close to consolidation doses of this drug used in the clinic; ref. 46). We determined the association between the degree of genotoxic stress and senescence induction by SA-β-gal activity assessed by flow cytometry after three days of Ara-C treatment, and observed little induction of senescence at 100 nmol/L of drug, but much stronger induction upon exposure to 1,000 nmol/L (Fig. 2A). In contrast, senescent cells were markedly diminished after exposure to 10,000 nmol/L Ara-C (Fig. 2A).
To model the nadir phase ex vivo and how leukemia regenerates after exposure to chemotherapy, we determined cell viability by flow cytometry (dye exclusion) over the course of 10 days after treating the cells with increasing doses of Ara-C for three days (Fig. 2B). There was very little effect on viability with the 100 nmol/L dose, but close to 100% cell death with the 10,000 nmol/L dose. In contrast, the intermediate dose yielded a significant fraction of cells (∼20%) that persisted in a viable state (Fig. 2B). To determine if these cells were end stage or could recover and regenerate the leukemic population, we performed serial viable cell counts over an extended time period by flow cytometry up to 30 days post Ara-C treatment (Fig. 2C). Cells treated with 100 nmol/L Ara-C manifested only a slight delay in their expansion, whereas the few cells remaining after the highest dose of Ara-C remained in a quiescent phase throughout the experiment. In contrast, those AMLs treated with 1,000 nmol/L Ara-C remained in a latent state until day 9, after which they were able to reexpand (Fig. 2C). These results suggest that induction of senescence is dose-dependent and reversible, allowing cells to potentially recover and restore the disease.
CIS as a Cellular Resilience Mechanism
To explore whether the senescence-like phenotype is a reversible resilience response as opposed to a process of selecting preexisting chemotherapy-resistant C12-FDGhi subpopulations, we first sorted C12-FDGhi vs. C12-FDGlo AML cells in three different patient-derived cases (Supplementary Fig. S3A). Exposure to increasing doses of Ara-C or daunorubicin did not reveal evidence of differential sensitivity in these C12-FDGhi cells (Supplementary Fig. S3B). To address this question in a more rigorous fashion, we next performed sequential cycles of Ara-C treatment and recovery in primary patient-derived AML cells, using a dose (500 nmol/L) that kills >90% of cells from this patient (Fig. 2D). These cells required approximately one to two months to recover after each Ara-C treatment, and these cycles were repeated a total of ten times (Fig. 2D). Notably, after the 10th cycle the AML cells manifested only a small increase in their EC50 (Fig. 2E). Moreover, Ara-C induced a similar degree of senescence in AML cells after the 10th recovery as compared with the initial parental cells (Fig. 2F).
In addition, we created an ex vivo AML clonal recurrence model in which single primary human cells were placed in culture before or after serial Ara-C treatment from the residual fraction of chemotherapy-persistent cells at nadir (Fig. 2G). Single cells from each time point formed a monoclonal population, thereby excluding genetic heterogeneity within the recovered population. This procedure revealed a modest and gradual increase in the fraction of surviving cells at nadir after each Ara-C recovery cycle (Fig. 2H), indicating that the single cell–derived clones are slightly more resilient to chemotherapy, consistent with progressive stress adaptation (47, 48) rather than chemotherapy resistance, as there was no major shift in the EC50 of previously exposed cells to Ara-C (Fig. 2H). Of note, consistent with previous data linking DNA methylation with population fitness in patients with AML (49), we observed increased abundance of methylcytosine in recovered cell populations (Fig. 2I). Collectively, these findings support that the senescence-like phenotype is a reversible resilience state instead of a manifestation of an inherently chemotherapy-resistant subpopulation that would have selected Ara-C–resistant clones (Fig. 2F).
Induction of Senescence-Like Phenotype Is Dose-Dependent Based on Chemotherapy Sensitivity
Patient-derived AML cells manifest significant variation in sensitivity (EC50) to Ara-C (Supplementary Fig. S4A). To explore the link between chemotherapy sensitivity and senescence, we next examined the senescence response at four different doses of Ara-C in three patient specimens with variable degree of sensitivity. Notably, the patient who was most insensitive to Ara-C manifested the greatest induction of senescence that peaked at 1,000 nmol/L dose of drug (Fig. 3A), but did not display senescence at the highest dose (where these cells were virtually all killed). The most sensitive case (AML 2741) peaked at 10 nmol/L, after which there was loss of senescent populations. The intermediate case manifested a weaker degree of senescence induction at lower doses than the resistant patient (Fig. 3B). Hence, in general we observed that higher concentrations of Ara-C induced increased SA-β-gal activity in AML cells until reaching a critical toxic dose that killed these cells (Fig. 3A and B).
Our coculture experiments suggested that acquisition of the senescence-like phenotype was associated with persistence and retention of leukemia-repopulating potential. To further explore if this was the case, we treated patient-derived AML cells in coculture with 1,000 nmol/L Ara-C for three days, after which the specimens were sorted into C12-FDGhi and C12-FDGlo viable (dye exclusion flow cytometry) fractions and plated on methylcellulose for colony-forming cell assays. Two of our cases were able to form colonies, and in both cases the senescence-like cells had significantly higher clonogenic potential (P < 0.001) compared with cells with low SA-β-gal activity after Ara-C treatment (Fig. 3C and D; Supplementary Fig. S4B–S4D). To determine if senescence was associated with true leukemia-initiating potential, we next treated patient-derived AML cells ex vivo with Ara-C 1,000 nmol/L or vehicle for three days in coculture, after which we sorted viable C12-FDGhi cells from the Ara-C–treated condition, as well as viable control cells, and implanted them in NSG mice (n = 5 mice each, Fig. 3E and F). Contrary to the expectation that senescence-like cells would be terminally nonreplicative and unable to reconstitute disease in vivo, we found that mice transplanted with senescence-like cells readily repopulated leukemia (Fig. 3F). We observed a trend to lower survival in mice transplanted with senescence-like cells as compared with untreated AML cells (P = 0.15; Fig. 3F). This observation is consistent with recent reports showing increased tumor repopulation properties of tumor cells that recover from CIS (28, 50).
Senescence-Like AML Cells Are Depleted of Leukemia Stem Cell Populations
Cancer stem cells are considered inherently resistant to cytotoxic drugs (2) and are hence expected to be enriched following chemotherapy. Given the ability of senescent cells to repopulate primary AMLs, we next examined whether LSCs are enriched in this population. For this, we used sorted viable C12-FDGhi senescence-like primary patient-derived AML cells collected three days after exposure to 1,000 nmol/L Ara-C (that killed the majority cells) as well as untreated viable control cells (Fig. 4A and B). These two populations were subjected to single-cell droplet-based RNA-seq. Plotting the variance of individual cell gene-expression profiles on 4,627 cells using UMAP embedding (51) demonstrated separation between the C12-FDGhi and untreated AML cells (Fig. 4C). As expected, projecting the bulk RNA-seq gene-expression signature derived from Ara-C–treated cells in Fig. 1E showed predominant enrichment among the sorted senescent cells (Supplementary Fig. S5A). Genes linked to SASP upregulated upon Ara-C treatment (e.g., cathepsins and IL8) were confirmed to be highly expressed in the senescent population (Supplementary Fig. S5B; Supplementary Table S3). On the other hand, MYC and genes associated with ribosome biosynthesis (e.g., RPL38) were among the most downregulated genes (Supplementary Fig. S5C). MYC target genes were depleted in AML cells that highly expressed senescence and diapause gene signatures (Supplementary Fig. S5D). Although there was some degree of heterogeneity among AML cells, we found significant inverse correlation of expression between MYC target genes and chemotherapy-induced stress genes (CISG) that we identified in this study (Supplementary Fig. S5E; Supplementary Table S1).
To further delineate subpopulations among these cells, we performed a k-nearest neighbors–based analysis, which yielded seven clusters (Fig. 4D; Supplementary Table S4) that were largely mutually exclusive between control and senescent cells, with the exception of senescence cluster 4, which consisted of ∼17% untreated control cells (Fig. 4E; Supplementary Fig. S5F). We next projected gene sets linked to senescence, diapause, or LSCs/hematopoietic stem cells (HSC) on the UMAP plots (Fig. 4F and G). As expected, genes associated with senescence and SASP were highly expressed within Ara-C–treated cells that compose clusters: 0, 3, 4, and 6 (Fig. 4F and H). Diapause genes were most highly expressed in cells composing clusters 4 and 6. Among untreated cells, only the small fraction contained in cluster 4 manifested strong association with senescence signatures, consistent with the presence of small numbers of senescent cells under basal conditions (Fig. 4E and H). Analysis of stem cell signatures showed the highest expression among control cell clusters 1, 2, and 5 (Fig. 4G and H). In contrast, there were reduced numbers of cells and lower expression of these signatures within senescent cell populations, although a sizable fraction of cells in clusters 3 and 0 still retained moderate expression of these gene sets (Fig. 4G and H). Single-cell RNA-seq (scRNA-seq) analysis of the senescence-like AML cells showed no enrichment of LSC markers that could explain the association with increased AML engraftment (Fig. 3F; Supplementary Fig. S5G). WNT/β-catenin pathways were reported to become de novo activated during senescence-associated reprogramming in cancer cells (28) and were likewise induced among cluster 6 posttreatment subpopulations compared with control cells (P < 0.0001; Fig. 4I and J). Altogether, we did not observe an enrichment for expression of LSC-associated genes nor an increase of the abundance of LSCs among chemotherapy-induced senescent cells.
To further confirm that we do not select for LSCs, we performed GSEA for sets of genes that are preferentially expressed or repressed in LSCs in the set of 15 patients with AML treated with Ara-C from Klco and colleagues (ref. 33; Fig. 1A). This analysis revealed significant enrichment of genes that are downregulated in stem cells in Ara-C–treated cells, with reciprocal depletion or lack of enrichment of genes that are induced in LSCs (Fig. 4K). GSEA also suggested induction of WNT/β-catenin programs after Ara-C exposure (Supplementary Fig. S5H), although not always significant, perhaps due to the cellular heterogeneity seen at the single-cell level (Fig. 4I and J). In addition, we treated an Ara-C–sensitive primary AML specimen (AML2741, EC50: ∼10 nmol/L) with Ara-C of 100 nmol/L (a toxic dose for this patient specimen) for three days, and then performed RNA-seq on the small number of purified viable Ara-C–persistent AML cells after treatment, as well as purified viable cells exposed to vehicle only (Fig. 4L). GSEA showed significant enrichment for senescence, diapause, and genes induced by Ara-C at 24 hours (Fig. 4M; Supplementary Fig. S5I). Genes downregulated in HSCs/LSCs were positively enriched in these Ara-C–persisting cells, whereas genes upregulated in HSCs/LSCs were significantly downregulated (Fig. 4M).
We next sought to examine which preexisting AML subpopulations might preferentially undergo senescence and treated six independent patient-derived cases (Supplementary Fig. S6A and S6B) that were measured for CD34, CD38, and C12-FDG by flow cytometry. Apart from the expected differences in the cellular composition between AML samples, we found that immature (CD34+CD38−) as well as mature AML cells (CD34−CD38+) can undergo senescence (Supplementary Fig. S6A and S6B).
Collectively, the data argue against selective enrichment for LSCs immediately following Ara-C treatment and suggest that survival post-chemotherapy may be linked instead to stochastic induction of senescence-like programs in cells independent of whether they correspond to LSC or bulk AML populations.
CIS Gene Signatures Suggest Potential Therapeutic Vulnerabilities
To gain deeper insight into biological functions affected by the AML chemotherapy senescence response, we merged the RNA-seq profiles obtained in the two different primary patient-derived AML specimens shown in Figs. 1 and 3, and identified the subset of genes most robustly differentially expressed after Ara-C exposure. This procedure yielded a gene-expression signature consisting of 262 upregulated and 26 downregulated genes (2-FC, q < 0.05; Fig. 5A; Supplementary Table S5). A number of the upregulated genes have been previously noted in SASP phenotypes such as IL8 and ECM modulators (matrix metalloproteinases, collagens, and cathepsin genes), which were also noted above. We also noted significant induction of diapause regulating genes such as LIF and EGF as well as robust downregulation of MYC (refs. 32, 45; Fig. 5A).
To further understand the biological functions induced in these cells upon CIS, we performed GSEA and a leading-edge network analysis using canonical pathways from the Molecular Signatures Database (MSigDB; http://software.broadinstitute.org/gsea/msigdb/search.jsp; Fig. 5B; Supplementary Table S6). Analysis of these gene networks strongly points to induction of inflammatory genes that are known to be induced by NFκB activation upon induction of SASP (52). Indeed, we observed significant enrichment for NFκB target genes among genes that are upregulated in primary AML cells exposed to Ara-C (Fig. 5C). Recent reports place ATR and ATM as upstream mediators of senescence-associated NFκB activation (53, 54). If the senescence-like phenotype observed in AML after chemotherapy is induced by ATR and is required for the resilience of these cells, then ATR-targeted therapy would be expected to prevent AML cells from achieving this protective state (Fig. 5D). Supporting this notion, we observed significant downregulation of genes suppressed by ATR, when examined by GSEA in Ara-C–treated AML cells [false discovery rate (FDR) = 0.02, Fig. 5E; ref. 55].
To determine whether ATR activity was induced by Ara-C in AML cells, we measured the abundance of phospho-CHK1 as a biochemical readout for ATR function. These experiments confirmed induction of CHK1 phosphorylation versus total CHK1 in AML specimens after exposure to Ara-C. This effect was severely impaired when cells were exposed to the specific ATR inhibitor VE821 (56) just prior to treatment with Ara-C for 24 hours (Fig. 5F). Because genetic deletion of ATR protein is lethal to dividing cells (57, 58), we examined the effect of catalytic inhibition with VE821, a selective ATR inhibitor chemically related to VX-970 (M6620), which is currently in clinical trials (59, 60). AML cells were pretreated for two hours with VE821 (2 μmol/L) or vehicle and then exposed to Ara-C or vehicle for three days, followed by C12-FDG flow cytometry. Strikingly, the VE821 severely impaired C12-FDG induction, suggesting that ATR activation plays a critical role in inducing the senescent/SASP phenotype in the AML context (Fig. 5G). If the senescent phenotype has a protective role, then it would be expected that VE821 would prevent survival of AML cells after exposure to Ara-C. For this, we exposed a panel of five primary AML specimens as well as the KG-1 AML cell line to Ara-C, VE821, or the combination of both for three days to assess cell viability one week after treatment (Fig. 5H). In half of the cases, Ara-C and VE821 killed roughly 50% of the AML cells, whereas the combination killed 95% of cells (Fig. 5H). However, in the other three cases, killing was supra-additive because these cells were relatively resistant to either Ara-C or VE821, yet still manifested >95% killing in the presence of the combination. These results were confirmed using a second, more potent ATR inhibitor BAY-1895344, which is also in clinical trials (ref. 61; Fig. 5I and J). In contrast, these drugs did not induce a marked killing effect in normal donor hematopoietic stem/progenitor cells (HSPC; Fig. 5J). Fordham and colleagues recently demonstrated that in vivo treatment with ATR inhibitor and chemotherapy also yields dramatically enhanced activity against patient-derived AML in mice (62). Altogether, these results suggest that ATR-targeted therapy strongly potentiates the efficacy of chemotherapy in AML by preventing these cells from being rescued through induction of the senescence resilience phenotype.
Senescence-Like AML Cells Are Enriched at Post-Induction Chemotherapy Nadir in Patients with AML
If it is true that AMLs enter a senescence-like state that enables them to persist after exposure to cytotoxic drugs, then we would expect that such cells would be enriched in the bone marrow of patients with AML after induction chemotherapy. To determine if this was the case, we investigated gene-expression profiles of residual AML cells that persist in the bone marrow of four human subjects approximately three weeks following induction therapy, as compared with their matched diagnostic specimens (Fig. 6A; ref. 4). This analysis revealed enrichment of our CISG_UP senescence and SASP signatures in the residual AML cells in three of these patients (Fig. 6B). In contrast, there was depletion of leukemia stem cell signatures in the patients.
We next determined the distribution of the senescence signature across the population of residual AML cells persisting 14 to 35 days post induction chemotherapy using scRNA-seq data collected from nine patients with AML, as compared with their matched diagnostic specimens (Fig. 6C; ref. 63). Analyzing the gene-expression profiles of myeloid, myeloid-like, HSPC, and HSPC-like cells collected from the bone marrow of patients with AML, we found generally depletion or no enrichment of LSCs in seven of nine patients with AML at nadir compared with diagnosis (Fig. 6D; Supplementary Fig. S7 and S8). On the other hand, we found around half of the patients (n = 4 with P < 0.05 and n = 1 with P = 0.066) exhibited significant enrichment for the CISG_UP signature at nadir (Fig. 6D). Similarly, we also observed that many cases displayed increased abundance of cells expressing SASP, senescence, diapause, inflammation, and β-catenin signatures (Fig. 6D; Supplementary Fig. S7). Of note, we found upregulation of several differentiation markers in these cases at nadir. This is coherent with the observation that differentiation markers can be induced by DNA-damaging agents (64). Still, we did not observe the same set of differentiation markers more highly expressed across different AML cases, which we expected for this heterogeneous disease (Supplementary Fig. S8).
Single-cell genotyping in one of these patients yielded sufficient cells to examine gene expression in genetically defined AML cells at diagnosis, day 20 after beginning induction therapy (<1% blasts), and a subsequent time point with increased blast counts at day 37 (3% blasts; Fig. 6E). This allowed us to exclude potential normal cells from our analysis (63). AML cells at day 37 revealed closer clustering to diagnosis samples as compared with nadir cells (Fig. 6F), whereas T cells did not manifest detectable transcriptional difference among the three time points (Fig. 6G; controlling for variability between specimens). We observed strong MYC target gene signature expression in AML cells at diagnosis and significantly lower levels at nadir (P < 1e−12), which were then significantly reversed at day 37 (P = 0.035; Fig. 6H). Conversely, we found significantly higher expression of senescence-related genes at nadir compared with diagnosis (P < 1e−12) and partial reversal by day 37 (P = 0.06; Fig. 6H). Most LSC/HSC signatures still showed significant reduction at day 37 compared with diagnosis, except for one of these signatures (Fig. 6I), whereas β-catenin/WNT pathway genes were enriched at nadir (day 20) compared with diagnosis (P = 0.035, respectively), and remained elevated as this patient was relapsing (P = 0.007; Fig. 6I).
Overall, the data indicate preferential enrichment of senescence-like resilient AML cells in patient specimens collected at nadir after induction therapy.
Although Often Depleted at Nadir, LSC Signatures Are Enriched at Relapse
Our data suggest that senescence signatures are enriched at nadir and are associated with persistence of AML cells that can repopulate the disease when they recover. To explore senescence and LSC/HSC signatures upon recovery of primary AMLs in vivo after nadir while excluding normal cells from our analysis, we established an AML relapse model using primary human AML specimens engrafted in NSG mice (Fig. 7A). After validation of engraftment of human blasts, mice were treated with a clinically relevant dose of Ara-C (60 mg/kg/day, equivalent to induction therapy in humans; ref. 5) for five consecutive days, which resulted in severe cytoreduction and nadir by day 8, followed by leukemic relapse occurring after 4 weeks (Fig. 7A). We performed gene-expression profiling on AML cells collected from NSG mice at three time points: just prior to treatment, at nadir (8 days after initiation of treatment), and after leukemia relapse (day 29 after initiation of treatment; Fig. 7B). We found that the genes composing the CISG_UP-induced signature were upregulated at nadir, but were partially reverted to their pretreatment levels at day 29 (Fig. 7C). We then engrafted four additional independent human primary AMLs into NSG mice, treated them with the same Ara-C regimen, and performed gene-expression analysis on AML purified cells pretreatment, at nadir, and upon relapse (Fig. 7D). Using GSEA on these gene-expression profiles revealed enrichment for the CISG_UP, SASP, senescence, and diapause signatures as well as partial enrichment for β-catenin at nadir (day 8) vs. pretreatment (day 0; Fig. 7D). In contrast, there was no enrichment for LSC signatures at day 8, which were instead often significantly depleted. However, when comparing gene-expression profiles at day 29 (relapse) versus day 8 nadir, we observed the opposite effect: depletion of CISG_UP, SASP, diapause, and β-catenin signatures, and enrichment for LSC profiles in at least three of the five cases. These data suggest senescent programs resolve as AML cells recover from CIS-like state, and as they emerge from this state, they feature enrichment for more stem-like transcriptional programming. These findings are consistent with a recent report showing that senescence is accompanied by epigenetic plasticity and reactivation of stem cell functionalities (28). To further examine this point, we performed RNA-seq analysis in 59 matched AML specimens collected at diagnosis versus relapse (Fig. 7E). GSEA demonstrated that AML cells exhibit a depletion for senescence gene signatures and an enrichment for LSC signatures at relapse (Fig. 7F and G). Relapsed AML compared with diagnosis samples demonstrated partial enrichment of WNT/β-catenin programs (Fig. 7G). Overall, these data suggest that survival after exposure to chemotherapy is most likely stochastic and involves a diapause/SASP-like resilience response, and that subsequent enrichment for stem-like features at relapse could reflect the consequence, rather than the cause, of leukemic survival after exposure to chemotherapy.
AML relapse occurs in a majority of patients after chemotherapy and portends incurable disease. Although selection of clones containing drug-resistance mutations (e.g., TP53) following chemotherapy is relatively rarely observed in relapsed AML, a number of studies suggest that relapse is explained by the presence of inherently more resistant LSCs as the origin of relapse (65, 66). Yet, most of these studies analyzed samples at diagnosis and relapse but not at nadir. Although LSC gene expression signatures are predictive of relapse (67, 68), it does not necessarily mean that these cells are enriched during exposure to chemotherapy. Indeed, while we also observed an enrichment of LSC gene signatures at relapse, we did not find LSC gene signatures in the residual fraction of chemotherapy-persistent leukemia cells at nadir. This is consistent with other recent studies reporting no enrichment of LSCs at nadir (4, 5). Instead, we found that AML cells surviving initial chemotherapy manifest a senescence-like resilient phenotype, with potential to repopulate leukemia. Previous studies analyzing the transient state of chemotherapy-persistent AML at nadir indicate the expression of genes related to SASP/inflammation and senescence-like cells (4, 5). Upon recovery these cells manifested positive enrichment of LSC gene signatures.
We suggest that LSC programming in many AML cases occurs during the senescence-like resilient phenotype, explaining how these stem cells are enriched at relapse. This hypothesis is supported by a recent study showing that doxorubicin-induced senescence reprograms murine leukemia and lymphoma cells to cancer stemness (28). We do not exclude that in certain AMLs with initially abundant LSCs, these cells could survive chemotherapy without major harm and repopulate the disease. Given the heterogeneity between patients with AML, it is also possible that other resilience mechanisms could exist that were not captured here. Yet, our data suggest that many AMLs enter a stress-adaptive and senescence-like resilient phenotype regardless of whether they are LSCs or not, and that this process plays a critical role in relapse, possibly through an acquired stemness reprogramming process mediated by the WNT/β-catenin pathway as described previously (28). Alternatively, adaptation may reflect either initial epigenetic heterogeneity or epigenetic plasticity as possible mechanisms of resilience. Fundamental to understanding this process is the inherent plasticity of cells to reprogram from one state to another. Importantly in this regard, cancer stem cell functionality is not a fixed property but rather can be induced under stress conditions (28, 69). Another layer of complexity in this context is that different AMLs require different doses of chemotherapy to induce a comparable senescence phenotype in which some refractory AMLs will only manifest overt senescence phenotypes when exposed to higher drug concentrations. Furthermore, AML is a heterogeneous disease and we cannot exclude that certain AML subtypes induce the senescence phenotype to a different degree than other subtypes.
Our findings are consistent with a report that investigated drug resistance in solid tumors and revealed that the drug-persistent state shares a diapause-like dormancy signature (70). The diapause principle in theory represents a long-term dormancy state. Almost half of senescence-associated genes are also expressed in quiescent cells, potentially indicating a mechanistic overlap between senescence and dormancy (71). The concept of a transient stress-induced endurance state is a common phenomenon described in developmental biology, whereby early embryos survive adverse conditions via transition into diapause (32). MYC controls cellular growth, and depletion of MYC can induce a diapause-like state (45, 72). Notably, translocations leading to constitutive expression of MYC are associated with better prognosis in certain cancers (e.g., Burkitt lymphoma; ref. 73), presumably in part because of the aggressive cell proliferation that incorporates more chemotherapy. Thus, suppression of MYC may represent a survival strategy of cells to endure stress. We have used the phrase “senescence-like phenotype” to describe our results. As noted, AML cells treated with chemotherapy induce a cell-cycle arrest and growth pause similar to replicative senescence, but unlike replicative senescence, the growth pause is not permanent. The observed senescence transcriptional profiles are most similar to SASP, a specific phenotype of the senescent state initially described in response to genotoxic stress (13). SASP is characterized by an inflammatory response that may contribute functionally to leukemic recovery, consistent with notions on regenerative properties of AML nadir cells described by Boyd and colleagues (4). Among SASP genes, we observed induction of IL8, which has been described to protect AML cells from the effects of chemotherapy (74).
Our data indicate that ATR plays a critical role in the CIS-resilient phenotype. This is consistent with previous reports showing that ATR can induce SASP through induction of NFκB in a p53-independent manner (53, 54). Indeed, we show for the first time that induction of senescence as well as tumor survival and persistence is dependent on ATR catalytic activity, through our experiments using chemotherapy treatment in primary human AML cells. Moreover, others have shown that treatment of primary AML cells with ATR inhibitor and chemotherapy is also remarkably effective in vivo (62). ATR inhibitors may thus represent a novel approach for targeting senescence resilience by preventing cells from harnessing this survival mechanism. It is intriguing to consider the potential for ATR inhibitors to potentiate the effect of drugs that directly target senescent cells such as BCL2 inhibitors, which have been used to enhance the efficacy of chemotherapy and other drugs in patients with AML (75, 76). ATR inhibitors are currently in clinical trials both as single agent and in combination with various chemotherapy drugs for solid tumors. Collectively, these and our data strongly point to the rationale for initiating clinical trials of ATR inhibitors in patients with AML. In summary, our findings suggest that AML relapse can be explained by entry of these cells into an ATR-dependent senescence-like resilient state of variable duration that endows them with superior fitness and enhanced ability to repopulate the disease. Although this effect appears to be independent of their initial stemness profiles, it may make these cells highly vulnerable to drugs such as ATR inhibitors, with potential to better eradicate these tumors.
Primary AML Samples
Primary human AML specimens were collected after written informed consent and obtained from peripheral blood or bone marrow aspiration according to the declaration of Helsinki for collection and use of sample materials. The studies were approved by the institutional review board at Weill Cornell Medicine (IRB protocol #0805009783). Information on the primary AML cases used in this study is described in Supplementary Table S7 and in more detail elsewhere (36, 49).
Culturing of Primary Patient-Derived AML
Primary AMLs were cultured ex vivo as previously described (36). In short, primary patient-derived AMLs were expanded on stromal feeder layers with cytokine support.
AML Cell Lines
Human myeloid leukemia cell lines were purchased from ATCC or DSMZ and maintained in Iscove's Modified Dulbecco's Medium (IMDM) containing 10% to 20% fetal bovine serum, 100 IU/mL penicillin, and 100 μg/mL streptomycin. Cell lines were validated via short tandem repeat DNA profiling and monitored regularly for Mycoplasma contamination.
Isolation of HSPCs
Mononuclear cells (MNC) were isolated from fresh human umbilical cord blood samples (New York Blood Center) using Ficoll (Atlanta Biologicals) density gradient centrifugation. After lysis of red blood cells, HSPCs were selected via immunomagnetic enrichment of CD34+ MNCs using CD34 MicroBead Kit and Automacs from Miltenyi Biotec. HSPCs were propagated in our ex vivo coculture model as previously described (36).
Drug Preparation and Treatment
Stock solutions of inhibitors against ATR (VE-821, BAY-1895344) derived from Selleckchem were prepared in DMSO and stored at −80°C. Cells were exposed to 2 μmol/L (VE-821) or 20 nmol/L (BAY-1895344) in a dilution factor that did not pass 0.3% DMSO content in growth medium during the whole treatment time. Cytarabine (Sigma-Aldrich), daunorubicin (EMD Millipore), and adriamycin/doxorubicin (Selleckchem) were prepared in water and kept protected from light at 4°C.
X-Gal–SA-β-Gal Activity Assay
The SA-β-gal staining kit from Cell Signaling Technology was used for SA-β-gal analyses and performed according to the manufacturer's protocol (BD Biosciences).
Fluorogenic SA-β-Gal Activity Assay
Cells cultured for a prolonged period were maintained in well plates that had PBS or plain media in the outer wells to reduce edge effects. Otherwise, cells cultured at outer wells exhibited stronger evaporation effects that increased intracellular stress and SA-β-gal activity. C12-FDG procedure was performed as previously described (37) using 20 μmol/L C12-FDG (Thermo Fisher Scientific). C12-FDG was prepared in prewarmed media and added to cells with gentle mixing of the cell suspension. Cells were protected from light and incubated for two hours at 37°C and 5% CO2 before flow cytometry measurement.
Cell viability was assessed by the propidium iodide (PI)–negative cell proportion, while cell growth was measured by absolute numbers of viable cells (PIneg) including forward scatter (FSC) and sideward scatter (SSC) size discrimination. Annexin V costained with PI was used for apoptosis analyses and performed according to the manufacturer's protocol (BD).
Patient-Derived Xenograft Relapse Model
MNCs extracted from patients were transplanted into NOD/SCID/IL2Rγnull (NSG) mice, which were obtained from The Jackson Laboratory. Mice were randomly assigned to treatment groups. The investigators were not blinded to allocation during experiments and outcome assessment. No power analysis was used to predetermine sample size. Ara-C administered using a physiologically relevant dose and schedule (60 mg/kg/day × 5 days) to AML-engrafted NSG mice induced a drop in peripheral blood leukemic cells and total body leukemic burden by one week after initiation of therapy. Human AML cells were collected from four bones and spleen after therapy. Cells from weeks 1 and 4 were purified using human anti-CD45 antibody and analyzed for gene expression using RNA-seq or Affymetrix Human Gene 1.0 ST arrays. All studies were conducted in accordance with the guidelines of and approved by the Institutional Animal Care and Use Committee at the institution where the work was performed.
In Vivo Leukemia Repopulation of Senescence-Like Cells
Untreated and treated AML cells with Ara-C for three consecutive days were stained with C12-FDG on the following day. Cells were washed and stained with sterile PI for cell viability discrimination. Cells were sorted with a FACSAria II system (BD) into FBS and centrifuged after collection was finished. Cell pellet was washed and resuspended in plain IMDM. Cells were transplanted into female NSG mice via tail-vein injection of 500,000 cells per condition per mouse.
Colony-Forming Cell Assay
After sorting, 5,000–25,000 cells/well of primary patient-derived AML were seeded in MethoCult (H4534) medium (STEMCELL Technologies) supplemented with cytokines (36). The MethoCult medium–cell mixture was plated into six-well plates with water supply in the inter-well chamber. Additionally, the six-well plates were placed inside a tub with extra water dishes to prevent evaporation. After three weeks of incubation at 37°C in 5% CO2, pictures were acquired using ChemiDoc Touch Imaging System (Bio-Rad) or STEMvision (STEMCELL Technologies) and colonies were counted using ImageJ software.
Lysates from leukemia cells were lysed using PBS containing 1% NP-40 and 1 mmol/L DTT supplemented with 1% protease/phosphatase inhibitor cocktail (cOmplete, Roche; Phosphatase Inhibitor Cocktail A, Santa Cruz Biotechnology) and were sonicated for 15 seconds using a tip sonicator (Branson Sonifier 450). Protein lysates were resolved by SDS-PAGE, transferred to PVDF membrane, and probed with the indicated primary antibodies: CHK1 (clone: 2G1D5, Thermo Fisher Scientific), phospho-CHK1 Ser345 (clone: 133D3; Cell Signaling Technology), p53 (SC-126, Santa Cruz Biotechnology), p16 (ab108349, Abcam), or β-actin (Sigma). Membranes were then incubated with a peroxidase-conjugated correspondent secondary antibody and detected using Pierce ECL chemiluminescent substrate (Thermo Fisher Scientific).
Depletion of Apoptotic Cells
When cell viability decreased below 50%, a depletion step was introduced for all relevant assays except for flow cytometry. Dead cells and cellular debris were removed via magnetic column depletion using the cell death removal kit (Miltenyi Biotec) according to the company's instructions.
Single-Cell Organoid AML Recurrence
Individual clones were established before and after killing the bulk leukemia following chemotherapy exposure. Briefly, parental and nadir-collected AML cells were serially diluted to a final concentration of one cell per two wells in a 96- or 384-well plate. Each well was checked and monitored that contained a single cell. Single cells were cultured using the ex vivo platform as previously described (36). Recovered clones were picked from the wells and subcultured in larger wells.
5-mC Immuno Dot Blot
Genomic DNA was extracted using the Purelink Genomic DNA Kit (Thermo Fisher Scientific) and quantified via NanoDrop (Thermo Fisher Scientific). Dilution of DNA samples was denatured with 0.1 mol/L NaOH at 95°C for 5 minutes. Subsequently, samples were rapidly chilled for 5 minutes on ice, spun down, and neutralized with 0.1 volume of 6.6 mol/L ammonium acetate. DNA samples were spotted onto a positively charged nylon membrane (Immobilon-NY+, EMD Millipore), air-dried for 15 minutes and UV cross-linked for 5 minutes. Membranes were blocked in PBS containing 0.05% Tween-20 and 5% nonfat milk powder (PBS-TM) for 1 hour at room temperature (RT). After blocking, membranes were washed in PBS-T for 3 × 5 minutes at RT and incubated with the 5-mC antibody (NA81, EMD Millipore) in a 1:1,000 dilution in PBS-TM for 1 hour. Subsequently, membranes were washed with PBS-T for 3 × 5 minutes at RT and incubated with a secondary antibody for 1 hour. After final washing, membranes were incubated in Pierce ECL chemiluminescent substrate (Thermo Fisher Scientific).
Single-Cell RNA-seq Library Preparation
AML cells were sorted with a FACSAria II system (BD) into FBS and washed with PBS after collection was finished. Sorted viable untreated AML cells and Ara-C–induced senescent AML cells were processed together on the same day on a 10 × Chromium instrument (10× Genomics) to generate barcoded single-cell gel beads-in-emulsion (GEM) according to the manufacturer's protocol. scRNA-seq libraries were prepared using the Chromium Single-Cell 3′ v3 Reagent Kit (10× Genomics) following the directions of the manufacturer. Briefly, the initial step consisted in performing an emulsion where individual cells were isolated into droplets together with gel beads coated with unique primers bearing 10× cell barcodes, unique molecular identifiers (UMI), and poly(dT) sequences. GEM-Reverse Transcription (53°C for 45 minutes, 85°C for 5 minutes; held at 4°C) was performed in a C1000 Touch Thermal cycler with 96-Deep Well Reaction Module (Bio-Rad). After reverse transcription reaction, GEMs were broken and the single-strand cDNA was cleaned up with DynaBeads MyOne Silane Beads (Thermo Fisher Scientific). The cDNA was amplified for 12 cycles (98°C for 3 minutes; 98°C for 15 seconds, 63°C for 20 seconds, 72°C for 1). Quality of the cDNA was assessed using an Agilent Bioanalyzer 2100, obtaining a product of about 1,400 bp. This cDNA was enzymatically fragmented, end repaired, A-tailed, subjected to a double-sided size selection with SPRIselect beads (Beckman Coulter), and ligated to adaptors provided in the kit. A unique sample index for each library was introduced through 12 cycles of PCR amplification using the indexes provided in the kit (98°C for 45 seconds; 98°C for 20 seconds, 54°C for 30 seconds, and 72°C for 20 seconds × 14 cycles; 72°C for 1 minute; held at 4°C). Indexed libraries were subjected to a second double-sided size selection, and libraries were then quantified using Qubit fluorometric quantification (Thermo Fisher Scientific). The quality was assessed on an Agilent Bioanalyzer 2100, obtaining an average library size of 455 bp. Libraries were pooled and normalized to 2 nmol/L and clustered on a HiSeq4000 high output mode on a paired-end read flow cell and sequenced for 28 cycles on R1 (10× barcode and the UMIs), followed by 8 cycles of I7 index (sample index), and 98 bases on R2 (transcript), with a coverage around 130M reads per sample. Primary processing of sequencing images was done using Illumina's Real-Time Analysis software.
Single-Cell RNA-seq Data Analysis
Single-cell expression was analyzed using the Cell Ranger Single Cell Software Suite (v3.0.2) to perform quality control, sample demultiplexing, barcode processing, and single-cell 3′ gene counting (77). Sequencing reads were aligned to the refdata-cellranger-hg19-3.0.0 transcriptome using the Cell Ranger suite with default parameters. Single cells from control and Ara-C–treated conditions, a total of 6,278 single cells with a mean of ∼3,200 genes per cell, were imported into Seurat 3 package (78) in R statistical software. Only genes detected in at least two cells were considered for further analysis. Apoptotic cells were filtered out by any cell containing >15% mitochondrial UMI counts. To detect changes in gene signatures more robustly, any cell with <1,500 genes was filtered out. To reduce outlier effects, cells with a total count of >40,000 UMIs or >6,000 genes per cell were filtered out. Overall, a total of 4,627 cells were kept for downstream analyses. After merging control and Ara-C treated cells using the merge function, cell expression levels were normalized and scaled using default settings in Seurat. Principal component analysis for dimensionality reduction was then run on the normalized gene–barcode matrix. UMAP (51) embedding was applied to visualize the cells in the two-dimensional space after selecting the first 20 principal components for the clustering analysis with a resolution parameter of 0.4. This graph-based clustering method relies on a clustering algorithm based on shared nearest neighbor (SNN) modularity optimization, which first calculates k-nearest neighbors to construct the SNN graph. Cluster-specific genes were identified by running the Seurat “FindMarkers/FindAllMarkers” function with the Wilcoxon rank-sum test. Signature scores were computed using the Seurat function “AddModuleScore” and the corresponding gene set. A positive score suggests that the module of genes in a particular cell is more highly expressed than randomly expected, given the average expression of this module across the population. To identify more robustly transcriptional programs for certain phenotypes (JAATINEN HSC; GAL LSC) that exhibited a large fraction (>1/3) of DEG both up- and downregulated, we collapsed the two sets into one score by taking the difference of the signature scores for up- and downregulated gene signatures. Gene signature sets were compiled from the MSigDB or elsewhere (described in the GSEA section below) and summarized in Supplementary Table S1. UMAP, dot, and violin/box plots for single-cell analysis were plotted using the Seurat package and R.
Single-Cell RNA-seq Analysis of Bone Marrow Samples of Patients with AML
Processed scRNA-seq data of bone marrow aspirates before and shortly after induction therapy (days 14–35, defined as nadir) including the annotation files were downloaded as tab-delimited text files from the Gene Expression Omnibus (GEO) series GSE116256 (ref. 63; Supplementary Table S8). The annotation file denoted the cell types for each cell and the expression matrix contained the number of UMIs for each gene (rows) and cell (columns). The files were converted into a Seurat object using the functions “CreateSeuratObject” and “AddMetaData” in the Seurat package. After normalization and scaling, gene enrichment scores were calculated using designated gene sets with the “AddModuleScore” function. For downstream analysis, we first focused on single cells composed of the following cell types in van Galen and colleagues (63): HSC, GMP, Prog, Prog-like, ProMono, Mono, Mono-like, ProMono-like, HSC-like, or GMP-like. To further focus on putative AML cells in the bone marrow, we used hematopoietic cells that carried AML-associated mutant genes identified by van Galen and colleagues (63), defined as: Prog-like, Mono-like, ProMono-like, HSC-like, or GMP-like cells.
Gene Expression of Persistent AML Cells in Patients
Gene-expression data were downloaded from the GEO series GSE75086 (Supplementary Table S8; ref. 4). AML samples were obtained from patients belonging to intermediate- and high-risk prognostic groups. Bone marrow aspirates of persistent disease were collected approximately three weeks following Ara-C-based chemotherapy treatment. Whenever leukemic blasts did not compose the majority of the MNCs, Boyd and colleagues (4) sorted leukemic blasts from both pre- and posttreatment cells based on CD45 intensity and side scatter gating.
High-quality RNA of treated samples was isolated using guanidinium thiocyanate-phenol-chloroform extraction (TRIzol; Thermo Fisher Scientific) with subsequent purification on silica-membrane columns using the Qiagen RNeasy Kit. Quality of RNA was validated using the 2100 Bioanalyzer system (Agilent). Samples were multiplexed and single-read 50 bp sequenced on the HiSeq 2500 sequencing system (Illumina). Reads were aligned to hg19 using the STAR aligner (79).
Bulk RNA-seq Analysis of Ex Vivo–Cultured AML Cells
Bulk RNA-seq of Clinical AML Samples at Diagnosis and Relapse
Collecting specimens and processing them for RNA-seq analysis is described in detail elsewhere (49). Information on patients that was previously published was deposited at GEO under GSE83533 and at dbGaP under accession number phs001027.v1.p1. Additional cases of previously unpublished AMLs used in this study were added to the same dbGaP number.
GSEA was performed using GSEA (34) software (http://www.broadinstitute.org/gsea/) and R software. Senescence, inflammation, diapause, mitotic cell cycle, and H/LSC-associated gene sets were obtained or compiled from previous studies (12, 15, 35, 55, 82) and the MSigDB (http://software.broadinstitute.org/gsea/msigdb/search.jsp). The diapause signature was defined by selecting DEG (2-fold log change and adjusted P < 0.05) in diapaused epiblasts versus preimplantation E4.5 epiblasts in Boroviak and colleagues (35). Compiled gene sets are listed in Supplementary Table S1. In general, given the lack of coherence in most expression data sets and the relatively small number of gene sets being analyzed, an FDR of 0.25 is reasonable in the setting of exploratory discovery. Published data samples used for data analysis are listed in Supplementary Table S8.
Gene Category Enrichment Analysis
Unsupervised pathway analysis was performed using the information-theoretic pathway analysis approach (83). Briefly, pathways that are informative about nonoverlapping gene groups were identified. Pathways annotations were used from the Biological Process annotations of the Gene Ontology database (http://www.geneontology.org). This pathway analysis estimates how informative each pathway is about the target gene groups and applies a randomization-based statistical test to assess the significance of the highest information values. We used the default significance threshold of P < 0.005. We estimated the FDR by randomizing the input profiles iteratively on shuffled profiles with identical parameters and thresholds, finding that the FDR was always less than 5%. For each informative pathway, we determined the extent to which the pathway was overrepresented in the target gene group, using the hypergeometric distribution.
Leading-Edge Network Analysis
GSEA was run on log2 FC ranking, correcting for individual AML cases using edgeR glm with TMM normalization (81, 84), using the Canonical Pathways and Hallmark signatures from MSigDB (34, 85). Leading-edge network analysis was performed using the Jaccard indices of leading edge of all pathways from GSEA with FDR <0.05.
Affymetrix Human Gene 1.0 ST arrays were processed with the oligo package in R using the RMA normalization method with CEL files derived from the primary AMLs collected on days 0, 8, and 29 post Ara-C treatment. IDs were annotated using the pd.hugene.1.0.st.v1 library. Differential regulation of genes was calculated for each case separately as the ratio of a sample at a given time point compared with the average of all time points.
Raw RNA-seq data of matched AML patient samples at diagnosis versus relapse were added to dbGaP accession number phs001027.v1.p1. Raw and processed data of patient-derived AML cells were deposited in GEO under accession number GSE146544.
C. Duy reports grants from Leukemia and Lymphoma Society (LLS), ECOG, and NCI/NIH during the conduct of the study. C. Meydan reports personal fees from Onegevity Health outside the submitted work. C.S. Mitsiades reports grants from NIH/NCI, LLS, and Ludwig Center at Harvard during the conduct of the study; other support from Takeda, personal fees from Fate Therapeutics and Ionis Pharmaceuticals, grants from Janssen/Johnson & Johnson, TEVA, EMD Serono/Merck KGA, AbbVie, Karyopharm, Arch Oncology, Sanofi, and Nurix outside the submitted work; and C.S. Mitsiades have been as coinventor in several patent applications pertinent to the broader theme of resistance of different types of neoplasias to various therapeutics. R.J. D'Andrea reports grants from Takeda Pharmaceuticals and other support from Calli Biotech outside the submitted work. C.E. Mason is a cofounder of Onegevity. A.M. Melnick reports grants from Janssen, Daiichi Sankyo, and Sanofi during the conduct of the study; personal fees from Epizyme, Constellation, BMS, and Exo-therapeutics, and other support from KDAC outside the submitted work. No other disclosures were reported.
C. Duy: Conceptualization, resources, data curation, formal analysis, supervision, investigation, visualization, methodology, writing–original draft, writing–review and editing. M. Li: Resources and validation. M. Teater: Data curation, investigation, and visualization. C. Meydan: Data curation and visualization. F.E. Garrett-Bakelman: Resources and data curation. T.C. Lee: Investigation. C.R. Chin: Visualization. C. Durmaz: Visualization. K.C. Kawabata: Validation. E. Dhimolea: Resources and investigation. C.S. Mitsiades: Resources and investigation. H. Doehner: Resources. R.J. D'Andrea: Resources. M.W. Becker: Resources. E.M. Paietta: Resources. C.E. Mason: Resources. M. Carroll: Resources, investigation, writing–review and editing. A.M. Melnick: Supervision, investigation, writing–review and editing.
We acknowledge the ECOG–ACRIN Cancer Research Group and the Weill Cornell Medicine Epigenomics Core for technical services. We thank Yaseswini Neelamraju for assistance with data submission to dbGap. This work is supported by the NCI of the NIH under the following award numbers: NIH/NCI R01CA198089 (A.M. Melnick and M. Carroll) and NIH/NCI UG1 CA233332 (A.M. Melnick). This work was supported by grants NIH R01 CA050947 (C.S. Mitsiades), CA196664 (C.S. Mitsiades), U01CA225730 (C.S. Mitsiades); Leukemia and Lymphoma Society (LLS) Scholar Award (C.S. Mitsiades); and Ludwig Center at Harvard (C.S. Mitsiades). The content is solely the responsibility of the authors and does not necessarily represent the official views of the NIH. This work was further supported by LLS SCOR 7013-17 (A.M. Melnick) and U10CA180820 (ECOG). C. Duy was supported by an LLS fellow award (LLS 5486) in partnership with The Jake Wetchler Foundation, and is a Forbeck scholar.