Fatty acid metabolism underlies venetoclax resistance in acute myeloid leukemia stem cells

Venetoclax with azacitidine (ven/aza) has emerged as a promising treatment regimen for acute myeloid leukemia (AML), with a high percentage of clinical remissions in newly diagnosed patients. However, approximately 30% of newly diagnosed patients and the majority of patients who have relapsed do not achieve remission with ven/aza. We previously reported that ven/aza efficacy is based on eradication of AML stem cells through a mechanism involving inhibition of amino acid metabolism, a pro- cess required in primitive AML cells to drive oxidative phosphorylation. Herein we demonstrate that resistance to ven/aza occurs via upregulation of fatty acid oxidation (FAO), which occurs either due to RAS pathway mutations or as a compensatory adaptation in relapsed disease. Utilization of FAO obviates the need for amino acid metabolism, thereby rendering ven/aza ineffective. Pharmacological inhibition of FAO restores sensitivity to ven/aza in drug-resistant AML cells. We propose inhibi- tion of FAO as a therapeutic strategy to address ven/aza resistance.

Outcomes for patients with AML have historically been poor and, despite extensive efforts, the development of improved therapies has proven difficult. In particular, therapeutically relevant targeting of disease-initiating leukemia stem cells (LSCs)1 has been a substantial challenge. Notably though, the BCL-2 inhibi- tor venetoclax was recently approved with a low-intensity chemo- therapy backbone for older patients newly diagnosed with AML, based on high response rates and durable remissions2. We showed that patients treated with venetoclax and the hypomethylating agent azacitidine led to significant decreases in the LSC population through a perturbation of LSC energy metabolism3,4. Specifically, ven/aza kills LSCs by decreasing amino acid uptake in this cell pop- ulation, resulting in decreased amino acid catabolism and a resul- tant decrease in oxidative phosphorylation (OXPHOS) in LSCs3.LSCs are uniquely dependent on OXPHOS5,6, and disruption of this process selectively targets the LSC population3,7–10.Despite promising outcomes with ven/aza in newly diagnosed AML4,11,12, subsets of patients are refractory to ven/aza and oth- ers relapse after initially responding. Therefore, characterizing and targeting the mechanisms that drive ven/aza resistance is important for improvement of AML therapy. Several characteristics of ven/aza resistance in AML have recently been described, including previous exposure to chemotherapy3,13, mutations in TP53 (ref. 14), changes in mitochondrial structure15, elevated expression of S100A8 and S100A9 (ref. 16), differentiation status17,18 and nicotinamide metabo- lism19. However, a conserved mechanism of resistance that can be targeted in patients with ven/aza-resistant AML has yet to be identified. The objective of our current study was to identify the underlying mechanism that mediates ven/aza resistance in LSCs, and to developmethods to therapeutically exploit such features. Specifically, using primary human LSCs we sought to determine: (1) what pathways are enriched in ven/aza-resistant LSCs, (2) whether these pathways could be used to identify patients with AML who are likely to be resistant to ven/aza and (3) whether targeting these enriched path- ways could resensitize resistant LSCs to ven/aza.

Results
Previous therapy and RAS mutations correlate with ven/aza resis- tance in patients with AML. We retrospectively evaluated 136 con- secutive patients with AML treated at the University of Colorado with ven/aza between January 2015 and October 2019. Of the vari- ables evaluated as potential predictors of response, the receipt of previous therapy (P = 0.0036), the presence of the PTPN11 muta- tion (P = 0.0348) and the presence of any RAS pathway gene muta- tion (PTPN11, KRAS, NRAS) (P = 0.0205) were the only significant variables (Supplementary Table 1). These factors also predicted shorter progression-free and overall survival (Fig. 1a,b).Ven/aza-resistant LSCs have higher levels of fatty acid metabo- lism. We next sought to determine whether ven/aza resistance observed in patients with AML could be modeled in vitro. In particular, we sought to evaluate the characteristics of the most primitive AML cell populations. To enrich for functionally defined LSCs, we employed labeling of specimens with CellROX, a fluo- rescent dye that measures cellular reactive oxygen species (ROS). We have previously shown that relatively low ROS levels enrich LSCs in specimens from both de novo and relapse/refractory (R/R) patients with AML5,20. Thus, for the purposes of this study, LSC sare defined as primary AML cells isolated by virtue of a ROS-low phenotype and absence of phenotypic characteristics found on normal lymphocytes. Since ven/aza activity is based on inhibition of OXPHOS3,4, we hypothesized that resistant cells would not show OXPHOS inhibition following drug treatment. As shown in Fig. 1, comparison of LSCs from specimens exhibiting differential ven/aza responsiveness demonstrates a direct correlation between viability (Fig. 1c) and OXPHOS inhibition (Fig. 1d).

All resistant specimens were derived from patients either newly diagnosed with RAS pathway mutations or those refractory or relapsed following exposure to conventional chemotherapy. As shown in Extended Data Fig. 2b, the presence of a PTPN11 mutation led to resistance to ven and ven/aza in vitro, corroborating the results shown in Fig. 1c. To further investigate the influence of RAS pathway mutations on metabolism, we used lentiviral gene transfer to transduce a primary AML specimen with PTPN11 E76A, representative of PTPN11 mutations found in mul- tiple patients with resistant disease from our cohort. Expression of PTPN11 E76A led to increased levels of basal respiration and gly- colysis (Extended Data Fig. 2c). In addition, ven/aza treatment of PTPN11 mutant cells failed to suppress oxygen consumption rate (OCR; Extended Data Fig. 2d), in agreement with the results shown in Fig. 1d.Because ven/aza inhibits OXPHOS in LSCs by decreasing amino acid uptake3, we sought to determine whether the ability to block amino acid uptake had been lost in drug-resistant LSCs. To examine this issue, LSCs from a sensitive AML specimen and three resistant AML specimens were treated with ven/aza for 4 h, incubated with stable isotope-labeled (SIL) amino acid for 1 h and intracellular SIL amino acids were measured by mass spectroscopy. As shown in Fig. 1e, SIL amino acids were comparably decreased in both sensi- tive and resistant LSCs.

These data demonstrate that ven/aza treat- ment retains the ability to inhibit amino acid uptake in resistant LSCs; however, unlike ven/aza-sensitive LSCs, this activity is not sufficient to decrease OXPHOS or induce LSC death. The absence of change in OXPHOS in the presence of decreased amino acid uptake suggests that resistant LSCs employ alterna- tive mechanisms to drive energy metabolism. Consistent with this hypothesis, LSCs from sensitive versus resistant specimens demon- strate significant differences in baseline metabolites as measured by liquid chromatography–mass spectrometry (LC–MS) spectro- photometry (Fig. 1f and Extended Data Fig. 1a). We note that three of the six resistant specimens shown in Fig. 1f have RAS pathway mutations, which we acknowledge may confer unique biology in comparison to non-RAS-resistant specimens. Nonetheless, because all resistant specimens share increased fatty acid metabolism, for the purposes of the present study we considered them as one group. To identify conserved metabolic changes that may contribute to drug resistance, a pathway analysis was performed on ven/aza-sensitive versus -resistant LSCs. This analysis revealed that the most upreg- ulated pathways include carnitine synthesis, fatty acid metabo- lism, fatty acid elongation in mitochondria and beta-oxidation of long-chain fatty acids (Extended Data Fig. 1b). Importantly, these findings are observed in LSCs that have RAS mutations and those that have been exposed to previous therapy. To further interrogate fatty acid metabolism in ven/aza resistance, we performed an unbi- ased lipidomics analysis21. Overall, no changes in global fatty acid levels were observed in ven/aza-resistant LSCs compared to ven/ aza-sensitive LSCs (Extended Data Fig. 2a). Furthermore, closer examination of individual fatty acids shows heterogeneity of lev- els between resistant and sensitive LSCs, with increased expression of polyunsaturated fatty acids in resistant LSCs (Extended Data Fig. 1c).

However, metabolites involved in fatty acid transport into the mitochondria were increased in ven/aza-resistant compared to ven/aza-sensitive LSCs, including l-carnitine and five of the six detected acyl-carnitines (Fig. 2a). These data suggest that increasedlevels of fatty acid transport may allow LSCs to compensate for loss of amino acid uptake following ven/aza treatment, resulting in ven/ aza resistance. Indeed, this is consistent with our previous studies demonstrating that relapsed LSCs increased FAO following amino acid depletion3. However, fatty acid transport into the mitochondria has not previously been described in ven/aza resistance.To determine whether ven/aza-resistant LSCs exhibited increased FAO, we performed SIL palmitate analysis in sensitive versus resistant LSCs (Fig. 2b–d). SIL palmitate levels were sig- nificantly increased and incorporated at significantly higher levels in multiple tricarboxylic acid cycle (TCA) metabolites, including citrate and malate (Fig. 2c). Further, SIL palmitate contributed to significantly higher levels of carnitines, including acetyl-carnitine, iso-butylryl carnitine and hexanoyl-carnitine, in ven/aza-resistant LSCs (Fig. 2d). Together, these data suggest that fatty acid transport plays a distinct role in LSC resistance to ven/aza and therefore may be a potential target to resensitization of resistant LSCs to ven/aza treatment.Inhibition of fatty acid metabolism resensitizes LSCs to ven/aza. Fatty acid metabolism within the mitochondria can be perturbed through multiple mechanisms.

Of particular interest in the context of AML studies, the BH3 protein MCL-1 has been shown to inter- act with very-long-chain acyl-CoA dehydrogenase (ACADVL) and regulate fatty acid metabolism22. Further, modulation of MCL-1 leads to decreases in oxidative phosphorylation23,24. Most impor- tantly, our recent studies have demonstrated that MCL-1 inhibi- tion can resensitize AML specimens to venetoclax18,25. We also saw altered levels of ACADVL in resistant LSCs (Extended Data Fig. 7d). Therefore, we hypothesized that fatty acid metabolism might be targeted by inhibition of ACADVL or MCL-1 in ven/ aza-resistant LSCs. To test this hypothesis, we first performed small interfering RNA-mediated knockdown of ACADVL in pri- mary ven/aza-resistant AML specimens and measured viability, colony-forming potential and OXPHOS with or without exposure to ven/aza. Knockdown of ACADVL (Extended Data Fig. 3) per se did not affect the viability or colony-forming potential of primary AML specimens (Fig. 3a,b). Further, as expected, ven/aza treatment per se had minimal effects on viability or colony-forming poten- tial (Fig. 3a,b). However, knockdown of ACADVL in combination with ven/aza restored ven/aza sensitivity with respect to viability and colony-forming potential. (Fig. 3a,b). Since ven/aza targets LSCs by decreasing OXPHOS, we next measured OXPHOS follow- ing knockdown of ACADVL in combination with or without 4 h of treatment with ven/aza. Although neither ACADVL knockdown nor ven/aza treatment per se significantly decreased OXPHOS, in combination they did decrease it (Fig. 3c).

These data suggest that ACADVL loss restores ven/aza sensitivity by decreasing OXPHOS. Based on the findings for ACADVL, we also investigated inhibi- tion of MCL-1 as a strategy to decrease OXPHOS and abrogate fatty acid metabolism in ven/aza-resistant LSCs. Inhibition of MCL-1 using small-molecule VU661013 (ref. 25) or a second MCL-1 inhibi- tor, S63845 (ref. 26), decreased viability and OXPHOS per se and in combination with azacitidine (Fig. 3d,e) in ven/aza-resistant LSCs. These data corroborate our previous findings18. We next determined whether MCL-1 inhibition would decrease OXPHOS by changing fatty acid metabolism. First, we measured steady-state metabolite levels following treatment with VU661013 plus azacitidine (VU/ aza). Carnitines and amino acids were decreased following MCL-1 inhibition (Extended Data Fig. 3e). Additionally, levels of TCA cycle intermediates, including citrate, malate, alpha-ketoglutarate and 2-hydroxyglutarate, were decreased following MCL-1 inhibi- tion (Extended Data Fig. 3d). These changes are specific to resistant LSCs. Measurement of metabolites in sensitive LSCs treated with VU/aza showed no decrease in these TCA metabolites or changes in carnitines (Extended Data Fig. 3d). Furthermore, the addition ofdown ACADVL in the presence of VU/aza did not significantly modify the effects on viability or OCR observed with either siRNA or drug treatment per se (Extended Data Fig. 3b,c).To further interrogate fatty acid metabolism following MCL-1 inhibition, we measured palmitate flux. SIL palmitate experiments demonstrated that MCL-1 inhibition resulted in decreased pal- mitate incorporation into TCA cycle intermediates (Fig. 3f,g and Extended Data Fig. 4c). The metabolites found to incorporate less palmitate following drug addition matched those found to beupregulated in resistant LSCs (Fig. 2).

These data indicate that ACADVL and its role in fatty acid metabolism can be targeted through MCL-1 inhibition.To investigate the biological consequences of FAO inhibition, the MCL-1 inhibitor VU661013 (ref. 25) was tested per se and in com- bination with azacitidine. Surprisingly, single-agent treatment withVU661013 was able to decrease the viability of ven/aza-resistant LSCs (Fig. 3d). These data were unexpected, as knockdown of ACADVL killed LSCs only in combination with ven/aza; therefore, we had anticipated that MCL-1 inhibition would target LSCs only in combination with venetoclax. Since ven/aza decreases amino acid levels, we determined whether MCL-1 inhibition could also decrease amino acid levels in LSCs. As shown in Extended Data Fig. 4a, treatment with VU/aza reduced global amino acid levels slightly. Further, amino acid flux into the TCA cycle was decreased following MCL-1 inhibition (Extended Data Fig. 4d). Overall, these data suggest that MCL-1 inhibition targets LSCs by decreasing both fatty acid and amino acid metabolism. To functionally assess effects on the LSC population, we treated ven/aza-resistant specimens with VU661013 and engrafted them into immune-deficient mice. The data demonstrate that MCL-1 inhibition significantly decreases engraftment of primary AML specimens (Fig. 3h and Extended Data Fig. 4e), suggesting that MCL-1 mediates energy metabolism processes in ven/aza-resistant LSCs.We next sought to determine whether inhibition of fatty acid transport into LSCs or into their mitochondria could decrease OXPHOS, and therefore kill ven/aza-resistant LSCs.

The fatty acid transporter CD36 and fatty acid mitochondrial transporters CPT1A and CPT1C were knocked down in primary ven/aza-resistant AML specimens using siRNAs. Knockdown of CD36, CPT1A and CPT1C (Extended Fata Fig. 5d) per se did not affect the viability or colony-forming potential of primary AML specimens (Fig. 4a,b). Further, as expected, ven/aza treatment per se had minimal effects on viability or colony-forming potential. However, knockdown of CD36, CPT1A or CPT1C in combination with ven/aza restored ven/aza sensitivity (Fig. 4a,b). Since ven/aza targets sensitive LSCs by decreasing OXPHOS, we measured OXPHOS following knock- down of CD36, CPT1A or CPT1C in combination with or with- out 4 h of treatment with ven/aza. Neither knockdown nor ven/aza treatment per se significantly decreased OXPHOS for any condition; however, the combination of knockdown and ven/aza decreased OXPHOS for all three pathways (Fig. 4c). Interestingly, there was no significant compensation between the two isoforms of CPT1, and knockdown of CPT1A and CPT1C together did not increase killing of resistant LSCs in the presence of ven/aza (Extended Data Fig. 5d). These data indicate that fatty acid transporter loss restores ven/aza sensitivity by decreasing OXPHOS.To assess the potential of pharmacological inhibition of fatty acid transport, we treated ven/aza-resistant primary AML specimens with CPT1 inhibitor etomoxir ± ven/aza and determined viability and OXPHOS levels. The combination of etomoxir with ven/aza reduced LSC viability (Fig. 4d) and OXPHOS (Fig. 4e).

Further, eto- moxir treatment reduced LSC viability (Extended Data Fig. 5e) and OXPHOS (Extended Data Fig. 5f) in combination with amino acid depletion, suggesting that the key activity of ven/aza in this drugcombination is reduction in amino acid levels. However, amino acid levels remained relatively unchanged with the combination of ven plus aza and etomoxir over venandaza separately (Extended Data Fig. 6a). In line with reductions in OXPHOS, measurement of TCA cycle metabolites in resistant LSCs showed significant decreases following treatment with etomoxir plus ven/aza compared to ven/ aza per se (Extended Data Fig. 6b). Further, reduced OXPHOS and viability could be accomplished with ven plus etomoxir without the addition of azacytidine (Extended Data Fig. 5j,k). Notably, the effects of etomoxir with ven/aza could be reversed by cotreatment with octanoic acid, a medium-chain fatty acid that does not require CPT1 for transport into the mitochondria, suggesting that the effect of etomoxir is due to its inhibitory effects on CPT1. The addition of octanoic acid in the presence of ven/aza plus etomoxir rescued both the viability of resistant LSCs and the reduction in OXPHOS (Extended Data Fig. 5g,h). In contrast, octanoic acid treatment per se did not have any effects on LSC viability following cytarabine treatment (Extended Data Fig. 5I), suggesting that the role of FAO is specifically related to ven/aza activityTo measure the effects of etomoxir with ven/aza on functional LSCs, a primary ven/aza-resistant AML specimen was treated ex vivo with etomoxir, ven/aza or their combination for 24 h and then transplanted into immune-deficient mice. The combina- tion treatment significantly reduced engraftment potential com- pared to ven/aza per se, demonstrating that etomoxir restores the ability of ven/aza to target functionally defined LSCs (Fig. 4f).

To model the therapeutic potential of etomoxir in combination with ven/aza, immune-deficient mice were transplanted with a primary AML specimen. At 6–8 weeks after establishment of the graft, ani- mals were treated for 2 weeks with etomoxir, ven/aza or etomoxir plus ven/aza. The combination therapy decreased leukemic burden within the bone marrow significantly better than either single ther- apy (Fig. 4g). Furthermore, phenotypically defined LSCs (CD45+/ CD34+/CD38–/CD123+) were significantly decreased in the triple drug combination group (Extended Data Fig. 6e). To determine whether etomoxir with ven/aza would affect normal hematopoi- etic stem and progenitor cells (HSPCs), mobilized peripheral blood specimens were treated with etomoxir, ven/aza or their combina- tion, and viability and colony-forming potential were assessed. Neither the single agents nor the combination decreased the viabil- ity or colony-forming potential (Extended Data Fig. 6c,d) of normal HSPCs, suggesting a therapeutic window to target LSCs without harming normal HSPCs. Lastly, addition of these drugs in vivo had no significant effects on mouse weight or liver pathology (Extended Data Fig. 6f,g).Prospective prediction of ven/aza resistance. Examination of the Beat AML dataset, which evaluates bulk primary AML speci- mens, reveals a link between mutations in the RAS pathway andfatty acid metabolism and oxidative phosphorylation gene sets (Fig. 5a).

Analysis of transcriptional profiles obtained from func- tionally derived LSCs from patients with relapse versus de novoincreased expression at relapse (Extended Data Fig. 7a–c). In addi- tion, the expression of multiple genes, including CD36 and CPT1a, correlates with decreased survival of patients with AML described in the The Cancer Genome Atlas (TCGA) dataset (Fig. 5c,d). Further, expression of CD36 and ACADVL at the transcript level is also higher in patients with the RAS mutant compared to wild type (WT; Extended Data Fig. 7e). Increased CD36 at the protein level was also found in resistant LSCs (Extended Data Fig. 7f). Together, these observations suggest that fatty acid metabolism plays a role in LSC therapy resistance and that such properties are evident before therapy.Finally, to understand the role of fatty acid metabolism in the response and progression of patients treated with ven/aza, we per- formed transcriptomic analysis on LSCs isolated from pretreatment baseline bone marrow specimens. Gene sets related to transla- tion, TCA cycle, lipid biosynthesis and fatty acid metabolism were increased in patients who progressed on ven/aza therapy (Fig. 5e). Gene set enrichment further confirms this finding, with increased enrichment scores for pathways involved in fatty acid metabolism (Extended Data Fig. 7g).

These data suggest that fatty acid metabolism can be transcriptionally profiled at initial diagnosis to predict the response to ven/aza of a patient with AML.To further investigate genes involved in fatty acid metabolism and ven/aza response at the single-cell level, we performed cellular indexing of transcriptomes and epitopes by sequencing (CITE-seq) on baseline samples of three patients who responded clinically to ven/aza and three who were refractory. The patients who were refractory showed enrichment for multiple genes involved in fatty acid metabolism, including CD36 and gene sets involved in fatty acid metabolism (Fig. 5f and Extended Data Fig. 7h,i). Expression of CD36 at the protein and transcript levels is higher in patients who were ven/aza sensitive versus patients who were ven/aza resistant, and corresponds to cells with the highest fatty acid metabolism gene set expression (Fig. 5f). These findings further suggest that transcriptional and surface phenotype at diagnosis can predict response to ven/aza.

Discussion
Although venetoclax-based regimens have shown promising clini- cal outcomes for patients with AML, therapy resistance is an emerg- ing problem. Recent studies have suggested that AML can develop resistance to venetoclax through various characteristics;3,14–17 how- ever, identifying conserved biological principles that contribute to therapeutic resistance has yet to be described. In this study we report two distinct mechanisms of therapeutic resistance: (1) relapsed or refractory disease status and (2) RAS pathway mutations. Mutations in RAS have previously been reported to mediate fatty acid metabo- lism in lung cancer29,30. Importantly, regardless of the mechanism by which ven/aza resistance is achieved, the ultimate consequence is increased fatty acid metabolism. Interestingly, TP53 mutations, which may result in suboptimal responses to venetoclax, have also been shown to increase the levels of many fatty acids in AML cell lines14. These data suggest that, instead of targeting each of the driv- ers of venetoclax resistance individually, a more widely applicable strategy to combat venetoclax resistance is potentially to target fatty acid metabolism.We propose two strategies to target fatty acid metabolism in ven/ aza-resistant LSCs. First, we demonstrate that inhibition of MCL-1 decreases fatty acid metabolism, probably through its relationship with ACADVL. Interestingly, MCL-1 inhibition per se or in com- bination with azacitidine decreases OXPHOS in LSCs, probably because MCL-1 inhibition also decreases amino acid metabolism.

MCL-1 inhibition is already under clinical investigation in AML because of its role in apoptosis25. Our study suggests that measur- ing the metabolic consequences of MCL-1 inhibition in AML cells isolated from patients treated with an MCL-1 inhibitor may also bea relevant biological endpoint to evaluate. The second strategy we propose is inhibition of fatty acid metabolism through targeting of CPT1. While etomoxir is known to have multiple off-target effects, our genetic (Fig. 4) and rescue data (Extended Data Fig. 5) suggest that the effects are probably due to the role of etomoxir in inhibiting CPT1. Unlike MCL-1 inhibitors, etomoxir is effective at targeting LSCs only in the presence of ven/aza, indicating that inhibition of both amino acid and fatty acid metabolism is needed to target ven/ aza-resistant LSCs. Interestingly, inhibition of fatty acid synthesis has been shown to potentiate the effects of the BCL-2/BCL-XL dual inhibitor, ABT-737, in AML through regulation of Bak-dependent mitochondrial permeability transition31. Further, inhibition of fatty acid uptake—specifically abrogation of CD36 function—sensitizes AML cells to conventional chemotherapy, including cytarabine32,33. These data suggest that increased fatty acid metabolism may be a more universal mechanism in the promotion of therapy resistance; therefore, identifying mechanisms to target fatty acid metabolism may have relatively broad implications for the treatment of patients with therapy-resistant AML.Our study also demonstrates that fatty acid metabolism may allow identification of patients who are likely to become (or be) resistant to ven/aza. Specifically, increased expression of genes involved in fatty acid metabolism correlates with a poor response to ven/aza.

Future studies designed to further evaluate the predictive potential of fatty acid gene expression are needed; however, if these findings are confirmed in a larger cohort of patients, it may be pos- sible to rationally determine which patients should receive ven/aza and which might benefit from other therapies. The other possibility is that a predictive tool may allow for rational trials designed to test whether addition of an MCL-1 inhibitor or etomoxir in addition to ven/aza up front may be beneficial for those patients predicted to have a poor response to ven/aza.In summary, these findings identify conserved biological prop- erties that contribute to ven/aza resistance in LSCs isolated from patients with primary AML. We also describe two strategies used to target ven/aza resistance and propose a predictive model to deter- mine which patients may be resistant to ven/aza therapy.Human specimens. Human AML specimens were obtained from patients via apheresis product, peripheral blood or bone marrow; mobilized peripheral blood was obtained from healthy donors. All patients gave informed consent for sample procurement on the University of Colorado tissue procurement protocol. The University of Colorado Institutional Review Board approved the retrospective analysis. See Supplementary Table 2 for additional details on human AML specimens.Baseline variables of interest (age, sex, race, previous treatment for AML, previous treatment for myelodysplastic syndrome (MDS), previous treatment for AML or MDS, relapsed status, refractory status, Southwest Oncology Group (SWOG) cyto category, complex status, monosomal status, flt3 mutation status, npm1 mutation status, idh mutation status, asxl1 mutation status, ptpn11 mutation status, RAS mutation status, RAS pathway mutation status, tp53 mutation status and European Leukemia Network (ELN) risk group) were summarizedfor all patients.

We then explored four general categories of outcome among patients—with definitional variations extending the total number of outcomes explored to nine: (1) response status (two definitions): complete remission (CR)/ complete remission with incomplete hematologic recovery (CRi)/partial respone (PR)/morphologic leukemia free state (MLFS) versus no response and CR/CRi versus PR/MLFS/no response; (2) refractory outcome (four definitions): refractory by ELN, refractory by ELN or by no evaluation because of disease, refractory by no CR/CRi, and refractory by CR/CRi or by no evaluation because of disease;(3) relapse in responders (CR/CRi/PR/MLFS); and (4) survival (two definitions): progression-free survival and overall survival. The relationship between baseline predictor variables and outcome variables was then investigated in two ways. For the binary outcome variables of response status, refractory outcome and relapse in responders, univariate logistic regression models were used to assess evidence of association between predictors and outcomes. Odds ratios and their 95% confidence intervals, as well as accompanying P values, were calculated for each variable. For survival outcomes, median progression-free survival and overall survival were calculated using Kaplan–Meier survival methods. To assess theeffect of the various predictors on survival outcomes, univariate Cox regression models were performed and hazard ratios and their 95% confidence intervals and accompanying P values were calculated.

The threshold level of significance in univariate logistic regression and Cox regression models was set at 0.05a priori. To explore the simultaneous effect of predictors on the various outcomes of interest, follow-up multivariate models were run: multivariate logistic regression models in the case of response status, refractory outcome and relapse in responders, and multivariate Cox regression models for survival outcomes.For each multivariate model, any predictor that achieved a univariate regression P ≤ 0.10 was included in the initial multivariate model for the outcome. Any predictor then achieving P ≥ 0.20 in this initial multivariate model was included in the final multivariate model for the outcome, where odds ratios and their 95% confidence intervals and P values were calculated. Analyses were performed using SAS v.9.4 (SAS Institute).Human specimen culture. When culturing was required, all samples were cultured in a base medium of MEM without amino acids and 5.5 mM glucose (My BioSource, no. MBS752807) supplemented with physiologic levels of amino acids (Carolina, no. 84-3700) and 10 nM human cytokines SCF (PEPROTech, no. 300-07), IL3 (PEPROTech, no. 200-03) and FLT3 (PEPROTech, no. 300-19) aspreviously described3. In addition, the medium was supplemented with low-density lipoprotein (Millipore, no. 437744), BIT (Stem Cell Technologies, no. 09500),β-ME (Gibco, no. 21985-023) and penicillin/streptomycin.Cell sorting and flow cytometry. Primary AML specimens were thawed, stained with CD45 (BD, no. 571875; dilution 1:40) to identify the blast population, CD19 (BD, no. 555413; dilution 1:20) and CD3 (BD, no. 557749; dilution 1:40) to exclude the lymphocyte populations, DAPI (EMD Millipore, no. 278298; dilution 500 nM) and CellROX deep red (Thermo Fisher, no. C10422; dilution 5 uM) and sorted using a BD FACSARIA.

ROS-low LSCs were identified as those with the 20% lowest ROS levels, and ROS-high blasts were identified as those with the highest 20% ROS levels, as previously described3–5,20. Flow cytometry was carried out on populations utilizing CD36 (BD, no. 555455; dilution 1:20).Global ultra-high-pressure liquid chromatography–mass spectrometry (UHPLC–MS) metabolomics. Approximately 100,000–500,000 ROS-low LSCs were sorted, and metabolomics analyses were performed via UHPLC–MS (Vanquish and Q Exactive, Thermo Fisher) as previously reported34. Briefly, cells were extracted in ice-cold methanol/acetonitrile/water (5:3:2 v/v/v) at a concentration of 2 million cells ml–1 of buffer. After vortexing for 30 min at 4 °C,samples were centrifuged at 12,000g for 10 min at 4 °C and supernatants processed for metabolomics analyses. Ten microliters of sample extracts was loaded onto a Kinetex XB-C18 column (150 × 2.1 mm2 interior diameter, 1.7 μm; Phenomenex). A 5-min gradient (5–95% B, phase A: water plus 0.1% formic acid and phase B: acetonitrile with 0.1% formic acid for positive ion mode; 0–100% B, phase A: 5% acetonitrile plus 5 mM ammonium acetate and phase B: 95% acetonitrile plus 5 mM ammonium acetate for negative ion mode) was used to elute metabolites. Themass spectrometer, scanned in full MS mode at 70,000 resolution in the range 65– 975 m/z, 4-kV spray voltage, 45 sheath gas and 15 auxiliary gas, operated in negative and then positive ion mode (separate runs). Metabolite assignment was performed against an inhouse standard library, as previously reported35.Approximately 100,000–500,000 ROS-low LSCs were sorted and metabolomics analyses were performed via UHPLC–MS (Vanquish and Q Exactive, Thermo Fisher) as previously reported.

Cells were extracted in ice-cold methanol at a concentration of 2 million cells ml–1 of buffer. After vortexing for15 min at 4 °C, samples were incubated without shaking for 15 min at −20 °C. Samples were then centrifuged at 18,213g for 10 min at 4 °C, followed by 1:1 dilution with 10 mM ammonium acetate. Ten microliters of sample extracts was loaded onto an HSS T3 column (150 × 2.1 mm2 i.d., 1.8 μm; Acquity). A 17-min gradient (25–99% B, phase A: 75:25 water:acetonitrile plus 5 mM ammonium acetate and phase B: 50:45:5 isopropanol:acetonitrile:water plus 5 mM ammonium acetate) was used to elute lipid metabolites. The mass spectrometer scannedin full MS mode at 70,000 resolution in the range 150–1,500 m/z, 4-kV spray voltage, 45 sheath gas and 15 auxiliary gas, operated in negative ion mode. Metabolite assignment was performed against an inhouse standard library, as previously reported35.MIn total, 500,000 ROS-low LSCs were sorted and incubated with stable isotope substrates, including uniformly 100-µM 13C, 15N-labeled amino acids (Cambridge Isotope Laboratories, no. MSK-A2-US-1.2) or 100 µM [13C16]palmitic acid (Sigma-Aldrich, no. 705573) where indicated. Metabolomics analyseswere performed via UHPLC–MS using the 5-min method as described in Global ultra-high-pressure liquid chromatography–mass spectrometry (UHPLC–MS) metabolomics and previously36.Patient samples were sorted and cultured without amino acids or drugs for 24 h. Viability was assessed by annexin V and 7AAD staining, followed by flow cytometry.Normal granulocyte–colony-stimulating factor mobilized (from bone marrow) peripheral blood samples were thawed, cultured under the conditions indicated for 24 h and CD34+ (BD, no. 572577) and CD45+ (BD, no. 571875) double-positive percentages were quantified by flow cytometry (FACsCaliber, BD).

CFU assays. Primary AML samples or normal granulocyte–colony-stimulating factor mobilized (from bone marrow) peripheral blood samples were cultured with the indicated drugs for 24 h before plating in human methylcellulose (R&D systems, no. HSC003). Colonies were counted 10–15 d after initial plating.Leukemia stem cell function was assessed by measuring engraftment of primary AML specimens following overnight culture with the indicated therapies and transplanting into NSGS mice. Engraftment was measured by flow cytometry for human CD45+ cells. All animal studies were done at the University of Colorado under Institutional Animal Care and Use Committee– approved protocol no. 308. The University of Colorado is accredited by the Association for Assessment and Accreditation of Laboratory Animal Care (ALAC), abides by the Public Health Service (PHS) Animal Assurance of Compliance and is licensed by the United States Department of Agriculture.To determine the effect of ven/aza, etomoxir or their combination in vivo, NSGS mice were transplanted with human AML leukemia cells (2 million per mouse) via tail vein injection . Six weeks after transplantation, mice were treated with either (1) 5 d on/2 d off with 100 mg kg–1 venetoclax and 3 d of 3 mg kg–1 azacitidine every other day, (2) 3 d per week every other day with 50 mg kg–1 etomoxir, (3) a combination of the two or (4) saline for 2 weeks, and the composition of residual leukemia cells within the bone marrow was then examined by flow cytometry.The extracellular flux assay kit XF96 (Agilent Technologies, no. 102417-100) was used to measure OCR. ROS-low LSCs were sorted, drug treated for 4 h and plated onto XF96 well plates. OCR was measured according to the manufacturer’s protocol and as previously described3,4.Protein lysates were loaded on a polyacrylamide gel.

Proteins were transferred to a polyvinylidene difluoride membrane using the minitrans-blot transfer system (Bio-Rad). To detect specific antigens, blots were probed with primary antibodies CPT1A (Cell Signaling, no. 12252; dilution 1:1,000), GAPDH (Santa Cruz, no. sc-32233; dilution 1:5,000) and CPT1C (abcam, no. ab123784; dilution 1:500) on a shaker at 4 °C, overnight, followed by 1 h of incubation at room temperature with horseradish peroxidase-conjugated secondary antibodies (Santa Cruz). Chemoluminescence was recorded using the automated Gel Doc XR system (Azure).Primary AML specimens were transfected with siRNA constructs targeting ACADVL, CPT1A, CPT1C, CD36 or a nontargeting scrambled siRNA (Dharmacon), following established protocols37. Specifically, 2 × 105 cells were electroporated using the Neon electroporator (Invitrogen) in Buffer T (R 1,600, V 10 ms, three pulses).Cells were prepared and stained according to the manufacturer’s protocol utilizing 19 TotalSeq A surface antibodies with oligo tags (Biolegend) stained for 30 min at 4 °C. Cells were loaded, with 3,000 captured via10x Genomics 3’ v.3 assay. Libraries were prepared as per the protocols of Biolegend and 10x Genomics and sequenced on a NovoSeq 6000 (Illumina). RNA sequencing data were initially processed using the Cell Ranger pipeline (v.3.1.0) from 10x Genomics. The count matrices thus generated were then analyzed using the Seurat package v.3.1.0 in R. Read counts were normalized to library size, scaled by 10,000, log transformed and filtered based on the following criteria: cells with <200 or >7,500 genes detected (to remove putative doublets) or with a proportion of unique molecular identifiers mapped to mitochondrial genes >0.25 were excluded from analysis; genes detected in fewer than fivecells were also excluded. Cell-to-cell variation in gene expression driven by the number of detected molecules and level of mitochondrial gene expression was regressed out using linear regression. Dimensionality reduction was performed with principal component analysis followed by Harmony alignment (https:// www.biorxiv.org/content/10.1101/461954v2) to remove batch effects.

Uniform manifold approximation and projection (UMAP) and graph-based clustering were performed using the first 20 Harmony components. Data are available at GEO.The TruSeq RNA Sample Preparation Kit V2 (Illumina) was used for next-generation sequencing library construction according to the manufacturer’s protocols. Amplicons were ∼200–500 base pairs in size. The amplified libraries were hybridized to the Illumina single-end flow cell and amplified using cBot (Illumina). Single-end reads of 100 nucleotides were generated for each sample and aligned to the organism-specific reference genome. Raw reads generated from the Illumina HiSeq2500 sequencer weredemultiplexed using configurebcl2fastq.pl v.1.8.4. Quality filtering and adapterremoval were performed using Trimmomatic v.0.32 with the following parameters: ‘SLIDINGWINDOW:4:20 TRAILING:13 LEADING:13 ILLUMINACLIP:adapters.fasta:2:30:10 MINLEN:15’. Processed/cleaned reads were then mapped to the UCSC hg19 genome build with SHRiMP v.2.2.3 at the following setting: –qv-offset 33 -all-contigs. Data are available at GEO.Statistical analyses were performed utilizing denoted tests in Graphpad Prism v.6.0 and v.7.0. Graphs and data are visualized as noted in figure legends. Detailed methods regarding survival graphs are included in Statistical analysis. Animal experiments were carried out based on previous power analysis and publications regarding AML xenograft models from groups utilizing the Lenth power calculator, and animal sample sizes of eight or greater based on expected s.d. for a two-sample t-test. Technical replicates from Seahorse experiments were excluded when outliers (pmol min–1 readings for OCR)were identified through an outlier analysis using Grubb’s test/electron-capture dissociation and technical issues with data collection, including cell loss. Seahorse experiments were done in technical replicates of five and multiple specimens from patients to account for technical issues with plates and collection of values. Values were removed in Figs. 4c and 5c and Extended Data Figs. 3c and 5b.Randomization was not applied to metabolism experiments or patients in the trial because this was a nonrandomized study. In animal experiments, the animals were randomized to injection and treatment groups. Investigators were not blinded to allocation during experiments and outcome assessment.

Reporting Summary. Further information on research design is available in the Nature Research Reporting Summary linked to this article. Patient-related clinical data not included in the paper were generated as part of a multicenter clinical trial (NCT02203773). A detailed description of the dose-escalation portion of the study has been published (Dinardo et al.)2. All DNA and RNA raw and analyzed sequencing data can be found at the GEO database and are available via accession numbers GSE156041 and GSE143363 (single-cell RNA-seq) and accession numbers GSE156008 and GSE155431 (bulk RNA-seq). The Beat AML dataset can be downloaded from the cBIOPortal (data listed in dataset denoted as Venetoclax OHSU, Nature 2018). Source data are provided with this paper. Other data supporting the findings of this study are available from the corresponding author on reasonable request.