C/EBPα Confers Dependence to Fatty Acid Anabolic Pathways and Vulnerability to Lipid Oxidative Stress–Induced Ferroptosis in FLT3-Mutant Leukemia

Acute myeloid leukemias (AML) are aggressive hematologic malignancies characterized by an aberrant proliferation of myeloid progenitors leading to an impairment of normal hematopoiesis. FLT3 mutations are detected in 30% of AML cases at diagnosis and are associated with adverse prognosis due to frequent relapse after intensive chemotherapy (1–4). Most FLT3 mutations are internal tandem duplication (ITD) in the juxta-membrane domain, whereas tyrosine kinase domain (TKD) point mutations are less frequently detected (1). Both ITD and TKD mutations result in constitutive activation of FLT3-dependent signaling pathways, supporting cell proliferation and survival through enhanced glucose, amino acid, and redox metabolism (5–9). Recently, FLT3 inhibitors (FLT3i) such as midostaurin or gilteritinib (GILT) were shown to improve therapeutic response in patients with FLT3-mutated AML when combined with intensive chemotherapy frontline or used as a monotherapy in the relapse setting (10, 11). However, resistance to FLT3i commonly appears, involving paracrine signaling and metabolic sensing from the bone marrow (BM) microenvironment (12–17), emergence of drug-resistant secondary FLT3 mutations, or clonal escape through activation of alternative pathways, as seen with RAS-activating mutations (17–20).

CCAAT-enhancer binding protein α (C/EBPα, encoded by the CEBPA gene) is a key transcription factor of myeloid differentiation (21). Different CEBPA mutations are detected in 5% to 9% of AML cases at diagnosis, with prognostic implications (1, 22). Moreover, CEBPA expression is downregulated in AML subtypes such as those harboring the RUNX1–CBF2T1 fusion gene, which participate in differentiation block, highlighting the importance of C/EBPα in leukemogenesis (23). Besides these hematopoietic functions, C/EBPα is a critical regulator of transcriptional programs related to glucose and lipid metabolism, promoting lipogenic and gluconeogenic gene expression programs in normal adipocytes and hepatocytes (24–26). However, the role of C/EBPα in cancer and leukemic metabolism has not been studied so far.

To investigate novel metabolic dependencies of highly metabolically active mutant FLT3 AML cells (5–9), we performed multiomics and functional approaches in vitro, in vivo using FLT3-mutant AML patient-derived xenograft (PDX) models, and in primary specimens from patients treated with GILT. We uncovered a novel role of C/EBPα on lipid biosynthesis downstream of FLT3. Specifically, we showed that FLT3 and C/EBPα are hierarchically activated in AML to induce the expression of fatty acid synthase (FASN), stearoyl-CoA desaturase (SCD), and fatty acid desaturase 2 (FADS2), promoting de novo biosynthesis of cytoprotective monounsaturated fatty acids (MUFA) and generation of highly unsaturated FA from the essential polyunsaturated FA (PUFA) linoleic acid precursor. Accordingly, inhibition of mutant FLT3 led to decreased C/EBPα and SCD expression and a relative increase in PUFA over MUFA, impacting the distribution of cellular lipid species. This sensitized leukemic cells to lipid peroxidation and induced their vulnerability to ferroptotic cell death. Finally, we demonstrated that ferroptosis induction synergized with GILT ex vivo and in vivo, indicating new therapeutic approaches for patients with FLT3-mutant AML.

To investigate changes in metabolic pathways of FLT3-activated AML cells, we analyzed differentially expressed genes (DEG) in FLT3-ITD–mutated MOLM-14 and MV4-11 cell lines, after FLT3i treatment or FLT3 depletion by RNA interference, and we defined an FLT3-ITD_UP cell line signature composed of 299 downregulated genes common to these conditions compared with the control cell lines (Supplementary Fig. S1A–S1D; Supplementary Table S1). This functionally defined signature was enriched in genes related to cell cycle and DNA/RNA processing, amino acid and redox metabolism pathways as previously described (5, 27–30), and to lipid metabolism including cholesterol and FA biosynthesis (Supplementary Fig. S1E; Supplementary Table S2). The same signatures were enriched in transcriptomes of FLT3-ITD–expressing cells from an MLL–AF9-driven murine AML model (GSE1639329; ref. 5; Supplementary Fig. S1F; Supplementary Table S2).

To generate FLT3-mutant models more relevant to patients with AML, we amplified primary leukemic cells from three different patients with AML (TUH06, TUH84, and TUH93) through serial transplantations in NSG mice (PDX^TUH06, PDX^TUH84, and PDX^TUH93; Fig. 1A; Supplementary Table S3). Using this approach, we enriched samples in leukemic blasts having a high variant allele frequency (VAF) of FLT3 mutation (Fig. 1A). After in vivo expansion in third recipients, we performed a transcriptomic analysis on human leukemic cells incubated ex vivo with FLT3i (Fig. 1A). Using gene set enrichment analysis (GSEA), we observed a depletion of gene signatures related to glucose, amino acids, mitochondria, and lipid metabolism in FLT3i-treated compared with control cells (Fig. 1B). Particularly, gene signatures related to lipid metabolism were strongly associated with FA and cholesterol biosynthesis, as observed in AML cell lines (Fig. 1B; Supplementary Table S4; Supplementary Fig. S1E). We further treated six FLT3-mutant PDXs in vivo with GILT after disease establishment (i.e., 8–50 weeks after xenotransplantation; Fig. 1C). The genomic landscape of these samples revealed frequent concurrent NPM1 and DNMT3A mutations, as expected (ref. 1; Fig. 1D; Supplementary Table S3). Moreover, we detected FLT3 D835Y and NRAS G12D mutations in TUH84 and TUH86, and TUH73 patient samples, respectively (Fig. 1D). After 1 week of treatment, GILT significantly reduced tumor burden in BM and spleen (SP) of four PDXs (Fig. 1E; Supplementary Fig. S2A–S2D). In PDX^TUH86 and PDX^TUH93, we observed significant tumor reduction in the SP but not in the BM after 7 days of GILT, whereas both compartments had decreased leukemia burden after 2 weeks of treatment in PDX^TUH93, defining early and late responders to GILT in vivo (Fig. 1D and E; Supplementary Fig. S2A–S2D). Notably, the presence of FLT3-TKD or NRAS mutations was not associated with delayed GILT response in these assays (Fig. 1D and E). We performed differential gene expression (DGE) analysis in human leukemic cells sorted from vehicle- and GILT-treated mice and showed a depletion of gene signatures related to glucose, amino acids, mitochondrial and lipid metabolism in the GILT-treated conditions (Fig. 1F; Supplementary Table S5). Among lipid-related gene signatures, FA and cholesterol biosynthesis pathways were the most prominently depleted by GILT treatment in vivo (Fig. 1F; Supplementary Table S5).

We next aimed to identify important effectors of lipid metabolism upon FLT3 inhibition in AML. To capture these variations, we performed a quantitative proteomic analysis in FLT3i-treated MOLM-14 and MV4-11 cells. First, we observed a downregulation of SCD involved in MUFA biosynthesis, and of fatty acid desaturase 2 (FADS2) involved in the biosynthesis of highly unsaturated FA from polyunsaturated FA (PUFA) precursors (ref. 31; Fig. 1G; Supplementary Table S6). Importantly, we found that the protein expression of transcription factor C/EBPα and C/EBPβ was decreased upon FLT3 inhibition (Fig. 1G; Supplementary Table S6), and we confirmed that FLT3i decreased the protein expression of C/EBPα by 38% in five FLT3-mutant AML PDX samples (Fig. 1H and I).

Next, we investigated the regulation of C/EBPα expression downstream of mutant FLT3 signaling. Focusing on samples having the 25% highest CEBPA expression in four independent cohorts of patients with AML (32–35), we observed an increased frequency of FLT3-ITD mutations (Supplementary Fig. S3A), but not of other signaling-related variants including FLT3-TKD, NRAS, or KRAS mutations compared with the lower quartile (Supplementary Fig. S3A–S3C). However, we found no clear association between FLT3-ITD status and CEBPA mRNA or protein expression in the same cohorts without filtering on high CEBPA mRNA expression (Supplementary Fig. S3D and S3E). To investigate protein stability after FLT3i treatment in AML cells, we annotated our proteomics data with a repository of quantified protein half-lives (36) and showed that proteins downregulated after quizartinib (QUIZ) treatment had significantly shorter half-lives compared with those increased after QUIZ (Supplementary Fig. S3F and S3G). These data suggested that QUIZ treatment induced a global inhibition of short-lived proteins.

Accordingly, we hypothesized that C/EBPα could have been subjected to a posttranslational regulation downstream mutant FLT3 signaling. C/EBPα could be phosphorylated downstream of FLT3-ITD activation (37, 38). We observed that treatment with FLT3i decreased both FLT3 signaling and C/EBPα phosphorylation at Ser-21 and also reduced the amounts of C/EBPα protein in MOLM-14 and MV4-11 cells (Supplementary Fig. S4A and S4B). Next, we hypothesized that upon FT3-ITD inhibition, unphosphorylated C/EBPα could be degraded by the ubiquitin/proteasome system. We conditionally expressed various CEBPA forms including untagged and tagged (HA) wild-type (WT), and phosphomimetic and nonphosphomimetic Ser-21 variants (S21D and S21A, respectively) in FLT3-ITD AML cell lines (Supplementary Fig. S5A and S5B). First, we observed that incubation with the proteasome inhibitor bortezomib led to the accumulation of HA-tagged C/EBPα after induction by doxycycline (dox; Supplementary Fig. S5C). Similarly, proteasome inhibition induced a significant and near-significant accumulation of WT or S21D forms, respectively, but had no impact on the expression of the S21A nonphosphomimetic variant of CEBPA (Supplementary Fig. S5D and S5E). This indicates that Ser-21 phosphorylation contributed to posttranslational regulation of C/EBPα expression in AML.

In FLT3-mutated PDX samples, we also confirmed that FLT3i decreased the expression of SCD and FADS2 as well as of sterol regulatory element-binding protein 1 (SREBP1) protein expression, which is a transcription factor involved in FA biosynthesis (ref. 39; Fig. 1H and I). In contrast, FLT3i decreased the mature form of SREBP1 but not of SREBP2 (known to be involved in cholesterol biosynthesis) in leukemic cell lines (Supplementary Fig. S6A and S6B). This highlights that FLT3 inhibition predominantly redirected lipid metabolism toward FA biosynthesis rather than cholesterol biosynthesis inhibition. Interestingly, FLT3i decreased the mRNA expression of SCD, FASN, and FADS2 but not the mRNA expression of CEBPA in primary AML samples and AML cell lines (Supplementary Fig. S7A and S7B). Furthermore, CEBPA overexpression resulted in a significant increase in these mRNA during GILT treatment compared with the empty vector (Supplementary Fig. S7C). The expression of SCD, FASN, and FADS2 was also increased in MOLM-14 cells expressing Ser-21–mutant compared with WT CEBPA forms (Supplementary Fig. S7D). Together, these results suggested that C/EBPα transcriptionally regulated the expression of FA biosynthesis regulators downstream of FLT3-ITD in AML.

To gain insight into the mechanisms whereby C/EBPα regulates lipid metabolism, we first analyzed public databases of C/EBPα chromatin immunoprecipitation (ChIP) sequencing and observed that C/EBPα more frequently occupied the promoter regions of SCD, FASN, SREBF1, and DHCR24 compared with randomly chosen genes (Supplementary Fig. S8A). Next, we generated various interfering RNA approaches to achieve C/EBPα depletion within the same range as observed after FLT3i treatment in terms of protein suppression and functional activity (Supplementary Fig. S8B–S8D). We performed and combined a bulk RNA sequencing (RNA-seq) and assay for transposase-accessible chromatin sequencing (ATAC-seq) after C/EBPα suppression by constitutive shRNA construct to compare genome accessibility based on C/EBPα expression in MOLM14 AML cells. We identified 31,707 peaks that were differentially accessible between shCTL and shC/EBPα MOLM14 FLT3-ITD AML cells (Bedtools intersect; Fig. 1J; Supplementary Table S7). We observed that C/EBPα depletion affected the chromatin accessibility within 10,860 genes, which include lipogenic genes such SCD, FASN, and FADS2 that are downregulated by C/EBPα invalidation and QUIZ treatment (Fig. 1K and L; Supplementary Fig. S8E; Supplementary Tables S7 and S8). As an example, several regions of open chromatin were identified in close proximity to the SCD locus in cells expressing C/EBPa while absent or decreased in shC/EBPα cells, thus confirming the role of C/EBPα on the regulation of SCD gene accessibility (Fig. 1M). These results established that C/EBPα expression confers a unique chromatin accessibility landscape to FLT3-ITD AML cells.

Collectively, this indicates that mutant FLT3 signaling sustains lipid metabolism through the key FA biosynthesis effector C/EBPα.

To study C/EBPα function in FLT3-mutant AML cells, we performed DGE analysis in MOLM-14 cells depleted in shCEBPA or overexpressing (OE) C/EBPα (CEBPA-OE) compared with their respective control (CTL) cells. We identified 198 significantly up- and downregulated genes after C/EBPα overexpression and depletion, respectively (Fig. 2A; Supplementary Table S9). This CEBPA_UP signature was enriched in genes related to the cell cycle, as well as glucose, FA, and cholesterol metabolism (Fig. 2B). Interestingly, we observed an enrichment of this CEBPA_UP signature in control compared with FLT3i-treated AML cell lines and primary AML cells, and in publicly accessible gene expression databases of FLT3-mutant compared with nonmutant AML cells (Fig. 2C, Supplementary Fig. S9A). Finally, we identified six genes in common between CEBPA_UP and FLT3-related gene signatures, and among them, four encoded key enzymes involved in FA (SCD and FASN) or cholesterol (DHCR7 and DHCR24) biosynthesis (Fig. 2D; Supplementary Table S10).

Next, we investigated intratumor transcriptional hetero­geneity and response to FLT3i in vivo using single-cell RNA-seq (scRNA-seq) on sorted human AML cells from vehicle- and GILT-treated mice of three FLT3-ITD PDX models (Fig. 2E). From bulk DNA sequencing in two of these PDXs, we observed that FLT3-ITD allele VAF were 46% and 81% in PDX^TUH93 and PDX^TUH110, respectively. This indicates that a majority of leukemic cells analyzed by scRNA-seq after in vivo exposure to GILT harbored an FLT3-ITD mutation (Supplementary Fig. S9B and S9C). Using Seurat and unsupervised hierarchical clustering analysis on each cell transcriptome, we uncovered distinct clusters defined by their transcriptional state (PDX^TUH93; Fig. 2F). The presence and proportion of cells in these states fluctuated between vehicle- and GILT-treated mice (Fig. 2G and H). First, we observed an enrichment of the FLT3-ITD_UP cell line signature in GILT-sensitive clusters from the vehicle conditions, highlighting the cells harboring an activation of FLT3-ITD (cluster #8 for PDX^TUH93; Fig. 2H and I; Supplementary Fig. S9D). This was associated with a concomitant enrichment of CEBPA_UP signature in the GILT-sensitive clusters in the vehicle conditions (Fig. 2J; Supplementary Fig. S9E). Interestingly, SCD, FASN, and FADS2 were coexpressed and enriched in cells with activation of FLT3-ITD in vehicle groups, and this cluster of coexpression was depleted upon GILT treatment (Fig. 2K). When independently applied to all PDXs, this consensus SCD coexpressed cluster was enriched in cells harboring an activation of both FLT3-ITD and CEBPA_UP signatures in vehicle groups (Fig. 2K and L; Supplementary Fig. S9F). Notably, enrichment in both the FLT3-ITD_UP cell line and CEBPA_UP signatures was not restricted to cells harboring a high expression of the MKI67 gene related to the cell cycle (Supplementary Fig. S9G and S9H). Similar results were observed for PDX^TUH84 (cluster #2) and PDH^TUH110 (cluster #3; Supplementary Figs. S10A–S10L and S11A–S11L).

Together, these transcriptional analyses showed the regulation of lipogenic gene expression by the FLT3–C/EBPα axis in FLT3-mutant leukemic cells.

We next investigated the role of C/EBPα in the antileukemic activity of FLT3i toward FLT3-mutant cells. First, we analyzed the transcriptomes of primary blasts from 21 patients with AML collected before (n = 20) and/or during (n = 17) GILT therapy [median duration: 54 (27–589) days], which include 16 paired pre-/post-GILT specimens (Fig. 3A; Supplementary Table S11). We observed a significant correlation between the CEBPA_UP signature and FLT3-ITD_UP signature and MKI67 expression across these samples (Supplementary Fig. S12A; Supplementary Table S12). Next, we stratified patient response based on the enrichment (AML patient group 1) versus the depletion (AML patient group 2) of CEBPA_UP signature after GILT treatment (Fig. 3A). We showed that FLT3-ITD–related signatures and MKI67 gene expression were upregulated in group 1 (Fig. 3B and C) and not modified in group 2 (Supplementary Fig. S12B and S12C). We further observed that all group 1 patients and the majority of group 2 patients displayed a low or high enrichment in the CEBPA_UP signature before GILT, respectively (Supplementary Fig. S12D). Interestingly, when focusing on group 1 patients, GSEA revealed an enrichment in FLT3-related signatures from independent studies (5, 8), along with MKI67 and CEBPA_UP signatures after the clinical failure of GILT (Supplementary Fig. S12E). These results suggested that CEBPA could be associated with clinical resistance to GILT in a subset of patients with FLT3-ITD AML.

Using dox-inducible shRNAs, we achieved a partial inhibition of C/EBPα expression in MOLM-14 cells (Fig. 3D; Supplementary Fig. S13A). Interestingly, the expression of enzymes involved into lipid metabolism, including SCD, FASN, and FADS2 decreased upon C/EBPα depletion (Fig. 3D; Supplementary Fig. S13A). In FLT3-ITD AML cell lines, C/EBPα depletion significantly reduced the sensitivity threshold to FLT3i (Fig. 3E; Supplementary Fig. S13B). Moreover, in three independent MOLM-14 cell line–derived xenograft (CLDX) assays, we observed that C/EBPα suppression significantly decreased leukemia burden and attenuated the surface expression of the differentiation markers CD14 and CD15 (Supplementary Fig. S13C–S13E) and markedly enhanced anti-AML efficacy of GILT in vivo (Fig. 3F and G). Finally, we harvested post-GILT residual disease from the PDX^TUH110 model and we transfected these cells with CTL or anti-CEBPA siRNAs ex vivo to perform secondary transplantation in NSG mice (Supplementary Fig. S13F). We observed that CEBPA suppression induced a 3-fold decrease in the propagation of leukemic cells into secondary recipient mice (Supplementary Fig. S13G). These results showed that C/EBPα participated in survival pathways downstream of FLT3-mutant signaling in AML.

Next, we observed that CEBPA-OE significantly induced SCD, FASN, and FADS2 expression in MOLM-14 cells, which was decreased after incubation of AML cells with FLT3i (Fig. 3H; Supplementary Fig. S13H). Notably, the cytotoxicity of FLT3i was significantly alleviated in FLT3-ITD AML cells overexpressing C/EBPα (Fig. 3I and J; Supplementary Fig. S13I). When transplanted into NSG mice, CEBPA-OE MOLM-14 cells showed a slight decrease in engraftment capacities and an increase in the expression of the differentiation markers CD14 and CD15 compared with the empty vector control (Fig. 3K). Finally, we overexpressed WT, S21A, or S21D CEBPA forms in MOLM-14 cells, and although the three isoforms significantly protected AML cells from FLT3i treatment, this effect was more marked in S21D-expressing AML cells (Fig. 3L; Supplementary Fig. S13J), suggesting that the inhibition of Ser-21 phosphorylation upon FLT3i impaired the prosurvival function of C/EBPα.

Collectively, these results highlight that C/EBPα through, at least in part, the regulation of lipid biosynthetic enzymes favors AML cell survival downstream of the mutant FLT3 signaling pathway.

To characterize lipid metabolism in FLT3-mutant AML, we performed global lipidomic analyses in MOLM-14 and MV4-11 cells treated with FLT3i. We observed a global decrease in glycerophospholipid (GP) categories, including phosphatidylethanolamines (PE) and phosphatidylcholines (PC), and an increased triglycerides (TG) content after FLT3i (Supplementary Fig. S14A; Supplementary Table S13). To evaluate if these lipid metabolic changes were driven by C/EBPα, we performed global lipidomic analyses in C/EBPα-depleted or C/EBPα-OE MOLM-14 cells treated with the FLT3i QUIZ or vehicle. Although C/EBPα suppression recapitulated the lipidomic profile observed upon FLT3i, C/EBPα-OE showed the opposite with increased PEs, PCs, and their derivatives [lyso-PCs (LPC) and lyso-PEs (LPE)] and decreased TG content, and strikingly this lipid distribution was abolished upon FLT3i (Fig. 4A; Supplementary Fig. S14B–S14H; Supplementary Table S13). As lipid droplets represent a reservoir for intracellular TG, we also performed electron microscopy, and we observed that FLT3 inhibition led to an increased number of lipid droplets in FLT3-mutant AML cell lines (Supplementary Information; Supplementary Fig. S15A and S15B). Similarly, C/EBPα depletion increased cellular neutral lipid content in MOLM-14 cells (Supplementary Fig. S15C). However, we observed a decreased expression of TG carriers including FABP5, SLC27A4, and SLC27A2 and an increase in FA transporters such as CD36 or SLC27A3 in leukemic cells after FLT3 or C/EBPα inhibition (Supplementary Fig. S15D). Accordingly, FLT3 or C/EBPα inhibition led to decreased lipid import through the plasma membrane in FLT3-mutant patient-derived cells and cell lines (Supplementary Fig. S15E). Finally, the sphingolipids ceramides (CER) and sphingomyelins (SM) were increased in FLT3-mutant cell lines after FLT3 or C/EBPα inhibition (Fig. 4A; Supplementary Fig. S14A–S14H). These results showed distinct regulation of glycerophospholipid and neutral lipids by the FLT3-ITD–C/EBPα axis in FLT3-mutant AML cells.

To follow the contribution of glucose and glutamine into FLT3-ITD–dependent FA synthesis, we performed untargeted lipidomic analysis using liquid chromatography coupled with high-resolution mass spectrometry (LC-HRMS) of MOLM-14 and MV4-11 incubated with FLT3i, and grown with stable uniformly labeled isotopologues of ¹³C₆-glucose (¹³C₆-Glc) or ¹³C₅-glutamine (¹³C₅-Gln; Fig. 4B). In this setting, incubation with FLT3i slightly decreased cell viability and density (Supplementary Fig. S16A). Moreover, the mean normalized enrichment of both ¹³C₆-Glc and ¹³C₅-Gln into newly synthetized FA was significantly decreased after FLT3i treatment (Fig. 4C; Supplementary Fig. S16B). Notably, isotopologue distribution showed a marked reduction of isotopes incorporation from 14:0 species, suggesting that fluxes of both glucose and glutamine oxidation to support MUFA biosynthesis were reduced upon FLT3 inhibition in FLT3-mutant leukemic cells (Fig. 4D; Supplementary Fig. S16C).

Collectively, targeted and isotope tracing lipidomics revealed that FLT3 regulated FA biosynthetic activities in FLT3-mutant AML cells in a C/EBPα-dependent manner.

FA biosynthesis mostly provides saturated fatty acids (SFA) and MUFAs, whereas PUFAs are available mostly by food intake and subjected to both desaturation and elongation through FADS and ELOVL, respectively (refs. 31, 40; Fig. 5A). SFAs and PUFAs are cytotoxic and can be stored into lipid droplets (40). Particularly, PUFAs form lipid complexes with reactive oxygen species (ROS), leading to lipid peroxidation products that trigger cell death (ref. 41; Fig. 5A). We hypothesized that inhibition of the FLT3–C/EBPα axis could induce changes in lipid metabolism that modify the membrane phospholipid distribution of SFA:MUFA:PUFA in FLT3-mutant AML cells (Fig. 5A). In lipidomic profiling experiments, we first observed a decreased incorporation of both SFAs and MUFAs into phospholipids, especially PC and PE after FLT3i or C/EBPα invalidation, along with an increase in PUFA levels (Fig. 5B and C; Supplementary Fig. S17A–S17E; Supplementary Table S14). On the contrary, C/EBPα-OE increased both SFAs and MUFAs into phospholipids that were abolished by FLT3i (Fig. 5B and C; Supplementary Fig. S17A–S17E; Supplementary Table S14). Accordingly, the PUFA:MUFA ratio in phospholipids was increased after FLT3 or C/EBPα inhibition and decreased after C/EBPα-OE in MOLM14 cells (Fig. 5D; Supplementary Fig. S18; Supplementary Table S15).

These results indicate that C/EBPα is essential to preserve lipid homeostasis through the generation of MUFAs and the control of PUFAs import and storage in FLT3-mutant AML cells.

As we showed that FLT3 inhibition increased cytotoxic PUFA and decreased protective MUFA, we next investigated whether this elevation of the PUFA/MUFA ratio that promotes membrane lipid peroxidation (42) could contribute to AML cell death upon FLT3 inhibition. We thus experimentally manipulated the PUFA/MUFA ratio via SCD inhibition or overexpression, glutathione peroxidase 4 (GPX4) inhibition, because this antioxidant enzyme specifically catalyzes the reduction of lipid peroxides using reduced glutathione (GSH; ref. 43) or direct MUFA (oleate) supply. We hypothesized that the PUFA/MUFA ratio will mechanistically link mutant FLT3 signaling and ferroptosis, a nonapoptotic form of cell death triggered by lipid peroxidation (ref. 41; Fig. 6A; Supplementary Fig. S19A).

In MOLM-14 and MV4-11 AML cell lines, the GPX4 inhibitor RSL3 (43) had antileukemic activity with IC₅₀ values of 86 nmol/L and 106.7 nmol/L, respectively, which was fully rescued by the synthetic antioxidant ferrostatin-1 (Fer-1; Supplementary Fig. S19B). As the dose–response curves of RSL3 were close to an “on/off” pattern, we arbitrarily defined noncytotoxic (100 nmol/L) and cytotoxic (250–500 nmol/L) RSL3 concentrations when we aimed to enhance or to attenuate the effects of RSL3 in AML cells, respectively. Next, we incubated MOLM-14 and MV4-11 cells with RSL3, and a MUFA (oleate) or vehicle (bovine serum albumin), and observed that oleate significantly rescued their viability, which correlated with reduced lipid oxidation in MOLM-14 and MV4-11 cells (Supplementary Information; Fig. 6B and C; Supplementary Fig. S19C). These results showed that ferroptotic cell death was actionable through GPX4 inhibition in FLT3-ITD AML cells.

To assess the specific role of mutant FLT3 signaling on lipid peroxidation (lipid-ROS), we used the Ba/F3 murine hematopoietic cell line constitutively expressing FLT3-ITD (Ba/F3-ITD, growth factor independent), which display exclusive addiction to FLT3-ITD signaling (44, 45), or nonmutated FLT3 allele (Ba/F3-WT, dependent on FLT3 ligand). We observed that RSL3 induced lipid-ROS formation in Ba/F3-ITD cells but not in Ba/F3-WT cells, which was markedly enhanced by coincubation with QUIZ, and attenuated by oleate supply (Supplementary Fig. S19D). In the same vein, we forced endogenous MUFA production by overexpressing SCD (SCD-OE), the rate-limiting enzyme of the last step of MUFA biosynthesis, in MOLM-14 and MV4-11 cell lines. We observed that SCD-OE significantly protected AML cells from FLT3i (Fig. 6D and E). This provided a functional validation that SCD acted downstream of FLT3-mutant signaling. Accordingly, we showed that SCD-OE markedly protected leukemic cells from lipid peroxidation and cell death induced by cytotoxic concentrations of RSL3 (Fig. 6F–H).

Finally, the depletion of SCD by two distinct lentivirally transduced shRNAs (Fig. 6I) markedly enhanced lipid peroxidation and cell death in FLT3-mutant AML cells when incubated with noncytotoxic concentrations of RSL3, as compared with their shCTL counterpart (Fig. 6J and K). Then, we used a small-molecule SCD inhibitor (MF-438; ref. 46), which demonstrated a moderate antileukemic activity as single agent and a synergistic effect when combined with FLT3i against FLT3-mutant AML cell lines and primary patient samples (Fig. 6L). In line with the effects of SCD depletion, lipid-ROS generation and cell death induced by noncytotoxic concentrations of RSL3 were markedly amplified by coincubation with MF-438 (Supplementary Fig. S19E and S19F).

These results showed that the relative PUFA/MUFA content determines the susceptibility of AML cells to ferroptotic cell death, which could represent a new vulnerability of FLT3-ITD leukemic cells.

We, therefore, wondered whether alterations in cellular lipid homeostasis induced by acute inhibition of mutant FLT3 signaling increased susceptibility to ferroptotic cell death, which could be exploited to define new therapeutic strategies for patients with AML. In FLT3-mutant cell lines, cell death observed after C/EBPα depletion was increased by RSL3, whereas C/EBPα overexpression partly protected leukemic cells from GPX4 inhibition (Fig. 7A). Moreover, although FLT3i or shFLT3 had a modest effect on lipid peroxidation, their combination with RSL3 markedly enhanced lipid peroxidation and cell death in FLT3-ITD cell lines (Supplementary Information; Fig. 7B and C; Supplementary Fig. S20A and S20B). Interestingly, Ba/F3-ITD cells had a limited sensitivity to RSL3 alone, but the incubation with QUIZ significantly sensitized Ba/F3-ITD cells but not Ba/F3-WT cells to RSL3 (Supplementary Fig. S20C). Next, we also observed a synergy between QUIZ and RSL3 in patient-derived AML cells ex vivo, and in MOLM-14 and MV4-11 cell lines (Fig. 7D; Supplementary Fig. S20D). Notably, we confirmed that FLT3i GILT increased lipid-ROS by 2.5-fold in vivo as assessed by anti–4-hydroxynonenal (4-HNE) staining in BM from three PDX models (Supplementary Fig. S20E). We observed that three other FLT3i (crenolanib, midostaurin, and sorafenib) enhanced the vulnerability of MOLM-14 and MV4-11 cells to RSL3 and that this effect was fully reverted by coincubation with Fer-1 (Supplementary Fig. S20F). On the other hand, the vulnerability to FLT3i was enhanced by two other GPX4 inhibitors FIN56 and ML210 that were abrogated by Fer-1 (Supplementary Fig. S20G and S20H). Finally, the SLC7A11 inhibitor erastin, which targets the antioxidant GSH thereby promoting ferroptosis (47), also synergized MOLM14, MV4-11, Ba/F3-ITD, or UT-7-ITD (but not Ba/F3-WT or UT-7-WT) cells to QUIZ treatment in vitro (Supplementary Fig. S20I). These results showed that FLT3 inhibition primed leukemic cells with mutant FLT3 oncogenic addiction to lipid oxidation and ferroptotic cell death.

We next investigated the cooperative lethality between redox stress induction and FLT3 inhibition in vivo. We depleted GPX4 by CRISPR/Cas9 in MOLM-14 cells (Supplementary Information; Supplementary Fig. S21A). As GPX4 knockout (KO) is lethal in our model, AML cells were grown in vitro with Fer-1 that efficiently prevented lipid oxidation and cell death in GPX4^KO, whereas Fer-1 removal promptly resulted in the opposed effect (Fig. 7E). Using a different approach by RNA interference, we similarly observed that GPX4 depletion induced leukemic cell death, which was further enhanced by FLT3i treatment (Supplementary Fig. S21B and S21C). We then performed CLDX assays using GPX4^KO or GPX4^WT (transduced with control CRISPR) MOLM-14 cells xenografted in immunodeficient NSG mice. To allow the engraftment of viable GPX4^KO AML cells, both GPX4^WT and GPX4^KO mice were treated with Fer-1 from the day of transplantation (Fig. 7F). After 10 days, Fer-1 was stopped at a stage of established disease, and mice were treated with GILT or vehicle for 1 week before the assessment of leukemia burden or followed-up for survival without further treatment (Supplementary information; Fig. 7F). We first observed an increased BM staining with anti–4-HNE in GPX4^KO mice, which was reinforced after GILT treatment (Fig. 7G). Moreover, human engraftment and total cell tumor burden were significantly reduced in GILT-treated GPX4^WT mice and vehicle-treated GPX4^KO mice, but residual disease was nearly eradicated in GPX4^KO mice treated with GILT (Fig. 7H and I). This marked tumor reduction was associated with an increased survival of GPX4^KO mice treated with GILT (Fig. 7J). Together, these results show that lipid redox stress induction primed FLT3-ITD AML cells to FLT3i in vitro and in vivo.

Finally, we and others previously showed that APR-246, a promising therapeutic agent in AML (48), induces AML cell killing (49, 50) by decreasing glutathione salvage and increasing lipid peroxidation (51–53). Using dose–response matrices, we observed a synergy between APR-246 and FLT3i on primary AML cells from three patients ex vivo (Supplementary Fig. S21D). Notably, APR-246 was also synergistic with other FLT3i, including QUIZ, midostaurin, and crenolanib in vitro, and enhanced lipid oxidation and cytotoxicity induced by shRNA-mediated FLT3 depletion in FLT3-ITD cell lines (Supplementary Fig. S21E–S21G). Next, we treated two established FLT3-ITD AML PDXs (PDX^TUH84 and PDX^TUH06) and MOLM-14 CLDX with GILT and/or APR-246 during 1 week (Supplementary Information; Supplementary Fig. S22A). APR-246 showed moderate activity alone, and a trend to enhanced activity of GILT in MOLM-14 CLDX experiments (Supplementary Fig. S22B). In PDX assays, we observed no significant APR-246 single-agent activity, but a significant increase in GILT effect by APR-246 in one PDX, which correlated to an increased lipid-ROS content (PDX^TUH84; Supplementary Fig. S22C–S22E). Notably, no toxicity was observed with APR-246 as assessed by quantification of murine hematopoietic cells and body weight (Supplementary Fig. S22F–S22H). Interestingly, APR-246 significantly decreased tumor growth in chick chorio-allantoic membrane (CAM)–based assays developed as an alternative MOLM-14 xenograft model, and this effect was moderately enhanced by coincubation with GILT (Supplementary Information; Supplementary Fig. S22H–S22J).

Collectively, these preclinical experiments uncover a sensitivity of FLT3-mutant AML cells to ferroptosis inducers that is unmasked by acute FLT3 inhibition in a C/EBPa–SCD-dependent manner, suggesting that this synthetic lethality could be exploited in future therapeutic strategies for patients with FLT3-mutant AML.

FLT3 is the most frequently mutated gene in AML, leading to an overactivation of signaling pathways downstream of this tyrosine kinase receptor. This results in a high proliferation rate of leukemic cells and frequent relapse after intensive chemotherapy. The practice-changing approval of FLT3i at different stages (first line, relapse, and post–allogeneic stem cell transplant maintenance) of AML therapy provides a significant survival gain for patients with AML (10, 11, 54). However, resistance to FLT3i is common and caused by multiple mechanisms including the selection of FLT3i-resistant clones harboring FLT3-TKD or RAS-activating mutations, or signaling or metabolic adaptations to FLT3i (19, 55, 56). The discovery of novel vulnerabilities is therefore an important goal in FLT3-mutant AML to uncover alternative therapeutic strategies targeting post-FLT3i residual disease and preventing the emergence of resistance.

In this study, we demonstrated that the biosynthesis of lipids (including FA, glycerophospholipids, sphingolipids, and triglycerides) is a critical vulnerability and dependency of FLT3-mutant AML mediated by the transcription factor C/EBPα. Among metabolic adaptations (5, 8), AML cells with FLT3-ITD mutations display activated FA oxidation and increased lysophospholipid amount (57). Moreover, CERs (a type of sphingolipids) are involved in cell-cycle arrest and apoptotic cell death in cancer models (58) and FLT3-ITD represses their biosynthesis in leukemia (59), suggesting that CER accumulation as observed in our lipidomic assays might contribute to leukemic cell death mediated by FLT3i. Furthermore, AML cell death induced by CER might be regulated by GPX4 activity, as previously described with the isoenzyme GPX1 (60), suggesting that intricate cell death modalities including apoptosis and ferroptosis could be triggered upon acute FLT3 inhibition dependent on impairment in lipid metabolism of mutant FLT3 AML cells. Interestingly, a high-fat diet promotes FLT3-dependent leukemia in the MLL–AF9 AML mouse model, highlighting a crucial role of lipid homeostasis with this oncogene (61). Interestingly, lipid dependency is not restricted to FLT3-mutant AML and is also observed in leukemia with other drivers such as oncogenic isocitrate dehydrogenase isoforms, which can redirect lipid metabolism toward FA oxidation to maintain mitochondrial oxidative phosphorylation by activating C/EBPα-driven transcriptional program (62). Using a multiomics approach in vitro and in vivo, we showed that C/EBPα is a novel essential regulator of FA biosynthesis through the activation of SCD, a rate-limiting enzyme for MUFA biosynthesis (63), downstream of FLT3-mutant receptors. This lipid dependency of FLT3-mutant AML was further observed at the single-cell level in vivo in PDX assays, highlighting the link between SCD, lipid peroxidation, ferroptosis priming, and post-FLT3i residual disease in vivo. In the same vein, acyl-CoA dehydrogenase involved in FA metabolism protects glioblastoma cells from toxic lipid oxidation, which favors their adaptive resistance to drug-induced stress (64, 65). Together, C/EBPα may account for the high reliance of leukemic cells on sustained FA catabolic (62) or anabolic (this present study) pathways.

Isotope tracing lipidomic experiments uncovered that FLT3 inhibition strongly reduced glucose and glutamine utilization to fuel FA biosynthesis, and that C/EBPα inhibition resulted in the same shift in lipid composition with decreased and increased amounts of phospholipids and neutral lipids, respectively. Based on our previous study (66), this accumulation of neutral lipids might be due to autophagy/lipophagy inhibition. Among lipid species, FLT3 and C/EBPα inhibition particularly reduced the amounts of MUFA through SCD inhibition along with increased PUFA import and storage as compensatory adaptation to FADS2 inhibition. Because imported PUFA are more susceptible to peroxidation, lipogenesis inhibitors increase the levels of peroxidation end products, render cancer cells more susceptible to oxidative stress–induced cell death, and act as chemotherapeutic sensitizers (67). Interestingly, periodic fasting-mimicking and calorie-restriction diets inhibit SCD, decrease MUFA, alter MUFA:SFA ratio, and affect tumor growth in mice (68, 69). Our study also showed that an increase in both PUFA:MUFA ratio and propensity to lipid oxidation as achieved by inhibition of the FLT3–C/EBPα–SCD axis represents a window of opportunity to trigger ferroptotic cell death (65). Accordingly, targeting the critical regulator of redox/ROS homeostasis GPX4 was synergistic with FLT3, C/EBPα, or SCD inhibition in FLT3-mutant AML cells.

Ferroptosis recently emerged as an iron-dependent, nonapoptotic form of regulated cell death particularly relevant to cancer therapy, with implications in response to chemotherapy, radiotherapy, and immunotherapy (70–72). Although we showed the potential of ferroptosis induction in FLT3-ITD AML in vivo using a genetic model of GPX4 suppression, currently available direct GPX4 inhibitors such as RSL3 showed limited selectivity and pharmacokinetic properties in vivo (73). In contrast, indirect GPX4 inhibition that might be achieved by inhibition of cyst(e)ine/selenocysteine import through SLC7A11 or LRP8 blockage could represent a promising approach to therapeutic ferroptosis induction in patients with cancer (52, 74). We also recently showed that the small-molecule APR-246 strongly induces ferroptosis through GSH depletion (52). Interestingly, APR-246 is now known as eprenetapopt and was successfully administrated to patients with myeloid neoplasms including AML in phase II clinical trials in association with azacitidine (48, 75). We, therefore, used APR-246 to trigger lipid oxidation in AML cells and observed a synergy between APR-246 and diverse FLT3i, including GILT, QUIZ, midostaurin, and crenolanib, compared with single agents in vitro. Moreover, APR-246 also enhanced GILT activity in vivo in FLT3-mutant AML PDX models. This further uncovered lipid oxidation as a new target in FLT3-mutant AML.

In summary, we describe a novel and protective role of C/EBPα from ferroptotic cell death in cancer through the transcriptional control of SCD-driven MUFA biosynthesis. A similar role of SCD was observed in cancer and ferroptosis (76). Our study also demonstrates that FLT3 inhibition causes metabolic changes in response to C/EBPα-SCD downregulation that induce lipid oxidation and therefore leukemia cell death, which is therapeutically actionable by ferroptosis inducers in AML cells (Fig. 7K).

Deidentified frozen primary specimens from patients with AML are from Toulouse University Hospital (TUH; Toulouse, France), Lyon University Hospital (Lyon, France), Barts Cancer Institute, Queen Mary University of London (London, UK), Peter MacCallum Cancer Centre, Royal Melbourne Hospital (Melbourne, Australia), and Perelman School of Medicine, University of Pennsylvania (Philadelphia, PA). A signed written informed consent for research use was obtained for each patient in accordance with the Declaration of Helsinki. Biobank collections are established according to the specific laws of each country, and transfer agreements for research applications were obtained after approbation by institutional review boards and ethics committees. Patient characteristics are summarized in Supplementary Tables S3 and S11. The cytogenetic and genomic landscape of these leukemia samples was obtained during routine diagnosis procedures. Samples from TUH were processed for large-scale genomic analysis using AmpliSeq for Illumina Myeloid Panel (Illumina), and FLT3-ITD mutations were also investigated by quantitative and fluorescent PCR (Eurofins Genomics), allowing the calculation of a ratio between ITD and WT FLT3 alleles.

MOLM-14 and MV4-11 human AML cell lines are short tandem repeat (STR; PCR-single-locus-technology, Promega PowerPlex21 PCR Kit, Eurofins Genomics) profiled periodically, and verified Mycoplasma negative yearly. MOLM-14 harbors a heterozygous FLT3-ITD mutation, whereas MV4-11 carries a homozygous FLT3-ITD mutation (77). AML cell lines were maintained at low passage number (8–10) in RPMI L-Glutamine (4 mmol/L) supplemented or minimum essential medium (MEM)-α glutamax supplemented with 10% fetal bovine serum (FBS, Invitrogen), 100 IU/mL penicillin and 100 μg/mL streptomycin (Life Technology) at 37°C with 5% CO₂, and split every 2 to 3 days before reaching confluence. After being thawed from the STR-authenticated batch, AML cell lines were passaged a maximum of 5 times (2 weeks) prior to use in experiments.

The Ba/F3 murine hematopoietic and UT-7 human AML cell lines were cultured in RPMI-1640 supplemented with 10% FBS, with 10 ng/mL murine IL3 (PeproTech, cat. # 213-13) or 2 IU/mL erythropoietin (Sigma, H5166), respectively. These cell lines are STR (PCR-single-locus-technology, Promega PowerPlex21 PCR Kit, Eurofins Genomics) profiled periodically, and verified Mycoplasma negative yearly. HEK 293T/17 (ATCC cat. # CRL-11268, RRID:CVCL_1926) cells were grown in Dulbecco's modified Eagle medium Glutamax (Life Technology) supplemented with 10% FBS, 100 IU/mL penicillin, and 100 μg/mL streptomycin (Life Technology).

Doxycycline (Sigma-Aldrich, cat. # D9891) was used at 1 μg/mL. RSL3 (cat. # HY-100218A), MF-438 (cat. # HY-15822), Fer-1 (cat. # HY-100579), GILT (cat. # HY-12432), and QUIZ (cat. # HY-13001) were all from MedChemExpress. Oleate [Oleic Acid-Albumin from bovine serum albumin (BSA), cat. # O3008] was from Sigma-Aldrich.

Animals were used in accordance with a protocol reviewed and approved by the Institutional Animal Care and Use Committee of Region Midi-Pyrenees (France). NOD/LtSz-SCID/IL-2Rγchain null (NSG) mice were produced at the Genotoul Anexplo platform in Toulouse, France, using breeders obtained from Charles River Laboratories. Mice (6–9 weeks old) were treated with busulfan (20 mg/kg for MOLM-14 and 30 mg/kg for PDXs) 24 hours before injection of leukemic cells. Leukemia samples were thawed at room temperature, washed twice in PBS1X, and suspended in Hank's Balanced Salt Solution (HBSS) at a final concentration of 0.2–10 × 10⁶ cells for PDXs and 1 × 10⁶ for cell line–derived xenograft (CLDX) per 200 μL of HBSS per mouse for tail-vein injection. PDX models: After engraftment (defined as >20/μL human leukemic cells in peripheral blood), mice were randomly assigned to the different treatment arms. PDXs were treated with vehicle or GILT after engraftment of human leukemic cells (8–18 weeks after AML cell transplantation). GILT was solubilized in water containing 0.5% methylcellulose 400cp (Sigma-Aldrich, cat. # M0262), mice were treated daily for 7 or 14 days by oral gavage with 30 mg per kg (mpk) GILT. CLDX models: MOLM-14 cells transduced with control or GPX4-targeting CRISPR-Cas9 guides were transplanted to NSG mice, which were treated from the day before transplant by intraperitoneal injection of 2 mpk Fer-1 for 10 days. Next, mice were treated with daily oral gavage with 10 mpk GILT or vehicle for 7 days.

NSG mice were euthanized in accordance with European ethics protocols. BM from tibias, femurs, and hips, and spleen were dissected and flushed in HBSS with 1% FBS. Mononuclear cells from BM and spleen were labeled with mCD45.1-PerCP-Cy5.5 (cat. #560580, RRID: AB_1727489), hCD45-APC (cat. #555485, RRID:AB_398600), and hCD44-PECy7 (cat. #560533, RRID:AB_1727483) or hCD33-PE (cat. #555450, RRID:AB_395843; all antibodies from BD Biosciences), and Annexin V–V500 (BD Biosciences; cat. #561501, RRID:AB_10694254) to determine the fraction of viable human blasts (AnxV-hCD45⁺mCD45.1−hCD44⁺ or AnnxVhCD45⁺mCD45.1−hCD33⁺ cells) using flow cytometry. Analyses were performed on a CytoFLEX flow cytometer with CytExpert software (Beckman Coulter, RRID:SCR_017217) and FlowJo 10.2 (Tree Star, RRID:SCR_008520). The number of AML cells/μL peripheral blood and the number of AML cells in total cell tumor burden (in BM and spleen) were determined by using CountBright beads (Invitrogen) using described manufacturer's protocol.

Human AML cells harvested from AML-xenografted mice were stained with Fixable Viability Stain 510 (BD Biosciences; cat. #564406, RRID:AB_2869572) and the following fluorescent-conjugated antibodies: hCD45-APCH7 (cat. #641399, RRID:AB_1645735) and hCD33-PE (cat. #555450, RRID:AB_395843; all antibodies from BD Biosciences) by BD FACSAria Fusion (Becton Dickinson) and FACSMelody (BD Biosciences).

Cells were thawed in a 37°C water bath and washed once in Iscove's modified Dulbecco's medium (IMDM) supplemented with 10% FBS and 10 μg/mL DNAse I. Next, cells were placed in IMDM glutamax supplemented with 20% of bovine serum albumin/insulin/transferrin (BIT) medium. BIT composition was: BSA (Sigma-Aldrich, cat. #A9418) 4 g/L, insulin 5 μg/mL (Sigma-Aldrich, cat. #I2643), and transferrin 60 μg/mL (Sigma-Aldrich, cat. #T8158), 50 ng/mL FLT3 ligand (cat. #300-19), 10 ng/mL IL6 (cat. #200-06), 50 ng/mL SCF (cat. #300-07), 25 ng/mL TPO (cat. #300-18), 10 ng/mL IL3 (cat. #200-03), and 10 ng/mL G-CSF (cat. #300-23; all cytokines from PeproTech), 100 IU/mL penicillin, and 100 μg/mL of streptomycin.

RNA was extracted using an RNeasy Mini Kit (Qiagen), and quality was evaluated with a Bioanalyzer 2100 (using Agilent RNA6000 nano chip kit), and 100 ng of total RNA was reverse transcribed using the GeneChip WT Plus Reagent Kit according to the manufacturer's instructions (Affymetrix, Thermo Fisher Scientific). Briefly, double strand cDNA was used for in vitro transcription with T7 RNA polymerase and 5.5 mg of Sens Target DNA was fragmented and labeled with biotin. cDNA were then hybridized to GeneChip Clariom S Human (Affymetrix) at 45°C for 17 hours, then washed on the fluidic station FS450 (Affymetrix), and scanned using the GCS3000 7G (Thermo Fisher Scientific). Scanned images were then analyzed with Expression Console software (Affymetrix, Thermo Fisher Scientific) to obtain raw data (.cel files) and metrics for quality controls.

Raw fluorescence intensity values were normalized using Robust Multiarray Average algorithm in R to generate the normalized data matrix by performing background correction, quantile normalization, and log₂ transformation of raw fluorescence intensity values of each gene. All quality controls and statistics were performed using Partek Genomics Suite software (Partek). Data were normalized using custom brainarray CDF files (v20 ENTREZG). To identify DEGs, we applied a classic analysis of variance (ANOVA) with a false discovery rate (FDR) permutation base for each gene. We created a matrix with only the significant ANOVA site and performed Z-scoring of rows. Hierarchical clustering by Pearson dissimilarity and average linkage and principal components analysis were conducted in an unsupervised fashion to control for experimental bias or outlier samples. We set a filter for those genes that displayed at least a ≥1.5- or ≤−1.5-fold difference in expression between groups and achieved an FDR of <0.05. Data were then interrogated for evidence of biological pathway dysregulation using the EnrichR user-friendly website (https://maayanlab.cloud/Enrichr/; refs. 78–80) and GSEA (Broad Institute). Enrichment rates were considered significant for P < 0.05

GSEA was performed using GSEA version 4.1.0 (https://www.gsea-msigdb.org/gsea/index.jsp, BroadInstitute, RRID:SCR_003199; refs. 81, 82). Gene signatures used in this study were from the Broad Institute database, literature, or in-house built. The following parameters were used: number of permutations = 1,000; permutation type = gene_set. Other parameters were left at default values.

RNA-seq library preparation was performed using Illumina TruSeq stranded mRNA protocol. The average number of reads per sample was 36 million. We mapped the reads to the human genome reference build hg38 using STAR aligner (STAR v2.6.1d; ref. 83). Detailed methods for RNA-seq quantification and DGE analysis are provided in the Supplementary Material.

Transcript-level counts and transcripts per million (tpm) were estimated using Salmon 1.5.2, with default parameters. R (R version 4.1.2) package tximport (version 1.20.0; ref. 84) was used to summarize transcript-level estimates to the gene-level estimates.

For plotting purposes, a variance stabilizing transformation [computed with the vst() function from DESeq2, version 1.34.0; ref. 85] was applied to the estimated counts from Salmon. As the samples were sequenced in different batches, the method removeBatchEffect (from the limma package, version 1.34.0; ref. 86) was further applied to remove the batch effect. For DGE analysis, raw counts were supplied to the R package DESeq2. As factors for the DGE analysis, we compared posttreatment versus pretreatment samples grouped according to the CEBPA signature expression level. Samples were labeled UP (or DOWN) if the CEBPA signature was upregulated (or downregulated) after treatment compared with the median of pretreatment samples. The sequencing batch number was included as a covariate to account for the batch effect. Hence, the resulting design formula supplied to DESeq2 was: treatment (post, pre) + CEBPA signature level (UP, DOWN) + batch. Post- versus pretreatment samples were compared for each CEBPA signature status (UP or DOWN). DEGs were defined as genes with Benjamini–Hochberg-adjusted Wald test P < 0.1. Genes were ranked according to their log₂ fold change in descending order and GSEA (81) was performed with the GSEA function from the clusterProfiler R package (version 4.2.2; ref. 87).

Cell pellets were resuspended in 100 μL of 0.1% RapiGest Surfactant (Waters) in 0.1M triethylammonium bicarbonate (TEAB). Samples were heated for 5 minutes at 95°C. Lysis was performed by sonication (6 × 30 seconds) at 70% amplitude and 0.5 pulse. Samples were kept for 30 seconds on ice between each cycle of sonication. Samples were centrifuged for 5 minutes at 17,000 × g. The remaining pellet was solubilized in 100 μL of 0.1% RapiGest, with 2 μL of 0.1M MgCl2 and 0.2 μL of benzonase for DNA digestion. Protein lysates were pooled together and protein concentration was measured by Bradford assay, and 25 μg of each sample was subjected to protein digestion as follows: sample volume was adjusted to 100 μL with 0.1M TEAB to obtain a final concentration of RapiGest 0.1%. Dithioerythritol (DTE; 2 μL) 50 mmol/L in distilled water was added, and the reduction was carried out at 60°C for 1 hour. Alkylation was performed by adding 2 μL of iodoacetamide (400 mmol/L in distilled water) for 1 hour at room temperature in the dark. Overnight digestion was performed at 37°C with 5 μL of freshly prepared trypsin (Promega; 0.1 μg/μL in TEAB 0.1M). To remove RapiGest, samples were acidified with TFA, heated at 37°C for 45 minutes, and centrifuged for 10 minutes at 17,000 × g. Supernatants were then desalted with a C18 microspin column (Harvard Apparatus) according to the manufacturer's instructions, completely dried under speed vacuum and stored at −20°C.

Samples were diluted in 25 μL of loading buffer (5% CH₃CN, 0.1% FA). Biognosys iRT peptides (1.5 μL) were added to each sample and 2 μL were injected into the column. Liquid chromatography–electrospray ionization–tandem mass spectrometry (LC-ESI-MS/MS) was performed on an Orbitrap Fusion Lumos Tribrid mass spectrometer (Thermo Fisher Scientific) equipped with an Easy nLC1200 liquid chromatography system (Thermo Fisher Scientific). Peptides were trapped on an Acclaim pepmap100, C18, 3 μm, 75 μm × 20 mm nano trap-column (Thermo Fisher Scientific) and separated on a 75 μm × 500 mm, C18 ReproSil-Pur (Dr. Maisch GmBH), 1.9 μm, 100 Å, home-made column. The analytical separation was run for 135 minutes using a gradient of H2O/FA 99.9%/0.1% (solvent A) and CH3CN/H2O/FA 80.0%/19.9%/0.1% (solvent B). Data-independent acquisition (DIA) was performed with MS1 full scan at a resolution of 60,000 (FWHM) followed by 30 DIA MS2 scan with variable windows. MS1 was performed in the Orbitrap with an AGC target of 1 × 10⁶, a maximum injection time of 50 ms, and a scan range from 400 to 1,240 m/z. DIA MS2 was performed in the Orbitrap using higher-energy collisional dissociation (HCD) at 30%. Isolation window was set to 28 m/z with an AGC target of 1 × 106 and a maximum injection time of 54 ms.

DIA raw files were loaded into Spectronaut v.14 (Biognosys) and analyzed by directDIA using default settings. Briefly, data were searched against the Human reference proteome fasta database (Uniprot, release 2020_09, 55471 entries). Trypsin was selected as the enzyme, with one potential missed cleavage. Variable amino acid modifications were oxidized methionine and deaminated (NQ). Fixed amino acid modification was carbamidomethyl cysteine. Both peptide precursor and protein FDR were controlled at 1% (q-value < 0.01). Single-hit proteins were excluded. For quantitation, the top 3 precursor areas per peptide were used, “only protein group specific” was selected as proteotypicity filter and normalization was set to “global normalization.” The quantitative analysis was performed with the MapDIA tool, using the precursor quantities extracted from Spectronaut output. No further normalization was applied. The following parameters were used: min peptides = 2, max peptides = 10, min correl = −1, Min_DE = 0.01, max_DE = 0.99, and experimental_design = replicate design. Proteins were considered to have significantly changed in abundance with an FDR ≤ 0.05 and an absolute fold change FC ≥ |1.5| (log₂FC ≥ |0.58|). FDR q-value was calculated using the two-stage step-up method of Benjamini, Krieger and Yekutieli (GraphPad Prism, RRID:SCR_002798). Proteins were excluded when more than one replicate was impaired per condition. If only one replicate was impaired, it was excluded and the protein was conserved.

Cells were lysed in NuPAGE LDS Sample Buffer supplemented with NuPAGE Sample Reducing Agent (Life Technologies) and heated for 5 minutes at 90°C. Then proteins were separated using 4% to 12% gradient polyacrylamide SDS–PAGE gels (Life Technologies) and electrotransferred to 0.2 μm nitrocellulose membranes (GE Healthcare). After blocking in Tris-buffered saline with 0.2% Tween and 5% BSA, membranes were blotted overnight at 4°C with the appropriate primary antibodies. Primary antibodies were detected using the appropriate horseradish peroxidase–conjugated secondary antibodies. Immunoreactive bands were visualized by enhanced chemiluminescence (Thermo Fisher Scientific, cat. # PI32209) with a Pxi camera (Syngene) using GeneSys software (Syngene, RRID:SCR_015770) or a fusion FX7 edge imaging System (Vilber) using Fusion FX software (Vilber). Protein levels were quantified using GeneTool software (Syngene) or Fusion FX software (Vilber) and normalized to nonvariable proteins such as Actin, GAPDH, or HSP90. Primary antibodies used were: Actin (cat. #MAB1501, RRID:AB_2223041) was purchased from Millipore; AKT (cat. #9272, RRID:AB_329827), phospho-AKT (Ser493; cat. # 9271, RRID:AB_329825), C/EBPα (cat. #8178, RRID:AB_11178517), FASN (cat. #3189, RRID:AB_2100798), FLT3 (cat. #3462, RRID:AB_2107052), phospho-FLT3 (Tyr591; cat. #3461, RRID:AB_331060), GAPDH (cat. #5174, RRID:AB_10622025), HSP90 (cat. #4874, RRID:AB_2121214), p44/p42 MAPK (ERK1/2; cat. #9102, RRID:AB_330744), phospho-p44/42 MAPK (ERK1/2; Thr202/Tyr204; cat. #9101, RRID:AB_331646), STAT5 (cat. #94205, RRID:AB_2737403), and phospho-STAT5 (Tyr694;cat. #9351, RRID:AB_2315225) were purchased from Cell Signaling Technology; FADS2 (cat. #ab72189, RRID:AB_2041175), GPX4 (cat. #ab41787, RRID:AB_941790), SREBP1 (cat. #ab28481, RRID:AB_778069), SREBP2 (cat. #ab30682, RRID:AB_779079), and SCD (cat. #ab19862, RRID:AB_445179) were purchased from Abcam.

We obtained human CEBPA, SCD, and FLT3, as well as CEBPA S21A and S21D, and FLT3-ITD mutant open-reading frames (ORF) through GeneArt gene synthesis technology (Thermo Fisher Scientific). CEBPA ORFs were cloned in pINDUCER21 (gift from Stephen Elledge and Thomas Westbrook, ref. 88); Addgene plasmid #46948; http://n2t.net/addgene:46948; RRID:Addgene_46948). SCD ORF was cloned in pLenti PGK Puro DEST 5W(w529–2; gift from Eric Campeau and Paul Kaufman, ref. 89); Addgene plasmid #19068; http://n2t.net/addgene:19068; RRID:Addgene_19068). FLT3 and FLT3-ITD ORFs were cloned in pLEX_307 (gift from David Root; Addgene plasmid #41392; http://n2t.net/addgene:41392; RRID: Addgene_41392).

We cloned shRNAs against SCD as well as control shRNA in the Tet-pLKO-puro vector (gift from Dmitri Wiederschain, Addgene plasmid #21915; http://n2t.net/addgene:21915; RRID: Addgene_21915) allowing the conditional expression of hairpins by doxycycline. Sequences of shRNAs are (5′ > 3′): shCTL, CAACAAGATGAAGAGCACCAA; shSCD#1, CGTCCTTATGACAAGAACATT; shSCD#2, GCACATCAACTTCACCACATT.

For FLT3 and CEBPA doxycycline-inducible invalidation, SMARTvector inducible human hEF1a-TurboGFP shRNA vectors were purchased from Horizon Discovery. FLT3: shFLT3: clone Id: V3IHSHEG_10441930, CTGGAGAATACCTACTTTT; nontargeting control hEF1a-TurboRFP, shCTL: Cat #VSC6573. CEBPA: pTRIPZ inducible vector (shCEBPA: RHS4696-200706732, clone ID: V2THS_14752, GGAAAGAATCAAGGAGG; shCTL: empty vector shRNA control, RHS4750).

For CEBPA constitutive invalidation, pLKO.1-puro constitutive vectors containing the following shRNA sequences were used (5′ > 3′): shRNA control, purchased from Sigma-Aldrich (SHC002 MISSION pLKO.1-puro nonmammalian shRNA control; CCGGCAACAAGATGAAGAGCACCAACTC); shRNAs against CEBPA, purchased from Sigma-Aldrich (SHCLNG-NM_004364, TRCN0000007306, and TRCN0000356198).

Briefly, we transfected HEK 293T/17 cells with different plasmids together with the packaging plasmids pMD2.G (gift from Didier Trono; Addgene plasmid #12259; http://n2t.net/addgene:12259; RRID:Addgene_12259) and psPAX2 (gift from Didier Trono; Addgene plasmid #12260; http://n2t.net/addgene:12260; RRID:Addgene_12260) encoding lentiviral proteins using Lipofectamine 2000 Transfection Reagent (Thermo Fisher Scientific). Twenty-four hours after cell transfection, the medium was removed, and the Opti-MEM culture medium (Life Technology) was added. HEK 293T/17 culture supernatants containing lentiviral particles were harvested 72 hours after transduction, filtered, and used for lentiviral infection of AML cells or stored at −80°C.

On the day of transduction, 2 × 10⁶ MOLM-14 or MV4-11 cells were seeded in a medium containing 8 μg/mL polybrene (Sigma-Aldrich, cat. #107689) and 2 mL of lentivirus-containing HEK 293T/17 supernatant. Three days after infection, transduced cells were selected using 1 μg/mL puromycin (LabForce) or sorted on an Astrios cell sorter (Beckman Coulter) dependent on the vector.

AML cells were seeded at 2 × 10⁵/mL in 100 μL of 10% FBS-supplemented RPMI prior to the addition of compounds. Cells were cultured in the presence of the test compounds for 14 to 96 hours at 37°C. Viability was quantified using the fluorescence-based Alamar Blue assay (Invitrogen, cat. #DAL1025). Alamar blue was added to each well in 10 μL aliquots. Fluorescence was then measured with a spectramax paradigm microplate reader from Molecular Devices. Fluorescence values were normalized to DMSO-treated controls for each sample or cell line investigated.

To investigate potential synergy, dose–response matrices of QUIZ and MF-438 or RSL3 were done and cell viability was calculated for every dose combination using the Synergy Finder 2.0 (https://synergyfinder.fimm.fi/; ref. 90) and compared with each agent alone. Calculations were based on the ZIP model. ZIP score interpretation: ZIP←10, antagonist effect; −10<ZIP<10, additive effect; ZIP>10, synergistic effect.

shCTL and shC/EBPα MOLM14 cells (in biological duplicates) were FACS-sorted, and ATAC-seq was performed as previously described (91) with minor modifications. Briefly, 50,000 cells were lysed in ice-cold lysis buffer, and the transposition reaction was performed using the Tn5 transposase at 37°C for 30 minutes. DNA was purified using the QIAGEN MinElute kit (QIAGEN). The libraries were prepared using Tagment DNA Enzyme and Buffer kits (Illumina), NEBNext High-Fidelity 2× PCR Master Mix (NEB, cat. #M0541S) with custom sequencing primers (92–95). The libraries were purified twice using AMPure XP beads (Beckman) following a double-sided protocol to remove primer dimers and large fragments. Libraries quality was assessed using the NGS High Sensitivity Kit on the Fragment Analyzer (Agilent Technologies). Only high-quality libraries were subsequently equimolarly pooled and sequencing was performed at the Pôle Technologique du CRCT – Plateau de Génomique et Transcriptomique (Inserm-UMR1037, Toulouse, France). The pool of libraries was quantified by qPCR using the KAPA Library Quantification Kit (Roche) to obtain an accurate quantification. Sequencing was then performed on one flowcell of the Illumina NextSeq 550 instrument (Illumina), using the NextSeq 500/550 High Output Kit v2.5 (150 cycles), and a paired-end 2 × 75 pb strategy. A minimum of 2 × 80 million raw reads were produced per sample.

Adapter trimming and quality controls were performed using Trimgalore v0.6.5 and FastQC v0.11.7. Reads were aligned to the hg38 reference genome by using Bowtie2 v2.2.9 with the splicing alignment feature switched off. The chromatin-accessible peaks were called using Macs2 callpeaks v2.2.7.1 (q-value <0.01). The CHIPIN method and bedtools suite were used for normalization and comparison.

After red blood cells lysis with EL buffer (Qiagen, 79217), 10 × 10⁶ cells from mice BM of PDXs treated with vehicle or GILT were stained with Fixable Viability Stain 510 (BD Biosciences; cat. #564406, RRID:AB_2869572) and the fluorescent-conjugated antibodies hCD45-APC (BD Biosciences; cat. #555485, RRID:AB_398600) and hCD33-PE (BD Biosciences; cat. #555450, RRID:AB_395843) in cell staining buffer (BioLegend, BLE420201) following to the manufacturer's instructions. Viable hCD45⁺CD33⁺ were sorted using BD FACSAria Fusion flow cytometer (Becton Dickinson) and gently resuspended at 1,500 cells/μL in PBS1X supplemented with 0.04% Ultrapure BSA (Invitrogen, cat. #AM2616).

Up to 10,000 cells per sample were encapsulated using the 10X Chromium controller (10X Genomics) following the Chromium Next GEM Single-Cell 3ʹ User Guide v3.1 (10X Genomics). Briefly, cells were encapsulated into oil droplets with barcoded Gel Beads and reagents to convert mRNA into complementary DNA (cDNA). cDNAs were then amplified, fragmented, and Illumina adapters were added during ligation. After performing a last index PCR, library quality was assessed using Tapestation 2200 (Agilent) and frozen before quantitation and sequencing. Barcoding was done on three levels: cell barcodes allow attribution of each sequence read to its cell of origin; unique molecular identifiers upstream to poly(d)T primers allow tagging of each original molecule to avoid amplification bias; and index allows pooling of different samples. scRNA-seq libraries were quantified with HSD1000 reagents on TapeStation 2200 system (Agilent) and pooled at equimolar concentrations within 200 to 700 bp. Pool libraries were sequenced at 1.8 pmol/L 1% phiX using a NextSeq 500 sequencer (Illumina), at a sequencing depth of 40,000 to 50,000 reads per cell. Sequencing was configured paired-end at 28 cycles for the first read, 91 cycles for the second one, and 8 cycles for a single index.

scRNA-seq output bcl2 files were converted to FASTQ format by using Cell Ranger v.6.1.1.1 (96). Count matrices from Cell Ranger were imported with the Seurat package v4.1.0 (97) into the R programming environment v4.1.2. Cells that were deemed of lower quality were removed, based on their number of unique genes (<200 or >7,500 unique genes) and the proportion of mitochondrial RNA (>20%). Gene counts were normalized with a regularized negative binomial regression, using the SCT normalization workflow from the Seurat package (98). In order to compare gene expression and signatures values after GILT treatment with those before treatment, the datasets from pre- and post-GILT treatment from each of the PDX samples were integrated following the Seurat dataset integration protocol (99). For dataset integration, we used 30 principal components and 3,000 genes. We then applied dimensional reduction where the 30 principal components were further summarized using uniform manifold approximation and projection (UMAP) dimensionality reduction. Cell clusters were identified using Louvain clustering within the Seurat workflow, the resolution parameter was set to 0.25. We identified the clustering of coexpressing genes by hierarchical cluster with the package nclust (https://gitlab.com/pwirapati/nclust). We computed the values of the signatures on SCT-transformed data with the AddModuleScore method from Seurat (100). The signature values were compared with UMAP and violin plots.

Cell lipid extracts were obtained from a frozen pellet of 4.10⁺⁶ harvested cells. Samples were resuspended in 70 μL of ultrapure water and then sonicated 5 times for 10 seconds using a sonication probe (vibra cell). At this step, 20 μL of each sample was withdrawn to further determine the total protein concentration (colorimetric quantification/Pierce BCA Protein Assay Kit, Thermo Fisher Scientific). Samples were extracted using an adapted method previously described (101). Briefly, volumes of CHCL3/MeOH 2:1 (v/v) were adjusted depending on total protein concentration at 3.5 μg/μL in addition to internal standards. Samples were vortexed for 60 seconds, sonicated for 30 seconds using an ultrasonic probe (Bioblock Scientific Vibra Cell VC 75,185, Thermo Fisher Scientific), and incubated for 2 hours at 4°C with mixing. According to volumes of CHCL3/MeOH 2:1 (v/v), H₂O was then added and samples were vortexed for 60 seconds before centrifugation at 15,000 × g for 15 minutes at 4°C. The upper phase (aqueous phase), containing gangliosides, lysoglycerophospholipids, and short-chain glycerophospholipids, was transferred into a glass tube and dried under a stream of nitrogen. The protein disk interphase was discarded and the lower lipid-rich phase (organic phase) was pooled with the dried upper phase and the mixture dried under nitrogen. All samples were resuspended in the same initial volume of CHCL3/MeOH 2:1 (v/v) and 40 μL of each extract was 5-fold diluted in a solution of MeOH/isopropanol/H2O 65:35:5 (v/v/v) before injection. A quality control (QC) sample was obtained by pooling 20 μL of each sample preparation. QC samples were injected every 10 samples in order to evaluate the signal variations of lipid species.

Lipidomic profiles were determined using an Ultimate 3000 liquid chromatography system (Thermo Fisher Scientific) coupled to a high-resolution Thermo Orbitrap Fusion (Thermo Fisher Scientific) equipped with an electrospray source (ESI). Chromatographic separation was performed on a Phenomenex Kinetex C8 column (150 × 2.1 mm, 2.6 μm) at 0.4 mL/minute, 60°C and using an injection volume of 10 μL. Mobile phases A and B were H₂O/MeOH 60:40 (v/v), 0.1% formic acid, and isopropanol/MeOH 90:10 (v/v), 0.1% formic acid in negative ionization mode, respectively. Ammonium formate (10 mmol/L) was added to both mobile phases in the positive ionization mode in order to detect glycerolipids and cholesteryl-esters under [M+NH4] + form. The gradient elution was solvent B was maintained for 2.5 minutes at 32%, from 2.5 to 3.5 minutes it was increased to 45% B, from 3.5 to 5 minutes to 52% B, from 5 to 7 minutes to 58% B, from 7 to 10 minutes to 66% B, from 10 to 12 minutes to 70% B, from 12 to 15 minutes to 75% B, from 15 to 19 minutes to 80% B, from 19 to 22 minutes to 85% B, and from 22 to 23 minutes to 95% B; from 23 to 25 minutes, 95% B was maintained; from 25 to 26 minutes solvent B was decreased to 32%, and then maintained for 4 minutes for column reequilibration. The mass resolving power of the mass spectrometer was 240,000 (FWHM) for MS experiments. Samples were analyzed in both positive and negative ionization modes. The ESI source parameters were as follows: the spray voltage was set to 3.7 kV and −3.2 kV in positive and negative ionization modes, respectively. The heated capillary was kept at 360°C, and the sheath and auxiliary gas flow were set to 50 and 15 (arbitrary units), respectively. Mass spectra were recorded in full-scan MS mode from m/z 50 to 2,000. After LC-HRMS analysis of samples and annotation of features, QC samples were reinjected for HCD MS/MS experiments in positive and negative ion modes on the same instrument set in targeted mode using inclusion lists. The isolation width was set at m/z 0.8, the stepped normalized collision energy was set at 20% ± 10% and the mass resolution was set at 17,500 FWHM at m/z 200. HCD mass spectra were inspected manually in order to confirm annotations.

Obtained raw files were converted to the mzML or mzXML format with the ProteoWizard software (102). Mass spectra were processed using the xcms R package deployed in the Workflow4Metabolomics ­Galaxy platform (ref. 103; https://workflow4metabolomics.org). Features were annotated by using an in-house database based on accurately measured masses and several filtration steps such as retention time windows depending on lipid classes and relative isotopic abundance (RIA) of lipid species in order to limit false annotations for lipidomics.

An amount of cells containing 10 μg of DNA was homogenized in 700 μL of water with a handheld sonicator and was mixed with 800 μL HCl(1M):CH3OH 1:8 (v/v), 900 μL CHCl3, 200 μg/mL of the antioxidant 2,6-di-tert-butyl-4-methylphenol (BHT, Sigma-Aldrich, cat. #B1378) and 3 μL of SPLASH LIPIDOMIX Mass Spec Standard (Avanti Polar Lipids, cat. #330707). After vortexing and centrifugation, the lower organic fraction was collected and evaporated using a Savant Speedvac spd111v (Thermo Fisher Scientific) at room temperature and the remaining lipid pellet was stored at −20°C under argon.

Just before mass spectrometry analysis, lipid pellets were reconstituted in 100% ethanol. Lipid species were analyzed by LC-ESI-MS/MS on a Nexera X2 UHPLC system (Shimadzu) coupled with hybrid triple quadrupole/linear ion trap mass spectrometer (6500+ QTRAP system; AB SCIEX). Chromatographic separation was performed on an XBridge amide column (150 mm × 4.6 mm, 3.5 μm; Waters) maintained at 35°C using mobile phase A [1 mmol/L ammonium acetate in water-acetonitrile 5:95 (v/v)] and mobile phase B [1 mmol/L ammonium acetate in water-acetonitrile 50:50 (v/v)] in the following gradient: (0–6 minutes: 0% B → 6% B; 6–10 minutes: 6% B → 25% B; 10–11 minutes: 25% B → 98% B; 11–13 minutes: 98% B → 100% B; 13–19 minutes: 100% B; 19–24 minutes: 0% B) at a flow rate of 0.7 mL/minute which was increased to 1.5 mL/minute from 13 minutes onward. SM, cholesterol ester, CER, dihydroceramide, hexosylceramide, and lactosylceramide were measured in positive ion mode with a precursor scan of 184.1, 369.4, 264.4, 266.4, 264.4, and 264.4, respectively. Triacylglyceride, diacylglyceride, and monoacylglyceride were measured in positive ion mode with a neutral loss scan for one of the fatty acyl moieties. PC, lysophosphatidylcholine, PE, lysophosphatidylethanolamine, phosphatidylglycerol, phosphatidylinositol, and phosphatidylserine were measured in negative ion mode by fatty acyl fragment ions. Lipid quantification was performed by scheduled multiple reactions monitoring, the transitions being based on the neutral losses or the typical product ions as described above. The instrument parameters were as follows: curtain gas = 35 psi; collision gas = 8 a.u. (medium); ionspray voltage = 5500 V and −4,500 V; temperature = 550°C; ion source gas 1 = 50 psi; ion source gas 2 = 60 psi; declustering potential = 60 V and −80 V; entrance potential = 10 V and −10 V; collision cell exit potential = 15 V and −15 V.

The following fatty acyl moieties were taken into account for the lipidomic analysis: 14:0, 14:1, 16:0, 16:1, 16:2, 18:0, 18:1, 18:2, 18:3, 20:0, 20:1, 20:2, 20:3, 20:4, 20:5, 22:0, 22:1, 22:2, 22:4, 22:5, and 22:6 except for TGs which considered: 16:0, 16:1, 18:0, 18:1, 18:2, 18:3, 20:3, 20:4, 20:5, 22:2, 22:3, 22:4, 22:5, and 22:6.

Peak integration was performed with the MultiQuantTM software version 3.0.3. Lipid species signals were corrected for isotopic contributions (calculated with Python Molmass 2019.1.1) and were quantified based on internal standard signals and adhering to the guidelines of the Lipidomics Standards Initiative (LSI; level 2 type quantification as defined by the LSI). Unpaired t test P values were calculated in Python StatsModels version 0.10.1. Lipid species were considered to have significantly changed in abundance with an FDR ≤ 0.05 and an absolute fold change FC ≥ |1.5| (log₂FC ≥ |0.58|). Lipid species were excluded when more than one replicate was impaired per condition. If only one replicate was impaired, it was excluded and the lipid species were conserved. PUFA:MUFA ratios for each phospholipid subclass were calculated using this formula:

MV4-11 and MOLM-14 cells were plated at 0.4 × 10⁶ cells per mL and treated with QUIZ in media supplemented with U-¹³C₆-glucose (5.6 mmol/L) or U-¹³C₅-glutamine (2 mmol/L) for 24 hours before sampling. Lipids corresponding to total FA from cell pellets of three million cells of three independent experiments were extracted according to Bligh and Dyer (104) in dichloromethane/water/methanol 2% acetic acid (2.5:2.5:2, v/v/v), in the presence of the internal standards glyceryl trinonadecanoate (4 μg). After centrifugation for 6 minutes at 2,500 rpm, the lipid extract was hydrolyzed in KOH (0.5 M in methanol, 1 mL) at 55°C for 30 minutes. A second extraction was performed with dichloromethane/water/methanol (2.5/2/1.5, v/v/v). The lipid extract was evaporated to dryness and dissolved in ethyl acetate (10 μL). Derivatization was performed with pentafluorobenzylbrommide in ACN (1%) and diisopropylethylamine in ACN (1%; 1/1, v/v).

FA methyl ester (1 μL) was analyzed on a gas chromatography Thermo Trace 1310 (Thermo Fisher Scientific) with mass spectrometer Thermo TSQ 8000 EVO (Thermo Fisher Scientific). Chemical ionization is used as ionization, helium as carrier gas, and an HP-5MS column (30 m × 0.25 mm, 0.25 μm) is used for separation. Oven temperature was programmed from 130°C to 232°C at a rate of 4°C/minute (8 minutes), from 232°C to 230°C at a rate of 10°C/minute (2 minutes). A full-scan method was used to detect compounds.

GC-MS analysis produced a mass spectrum for each FA, which contains the abundance of each isotopologue. For each FA, the lightest (unlabeled) isotopologue is denoted M + 0, whereas the isotopologue with 1 atom [¹³C] PFB-palmitate (M+1) has a mass of 256.3, etc. Isotopic clusters were obtained by integrating gas chromatographic signals for each isotopologue. Isotopologue distributions were obtained from the corresponding isotopic clusters after correction for the natural abundance of carbon and nontracer elements (oxygen and hydrogen) using IsoCor software (ref. 105; https://isocor.readthedocs.io/en/latest/index.html), and purity of the tracer was corrected assuming 99% ¹³C-purity. Finally, the ¹³C-enrichment, which represents the mean content in tracer atoms (¹³C) within the molecule, was calculated from the corresponding IDs.

We measured lipid peroxide production using 2 μmol/L C11-BODIPY (581/591; Invitrogen, cat. #D3861) according to the manufacturer's instructions. We used either Annexin V APC (cat. #550474), FITC (cat. #556420), BV421 (cat. #563973), or V500 (cat. #561501), all from BD Biosciences), or propidium iodide (Invitrogen, cat. #P1304MP), or DRAQ7 (Thermo Fisher Scientific, cat. #D15106) to measure cell viability, according to the manufacturer's instructions. Data were collected using a C6 Accuri flow cytometer (Becton Dickinson) with CFlow Plus software or using CytoFLEX flow cytometer (Beckman Coulter). We used CFlow Plus (Becton Dickinson) or FlowJo software (Becton Dickinson, RRID:SCR_008520) for analysis of at least 10⁴ cells per condition.

Femurs, tibias, and spleens of mice were fixed for 24 hours in 4% paraformaldehyde. Decalcification was carried out using 15% formic acid at 4°C for 4 hours, followed by a second fixation in 4% paraformaldehyde for 24 hours. Samples were paraffin embedded and then sliced using a microtome. Serial sections (4 μm) were analyzed by IHC using a polyclonal anti–4-HNE antibody (Thermo Fisher Scientific, cat. #BS-6313R) according to the manufacturer's instructions. Briefly, slides have been heated in a microwave submersed in 1× citrate unmasking solution until boiling, followed with 10 minutes at sub-boiling temperature (95°–98°). After cooling and washing, sections were blocked for 1 hour at room temperature, and incubated with anti–4-HNE antibody (1/100 dilution) overnight at 4°C. After removing antibody solution, sections were covered with the SignalStain Boost IHC Detection Reagent (HRP, Rabbit, Cell Signaling Technologies, cat. #8114) for 30 minutes. Next, stained slides were imaged using the Widefield scanner Axioscan.Z1 slide scanner (Zeiss), and quantification of the 4-HNE signal was done using the QuPath software (106).

Statistical analyses were conducted using GraphPad Prism software v9.2.0 (RRID:SCR_002798). For in vitro studies, statistical significance was determined by the two-tailed unpaired Student t test with Welch correction, or paired t test or ratio paired t test when appropriate. For in vivo studies, statistical significance was determined by the nonparametric Mann–Whitney test. Unless otherwise indicated, all data represent the mean ± SD from at least three independent experiments. *, P < 0.05; **, P < 0.01; ***, P < 0.001.

Further information, resources, and reagents are available upon request. Inquiries should be directed to and will be fulfilled by Jean-Emmanuel Sarry ([email protected]) and Jerome Tamburini ([email protected]).

Microarray, bulk RNA-seq, scRNA-seq, and ATAC-seq data are publicly available in the Gene Expression Omnibus at GSE226461, GSE227839, GSE227874, and GSE227400, respectively. All additional data reported in this article will be shared by the lead contacts upon request. Any additional information required to reanalyze the data reported in this article is available from the lead contacts upon request.

F. Simonetta reports grants from Kite/Gilead, Novartis, and Bristol Myers Squibb/Celgene and personal fees from Bristol Myers Squibb/Celgene, Kite/Gilead, Incyte, AstraZeneca, Neovii, and Janssen outside the submitted work. V. Dembitz reports grants from Cancer Research UK during the conduct of the study. N.S. Anstee reports royalties from the Walter and Eliza Hall Institute of Medical Research outside the submitted work. A.H. Wei reports grants from AbbVie, Novartis, Bristol Myers Squibb, Syndax, AstraZeneca, and Amgen, personal fees from the Walter and Eliza Hall Institute of Medical Research, grants and personal fees from Servier, and other support from Janssen and Astex outside the submitted work. S. Bertoli reports personal fees from AbbVie, Astellas, Bristol Myers Squibb/Celgene, Pfizer, and Jazz Pharmaceuticals outside the submitted work. S. Tavitian reports other support from Novartis and Servier outside the submitted work. C. Récher reports grants and personal fees from AbbVie, Astellas, Bristol Myers Squibb, and Jazz Pharmaceuticals, grants from Iqvia, and personal fees from Novartis, Servier, and Takeda outside the submitted work. J.V. Swinnen reports a patent for phospholipid profiling issued. G. Roué reports grants from Interreg-POCTEFA, Instituto de Salud Carlos III, Ministerio de Ciencia e Innovación, Gilead Sciences, TG Therapeutics, Kancera AB, and Onconova Therapeutics outside the submitted work. A.S. Dejean reports grants from ANR Emetic during the conduct of the study. J.-E. Sarry reports personal fees from Servier outside the submitted work. No disclosures were reported by the other authors.

M. Sabatier: Conceptualization, resources, data curation, formal analysis, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. R. Birsen: Conceptualization, resources, data curation, formal analysis, validation, investigation, visualization, methodology, writing–original draft, writing–review and editing. L. Lauture: Conceptualization, resources, data curation, formal analysis, validation, investigation, visualization, methodology. S. Mouche: Conceptualization, resources, data curation, formal analysis, validation, investigation, visualization, methodology. P. Angelino: Resources, data curation, software, investigation, visualization. J. Dehairs: Conceptualization, resources, data curation, formal analysis, validation, investigation, visualization, methodology. L. Goupille: Investigation. I. Boussaid: Investigation, methodology. M. Heiblig: Resources, formal analysis, investigation. E. Boet: Data curation, formal analysis, investigation. A. Sahal: Resources, formal analysis, investigation. E. Saland: Conceptualization, resources, data curation, formal analysis, validation, investigation, visualization, methodology. J.C. Santos: Investigation. M. Armengol: Investigation. M. Fernández-Serrano: Investigation. T. Farge: Resources, data curation, software, investigation. G. Cognet: Resources, investigation. F. Simonetta: Resources, formal analysis, validation, investigation, writing–review and editing. C. Pignon: Resources, data curation, formal analysis, validation, investigation, visualization. A. Graffeuil: Resources, validation, investigation, visualization, writing–review and editing. C. Mazzotti: Resources, formal analysis, validation, investigation, visualization. H. Avet-Loiseau: Resources, validation, investigation, visualization. O. Delos: Resources, formal analysis, validation, investigation, visualization. J. Bertrand-Michel: Resources, data curation, formal analysis, investigation, visualization. A. Chedru: Resources, formal analysis, validation, investigation. V. Dembitz: Resources. P. Gallipoli: Supervision. N.S. Anstee: Resources. S. Loo: Resources. A.H. Wei: Supervision. M. Carroll: Resources. A. Goubard: Resources. R. Castellano: Resources, investigation. Y. Collette: Supervision. F. Vergez: Resources, data curation. V. Mansat-De Mas: Resources, data curation, formal analysis, validation, investigation, visualization, writing–review and editing. S. Bertoli: Resources, formal analysis, investigation. S. Tavitian: Resources, investigation. M. Picard: Resources, validation, investigation. C. Récher: Resources, formal analysis, validation, investigation. N. Bourges-Abella: Supervision. F. Granat: Supervision. O. Kosmider: Resources, formal analysis, validation, investigation, visualization. P. Sujobert: Resources, data curation, investigation, visualization. B. Colsch: Resources, formal analysis, validation, investigation, visualization, writing–review and editing. C. Joffre: Data curation, investigation, visualization, project administration, writing–review and editing. L. Stuani: Data curation, validation, investigation, visualization, writing–review and editing. J.V. Swinnen: Data curation, formal analysis, validation, investigation, visualization, project administration, writing–review and editing. H. Guillou: Resources, data curation, validation, visualization, writing–review and editing. G. Roué: Supervision. N. Hakim: Formal analysis, investigation. A.S. Dejean: Formal analysis, supervision, validation. P. Tsantoulis: Resources, data curation, validation, writing–review and editing. C. Larrue: Resources, data curation, supervision, validation. D. Bouscary: Conceptualization, resources, supervision, funding acquisition, validation, writing–original draft, project administration, writing–review and editing. J. Tamburini: Conceptualization, resources, supervision, funding acquisition, validation, investigation, visualization, writing–original draft, project administration, writing–review and editing. J.-E. Sarry: Conceptualization, resources, supervision, funding acquisition, validation, visualization, writing–original draft, project administration, writing–review and editing.

We thank all members of mice core facilities (UMS006, ANEXPLO, Inserm, Toulouse) in particular Marie Lulka, Christine Campi, and Cédric Baudelin for their support and technical assistance. We thank Anne-Marie Benot, Muriel Serthelon, and Stéphanie Nevouet for their daily help about the administrative and financial management of our Team. We thank the Geneva University Medical School electron microscopy (PFMU), flow cytometry, bioimaging, genomics (iGE3), proteomics, Reader Assay Development and Screening (READS) and zootechnie core facilities for technical support. The authors also wish to thank the Barts Cancer Institute tissue bank for sample collection and processing. This research was supported by Cancer Research UK (PG, Advanced Clinician Scientist fellowship, C57799/A27964). This work was supported by grants from the Laboratoire d'Excellence Toulouse Cancer (TOUCAN and TOUCAN2.0; contract ANR11-LABEX), INCA (PLBIO 2020-010, DIALAML), the Fondation Toulouse Cancer Santé, the Ligue National de Lutte Contre le Cancer, the European Regional Development Fund [ERDF; the Interreg V-A Spain-France-Andorra (POCTEFA) program, EFA360/19], the Fondation ARC, and the association GAEL. M. Sabatier and R. Birsen were fellows from the Fondation ARC contre le Cancer. L. Lauture is supported by a PhD fellowship from the Ligue National de Lutte Contre le Cancer. A. Sahal is a fellow from the European Regional Development Fund (ERDF) through the Interreg V-A Spain-France-Andorra (POCTEFA) program, project PROTEOblood (EFA360/19). This work was also supported by grants from the Dr. Henri Dubois-Ferrière, Dinu Lipatti Fondation and Geneva university hospital private Foundation. This work was also supported by funding from Ligue Genevoise contre le Cancer, Fondation Copley May, Fondation Medic and Fondation Pastré through the Translational Research Center for Hemato-Oncology (University of Geneva, Faculty of Medicine, Geneva, Switzerland). METATOUL-lipidomics core facility is a member of the MetaboHUB French infrastructure (ANR-INBS-0010). Team J.-E. Sarry and Team D. Bouscary are members of OPALE Carnot Institute, The Organization for Partnerships in Leukemia.

The publication costs of this article were defrayed in part by the payment of publication fees. Therefore, and solely to indicate this fact, this article is hereby marked “advertisement” in accordance with 18 USC section 1734.