Re-programming energy metabolism is a hallmark of cancer [1]. Acute myeloid leukemia (AML) permits analysis of the impact of a systemic cancer on the metabolism over time. Oncometabolites have been demonstrated in 5–20 % of AML-harboring mutations in Isocitrate Dehydrogenase (IDH) genes [2, 3], and the metabolic profile of AML cell lines can be employed to investigate drug treatments [4, 5]. Moreover, previous studies linked AML with perturbation of metabolic pathways including glucose metabolism [6, 7]. Nuclear magnetic resonance (NMR) is a rapid, highly reproducible cost-effective analytical tool to profile metabolic fluctuations, requiring minimal sample manipulation and well suited for automation and high-throughput purposes, thus ideal for untargeted analysis in clinical applications [8–11].

To study the metabolic changes associated with AML in patients, we enrolled nine newly diagnosed patients in a prospective trial (METAM-02 trial, Additional file 1) and monitored the metabolic trajectory of blast clearance during the first two cycles of intensive chemotherapy (CT) (Fig. 1a). Peripheral blood (PB), bone marrow (BM), and urine samples were collected prior, during, and after induction and consolidation CT, resulting in nine time points (tpH) (Fig. 1b). At the final time point (tp9H), all patients had achieved complete remission.

Fig. 1
figure 1

Metabolic changes in AML patients undergoing intensive chemotherapy. a Patient characteristics. CR1 first complete remission. *UPN#1 developed a lethal fungal pneumonia while in remission after the first consolidation CT cycle. **UPN#3 underwent allogeneic hematopoietic stem cell transplantation in CR1 and subsequently relapsed. b Outline of the study design. c Superposition of representative 1H Carr-Purcell-Meiboom-Gill (CPMG) spectra of BM at tp1H (black line) and tp9H (red line) acquired at 37 °C on a Bruker Avance 600-MHz spectrometer. The 1H CPMG NMR experiment is based on a pulse sequence that strongly reduces the NMR signals deriving from large molecules; herewith, molecules with high molecular weight are essentially invisible in the 1H spectrum, thus facilitating spectra interpretation and small molecule identification in the presence of large proteins and lipoproteins. Peaks correspond to the different metabolites: 1—high-, low-, and very low-density lipoproteins (HDL, LDL, VLDL) CH3; 2—isoleucine; 3—leucine; 4—valine; 5—3-aminoisobutyrate; 6—3-hydroxybutyrate; 7—LDL/VLDL CH2; 8—lactate; 9—alanine; 10—lipids CH2CH2CO; 11—acetate; 12—lipids CH2C=C; 13—N-acetyl-glycoproteins (NAG) NHCOCH3; 14—glutamine; 15—lipids CH2CO; 16—acetoacetate; 17—citrate; 18—lipids C=CCH2C=C; 19—creatinine; 20—creatine; 21—creatine phosphate; 22—glucose; 23—glycerol of lipids CHOCOR; 24—α glucose; 25—poly-unsaturated fatty acids (UFA); 26—tyrosine; 27—phenylalanine; 28—histidine; 29—formate. d OPLS-DA score plot for pooled PB and BM samples collected at diagnosis and typing positive (blue circles; n = 6) or negative (green circles; n = 12) for missense mutations in the IDH1/2 genes. OPLS-DA with N = 18, CV ANOVA p = 0.059, R 2 = 0.93, and Q 2 = 0.615. The area under the curve (AUC) of the ROC analysis was 0.86 (p < 0.001). e Metabolites discriminating AML patients with or without IDH gene mutations. Loadings indicate how much the variables, i.e., the metabolites, contribute to the model. Shown are loadings with jack-knifed confidence interval. The metabolites significantly contributing to the model were selected based on variable importance in projection >1 and jack-knifed confidence interval of loadings not crossing the zero line. Positive loading values (blue bars) indicate the metabolites increased in patients with mutated IDH, while negative values (green bars) are associated with increased levels in patients with wild-type IDH. Note that the concentration levels of the classical IDH mutation oncometabolite (R)-2-hydroxyglutarate were below the NMR detection limit and that because of the overlap of the resonances associated to CH2 groups of high-, low- and very low-density lipoproteins (HDL, LDL, VLDL) and the CH2C=C groups of different lipid molecules it was not possible to establish their specific contribution to the model. f OPLS-DA score plot for pooled PB and BM samples of patients at diagnosis (tp1H, blue circles; n = 18) vs remission after chemotherapy (pooling samples after induction, tp5H, and after first consolidation, tp9H, green circles; n = 29). OPLS-DA model with N = 47, CV-ANOVA p = 0.006, R 2 = 0.99, and Q 2 = 0.671. The AUC of the ROC analysis was 0.73 (p < 0.001). g Metabolites discriminating AML patients at diagnosis and in remission after chemotherapy. Positive loading values (blue bars) indicate the metabolites increased in AML patients at tp1H, while negative values (green bars) are associated with increased metabolite levels at tp5H + tp9H

Full size image

Seventy-four PB, 33 BM, and 75 urine samples were analyzed by NMR spectroscopy (Fig. 1c and Additional file 1). First, based on results from unsupervised statistical inspection, we decided to pool BM samples with the respective PB and observed that despite the small number of cases, supervised orthogonal partial least squares-discriminant analysis (OPLS-DA) allowed clustering of the samples at diagnosis (tp1H) according to the presence or absence of mutations in IDH genes (Fig. 1d, e), supporting the notion that these mutations have an overall impact on the metabolome [2, 3]. Moreover, OPLS-DA discriminated the metabolic differences between patients at disease diagnosis and those in remission (tp5H and tp9H), highlighting 30 metabolites significantly contributing to the model (Fig. 1f, g).

To pinpoint the most relevant ones for AML biology, we integrated our study with a mouse-human AML model, allowing us to follow the metabolic changes associated with disease progression. Six littermate NOD/SCID γ-chain-null (NSG) mice were infused with 2 × 106 primary AML blasts (from patient #3). PB samples were collected over time from the same animals before and after AML infusion for a total of nine time points (tpM) (Fig. 2a). AML became detectable (blast count > 1 cell/μl) in the PB at tp5M for all infused mice; thereafter, an exponential growth started (Fig. 2b). We thus compared samples before AML infusion (tp1M and tp2M) and in the exponential phase of disease (tp5M and tp6M). Possibly due to the lower number of variables present in the mouse model, OPLS-DA (Fig. 2c) pinpointed only 22 metabolites accounting for AML progression, mainly involved in pathways related to energy and fatty acid metabolism (Fig. 2d). The increased lactate level in the plasma is a typical hallmark of cancer, characterized by an aerobic glycolytic shift (“Warburg effect”) [1, 12]. Lower plasma levels of cholesterol, lipids, and unsaturated fatty acids were found in AML mice than in healthy controls, in agreement with previous reports [6, 13, 14], highlighting an involvement of fatty acid metabolisms in AML, possibly related to the demand for lipids and cholesterol in tumor proliferation. Also, the increase of glutamine in AML mice is consistent with cancer since glutamine is a major carbon and nitrogen source for tumor cell proliferation [15].

Fig. 2
figure 2

Metabolic profile of the AML mouse-human model. a Design of the study. b Leukemic cell count (human CD33+, CD45+ values) over time. c OPLS-DA score plot for healthy mice (tp1M and tp2M; green circles; n = 12) vs mice at tp5M and tp6M (blue circles; n = 12) showing an AML metabolic signature at an early stage of AML. OPLS-DA model with N = 24, CV-ANOVA p = 0.002, R 2 = 0.99, and Q 2 = 0.819. The area under the curve (AUC) of the ROC analysis was 0.96 (p < 0.001). d Metabolites discriminating healthy and human AML-engrafted mice. Positive loading values (blue bars) indicate the metabolites increased in AML mice, while negative values (green bars) are associated with increased levels in healthy mice

Full size image

Combining the two longitudinal approaches to trace AML in patients and mice, we narrowed our screen to seven metabolites representative of AML tumor burden: leucine, valine, alanine, glutamine, citrate, acetoacetate, and lipids CH2CO. We interpreted this set of metabolites as a dynamic fingerprint of AML evolution that could be followed in a mirror-like trajectory in both studies (Fig. 3).

Fig. 3
figure 3

Heat maps of human or mouse plasma tracing the trajectory of the seven metabolites associated with AML evolution in both the patients and mice. a Heat maps of human BM and PB tracing the trajectory of the seven metabolites (averaged normalized metabolite area) associated with AML evolution in both the mice and patients. b Heat map of mouse PB depicting the trajectory of the seven metabolites (averaged normalized metabolite area of the six mice) associated with AML evolution in both the mice and patients. Note that leucine, valine, alanine, citrate, and acetoacetate increased upon AML progression until tp8M and then decreased at tp9M. This reverse effect supported our hypothesis about a metabolic change due to overall systemic failure at a late stage of disease and not to a specific effect of AML

Full size image

Collectively, our results highlight how NMR-based metabolomics might provide valuable and non-redundant information on the systemic effects of leukemia, to be consolidated in larger cohorts, and integrated in more comprehensive system biology approaches. Although the limited number of cases investigated in this study does not allow a meaningful comparison to previous reports [6, 7], the technical robustness and reproducibility of NMR, together with the minimal sample manipulations and invasiveness of this technique, appear well suited to perform longitudinal studies in clinical settings. Moreover, we first provide experimental evidence on how patient-derived xenografts can be integrated into metabolomic studies, complementing and narrowing the screening of clinically relevant metabolites.