Publicly available clinical and transcriptomic data of three adult AML cohorts whose patients were treated with intensive chemotherapy were included in this study: TCGA-AML (n = 121; Illumina HiSeq 2000) (10), Beat AML (n = 219; Illumina HiSeq 2500) (11), and HOVON-GSE6891 (n = 405; Affymetrix Human Genome U133 Plus 2.0 Array) (12, 13). The normalized gene expression and clinical data for the TCGA cohort were retrieved from the FireBrowse data portal (www.firebrowse.org), whereas for the Beat AML cohort we used the Beat AML Vizome data portal (www.vizome.org) and the cBioPortal for Cancer Genomics (www.cbioportal.org). The normalized gene expression data were retrieved from the Gene Expression Omnibus database (www.ncbi.nlm.nih.gov/geo/) for the HOVON cohort.

The AML (HL-60, MOLM-13, U-937, THP-1, OCI-AML-2, OCI-AML-3, and NB-4) and stromal (MS5) cell lines were originally obtained from the American Type Culture Collection (ATCC) and cultured in, respectively, Roswell Park Memorial Institute (RPMI) medium (Lonza 12–115F; Basel, Switzerland) and alpha minimum essential medium (α-MEM) (Lonza) supplemented with 10% fetal bovine serum (FBS; F7524; Sigma Aldrich, St. Louis, MO, USA), in a humidified 5% CO₂ atmosphere at 37°C. AraC-resistant AML cell lines were generated previously (18).

Patient-derived AML samples taken at time of diagnosis ( Supplementary Table 3 for clinical characteristics) were thawed from our cryobank [stored after informed consent and following approval by the Medical ethics committee of the University Medical Center Groningen (MCG) in accordance with the Declaration of Helsinki protocol; code NL43844.042.13, 6 January 2014] and cultured, as previously described, in Gartner’s medium on top of a MS5 support layer (19). After 2–3 days of acclimatization, only the AML samples with a cell viability above 80% were used for further experiments.

To determine the extent of mannosylation, AML cells (5 × 10⁴ in 100 µL of RPMI plus 10% FBS) were incubated with 25 µg/mL of ConA-488 or ConA-647 (C11252/C21421; Thermo Fisher Scientific) for 1 h at 4°C. After washing with phosphate-buffered saline (PBS), the ConA-488/ConA-647 fluorescent intensity was determined using flow cytometry and accessory software (CytoFLEX; Beckman Coulter, Brea, CA, USA/BD, Accuri Franklin Lakes, NJ, USA). The same protocol was used for the lens culinaris agglutinin (LCA) staining using 125 µg/mL of LCA-fluorescein isothiocyanate (FITC; L32475; Thermo Fisher Scientific). For the patient-derived samples the cells were counter stained for CD45 (catalog number 21810456; ImmunoTools GmbH, Friesoythe, Germany) to enable gating on myeloblasts (see Supplementary Figure 2A ).

For the fluorescent images, a similar protocol as for flow cytometry was used. After washing, cells were spotted on a microscope slide using a cytocentrifuge (CytoSpin 3; Shandon, UK) and analyzed using the EVOS™ Cell Imaging System (EVOS-FL; Thermo Scientific).

To determine cytarabine (AraC) sensitivity, the AML cells were plated in a 48-wells plate (5 × 10⁴ in 200 µL of RPMI plus 10% FBS) and treated with 500nM AraC (hospital pharmacy, UMCG; Cytarabine Accord solution (100 mg/mL); RVG 112666) and incubated for 72 h at 37°C. Subsequently, cell viability was assessed using the MTS assay (CellTiter 96^® AQueous One Solution Cell Proliferation, G3580; Promega, Madison, WI, USA). In brief, MTS was added [7.5% v/v (volume to volume)] to each well and incubated at 37°C. After sufficient color development a read-out was performed at OD_{490 nm} (MultiSkan Sky of Thermo Fisher Scientific). The cell viability was calculated by subtracting OD_{490 nm} of the dead control of each value (7.5% dead mix, consisting of 10% Triton X in 70% ethanol) and calculating the viability as a percentage of the untreated control (treated/untreated x 100%). It is noteworthy that the ConA binding toward the AML cell lines was determined at the time of plating for the MTS assays to ensure the correct correlation analysis. For the tunicamycin experiments (T7765; Sigma Aldrich, Steinheim, Germany), the cells were pre-incubated with 50ng/mL of tunicamycin for 72 h prior to treatment with AraC. For the patient-derived samples, the cells were harvested after 72 h of incubation (same protocol as cell lines) and stained with anti-CD45-allophycocyanin (APC) and annexin-V-FITC (catalog number 31490013; ImmunoTools) in binding buffer (10 x annexin V-binding buffer; BD Biosciences, San Jose, CA, USA).

To determine whether AML cells with high levels of ConA binding survived AraC treatment, cells were prestained with ConA-647 using a low non-toxic dose following the protocol as used for fluorescent imaging. Subsequently the cells were plated, as described above, and after 72 h of incubation, fluorescent intensity of the surviving cells was measured using flow cytometry.

To identify which glycosylation signatures of AML cells could possibly be associated with AraC sensitivity, glycosylation-related genes (n = 271; Supplementary Table 1 ) were analyzed for their prognostic role regarding the overall survival of AML patients that were intensively treated with AraC. Here, high expression of 21 glycosylation-related genes was significantly associated with poor overall survival in AML patients, independently of confounders ( Supplementary Table 2 ). After excluding genes without a direct role in the addition or removal of specific carbohydrates (n = 7), 13 genes were left, among which six genes were directly involved in mannosylation, i.e., MPI, POMT2, RPN1, ALG2, ALG1L2, and ALG12 ( Figure 1A ), predominantly of N-glycans (five out of six; Figure 1B ). Notably, MAN1A2, MAN2A1, and EDEM3, three mannosidases that remove mannoses from N-glycans, were borderline significant after the correction for confounders (with a p-value of, respectively, 0.05, 0.08, and 0.09) among the genes that were associated with better overall survival (data not shown). Together, these data suggest that high levels of mannosylation are associated with worse clinical outcome in AML patients that are treated with AraC-based protocols.

We hypothesized that the worse survival of patients that express higher levels of glycosyltransferases that couple mannoses is caused by a reduced sensitivity to AraC. To confirm this association, we used PGC nano-LC-MS/MS to determine the levels of mannosylated N-glycans in parental compared to AraC-reistant cell line pairs of THP-1 and MOLM-13 cells.

In both THP-1^par and THP-1^Arac-Res cell lines, 22 N-glycans could be identified, whereby oligomannose accounted for 6 of the 22 assigned N-glycans that contain α-mannose ( Figure 2A ). Furthermore, two out of three hybrid N-glycans with composition of N3H6 and N3H6S1 were detected. In addition, there were three paucimannose and 10 complex N-glycans detected in the PGC LC-MS/MS analysis. Also quantitatively, the most prevalent structures were oligomannose glycans, varying from 69% to 75%, whereby the relative abundance was slightly but significantly higher in the resistant cells ( Figure 2B ), as overall glycosylation seemed to be increased in the resistant cells. To estimate the absolute abundance, the intensity of N-glycans carrying α-mannose was normalized to the intensity of the spiked internal standard DP7, revealing a two times higher expression level of α-mannose-containing N-glycans in THP-1^Arac-Res (~ 3 × 10⁷ per cell) compared to THP-1^par cells (~ 1.5 × 10⁷ per cell) ( Figure 2C ).

A similar glycan profile was observed for the MOLM-13 cell line pair ( Supplementary Figures 1A, B ), which showed a 1.9 times higher expression level of α-mannose-containing glycans in MOLM-13^Arac-Res (~ 3.8 × 10⁷ per cell) compared to in MOLM-13^par cells (~ 1.9 × 10⁷ per cell) ( Figure 2C ). Together, these results demonstrate a significant increase in the expression of α-mannose-containing N-glycans in AraC-resistant cell lines compared to parental cell lines, which is in line with the hypothesis that AML cells that do not respond to therapy are more strongly mannosylated.

To use the increased expression of mannoses to predict AraC sensitivity, fluorescently labeled concanavalin A (ConA) was selected, a lectin that predominantly recognizes high mannose N-glycans (23, 24). As expected, ConA-binding was increased toward AraC-resistant cells, both on the cell membrane and in intracellular compartments ( Figures 3A, B ). It is noteworthy that ConA binding was increased in AraC-resistant cells on the cell membrane, as well as intracellular compartments ( Figure 3C ). In addition, lens culinaris agglutinin (LCA), a lectin that also recognizes α-mannoses, also bound more strongly to THP-1^AraC-Res cells ( Supplementary Figure 1B ). To determine if increased N-glycan mannosylation directly impacts the chemosensitivity, both MOLM-13^par and MOLM-13^Arac-Res cells were treated with the N-glycan inhibitor tunicamycin prior to AraC treatment. This increased the sensitivity for both cell lines toward AraC, with re-sensitizing MOLM-13^Arac-Res to AraC, with the same level of cell death as induced by AraC in parental cells ( Figure 3D ). Taken together, AraC-resistant AML cells express more mannosylated N-glycans that can be determined by the α-mannose-binding lectin ConA.

To determine whether ConA staining also predicts AraC sensitivity in non-resistant AML, a panel of cell lines (n = 7) was stained with ConA and treated with AraC. Both the extent of ConA staining ( Figure 4A ) as well as the AraC sensitivity ( Figure 4B ) differed between the cell lines. As expected, the extent of ConA binding significantly correlated with AraC sensitivity, whereby cells with higher levels of ConA binding were less sensitive toward AraC treatment ( Figure 4C ). Furthermore, the cell lines clustered together in “low”, “intermediate”, and “high” binders ( Figure 4C ). It is noteworthy that highly mannosylated cells within one and the same cell line were also less sensitive toward AraC than the cells with fewer mannoses. This was demonstrated by a dose-dependent increase in the fluorescent signal of AraC treated AML cells that were prestained with ConA, reflecting that only the strong binders survived the treatment ( Supplementary Figure 1C ). A correlation analysis between ConA binding and gene expression of the cell lines in the CCLE dataset revealed that ConA binding positively correlated with the expression of genes associated with adverse-risk AMLs, i.e., CD109 and EGR1 ( Figure 4D ) (25, 26). Furthermore, GSEA analysis demonstrated a correlation between ConA binding and cholesterol-related pathways, TP53 mutations, and SREBF/SREBP biology ( Figure 4D ), which are processes usually enriched in AML patients with a poor response to intensive chemotherapy, as demonstrated by us and others (27–29). Thus, ConA staining can predict AraC sensitivity in AML cell lines and is associated with gene expression patterns associated with adverse-risk and chemoresistant AML.

To make a first step toward clinical translation, two patient-derived AML samples, which were obtained upon diagnosis, were defrosted from our cryodatabase. The ConA staining revealed that AML-1 expressed more α-mannose-containing N-glycans than AML-2 ( Figure 4E ), which would mean, based on our previous findings, that AML-1 would be less sensitive toward AraC treatment than sample AML-2. Indeed, AML-1 hardly responded to AraC treatment, whereas AML-2 was sensitive ( Figure 4F ). Furthermore, within the viable cells, AML-2 had a 2.8-fold higher percentage in annexin-V-positive cells than AML-1 (33 vs. 12%) upon treatment with 1,000 nM AraC ( Supplementary Figure 2B ).

Together, we propose that ConA may be used as biomarker to predict AraC sensitivity in AML, whereby AML cells that strongly bind ConA, and thus express high levels of α-mannose-containing N-glycans, are less sensitive or resistant toward AraC treatment ( Figure 4G ).