In this study, we set out to expand our understanding of protein and lipid glycosylation and their regulation in AML to obtain insights that may affect the development of diagnostic or therapeutic approaches. To this end, we retrieved glycomics and transcriptomics data from multiple studies dealing with both cell line models and primary blasts (Fig. 1). Employing the well-defined and comprehensively assessed cell lines as a stepping stone, we intended to define glycosylation of AML blasts as well as its transcriptional regulation on the level of GSTs and TFs.

Fig. 1

Overview of the conducted study. AML glycosylation was explored on the level of glycomics (GPST datasets) and transcriptomics (GSE and DepMap datasets). Based on the depicted datasets originating from cell lines and primary cells we sought to explore cellular glycosylation, involved GSTs, and responsible TFs

Integration of glycan classes in AML cell lines and their association with the FAB classification

The bioinformatic evaluation of the AML glycomic landscape was initiated by relying on cell line models, as the most detailed glycomics and transcriptomics data were available for these cells. Recently, we reported on the in-depth glycan characterization of numerous AML cell lines, i.e., N-, O-, and GSL-glycosylation [13, 14]. To obtain a holistic picture of cellular glycosylation of AML cell lines, to explore cooperative trends, and to determine unique features, we integrated available data [13, 14] on these three glycan classes. To this end, glycans were grouped into glycosylation features as specified in Additional file 2: Table S1 (N-glycans), 2 (O-glycans), and 3 (GSL-glycans). The compiled quantitative information is available in theses supporting tables. The relative abundances of the different glycomic features across 19 AML cell lines were explored by PCA, which pointed towards clustering of cell lines by their FAB classification as visualized in the score plot (Fig. 2a). In particular, cell lines from the M5 class (acute monocytic leukemia cell lines: AML-193, MOLM-13, U-937, THP-1, EOL-1, and MV4-11) showed an apparent grouping driven by relatively high expression of (s)Lex/a across all three glycan classes (Fig. 2b). In addition, some class-specific glycan features contributed to the clustering of M5 cell lines, namely O-glycan associated core 2 and sulfation, N-acetyllactosamine (LacNAc), T antigen, as well as GSL-glycan-associated α-2,6 sialylation. Besides, the cell line M-07e, which is derived from the M-07 (RRID:CVCL_D630) cell line andbelongs to the M7 class (acute megakaryoblastic leukemia), showed similar glycosylation features as the M5 cell lines. However, as this was the only M7 cell line included, it remains unclear as to whether this is a common pattern of M7 cells or unique for this particular cell line. Cell lines of the M6 subtype (acute erythroid leukemia: HEL, HEL 92.1.7, TF-1, KG-1a, and KG-1) clustered and were separated from the M5 cell lines in the score plot (Fig. 2a). Above all, different forms of sialylation were drivers of the clustering of these M6 cell lines, i.e., O- and GSL-glycan-associated α-2,8 sialylation, Neu5Gc on GSLs, α-2,3/6 sialylation on N-glycans, and α-2,6 sialylation on O-glycans, (Fig. 2b). The two sister M6 cell lines (KG-1a and KG-1) clustered particularly close due to the high expression of (α-2,3/6) sialylation on N-glycans and their high antennarity as well as ganglioside GSL-glycans. To further facilitate the comparison of M5 and M6 subtypes, we visualized the z-transformed data of the most important glycan features (Additional file 2: Table S5) in two radar plots (Fig. 2c) revealing differences in (s)Lex/a expression and various types of sialylation in a glycan class-specific manner.

Fig. 2

Glycomic overview of various AML cell lines. a PCA of glycosylation features derived from glycomics data of 19 AML cell lines. Individual cell lines are annotated and colored by their FAB classifications as assigned earlier [13]. b The associated score plot depicts considered glycan features, which are linked to their respective glycan class (N-, O-, and GSL) by color (purple, orange, and green) and symbol (triangle, square, and circle). In addition, arrows indicate features that are linked to a specific type of fucosylation. c Radar plots are showing the differences in glycosylation features between AML classes M5 and M6. Again, these features are subdivided into their respective classes based on color and symbols. Data on all AML cell lines were z-transformed prior to visualizing differences between FAB classes in these radar plots. d Spearman correlation of selected glycosylation features between the different glycan classes. Thick connective lines indicate a good correlation whereas thin connective lines show less correlation. Correlation values are depicted

While the three cell lines of the M4 class (acute myelomonocytic leukemia) showed less clear grouping, they were in part characterized by H antigen expression on O- and GSL-glycans. For the M2 subtype, three cell lines were typed with no clear common glycan signatures becoming apparent.

To investigate whether the same glycan features show distinct expression across the investigated glycan classes, we performed a correlation analysis thereof (Fig. 2d). Common features across glycan classes included (s)Lex/a and α-2,3/6 sialylation, whereas H antigen and α-2,8 sialylation were only found on O- and GSL-glycans. Between O- and GSL-glycans, positive correlations were found for H antigens (r = 0.36), (s)Lex/a (r = 0.48), and α-2,8 sialylation (r = 0.69). In contrast, weak negative correlations were found for α-2,6 sialylation on GSL-glycans compared to N- and O-glycans (r = − 0.12 and − 0.36, respectively). Between N- and GSL-glycans, α-2,3 sialylation showed a positive correlation (r = 0.55).

Exploring the correlation of glycomic features and GST expression in AML cell lines

Biosynthesis of glycans and resulting cellular glycomic features are shaped by GSTs. To explore associations of specific GSTs and related glycosylation features, we performed a Spearman correlation analysis between glycomics and transcriptomics data (Additional file 2: Tables S5, S6, and S7). Previously, we investigated the associations of GSTs with glycomic features specific to individual glycan classes [13, 14]. For this study, we focused on glycan traits shared between the three glycan classes, i.e., α-2,3/6/8 sialylation and blood group antigens (Fig. 3). For this purpose, glycan traits, e.g., sLex/a or α-2,3 sialylation, were calculated by summing up the relative abundances of glycans that carry the respective trait. If multiple of these epitopes were present on one glycan, the relative abundance of this glycan was multiplied by the number of epitopes identified. Glycan traits were integrated across glycan classes, i.e., N-, O-, and GSL-glycans, by summing up their values throughout the three classes in order to observe the impact of specific GSTs on the global glycomic phenotype of AML blasts (Additional file 2: Tables S1, S2, S3, S4, and S5). With respect to α-2,3 sialylation, no significant correlations were found for ST3GALs in the different glycan classes. Of note, ST3GAL1, encoding a GST involved in terminal sialylation of glycoproteins and glycolipids, showed a trend toward a positive correlation with α-2,3 sialylation on both O-glycans and GSL-glycans. Similarly, ST3GAL2, which encodes a GST thought to be specifically responsible for the production of terminal sialylation of glycolipids, positively correlated with α-2,3 sialylation on GSL-glycans. Next, correlations were investigated for GSTs responsible for the formation of α-2,6 sialylation. Strong correlations were found between ST6GALNAC1/3 and α-2,6 sialylation on O-glycans. Despite its preferential activity on glycolipids [43, 44], ST6GALNAC6 expression negatively correlated with GSL-associated α-2,6 sialylation. As for α-2,8 sialylation, ST8SIA6 significantly correlated with this trait on O- and GSL-glycans, whereas this feature was absent on N-glycans.

Fig. 3

Correlation of glycomic features on N-, O-, and GSL-glycans with GST expression. Coefficients were obtained by Spearman correlation and are colored as indicated in the right key bar. Integrated glycosylation traits were obtained by summing up values of class-specific glycosylation traits. Significant values are marked with * (p ≤ 0.05), ** (p ≤ 0.01), and *** (p ≤ 0.001). Correlation coefficients and p-values are listed in the Additional file 2: Table S7

Finally, blood group antigens were assessed. Specifically, FUT1-7 and FUT9 were included in the analysis, whereas FUT8 was excluded as it is not involved in blood group antigen synthesis. FUT1, which is responsible for catalyzing the transfer of a fucose to a terminal galactose residue of glycoconjugates in α-1,2 linkage to form H antigens, showed a significantly positive correlation with global H antigen expression, whereas its expression on O- and GSL-glycans individually did not reach the significance threshold. Concerning (s)Lex/a expression, the strongest positive correlations were found for FUT7 and N- and O-glycosylation. Global expression of (s)Lex/a on N-, O-, and GSL-glycans, correlated significantly positively with FUT7 expression. In addition, FUT4 and FUT9 appeared to be significantly associate with (s)Lex/a expression on N-glycans. Interestingly, overall (s)Lex/a expression seems to negatively correlate with FUT1 and FUT6 expression.

Integrated glycosylation features reveal key transcription factors in AML cell lines

To explore the possible transcriptional regulation of glycosylation throughout all glycan classes, we correlated glycomic features and GST expression with a set of selected TFs (Additional file 2: Tables S8 and S9). Specifically, we focused on hematopoietic TFs, which are guiding normal hematopoiesis and differentiation and were shown to be specifically altered in many cases of AML and affected by mutation in specific types of AML [17, 45]. As illustrated in the correlation heatmap in Fig. 4, the (s)Lex/a antigen for the integrated as well as separate glycoprotein classes showed a positive correlation with SPI1, CEBPA, and MYB. In line with these observations, FUT4 and FUT7 also exhibited positive correlations with SPI1 and CEBPA. FUT9 appeared to follow a similar expression pattern but did not meet the significance criteria. Interestingly, (s)Lex/a expression on N-linked glycoproteins showed a distinct correlation compared to sLex/a expression on glycolipids. Concerning the H antigen, GATA1 showed a significant correlation with integrated glycan classes as well as with its expression on GSL-glycans.

Fig. 4

Correlation of glycosylation features of N-, O-, and GSL-glycans with the expression of selected TFs in AML cell lines. Correlation coefficients were obtained by Spearman analysis and are indicated by color as indicated in the legend. Of note, due to rather weak correlations of ST6GALs and glycomics data, we did not include these GSTs in our overview. Significant values are marked with * (p ≤ 0.05), ** (p ≤ 0.01,) and *** (p ≤ 0.001). Correlation coefficients and p-values are listed in the Additional file 2: Table S9

Intriguingly, for sialylation, depending on its linkage and involved glycan class, distinct and sometimes opposing correlations with TFs were observed. For instance, while α-2,6 sialylation on GSLs correlated with CEBPA and SPI1, an opposing correlation was observed for N- and O-glycans. Above all, highly significant positive correlations were found between α-2,8 sialylation and several TFs, a finding that was well reflected by sialyltransferase ST8SIA1 and ST8SIA6 expression. Interestingly, correlations of selected sialyltransferases and TFs showed a simpler picture: almost all investigated sialyltransferases exhibited significant correlation with either SPI1/CEBPA (negative correlation) or RUNX1/GATAs/TAL1/CBFA2T3 (positive correlation).

AML cell lines show distinct glycomic and transcriptomic signatures of M5 and M6 FAB classes

Sufficiently high numbers of cell lines belonging to either M5 or M6 were available to clearly associate FAB classification and glycomic profiles. In Fig. 5, we depict a detailed glycomic overview of these FAB classes integrating glycomics and transcriptomics data, including upstream GST and TF expression, which pinpoint the glycomic regulation of these AML subclasses resulting in their distinction. As mentioned earlier, a hallmark of the M5 subtype is its elevated (s)Lex/a levels, which may be dependent mostly on FUT7 expression and upstream SPI1/CEBPA expression. In contrast, the M6 subtype is characterized by especially high sialylation. To this end, ST3GAL2, ST6GALNAC1/4 and ST8SIA6 appeared to be the most correlated GSTs across all glycan classes, which in turn strongly correlated with the expression of RUNX1, GATA1/2/3, MECOM, TAL1, and CBFA2T3.

Fig. 5

Differences in glycan signatures of M5 and M6 AML cell lines as well as corresponding GST and TF expression. M5 and M6 classes are presented as grey and brown rectangles, respectively. GSTs displayed in the figure present a positive correlation with the corresponding glycosylation feature. The underlined TFs correlate with the glycosylation features. The underlined TFs colored in red are positively correlated with GSTs

GST and hematopoietic TF expression in primary AML cells

To explore the glyco-code of AML beyond the cell line models, we obtained transcriptomic data of AML primary cells from several previous studies and analyzed GST and TF expression as well as their associations (Additional file 2: Tables S10, S11, and S12). Based on GSTs and TFs that showed significant correlations in the cell line model (Fig. 4), we explored whether similar GST and TF correlation patterns existed in AML cell lines versus primary cells (Fig. 6a). After finding a good agreement as indicated by a modified RV2 coefficient [40] of 0.49, correlations between GSTs and TFs in primary cells and cell lines were visualized by heat maps (Fig. 6b). In line with our previous findings in AML cell lines, FUT4/7 showed a significantly positive correlation with SPI1 and CEBPA in primary blasts. Importantly, the broad positive correlations of sialyltransferases with RUNX1, GATA1/2/3, TAL1, and CBFA2T3 were also largely observed in the data obtained from primary patient material.

Fig. 6

GST and TF expression in primary AML cells. a Determination of the matrix correlation coefficient RV2 (0.49) between expression patterns observed in cell lines and primary samples. b Spearman correlation of selected GSTs with TFs in AML cell lines (left) and primary AML cells (right). c Comparison of the expression of selected GSTs and TFs in primary AML cells grouped by FAB classification. Significances were assessed by one-way ANOVA followed by a Tukey post-hoc test. Significant values are marked with * (p ≤ 0.05), ** (p ≤ 0.01), *** (p ≤ 0.001), and **** (p ≤ 0.0001)

Prompted by these good agreements, we sought to substantiate the observation that M5 and M6 represent distinctly glycosylated AML subtypes by looking into FAB-grouped primary cell data. Moreover, by relying on a compiled dataset of 640 FAB-classified patients we were curious to see how the expression patterns in these two subtypes appear in the context of M0-4 subtypes (Fig. 6c). The M7 subclass was excluded for this analysis as only two patients appertained to this rather rare class of AML in our datasets. In total, out of our pre-defined genes of interest, we could obtain information on 14 genes from this dataset, which are depicted in Fig. 6c, Additional file 1: Fig. S1, and Additional file 2: Table S13. In clear agreement with our previous findings, FUT7 showed high expression in M5 cells compared to M6. In accordance, FUT4 levels were significantly higher in M5 cell lines compared to M6 (Additional file 1: Fig. S1). Moreover, both ST6GALNAC4 and ST3GAL2 were elevated in M6 compared to M5; a finding that fits well with the results obtained from AML cell lines. Surprisingly, ST3GAL4, a sialyltransferase associated with biosynthesis of the sLex epitope, was not elevated in M6 but downregulated compared to most other subtypes. Also, M3 is characterized by elevated levels of this specific sialyltransferase. Based on ST3GAL1 expression, the M0 subclass, for which we did not obtain suitable cell line data, seems to be a FAB subtype with a rather high sialylation profile comparable to the M6 subtype (Additional file 1: Fig. S1). In addition, TF expression in primary cells was largely in agreement with the cell line models. CEBPA, a transcription factor highly correlated with both fucosyl- and sialyltransferases, was significantly upregulated in M5 blasts compared to M6. Again, the M3 subtype seems to resemble the phenotype of M5 also on the level of TFs. As anticipated and in contrast to CEBPA, the transcription factors GATA1 and TAL1 were highly expressed in M6. This characteristic seems to be rather unique to M6 cells compared to all other FAB classes.