AML-stroma proteome unveils a protein network mainly associated with metabolic function

BMSCs support leukemogenesis and allow leukemic cells to evade chemotherapy-induced cell death; however, the protein-signaling network leading to leukemic growth and cell survival remains unknown. To unravel early (pre-apoptotic) changes in protein abundance regulated by leukemia-stroma interactions, we cultured primary AML (n = 14) in the absence (AML Mono) or presence of HS-5 stroma cells (AML Cocu) and measured an untargeted global proteomic profile using LC-MS/MS (Fig. 1A, Supplementary Table 1, and Supplementary Dataset 1). A summary of patient clinical information such as gender, age, cytogenetics, and mutation status is presented in Supplementary Table 1. Overall, the average protein counts in each sample AML Mono (2359 ± 430) and AML Cocu (2540 ± 333) were equally represented. A total of 2231 unique proteins were identified and quantified. We determined the differentially expressed proteins with a fold change of >1.5 (FDR 0.05 cutoff) between AML Cocu and AML Mono (Fig. 1B) of which 114 proteins were significantly differentially regulated. In comparison to HS-5 monoculture, a minimal overlap of proteins was unique to stroma. To gauge molecular responses triggered by AML-stroma interaction, gene set analysis shows an upregulation of membrane and cell differentiation markers including ICAM1, ITGB1, SPN, ANPEP, and CD44. The differentially regulated proteins were also subjected to DAVID and ClueGo for enriched molecular pathways. Thirty-eight of 114 proteins are associated with upregulation of metabolic processes and associated with enrichment of nucleic acid metabolism, translation/protein biosynthesis, mitochondrial matrix, glycolysis and gluconeogenesis, fatty acid biosynthesis, and signaling receptor activity (Fig. 1C). Rate limiting NAD biosynthesis factor NAMPT, related to energy metabolism, was the most stroma-dependent upregulated protein across the 14 AML samples. Down regulated proteins (25/144) were enriched for NADP, oxidoreductase, mitochondrion, and lipid metabolism. The primary AML cells stimulated by stroma interaction reveal a proteomic network predominately representing a signature of enriched metabolic pathways including an upregulation of early membrane signaling molecules.

Fig. 1: The proteomic profiles of AML samples are metabolically impacted by the interactions with BMSC.

A A schematic diagram depicts how AML protein extracts were created from 14 primary AML cultured overnight in the presence and absence of HS-5 stroma cells, referred to as AML Mono and AML Cocu, respectively. Both AML Mono (n = 14) and AML Cocu (n = 14) were collected and protein lysates were processed for LC-MS/MS to detect proteins involved in signaling pathways stimulated by stroma. B A volcano plot depicts the differential protein expression in AML Cocu with a ≥1.5 fold change relative to AML Mono. Upregulated, downregulated, and proteins below the cutoff of 1.5 fold change are displayed in red, blue, and gray, respectively. C A bar plot depicts top biological pathways that are activated in primary leukemic cells in response to stroma interaction (Supplementary Dataset 1).

Stoma-dependent protection of leukemic cells was most evident with HDAC inhibitor treatment

Recurrent mutations of epigenetic modifiers that appear in pre-relapse and post-relapse AML cases indicate the importance of understanding epigenetic modalities in drug resistance [34, 35]. AML-stroma interactions also lead to increased drug resistance that might be in part dependent on deregulation of leukemic epigenetic pathways. Thus to further understand pro-survival/anti-apoptotic response in AML, we modulated epigenetic pathways to give insight into how leukemia becomes resistant through stroma interactions [36, 37]. To this end, we performed an epigenetic drug screen targeting a wide array of epigenetic modifiers (80-epigenetic compound library). First, proliferative growth was measured for each compound in 9 leukemia cell lines. The compound library was then clustered by drug target function, which included histone deacetylase inhibitors (HDACi), histone methyltransferases (HTMi), bromodomain (BETi), DNA methylase (DNMTi), and sirtuins (SIRTi). Inhibitors targeting histone deacetylases significantly reduced cell growth across all 9 leukemia cell lines relative to untreated cells, indicating that targeting histone acetylation may achieve the greatest effect on leukemic subtypes (Supplementary Fig. 1).

To model leukemia-stroma interactions, we then determined the effect of apoptosis on leukemic KG1a cells in monoculture (KG1a Mono) versus KG1a coculture with stroma HS-5 cells (KG1a-HS-5 Cocu). We treated KG1a Mono and KG1a-HS-5 Cocu again with the 80 epigenetic compound library for 48 h (Fig. 2A, B). Treatment with histone deacetylase inhibitors (10 μM), in particular Apicidin (C3), M344 (A8), CBHA (H7), SAHA (C11), and Oxamflatin (B8) significantly induced apoptosis in KG1a monocultures, whereas in coculture apoptosis was reduced by 10–50%. Of the 30 HDACi compounds surveyed, these five compounds primarily target HDAC Class I. Stroma-induced protection of leukemia cells was most evident with Apicidin (Api) treatment, which effectively induced apoptosis in monocultured leukemia (84%) than in cocultured leukemic cells (31%, Fig. 2A, B). The protective properties of stroma were further verified by coculture and treatment of five HDACi (A8, B8, C3, C11, and H7) in other leukemia cell lines and primary AML (n = 10, Supplementary Fig. 2). In 6 of 10 AML primary cocultures, at least 40% apoptosis levels were reached in monoculture with HDACi treatment but exhibited significantly reduced apoptosis levels in coculture. These results indicate that cocultured leukemic cells treated with HDACi may be protected from HDACi cytotoxicity in a stroma-dependent manner.

Fig. 2: The drug sensitivity differences between AML Mono and AML Cocu identify that HDACi compounds promote stroma-dependent protection of leukemia.

A A schematic diagram is depicting the flow cytometer gating approach for monitoring apoptosis levels in KG1a Mono and KG1a-HS-5 Cocu. CD71-APC was used to label stroma (HS-5 cells) thus CD71-APC negative cells trace leukemia cells. Annexin V-FITC was used to determine the extent of apoptosis. An example is given of stroma protection of leukemia cells treated for 48 h with HDACi, Api (10 µM). The apoptosis levels of KG1a Mono (4%) and KG1a-HS-5 Cocu (21%) untreated are compared with KG1a Mono HDACi treated at 81% versus KG1a-HS-5 Cocu at 17.7%, respectively. B Summary of apoptosis levels in KG1a Mono and KG1a Cocu following treatment with an 80 epigenetic compound drug library. The x and y-axes represent the percent of apoptosis levels in KG1a Cocu and KG1a Mono, respectively. The colors indicate drug classes and were at a 10 μM concentration for each compound. The plot depicts the significance of at least 3–6 replicate measurements of the data where the labeled points indicate ≤0.05 P Value determined by a paired t-test. The gray dash distinguishes the slope of 1 where less than 1 signifies % apoptosis coculture is less than monoculture (stroma-dependent protection trend). Supplementary Dataset 2 lists the compound description used in this study.

Identification of phosphorylation events as stroma protective effects

The marked differences between HDACi sensitivity in leukemia monocultures compared to cocultures led us to investigate signaling mechanisms associated with cellular regulation through phosphorylation events. To this end, we applied mass spectrometry (MS)-based discovery phosphoproteomics to define the earliest leukemia-stroma activated pathways under protective coculture conditions. We treated leukemia monoculture and leukemia-stroma coculture with HDACi for 5 h, a pre-apoptotic phase with apicidin, (termed KG1a+Api and Coculture+Api, respectively), and used untreated monocultures and untreated coculture as reference signals. Multiplex isotope labeled lysates coupled with LC-MS/MS analysis was used to quantify PTMs of leukemia-stroma cocultures in comparison with monoculture cells including HDACi treated and untreated (Fig. 3A and Table 1). Relative quantification was only compared within the light, medium, and heavy signals and cross-validated across three independent labeling mixes (Mix 1, 2, or 3, isotope label scheme described in Table 1). The multiplex isotope labeling approach identified 2381 unique phosphosites from 1319 proteins with 684 phosphosites exceeding a log2-fold change of ±2 (Fig. 3B and Table 1). In order to determine the differentially phosphorylated sites leading to stroma-dependent protection of leukemic cells, relative quantification of phosphosite intensities of KG1a+Api (baseline) were subtracted from Coculture+Api (Mix 2). To identify highly relevant phosphosites we then surveyed the top 1% differentially phosphorylated sites (log2-fold change higher than >5 or lower than <5, Fig. 3C). Interestingly, two highly ranked phosphosites involving acetylation metabolism were identified: acyl-CoA synthetase 2 (ACSS2), which was up-regulated at phosphosite S30 (fold change: 6), and acetyl-CoA carboxylase alpha (ACACA), which was down-regulated at phosphosite S80 (fold change: −6.8) relative to KG1a+Api. Cross-validation in Mix 1 of ACSS2 (S30) indicated a >2-fold increase in intensity compared with Cocu+Api and Cocu relative to HS-5 Mono. Protein blot using phospho-ACACA (S80) antibody supports in validating these results in a decrease ACACA (S80) expression in Cocu (±Api) relative to KG1a and KG1a+Api (Supplementary Fig. 3A, B). Enzymatically, ACSS2 mediates cytosolic acetyl-CoA production from acetate and coenzyme A and the phosphorylation of S30 affects insulin signaling [38], whereas acetyl-CoA carboxylase alpha (ACACA) catalyzes the carboxylation of acetyl-CoA to malonyl-CoA in fatty acid synthesis [39] and its function is suppressed by phosphorylation of S80 in mice [40].

Fig. 3: The leukemia-stroma phosphoproteome reveals biological pathways leading to stroma-dependent protection of leukemia cells.

A Discovery phosphoproteomic tools were used to identify and determine relative amounts of phosphorylated phosphosites in KG1a Mono, HS-5 Mono, and KG1a-HS-5 Cocu. Briefly, following a 24h-culturing period, cells were treated with HDACi Api at 1 μM for 5 h, a sub-apoptotic time point, and then harvested to create protein lysates. Protein extraction of each lysate, digestion, and isotope labeling was performed on each experimental condition. To compare phosphosite peptide intensities between culturing conditions, a multiplex labeling scheme of light, medium, and heavy-labeled lysates were used. Light, medium, heavy labeled lysates were pooled into a single mix termed Mix 1, 2, or 3. Each mix was then enriched for phosphopeptides using TiO2. LC-MS/MS resolved the enriched phosphopeptides. This approach allows for direct comparison of phosphopeptide intensities within a mix consisting of light, medium, heavy isotope tags. B Phosphopeptide intensities between Coculture+Api in comparison with KG1a+Api were compared to depict upregulated phosphopeptides in increasing red hues and downregulated phosphopeptides in increasing blue hues. A scatter plot depicts log2-transformed values of phosphopeptides (n = 2381). Phosphosites having a missing value (no expression or undetected) were given a minimum intensity value for determining differential phosphorylation (log2-transformed LFQ values). C The proteome of KG1a Mono, HS-5 Mono, and KG1a-HS-5 Cocu label-free was determined in parallel to cross-validate enriched phosphopeptides. The top 1% phosphopeptide fold change (Coculture+Api minus KG1a+Api) is depicted by gene name and phosphopeptide fold change (blue). The corresponding protein abundance is plotted with the phosphopeptide in the stacked bars (gray). The fold changes are relative to KG1a+Api. D Differentially phosphorylated phosphosites with at least log2-transformed difference of 3 (309 phosphosites) compared with monoculture KG1a+Api were uploaded to ClueGo to determine enriched functionally grouped biological networks, P ≤ 0.05. The node size of the biological theme correlates with increasing significance. Functionally related groups overlap with color. For clarity, a subset pathway is depicted with gene name (circle indicates hyperphosorylation and triangle indicates hypophosphoryation) demonstrating the possible signal transduction of glycolysis proteins (red nodes) linking with AMPK signaling (blue nodes) in response to stroma-dependent protection of leukemia. This network highlights regulated genes mapped to the significantly enriched phosphosites, including ACSS2 (S30) and ACACA (S80). The complete biological pathways (≤0.05 P Value) of enriched phosphopeptides impacted by stroma-dependent protection of leukemia are depicted in Table 2.

To assess how the phosphorylation profiles were associated with protein abundance, we measured and compared protein abundance in KG1a monoculture, HS-5 monoculture, and KG1a-HS-5 coculture. Relative quantification of protein abundance was determined by subtracting KG1a protein abundance intensities (baseline) from KG1a-Hs5. Thus, we uncovered 5.8-fold higher ACSS2 protein levels in coculture relative to KG1a monoculture having no detection value and were indeed in agreement with phosphorylation levels (Fig. 3C). Therefore, it is likely that up-regulation of ACSS2 (S30) in coculture stems from within the stroma cellular compartment. In a similar vein, ACACA protein levels were 2.3-fold lower in coculture compared to KG1a monoculture, mirroring the decrease in phosphorylation levels (Fig. 3C).

To gain a broader understanding of biological processes and pathways, we performed enrichment analyses of 3-fold differentially up- and down-regulated phosphoproteins using ClueGo with KEGG, Reactome, and gene ontology (GO) categories as resource databases. The most affected networks and pathways were PKB-dependent events, PI3K, mTOR signaling, glycolysis, EGF/EGFR, and AMP-activated protein kinase (AMPK) signaling (P ≤ 0.03, Table 2 and Fig. 3D, highlighted AMPK/mTOR signaling). Key phosphoproteins in AMPK/PI3K/mTOR signaling were stimulated by stroma-dependent protection with HDACi treatment of leukemia for which its protein network kinase activities have been previously shown to protect leukemia cells by suppressing oxidative stress in the bone marrow microenvironment [41]. Taken together, phosphorylation events leading to induce stroma-dependent protection of leukemia were identified by modulation with HDACi treatment.

Table 2 Significantly enriched biological pathways of top differentially phosphorylated phosphosites determined by Cytoscape ClueGo with Reactome, KEGG, and WikiPathways as ontology sources of stroma-dependent protection of leukemia cells.CRISPR-edited ACSS2 knockout in stroma modulates leukemia in an ACSS2-dependent manner

The phosphoproteome screen indicated that the acetyl-CoA processing enzymes ACSS2 and ACACA might be metabolically supporting leukemia through stroma interactions. Due to the up-regulation of the ACSS2 (S30) signal from cocultured stroma cells, we wanted to address the ACSS2 functional role of in a stroma-dependent manner. An antibody to follow ACSS2 (S30) phosphorylation was not commercially available therefore another approach was taken, single-guide RNAs (sgRNAs) with CRISPR/Cas9 gene editing were designed to target the phosphosite S30 of ACSS2 which impacted ACSS2 expression in HS-5 stroma cells (termed HS-5 ACSS2-KO, Supplementary Fig. 4A). Surprisingly, HS-5 ACSS2-KO cells had a greater proliferative potential relative to wild type cells implicating a loss of ACSS2 could change proliferative rates in non-cancer cells (Supplementary Fig. 4B).

To test whether the loss of ACSS2 in stroma cells (HS-5 ACSS2-KO) impacted the viability of leukemic cells, we first examined apoptosis levels by treating leukemia cocultures with histone deacetylase inhibitors (Api, CBHA, and SAHA), with only API and SAHA treatment resulting in partial restoration of apoptosis in KG1a cocultured with HS-5 ACSS2-WT (Supplementary Fig. 4C). This result indicated that compensating metabolic pathways in HS-5 ACSS2-KO cells contributing to this path of proliferative growth and resistance may be involved [42].

ACSS2 is known to be important for tumor metabolism in hypoxic environments [43,44,45]. Likewise, in this study, AML-stroma proteomic analysis indicated upregulation of key metabolic pathways. We then wanted to examine the potential gain of metabolic function of leukemic cells gained from stroma contact including the role of ACCS2. In order to measure metabolic potential, OCR of leukemic cells was measured live in real-time. Leukemic cells were pre-exposed to HS-5 ACSS2-WT or to HS-5 ACSS2-KO overnight to allow cell-cell contacts then leukemic cells were gently removed, re-seeded in RPMI assay media for recalibration. Any contaminating stroma cells re-adhered to plastic and then only suspended leukemic cells were transferred to an assay plate at equal cell densities and subjected to hypoxia for 45 min for downstream mitochondrial respiration measurements. Leukemic monocultures were cultured in parallel as a control. The saturation oxygen levels (maximal OCR) were compared to the basal OCR to distinguish differences in mitochondrial respiration in leukemic cultures. Apparent steady rate increases of OCR in KG1a (2-fold), OCI-AML3 (1.5-fold) and K562 (1.5-fold) pre-exposed to HS-5 ACSS2-WT relative to monocultures were observed (Fig. 4A). In contrast, KG1a, OCI-AML3 and K562 pre-exposed to HS-5 ACSS2-KO had a markedly diminished OCR (2.7-fold, 1.7-fold, and 1.5-fold, respectively, Supplementary Fig. 4D) compared to leukemia cells pre-cultured with HS-5 ACSS2-WT. The results from these analyses indicate that leukemic cells pre-exposed to wild-type stroma gain metabolic respiratory capacity in a hypoxic environment dependent on ACSS2.

Fig. 4: ACSS2 and metabolic substrate acetate modulate leukemic proliferation and metabolic fitness.

A Seahorse XF Cell Mito Stress assay was used for assessing mitochondrial respiration of leukemia cells cultured overnight with stroma cells that were CRISPR-edited for ACSS2 (HS-5 ACSS2-KO) and compared with HS-5 WT. Leukemia cells (KG1a, OCI-AML3, or K562), monocultures, and cocultures were gently removed, equilibrated with modified Seahorse XF Cell Mito Stress RPMI buffer in a non-CO2 chamber at 37 °C, and re-platted into the provided manufacturer’s microwell plate to measure mitochondrial respiration OCAR output with Seahorse XFe96 instrument. Leukemia monoculture (red), leukemia pre-cultured with HS-5 (blue), and leukemia pre-cultured with HS-5 ACSS2-KO (clone 1, green or clone 2, purple) OCR is displayed at 5 min time points (mean ± SEM). Key parameters of oxidative respiration (mitochondrial respiration) are displayed as changes in rate mode at basal respiration, ATP-linked production, maximal respiration, and spare respiratory capacity by sequential injections of Oligo, FCCP, Rot/Ant. B Exogenous acetate treatment improved viability in primary AML blast specimens significantly in 6 of 9 specimens (paired t-test, P ≤ 0.05, indicated by an asterisk). C mRNA expression of ACSS1, ACSS2, and ACACA was measured in AML blasts (n = 46, solid circles). Measured in triplicate by standard SYBR Green quantitative RT PCR, plotted values display the relative mRNA expression measured relative to GAPDH and HS-5 (positive control). D Proliferative growth of OCI-AML3 WT and OCI-AML3 ACSS2-KO monocultures treated with acetate at 12.5 μM and measured at 0, 24, and 48 h (left plot). K562 WT and K562 ACSS2-KO were treated as OCI-AML3 (right plot). A solid black line depicts the proliferative growth of untreated WT cells while the black dashed line depicts the proliferative growth with acetate treatment. The blue lines depict the proliferative growth of OCI-AML3 ACSS2-KO cells untreated (solid line) and acetate treated (dashed). The red lines depict the proliferative growth of K562 ACSS2-KO cells untreated (solid line) and acetate treated (dashed).

The ACSS2 substrate acetate increases cell viability in leukemia

Extracellular acetate, a precursor of acetyl-CoA, has also been shown to contribute substantially to the increase in acetyl-CoA concentrations under hypoxic conditions through ACSS2 activity [45,46,47,48]. Therefore, we subjected cultured leukemic cell lines (KG1a, OCI-AML3, and K562) to an increasing concentration of acetate (0.5–5 μM) for 48 h to determine if leukemic proliferative growth was affected. The impact of acetate treatment on proliferative growth was 1.2- to 1.6-fold greater than on untreated leukemic cells (Supplementary Fig. 4E). To further validate this finding, we cultured primary AML blasts (n = 9) with exogenous acetate (12.5 µM; 48 h), which also resulted in a proliferative growth advantage of 1.2–1.4-fold in comparison to untreated AML blasts (Fig. 4B).

From the proteomic screen, we uncovered that KG1a ACSS2 protein abundance was undetected (confirmed by RT PCR), although acetate treatment enhanced proliferative growth in monocultures. We reasoned that ACSS1, an important paralog of ACSS2, would function as an alternate gene in processing acetate into acetyl-CoA in the absence of ACSS2 [44, 49]. Therefore, we probed ACSS1 mRNA expression in primary AML and, indeed, found that 32 of the 56 specimens expressed ACSS1 at a staggering ~log2 fold change of 32 higher than ACSS2 which was expressed in 8 of 56 specimens (Fig. 4C). Half of the cohort expressed elevated levels of ACACA relative to HS-5. Thus, surveying primary AML ACSS1 and ACSS2 expression profiles and the effects of exogenous acetate enhancement on proliferative growth yields the possibility that leukemic cells may uptake acetate from exogenous sources such as the microenvironment and process the metabolite through ACSS1or ACSS2.

Given that HDAC inhibitors were used to model drug resistance of AML, we examined acetylation profile of KG1a cells with histone mark H3K9, a post-translation modification associated with active transcription, to ask whether HDACi or acetate treatment transmits epigenetic modulation. By flow cytometry, we treated KG1a with HDACi (Api, 1 µM) for 48 h and measured intracellular levels of histone protein H3 and acetyl-modified histone residue, H3K9. We found that HDACi-treated KG1a cells resulted in increased expression of H3 and H3K9ac relative to unstained cells (Supplementary Fig. 5A, B). KG1a cells were also treated with acetate (48 h, 10 µM) resulted in a similar pattern of increased H3 and H3K9ac implicating that leukemic cell exposure to HDACi or exogenous acetate may play a vital role in the overall histone acetylation landscape of leukemic cells as previously shown in other cancer types [50, 51].

Leukemic proliferative growth stimulated by exogenous acetate is dependent upon ACSS2

To test whether acetate uptake in leukemic cells was dependent on ACSS2 expression, ACSS2 single-guide RNAs (sgRNAs) were again utilized to knockout ACSS2 expression in two leukemia cell lines expressing ACSS2, OCI-AML3, and K562. Monocultures of leukemic wild-type (WT) cells treated with exogenous acetate elicited a proliferative growth advantage compared to untreated cells, however proliferation was significantly diminished in ACCS2-KO leukemic cells (Fig. 4D). Our data support the view ACSS2 expression in leukemia is aimed at processing exogenous acetate to support proliferative growth.

High expression of ACSS1/ACSS2 predicts poor overall survival in AML

We further explored the clinical impact of ACSS1, ACSS2, and ACACA expression in AML in terms of overall survival. We reasoned that the expression of ACSS1 and ACSS2 would carry clinical relevance based on biochemical function. Therefore, we examined the mRNA expression of ACSS1 and ACSS2 using AML TCGA RNA-seq dataset [30] that included different cytogenetic AML subtypes (n = 156). We defined two groups of AML samples using patients with highest gene expression values of ACSS1 (n = 10) and highest gene expression values of ACSS2 (n = 10) denoted as top ACSS1/2 samples (n = 20). These cases displayed inferior outcome survival (OS) probability compared to the remaining samples (top ACSS1/2 median OS was 9.2 months versus Others with median 16.4 months; p = 0.028, Fig. 5A). We found 1225 genes differentially expressed between these two groups (Fig. 5B). Using K-means clustering to statistically partition these expression patterns into two groups matching the expression pattern of 1225 differentially expressed genes, we obtained a broader group of patients with ACSS1/2-high expression levels (Group 2, n = 60) and with ACSS1/2-low expression levels (Group 1, n = 96, Fig. 5B). K-means clustering remained statistically significant in further partitioning into 4 groups. Interestingly, the two-group cluster also displayed different survival probabilities with an inferior outcome for patients in ACSS1/2-high (Group 2, median OS of 7.8 months versus Group 1 median OS of 24.8 months, p = 7E-05, Fig. 5C). The up-regulated genes include enrichment of 83 (14%) mitochondrial proteins (Fig. 5D and listed in Supplementary Dataset 5). ACSS1/2-high patients were significantly older and encompassed almost all the molecular subtypes, and significantly harbored high molecular risk AML cases including TP53 mutated AML. ACSS1/2-high was of prognostic significance in a uni- and multivariate analysis (Supplementary Table 2A, B) and remains significant when including only intensively treated patients in the ACSS1/2-high group (p = 0.042).

Fig. 5: ACSS2 or ACSS1 high expression in the AML TCGA cohort predicts poor prognosis and relates to a distinct metabolic signature.

A Kaplan-Meier estimation of the overall survival shows the poor outcome of patients in the top ACSS1/2 expression (red line, n = 20) compared to others (blue line, n = 136, log-rank test). Dotted lines mark the 95% confidence bounds. B Heatmap representation of differentially expressed genes in the TCGA AML dataset. Patients were plotted in the columns with the differential expressed genes in the rows. Top 20 ACSS1/ACSS2 expression samples (n = 20 patients) were plotted in the horizontal rows above the heat map. Groups 1 and 2 show the division of the TCGA cohort into two groups using unsupervised clustering of this gene expression matrix (K-means for 2 groups) resulting in 1226 differentially expressed genes. Genes were classified as high or low according to the average expression of a gene in Group 2. C Kaplan-Meier estimation of the overall survival shows the poor outcome of patients in Group 2 (red line) compared to Group 1 (blue line, log-rank test). Dotted lines mark the 95% confidence bounds. D Functional annotation and enrichment cluster score from DAVID Bioinformatic Resources of genes up and downregulated with ACSS1/2 high expression are plotted.