GFPT2-Expressing Cancer-Associated Fibroblasts Mediate Metabolic Reprogramming in Human Lung Adenocarcinoma

Cancer is known to have altered metabolism through the glycolysis pathway to meet demands for tumor growth. This phenomenon, termed the “Warburg effect” (1) is widely accepted, yet alterations in tumor metabolism are not restricted to enabling tumor growth but also promoting tumor invasion and metastatic progression (2). This broader perspective has established metabolic reprogramming as a “cancer hallmark” (2, 3). Recent studies are revealing substantial metabolic heterogeneity in tumors (4, 5). Moreover, increasing consideration is being given to metabolic reprogramming of stromal cells comprising the tumor microenvironment (TME; ref. 6). For example, cancer-associated fibroblasts (CAF), a major component in tumor stroma, have been reported to have increased glycolysis and produce high-energy nutrients that facilitate biogenesis in malignant cells, a process referred to as the “reverse Warburg effect” (7). In support of these findings, studies have reported that malignant cells and CAFs express different monocarboxylate transporters (MCT) for the consumption and production of lactate (8). In this study, we provide new evidence of prognostically significant changes in the tumor stroma related to metabolic reprogramming of human non–small cell lung carcinoma.

Lung cancer remains the number one cause of cancer mortality in the United States, where non–small cell lung cancer (NSCLC) constitutes around 85% of lung cancer cases (9). NSCLC has two major histology subtypes, namely adenocarcinoma and squamous cell carcinoma (SCC). Adenocarcinoma and SCC are known to differ in their cell of origin and distribution across the lung (10–12). They also differ in terms of glucose uptake as measured by ¹⁸fluoro-2-deoxy-D-glucose PET (¹⁸FDG-PET) scans: SCC has been associated with higher levels of uptake than adenocarcinoma. Consistent with these findings, SCC has higher expression of the glycolysis markers GLUT1 (SLC2A1), CA9, and MCT1 (SLC16A1) (13). Even though higher glucose uptake in general is associated with more aggressive disease, SCC has been associated with better survival outcomes than adenocarcinoma among symptomatically detected patients (13). However, it remains unclear how histology-related differences in glucose metabolism are related to cancer progression.

We assembled a study cohort of patients with NSCLC for whom resected tumor specimens were acquired for transcriptomic analysis. To relate gene expression to metabolism, we also acquired a preoperative ¹⁸FDG-PET uptake feature, namely the maximum standardized uptake value (SUV_max), which is a common clinical measure of glucose uptake in human tumors in vivo. We found that genes associated with glucose uptake in adenocarcinoma versus SCC cases were functionally enriched for invasion versus growth, respectively. From the transcriptome of flow-sorted NSCLC, we identified genes correlated with glucose uptake that were predominately expressed in a single cell-type comprising the tumor microenvironment. For SCC, the majority of these genes were expressed in malignant cells, whereas they were expressed in stromal cells of adenocarcinoma. In adenocarcinoma, we focused on glutamine-fructose-6-phosphate transaminase 2 (GFPT2) because we found it to be a prognostically significant glucose-related metabolic gene that was predominately expressed in the tumor stroma. GFPT2 is a rate-limiting gene of the hexosamine biosynthesis pathway (HBP), known to glycosylate proteins yet underreported in its relevance to cancer. Our analysis identifies a significant role for GFPT2 in adenocarcinoma CAFs, in association with extracellular matrix remodeling mediated through the hexosamine biosynthesis pathway. Overall, our findings provide new evidence of the histology-specific tumor stromal functional properties associated with metabolic reprogramming in NSCLC and potential new therapeutic avenues.

We performed an integrative analysis of primary NSCLC with data obtained on medical imaging (¹⁸FDG-PET), tissue microarray, transcriptomics, protein expression, and survival outcomes across multiple cohorts and studies, as detailed below. Briefly, our primary cohort is our radiogenomics (RG) cohort (n = 130), on which we relate SUV_max and bulk tumor gene expression. In addition, we assembled a companion TME NSCLC cohort (n = 40), from which we used FACS to sort major cell types of the tumor microenvironment in order to determine which cell type, if any, predominately expressed a specific gene of interest. In addition, we constructed a separate validation cohort, tissue microarray analysis (TMA) cohort (n = 211), to validate the specific findings associated with glucose uptake and survival outcomes on the protein level. A workflow is provided in Supplementary Material S1; Supplementary Fig. S1. Throughout our analysis, we focus on various gene sets, including: genes correlated with glucose uptake are defined as genes significantly correlated to SUV_max in our RG cohort; cell-type–specific genes are defined as genes predominately expressed in one cell type comprising the tumor microenvironment, derived from our TME cohort; prognostically significant genes as determined by PRECOG database (14); and metabolic genes, which were collated from metabolic-related pathways in Kyoto Encyclopedia of Genes and Genomes (15) and HumanCyc (16). For the metabolic genes, we focused only on genes associated with glucose uptake through the glucose transporter gene SLC2A1 (GLUT1), which shunt glucose into one of the three major pathways (17): (i) the glycolysis pathway for energy production; (ii) the pentose phosphate pathway (PPP) for biomass production; and (iii) the HBP for protein glycosylation.

With Institutional Review Board (IRB) approval in accordance with U.S. Common Rule, we studied 130 patients with NSCLC who underwent curative surgical resection at Stanford University Hospital (Stanford, CA) or Palo Alto VA Medical Center (Palo Alto, CA) between year 2008 and 2015 and had a preoperative PET/CT imaging. We refer to this cohort as our RG cohort. Patient demographics of the study cohort were summarized in Table 1. Ninety-six patients with adenocarcinoma and 31 patients with SCC were included in the analysis (three cases were excluded due to undefined histology). Tissue samples ranged from 30 to 100 mg and were flash-frozen. Transcriptomic data were obtained through RNA sequencing (RNA-seq) with data alignment and expression estimation performed via Centrillion Biosciences, Inc (Supplementary Material S2). Preoperative PET/CT scans for tumor FDG uptake were paired to the tumor tissue specimens from which the RNA-seq data were generated. SUV_max was obtained by re-reading all 130 cases on MimVista software, guided by surgically reported information on excised lobe and slice to ensure the RNA-seq data and the annotated lesion were matched. Note that a preliminary analysis relating the transcriptome and the ¹⁸FDG-PET features on the first 26 of 130 patients comprising our current RG cohort was reported by us and others (18, 19). The current RG cohort RNA-seq data and imaging data are available at http://wiki.cancerimagingarchive.net/display/Public/NSCLC+Radiogenomics. A list of deidentified IDs of the patients used in the RG cohort is included in Supplementary Material S3.

Genes detected in fewer than 50% samples were removed. The estimated gene expression profile using FPKM (fragments per kilobase of transcript per million mapped reads) was log-transformed, and genes with variances in the lower 50% quartile were filtered out. RNA-seq profiling and data processing were performed in three batches through year 2014 to 2015. We observed significant batch effect between samples processed in 2014 and 2015 using principal component analysis (Supplementary Material S4; Supplementary Fig. S2) and applied COMBAT to adjust the batch effect (20). Differentially expressed genes (DEG) between adenocarcinoma and SCC samples were identified using R package “multtest” based on permutation multiple testing (21). A false discovery rate (FDR) of 5% was used to assess statistical significance.

Correlation between glucose uptake, measured in terms of SUV_max, and individual gene expression was assessed by Spearman rank test on adenocarcinoma and SCC samples separately. To reduce false-positive correlations, we used Significance Analysis of Microarray (SAM; ref. 22) to calculate an FDR for each gene based on permutation testing using gene expression as features and SUV_max as a continuous outcome variable. For adenocarcinoma, significant correlations were defined with FDR lower than 10%. For SCC, due to the relatively smaller sample size, we increased the FDR threshold for significance to 20%. Genes correlated to glucose uptake were clustered using R package Weighted Correlation Network Analysis (WGCNA; ref. 23) for adenocarcinoma and SCC separately. To identify the relevant biological functions of the gene clusters, we investigated the enriched gene ontology categories using MsigDB (24) and ToppFun (25). We focused on cancer-related hallmark gene sets, biological processes, and cellular components with FDR less than 5%.

With IRB approval in accordance with U.S. Common Rule, we studied 40 patients with NSCLC who underwent curative surgical resection at Stanford University Hospital or Palo Alto VA Medical Center between year 2008 and 2015. Hereon, we refer to this cohort as our TME cohort; this cohort enabled us to identify the cell-type–specific contributions to metabolic reprogramming underlying the tumor microenvironment. For each tumor sample in this cohort, we purified malignant and stromal cells. We isolated immune cells (CD45⁺, EPCAM⁻), endothelial cells (CD31⁺, CD45⁻, EPCAM⁻), malignant cells (EPCAM⁺, CD45⁻), and fibroblasts (CD10⁺, CD45⁻, EPCAM⁻, CD31⁻), via flow cytometry as described elsewhere (14). We performed RNA-seq for each cell type for every patient sample. We performed differential analysis on four cell types using SAM (22) to identify specific cell types in which the SUV_max-correlated genes were uniquely overexpressed with FDR less than 5%.

We examined the protein expression of GFPT2 (aka GFAT2) and glucose transporter SLC2A1 (aka GLUT1) by IHC using a NSCLC tissue microarray (n = 211). With IRB approval in accordance with U.S. Common Rule, patient samples were retrieved from surgical pathology archives at Department of Pathology, Stanford Medical Center (26). SUV_max data were obtained from linked clinical database STRIDE. Of the 211 patients on the TMA, 52 patients with adenocarcinoma had a ¹⁸FDG-PET scan for which SUV_max was reported. The GFAT2 antibody was validated on controls as shown in Supplementary Material 5; Supplementary Fig. S3. TMA staining intensity was assessed by qualitative ordinal scoring as 0 (negative), 1 (low), 2 (moderate), and 3 (high) for malignant cells and fibroblasts separately. An overall score was given to each sample based on the higher score between the malignant cells and fibroblasts. SUV_max was predicted using a general linear regression model based on staining scores for GFAT2 and GLUT1 as covariates. Survival analysis was performed using R “survival” package on all the patients with adenocarcinoma using GFAT2 overall scores and fibroblast scores. Risk groups were defined by low GFAT2 expression (score 0 and 1) and high GFAT2 expression (score 2 and 3). Cox proportional hazards regression was used to obtain the P value, HR, and confidence interval (CI). A multivariate Cox model was used to obtain the prognostic significance of GFAT2 adjusted by age, stage, tumor size, and gender.

Publicly available The Cancer Genome Atlas (TCGA) RNA-seq data on adenocarcinoma (n = 511) and SCC (n = 501) was used as an external cohort to verify DEGs between adenocarcinoma and SCC based on the RG cohort and to further analyze the relationship among genes correlated to glucose uptake identified in our RG cohort. TCGA gene expression profile was measured using Illumina HiSeq 2000 RNA-seq by the University of North Carolina genome characterization center, and RSEM normalized values were used for gene-level transcription estimates.

To identify potential genomic drivers of altered glucose uptake, we analyzed TCGA copy number variation (CNV) data of adenocarcinoma and SCC from UCSC Xena (27). CNV gene level data were processed by TCGA Firehouse pipeline (28, 29). In the UCSC Xena CNV data, 29 significant focal amplification and 46 significant focal deletion genomic regions were identified for adenocarcinoma, while 30 significant focal amplification and 53 significant focal deletion genomic regions were identified for SCC using GISTIC (30).

To validate our findings in CAFs, we used the Affymetrix microarray gene expression dataset (GSE60880; ref. 31) of normal fibroblast (NF) derived from normal human lung tissue treated with TGFβ. The NFs were incubated with TGFβ for 0.5, 1, 2, and 8 hours. CAF marker genes were used to assess the transformation of the NFs. We compared the expression of genes associated with glucose metabolism and SUV_max correlated genes enriched for epithelial-to-mesenchymal transition (EMT) between the NFs and the CAF-like cells.

To validate our findings on EMT, we analyzed the Affymetrix microarray gene expression dataset (GSE49644; ref. 32) of three cell lines (A549, HCC827, and NCI-H358) before and after TGFβ treatment. Three replicates were included for each condition in each cell line. Replicates for each condition were normalized and merged for meta-analysis. Permutation multiple test (multtest) was applied to identify genes that changed significantly before and after induced EMT with FDR < 0.05. In particular, we investigated genes involved in glucose metabolism and SUV_max correlated genes functionally enriched for EMT.

To investigate the relation between HBP rate-limiting gene GFPT2 (coding protein GFAT2) and EMT at protein level, we used Western blots with human lung adenocarcinoma cells (HCC827). To induce EMT, cells were cultured in the presence of 2% FBS and 10 ng/mL TGFβ up to 10 days. Fresh media containing TGFβ were replenished every 2 days, and cells were reseeded if needed before reaching confluency. Levels of histone H3 were used as an internal standard for equal loading. The blots were incubated with E-cadherin, vimentin as well as GFAT2. Additional experimental details were included in Supplementary Material S6.

In our RG cohort, we identified 657 DEGs between adenocarcinoma and SCC (FDR < 5%; Supplementary Material S7). Genes more highly expressed in adenocarcinoma were enriched for processes related to extracellular matrix remodeling, whereas genes more highly expressed in SCC were enriched for cell growth and proliferation processes; these findings were validated in TCGA (Fig. 1A and B; Supplementary Material S8; Supplementary Table S1). DEGs in glucose-driven metabolic pathways differed by histology (Fig. 1C). SCC showed higher expressed DEGs in the glycolysis and PPP, whereas adenocarcinoma showed higher expressed DEGs in the HBP. Because SCC had higher SUV_max than adenocarcinoma (Supplementary Material S9; Supplementary Fig. S4), we confirmed that the metabolic DEGs were histology specific and not reflective of differences in SUV_max (Supplementary Material S9; Supplementary Table S2).

Among metabolically associated DEGs more highly expressed in SCC were: (i) SLC2A1 (FDR < 0.0001), the glucose transporter I; (ii) G6PD (FDR = 0.0002), the gatekeeper gene of PPP; and (iii) PGD (FDR < 0.0001), an indicator of higher production of NADPH, an essential reductant in anabolic reactions. These genes have been associated with cell proliferation and growth (33), supporting the Warburg effect. Among metabolically associated DEGs more highly expressed in adenocarcinoma were: (i) GFPT1 (FDR < 0.0001), the rate-limiting gene of HBP, and protein glycosylation genes (ii) OGT (FDR = 0.004) and (iii) GALE (FDR < 0.0001); these genes suggest a unique role for HBP in adenocarcinoma that extends beyond the Warburg effect.

We identified genes correlated with SUV_max in adenocarcinoma and SCC samples separately from our RG cohort. In adenocarcinoma (n = 96), 169 genes were correlated to glucose uptake (FDR < 10%; Supplementary Material S10), of which 96 (57%) were found to be prognostic using PRECOG database (FDR < 5%; ref. 14). Using WGCNA to cluster the 169 genes based on their bulk gene expression profiles in adenocarcinoma samples in our RG cohort, we obtained 6 gene clusters. We performed functional enrichment on each cluster in adenocarcinoma (Fig. 2A), and we identified a gene cluster that was highly enriched for EMT (FDR = 3 × 10⁻³⁸). This particular cluster was also enriched for the extracellular matrix (FDR = 9 × 10⁻²⁶), suggesting a strong stromal factor. Interestingly, GFPT2, a rate-limiting gene in HBP, belonged to the gene cluster enriched for EMT and ECM. In SCC (n = 31), 141 genes were correlated with SUV_max (FDR < 20%), most of which were negatively correlated with SUV_max (Fig. 2B; Supplementary Material S11). Using WGCNA, we obtained two gene clusters based on bulk gene expression profiles in SCC samples. Functional enrichment analysis showed that SUV_max-correlated gene clusters in SCC were related to cell development (FDR = 6 × 10⁻³) and chemical homeostasis (FDR = 4 × 10⁻⁵). Among glucose-driven metabolic pathways in adenocarcinoma, we identified genes correlated to glucose uptake in the glycolysis pathway, including glucose transporter I (SLC2A1), PFKP, GAPDH, and lactate transporter gene (SLC16A1) and GFPT2 (Fig. 2C). Our findings were consistent with a previous study (19) that reported an association between increased glucose uptake and EMT in NSCLC based on bulk tumor gene expression and PET SUV_max; our findings extend the results of that study by showing that the association is histology specific and holds for adenocarcinoma, but not SCC.

In adenocarcinoma, among the genes correlated to glucose uptake, 55 genes were predominately expressed in a single cell-type of the tumor microenvironment: 25 in fibroblasts, 11 in immune, 17 in malignant cells, and 2 in endothelial cells, and several were shared across multiple compartments (Fig. 3A). Of the 55 adenocarcinoma genes predominately expressed in a single cell-type, 30 were prognostically significant, including GFPT2 (PRECOG, FDR < 5%). The vast majority of these prognostically significant genes were secreted factors (Supplementary Material S10), likely facilitating cell–cell cross-talk associated with disease progression. In SCC, 28 genes correlated to glucose uptake were largely confined to the malignant compartment (Fig. 3B), suggesting that glucose uptake is more associated with a reactive stromal in adenocarcinoma than SCC.

In our study, GFPT2 was the only prognostically significant gene in adenocarcinoma correlated to glucose uptake, which was predominately expressed in a single stromal compartment and associated with a glucose-driven metabolic pathway (Supplementary Material S10). We confirmed the prognostic significance of GFPT2 in TCGA (P = 0.002; HR = 1.27; CI, 1.1–1.5) for adenocarcinoma. Interestingly, GFPT2, a rate-limiting gene of HBP, showed the strongest expression in CAFs (Fig. 3C). To confirm the CAF enrichment of this sorted subpopulation in our TME cohort, we verified that this cell subpopulation uniquely overexpressed common CAF marker genes (ACTA2, FAP, FGF1, PDGFRB, COL1A1, and FN1; Fig. 3D). In TCGA, GFPT2 expression was strongly correlated to secreted glycoproteins that were also correlated to glucose uptake (Fig. 4A); many of these genes coding these glycoproteins are expressed by the CAF compartment in our TME cohort and were associated with EMT and extracellular matrix (ECM). This finding is consistent with a previous study that showed EMT could induce aberrant glycosylation through HBP activation (34). To validate the role of GFPT2 in EMT, we showed that at the protein level, GFAT2 (GFPT2-coded protein) increased in adenocarcinoma cells following TGFβ-induced EMT (Fig. 4B).

We validated the GFPT2-SUV_max association in CAFs in an external cohort based on TMA (Fig. 5A). We found that the TMA overall scores of protein GFAT2 (coded by GFPT2 gene) expression were predictive of SUV_max. Moreover, in the TMA, we observed that CAFs were more likely to express GFAT2 than malignant cells, but if malignant cells expressed GFAT2, then GFAT2 was expressed in the CAFs at a similar or higher level. Because glucose uptake is commonly associated with GLUT1 (SLC2A1) expression, we evaluated the predictive significance of both GLUT1 and GFAT2 for SUV_max. Interestingly, GFAT2 in CAFs alone and GLUT1 in malignant cells were both predictive of SUV_max (Supplementary Material S12; Supplementary Fig. S5). A side-by-side comparison of GLUT1 and GFAT2 in a representative TMA sample with high SUV_max showed GLUT1 confined to the malignant cells and GFAT2 to the fibroblasts (Fig. 5B). A representative case from our RG cohort with relatively high SUV_max expression showed that the GFPT2-expressing CAFs were located at the invasive edge of the tumor (Fig. 5C). We also found that GFAT2 scores overall and fibroblasts only were highly prognostic among patients with adenocarcinoma in the TMA cohort (n = 211) adjusted for age, stage, tumor size, and sex (Fig. 5D; Supplementary Material S12; Supplementary Table S3). In particular, prognostic significance of GFAT2 fibroblast score implicates the importance of reactive stromal as a potential therapeutic target for adenocarcinoma.

In several of our TMA samples, we also found GFAT2 expressed in the malignant cells. To investigate the genomic significance of this finding, we observed that GFPT2 gene is located on chromosome 5q region that was recurrently amplified in TCGA adenocarcinoma, but recurrently deleted in TCGA SCC (Supplementary Material S13; Supplementary Fig. S6). Interestingly, numerous genes correlated to glucose uptake are located in the same genomic region, including VCAN, coding extracellular matrix protein versican, and TGFBI and FBN2 that are both related to TGFβ signaling, a signaling pathway known to play an important role in EMT (35).

Because GTPF2 is a known driver of HBP and HBP has been associated with EMT, we hypothesized that GFPT2 expression would increase with TGFβ treatment in both NFs and malignant cells. Our rationale was based on the observations that TGFβ transforms NF to CAF-like cells and induces EMT in adenocarcinoma. We found TGFBI, known to be induced by TGFβ, was also correlated to glucose uptake (Supplementary Material S10). We investigated the expression of glucose metabolism genes and EMT genes associated with glucose uptake in human lung NFs before and after TGFβ treatment (GSE60880). Following TGFβ treatment, the NFs were transformed to CAF-like cells based on the gene expression of common CAF markers (Supplementary Material S14; Supplementary Fig. S7). Consistent with our hypothesis, the HBP genes were more expressed in the CAF-like cells than the NFs (Fig. 6A); in comparison, there were little to no change in genes related to glycolysis, PPP, and TCA cycle, suggesting no change in genomic regulation of these metabolic pathways with the transformation of NFs to CAF-like cells. Also, many EMT genes associated with glucose uptake were upregulated in the CAF-like cells (Fig. 6A).

With TGFβ treatment, NSCLC adenocarcinoma cell lines (GSE49644) have been shown to undergo EMT. We analyzed this transformation in terms of changes in glucose metabolism–related genes and EMT genes correlated to glucose uptake. We found that glucose metabolism genes associated with energy production and cell proliferation (glycolysis, PPP, TCA cycle) were mostly unchanged or reduced after TGFβ induced EMT, whereas several HBP genes and most EMT genes correlating glucose uptake were increased (Fig. 6B). This is consistent with our prior observation that GFAT2 (GFPT2-coded protein) increased adenocarcinoma cells following TGFβ-induced EMT (Fig. 4B).

In parallel with the cell line analyses, we analyzed expression of the metabolic genes and the EMT genes associated with glucose uptake across the different cell-specific compartments of adenocarcinoma in our TME cohort; we found metabolic genes associated with energy production and cell proliferation (glycolysis, PPP, TCA cycle) were strongly expressed in malignant cells, whereas HBP and EMT genes associated with glucose uptake were mostly expressed among the CAFs (Fig. 6C). Taken together, these findings indicate that altered glucose metabolism in tumor stroma through HBP is related to processes associated with EMT, possibly facilitating cancer invasion.

The best-characterized metabolic behavior of cancer is the Warburg effect; however, tumor metabolic reprogramming is not restricted to support tumor growth and proliferation, and it is not restricted to malignant cells. In this analysis, we provided evidence of metabolic reprogramming in the tumor stroma of adenocarcinoma that is prognostically significant. We applied a radiogenomic analysis, which integrated data on glucose uptake, as measured on ¹⁸FDG-PET scans, and transcriptomics of the bulk and flow-sorted human tumors. Our analysis revealed surprising differences in glucose metabolism between adenocarcinoma and SCC, the two major histologic subtypes of NSCLC. Compared with adenocarcinoma, SCC had more highly expressed genes in glycolysis and PPP, which are associated with Warburg effect, whereas adenocarcinoma had more highly expressed genes in the HBP, which have not been associated with the Warburg effect. For adenocarcinoma, genes correlated to glucose uptake were enriched for EMT and ECM. The glucose uptake and EMT was previously reported in NSCLC in the first 26 patients of our RG cohort (19), but we found it held only for adenocarcinoma, not SCC, given the larger size of current RG cohort (n = 130). For SCC, genes associated with glucose uptake were functionally enriched for tumor growth. Using cell-type–specific transcriptomics of the NSCLC TME, we found a stronger association of altered glucose uptake in the tumor stroma of adenocarcinoma compared with that of SCC. Taken together, our analysis confirmed a prior report that metabolic reprogramming in NSCLC is histology-specific (13) and extended this finding to show that the molecular pathways and cell types associated with glucose uptake were also histology specific.

We explored the significance of GFPT2 as a mediator of tumor metabolic programming in the stromal compartment of adenocarcinoma and a predictor of survival outcomes. GFPT2 was found to be prognostically significant in PRECOG, TCGA, and our TMA cohort for adenocarcinoma. GFPT2 and its isoform, GFPT1, are both rate-limiting genes of HBP; however, they have significant differences. GFPT1 is expressed in many cell types, but GFPT2 expression is thought to be more restricted to specific cell types (36). In a pathologic context, GFPT2 has been associated with diabetes, with few reports relevant to cancer, whereas GFPT1 has been well studied in the context of both diabetes and cancer (37). In our analysis, we found GFPT1 was more highly expressed in adenocarcinoma relative to SCC, but GFPT2 was correlated to glucose uptake in adenocarcinoma, suggesting that GFPT2 (not GFPT1) is associated with adenocarcinoma progression. We validated the correlation of GFPT2 and SUV_max on TMA of adenocarcinoma. The TMA also supported our finding that GFPT2 is largely expressed in CAFs but GFPT2 can be expressed in malignant cells. Because it was not common to find GFPT2-expressing malignant cells in the absence of GFPT2-expressing in CAFs, it is possible that GFPT2 expression in CAFs may be a precursor to GFPT2 expression in malignant cells for adenocarcinoma. Also, because GFPT2 is recurrently amplified in adenocarcinoma and deleted in SCC, adenocarcinoma may have a genomic propensity toward the HBP through GFPT2 among the malignant cells.

In adenocarcinoma, when we jointly analyzed SLC2A1 (GLUT1) expression in malignant cells and GFPT2 expression in CAFs, we found that GFPT2 expression in CAFs was an independent predictor of SUV_max, suggesting that the GFPT2-expressing CAFs may have increased glucose uptake. Prior work has shown that when NFs were transformed to CAF-like cells, through TGFβ treatment, glucose uptake increased (38). Interestingly, we did not observe any changes in the expression of GLUT1 when comparing NF and CAFs. Moreover, we did not observe significant GLUT1 expression in CAFs on TMA, even at high SUV levels (analysis not shown). It is possible that glucose uptake in CAFs may be altered through other regulatory mechanisms. We did observe that the insulin-regulatory gene IGFBP5 was correlated to glucose uptake and uniquely expressed in CAFs (Supplementary Material S10), suggesting that it may have a role in regulating glucose uptake (39).

Our findings on GFPT2 and the HBP provide an important link between altered metabolic reprogramming and altered cellular signaling in adenocarcinoma. HBP uses glycolytic intermediates to generate UDP-GlcNAc and UDP-GalNAc, which serve as the substrates for biosynthesis of glycoproteins, including many associated with promoting EMT and cancer invasion (40). For example, elevated O-glycosylated fibronectin (FN1) could induce EMT in human lung carcinoma cells (41); in our analysis, FN1 was among the genes correlated with glucose uptake and uniquely expressed by CAFs and one of many EMT genes associated with glucose uptake. Interestingly, because only 2% to 5% of glucose is typically shunted to the HBP (42), it is possible that a small absolute increase of glucose to the HBP could represent a large percentage change for HBP and have a dramatic effect. Moreover, if small changes in glucose flux to the HBP promote invasion, this could explain why adenocarcinoma has poorer prognosis for lower overall levels of glucose uptake compared with SCC. Understanding the impact of small alterations in glucose uptake to the HBP through carbon-labeled glucose flux analysis, in both the malignant and stromal compartment of tumors, promises to provide new insights into tumor metabolic reprogramming beyond the Warburg effect.

Although our study provides new insights with respect to tumor metabolic behavior, it has its limitations. First, the sample size of SCC in our RG cohort was small (n = 31); therefore, our SUV_max-gene correlation analysis in SCC had a relatively large FDR. This limited our ability to identify statistically significant genes in SCC that could be driving glucose metabolic alterations and cellular processes. Despite this limitation, we can conclude that the SCC likely does not involve EMT-SUV_max correlated genes because when we repeatedly subsampled the adenocarcinoma cases to match the sample size of the SCC cases, we still found the EMT association held with a P value <5% in adenocarcinoma (analysis not shown). Second, our analysis was limited to genes associated with maximum SUV, but other features of glucose uptake measured on PET, such as the total lesion glycolysis and SUV variation, are worthy of pursuit and may reveal additional insights (18). Third, on our TME cohort, the malignant cells were restricted to the EPCAM-positive subpopulation, and thus a more mesenchymal malignant cell type might be lost that could exhibit a higher expression of GFPT2 as found in TGFβ-induced EMT in adenocarcinoma (Fig. 4B). Moreover, because it has been shown that tumor cells that undergo EMT downregulate EPCAM and upregulate CD10 (43, 44), our CAFs sorted based on CD10⁺, CD45-, EPCAM⁻, and CD31⁻ markers could potentially include mesenchymal malignant cells, although any contamination by malignant cells seems minimal because our CAFs express canonical CAF markers (Fig. 3D). Finally, the TMA analysis was limited to 6 mm. Further analysis of whole slide images would provide a more comprehensive assessment of GFPT2 spatial distribution in the TME.

Our work highlighted CAFs as a major contributor to the metabolic reprogramming of the TME in adenocarcinoma; however, the endothelial and immune cells were also implicated in our analysis and deserve consideration going forward. Much of our focus was on GFPT2, but our analysis provided many more SUV_max-associated genes that deserve further exploration. We focused on GFPT2 because it is a metabolic enzyme that has been underreported in the cancer context. Its relevance in adenocarcinoma suggests that GFPT2 deserves more consideration in cancers and possibly beyond its role in the HBP. Our work provides a spotlight on the HBP for promoting tumor invasion in adenocarcinoma and may explain why the lower glucose uptake for adenocarcinoma relative to SCC carries worse prognosis. A more in-depth analysis of this behavior is warranted. Finally, our findings have therapeutic implications and suggest that hexosamine biosynthesis pathway and glycosylation inhibitors (such as ST060266 and tunicamycin; ref. 37) could be effective for adenocarcinoma, whereas SCC may be more effectively targeted by PPP and glycolytic inhibitors (such as 6-AN and 2-DG; ref. 45).

In summary, our integrative analysis showed that glucose uptake associated with GFPT2-expressing CAFs was prognostic for adenocarcinoma. In SCC, glucose uptake was associated with glycolysis and higher proliferation potential. In adenocarcinoma, we found a stromal component to glucose uptake implicating CAFs, endothelial, and immune cells. In adenocarcinoma, we focused our analysis on GFPT2 and its relation to hexosamine biosynthesis pathway that could bridge the altered glucose metabolism with protein glycosylation. These insights can provide a therapeutic approach for targeting tumor–stromal interactions associated with disease progression by disrupting the unique metabolic tumor microenvironment of adenocarcinoma.

M. Diehn reports receiving a commercial research grant from Varian Medical Systems, has ownership interest (including patents) in CiberMed, and is a consultant/advisory board member for Roche and CiberMed. S. Napel is a consultant/advisory board member for Carestream, Inc., Fovia, Inc., Radlogics, Inc, and EchoPixel, Inc. No potential conflicts of interest were disclosed by the other authors.

Conception and design: W. Zhang, S. Bakr, M. Diehn, S.K. Plevritis

Development of methodology: W. Zhang, G. Bouchard, S. Napel, S.K. Plevritis

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): G. Bouchard, M. Shafiq, J.B. Shrager, K. Ayers, S. Bakr, R.B. West, M. van de Rijn, S. Napel, S.K. Plevritis

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): W. Zhang, A. Yu, M. Shafiq, A.J. Gentles, M. Diehn, A. Quon, V. Nair, S.K. Plevritis

Writing, review, and/or revision of the manuscript: W. Zhang, G. Bouchard, A. Yu, M. Shafiq, J.B. Shrager, S. Bakr, A.J. Gentles, M. Diehn, A. Quon, R.B. West, V. Nair, M. van de Rijn, S. Napel, S.K. Plevritis

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): M. Jamali

Study supervision: M. Diehn, V. Nair, S. Napel, S.K. Plevritis

We are grateful to Dr. Ann Leung for providing a valuable review of our manuscript. This work was supported by the NIH grants R01 CA160251 and U01 CA154969.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.