Altered Gene Expression Along the Glycolysis– Cholesterol Synthesis Axis Is Associated with Outcome in Pancreatic Cancer Joanna M

Published OnlineFirst September 3, 2019; DOI: 10.1158/1078-0432.CCR-19-1543

CLINICAL CANCER RESEARCH | PRECISION MEDICINE AND IMAGING

Altered Gene Expression along the Glycolysis– Cholesterol Synthesis Axis Is Associated with Outcome in Pancreatic Cancer Joanna M. Karasinska1, James T. Topham1, Steve E. Kalloger1,2, Gun Ho Jang3, Robert E. Denroche3, Luka Culibrk4, Laura M. Williamson4, Hui-Li Wong5, Michael K.C. Lee5, Grainne M. O’Kane6, Richard A. Moore4, Andrew J. Mungall4, Malcolm J. Moore5, Cassia Warren1, Andrew Metcalfe1, Faiyaz Notta3, Jennifer J. Knox6, Steven Gallinger3,6, Janessa Laskin4,5, Marco A. Marra4,7, Steven J.M. Jones4,7, Daniel J. Renouf1,5,8, and David F. Schaeffer1,2,9

ABSTRACT ◥ Purpose: Identification of clinically actionable molecular Results: On the basis of the median normalized expression of subtypes of pancreatic ductal adenocarcinoma (PDAC) is key glycolytic and cholesterogenic genes, four subgroups were iden- to improving patient outcome. Intertumoral metabolic het- tified: quiescent, glycolytic, cholesterogenic, and mixed. Glyco- erogeneity contributes to cancer survival and the balance lytic tumors were associated with the shortest median survival in between distinct metabolic pathways may influence PDAC resectable (log-rank test P ¼ 0.018) and metastatic settings (log- outcome. We hypothesized that PDAC can be stratified into rank test P ¼ 0.027). Patients with cholesterogenic tumors had the prognostic metabolic subgroups based on alterations in the longest median survival. KRAS and MYC-amplified tumors had expression of genes involved in glycolysis and cholesterol higher expression of glycolytic genes than tumors with normal or synthesis. lost copies of the oncogenes (Wilcoxon rank sum test P ¼ 0.015). Experimental Design: We performed bioinformatics analysis Glycolytic tumors had the lowest expression of mitochondrial of genomic, transcriptomic, and clinical data in an integrated pyruvate carriers MPC1 and MPC2. Glycolytic and cholestero- cohort of 325 resectable and nonresectable PDAC. The resectable genic gene expression correlated with the expression of prognos- datasets included retrospective The Cancer Genome Atlas tic PDAC subtype classifier genes. (TCGA) and the International Cancer Genome Consortium Conclusions: Metabolic classification specific to glycolytic and (ICGC) cohorts. The nonresectable PDAC cohort studies included cholesterogenic pathways provides novel biological insight into prospective COMPASS, PanGen, and BC Cancer Personalized previously established PDAC subtypes and may help develop OncoGenomics program (POG). personalized therapies targeting unique tumor metabolic profiles.

specific cellular tumor progression pathways contribute to PDAC Introduction prognostic stratification is needed to enable customized treatment The 5-year survival rate in pancreatic ductal adenocarcinoma design and novel therapeutics development. (PDAC) is less than 10% and remains one of the lowest in all Oncogene-driven metabolic adaptations allow cancer cells to sur- cancers. Effective therapy is limited by the treatment-refractory vive and thrive in the tumor microenvironment (9). A pan-cancer nature of PDAC and a short supply of clinically validated biomar- analysis of global metabolic pathways showed that tumor metabolic kers capable of predicting treatment response (1). Emerging molec- heterogeneity is associated with survival, somatic driver mutations, ular subtypes of PDAC have defined intertumoral heterogeneity and tumor subtypes (10), but whether heterogeneity in distinct met- at the genome and transcriptome levels (2–6), driving efforts to abolic pathways can be used to stratify PDAC into clinically relevant identify clinically relevant biomarker signatures and actionable subgroups has not been well established. A vast majority of PDACs genomic alterations (7, 8). However, a better understanding of how harbor oncogenic KRAS and loss-of-function TP53 mutations (6), in

1Pancreas Centre BC, Vancouver, British Columbia, Canada. 2Department of Clinical trial information: Personalized OncoGenomics (POG) Program of Pathology and Laboratory Medicine, University of British Columbia, Vancouver, British Columbia: Utilization of Genomic Analysis to Better Understand Tumour British Columbia, Canada. 3Ontario Institute for Cancer Research, Toronto, Heterogeneity and Evolution (NCT02155621); Prospectively Deﬁning Metastatic Ontario, Canada. 4Canada's Michael Smith Genome Sciences Centre, Vancouver, Pancreatic Ductal Adenocarcinoma Subtypes by Comprehensive Genomic British Columbia, Canada. 5Division of Medical Oncology, BC Cancer, Vancouver, Analysis (PanGen; NCT02869802); Comprehensive Molecular Characterization British Columbia, Canada. 6University Health Network, University of Toronto, of Advanced Pancreatic Ductal Adenocarcinoma for Better Treatment Selection Toronto, Ontario, Canada. 7Department of Medical Genetics, University of British (COMPASS; NCT02750657). Columbia, Vancouver, British Columbia, Canada. 8Department of Medicine, Corresponding Author: David F. Schaeffer, University of British Columbia, University of British Columbia, Vancouver, British Columbia, Canada. 9Division Vancouver General Hospital, 910 West 10th Avenue, Vancouver V5Z 1M9, of Anatomic Pathology, Vancouver General Hospital, Vancouver, British Colum- Canada. Phone: 604-875-4480; Fax: 604-875-5707; E-mail: bia, Canada. [email protected] Note: Supplementary data for this article are available at Clinical Cancer Clin Cancer Res 2019;XX:XX–XX Research Online (http://clincancerres.aacrjournals.org/). doi: 10.1158/1078-0432.CCR-19-1543 J.M. Karasinska, J.T. Topham, S.E. Kalloger, D.J. Renouf, and D.F. Schaeffer contributed equally to this article. 2019 American Association for Cancer Research.

AACRJournals.org | OF1

Karasinska et al.

NCT02750657) studies at the BC Cancer Agency (PanGen, POG) Translational Relevance and Ontario Institute for Cancer Research (COMPASS) as Pancreatic ductal adenocarcinoma (PDAC) has one of the described previously (7, 29). The PanGen and POG studies were lowest survival rates of all cancers due in part to a limited approved by the University of British Columbia Research Ethics knowledge of clinically relevant tumor subtypes that would facil- Committee (REB# H12-00137, H14-00681, H16-00291) and itate better treatment stratification and the development of new the COMPASS study was approved by the University Health therapies targeting unique molecular signatures. Tumor metabolic Network Research Ethics Board (REB# 15-9596). The studies were heterogeneity contributes to clinical outcome in cancer and repre- conducted in accordance with international ethical guidelines. sents a potential avenue for the development of personalized Written informed consent was obtained from each patient prior treatment strategies. How alterations in distinct metabolic path- to genomic profiling. ways influence PDAC outcome is not well known. We profiled the expression of glycolytic and cholesterogenic genes in 325 resectable Whole genome and transcriptome sequencing and non-resectable PDAC patients and identified distinct sub- POG and PanGen samples were subjected to whole genome and fl groups associated with differences in survival and known prog- transcriptome sequencing as described previously (30). Brie y, nostic pancreatic tumor subtypes. Our findings demonstrate that fresh tumor biopsies were sequenced to a depth of approximately distinct metabolic gene expression pathways may provide a func- 80 , with approximately 200 million reads generated for transcriptional correlate to transcriptome-based pancreatic cancer subtypes, tomes. RNA sequencing (RNA-seq) libraries were prepared using fi which may enable the development of subtype-specific treatment inactive magnetic bead-based mRNA puri cation. RNA-seq was strategies targeting unique metabolic vulnerabilities. performed on COMPASS samples as described previously (7). TCGA (PAAD-US) and ICGC (PACA-CA) data Normalized RNA-seq data (sequence-based gene expression; GRCh37) for all available PAAD-US (n ¼ 142) and PACA-CA addition to prevalent hypoxia (11), which are known inducers of the (n ¼ 234) samples were downloaded from the ICGC data portal – glycolytic pathway in cancer (12 15), and glycolysis contributes to (dcc.icgc.org/) on November 8, 2018 (ICGC data release 27). PACA- tumor progression and chemoresistance in PDAC (13, 16, 17). The CA samples were filtered to exclude any samples labeled as cell lines, effects of glycolysis on tumor progression can be diminished by xenografts, metastatic, normal, or non-laser microdissected enriched. diverting the metabolite pyruvate from conversion to lactate in part PAAD-US samples were filtered to exclude non-PDAC samples as through transport into the mitochondria via the activity of the outlined in a previous study (6). mitochondrial pyruvate complex (MPC), comprised of pyruvate car- Somatic mutation data [those with both copy number variants – rier 1 and 2 (MPC1 and MPC2; refs. 18 20). Reduced MPC activity is (CNV) and single nucleotide variants/indels (SNV/indels) available; associated with poor prognosis in some cancer types (20). Pyruvate is a GRCh37] for all filtered samples with RNA-seq data available (PAAD- metabolic intermediate for the tricarboxylic cycle, providing the US n ¼ 60, PACA-CA n ¼ 86) were downloaded from the ICGC data precursor citrate for lipogenesis including cholesterol and free fatty portal on November 8, 2018. acid biosynthesis (9, 21). Oncogene-induced activation of the mevalonate pathway, which is essential for de novo cholesterol synthesis and RNA-seq data processing contributes to tumor growth (22), supports the use of pathway POG RNA-seq data were aligned (GRCh37) as described previ- inhibitors such as statins in cancer. However, studies on the association ously (30). COMPASS RNA-seq reads were trimmed to 75 bp using of statins with cancer risk or survival have produced mixed fastx (hannonlab.cshl.edu/fastx_toolkit/; parameters: -l 75 -Q33) results (23–26), and statin response heterogeneity may be linked to and aligned to GRCh37 using BWA-mem v.0.7.6 (31) and JAGuaR distinct tumor molecular signatures (22, 24, 27). The association of v.2.2 (default parameters; ref. 32). Raw read counts for POG MPC1 and MPC2 expression with cancer outcomes (20) raises the and COMPASS samples were assigned to Ensembl 75 genes using possibility that intertumoral differences in pyruvate flux and the Subread v.1.4.6 (parameters: -T 3 -s 1 -C -t ‘gene‘;ref.33).POG balance between glycolysis and cholesterol synthesis could regulate and COMPASS gene expression values were RPKM normalized and þ tumor aggressiveness. PDAC cell lines have distinct glycolytic and log-transformed (log10(RPKM 1)). PAAD-US normalized gene lipogenic profiles which affect their response to metabolic agents (28); expression values were converted to transcripts per million (TPM) 6 þ however, whether gene expression heterogeneity in discrete metabolic and log-transformed (log10((normalized_count 1e ) 1)). PACA- pathways influences clinical outcome or underlies actionable meta- CA normalized gene expression values were log-transformed þ fi bolic vulnerabilities in PDAC is not completely understood. (log10(normalized_count 1)). All samples were ltered to exclude Here, we stratify PDAC into subgroups based on the expression those with tumor content values <30%. patterns of glycolytic and cholesterogenic genes, and report their association with survival, mutational, and prognostic gene expression Batch correction signatures. Data from POG, COMPASS, PAAD-US, and PACA-CA cohorts were batch-corrected using gene-wise location scaling within each Materials and Methods cohort (34). Principal component analysis of the top 25% (n ¼ 4,534) most variable genes confirmed a reduction in between-sample batch fi Patients were enrolled in the ongoing Prospectively De ning effects after normalization (Supplementary Fig. S1). Metastatic Pancreatic Ductal Adenocarcinoma Subtypes by Comprehensive Genomic Analysis (PanGen, NCT02869802), Metabolic subgrouping the BC Cancer Personalized OncoGenomics program (POG, Genesbelongingtomolecularsignaturesdatabase(mSigDB; NCT02155621), and Comprehensive Molecular Characterization ref. 35) gene sets “REACTOME_GLYCOLYSIS” (n ¼ 29) and of Advanced PDAC For Better Treatment Selection (COMPASS, “REACTOME_CHOLESTEROL_BIOSYNTHESIS” (n ¼ 24) were

OF2 Clin Cancer Res; 2019 CLINICAL CANCER RESEARCH

Metabolic Gene Expression Heterogeneity and Survival in PDAC

used as glycolytic and cholesterogenic genes, respectively. Consen- alternate copy number fits for each set of clusters. We then fitted a sus clustering was performed on glycolytic and cholesterogenic linear model to each set of selected clusters and computed the penalty genes using ConsensusClusterPlus (36) v1.38 (parameters: reps ¼ for each model. The penalty is an empirically derived function of the 100, pItem ¼ 0.8, pFeature ¼ 1). Ward.D2 and Euclidean distances standard error of fit, proportion of data explained by fit, and the were used as the clustering algorithm and distance metric, respec- relative proportion of predicted versus observed copy number states. tively, with k ¼ 3. Median expression levels of coexpressed glyco- Next, allelic frequencies of heterozygous germline SNPs within the lytic and cholesterogenic genes were used to assign quiescent selected clusters were used to determine the absolute copy number of (glycolytic 0, cholesterogenic 0), glycolytic (glycolytic > 0, each cluster. Assuming that all reads mapping to homozygously cholesterogenic 0), cholesterogenic (glycolytic 0, cholestero- deleted regions originate from normal tissue contamination, the genic > 0), and mixed (glycolytic > 0, cholesterogenic > 0) metabolic number of reads mapping to homozygous regions was proportional subgroups to each sample. to the normal contamination. This value was used to compute tumor content. When tumor content was not able to be predicted compu- Classification of preexisting PDAC subtypes tationally, the pathology-measured value was used. Samples were classified according to Collisson (3), Bailey (2), and Tumor content values for COMPASS and PACA-CA samples were Moffitt (4) subtypes using consensus clustering, as outlined by a determined as described previously (7). Tumor content (ABSOLUTE; previous study (7). Collisson subtyping was based on the 62 gene ref. 37) for PAAD-US samples were downloaded from the GDC data signature in the original publication (3). Bailey subtyping was based portal on February 1, 2016. For the pan-TCGA analysis, tumor content on genes from differential expression analysis results from the values were obtained from a previous publication (38). original publication (2), which were filtered for genes with an P < > adjusted 0.05 and log2 fold change 0, resulting in 240 ADEX, Mutation analysis 1,061 squamous, 268 progenitor, and 370 immunogenic genes. All mutation data were derived from hg19. SNV/indels and Moffitt subtyping was based on genes belonging to the 50 gene CNVs in POG, COMPASS, and PACA-CA samples were identified signature in the original publication (4). For each subtyping pro- as described previously (7, 30). Relative to tumor ploidy, DNA cedure, samples underwent consensus clustering based on each segments with copy state 4and1 were considered amplifica- classifiers genes, followed by semi-automatic subtype assignment tions and deletions, respectively. PAAD-US mutation data were based on gene expression patterns (7). downloaded February 1, 2016, from the GDC data portal. PAAD- US copy number events were filtered for those with at least 10 MPC1/2 analysis supporting probes and a segment mean >0.2 (amplifications) or When comparing MPC1/2 mRNA levels between metabolic <0.2 (deletions), as recommended by a previous study (39). subtypes, Levene and Bartlett tests confirmed that differences in Coordinatesofcopynumbereventsweremappedtogenecoding variance were not statistically significant (P < 0.05) between groups, regions using Bedtools v2.26. For contingency analyses, SNVs and and Shapiro–Wilk tests followed by BH test correction confirmed CNVs were tested individually for each gene. For each subgroup, that MPC1/2 expression values within each group was not signif- each of the 12 genes were tested, computing a Fisher exact test to icantly (adjusted P < 0.05) different from a normal distribution. determine whether loss-of-function mutations or copy number Levene, Bartlett, Shapiro–Wilk, one-way ANOVA, and Tukey amplifications/deletions were enriched in that subtype. Resulting honestly significant difference (HSD) tests were performed in R P values were subjected to BH test correction. v3.3.2. To identify genes positively and negatively correlated with MPC1/2 Pan-TCGA RNA-seq analysis expression levels, all genes assayed by RNA-seq (n ¼ 16,733) were then RNA-seq data [RNA-seq by expectation maximization (RSEM); tested for correlation (Spearman) with either MPC1 or MPC2,andP GRCh37] for all non-PAAD TCGA samples was downloaded from values were subjected to Benjamini–Hochberg (BH) multiple test the GDC data portal on November 23, 2015. Samples were filtered correction. Genes were identified as being positively or negatively for those belonging to cancer types with at least 100 samples, correlated with both MPC1/2 if correlations with both genes were leaving 17 cancer types. Expression values were log-transformed fi P < r > n ¼ þ signi cant (threshold of adjusted 0.01) and had 0( 713) (log10(RSEM 1)) and batch correction was applied using gene-wise or r < 0(n ¼ 303), respectively. location scaling within each cancer type. To identify pathways enriched among genes positively and nega- For each cancer type individually, we repeated consensus clustering tively correlated with both MPC1/2, we performed a comprehensive (ConsensusClusterPlus, parameters: reps ¼ 100, pItem ¼ 0.8, pFeature gene set enrichment analysis (GSEA) on both groups of genes. GSEA ¼ 1; Ward.D2 and Euclidean distance, k ¼ 3) using expression values was performed in R using hypergeometric tests followed by BH of genes belonging to the “REACTOME_GLYCOLYSIS” and “REAC- multiple test correction, and all 18,026 gene sets downloaded from TOME_CHOLESTEROL_BIOSYNTHESIS” gene sets. For each clus- mSigDB (35) were tested. ter, the proportion of glycolytic and cholesterogenic genes was calculated, and clusters were considered “core” clusters if they consisted Calculation of tumor content of >50% of either gene set. For cancer types with more than one core For POG samples, a novel statistical modeling algorithm was used to cluster of the same gene set, the most homogenous cluster was determine tumor purity from whole genome sequencing data (Culibrk considered core. Cancer types that did not have core glycolytic and and colleagues., in preparation). Briefly, the germline genomes were cholesterogenic clusters of at least 75% homogeneity were omitted divided into segments with equal coverage at a threshold. Next, from further analysis (BRCA, CESC, KIRP, PRAD, SKCM, and somatic read counts within these bins were enumerated and clusters UCEC), leaving 9 cancer types (Supplementary Fig. S2). Metabolic were identified by univariate density approximation following kernel subtypes were determined for each individual cancer type based on density estimation. Each cluster corresponded to a distinct copy median values of their respective core glycolytic and cholesterogenic number state within the genome. We generated models that described genes.

AACRJournals.org Clin Cancer Res; 2019 OF3

Karasinska et al.

Survival analysis reprogramming has been associated with changes in the relevant Kaplan–Meier plots were generated using R packages “survival” metabolites (10). Our results indicate that there are multiple v.2.4.2 (40) and “survminer” v.0.4.2 (41). Samples with an overall metabolic phenotypes relevant to the glycolysis-cholesterol synthe- survival of less than one month were omitted from survival analyses. sisaxisinPDAC,wheretumorswithhigherratesofglycolysisbut lower cholesterol synthesis may be more aggressive or less sensitive Sequencing data availability to chemotherapy than tumors with a more cholesterogenic Genomic data generated within the POG and COMPASS phenotype. studies are being submitted to the European Genome-phenome Archive (EGA) under accession numbers #EGAS00001001159 and Association of metabolic subtypes with tumor genomic profiles #EGAS00001002543, respectively. and known PDAC subtypes Molecular events such as oncogenic mutant KRAS and MYC amplification, and mutations in TP53 are capable of driving Results metabolic reprogramming in cancers including PDAC (13, 43, 44). Dual analysis of glycolytic and cholesterogenic gene expression To determine oncogenic events across the different metabolic sub- identifies four distinct subgroups of PDAC types, we investigated the frequency of SNV, indels, and CNV To stratify PDAC tumors based on their relative expression levels of affecting genes frequently mutated in PDAC (2, 6, 7) across glycolytic and cholesterol synthesis pathway genes, we utilized RNA- the metabolic subtypes (Fig. 2A). Although differences in the seq data from resectable and nonresectable PDAC patient tumors. The frequency of mutations in each gene were not significantly different resectable datasets included TCGA PAAD-US and ICGC PACA-CA. across subtypes (Fisher exact test with BH test correction, adjusted The non-resectable cohort studies included COMPASS, PanGen and P > 0.05), we noted a significant increase in median glycolytic gene POG. To enrich for tumor cell-specific mRNA, samples were filtered to expression in samples with both KRAS and MYC amplification exclude those with tumor content < 30% (Supplementary Fig. S3), (Wilcoxon rank sum test P ¼ 0.015; Fig. 2B). There was a moderate leaving a total of 325 patient samples (TCGA n ¼ 61, ICGC n ¼ 148, correlation of the expression of glycolytic genes with KRAS (Spear- COMPASS n ¼ 90, PanGen/POG n ¼ 26). Genes belonging to man correlation r ¼ 0.27, P ¼ 6.6e7) and a weak correlation Reactome gene sets “glycolysis” (n ¼ 29) and “cholesterol biosyn- with MYC (Spearman correlation r ¼ 0.15, P ¼ 0.007) expression thesis” (n ¼ 24) were selected for analysis. Previous studies have (Fig. 2C). These findings are compatible with the notion that KRAS demonstrated heterogeneity in metabolic gene expression, includ- and MYC drive tumor metabolism toward glycolysis in PDAC and ing isoenzymes within specific pathways, between different cancer suggest that tumors with KRAS and MYC copy gain may be more types (10, 42). To aid in selecting genes co-regulated within each dependent on glucose utilization and vulnerable to glycolytic pathway and relevant to PDAC biology, we used consensus clus- inhibition. tering to identify two groups of robustly co-expressed glycolysis Gene expression analysis of other PDAC metabolic reprogram- (n ¼ 14) and cholesterol synthesis (n ¼ 15) pathway genes to be ming drivers revealed increased expression of HIF1A and associated used for metabolic subtyping (Fig. 1A). Median expression levels of glycolysis genes LDHA and SLC16A3 in the glycolytic group and coexpressed glycolytic and cholesterogenic genes were calculated increased expression of the sterol synthesis transcriptional activator for each sample and used in assigning one of four metabolic profiles SREBF2 in the cholesterogenic group (Supplementary Fig. S6). The specifically relevant to these two pathways: quiescent, glycolytic, expression of PRKAA1, which encodes one of the catalytic subunits cholesterogenic and mixed (Fig. 1B). Expression levels of glycolytic of AMP-activated protein kinase (AMPK), was reduced in glycolytic and cholesterogenic genes across the metabolic subgroups are tumors. The cholesterogenic tumors had reduced expression of visualized in Fig. 1C. The quiescent phenotype defined the largest AKT2 and AKT3,aswellasGLS, the gene encoding glutaminase. group of cases (101/325; 31.1%), followed by mixed (79/325; 24.3%), This indicates that components of the PI3K/AMPK and glutamine cholesterogenic (73/325; 22.5%) and glycolytic (72/325; 22.2%). metabolism pathways are uniquely dysregulated in glycolytic and Differences in the proportion of samples belonging to each subtype cholesterogenic tumors. PDAC has been associated with reduced were not statistically significant (Fisher exact test P ¼ 0.15) between uptake and utilization of branched-chain amino acids (45) and resectable and metastatic samples. In order to determine if stroma KRAS-dependent upregulation of the pentose phosphate and nucle- type impacts the metabolic classification, we performed clustering otidemetabolismpathways(13).Analysisofthesenetworksdem- analysis of the cases using normal and activated PDAC stroma gene onstrated reduced expression of genes involved in amino acid signatures (Supplementary Fig. S4; ref. 4). There were no significant catabolism, nucleotide metabolism and pentose phosphate path- differences in the distribution of metabolic subgroups across the ways in the quiescent subgroup (Supplementary Fig. S7), suggesting two stroma subtypes. The glycolytic profile was associated with the that these tumors may overall represent a PDAC subtype with low shortest median survival in both resectable (log-rank test P ¼ 0.018) metabolic activity. and metastatic (log-rank test P ¼ 0.027) PDAC (Fig. 1D), consis- Previous studies have identified PDAC gene expression signa- tent with the role of glycolysis in tumor progression (9). Interest- tures associated with survival (2–4). Subtypes associated with poor ingly, a survival benefit was observed in cases with increased outcome include the basal-like (Moffitt and colleagues; ref. 4), expression of cholesterol synthesis genes. Cholesterogenic cases squamous (Bailey and colleagues; ref. 2), and quasi-mesenchymal had the longest median overall survival in the resectable (log- (Collisson and colleagues; ref. 3) classifications. To investigate rank test P ¼ 0.0031 vs. glycolytic, P ¼ 0.043 vs. mixed, P ¼ whether expression patterns across the glycolytic-cholesterogenic 0.025 vs. quiescent) and metastatic (log-rank test P ¼ 0.011 vs. axis could underlie the differences between previously established glycolytic) groups. Mixed cases had a significantly better outcome subtypes, we determined the various PDAC subtypes for each than glycolytic cases in the metastatic group (log-rank test P ¼ sample (7) and investigated their degree of overlap with the 0.039). Cholesterogenic cases had the longest relapse interval in metabolic phenotypes (Fig. 3A). The quiescent group included resected PDAC (Supplementary Fig. S5). Metabolic gene expression predominantly classical (Moffitt and colleagues) cases (72.3%) and

OF4 Clin Cancer Res; 2019 CLINICAL CANCER RESEARCH

Metabolic Gene Expression Heterogeneity and Survival in PDAC

Metabolic subgroup A Cluster B Cholesterogenic Mixed 12 3 Quiescent Glycolytic Pathway 2 Glycolytic Cholesterogenic Cluster Consistency 1 0 0.4 0.8

Coexpressed 0 Cholestero- Genes genic Genes Median cholesterogenic gene expression ( Z -score) −1 Coexpressed Glycolytic Genes −1 01 Genes Median glycolytic gene expression (Z-score)

C Gene type Glycolytic Cholesterogenic Gene expression (Z-score) −2 024 Quiescent Glycolytic Cholesterogenic Mixed GAPDH ALDOA PKM ENO1 TPI1 PGK1 GPI PGAM1 PFKP PFKFB3 ENO2 PPP2R5D PFKM PFKFB4 FDPS FDFT1 DHCR24 EBP IDI1 MVD HMGCS1 SQLE NSDHL DHCR7 HMGCR LSS SC5D MVK HSD17B7 PDAC samples D Resectable PDAC Metastatic PDAC 100

50 Percentage of

patients alive (%) 25

P = 0.018 P = 0.027 0 0 24 48 72 96 120 144 0122436 Time (months) Number at risk 561340000 29 10 2 0 40431000 27 5 0 0 36 15 7 1 1 1 1 21 8 2 1 541851000 16 6 0 0

Figure 1. Stratification of PDAC tumors based on expression of glycolytic and cholesterogenic genes. A, Heatmap depicting consensus clustering solution (k ¼ 3) for glycolytic and cholesterogenic genes in resected and metastatic PDAC samples (n ¼ 325). B, Scatter plot showing median expression levels of coexpressed glycolytic (x-axis) and cholesterogenic (y-axis) genes in each PDAC sample. Metabolic subgroups were assigned on the basis of the relative expression levels of glycolytic and cholesterogenic genes. C, Heatmap depicting expression levels of coexpressed glycolytic and cholesterogenic genes across each subgroup. D, Kaplan–Meier survival analysis of patients with resectable (left) and metastatic (right) PDAC stratified by metabolic subgroup. Log-rank test P values are shown. contained the highest frequency (Fisher's exact test with BH test 6.3e12) and quasi-mesenchymal (59.7%, adjusted P ¼ 2.3e10). correction) of ADEX (17.8%, adjusted P ¼ 0.015) and exocrine-like The cholesterogenic subgroup was characterized by the lowest (23.8%, adjusted P ¼ 4.9e-3) cases (Fig. 3B). The Collisson and proportion of either of the poor prognosis signatures (basal-like: colleagues’ exocrine-like signature is associated with high expres- 5.5%, adjusted P ¼ 4.9e7; quasi-mesenchymal: 5.5%, adjusted P ¼ sion of genes encoding digestive enzymes (3), suggesting that the 2.1e6; squamous: 2.7%, adjusted P ¼ 6.2e7), and highest pro- quiescent group may include a higher proportion of cases with a portion of progenitor cases (83.6%, adjusted P ¼ 1.9e5). This distinct cell type of origin, specifically those involved in digestive finding is in line with the progenitor subtype reported to be enzyme secretion. Consistent with the glycolytic phenotype con- enriched in genes involved in steroid hormone biosynthesis (2), ferring the worst outcome, the majority of these cases were basal- a pathway downstream of cholesterol synthesis. Glycolytic gene like (62.5%, adjusted P ¼ 9.6e11), squamous (58.3%, adjusted P ¼ expression was positively correlated with Moffit and colleagues’

AACRJournals.org Clin Cancer Res; 2019 OF5

Karasinska et al.

A Quiescent Glycolytic Cholesterogenic Mixed KRAS SNV/INDEL Type TP53 Missense Nonsense SMAD4 Frameshift INDEL CDKN2A Inframe KMT2C INDEL Other KMT2D Wild type

RNF43

ARID1A CNV type Gain GNAS Loss Neutral TGFBR2 NA RREB1

MYC Samples B C KRAS MYC KRAS/MYC copy status P = 6.6e-7 P = 0.007 ρ ρ Both vs. neither P = 0.015 = 0.27 = 0.15

1 1

0 0 Median glycolytic −1 gene expression ( Z -score)

−1

−20 2 −2024

Median glycolytic gene expression ( Z -score) Gene expression (Z-score) Neither One Both amplified amplified amplified

Figure 2. Mutational landscape across metabolic subgroups of PDAC. A, Oncoprint depicting the distribution of somatic mutation (SNV/indel) and copy number variation (CNV) events affecting frequently mutated genes in PDAC across the metabolic subtypes. B, Box plot illustrating median expression of glycolytic genes in samples with KRAS and/or MYC copy number ampliﬁcation. C, Scatter plot depicting the correlation between median glycolytic gene expression and KRAS (left) and MYC (right) expression.

basal-like genes (Spearman correlation r ¼ 0.42, P ¼ 2.7e15), tumor glycolytic activity and lactate generation (20). To investigate the while negatively correlated with Moffit and colleagues’ classical relationship between MPC1 and MPC2 and the glycolytic and cho- genes (Spearman correlation r ¼0.28, P ¼ 3.6e7; Fig. 3C). In lesterogenic phenotypes, we compared mutation frequencies and contrast, cholesterogenic gene expression was positively correlated expression levels of both genes across the metabolic subgroups. We with Moffit and colleagues’ classical genes (Spearman correlation noted a paradoxical relationship between CNVs in each gene, with r ¼ 0.35, P ¼ 1.2e10),andshowedaweaknegativecorrelation CNVs affecting MPC1 being exclusively deletions while the majority with Moffit and colleagues’ basal-like genes (Spearman correlation (16/17) of CNVs affecting MPC2 were amplifications (Fig. 4A). The r ¼0.15, P ¼ 0.006). Taken together, these data implicate a role expression levels of both MPC1 and MPC2 were found to be signif- for distinct tumor metabolism pathways as mechanisms contrib- icantly different (one-way ANOVA P ¼ 6.7e4 and 8.8e9, respec- uting to the prognostic effects of the known PDAC subtypes, and tively) across metabolic subgroups. Both MPC1 and MPC2 gene identify glycolysis and cholesterol synthesis as potential targetable expression was significantly reduced in glycolytic samples compared metabolic vulnerabilities in the different subtypes. with each of the other groups (Tukey's honestly significant differences (HSD) test adjusted P < 0.01) and the expression of MPC2 was MPC complex as a potential regulator of the tumor glycolysis– significantly increased in cholesterogenic cases compared to quiescent cholesterol synthesis axis samples (Tukey HSD test adjusted P ¼ 0.006; Fig. 4B), suggesting that The MPC complex regulates mitochondrial pyruvate flux and dysregulation of mitochondrial pyruvate transport at the mRNA level suppression of MPC1 and MPC2 expression in cancer cells promotes may contribute to the metabolic tumor subtypes. Consistent with the

OF6 Clin Cancer Res; 2019 CLINICAL CANCER RESEARCH

Metabolic Gene Expression Heterogeneity and Survival in PDAC

Moffitt subtype (inner) Basal-like Classical

Mixed Quiescent Bailey subtype (middle) Squamous

PDAC Immunogenic samples Progenitor

Cholester- ADEX ogenic Glycolytic

Collisson subtype (outer) Quasi-mesenchymal Exocrine-like

Classical

BC Moffitt subtypes Bailey subtypes Glycolytic genes Cholesterogenic genes P = 2.7e-15; ρ = 0.42 P = 0.006; ρ = -0.15 * 75 *** 1 *** 50 ** 0 25 * −1 ** * ** 0 basal-like genes ( Z -score) Median expression of Moffitt Collisson subtypes P = 3.6e-7; ρ = -0.28 P = 1.2e-10; ρ = 0.35 genic Mixed

* Quiescent Glycolytic 75 Cholestero- 1

Percentage of samples (%) *** Moffitt subtype Collisson subtype 50 Basal-like Quasi- 0 mesenchymal * Classical 25 * Exocrine-like Bailey subtype −1 Classical * ** Squamous

0 classical genes ( Z -score) Immunogenic *** adj. P < 1e-10 Median expression of Moffitt −2 ed x ** adj. P < 1e-5 −1 01 −1 012 genic Mi Progenitor Median expression (Z-score) Quiescent Glycolytic * adj. P < 0.05 Cholestero- ADEX

Figure 3. Alignment of PDAC metabolic subgroups with known gene expression subtypes. A, Overlay of metabolic profiles (inner ring) with PDAC expression subtypes (outer rings) based on the Moffitt and colleagues’ (4), Bailey and colleagues’ (2), and Collisson and colleagues’ (3) classifications. B, Bar plots illustrating the proportion of PDAC expression subtypes across each metabolic subgroup. C, Scatter plot depicting reciprocal correlations between expression of glycolytic and cholesterogenic genes and genes associated with the Moffitt and colleagues’ (4) basal-like and classical subtypes. hypothesis that cytoplasmic pyruvate is shuttled toward lactate gen- Table S1). GSEA of positively correlated genes revealed significant eration in glycolytic tumors, LDHA levels were higher in tumors with association (hypergeometric test, BH-adjusted P < 0.05) with the increased expression of glycolytic genes (Supplementary Fig. S6). oxidative phosphorylation pathway. Pathways enriched among neg- To search for cellular pathways associated with MPC1/2 expression, atively correlated genes included those involved in poorly differenti- we performed a comprehensive correlation analysis between MPC1/2 ated tumors (adjusted P ¼ 2.3e11), hypoxia (adjusted P ¼ 8.5e10), and all other genes assayed (n ¼ 16,733). A total of 713 and 303 genes TGFb signaling (adjusted P ¼ 1.0e8) and E-cadherin activity were positively and negatively (Spearman correlation BH-adjusted P < (adjusted P ¼ 1.3e9; Fig. 4D, Supplementary Tables S2 and S3). 0.01) correlated with MPC1/2, respectively (Fig. 4C, Supplementary These data suggest that MPC activity is involved in cellular networks

AACRJournals.org Clin Cancer Res; 2019 OF7

Karasinska et al.

A B MPC1 MPC2 Quiescent Glycolytic Cholestero- Mixed Glyco. vs choles. P = 0.0009 Glyco. vs choles. P = 5.0e-9 genic Glyco. vs mixed P = 0.007 Glyco. vs mixed P = 2.9e-5 Glyco. vs quies. P = 0.008 Glyco. vs quies. P = 0.002 MPC1 Choles. vs quies. P = 0.006 4 MPC2

SNV type CNV type 2 Missense Amplification Wildtype Loss Neutral 0

Gene expression (z-score) −2

Mixed Mixed Quiescent Glycolytic Quiescent Glycolytic

Cholesterogenic Cholesterogenic

C D Positively correlated with MPC1 and MPC2 (n = 713 genes) Genes positively correlated Negatively correlated with MPC1 and MPC2 (n = 303 genes) with MPC1 and MPC2 Other (n = 15, 717 genes) (Hallmark) oxidative phosphorylation 72/713 (10%)

(GO) small molecule metabolic process 205/713 (29%)

0.50 (GO) mitochondrion 181/713 (25%)

(GO) mitochondrial part 132/713 (19%)

(Mootha) human mitochondrial genes 87/713 (12%) 0.25 (GO) oxidation reduction process 121/713 (17%)

010203040

0 Genes negatively correlated with MPC1 and MPC2 (Rickman) poorly differentiated tumours 34/303 (11%) Correlation with MPC2 ( ρ ) −0.25 (Elvidge) up-regulated during hypoxia 22/303 (7%) (Onder) CDH1 targets 35/303 (12%)

(Koinuma) SMAD2/3 targets 45/303 (15%)

(Johnstone) PARVB targets 29/303 (10%) −0.25 0 0.25 0.50 Correlation with MPC1 (ρ) (Elvidge) HIF1A targets 14/303 (5%) 0369

-Log10 adjusted P value

Figure 4. Association of MPC1 and MPC2 expression with PDAC metabolic subgroups and cell signaling pathways. A, Oncoprint showing the distribution of MPC1 and MPC2 SNVs and CNVs across the metabolic groups. Only one case was found with an SNV in MPC2. B, Box plots illustrating signiﬁcant (one-way ANOVA P < 0.001) differences in expression levels of MPC1 and MPC2 across PDAC metabolic subgroups. Glycolytic cases have the lowest levels of MPC1 and MPC2 mRNA (Tukey HSD test adjusted P < 0.05). C, Scatter plot depicting the correlations between MPC1 (x-axis) and MPC2 (y-axis) and each of 16,733 genes. A total of 713 and 303 genes were found to be positively (Spearman correlation BH-adjusted P < 0.01; r > 0) and negatively (adjusted P < 0.01; r < 0) correlated with both MPC1 and MPC2 expression, respectively. D, Top 6 most signiﬁcantly enriched (hypergeometric test BH-adjusted P < 0.05) gene sets among genes positively (top) and negatively (bottom) associated with MPC1/2 expression. Genes negatively associated with MPC1/2 are enriched for gene sets related to hypoxia and cell adhesion.

associated with tumor progression in PDAC, possibly in part by may influence outcomes (10, 42). To determine the relevance of influencing the balance between tumor glycolytic and cholesterogenic glycolytic and cholesterogenic gene expression subtyping in other activity. organ sites, we repeated the consensus clustering analysis of glycolytic and cholesterogenic gene expression in 17 TCGA cancer types (7,375 Relevance of glycolytic and cholesterogenic gene clusters in samples with tumor content 30%; ref. 38). We found discrete clusters other cancer types of coexpressed pathway-specific genes in 9 cancer types (Supplemen- Different cancer types have unique metabolic signatures driven by tary Fig. S2). Although many of the genes were consistently coex- mutational landscape and organ-specific enzyme expression, which pressed in most tumor types, the appearance of some genes in

OF8 Clin Cancer Res; 2019 CLINICAL CANCER RESEARCH

Metabolic Gene Expression Heterogeneity and Survival in PDAC

A Cancer-specific Glycolytic Ambiguous Gene set (upper) Glycolytic Gene cluster Cholesterogenic Cholesterogenic

Gene set HNSC LIHC LUAD COAD LGG BLCA LUSC OV GBM

OA GPI IDI2 IDI1 LSS TPI1 PKM EBP MVK MVD LBR ENO1 PGK1PFKPENO2 PFKM PKLR PFKL FDPS SQLE SC5D ENO3 FDFT1 PMVK PGAM1 ALD ALDOC NSDHL DHCR7 GGPS1 ALDOBPGAM2 GAPDH PFKFB4PFKFB3 MSMO1 HMGCR TM7SF2PFKFB2 PFKFB1 PPP2CA PPP2CB DHCR24 GAPDHS PPP2R1A PPP2R5D CYP51A1 HSD17B7HMGCS1 PPP2R1B

BCMetabolic subtype (upper) Correlation (lower) Glycolytic Cholesterogenic Positive n.s. + Mixed + Cholesterogenic Quiescent Mixed Negative + Glycolytic + Quiescent 100 LGG 100 +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ ++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ + n = 111 +++++++++++++++++++++++++++++++++++++++++++ + n = 92 +++++++++++++++++++++++++++++++++++++++++++++++++ 50 +++++++++++++ + n = 90 75 ++++++++++ ++ +++ + +++ + n = 163 +++++ ++ ++++ + + samples (%) ++++ + Percentage of ++++ ++ 0 50 ++ + ++ + +++++ + 0.4 + Glycolytic 25 ++ exp. vs. Moffitt ++ + 0.2 P = 0.0076 basal-like exp. 0 Percentage of patients alive (%)

0 Correlation ( ρ ) 0 24 48 72 96 120 144 168 192

0.1 Number at risk Glycolytic 163 56 25 13 3 3 0 0 0 exp. vs. 0 90 34 19 12 9 5 4 1 1 92 30 12 9 3 2 1 1 0 KRAS exp. −0.1 111 36 17 6 3 2 2 0 0 Time (months) Glycolytic 0.2 LIHC exp. vs. 100 +++++++++++++++++++++++++ + n = 45 MYC exp. 0 ++++++++++++++++++++++++++++++++++++++++++++++++++++++++ + +++++++++++++++++++++ + n = 109 ++ ++ n = 64 −2 +++ + ++++++ + 75 ++++ ++ ++++ ++++ + n = 72 ++ +++++ +++++++++++ ++ ++ +++ ++ + +++ +++ 0 glycolytic group ( Z -score) +++ +++++++++ 50 +++ ++ + MPC1 exp. Mean expression in + − ++++ in glycolytic 0.2 +++ ++ + group vs. rest 25 + -0.4 + P = 0.018 0 Percentage of patients alive (%) MPC2 exp. 0 0 12 24 36 48 60 72 84 96 108 120 in glycolytic −0.2 group vs. rest 72 47 27 19 15 8 6 1 0 0 0 −0.4 64 37 21 16 9 5 4 2 2 1 1 109 70 44 31 21 11 4 1 1 0 0 45 27 15 10 6 3 1 1 1 0 0 D OV A Time (months) GBM LIHC LGG O HNSC LUAD BLCA C LUSC

Figure 5. Glycolytic and cholesterogenic gene proﬁling of other cancer types. A, Heatmap depicting which glycolytic and cholesterogenic genes were robustly coexpressed when consensus clustering was applied to each individual cancer type. B, Bar plots depicting the proportions of metabolic subgroups across the various cancer types (top) and correlation between glycolytic subgroups and expression of basal-like genes (4), KRAS, MYC,andMPC1/2 in each cancer type (bottom). Median glycolytic gene expression was positively (Spearman r > 0, BH-adjusted P < 0.05) correlated with basal-like gene expression in all cancer types. The correlation between MPC1/2 expression and the glycolytic subgroup was measured using Wilcoxon rank sum tests followed by BH correction. C, Kaplan–Meier survival analysis curves showing differences in median overall survival across metabolic subgroups in LGG (log-rank P ¼ 0.0076) and LIHC (P ¼ 0.018).

AACRJournals.org Clin Cancer Res; 2019 OF9

Karasinska et al.

coexpressed glycolytic and cholesterogenic pathways was unique to signature were associated with shortest overall survival corroborates only a few cancers, indicating cell type-specific contributions of some theroleofglycolysisintumoraggressivenessinPDAC(16). genes to the metabolic programs in each cancer (Fig. 5A). Glycolytic Cholesterol and its metabolites stimulate cancer cell growth (46) gene expression showed significant (Spearman correlation BH- and the tumor suppressor function of AMPK is mediated in part by adjusted P < 0.05) positive correlation with the expression of basal- the suppression of sterol synthesis (47). However, the finding that like genes from the Moffitt and colleagues’ PDAC classification PDAC cases with increased expression of cholesterogenic genes had system (4) in all cancer types, demonstrating that the basal-like gene better outcome indicates that while both glycolysis and cholesterol signature is also associated with increased glycolytic activity across a synthesis support cancer cell growth, they do not act synergistically greater spectrum of tumor types (Fig. 5B). For some cancer types, we to accelerate tumor progression in PDAC. Furthermore, the cor- also found a positive correlation (BH-adjusted P < 0.05) between relation of gene expression heterogeneity along the glycolysis– median glycolytic gene expression and either KRAS (LGG, COAD, cholesterol synthesis axis and prognostic subtypes of PDAC LUSC) or MYC (GBM, HNSC, LUAD, BLCA, LUSC) expression. indicates that different therapeutic strategies targeting metabolic Similar to PDAC, the expression of either MPC1 (HNSC, LUSC) or vulnerabilities, for example, those targeting tumor dependency MPC2 (BLCA, COAD), or both (LUAD), was significantly (Wilcoxon on glucose or cholesterol synthesis, could have clinical benefitin rank sum test, BH-adjusted P < 0.05) attenuated (fold change < 0) in subsets of PDAC patients. the glycolytic groups (Fig. 5B), further strengthening the notion that A survival benefit associated with the cholesterogenic profile changes in mRNA levels of these genes contributes to tumor could be an indirect effect due to increased pyruvate shuttling into glycolysis. Significant differences in survival across the four met- the mitochondria and reduced lactate production, a known medi- abolic subtypes were observed in LGG (log-rank P ¼ 0.0076) and ator of tumor growth (9). This hypothesis is supported by the LIHC (P ¼ 0.018, Fig. 5C). In LGG, subtypes with low expression of observation that the levels of MPC1 and MPC2 were lowest in cholesterogenic genes had shorter survival times than those with tumors of the glycolytic subtype. Other potential mechanisms higher expression of cholesterogenic genes (median overall survival: contributing to the outcome in cholesterogenic tumors could quiescent: 63.5 months, glycolytic: 73.5 months, cholesterogenic: involve the anti-tumor effects of cholesterol metabolites and deri- 95.6 months, mixed: 146.1 months). In LIHC, the difference in vatives. Oxysterols are cholesterol metabolites that activate liver-X- survival was observed across the glycolytic but not cholesterogenic receptors (LXR), which were shown to have tumor anti-proliferative gene expression gradient with glycolytic and mixed cases having and proapoptotic effects (48), and there is currently a clinical trial of inferior outcome (median overall survival: quiescent: 80.7 months, an LXR agonist in advanced cancer (NCT02922764). Dendrogenin glycolytic: 41.8 months, cholesterogenic: 70.1 months, mixed: A is a recently identified cholesterol derivative formed from cho- 45.1 months). These results indicate that tumor metabolic depen- lesterol-5,6-epoxides and histamine with LXR activating and anti- dencies may differ according to distinct genomic signatures and tumorproperties(49). tumor microenvironment factors that are specific to cancer type. PDAC tumors with both KRAS and MYC amplifications Forexample,drivermutationsinIDH1 and IDH2 encoding met- had the highest expression of glycolytic genes, indicating a poten- abolic enzymes are prevalent in LGG, which may uniquely affect cell tial glycolytic dependency and vulnerability to pharmacologic metabolism and tumor progression. glycolytic inhibition. One treatment strategy to target glycolysis- dependent tumors, such as those with KRAS and MYC amplification, may be to shift the metabolic phenotype towards the activa- Discussion tion of the sterol synthesis pathway. The activity of the MPC Continuous progress in the understanding of clinically relevant complex regulates pyruvate distribution in tumor cells (20) and we tumor subtypes is needed to accelerate the development of personal- show that the expression of MPC1 and MPC2 is negatively cor- ized treatment in PDAC. We show that PDAC tumors have distinct related with cellular gene programs linked to an aggressive tumor metabolic profiles based on the expression of genes involved in phenotype. These findings establish the MPC complex as a poten- glycolysis and cholesterol synthesis, which influence disease outcome tial target for altering tumor metabolic activity, whereby increasing and provide functional context to previously identified gene expres- MPC1 and MPC2 activity in glycolytic PDAC cases may reduce the sion subtypes. effects of tumor glycolysis by shifting the tumor into a cholestero- There is significant molecular heterogeneity in PDAC giving rise genic profile. We observed that MPC1 deletion was a frequent event to distinct tumor subclasses based on structural variations, chro- in PDAC, consistent with other cancer types (20), while MPC2 mosomal rearrangement events, epigenetic modification and gene CNVs were mostly amplified.AlthoughbothMPC1andMPC2 expression signatures (2–6), leading to a growing interest into form a functional heterocomplex, MPC2, but not MPC1, can also translating this information into clinical practice for outcome form independently functional homooligomers (50). Thus, the prognostication and treatment response prediction, as well as the increased MPC2 expression in the cholesterogenic subtype raises development of personalized therapies based on each tumor's the possibility that increased presence of MPC2 homooligomers unique molecular signature. PDAC transcriptome subtypes are may facilitate mitochondrial pyruvate uptake in these tumors. Our associated with survival (2–4, 6), but have yet to make a significant results also identify MPC2 as a potential target for stimulating impact on clinical management and the development of new mitochondrial pyruvate transport, which may be particularly rel- treatment strategies. The association between glycolytic and cho- evant in tumors with MPC1 copy loss. lesterogenic gene expression and expression levels of basal-like and Overall, tumor type-specific metabolic heterogeneity demon- classical genes provides a biological significance to PDAC subtypes strates that some reprogrammed metabolic pathways are less det- and supports targeting tumor metabolic plasticity as a means to rimental than others to clinical outcome depending on the reprogram an aggressive tumor type. cancer type, which may be exploited for the development of Glycolysis facilitates tumor progression, immune escape, and che- more precise therapeutic strategies targeting unique metabolic moresistance (9). The finding that tumors with a glycolytic gene dependencies.

OF10 Clin Cancer Res; 2019 CLINICAL CANCER RESEARCH

Metabolic Gene Expression Heterogeneity and Survival in PDAC

Disclosure of Potential Conflicts of Interest Acknowledgments J.J. Laskin reports receiving speakers bureau honoraria from Roche Canada and is We gratefully acknowledge the participation of patients and their families, and the an advisory board member/unpaid consultant for Takeda. D.J. Renouf reports POG and PanGen teams. The results published here are in part based upon data receiving speakers bureau honoraria from Celgene, Servier, Taiho, Bayer, and Ipsen. generated by The Cancer Genome Atlas managed by the NCI and NHGRI (http:// No potential conflicts of interest were disclosed by the other authors. cancergenome.nih.gov), as well as data generated by the International Cancer Genome Consortium (https://icgc.org/). This research was supported through phil- Authors’ Contributions anthropic donations received through the BC Cancer Foundation (project B20POG), as well as funding provided by the Terry Fox Research Institute (project 1078), Conception and design: J.M. Karasinska, J.T. Topham, S.E. Kalloger, D.J. Renouf, Ontario Institute for Cancer Research (PanCuRx Translational Research Initiative), D.F. Schaeffer Pancreatic Cancer Canada and Genome British Columbia. We acknowledge con- Development of methodology: J.M. Karasinska, J.T. Topham, S.E. Kalloger, tributions toward equipment and infrastructure from Genome Canada and Genome L. Culibrk, A.J. Mungall, D.J. Renouf BC (projects 202SEQ, 212SEQ, 12002), Canada Foundation for Innovation (projects Acquisition of data (provided animals, acquired and managed patients, provided 20070, 30198, 30981, 33408), and the BC Knowledge Development Fund. M.A. Marra facilities, etc.): L.M. Williamson, H.-L. Wong, M.K.C. Lee, G.M. O’Kane, R.A. Moore, acknowledges infrastructure investments from the Canada Foundation for Innova- A.J. Mungall, M.J. Moore, F. Notta, J.J. Knox, S. Gallinger, J. Laskin, M.A. Marra, tion and the support of the Canada Research Chairs. D.J. Renouf is a recipient of the D.J. Renouf MSFHR Health Professional-Investigator Award and D.F. Schaeffer is a recipient of Analysis and interpretation of data (e.g., statistical analysis, biostatistics, the VCHRI Investigator Award. computational analysis): J.M. Karasinska, J.T. Topham, S.E. Kalloger, G.H. Jang, M.K.C. Lee, M.J. Moore, C. Warren, J. Laskin, M.A. Marra, S.J.M. Jones, D.J. Renouf Writing, review, and/or revision of the manuscript: J.M. Karasinska, J.T. Topham, The costs of publication of this article were defrayed in part by the ’ S.E. Kalloger, G.H. Jang, R.E. Denroche, H.-L. Wong, M.K.C. Lee, G.M. O Kane, payment of page charges. This article must therefore be hereby marked R.A. Moore, A.J. Mungall, C. Warren, A. Metcalfe, J.J. Knox, S. Gallinger, J. Laskin, advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate S.J.M. Jones, D.J. Renouf, D.F. Schaeffer this fact. Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): J.T. Topham, R.E. Denroche, G.M. O’Kane, R.A. Moore, A.J. Mungall, M.J. Moore Received May 10, 2019; revised July 11, 2019; accepted August 28, 2019; Study supervision: D.J. Renouf, D.F. Schaeffer published first September 3, 2019.

References 1. Garrido-Laguna I, Hidalgo M. Pancreatic cancer: from state-of-the-art treat- 15. Guillaumond F, Leca J, Olivares O, Lavaut M-N, Vidal N, Berthezene P, et al. ments to promising novel therapies. Nat Rev Clin Oncol 2015;12:319–34. Strengthened glycolysis under hypoxia supports tumor symbiosis and hexosa- 2. Bailey P, Chang DK, Nones K, Johns AL, Patch AM, Gingras MC, et al. Genomic mine biosynthesis in pancreatic adenocarcinoma. Proc Natl Acad Sci U S A 2013; analyses identify molecular subtypes of pancreatic cancer. Nature 2016;531:47– 110:3919–24. 52. 16. Baek G, Tse YF, Hu Z, Cox D, Buboltz N, McCue P, et al. MCT4 3. Collisson EA, Sadanandam A, Olson P, Gibb WJ, Truitt M, Gu S, et al. Subtypes defines a glycolytic subtype of pancreatic cancer with poor prognosis of pancreatic ductal adenocarcinoma and their differing responses to therapy. and unique metabolic dependencies. Cell Rep 2014;9:2233–49. Nat Med 2011;17:500–3. 17. Shukla SK, Purohit V, Mehla K, Gunda V, Chaika NV, Vernucci E, et al. 4. MoffittRA,MarayatiR,FlateEL,VolmarKE,LoezaSG,HoadleyKA,etal. MUC1 and HIF-1alpha signaling crosstalk induces anabolic glucose metab- Virtual microdissection identifies distinct tumor- and stroma-specificsub- olism to impart gemcitabine resistance to pancreatic cancer. Cancer Cell types of pancreatic ductal adenocarcinoma. Nat Genet 2015;47:1168–78. 2017;32:71–87. 5. Waddell N, Pajic M, Patch A-M, Chang DK, Kassahn KS, Bailey P, et al. Whole 18. Bricker DK, Taylor EB, Schell JC, Orsak T, Boutron A, Chen Y-C, et al. A genomes redefine the mutational landscape of pancreatic cancer. Nature 2015; mitochondrial pyruvate carrier required for pyruvate uptake in yeast, Drosophila, 518:495–501. and humans. Science 2012;337:96–100. 6. Raphael BJ, Hruban RH, Aguirre AJ, Moffitt RA, Yeh JJ, Stewart C, et al. 19. Herzig S, Raemy E, Montessuit S, Veuthey J-L, Zamboni N, Westermann B, et al. Integrated genomic characterization of pancreatic ductal adenocarcinoma. Identification and functional expression of the mitochondrial pyruvate carrier. Cancer Cell 2017;32:185–203. Science 2012;337:93–6. 7. Aung KL, Fischer SE, Denroche RE, Jang GH, Dodd A, Creighton S, et al. 20. Schell John C, Olson Kristofor A, Jiang L, Hawkins Amy J, Van Vranken Genomics-Driven precision medicine for advanced pancreatic cancer: early Jonathan G, Xie J, et al. A role for the mitochondrial pyruvate carrier as a results from the COMPASS trial. Clin Cancer Res 2018;24:1344–54. repressor of the Warburg effect and colon cancer cell growth. Mol Cell 2014;56: 8. Aguirre AJ, Nowak JA, Camarda ND, Moffitt RA, Ghazani AA, Hazar-Rethinam 400–13. M, et al. Real-time genomic characterization of advanced pancreatic cancer to 21. Baggetto LG. Deviant energetic metabolism of glycolytic cancer cells. Biochimie enable precision medicine. Cancer Discov 2018;8:1096–111. 1992;74:959–74. 9. DeBerardinis RJ, Lum JJ, Hatzivassiliou G, Thompson CB. The biology of cancer: 22. Mullen PJ, Yu R, Longo J, Archer MC, Penn LZ. The interplay between cell metabolic reprogramming fuels cell growth and proliferation. Cell Metab 2008;7: signalling and the mevalonate pathway in cancer. Nat Rev Cancer 2016;16:718. 11–20. 23. Dale KM, Coleman CI, Henyan NN, Kluger J, White CM. Statins and cancer risk: 10. Peng X, Chen Z, Farshidfar F, Xu X, Lorenzi PL, Wang Y, et al. Molecular a meta-analysis. JAMA 2006;295:74–80. characterization and clinical relevance of metabolic expression subtypes in 24. Harshman LC, Wang X, Nakabayashi M, Xie W, Valenca L, Werner L, et al. human cancers. Cell Rep 2018;23:255–69. Statin use at the time of initiation of androgen deprivation therapy and time to 11. Koong AC, Mehta VK, Le QT, Fisher GA, Terris DJ, Brown JM, et al. Pancreatic progression in patients with hormone-sensitive prostate cancer. JAMA Oncol tumors show high levels of hypoxia. Int J Radiat Oncol Biol Phy 2000;48:919–22. 2015;1:495–504. 12. Kerr EM, Gaude E, Turrell FK, Frezza C, Martins CP. Mutant Kras copy number 25. McDougallJA,MaloneKE,DalingJR,Cushing-HaugenKL,PorterPL,LiCI. defines metabolic reprogramming and therapeutic susceptibilities. Nature 2016; Long-term statin use and risk of ductal and lobular breast cancer among 531:110–3. women 55 to 74 years of age. Cancer Epidemiol Biomarkers Prev 2013;22: 13. Ying H, Kimmelman AC, Lyssiotis CA, Hua S, Chu GC, Fletcher-Sananikone E, 1529–37. et al. Oncogenic Kras maintains pancreatic tumors through regulation of 26. Nielsen SF, Nordestgaard BG, Bojesen SE. Statin use and reduced cancer-related anabolic glucose metabolism. Cell 2012;149:656–70. mortality. N Engl J Med 2012;367:1792–802. 14. Bensaad K, Tsuruta A, Selak MA, Vidal MNC, Nakano K, Bartrons R, et al. 27. Goard CA, Chan-Seng-Yue M, Mullen PJ, Quiroga AD, Wasylishen AR, TIGAR, a p53-inducible regulator of glycolysis and apoptosis. Cell 2006;126: Clendening JW, et al. Identifying molecular features that distinguish fluvasta- 107–20. tin-sensitive breast tumor cells. Breast Cancer Res Treat 2014;143:301–12.

AACRJournals.org Clin Cancer Res; 2019 OF11

Karasinska et al.

28. Daemen A, Peterson D, Sahu N, McCord R, Du X, Liu B, et al. Metabolite 39. Laddha SV, Ganesan S, Chan CS, White E. Mutational landscape of the profiling stratifies pancreatic ductal adenocarcinomas into subtypes with distinct essential autophagy gene BECN1 in human cancers. Mol Cancer Res 2014;12: sensitivities to metabolic inhibitors. Proc Natl Acad Sci U S A 2015;112:E4410–7. 485–90. 29. Owen DR, Wong HL, Bonakdar M, Jones M, Hughes CS, Morin GB, et al. 40. Therneau TM. A Package for Survival Analysis in R. 2015. Available from: Molecular characterization of ERBB2-amplified colorectal cancer identifies https://CRAN.R-project.org/package¼survival. potential mechanisms of resistance to targeted therapies: a report of two 41. Kassambara A. Drawing Survival Curves using “ggplot2”. 2018. Available from: instructive cases. Cold Spring Harb Mol Case Stud 2018;4:pii:a002535. https://cran.r-project.org/web/packages/survminer/index.html. 30. Jones MR, Lim H, Shen Y, Pleasance E, Ch'ng C, Reisle C, et al. Successful 42. Hu J, Locasale JW, Bielas JH, O'Sullivan J, Sheahan K, Cantley LC, et al. targeting of the NRG1 pathway indicates novel treatment strategy for metastatic Heterogeneity of tumor-induced gene expression changes in the human met- cancer. Ann Oncol 2017;28:3092–7. abolic network. Nat Biotechnol 2013;31:522. 31. Li H, Durbin R. Fast and accurate short read alignment with Burrows-Wheeler 43. Son J, Lyssiotis CA, Ying H, Wang X, Hua S, Ligorio M, et al. Glutamine supports transform. Bioinformatics 2009;25:1754–60. pancreatic cancer growth through a KRAS-regulated metabolic pathway. Nature 32. Butterfield YS, Kreitzman M, Thiessen N, Corbett RD, Li Y, Pang J, et al. JAGuaR: 2013;496:101–5. junction alignments to genome for RNA-seq reads. PLoS One 2014;9:e102398. 44. Zhang C, Liu J, Liang Y, Wu R, Zhao Y, Hong X, et al. Tumour-associated mutant 33. Liao Y, Smyth GK, Shi W. featureCounts: an efficient general purpose program p53 drives the Warburg effect. Nat Commun 2013;4:2935. for assigning sequence reads to genomic features. Bioinformatics 2014;30:923– 45. Mayers JR, Torrence ME, Danai LV, Papagiannakopoulos T, Davidson SM, 30. Bauer MR, et al. Tissue of origin dictates branched-chain amino acid metab- 34. Lazar C, Meganck S, Taminau J, Steenhoff D, Coletta A, Molter C, et al. Batch olism in mutant Kras-driven cancers. Science 2016;353:1161–5. effect removal methods for microarray gene expression data integration: a 46. Kuzu OF, Noory MA, Robertson GP. The Role of Cholesterol in Cancer. survey. Brief Bioinform 2013;14:469–90. Cancer Res 2016;76:2063–70. 35. Liberzon A, Subramanian A, Pinchback R, Thorvaldsdottir H, Tamayo P, 47. Shackelford DB, Shaw RJ. The LKB1–AMPK pathway: metabolism and growth Mesirov JP. Molecular signatures database (MSigDB) 3.0. Bioinformatics control in tumour suppression. Nat Rev Cancer 2009;9:563. 2011;27:1739–40. 48. Bovenga F, Sabba C, Moschetta A. Uncoupling nuclear receptor LXR and 36. Wilkerson MD, Hayes DN. ConsensusClusterPlus: a class discovery tool with cholesterol metabolism in cancer. Cell Metabolism 2015;21:517–26. confidence assessments and item tracking. Bioinformatics 2010;26:1572–3. 49. Segala G, David M, de Medina P, Poirot MC, Serhan N, Vergez F, et al. 37. Carter SL, Cibulskis K, Helman E, McKenna A, Shen H, Zack T, et al. Absolute Dendrogenin A drives LXR to trigger lethal autophagy in cancers. quantification of somatic DNA alterations in human cancer. Nat Biotechnol Nat Commun 2017;8:1903. 2012;30:413–21. 50. Nagampalli RSK, Quesnay~ JEN, Adamoski D, Islam Z, Birch J, Sebinelli HG, et al. 38. Aran D, Sirota M, Butte AJ. Systematic pan-cancer analysis of tumour purity. Human mitochondrial pyruvate carrier 2 as an autonomous membrane trans- Nat Commun 2015;6:8971. porter. Sci Rep 2018;8:3510.

OF12 Clin Cancer Res; 2019 CLINICAL CANCER RESEARCH

Altered Gene Expression along the Glycolysis−Cholesterol Synthesis Axis Is Associated with Outcome in Pancreatic Cancer

Joanna M. Karasinska, James T. Topham, Steve E. Kalloger, et al.

Clin Cancer Res Published OnlineFirst September 3, 2019.

Updated version Access the most recent version of this article at: doi:10.1158/1078-0432.CCR-19-1543

Supplementary Access the most recent supplemental material at: Material http://clincancerres.aacrjournals.org/content/suppl/2019/08/31/1078-0432.CCR-19-1543.DC1

E-mail alerts Sign up to receive free email-alerts related to this article or journal.

Reprints and To order reprints of this article or to subscribe to the journal, contact the AACR Publications Subscriptions Department at [email protected].

Permissions To request permission to re-use all or part of this article, use this link http://clincancerres.aacrjournals.org/content/early/2019/10/21/1078-0432.CCR-19-1543. Click on "Request Permissions" which will take you to the Copyright Clearance Center's (CCC) Rightslink site.