Molecular Correlates of Metastasis by Systematic Pan-Cancer Analysis

Home , CFHR4, GABRG1, OR2M3, POF1B, REG3G, SPEF2

Published OnlineFirst November 6, 2018; DOI: 10.1158/1541-7786.MCR-18-0601

Genomics Molecular Cancer Research Molecular Correlates of Metastasis by Systematic Pan-Cancer Analysis Across The Cancer Genome Atlas Fengju Chen1, Yiqun Zhang1, Sooryanarayana Varambally2,3, and Chad J. Creighton1,4,5,6

Abstract

Tumor metastasis is a major contributor to mortality of signature that was distinctive from those of the other cancer cancer patients, but the process remains poorly understood. types. Functional categories of genes enriched in multiple Molecular comparisons between primary tumors and metas- cancer type–specific metastatic overexpression signatures tases can provide insights into the pathways and processes included cellular response to stress, DNA repair, oxidation– involved. Here, we systematically analyzed and cataloged reduction process, protein deubiquitination, and receptor molecular correlates of metastasis using The Cancer Genome activity. The TCGA-derived prostate cancer metastasis signa- Atlas (TCGA) datasets across 11 different cancer types, these ture in particular could define a subset of aggressive primary data involving 4,473 primary tumor samples and 395 tumor prostate cancer. Transglutaminase 2 protein and mRNA were metastasis samples (including 369 from melanoma). For each both elevated in metastases from breast and melanoma can- cancer type, widespread differences in gene transcription cers. Alterations in miRNAs and in DNA methylation were also between primary and metastasis samples were observed. For identified. several cancer types, metastasis-associated genes from TCGA comparisons were found to overlap extensively with external Implications: Our findings suggest that there are different results from independent profiling datasets of metastatic molecular pathways to metastasis involved in different tumors. Although some differential expression patterns asso- cancers. Our catalog of alterations provides a resource for ciated with metastasis were found to be shared across multiple future studies investigating the role of specificgenesin cancer types, by and large each cancer type showed a metastasis metastasis.

Introduction the factors governing the cancer spread and establishment at secondary locations remains poorly understood (3). Only a Metastases are formed by cancer cells that have left the small fraction of cancer cells from the primary tumor may go on primary tumor mass to form new colonies at sites throughout to successfully establish distant, macroscopic metastasis, and the human body (1). Tumor metastasis remains a major con- although the tumor microenvironment is understood to play tributor to deaths of cancer patients (2). Metastasis is a multi- an important role (3), the molecular state of the cancer cells in a step process, which includes localized invasion, intravasation macroscopic metastasis may widely differ from that of the into lymphatic or blood vessels, traversal of the bloodstream, cancer cells in the associated primary tumor. extravasation from the bloodstream, formation of microme- Molecular comparisons between primary tumors and meta- tastasis, and colonization (1, 2). The process of metastasis and stases can potentially provide insights into the pathways and processes involved with cancer disease progression (4, 5). Numer- ous independent studies have carried out gene expression pro- fi 1Division of Biostatistics, Dan L. Duncan Comprehensive Cancer Center Baylor ling of metastasis versus primary cancer for individual cancer College of Medicine, Houston, Texas. 2Comprehensive Cancer Center, University types (4–18). In addition to individual studies by cancer type, of Alabama at Birmingham, Birmingham, Alabama. 3Department of Pathology, "pancancer" molecular analyses would allow for examining Molecular and Cellular Pathology, University of Alabama at Birmingham, similarities and differences among the molecular alterations that 4 Birmingham, Alabama. Department of Bioinformatics and Computational may be associated with metastasis across diverse cancer types. Biology, The University of Texas MD Anderson Cancer Center, Houston, Texas. 5Human Genome Sequencing Center, Baylor College of Medicine, Houston, The recently published "MET500" dataset includes transcriptome fi Texas. 6Department of Medicine, Baylor College of Medicine, Houston, Texas. pro ling data for metastasis samples from approximately 500 patients, involving over 30 primary sites, and biopsied from over Note: Supplementary data for this article are available at Molecular Cancer Research Online (http://mcr.aacrjournals.org/). 22 organs (19); however, the MET500 dataset does not include any data on primary cancers. The Cancer Genome Atlas (TCGA), F. Chen and Y. Zhang are the co-first authors of this article. a large-scale initiative to comprehensively profile over 10,000 Corresponding Author: Chad J. Creighton, Baylor College of Medicine, One cancer cases at the molecular level, includes data on some meta- Baylor Plaza, Houston, TX 77030. Phone: 713-798-2264; Fax: 713-798-2716; stasis samples as well as on primary samples. Other than the E-mail: [email protected] TCGA-sponsored melanoma marker study (20), the metastasis doi: 10.1158/1541-7786.MCR-18-0601 samples were not featured in the respective marker analyses by 2018 American Association for Cancer Research. cancer type that were led by TCGA network, as the project as a

476 Mol Cancer Res; 17(2) February 2019

Pan-Cancer Metastasis Correlates

whole was focused on primary disease. The advantages of ana- Differential analyses by molecular feature lyzing TCGA data for metastasis-associated molecular correla- For mRNA, miRNA, and RPPA data platforms, differential tions include the multiple cancer types having been profiled on expression between comparison groups was assessed using Pear- a common platform that involves multiple levels of molecular son correlation on log-transformed values (base 2). For cancer data in addition to mRNA expression. types with more than one metastasis profile, the Pearson corre- In this study, we systematically analyzed and cataloged molec- lation P value is equivalent to a t test; for cancer types with just one ular correlates of metastasis using TCGA datasets, across 11 metastasis profile, significant genes in effect represented outliers different cancer types for which metastasis versus primary data with large differences at the edge or outside of the distribution as were available. Molecular profiling data platforms analyzed defined by the primary samples. Differential analyses between included mRNA expression, protein expression, miRNA expres- metastasis and primary by alternate methods for RNA-seq data sion, and DNA methylation. Significantly altered genes, as iden- were found to be largely concordant with results by the Pearson tified in a given cancer type, were compared across the other cancer method (Supplementary Fig. S1). For DNA methylation platform, types, as well as across results from other profiling datasets from differential expression between comparison groups was assessed studies outside of TCGA. using Pearson correlation on logit-transformed values (natural log). For SKCM datasets, a linear regression model was also carried Materials and Methods out for each gene, with dependent variable (continuous variable) of expression and with independent variables: metastasis/ TCGA patient cohort primary (categorical variable) þ estimated tumor purity (ref. 24; Results are based upon data generated by TCGA Research continuous variable). FDRs were estimated using the method of Network (https://gdc.cancer.gov/). Molecular data were aggre- Storey and Tibshirini (25). For selecting top features for a given gated from public repositories. Tumors analyzed in this study data platform and cancer type, FDR < 10% was used as a cutoff; for spanned 11 different TCGA projects, each project representing a SKCM datasets, top features were also significant with P < 0.05 for specific cancer type, listed as follows: Breast invasive carcinoma linear model incorporating tumor purity as a covariate. Visuali- (BRCA); Cervical squamous cell carcinoma (CESC) and endo- zation using heat maps was performed using both JavaTreeview cervical adenocarcinoma; Colorectal adenocarcinoma (CRC, (version 1.1.6r4; ref. 26) and matrix2png (version 1.2.1; ref. 27). combining COAD and READ projects); Esophageal carcinoma R software (version 3.1.0) was used for generation of box plots. (ESCA); Head and Neck squamous cell carcinoma (HNSC); Pancreatic adenocarcinoma (PAAD); Pheochromocytoma and Pathway and network analyses Paraganglioma (PCPG); Prostate adenocarcinoma (PRAD); Enrichment of gene ontology (GO) annotation terms within Sarcoma (SARC); Skin Cutaneous Melanoma (SKCM); and the sets of differentially expressed genes was evaluated using Thyroid carcinoma (THCA). Cancer molecular profiling data SigTerms software (28) and one-sided Fisher exact tests, with were generated through informed consentaspartofpreviously FDRs estimated using the method of Storey and Tibshirini (25). published studies and analyzed in accordance with each orig- Protein interaction network analysis used the entire set of human inal study's data use guidelines and restrictions. Metastasis protein–protein interactions cataloged in Entrez Gene (down- versus primary samples were inferred using the TCGA sample loaded June 2017). Entrez Gene interactions with yeast two- code ("06" vs. "01," respectively), which is the two digit code hybrid experiments providing the only support for the interaction following the TCGA legacy sample name (e.g., metastasis were not included in the analysis. Graphical visualization of sample "TCGA-V1-A9O5-06" and primary sample "TCGA- networks was generated using Cytoscape (29). ZG-A9L9-01"). Analysis of external expression profiling datasets Datasets We examined the following external gene expression profiling RNA sequencing (RNA-seq) data were obtained from The datasets of metastasis versus primary samples (listed by Gene Broad Institute Firehose pipeline (http://gdac.broadinstitute. Expression Omnibus or ArrayExpress accession number): BRCA org/). All RNA-seq samples were aligned using the by UNC studies E-MTAB-4003 (8), GSE100534 (9), and GSE110590 (5); RNA-seq V2 pipeline (21). Expression of coding genes was CRC studies GSE50760 (10), GSE22834 (11), and GSE41258 quantified for 20,531 features based on the gene models (12); PAAD studies GSE42952 (13) and GSE19281 (14); PRAD defined in the TCGA Gene Annotation File (GAF). Gene expres- studies GSE21034 (7), GSE3933 (6), and GSE6099 (4); SKCM sion was quantified by counting the number of reads over- studies GSE65904 (15), GSE17275 (16), and GSE46517 (17); lapping each gene model's exons and converted to reads per and THCA study GSE60542 (18). Differential expression between kilobase mapped (RPKM) values by dividing by the transcribed comparison groups was assessed using t test on log-transformed gene length defined in the GAF and by the total number of values (base 2). For the purposes of comparing results of external reads aligned to genes. Proteomic data generated by reverse- datasets with TCGA metastasis signatures, where multiple expres- phase protein array (RPPA) across 7,663 patient tumors sion array features referred to the same gene, the feature with the ("level 4" data) were obtained from The Cancer Proteome Atlas smallest P value for differences between metastasis and primary (http://tcpaportal.org/tcpa/; ref. 22). The miRNA-seq dataset tumors (either direction) was used to represent the gene. For was obtained from TCGA PanCanAtlas project (https://gdc. patient survival associations involving the TCGA PRAD metastasis cancer.gov/about-data/publications/pancanatlas; ref. 23), signature, we examined external gene expression profiling data- which involved batch correction according to Illumina GAIIx sets of primary prostate cancer from Taylor and colleagues or HiSeq 2000 platforms. DNA methylation profiles for 450K (GSE21034; ref. 7), Sboner and colleagues (GSE16560; ref. 30), Illumina array platform were obtained from The Broad Institute and Nakagawa and colleagues (GSE10645; ref. 31), assigning a Firehose pipeline (http://gdac.broadinstitute.org/). metastasis signature score to each external tumor profile using our

www.aacrjournals.org Mol Cancer Res; 17(2) February 2019 477

Chen et al.

previously described "t score" metric (21); log2-transformed expressed mRNAs (genes) in metastasis greatly exceeded the values within each dataset were for normalized to SDs from the chance expected (Fig. 1; Supplementary Data S2 and S3). Using median across the primary sample profiles. In the same way, the t a FDR cutoff of 10%, the numbers of top significant genes ranged score metric was also used in applying the TCGA metastasis gene from 43 for PCPG to 10,084 for SKCM, with the other cancer types signature for a given cancer type to the primary sample mRNA having between 178 and 1,205 top genes. For cancer types with profiles in TCGA for that cancer type. only one metastasis sample, significant genes in effect represented We also examined tissue-specific mRNA signatures, to deter- outliers with large differences at the edge or outside of the mine whether these might overlap with the cancer metastasis– distribution as defined by the primary samples (Supplementary specific mRNA signatures that were identified. Gene expression Fig. S1). The limitations with metastasis signature as defined by a data (TPM values) from GTEx Analysis version 7 release were single sample would include false negatives (e.g., in cases where obtained from the GTEx Portal (https://www.gtexportal.org). the distributions between primary and metastasis would overlap) Genes with average TPM values greater than 5 units across the and questions as to the generalizability of the signature to other normal tissue samples were used in this analysis, which involved metastasis cases, where the latter may be partially addressable by 12,769 unique genes in total. Using log-transformed values, for comparisons with results from external datasets (see below). We each tissue in GTEx dataset that would be associated with one of examined differences involving estimated tumor purities (24), as the cancer types analyzed in the present study (breast, BRCA; skin, gene expression patterns in cancer can reflect noncancer as well as SKCM; cervix/uteri, CESC; colon, CRC; esophagus, ESCA; muscle, cancer cells (32). Of all the cancer types examined, only SKCM SARC; nerve, PCPG; pancreas, PAAD; prostate, PRAD; and thy- showed significantly lower tumor purity in metastasis versus roid, THCA), the top 500 genes positively correlated with that primary (P ¼ 7.6E-7, t test, Supplementary Fig. S2). Using linear tissue as compared with all other tissues were determined (t test models incorporating purity as a covariate, on the order of 8,038 using log-transformed data). For a given cancer type, both the genes remained significantly differentially expressed in SKCM, of genes overexpressed in metastasis and the genes underexpressed the above 10,084 genes (Fig. 1). in metastasis were each compared with the set of tissue-specific Although some differential expression patterns associated with mRNA markers from GTEx corresponding to that cancer type, with metastasis were found to be shared across multiple cancer types, the significant overlap determined using one-sided Fisher exact by and large each cancer type showed a metastasis signature that tests. In the same way, we examined GTEx-derived markers of was distinctive from those of the other cancer types. In comparing tissues representing common sites of metastasis (adrenal gland, the respective expression signatures of metastasis from each brain, liver, and lungs) for significant overlap with TCGA-derived cancer type to each other, some amount of gene-set overlap was metastasis overexpressed genes. observed (Fig. 2A). In a number of cases, the overlap in signatures between any two cancer types was statistically significant, even if Statistical analysis the overlap itself involved a fraction of genes (e.g., on the order of All P values were two-sided unless otherwise specified. 10%). A set of 821 genes were found significant (FDR < 10%) with same direction of change for two or more cancer types (Fig. 2B). Of Results these genes, 65 were significant for three or more cancer types, TCGA cohort of primary and metastasis samples including genes with previously demonstrated functional roles in Our study utilized 4,473 primary tumor samples and 395 metastasis such as EPL3 (33), MYCNOS (34), and FOXF2 (35). tumor metastasis samples, involving 4,839 human cancer cases Just 8 genes (BEND4, CD5L, CELA1, CLEC4M, CYP17A1, representing 11 different major types, for which TCGA generated DCAF8L2, FAM151A, and SPIC) were overexpressed in metastasis data on one or more of the following molecular characterization (FDR < 10%) for four or more cancer types. We furthermore platforms (Supplementary Data S1): RNA-seq (4,446 primaries examined whether any of the metastasis signature genes (consid- and 393 metastases), RPPA (3,194 and 267), miRNA sequencing ering overexpressed and underexpressed gene sets separately) (4,350 and 378), and DNA methylation arrays (3,913 and 391). would be enriched for normal tissue-specific mRNA markers Of the cancer types studied, TCGA SKCM data involved the most associated with the given cancer type (as obtained using GTEx metastasis samples (n ¼ 369), followed by THCA (n ¼ 8), and data). Of 10 different tissue-specific marker gene sets, only a BRCA (n ¼ 7); CESC, HNSC, and PCPG cancer types each involved nominally significant association (P < 0.001, one-sided Fisher two metastasis samples; CRC, ESCA, PAAD, PRAD, and SARC each exact test) was observed between SKCM metastasis underex- involved one metastatic sample. Just 29 of the 395 metastasis pressed genes and gene markers associated with GTEx mRNA samples had a primary pair from the same patient, and so markers of normal skin tissues. unpaired analyses between primary and metastasis were made Functional categories of genes represented by the cancer type– the focus of this study. In terms of somatic DNA copy by SNP array specific metastasis expression signatures were examined using the platform, only SKCM metastasis samples had available data, with GO annotation terms (Supplementary Data S4). Specific GO term no data generated on primaries. Somatic mutation calls by whole- categories were found enriched within the corresponding metas- exome sequencing were considered too sparse for carrying out tasis signatures of multiple cancer types (Fig. 3A). Significantly comparisons within each cancer type, with the exception of enriched GO terms (FDR < 10% using one-sided Fisher exact tests) SKCM, whose data have been studied previously (20). found with the metastasis overexpressed genes for at least three cancer types included "cellular response to stress," "DNA repair," Differential mRNA patterns associated with metastasis by TCGA "oxidation–reduction process," "protein deubiquitination," and cancer type "receptor activity," and significant GO terms within the underex- We first set out to define differentially expressed mRNAs (based pressed genes for at least three cancer types included "extracellular on RNA-seq platform) between primary and metastasis samples region," "proteolysis," and "regulation of locomotion." We took for each cancer type. For each cancer type, the top differentially the genes related to receptor activity and genes high in metastasis

478 Mol Cancer Res; 17(2) February 2019 Molecular Cancer Research

Pan-Cancer Metastasis Correlates

(FDR < 10%) for at least one cancer type, and we integrated these using TCGA data, with metastasis expression signatures obtained with public databases of protein–protein interactions to generate from external datasets made available by previously published a protein interaction network (Fig. 3), which allowed us to studies. We examined 15 external gene expression profiling data- visualize the potential relationships involving these genes. sets of metastasis versus primary samples, involving six cancer Although most of the genes in this network involved SKCM, a types (BRCA, CRC, PAAD, PRAD, SKCM, and THCA). For each of number of other genes involved a trend (P < 0.05, Pearson on log- the cancer types surveyed, a significant number of genes where transformed data) of higher expression in metastasis in two found to overlap with the results of at least one external dataset of or more cancer types, and 10 genes in the network were high the given cancer type, for either the metastasis overexpressed or (P < 0.05) in three or more cancer types: CR1, CR2, GP1BA, underexpressed genes (Fig. 4A). Perhaps, in part, because the CRC GRID2, GRM7, LHCGR, LRP2, MED14, P2RX2, and PTPRH. and THCA metastasis signatures each involved fewer genes, the Similar types of interaction networks were also generated involv- CRC overexpressed genes showed some overlap but not a signif- ing genes related to oxidation–reduction process or protein icant overlap with CRC overexpressed genes from external data- deubiquitination (Supplementary Fig. S3). Genes involved in the sets, and THCA underexpressed genes by TCGA did not show immune checkpoint pathway were also examined in TCGA significant overlap with external dataset results. For each cancer metastasis profiles (Supplementary Fig. S4), with these being type, on the order of 35%–70% of genes comprising the corre- elevated across SKCM metastasis samples as expected (20), as sponding TCGA metastasis signature showed a similar significant well as elevated in a portion of metastasis samples from other trend (P < 0.05) in at least one external dataset of that cancer type cancer types. (Fig. 4B). Notably, the external datasets often involved different sites of Metastasis-associated mRNA patterns as observed in datasets metastasis for a given cancer type; for example, the external PRAD external to TCGA datasets involved samples taken from various sites including To help assess their generalizability, we compared the gene lymph node, bone, lungs, testes, and brain (4, 6, 7), implying expression signatures of metastasis, as defined for each cancer type that the TCGA PRAD signature, while derived from a single

BRCA BRCA CESC CESC CRC CRC Primary Metastasis AFM Primary Metastasis Primary Metastasis n = 1095 n = 7 AKR1C4 n = 304 n = 2 n = 379 n = 1 Differential expression C7 (change from primary) CASC1 C9 C7 1/3-fold 3-fold CRIP3 GABRB2 GRK1 DUXA KIF26B CCNE1 KRAS F10 MMP14 HTR5A LYRM5 ADAM6 OR2T35 MXRA5 LOC150185 Genes overexpressed in metastasis PCID2 CD2 PTPRD and focally amplified in pancancer analysis PROZ RNF144A NLRP4 FNTA SERPIND1RSAD2 OR2G3 GZMB NEGR1 Genes underexpressed in metastasis SLC22A7 SH3PXD2B and focally deleted in pancancer analysis COL10A1 SOX11 OR5AU1 NKX2-3 COL17A1 VCAN ZIM2 PDE4D 547 Genes (FDR < 0.1) CYS1 RHOBTB3 VSNL1 271 Genes (FDR < 0.1) HAPLN1 180 Genes (FDR < 0.1) F2RL2 XG SLC16A10 FGF18 ZNF544 PDE4D

AIFM3 UBE2M MAP3K7 B4GALT7 UFD1L MCC AARS2 CAPSL WDR70 MCTP1 ESCA ESCA ABCC10 SLC27A5 HNSC HNSC CDC45 ARID1A MDN1 PAAD PAAD SKCM SKCM Primary Metastasis CAPN11 SLC35B2 Primary Metastasis ARL10 Primary Metastasis Primary Metastasis n n n n CHMP2A NT5C2 n n n n = 184 = 1 CNPY3 SRF = 520 = 2 CNPY3 ATG2B ODZ2 = 178 = 1 = 104 = 367 CUL7 TBCC COMT BDKRB1 PAM CUL9 TJAP1 DDX41 BDKRB2 PAPOLA HSP90AB1TMEM151B DGCR6L BTRC PPP1R2P3 KLHDC3 TMEM63B DGCR6 BVES RASA1 MEA1 TRIM28 EDDM3B C19orf26 RCOR1 MGC2752 TRIT1 FBXO4 CASP8AP2 REV3L INSM2 MYCL1 XPO5 FLJ35024 CCNK ROCK2 KCTD1 YIPF3 NKX2-1 ISOC2 CDC42BPB SIK3 ZIM2 NSD1 ZBTB45 MRPL40 CPEB4 SLC16A1 OR2M4 ZNF318 NLRP9 DENND2C SLK ZNF805 POLH ZNF324B PRODH DICER1 SYNCRIP F2RL2 POLR1C ZNF324 RNF32 ENTPD7 TECPR2 NKX2-3

173 Genes (FDR < 0.1) PPP2R5D ZNF496 RPL37 F2RL2 TJP1 178 Genes (FDR < 0.1) 1205 Genes (FDR < 0.1) NUDCD2 8038 Genes (FDR < 0.1) RPL7L1 ZNF497 RTDR1 GTF3C4 TTC37 SFTA3 ZNF551 SPEF2 HOMER1 WDR20 STXBP6 THAP7 HSP90AA1 YY1 TMEM191A LDB1 ZMYND11 TRMT2A MAN2A1

PCPG PCPG PRAD PRAD AFP PAIP1 GABRG3 SARC SARC THCA THCA Primary Metastasis Primary Metastasis AKR1C1 SKP2 GLB1L3 Primary Metastasis Primary Metastasis n = 179 n = 2 n = 497 n = 1 AKR1C3 SPEF2 GLOD4 n = 259 n = 1 C13orf35 n = 503 n = 8 AKR1C4 WDR70 KCNMB1 CD274 AKR1C4 C1orf229 ZNF124 LOC100128239 EHF C5orf28 ZNF280A LOXL4 FLJ44054 C6 C5orf34 ZNF324B MTERFD2 LOC338588 CCNE1 ZNF669 NCAM1 NKX2-1 DNMT1 ZNF695 NDN OR1C1 E2F3 ANO7 NSA2 OR2L2 ENSA BOK PDE8B OR2M1P FBXO4 CECR6 PER3 OR2M3 OR2L8 G6PD CNNM1 PPP1R7 OR2T8 POTEG OR2W5 KPNA2 CNTNAP3 SCAMP1 OR6F1 LETM2 COX15 SEMA4G PGC SBK2 LGSN DMGDH SLC7A8 TMEM190 43 Genes (FDR < 0.1)

185 Genes (FDR < 0.1) RFPL4A 192 Genes (FDR < 0.1) FAM49A 1013 Genes (FDR < 0.1) MRPS30 DUSP1 TMEM18 SBK2 OPLAH EPHA7 TRIM8 SFTA3 tAKR OR4K15 FAM20C TSLP OR4L1 FOXC1 WDR41 ZNF658

Figure 1. Top differentially expressed mRNAs in metastasis versus primary samples for each of 11 cancer types in TCGA. Top genes for each cancer type were selected using Pearson correlation (on log-transformed values) with Storey and Tibshirini estimate of FDR of <10% (for SKCM, FDR < 10% and significant with P < 0.05 for linear model incorporating tumor purity as a covariate). Yellow, high expression relative to the average of primary samples; blue, low expression. Genes listed individually are either overexpressed in metastasis and focally amplified in a previous pan-cancer analysis (44) or underexpressed in metastasis and focally deleted in pan-cancer analysis. For SKCM, hundreds of genes involved regions of focal amplification or deletion, and so these are not listed here.

www.aacrjournals.org Mol Cancer Res; 17(2) February 2019 479

Chen et al.

metastasis sample, would not be specific to a single site. Similarly, tigated the corresponding metastasis expression signatures in breast metastasis in the GSE110590 dataset (5) involved a num- primary tumors. TCGA expression profiles of primary tumors ber of different sites, with the TCGA BRCA metastasis signature were each scored for manifestation of the metastasis signature. being manifested in samples from most of these sites (Fig. 4C). Out of nine cancer types for which pathologic stage or grade Furthermore, we examined GTEx-derived markers of tissues repre- information were provided, five (CSEC, HNSC, PRAD, SKCM, and senting common sites of metastasis (adrenal gland, brain, liver, THCA) showed some statistical trend for positive correlation and lung) for significant overlap with TCGA-derived metastasis between the signature score and stage or grade across primary overexpressed genes; after multiple testing correction (25), only cancers (one-sided P 0.05, Pearson, Fig. 5A). This association the GTEx liver signature was found to significantly overlap with was notably strongest for PRAD cancer type (P < 1E-30) to the metastasis genes associated with BRCA (P < 1E-7, one-sided Fisher extent that clear differences in time to adverse events between exact test, with 20 of the 342 BRCA metastasis overexpressed genes patients with primary prostate tumors manifesting the PRAD also included in the top 500 genes highly expressed in normal metastasis signature as compared with the rest of the patients liver), but not with genes from the other cancer types. were observable, when applying the signature to profiles from Previous studies have suggested that a subset of primary tumors multiple external cohorts (Fig. 5B). In addition, in another pros- resemble metastatic tumors with respect to gene expression pat- tate cancer dataset, consisting of primary prostate cancer samples terns (36). For each cancer type in our TCGA cohort, we inves- from patients for whom the early onset of metastasis following

A Significance of overlap BRCA CESC CRC ESCA HNSC PAAD PCPG PRAD SARC SKCM THCA BRCA CESC CRC ESCA HNSC PAAD PCPG PRAD SARC SKCM THCA # Genes # Genes

BRCA 342 17 12 2 5 15 0 23 7 53 18 BRCA 205 2 4 0 9 2 0 10 0 70 0 P value ( − log CESC 244 17 7 8 3 12 3 9 6 31 10 CESC 27 2 3 0 3 0 0 1 0 8 0 0 CRC 84 12 7 1 0 4 1 2 2 13 0 CRC 96 4 3 0 13 4 0 3 0 12 0 ESCA 171 2 8 1 10 7 1 3 11 32 0 ESCA 2 0 0 0 1 0 0 0 0 1 0 HNSC 537 5 3 0 10 2 2 11 1 20 6 HNSC 668 9 3 13 1 1 0 5 0 55 1 5 PAAD 154 15 12 4 7 2 0 4 2 18 4 PAAD 24 2 0 4 0 1 0 1 0 6 0

PCPG 37 0 3 1 1 2 0 0 1 0 1 PCPG 6 0 0 0 0 0 0 0 0 2 0 10

PRAD 693 23 9 2 3 11 4 0 4 210 19 PRAD 320 10 1 3 0 5 1 0 0 86 0 10 ) SARC 182 7 6 2 11 1 2 1 4 14 1 SARC 3 0 0 0 0 0 0 0 0 2 0 SKCM 4088 53 31 13 32 20 18 0 210 14 42 SKCM 3950 70 8 12 1 55 6 2 86 2 0 THCA 184 18 10 0 0 6 4 1 19 1 42 THCA 8 0 0 0 0 1 0 0 0 0 0 FDR > 0.1 Genes overexpressed in metastasis (FDR < 0.1) Genes underexpressed in metastasis (FDR < 0.1)

B 821 Genes (significant for two or more cancer types) Differential expression BRCA t statistic CESC Low met CRC -6 ESCA -4 HNSC -2 0 PAAD 2 PCPG 4 PRAD 6 SARC High met SKCM THCA P > 0.05 F3 DIO2 RTP1 ARSJ PRM2 MMP3 F2RL2 KRT17 KRT14 PTPLA TPSB2 CELA1 FOXF1 FOXF2 AZGP1 PRSS8 AP1M2 CNTN6 OR7C1 LHCGR NKX2-1 FAM89A CYP1A2 TPSAB1 OR7E5P FAM180A DCAF8L2 KRTAP5-7 ADCYAP1R1 VPS13A-AS1 C7 SPIC ELP3 DLK1 CD5L STAR RAD1 GRID2 ATP4A SPEF2 SSTR3 NR5A1 FCER2 GP1BA CFHR1 BEND4 AMHR2 FCAMR CT45A6 ADIPOQ HSD3B2 AKR1C4 LRRTM3 CLEC4G CLEC4M GABRA2 C13orf30 GABRG1 FAM151A FAM163B MYCNOS CYP17A1 FAM129C SIGLEC11 LOC647309

Figure 2. Genes shared among the cancer type–specific metastasis mRNA signatures. A, For both the genes overexpressed in metastasis for at least one cancer type (left, genes from Fig. 1) and the genes underexpressed in metastasis for at least one cancer type (right, genes from Fig. 1), the numbers of overlapping genes between any two cancer types are indicated, along with the significance of overlap (using colorgram, by one-sided Fisher exact test). B, Heatmap of differential t statistics (Pearson correlation on log-transformed data), by cancer type, comparing metastasis versus primary (red, higher in metastasis; white, not significant with P > 0.05), for 821 genes significant for two or more cancer types (FDR < 10%, for SKCM; FDR < 10% and significant with P < 0.05 for linear model incorporating tumor purity as a covariate). Genes significant for three or more cancer types are indicated by name.

480 Mol Cancer Res; 17(2) February 2019 Molecular Cancer Research

Pan-Cancer Metastasis Correlates

A Genes overexpressed in metastasis Genes underexpressed in metastasis BRCA 4 2 14 6 2 2 1 7 0 29 0 0 2 0 2 50 44 36 4 46 44 29 3 23 23 16 28 24 42 20 85 4 GO term enrichment CESC 4 3 8 7 3 4 3 4 1 8 1 1 1 2 5 37 34 34 4 5 2 5 3 0 0 3 2 0 5 3 6 1 0 P value ( − log CRC 5 2 3 2 0 1 0 5 2 3 2 2 2 2 7 13 12 11 0 24 17 11 0 6 6 8 17 8 23 11 36 9 ESCA 5 6 11 8 3 5 7 8 4 4 0 0 2 1 6 18 18 18 0 0 0 0 0 0 0 0 0 0 0 0 0 0 HNSC 35 36 61 34 24 43 10 40 16 41 19 19 22 22 41 18 17 14 3 38 31 11 1 33 33 25 67 68 108 42 166 35 5 PAAD 2 3 10 2 2 0 1 2 1 18 0 0 1 2 6 24 23 20 0 5 3 2 0 3 3 1 0 3 3 1 4 0 PCPG 0 0 0 0 0 0 7 0 0 0 0 0 0 0 0 8 8 8 0 0 1 0 0 1 1 0 1 1 1 1 3 0 10

PRAD 41 75 101 83 56 40 2 45 35 37 27 48 41 29 28 1 28 27 5 1 10 9 9 29 18 48 18 89 7 ) 27 28 18 10 SARC 5 2 4 3 2 3 6 9 1 3 4 3 7 3 6 26 18 17 0 0 0 0 0 0 0 0 0 0 1 1 1 0 SKCM 187 226 408 236 155 213 1 215 72 144 78 77 97 73 284 267 211 178 16 400 307 68 9 152 146 88 244 264 474 147 784 142 THCA 1 2 6 5 2 1 3 1 0 8 0 0 0 0 0 28 25 24 0 1 0 0 0 0 0 0 2 0 2 0 4 2 P

11 > 0.05 40 # Genes 86 289 616 597 420 792 579 697 703 759 472 780 846 261 895 294 287 376 264 955 1139 1703 1312 1424 2471 4238 1448 1572 1313 1205 Proteolysis DNA repair keratin filament Receptor activity Peptidase activity Extracellular space Extracellular region DNA metabolic process Protein deubiquitination Regulation of locomotion Regulation of localization Protein catabolic process Signaling receptor activity Cellular response to stress Hemidesmosome assembly Oxidation-reduction process Cell-substrate junction assembly Macromolecule catabolic process Extracellular structure organization Protein modification by small protein... Proteasomal protein catabolic process Intracellular ribonucleoprotein complex Ubiquitin/proteasome catabolic process Single-organism developmental process Negative regulation of cell cycle process Positive reg. of cellular comp. movement Single-organism membrane organization Positive reg. of multicellular organismal... Cellular macromolecule catabolic process Transmembrane signaling receptor activity Cellular response to DNA damage stimulus Peptidase activity, acting on L-amino acid...

B Protein–protein interaction network involving genes associated with GO:receptor activity

TGFBR3 ACVR2A NR2C2 NR2F1 AHR HNF4G HNF4A GRIK5 ARNT CALCRL CRCP ASGR1 TLR4 TLR5 GRID2 DLG3 MED13 CFI TSHR ASGR2 TNFRSF4 CD226 TLR1 ERBB4 MED17 SEC63 TLR10 AGER GRIA2 NR0B2 MED30 GLP1R LHCGR DERL1 TNFRSF9 ITGB2

KCNJ1 ESRRG NR5A2 SEC62 SORL1 MERTK BMPR2 GPR183 TRAM1 BMPR1A CFTR KCNH2 GPC6 NRP1 PTPRR EPHB6 CD80 MED14 MED1 PPARA CELSR2 NCAM1 RYK CD86 PPARG STRA6 EPNA3 LRP2 EFNB2 INTS6 CD28 PIGR ESRRB NR4A2 THRAP3 TFRC GRM5 CD79B FLT1 ATP6AP2 NTRK1 INSRR NR3C1 XPR1 FCGR2C KDR ITGB1 CD79A CD3G AMOT ITGA4 IL2RA INSR PTPRC NR2F2 CD4 FZD4 DDR2 CXCR5 CD36 CD46 ITGAV ITGB6 VTN DLG1 CXCR4 CD1D IFNAR1 CALM3 GRM7 CNR1 FZD5 CD55 HCRTR1 ADRB1 GABRR1 CSF1R NPC1 CR2 GABRR2 F2R F2 KCNQ5 CD47 LIFR ANTXR2 MET CD97 GP1BA LRP6 CR1 P2RX2 ABCC9 IL6ST FLT4 OSMR PTPRH P2RX4 KCNJ8 CUL5 GHR PTPRB

Higher expression in metastasis versus primary for given cancer type

BRCA CESC CRC ESCA HNSC PAAD PCPG PRAD SARC SKCM THCA

Figure 3. Functional gene classes shared among the cancer type–specific metastasis mRNA signatures. A, Left, GO terms significantly enriched for at least three cancer types (enrichment for cancer type defined as FDR < 10% using one-sided Fisher exact test) within the respective sets of genes overexpressed in metastasis (based on the gene sets represented in Fig. 1); right, GO terms significantly enriched for at least three cancer types within the respective sets of genes underexpressed in metastasis. For both sets of enriched GO terms, the numbers of genes involved for each cancer type and overall significance of enrichment (by colorgram; black, highly significant) are indicated. B, Protein–protein interaction network involving genes overexpressed in metastasis, with focus on genes involved in receptor activity. Nodes represent genes with GO annotation "receptor activity" and which were found overexpressed in metastasis for at least one cancer type (FDR < 10%, for SKCM; FDR<10% and significant with P < 0.05 for linear model incorporating tumor purity as a covariate). Nodes are colored according to the individual cancer types in which atrend(P < 0.05, Pearson correlation on log-transformed data) of higher expression in metastasis versus primary samples was observed. A line between two nodes signifies that the corresponding protein products of the genes can physically interact (according to the literature, from Entrez Gene interactions database). www.aacrjournals.org Mol Cancer Res; 17(2) February 2019 481

Chen et al.

Significance of overlap P > 0.05 P value (-log ) 10 0 30 A 15

Genes overexpressed in metastasis Genes underexpressed in metastasis

BRCA CRC PAAD PRAD SKCM THCA BRCA CRC PAAD PRAD SKCM THCA External External datasets datasets # Genes E-MTAB-4003 GSE100534 GSE110590 GSE50760 GSE22834 GSE41258 GSE42952 GSE19281 GSE21034 GSE3933 GSE6099 GSE65904 GSE17275 GSE46517 GSE60542 E-MTAB-4003 GSE100534 GSE110590 GSE50760 GSE22834 GSE41258 GSE42952 GSE19281 GSE21034 GSE3933 GSE6099 GSE65904 GSE17275 GSE46517 GSE60542 # Genes BRCA 234 20 26 158 127 135 128 84 54 21 47 20 3 65 50 342 BRCA 144 30 101 60 61 63 83 47 42 40 16 55 8 54 27 205 CRC 34 5 1 6 24 12 9 6 17 0 4 3 0 12 4 84 CRC 45 13 34 25 31 41 35 25 32 29 13 11 4 22 5 96 PAAD 80 16 8 17 48 32 37 14 26 11 12 10 1 30 13 154 PAAD 13 2 7 11 6 7 15 8 7 6 2 6 0 5 5 24 PRAD 296 112 162 45 263 95 133 112 354 213 106 39 22 245 167 693 PRAD 116 33 74 67 92 62 79 60 178 101 41 40 8 67 63 320 TCGA SKCM 1174 453 203 246 1785 809 595 597 648 608 366 745 172 1647 1026 4088 TCGA SKCM 1180 253 317 300 1128 681 425 580 387 412 283 1106 135 1124 705 3950 THCA 102 9 6 35 51 46 37 18 44 6 14 17 1 34 69 184 THCA 3 1 4 2 1 2 3 0 3 5 0 0 1 0 2 8 # Genes # Genes 432 398 2113 6871 2313 3921 1369 6580 4037 3102 2350 4873 2866 1990 1549 4076 2681 6079 1927 4609 1551 4679 3656 2620 2744 2990 3363 1876 2607 2533

B Genes overexpressed in metastasis ** Genes underexpressed in metastasis **

2,072 1,789** 2,072 1,789 One or more external datasets (P < 0.05) One or more external datasets (P < 0.05) 524 Two or more external datasets (P < 0.05) ** Two or more external datasets (P < 0.05) 500 ** 500 Overlap P < =0.0001 465 Overlap P < =0.0001 * 440 * ** Overlap P < 1E-15 ** Overlap P < 1E-15 400 400 ** 300 300 ** 249 ** ** 220 200 184 200 165 ** ** ** ** 96 * Number of overlapping genes Number of overlapping genes 84 100 69 100 * 46 61 ** 30 32 * 27 16 4 5 7 2 0 0 BRCA CRC PAAD PRAD SKCM THCA BRCA CRC PAAD PRAD SKCM THCA TCGA TCGA # genes 342 81 154 693 4088 184 # genes 205 90 24 320 3950 8 A1CF F11 AGXT2L1 ADAM12 ABI2 ADH1B C11orf41 ADAMDEC1 C7orf60 ALDH1A3 AKR1B10 LAD1 ACSM2B F9 APC2 AURKA APPBP2 BEND4 COL11A1 ARSJ CPA3 APH1B CD207 LAMA3 ADRA1A FAM151A ARSF AURKB BRWD1 BLK CTHRC1 COL12A1 CRISPLD1 ATAD1 CDALCN2 APOA2 FAM9B ASB4 C7orf49 C1orf25 C4orf7 F2RL2 COL14A1 F2RL2 CTBS CDS1 MAP7 APOC3 FSHR ATP2B3 CDCA8 C1orf9 CCR7 FOXF2 CSRP2 FOXF1 CTSO CELSR2 MAPK13 ART4 G6PC CLRN1 CENPF CD86 CDK5R1 GJB2 CYTIP HIGD1B DHRS7 CLCA4 MYO1C ASPDH HRG DHCR7 CIT ELK4 CR2 GRP DIO2 ITIH5 ESR1 COL17A1 NDEL1 ASPG KNG1 FCN2 DNMT3B EPB41L2 FAM129C KLK4 FBLN1 MS4A2 FOXC1 CRABP2 NMU C14orf105 MBOAT4 FDPS DYRK2 FNDC3A FCER2 LRRC15 FOXF1 MYLIP MEIS2 CSTAOTUB2 C3orf27 NR1H4 GRIK1 FANCD2 NNT FCRL1 MFAP5 FOXF2 MYO1B NR4A1 DHRS1 PDZK1IP1 C3P1 OR5AK2 KCTD1 FGD1 NUCB2 GP1BA MMP13 HOXB5 NKX2-3 PDE8B DIO2 PI3 C8A PDILT SRRM4 GDAP1 OGT HOTAIR MMP3 HOXB7 SHISA2 RPS27L DSC3 POF1B C9 PLG ST8SIA5 GPI P4HA1 HOXC10 PDGFRL IGFBP6 STXBP6 SEC62 EHF PPP4R1 CA5A PPY SULT2A1 GTSE1 PIK3CA HOXC6 POSTN LTBP1 TPSAB1 SPCS3 EPHA1 S100A14 CASR REG3G HNRNPUL1 PLOD2 HOXC9 PPAPDC1A MAP2K4 TPSB2 ZMAT1 EPS8L1 S100A9 CFHR4 SHISA3 IPO9 PLS1 LILRA4 SFRP2 PDE4D VOPP1 EVPL SCEL CLEC4M SLC17A2 KIF23 PPM1B LOC283663 SGIP1 PRR16 EXPH5 SERPINB13 Selected top overlapping genes CRP SLC2A2 LIG1 PTPRO MS4A1 Selected top overlapping genes TNFAIP6 RAC2 FGFBP1 SFN CRYGD TCF21 NCAPH RARB P2RX5 VSNL1 SGIP1 GJA1 SLPI CYP11B1 TEX13B NHSL1 RBM26 PAX5 WNT2 SULF1 IMPA2 SNAI2 CYP3A4 TMEM225 PEG10 SENP6 PDE3B SVIL KCNK1 SPINT2 DAO ZDHHC19 SKP2 SFRS7 RASGRP2 TACC1 KLK13 SPRR3 DBH TMEM65 SLC33A1 STAP1 TNFSF11 KLK6 ST14 TOP2A SMARCAL1 TCL1A TPSAB1 KLK7 STMN2 TPX2 UBA5 TIMD4 TPSB2 KLK8 TP63 UBE2T ZMYM2 TXK TWSG1 KRT16 TTC15 ZMYM4 VPREB3 WLS ZCCHC18 C GSE110590 Dataset, breast cancer metastases ** ** * ***

t statistic) 5

0 similarity score ( P -5 * < 0.05, versus primary TCGA-BRCA metastasis gene signature ** P < 0.01, versus primary a -10 a Skin Skin Dura Liver Liver Liver Liver Liver Liver Liver Liver Liver Liver Liver Liver Liver Liver Lung Lung Lung Lung Lung Lung Lung Lung Lung Lung Lung Lung Lung Lung Lung Brain Brain Brain Brain Brain Brain Brain Brain Brain Brain Bone Bone Bone Pleur Pleur Chest Chest Ovary Spinal Spinal Spinal Kidney Primary Primary Primary Primary Primary Primary Primary Primary Primary Primary Primary Primary Primary Primary Primary Adrenal Adrenal Adrenal Adrenal Mediastn Pancreas Pancreas Softtissue Lymph node Lymph node Lymph node Lymph node Lymph node Lymph node

Figure 4. Significance of overlap between TCGA metastasis mRNA signatures and metastasis mRNA signatures from datasets external to TCGA. A, For both the genes overexpressed in metastasis for a given cancer type (left) and the genes underexpressed in metastasis for a given cancer type (right), the numbers of overlapping genes between the TCGA mRNA signatures (rows, signatures from Fig. 1) and the genes over- or underexpressed in metastasis (P < 0.05, t test) in the indicated external datasets from previously published gene expression profiling studies (columns), along with the corresponding significances of overlap (using colorgram, by one-sided Fisher exact test, x2 test for TCGA SKCM gene sets). (Continued on the following page.)

482 Mol Cancer Res; 17(2) February 2019 Molecular Cancer Research

Pan-Cancer Metastasis Correlates

radical prostatectomy was recorded (37), PRAD metastasis sig- (Fig. 6E; Supplementary Data S6). For all cancer types except SARC nature scores were significantly elevated (P < 1E-9) in the early and THCA, significant inverse correspondences between methyl- onset group (Fig. 5C). ation and expression results were observed (P < 0.05, one-sided Fisher exact test or c2 test), either involving genes overexpressed Molecular patterns associated with metastasis involving and with lower associated methylation in metastasis or involving protein, miRNA, and methylation genes underexpressed and with higher associated methylation in We went on to examine the protein, miRNA, and DNA metastasis. The significantly overlapping results involved, for methylation datasets in TCGA, to define the differentially example, 2,730 genes for SKCM (both overexpressed and underexpressed features between primary and metastasis samples expressed genes, with inverse patterns of DNA methylation), 66 for each cancer type. RPPA proteomic data involved 218 fea- genes for PRAD, 43 genes for HNSC, and 33 genes for CESC. tures and four cancer types (BRCA, PCPG, SKCM, and THCA) fi with metastasis pro les. For SKCM, a large portion of RPPA Discussion features examined were differentially expressed in metastasis (94 features at FDR < 10%, Pearson correlation on log-trans- Our study of TCGA data on cancer metastasis samples had three formed data, 83 features significant after corrections for tumor overall objectives: (i) to obtain a preliminary global view of purity; Supplementary Data S5), analogous to results from metastasis versus primary molecular differences across several mRNA expression. No RPPA features with globally significant cancer types, (ii) to provide a resource for future studies investi- (FDR<10%) were found for PCPG or THCA, likely, in part, due gating the role of specific genes in metastasis, and (iii) to help to limited sample power. For BRCA, one protein feature, provide direction for future genomics studies of metastasis, for transglutaminase 2, was elevated in metastasis and globally example, by showcasing the utility of examining molecular dif- significant at FDR < 10% (FDR < 1E-12), corresponding to ferences across cancer types and across other molecular profiling mRNA-level differences (Fig. 6A). Transglutaminase 2 protein platforms, in addition to RNA-seq. A clear limitation of this study and mRNA were also elevated in SKCM (Fig. 6A) and the involves the limited number of metastasis samples profiled as part protein is known to promote metastasis (38). For most cancer of TCGA consortium, as the main focus of TCGA was to examine types, widespread differences in miRNA expression between genomic and molecular patterns of primary rather than of met- metastasis and primary were observed (Fig. 6B; Supplementary astatic cases. For several of the cancer types examined, this Data S5). Most of the significant miRNAs detected were over- limitation is mitigated somewhat by comparing the results from expressed versus underexpressed in metastasis, with 85 over- TCGA transcriptomic data to existing data from other studies, expressed miRNAs and 12 underexpressed miRNAs significant thereby demonstrating the relevance of differential gene patterns (FDR < 10%) in two or more cancer types, and with 17 over- as observed across sample cohorts. Our results would support the expressed miRNAs significantinthreeormorecancertype(Fig. need for future multiplatform-based and pancancer genomic 6C). For a number of cancer types, mRNA–miRNA pairings, as studies profiling larger numbers of metastasis with primary sam- defined by both a previously identified miRNA–target interac- ples, which would allow us to further define and refine the tion (as cataloged by miRTarBase; ref. 39) and significant molecular signatures of metastasis as put forth in this study. differential expression in metastasis for both mRNA and Nevertheless, our study demonstrates that even on the basis of miRNA (in opposite directions), could also be identified (Sup- a single metastasis sample, there would be molecular information plementary Data S5). contained here, representing real biological differences that may Using TCGA data from DNA methylation arrays, we examined involve at least some metastasis cases for a given cancer type. 150,253 CpG Island probes, finding widespread differences in Our study has identified widespread molecular differences in methylation between metastasis and primary samples for each metastasis versus primary tumors for 11 different cancer types, cancer type studied (Fig. 6D; Supplementary Data S6). The num- with each cancer type having a signature of metastasis that is bers of top significant methylation features (FDR < 10%, Pearson distinct from that of the other cancer types. This would suggest on logit-transformed data) ranged from 163 for THCA to 27,530 that there are different molecular pathways to metastasis involved for SKCM (after corrections for tumor purity), with the other in different cancers. Our findings would seemingly differ with cancer types having between 441 and 6,611 top features. those of two early studies of gene expression patterns of metas- As increased methylation of regulatory regions in proximity to tasis, one from Ramaswamy and colleagues (36), which defined a genes can lead to epigenetic silencing, we integrated DNA meth- single 128-gene signature of metastasis across multiple cancer ylation results with mRNA expression results, defining sets of types (lung, breast, prostate, colorectal, uterus, ovary, etc.), and genes associated with both altered methylation and expression one from Weigelt and colleagues (40), which could not find any

(Continued.) B, For each indicated cancer type, numbers of genes overlapping between the TCGA metastasis signature genes (left, genes overexpressed in metastasis; right, genes underexpressed in metastasis) and the genes significantly high or low in metastasis (P < 0.05, t test) in the published external datasets corresponding to the given cancer type. Significance of overlap (by one-sided Fisher exact test; x2 test for SKCM genes) is indicated for TCGA genes found in one or more external datasets (blue bars) and in two or more external datasets (red bars). Selected top genes overlapping between TCGA and results from other datasets are listed (BRCA overexpressed: TCGA P < 1E-6 and P < 1E-6 for E-MTAB-4003 dataset; BRCA underexpressed: TCGA P < 1E-6 and P < 1E-6 for E-MTAB-4003 dataset; CRC underexpressed: TCGA FDR < 10% and P < 0.05 for two or more external datasets; PAAD overexpressed: TCGA FDR < 10% and P < 0.01 for GSE42952 dataset; PAAD underexpressed: TCGA FDR < 10% and P < 0.05 for one or more external datasets; PRAD overexpressed: TCGA FDR < 10% and P < 0.01 for all three external datasets; PRAD underexpressed: TCGA FDR < 10% and P < 0.05 for all three external datasets; SKCM overexpressed: TCGA FDR < 10% and P < 0.05 for all three external datasets; SKCM underexpressed: TCGA FDR < 10% and P < 0.05 for all three external datasets; THCA overexpressed: TCGA FDR < 10% and P < 0.001 for GSE60542 dataset; P values by Pearson correlation or t test on log-transformed data). C, TCGA-BRCA metastasis gene expression signature similarity score (t statistic as derived from the "t score" metric; refs. 21, 45), as applied to the sample profiles in the GSE110590 breast cancer metastases RNA-seq dataset (5). For selected groups of metastasis according to site, comparisons with the primary group are indicated (t test as applied to the signature t scores).

www.aacrjournals.org Mol Cancer Res; 17(2) February 2019 483

Chen et al.

A C GSE46691, Prostate cancer 31 * * P < =0.05 10 Gleason > 7 4

( P value)] Gleason < =7 10 P < 1E-9 5 3 * t statistic) * 0 2 *

* similarity score ( -5 1 TCGA-PRAD metastasis gene signature

Association of metastasis signature scoring Primary cancers Primary cancers that did not develop that developed

with aggressive primary cancer [-log 0 C metastases metastases CRC n n BRCA CESC ESCA HNS PAAD PRAD SKCM THCA ( = 333) ( = 212)

B GSE21034, Prostate cancer GSE16560, Prostate cancer GSE10645, Prostate cancer 1 1 1 0.9 Other (n = 93) 0.9 Log rank P < 1E-5 0.9 0.8 0.8 0.8 Other (n = 397) 0.7 0.7 0.7 0.6 0.6 0.6 n 0.5 0.5 Other ( = 187) 0.5 0.4 PRAD 0.4 0.4 PRAD metastasis metastasis PRAD signature high (n = 199) 0.3 signature high 0.3 Censored 0.3 0.2 n Censored 0.2 metastasis 0.2 Survival probability ( = 47) Survival probability Survival probability P signature high P 0.1 Log rank < 1E-6 0.1 (n = 94) 0.1 Log rank < 1E-16 Censored 0 0 0 0 50 100 150 0 50 100 150 200 250 300 0 50 100 150 200 250 Recurrence-free survival (months) Disease-specific survival (months) Disease-specific survival (months)

Figure 5. For specific cancer types, gene expression signatures of metastasis found present within a subset of primary samples and associated with more aggressive disease. A, For each of the indicated cancer types, the corresponding TCGA metastasis gene signature was applied to the primary sample mRNA profiles for that cancer type; across the primary samples, the metastasis signature similarity scores (t statistic as derived from the "t score" metric, refs. 21, 45) were correlated with the cancer stage or grade (Gleason grade for PRAD, pathologic stage for the other cancer types). One-sided P values indicate the Pearson correlation between the signature score and stage or grade (numerical 1–4 for stage, 6–10 for Gleason grade). B, For each of three independent mRNA expression profiling datasets of primary prostate cancer (7, 30, 31), differences in survival between patients with tumors manifesting the TCGA-PRAD metastasis signature (top third of signature similarity scores across the samples) and the other patients. P values by log-rank test. C, The TCGA PRAD metastasis gene signature was applied to the primary sample mRNA profiles for the GSE46691 prostate cancer dataset (37) consisting of primary prostate cancer samples from patients for which the early onset metastasis following radical prostatectomy was recorded. Box plot represents 5%, 25%, 50%, 75%, and 95%. P value for differences in signature scores between groups with or without metastasis by t test.

global significant differences over chance expected between pri- consistent patterns across multiple datasets and studies, we may mary and metastasic breast cancer samples. Studies subsequent to place the most confidence in these gene patterns, at least given the the Weigelt and colleagues' study have been able to define currently available data. widespread differences associated with breast cancer metastasis The results of this study (e.g., as provided in the Supplementary versus primary tumors (5, 8, 9). Interestingly, when surveying Materials and Methods) may serve as a resource for future studies TCGA data, none of the Ramaswamy signature genes showed investigating the role of specific genes in metastasis. The various consistent high or low expression patterns in metastasis across the gene signatures of metastasis, as identified in each cancer type by different cancer types (Supplementary Data S2). Although the TCGA data, may be mined to help identify candidates for func- Ramaswamy study found that a subset of primary tumors from tional studies. For cancer types with only a single metastasis various cancer types expressing the 128-gene metastasis signature sample with TCGA, there would be potential limitations with were associated with worse outcome, we find in this study that the associated metastasis signature, including questions as to the aggressive prostate cancers, in particular, appear to express a generalizability of the signature to other metastasis cases. Inte- metastasis signature pattern, but other cancer types, such as breast gration of TCGA results with results of external public datasets can cancer, do not show a similar phenomenon. One salient feature of considerably strengthen the metastasis associations as identified this study was to survey available data from multiple external for specific genes. For example, the TCGA PRAD metastasis sources, in addition to TCGA data. Where genes are found to show signature was based on a single metastasis profile, but this

484 Mol Cancer Res; 17(2) February 2019 Molecular Cancer Research

Pan-Cancer Metastasis Correlates

A 5.0 5.0 P P = 0.03 P P < 1E-14 4 < 0.01 < 1E-12 10 2.5 2 2.5

5 0.0 0 0.0

0 -2 -2.5 -2.5 TGM2 mRNA (normalinzed) mRNA (normalinzed) TGM2 mRNA

Transglutaminase 2 protein (norm.) BRCA BRCA BRCA BRCA Transglutaminase 2 protein (norm.) SKCM SKCM SKCM SKCM Primary Metastases Primary Metastases Primary Metastases Primary Metastases

B C 200 Higher in met. 180 Lower in met. 160 Differential expression miR-134-5p miR-127-3p miR-409-3p miR-199b-5p miR-382-5p miR-379-5p miR-200a-3p miR-200a-5p miR-200b-3p miR-200b-5p miR-429 miR-205-5p miR-105-3p miR-1224-5p miR-3923 miR-424-3p miR-202-5p miR-507 miR-506-3p miR-514a-5p miR-508-5p miR-508-3p miR-509-3p miR-514a-3p miR-509-3-5p miR-513a-5p miR-513b-5p miR-513c-5p miR-514b-5p 140 t statistic BRCA 120 Low met CESC -6 100 CRC -4 80 ESCA -2 0 HNSC 60 2 PAAD 4 40 PCPG 6 20 PRAD High met # Significant microRNAs (FDR < 0.1) 0 SARC SKCM P > 0.05 A M CRC AAD THCA BRCA CESC ESC HNSC P PCPG PRAD SARC SKC THCA Any two Any three

D E 2,169 1,874 Methylation higher in met. 22,150 Methylation higher in met. Methylation higher in met. 22,150 2,169 ** 1,874 ** 856 585 Methylation lower in met. Methylation lower in met. 585 Methylation lower in met. 7,000 856 ** Overlap P < 0.05 50 Overlap P < =0.01 ** * P * P 6,000 35 ** Overlap < 0.01 45 ** Overlap < 0.0001 30 40 5,000 ** 25 ** 35 4,000 30 20 25 3,000 * 15 20 ** ** 2,000 10 ** 15 * 10 1,000 5 * ** * * 5 * * ** # Significant probes (FDR < 0.1) 0 0 0 Overlapping genes, over-expr. in met. A C A C in met. Overlapping genes, under-expr. A C C CRC CRC AAD CRC BRCACESC ESC HNSC PAAD PCPG PRAD SAR SKCM THCA BRCACESC ESC HNSC P PCPG PRAD SAR SKCM THCA BRCACESC ESC HNS PAAD PCPG PRAD SAR SKCM THCA

Figure 6. Molecular correlates of metastasis by protein, miRNA, or DNA methylation profiling. A, Transglutaminase 2 protein and mRNA (TGM2 gene) were significantly elevated in TCGA BRCA metastases as well as in TCGA SKCM metastases. Box plots represent 5%, 25%, 50%, 75%, and 95%. P values by t test on log-transformed data. B, Numbers of significant miRNAs between metastasis ("met.") and primary samples for each cancer type (FDR < 10%, based on Pearson correlation using log-transformed values; for SKCM, FDR < 10% and significant with P < 0.05 for linear model incorporating tumor purity as a covariate), along with the numbers or miRNAs significant (FDR < 10%) for two or three cancer types. C, Heatmap of differential t statistics (Pearson correlation on log-transformed data), by cancer type, comparing metastasis versus primary (red, higher in metastasis; white, not significant with P > 0.05), for 29 miRNAs that were either significantly overexpressed for three or more cancer types (FDR < 10%) or significantly underexpressed for two or more cancer types. D, Numbers of significant DNA methylation array probes located within CpG Islands (by Illumina 450K array, 150K CpG Island probes) between metastasis and primary samples for each cancer type (FDR < 10%, based on Pearson correlation using logit-transformed values; for SKCM, FDR < 10% and significant with P < 0.05 for linear model incorporating tumor purity as a covariate). E, For each cancer type, numbers of genes overlapping between the RNA-seq and DNA methylation results (FDR < 10% for each platform, with top features in SKCM corrected for tumor purity as described above). Left represents genes overexpressed in metastases, and right plot represents genes underexpressed in metastases. Significance of overlap by one-sided Fisher exact test (x2 test for SKCM results). signature showed highly significant overlap with results with each Although successfully defining molecular signatures of metas- of the three independent profiling datasets of prostate cancer tasis across several different cancer types, this study points to the metastasis versus primary disease (4, 6, 7). The TCGA PRAD need for more molecular data on metastasis in human tumors. signature could also define a subset of aggressive primary prostate Much could be gained by generating molecular data on larger cancer. Integration between mRNA data and data from other numbers of metastasis and primary cancers, using multiple platforms in TCGA may also be used to select genes of particular "omics" data platforms, in addition to mRNA expression profil- interest, such as the genes showing concordant alterations involving. The global molecular patterns involved in metastasis would ing both expression and DNA methylation. Genes that appear entail proteomic and DNA methylation levels, in addition to significant in multiple cancer types, including genes encoding cell transcriptomic levels. For many cancer types with metastasis data receptors, may also be of interest for further investigation. in TCGA, few or no relevant external molecular profiling datasets

www.aacrjournals.org Mol Cancer Res; 17(2) February 2019 485

Chen et al.

were found to be available. For cancer types where a large number Disclosure of Potential Conflicts of Interest of expression outliers could be associated with a single metastasis No potential conflicts of interest were disclosed. sample profile, profiling more metastasis cases would enable us to define more robust molecular signatures that would presumably Authors' Contributions be generalizable to the disease as a whole. Profiling larger num- Conception and design: C.J. Creighton bers of cases would also allow for paired analyses by patient Development of methodology: C.J. Creighton between primary and metastasis samples, as well as offering the Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): F. Chen, Y. Zhang, C.J. Creighton possibility of subtype discovery within metastatic tumors, accord- Writing, review, and/or revision of the manuscript: S. Varambally, ing to differential patterns being found within some but not all C.J. Creighton metastasis cases. Molecular data from human tumors may be Study supervision: C.J. Creighton combined with molecular data from experimental models of metastasis (41) to identify genes common to both, which may Acknowledgments help pinpoint critical targets relevant in both the laboratory and This work was supported, in part, by NIH grant P30CA125123 human disease settings. The top gene correlates of metastasis by (to C.J. Creighton). and large do not appear to represent canonical oncogenes (32, 42) The costs of publication of this article were defrayed in part by the payment of or frequent targets of point mutation (43), but rather appear page charges. This article must therefore be hereby marked advertisement in indicative of complex processes at work involving multiple inter- accordance with 18 U.S.C. Section 1734 solely to indicate this fact. nal and external factors. The molecular signatures of metastasis for each cancer type have the potential to lead to new discoveries into Received June 7, 2018; revised July 19, 2018; accepted October 26, 2018; the disease process. published first November 6, 2018.

References 1. Weinberg RA. The Biology of Cancer. New York, NY: Garland Science; profiling: Prediction of survival outcome and benefit from molecular 2006. targeted therapy. Oncotarget 2015;6:12297–309. 2. Steeg P. Targeting metastasis. Nat Rev Cancer 2016;16:201–18. 16. Martins W, Esteves G, Almeida O, Rezze G, Landman G, Marques S, 3. Jiang W, Sanders A, Katoh M, Ungefroren H, Gieseler F, Prince M, et al. et al. Gene network analyses point to the importance of human Tissue invasion and metastasis: molecular, biological and clinical perspec- tissue kallikreins in melanoma progression. BMC Med Genomics tives. Semin Cancer Biol 2015;35:S244–75. 2011;4:76. 4. Tomlins S, Mehra R, Rhodes D, Cao X, Wang L, Dhanasekaran S, et al. 17. Kabbarah O, Nogueira C, Feng B, Nazarian R, Bosenberg M, Wu M, et al. Integrative molecular concept modeling of prostate cancer progression. Integrative genome comparison of primary and metastatic melanomas. Nat Genet 2007;39:41–51. PLoS One 2010;5:e10770. 5. Siegel M, He X, Hoadley K, Hoyle A, Pearce J, Garrett A, et al. Integrated 18.TarabichiM,SaiseletM,Tresallet C, Hoang C, Larsimont D, Andry G, RNA and DNA sequencing reveals early drivers of metastatic breast cancer. et al. Revisiting the transcriptional analysis of primary tumours J Clin Invest 2018;128:1371–83. and associated nodal metastases with enhanced biological and sta- 6. Lapointe J, Li C, Giacomini C, Salari K, Huang S, Wang P, et al. Genomic tistical controls: application to thyroid cancer. Br J Cancer 2015;112: profiling reveals alternative genetic pathways of prostate tumorigenesis. 1665–74. Cancer Res 2007;67:8504–10. 19. Robinson D, Wu Y, Lonigro R, Vats P, Cobain E, Everett J, et al. Integrative 7. Taylor B, Schultz N, Hieronymus H, Gopalan A, Xiao Y, Carver B, et al. clinical genomics of metastatic cancer. Nature 2017;548:297–303. Integrative genomic profiling of human prostate cancer. Cancer Cell 20. Cancer Genome Atlas Network. Genomic classification of cutaneous 2010;18:11–22. melanoma. Cell 2015;161:1681–96. 8. Lawler K, Papouli E, Naceur-Lombardelli C, Mera A, Ougham K, Tutt A, 21. Cancer Genome Atlas Research Network. Comprehensive molecular et al. Gene expression modules in primary breast cancers as risk factors for characterization of clear cell renal cell carcinoma. Nature 2013;499: organotropic patterns of first metastatic spread: a case control study. Breast 43–9. Cancer Res 2017;19:113. 22. Zhang Y, Kwok-Shing Ng P, Kucherlapati M, Chen F, Liu Y, Tsang Y, et al. A 9. Schulten H, Bangash M, Karim S, Dallol A, Hussein D, Merdad A, et al. Pan-cancer proteogenomic atlas of PI3K/AKT/mTOR pathway alterations. Comprehensive molecular biomarker identification in breast cancer brain Cancer Cell 2017;31:820–32. metastases. J Transl Med 2017;15:269. 23. Hoadley K, Yau C, Hinoue T, Wolf D, Lazar A, Drill E, et al. Cell-of-origin 10. Kim S, Kim S, Kim J, Roh S, Cho D, Kim Y, et al. A nineteen gene-based risk patterns dominate the molecular classification of 10,000 tumors from 33 score classifier predicts prognosis of colorectal cancer patients. Mol Oncol types of cancer. Cell 2018;173:291–304. 2014;8:1653–66. 24. Aran D, Sirota M, Butte A. Systematic pan-cancer analysis of tumour purity. 11. Lin A, Chua M, Choi Y, Yeh W, Kim Y, Azzi R, et al. Comparative profiling of Nat Commun 2015;6:8971. primary colorectal carcinomas and liver metastases identifies LEF1 as a 25. Storey JD, Tibshirani R. Statistical significance for genomewide studies. prognostic biomarker. PLoS One 2011;6:e16636. Proc Natl Acad Sci USA 2003;100:9440–5. 12. Sheffer M, Bacolod M, Zuk O, Giardina S, Pincas H, Barany F, et al. 26. Saldanha AJ. Java Treeview–extensible visualization of microarray data. Association of survival and disease progression with chromosomal insta- Bioinformatics 2004;20:3246–8. bility: a genomic exploration of colorectal cancer. Proc Natl Acad Sci U S A 27. Pavlidis P, Noble W. Matrix2png: a utility for visualizing matrix data. 2009;106:7131–6. Bioinformatics 2003;19:295–6. 13. Van den Broeck A, Vankelecom H, Van Eijsden R, Govaere O, Topal B. 28. Creighton C, Nagaraja A, Hanash S, Matzuk M, Gunaratne P. A bioinfor- Molecular markers associated with outcome and metastasis in human matics tool for linking gene expression profiling results with public pancreatic cancer. J Exp Clin Cancer Res 2012;31:68. databases of microRNA target predictions. RNA 2008;14:2290–6. 14. Barry S, Chelala C, Lines K, Sunamura M, Wang A, Marelli-Berg F, et al. 29. Shannon P, Markiel A, Ozier O, Baliga N, Wang J, Ramage D, et al. S100P is a metastasis-associated gene that facilitates transendothelial Cytoscape: a software environment for integrated models of biomolecular migration of pancreatic cancer cells. Clin Exp Metastasis 2013;30: interaction networks. Genome Res 2003;13:2498–504. 251–64. 30. Sboner A, Demichelis F, Calza S, Pawitan Y, Setlur S, Hoshida Y, et al. 15. Cirenajwis H, Ekedahl H, Lauss M, Harbst K, Carneiro A, Enoksson J, et al. Molecular sampling of prostate cancer: a dilemma for predicting disease Molecular stratification of metastatic melanoma using gene expression progression. BMC Med Genomics 2010;3:8.

486 Mol Cancer Res; 17(2) February 2019 Molecular Cancer Research

Pan-Cancer Metastasis Correlates

31. Nakagawa T, Kollmeyer T, Morlan B, Anderson S, Bergstralh E, Davis B, et al. 38. Huang L, Xu A, Liu W. Transglutaminase 2 in cancer. Am J Cancer Res A tissue biomarker panel predicting systemic progression after PSA recur- 2015;5:2756–76. rence post-definitive prostate cancer therapy. PloS One 2008;3:e2318. 39. Hsu S, FM L, Wu W, Liang C, Huang W, Chan W, et al. miRTarBase: a 32. Chen F, Zhang Y, Gibbons D, Deneen B, Kwiatkowski D, Ittmann M, et al. database curates experimentally validated microRNA-target interactions. Pan-cancer molecular classes transcending tumor lineage across 32 cancer Nucleic Acids Res 2011;39:D163–9. types, multiple data platforms, and over 10,000 cases. Clin Cancer Res 40. Weigelt B, Glas A, Wessels L, Witteveen A, Peterse J, van't Veer L. Gene 2018;24:2182–93. expression profiles of primary breast tumors maintained in distant metas- 33. Delaunay S, Rapino F, Tharun L, Zhou Z, Heukamp L, Termathe M, tases. Proc Natl Acad Sci U S A 2003;100:15901–5. et al. Elp3 links tRNA modification to IRES-dependent translation of 41. Gibbons D, Lin W, Creighton C, Zheng S, Berel D, Yang Y, et al. Expression LEF1 to sustain metastasis in breast cancer. J Exp Med 2016;213: signatures of metastatic capacity in a genetic mouse model of lung 2503–23. adenocarcinoma. PLoS One 2009;4:e5401. 34. Zhao X, Li D, Pu J, Mei H, Yang D, Xiang X, et al. CTCF cooperates with 42. Hanahan D, Weinberg R. The hallmarks of cancer. Cell 2000;100:57–70. noncoding RNA MYCNOS to promote neuroblastoma progression 43. Lawrence M, Stojanov P, Mermel C, Robinson J, Garraway L, Golub T, et al. through facilitating MYCN expression. Oncogene 2016;35:3565–76. Discovery and saturation analysis of cancer genes across 21 tumour types. 35. Wang Q, Kong P, Li X, Yang F, Feng Y. FOXF2 deficiency promotes Nature 2014;505:495–501. epithelial-mesenchymal transition and metastasis of basal-like breast 44. Zack T, Schumacher S, Carter S, Cherniack A, Saksena G, Tabak B, et al. Pan- cancer. Breast Cancer Res 2015;17:30. cancer patterns of somatic copy number alteration. Nat Genet 2013; 36. Ramaswamy S, Ross K, Lander E, Golub T. A molecular signature of 45:1134–40. metastasis in primary solid tumors. Nat Genet 2003;33:49–54. 45. Creighton C, Hernandez-Herrera A, Jacobsen A, Levine D, Mankoo P, 37. Erho N, Crisan A, Vergara I, Mitra A, Ghadessi M, Buerki C, et al. Discovery Schultz N, et al. Integrated analyses of microRNAs demonstrate their and validation of a prostate cancer genomic classifier that predicts early widespread influence on gene expression in high-grade serous ovarian metastasis following radical prostatectomy. PLoS One 2013;8:e66855. carcinoma. PLoS One 2012;7:e34546.

www.aacrjournals.org Mol Cancer Res; 17(2) February 2019 487

Molecular Correlates of Metastasis by Systematic Pan-Cancer Analysis Across The Cancer Genome Atlas

Fengju Chen, Yiqun Zhang, Sooryanarayana Varambally, et al.

Mol Cancer Res 2019;17:476-487. Published OnlineFirst November 6, 2018.

Updated version Access the most recent version of this article at: doi:10.1158/1541-7786.MCR-18-0601

Supplementary Access the most recent supplemental material at: Material http://mcr.aacrjournals.org/content/suppl/2018/11/06/1541-7786.MCR-18-0601.DC1

Cited articles This article cites 44 articles, 8 of which you can access for free at: http://mcr.aacrjournals.org/content/17/2/476.full#ref-list-1

Citing articles This article has been cited by 3 HighWire-hosted articles. Access the articles at: http://mcr.aacrjournals.org/content/17/2/476.full#related-urls

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 Department at Subscriptions [email protected].

Permissions To request permission to re-use all or part of this article, use this link http://mcr.aacrjournals.org/content/17/2/476. Click on "Request Permissions" which will take you to the Copyright Clearance Center's (CCC) Rightslink site.