Epithelial–mesenchymal transition-associated secretory phenotype predicts survival in lung cancer patients

Summary

A gene signature derived from EASP strongly correlated with nodal metastasis, advanced tumor stage and poor differentiation and predicted survival in three independent lung cancer data sets, including early-stage patients. It may serve as a prognostic signature to identify high-risk early-stage patients.

Abstract

In cancer cells, the process of epithelial–mesenchymal transition (EMT) confers migratory and invasive capacity, resistance to apoptosis, drug resistance, evasion of host immune surveillance and tumor stem cell traits. Cells undergoing EMT may represent tumor cells with metastatic potential. Characterizing the EMT secretome may identify biomarkers to monitor EMT in tumor progression and provide a prognostic signature to predict patient survival. Utilizing a transforming growth factor-β-induced cell culture model of EMT, we quantitatively profiled differentially secreted proteins, by GeLC-tandem mass spectrometry. Integrating with the corresponding transcriptome, we derived an EMT-associated secretory phenotype (EASP) comprising of proteins that were differentially upregulated both at protein and mRNA levels. Four independent primary tumor-derived gene expression data sets of lung cancers were used for survival analysis by the random survival forests (RSF) method. Analysis of 97-gene EASP expression in human lung adenocarcinoma tumors revealed strong positive correlations with lymph node metastasis, advanced tumor stage and histological grade. RSF analysis built on a training set (n = 442), including age, sex and stage as variables, stratified three independent lung cancer data sets into low-, medium- and high-risk groups with significant differences in overall survival. We further refined EASP to a 20 gene signature (rEASP) based on variable importance scores from RSF analysis. Similar to EASP, rEASP predicted survival of both adenocarcinoma and squamous carcinoma patients. More importantly, it predicted survival in the early-stage cancers. These results demonstrate that integrative analysis of the critical biological process of EMT provides mechanism-based and clinically relevant biomarkers with significant prognostic value.

Introduction

Lung cancers are the leading cause of cancer-related deaths worldwide. The advances made in the last decade in diagnosis and treatment have not been translated into significant improvements in overall 5 year survival rates. The tumor-node-metastasis (TNM) staging system combined with pathologic diagnosis has remained the major tool for medical decision making and predicting patient survival (1,2). However, accumulating evidence suggests that though patients with identical histology, differentiation, location and stage at diagnosis are treated by similar therapy, survival is most heterogeneous indicating that the current methods of tumor classification and staging are not sufficient for selecting the best treatment choices and defining prognosis. About 30–55% of early-stage patients who are treated primarily by surgery will have recurrence within 3 years. Recent randomized clinical trails revealed a significant survival advantage in patients receiving chemotherapy after complete resection in the stage IB–IIIA categories (3–6). This trend indicates a need to explore alternative indicators to understand the underlying prognosis of a given patient, to identify the early-stage patients at greatest risk of relapse and decide on appropriate treatment strategy to increase patient survival.

Advances in microarray analysis and proteomics have stimulated research in molecular prognostics and provided alternatives to the traditional TNM-based system for more precise classification and prognostication of human cancers. In different studies, the gene expression signatures derived from tumors at the time of diagnosis have shown promise in predicting long-term patient outcome and outperformed the standard pathologic TNM staging in stratifying breast cancer patients (7–11) and lung cancer patients (12) into high- and low-risk groups. These studies clearly established the fact that tumors contain informative signatures which can be highly valuable in prognosis and estimation of biological response of tumors to therapy over time (7–9). Based on this information, it is possible to identify the subsets of patients who are at high risk of mortality at the time of diagnosis. As a proof of this concept, recently a 70-gene signature was approved for clinical application as a prognostic classifier of breast cancers (10,11). Efforts to identify an analogous gene expression-based prognostic signature of non-small-cell lung cancer have been promising but yet to reach the stage of clinical application (12,13). The predominant approach in all the prognostic signatures to date has been identifying these signatures based on their differential expression between good and poor prognosis groups.

In this study, we attempted an alternative approach to develop a prognostic gene signature based on the cellular process of epithelial–mesenchymal transition (EMT), which plays a critical role in tumor progression. EMT is considered as an initiating event for distant dissemination of tumor cells; proteins secreted during this process may serve as potential biomarkers for patient prognosis. Utilizing a transforming growth factor (TGF)-β-induced EMT model, we quantitatively profiled differentially secreted proteins using label-free 1D gel electrophoresis followed by nanoliquid chromatography, coupled to tandem mass spectrometry (GeLC-MS/MS) analysis of the conditioned media of A549 lung adenocarcinoma cells cultured in the presence or absence of TGF-β. By integrating the secretome with the transcriptome from our earlier study (14), we identified a 97-gene EMT-associated secretory phenotype (EASP) that showed strong correlation to differentiation and stage and predicted survival of lung adenocarcinoma patients in training and independent test sets. We further refined this to a 20-gene signature (rEASP), which performed equally well in predicting survival particularly, in early-stage (stages I and II) adenocarcinomas as well as in squamous cell carcinomas of lung and stratified the lung cancer patients into low-, medium- and high-risk groups with distinct survival times.

Materials and methods

Cell culture

The A549 human lung adenocarcinoma cell line was obtained from the American Type Culture Collection (Manassas, VA) and maintained in RPMI-1640 medium with glutamine, supplemented with 10% fetal bovine serum, penicillin and streptomycin and tested for mycoplasma contamination. All tissue culture media and media supplements were purchased from Life Technologies (Gaithersburg, MD). The porcine TGF-β1 was purchased from PeproTech (Rocky Hill, NJ). In all experiments, cells at 40–50% confluency were serum starved for 24h and treated with TGF-β (5ng/ml) for 72h. At the end, conditioned media collected was centrifuged at 2000g for 20 min and filtered through 0.2 µm filter to remove the intact cells and debris and stored at −80°C until further processing. Cells in the culture dishes were lysed in radioimmunoprecipitation assay buffer and processed for western immunoblotting for assessing the expression of epithelial and mesenchymal markers. Protein concentrations were determined using the BCA Protein Assay Reagent from Pierce (Rockford, IL).

Sample preparation, sodium dodecyl sulfate–polyacrylamide gel electrophoresis and in-gel digestion

Seven milliliters of conditioned media from control and TGF-β-treated cells from two independent biological replicates was buffer exchanged into 25mM ammonium bicarbonate and the volume reduced to 100 µl using a 10kDa molecular weight cutoff filter (Millipore). Half of each replicate (~20 µg protein from controls and 10 µg protein from TGF-β treatment) was solubilized in loading buffer and resolved using Novex 4–12% gradient gels (Invitrogen Life Technologies, Carlsbad, CA). Each lane was manually excised into 40 equal slices and each slice was transferred to a well of a 96-well plate. Proteins in each gel slice were robotically reduced with 10mM dithiothreitol, alkylated with 50 mM iodoacetamide and digested with 160 ng trypsin (ProGest, Genomic Solutions, Ann Arbor, MI). Tryptic peptides were analyzed following acidification with 0.5% formic acid to a final pH 3.8. The volume of peptide mixture for each band was 40 µl.

Data-dependent LC/MS/MS

Thirty microliters of each digested gel slice was analyzed using nano-LC/MS/MS on a LTQ Orbitrap XL tandem mass spectrometer (ThermoFisher, San Jose, CA). Sample was loaded onto an IntegraFrit (New Objective, Woburn, MA) 75 µm × 3cm vented column packed with 0.5mm Jupiter C12 material (Phenomenex, Torrance, CA) at 10 µl/min. Peptides were eluted with a 50min gradient (0.1–30% B in 35min, 30–50% B in 10min and 50–80% B in 5min where A = 99.9% H₂O, 0.1% acetonitrile in 0.1% formic acid and B = 80% acetonitrile, 20% H₂O in 0.1% formic acid) at 300 nl/min using a NanoAcquity HPLC pump (Waters, Beverley, MA) over a 75 µm × 15cm IntegraFrit analytical column packed also with Jupiter C12 material. The column was coupled to a 30 µm ID × 3cm stainless steel emitter (ThermoFisher). Mass spectrometry (MS) was performed in the Orbitrap at 60 000 full width at half maximum resolution and MS/MS was performed in the LTQ on the top six ions in each MS scan using the data-dependent acquisition mode. Normalized collision energy was set at 35% and 1 microscan was used with automatic gain control implementation. Automatic gain control enables the trap to fill with ions to the set ion target values. Target values for MS and MS/MS were 5 × 10⁴ and 1.5 × 10³ counts, respectively. Dynamic exclusion and repeat settings ensured each ion was selected only once and excluded for 30 s thereafter.

Data processing

Data were processed using the MaxQuant v1.0.13.8 software (15), which provides protein identifications at a target false discovery rate. This version of MaxQuant utilizes a locally stored copy of the Mascot search engine (version 2.2; Matrix Science, London, UK) and data were searched against the International Protein Index (IPI) Human v3.53 protein database. Search parameters were: product ion mass tolerance 0.5Da, two missed cleavages allowed, fully tryptic peptides only, fixed modification of carbamidomethyl cysteine, variable modifications of oxidized methionine, N-terminal acetylation and pyroglutamic acid on N-terminal glutamine. Selected MaxQuant parameters were: ‘singlets’ mode, peptide, protein and site false discovery rate 1%, minimum peptide length of five amino acids, minimum of one unique peptide per protein. Proteins identified by this analysis are summarized in Supplementary Table S1, available at Carcinogenesis Online. In MaxQuant, the quantitative measure of each protein is based on the sum of the chromatographic peak area of each peptide matched, termed ‘intensity’. For each protein, a log2 ratio of expression is determined by comparing the average intensity for that protein between the replicates of TGF-β-treated and controls. A protein is determined as differentially expressed if it has >2-fold change in either direction. Log2 ratio >1 is considered as upregulation and <1 is considered as downregulation.

Annotation of secreted proteins and mapping to gene expression

Proteins were annotated as secreted using multiple different bioinformatic tools including SecretomeP (16) for non-classical and leaderless secreted proteins, TMHMM, an HMM-based method for prediction of transmembrane domains (17), SignalP package that detects signal peptides and predicts classical secreted proteins (18), PSORT II that predicts the protein subcellular localization (19) and Secreted Protein Database (SPD) (20). All these predictions are incorporated into Supplementary Table S2, available at Carcinogenesis Online. Others were annotated as secreted proteins based on reported empirical evidence and Gene Ontology (GO) analysis.

Entrez gene identifiers corresponding to the IPI accession numbers of identified proteins were obtained using human IPI cross reference data (‘IPI.genes.HUMAN’ for IPI human release 3.65 from ftp://ftp.ebi.ac.uk/pub/databases/IPI/current). Entrez gene identifiers were used to obtain the corresponding probe set identifiers for the associated arrays from the Affymetrix annotation. Following the above protocol, all the annotated secreted proteins were mapped to our previously published TGF-β-induced EMT time course gene expression data set (GSE 17708) from the same cell line at identical conditions (14). To match the secretome, differentially expressed genes only at 72h time point (5057 probes corresponding to 3397 genes) were used for mapping. Some probes that are identified as differentially expressed but with no assigned gene symbol were excluded.

Gene set enrichment and hierarchical clustering analysis

ConceptGen (http://conceptgen.ncibi.org) is a concept and gene set enrichment analysis tool (14). It will test a given list of genes for overlap and its significance with a specified concept or gene set, which includes GO, direct protein interactions, transcriptional regulation, miRNA targets and gene expression data sets. Using this tool, we performed GO cellular component, cellular process and KEGG pathways enrichment analysis for the 97-gene EASP. Statistically significant (P < 0.001) concepts are presented as network graphs with nodes representing concepts or gene sets and edges representing statistical significance of enrichment.

For clustering, the lists of oncogenic pathways included in the analysis were compiled from the KEGG database, except for embryonic stem cell (ESC) list, which was based on Ben-Porath et al. and Hassan et al. studies (21,22). The expression value for each pathway, including EASP, is the arithmetic mean of all genes in that pathway, giving a single value for each pathway in a given sample. Hierarchical clustering of the Shedden et al. 442 lung adenocarcinoma tumors (23) was performed for indicated oncogenic pathways along with EASP using TreeView (http://www.eisenlab.org/eisen/?page_id=42), and correlations are presented as a heat map with columns representing individual tumors and rows representing the arithmetic mean of a pathway.

Primary tumor-derived gene expression data sets and patient characteristics

Four published Affymetrix microarray data sets representing 908 lung tumors were used in the EASP survival analysis. The CEL files of microarray data were normalized using Robust Multi-array Average method (24). Shedden et al. 442 lung adenocarcinomas (Shedden) were used as training set (23). The other three data sets were used as test sets which included Bild et al. 111 adenocarcinomas and squamous cell carcinoma data set (Bild) (25), Okayama et al. 226 early-stage (stages 1 and 2) adenocarcinoma data set (Okayama) (26) and Raponi’s 129 squamous cell carcinoma data set (Raponi) (27). The patient characteristics and clinical information for these four data sets are provided in Table II. The primary end point was 5 year survival.

Statistical analysis method

The random survival forests (RSF) developed in R package by Ishwaran et al. (28,29) was employed for the EASP survival analysis of the four microarray data sets of lung cancer, as described before (12). Briefly, The RSF is an ensemble tree method for analysis of right-censored survival data. Each decision tree of forests was grown by splitting patients by comparing survival differences via log-rank test based on a randomly selected subset of variables at each node. The 1000 trees were grown for each RSF. Once trees were built, test sets were dropped down to the trees for prediction. The cumulative hazard function was derived from each tree, and an ensemble cumulative hazard function, an average over 1000 survival trees, was determined. Mortality was obtained as a weighted sum over ensemble cumulative hazard function, weighted by the number of individuals at risk at the different time points. Higher mortality values imply the higher risk. We used mortality as risk index to separate patients into three risk groups (high, medium and low risk, one-third each group) and presented Kaplan–Meier survival curves for each group. Each tree provides a measure of its predictive error as described by Ishwaran et al. (28,29), with smaller number indicating a better tree. The prediction error is calculated by C-index (i.e. the Harrell’s concordance index) in the out-of-bag data which were not used for building a tree each time.

Variable importance scores (VIMPS) for all the variables used to grow trees were also generated. Large VIMPS indicate variables as good predictors for outcome, whereas zero or negative values identify non-predictive. These scores were used to refine the 97-gene EASP to the 20-gene rEASP.

Cox proportional hazards regression model, Kaplan–Meier survival curve and log-rank test were used for survival analysis of individual genes or mortality index derived from RSF. The t-test was used to assess the difference of mean expression of EASP signature in clinical and pathological groups including stage, differentiation and nodal status.

Data reporting guidelines

In preparing this manuscript, we followed Molecular & Cellular Proteomics J ournal-recommended guidelines for reporting proteomics data (http://www.mcponline.org/site/misc/PhialdelphiaGuidelinesFINALDRAFT.pdf), the Journal of Clinical Oncology-recommended REMARK guidelines for reporting tumor marker prognostic studies (30,31) and lock-down of the fully specified classifier before application to the first of the four tests of specimens, as mandated by the Institute of Medicine (IOM) (32) and by Hayes et al. (33).

Results

Quantitative identification of differentially secreted proteins during EMT

A549 lung adenocarcinoma cells were cultured in the serum-free media, stimulated with TGF-β for 72h to induce EMT and the conditioned media were collected from control and TGF-β-treated cells for the analysis of differentially secreted proteins. Induction of EMT was confirmed by assessing E-cadherin, N-cadherin and vimentin expression in the cells by western immunoblotting as described before (34,35) (data not shown). Proteins in the conditioned media from two different biological replicates were fractionated by sodium dodecyl sulfate–polyacrylamide gel electrophoresis. Each lane on the gel was cut into 40 slices. Proteins in each gel slice were subjected to trypsin digestion and analyzed by LC-MS/MS on a LTQ Orbitrap mass spectrometer. The resulting MS/MS spectra were analyzed for protein identification and quantitation using MAXQUANT as described under Materials and methods. We identified a total of 2410 proteins, of which 1647 (70%) proteins were annotated as secreted using the multiple data bases and strategies described in Materials and methods (Supplementary Table S2, available at Carcinogenesis Online). With the criteria of at least 2-fold change, we identified 136 proteins as increased in secretion (log2 ratio >1) and 94 proteins as decreased in secretion (log2 ratio < −1) during EMT.

Among the differentially secreted proteins, we observed various categories of proteins including increased secretion of proteases (MMP2, MMP9, BMP1), extracellular matrix components (collagens, fibronectin, versican and SPARC), cytokines (CTGF) and cell surface receptors (mucins, CD59) that are consistent with the migratory, invasive and immune evasive abilities conferred by EMT and their regulation by TGF-β.

EMT-associated secretory phenotype

To identify a gene signature that is representative of EMT and may serve as a reliable biomarker for patient prognosis, we integrated the differentially secreted protein profile with the corresponding gene expression profile we published earlier (14), from the same cell line and under identical conditions. To match with the secretome, differentially expressed genes only at 72h time point were used for integration from the time course data set. Because the goal is to derive a measurable signature, only proteins whose secretion is induced during EMT were considered. By integrating gene and protein expression, we identified 97 genes that are upregulated at mRNA level by at least 2-fold (P > 0.01) and increased in secretion at the protein level by at least 2-fold, irrespective of P value, and defined them as EASP (Table I).

Table I.

List of genes that constitute EASP with corresponding fold change for gene and protein expression

Gene symbol	Entrez ID	Protein accession ID	Gene title	Fold change (TGF-β/control)
				Microarray	Secretome
ADAM19	8728	IPI00011901	A disintegrin and metalloproteinase domain 19 (meltrin beta)	17.47	26.78
ANGPTL4	51129	IPI00153060	Angiopoietin-like 4	19.67	2.33
AP1S2	8905	IPI00909244	Adaptor-related protein complex 1, sigma 2 subunit	2.15	3.23
ARPC4	10093	IPI00925052	Actin-related protein 2/3 complex, subunit 4, 20 kda	2.03	1.41
BMP1	649	IPI00014021	Bone morphogenetic protein 1	4.68	17.90
BPGM	669	IPI00215979	2,3-bisphosphoglycerate mutase	3.85	4.23
CD151	977	IPI00298851	Cd151 antigen	1.81	2.09
CD59	966	IPI00011302	Cd59 antigen	3.99	2.53
CHST11	50515	IPI00099831	Carbohydrate (chondroitin 4) sulfotransferase 11	4.62	4.00
CHST3	9469	IPI00306853	Carbohydrate (chondroitin 6) sulfotransferase 3	4.30	3.61
COL1A1	1277	IPI00297646	Collagen, type i, alpha 1	10.11	3.44
COL4A1	1282	IPI00743696	Collagen, type iv, alpha 1	62.47	5.41
COL4A2	1284	IPI00306322	Collagen, type iv, alpha 2	15.99	5.24
COL4A3	1285	IPI00010360	Collagen, type iv, alpha 3 (good pasture antigen)	3.18	4.57
COL7A1	1294	IPI00025418	Collagen, type vii, alpha 1	3.42	1.35
CRIP2	1397	IPI00921911	Cysteine-rich protein 2	2.86	19.22
VCAN	1462	IPI00215628	Chondroitin sulfate proteoglycan 2 (versican)	2.97	1.89
CTGF	1490	IPI00020977	Connective tissue growth factor	4.96	4.37
CXCL12	6387	IPI00719836	Chemokine (c-x-c motif) ligand 12 (stromal cell- derived factor 1)	5.56	17.71
CYR61	3491	IPI00299219	Cysteine-rich, angiogenic inducer, 61	5.66	2.16
DSC2	1824	IPI00025846	Desmocollin 2	4.12	3.15
ECM1	1893	IPI00645849	Extracellular matrix protein 1	2.09	3.64
EFNA1	1942	IPI00025840	Ephrin-a1	1.72	1.14
EIF4EBP1	1978	IPI00002569	Eukaryotic translation initiation factor 4e-binding protein 1	1.94	20.54
EPHB2	2048	IPI00021275	Eph receptor b2	8.07	1.76
FHL2	2274	IPI00396967	Four and a half lim domains 2	3.15	3.31
FN1	2335	IPI00845263	Fibronectin 1	5.14	3.15
FST	10468	IPI00021081	Follistatin	6.45	1.30
FSTL1	11167	IPI00029723	Follistatin-like 1	5.14	2.18
FSTL3	10272	IPI00025155	Follistatin-like 3 (secreted glycoprotein)	3.98	2.21
C11orf41	25758	IPI00852979	G2 protein	3.98	20.18
GALNT2	2590	IPI00004669	Udp-n-acetyl-alpha-d-galactosamine:polypeptide n-acetylgalactosaminyltransferase 2 (galnac-t2)	2.78	1.69
GSN	2934	IPI00646773	Gelsolin (amyloidosis, finnish type)	3.99	2.70
HMGA2	8091	IPI00005996	High mobility group at-hook 2 /// high mobility group at-hook 2	4.20	3.13
HMOX1	3162	IPI00215893	Heme oxygenase (decycling) 1	3.41	17.36
HSPB1	3315	IPI00025512	Heat shock 27 kda protein 1	2.28	2.42
IGF1	3479	IPI00433029	Insulin-like growth factor 1 (somatomedin c)	4.65	1.11
IGF2	3481	IPI00215977	Insulin-like growth factor 2 (somatomedin a)	1.03	1.87
IGFBP5	3488	IPI00029236	Insulin-like growth factor-binding protein 5	21.60	7.17
IGFBP7	3490	IPI00016915	Insulin-like growth factor-binding protein 7	63.94	5.33
IL11	3589	IPI00025820	Interleukin 11	39.43	5.57
INHBA	3624	IPI00028670	Inhibin, beta a (activin a, activin ab alpha polypeptide)	24.32	25.47
ITGA2	3673	IPI00013744	Integrin, alpha 2 (cd49b, alpha 2 subunit of vla-2 receptor)	5.13	1.15
ITGA3	3675	IPI00290043	Integrin, alpha 3 (antigen cd49c, alpha 3 subunit of vla-3 receptor)	1.90	1.80
JAG1	182	IPI00099650	Jagged 1 (Alagille syndrome)	3.10	1.87
MGC17330	113791	IPI00298388	Phosphoinositide-3-kinase interacting protein 1	2.53	20.35
KIAA1797	54914	IPI00748360	Kiaa1797	1.70	19.36
LAMC2	3918	IPI00015117	Laminin, gamma 2	20.02	7.45
LEFTY2	7044	IPI00010893	Left-right determination factor 2	2.23	23.89
LIF	3976	IPI00009720	Leukemia inhibitory factor (cholinergic differentiation factor)	2.71	2.22
XYLT1	64131	IPI00183487	Hypothetical protein loc283824	12.13	23.61
LTBP1	4052	IPI00784258	Latent transforming growth factor beta-binding protein 1	2.86	2.25
LTBP2	4053	IPI00292150	Latent transforming growth factor beta-binding protein 2	9.42	4.66
LTBP3	4054	IPI00073196	Latent transforming growth factor beta-binding protein 3	3.87	1.55
LTBP4	8425	IPI00873371	Latent transforming growth factor beta-binding protein 4	3.01	2.14
PIK3IP1	113791	IPI00298388	Hgfl gene /// hgfl gene	2.53	20.35
MMP1	4312	IPI00008561	Matrix metalloproteinase 1 (interstitial collagenase)	5.81	6.27
MMP10	4319	IPI00013405	Matrix metalloproteinase 10	20.03	25.23
MMP2	4313	IPI00027780	Matrix metalloproteinase 2	10.86	8.36
MMP9	4318	IPI00027509	Matrix metalloproteinase 9	1.34	22.00
MRC2	9902	IPI00005707	Mannose receptor, c type 2	4.97	1.47
NPC2	10577	IPI00301579	Niemann-pick disease, type c2	2.88	3.33
NPTX1	4884	IPI00220562	Neuronal pentraxin i	7.29	7.41
NRG1	3084	IPI00221375	Neuregulin 1	3.27	2.19
PAWR	5074	IPI00001871	Prkc, apoptosis, wt1, regulator	2.34	20.17
PCDH1	5097	IPI00872579	Protocadherin 1 (cadherin-like 1)	4.48	22.09
PDGFB	5155	IPI00000044	Platelet-derived growth factor beta polypeptide	4.86	18.85
PDLIM2	64236	IPI00007983	Pdz and lim domain 2 (mystique)	2.74	16.79
PGRMC2	10424	IPI00005202	Progesterone receptor membrane component 2	1.85	2.20
PLAT	5327	IPI00019590	Plasminogen activator, tissue	4.41	19.85
PLAUR	5329	IPI00010676	Plasminogen activator, urokinase receptor	3.07	2.22
PLOD2	5352	IPI00337495	Procollagen-lysine, 2-oxoglutarate 5-dioxygenase 2	2.81	2.03
PLSCR3	57048	IPI00216127	Phospholipid scramblase 3	2.25	3.22
PPP1R14B	26472	IPI00398922	Protein phosphatase 1, regulatory (inhibitor) subunit 14b	2.11	1.55
HTRA1	5654	IPI00003176	Protease, serine, 11 (igf binding)	2.63	3.00
PTPRK	5796	IPI00470937	Protein tyrosine phosphatase, receptor type, k	6.19	1.64
RSU1	6251	IPI00847168	Ras suppressor protein 1	2.00	1.41
SCG2	7857	IPI00009362	Secretogranin ii (chromogranin c)	31.96	7.63
SEMA3C	10512	IPI00019209	Sema domain, immunoglobulin domain (ig), short basic domain, secreted, (semaphorin) 3c	4.87	3.70
SERPINE1	5054	IPI00007118	Serine (or cysteine) proteinase inhibitor, clade e (nexin, plasminogen activator inhibitor type 1), member 1	41.73	4.13
SERPINE2	5270	IPI00009890	Serine (or cysteine) proteinase inhibitor, clade e (nexin, plasminogen activator inhibitor type 1), member 2	17.90	4.36
SPOCK1	6695	IPI00005292	Sparc/osteonectin, cwcv and kazal-like domains proteoglycan (testican)	70.53	5.84
STC1	6781	IPI00005564	Stanniocalcin 1	10.09	3.99
TAGLN	6876	IPI00216138	Transgelin	13.35	6.44
TAGLN2	8407	IPI00647915	Transgelin 2	2.24	1.17
TAX1BP3	30851	IPI00005585	Tax1 (human t-cell leukemia virus type i)-binding protein 3	1.83	3.54
TGFB1	7040	IPI00000075	Transforming growth factor, beta 1	2.57	2.22
TGFBR1	7046	IPI00005733	Transforming growth factor, beta receptor i	5.72	2.72
THBS1	7057	IPI00296099	Thrombospondin 1	16.73	4.71
TIMP2	7077	IPI00027166	Tissue inhibitor of metalloproteinase 2	4.72	2.54
TLL2	7093	IPI00465231	Tolloid-like 2	2.59	2.15
TNFAIP6	7130	IPI00303341	Tumor necrosis factor, alpha-induced protein 6	5.03	7.48
TNFRSF12A	51330	IPI00010277	Tumor necrosis factor receptor superfamily, member 12a	5.10	1.83
TP53I3	9540	IPI00384643	Tumor protein p53 inducible protein 3	3.35	2.47
TUBA4B	80086	IPI00017454	Tubulin, alpha 4	2.24	2.20
ULBP2	80328	IPI00018860	Ul16-binding protein 2	5.18	3.51
VEGFA	7422	IPI00012567	Vascular endothelial growth factor	4.66	2.86

Twenty genes that are part of rEASP are in bold.

For functional interpretation, EASP was subjected to gene set enrichment analysis using ConceptGen (14). Analysis for cellular components has associated EASP with extracellular matrix, proteinaceous extracellular, collagen, basement membrane, matrix space, matrix part and matrix region part (Figure 1A), consistent with their annotation as secretory proteins. More importantly, enrichment analysis for biological processes has associated EASP with the cellular processes including cell adhesion, motility, actin cytoskeleton reorganization, coagulation, acute inflammatory response, proteolysis and response to wounding and external stimuli (Figure 1A), consistent with the biology of EMT. Moreover, this also demonstrates that EASP is a true representation of EMT and may serve as reliable biomarker to track EMT.

Fig. 1.

(A) Analysis of EMT secretome for enriched biological processes and cellular components. All the genes listed in EASP signature (Table I) were uploaded to ConceptGen (http://conceptgen.ncibi.org) tool. The enriched GO cellular components (P < 0.0001) and biological processes (P < 0.0001) are shown. Each node size is proportional to the number of genes in the enriched process or pathway and each edge represents a statistically significant enrichment with defined P value. (B) Cluster analysis of EASP signature and other cancer-related pathways in 442 lung adenocarcinomas. Column represents individual cancer sample, row represents different pathway (mean value was used for each specific pathway). Ccpos, positive cell cycle gene; ESC1_380, embryonic stem cell 380 genes. Red indicated higher expression and green is lower expression.

To assess the correlation of EASP with other known oncogenic pathways, we performed hierarchical clustering of 442 lung adenocarcinomas based on their mean gene expression of the indicated pathway. Clustering analysis yielded two distinct lung adenocarcinoma tumor groups with 50% tumors demonstrating higher expression of all pathways. Mean EASP expression pattern correlated with mean gene expression of all the oncogenic pathways tested. These include NF-κβ, antiapoptosis, JAK-STAT, Notch, AKT, WNT pathways and ESC signature (Figure 1B). All these pathways are known to be deregulated in lung adenocarcinomas and were implicated in the regulation of EMT.

Correlation of EASP with clinical variables

We determined the ability of the EASP signature to stratify the patients based on tumor stage, differentiation and nodal status using the gene expression data derived from the Shedden et al. 442 lung adenocarcinoma patients (23) (Figure 2A). The EASP signature was able to identify the patients with well-differentiated tumors from moderately and poorly differentiated tumors (P < 0.001). Similarly, the EASP signature was able to separate patients with stage I tumors from stage II and III (P ≤ 0.01). Furthermore, the EASP signature expression is high in patients with positive nodal status (N1–2) compared with patients with negative nodal status (N0). Together, these results indicate the potential clinical utility of EASP in predicting aggressive tumor behavior.

Fig. 2.

(A) Correlation of EASP expression with differentiation, stage and nodal status. All the patients in Shedden data set (n = 442) were classified into different subgroups based on tumor differentiation, tumor stage or lymph node status recorded at the time of diagnosis. The mean expression of EASP signature (mean centered value) in tumors is significantly different in differentiation (well vs poor), tumor stage (stage 1 vs stage 3) and lymph node status (N0 vs N1–2). (B) Testing of 97-gene EASP signature as a predictor of patient survival in lung cancer: lung cancer patients from Bild, Okayama and Raponi data sets were stratified using 97-gene EASP signature into low-, medium- and high-risk groups (one-third in each group) based on survival analysis by RSF model built on Shedden 442 training set. Shown are the Kaplan–Meir survival curves based on mortality. The log-rank P values compare the groups stratified by EASP signature.

EASP stratifies lung cancer patients into low-, medium- and high-risk groups with distinct survival

In order to investigate whether the 97-gene EASP signature could predict the overall survival in non-small-cell lung cancer patients, the Shedden data set (n = 442) was used as training set. As detailed in Materials and methods, a mathematical model based on an RSF algorithm was built in the training set to predict the prognostic significance of EASP with stage, age and sex included. After locking down the model, it was tested in three independent publicly available lung cancer data sets, Bild et al. (n = 111) (25), Okayama et al. (n =226) (26) and Raponi et al. (n = 129) (27). These cohorts include lung adenocarcinoma and squamous cell carcinoma patients. The prediction error rates were 33.6, 30.0 and 36.7%, respectively, for the Bild, Okayama and Raponi data sets (Supplementary Table S3, available at Carcinogenesis Online). We tested the usefulness of RSF predictors using a univariate Cox model with the mortality index as a continuous measure. The RSF prediction was significant for the Bild test set (likelihood ratio test [LRT] P = 0.00008), Okayama test set (LRT P = 0.005) and Raponi test set (LRT P = 0.02). In all three test sets, low-, medium- and high-risk groups were clearly separated by mortality index (Figure 2B). The hazard ratios (HRs) were 1.00, 2.17 and 3.16 for the Bild data set (log-rank test, P = 0.003); 1.00, 2.60 and 5.37 for Okayama data set (log-rank test, P = 0.002) and 1.00, 0.96 and 2.61 for the Raponi data set (log-rank test, P = 0.002), for low-, medium- and high-risk groups, respectively (Supplementary Table S3, available at Carcinogenesis Online).

Refining the 97-gene EASP to the 20-gene rEASP

In order to refine the EASP into a smaller gene subset that remains as effective in predicting survival of lung cancer patients, the 97-gene EASP was filtered based on the VIMPS generated by RSF analysis of the Shedden et al. 442 data set. Higher VIMP values indicate variables with predictive ability, whereas zero or negative values identify non-predictive variables. This resulted in a refined EASP (rEASP) comprising of top 20 genes with higher VIMPS (Supplementary Figure S1, available at Carcinogenesis Online). To assess whether the 20-gene rEASP performs as well as the 97-gene EASP signature in predicting patient survival, another model based on RSF algorithm was built to predict the prognostic significance of rEASP with stage, age and sex included, using Shedden data set (n = 442) as training set. Next, we tested the predictive power of this 20-gene rEASP model in all three independent test sets described above (Bild, Raponi and Okayama) with clinical information in Table II. The prediction error rates were 35.6, 29.9 and 36.4%, respectively for the Bild, Okayama and Raponi data sets (Table III). We tested the usefulness of RSF predictors using a univariate Cox model with the mortality index as a continuous measure. The RSF prediction was significant for the Bild test set (LRT P = 0.002), Okayama test set (LRT P = 0.0004) and Raponi test set (LRT P = 0.007). In all three test sets, low-, medium- and high-risk groups were clearly separated by mortality index (Figure 3). The HRs were 1.00, 1.54 and 2.34 for the Bild data set (log-rank test, P = 0.03); 1.00, 2.37 and 3.75 for Okayama data set (log-rank test, P = 0.002) and 1.00, 2.38 and 2.79 for the Raponi data set (log-rank test, P = 0.02), for low-, medium- and high-risk groups, respectively (Table III).

Table II.

Clinical characteristics of samples used in this study

Data set	Shedden set	Bild set	Raponi set	Okayama set
Sample number	442	111	129	226
Type of cancer	Ad	58 Ad/53 SCC	SCC	Ad
Age average	64.4	64.8	67.5	59.6
Gender
Female	219	48	48	121
Male	224	63	82	105
Stage
Stage I	276	67	73	168
Stage II	105	18	34	58
Stage III	59	21	23	0
Differentiation
Well	60	NA	15	NA
Moderate	209	NA	76	NA
Poor	167	NA	39	NA
Dead (5 year)	188	58	52	32
Alive	255	53	78	194
Adjuvant therapy
Yes	109		48
No	330		69	204
Unknown	3	111	12	22

Ad, adenocarcinomas; NA, not available; SCC, squamous cell cancer.

Adjuvant therapy includes chemo- and/or radiotherapy.

Table III.

Prediction results of 20-gene rEASP signature on three test sets

	RSFa	Cox modelb	HR	Log rank testc	P
	Test error rate	P		95% Confidence interval
Okayama test set (n = 226)
	29.9%	0.0004
Low risk			1		0.002
Medium risk			2.37	1.03–5.46
High risk			3.75	1.70–8.28
Raponi test set (n = 129)
	36.4%	0.007
Low risk			1		0.02
Medium risk			2.38	1.12–5.10
High risk			2.79	1.32–5.90
Bild test set (n = 111)
	35.6%	0.002
Low risk			1		0.03
Medium risk			1.54	0.78–3.02
High risk			2.34	1.24–4.42

^aRSF prediction model built from the 442 training set including 20 genes, age, gender and stage.

^bMortality risk index (MRI) as continuous value, LRT was used in univariate Cox model.

^cMRI separated test patients to three risk groups (low, medium and high-risk, one-third in each group).

Fig. 3.

Testing of 20-gene rEASP signature as a predictor of patient survival in lung cancer: lung cancer patients from Bild’s, Okayama’s and Raponi’s data sets were stratified using 20-gene rEASP signature into low-, medium- and high-risk groups based on survival analysis by RSF model built on Shedden 442 training set. Shown are the Kaplan–Meir survival curves based on mortality. The log-rank P values compare the groups stratified by EASP signature.

Discussion

Contrary to the perception that tumor metastasis progresses in a linear and step-wise fashion, recent evidence suggests that a subset of tumors harbor molecular alterations at an early stage that are indicative of bad prognosis and poor patient survival (36). This demonstrates the importance of identifying molecular changes at an early stage that dictate clinical behavior. The current system of TNM staging cannot identify such changes. There is an urgent need to develop prognostic tests that can predict recurrence and identify high-risk patients at an early stage when they would benefit from adjuvant therapy (5,37–44). Demonstrating the utility of such a prognostic test, a 70-gene signature (10) (MammaPrint; Agendia, The Netherlands) has been approved by the Food and Drug Administration for breast cancer patients (32). A 21-gene signature (45) (Oncotype DX; Genome Health, CA) is approved for breast cancers, with analogous signatures under development for prostate and colon cancers. Even though multiple gene, protein, autoantibodies and miRNA-based profiles have been proposed for lung cancer prognosis, none to date has been approved for clinical use.

The predominant approach in deriving most of the prognostic signatures has been profiling differentially occurring molecular changes between good versus bad outcome groups, without any consideration to the underlying tumor biology. Here, we adopted a new, mechanism-based approach to identify predictive biomarkers by profiling the complex cellular process of EMT, which is implicated in the initiation of tumor metastasis. The rationale behind this approach is that identifying proteins secreted during the course of a critical biological process that promotes a metastatic phenotype would provide relevant, reliable and robust prognostic biomarkers. The other novel aspect of this approach is that one can measure EASP at the mRNA level and also at the protein level, because the biomarkers of EASP are based on the strong concordant expression of both mRNA and protein. Given that EMT is the initiating event for metastasis and may result in the dissemination of tumor cells, measuring EASP in the primary tumor, tumor cells in bone marrow compartment and circulating tumor cells may allow the ability to track disease progression.

Consistent with the functional attributes conferred by EMT and its regulation by TGF-β, we identified proteins that are implicated in tumor cell adhesion, migration, invasion, immune evasive mechanisms, extracellular matrix components and tumor-stromal interactions. Gene set enrichment analysis of the 97-gene EASP, which is the subset of all up regulated proteins and their mRNAs, also identified biological processes that are reflective of EMT and TGF-β biology. Similarly, even the proteins in rEASP are representative of the functional EMT phenotype. Furthermore, clustering analysis of EASP with key oncogenic pathways showed a similar correlation to the expression of various pathways that are deregulated in lung adenocarcinomas. These include NF-κβ, antiapoptosis, JAK-STAT, PTEN, AKT, WNT, Notch, Hedgehog and EGFR signaling pathways (46,47). Most importantly, the correlation of EASP expression with ESC signature is consistent with the recent finding that the ESC signature is associated with poor prognosis and worse overall survival in lung adenocarcinoma patients (22). This correlation is also consistent with finding that EMT may confer stem cell-like properties to breast cancer cells (48). Together, these observations demonstrate that EASP not only reflects the heterogeneity and complexity associated with oncogenesis of lung cancer, but also demonstrates the significance and relevance of EASP biomarkers to the underlying biology of tumor metastasis.

Consistent with its prognostic significance, EASP distinguished well from moderate or poorly differentiated tumors and stage 1 from stage 2 and 3 patients. Most importantly, it was strongly correlated with positive lymph node status, which is an important prognostic factor that influences the therapeutic decision making and probability of lung cancer recurrence. To test the clinical utility of EASP, the RSF analysis-based survival model was built and trained on 442 primary lung adenocarcinoma tumor-derived gene expression data set, the largest lung cancer gene expression data set available with pathological, clinical and treatment annotations (23). Using VIMPS from the training set, we refined the EASP into a subset of 20 genes (rEASP) with highest VIMPS and tested its prognostic significance in three independent lung cancer data sets. Because the EASP was derived from an adenocarcinoma cell line, we asked whether EASP is specific to adenocarcinomas or does it have any relevance to other subtypes of lung cancer. To address this, we selected the Okayama et al. data set of only adenocarcinomas (n = 226), the Bild et al. data set of adenocarcinomas (n = 58) and squamous cell carcinomas (n = 53) and the Raponi et al. data set of only squamous cell carcinomas (n = 129) for independent testing. Interestingly, in all three independent test sets, both EASP- and rEASP-based models were able to stratify patients into low-, medium- and high-risk groups with clearly separated mortality indexes and distinct HRs. This demonstrates the relevance of these models even for lung squamous cell cancer. It is important to note that both EASP and rEASP performed very well in the Okayama et al. data set, which is comprised of only early-stage patients (stages I and II). Because rEASP predicted the survival of early-stage patients, it might serve as an ideal prognostic signature to identify the high-risk early-stage patients who might benefit the most from adjuvant therapy. In earlier studies, we identified multiple inhibitors of EMT in lung cancer (49,50), and it might be beneficial to test these agents or other EMT blockers as adjuvants for patients with high EASP expression.

In conclusion, this study demonstrates the importance of a mechanism-based approach that integrates multiple omics data sets, to identify clinically relevant biomarkers for patient prognosis. Biomarkers rooted in underlying molecular and cellular biology of the tumors may provide very useful and actionable information for patient care.

Supplementary material

Supplementary Tables S1–S3 and Figures S1 can be found at http://carcin.oxfordjournals.org/

Funding

National Institutes of Health/ National Cancer Institute (CA132571-01); American Cancer Society (RSG-CSM-116801 to V.G.K.); NIH (RM-08-029, P30U54ES017885 to G.S.O).

Conflict of Interest Statement: None declared.

Glossary

Abbreviations:

EASP

EMT-associated secretory phenotype

EMT

epithelial–mesenchymal transition

ESC

embryonic stem cell

GO

Gene Ontology

HR

hazard ratio

IPI

International Protein Index

LC

liquid chromatography

LRT

likelihood ratio test

MS

mass spectrometry

MS/MS

tandem mass spectrometry

RSF

random survival forests

TGF

transforming growth factor

TNM

tumor-node-metastasis

VIMPS

variable importance scores.