Oncogene (2007) 26, 4600–4608 & 2007 Nature Publishing Group All rights reserved 0950-9232/07 $30.00 www.nature.com/onc Identification of a metastasis signature and the DLX4 homeobox as a regulator of metastasis by combined transcriptome approach

S Tomida1,5, K Yanagisawa1,5, K Koshikawa1, Y Yatabe2, T Mitsudomi3, H Osada4 and T Takahashi1

1Division of Molecular Carcinogenesis, Center for Neurological Diseases and Cancer, Nagoya University Graduate School of Medicine, Nagoya, Japan; 2Department of Pathology and Molecular Diagnostics, Aichi Cancer Center Hospital, Nagoya, Japan; 3Department of Thoracic Surgery, Aichi Cancer Center Hospital, Nagoya, Japan and 4Division of Molecular Oncology, Aichi Cancer Center Research Institute, Nagoya, Japan

Although widespread metastasis is the major cause of Introduction human -related deaths, its underlying mechan- ism remains largely unclear. Our genome-wide compari- Cancer metastasis is the major cause of cancer-related son of the expression profiles of a highly metastatic lung deaths. It has long been understood that a minute cancer cell line, NCI-H460-LNM35 (LNM35), and its fraction of primary cancer cells acquire multiple genetic parentalclone,NCI-H460-N15 (N15), resultedin the and/or epigenetic changes in a multi-step manner, which identification of a cancer metastasis signature composed eventually confer the ability to metastasize to distant of 45 . Through ontology analysis, our study organs (Fidler, 2003). This concept of an evolution of also provided insights into how this 45-gene metastasis metastatic clones within the primary tumor site was signature may contribute to the acquisition of metastatic proposed mostly on the basis of the findings in potential. By applying the signature to datasets of human experimental models, but a provocative study by cancer cases, we could demonstrate significant associa- Ramaswamy et al recently provided intriguing evidence tions with a subset of cases with poor prognosis not only for the possible acquisition of metastatic potential early for the two datasets of cancers of the lung but also for in carcinogenesis, which may consequently confer cancers of the breast. Furthermore, we were able to show metastatic capabilities to the bulk of primary tumors that enforced expression of the DLX4 homeobox gene, (Ramaswamy et al., 2003; Pantel and Brakenhoff, 2004). which was identified as a gene with significant down- Considerable advances have been attained in the regulation in LNM35 as well as with significant associa- understanding of the molecular carcinogenesis of lung tion with favorable prognosis for lung cancer patients, cancer (Osada and Takahashi, 2002), but it still remains markedly inhibited in vitro motility and invasion as well as the leading cause of cancer death in economically in vivo metastasis via both hematogenous and lympho- developed countries including Japan (Jemal et al., genous routes. Taken together, these findings indicate that 2005). The long-term survival rate continues to be our combined transcriptome analysis is an efficient unsatisfactory, and no more than 50% of the cases who approach in the search for genes possessing both clinical successfully undergo potentially curative resection sur- usefulness in terms of prognostic prediction in human vive for more than 5 years after surgery, the remaining cancer cases and clear functional relevance for studying cases eventually suffering widespread metastases or local cancer biology in relation to metastasis. recurrence. Unfortunately, very little is known about Oncogene (2007) 26, 4600–4608; doi:10.1038/sj.onc.1210242; how lung cancer cells give rise to distant metastasis. published online 29 January 2007 Identification of molecules with a crucial role in distant spread is therefore of the utmost importance to reduce Keywords: lung cancer; metastasis; expression profile; the intolerable number of deaths owing to this ONCOGENOMICS DLX4; devastating disease. To this end, we previously established a highly metastatic subline, NCI-H460-LNM35 (hereafter re- ferred to as LNM35), which consistently and sponta- neously metastasizes through both hematogenous and lymphogenous routes when injected either subcuta- Correspondence: T Takahashi, Division of Molecular Carcinogenesis, Center for Neurological Diseases and Cancer, Nagoya University neously or orthotopically (Kozaki et al., 2000). In the Graduate School of Medicine, 65 Tsurumai-cho, Showa-ku, Nagoya present study, we aimed at identifying a metastasis 466-8550, Japan. signature with both functional and clinical relevance, E-mail: [email protected] first selecting genes with abundant and differential 5 These two authors contributed equally to this work. The microarray expression between LNM35 and N15 cells, and then data were deposited in GEO (GSE4705 and GSE4716). Received 26 September 2006; revised 21 November 2006; accepted 28 questioning whether they could be used to discrimi- November 2006; published online 29 January 2007 nate lung cancer patients in terms of postoperative Metastasis signature and DLX4 S Tomida et al 4601 prognoses. Finally, we determined whether the DLX4 N15 cell lines, we compared the expression profiles of homeobox gene, one of the metastasis signature genes 11 168 probes. A total of 4849 probes remained after the with uncharacterized features in relation to metastasis, first filtering, which omitted probes with expression could functionally play a role in metastasis-related values below 0.1. Two-hundred and forty-nine probes abilities both in vitro and in vivo. further passed the 2.0-SD variability threshold filtering, and 60 probes corresponding to 45 unique genes also met the ‘robustness criteria’. Twenty-seven genes were Results overexpressed, whereas 18 were significantly reduced in association with the acquisition of a metastatic capa- Identification of genes associated with the acquisition of bility during in vivo selection of LNM35 (Table 1). The metastatic potential in LNM35 27 examples overexpressed in LNM35 included various In order to identify genes differentially expressed genes previously suggested to play a functional role in between highly metastatic LNM35 and low metastatic metastasis such as chemokine (C-X-C motif)ligand

Table 1 Genes differently expressed in LNM35 vs. N15 Rank UGcluster Genesymbol Description Ratio

Genes upregulated in LNM35 1 Hs.789 CXCL1 CXCL1: Chemokine (C-X-C motif)ligand 1 ( growth stimulating 17.4 activity, alpha) 2 Hs.509585 Transcribed sequences 9.2 3 Hs.193235 CPLX2 Complexin 2 8.6 4 Hs.518464 PSMD2 Proteasome (prosome, macropain)26S subunit, non-ATPase, 2 6.2 5 Hs.478481 ALG3 Asparagine-linked glycosylation 3 homolog (yeast, 6.1 alpha-1,3-mannosyltransferase) 6 Hs.503749 H2-ALPHA Alpha-tubulin isotype H2-alpha 5.9 7 Hs.533059 TUBB Tubulin, beta 5.5 8 Hs.522584 TMSB4X Thymosin, beta 4, X-linked 4.7 9 Hs.74368 CKAP4 Cytoskeleton-associated protein 4 4.2 10 Hs.234074 DNER Delta-notch-like EGF repeat-containing transmembrane 3.9 11 Hs.509736 HSPCB Heat shock 90 kDa protein 1, beta 3.7 12 Hs.533273 UBE1 Ubiquitin-activating E1 (A1S9 T and BN75 temperature sensitivity 3.1 complementing) 13 Hs.196176 ECH1 Enoyl Coenzyme A hydratase 1, peroxisomal 3.0 14 Hs.14074 MRPS22 Mitochondrial ribosomal protein S22 2.9 15 Hs.75318 TUBA1 Tubulin, alpha 1 (testis specific)2.8 16 Hs.349695 TUBA2 Tubulin, alpha 2 2.8 17 Hs.77269 GNAI2 Guanine nucleotide binding protein (G protein), alpha inhibiting activity 2.7 polypeptide 2 18 Hs.76095 IER3 Immediate early response 3 2.7 19 Hs.516539 HNRPA3 Heterogeneous nuclear ribonucleoprotein A3 2.5 20 Hs.350966 PTTG1 Pituitary tumor-transforming 1 2.5 21 Hs.82890 DAD1 Defender against cell death 1 2.5 22 Hs.155218 HNRPUL1 Heterogeneous nuclear ribonucleoprotein U-like 1 2.4 23 Hs.499709 SLC16A9 Solute carrier family 16 (monocarboxylic acid transporters), member 9 2.3 24 Hs.444986 METAP2 Methionyl aminopeptidase 2 2.3 25 Hs.105465 SNRPF Small nuclear ribonucleoprotein polypeptide F 2.3 26 Hs.519880 TFAP2A Transcription factor AP-2 alpha (activating enhancer binding protein 2 alpha)2.3 27 Hs.125898 GNAS GNAS complex 2.3

Genes downregulated in LNM35 28 Hs.150580 SUI1 Suppressor of initiator codon mutations 0.5 29 Hs.503787 DARS Aspartyl-tRNA synthetase 0.5 30 Hs.406013 KRT18 Keratin 18 0.5 31 Hs.461086 CDH1 Cadherin 1, type 1, E-cadherin (epithelial)0.4 32 Hs.390738 FLJ20366 Hypothetical protein FLJ20366 0.4 33 Hs.524430 NR4A1 Nuclear receptor subfamily 4, group A, member 1 0.4 34 Hs.172648 DLX4 Distal-less homeobox 4 0.4 35 Hs.511915 ENO2 2 (gamma, neuronal)0.4 36 Hs.348883 FOXC1 Forkhead box C1 0.4 37 Hs.418123 CTSL Cathepsin L-like 3 0.3 38 Hs.534360 TRIP6 Thyroid hormone receptor interactor 6 0.3 39 Hs.28988 GLRX Glutaredoxin (thioltransferase)0.3 40 Hs.531081 LGALS3 Lectin, galactoside-binding, soluble, 3 (galectin 3)0.3 41 Hs.433732 CLK1 CDC-like kinase 1 0.3 42 Hs.28491 SAT Spermidine/spermine N1-acetyltransferase 0.2 43 Hs.531176 SARS Seryl-tRNA synthetase 0.2 44 Hs.404321 GARS Glycyl-tRNA synthetase 0.2 45 Hs.497599 WARS Tryptophanyl-tRNA synthetase 0.1

Oncogene Metastasis signature and DLX4 S Tomida et al 4602 Table 2 Gene Ontology (GO)terms significantly related to the metastasis signature GO aspect GO term GO accession P

Biological processes Microtubule polymerization GO:0046785 o0.0001 Microtubule-based movement GO:0007018 o0.0001 Protein biosynthesis GO:0006412 0.0020 Morphogenesis GO:0009653 0.0029 Molecular functions GTPase activity GO:0003924 o0.0001 activity GO:0016874 0.0012 GTP binding GO:0005525 0.0015 Cellular component Tubulin GO:0045298 o0.0001

1 (CXCL1)(Bobrovnikova-Marjon et al., 2004; dataset of 50 surgically treated cases of non-small cell Gallagher et al., 2005), pituitary tumor-transforming lung cancers (NSCLCs), acquired in our previous study, 1 (PTTG1)(Heaney et al., 2000; Bernal et al., 2002), using the same microarray platform (Aichi dataset) immediate early response 3 (IER3)(Goswami et al., (Tomida et al., 2004). Unsupervised hierarchical cluster- 2004; Nicholson et al., 2004), defender against cell death ing, using the 45 metastasis-signature gene set, clearly 1 (DAD1)(Ayala et al., 2004; Goswami et al., 2004; differentiated the 50 NSCLC cases into two distinct Kim et al., 2006)and methionyl aminopeptidase 2 clusters, cluster 1 and cluster 2 (Figure 1a). It was (METAP2)(Dasgupta et al., 2005; Morowitz et al., intriguing that patients in cluster 1 showed significantly 2005). Similarly, genes with higher expression in N15 more unfavorable prognosis than those in cluster 2 also included some previously implicated in metastasis (Figure 1b; P ¼ 0.0234, log-rank test). It is worth noting such as cadherin 1 (E-cadherin)(Cavallaro and Christo- that 18 of 22 genes (82%)clustered in the upper branch fori, 2004)and lectin galactoside-binding soluble 3 of the side tree were derived from LNM35-over- (galectin-3)(Liu and Rabinovich, 2005). expressed genes (indicated in yellow), whereas 14 of 23 genes (61%)in the lower branch of the side tree were derived from N15-overexpressed/LNM35-downregu- Characterization of the 45-gene metastatic signature lated genes (indicated in green)( P ¼ 0.0058, Fisher’s using gene ontology terms exact test). In addition, multivariate analysis revealed In order to gain functional insight into how the that metastasis signature defined cluster to be a metastasis signature genes might contribute in the significant prognostic factor independent of disease acquisition of metastatic potential in LNM35, we stage in terms of five-year postoperative survival in this therefore employed our Gene Ontology (GO)term dataset (Table 3). These findings suggested that the identifier of functions and other characteristics that are 45- signature, which was captured as a utilized in association with certain phenotypes of metastasis signature through analysis of our metastasis interest (Takeuchi et al., 2006). With this identifier, model, may also play a role in human lung cancer we found that GO terms related to four biological metastasis. It should be noted that the 45-gene meta- processes, three molecular functions and one cellular stasis signature did not overlap with our previous component were significantly more frequently observed 25-gene prognosis prediction classifier, which was than expected (P 0.005), when the frequency of a GO o constructed specifically for analyzing the Aichi dataset, term appearing in the 45-gene metastatic signature was possibly because of the use of distinct filtering criteria compared with that in the entire set of 8644 unique (Tomida et al., 2004). However, clusters 1 and 2 in genes on the microarrays used in this study (Table 2). Figure 1a predominantly include cases with predictions The acquisition of metastatic potential in LNM35 was of fatal and favorable prognoses, which seems to show consequently suggested to be associated with changes in consistency with the notion that most lung cancer- microtubule-related biological processes (P 0.0001 for o related deaths are attributable to metastasis. microtubule polymerization and P 0.0001 for micro- o The general applicability of the 45-gene metastasis tubule-based movement)and GTP-related molecular signature was further examined with the aid of an functions (P 0.0001 for GTPase activity and P o ¼ additional set of 62 Dana-Faber/MIT (Boston)stage 0.0015 for GTP binding). In addition, it is worth noting I/II lung adenocarcinoma cases, which includes those that morphogenesis (P 0.0029)was selected as a ¼ with distinctive postoperative prognoses, i.e., either statistically significant term. ‘alive without recurrence for at least 4 years’ or ‘dead with recurrence’ (Ramaswamy et al., 2003). Again, we Clinical significance of the 45-gene metastatic signature observed the presence of two major branches with for postoperative prognosis for human lung and distinctly different postoperative survival (P ¼ 0.0554, breast cancers Figure 2a). The applicability of the signature to other As the next step towards investigating the relevance and solid tumors was similarly examined by using a Nether- generality of the metastasis signature, we asked whether lands Cancer Institute (Amsterdam)dataset of breast the 45-gene metastasis signature identified in our meta- cancer as well as Dana-Faber/MIT (Boston)datasets of stasis model could be applied to human lung cancer medulloblastoma and diffuse large B-cell specimens to predict clinical behavior. We analyzed a (Ramaswamy et al., 2003), all of which contained

Oncogene Metastasis signature and DLX4 S Tomida et al 4603

Figure 1 Kaplan–Meier survival curves based on favorable/unfavorable categorization of hierarchical clustering analysis using the 45-gene metastasis signature. (a)Analysis of the Aichi dataset comprising 50 NSCLC cases, showing clear distinctions between the two resulting clusters. Columns represent individual patients and rows represent genes constituting the 45-gene metastasis signature. Yellow and green boxes on the left indicate upregulated and downregulated expression in LNM35, respectively, in comparison with N15. Boxes below indicate the results of prognosis prediction (red for fatal and blue for favorable), using our previous 25-gene prognosis classifier for NSCLC constructed with the Aichi dataset. (b)Kaplan-Meier survival curves of cluster 1 and cluster 2 for Aichi 50 NSCLC cases. information of either overall survival or time to 5-year clinical follow-up was analyzed by means of metastasis. The Amsterdam dataset of 78 cases with unsupervised hierarchical clustering analysis. Our me- stage I primary breast adenocarcinomas with at least tastasis signature again clearly showed two major

Oncogene Metastasis signature and DLX4 S Tomida et al 4604 Table 3 Multivariate Cox regression analysis of potential prognostic factors for survival 5 years after surgery Variables Unfavorable/Favorable Hazard Ratio 95% CIa P

Age >63/p63 1.21 0.43–3.38 0.72 Sex Male/Female 4.64 0.56–38.6 0.16 Smoking Current & Former/Never 0.49 0.07–3.44 0.47 Stage II-III/I 2.15 1.16–3.98 0.014 Histology Squamous/Non-squamous 0.89 0.27–2.98 0.86 Metastasis signature Cluster 1/Cluster 2 3.12 1.06–9.17 0.039

a95% CI, 95% confidence interval.

Figure 2 Analysis of Boston dataset of 62 lung adenocarcinomas (a)and Amsterdam dataset of 78 breast cancers ( b)as well as Boston datasets of 60 medulloblastomas (c)and 58 diffuse large B-cell ( d), showing significantly different prognosis in solid tumors but not in hematologic malignancy, according to the clusters based on the expression profiles of the 45-gene metastasis signature.

groups with a significant difference in postoperative by applying permutation t-test analysis to the Aichi prognosis (P ¼ 0.0173, Figure 2b). We also observed a dataset used for hierarchical clustering analysis, and similar difference between two major clusters in the found that the distal-less homeobox 4 (DLX4)was the hierarchical clustering analysis of the Boston dataset most significantly favorable prognosis-associated gene of 60 medulloblastomas with at least 5-year clinical (P ¼ 0.0234)(Supplementary Table).As the functional follow-up after multimodality treatment (P ¼ 0.0507, relationship between DLX4 and metastasis has not yet Figure 2c). Interestingly, however, the analysis of the been explored, we further investigated whether augmen- Boston dataset of 58 diffuse large B-cell lymphomas did tation of the expression level of DLX4 could affect the not identify any statistically significant differences invasive and metastatic nature of LNM35. (P ¼ 0.876, Figure 2d). Our microarray analysis revealed that low-metastatic N15 cells expressed the favorable-prognosis-associated Identification of DLX4 as a metastasis signature gene gene, DLX4, at a level 2.5-fold higher than highly with the functional involvement in metastasis of LNM35 metastatic LNM35 cells, which was also confirmed at The purpose of the combined transcriptome analysis is the protein level by Western blot analysis showing to identify genes that are both clinically useful for readily detectable expression of the DLX4 protein in prognosis for human cancer cases and functionally N15 cells in contrast to negligible expression in LNM35 relevant for studying cancer biology in relation to cells (Figure 3a). A DLX4 expression construct was metastasis. We therefore performed in silico screening then stably transfected into LNM35, yielding DLX4-

Oncogene Metastasis signature and DLX4 S Tomida et al 4605

Figure 3 Introduction of DLX4 into the highly metastatic LNM35 cells, showing inhibition of the invasive and metastatic phenotype. (a)DLX4 protein expression in LNM35, DLX4-transfected LNM35 clones 17 and 18 as well as in the low-metastatic N15, a parental clone of LNM35. (b)Comparison of in vitro growth characteristics. (c and d)Comparison of in vitro motility and invasion. Marked reduction is evident in DLX4-transfected clones 17 and 18, but not in vector control (VC1), to levels comparable with those for N15. Significant reduction of in vivo metastasis via both hematogenous (e)and lymphogenous ( f)routes with DLX4-transfected LNM35 clones 17 and 18. transfected LNM35 cell clones stably expressing DLX4 at indicating that it is generic in nature. Interestingly, a level comparable to N15 cells. Their morphologies and however, this signature did not appear to be related to growth rates were not noticeably altered (Figure3b and outcome for diffuse large B-cell lymphomas, which is data not shown for morphologies), but DLX4 transfec- consistent with the idea that metastasis of hematopoietic tants had significantly reduced capabilities in terms of tumors may employ a mechanism or mechanisms both motility and invasion in vitro (Figures 3c and d). specific for navigating the hematologic and lymphoid Motility of two DLX4 transfectants was significantly cells (Ramaswamy et al., 2003). It is of note that this reduced down to less than 30% of the LNM35 level, and metastasis signature overlaps partly with that identified invasiveness of both transfectants was similarly reduced. by Ramaswamy et al. (2003)through comparisons of Furthermore, the introduction of DLX4 into the highly expression profiles between primary and metastatic sites metastatic LNM35 cells clearly abrogated metastasis of human cancers. For example, PTTG1 (also known as in vivo, resulting in the marked reduction in the abilities securin)and small nuclear ribonucleoprotein (SNRPF) to form metastases via both hematogenous and lympho- were upregulated in common, and downregulation of genous routes (Figures 3e and f). nuclear hormone receptor (NR4A1)was included as metastatic phenotype-associated genes in both studies. A few previous expression profiling studies have used Discussion paired metastatic and non-metastatic cell lines for identifying a gene set involved in the process of It is well accepted that the acquisition of metastatic metastasis. As expected, they identified partially over- phenotype, the most deadly manifestation, is a highly lapping gene sets, but clearly distinct sets of genes were complex process involving multiple genes. Our study also selected depending on the models used by each clearly showed our 45-gene metastasis signature, se- research group. One at Memorial Sloan-Kettering lected on the basis of differential expression between Cancer Center found a gene set associated with either highly metastatic- and low-metastatic LNM35 sublines, bone or lung metastasis in patients using a to be significantly associated with postoperative prog- highly metastatic clone of the MDA-MB-231 human nosis in two independent sets of lung cancers, the Aichi breast cancer cell line, and certain genes such as CXCL1 and Boston datasets. Furthermore, the 45-gene metas- were shared with our gene set (Minn et al., 2005). tasis signature was also significantly associated with Similarly, our metastasis signature gene set contains prognosis for breast cancers and medulloblastomas, related to the ubiquitin-proteasome pathway, including

Oncogene Metastasis signature and DLX4 S Tomida et al 4606 UBE1 and PSMD2, whereas others like USP22 and the 45-gene metastasis signature, including the two as RNF2 were included in the 11-gene signature identified yet uncharacterized genes, should provide a clue to by Glinsky et al. (2005)through analysis of highly better understanding of processes underlying tumor metastatic cell line. These findings progression, and may ultimately lead to the develop- suggest that certain functions are shared in common ment of novel treatment or preventative approaches. between organs for acquiring metastatic phenotype, but at the same time there are also many other constituents that are not shared. Materials and methods Our GO term analysis also sheds light on the constituents of the 45-gene metastasis signature as a Cell lines and animals whole. It is interesting that the microtubule-related Establishments of LNM35, a highly metastatic subline of the biological process was found to be significantly asso- NCI-H460 human large cell lung cancer cell line, and its parent, ciated with the acquisition of metastatic potential, N15, have been reported previously (Kozaki et al., 2000). They considering that the involvement of microtubules in cell were maintained in Rosewell’s Park Memorial Institute media extensions that participate in active cell migration is well (RPMI)1640 medium, supplemented with 10% fetal bovine established, and that some of the agents that bind to serum, as described earlier (Kozaki et al., 2000). Five-week-old female severe combined immunodeficient (SCID)mice were tubulin and disrupt microtubule dynamics are in clinical purchased from CLEA Japan, Inc. (Tokyo, Japan)and use for cancer treatment (Jordan and Wilson, 2004). maintained under specific-pathogen-free conditions. Although the GO term selected as being significant is associated with GTPase-related activities, GTPases are known to be involved in cell morphogenesis through the Microarray data acquisition and analyses Two sets of membrane cDNA microarrays (GeneFilter induction of specific types of actin cytoskeleton and the Human Microarrays Release I and Release II; Invitrogen, alignment and stabilization of microtubules. In addi- Carlsbad, CA, USA), containing a total of 11 168 spots tion, the selection of morphogenesis as a statistically corresponding to 8644 independent genes, were analyzed in significant term is intriguing, as tumors are often viewed duplicate using five microgram of total RNA as detailed in our as corrupt forms of normal developmental processes, previous report, with microarray data acquisition as detailed and genes that play an important role in embryonic previously (Tomida et al., 2004). The raw data were re-scaled development are frequently found to be altered in cancer to account for the differences in individual hybridization (Kang and Massague, 2004). Epithelial-mesenchymal intensities as follows. First, array spots that showed expression transition (EMT)is a vital step in morphogenesis during values below 0.1 were omitted from the analysis before embryonic development, and it is also thought to be normalization. Then, expression levels for each gene were normalized within each of the independent hybridizations by involved in the conversion of early stage tumors into using loess nonlinear normalization in the statistical package invasive malignancies (Kang and Massague, 2004). R, available at http://www.R-project.org (Gentleman et al., In this connection, it is worth noting that loss of E- 1994). The average expression values of LNM35 were cadherin, a hallmark of EMT in cancer, appears to have compared to those of N15, and array spots that met the occurred in LNM35 as part of the process of acquisition following ‘robustness criteria’ were selected; expression values of metastatic potential during in vivo selection. of at least 1.0 in the two independent hybridizations in LNM35 We further demonstrated that the DLX4 homeobox and/or N15 as well as differential expression at the level of gene, which was reduced in LNM35 and was most more than 2.0-SD of that considering all the spots. Averaged significantly associated with favorable prognosis in lung expression values of probes corresponding to the same unique cancer patients, could reduce motility and invasion gene were used for further analysis. in vitro as well as their metastasis in vivo through both hematogenous and lymphogenous routes. Our findings, Hierarchical clustering and survival analysis therefore, indicate that DLX4 plays a functional role Expression profiles of the 45-gene metastasis signature in 50 affecting invasive and metastatic capabilities, and is not NSCLC cases were extracted from our previous expression a mere surrogate. Homeobox genes of the Distal-less profiling dataset (Tomida et al., 2004), which is available at GEO as GSE4705 and GSE4716. The microarray datasets of (Dlx)family are expressed in vertebrate embryos at the Boston lung cancers, Amsterdam breast cancers and Boston contact of epithelial cell layers and adjacent mesench- medulloblastomas and diffuse large B-cell lymphomas, de- ymal cells, playing a role in epithelial-mesenchymal cell scribed in a paper by Ramaswamy (Ramaswamy et al., 2003), interactions (Quinn et al., 1998). In this context it is of are publicly available at http://www.broad.mit.edu/cgi-bin/ particular interest that DLX4 has also been suggested cancer/datasets.cgi. as datasets B, C, E and F, respectively. to regulate trophoblast invasion (Quinn et al., 1998). The GPL91 platform file, available at GEO (http:// We should also mention that a CpG island in the www.ncbi.nlm.nih.gov/geo/), was used to annotate each 50 upstream region of DLX4 was recently reported to be Affymetrix probe set ID of dataset B with the UniGene ‘Gene methylated in breast cancer cell lines (Miyamoto et al., Symbol’. We used the UniGene Gene Symbols for our cross- 2005). platform mapping of genes, and 40 genes in dataset B were mapped from among the 45-gene set. These genes were then In conclusion, expression-profiling analyses of a meta- used for hierarchical clustering of the Boston lung data set. stasis model, in combination with human cancer speci- The gene symbols contained in the ‘Description’ column mens, appear to be a highly relevant way to investigate of dataset C were similarly used for mapping, with the result genes of both functional and clinical importance. that 41 genes in dataset C were mapped from among the Functional clarification of additional components of 45-gene set, and then used for the Amsterdam breast data set.

Oncogene Metastasis signature and DLX4 S Tomida et al 4607 For datasets E and F, the GPL80 platform file was used for USA), according to the manufacturer’s instructions. Trans- cross-platform mapping of the genes, resulting in the mapping fected cells were selected with 1 mg/ml G418, and Western blot of 28 genes from among the 45-gene set. Averaged expression analyses were performed with an anti-DLX4 polyclonal values of probes corresponding to the same unique gene were (Santa Cruz Biotechnology Inc., Santa Cruz, CA, further analyzed. The CLUSTER (Eisen et al., 1998)program USA)and a horseradish peroxidase-conjugated secondary was used for average linkage hierarchical clustering of both antibody (Cell Signaling Technology, Beverly, MA, USA). genes and cases by means of median centering and normali- zation. The results were displayed with the aid of TREEVIEW (Eisen et al., 1998), and the resultant two clearly distinct In vitro motility and invasion assay and in vivo spontaneous clusters containing predominantly LNM-35-upregulated and metastasis assay downregulated genes were termed ‘Cluster 1’ and ‘Cluster 2’, To quantify in vitro motility and invasion, transwell-chamber respectively. The Kaplan–Meier method was used to estimate culture systems were employed. The upper surface of 6.4 mm- survival as a function of time, and survival differences were diameter filters with 8 mm pores (BD Biosciences, Bedford, analyzed with the log-rank test. Cox proportional hazards MA, USA)was coated with 0.1 ml of 0.1 mg/ml Matrigel modeling was performed to identify which independent factors (Collaborative Research Inc., Bedford, MA, USA)for the might jointly have a significant effect on survival. All the invasion assay, and the upper chambers were filled with 0.4 ml statistical analyses were performed with Stata software of serum-free RPMI 1640 medium and placed on culture plates 4 (version 7; Stata Corp, College Station, TX, USA). with 24 wells containing 1 ml of the medium. Cells (1.0 Â 10 cells for the motility assay, 1.0 Â 105 cells for the invasion assay)in 0.1 ml of serum-free RPMI 1640 medium were then Bioinformatic analyses added to the upper chambers and cultured for 24 h. The filters et al GO (Ashburner .,. 2000)analysis was employed to were fixed with 70% ethanol and stained with Giemsa, and the highlight functionally distinct biological features of a gene cells on the lower surface of the filters were counted in set associated with the acquisition of invasive and metastatic triplicate. Three independent experiments gave similar results, capabilities in LNM35, as described previously (Takeuchi and the representative results are shown for each analysis. et al ., 2006). Briefly, database files used for this GO analysis Spontaneous metastasis assay was performed as follows. were downloaded from the UniGene ftp site. Eventually, 5346 1.0 Â 107 cells in 0.1 ml of serum-free RPMI 1640 medium were unique genes on the microarray were linked to about 30 000 injected into subcutaneous tissue of the right abdominal wall GO terms by parsing the database files including Hs.seq.all, of 6-week-old female SCID mice. Three mice each were Hs.data and LL_tmpl with the program written by Perl. These transplanted with LNM35, N15 and VC1, respectively, terms were subjected to Fisher’s exact test to identify which whereas five mice were subcutaneously injected with the GO terms were over- or under-represented in a gene set of LNM35-DLX transfectant. One mouse injected with interest. LNM35-DLX4-19 cells died before the end of the 40-day t Permutation -test analysis was used for selecting genes that observation period because of peritonitis carcinomatosa. Forty were significantly different in the favorable and unfavorable days after injection, the mice were sacrificed by cervical prognosis patient subgroups. We then randomly permuted dislocation under deep anesthesia and their lungs, lymph nodes the class labels among the samples and re-computed the and subcutaneous tumors were resected, weighed and fixed t -statistics. This randomization was repeated 1000 times, and with 4% formaldehyde. The lung-metastatic nodules were P the -value for the observed class label was calculated from the examined and counted under a dissecting microscope. All the t distribution of 1000 sets of -statistics. results presented in this study were obtained by averaging these data. Construction and transfection of the DLX4 expression vector and Western blot analyses The IMAGE clone (clone ID 3907376)containing a full-length Acknowledgements DLX4 cDNA was purchased from Invitrogen. The DLX4 expression vector was constructed by inserting the SmaI- We thank Keishi Katoh for his helpful discussion on digested DNA fragment from IMAGE clone into the EcoRV bioinformatic analysis. This work was supported by a Grant- site of pcDNA3 (pcDNA3-DLX4)and then sequenced. in-Aid for Scientific Research on Priority Areas from the LNM35-DLX4 stable transfectants (clones 17 and 19)were Ministry of Education, Culture, Sports, Science and Techno- generated by transfection of 2 mg of pcDNA3-DLX4, using the logy of Japan and a Grain-in-Aid for Scientific Research (B) FuGENE 6 reagent (Roche Applied Science, Indianapolis, IN, from the Japan Society for the Promotion of Science.

References

Ashburner M, Ball CA, Blake JA, Botstein D, Butler H, Bobrovnikova-Marjon EV, Marjon PL, Barbash O, Vander Cherry JM et al. (2000). Gene ontology: tool for the Jagt DL, Abcouwer SF. (2004). Expression of angiogenic unification of biology The Gene Ontology Consortium. Nat factors vascular endothelial growth factor and interleukin-8/ Genet 25: 25–29. CXCL8 is highly responsive to ambient glutamine availa- Ayala G.E, Dai H, Ittmann M, Li R, Powell M, Frolov A et al. bility: role of nuclear factor-kappaB and activating protein-1. (2004). Growth and survival mechanisms associated Cancer Res 64: 4858–citepn>4869. with perineural invasion in prostate cancer. Cancer Res 64: Cavallaro U, Christofori G. (2004). Cell adhesion and 6082–6090. signalling by cadherins and Ig-CAMs in cancer. Nat Rev Bernal JA, Luna R, Espina A, Lazaro I, Ramos-Morales F, Cancer 4: 118–132. Romero F et al. (2002). Human securin interacts with p53 Dasgupta B, Yi Y, Hegedus B, Weber JD, Gutmann DH. and modulates p53-mediated transcriptional activity and (2005). Cerebrospinal fluid proteomic analysis reveals apoptosis. Nat Genet 32: 306–311. dysregulation of methionine aminopeptidase-2 expression

Oncogene Metastasis signature and DLX4 S Tomida et al 4608 in human and mouse neurofibromatosis 1-associated glioma. a human lung cancer cell line NCI-H460-LNM35 Cancer Res 65: 9843–9850. with consistent lymphogenous metastasis via both subcuta- Eisen MB, Spellman PT, Brown PO, Botstein D. (1998). neous and orthotopic propagation. Cancer Res 60: Cluster analysis and display of genome-wide expression 2535–2540. patterns. Proc Natl Acad Sci USA 95: 14863–14868. Liu FT, Rabinovich GA. (2005). Galectins as modulators of Fidler IJ. (2003). The pathogenesis of cancer metastasis: the tumour progression. Nat Rev Cancer 5: 29–41. ’seed and soil’ hypothesis revisited. Nat Rev Cancer 3: Minn AJ, Gupta GP, Siegel PM, Bos PD, Shu W, Giri DD 453–458. et al. (2005). Genes that mediate breast cancer metastasis to Gallagher PG, Bao Y, Prorock A, Zigrino P, Nischt R, Politi V lung. Nature 436: 518–524. et al. (2005). Gene expression profiling reveals cross-talk Miyamoto K, Fukutomi T, Akashi-Tanaka S, Hasegawa T, between melanoma and fibroblasts: implications for host- Asahara T, Sugimura T et al. (2005). Identification of tumor interactions in metastasis. Cancer Res 65: 4134–4146. 20 genes aberrantly methylated in human breast cancers. Gentleman RC, Lawless JF, Lindsey JC, Yan P. (1994). Multi- Int J Cancer 116: 407–414. state Markov models for analysing incomplete disease Morowitz MJ, Barr R, Wang Q, King R, Rhodin N, Pawel B history data with illustrations for HIV disease. Stat Med et al. (2005). Methionine aminopeptidase 2 inhibition is an 13: 805–821. effective treatment strategy for in preclinical Glinsky GV, Berezovska O, Glinskii AB. (2005). Microarray models. Clin Cancer Res 11: 2680–2685. analysis identifies a death-from-cancer signature predicting Nicholson BE, Frierson HF, Conaway MR, Seraj JM, therapy failure in patients with multiple types of cancer. Harding MA, Hampton GM et al. (2004). Profiling the J Clin Invest 115: 1503–1521. evolution of human metastatic . Cancer Res Goswami S, Wang W, Wyckoff JB, Condeelis JS. (2004). 64: 7813–7821. Breast cancer cells isolated by chemotaxis from primary Osada H, Takahashi T. (2002). Genetic alterations of multiple tumors show increased survival and resistance to chemother- tumor suppressors and oncogenes in the carcinogenesis and apy. Cancer Res 64: 7664–7667. progression of lung cancer. Oncogene 21: 7421–7434. Heaney AP, Singson R, McCabe CJ, Nelson V, Nakashima M, Pantel K, Brakenhoff RH. (2004). Dissecting the metastatic Melmed S. (2000). Expression of pituitary-tumour trans- cascade. Nat Rev Cancer 4: 448–456. forming gene in colorectal tumours. Lancet 355: 716–719. Quinn LM, Latham SE, Kalionis B. (1998). A distal-less class Jemal A, Murray T, Ward E, Samuels A, Tiwari RC, Ghafoor A homeobox gene, DLX4, is a candidate for regulating et al. (2005). Cancer statistics, 2005. CA Cancer J Clin 55: epithelial-mesenchymal cell interactions in the human 10–30. placenta. Placenta 19: 87–93. Jordan MA, Wilson L. (2004). Microtubules as a target for Ramaswamy S, Ross KN, Lander ES, Golub TR. (2003). A anticancer drugs. Nat Rev Cancer 4: 253–265. molecular signature of metastasis in primary solid tumors. Kang Y, Massague J. (2004). Epithelial-mesenchymal transi- Nat Genet 33: 49–54. tions: twist in development and metastasis. Cell 118: Takeuchi T, Tomida S, Yatabe Y, Kosaka T, Osada H, 277–279. Yanagisawa K et al. (2006). Expression profile-defined Kim YH, Girard L, Giacomini CP, Wang P, Hernandez- classification of lung adenocarcinoma shows close relation- Boussard T, Tibshirani R et al. (2006). Combined micro- ship with underlying major genetic changes and clinico- array analysis of small cell lung cancer reveals altered pathologic behaviors. J Clin Oncol 24: 1679–1688. apoptotic balance and distinct expression signatures of Tomida S, Koshikawa K, Yatabe Y, Harano T, Ogura N, MYC family gene amplification. Oncogene 25: 130–138. Mitsudomi T et al. (2004). Gene expression-based, indivi- Kozaki K, Miyaishi O, Tsukamoto T, Tatematsu Y, Hida T, dualized outcome prediction for surgically treated lung Takahashi T. (2000). Establishment and characterization of cancer patients. Oncogene 23: 5360–5370.

Supplementary Information accompanies the paper on the Oncogene website (http://www.nature.com/onc).

Oncogene