| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
Regular Articles |
Department of Medical Oncology, Dana Farber Cancer Institute [R. C. O., C. W. B., K. K., M. D., C. C., N. E. S., L. C.], Department of Biostatistics, Harvard School of Public Health [A. S., X. Z., W. H. W.], Belfer Cancer Genomics Centre, Dana Farber Cancer Institute [M. D., C. C.], and Department of Dermatology, Harvard Medical School [L. C.], Boston, Massachusetts 02115; Department of Automation, Tsinghua University, Beijing 100084, China [X. Z.]; and Department of Statistics, Harvard University, Cambridge, Massachusetts 02138 [W. H. W.]
 |
ABSTRACT |
|---|
 Top ABSTRACT INTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONREFERENCES |
|---|
|
INTRODUCTION |
|---|
|
TopABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONREFERENCES |
|---|
Recent advances in microarray technology have revolutionized our ability to catalogue chromosomal imbalances and gene expression patterns present in tumors. Microarray-based expression studies have established the feasibility of tumor classification based on gene expression patterns (4 , 6 , 7, 8, 9, 10, 11, 12) . In addition, profiling studies in human tumors, such as melanomas, have implicated tumor suppressor genes (13) and oncogenes with plausible links to metastasis (14) , suggesting that genome-wide expression profiles might serve as a means of identifying genes and pathways governing specific cancer-relevant processes. Although array-CGH has been highly effective in cataloging recurrent CNAs associated with a number of different tumor types in humans (15, 16, 17, 18, 19, 20) and in engineered cancer-prone mouse strains (21, 22, 23, 24) , array-CGH analyses have not yet been assessed as an approach to segregate tumors on the basis of differing etiologies or genetics or to identify those CNAs with potential links to specific tumor biological processes. To investigate whether array-CGH profiles can be used for these purposes, we compared the scope and nature of the chromosomal CNAs by array-CGH in phenotypically and histopathologically homogeneous melanomas arising spontaneously or after exposure to UVB in mouse models.
A Tyr-RAS transgenic model of melanoma has demonstrated that activated RAS mutation can cooperate with loss of p16INK4a, p19ARF, or both to drive melanoma formation.5 To explore the role of neonatal UV exposure in melanoma development, we subjected Tyr-RAS+ p16Ink4a -/- and Tyr-RAS+ p19ARF-/- animals to a single erythrogenic dose of UVB on postnatal day 1 and compared the latency and incidence of melanoma in these UV-treated animals with that in the untreated cohort (26) . We found that UV exposure greatly accelerated melanoma formation in Tyr-RAS+ p19ARF-/- animals but not in Tyr-RAS+ p16Ink4a -/- mice. Despite different germ-line mutations (p16INK4a or p19ARF deficiency) and etiologies (spontaneous versus UV treated), the resultant melanomas were histopathologically indistinguishable and frequently sustained mutations that inactivated the Rb or p53 pathways in the p19ARF or p16INK4a mutant mice, respectively (26) . Given their common tumor biological end points, these highly related melanomas from genetically equivalent Tyr-RAS+ p19ARF-/- mice, differing only in their history of UV exposure, provided a stringent system in which to test the hypothesis that profiles of CNAs contain discriminatory information on tumor etiology and pathogenesis, and such unbiased information can uncover molecular events underlying etiologic difference and lead to new mechanistic insights.
|
MATERIALS AND METHODS |
|---|
|
TopABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONREFERENCES |
|---|
Labeling and Hybridization to Arrays.
Array-CGH experiments were performed with SpectralChip arrays (Spectral Genomics, Houston, TX) according to the manufacturer’s protocol. These arrays comprise 976 mapped mouse BACs from the RPCI11 BAC library. One µg of tumor and normal genomic DNA was labeled with Cy3-cCTP or Cy5-dCTP (Amersham Pharmacia Biotech). Fluor-reversal experiments were performed in all cases. Hybridizations were performed in sealed chambers for 20 h at 60°C. After hybridization, slides were rinsed in 2 x SSC, 0.5% SDS at room temperature; washed for 40 min in 2 x SSC, 0.5% SDS at 65C; washed for 10 min in 0.2 x SSC at 65°C; and washed for 1 min in 0.2 x SSC at 2°C. Slides were then air dried by centrifugation before imaging.
Image and Data Analysis.
Sixteen-bit TIF images were collected using an Axon 4000B scanner and processed initially using GenePix Pro. Subsequently, custom software was used to exclude spots that demonstrated low signal:noise ratios or poor spot morphology. Data were normalized to the ratio of medians of Cy3 and Cy5, and the normalized values converted to log2 ratios to weight gains and losses equally. Dye swap experiments were merged to calculate the mean and SD for each BAC.
Clustering and Classification.
Cluster (http://rana.ibl.gov/) was used to cluster log2 values for array-CGH data processed as described above, using all BACs. Subsequently, BACs were excluded if the processing software generated at least one "missing value" for any tumor. Under this condition, 874 BACs were included in additional hierarchical clustering and classification experiments. The data were standardized so that BACs had common mean and variance.
A method based on the linear SVM (27
, 28)
was used to classify tumors using their array-CGH profiles for the remaining 874 BACs. Given a training set, in this case, tumors whose UV status is known, SVM builds a linear classifier:
 |
A software package, written by X. Z. and W. H. W., was used to implement the recursive classification and feature selection procedure introduced in "Results." This package incorporates the SVMTorch (29) software. Initially, SVMTorch is used to build an SVM on the 874 BACs identified above. The BACs with the largest contribution to the classifier are then selected. The contribution for a given BAC is measured by the difference between its two class means weighted by its coefficient in the SVM. A new SVM is generated on the reduced set, and a subset of these BACs is selected using the same procedure. This is done recursively, resulting in a series of nested subsets of BACs and a series of SVM models built using these BACs. For the current purpose of evaluating the feasibility of classification, we arbitrarily set the number of selected BACs in the above successive stages to be 500, 200, 100, 50, 30, and 20.
Cross-validation is implemented by removing one tumor sample from the data set and performing the above procedure on the remaining 37 samples. The series of models obtained at different selection levels are then applied to the excluded sample to predict its UV status. This process is duplicated for each tumor in the sample. The accuracy of the predictions across experiments is summarized for each size of classifier.
qPCR and RT-PCR.
Primers for PCR and RT-PCR corresponded to the primers for the indicated MIT markers or were designed from the 3'untranslated region of the indicated genes using Primer 3. Reverse transcription was performed on 3 µg of total RNA from each tumor using the Superscript reverse transcriptase kit (Stratagene) and an oligo-dT primer according to the manufacturer’s specifications in 20 µl of total volume. PCR was performed using 100 ng of genomic DNA or 1 µl of reverse transcriptase reaction mix with marker- or gene-specific primers for 17 cycles of 95°C for 30 s, 56°C for 1 min, and 72°C for 45 s. PCR products were run on a 1% agarose Tris-borate EDTA gel and transferred to Hybond N+. cdk6 and glyceraldehyde-3-phosphate dehydrogenase were not visible on eithidium bromide-stained gels and only detected by hybridization to specific probes. Probes were generated from pooled normal genomic DNA from the mice of origin for each tumor by PCR using the same primers as for the quantitative PCR analysis. Random-primed 32P-labeled probes were hybridized for 2 h at 65°C in RapidHyb (Amersham Pharmacia). Quantification was performed using a phosphorimager.
Mapping of Genes and Markers.
Physical map positions of genes and markers were obtained from the Ensembl mouse genome database v4.1.1, January 31, 2002.
|
RESULTS AND DISCUSSION |
|---|
|
TopABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONREFERENCES |
|---|
. For each tumor, regional chromosomal CNAs in tumor genomic DNA were compared with somatic DNA from the same mouse. The threshold for defining a CNA was based on the distribution of log ratios in self-self hybridizations (data not shown). Regions were scored as significantly increased/decreased if the CGH ratio was 4 SDs above/below the mean (+/- 0.35, log base 2).
|
). This global difference in genomic profiles suggests that UV exposure significantly affected the nature of accumulated mutations in p19ARF-/- melanomas and that different patterns of CNAs are likely to be identified in UV versus non-UV tumors. In contrast, UV treatment did not alter the frequency of CNAs in melanoma from Tyr-RAS+ p16Ink4a -/- animals (Table 1)
, consistent with the lack of biological impact by UV exposure in p16INK4a-/- animals (as measured by incidence, latency, and clinical behavior of melanoma, as well as status of the pRB and p53 pathways in the tumors; Ref. 26
). We first used a method based on the linear SVM to classify the non-UV and UV-induced p19ARF-/- tumors based on their array-CGH profiles (see "Materials and Methods"; Refs. 27 and 28 ). Thirty-eight p19ARF-/- tumors with high quality array-CGH data were used for this part of the analysis. Only 874 BACs showed valid readings in all 38 of these tumors. This filtering procedure eliminated the need to make assumptions about the causes of the low information content in the discarded tumors or BACs. Hierarchical clustering of these 38 tumors using 874 BACs was virtually identical to that based on the full set of 931 BACs (data not shown), suggesting that the ability to classify tumors is not substantially reduced by this filter.
Because the number of descriptive spots (BACs) exceeded the number of tumor samples tested, we reduced the number of BACs considered in the classification scheme by feature selection. In designing a feature selection procedure to form classifiers based on smaller numbers of BACs, an optimal SVM model was built using all available BACs, and the BACs were ranked according to their contribution to this model. A subset of the BACs that made the largest contribution to this model was selected, and a new SVM model was built with these selected BACs. The selected BACs were then reranked according to their contribution to the new model, and the selection was repeated. For each model, the performance of SVM can be assessed. For our current study, we have arbitrarily chosen the numbers of selected features (BACs) for this recursive procedure to be 500, 200, 100, 50, 30, and 20. As shown in Fig. 1
, the 874-features classifier achieved a higher accuracy in the UV-treated cohort alone, whereas the 20-feature classifier achieved a higher accuracy in the non-UV cohort. However, the overall classification accuracy (combined for both cohorts) is similar for both. The observation that a small-feature classifier was able to achieve similar accuracy as the large data set suggests the likelihood of identifying key stereotypical changes that may serve as potential prognostic/predictive markers.
|
shows the results of this validation procedure for the 38 tumors, where the data are standardized so that each BAC has common mean and variance. This analysis revealed that we were able to achieve a maximum overall accuracy of classification of 76.5%. To assess the significance of this finding, we repeated the cross-validation procedure on 100 random permutations of the UV labels for all of the 38 tumors. We found that none of the 100 runs of R-SVM permutation test achieved a better cross-validation error rate than that observed in the actual classification, indicating that the 76.5% accuracy is significant. In other words, UV exposure history can be predicted retrospectively based on the array-CGH profiles of the resulting tumors. This finding is particularly encouraging in view of the fact that detailed histopathological and classical analyses were not able to distinguish tumors with or without UV exposure history, suggesting that array-CGH profiles reveal aspects of cancer etiology that are not detectable by standard approaches. Finally, it is unlikely that array-CGH merely identified a UV signature profile, because SVM analysis was unable to distinguish the UV status of melanomas arising from Tyr-RAS+ p16INK4a-/- animals (data not shown). This lack of distinction on the genomic level is consistent with the absence of any observed UV effect in melanoma genesis on the p16Ink4a-/- background (26)
. In summary, the computational analysis of array-CGH profiles was capable of correctly classifying tumors that exhibited differences in biological behavior (p19ARF-/- UV versus non-UV) but could not differentiate among tumors with similar biological behavior (p16Ink4a-/- UVversus non-UV). Having shown that the array-CGH profiles contain statistically significant information to predict the UV etiology, we asked whether we could identify a subset of CNAs that are most informative at telling the two classes of tumors apart and whether such signature-CNAs can provide clues to the mechanism of UV’s melanoma-promoting activity. SVM is a powerful tool for the identification of classifying information among linear combinations of BACs, but it is not designed to test the classification power of individual BACs. Because most loci identified by array-CGH at the current resolution are represented by changes in a single BAC, we investigated an approach that could be applied to individual BACs to identify ones with the greatest classification power.
For this exercise, we chose to focus on regions of gain rather than loss attributable to the greater sensitivity of array-CGH to copy number increase, which can be many-fold, than typical single copy losses in regions of deletion. BACs were scored as "amplified or gained" if the CGH ratio was 4 SDs above the mean log ratio for self-self hybridization controls (>0.35, log base 2). For each BAC, we calculated the difference in the number of tumors that show copy number increase (i.e., "amplification" or "gain") in the UV versus non-UV cohorts. Next, we rank ordered all of the informative BACs by the magnitude of such difference and listed the top 20 in Table 2
. In other words, the BACs listed are those with the greatest classification power (i.e., most informative in separating the UV and non-UV tumors) and may thus point to gene(s) governing biological differences between the two tumor classes. It is also noteworthy that these top classifying features are found not only in the cohort with significant higher frequency of CNAs (e.g., the spontaneous cohort) but also in the cohort with low frequency of CNAs. This indicates that the primary driving force of this classification is recurrence of specific CNAs, and significant predictors are not dependent on the frequency of CNAs.
|
. Specifically, 5 of the top 20 BACs localized to chromosome 5, and in total, these 5 BACs were amplified 28 times in 19 UV-induced tumors (indicating that the regions of amplification involved more than one BAC in some tumors; Fig. 2
) versus only once among the non-UV tumors (Table 2)
. Moreover, three of these five BACs (D5MIT249, D5MIT102, and D5MIT346) were localized to a contiguous region at the proximal end of chromosome 5 (Table 2
and Fig. 2
). Gain of the proximal end of chromosome 5 was apparent even when the log ratio threshold for defining amplification was raised or lowered by 50% (data not shown). It is also worth noting that the same three BACs from proximal chromosome 5 were identified in the list of top 20 generated when BACs were rank ordered based on metrics that account for both amplifications and deletions, including T-score and the related correlation metric defined in Golub et al. (data not shown; Ref. 8
). From these data, we surmised that gain of a gene or genes residing on proximal chromosome 5 represented an important genetic event in the development of UV-induced melanomas in this model system.
|
. Bioinformatic analysis of the mouse genome sequence (Ensembl mouse genome database v4.1.1, Jan. 31, 2002) revealed three annotated genes residing within the minimal region, namely, cdk6, cyp51, and ccm1.
The finding of cdk6 within this locus is remarkable in view of recent reports demonstrating that components of the Rb pathway are critical targets of UV’s melanoma-promoting action (26)
. Indeed, in this p19ARF-deficient model, a frequent acquired lesion in non-UV melanomas is loss of p16Ink4a expression or function. In contrast, p16Ink4a expression and function is retained in a greater proportion of the UV-treated tumors, which instead appear to preferentially amplify the cdk6 locus on chromosome 5 (Fig. 3A)
. Moreover, cdk6 amplification and p16Ink4a loss were mutually exclusive (Fig. 3A
; Ref. 26
). The cdk6 gene amplification was confirmed in each tumor that showed proximal chromosome 5 amplification by array-CGH (Fig. 3B)
. cdk6 amplification was also identified in two tumors in which none of the proximal chromosome 5 array-CGH markers were amplified (tumors 16 and 19; Fig. 3B
). This more focal amplification strongly suggests that cdk6 is the target of the amplification event. Moreover, cdk6 gene amplification is associated with cdk6 overexpression as determined by quantitative RT-PCR, which showed increased cdk6 mRNA levels in UV-treated tumors that exhibited cdk6 amplification by comparison with normal melanocytes and by comparison with non-UV-treated tumors (Fig. 3C)
. In addition, cdk6 is overexpressed in UV-treated tumors in the absence of amplification of proximal chromosome 5 or the cdk6 gene, suggesting that UV treatment can enhance cdk6 expression through multiple mechanisms (tumor 9; Fig. 3, B and C
). In contrast, cdk6 amplification was not detected in tumors arising in p16Ink4a-/- mice spontaneously or as a result of UV exposure (data not shown). In conclusion, the combined use of computational and array-CGH methods in this study has established that genome-wide CNA patterns can be used for tumor classification. Furthermore, prioritization of key features distinguishing between two subclasses of melanoma in this study has revealed cdk6 as an important target associated with UV exposure. This is consistent with recent reports identifying components of the Rb pathway, including Cdk6, as critical targets of UV’s melanoma-promoting activity (26)
.
|
Additionally, there are several distinct advantages to array-CGH over expression profiling. First, the increased stability of DNA over RNA may permit the use of archival or poor quality tumor specimens from which RNA cannot be obtained, as well as samples with high degrees of stromal contamination where laser capture micro-dissection is necessary. In addition, RNA expression profiling is limited in its ability to analyze complex heterogeneous tissues because of the lack of a "normal" control, particularly in cases of tumors where cells of origin are yet to be identified. In contrast, diploid somatic tissue is generally available from cancer patients or normal pool. Finally, array CGH measures information about a tumor not necessarily captured by expression profiling, e.g., cytogenetic complexity, regardless of the specific genetic lesions identified, has been shown to correlate with adverse outcome in both solid and hematological malignancies (5 , 30 , 31) . In summary, through the analysis of array-CGH patterns, we were able to distinguish genetically and histopathologically homogeneous tumors, predict their UV-induced versus spontaneous etiologies, and identify novel pathogenic lesions, such as cdk6 amplification in these tumors.
|
ACKNOWLEDGMENTS |
|---|
|
FOOTNOTES |
|---|
1 Supported in part by grants to L. C. (NIH UO1 CA84313, the Claudia Adams Barr Program in Cancer Research, and the Rockefeller Brothers Fund) and by a grant to W. H. W. (NIH P20 CA96470). L. C. is a Charles E. Culpeper Medical Scholar. 
2 R. C. O. and C. W. B. contributed equally to this work.
3 To whom requests for reprints should be addressed, at Dana-Farber Cancer Institute, Harvard Medical School, 44 Binney Street (M413), Boston, MA 02115. Phone: (617) 632-6091; Fax: (617) 632-6069.
4 The abbreviations used are: CNA, chromosomal numerical alteration; SVM, support vector machine; CGH, comparative genomic hybridization; LOH, loss of heterozygosity; Rb, retinoblastoma; cdk, cyclin-dependent kinase; BAC, bacterial artificial chromosome; Tyr-RAS, tyrosinase-H-RASV12G; RT-PCR, reverse transcription-PCR.
5 L. Chin et al., unpublished observations.
Received 3/11/03. Revised 5/16/03. Accepted 6/25/03.
|
REFERENCES |
|---|
|
TopABSTRACTINTRODUCTIONMATERIALS AND METHODSRESULTS AND DISCUSSIONREFERENCES |
|---|
This article has been cited by other articles:
|
|
 |
|
  F. Rapaport, E. Barillot, and J.-P. Vert Classification of arrayCGH data using fused SVM Bioinformatics, July 1, 2008; 24(13): i375 - i382. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 M. Stark and N. Hayward Genome-Wide Loss of Heterozygosity and Copy Number Analysis in Melanoma Using High-Density Single-Nucleotide Polymorphism Arrays Cancer Res., March 15, 2007; 67(6): 2632 - 2642. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 L. Chin, L. A. Garraway, and D. E. Fisher Malignant melanoma: genetics and therapeutics in the genomic era. Genes & Dev., August 15, 2006; 20(16): 2149 - 2182. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 M. Guillaud-Bataille, A. Valent, P. Soularue, C. Perot, M.-d.-M. Inda, A. Receveur, S. Smaili, H. R. Crollius, J. Benard, A. Bernheim, et al. Detecting single DNA copy number variations in complex genomes using one nanogram of starting DNA and BAC-array CGH Nucleic Acids Res., July 29, 2004; 32(13): e112 - e112. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 P. L. Paris, A. Andaya, J. Fridlyand, A. N. Jain, V. Weinberg, D. Kowbel, J. H. Brebner, J. Simko, J.E. V. Watson, S. Volik, et al. Whole genome scanning identifies genotypes associated with recurrence and metastasis in prostate tumors Hum. Mol. Genet., July 1, 2004; 13(13): 1303 - 1313. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
| Cancer Research | Clinical Cancer Research |
| Cancer Epidemiology Biomarkers & Prevention | Molecular Cancer Therapeutics |
| Molecular Cancer Research | Cancer Prevention Research |
| Cancer Prevention Journals Portal | Cancer Reviews Online |
| Annual Meeting Education Book | Meeting Abstracts Online |
