Originally published as JCO Early Release 10.1200/JCO.2005.05.0229 on September 11 2006

This Article Abstract Full Text (PDF) Purchase Article View Shopping Cart Alert me when this article is cited Alert me if a correction is posted Services Email this article to a colleague Similar articles in this journal Similar articles in PubMed Alert me to new issues of the journal Save to my personal folders Download to citation manager Rights & Permissions Citing Articles Citing Articles via HighWire Citing Articles via Google Scholar Google Scholar Articles by Barrier, A. Articles by Dudoit, S. Search for Related Content PubMed PubMed Citation Articles by Barrier, A. Articles by Dudoit, S. Related Articles Related Correspondence Social Bookmarking                            What's this?

Stage II Colon Cancer Prognosis Prediction by Tumor Gene Expression Profiling

Alain Barrier, Pierre-Yves Boelle, François Roser, Jennifer Gregg, Chantal Tse, Didier Brault, François Lacaine, Sidney Houry, Michel Huguier, Brigitte Franc, Antoine Flahault, Antoinette Lemoine, Sandrine Dudoit

From the Service de Chirurgie digestive, Hôpital Tenon, Assistance Publique—Hôpitaux de Paris; Epidemiologie, Systèmes d’information et Modélisation (U707), INSERM; UMR-S 707, Université Pierre et Marie Curie; Service de Biochimie, Hôpital Tenon, Assistance Publique–Hôpitaux de Paris, Paris; Micro-environnement et physiopathologie de la différenciation (U602), INSERM, Villejuif; Service d’Anatomie Pathologique, Hôpital Ambroise Paré, Assistance Publique Hôpitaux de Paris, Boulogne; Université Versailles Saint Quentin, Boulogne, France; Division of Biostatistics, School of Public Health, University of California Berkeley, Berkeley; J. David Gladstones Institute, University of California–San Francisco, San Francisco, CA

Address reprint requests to Alain Barrier, MD, Service de Chirurgie digestive, Hôpital Tenon, 4 rue de la Chine, 75020 Paris, France; e-mail: alain.barrier{at}tnn.ap-hop-paris.fr or barrier{at}stat.Berkeley.edu

	ABSTRACT

TOP ABSTRACT INTRODUCTION PATIENTS AND METHODS RESULTS DISCUSSION Appendix Authors' Disclosures of... Author Contributions GLOSSARY REFERENCES

 
PURPOSE: This study mainly aimed to identify and assess the performance of a microarray-based prognosis predictor (PP) for stage II colon cancer. A previously suggested 23-gene prognosis signature (PS) was also evaluated.

PATIENTS AND METHODS: Tumor mRNA samples from 50 patients were profiled using oligonucleotide microarrays. PPs were built and assessed by random divisions of patients into training and validation sets (TSs and VSs, respectively). For each TS/VS split, a 30-gene PP, identified on the TS by selecting the 30 most differentially expressed genes and applying diagonal linear discriminant analysis, was used to predict the prognoses of VS patients. Two schemes were considered: single-split validation, based on a single random split of patients into two groups of equal size (group 1 and group 2), and Monte Carlo cross validation (MCCV), whereby patients were repeatedly and randomly divided into TS and VS of various sizes.

RESULTS: The 30-gene PP, identified from group 1 patients, yielded an 80% prognosis prediction accuracy on group 2 patients. MCCV yielded the following average prognosis prediction performance measures: 76.3% accuracy, 85.1% sensitivity, and 67.5% specificity. Improvements in prognosis prediction were observed with increasing TS size. The 30-gene PS were found to be highly-variable across TS/VS splits. Assessed on the same random splits of patients, the previously suggested 23-gene PS yielded a 67.7% mean prognosis prediction accuracy.

CONCLUSION: Microarray gene expression profiling is able to predict the prognosis of stage II colon cancer patients. The present study also illustrates the usefulness of resampling techniques for honest performance assessment of microarray-based PPs.

	INTRODUCTION

TOP ABSTRACT INTRODUCTION PATIENTS AND METHODS RESULTS DISCUSSION Appendix Authors' Disclosures of... Author Contributions GLOSSARY REFERENCES

 
Despite numerous clinical trials, the benefit of adjuvant chemotherapy for stage II colon cancer patients has never been proved in a randomized study. In most meta-analyses, there is a trend towards a benefit, but statistical significance is not reached.¹ Including all stage II colon cancer patients in a randomized trial is debatable. Even if a properly designed study, comprising thousands of patients, demonstrated a significant benefit of adjuvant chemotherapy, it may not be logical to conclude that this treatment should be administered to all patients. Indeed, such a conclusion would not take into account that three fourths of patients are cured by surgery alone and that the approach would lead to administering to all patients a treatment that would be useful for only a few. Another, more rational, approach would be to identify a subgroup of patients at high risk of recurrence, thus more likely to benefit from adjuvant chemotherapy, and to include only these selected patients in a randomized trial. This presupposes finding accurate prognosis predictors (PPs) for stage II colon cancer patients.

As for several types of malignant tumors (breast carcinomas,^2,3 lung carcinomas,^4,5 lymphomas^6,7), microarray gene expression profiling has been reported to accurately predict the prognosis of stage II colon cancer.⁸ In their report, Wang et al⁸ identified, from a set of 38 patients, a 23-gene prognosis signature (PS) that was validated on an independent set of 36 patients and yielded a 78% prognosis prediction accuracy.

Fifty stage II colon cancer patients, with the same postoperative treatment (no adjuvant chemotherapy) but with different outcomes (25 patients developed a metachronous metastasis, whereas the other 25 remained disease free for at least 5 years), were included in the present study. Tumor samples were profiled using the Affymetrix (Santa Clara, CA) HGU133A GeneChip, with the following aims: (1) to identify a microarray-based PP and assess its performance in terms of accuracy, sensitivity, and specificity; and (2) to assess the prognosis prediction performance of the 23-gene PS proposed by Wang et al.⁸

	PATIENTS AND METHODS

TOP ABSTRACT INTRODUCTION PATIENTS AND METHODS RESULTS DISCUSSION Appendix Authors' Disclosures of... Author Contributions GLOSSARY REFERENCES

 
Patients and Samples
Fifty patients (27 male, 23 female; mean age, 71 years) operated on for a stage II colon adenocarcinoma between 1996 and 2000 were included in this study. The main patient and tumor characteristics are given in Table 1. None of the patients had emergency surgery or received any adjuvant chemotherapy. Twenty-five patients developed a distant metastasis (liver in 22 patients, lung in five patients) in the follow-up, and 21 within 3 years of surgery. The mean time to recurrence was 27 months (range, 14 to 52 months). The other 25 patients remained disease free for at least 60 months, with mean follow-up of 79 months (range, 60 to 101 months).

Gene expression measures were computed using the Robust Multichip Average method implemented in the Bioconductor R package rma (http://www.bioconductor.org) and described in Irizarry et al.¹⁰ Gene expression measures are available at http://www.u707.jussieu.fr/boelle/genechips/index.html and http://www.stat.berkeley.edu/sandrine.

Data Analysis
Prognosis prediction method. For a given split of patients into a training set (TS) and a validation set (VS), a 30-gene PP was built on the TS and its performance assessed on the VS as follows.

Step 1. Gene expression measures were compared in recurrent and nonrecurrent TS patients by computing two-sample equal-variance t statistics for each of the 22,283 genes. A PS was defined in terms of the expression measures of the 30 genes with the largest absolute t statistics.

Step 2. A PP was constructed by applying diagonal linear discriminant analysis (DLDA) to the 30-gene PS of the TS patients.^11,12

Step 3. The 30-gene PP from Step 2 was applied to predict the prognoses of the VS patients.

Step 4. The predicted and actual prognoses (recurrence or no recurrence) of VS patients were compared to obtain the following three measures of prognosis prediction performance: accuracy (proportion of correctly predicted prognoses), sensitivity (proportion of correctly predicted recurrences), and specificity (proportion of correctly predicted nonrecurrences).

Validation procedure: Single-split validation. Two schemes were considered for dividing patients into TS and VS: single-split validation and Monte Carlo cross validation.

Patients were randomly divided into two groups of equal size, group 1 and group 2. Group 1 and group 2 were used as TS and VS, respectively. A 30-gene PP was built on group 1 patients and its performance assessed on group 2 patients.

Validation procedure: Monte Carlo cross validation. For Monte Carlo cross validation (MCCV), 16 different values for the TS size n₀ were considered: n₀ = 10,12,...,40. For each choice of n₀, the 50 patients were repeatedly and randomly divided into 100 TS of size n₀ and corresponding VS of size 50-n₀. For each TS/VS split, a 30-gene PP was identified on TS patients and applied to VS patients as described herein. This yielded, for each value of the TS size n₀, 100 30-gene PSs and 100 measures of prognosis prediction performance. The gene compositions of the 100 PSs were compared. Graphical and numerical summaries (eg, minimum, maximum, and mean) of the distributions of prognosis prediction accuracies, sensitivities, and specificities for the 16 x 100 = 1,600 TS/VS splits were obtained.

Performance Assessment of the 23-Gene PS
The prognosis prediction performance of the 23-gene PS of Wang et al⁸ was assessed based on the same 16 x 100 random TS/VS splits of patients as for the 30-gene PS. For a given TS/VS split, a PP was obtained by applying DLDA to the 23-gene PS of the TS patients. This 23-gene PP was then applied to predict the prognoses of the VS patients. Predicted and actual prognoses (recurrence or no recurrence) of VS patients were compared.

Proposal of a 30-Gene PS
An overall 30-gene PS was identified based on all 50 patients, by comparing the expression measures of recurrent and nonrecurrent patients for each of the 22,283 genes using two-sample equal-variance t statistics and selecting the 30 genes with the largest absolute t statistics.

	RESULTS

TOP ABSTRACT INTRODUCTION PATIENTS AND METHODS RESULTS DISCUSSION Appendix Authors' Disclosures of... Author Contributions GLOSSARY REFERENCES

 
Single-Split Validation
A 30-gene PS and corresponding PP were identified on the 25 group 1 patients. Applied to the 25 group 2 patients, this 30-gene PP yielded 80% accuracy, 75% sensitivity, and 85% specificity.

MCCV
For each of the 16 values of the TS size n₀, the 100 random splits of patients into a TS and a VS each yielded a 30-gene PP and corresponding measures of prediction performance on the VS (accuracy, sensitivity, specificity). Numerical summaries of the distributions of prognosis prediction performance measures for the 16 x 100 TS/VS splits were 76.3% mean accuracy (range, 52.5% to 100.0%), 85.1% mean sensitivity, and 67.5% mean specificity. Prognosis prediction performance improved with TS size (Figs 1A and 1B). For TS of size 40, mean accuracy, sensitivity, and specificity were 82.7%, 92.0%, and 73.4%, respectively. Sensitivity was higher than specificity for all TS sizes.

View larger version (15K): [in this window] [in a new window]   Fig 1. Monte Carlo cross validation. Prognosis prediction performance of 30-gene prognosis signatures. (A) Mean, minimum, and maximum prognosis prediction accuracies as a function of the training set (TS) size that were observed for the 100 random splits of patients; (B) mean accuracy, sensitivity, and specificity as a function of the TS size that were observed for the 100 random splits of patients.

View larger version (14K): [in this window] [in a new window]   Fig 2. Monte Carlo cross validation. 30-gene prognosis signature composition. (A) The number of genes that were included in at least one of the 100 signatures as a function of the training set (TS) size; (B) the number of genes that were included in at least 10, 25, 50, 75, and 100 of the 100 signatures as a function of the TS size.

View larger version (25K): [in this window] [in a new window]   Fig 3. Monte Carlo cross validation. Prognosis prediction (PP) accuracy of the 23-gene prognosis signature.8 (A) Mean accuracy of the 23-gene prognosis signature (PS)8 (blue line) and 30-gene prognosis signature (red line) as a function of the training set (TS) size; (B) relative performance of the 23- and 30-gene PS for each of the 100 random TS/validation set (VS) splits of patients as a function of the TS size.

	DISCUSSION

TOP ABSTRACT INTRODUCTION PATIENTS AND METHODS RESULTS DISCUSSION Appendix Authors' Disclosures of... Author Contributions GLOSSARY REFERENCES

The results from MCCV clearly suggest the possibility to use gene expression profiling to predict the prognosis of stage II colon cancer patients. For the 16 x 100 30-gene PPs, the mean prognosis prediction accuracy was 76.3%; moreover, none of these 1,600 PPs yielded an accuracy lower than 50%. Mean sensitivity was higher than mean specificity (85.1% v 67.5%); this finding is of interest because the practical problem for stage II colon cancer patients, which underlies the present study, is the identification of the minority of these patients at high risk of metastatic recurrence, thus more likely to benefit from adjuvant chemotherapy. Performance consistently increased with TS size to reach a maximum of 82.7% accuracy, 92.0% sensitivity, and 73.4% specificity for TS of size 40. This suggests that, as expected, additional gains in performance could be obtained with predictors built on larger numbers of patients.

MCCV also revealed great variability in PS composition and PP performance between random splits of patients. This variability, which has been previously reported,^14-16 outlines the weakness of studies based on a unique split of patients.

For a given TS and VS size, the range of observed accuracies was wide: 20% for the largest VS size, 40% for the smallest VS size. This suggests that the results of studies based on single-split validation should be interpreted with caution, because there is a risk to obtain overoptimistic performance estimates. In their report, Michiels et al used multiple random splits of patients from seven previously published studies,^{2,4,5,7,17-19} and concluded that five of these studies did not classify patients better than did chance.¹⁴

PS composition was highly variable, especially for TS of small sizes; with TS of size 10, more than 2,200 different genes were included in the 100 30-gene signatures, meaning that the vast majority of these genes were selected only once. Variability of PS composition was also observed for larger TS, but it concerned only a subset of genes; with TS of size 40, 280 different genes were included in the 100 30-gene signatures, but 12 of these genes were constantly, or almost constantly, selected.

The findings from MCCV strongly suggest that a unique PP does not exist and that many PSs lead to PPs with similar performances. This conclusion is consistent with the well-known fact that, especially for high-dimensional prediction problems, many models yield the same fit.

In the present study, the following two main choices were made for building prognosis predictors: (1) the number of genes to include in the prognosis signature was set to 30, on the basis of previous results⁹; and (2) prognosis predictors were constructed using DLDA, since DLDA was shown to be competitive with more complex techniques.^11,12 Both of these choices were somewhat arbitrary, and many other gene selection methods and classifiers could have been used. To determine the influence of our choices on results, we have reproduced exactly the same MCCV analysis as above with 30-gene t statistic–based prognosis signatures and nearest neighbor classifiers¹¹ (Fig A2, online only), and with DLDA based on prognosis signatures including various numbers of genes (from 10 to 200; Fig A3, online only). The results of these supplementary analyses suggest a moderate influence of the length of the PS and the choice of classifier on PP performance.

The second aim of the present study was to assess the performance of the PS proposed by Wang et al.⁸ These authors built from a TS of 36 patients a PP based on the expression measures of 23 genes, and applied this PP to a VS of 38 patients, with a 78% prognosis prediction accuracy. Interestingly, this 23-gene PS led to fairly accurate predictors for the prognosis of our patients (overall mean accuracy of 67.1% and a mean accuracy > 70% for TS of > 30). To our knowledge, this is the first time that a PS proposed by one research team is successfully validated by another research team. Since we used the same 1,600 random splits of patients, we were able to directly compare the performance of the 23-gene PP and the 30-gene PP. The mean prognosis prediction accuracy was 76.3% for the 30-gene PP, and 67.1% for the 23-gene PP; for 1,190 (74.4%) of the 1,600 splits, the accuracy of the 30-gene PP was higher than that of the 23-gene PP. We hypothesized that the observed differences in accuracy between 30-gene and 23-gene PP were mainly caused by the different criteria used to classify patients into the disease-free group (disease status after 5 years in our study v 3 years for Wang et al⁸). This hypothesis was confirmed by results of an additional study in which we considered the 3-year status of our patients (Fig A4, online only).

MCCV allowed honest performance assessment of a prognosis prediction procedure, but did not lead to the identification of a unique prognosis signature and corresponding prognosis predictor. Instead, MCCV suggested that many combinations of genes could lead to PP with similar performances. Despite these findings, it seemed of interest to propose a single prognosis predictor that could be used by others. Since we applied on the whole set of 50 patients the same gene selection method than in MCCV, performance estimates of the proposed 30-gene PP are provided by results of MCCV. From a statistical point of view, all 30 genes do not have the same value. Two groups might be distinguished: a "stable" group of 12 genes, and a "variable" group of 18 genes. Seven genes had a permutation-based step-down maxT-adjusted P value of .0001; they were selected on average 70% of the times by MCCV, and constantly with large TS. Five additional genes had an adjusted P value lower than .002; they were selected on average 50% of the times by MCCV, and almost constantly for large TS. It would be of interest to assess the performance of a "reduced" prognosis predictor containing these 12 "stable" genes. From a biologic point of view, the presence of 10 genes encoding ribosomal proteins in our proposed 30-gene prognosis signature is of particular interest. All 10 genes were overexpressed in patients who remained disease free. More remarkably, five of these 10 genes were among the seven genes with the lowest adjusted P values (.0001) and were the five genes selected most often by MCCV. The best-known function shared by ribosomal proteins is their role in the assembly of ribosomal subunits, and, as a result, their role in translation. Individual ribosomal proteins have been implicated in a wide variety of biologic functions, including cell cycle progression, apoptosis, and DNA damage responses.^20-23 It has also been suggested that their role in these processes may arise independently of their role in the ribosome itself. Our data raise the possibility that some ribosomal protein genes could play a role in tumor invasion, the latter being favored by their decreased transcription.

In conclusion, the present study suggests the possibility of using functional genomic approaches to predict the prognosis of stage II colon cancer patients, thereby identifying a subgroup of patients at high risk of metastatic recurrence and thus more likely to benefit from adjuvant chemotherapy. At this point, it seems premature to claim that the decision to give patients a postoperative treatment should be based on their gene expression profiles. More rationally, we propose the use of gene expression profiling to select stage II patients to include in future studies aiming to assess the potential benefits of adjuvant chemotherapy. The present study also suggests the usefulness of resampling techniques for honest performance assessment of microarray-based prognosis predictors.

	Appendix

TOP ABSTRACT INTRODUCTION PATIENTS AND METHODS RESULTS DISCUSSION Appendix Authors' Disclosures of... Author Contributions GLOSSARY REFERENCES

 
Table A1 and Figure A1 present the number of selections (of 100 training set [TS]/validation set [VS] splits by Monte Carlo cross validation [MCCV]) for each of the 30 genes of the prognosis signature (PS) that were identified on the whole set of 50 patients, for 16 different TS sizes n₀ = 10, 12, ..., 40.

The same 1,600 random TS/VS splits of patients were used as in the main MCCV analysis (100 random splits for each of 16 possible values for the TS size n₀). As before, 30-gene PS were identified based on the 30 genes with the largest absolute t statistics. However, prognosis predictors were built using nearest neighbor classification,^11,22 rather than diagonal linear discriminant analysis (DLDA).

Table A2 gives, for each possible value of n₀, the minimum, the mean, and the maximum prognosis prediction accuracies obtained with DLDA and nearest neighbor classification.

Results obtained with both classifiers were quite similar.

Figure A2 shows, as a function of n₀, the minimum, the mean, and the maximum prognosis prediction accuracies that were observed with the 100 splits of patients.

The same 1,600 random TS/VS splits of patients were used as in the main MCCV analysis (100 random splits for each of 16 possible values for the TS size n₀). DLDA was used for prognosis prediction. Prognosis signatures were based on the m genes with the largest absolute t statistics (m = 10, 20, 30, 40, 50, 100, 150, and 200).

Figure A3 shows, as a function of the TS size n₀, the mean accuracy that was observed with the 100 m-gene signatures. For all values of m, the mean accuracy of the m-gene prognosis predictors increased with the TS size. The overall mean accuracy, over the 1,600 PP, slightly decreased as m increased (77.4% for m = 10; 76.3% for m = 30; 72.7% for m = 200). For small TS sizes, the mean accuracy over the 100 PP slightly increased with m (with TS = 10, 64.1% for m = 10; 65.5% for m = 30; 67.4% for m = 200). For large TS sizes, the mean accuracy over the 100 PP decreased as m increased (86.0% for m = 10; 82.7% for m = 30; 75.8% for m = 200).

A different binary end point was considered: the status of disease 3 years after surgery (v 5 years after surgery as in the main study). Of the 50 patients, 29 patients remained disease free, while the other 21 developed a recurrence (four patients recurred during the fourth and fifth postoperative years).

MCCV was performed exactly as in the main analysis. The patients were randomly split into 1,600 TSs and VSs. Sixteen TS sizes were considered (n₀ = 10, 12, ..., 40). For each value of n₀, 100 random TS/VS splits were performed.

Figure A4A shows, as a function of the TS size n₀, the mean 3-year prognosis prediction accuracies that were observed with the 30-gene PS and the 23-gene PS. These mean accuracies were very close and increased with TS sizes: from 65% and 66% for the 30-gene PS and the 23-gene PS, respectively, with TS of size 10, to 71% and 74% for the 30-gene PS and the 23-gene PS, respectively, with TS of size 40. Results obtained with the 23-gene PS and 3-year disease status were quite similar to those of the main study for the 23-gene PS and 5-year disease status: the overall mean prognosis prediction accuracy was 68.6% (v 67.1%); the mean prognosis prediction accuracy varied between 64.0% for the smallest TS size (v 62.0%) and 70.7% for the largest TS size (v 71.3%). In contrast, 3-year prognosis prediction accuracies for the 30-gene PS were inferior to the main study’s 5-year prognosis prediction accuracies for the 30-gene PS and were very close to the 3-year prognosis prediction accuracies for the 23-gene prognosis signatures: the overall mean prognosis prediction accuracy was 70.7% (v 76.3%); the mean prognosis prediction accuracy varied between 66.0% for the smallest TS size (v 65.5%) and 74.5% for the largest TS size (v 82.7%).

Figure A4B shows, as a function of the TS size n₀, the relative 3-year prognosis prediction performance of the 23-gene PS and the 30-gene PS for each of the 100 random TS/VS splits of patients. For 752 (47.0%) of 1,600 splits, the 30-gene PS performed better than the 23-gene PS; for 496 splits (31.0%), the 23-gene PS performed better than the 30-gene PS; for 352 splits (22.0%), the 30-gene and 23-gene PS lead to the same prognosis prediction accuracy.

View larger version (37K): [in this window] [in a new window]   Fig A1. Monte Carlo cross validation. Number of selections for each of the genes of the 30-genes prognosis signature as a function of the training set size.

View larger version (14K): [in this window] [in a new window]   Fig A2. Mean, minimum, and maximum prognosis prediction accuracies as a function of the training set size that were observed for the 100 random splits of patients. Nearest neighbor classification was used instead of diagonal linear discriminant analysis, using the same random split of patients as in Fig 1.

View larger version (18K): [in this window] [in a new window]   Fig A3. Prognosis prediction accuracy as a function of the number of genes in the prognosis signature and TS size. The analysis used diagonal linear discriminant analysis and the same random split of patients as in Fig 1.

View larger version (25K): [in this window] [in a new window]   Fig A4. Prognosis prediction accuracy using survival at 3 years as the end point. The 30-gene prognosis signature (PS) and 23-gene PS8 were compared for (A) accuracy and (B) relative performance.

	Authors’ Disclosures of Potential Conflicts of Interest

TOP ABSTRACT INTRODUCTION PATIENTS AND METHODS RESULTS DISCUSSION Appendix Authors' Disclosures of... Author Contributions GLOSSARY REFERENCES

	Author Contributions

TOP ABSTRACT INTRODUCTION PATIENTS AND METHODS RESULTS DISCUSSION Appendix Authors' Disclosures of... Author Contributions GLOSSARY REFERENCES

 

Conception and design: Alain Barrier, Pierre-Yves Boelle, Antoine Flahault, Antoinette Lemoine, Sandrine Dudoit Provision of study materials or patients: Alain Barrier, François Roser, Jennifer Gregg, Chantal Tse, Didier Brault, François Lacaine, Sidney Houry, Michel Huguier, Brigitte Franc Collection and assembly of data: Alain Barrier, François Roser, Jennifer Gregg, Chantal Tse, Didier Brault, Sandrine Dudoit Data analysis and interpretation: Alain Barrier, Pierre-Yves Boelle, Antoine Flahault Manuscript writing: Alain Barrier, Antoine Flahault, Sandrine Dudoit Final approval of manuscript: Alain Barrier, Pierre-Yves Boelle, François Roser, Jennifer Gregg, Chantal Tse, Didier Brault, François Lacaine, Sidney Houry, Michel Huguier, Brigitte Franc, Antoine Flahault, Antoinette Lemoine, Sandrine Dudoit

 

	GLOSSARY

TOP ABSTRACT INTRODUCTION PATIENTS AND METHODS RESULTS DISCUSSION Appendix Authors' Disclosures of... Author Contributions GLOSSARY REFERENCES

 

Diagonal linear discriminant analysis:

A mathematical form of classifier that combines the component features by a weighted linear average. With gene expression based classifiers, the components are generally the logarithm of expression level of the selected genes. The weights are based on the degree of differential expression of the individual genes among the classes.

Monte Carlo cross validation:

A method used for assessing variability in the performance of a classifier, by repeating split sample validation with random allocation to training and validation sets.

Training set:

Samples used in a developmental study to define a classifier. The classifier can be internally validated in the test set of samples; those that were not used to develop the classifier.

Validation set:

Samples used in evaluating performance of a classifier. The validation set is formed by the units not used in developing the classifier (ie, the training set and test set).

	NOTES

 
published online ahead of print at www.jco.org on September 11, 2006.

Supported by the J. David Gladstone Institutes and General Clinical Research Center at San Francisco General Hospital, and by a grant from the California Institute for Quantitative Biomedical Research, University of California Berkeley (A.B.).

Terms in blue are defined in the glossary, found at the end of this article and online at www.jco.org.

Authors’ disclosures of potential conflicts of interest and author contributions are found at the end of this article.

	REFERENCES

TOP ABSTRACT INTRODUCTION PATIENTS AND METHODS RESULTS DISCUSSION Appendix Authors' Disclosures of... Author Contributions GLOSSARY REFERENCES

 
1. Figueredo A, Charette ML, Maroun J, et al: Adjuvant therapy for stage II colon cancer. A systematic review from the Cancer Care Ontario Program in evidence-based care’s gastrointestinal cancer disease site group. J Clin Oncol 22:3395-3407, 2004[Abstract/Free Full Text]

2. Van’t Veer LJ, Dai H, Van De Vijver MJ, et al: Gene expression profiling predicts clinical outcome of breast cancer. Nature 415:530-536, 2002[CrossRef][Medline]

3. Van de Vijver MJ, Yudong DH, Van’t Veer LJ, et al: A gene-expression signature as a predictor of survival in breast cancer. N Engl J Med 347:1999-2009, 2002[Abstract/Free Full Text]

4. Beer DG, Kardia SLR, Huang C, et al: Gene expression profiles predict survival of patients with lung adenocarcinoma. Nat Med 8:816-824, 2002[Medline]

5. Bhattacharjee A, Richards WG, Staunton J, et al: Classification of human lung carcinoma by mRNA expression profiling reveals distinct adenocarcinoma subclasses. Proc Natl Acad Sci U S A 98:13790-13795, 2001[Abstract/Free Full Text]

6. Shipp MA, Ross KN, Tamayo P, et al: Diffuse large B-cell lymphoma outcome prediction by gene-expression profiling and supervised machine learning. Nat Med 8:68-74, 2002[CrossRef][Medline]

7. Rosenwald A, Wright G, Chan WC, et al: The use of molecular profiling to predict survival after chemotherapy for diffuse large B-cell lymphoma. N Engl J Med 346:1937-1947, 2002[Abstract/Free Full Text]

8. Wang Y, Jatkoe T, Zhang Y, et al: Gene expression profiles and molecular markers to predict recurrence of Dukes’ B colon cancer. J Clin Oncol 22:1564-1571, 2004[Abstract/Free Full Text]

9. Barrier A, Lemoine A, Boelle PY, et al: Colon cancer prognosis prediction by gene expression profiling. Oncogene 24:6155-6164, 2005[CrossRef][Medline]

10. Bolstad BM, Irizarry RA, Gautier L, et al: Preprocessing high-density oligonucleotide arrays, in Gentleman R, Carey VJ, Huber W, et al (eds): Bioinformatics and Computational Biology Solutions Using R and Bioconductor. New York, NY, Springer, 2005, pp 13-32

11. Dudoit S, Fridlyand J, Speed TP: Comparison of discrimination methods for the classification of tumors using gene expression data. J Am Statist Assoc 97:77-87, 2002[CrossRef]

12. Dudoit S, Fridlyand J: Classification in microarray experiments, in Speed T (ed): Statistical Analysis of Gene Expression Microarray Data. Boca Raton, FL, Chapman & Hall/CRC, 2003, pp 93-158

13. Pollard KS, Dudoit S, van der Laan MJ: Multiple testing procedures: The multtest package and applications to genomics, in Gentleman R, Carey VJ, Huber W, et al (eds): Bioinformatics and Computational Biology Solutions Using R and Bioconductor. New York, NY, Springer, 2005, pp 249-271

14. Michiels S, Koscielny S, Hill C: Prediction of cancer outcome with microarrays: A multiple random validation strategy. Lancet 365:488-492, 2005[CrossRef][Medline]

15. Barrier A, Boelle PY, Lemoine A, et al: Gene expression profiling of nonneoplastic mucosa may predict clinical outcome of colon cancer patients. Dis Colon Rectum 48:2338-2248, 2005

16. Dudoit S, van der Laan MJ: Asymptotics of cross-validation risk estimation in estimator selection and performance assessment. Statist Methodol 2:131-154, 2005

17. Yeoh EJ, Ross ME, Shurtleff SA, et al: Classification, subtype discovery, and prediction of outcome in pediatric acute lymphoblastic leukemia by gene expression profiling. Cancer Cell 1:133-143, 2002[CrossRef][Medline]

18. Pomeroy SL, Tamayo P, Gaasenbeek M, et al: Prediction of central nervous system embryonal tumour outcome based on gene expression. Nature 415:436-442, 2002[CrossRef][Medline]

19. Iizuka N, Oka M, Yamada-Okabe H, et al: Oligonucleotide microarray for prediction of early intrahepatic recurrence of hepatocellular carcinoma after curative resection. Lancet 361:923-929, 2003[CrossRef][Medline]

20. Sonenberg N: Translation factors as effectors of cell growth and tumorigenesis. Curr Opin Cell Biol 5:955-960, 1993[CrossRef][Medline]

21. Chen FW, Ioannou YA: Ribosomal proteins in cell proliferation and apoptosis. Int Rev Immunol 18:429-448, 1999[Medline]

22. Lohrum MA, Ludwig RL, Kubbutat MH, et al: Regulation of HDM2 activity by the ribosomal protein L11. Cancer Cell 3:577-587, 2003[CrossRef][Medline]

23. Volarevic S, Stewart MJ, Ledermann B, et al: Proliferation, but not growth, blocked by conditional deletion of 40S ribosomal protein S6. Science 288:2045-2047, 2000[Abstract/Free Full Text]

Submitted November 23, 2005; accepted June 7, 2006.

CiteULike    Complore    Connotea    Del.icio.us    Digg    Facebook    Reddit    Technorati    Twitter    What's this?

Related Correspondence

• Molecular Biology for Stage II Colorectal Cancer: The Jury Is Still Out
  Stefano Cascinu, Alberto Zaniboni, Mario Scartozzi, and Fausto Meriggi
  JCO 2007 25: 2861 [Full Text]
• Molecular Biology for Stage II Colorectal Cancer: The Jury Is Still Out
  Stefano Cascinu, Alberto Zaniboni, Mario Scartozzi, and Fausto Meriggi
  JCO 2007 25: 2861 [Full Text]

This article has been cited by other articles:

				H.-J. Lenz Established Biomarkers for Colon Cancer ASCO Educational Book, January 1, 2009; 2009(1): 215 - 219. [Abstract] [Full Text] [PDF]

				K. S. Garman, C. R. Acharya, E. Edelman, M. Grade, J. Gaedcke, S. Sud, W. Barry, A. M. Diehl, D. Provenzale, G. S. Ginsburg, et al. A genomic approach to colon cancer risk stratification yields biologic insights into therapeutic opportunities PNAS, December 9, 2008; 105(49): 19432 - 19437. [Abstract] [Full Text] [PDF]

				W. L. Allen, V. M. Coyle, P. V. Jithesh, I. Proutski, L. Stevenson, C. Fenning, D. B. Longley, R. H. Wilson, M. Gordon, H.-J. Lenz, et al. Clinical Determinants of Response to Irinotecan-Based Therapy Derived from Cell Line Models Clin. Cancer Res., October 15, 2008; 14(20): 6647 - 6655. [Abstract] [Full Text] [PDF]

				T. Schepeler, J. T. Reinert, M. S. Ostenfeld, L. L. Christensen, A. N. Silahtaroglu, L. Dyrskjot, C. Wiuf, F. J. Sorensen, M. Kruhoffer, S. Laurberg, et al. Diagnostic and Prognostic MicroRNAs in Stage II Colon Cancer Cancer Res., August 1, 2008; 68(15): 6416 - 6424. [Abstract] [Full Text] [PDF]

				T. Nakamura, T. Kuwai, Y. Kitadai, T. Sasaki, D. Fan, K. R. Coombes, S.-J. Kim, and I. J. Fidler Zonal Heterogeneity for Gene Expression in Human Pancreatic Carcinoma Cancer Res., August 15, 2007; 67(16): 7597 - 7604. [Abstract] [Full Text] [PDF]

				S. Cascinu, A. Zaniboni, M. Scartozzi, and F. Meriggi Molecular Biology for Stage II Colorectal Cancer: The Jury Is Still Out J. Clin. Oncol., July 1, 2007; 25(19): 2861 - 2861. [Full Text] [PDF]

				A. Barrier, P.-Y. Boelle, S. Houry, S. Dudoit, and A. Lemoine In Reply J. Clin. Oncol., July 1, 2007; 25(19): 2862 - 2863. [Full Text] [PDF]

				M. Del Rio, F. Molina, C. Bascoul-Mollevi, V. Copois, F. Bibeau, P. Chalbos, C. Bareil, A. Kramar, N. Salvetat, C. Fraslon, et al. Gene Expression Signature in Advanced Colorectal Cancer Patients Select Drugs and Response for the Use of Leucovorin, Fluorouracil, and Irinotecan J. Clin. Oncol., March 1, 2007; 25(7): 773 - 780. [Abstract] [Full Text] [PDF]

This Article Abstract Full Text (PDF) Purchase Article View Shopping Cart Alert me when this article is cited Alert me if a correction is posted Services Email this article to a colleague Similar articles in this journal Similar articles in PubMed Alert me to new issues of the journal Save to my personal folders Download to citation manager Rights & Permissions Citing Articles Citing Articles via HighWire Citing Articles via Google Scholar Google Scholar Articles by Barrier, A. Articles by Dudoit, S. Search for Related Content PubMed PubMed Citation Articles by Barrier, A. Articles by Dudoit, S. Related Articles Related Correspondence Social Bookmarking                            What's this?