The RNA extraction and cDNA preparation in these studies was described in their original publications. The ovarian cancer dataset from Bild et al (5) were assayed with Affymetrix U133A (retrieved with record GSE3149 from Gene Expression Omnibus). Two colon cancer datasets were all generated with Affymetrix U133A arrays (7,6). The lung adenocarcinoma datasets from Shedden et al (8) were generated with Affymetrix U133A.

The ovarian cancer cohort (n = 124) from Bild et al. (5) was used to explore whether the 28-gene signature reveals molecular portraits common in breast cancer and ovarian cancer. To avoid over-fitting in the validation, the data set was randomly partitioned into a training set (n = 82) and a test set (n = 42). The 28 gene predictors were fitted in a Cox hazard proportional model on the training set, and a survival risk score was generated for each patient. A high risk score represents a high probability of post-operative treatment failure, and similarly for a low risk score. The median of the survival risk scores in the training set was used as the cutoff point to stratify patients into different prognostic groups. A patient with a risk score higher than median risk score was classified into poor-prognosis group; whereas a patient with a lower risk score was classified into good-prognosis group. The same cutoff value and prognostic model were applied to patient stratification in the test set.

The matching genes in the 28-gene signature were identified with Affymetrix IDs. Twenty-five common genes were found in each colon cancer cohort. If a gene has multiple probes, the average expression of multiple probes was used in the classification. The patient cohort from Barrier et al (6) was used as training set (n = 50), while the cohort from another study by Barrier et al (7) was used as an independent validation set (n = 24). A training model was built with the 25 signature genes to classify recurrence in colon cancer patients using a Linear Discriminant Analysis function in SAS 9.1. A 10-fold cross validation was used to evaluate the performance of the training model. This training model was used to predict tumor recurrence in each patient in the validation set.

The patient samples collected from the University of Michigan Cancer Center (UM) and Moffitt Cancer Center (HLM) form the training set (n = 256), whereas the samples obtained from Memorial Sloan-Kettering Cancer Center (MSK, n = 104) and the Dana-Farber Cancer Institute (DFCI, n = 82) constitute an independent validation set. Gene symbols were used to find the matching genes in the signature. In the training set (UM and HLM cohorts), a Cox proportional hazard model was constructed by using the matching genes as covariates to predict lung cancer survival after the initial treatment. A risk score was generated for each patient in this cohort. Based on the distribution of the risk scores in the training set, a cutoff point representing the peak value in the histogram was identified to stratify patients into high- or low-risk groups. This cutoff risk score and the training model were applied in prognostic categorization in the validation set (MSK and DFCI cohorts).

Patient survival rates were assessed with Kaplan-Meier analysis using log-rank tests. Cox hazard proportional model was used to generate a risk score for each ovarian cancer patient based on the 28-gene signature. All statistical analyses were performed with software package R (9).

Ovarian cancer is a common malignancy in women, whose prognosis is bleak due to a usually advanced disease stage at the time of diagnosis. Common genetic risk factors of susceptibility to breast and ovarian cancer have recently been proposed (2). To explore whether the 28-gene signature reveals common molecular features affecting breast and ovarian cancer survival, an ovarian cancer cohort from Bild et al (5) was analyzed. This ovarian cancer cohort (n = 124) was randomly split into a training set (n = 82) and a test set (n = 42). A Cox model was built on the training set using the signature genes as covariates. A survival risk score was generated for each patient. The median of the risk scores in the training set was identified as a cutoff point for patient stratification. Patients with a risk score greater than the cutoff were stratified into the high-risk group, and otherwise, into the low-risk group. In the prognostic model evaluation, the high- and low-risk groups had significantly (log-rank P < 0.0001) different relapse-free survival in the training cohort in Kaplan-Meier analysis (Fig. 1A). This training model and stratification scheme were applied to the test set, and generated significant prognostic stratification (log-rank P = 0.0075) in Kaplan-Meier analysis (Fig. 1B). The details of the prognostic Cox model were provided in http://www.hsc.wvu.edu/mbrcc/fs/GuoLab/pdfs/Shruti_Rathnagiriswaran_Thesis.pdf. These results indicate that the 28-gene signature reflects common biological processes involved in breast and ovarian cancer metastases and relapse. The 28-gene signature could indentify more aggressive ovarian cancers that were more likely to develop recurrence after surgical resections and initial treatment. Therefore, the high risk patients defined with this gene signature might benefit from second line chemotherapy.

In order to extent the potential usefulness of the 28-gene prognostic signature, we explored its value for predicting clinical outcome in patients with stage II colon cancer. To construct a molecular classifier to predict colon cancer recurrence, 50 patients with stage II colon adenocarcinoma (6) were used as training cohort. Twenty-five genes within the 28-gene signature were identified from the DNA microarray data. These signature genes were used to classify recurrence in each patient with the Linear Discriminant Analysis algorithm. The performance of the classifier was evaluated in a 10-fold cross validation on the training set (Table 1). The prognostic signature correctly predicted recurrence in 94% (47/50) of patients, with a sensitivity of 100% (25/25) and a specificity of 88% (22/25). The model identified in the training cohort was applied to predict recurrence in each patient in the validation set (n = 24) with a patient cohort retrieved from Barrier et al (7). In the validation, the prognostic signature correctly predicted recurrence in 75% (18/24) patients, with a sensitivity of 80% (8/10) and a specificity of 71.43% (10/14). Both cohorts contained only stage II lymph node negative colon adenocarcinomas. These results indicate that the 28-gene prognostic signature provides independent prognostic information in addition to tumor stage. Once validated in larger, independent cohorts this signature could be potentially used to select lymph node-negative patients for receiving adjuvant chemotherapy.

To explore the clinical relevance of the 28-gene prognostic signature for the prognostication of patients with non-small cell lung cancer, the lung adenocarcinoma cohorts (UM and HLM) retrieved from Shedden et al (8) were used as a training set (n = 256). A Cox model of overall survival was constructed based on the 28-gene signature, with each gene variable as a covariate. A survival risk score was generated for every patient, with a higher risk score representing a greater probability of treatment failure (i.e., death). Based on the histogram representing distribution of gene expression-defined risk scores in this cohort (Fig. 2A), a cutoff value of −0.75, the peak value in the histogram, was used to stratify patients into high- and low-risk groups. This cutoff value represents the linear additive expression levels of all the signature genes in lung cancer patients. This stratification separated patients into two groups with distinct overall survival (log-rank P < 3.91e-5) in Kaplan-Meier analysis (Fig. 2B). This cutoff risk score and training model were applied to the validation set (MSK and DFCI, n = 186). The 28-gene signature generated borderline significant prognostic categorization in the validation set (log-rank P = 0.07; Fig. 2C) in Kaplan-Meier analysis. In all studied lung adenocarcinoma cohorts, the low-risk groups had 73.54% to 82.15% of 2.5-year postoperative survival rate, representing a significantly better prognosis compared with the corresponding high-risk groups for which the 2.5-year survival was ranging from 53.76% to 63.51%. As the majority of non-small cell lung cancer recurrence occurs within 2 years after surgery (13), these results indicate that the 28-gene prognostic signature could be used to predict postoperative survival in non-small cell lung cancer patients.