Discovery and validation of a prostate cancer genomic classifier that predicts early metastasis following radical prostatectomy

Abstract

Purpose: Clinicopathologic features and biochemical recurrence are sensitive, but not specific, predictors of metastatic disease and lethal prostate cancer. We hypothesize that a genomic expression signature detected in the primary tumor represents true biological potential of aggressive disease and provides improved prediction of early prostate cancer metastasis.

Methods: A nested case-control design was used to select 639 patients from the Mayo Clinic tumor registry who underwent radical prostatectomy between 1987 and 2001. A genomic classifier (GC) was developed by modeling differential RNA expression using 1.4 million feature high-density expression arrays of men enriched for rising PSA after prostatectomy, including 213 who experienced early clinical metastasis after biochemical recurrence. A training set was used to develop a random forest classifier of 22 markers to predict for cases--men with early clinical metastasis after rising PSA. Performance of GC was compared to prognostic factors such as Gleason score and previous gene expression signatures in a withheld validation set.

Results: Expression profiles were generated from 545 unique patient samples, with median follow-up of 16.9 years. GC achieved an area under the receiver operating characteristic curve of 0.75 (0.67-0.83) in validation, outperforming clinical variables and gene signatures. GC was the only significant prognostic factor in multivariable analyses. Within Gleason score groups, cases with high GC scores experienced earlier death from prostate cancer and reduced overall survival. The markers in the classifier were found to be associated with a number of key biological processes in prostate cancer metastatic disease progression.

Conclusion: A genomic classifier was developed and validated in a large patient cohort enriched with prostate cancer metastasis patients and a rising PSA that went on to experience metastatic disease. This early metastasis prediction model based on genomic expression in the primary tumor may be useful for identification of aggressive prostate cancer.

Conflict of interest statement

Competing Interests: The authors have read the journal's policy and have the following conflicts: NE, AC, IV, MG, CB, ZH, BZ, TS, TT, and ED are employees of GenomeDx Biosciences Inc. ED and TT own stock in GenomeDx Biosciences Inc. ED has received research funding from GenomeDx Biosciences Inc. and the National Research Council - Industrial Research Assistance Program. PB has received research funding from GenomeDx Biosciences Inc. GK has received research funding from Beckman Coulter. KB, EB, RC, SF, RK, RJ, and TK have declared that no competing interests exist. This does not alter the authors adherence to all the PLOS ONE policies on sharing data and materials.

Figures

Figure 1. Consort diagram.

Study breakdown into cases and controls. Training and validation sets are shown.

Figure 2. Multidimensional scaling plot of (A) the training and (B) the validation sets.

Controls are indicated in blue and cases in red. In both the training and validation sets the controls tend to cluster on the left of the plot and the cases on the right of the plot. In this manner, most of the biological differences are expressed in the first dimension of the scaling. Random forest proximity [http://www.stat.berkeley.edu/~breiman/] was used to measure the 22 marker distance between samples.

Figure 3. Performance of classifiers and individual clinicopathologic variables.

For each predictor, the AUC obtained in the training and validation sets, as well as the 95% Confidence Interval for this metric is shown. CC: clinical-only classifier. GC: genomic classifier. GCC: combined genomic-clinical classifier.

Figure 4. Score distributions of multivariable classifiers in cases and controls in validation set.

Distributions of scores are plotted for A) CC B) GC and C) GCC for controls and cases. Median scores and 95% confidence intervals are represented by a horizontal black line and notches, respectively. Non-overlapping notches indicate that differences in the distribution of scores between cases and controls are statistically significant. Outliers are represented as points beyond the boxplot whiskers.

Figure 5. Distribution of GC scores among pathologic GS categories in validation.

GC scores are plotted with a jitter so as to more easily differentiate the patients among each pathologic GS (x-axis) groups. Case (red) and controls patients (blue) are shown for each category. The dashed black line indicates the GC cutoff of 0.5. Trends show the patients with high GC scores tend to have high GS as well.

Figure 6. Kaplan Meier estimates for all Cases with (A) PCSM and (B) OS endpoints.

Cases were separated into high (>0.5) or low risk according to GC score. Log-rank p-values are shown in the upper right corner. Time to PCSM and OS is measured from BCR in years.

Figure 7. Performance of external signatures in training and validation sets.

For each signature, the institution associated to it, year of publication, lead author, the AUC obtained in the training and validation sets, as well as the 95% Confidence Interval for this metric is shown.