Gene expression profiling predicts clinical outcome of prostate cancer (original) (raw)
Identification of molecular signatures distinguishing subgroups of prostate cancer patients with distinct clinical outcome after therapy. To identify the outcome-predictor signatures, we used as a training data set the expression analysis of 12,625 transcripts in 21 prostate tumor samples obtained from patients with prostate cancer who had distinct clinical outcome after therapy. Using biochemical evidence of relapse after therapy as a criterion of treatment failure, 21 patients were divided into two subgroups representing prostate cancer with recurrent (8 patients) and nonrecurrent (13 patients) clinical behavior (14). This analysis identified a set of 218 genes (91 upregulated and 127 downregulated transcripts) differentially regulated in tumors from patients with recurrent versus nonrecurrent prostate cancer at the statistically significant level (P < 0.05) defined by both t tests and Mann-Whitney tests (Supplemental Table 3S).
To reduce the number of hits in potential outcome predictor clusters and to identify transcripts of potential biological relevance, we compared the expression profile of 218 genes to the expression profiles of transcripts differentially regulated in multiple experimental models of human prostate cancer (19) in the search for genes with consistently concordant expression patterns across multiple data sets and various experimental conditions (see signature discovery and validation protocol presented in the supplemental data for details). We identified several small gene clusters exhibiting highly concordant pattern of expression (Pearson correlation coefficient: r > 0.95) in clinical and experimental samples. We evaluated the prognostic power of each identified cluster of coregulated transcripts based on ability to segregate the patients with recurrent and nonrecurrent prostate tumors into distinct subgroups and selected a single best-performing cluster for each binary condition specified in Supplemental Table 4S (Figure 1; Tables 1 and 2). To assess a potential prognostic relevance of individual gene clusters, we calculated a Pearson correlation coefficient for each of 21 tumor samples by comparing the expression profiles of individual samples to the “average” expression profile of recurrent versus nonrecurrent tumors and expression profiles of relevant experimental samples (Table 1 and Figure 1). Based on expected correlation of expression profiles of identified gene clusters with recurrent clinical behavior of prostate cancer, we named the corresponding correlation coefficients calculated for individual samples the phenotype association indices (PAIs).
PAIs defined by the expression profile of the prostate cancer recurrence predictor signature 1 for 21 prostate carcinoma samples constituting a signature discovery (training) data set. Prostate tumor samples were taken from the patients at the time of surgery and subjected to a microarray gene expression analysis as described in Methods. Note that all samples derived from tumors of patients who subsequently manifested a biochemical relapse of disease have positive PAI values, whereas 12 of 13 samples obtained from patients who remained disease-free have negative PAI values. See text for details.
Gene expression signatures associated with recurrent prostate cancer
Prostate cancer recurrence prediction accuracy in a good-prognosis and a poor-prognosis subgroup of patients defined according to whether they had a good-prognosis or a poor-prognosis signature
Using this strategy, we identified several gene clusters (Tables 1 and 2) discriminating with 86–95% accuracy human prostate tumors exhibiting recurrent or nonrecurrent clinical behavior (Figure 1; Tables 1 and 2). Figure 1 illustrates application of the five-gene cluster to characterize clinical prostate cancer samples according to their propensity for recurrence after therapy. The expression pattern of the genes in the recurrence predictor cluster was analyzed in each of 21 separate clinical samples. The analysis produces a quantitative PAI (plotted on the y axis) for each of the 21 clinical prostate cancer samples. Tumors that are likely to recur are expected to have positive PAIs reflecting positive correlation of gene expression with metastasis-promoting orthotopic xenografts, while those that are unlikely to recur are expected to have negative association indices.
Figure 1 shows the PAIs for eight samples from patients who later had recurrence as bars 1–8, while the association indices for 13 samples from patients whose tumors did not recur is shown as bars 11–23. Eight of the eight samples (or 100%) from patients who later experienced recurrence had positive PAIs and so were properly classified. Twelve of the 13 samples (or 92.3%) from patients whose tumors did not recur had negative PAIs and so were properly classified as nonrecurrent tumors. Thus, overall, 20 of the 21 samples (or 95.2%) were properly classified using a five-gene recurrence predictor signature. Two alternative clusters identified using this strategy showed similar sample classification performance (Tables 1 and 2).
To further evaluate the prognostic power of identified gene expression signatures, we performed the Kaplan-Meier survival analysis using as a clinical end-point DFI after therapy in prostate cancer patients with positive and negative PAIs. The Kaplan-Meier survival curves showed a highly significant difference in the probability that prostate cancer patients would remain disease-free after therapy between the groups with positive and negative PAIs defined by the signatures (Figure 2, A–C), suggesting that patients with positive PAIs exhibit a poor-outcome signature whereas patients with negative PAIs manifest a good-outcome signature. The estimated hazard ratio for disease recurrence after therapy in the group of patients with positive PAIs as compared with the group of patients with negative PAIs defined by the recurrence predictor signature 3 (Table 1) was 9.046 (95% confidence interval of ratio, 3.022–76.41; P = 0.001). Eighty-six percent of patients with the positive PAIs had a disease recurrence within 5 years after therapy, whereas 85% of patients with the negative PAIs remained relapse-free for at least 5 years (Figure 2C). Based on this analysis, we propose to identify the group of prostate cancer patients with positive PAIs as a poor-prognosis group and the group of prostate cancer patients with negative PAIs as a good-prognosis group.
Kaplan-Meier analysis of the probability that patients would remain disease-free among 21 prostate cancer patients constituting a signature discovery group according to whether they had good-prognosis or poor-prognosis signatures defined by the recurrence predictor signature 1 (A), recurrence predictor signature 2 (B), recurrence predictor signature 3 (C), and the recurrence predictor algorithm, which takes into account calls from all three signatures (D). The cut-off values for each marker were identified through the detailed analysis of behavior of log-rank test P values across the range of the measurements for each marker. We selected the prognosis discrimination cut-off value for each signature based on highest level of statistical significance in patients’ stratification into poor- and good-prognosis groups as determined by the log-rank test (lowest P value and highest hazard ratio; see Supplemental Table 6S). CI, confidence interval.
Theoretically, the recurrence predictor algorithm based on a combination of signatures should be more robust than a single predictor signature, particularly during the validation analysis using an independent test cohort of patients. Next, we analyzed whether a combination of the three signatures would perform in a patient’s classification test with accuracy similar to that of the individual signatures. We found that the cut-off value of PAIs greater than 0.2 scored in two of three individual clusters allowed achievement of the 90% recurrence prediction accuracy (Table 2). This recurrence predictor algorithm correctly identified 88% of patients with recurrent disease and 92% of patients with nonrecurrent disease (Table 2). The Kaplan-Meier survival analysis (Figure 2D) showed that the median relapse-free survival after therapy of patients in the poor-prognosis group was 26 months. All patients in the poor-prognosis group had a disease recurrence within 5 years after therapy, whereas 92% of patients in the good-prognosis group remained relapse-free for at least 5 years. The estimated hazard ratio for disease recurrence after therapy in the poor-prognosis group of patients as compared with the good-prognosis group of patients defined by the recurrence predictor algorithm was 20.32 (95% confidence interval of ratio, 6.047–158.1; P < 0.0001).
Validation of the outcome predictor signatures using independent clinical data set. To validate the potential clinical utility of identified molecular signatures, we evaluated the prognostic power of signatures applied to an independent set of 79 clinical samples obtained from 37 patients with prostate cancer who developed recurrence after the therapy and 42 patients who remained disease-free (Supplemental Table 1S). The Kaplan-Meier survival analysis demonstrated that all three recurrence predictor signatures segregate prostate cancer patients into subgroups with a statistically significant difference in the probability of remaining relapse-free after the therapy (Table 3). Interestingly, application of the recurrence predictor algorithm (the cut-off value of PAIs greater than 0.2 scored in two of three individual clusters) appears to perform better than individual signatures in a patient’s stratification test using an independent data set (Table 3).
Stratification of 79 prostate cancer patients based on recurrence predictor signatures into poor- and good-prognosis groups at the time of diagnosis
The Kaplan-Meier survival analysis (Figure 3A) showed that the median relapse-free survival after therapy of patients in the poor-prognosis group defined by the recurrence predictor algorithm was 34.6 months. Sixty-seven percent of patients in the poor-prognosis group had a disease recurrence within 5 years after therapy, whereas 76% of patients in the good-prognosis group remained relapse-free for at least 5 years. The estimated hazard ratio for disease recurrence after therapy in the poor-prognosis group as compared with the good-prognosis group of patients defined by the recurrence predictor algorithm was 4.224 (95% confidence interval of ratio, 2.455–9.781; P < 0.0001). Overall, the application of the recurrence predictor algorithm allowed accurate stratification into the poor-prognosis group 82% of patients who failed the therapy within 1 year after prostatectomy. The recurrence predictor algorithm seems to demonstrate more accurate performance in patient’s classification compared with the conventional markers of outcome such as preoperative PSA level or RP Gleason sum (Figures 3 and 4; Table 4).
Kaplan-Meier analysis of the probability that patients would remain disease-free among 79 prostate cancer patients constituting a signature validation group for all patients (A), patients with high (B) or low (C) preoperative PSA levels in blood according to whether they had good-prognosis or poor-prognosis signatures defined by the recurrence predictor algorithm, or whether they had high or low preoperative PSA level in the blood (D). Preoperative PSA level of 7.8 ng/ml was used as a cut-off discrimination level for patients’ stratification into poor- and good-prognosis subgroups. The cut-off values for each marker were identified through the detailed analysis of behavior of log-rank test P values across the range of the measurements for each marker. We selected the prognosis discrimination cut-off value for each marker based on highest level of statistical significance in patients’ stratification into poor- and good-prognosis groups as determined by the log-rank test (lowest P value and highest hazard ratio).
Kaplan-Meier analysis of the probability that patients would remain disease-free among prostate cancer patients with Gleason sum 6 and 7 tumors (A) and patients with Gleason sum 8 and 9 tumors (B) according to whether they had good-prognosis or poor-prognosis signatures defined by the recurrence predictor algorithm or whether they had Gleason sum 8 and 9 or Gleason sum 6 and 7 prostate tumors (C).
Prostate cancer recurrence prediction accuracy in poor-prognosis and good-prognosis subgroups of patients defined by a gene expression–based recurrence predictor algorithm alone or in combination with established biochemical and histopathological markers of outcome
Recurrence predictor signatures provide additional predictive value over conventional markers of outcome. Next, we thought to determine whether the application of the recurrence predictor signatures would provide an additional predictive value when combined with conventional markers of outcome such as preoperative PSA level and Gleason score. Both preoperative PSA level and RP Gleason sum were significant predictors of prostate cancer recurrence after therapy in the validation cohort of 79 patients (Figure 3D and Figure 4C).
The Kaplan-Meier survival analysis (Figure 3D) showed that the median relapse-free survival after therapy of patients in the poor-prognosis group defined by the high preoperative PSA level was 49.0 months. Sixty percent of patients in the poor-prognosis group had a disease recurrence within 5 years after therapy, whereas 73% of patients in the good-prognosis group remained relapse-free for at least 5 years. The estimated hazard ratio for disease recurrence after therapy in the poor-prognosis group as compared with the good-prognosis group of patients defined by the preoperative PSA level was 2.551 (95% confidence interval of ratio, 1.344–4.895; P = 0.0043). Prediction of the outcome after therapy based on preoperative PSA level, however, accurately stratified into poor-prognosis group only 65% of patients who failed the therapy within 1 year after prostatectomy (Table 4).
Next, we set out to determine whether the application of the recurrence predictor algorithm would identify subgroups of patients with distinct clinical outcome after therapy in both high and low PSA-expressing groups, thus adding additional predictive value to the therapy outcome classification based on preoperative PSA level alone.
In the group of patients with high preoperative PSA level (Figure 3B), the median relapse-free survival after therapy of patients in the poor-prognosis subgroup defined by the recurrence predictor algorithm was 36.2 months. Seventy-three percent of patients in the poor-prognosis subgroup had a disease recurrence within 5 years after therapy. Conversely, 73% of patients in the good-prognosis subgroup remained relapse-free for at least 5 years. The estimated hazard ratio for disease recurrence after therapy in the poor-prognosis subgroup as compared with the good-prognosis subgroup of patients defined by the recurrence predictor algorithm was 4.315 (95% confidence interval of ratio, 1.338–7.025; P = 0.0081).
In the group of patients with low preoperative PSA level (Figure 3C), the median relapse-free survival after therapy of patients in the poor-prognosis subgroup defined by the recurrence predictor algorithm was 42.0 months. Fifty-three percent of patients in the poor-prognosis subgroup had a disease recurrence within 5 years after therapy, whereas 92% of patients in the good-prognosis subgroup remained relapse-free for at least 5 years. The estimated hazard ratio for disease recurrence after therapy in the poor-prognosis subgroup as compared with the good-prognosis subgroup of patients defined by the recurrence predictor algorithm was 6.247 (95% confidence interval of ratio, 2.134–24.48; P = 0.0015). Overall, combination of the recurrence predictor algorithm and preoperative PSA level allowed accurate stratification into poor-prognosis group 88% of patients who failed the therapy within 1 year after prostatectomy (Table 4).
RP Gleason sum is a significant predictor of relapse-free survival in the validation cohort of 79 prostate cancer patients (Figure 4C). The Kaplan-Meier survival analysis (Figure 4C) demonstrates that the median relapse-free survival after therapy of patients with the RP Gleason sum 8 and 9 was 21.0 months, thus defining the poor-prognosis group based on histopathological criteria. Seventy-four percent of patients in the poor-prognosis group had a disease recurrence within 5 years after therapy, whereas 69% of patients in the good-prognosis group (RP Gleason sum 6 and 7) remained relapse-free for at least 5 years. The estimated hazard ratio for disease recurrence after therapy in the poor-prognosis group as compared with the good-prognosis group of patients defined by the RP Gleason sum criteria was 3.335 (95% confidence interval of ratio, 2.389–13.70; P < 0.0001). RP Gleason sum-based outcome classification accurately stratified into the poor-prognosis group only 47% of patients who failed the therapy within 1 year after prostatectomy (Table 4).
In the group of patients with RP Gleason sum 6 and 7 (Figure 4A), the median relapse-free survival after therapy of patients in the poor-prognosis subgroup defined by the recurrence predictor algorithm was 61.0 months. Fifty-three percent of patients in the poor-prognosis subgroup had a disease recurrence within 5 years after therapy, whereas 77% of patients in the good-prognosis subgroup remained relapse-free for at least 5 years. The estimated hazard ratio for disease recurrence after therapy in the poor-prognosis subgroup as compared with the good-prognosis subgroup of patients defined by the recurrence predictor algorithm was 3.024 (95% confidence interval of ratio, 1.457–8.671; P = 0.0055).
In the group of patients with RP Gleason sum 8 and 9 (Figure 4B), the median relapse-free survival after therapy in the poor-prognosis subgroup defined by the recurrence predictor algorithm was 11.5 months. One hundred percent of patients in the poor-prognosis subgroup had a disease recurrence within 5 years after therapy, whereas 67% of patients in the good-prognosis subgroup remained relapse-free for at least 5 years. The estimated hazard ratio for disease recurrence after therapy in the poor-prognosis subgroup as compared with the good-prognosis subgroup of patients defined by the recurrence predictor algorithm was 6.143 (95% confidence interval of ratio, 1.573–13.49; P = 0.0053). Overall, a patient’s classification using a combination of the recurrence predictor algorithm and RP Gleason sum allowed an accurate stratification into the poor-prognosis group 82% of patients who failed the therapy within 1 year after prostatectomy (Table 4). Based on this analysis, we concluded that application of the recurrence predictor algorithm appears to provide an additional predictive value to the therapy outcome classification based on established markers of prostate cancer outcome.
In the Cox multivariate proportional hazard analysis, we included in the model four variables that were identified as statistically significant prostate cancer recurrence predictors in a cohort of 79 patients using a univariate survival analysis: the preoperative PSA levels, the age of patients at the time of diagnosis, the RP Gleason sum, and the gene expression–based recurrence predictor algorithm. The multivariate analysis demonstrated that both the recurrence predictor algorithm (P = 0.0001) and RP Gleason sum (P = 0.0306) remained statistically significant prognostic markers, whereas preoperative PSA levels (P = 0.0593) and age (P = 0.0956) failed to achieve the P = 0.05 level of significance in patients’ stratification (Supplemental Table 7S). Interestingly, the recurrence predictor algorithm [HR = 4.0124 (1.9985–8.0556); P = 0.0001] (HR, hazard ratio) appears to perform better in selecting poor-prognosis patients than the RP Gleason sum [HR = 1.5367 (1.040–2.2684); P = 0.0306].
Recurrence predictor signatures provide additional predictive value over outcome prediction based on a multiparameter nomogram. Classification nomograms are generally recognized as the most efficient clinically useful models currently available for prediction of the probability of relapse-free survival after therapy of individual prostate cancer patients (4–7). We applied the Kattan nomogram using multiple postoperative parameters (7) for prognosis prediction classification in the test group of 79 prostate cancer patients.
The Kaplan-Meier survival analysis (Figure 5A) showed that the median relapse-free survival after therapy of patients in the poor-prognosis group defined by the Kattan nomogram was 33.1 months. Seventy-two percent of patients in the poor-prognosis group had a disease recurrence within 5 years after therapy, whereas 81% of patients in the good-prognosis group remained relapse-free for at least 5 years. The estimated hazard ratio for disease recurrence after therapy in the poor-prognosis group as compared with the good-prognosis group of patients defined by the Kattan nomogram was 3.757 (95% confidence interval of ratio, 2.318–9.647; P < 0.0001). Prediction of the outcome after therapy based on Kattan nomogram accurately stratified into the poor-prognosis group 71% of patients who failed the therapy within 1 year after prostatectomy (Table 4).
Kaplan-Meier analysis of the probability that patients would remain disease-free among 79 prostate cancer patients constituting a signature validation group for all patients (A), patients with poor prognosis (B), or good prognosis (C), defined by the Kattan nomogram according to whether they had a good-prognosis or poor-prognosis signatures defined by the recurrence predictor algorithm (B and C) or whether they had poor or good prognosis defined by the Kattan nomogram (A).
Next, we thought to determine whether the application of the recurrence predictor algorithm would identify subgroups of patients with distinct clinical outcome after therapy in both poor- and good-prognosis groups defined by the Kattan nomogram, thus adding additional predictive value to the therapy outcome classification based on nomogram alone.
In the poor-prognosis group of patients defined by the Kattan nomogram the application of the recurrence predictor algorithm appears to identify two subgroups of patients with statistically significant difference in the probability of remaining relapse-free after therapy (Figure 5B). Median relapse-free survival after therapy of patients in the poor-prognosis subgroup defined by the recurrence predictor algorithm was 11.5 months compared with median relapse-free survival of 71.1 months in the good-prognosis subgroup (Figure 5B). Eighty-nine percent of patients in the poor-prognosis subgroup had a disease recurrence within 5 years after therapy. Conversely, 50% of patients in the good-prognosis subgroup remained relapse-free for at least 5 years. The estimated hazard ratio for disease recurrence after therapy in the poor-prognosis subgroup as compared with the good-prognosis subgroup of patients defined by the recurrence predictor algorithm was 3.129 (95% confidence interval of ratio, 1.378–7.434; P = 0.0068).
Similarly, in the good-prognosis group of patients identified based on application of the Kattan nomogram, the recurrence predictor algorithm seems to define two subgroups of patients with a statistically significant difference in the probability of remaining relapse-free after therapy (Figure 5C). Median relapse-free survival after therapy of patients in the poor-prognosis subgroup defined by the recurrence predictor algorithm was 64.8 months. Forty-one percent of patients in the poor-prognosis subgroup had a disease recurrence within 5 years after therapy. Conversely, 87% of patients in the good-prognosis subgroup remained relapse-free for at least 5 years. The estimated hazard ratio for disease recurrence after therapy in the poor-prognosis subgroup as compared with the good-prognosis subgroup of patients defined by the recurrence predictor algorithm was 4.398 (95% confidence interval of ratio, 1.767–18.00; P = 0.0035). Overall, combination of the recurrence predictor algorithm and Kattan nomogram allowed accurate stratification into the poor-prognosis group 82% of patients who failed the therapy within 1 year after prostatectomy (Table 4).
Recurrence predictor algorithm defines poor- and good-prognosis subgroups of patients diagnosed with the early stage prostate cancer. Identification of subgroups of patients with distinct clinical outcome after therapy would be particularly desirable in a cohort of patients diagnosed with early-stage prostate cancer. Next, we investigated whether recurrence predictor signatures would be useful in defining subgroups of patients diagnosed with early-stage prostate cancer and having a statistically significant difference in the likelihood of disease relapse after therapy.
In the group of patients diagnosed with the stage 1C prostate cancer (Figure 6A), the median relapse-free survival time after therapy in the poor-prognosis subgroup defined by the recurrence predictor algorithm was 12 months. In contrast, the median relapse-free survival time after therapy in the good-prognosis group was 82.4 months. Seventy-seven percent of patients in the poor-prognosis subgroup had a disease recurrence within 5 years after therapy. Conversely, 81% of patients in the good-prognosis subgroup remained relapse-free for at least 5 years. The estimated hazard ratio for disease recurrence after therapy in the poor-prognosis subgroup as compared with the good-prognosis subgroup of patients defined by the recurrence predictor algorithm was 5.559 (95% confidence interval of ratio, 2.685–25.18; P = 0.0002).
Kaplan-Meier analysis of the probability that patients would remain disease-free among prostate cancer patients with stage 1C tumors (A) and patients with stage 2A tumors (B) according to whether they had a good-prognosis or poor-prognosis signatures defined by the recurrence predictor algorithm.
In the group of patients diagnosed with the stage 2A prostate cancer (Figure 6B), the median relapse-free survival after therapy in the poor-prognosis subgroup defined by the recurrence predictor algorithm was 35.4 months. Eighty-six percent of patients in the poor-prognosis subgroup had a disease recurrence within 5 years after therapy, whereas 78% of patients in the good-prognosis subgroup remained relapse-free for at least 5 years. The estimated hazard ratio for disease recurrence after therapy in the poor-prognosis subgroup as compared with the good-prognosis subgroup of patients defined by the recurrence predictor algorithm was 7.411 (95% confidence interval of ratio, 2.220–40.20; P = 0.0024). Based on this analysis we concluded that application of the recurrence predictor algorithm seems to provide potentially useful clinical information in stratification of patients diagnosed with the early-stage prostate cancer into subgroups, with statistically significant difference in the likelihood of disease recurrence after therapy.