Abstract
Purpose: Stratifying patients who have a high risk of prostate cancer recurrence following prostatectomy can potentiate the use of adjuvant therapy at an early stage. Inflammation has emerged as a mediator of prostate cancer metastatic progression. We hypothesized that chemokines can be biomarkers for distinguishing patients with high risk for biochemical recurrence of prostate cancer.
Experimental Design: In a nested case-control study, 82 subjects developed biochemical recurrence within 5 years of prostatectomy. Prostate tissues from 98 age-matched subjects who were recurrence-free following prostatectomy in the same period were the controls. A high-throughput lectin-based enrichment of prostate tissue enabled multiplex ELISA to identify the expression of three chemokines to discriminate the two patient populations.
Results: The expression of CX3CL1 and IL-15 in prostate tissue was associated with 5-year biochemical recurrence-free survival following prostatectomy. However, the expression of chemokine ligand 4 (CCL4) was associated with biochemical recurrence. Multivariable logistic regression model combining preoperative prostate-specific antigen, Gleason score, surgical margin, and seminal vesicle status with the three chemokines doubled the specificity of prediction at 90% sensitivity compared with use of the clinicopathologic variables alone (P < 0.0001). Survival analysis yielded a nomogram that supported the use of CX3CL1, IL-15, and CCL4 in predicting 1-, 3-, and 5-year recurrence-free survival after prostatectomy.
Conclusions: Each of the three chemokines can serve as independent predictors of biochemical recurrence. However, the combination of chemokine biomarkers plus clinicopathologic variables discriminated prostatectomy subjects for the probability of biochemical recurrence significantly better than clinicopathologic variables alone.
Current therapeutic options for prostate cancer patients involve surgical or radiation-based ablation of the prostate. Although this can be curative for many patients, up to 35% of prostatectomy patients develop biochemical recurrence of prostate cancer. The high-risk prostate cancer patient population often dies from the disease due to metastatic progression. Early detection of prostate cancer likely to develop biochemical recurrence can lead to proactive use of adjuvant therapeutic options before frank biochemical recurrence. As tumor-associated inflammatory changes can regulate prostate cancer metastatic progression, we looked to chemokines in prostatectomy tissue as potential indicators of future biochemical recurrence. We identified three chemokines that independently predict biochemical recurrence status within 5 years of prostatectomy in a nested case-control study. The differentially expressed chemokines further supplemented known clinicopathologic variables to provide high sensitivity and specificity. The multivariable prediction models generated resulted in a nomogram that provides superior prediction capacity.
There is a large disparity between the number of newly diagnosed cases of prostate cancer every year and the number of men who die of metastatic progression of the disease (1). As a consequence, although prostate cancer is the second leading cause of cancer-related mortality in men in the United States, overtreatment of the disease is a concern. The challenge has been to determine which patients harbor high-risk disease requiring aggressive/curative therapy and which patients harbor indolent disease that could be managed with active surveillance. Traditionally, a combinatorial assessment of various clinical parameters is used to risk-stratify patients. Although these assessments have their utility and have undergone many different revisions and modifications over time, they are still unable to distinguish the 80% of patients that may not have any clinical consequences from their prostate cancer (2). Nonetheless, monotherapy such as surgical or radiation ablation of the prostate may not be curative for many patients. In such patients with high-risk disease, adjuvant therapy (e.g., antiandrogens, chemotherapy) is common following monotherapy. However, because adjuvant therapies often provide only temporary suspension of disease progression, the poor efficacy may be due to their administration after the development of frank biochemical prostate cancer recurrence (3–5). Because a majority of patients remain disease-free after prostatectomy (75-80%), the identification of candidates potentially benefiting from adjuvant therapy is important given the potential for decreased quality of life and added morbidity from such intervention (6). Nomograms, such as the one described by Kattan et al., are commonly used to predict disease recurrence after prostatectomy and involve multiple criteria that include the pretreatment prostate-specific antigen (PSA), prostate capsule invasion, pathologic Gleason score, surgical margin status, seminal vesicle involvement, and lymph node involvement (7). However, the sensitivity of predicting biochemical recurrence by such criteria can be improved (7).
Inflammatory chemokines in the tumor microenvironment can regulate the fate of tumor progression (8, 9). We hypothesized that such tissue chemokines can be strong biomarker candidates for distinguishing patients with a high risk for biochemical recurrence or metastatic progression of prostate cancer. Inflammatory cell recruitment in prostate cancer has emerged as a modulator of metastatic progression (10, 11). We sought to identify factors that mediate inflammatory cell recruitment because the immune response to cancer can result in tumor cell ablation and provide growth factors to further stimulate tumor progression and motility. The nested case-control study described identify the differential expression of two chemokine biomarkers in prostatic tissue that support greater detection sensitivity for the biochemical recurrence of prostate cancer alone following prostatectomy. However, incorporating clinicopathologic parameters plus three chemokine biomarkers provided superior prediction of biochemically recurrent disease than clinicopathologic parameters alone. The nomogram developed based on the modeling results illustrate the power of the combined survival prediction. The model has the potential to improve patient risk stratification after primary local treatment for prostate cancer and enable earlier decision-making for possible secondary therapeutic options.
Patients, Materials, and Methods
Patient selection
This study was conducted in accordance with the Vanderbilt University Institutional Review Board. The digital medical record of 660 subjects was retrospectively examined using the Vanderbilt University Urologic Surgery registry of radical prostatectomies done between 1998 and 2002. Several of these patients were excluded for reasons that included availability of at least 5-year follow-up data, availability of archived fresh frozen peripheral zone tissue, and records of presurgical hormone ablation therapy. Patients who had undergone hormone ablation therapy at any point before surgery or the demonstration of biochemical recurrence were excluded. Biochemical recurrence following prostatectomy was defined as PSA ≥0.2 ng/mL confirmed at least once with another PSA at least 2 wk apart, and associated with two consecutive subsequent increases in PSA level. Ultimately, for this nested study, we focused on 82 subjects who developed biochemical recurrence within 5 y of prostatectomy and an age-matched control group of 98 subjects who were free of recurrence within the same time frame. The mean age for the subjects was 60 y (43-72 y). All subjects were annotated based on age, race, presurgical serum PSA, pathologic Gleason score, pathologic stage, extracapsular involvement, seminal vesicle involvement, surgical margin status, and detection of biochemical recurrence (Table 1).
Clinical and pathologic stratification of the biochemical recurrent and recurrence-free subject groups (n = 180)
. | Overall (n = 180) . | Recurrence-free (n = 98) . | Recurrent (n = 82) . | P . | ||||
---|---|---|---|---|---|---|---|---|
Age | 60.0 (56.0, 66.0) | 60.0 (55.3, 66.0) | 60.5 (56.0, 66.0) | 0.834* | ||||
Race | ||||||||
White | 92% (165) | 88% (86) | 96% (79) | 0.093† | ||||
Black | 7% (13) | 10% (10) | 4% (3) | |||||
Others | 1% (2) | 2% (2) | 0% (0) | |||||
Extracapsular involvement | 42% (76) | 29% (28) | 59% (48) | <0.001† | ||||
Pos. margin | 28% (50) | 11% (11) | 48% (39) | <0.001† | ||||
Seminal vesicle involvement | 16% (28) | 2% (2) | 32% (26) | <0.001† | ||||
Lymph node involvement | 5% (9) | 0% (0) | 11% (9) | <0.001† | ||||
Preoperative PSA | 6.3 (4.8, 9.1) | 5.7 (4.6, 7.4) | 7.3 (5.1, 13.4) | <0.001† | ||||
Gleason | ||||||||
5-6 | 24% (44) | 34% (33) | 13% (11) | <0.001† | ||||
7 | 58% (105) | 59% (58) | 57% (47) | |||||
8-9 | 18% (31) | 7% (7) | 30% (24) | |||||
Clinical stage | ||||||||
T1c | 63% (114) | 68% (67) | 57% (47) | 0.092 | ||||
T2a | 25% (45) | 24% (24) | 26% (21) | |||||
T2b+ | 12% (21) | 8% (7) | 17% (14) |
. | Overall (n = 180) . | Recurrence-free (n = 98) . | Recurrent (n = 82) . | P . | ||||
---|---|---|---|---|---|---|---|---|
Age | 60.0 (56.0, 66.0) | 60.0 (55.3, 66.0) | 60.5 (56.0, 66.0) | 0.834* | ||||
Race | ||||||||
White | 92% (165) | 88% (86) | 96% (79) | 0.093† | ||||
Black | 7% (13) | 10% (10) | 4% (3) | |||||
Others | 1% (2) | 2% (2) | 0% (0) | |||||
Extracapsular involvement | 42% (76) | 29% (28) | 59% (48) | <0.001† | ||||
Pos. margin | 28% (50) | 11% (11) | 48% (39) | <0.001† | ||||
Seminal vesicle involvement | 16% (28) | 2% (2) | 32% (26) | <0.001† | ||||
Lymph node involvement | 5% (9) | 0% (0) | 11% (9) | <0.001† | ||||
Preoperative PSA | 6.3 (4.8, 9.1) | 5.7 (4.6, 7.4) | 7.3 (5.1, 13.4) | <0.001† | ||||
Gleason | ||||||||
5-6 | 24% (44) | 34% (33) | 13% (11) | <0.001† | ||||
7 | 58% (105) | 59% (58) | 57% (47) | |||||
8-9 | 18% (31) | 7% (7) | 30% (24) | |||||
Clinical stage | ||||||||
T1c | 63% (114) | 68% (67) | 57% (47) | 0.092 | ||||
T2a | 25% (45) | 24% (24) | 26% (21) | |||||
T2b+ | 12% (21) | 8% (7) | 17% (14) |
NOTE: For continuous variables, a (b, c) represent the median a, lower quartile b, and the upper quartile c. Numbers after percentages are frequencies in parentheses.
Wilcoxon test.
Fisher's test.
Sample preparation and analysis
The tissue samples were derived from a tumor bank of frozen cores of the prostate from patients after radical prostatectomy for adenocarcinoma of the prostate. Eight 4-mm-diameter cores were taken from fresh prostates removed during prostatectomy and snap frozen in liquid nitrogen. Four of these cores were from the peripheral zone, where the majority of the prostate cancer originates. The frozen cores, determined by gross assessment of tumor involvement, were dissected longitudinally into three sections with the outside sections used for duplicate protein isolation (Fig. 1A). The central piece was paraformaldehyde fixed and paraffin embedded for histologic evaluation by a pathologist.
The histologic evaluation of the frozen prostate cores. A, each tissue core was cut longitudinally into thirds for duplicate chemokine enrichment and histology analysis, respectively. Subjects B and C developed biochemical recurrence, but subject D was free of recurrence for the 5 y following prostatectomy.
The histologic evaluation of the frozen prostate cores. A, each tissue core was cut longitudinally into thirds for duplicate chemokine enrichment and histology analysis, respectively. Subjects B and C developed biochemical recurrence, but subject D was free of recurrence for the 5 y following prostatectomy.
Samples were homogenized in lysis buffer [100 mmol/L Tris (pH 7.2), 500 mmol/L NaCl], sonicated, and centrifuged for 5 min at 1,000 × g. Glycosylated proteins were purified using the glycoprotein isolation kit, WGA (Pierce Biotechnology, Inc.), according to the manufacturer's instructions. All procedures except for binding to the wheat germ agglutinin resin (Pierce Biotechnology, Inc.) were done on ice in siliconized microcentrifuge tubes. The expression levels of 30 chemokines of the resulting samples (100 μL) were measured by LINCO Research, Inc., using the human chemokine multiplex antibody-array for TGF-β1, IL-1α, IL-1β, IL1α receptor, IL-2, IL-4, IL-5, IL-6, IL-7, IL-8, IL-10, IL-12p40, IL-12p70, IL-13, IL-15, IL-17, EGF, TGF-α, CX3CL1, CCL2, sCD40L, IP-10, VEGF, RANTES, GM-CSF, G-CSF, IFN-γ, MIP1α, MIP1β, and Eotaxin. The expression levels were normalized to total protein.
Statistical analysis
Baseline patient characteristics. Baseline demographic and clinical variables for all patients were assessed using Wilcoxon rank sum tests for continuous variables and Fisher's exact tests for categorical variables. Biomarker expression levels underwent logarithmic transformation to stabilize variance. Values below the detection limit were imputed with half of the minimum detected value for the particular biomarker.
Candidate chemokine selection. As there were 31 candidate chemokines, it was necessary to reduce the number of candidate variables before regression models could be considered. Exploratory data analyses were conducted on 36 (21 recurrent and 15 recurrence-free) prostatectomy subjects. The strength of marginal relationship to the response was used to eliminate variables that showed only a very weak relation. Spearman's rank correlations and Wilcoxon rank sum tests were used for initial screening. A receiver operating characteristics (ROC) curve was constructed for each chemokine and the area under the ROC curve (AUC) was used to compare how strongly chemokines were related to the recurrence. Finally, a backward elimination model building strategy on 1,000 bootstrapped data was used to choose variables for further consideration. Using this bootstrap sample, a full model with all the variables in consideration was constructed. Then, the variable with the largest P value (Wald test) was dropped and a new model was fitted with one fewer variable. Variable elimination continued until all the remaining variables showed a P value of <0.5. The number of retentions of each variable in 1,000 iterations of the final model were used to select the variables for further consideration. Pair-wise Spearman's rank correlations among the candidates were also considered in selecting these variables. All the information from these multiple analyses was used to select the candidate chemokines for the next phase of the data collection and model building.
Logistic regression. Demographic and clinical variables for the logistic regression model were selected based on their relation to the outcome variable (biochemical recurrence) and interrelations among the covariates. Two models were constructed: (a) The three selected chemokine variables with the clinicopathologic variables made up one model. (b) We also fitted a clinical variable–only simpler model without any chemokine variables. These two models were compared using a likelihood ratio test. From each model, the predicted probabilities of recurrence were computed and the ROC curves were constructed. Comparison of the ROC curves was conducted with the integrated discrimination improvement (12). An estimate of the odds ratio with a confidence interval was reported for each variable based on the developed model.
Survival analysis. To visualize the association between biochemical recurrence-free survival and each chemokine marker, the product limit estimator was computed for the two groups defined as above and below the median, and the log-rank tests were used to assess the difference of the recurrence-free survival between the two groups illustrated by Kaplan-Meier plots. A Cox proportional hazard regression was used to model the recurrence-free survival. The proportionality of the hazard ratio was assessed graphically and numerically using Schoenfeld's partial residuals (13). The effects of predictors in the model were presented with individual hazard ratios. A nomogram equating each of the predictors to the probabilities of 1-, 3-, and 5-year recurrence-free survival is also presented. All analyses were carried out with R version 2.7.0 (14).
Results
Chemokines support prediction of biochemical recurrence. The histology of prostatic tissue cores from prostatectomy subjects had focal adenocarcinoma and HGPIN as shown in Fig. 1. However, the histologic patterns were not predictive of future progression to biochemical recurrence. There was no statistical difference in the age, race, and clinical stage of the subjects in the recurrent and recurrence-free groups (Table 1). Because chemokines and growth factors that influence metastatic progression are commonly in low abundance, we developed a methodology to enrich such factors from tissue lysates. Following lectin enrichment of 0.03 to 0.05 g (wet weight) tissue, 31 chemokines were screened by multiplex ELISA for a panel of inflammatory chemokines. Importantly, little to no signal was detected for these markers if wheat germ agglutinin–mediated enrichment process was not done on the tissue extracts. The use of another lectin, concanavalin A–mediated enrichment was ineffective in chemokine detection (data not shown). The results of the initial screening of 36 prostatectomy subjects (21 recurrent and 15 recurrence-free) were analyzed by Spearman's rank correlation and Wilcoxon rank sum tests. The common inflammatory factors found to be differentially expressed by both tests in biochemically recurrent and recurrence-free subjects included CX3CL1, IL-12, IL-15, IL-4, and CCL4. Of this group, CCL4 (macrophage inflammatory protein-1β) was up-regulated in patients that developed biochemical recurrence within 5 years following prostatectomy. However, CX3CL1 (fractalkine), IL-12, IL-15, and IL-4 were predominantly down-regulated in such high-risk patients. Interestingly, each of the factors has been implicated biologically to the progression of prostate cancer in the past (15–19). The AUC for individual chemokines and backward elimination bootstrap methods were correlated to biochemical recurrence in finally selecting CX3CL1, CCL4, and IL-15 for further analyses (data not shown). CX3CL1 was the best predictor of recurrence with all the selection methods, and the other two were among the favorable ones with multiple methods.
Validation of chemokine biomarkers in model development. We sought to develop a multivariable logistic regression model for the prediction of the probability of biochemical recurrence following prostatectomy. We considered all predictor variables available to us that were either numerical or categories for logistic regression analysis. In addition to the three chemokine markers, the candidate clinical and surgical variables we considered were pathologic Gleason score, preoperative PSA, surgical margin status, seminal vesicle involvement, clinical stage, extracapsular involvement, and lymph node metastasis. However, for our population, the clinical stage categorized to T1c, T2a, and T2b or greater did not provide any discrimination regarding the recurrence status of the subjects. Extracapsular involvement was also dropped as a covariate as it was highly correlated with surgical margin (Spearman's rank correlation = 0.47); there were only 46 cases (25%) for which these two variables do not agree. The limited number of subjects having positive lymph node tumor involvement (n = 9) provided little information and was also not considered in further analysis. A fitted logistic regression model of our population suggested two chemokines, CCL4 (P = 0.011) and CX3CL1 (P < 0.0001), along with surgical margin (P = 0.008), seminal vesicle involvement (P = 0.016), preoperative PSA (P = 0.041), and Gleason score (P = 0.048) to be significant factors, whereas IL-15 (P = 0.66) was not. Table 2 summarizes the fitted logistic regression model in terms of the odds ratios.
Odds ratios and confidence intervals for the logistic regression model
. | Reference . | . | Odds ratio (95% CI) . | P . | |
---|---|---|---|---|---|
Gleason score | 6 | 8 | 2.60 (1.01-6.71) | 0.048 | |
Preoperative PSA | 5 | 9 | 1.54 (1.02-2.34) | 0.041 | |
Surgical margin | Negative | Positive | 3.64 (1.41-9.42) | 0.001 | |
Seminal vesicle involvement | Negative | Positive | 7.15 (1.44-35.5) | 0.016 | |
CCL4 | 0.003 | 0.04 | 2.16 (1.19-3.91) | 0.011 | |
CX3CL1 | 0.01 | 0.07 | 0.35 (0.21-0.58) | <0.0001 | |
IL-15 | 0.004 | 0.06 | 0.88 (0.51-1.55) | 0.66 |
. | Reference . | . | Odds ratio (95% CI) . | P . | |
---|---|---|---|---|---|
Gleason score | 6 | 8 | 2.60 (1.01-6.71) | 0.048 | |
Preoperative PSA | 5 | 9 | 1.54 (1.02-2.34) | 0.041 | |
Surgical margin | Negative | Positive | 3.64 (1.41-9.42) | 0.001 | |
Seminal vesicle involvement | Negative | Positive | 7.15 (1.44-35.5) | 0.016 | |
CCL4 | 0.003 | 0.04 | 2.16 (1.19-3.91) | 0.011 | |
CX3CL1 | 0.01 | 0.07 | 0.35 (0.21-0.58) | <0.0001 | |
IL-15 | 0.004 | 0.06 | 0.88 (0.51-1.55) | 0.66 |
NOTE: Odds ratios for continuous variables represent change from the lower quartile to upper quartile except for Gleason score whose lower and upper quartiles were both 7.
Abbreviation: 95% CI, 95% confidence interval.
The logistic regression model that used only the clinical variables (i.e., Gleason score, preoperative PSA, seminal vesicle involvement, and surgical margin status) was compared with the model combining clinical and chemokine variables. The improvement by the three biomarkers was highly significant (likelihood ratio test P <0.0001). Figure 2 shows the ROC curves obtained by the predicted values using the two models. Improvement in area under the curve (AUC) was 7.1 percentage points (from 80.6% to 87.7%). Integrated discrimination improvement was estimated to be 0.116 (P < 0.0001), supporting the statistical significance of the improvement (12). The addition of the chemokine biomarkers to the clinical variables provided little improvement in predictive ability up to ∼80% sensitivity. However, given a sensitivity of 90%, the clinical variables alone provided a specificity of only 36% (95% confidence interval, 20-58%) compared with the addition of the chemokines that provided a specificity of 72% (95% confidence interval, 55-84%). The addition of the chemokine markers to the clinical variables doubled the specificity at 90% sensitivity (P = 0.02) according to the ROC analysis. The chemokines, particularly CCL4 and CX3CL1, supported the dichotomous prediction of biochemical recurrence and recurrence-free survival following prostatectomy.
ROC curves for the prediction models affected by the chemokines (CCL4, CX3CL1, and IL-15). Predicted probability of recurrence for each subject was computed from logistic regression models, including preoperative PSA, surgical margin status, seminal vesicle invasion status, and pathologic Gleason score, with and without chemokines. Specificity and sensitivity were computed at each possible cutoff on the predicted probability for the two models. The AUC values were compared for the two models.
ROC curves for the prediction models affected by the chemokines (CCL4, CX3CL1, and IL-15). Predicted probability of recurrence for each subject was computed from logistic regression models, including preoperative PSA, surgical margin status, seminal vesicle invasion status, and pathologic Gleason score, with and without chemokines. Specificity and sensitivity were computed at each possible cutoff on the predicted probability for the two models. The AUC values were compared for the two models.
Analysis of recurrence-free survival. To define the efficacy of the markers in predicting recurrence-free survival, the same clinical and chemokine variables as in the logistic regression were considered for a Cox proportional hazard regression analysis. CCL4, CX3CL1, and IL-15 proved to individually serve as highly significant markers for biochemical recurrence status by the Kaplan-Meier method (Fig. 3). Using a dichotomous median split for the upper and lower concentration range for tissue chemokine expression, CX3CL1 exhibited the best prediction ability (P < 0.0001) followed by CCL4 (P < 0.001) and IL-15 (P = 0.003). The proportional hazard assumption was tested with scaled Schoenfeld residuals (20). There was no evidence of violation, as the χ2 tests for trend were not significant for any of the seven variables (surgical margin status, seminal vesicle involvement, Gleason score, preoperative PSA, CCL4, and CX3CL1, and IL-15; P values ranging from 0.46 to 0.90). The effect of the covariates in the multivariable Cox regression model on recurrence-free survival was summarized with the hazard ratios with 95% confidence intervals in Table 3. CCL4, CX3CL1, preoperative PSA, and surgical margin were significant factors (Fig. 4). However, Gleason score, seminal vesicle involvement, and IL-15 were not significant in the overall survival multivariable model.
The Kaplan-Meier estimates of the recurrence-free survival based on chemokine expression. The patients were separated into two groups, divided at median tissue level for CCL4 (A), CX3CL1 (B), and IL-15 (C). The two groups were discriminated by the median respective chemokine expression concentration indicated: Those above the median were termed upper half, whereas those below the median were termed lower half. The recurrence-free survival probabilities were estimated by the Kaplan-Meier method and the differences were tested using the log-rank test. Each of the dichotomous chemokine expression levels supported statistically significant differences in biochemical recurrence-free survival.
The Kaplan-Meier estimates of the recurrence-free survival based on chemokine expression. The patients were separated into two groups, divided at median tissue level for CCL4 (A), CX3CL1 (B), and IL-15 (C). The two groups were discriminated by the median respective chemokine expression concentration indicated: Those above the median were termed upper half, whereas those below the median were termed lower half. The recurrence-free survival probabilities were estimated by the Kaplan-Meier method and the differences were tested using the log-rank test. Each of the dichotomous chemokine expression levels supported statistically significant differences in biochemical recurrence-free survival.
Cox proportional hazard regression
. | Reference . | . | Hazard ratio (95% CI) . | P . | |
---|---|---|---|---|---|
Gleason score | 6 | 8 | 1.63 (0.91-2.92) | 0.10 | |
Preoperative PSA | 5 | 9 | 1.35 (1.11-1.63) | 0.0025 | |
Surgical margins | Negative | Positive | 1.81 (1.09-1.63) | 0.023 | |
Seminal vesicle involvement | Negative | Positive | 1.70 (0.91-3.16) | 0.095 | |
CCL4 | 0.003 | 0.04 | 1.34 (1.01-1.78) | 0.040 | |
CX3CL1 | 0.01 | 0.07 | 0.60 (0.47-0.76) | <0.0001 | |
IL-15 | 0.004 | 0.06 | 0.77 (0.57-1.05) | 0.095 |
. | Reference . | . | Hazard ratio (95% CI) . | P . | |
---|---|---|---|---|---|
Gleason score | 6 | 8 | 1.63 (0.91-2.92) | 0.10 | |
Preoperative PSA | 5 | 9 | 1.35 (1.11-1.63) | 0.0025 | |
Surgical margins | Negative | Positive | 1.81 (1.09-1.63) | 0.023 | |
Seminal vesicle involvement | Negative | Positive | 1.70 (0.91-3.16) | 0.095 | |
CCL4 | 0.003 | 0.04 | 1.34 (1.01-1.78) | 0.040 | |
CX3CL1 | 0.01 | 0.07 | 0.60 (0.47-0.76) | <0.0001 | |
IL-15 | 0.004 | 0.06 | 0.77 (0.57-1.05) | 0.095 |
NOTE: Multivariable Cox proportional hazard ratios were computed to determine predictors of biochemical recurrence-free survival. The hazard ratios were computed for a change from the lower quartile to upper quartile in continuous variables. For both surgical margins and seminal vesicle involvement, negative is the reference group.
Cox proportional hazard regression. Multivariable Cox proportional hazard regression for biochemical recurrence-free survival showed that preoperative PSA, surgical margin, CCL4, and CX3CL1 were significant predictors of recurrence-free survival. For each predictor variable, the vertical bars illustrate the hazard ratio estimate and the gray horizontal bars represent the respective 95% confidence intervals described in Table 3. The hazard ratios were computed for a change from the lower quartile to upper quartile in continuous variables, namely Gleason score 6 to 8, preoperative PSA 5 to 9, CCL4 0.003 to 0.04, CX3CL1 0.01 to 0.07, and IL-15 0.004 to 0.06. For both surgical margins and seminal vesicle involvement, negative is the reference group.
Cox proportional hazard regression. Multivariable Cox proportional hazard regression for biochemical recurrence-free survival showed that preoperative PSA, surgical margin, CCL4, and CX3CL1 were significant predictors of recurrence-free survival. For each predictor variable, the vertical bars illustrate the hazard ratio estimate and the gray horizontal bars represent the respective 95% confidence intervals described in Table 3. The hazard ratios were computed for a change from the lower quartile to upper quartile in continuous variables, namely Gleason score 6 to 8, preoperative PSA 5 to 9, CCL4 0.003 to 0.04, CX3CL1 0.01 to 0.07, and IL-15 0.004 to 0.06. For both surgical margins and seminal vesicle involvement, negative is the reference group.
We computed a nomogram from the Cox proportional hazard regression model that connects each predictor and the probabilities of 1, 3, and 5-year recurrence-free survival (Fig. 5). The contributions of CX3CL1 and preoperative PSA were further shown to be important in predicting recurrence-free survival. We also noted that IL-15, although not statistically significant on multivariable analysis, contributed to the prediction of recurrence-free survival comparable with CCL4, seminal vesicle involvement, and surgical margins within the nomogram. Together, the survival model suggested the use of surgical margin status, seminal vesicle involvement, Gleason score, preoperative PSA, CCL4, and CX3CL1, and IL-15 for predicting biochemical recurrence following prostatectomy.
Nomogram from the Cox proportional hazard regression model. The Cox proportional hazard regression model was used to create a prediction model for 1-, 3-, and 5-y recurrence-free survival. A value in each predictor variable corresponds to a point scale (top). The sum of the individual predictor variable points corresponds to the probability of 1-, 3-, and 5-y recurrence-free survival (bottom).
Nomogram from the Cox proportional hazard regression model. The Cox proportional hazard regression model was used to create a prediction model for 1-, 3-, and 5-y recurrence-free survival. A value in each predictor variable corresponds to a point scale (top). The sum of the individual predictor variable points corresponds to the probability of 1-, 3-, and 5-y recurrence-free survival (bottom).
Discussion
In most cases, patients with clinically localized prostate cancer treated locally with modalities such as surgery or radiation therapy will be cured of their disease. However, a proportion of men will harbor microscopic localized or metastatic residual disease. These patients will ultimately develop biochemical recurrence of disease and, eventually, are at risk of developing clinical metastatic progression and death from their prostate cancer. The risk of progression after local therapy is generally estimated based on available clinicopathologic variables. For example, after radical prostatectomy, it is estimated based on both clinical data, such as presurgical PSA, as well as pathologic data such as extracapsular extension, seminal vesicle involvement, surgical margin status, and the Gleason score. These approaches described by Amico et al., Kattan et al., and others have improved our ability to risk-stratify patients through a continuous variable, taking into account these types of variables with high specificity yet with relatively low sensitivity (21–23). Due to the lower sensitivity of such available risk-stratifying measures, it has been difficult to justify the routine use of adjuvant therapy before frank biochemical recurrence. Several randomized phase III trials have tested the use of routine, adjuvant radiation therapy after prostatectomy in men with high-risk disease estimated using only clinicopathologic parameters (24–26). These studies found significant improvement in biochemical recurrence-free survival in men with adjuvant radiation therapy but found no improvement in overall survival (27). An inability to identify men at high risk may have contributed to the equivocal outcome. This was in no small part due to deficiencies in our ability to accurately risk-stratify patients using only clinicopathologic variables. With the addition of the chemokines identified in this study to more standard clinicopathologic variables, we were able to predict the risk of recurrence among men who after prostatectomy with greater accuracy than by clinicopathologic parameters alone. The logistic regression model that included the three chemokines improved the specificity from 36% to 72% at 90% sensitivity, when compared with clinicopathologic parameters alone (Fig. 2). This corresponded with a significant improvement in the AUC for the model that included all three chemokines to 87.7% versus 80.6% for only clinicopathologic variables.
The chemokines identified through the lectin-based enrichment method have specific positive and negative biological roles in prostate cancer progression. All eukaryotic organisms glycosylate proteins exposed to the extracellular space. Because lectins specifically bind such glycosylation groups, we processed the prostatectomy specimens with wheat germ agglutinin resin in batch and were able to observe the expression of a number chemokines otherwise not detectable. Inflammatory cells are emerging as potential mediators of cancer metastasis (11). Both CX3CL1 and CCL4 can recruit natural killer cells, T cells, and monocytes. Based on our data, however, the two chemokines seem to reflect an opposite status of prostate cancer recurrence. Accordingly, apart from similar inflammatory recruitment characteristics, CCL4 has direct proliferative and migration effects on prostate cancer cells in vitro (28). Conversely, CX3CL1 is reported to reduce migration of prostate cancer cells in culture (29). Finally, IL-15 can prevent prostate cancer progression by supporting natural killer cell function in vivo (15, 30). CX3CL1 and CCL4 overwhelmingly supported prediction of prostate cancer biochemical recurrence under all univariant and multivariant criteria tested. Although the univariable analysis suggested IL-15 to be significant, as illustrated by Kaplan-Meier plot (Fig. 3), the overall multivariable models (logistic regression and Cox proportional hazard) did not indicate statistical significance in predicting recurrence-free survival (Tables 2 and 3). Nevertheless, specific multivariable 1-, 3-, and 5-year survival analysis supported the importance of IL-15 as a predictive factor, as shown in the nomogram (Fig. 5). The significant contribution of each of the chemokine biomarkers needs to be taken in context of clinicopathologic parameters. The well-recognized predictors for biochemical recurrence, like Gleason scores and seminal vesicle involvement, were greater in the population studied (Table 1); however, their predictive significance was diminished when compared with the chemokine biomarkers in a multivariable analysis (Table 3; Fig. 4). Together, the data argue for considering multiple biomarkers in the prediction of those prostate cancers likely to have the greatest clinical significance, especially when coupled to clinicopathologic parameters.
We have shown in this report that a cumulative evaluation of the expression by a continuous variable was able to accurately predict biochemical recurrence after radical prostatectomy. Using a model incorporating chemokines and clinicopathologic variables developed in this study, men could be identified who were at substantially higher risk of recurrence after prostatectomy. Whereas local recurrence can potentially be treated with prostatic bed salvage radiation, systemic disease is treated with hormonal therapy, of which the exact timing is controversial. However, of the 82 subjects with biochemical recurrence, 44 (54%) showed elevated PSA within 1 year of surgery. These apparent early recurrence subjects likely had already developed local or distant metastasis before surgery. In an era where there is growing concern that many prostate cancers are overtreated, the use of this model could support the clinical determination of those prostate cancers that would progress to symptomatic disease and ultimately death, as well as potentially provide novel biological targets to inhibit the metastatic progression. At the same time, it has the potential to identify those patients who harbor more indolent disease that could be managed by active surveillance and may be spared the potential morbidity of aggressive local therapy. In principle, the same model that was generated in this study could be done on diagnostic biopsy specimens and incorporated into a model using clinical variables such as the PSA, clinical Gleason score, and clinical stage to more accurately predict a patient's disease risk before any therapy. It should be noted, however, that this idea was not formally tested in this study as the specimens were obtained on freshly removed prostates after prostatectomy. Nevertheless, the cores used here were of similar nature as would be obtained at the time of diagnostic biopsy of the prostate. Notably, many of the tissues processed had no evidence of adenocarcinoma in the sample (Fig. 1). As the biomarkers are secreted factors, it seems that actual tumor sampling is not particularly necessary. However, more studies are needed to know the extent of adjacency required for positive prediction ability. Therefore, the same techniques used in this study can be readily transferred to the clinical setting at the time of prostate biopsy. Previously reported blood or tissue biomarker analyses have not approached the sensitivity and specificity of the prediction achieved by the model described here (31–37).
In summary, we have shown that prostate tissue levels of three chemokines—CCL4, CX3CL1, and IL-15—are predictive of biochemical recurrence in men who have undergone radical prostatectomy for adenocarcinoma of the prostate with high specificity and sensitivity. Further, we have shown that a nomogram that incorporates these three chemokines with other established clinicopathologic variables is a better predictor of outcome than using clinicopathologic variables alone. This suggests that CCL4, CX3CL1, and IL-15 are biomarkers of prostate cancer recurrence after radical prostatectomy.
Disclosure of Potential Conflicts of Interest
No potential conflicts of interest were disclosed.
Grant support: NIH Institutional Research and Academic Career Development Award (5K12GM068543; A.E. M'Koma) and the T.J. Martel Foundation.
The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.
Note: D.L. Blum, T. Koyama, and A.E. M'Koma contributed equally to this work.
Acknowledgments
We thank Brenda Hughes, Indrani Bhowmick, and Dr. Jheelam Banerjee for their efforts in gathering patient information; Dr. Bruce Roth for critical discussions; and Susan and Luke Simons for supporting the study (N.A. Bhowmick).