Bone turnover biomarkers identify unique prognostic risk groups in men with castration resistant prostate cancer and skeletal metastases: results from SWOG S0421

Abstract

Background.

Skeletal metastases often occur in men with castration-resistant prostate cancer (CRPC) where bone biomarkers are prognostic for overall survival (OS). In those with highly elevated markers, there is preferential benefit from bone-targeted therapy. In the phase IIIS0421 docetaxel+/− atrasentan trial, clinical covariates and bone biomarkers were analyzed to identify CRPC subsets with differential outcomes.

Subjects and Methods.

Markers of bone resorption [N-telopeptide-NTx; pyridinoline-PYD] and formation [C-terminal collagen propeptide-CICP; bone alkaline phosphatase-BAP] were measured in pre-treatment sera. Bone biomarkers and clinical covariates were included in a Cox model for OS; bone markers were added in a stepwise selection process. Receiver operating characteristic(ROC) curves were constructed for risk factor models +/− bone markers. Significant variables were allowed to compete in a classification and regression tree (CART) analysis. Hazard ratios(HR) were calculated by comparing OS in each of the terminal nodes to a reference group in a Cox model.

Results.

750 patients were included. Each bone marker significantly contributed to the risk factor-adjusted OS Cox model, with higher levels associated with worse OS. BAP (HR=1.15,p=0.008), CICP (HR=1.27,p<0.001), and PYD (HR=1.21,p=0.047) in combination were significantly associated with OS. Prognostic accuracy was improved by addition of bone markers to clinical covariates. CART analysis selected CICP, BAP, hemoglobin, and pain score for the final OS model, identifying five prognostic groups.

Conclusions.

Elevated serum bone biomarker levels are associated with worse OS in bone-metastatic CRPC. Bone biomarkers can identify unique prognostic subgroups. These results further define the role of bone biomarkers in the design of CRPC trials.

Background

Metastatic castration resistant prostate cancer (CRPC) is the terminal state of the prostate cancer disease trajectory.^1,2 Of the 26,700 American men who are estimated to succumb to prostate cancer in 2017³, the majority is attributable to CRPC. In men with CRPC, skeletal metastasis is often observed and is a frequent source of morbidity such as bone pain and fracture.⁴ In these patients, the homeostatic balance between bone formation and resorption is frequently disrupted, with predominance of osteoblastic activity manifesting as sclerotic bony disease. Furthermore, the concurrent use of androgen deprivation therapy enhances bone turnover in these patients, resulting in osteopenia and osteoporosis.

Markers of bone turnover can be clinically assessed using circulating biomarkers in serum.⁵ Our group previously evaluated these biomarkers in the context of a large placebo-controlled prospective phase III trial of the bone-targeted agent endothelin-A antagonist atrasentan in combination with docetaxel.⁶ We reported that elevated levels of these blood-based bone biomarkers have significant independent prognostic value for overall survival in men with CRPC.⁷ Importantly, we identified a subset of CRPC patients with highly elevated markers who preferentially benefit from bone-targeted therapy, providing a potential pathway for a precision medicine approach in this disease.⁷

In order to further refine the potential role of bone biomarkers in the clinical evaluation of men with CRPC and in the design of trials testing bone targeted therapies in these patients, we assessed the individual and collective contributions of each of the bone resorption and formation markers to OS in the context of baseline covariates. Furthermore, we sought to identify unique subsets of patients defined by their baseline bone marker levels and clinical features using classification and regression tree methods.

Methods

S0421 was a two-arm, randomized phase III trial, open label for docetaxel and double-blind placebo-controlled for atrasentan (ClinicalTrials.gov NCT00134056). Patients were assigned to receive docetaxel 75mg/m² intravenously over 1 hour every 21 days with prednisone 10 mg orally daily with or without atrasentan 10 mg daily orally. The trial was placebo-controlled for atrasentan. As previously reported, atrasentan failed to improve progression-free survival or OS in the overall population.⁶ Patients registered to S0421 were consented to provide serial serum specimens for the bone marker studies. All patients in this trial had metastatic CRPC with imaging evidence of bone metastasis. Bisphosphonate therapy was permitted but must have commenced before registration; initiation of bisphosphonates was not permitted within the first four cycles of study therapy. The study protocol was approved by the institutional review board or the National Cancer Institute Central Institutional Review Board or both. Blood (serum) samples were collected using standard venipuncture techniques. Fifteen milliliters of whole blood were drawn pretreatment (after registration but before receiving the first dose of protocol therapy) and on the day of docetaxel infusion in weeks 4, 7, and 9. Whole blood was collected in red-top vacutainer tubes and allowed to clot for approximately 30 minutes. Serum was separated from cells within 45 to 60 minutes of venipuncture by centrifugation at 3000× for 10 minutes. Serum was equally aliquoted into four cryotubes and shipped to the SWOG biobank at Nationwide Children’s Hospital in Columbus, OH. Specimens were subsequently shipped to the USDA lab at UC Davis for bone marker analysis.

Markers for bone resorption (N-telopeptide [NTx] and pyridinoline [PYD]) and bone formation (C-terminal collagen propeptide [CICP] and bone alkaline phosphatase [BAP]) were measured in pre-treatment sera collected from men enrolled in the SWOG S0421 trial, as previously described.⁷ Briefly, CICP was measured by a sandwich enzyme-linked immunosorbent assay (Quidel Corp, San Diego, CA) using a microtiter plate coated with monoclonal anti-C1CP antibody. Bone-specific alkaline phosphatase (BAP) activity in serum was measured using the Microvue BAP enzyme-linked immunosorbent assay (Quidel Corp) with a monoclonal anti-BAP antibody coated on a microtiter plate to capture BAP in the sample. N-telopeptide (NTx) was measured by a competitive enzyme-linked immunosorbent assay (Wampoles Laboratories, Princeton, NJ) using a 96-well microplate. Pyridinoline (PYD) was measured in serum using a competitive enzyme immunoassay in a microtiter plate format (Quidel Corp).

Bone markers were first added individually to a multivariate Cox regression model for OS that contained traditional risk factors of CRPC with a low fraction of missing data. Then, holding all risk factors in the Cox model constant, each bone marker was evaluated univariately and in a multivariate fashion. (A criteria to stay of p< 0.05 was used to select which combination of the four bone markers best contributes to the model. Receiver operating characteristic (ROC) curves were estimated for the traditional risk factor model predicting survival at 24 months, and also with the addition of the multivariate bone markers. The AUC was compared between the two models. Classification of patients by their survival status at 24 months and their predicted probability of death (low, medium/low, medium/high, or high) from the logistic regression model for the risk factor model and the multivariate bone marker model were compared to assess re-classification of risk based on the addition of bone marker data.

All four bone markers (BAP, CICP, NTx, PYD) and all risk factors used in the OS Cox model were allowed to compete as input variables in a regression tree analysis, with overall survival as the outcome. The binary partitioning of the bone marker levels in the regression tree was performed by the C-TREE function in the R package party. This function uses recursive binary partitioning to choose the cut point of quantitative variables, such as bone marker measures and HgB levels. The optimal binary split is identified as the observed bone marker measure where the logrank test statistic is maximized. A final classification and regression tree (CART) was constructed. C-TREE was used to construct the regression tree due to its permutation based significance testing which avoids selection bias towards input variables with many possible cut-points. Since this method utilizes permutation tests, and splitting is based on statistical stopping rules, pruning was not conducted. Hazard ratios and corresponding 95% CIs were calculated by comparing overall survival of patients in each of the nodes 2–5 relative to node 1 (reference group) in a Cox proportional hazards model. Kaplan-Meier curves for each terminal node were constructed.

Results

S0421 registered 1038 eligible patients with CRPC. Of these, 855 submitted serum for the bone biomarker studies. Of 855 men, 778 had usable specimens at baseline. In total, 750 patients with evaluable bone marker and clinical data were included. Patient characteristics for this cohort are summarized in Table 1.The following risk factors were considered of interest, were included in the multivariate OS Cox models and were candidates in regression tree analyses: age (in years), performance status (0–1 vs 2), hemoglobin (Hgb, g/dL), type of progression determining unresponsiveness to hormone therapy at baseline (measurable disease /non-measurable disease vs rising PSA), worst pain score at study entry as measured by the Brief Pain Inventory (<4 vs ≥4), race (black vs other), PSA at study entry (ng/mL), visceral disease present (yes vs no) and treatment arm (placebo vs atrasentan). Although treatment arm was not significant in the model (p=0.90), we chose to keep it for completeness.

Table 1:

Patient Characteristics

Variable	N (%)
Total Patients	750
Race
Black	100 (13)
Other	15 (2)
Unknown	13 (2)
White	622 (83)
Type of Progression
Measureable/Evaluable	606 (81)
PSA only	144 (19)
Bisphosphonate Usage
No	293 (39)
Yes	457 (61)
Worst pain score
< 4	446 (59)
≥ 4	304 (41)
Extraskeletal metastases
No	337 (45)
Yes	413 (55)
Performance Status
0	333 (44)
1	356 (48)
2	61 (8)
Gleason Score
≤ 6	92 (12)
7	214 (29)
≥ 8	415 (55)
Missing	29 (4)
Treatment arm
Atrasentan	372 (50)
Placebo	378 (50)
Age at registration
Mean (Standard deviation)	69 (9)
Median (Interquartile Range)	69 (63–76)
Baseline PSA
Mean (Standard deviation)	249 (643)
Median (Interquartile Range)	77 (25–208)

Each bone marker significantly contributed to the risk factor Cox model univariately, with higher levels associated with worse OS. A selection model adjusted for clinical risk factors showed that the best combination of bone markers was with BAP (Hazard Ratio [HR]=1.15; 95% CI (1.04,1.27); p=0.008), CICP (HR=1.27; 95% CI (1.11, 1.45); p<0.001), and PYD (HR=1.21; 95% CI (1.00, 1.47); p=0.047).

In ROC analysis, the AUC of clinical risk factors for predicting 24 month OS was found to be 0.73; this modestly improved with the addition of CICP (AUC=0.76) or BAP (AUC 0.75) or combination BAP/CICP/PYD (AUC=0.76). These results are shown in Table 2 and the multivariate bone marker model in Figure 1. In Table 3, patients were cross-classified by their survival status at 24 months and the predicted probabililty (4 categories) of their death at that time point both for the risk factor model and the model that also included bone markers. . The addition of bone markers to the model correctly shifted some of the men who had died into a high probability category and conversely, some of those who were alive were shifted to a lower probability category. Approximately 9% more men who had died by 24 months were re-assigned to the highest risk category and approximately 10 % more men who were alive at 24 months were re-assigned to the lowest risk category when the 3 bone marker combo (BAP + CICP + PYD) was added to the predictive model.

Table 2.

Logistic Regression and Receiver Operating Characteristic (ROC) Analysis for Overall Survival

Bone marker	Area Under the ROC	95% Confidence Interval	P-value (vs. reference group)
Risk factors (RF), No bone markers*	0.73	0.69, 0.76	N/A
RF + Bone alkaline phosphatase (BAP)	0.75	0.72, 0.79	0.0038
RF + C-terminal collagen propeptide (CICP)	0.76	0.72, 0.79	0.0025
RF + Pyridinoline (PYD)	0.73	0.69, 0.77	0.57
RF + N-telopeptide (N-Tx)	0.74	0.71, 0.78	0.07
RF + (BAP+CICP+ PYD)	0.76	0.73, 0.80	0.0013

N= 734 patients (285 alive, 449 dead at 2 years); 16 patients who were lost to follow-up prior to 2 years were excluded from this analysis

^*Reference group: model contains all of the risk factors included in Cox model

Abbreviations: N/A = not applicable

Figure 1:

Receiver Operating Characteristic (ROC) Curves for clinical covariates with (red curve) and without (blue curve) the three significant bone markers predicting 2-year survival.

Table 3:

Impact of Bone Markers on Predicted Probability of Death, Stratified by Survival Status at 24 Months

Multivariate Logistic Model, Risk Factors Only*
	Model-based Predicted Probability of Death Categories
	LOW	MED/LOW	MED/HIGH	HIGH	TOTAL
	< 45 %	45–62 %	63–79 %	>79%	N=734
Deceased by 24 months (n, %)	70 15.6 %	98 21.8 %	143 31.9 %	138 30.7 %	449 100 %
Alive at 24 months (n, %)	105 36.9 %	95 33.3 %	67 23.5 %	18 6.3 %	285 100 %
When Bone Markers are Added to the Model**
	Model-based Predicted Probability of Death Categories
	LOW	MED/LOW	MED/HIGH	HIGH
	< 45 %	45–62 %	63–79 %	>79%
Deceased by 24 months (n, %)	64 14.3 %	96 21.4 %	110 24.5 %	179 39.9 %	449 100%
Alive at 24 months (n, %)	133 46.7 %	69 24.2 %	66 23.2 %	17 5.9 %	285 100%

CART analysis selected CICP, BAP, Hgb, and pain score for the final model (Figure 2), and identified five prognostic groups with differential OS outcomes (Table 4 and Figure 3). Kaplan-Meier curves for each terminal node were constructed to demonstrate the OS differences across each identified subset. Patients in Node 1 had a Hgb <= 11.3 g/dL. Patients in Nodes 2–5 all had Hgb levels above 11.3 g/dL. Patients in Nodes 3–5 all had CICP levels > 6.8 ng/mL (only patients in Node 2 had CICP <= 6.8 ng/mL). Patients in Nodes 4–5 had a pain score of <4 while patients in node 3 had a pain score >= 4. Finally, patients in Node 4 had BAP levels <= 90.9 u/L while those in Node 5 had BAP > 90.9 u/L. Node 1 had the worst outcome (median OS of 12.4 months), while Node 2 had the best median OS of 31.6 months. Patients in Node 2 and Node 4 (median OS of 27.1 months) had significantly longer OS than patients in Node 1 (both comparisons p<0.001).

Figure 2:

Final Classification and Regression Tree (CART).

Table 4: Prognostic Risk Groups identified by classification and regression tree analysis.

CART analysis selected CICP, BAP, HGB, and pain score for the final model, and identified five prognostic groups with differential OS outcomes.

NODE	Description				N	Median OS (months)	HR	95% CI		p-value
1	HGB ≤ 11.3				177	12.4	REF	REF	REF	REF
2	HGB > 11.3	CICP ≤ 6.8			191	31.6	0.28	0.22	0.36	< 0.001
3		CICP > 6.8	worst pain ≥ 4		144	15.3	0.79	0.63	0.99	0.04
4			worst pain  < 4	BAP ≤ 90.9	140	27.1	0.34	0.27	0.44	< 0.001
5				BAP > 90.9	98	17.1	0.68	0.53	0.88	0.003

Figure 3: Kaplan Meier survival curves for each terminal node of the classification and regression tree.

Kaplan-Meier curves for each terminal node were constructed to demonstrate the survival differences across each identified subset. Node 1 had the worst outcome (median OS of 12.4 months), while Node 2 had the best median OS of 31.6 months (see Table 4).

The effect of treatment within each node group was evaluated with an interaction term. The addition of atrasentan to docetaxel did not alter OS in any of the nodes (4 degreee of freedom residual Chi-square, p=0.53), a finding in keeping with the negative primary outcome for S0421.

Discussion

Precision cancer care, where patients with molecularly defined subsets of tumors are treated with a specific targeted agent that is predicted to yield a high tumor response rate, is now an established paradigm in many solid tumors such as non-small cell lung cancer, colorectal cancer, and malignant melanoma, among others. For example, in patients with metastatic non-small cell lung cancer of the adenocarcinoma subtype, it is standard-of-care to assess the status of selected oncogenic drivers to define the optimal systemic therapy approach.⁸ However, such an approach remains investigational for the vast majority of patients with CRPC.⁹ One exception is the ongoing active investigation into the predictive value of DNA damage repair gene alterations for treatment with PARP inhibitors or platinum-based chemotherapy^10,11; however, as of this writing, there is still no regulatory approval for any systemic therapy for prostate cancer that is defined by unique molecular subsets. Existing biomarkers in the CRPC context such as prostate specific antigen (PSA) levels or circulating tumor cells (CTCs) are currently employed to refine prognostication but not to assign specific treatments.

Most men with metastatic CRPC will have skeletal involvement at some point in their disease trajectory. Thus, there has been high interest in the drug development community to pursue therapies directed towards this patient population. As a result, many systemic agents directed towards bony metastatic disease have become commercially available over the past decade. These include drugs that reduce skeletal-related events such as bone pain or fracture (e.g., zoledronic acid, denosumab) as well as radio-isotopes (e.g., Radium 223).¹² Many other bone-targeted therapies have failed to yield sufficient activity in randomized phase III trials, including endothelin antagonists (e.g., atrasentan, zibotentan) and SRC-inhibitors (e.g., dasatinib), to warrant regulatory approval.^6,12,13

There are still no practical means in the clinic to identify biomarker-defined subsets of patients that can guide bone-targeted therapy selection. Interestingly, skeletal metastatic disease can be clinically assessed and monitored using blood- or urine-based bone turnover biomarkers. This context offers an opportunity to define patient subsets responsive to bone-targeted therapies. For example, patients with bone metastases treated with denosumab were shown to have rapid and sustained reduction in bone turnover biomarkers including urinary N-telopeptide regardless of prior bisphosphonate therapy.¹⁴ Similar effects were seen in a subset of men with bony metastatic prostate cancer.¹⁵ Even in patients with non-metastatic prostate cancer, denosumab was shown to induce rapid inhibition of bone turnover biomarkers.¹⁶ The radioisotope Radium-223 has also been shown to significantly reduce bone alkaline phosphatase levels in men with bone-metastatic CRPC. However, marker decline did not appear to be a surrogate for survival.¹⁷ In contrast, our data from S0421 CRPC subjects treated with docetaxel/prednisone +/− atrasentan not only reinforced the strong prognostic value of baseline circulating markers of bone metabolism, but showed that in an enriched subset of men with the highest bone marker levels, atrasentan can improve survival.⁷ In addition, our updated CART analysis incorporating bone biomarkers with baseline clinical covariates (CICP, BAP, Hgb, and pain score) identified five prognostic subgroups of CRPC patients with differential survival outcomes. It must be noted that although the primary analysis on the prognostic role of each of the four bone biomarkers was prospectively pre-specified at the time of S0421 initiation, the subsequent ROC and CART analyses reported here were performed post-hoc and should be considered a limitation of the study.

We believe that these data can be employed by clinical investigators in the design and conduct of future CRPC trials either as an enrichment strategy for bone-targeted therapies or as a stratification factor at randomization. Presently, our group is evaluating the prognostic and predictive role of these bone biomarkers in the context of metastatic hormone sensitive prostate cancer as part of SWOG S1216, a phase III trial of androgen deprivation therapy with or without the CYP17 inhibitor orteronel.¹⁸