EURURO Vol. 73 No. 1

inflection point was around 11

–

12 yr after which the GPS

appeared to be less predictive of PCD. To account for this, a

model was fitted with censoring at 10 yr and the PH

assumption was valid for the censored model (

= 0.19 for

test for nonproportionality). GPS was more strongly

associated with time to PCD within 10 yr (HR per 20 GPS

units = 4.43; 95% CI: 2.66

–

7.34;

0.001).

3.4.

GPS as a predictor of BCR

The cohort included 117 patients with BCR (weighted

proportion = 23%) and 142 with non-BCR. GPS was strongly

associatedwith time to BCR (HR/20 units = 2.50; 95% CI 1.62

–

3.85;

0.0001) in univariable analysis, and was significant

after adjusting for diagnostic PSA, clinical T stage, and central

biopsy GS (HR/20 units = 1.96; 95% CI: 1.28

–

3.03;

= 0.002).

GPS remained significant after adjusting for clinical risk

groups with HR 2.11 (

0.001;

Table 2

). The linearity and PH

assumptions were valid in the Cox PH model for BCR.

3.5.

Contribution of gene groups to the prediction of outcome

The 12 cancer-related genes in the GPS represent four

biological pathways: androgen signaling, stromal response,

cellular organization, and proliferation. Downregulation of

androgen signaling and upregulation of stromal response

gene groups were most strongly associated with time to

metastases, time to PCD, and time to BCR, but each of the

four gene groups contributed to the prediction of the end

points

( Fig. 3

3.6.

Clinical utility of the GPS assay

Five- and 10-yr risk profile curves were generated to

describe risk prediction based on the continuous GPS for

metastasis, PCD, and BCR

( Fig. 3

and Supplementary Fig. 1).

GPS provided additional individualized prognostic informa-

tion beyond NCCN risk stratification. It is noteworthy that no

distant metastases or PCD were observed in low- and

intermediate-risk patients with GPS values

20. Conversely,

intermediate-risk patients with GPS

40 had a 5-yr risk of

distant metastasis

–

free survival similar to clinically high-

risk patients (84% vs 85%; Supplementary Fig. 2).

In ROC analysis of metastases and PCD at 10 yr,

incorporation of GPS to CAPRA also improved the c-statistic

from 0.65 (CAPRA alone) to 0.73 (GPS + CAPRA) for

metastases, and from 0.78 (CAPRA alone) to 0.84 (GPS

+ CAPRA) for PCD, which was a relative increase in

c-statistic by 53% compared with CAPRA risk stratification

for 10-yr metastases, and by 21% for 10-yr PCD

( Fig. 4 A

and

4B). Similarly, incorporation of GPS improved the c-statistic

from 0.66 for NCCN alone to 0.75 (GPS + NCCN) for

metastases (

0.001), and from 0.71 (NCCN alone) to

0.81 (GPS + NCCN) for PCD (

0.001). Thus, when GPS was

combined with NCCN, the c-statistic increased by 56%

compared with NCCN risk stratification for 10-yr metasta-

ses, and by 48% for 10-yr PCD (data not shown).

GPS was predictive of time to metastases, time to PCD,

and time to BCR within clinically relevant subsets of

patients by NCCN risk groups, age, race, and central biopsy

GS (Supplementary Tables 1

–

3, and Supplementary Fig. 3).

GPS was also a significant predictor of all three end points in

the subset of patients who underwent 12 core biopsies

(52% weighted proportion of full study cohort), after

adjusting for PSA density (

0.004 for all; data not shown).

Discussion

The adoption of new prognostic biomarkers for PCa requires

a high level of evidence, including analytic and clinical

validation and prospective support of clinical utility

[19]

. The GPS has previously been validated as a predictor

[(Fig._2)TD$FIG]

Fig. 2

–

GPS distribution by (A) NCCN risk group and (B) race. GPS = Genomic Prostate Score; NCCN = National Comprehensive Cancer Network; VL = very

low.

E U R O P E A N U R O L O GY 7 3 ( 2 0 18 ) 12 9

–

13 8

133