Copy number alteration burden predicts prostate cancer relapse

To generate an independent cohort of prostate cancer patients with molecular and clinical follow-up data, we analyzed genome-wide CNAs in 104 prostatectomy cases from Memorial Sloan–Kettering Cancer Center. This cohort, hereafter referred to as the contemporary cohort, had a median follow-up time of 6 y (Table S1 and Dataset S1). We also updated the full clinical and outcome information from our initial cohort (14) of 181 prostatectomy cases (Datasets S2–S4), including 168 cases with high-resolution CNA data (Dataset S2) and a median follow-up of 8 y (Table S1). This initial cohort has served as a widely used prostate cancer resource with copy number, gene expression, and clinical and outcome annotation (7, 15–20); it now includes metastasis as a definitive clinical endpoint in a significant number of cases. The initial and contemporary cohorts have BCR incidence rates of 43 and 53 per 1,000 person-years and metastasis incidence rates of 15 and 5 per 1,000 person-years, respectively. These cohorts have BCR rates of 27% (n = 46) and 23% (n = 24), respectively, and metastasis rates of 11 (n = 19) and 3% (n = 3), respectively, representing one of few large prostate oncogenomic resources with substantial outcome data.

The clinical characteristics of the two independent prostate cancer cohorts are similar (Fig. S1 and Table S1). Pretreatment PSA, age, and lymph node and seminal vesicle invasion did not significantly differ between cohorts (Table S1). The rate of extracapsular extension was higher in the contemporary cohort (Table S1) (P value = 0.05) whereas positive surgical margins were more prevalent in the initial cohort (Table S1) (P value = 0.04). The Gleason score distribution differed between the two cohorts (P value = 0.022), with the contemporary cohort having a higher percentage of Gleason 7 cases and the initial cohort having a higher percentage of Gleason 6 and 8 cases. Nonetheless, these clinical differences did not translate into outcome risk differences. The two cohorts were not significantly different in their risk of BCR, determined by the Stephenson postoperative nomogram (21), which combines clinical and pathological variables such as age, PSA, Gleason grade, and pathological stage (Fig. S1 and Table S1) (P value = 0.8).

The landscape of CNAs observed in the contemporary cohort is highly representative of the gross pattern of aberrations in prostate cancer genomes, including in our initial cohort, that of Barbieri et al. (22), and the interim Cancer Genome Atlas (TCGA) prostate cancer cohort (Fig. 1A); these include frequent genomic losses on chromosomes 6p, 8p, 13q, and 16p, genomic gains of 7 and 8q, and established focal alterations spanning phosphatase and tensin homolog (PTEN), RB1, and tumor protein p53 (TP53) among others (Fig. 1B). The level of CNAs in the new cohort was lower on average than earlier cohorts (Fig. 1), possibly reflecting differences in cohort selection criteria. Nonetheless, the contemporary cohort possesses canonical alterations common to prostate cancer.

In our initial prostate cancer cohort, some CNA patterns identified by hierarchical clustering were previously found to be correlated with the risk of BCR (14). Clinical application of this finding, however, is complicated by the analytical challenge of assigning individual patients to a risk group based on a clustering analysis derived from a different population.

We therefore asked whether CNA burden, a simple metric of CNA level defined as the percent of the autosomal tumor genome bearing CNAs, could be used as an informative measure of CNA. CNA burden represents the level of copy number gains and losses across the genome in a tumor. To calculate CNA burden for a sample, segments of copy number gains and losses were determined (23), and their total genomic length was summed and calculated as a percentage of the size of the autosomal genome. In primary prostate cancer genomes, CNA burden ranged from less than 0.1% to greater than 50%, with an average of 5% and 4% for the initial and contemporary cohorts, respectively. This finding differs significantly from the higher CNA burden of metastatic prostate cancers (∼32%), as well as of other urogenital neoplasms, such as bladder and renal cell carcinomas (Fig. 1C).

CNA burden, when examined as a continuous variable, was significantly associated with BCR in both the initial and contemporary cohorts [hazard ratios (HR) = 1.09 and 1.05, respectively, for each percent change in CNA burden; P values ≤ 0.0001 and 0.008, respectively] (Table 1). We additionally sought to determine whether CNA burden was associated with BCR when patients were separated into groups based on discrete levels of CNA in their tumor genomes. As a starting point for this analysis, we examined the CNA burden from our earlier hierarchical clustering analysis, which segregated tumors into groups with low, intermediate, and high CNA burdens, independent of clinical outcome (14). We reasoned that the mean CNA burden in the low and intermediate alteration clusters (1.34% and 5.41%, respectively) would serve as logical thresholds for further stratified analysis. These thresholds are not based on risk categories from the prior group, but rather copy number clusters indicative of the identity and level of CNA. Stratification at the intermediate CNA burden threshold (5.41% CNA) was associated with BCR in both the initial and contemporary cohorts, with HRs of 1.99 and 3.85, respectively (P values = 0.021 and 0.001) (Fig. 2 A and B and Table 1). Similarly, this finding holds true for patients stratified by low CNA burden threshold (1.34% CNA) with HRs of 2.03 and 2.42 for the initial and contemporary cohorts, respectively (P values = 0.048 and 0.037, respectively) (Fig. S2). Patients in the initial cohort, with a low CNA burden (<1.34% CNA), had a 13% risk of recurrence within 5 y, compared with a 29% risk of BCR for those with a higher CNA burden (≥1.34% CNA) estimated by Kaplan–Meier. Similarly, patients in the contemporary cohort with a low CNA burden had a 15% risk of BCR, compared with a 38% risk for those with a higher CNA burden (≥1.34% CNA). If cases were dichotomized on the basis of their cohort’s CNA burden quartiles rather than the independently defined CNA cluster medians, CNA burden is still significantly associated with BCR in both cohorts and with metastasis in the initial cohort (insufficient metastatic events in the contemporary cohort) by Cox proportional hazards regression (Table S2). For groups with CNA burden above and below the 75th percentile of their cohort, for example, CNA burden is associated with BCR with HRs of 2.65 and 4.43, respectively, for the initial and validation cohorts (P values = 0.001 and < 0.001) (Fig. S3 and Table S2).

The longer clinical follow-up now available for the initial cohort allowed us to analyze—to our knowledge for the first time—the association between CNA burden and metastasis as a definitive endpoint (Fig. 2C and Table 1). CNA burden was significantly associated with metastasis in the initial cohort as a continuous variable in a univariate Cox proportional hazards regression model (HR = 1.10; P value = 0.0004) (Table 1) and neared significance as a stratified variable (Cox proportional hazards regression P value = 0.0524, Mantel–Cox log-rank P value = 0.044) (Table 1 and Fig. 2C). Because of the shorter available follow-up time, the contemporary cohort possessed too few metastatic events (n = 3) for this analysis.

We next assessed whether CNA burden was prognostic in multivariate models when combined with established clinical variables, such as Gleason score and nomogram score. Gleason score, together with PSA level and clinical stage, is used in the preoperative setting to guide whether to treat or conservatively monitor an individual. CNA burden was significantly associated with BCR in both cohorts when corrected for pretreatment PSA (HR = 1.09 and 1.04 for each percent change in CNA burden, respectively, for initial and contemporary cohorts; P values < 0.001 and 0.025) (Table 2). When adjusted for biopsy Gleason score in multivariate models, CNA burden also remained associated with BCR in both cohorts (HR = 1.08 and 1.05, respectively, for initial and contemporary cohorts; P values < 0.0001 and 0.026) (Table 2). Moreover, when biopsy Gleason score categories included separate Gleason 3+4 and 4+3 groups, CNA burden was still associated with BCR in multivariate models with both cohorts (HR = 1.09, P value <0.001, for the initial cohort; HR = 1.05, P value = 0.041, for the contemporary cohort). This result was also true for metastasis in the initial cohort (HR = 1.07, P value = 0.036) where there are sufficient events for analysis.

We next examined the prognostic significance of CNA burden when corrected for postoperative variables, such as pathology Gleason score and the Stephenson nomogram. CNA burden was associated with BCR in both cohorts after correcting for pathology Gleason score (HR = 1.04 and 1.05 per 1% change in CNA burden, respectively, for initial and contemporary cohorts; P values = 0.043 and 0.026) (Table 2). In contrast, the postoperative Stephenson nomogram relies on additional postoperative variables, such as extracapsular extension, lymph node involvement, surgical margins, and treatment year to assess postoperative recurrence risk (21). Although the addition of CNA burden did not result in an increase in the concordance index from 0.812 and 0.817 in the initial or contemporary cohorts, respectively, CNA burden was still associated with BCR in the contemporary cohort when combined with the Stephenson 5-y nomogram (HR = 1.05, P value = 0.011) (Table 2), though not in the initial cohort.

We then investigated whether the association of CNA burden with cancer relapse can be attributed to specific focal CNAs that are known to be prognostic. In prior analysis of the initial cohort, we tested all focal CNAs and found few to be individually associated with BCR (14). PTEN loss and MYC gain are the principal focal CNAs that have consistently been found to be associated with prostate cancer recurrence, especially when combined with other clinical variables (24). We therefore examined the prognostic significance of CNA burden when combined with PTEN loss, MYC gain, and TP53 loss in the initial cohort, which had sufficient recurrence and metastatic events for analysis. PTEN copy number loss was positively associated with disease relapse in our cohort, whereas TP53 loss and MYC gain were not. CNA burden remained significantly associated with BCR [P value = 0.001, HR = 1.09, 95% confidence interval (CI) 1.05–1.13, for each percent CNA burden as a continuous variable] when adjusted for PTEN copy number loss, which is independently associated with BCR (P value = 0.001, HR = 3.04, 95% CI 1.55–6.00 in multivariate analysis with CNA burden as a continuous variable). CNA burden remained significantly associated with BCR (P value = 0.001, HR = 1.07, 95% CI 1.03–1.11, for each percent CNA burden as a continuous variable) after additional correction for biopsy Gleason score in addition to PTEN loss, but not after correction for pathology Gleason score. PTEN copy number loss was not significantly associated with metastasis. CNA burden therefore captures information conveyed by CNAs beyond that of these major focal CNAs.

Gleason 7 tumors fall in an intermediate-risk category, and there is a broad need for new approaches to discriminate indolent from aggressive cancers in this group. We found a wide range of CNA burden across Gleason 7 cancers in both cohorts with ranges of 0.05–25% in the initial cohort and 0.003–50% in the contemporary cohort (Fig. S4). In Gleason 7 tumors, we found that CNA burden was significantly associated with BCR in both the initial and contemporary cohorts (HR = 1.08 and 1.06, respectively; P values = 0.011 and 0.017, respectively) (Fig. 3 and Table 3). CNA burden remained significantly associated with BCR in Gleason 7 disease after correcting for pretreatment PSA in the initial and contemporary cohorts (P values = 0.006 and 0.019, respectively) (Table 3). Moreover, CNA burden was associated with BCR after adjusting for the Stephenson 5-y nomogram in the contemporary cohort (P value = 0.014) (Table 3) and approached significance in the initial cohort (P value = 0.062) (Table 3), although it did not improve the concordance index above 0.7 achieved with the nomogram alone. Too few metastases were available in the Gleason 7 subpopulation for analysis.

Recent work has established the prognostic significance of gene expression-based signatures in prostate cancers, including a cell cycle progression (CCP) signature recently developed for clinical use (8). We compared a nonproprietary version of the RNA-based CCP signature to the DNA-based CNA burden within our initial cohort, for which both DNA copy number and RNA expression data had been generated. Both the CCP signature and CNA burden were independently associated with BCR, with HRs of 1.124 and 1.097, respectively (P value = 0.05 and <0.001, initial cohort cases with both expression and CNA data, n = 113) (Table S3). CNA burden therefore contributes independent information beyond that represented by the CCP expression signature with a comparable HR for each percent CNA burden or percent CCP signature score. CNA burden outperformed the CCP signature for association with BCR in Gleason 7 cancers (Table S3).

To have utility in clinical risk prediction, CNA burden must be measurable at biopsy. We therefore tested the feasibility of assessing CNA burden in FFPE biopsy samples by whole-genome sequencing. We performed proof-of-principle low-pass whole-genome sequencing (between one- to threefold coverage) with varying low-input quantities of FFPE biopsy DNA from four tumors and their adjacent benign controls (Fig. 4 and Dataset S5). Concurrently, we also analyzed these samples by high-resolution array comparative genomic hybridization (aCGH) for comparison (Fig. 4). The pattern and extent of CNAs identified were consistent between aCGH and whole-genome sequencing in even the challenging low-input FFPE DNA setting, spanning from concordant focal deletions (3p13 spanning FOXP1, RYBP, and SHQ1) to arm-level gains and losses of significance to prostate cancer biology (7, 8p, 8q, and 13q). In addition to concordance among discrete CNAs, the amplitudes—and therefore their clonal representation—was also highly consistent (Fig. 4). These results illustrate that whole-genome sequencing can achieve copy number profiling comparable to aCGH and enables characterization of lower tumor burden FFPE biopsies.

A total of 104 tumor and matched normal samples were obtained from patients treated by prostatectomy at the Memorial Sloan–Kettering Cancer Center with Institutional Review Board approval. BCR was defined as an increase in PSA of ≥0.2 ng/mL on two occasions. The 2005 Stephenson nomogram for postoperative risk of BCR at 5 y was reported (21). Patient follow-up data were updated through March 2013.

Snap-frozen samples containing greater than 70% tumor cell content were dissected from their frozen blocks. Tumor and pooled reference (Promega) DNA was extracted, labeled, and analyzed by aCGH. Agilent 1M feature arrays were used for the contemporary cohort and the 244K arrays for biopsy cases.

Whole-genome sequencing libraries were prepared using 100-ng or 250-ng input FFPE biopsy dsDNA (picogreen) from four patients. Paired-end 100-bp libraries were generated (KAPA LTP kit, Kappa Biosystems) and sequenced by HiSEq. 2000.

For aCGH data, all copy-number array data from patients in the contemporary cohort were quantified, normalized, segmented, and analyzed with RAE, as previously described (14). To assess the percent of autosomal genome affected by CNA in both the initial and contemporary cohorts, the per-tumor parameterization by RAE was used (23). For the analyses described here, CNAs are those segments whose value of A₀ or D₀ (gains and losses or more, respectively) were greater than or equal to 0.75. The total genomic territory spanned by these contiguous segments was summed and a percent was generated using the size of the autosomal human genome.

For sequencing data, reads were aligned to hg19 of the human genome with BWA. The aligned first read-in pair was retained and coverage calculated in 100-kb nonoverlapping windows of the autosomal genome. In each genome, a locally weighted polynomial regression was fit between read coverage and G+C% content in alignable regions not overlapping known gaps in the reference assembly. G+C%-content normalization using the loess fit to normalize read counts in each independent tumor and matched normal genome before segmentation. A pseudocount was added to read counts genome-wide in both samples, the ratio was normalized by library size, and segmented with circular binary segmentation. Individual samples were then analyzed with RAE to determine CAN, as described above.

For cohort characteristics, P values were determined by Wilcoxon rank sum for continuous variables and by Fisher’s exact test for categorical variables. We tested for a univariate association between CNA and BCR or metastasis in the initial and contemporary cohorts separately. CNA was then added to a multivariable Cox proportional hazards-regression model for the risk of BCR. The proportional hazards model was not checked, as formal testing of the proportional hazards assumption is of limited utility; it is known that for outcomes, such as BCR or metastases after treatment, that the proportional hazards assumption is not met with long follow-up time. For Kaplan–Meier analyses, follow-up time was censored to 14.5 y and the Mantel–Cox log-rank significance reported. For outcome analyses performed using a stratification of cases on the basis of their CNA burden, we grouped case cut-points of 5.41% and 1.34%, which represent the mean CNA burden from patient genomes that defined the intermediate and low alteration clusters reported previously (14) (clusters 1, 3, and 4 vs. 2, respectively). To assess the discriminative accuracy of the models developed, we calculated Harrell’s C concordance statistic after 10-fold cross-validation.

The nonproprietary CCP signature score was defined as the average of the expression level (Affymetrix) of the 30 CCP genes with available expression data in our cohort.

Study array data were deposited in the National Center for Biotechnology Information Gene Expression Omnibus under accession no. GSE54691. The data and annotation also available via the Prostate Cancer Genomics Data Portal: http://cbio.mskcc.org/prostate-portal.