Neoadjuvant-Intensive Androgen Deprivation Therapy Selects for Prostate Tumor Foci with Diverse Subclonal Oncogenic Alterations

The androgen receptor (AR) plays a central role in prostate cancer, and most patients with metastatic prostate cancer respond initially to androgen deprivation therapy (ADT), but they invariably relapse despite castrate systemic androgen levels (castration-resistant prostate cancer, CRPC) with tumors that express high levels of AR and AR-regulated genes. Increased intratumoral androgen synthesis is a major mechanism driving AR in CRPC (1–5), and responses can be obtained by further suppression of androgen synthesis using agents such as abiraterone that block CYP17A1 or by more effective AR antagonists such as enzalutamide. Based on the hypothesis that early use of ADT may improve outcomes, previous clinical trials have examined ADT with a GnRH agonist prior to radical prostatectomy (RP) in men with clinically localized disease. Although this neoadjuvant GnRH agonist therapy was generally found to reduce tumor volume, complete pathologic responses were rare and there was no clear improvement in survival (6). Significantly, although treatment with a GnRH agonist decreases serum androgens, substantial levels persist in the prostate due to increased conversion of adrenal-derived DHEA to testosterone and DHT by normal prostate as well as by tumor cells (7, 8).

Based on these observations, we hypothesized that efficacy in previous neoadjuvant ADT trials may have been compromised due to inadequate suppression of intraprostatic androgens, and that responses may be improved by more robust suppression of androgen synthesis. To test this hypothesis, we conducted a phase II trial of neoadjuvant leuprolide for 24 weeks in combination with abiraterone acetate and prednisone (referred to subsequently as leuprolide plus abiraterone) for 12 or 24 weeks prior to RP. As reported recently, we confirmed that the addition of abiraterone further markedly reduced intraprostatic androgen levels and appeared to improve responses relative to historic controls using single-agent GnRH agonists (9). Nonetheless, residual prostate cancer was found in the majority of patients, with only a small number of patients demonstrating complete pathologic responses. Moreover, substantial nuclear and cytoplasmic AR expression was detected by immunohistochemistry (IHC) in most cases, suggesting that AR activity still may persist and contribute to residual disease. In this study, we have exploited RP specimens containing residual tumor from this trial in order to identify possible mechanisms of resistance to the combination leuprolide–abiraterone therapy.

RP tissue of trial cases was acquired as part of our clinical trial (NCT00924469, DF-HCC study 09-107). The studies were conducted in accordance with recognized ethical guidelines (U.S. Common Rule), were approved by the DF/HCC Institutional Review Board (IRB), and written informed consent from the patients was obtained. For matched untreated controls, tissue from RPs performed between 2010 and 2016 at Beth Israel Deaconess Medical Center (BIDMC) under standard of care was collected and deidentified in accordance with BIDMC IRB protocol #2010-P-000254. All trial patients received 24 weeks of leuprolide prior to RP, and those patients received either 12 weeks or 24 weeks of simultaneous treatment with abiraterone acetate plus prednisone leading up to surgery. Control cases matched the trial cases on pre-RP characteristics including ethnicity, serum prostate-specific antigen (PSA) level, biopsy tumor grade, number of positive cores, and maximum cancer percentage on positive cores. Clinicopathologic characteristics of the two cohorts are listed in Table 1. Tumor was graded using both new Grade Group system and Gleason Score as recommended by the 2014 International Urological Pathology consensus (10). Prostate tissue was fixed in formalin, processed, and embedded in paraffin using standard methods. For each case, one tissue block with the largest dimensions of the index tumor was selected for IHC and ISH.

IHC stains were scored semiquantitatively as follows. Ki-67 IHC score was defined as the highest Ki-67 percentage score per high power field across the tumor sections. ERG was scored as 0 (negative) and 1 (positive). PTEN loss scores were derived as follows: (1) PTEN binary score was scored as 0 (retained, or reduced staining, or negative staining in <5% tumor cells) and 1 (negative staining in ≥5% tumor cells); (2) percentage score was assigned based on percentage of tumor cells with negative staining as 0 (no negative staining), 1 (1%–9 %), 2 (10%–49%), and 3 (≥ 50%); (3) final PTEN loss IHC score (0–3) was derived by the binary score multiplied by the percentage score. pAkt Ser473 was scored as 0 (negative), 1 (weak), 2 (moderate), and 3 (strong), based on the most predominant intensity pattern. Synaptophysin and chromogranin were scored as 0 (negative), 1 (<1%–4%), 2 (5%–24%), 3 (25%–49%), and 4 (≥50%). For AR, AR-V7, PSA, and NKX3.1, IHC scores were derived as follows: (1) immunointensity score was assigned as 0 (negative), 1 (weak), 2 (moderate), and 3 (strong), based on the most predominant intensity pattern; (2) percentage score was assigned based on percentage of tumor cells demonstrating the most predominant intensity pattern as 0 (negative), 1 (1%–9 %), 2 (10%–49%), and 3 (≥ 50%); (3) final IHC score (0–9) was derived by the immunointensity score multiplied by the percentage score. All immunostains were evaluated by a board-certified pathologist (H. Ye) and a trained physician investigator (L. Montaser-Kouhsari or C. Calagua).

Stata 12.1 (StataCorp) was used for statistical analysis. Characteristics between groups were compared using Pearson χ² and Fisher exact tests for categorical variables, and Wilcoxon rank-sum test for nonparametric continuous and ordinal variables. A P < 0.05 was considered statistically significant.

Samples were stained with hematoxylin and eosin and PIN-4, and serial sections were stained with Paradise Plus stain and laser capture microdissected (LCM) as previously described (11). Regions of residual or untreated tumor (or accompanying regions not involved with tumor) were identified by a board-certified pathologist (H. Ye or R.T. Lis), and images of capture tissue were verified prior to extraction. RNA and DNA were extracted simultaneously using the AllPrep DNA/RNA FFPE Kit (Qiagen) with modifications as described previously (12). Fifty ng of total RNA from LCM tissue was amplified and prepared into sense-strand cDNA as previously described (12). Labeled libraries were hybridized onto Affymetrix Human Transcriptome Array 2.0 GeneChips. In addition to foci of residual tumor, three foci of tumor from untreated cases were microdissected as control cases and used in conjunction with microarray data from 26 foci of similarly-processed samples (12) for a total of 29 control samples.

A total of 10 to 100 ng of genomic DNA from LCM tissue was prepared into libraries using the HyperPrep Kit (KAPA) modified to include reagents from the Agilent SureSelectXT2 Kit, to enable low-input amounts of formalin-fixed and paraffin-embedded (FFPE)–derived DNA. Following sonication (Covaris), clean-up, and end-repair, partial SureSelect Agilent linkers were ligated to the DNA at a molar ratio of 10:1. PCR was performed until 1 μg of precapture library was generated. Hybridization to the SureSelect Exome V5 was performed in solution for 48 hours, and additional amplification was performed to generate at least 10 μL of a 10 nmol/L library. Libraries were quantified with the KAPA Illumina QC Kit, pooled in equimolar ratios, and sequenced on an Illumina HiSeq2500. Alignment, deduplication, local realignment, and quality score recalibration were performed with Burrows-Wheeler Aligner, PICARD, and Genome Atlas Toolkit. Mutation calling and annotation were performed using muTect2 and Oncotator (13, 14), and mutations were confirmed by VarScan (15).

Sequencing and microarray raw data were deposited into dbGaP, accession phs001399.v1.p1. Gene expression data were deposited into GEO, accession GSE102124. Methods for machine-aided IHC quantification, criteria for calling genomic alterations, and criteria for calling differentially expressed genes and pathways are provided in Supplementary Data.

As our study (NCT00924469) recently reported (9), although neoadjuvant leuprolide plus abiraterone substantially reduced tumor volumes, 52 (93%) of 56 trial patients still had foci of residual tumor detected in their RP specimens. Among these 52 patients, residual tumors had a median cross-sectional greatest dimension of 1.5 cm [interquartile range (IQR), 1.0–2.5 cm] and median residual cancer burden (RCB) of 0.17 cm³ (IQR, 0.02–0.86 cm³). Residual tumors measuring >0.5 cm in the greatest dimension were found in 48 (86%) of patients. In most cases, the residual tumors were pathologically acinar adenocarcinomas, not otherwise specified (NOS; Supplementary Fig. S1A), except two cases that contained a component of adenocarcinoma with Paneth cell–like neuroendocrine differentiation (NED; Supplementary Fig. S1B). These two cases had tumor sizes of 2.5 and 1.5 cm in the greatest dimension and had RCB of 1.48 and 0.19 cm³, respectively. The tumor components with Paneth cell–like NED expressed neuroendocrine markers including synaptophsyin and chromogranin, with reduced expression of AR and AR targets (NKX3.1 and PSA) in comparison with the coexisting component of adenocarcinoma, NOS (Supplementary Fig. S1B).

For the 49 trial cases that had available tissue for further IHC characterization, we examined IHC profiles of residual tumors to assess cell proliferation and AR activity. Adequate tissue from baseline core biopsies was not available in most cases, so trial cases were compared with 49 untreated control cases with matched baseline clinicopathologic characteristics (Supplementary Table S1). Cell proliferation rate, measured by the Ki-67 index on IHC (Fig. 1A), was significantly reduced (P < 0.0001, Wilcoxon rank-sum test) in treated tumors compared with untreated controls (Fig. 1B). Among the treated cases, the 4 cases with the highest Ki-67 indices were in the 24-week abiraterone arm (cases 513, 601, 609, and 006; Fig. 1C). In some cases, spatially separated tumor foci of the same index tumor demonstrated markedly diverse Ki-67 proliferation indices (such as case 513, depicted in Fig. 1A). AR expression with variable nuclear and cytoplasmic levels was readily detected in 48 (98%) of residual tumors (Fig. 1D; see also Supplementary Fig. S1B). However, nuclear AR was significantly reduced compared with untreated controls, consistent with effects of the intense ADT (P = 0.0175, Wilcoxon rank-sum test). Nuclear AR IHC scores in the 24-week abiraterone arm showed a trend toward reduction compared with that of the 12-week arm (Fig. 1D).

Recent studies indicate that the emergence of CRPC and resistance to abiraterone and enzalutamide are associated with increased expression of AR splice variants that delete the ligand binding domain, with AR-V7 being the most commonly expressed variant (16). ISH with probes against the AR-V7 mRNA showed focal increases in AR-V7 expression, which were concordant with increases in total AR mRNA as indicated by a probe recognizing a region (nucleotides 1786–2998) shared by the full-length AR and AR variants (Fig. 1E). Immunostaining with an AR-V7 antibody similarly demonstrated focal areas of high AR-V7 expression in the treated tumors (Fig. 1F). However, focal staining with the AR-V7 antibody was also observed in the control untreated tumors, and although there was a trend toward increased AR-V7 by IHC in the treated versus control untreated tumors, the increase was not significant (Fig. 1G).

Expression of the tightly AR-regulated PSA (KLK3) protein was significantly reduced in the treated tumors, suggesting decreased AR activity (P = 0.0302, Wilcoxon rank-sum test; Fig. 1H). In contrast, expression of the AR-regulated protein NKX3.1 was upregulated in treated cases (P < 0.0001, Wilcoxon rank-sum test; Fig. 1I). ERG immunostaining to detect expression of the AR-regulated TMPRSS2:ERG fusion gene also was positive in comparable numbers of treated versus control tumors (Supplementary Table S1), although treatment was associated with reduced ERG intensity in treated fusion-positive tumors (consistent with lower AR activity). Overall, these findings indicated that reduced, but still substantial, AR activity persisted in the residual tumor foci.

We reported previously that phosphorylation of ErbB3 was increased in a subset of these residual tumor foci, suggesting that increased ErbB2 activity was a mechanism contributing to resistance (17). To explore further oncogenic mechanisms of resistance, we carried out IHC for additional proteins or pathways known for their enrichment in metastatic CRPC (18). PTEN loss was observed in 47% of residual tumors, as defined by negative PTEN staining in >5% of tumor cells (Fig. 2A; Table 1). In 20% of residual tumors, this PTEN loss was identified in 5% to 50% of tumor cells, indicating PTEN loss was subclonal in those cases. Not surprisingly, PTEN loss co-occurred with increased phosphorylation of Akt (Ser473), an indicator of PI 3-kinase pathway activation (Fig. 2A–C). PTEN loss also was positively associated with Ki-67 proliferation index (Fig. 2D). Compared with untreated tumors, however, treated tumors did not show a significant increase in PTEN loss, either by proportion of cases with PTEN loss or by proportion of tumor cells with PTEN loss (see Table 1). Tumor foci with pRb loss were observed in 3 of 44 (7%) residual tumors examined, with all three cases demonstrating focal/subclonal pRb loss (Fig. 2E). Of these three cases (cases 505, 513, and 654), one had an outlier high Ki-67 proliferation index on IHC (case 513, Ki-67 index 50%, see Fig. 1B and C).

As long-term ADT promotes NED in metastatic CRPC, we also examined the influence of the neoadjuvant ADT on NED in tumors that were morphologically acinar adenocarcinomas (as noted above, two cases contained a component with Paneth cell–like NED that were synaptophysin positive, see Supplementary Fig. S1B). Synaptophysin expression in scattered individual cells or foci-specific groups of cells was present not only in treated tumors, but also in matched untreated controls (Supplementary Fig. S1C). There was no significant difference in synaptophysin or chromogranin IHC scores between treated and untreated tumors (Pearson χ² test). Further, we compared the proportions of synaptophysin-positive cells in treated and untreated tumors in all cases that demonstrated ≥5% synaptophysin immunoreactivity by performing computational quantitative analysis on digitalized synaptophysin-stained slides. Consistent with a treatment effect, the results showed more synaptophysin-positive cells in untreated cases, but there was no significant enrichment in treated residual tumors compared with untreated controls as a proportion of tumor cells present (Supplementary Fig. S1D and S1E).

Taken as a group, these studies indicate that residual tumors had significantly lower Ki-67 and reduced expression of nuclear AR and AR target genes, but no significant increases in AR-V7 expression, PTEN loss, or markers of NED. However, there was substantial intratumoral heterogeneity in residual tumors, as indicated by distinct Ki-67 proliferation indices among foci, focal increases in AR-V7, subclonal losses of PTEN, focal losses of pRb, and foci-specific NED in subsets of cases. These findings indicated that spatially separated tumor foci may harbor different drivers of resistance.

Together, the above results suggested that diverse mechanisms may be contributing to resistance in individual tumors, especially in cases where there were morphologically or immunohistochemically distinct residual tumor foci. To further address mechanisms of resistance, in a series of trial cases, we used LCM to purify residual tumor in one to three discrete tumor foci from the FFPE RP specimens, in conjunction with a distant area that did not contain tumor. Clinicopathologic characteristics of the samples that were microdissected are reported in Supplementary Table S1.

Both RNA and DNA were then extracted from each microdissected tumor focus and corresponding uninvolved benign focus for transcriptome analysis and whole-exome sequencing (WES, see below). In cases where adequate RNA was obtained from the microdissected tumor foci, global gene expression was examined on GeneChip Human Transcriptome 2.0 arrays (Affymetrix) and compared with RNA extracted and analyzed previously from similarly microdissected Gleason score 7–9 (with coexisting Gleason pattern 3 and cribriform Gleason pattern 4) control untreated tumors (12). Unsupervised hierarchical clustering based on the 1,000 genes with the most variance separated the treated from the control untreated tumor foci (Fig. 3A). Interestingly, in some cases, tumor foci from the same case did not cluster together, consistent with heterogeneity noted by IHC above.

We next used the gene expression data to assess for proliferation and AR signaling in the residual tumor foci. The proliferation rate, based on a 131-gene index (19), indicated that proliferation was decreased in the residual tumors, consistent with the Ki-67 IHC data (Fig. 3B). Also consistent with the Ki-67 IHC, this proliferation index was significantly lower in the cases treated for 12 weeks versus 24 weeks with abiraterone (Fig. 3C). Three of the 5 foci that were highest for the 131-gene proliferation single-sample gene set enrichment analysis (ssGSEA) score (513.D16, 513.D15, and 006.C21) were from cases with outlier high Ki-67 indices. Two of the four foci with the highest proliferation rate based on ssGSEA had outlier low levels of RB1 expression based on the transcriptome data (505.C20, 507.C16), suggesting that decreased RB1 expression may be a driver of increased proliferation in these residual tumors. Indeed, there was a significant negative correlation (r = –0.56) between RB1 mRNA levels and proliferation index, suggesting that RB1 downregulation was a major factor driving proliferation (Fig. 3D).

To assess AR activity, we applied ssGSEA with a panel of 267 androgen-regulated genes (20). AR activity overall was lower in the residual tumor foci versus control untreated tumors (Fig. 3E), consistent with the IHC results. However, it was still substantial, being equivalent to levels in metastatic CRPC and greater than in a panel of neuroendocrine-like prostate cancers (Fig. 3F). AR activity was lower after 24 weeks versus 12 weeks of abiraterone (Fig. 3G). Surprisingly, we observed a negative correlation (r = –0.54) between the ssGSEA scores for AR activity and proliferation (Fig. 3H). To determine whether this was unique to the neoadjuvant setting, we similarly examined other data sets of primary prostate cancer and metastatic CRPC. In the prostate The Cancer Genome Atlas (TCGA) and MSKCC 2010 primary prostate cancer datasets, there were no significant correlations between the AR and proliferation signatures (Supplementary Fig. S2A and S2B; refs. 21, 22). Similarly, in the Stanbrough and Stand Up to Cancer metastatic CRPC datasets, there were no statistically-significant correlations (Supplementary Fig. S2C and S2D; refs. 1, 18), whereas there was a negative correlation (as noted previously) in the Fred Hutchinson Cancer Research Center dataset (Supplementary Fig. S2E; ref. 23).

Together, these observations indicated that although a basal level of AR activity may be necessary, further increases do not drive proliferation in the neoadjuvant or metastatic setting. Consistent with this conclusion, and with decreased RB1 expression being a driver of proliferation, there was a positive correlation (r = 0.67) between the AR activity ssGSEA score and RB1 expression (Fig. 3I). We further found that decreased RB1 expression was associated with increased expression of genes that are negatively regulated by pRb, based on an 80-gene RB1 loss signature derived from CRPC, although the correlation did not reach statistical significance (Fig. 3J; ref. 24). Finally, although RB1 loss has been associated with NED in metastatic CRPC (25), we did not observe any substantial negative correlation between expression of RB1 and CHGA (r = –0.12), CHGB (r = –0.12), or SYP (r = –0.17). Moreover, RB1 expression was not decreased in the two tumor foci microdissected from a case with Paneth cell–like NED (case 004, log₂ array intensity of RB1 mRNA = 7.77 and 7.95).

We next broadly examined gene expression in the residual tumor foci versus control untreated tumors. As shown in the Volcano plot, only a relatively small set of genes were significantly overexpressed in the residual tumor foci (Fig. 4A). Genes that were increased by at least 2-fold are shown in Supplementary Table S2. Among these, the lncRNA SCHLAP1 has been implicated in prostate cancer and is associated with more aggressive disease (26). Alterations in AR coactivators, corepressors, or other modulators may also be contributing to reconstitution of AR activity in some cases. However, none of the genes in Supplementary Table S2 are established AR coactivators. We also examined expression of a series of genes implicated previously as AR coregulators in CRPC (27), and did not identify any that were significantly increased in the treated versus control tumors (Supplementary Table S3).

Interestingly, although AR mRNA is markedly increased in metastatic CRPC with frequent AR gene amplification (1, 28), it was not consistently increased in the residual tumors. However, there was a positive correlation (r = 0.55) between AR mRNA and AR activity ssGSEA score (which itself does not include AR expression), indicating that AR level was contributing to residual AR activity in treated cases (Supplementary Fig. S3A). Genes involved in androgen synthesis, and particularly AKR1C3, are also increased in metastatic CRPC (1, 2). However, expression of AKR1C3 had no association with AR activity (Supplementary Fig. S3B), and expression of CYP17A1 (encoding the enzyme targeted by abiraterone) was negatively correlated with AR activity (Supplementary Fig. S3C).

Interestingly, NR3C1 (glucocorticoid receptor) expression, which may drive a subset of AR-regulated genes in advanced CRPC (29), was positively correlated (r = 0.52) with AR activity (Supplementary Fig. S3D). Nuclear GR expression by IHC also was generally significantly increased in the treated tumors compared with untreated controls (P = 0.0216, Wilcoxon rank-sum test), with no difference between the 24-week and 12-week arms (Supplementary Fig. S3E). Consistent with increased GR protein expression, a comprehensive 121-gene GR activity signature, which has considerable overlap with AR target genes (29), was increased in the treated versus untreated cases, although the difference was not statistically significant (P = 0.0709, Supplementary Fig. S3F). Moreover, when only the 67 genes that did not overlap with AR activity were retested, the treated cases had significantly greater enrichment for GR activity (P < 0.0001, Supplementary Fig. S3G). Together, these findings indicate that GR activity is generally increased in the treated tumors, particularly toward GR-specific genes and also contributes to the expression of AR-regulated genes.

In contrast to the relatively small number of genes that were increased in the residual tumors, a much larger set of genes were expressed at higher levels in the control tumors, with approximately 300 genes being at least 2-fold higher in the control versus treated cases (Supplementary Table S4). As expected, many of these are AR-regulated genes. Pathway analysis indicated that the most significant alterations in the residual tumors were decreases in genes involved in the endoplasmic reticulum stress response, the unfolded protein responses, and lipid biosynthesis (Fig. 4B). This finding is consistent with decreased AR transcriptional activity, as a major function of AR is to drive protein and lipid synthesis (30–32). Assessment for upstream regulators of these gene expression changes confirmed a central role for decreased AR activity (Fig. 4C). It also indicated that there was decreased activity of XBP1 (a positive regulator of the ER-unfolded protein response; refs. 33, 34) and NFE2L2 (activated in response to oxidative stress; ref. 35), consistent with decreased endoplasmic reticulum stress (Fig. 4C). Interestingly, although XBP1 activity primarily is regulated at the level of splicing, XBP1 mRNA levels were positively correlated with AR activity, consistent with AR activity being a major regulator of these metabolic alterations and possibly indirectly increasing XBP1 by increasing metabolic stress (Fig. 4D). Alternatively, the XBP1 gene may be directly regulated by AR in CRPC (36).

As noted above, DNA from microdissected tumor foci was in parallel examined by WES. Somatic copy-number alterations (SCNA) were identified based on segmented WES read depth (Supplementary Fig. S4A–S4O). Consistent with a common clonal origin, many SCNAs were shared between two foci from the same RP specimen (Fig. 5A; Supplementary Table S5). This also held true for case 518, in which we microdissected tumor and sequenced DNA from three foci. However, many others were unique to one focus, indicating they were likely subclonal in the initial tumor prior to therapy but were clonal in the individual tumor focus.

We next used somatic copy number, allele frequency, and estimated LCM purity to infer which mutations and small indels were likely to be clonal in each focus, and to identify those that were shared versus unique in foci from the same RP specimen. In case 518 where we dissected three foci, as well as the cases in which two tumor foci were isolated, we identified multiple shared somatic mutations that we presume to be truncal or present on a large branch of the tumor, confirming a common clonal origin (see Supplementary Table S6). Significantly, in those cases with 2 or 3 microdissected foci, we also identified many somatic mutations or small indels that were present in high frequency in one focus, but were not identified in the other focus (Supplementary Table S7). As with the SCNAs, these were presumably subclonal in the untreated tumor and reflect focal expansion of the corresponding subclones. Finally, in three cases, we microdissected and analyzed only a single focus of tumor and identified high-confidence mutations and indels with potential driver status, which may have been truncal or subclonal in the untreated tumor (see Supplementary Table S8).

We had reported previously that increased progesterone in men with metastatic CRPC treated with abiraterone could select for the T878A-mutant AR that can be strongly activated by progesterone and identified the T878A mutation in one of two foci from the initial patient analyzed from this trial (37). Among the 18 cases examined here, we found the T878A mutation in only 1 additional patient (case 006, wherein only one focus was analyzed). We also found an AR gain in one of two foci from case 506 (focus C11). This apparent alteration rate for AR of approximately 10% is far less than the reported rate of approximately 60% in mCRPC (18), possibly reflecting less of a growth advantage for tumor cells with these AR alterations with upfront use of aggressive ADT. Alternatively, as there would presumably be no selective advantage for cells with AR gene amplification or mutations prior to ADT, their precursor frequency would be very low, and it may take a more prolonged period of ADT for these cells to emerge and become dominant clones.

To identify further potential driver alterations to tumor-suppressor genes or oncogenes, we next filtered genomic events to large chromosomal changes and a refined list of 727 cancer-related genes. The results are summarized for six cases in Fig. 5B, and the remaining 9 cases for which 2 foci were analyzed are depicted in Supplementary Fig. S5. For each Euler diagram, the size of the circles for each focus reflects the total number of somatic mutations, small indels, and copy-number alterations that were identified in each focus, as shown in Supplementary Table S9. Overall, genomic alterations as analyzed in these residual tumor foci presented a pattern of sequential tumor-suppressor losses, consistent with spatially distinct foci emerging from subclones with unique oncogenic alterations that were cooperating with truncal events acquired earlier in a common precursor.

Foci from several cases showed discordance for PTEN loss. Three cases (505, 512, and 518) had deletions of one PTEN allele in only one focus, and one case (520) had independent single-copy PTEN losses (Fig. 5B; Supplementary Fig. S5). Case 003 had a shared deletion of one PTEN allele, which likely emerged earlier during tumorigenesis, and only focus C12f2 from this case lost the second PTEN allele (Supplementary Fig. S5). In addition, the C12f2 focus of case 003 also harbored single copy losses of TP53, CHD1, and BRCA1 that were not detected in 003 focus C12f1. In case 516, one focus (C7) exhibited homozygous losses of PTEN and ZBTB16 (PLZF), whereas both genes were intact in the other focus (C12; Fig. 5B). It should be noted that although both foci from this case shared alterations, a chromosome 8p loss and 8q gain (encompassing NKX3.1 and MYC, respectively), we did not identify any shared mutations or additional shared SCNAs between these foci, indicating that they may reflect independent tumors. Case 513 had biallelic inactivation of PTEN in both foci examined, consisting of a single copy loss and a splice site mutation on the remaining copy (Fig. 5B). Case 513 also harbored a homozygous deletion of RB1 in both foci. Consistent with this being a truncal loss, case 513 had the highest level of proliferation based on Ki-67 expression and the greatest residual tumor volume.

In addition to case 513, RB1 losses were found in 6 other cases (004, 006, 501, 505, 508, and 518), and these losses were discordant among foci in 3 cases (505, 508, and 518). Case 508 had a single copy RB1 loss only in focus C6. This focus also had a homozygous deletion of CHD1 as well as additional alterations that were not present in the other focus, and had a much greater overall mutational burden than focus C5 (Fig. 5B). Case 518 demonstrated a truncal loss of one copy of RB1 detected in all three foci, but only focus C12 harbored a deletion of the second copy of RB1 (Fig. 5B). All three foci shared gains to chromosome 3q encompassing PIK3CA and PIK3CB. Interestingly, foci C8C9 and C10 shared a frameshift deletion to ATM and a copy-neutral LOH event to chromosome 11, resulting in biallelic inactivation of ATM. Focus C8C9 also harbored a high-level amplification of HSP90AB1, which was not observed in other foci.

In case 505, the foci shared a loss of chromosome 13q (encompassing RB1 and BRCA2), with only focus C20 harboring an additional loss on the other 13q, resulting in homozygous loss of RB1 and BRCA2 (Fig. 5B). Interestingly, biallelic loss of TP53 was observed as a truncal event in this case, as both the C17 and C20 foci harbored single copy losses of chromosome 17p with a shared p.W53* nonsense mutation on the remaining copy. Significantly, although focus C20 was RB1/TP53 null, there was no evidence of NED based on morphology or transcriptome profiling. Conversely, in case 004 where both foci showed NED, there was only a shared single copy loss of RB1 and no detectable alteration in TP53 (Fig. 5B).

Overall, we observed homozygous loss of PTEN that appeared to be truncal in one case (513), and homozygous losses in one focus from cases 003 and 516. For TP53, we observed truncal homozygous loss in one case (505), a truncal single copy loss in case 513, and single copy loss in one focus from case 003. These frequencies for homozygous inactivation of PTEN and TP53 are very similar to the frequencies reported in the primary prostate cancer TCGA dataset (∼17% and 6%, respectively; ref. 21), but less than the frequency reported in the prostate cancer SU2C dataset (∼38% and 31%, respectively; ref. 18). In contrast, although RB1 loss appears to be uncommon in primary prostate cancer (<∼1% in TCGA; ref. 21), we found deep deletions to RB1 in four foci from three cases (505, 513, and 518) and single copy RB1 losses in 9 foci. This enrichment for biallelic inactivation of RB1 in these cases versus untreated primary prostate cancer was statistically significant (Fig. 6A), and was associated with lower expression of RB1 mRNA (Fig. 6B). This finding and the negative correlation between RB1 mRNA and proliferation (see Fig. 3D) suggest that the intensive ADT is selecting for tumor foci that have decreased expression of RB1 mRNA through genetic or epigenetic mechanisms as a major mechanism for bypassing the G₁–S block in response to androgen deprivation.

In addition to these cases with deep deletions of tumor-suppressor genes, we found shallow deletions (indicative of single copy losses) in multiple other foci. For example, case 004 (Paneth cell–like morphology) had shared shallow deletions of APC, CHD1, RAD50, RB1, and CDKN1B, whereas one focus from this case had additional single copy losses of chromosome 8p, MSH2, and MSH6 (which were associated with an increased mutational burden; see Fig. 5B). Although the significance of these single copy losses is less clear, many of these tumor suppressors demonstrate haploinsufficiency, and their loss also may cooperate with other epigenetic mechanisms or complimentary alterations in other genes (such as losses to RB1 and CDKN1B in case 004) to contribute to tumor survival and growth. Overall, these data indicate that the intense ADT selected for discrete foci of tumor that contained genomic alterations that contribute to resistance, which may contribute to the eventual emergence of metastatic disease (Fig. 6C).

This clinical trial was conducted to determine whether more intensive neoadjuvant ADT could be obtained with the addition of abiraterone acetate plus prednisone (referred to herein as abiraterone) to leuprolide, to determine the extent to which this could mediate regression of primary prostate cancer in men with intermediate- to high-risk disease, and to identify mechanisms that may be contributing to therapy resistance. As reported previously, the addition of abiraterone to leuprolide effectively reduced intraprostatic androgens, and a small number of complete or near-complete responses were obtained, but substantial residual tumor was found in most cases (9). Analysis of these residual tumors by histology and IHC showed that most were adenocarcinoma, NOS, and that the extent of NED was not increased relative to control untreated tumors. Consistent with this finding, nuclear AR was expressed at intermediate-to-high levels in most cases, and transcriptome profiling showed that substantial AR transcriptional activity persisted in the residual tumor foci. This AR activity was positively associated with AR and NR3C1 (GR) mRNA levels, but not with expression of AKR1C3 that drives increased intratumoral androgen synthesis from adrenal gland precursors in metastatic CRPC. Distinct from mCRPC, we similarly did not observe increased expression of the AR-V7 splice variant, which correlates with resistance to ADT in metastatic CRPC (16). Although this does not preclude expression of other variants we did not measure, it suggests that increased AR splice variant expression is a longer-term adaptation or there is less selective pressure favoring their expression in the prostate versus at metastatic sites.

Significantly, AR transcriptional activity was not associated with increased proliferation (and was instead negatively correlated), indicating that AR reactivation beyond some basal levels did not provide a clear growth advantage for residual tumor cells, and consistent with the apparent lack of selection for increased intratumoral androgen synthesis. Interestingly, analyses of primary prostate cancer and metastatic CRPC datasets similarly indicate proliferation is not associated with increased AR signaling, with a negative correlation in some cases (23). The somewhat more robust negative correlation in these neoadjuvant cases versus metastatic CRPC may in part reflect more dependence for proliferation on AR-regulated growth factors from prostatic stromal cells, which in turn are also affected by ADT. Moreover, in the absence of growth factors or other factors driving proliferation, the AR in residual tumor cells with higher AR activity may be driving metabolic pathways related to differentiated functions, causing endoplasmic reticulum stress and decreasing proliferation.

In contrast to AR activity, increased proliferation in residual tumors was associated with decreased RB1 mRNA. Prior studies showed that pRb inactivation could circumvent the G₁–S block mediated by androgen deprivation (38) and was associated with progression to CRPC after castration therapy (39), and RB1 loss is emerging as a mechanism of resistance in advanced CRPC after treatment with abiraterone and enzalutamide (18, 40–42). Interestingly, RB1 loss in these latter advanced tumors is associated with NED (25, 43, 44), but was not associated with NED in these neoadjuvant-treated tumors. Significantly, we found deep deletions to RB1 in 4 foci (and shallow deletions in another 9 foci), which is significantly greater than the incidence in the TCGA dataset of primary prostate cancer (21), but comparable with the rate of homozygous RB1 inactivation in mCRPC (∼9%; ref. 18). These deep RB1 deletions we observed were associated with decreased RB1 mRNA, but some foci with relatively low RB1 mRNA did not have apparent RB1 losses. Focal genomic RB1 losses that are not apparent based on segmented WES read depth may be occurring in these latter cases, although epigenetic mechanisms could also be contributing to decreased RB1 expression. In any case, these findings suggest that subclonal RB1 loss may be more common than previously appreciated in intermediate- to high-risk primary prostate cancer and may be an earlier event, independent of NED, in the development of initial resistance to aggressive ADT (see Fig. 6C).

Recent studies from others and us have clearly shown that primary prostate cancer, like many other cancers, can have extensive microheterogeneity (12, 45–49). However, the contribution of this microheterogeneity in the primary tumor to metastatic disease and therapy resistance has not been clear. Significantly, in addition to RB1 losses, we identified multiple established or likely oncogenic alterations that were present in only one of two tumor foci analyzed. This indicates that they were subclonal in the tumor prior to therapy, and that they likely contributed to the persistence and expansion of the corresponding tumor focus. In men who relapse after RP, we suggest that these foci with subclonal oncogenic alterations may be the origin of tumor cells that give rise to metastatic CRPC. This hypothesis is being tested in long-term follow-up studies.

Significantly, as noted above, AR activity was negatively correlated with proliferation and was found to be similarly inversely associated with proliferation in a recent study of advanced mCRPC (23). Nonetheless, it is important to note that substantial AR expression and activity persisted in most foci of residual tumor, and similarly persists in most cases of mCRPC (1, 50). Therefore, although this AR activity is clearly not sufficient to drive proliferation, and higher levels may suppress proliferation, it may nonetheless be necessary for tumor cells to maintain some basal level of AR activity in order to survive. Moreover, AR in metastatic foci may be driving expression of growth factors that are stromal derived in the primary tumors, making the metastatic tumors more dependent on intrinsic AR activity. Hence, our observations are consistent with the recently reported LATITUDE (51) and STAMPEDE (52) studies, in which survival benefits were observed for the earlier treatment of hormone-sensitive metastatic prostate cancer with abiraterone. In the case of localized disease, we hypothesize that more robust responses might still be obtained through therapies that further ablate AR activity, although it is not clear whether even complete AR annihilation would be effective in tumors that have lost RB1. The efficacy of further AR suppression is currently being studied in trials that combine neoadjuvant leuprolide and abiraterone with direct AR antagonists.

E.M. Van Allen reports receiving commercial research grant from Novartis and BMS; has ownership interest (including stock, patents, etc.) in Genome Medical, Syapse, and Tango Therapeutics; and is a consultant/advisory board member for Invitae, Tango Therapeutics, and Genome Medical. B. Montgomery reports receiving commercial research grant from Janssen and Astellas. P.S. Nelson is a consultant/advisory board member for Astellas and Janssen. M.-E. Taplin is a consultant/advisory board member for Janssen. S.P. Balk is a consultant/advisory board member for Sanofi and has expert testimony in Astellas. No potential conflicts of interest were disclosed by the other authors.

Conception and design: A.G. Sowalsky, H. Ye, G.J. Bubley, P.S. Nelson, M.-E. Taplin, S.P. Balk

Development of methodology: A.G. Sowalsky, H. Ye, E.M. Van Allen, M. Loda, F. Ma, O.S. Voznesensky, M.-E. Taplin, S.P. Balk

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): H. Ye, M. Bhasin, E.M. Van Allen, M. Loda, R.T. Lis, J.W. Russo, R.J. Schaefer, O.S. Voznesensky, G.J. Bubley, B. Montgomery, E.A. Mostaghel, P.S. Nelson, M.-E. Taplin, S.P. Balk

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): A.G. Sowalsky, H. Ye, M. Bhasin, E.M. Van Allen, M. Loda, R.T. Lis, L. Montaser-Kouhsari, F. Ma, B. Montgomery, M.-E. Taplin, S.P. Balk

Writing, review, and/or revision of the manuscript: A.G. Sowalsky, H. Ye, M. Bhasin, E.M. Van Allen, M. Loda, G.J. Bubley, B. Montgomery, E.A. Mostaghel, P.S. Nelson, M.-E. Taplin, S.P. Balk

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): R.T. Lis, L. Montaser-Kouhsari, C. Calagua, O.S. Voznesensky, Z. Zhang, M.-E. Taplin, S.P. Balk

Study supervision: H. Ye, M.-E. Taplin, S.P. Balk

The authors gratefully acknowledge the patients and the families of patients who contributed to this study. The authors acknowledge the technical support of the Beth Israel Deaconess Molecular Medicine Core and the Harvard Medical School Biopolymers Facility. Portions of this research were conducted on the Orchestra High Performance Compute Cluster at Harvard Medical School and utilized the computational resources of the NIH HPC Biowulf cluster.

This work was supported by NIH grants (DH/HCC SPORE P50 CA090381 to A.G. Sowalsky, H. Ye, M. Loda, M.-E. Taplin, G.J. Bubley, and S.P. Balk; Pacific Northwest SPORE P50 CA097186 to P.S. Nelson, B. Montgomery, and E.A. Mostaghel; P01 CA163227 to E.A. Mostaghel, P.S. Nelson, and S.P. Balk); Prostate Cancer Foundation Challenge Awards (E.A. Mostaghel, P.S. Nelson, M.-E. Taplin, and S.P. Balk); Prostate Cancer Foundation Young Investigator Awards (A.G. Sowalsky, H. Ye, E.M. Van Allen, and E.A. Mostaghel); Department of Defense Prostate Cancer Research Program (W81XWH-13-1-0267 and W81XWH-16-1-0433 to A.G. Sowalsky; W81XWH-16-1-0431 to S.P. Balk and W81XWH-16-1-0432 to M.-E. Taplin); and the Intramural Research Program of the NIH, National Cancer Institute (A.G. Sowalsky). The original neoadjuvant trial (COU-AA-201) was funded by Janssen Research & Development.