To characterise the subclonal architecture of androgen-deprived metastatic prostate cancer, we performed whole genome sequencing (WGS) of 51 tumours from 10 patients to an average sequencing depth of 55X, including multiple metastases from different anatomic sites in each patient and, in 5 cases, the prostate tumour (Supplementary Table 1). We identified a set of high-confidence substitutions, insertions/deletions, genomic rearrangements and copy number changes present in each tumour sample (Extended Data Figure 1 and Supplementary Information, Section 3). To portray the populations of tumour cells within each patient, we employed an n-dimensional Bayesian Dirichlet process to group clonal and subclonal mutations, i.e. those mutations present in all or a fraction of tumour cells within a sample, respectively. The fraction of tumour cells carrying each mutation was calculated from the mutant allele fraction, taking into account the tumour purity and local copy number state, as described previously^2,11. Each of the mutations assigned to a single cluster is present in a fixed proportion of cells in each sample and hence belongs to a separate subclone, i.e. a genetically distinct population of cells.

By plotting the cancer cell fractions of mutations from pairs of samples, we determined the clonal relationship between the constituent subclones and found evidence for polyclonal seeding of metastases, the most striking example of which is seen in patient A22 (Figure 1). Each of the plots in Figure 1a contains a cluster of mutations at (1,1), indicative of truncal mutations that were present in the most recent common ancestor (MRCA) of both metastases. However, in many of the plots, there are additional clusters at subclonal proportions in both samples plotted. For example, the cluster of mutations indicated by the purple circles in Figure 1a are present in 40% of cells in A22-G, 62% of cells in A22-H, 37% of cells in A22-J and 92% of cells in A22-K. A metastasis seeded by a single cell must carry a set of mutations present in all tumour cells, representing the complement of lesions in that founding cell. In some cases, this set of mutations will be subclonal in the originating site. However, mutation clusters present subclonally in two or more samples can only occur as the result of multiple seeding events by two or more genotypically distinct cells. A graphic illustration of the clonal and subclonal clusters and their representation in all of the 10 samples from A22 is shown in Figure 1b. Where one subclone is present in the same or a lower fraction of cells than a second subclone in all samples, the subclones are represented as nested ovals when required by the pigeonhole principle (Supplementary Information, Section 4b). In contrast, clusters whose relative cancer cell fractions are reversed in different samples represent branching subclones and are shown as disjoint ovals. The full lineage relationship between the subclones can be depicted in the form of a phylogenetic tree whose branch lengths are proportional to the number of substitutions in the corresponding subclone (Figure 1c).

Open in a separate window

Figure 1

n-D Dirichlet process clustering reveals widespread polyclonal seeding in A22

(a) For pairs of metastases, cancer cell fractions (CCF), i.e. the fraction of cells within a sample containing a mutation, are plotted for all the substitutions detected in the WGS data. Red density areas off the axes and with CCF >0 and <1 reveal the existence of mutation clusters present at subclonal levels in more than one metastatic site. Mutation clusters for each sample are indicated with circles coloured according to the subclone they correspond to (Supplementary Table 3). The centre of each circle is positioned at the CCF values of the subclone in the two samples. The clusters at (1,1) correspond to the mutations present in all the cells in both sites (CCF=1) while those on axes refer to sample-specific subclones. For example, light blue and dark green clusters absent from sample A are positioned on the y-axis when H is compared to A but are moved to (0.60,0.08) and (0.60,0.88) when H is compared to K. (b) Each subclone detected in A22 is represented as a set of colour-coded ovals across all organ sites (Supplementary Table 3). Each row represents a sample, with ovals in the far left column nested if required by the pigeonhole principle (SI). The area of the ovals is proportional to the CCF of the corresponding subclone. Subclonal mutation clusters are shown with solid borders. Oval plots are divided into three types: trunk (CCF=1 in all samples), leaf (specific to a single sample) and branch (present in >1 sample and either not found in all samples or subclonal in at least one). (c) Phylogenetic tree showing the relationships between subclones in A22. Branch lengths are proportional to the number of substitutions in each cluster. Branches are annotated with samples in which they are present and with oncogenic/putative oncogenic alterations assigned to that subclone (LOH: Loss of Heterozygosity). (d) Subclone colour key.

In 5/10 cases (A34, A22, A31, A32, A24), we found clusters of mutations present subclonally across multiple metastases, suggesting that polyclonal seeding between different organ sites is a common occurrence in metastatic prostate cancer (Figure 2). Mutations selected from these clusters (181-429 mutations per patient) were validated by deep sequencing (median coverage 471X) of additional aliquots of DNA from each WGS sample and extra metastatic and/or prostate samples, confirming these findings (Extended Data Figures 2-​-7,7, Extended Data Table 1 and Supplementary Information, Section 4e).

Open in a separate window

Figure 2

Subclonal structure within 10 metastatic lethal prostate cancers

All the subclones identified in the whole genome sequenced samples are shown as phylogenetic trees and oval plots (as described in Figure 1). Patients with polyclonal seeding (A34, A22, A31, A32 and A24) are on the right (amp: amplification).

Analysis of known driver events found in the subclones provides important insights into polyclonal spread of prostate cancer during therapy. Androgen-deprivation therapy (ADT) is the standard of care for metastatic prostate cancer and initially induces tumour regression in most patients. However, ADT inevitably results in castration-resistance through various mechanisms, including androgen receptor (AR) amplification, increased AR sensitivity as a result of mutation, AR phosphorylation and bypass of the AR pathway^12,13. It is currently unknown whether castration resistance is generally acquired via a single event or more commonly appears in multiple cells independently. Two of the subclones implicated with polyclonal seeding in A22 carry different oncogenic alterations associated with ADT resistance, suggesting that clonal expansion has been driven by distinct resistance mechanisms: MYC amplification¹⁴ in the purple cluster and a pathogenic AR substitution¹⁵ in the mid blue cluster. Overall, in all five patients with polyclonal seeding, subclones carrying either alterations in AR or genes involved in AR signalling (such as FOXA1), or alternative mechanisms of castration resistance such as MYC amplification and CTNNB1 mutation¹⁶, were found to have re-seeded multiple sites. This suggests that the tumour cell populations with a significant survival advantage are not confined within the boundaries of an organ site but can successfully spread to and reseed other sites (Figure 2).

Precise relationships between metastatic sites reveal the patterns of metastasis-tometastasis seeding. In all 7 cases for which the prostate tumour was sequenced (A10, A22, A29, A31 and A32; by targeted deep sequencing in A21 and A34), multiple metastases were more closely related to each other than any of them were to the primary tumour (Figure 2; Extended Data Figures 2-​-55 and ​and7;7; Supplementary Information, Section 4e). In the 5 cases with polyclonal seeding, this relationship resulted from multiple subclones shared subclonally by different metastases, raising the possibility of interclonal co-operativity, in agreement with recent studies using mouse models^10,17, or remodelling of metastatic niches by initial colonising prostate cancer clones, making them attractive habitats that other clones can colonise later¹⁸. Further, for those patients where multiple metastases from the same tissue type were analysed (A22, A34, A21), metastases located in the same tissue are more closely related than those in different tissues, as previously observed in pancreatic cancer¹⁹. Intriguingly, samples within close physical proximity were often more similar to each other than to more distant samples. This raises the question whether the similarity between metastases in the same tissue type arises as a result of geographical proximity or from tissue-specific seeding.

In order to explore further the relationships between samples, we considered the order of acquisition of mutations. Starting from the MRCA, we observe the accumulation of additional clusters of mutations representing subsequent ‘selective sweeps’²⁰. Phylogenetic trees give clear pictures of the order of events, allowing the creation of ‘body maps’ that represent emergence and movement of clones from one site to another (Figure 3). The observed representation of subclones across different sites may be explained by two different patterns of spread: linear and branching. A22 demonstrates both patterns (Figure 3a). The red and light green subclones are present in all metastases and indicate linear spread from the prostate to the seminal vesicle and thence to the remaining metastases. The remaining inter-site subclones have a more complex pattern demonstrating the emergence of branching lineages, each with demonstrated metastasis-to-metastasis seeding. The stepwise accumulation of clonal mutations in A21, on the other hand, displays a simple linear pattern of metastasis-to-metastasis spread (Figure 3b). Finally, in A24, a period of sequential metastasis-to-metastasis spread was followed by parallel polyclonal spread of subclones between multiple metastases (Figure 3c). Overall, these patterns of seeding from one metastasis to the next are seen in 8 of the 10 patients (all but A12 and A29). We cannot formally exclude an alternative explanation for the observed patterns, that each of these metastases has seeded from an undetected subclone in the primary tumour. However, targeted re-sequencing of a subset of mutations failed to detect any such subclones, despite a median sequencing depth of 471X (Supplementary Information, Section 4e).

Open in a separate window

Figure 3

Metastasis-to-metastasis seeding occurs either by a linear or a branching pattern of spread

Body maps show the seeding of all tumour sites from (a) A22, (b) A21 and (c) A24. Sites shown include samples subject to targeted sequencing (A22-L, A24-F, A24-G) in addition to WGS samples. Seeding events are represented with arrows colour-coded according to Supplementary Table 3 and with double-heads when seeding could be in either direction. When the sequence of events may be ordered from the acquisition of mutations, arrows are numbered chronologically. Subclones on branching clonal lineages are labelled with the same number but with different letters, e.g. 4a & 4b. See Supplementary Information Section 4e for a detailed discussion of the body map in these cases.

Mutations found subclonally in the prostate tumour but clonally in all metastases expose the metastasizing subclone in four cases: A22, A29, A31 and A32. In each of these patients, phylogenetic reconstruction indicates that the metastases are derived from a minor subclone, encompassing <50% of tumour cells. In three cases (A32, A10 and A34), more than one subclone from the primary tumour was involved in seeding of metastases, indicating that multiple subclones achieved metastatic potential (Supplementary Information, section 4e). In the case of A31 and A32, driver alterations that could confer selective advantage on the metastasising subclone(s) were identified (Figure 2). In A32, both copies of TP53 as well as one copy of PTEN, RB1 and CDKN1B²¹ were inactivated early in tumour evolution (Figure 2). Additional aberrations occurred separately in the purple and mid blue subclones to achieve homozygous inactivation of these tumour suppressor genes via independent mechanisms (Supplementary Information, section 4e). In A31, a PPP2R5A deletion and an AR duplication occurred in the metastasising subclones (purple or orange) while, interestingly, the pink cluster, displaying many important oncogenic alterations including events affecting TP53 and MLL3, showed no evidence of metastatic spread (Figure 2, Extended data Figures 3a and ​and8a8a).

Annotation of oncogenic/putative oncogenic alterations (Supplementary Information, section 4c; Supplementary Table 2; Extended Data Table 2) on the phylogenetic trees provides some insight into the sequence of oncogenic events that take place during metastatic progression under ADT. The tumour cells in each patient share a common clonal origin (Figure 2, grey clusters). In all patients but one (A34), this mother clone represents the largest cluster of mutations (range 40-90% of all mutations) and contains the majority of driver mutations (Figures 2 and 4a-b) similar to previous observations in pancreatic cancer²². In contrast, oncogenic alterations disrupting genes important for AR signalling were rarely on the trunk. All patients had at least one alteration directly affecting the AR locus or genes involved in AR signalling, with widespread heterogeneity and convergent evolution observed across multiple samples from the same patient.

Open in a separate window

Figure 4

Drivers of tumorigenesis are truncal while drivers of castration resistance are convergent

(a) Proportion of trunk, branch and leaf mutations in each sample. (b) Heatmap of oncogenic alterations present on the trunk (top) or off the trunk, i.e. on branches or leaves (bottom). Alterations in oncogenes and tumour suppressors are shown in red and blue, respectively, with shade indicating the number of events in that patient. Focal deletions and substitutions/indels are shown with crosses and stars, respectively. Double crosses indicate homozygous deletions resulting from deletions of both alleles. (c) Continuous selective pressure on AR signalling is observed in the form of multiple rearrangements resulting in multiple copy number increases at the AR locus within the same patient. Chromosomal rearrangements are plotted on top of the genome-wide copy number for each of the 4 WGS samples from A24. Rearrangements are coloured according to the colour code in Supplementary Table 3. Arcs above and below the top vertical line indicate deletion and tandem duplication events, while arcs above and below the second vertical line are head-to-head and tail-to-tail inversions, respectively.

In the great majority of cases, aberrations in AR signalling seem to have occurred after metastatic spread, although A21 and A24 are exceptions. The former has a large tandem duplication including the AR locus present in all samples, suggesting this was an early event. The latter harbours a truncal T878A mutation, which was also detected in two additional metastases (A24-F and A24-G, interrogated by targeted sequencing). Interestingly, though, a series of complex rearrangements between chromosomes 2 and X resulting in AR amplification was not detected in these samples (Figure 4c). Since such amplification is selected for by ADT²³, it is likely that spread from the falciform ligament (A24-G) to the right axillary lymph node (A24-A) took place after ADT, which commenced 2 years and 9 months prior to death (Figure 3c). Across the whole cohort, only one out of 17 AR amplifications was truncal, with the remainder present only in a subset of metastases. Furthermore, in five patients, copy number had increased on more than one occasion within the same sample (Figure 4c and Extended Data Figure 8) implying continuous selective pressure on the AR pathway, in line with recent reports of persistent AR signalling in castration resistant prostate cancer¹⁵.

Our analyses allow us to view with unprecedented clarity the genomic evolution of metastatic prostate cancer, from initial tumorigenesis through the acquisition of metastatic potential to the development of castration resistance. A picture emerges of a diaspora of tumour cells, sharing a common heritage, spreading from one site to another, while retaining the genetic imprint of their ancestors. After a long period of development prior to the most recent complete selective sweep, metastasis usually occurs in the form of spread between distant sites, rather than as separate waves of invasion directly from the primary tumour. This observation supports the ‘seed and soil’ hypothesis in which rare subclones develop metastatic potential within the primary tumour⁷, rather than the theory that metastatic potential is a property of the primary tumour as a whole^24,25. Transit of cells from one host site to another is relatively common, either as monoclonal metastasis-tometastasis seeding or as polyclonal seeding. Clonal diversification occurs within the constraining necessity to bypass ADT, driving distinct subclones towards a convergent path of therapeutic resistance. However, the resulting resistant subclones are not constrained to a single host site. Rather, a picture emerges of multiple related tumour clones competing for dominance across the entirety of the host.