Integrated signatures define mutational processes in prostate cancer

Prostate cancer is the second most common cancer globally², and the fifth leading cause of cancer death among men³. Despite its high prevalence and substantial burden on patients and society, the aetiology underlying the disease’s clinical heterogeneity, from indolent to highly lethal, remains poorly understood. Prostate cancer is influenced by family history, inherited germline mutations and patient ancestry⁴. Mutations in DNA damage repair (DDR) genes such as BRCA1 or BRCA2, associated with homologous recombination deficiency (HRD), are present in approximately 2% of localized disease and 15% of metastatic castration-resistant disease⁵. Other clinically relevant but rare DDR pathways (reviewed previously⁶) include mismatch repair deficiency (MMRd), identified in 2–3% of primary prostate cancer^7,8, and CDK12 inactivation, associated with a high burden of somatic tandem duplications^9,10,11. As a hormone-driven disease, androgen signalling is pivotal, mediated by activation of the steroid-binding androgen receptor (AR) transcription factor¹. AR activation by testosterone and dihydrotestosterone causes widespread changes in transcriptional activation of AR-target genes, also termed the AR cistrome^12,13. Moreover, both single-nucleotide variants (SNVs) and structural variants (SVs) accumulate at AR-binding sites (ARBSs)¹⁴, with the latter associated with chromoplexy¹⁵, creating chains of interchromosomal complex SVs (cSVs).

Beyond these specific contributions, the broader landscape of oncogenic mutational processes and their underlying aetiologies remains largely undefined. To characterize mutational processes, mathematical methods are commonly used that identify mutational signatures, or footprints, that correspond to each specific mutational process. Methods have been developed to identify signatures based on SNVs and insertion–deletions (indels)¹⁶ and, more recently, copy-number alterations (CNAs)¹⁷ and cSVs¹⁸. Previous studies have mainly investigated these mutational signatures separately, identifying a few common mutational processes in prostate cancer including clock-like mutagenesis (SBS1 and SBS5), HRD (SBS3 and ID6), MMRd (SBS15, SBS21 and SBS44)^16,19,20,21 and complex genomic configurations, mainly chromoplexy (approximately 20%)^15,18 and chromothripsis (20–40%)^22,23. Despite these important findings, a fundamental appreciation and understanding of the aetiologies and mutational processes driving prostate cancer remains unexplored.

The Pan Prostate Cancer Group (PPCG) has assembled a clinically well-annotated and diverse collection of whole-genome sequencing (WGS) data from fresh-frozen tumour specimens from 1,001 donors. The WGS data have been reanalysed with standardized pipelines annotating high-accuracy germline and somatic variant calls and driver mutations²⁴. Here we present a comprehensive and integrative mutational signature analysis to describe the mutational processes that influence and shape the clinical outcome of primary prostate cancer.

The PPCG cohort, comprising 1,209 prostate samples from 1,001 prostate cancer donors, provides the most comprehensive resource to date to investigate the footprints of somatic mutational processes in prostate cancer genomes (Fig. 1). Here we examined one representative primary tumour sample (n = 959) from each PPCG donor (Extended Data Fig. 1), with ages ranging from 32 to 88 years (mean, 60 years), and a median follow-up of 7 years. Of the 903 patients with follow-up information, 132 (14.6%) developed metastatic disease (median time to metastasis, 6 years; 24 patients had no available metastatic follow-up information)²⁴.

The genomic landscape comprised 4,678,132 somatic SNVs (range, 51–292,073), 291,531 indels (range, 1–62,203), 74,386 SVs (range, 0–876) and 55,279 CNAs (range, 0–1,023) (Fig. 1). We used a curated set of germline and somatic driver mutations to annotate patient tumour samples with mutations in known DDR genes, including those involved in homologous recombination (HR), base excision repair (BER), nucleotide excision repair (NER), non-homologous end-joining (NHEJ) and mismatch repair (MMR), tallying 21% of the cohort in total, in agreement with previous estimates^25,26.

To shed light on the mutational processes affecting prostate cancer genomes, we extracted four types of signatures de novo: single base substitution (SBS), indel (ID), cSV and copy-number (CX) signatures. In total, we extracted 22 SBS, 10 ID, 8 CX and 6 cSV signatures (Fig. 2a,b and Extended Data Figs. 2 and 3). Two of the extracted SBS signatures are considered artefactual according to the COSMIC signature annotations (SBS52 and SBS54) and were excluded from downstream analyses. Of the 899 samples with available copy-number profiles, we identified 321 (37.4%) as having CNA levels below the threshold required to identify a discernible pattern of chromosomal instability¹⁷ (CIN) (<20 CNAs; Supplementary Methods), and were therefore classified as copy-number stable (hereafter, non-CIN).

Clock-like mutational processes, represented by SBS1, SBS5 and SBS40, were the most prevalent across the cohort as expected^16,27 (Extended Data Fig. 3c). Other established mutational processes detected included MMRd (2%), linked with the signatures SBS15, SBS21 and SBS44 (Fig. 2c) and MSH2 mutations (Fig. 2d); reactive oxygen species (ROS) and BER-related mutagenesis (13.1%), represented by the SBS18 signature; and APOBEC mutagenesis (1.44%), caused by cytidine deaminase enzymes that are part of a viral defence mechanism, and associated with the SBS2 and SBS13 signatures¹⁶ (Fig. 2c). While the prevalence of APOBEC mutagenesis is consistent with previous studies¹⁶, we note that higher prevalences have been reported in other studies²⁸ (Supplementary Note 1).

We also found evidence for two additional SBS and ID signatures with limited previous evidence: ID83C, a signature with single-base deletions in homopolymers detected in 1.3% of samples; and SBS96D, which was detected in 3.9% of samples (Extended Data Fig. 3, further details are provided in Supplementary Note 2).

Motivated by the diverse SV patterns and associated genomic features found in prostate cancer (Supplementary Note 3 and Supplementary Figs. 1–5), we constructed prostate-specific cSV signatures de novo. In brief, we performed unsupervised clustering using multivariate analysis to identify cSV signatures based on a recent pan-cancer study approach¹⁸. In total, we identified six stable cSV signatures (Fig. 2a,b and Supplementary Table 1), which we tested for association with genomic features to assign putative aetiologies. This included somatic mutations in driver genes, deficiency in five DDR pathways, whole-chromosome CNAs, focal amplifications, WGD, tandem duplications, chromothripsis, chromoplexy and the unique patterns defining each individual signature. We also computed pairwise correlations between the signature types to determine evidence of convergence on an aetiology (Fig. 2c–e and Extended Data Fig. 4).

cSV signature 1 (cSV1) encoded a characteristic pattern of deletions in late-replicating regions, chromothripsis, inversions and local 2-jumps, suggesting replication stress as the putative aetiology (Fig. 2a). Moreover, cSV1 was associated with somatic alterations of MYC and SPOP, and with high activity levels of CX8 and CX11 (Fig. 2c–e), signatures linked to mid- and high-level amplifications through replication stress¹⁷. By contrast, cSV2 had characteristic patterns linked with SVs near ARBS and with chromoplexy formation^15,18, suggesting AR-driven transcriptional activity as the aetiology. cSV3 (unknown aetiology) was also associated with chromoplexy but characterized by deletions in early-replicating regions. cSV4 and cSV5 stratified patients by different sizes of tandem duplications and tandem duplicator phenotype (TDP)²⁹, and cSV4 was concurrently observed with biallelic inactivation of CDK12 (Fig. 2d). Lastly, there was compelling evidence supporting HRD in tumours with cSV6, a complex signature characterized by a composite footprint of unbalanced translocations and foldback inversion. cSV6 was significantly correlated with the HRD-related mutational signatures SBS3, ID6 and CX3 (Fig. 2c), with deficiency in the HR pathway through biallelic inactivation of canonical genes (Fig. 2e), including BRCA2 mutations (Fig. 2d). As expected, BRCA2 mutations were also correlated with elevated activity of other well-established HRD signatures (SBS3, ID6, ID8 and CX3; Fig. 2d).

In summary, we identified the full spectrum of mutational signatures, including six novel cSV signatures, supporting the presence of several mutational processes in prostate cancer genomes.

Somatic alterations are not uniformly distributed across the genome but exhibit topographical biases—often in a tissue-specific manner^30,31. To investigate the topographical biases of mutational processes in prostate cancer, we integrated our comprehensive set of signatures with a diverse range of common and prostate-specific genomic features.

We first tested the proximity of simple and complex SVs to SBS signatures, which showed a strong and distance-dependent co-occurrence of SVs with SNVs linked with APOBEC activity (false-discovery rate (FDR)-corrected P = 1.9 × 10⁻¹³, Kolmogorov–Smirnov test; Fig. 3a and Supplementary Fig. 6a) in the subset of patients with APOBEC activity (SBS2/SBS13, n = 13), as expected^19,32. Across the different types of cSV, we identified an enrichment of chromothripsis breakpoints near APOBEC-associated SNVs (FDR-corrected P = 3.03 × 10⁻³, Kolmogorov–Smirnov test; Fig. 3b and Supplementary Fig. 6b,c), suggesting APOBEC-mediated editing at chromothriptic breakpoints. In support, a recent study used an APOBEC3B-knockout cell line to demonstrate that such APOBEC mutagenesis at chromothriptic breakpoints depends on the cytidine deaminase APOBEC3B³³. Next, we examined the impact of replication and transcription. Highly transcribed genes (fourth quartile expression) and early-replicating regions were strongly enriched for both simple and complex SVs, including templated insertions, chromoplexy and tandem duplications (Fig. 3c), probably arising through fork-stalling and template switching or transcription–replication collision¹⁸. Finally, we examined AR binding to DNA, another well-known mutagenic process that can cause SVs in prostate cancers¹⁴. We found chromoplexy events to be specifically and uniquely enriched at tumour-specific ARBSs (TARBSs)¹² (Fig. 3c). By contrast, chromothripsis, was highly enriched at late-replicating regions, which are on average gene poor and transcriptionally inactive.

The mutagenicity of AR binding was not restricted to SVs. In agreement with previous studies^14,34, we found enrichment of SNVs at ARBSs (Supplementary Note 4 and Supplementary Figs. 7–11). This occurred primarily in tumours exceeding a minimal AR activity threshold, suggesting additional factors contribute beyond this level. SNVs at ARBS may potentially disrupt AR binding and therefore contribute to cancer-promoting gene deregulation¹⁴. We next examined whether AR binding was linked with specific mutational signatures at ARBS, distinguishing between those enriched in tumours (TARBS) and those enriched in normal prostate (NARBS)³⁴. We found several lines of evidence for AR binding associated with an increase in clock-like mutational signatures in tumours relative to their genome-wide activity. These included SBS5, SBS40a, ID5, SBS3 and SBS8 (fold change (FC) = 1.42, 1.47, 1.35, 1.40, 1.34; FDR-corrected P = 4.5 × 10⁻⁷⁸, 2.8 × 10⁻⁶⁹, 9.8 × 10⁻⁵, 3.8 × 10⁻²², 1.4 × 10⁻¹⁹, respectively, RM2 method³⁵; Fig. 3d and Supplementary Note 4). Notably, while two of the clock-like mutational signatures were enriched at TARBS (SBS5 and SBS40a), we found a depletion of the clock-like signature SBS1 compared with both NARBS and genome background (Fig. 3e). SBS1 is associated with spontaneous deamination of 5-methylcytosine (5mC) at CpG dinucleotides leading to C to T transitions³⁶, and regulatory regions are enriched for CpG sites, where 5mC is associated with transcriptional repression³⁷. We hypothesized that CpG methylation differences at ARBS could underpin the observed depletion of SBS1 signature. To this end, we performed Oxford Nanopore long-read sequencing of eight tumours in our cohort to jointly identify DNA nucleotide context and methylation changes, which showed an enrichment of unmethylated CpGs at TARBS compared to NARBS (average FC = 2.07, P = 0.039, Wilcoxon signed-rank test; Fig. 3f). This suggests that the AR mutagenicity at enhancers and promoters is conditioned on existing CpG methylation levels. In agreement, AR binding depends on the pioneering transcription factor FOXA1³⁸, which first binds to specific enhancers and promoters and recruits the demethylase TET1, therefore allowing AR to bind to these ectopic locations³⁹. This provides a conceptual explanation for the observed depletion of SBS1 mutations at TARBS, probably due to open active chromatin and unmethylated CpG sites at ARBS in tumours (Fig. 3g).

We next investigated the influence of replication timing, derived from prostate cancer cell lines⁴⁰, on both mutational signatures and AR mutagenicity. We observed SNVs in late replicating regions to be 1.73-fold enriched (P = 4.9 × 10⁻³²³, Poisson test) compared with early-replicating regions, consistent with previous studies⁴¹. Most mutational signatures displayed similar enrichment in late replicating regions, with notable enrichment of the ROS-associated SBS18 in late replicating regions (median, 4.33-fold enrichment, P = 4.9 × 10⁻³²⁴, Poisson test; Extended Data Fig. 5a). By contrast, we found a relative enrichment of TARBS, and to a lesser extent NARBS, in early-replicating regions (Fig. 3h and Supplementary Fig. 10), compatible with the association with hypomethylated open active chromatin regions, generally found enriched in early-replicating regions⁴². We observed a near-uniform distribution of mutations across replication timing zones at TARBS with strong depletion of SBS1, which we propose is due to the epigenetic landscape of prostate cancer, with TARBS found preferentially at hypomethylated open active chromatin.

In summary, we found a broad impact of replication and epigenetic contexts in defining the topographical biases of both cSVs and SNVs, with AR-mediated somatic alterations strongly linked with early replication and active transcription and relative depletion of the ubiquitous SBS1 signature.

Mutational processes can cause highly complex footprints in the genome, simultaneously involving signatures of SNVs, indels, SVs and CNAs. We therefore reasoned that combining all four levels of mutational signature (SBS, ID, CX and cSV) would better characterize the major mutational processes operating on prostate cancer genomes. For example, while processes such as chromothripsis and cytosine deamination can be potentially identified with a single signature type, robust identification of most mutational processes depends on multiple complementary mutational signatures, for example, HRD (SBS, ID, CX and cSV) and MMRd (multiple SBS and ID signatures).

To this end, we performed hierarchical clustering of the four levels of mutational signature from 521 samples of patients with CIN with activities quantified for all four mutational signature types (Extended Data Fig. 1), identifying eight clusters, which we term IMFs (Fig. 4a,b, Supplementary Figs. 12–13 and Extended Data Fig. 6). These IMFs represent the key mutational processes operating in prostate cancer. We quantified IMF activities within each tumour using linear combination decomposition. Using IMF activities, we pursued multiple association analyses to characterize the IMF aetiologies, including mutated genes (Fig. 4c), individual mutational signatures, specific chromosomal alterations, DDR pathway deficiencies, HRDetect⁴³ (HR deficiency score) and MSIsensor-pro⁴⁴ (MMR deficiency score), as well as tumour mutational burden (TMB), whole-genome duplication (WGD), cell cycle proliferation, AR pathway activity and hypoxia scores (Extended Data Fig. 7 and Supplementary Fig. 14).

IMF1 (MMRd) activity correlated with MMRd-related mutational signatures SBS15, SBS21 and ID2 (Spearman’s rho = 0.14, 0.13 and 0.69, FDR-corrected P = 0.006, 0.01 and 1.45 × 10⁻⁷², respectively; Extended Data Fig. 7a). IMF1 had higher activity in cases with microsatellite instability (Cohen’s d = 1.53), mutations in MSH2 (Cohen’s d = 1.32) and deficiency in the MMR pathway (Cohen’s d = 0.65; Extended Data Fig. 7d). Together, this supported an aetiology of MMRd for IMF1. Tumours with high IMF1 exhibited a near-uniform mutation rate across the genome (Extended Data Fig. 5b and Supplementary Fig. 8b).

IMF2 (ROS mutagenesis) was significantly associated with SBS18 (Spearman’s rho = 0.32, FDR-corrected P = 1.56 × 10⁻¹²; Extended Data Fig. 7a), a signature linked with faulty BER and DNA damage induced by ROS⁴⁵. In agreement, IMF2 was anti-correlated with hypoxia score (Spearman’s rho = −0.13, FDR-corrected P = 0.051; Extended Data Fig. 7b). IMF2 was also associated with SBS8 (Spearman’s rho = 0.32, FDR-corrected P = 1.56 × 10⁻¹²; Extended Data Fig. 7a), a signature with unknown aetiology but speculated to involve NER⁴⁶, a repair pathway that can be used in response to ROS-based damage⁴⁷ and late-replicating errors⁴⁸. Both SBS8 and SBS18 were strongly enriched in late replicating regions (Extended Data Fig. 5a).

IMF3 (AR-mediated mutagenesis) displayed several marks of AR-mediated mutagenesis, including significant correlation with cSV2 and cSV3 (Spearman’s rho = 0.34 and 0.48, FDR-corrected P = 4.02 × 10⁻¹⁵ and 1.75 × 10⁻³⁰; Extended Data Fig. 7a), and elevated mutation rate at TARBS (FDR-corrected P = 1.61 × 10⁻¹⁰, RM2 negative binomial test³⁵; Supplementary Fig. 8b). IMF3 activities were higher in cases with ETS fusions (Cohen’s d = 0.75, FDR-corrected P = 1.65 × 10⁻¹⁵, Welch’s t-test; Fig. 4c), with PTEN mutations (Cohen’s d = 0.55, FDR-corrected P = 1.93 × 10⁻⁵, Welch’s t-test) and with mutations in one of two key AR regulators TBL1XL1⁴⁹ or ZBTB16⁵⁰ (Cohen’s d = 1.67, 1.17; FDR-corrected P = 0.02, 0.002; Welch’s t-test). We also found correlations with clock-like SBS1 and SBS40a signatures (Spearman’s rho = 0.51 and 0.50, FDR-corrected P = 3.21 × 10⁻¹⁶, 1.92 × 10⁻⁹; Extended Data Fig. 7a).

IMF4 (mitotic defects and replication stress) activity was significantly higher in cases with mutations in genes involved in the mitotic machinery, including CDC16⁵¹ (homozygous loss) and MZT2A/B⁵² and TUBA3D/E⁵³ (homozygous losses, located in close proximity) (FDR-corrected P < 0.05, Welch’s t-test; Fig. 4c). Moreover, IMF4 displayed CNAs linked to replication stress with CX4 (associated with WGD), CX5 (impaired HR plus replication stress) and CX11 (mid-level amplifications through replication stress)¹⁷, suggesting faulty replicated DNA after replication stress that persists into mitotic defects⁵⁴. IMF4 also correlated with SBS37 (Spearman’s rho = 0.35, FDR-corrected P = 7.07 × 10⁻¹⁶), a signature with unknown aetiology that accumulates in late-replicating regions¹⁶ (Extended Data Fig. 7a), and showed the highest late-to-early mutation rate ratio among all mutational footprints (Extended Data Fig. 5b).

IMF5 (non-canonical impaired HR) had significantly higher activity in tumours with mutations in CDK12 (Cohen’s d = 1.87, P = 0.02, Welch’s t-test; Fig. 4c), which has been shown to phenocopy impaired HR in the absence of mutations in BRCA1/2⁵⁵. This suggests an aetiology of impaired HR, which we term non-canonical to differentiate from canonical impaired HR, which typically involves enrichment in BRCA1/2-mutant cases and a different pattern of DNA damage. In support, IMF5 activities were significantly correlated with cSV4, characterized by tandem duplications, and CX2, linked to impaired HR (Extended Data Fig. 7a). Notably, this mutational process was enriched in patients of African ancestry (P = 5.36 × 10⁻⁵, n = 32, Fisher’s test; Supplementary Fig. 15), consistent with a previous study finding more tandem duplications in this group of patients⁵⁶. We also found IMF5 to associate with increased hypoxia and TMB, and decreased AR pathway activity (Extended Data Fig. 7).

IMF6 (replication stress) showed significant correlation with cSV1 (Spearman’s rho = 0.78, FDR-corrected P = 2.39 × 10⁻¹⁰⁴; Extended Data Fig. 7a), a signature linked with chromothripsis and genomic instability in late-replicating regions, proliferation rate and with AR pathway activity and hypoxia (Spearman’s correlation, FDR-corrected P < 0.05; Extended Data Fig. 7b). Moreover, IMF6 exhibited increased activity in cases with mutations in SPOP and WNT-pathway genes APC and SOX17 (FDR-corrected P < 0.05, Welch’s t-test; Fig. 4c). SPOP is a ubiquitin ligase, involved in preventing replication stress and genomic instability, and SPOP-mutant tumours are associated with dysregulation of PI3K, AR and WNT signalling⁵⁷.

IMF7 (APOBEC mutagenesis) was correlated with APOBEC-associated signatures SBS2 and SBS13, and with impaired HR-related mutational signatures SBS3 and ID6 (Spearman’s correlation, FDR-corrected P < 0.05; Extended Data Fig. 7a), but showed no elevated activity in BRCA2-mutant cases. We examined whether these two distinct mutational processes tended to occur simultaneously or separately in the same tumour samples. To explore this, we evaluated the clonality of mutations assigned to SBS2/13 or SBS3/ID6 to test which signatures may have occurred preferentially early in tumour evolution (clonal) or later (subclonal). We found that APOBEC-related mutations often occurred clonally, while impaired HR mutations appeared subclonally (Supplementary Fig. 16), suggesting that APOBEC mutagenesis was the main aetiology and that SBS3/ID6 represented separate, subsequent processes.

IMF8 (impaired HR) was correlated with mutational signatures SBS3, ID6, cSV6 and CX3 related to impaired HR (Spearman’s FDR-corrected P < 0.05; Extended Data Fig. 7a). IMF8 activity was also significantly higher in cases with BRCA2 mutations, mutations in the HRD pathway and high HRdetect scores (FDR-corrected P < 0.05, Welch’s t-test; Fig. 4c and Extended Data Fig. 7). Together, this evidence suggests an aetiology of canonical impaired HR.

To determine the dominant mutational process shaping each prostate cancer, we assigned samples to the IMF with the highest activity. Extended Data Table 1 summarizes the distribution of dominant mutational processes across PPCG tumours, highlighting AR, replication stress and ROS-mediated mutagenesis as the most recurrent mutational processes (39%, 28% and 21%, respectively; Extended Data Table 1).

Among the 321 non-CIN tumour samples (without CX but with available SBS and ID signatures), we reliably assigned a known mutational process to 244 of these non-CIN tumours, primarily clock-like mutational processes (Supplementary Note 5 and Supplementary Figs. 17–20). Of the remaining tumours without a detectable mutational process, 62 had low tumour purity (<20%) and low somatic variant burden. Overall, this results in 837 tumours (85%) with ascertainable signatures in our cohort (Supplementary Fig. 21).

In summary, by integrating four different types of mutational signature, we identified eight distinct mutational processes that, in combination with a subset found in non-CIN tumours, underpin 85% of primary prostate cancers in our dataset.

We hypothesized that the distinct mutational processes captured by the IMFs may help explain differences in the age-at-diagnosis and clinical outcomes observed in patients with prostate cancer. 

Previous studies have shown distinct mutational patterns in early-onset prostate cancer (55 years of age or younger at time of diagnosis)^19,58, but no specific mutational processes have been identified in patients with the late-onset disease. We found that IMF6 (replication stress) was predominant in late-onset tumours, whereas IMF8 (HRD) was enriched in early-onset cases (FDR-corrected P = 3.58 × 10⁻⁴ and P = 0.0107, respectively, Fisher’s exact test; Extended Data Fig. 8a). Age was strongly correlated with cSV1, a key signature defining IMF6 (Extended Data Fig. 8b–d and Supplementary Figs. 22 and 23), and late-onset tumours frequently had alterations in replication-stress genes, including somatic variants in SPOP (n = 39) and RIF1 (n = 5), germline variants in RECQL4 (n = 4) and somatic MYC amplifications (n = 40; Extended Data Fig. 8e). These findings support the presence of a unique subset of patients with late-onset prostate cancer dominated by replication-stress processes.

To dissect the clinical heterogeneity of prostate cancer, we performed multivariate Cox proportional hazard analyses to compare metastasis-free survival (MFS) among patients according to their dominant IMF (Fig. 5a, Extended Data Table 1 and Supplementary Table 2), using patients with low chromosomal instability (non-CIN cases) as the reference group in the survival analyses given their reduced risk of disease progression compared with patients with genomically unstable tumours (Extended Data Fig. 9). Tumours primarily driven by either ROS-mediated mutagenesis (IMF2; hazard ratio = 5.26, P = 1.4 × 10⁻⁴, Wald test), AR-mediated mutagenesis (IMF3; hazard ratio = 4.05, P = 0.0019, Wald test), non-canonical impaired HR (IMF5; hazard ratio = 7.17, P = 0.006) or canonical impaired HR (IMF8; hazard ratio = 5.8, P = 4.1 × 10⁻⁴, Wald test) presented a high risk of metastasis, independent of age at diagnosis, Gleason grade group, replication timing ratio, tumour stage and TMB. MMRd (IMF1) and replication stress (IMF4 and IMF6) did not confer increased risk of metastasis compared with non-CIN cases. Notably, we found that IMFs were also able to risk-stratify patients within intermediate-risk Gleason grade groups GG2 or GG3, with GG2 tumours showing significantly increased risk with IMF8, and GG3 tumours with IMF2, IMF3 or IMF8 (Supplementary Fig. 24). Increased late-to-early replication timing ratio was associated with a decreased risk of metastasis (Fig. 5a), and with decreased biochemical recurrence risk in an external cohort (Supplementary Fig. 25). A similar Cox model analysis using individual mutational signatures did not reveal strong risk associations (Supplementary Figs. 26 and 27).

To determine whether intratumoural heterogeneity was also reflected in the dominant mutational processes, we analysed the mutational processes in an expanded series of tumour regions from four patients from PPCG with a high degree of intratumoural heterogeneity based on pathology reports (Fig. 5b and Supplementary Fig. 28). One patient (PPCG0851) had two genetically unrelated and spatially confined tumour foci with entirely distinct mutational processes. Two further patients (PPCG0056 and PPCG0440) exhibited both shared and regionally localized mutational processes, whereas all five analysed tumour samples in another patient (PPCG0058) were dominated by IMF3. Notably, we identified a tumour area (T1) defined by IMF2 in PPCG0440, a mutational process linked to ROS mutagenesis and altered oxygen metabolism.

We next sought to investigate whether the IMFs could also inform therapy selection in metastatic disease, in which the clinical need for robust predictive biomarkers is most urgent. To examine this, we used a WGS cohort of metastatic castration-resistant prostate cancers (mCRPC) from the Hartwig Medical Foundation study⁵⁹. mCRPC is an advanced, lethal stage of prostate cancer where tumours progress despite androgen deprivation therapy (ADT). Current management relies on the sequential or combined use of AR pathway inhibitors (ARPIs) and taxane-based chemotherapies on a continuous ADT backbone. As the transition to second-line treatment often involves an empirical switch between these agents, we used this setting to evaluate the predictive power of IMFs. This context enabled us to control for first-line therapy and determine whether IMFs could independently predict response to the subsequent ARPI or taxane.

Overall, 240 patients from HMF with sufficient data were available for our analyses (Methods). Using these data, we conducted a target trial emulation design to mimic phase III randomized controlled biomarker trials⁶⁰, with patients stratified by median IMF activity. Patients were randomized in silico to an ARPI- or taxane-based regimen, with time to treatment failure (TTF) as the primary end point. Power analyses supported testing of IMF1, IMF4, IMF5 and IMF6 (Supplementary Tables 3 and 4), with IMF6 subsequently showing predictive capacity (Fig. 5c and Supplementary Table 5). In patients predicted to be sensitive (IMF6 > median), ARPI treatment was associated with a significantly reduced risk of treatment failure compared to taxane (hazard ratio = 0.21, 95% confidence interval (CI) = 0.05–0.79, P = 0.021, Wald test), whereas no significant difference was observed among patients predicted to be resistant.

In summary, IMFs not only stratify clinical outcomes but also show potential to predict therapeutic response in the metastatic setting, supporting their relevance for patient stratification and disease management. As our trial emulation analyses are retrospective and observational, with limited subgroup sizes, these findings should be viewed as hypothesis generating and warrant prospective biomarker-driven validation.

The complete landscape of mutational processes in prostate cancer and how they contribute to clinical outcomes has remained unclear. Efforts have focused on characterizing and risk-stratifying patients with HRD owing to the increased risk and therapeutic potential. However, fewer than 3% of primary prostate cancers exhibit a deficiency in this repair mechanism⁵. Here we performed a large and comprehensive examination of the mutational processes shaping the genomes of 961 primary prostate cancers with diverse histopathological stages and clinical outcomes. We leveraged all genomic variants captured by WGS to dissect the genomic footprints caused by different mutational processes within the tumour. Through integrated analyses of multiple signature modalities including six novel cSV signatures, we identified eight mutational processes in prostate cancer that, together with clock-like mutational processes in non-CIN tumours, were able to explain the mutational processes underpinning genomic alterations in 85% of primary prostate cancers.

A key driver in prostate cancer is AR, the activity of which is mutagenic^15,58. We found that mutations in more than a third of prostate cancer genomes with CIN were largely attributed to IMF3 (Extended Data Table 1), linked with AR-mediated mutagenesis. We observed substantial biases in the regional distribution of both SNVs and SVs, and we provide evidence that this is mainly driven by differences in replication timing, DNA methylation and AR binding, except in tumours with MMRd, which showed a nearly uniform distribution of SNVs, consistent with replication-coupled MMR with higher efficiency in early-replicating regions⁶¹. Signatures of MMRd, associated with replicative slippage and formation of microsatellite instability were detected with a prevalence similar to previous studies and with a dominance of mutations in MSH2⁸. Although IMF1 (MMRd) tumours exhibited CIN, we also found signatures of MMRd in chromosomally stable tumours.

Our study also identified a pervasive role for replication stress in prostate cancer, with dysfunctional DNA replication as the dominant mutational process in 28% of prostate cancers with CIN (IMF4 and IMF6). Although prostate cancer is generally considered to grow slowly, it may follow a pattern of punctuated equilibrium, where long periods of slow growth and relative genomic stability are interrupted by short episodes involving accelerated proliferation, genomic instability and replication stress^15,62,63,64. Similar findings have been made in breast cancer, another hormone-sensitive cancer⁶⁵.

Although our study was not designed to develop a biomarker for risk stratification, we found that the IMFs were able to stratify primary prostate cancers that developed metastases. Using an approach to emulate phase III randomized control biomarker trials that we recently pioneered⁶⁰, we showed that IMF6 activity was predictive of reduced risk of treatment failure following ARPI. IMF6 is linked to SPOP mutations and replication stress, which corroborates recent findings of improved ARPI response in SPOP-mutated mCRPC⁶⁶ and may extend the scope of prediction beyond isolated SPOP-mutation-based approaches. Notably, only seven patients in the HMF cohort had SPOP alterations, two of whom received these therapies at second line, further supporting that the predictive power of IMF6 is not limited to, but extends beyond, SPOP-mutated tumours. While these findings are encouraging towards addressing an urgent clinical unmet need, more extensive and well-powered prospective biomarker-driven studies are warranted.

Our study is not without limitations. For example, our approach is focused on identifying mutational processes causing alterations of the genome. Across the cohort, we identified 8.6% of primary prostate cancer samples without any detectable mutational processes, and found 22% to be non-CIN with primarily clock-like mutations. Several known oncogenic mutational processes involve dysregulation of epigenetic and/or transcriptional programs, and it is interesting to speculate that a subset of primary prostate cancers might be driven by such non-genetic processes. In support, a recent study found that epigenetic modifications, mediated by transcriptional silencing by Polycomb group proteins, could promote cancer initiation in an animal model system⁶⁷.

Together, our findings highlight that the complex nature of mutational processes, which often involve multiple types of genomic alterations (SNVs, indels, CNAs and SVs), requires a comprehensive integrative approach for accurate identification. We provide compelling evidence that such an approach is essential for deciphering the mutational processes shaping prostate cancers and for identifying patients developing more aggressive disease. Looking ahead, our study supports the implementation of WGS in clinical practice to improve risk stratification and potentially drive the development of predictive biomarkers for targeted therapy response.

PPCG cohort, WGS and variant calling

We used data from the PPCG consortium of primary prostate cancer samples from a total of 1,001 prostate cancer donors with 1,172 tumour samples. Informed ethical consent was obtained at clinical follow-up, and was consistent with local research ethics and International Cancer Genome Consortium (ICGC) guidelines (https://icgc.org/). Ethical approval was obtained from local research ethical committees (details are provided in the Supplementary Methods).

Analysis of SVs from genomic data

Simple and complex SV classification method

We conducted SV classification on the PPCG cohort using the cSVc tool¹⁸, including also SVs near TARBS and chromothripsis, both abundant in prostate cancer. First, exact breakpoint estimation from our WGS short-read sequencing data was conducted by using split-read information. Specifically, for each sample, we used the median of soft-clipped reads from the corresponding tumour BAM file to obtain corrected breakpoint positions. Next, we used ClusterSV¹⁸ to obtain clusters of SVs from the corrected SV data and we also used Battenberg⁶⁸ to obtain segmented CN files from tumour and normal read coverage files. For each sample, we generated CN files segmented by corrected SV breakpoints by using the corrected SV data, the CN segmentation file, CN coverage and the tumour BAM file. Finally, we conducted simple and complex SV classification (chromothripsis excluded) using the CN-SV segmentation file together with the corrected SV file, ClusterSV file and information about ploidy and purity.

To classify chromothripsis on the PPCG cohort, we used Shatterseek v.0.4, using SV and CN data. We used Shatterseek’s recommended cut-off criteria to obtain high-confidence calls⁶⁹. Multiple-testing correction was performed on the P values of three statistical tests indicative of chromothripsis (breakpoint enrichment test, exponential clustering test and the fragment joins test) and we used a q-value cut-off of 0.2 for the chromothripsis calls on each chromosome to confine the final high confidence call set.

To obtain the final SV dataset we merged the chromothripsis call set with the remaining simple and complex SV call set. Specifically, we merged calls only if a minimum of 90% of SV call sets spanning the chromothripsis cluster were marked as Complex Unclassified.

Tandem duplication detection

The presence of TDP samples was assessed by using three criteria described previously⁷⁰. The criteria involve the proportion of tandem duplications, the total tandem duplication count and a TDP score⁷¹:

For a given sample, Obs_i denotes the observed tandem duplications for chromosome i, Exp_i denotes the expected tandem duplications for chromosome i and TD is the total number of tandem duplications. We calculated Exp_i by counting the tandem duplications per chromosome among all samples and dividing this by the sample size. Together, we classified TDP samples when the following criteria were met: tandem duplication proportion > 20%, total tandem duplication count > 50, and TDP score > −1.

Enrichment analysis of SV classes at specific genomic features were performed using a permutation testing framework as described in the Supplementary Methods.

De novo extraction and assignment of mutational signatures

SBS and ID signatures

For the de novo extraction of SBS and ID signatures we considered high-accuracy SNVs and stringently filtered (Tier 1) high-accuracy indels called and quality controlled as described previously²⁴. SBS signatures were extracted de novo using SigProfilerExtractor (v.1.1.23)⁷². De novo extraction of SBS signatures (96 channels) and ID signatures (83 channels) was performed for 3 to 25 signatures for SBS and 1 to 25 signatures for ID using the following parameters: nmf_replicates=500;nmf_init = “random”;min_nmf_iterations=10,000;max_nmf_iterations=1,000,000;export_probabilities=True;make_decomposition_plots=True;get_all_signature_matrices=True. We then decomposed every set of signatures (3 to 25 signatures for SBS and 1 to 25 signatures for ID) independently to known COSMIC signatures (v.3.4 for SBS and v.3.3 for ID) using the default parameters of the decompose_fit function of SigProfilerExtrator. To find the optimal number of SBS or ID signatures, respectively, we investigated each obtained set of signatures based on a number of features (Supplementary Methods).

Structural variant signatures (cSV)

We extracted SV signatures by incorporating a non-negative matrix factorization (NMF) framework onto specific feature counts of the SV data. For deletions and tandem duplications, we grouped the SV classes into four different size groups: 0–50 kb, 50–500 kb, 500 kb to 5 Mb and >5 Mb. Each of these groups was further subdivided on the basis of breakpoint occurrence at distinct replication timing regions (early, mid and late) by applying bedtools pairToBed between SV data and replication timing data (see the ‘Enrichment analysis of SV classes at specific genomic features’ section of the Supplementary Information for details about assessment of replication timing status). In the same manner, we also incorporated a group that captured deletions and tandem duplications occurring at fragile sites (see the ‘Enrichment analysis of SV classes at specific genomic features’ section of the Supplementary Information for details about fragile site data). For balanced inversions, and the two local 2-jump classes Dup-invDup and Loss-invDup, we created two groups based on size with a 100 kb threshold. Templated insertion groups (chains, cycles and bridges) were divided by a 5 kb size threshold. Lastly, we included two novel features compared to a previous study¹⁸, including chromothripsis and SV breakpoints in close proximity to TARBS, a unique feature of prostate cancer genomes. We counted SV breakpoints occurring at TARBS regardless of SV class (see the ‘Enrichment analysis of SV classes at specific genomic features’ section of the Supplementary Information for details about TARBS data). These groups were converted into a matrix that was used as input for the NMF (NMF package in R)⁷³. For finding the optimal number of SV signatures, we applied a procedure used for finding the optimal number of rearrangement signatures in the Palimpsest R package⁷⁴. In brief, based on 100 iterations, we determined the index for which the cophenetic values had the steepest decrease and used that index as the number of signatures. We normalized the SV signature values by dividing each value in a row with the row sum and subsequently visualized these values with ggplot2. We also extracted cSVs from WGS data of the HMF-Prostate cohort to quantify cSV signatures (see the ‘Processing genomic data from the HMF-Prostate cohort’ section).

CX signatures

CX signatures were identified and quantified from absolute copy-number profiles as previously described¹⁷ (Supplementary Methods).

We computed the genome-wide distributions of the five fundamental copy-number features for each sample (the segment size, the difference in copy-number between adjacent segments, the lengths of oscillating copy-number segment chains, the breakpoint counts per 10 Mb, and the breakpoint counts per chromosome arm). Feature distributions observed in the cohort were then deconstructed into a total of eight signatures by applying a NMF. All of these signatures were already present in the pan-cancer compendium of CX signatures¹⁷ (cosine similarity > 0.85). The final signature activities for all samples were computed using the linear combination decomposition function from YAPSA⁷⁵ (R package, v.1.12.0) on the feature component-distributions given a predefined signature matrix. To ensure the robustness of the signature activities and to enable trust in small signature activities, we identified signature-specific thresholds by following the same procedure as in our previous work¹⁷. We next set activities to zero if they were below the signature-specific threshold. The sample-by-activity vector was finally normalized per sample. This approach was also applied to quantify CX signatures in the HMF-Prostate cancer cohort (see the ‘Processing genomic data from the HMF-Prostate cohort’ section).

Identifying integrated mutational footprints

We performed clustering of patient tumours defined by the complete spectrum of signatures to characterize the main mutational processes driving prostate cancer. For each sample, we combined the activities of each of the different types of mutational signatures into a single vector for clustering input, whereas only signatures being active in any of the clustered samples were considered. The data were scaled and centred before clustering, which was performed independently for samples with CIN and without CIN (non-CIN). Given that all tumours share a core set of clock-like mutational signatures³⁶ complemented with more distinct mutational processes during tumorigenesis, we opted to use hierarchical clustering given the natural hierarchical structure of the data. We tested a broad range of clustering parameters and empirically determined that the Ward-D2 algorithm in combination with Manhattan distance best identified known biological features of our cohort. This approach has also been chosen in other studies^76,77 to prioritize robust identification of known biology. In the case of clustering CIN samples, we excluded samples with fewer than 100 SNVs, fewer than 10 indels, or fewer than 5 SVs. We identified eight clusters of CIN samples (Fig. 4b) based on (1) the obtained minimum, median and maximum cluster stability metrics (Supplementary Fig. 12); (2) the silhouette scores in samples (Supplementary Fig. 13); and (3) biological expert knowledge. Each of these clusters characterized a distinct mutational process represented by the integration of mutational signatures. Thus, we referred to these processes as IMFs. In the case of non-CIN samples, we excluded samples with less than 100 SNVs, or less than 10 indels (244 cases; Supplementary Fig. 17). As performed for CIN samples, we explored cluster stability metrics to select 4 clusters for non-CIN samples (Supplementary Figs. 18 and 19).

To quantify the activity of the eight IMFs in a given tumour, we first generated an IMF definition matrix, where each IMF was defined across the 46 multi-signature space (22 SBS, 10 ID, 8 CX and 6 cSV signatures). This was done by averaging the multi-signature activities across all samples within each cluster (Extended Data Fig. 6). Then, given the IMF definitions, we applied the linear combination decomposition function from YAPSA⁷⁵ to the multi-signature vector for each tumour to compute the activity of each of the eight IMFs in the tumour sample. The resulting IMF activity vector is then normalized to sum to 1.

After quantifying the activities of the IMFs, each sample was classified according to its dominant mutational process, defined as the IMF with the highest activity. This stratification was applied for testing the potential of the IMF as biomarkers for predicting metastasis occurrence and treatment response (see the ‘Predicting clinical outcomes in cases of primary prostate cancer’ section).

Linking mutational processes with genomic features

Here we performed association analyses between activities of the integrated mutational footprints and different genomic features. We also correlated each individual mutational signature (SBS, ID, cSV and CX signatures) with different genomic features (detailed below). Only samples with available calls for a specific mutational signature and genomic feature were included in each analysis.

Spearman’s correlation was applied to test the correlation of IMF activities with the 44 individual mutational signatures. All P values were corrected for multiple testing using the Benjamin–Hochberg method (Extended Data Fig. 7a).

Selection of 1,747 genes of interest

We compiled a list of 1,747 genes by integrating and curating the annotations provided in refs. ^78,79,80,81. This set of genes included driver genes recurrently altered in prostate cancer^21,82, known pan-cancer driver genes, as well as genes reported to be involved in DNA damage response and repair (DDR) pathways, DNA replication, the cell cycle and chromatin organization. For all 1,747 genes, we identified their transcriptional start sites (TSS) on a GRCh37 reference genome using ENSEMBL. We selected the canonical TSS for each gene.

Mutation status of genes of interest

For each gene, we integrated germline and somatic alterations for classifying samples as mutant using the following criteria.

Oncogenes: samples with amplifications and somatic mutations that did not result in the truncation of the oncogene were classified as mutant. For oncogenes located in autosomal chromosomes (that is, SPOP, IDH1 and MYC), amplifications were considered in cases in which the copy-number value at the transcriptional start site exceeded the sample ploidy by at least two copies. For oncogenes located in sex chromosomes (AR), the threshold was adjusted to require a gain of at least two copies more than the ploidy − 1. Given that the amplification of the enhancer region is frequently observed to be amplified after ADT treatment⁸³, we scanned a genomic region consisting of 0.5 Mb before the TSS and 1 Mb after the coding region of the AR oncogene.

Tumour suppressor genes: only samples with biallelic inactivation or homozygous deletion (copy-number value < 0.5) of the gene were considered as mutant, with the following exceptions: (1) for TP53, samples with germline variants or single hotspot driver mutations; (2) for CDK12, samples with single hotspot driver mutations; and (3) for DDR-related genes, samples with germline variants due to its impact on cancer risk. For BRCA1 mutants, we also evaluated the presence of reversion mutations rescuing the HRD phenotype. Rescue was defined by a mutation in either TB53BP1, RIF1 or MAD2L2^84,85. None of the BRCA1 mutant samples presented reversion mutations.

To test for differences in signature activities between samples with and without mutations in any of the 1,747 key genes, we divided the samples into two groups: one group with a mutation and one without. We then used Welch’s t-test to perform a test of equality of means between the two groups. Only positive associations were shown to mitigate compositional data effects. To ensure high-confidence associations, we discarded gene associations lacking a moderate/large effect size (Cohen’s d < 0.5). Finally, we further filtered out genes with less than three samples mutated for IMF associations (Fig. 4c).

In the case of individual mutational signatures, we limited the association analysis to well-known prostate cancer genes^21,82, filtered out genes with less than ten samples mutated, and the analysis was performed for the whole cohort and also after stratifying samples on the basis of age at diagnosis (Fig. 2d and Supplementary Fig. 29).

Deficiency in DDR mechanisms

Here we evaluated differences in signature activity between samples deficient and proficient for five different DDR mechanisms: HR, MMR, NHEJ, NER and BER. To do so, we first evaluated the mutation status of genes involved in each DDR pathway. Samples with inactivation of at least one of the DDR-related genes were classified as deficient; otherwise, as proficient. The genes evaluated for each repair pathway were as follows: HR: BRCA1, BRCA2, PALB2; MMR: MLH1, MSH2, MSH6, PMS2; BER: MUTYH, NTHL1, POLD1, POLE, WRN; NHEJ: NBN, RAD50, ATM; and NER: DDB2, ERCC2, ERCC3, ERCC4, ERCC5, POLD1, POLE, XPC, XPA, CUL3.

For each repair mechanism, Welsh’s t-tests were performed to compare signature activities between deficient and proficient samples. P values were corrected for multiple testing using the Benjamini–Hochberg method. Cohen’s d was computed to assess the size of the difference between groups (Fig. 2e, Supplementary Fig. 22 and Extended Data Fig. 7).

Assigning mutational signatures to mutations

For each mutation type in a sample, SigProfilerExtractor⁷² outputs the posterior probability of a mutation being generated from a signature given the mutation type. The probability is computed by SigProfilerExtractor from the estimated exposure of each signature in the sample, trinucleotide profilers of the de novo extracted signatures, and the abundance of each mutation type in the sample. The SigProfilerMatrixGenerator package was used for classifying SNVs into 96 categories and for classifying indels into 83 categories based on the sequence context⁸⁶. We then assigned mutations to the signature with the highest probability given their mutation type, as previously done in other studies^87,88,89. We note that flat signatures, such as SBS5 and SBS40a, are intrinsically difficult to distinguish because their trinucleotide profiles have low information content and high similarity to one another^16,90.

WGD

To test whether the WGD status associated with the activity of a signature, we compared activities between samples with and without WGD by applying a Welch’s t-test. Only samples with activities higher than zero were included in the analysis. P values were corrected for multiple testing using the Benjamini–Hochberg method. Cohen’s d was computed to assess the size of the difference between groups (Extended Data Fig. 4b).

Whole-chromosome CNAs

Copy-number segments were classified as a whole-chromosome alteration if it was larger than 95% of the whole chromosome length covered and had a copy-number value greater or smaller than 2. We then compared signature activities between samples with and without whole-chromosome CNAs using a Welch’s t-test. P values were corrected for multiple testing using the Benjamini–Hochberg method. Cohen’s d was computed to assess the size of the difference between groups.

Focal amplifications

Focal amplifications were defined as segments with a copy-number value equal or higher than 8. We then compared the activities of signatures between samples with and without focal amplifications using a Welch’s t-test. P values were corrected for multiple testing using the Benjamini–Hochberg method. Cohen’s d was computed to assess the size of the difference between groups.

Microsatellite instability

Microsatellite instability (MSI) was evaluated using MSIsensor-pro v.1.3.0 with the default parameters⁴⁴. In brief, MSIsensor-pro scan was first executed on the reference genome to obtain homopolymers and microsatellites. Next, MSIsensor-pro was applied on tumour and normal BAM files with the homopolymer and microsatellite information file, and with coverage threshold set to 15 and coverage normalization set to 1. Samples with a score higher than 10 were considered as MSI-Hi.

HRDetect

We used HRDetect⁴³ to detect the presence of samples with HRD. In brief, we combined all four mutation data type (SV, CNA, SNV and indels) as input for HRDetect to infer six HRD-associated features (microhomology deletions, SBS3 and SBS8 signature, SV3 and SV5 signature). These features were then used to compute an HRD probability score using the default parameters. We considered all samples with a probability score > 0.7 to have HRD.

Kataegis

The detection of kataegis was performed as described previously²⁴. We counted the number of kataegis events per sample and tested the Spearman correlation to the signature activities. Multiple-testing correction was performed according to the Benjamini–Hochberg method.

Cell cycle progression scores

RNA-sequencing (RNA-seq) data were processed and harmonized before downstream analysis as described previously²⁴. For this analysis, we used the high-confidence set of gene expression data (tier 1) RUV-III PRPS-normalized data (v.1.4/K10). Cell cycle progression (CCP) scores, also known as Prolaris scores, were computed for each sample. For each of the 31 genes in the CCP signature⁶³, log₂-normalized expression values were centred and scaled relative to the median expression of the specific gene across all samples. The resulting values were then summed to produce a single CCP score per sample. All 31 signature genes were available in the tier 1 data release used for this analysis.

AR pathway activities

The pathway scores were computed from AR and prostate cancer-related gene sets using the singscore package (https://doi.org/10.18129/B9.bioc.singscore) v.3.21 as described in our companion paper²⁴ using the singscore HALLMARK_ANDROGEN_RECEPTOR as a representation of AR activity.

Hypoxia scores

A hypoxia score was computed as described in our companion paper²⁴ using the ‘Buffa’ hypoxia signature⁹¹. In brief, the hypoxia score was computed for each RNA-seq sample by median-centring gene expression across the cohort, scaling to unit variance and summing across signature genes to yield a single hypoxia score.

ARBS mutation rate enrichment

Chromatin immunoprecipitation–sequencing (ChIP–seq) data of prostate tumour-specific AR binding (TARBS, 9,181 sites), normal prostate-specific AR binding (NARBS, 2,690 sites), FOXA1 (19,735 sites) and HOXB13 (66,104 sites) were obtained from ref. ¹². TARBS were overlapped with FOXA1 ChIP–seq peaks and HOXB13 ChIP–seq peaks to obtain co-binding sites (Supplementary Fig. 11a). Assay for transposase-accessible chromatin using sequencing (ATAC–seq; 111,884 sites) data were obtained from ref. ⁹². ATAC–seq peaks were overlapped with TARBS to obtain TARBS⁺ ATAC–seq sites. ATA–seq peaks without TARBS overlap were sampled without replacement by selecting sites that match the ATAC–seq signal of each TARBS⁺ ATAC–seq site—this forms a set of TARBS⁻ ATAC–seq sites with matching site number and ATAC–seq signal distribution for comparison to TARBS⁺ ATAC–seq sites (Supplementary Fig. 7c). Similarly, TARBS were downsampled to match the number and ATAC–seq signal of NARBS for an additional validation (Supplementary Fig. 7b). Chromatin state data were obtained from ref. ³⁴. Enhancer regions were overlapped with TARBS to obtain a set of TARBS⁺ enhancers, and enhancers without TARBS overlapped were sampled to the same number as TARBS⁻ enhancers for comparison (Supplementary Fig. 7d).

To analyse mutation rate uniformly, we defined ARBS to be 400 bp regions using the midpoint of each ChIP–seq peak as the midpoint of each site. 400 bp is the median of ChIP–seq peak length. We also observed that the mutation rate starts to increase around the 400 bp boundary. 400 bp flanks on both sides of the defined ARBS were used as control. A negative binomial regression model, RM2 (ref. ³⁵), was used to compare ARBS mutation rate to flanking regions. Sequence trinucleotide ratio and megabase-scale background mutation rate were accounted for as covariates in the model. Mutation rate of SBS and ID signatures was calculated by considering subsets of mutations assigned to individual signatures. ARBS mutation rate analysis was performed on samples with at least 100 SNVs, excluding samples with all SNVs attributed to artifact signatures (n = 959).

ARBS methylation level

Eight samples were sequenced on the Oxford Nanopore Promethion N24 machine. Methylation levels were called using the modkit package (https://github.com/nanoporetech/modkit) and methrix package (https://github.com/CompEpigen/methrix). Uncovered sites and sites overlapping SNPs were removed, as well as sites with a coverage below 20× in 6 out of the 8 samples. In each sample, the methylation level of each CpG site was calculated as the number of methylated reads divided by the total reads. The get_region_summary function in methrix was then used to generate average methylation levels of CpG sites located in TARBS and NARBS in each sample. Two-sided Wilcoxon rank-sum tests were used to compare TARBS and NARBS methylation levels.

Replication timing–mutation rate association

Early, mid and late replication timing regions were defined as described in the ‘Enrichment analysis of SV classes’ section of the Supplementary Information. For each sample, the mutation rates of each replication timing were calculated as the number of SNVs and indels per Mb. replication timing associations were calculated as the ratio of late to early mutation rates.

To account for sequence-context biases that may differ across replication timing regions, we generated an expected background using SigProfilerSimulator. Within each sample, SNVs and indels were randomly assigned to a new genomic position that preserved their original context (trinucleotide for SNVs and indel-context class for indels). The expected mutation rates for replication timing regions were then computed from the simulated mutations (SNVs + indels per Mb), and a simulated late-to-early replication timing ratio was derived for each sample. Corrected replication timing ratios were obtained by dividing the observed late-to-early ratio by the corresponding simulated ratio. When estimating replication timing association of individual samples, two screening steps were implemented to avoid unreliable estimation of replication timing association due to low mutation number. Firstly, samples with less than 100 mutations were excluded, secondly, samples with less than 10 mutations in either early or late replication timing regions were excluded (n = 959).

Replication timing association of SBS and ID signatures of individual samples was estimated by considering subsets of mutations assigned to each signature as described in the signature assignment section. The same mutation number screening procedure was applied. For background correction in signature analyses, Sigprofiler simulated mutations were assigned to their most probable mutational signature using the same procedure applied to observed mutations and signature-specific corrected replication timing ratios were computed analogously.

Owing to low mutation counts in individual samples when restricting regions to TARBS or NARBS, regional mutation rates were estimated by summing mutations from all samples. A genome mutation rate for all samples was also calculated with the same method applied to compare with TARBS and NARBS. TARBS/NARBS replication timing mutation rates for each IMF were calculated by summing mutations from all samples belonging to each IMF.

Signature timing

To investigate the temporal dynamics of APOBEC- and HRD-associated mutagenesis in prostate tumours with evidence of IMF7 activity, we requantified the contributions of SBS2, SBS13 (APOBEC), SBS3 and ID6 (HRD) in clonal and subclonal epochs. For each donor, clonal and subclonal mutations were analysed separately and signature activities (A_clonal and A_subclonal, respectively) were estimated using non-negative least squares (NNLS; R nnls package v.1.5), following the framework applied in PCAWG⁹³. To compare activity between epochs, we calculated a fold change (FC) statistic for each signature, defined previously⁹³:

where A_clonal and A_subclonal represent the re-estimated activities in the clonal and subclonal epochs, respectively. This statistic reflects dynamic changes in signature activity, with FC > 1 indicating higher activity in the subclonal period.

Predicting clinical outcomes in cases of primary prostate cancer

Preparation of clinical data

The collection and harmonization of clinical data from the PPCG cohort is described previously²⁴ (Supplementary Table 5). Histopathological data included tumour stage and Gleason grade group. Tumour stages were simplified by omitting additional stage identifier after the main identifier (that is, stage T3b was simplified to stage T3). For survival analyses, tumour stage was then categorized into three groups (T2, T3 and T4); Gleason grade group (GG) 1 to 5 was divided into two groups (≤2 and >2); and age at diagnosis was split into two categories (≤55 and >55 years). MFS was used as the primary end point for testing prognostic utility of mutational scenarios. MFS was defined as the number of days from the date of sample collection and to the date of first metastasis.

Survival analyses

Kaplan–Meier estimates (function survfit from the survival⁹⁴ R package) and two-sided Cox proportional hazard models (function coxph from the survminer⁹⁵ R package) were used for survival analysis across patients based on their dominant integrated mutational footprint (IMF; see the ‘Identifying integrated mutational footprints’ section). Cox proportional hazard models were corrected by age at diagnosis, tumour stage, TMB, late-to-early mutation rate and Gleason grade. Wald test was used to evaluate statistical significance of regression coefficients obtained from Cox proportional hazard models.

Given the well-known clinical impact of CIN⁹⁶, we first evaluated differences in MFS between tumours with (≥20 CNAs) and without (<20 CNAs) detectable CIN (Extended Data Fig. 1). As expected, tumours without CIN showed longer MFS times compared to tumours with CIN. Therefore, tumours without CIN were used as reference in the survival analysis comparing MFS across dominant IMFs (Extended Data Table 1).

Furthermore, survival analyses were also performed based on the activities of individual signatures within each tumour. Two-sided Cox proportional hazard models were performed for the whole cohort, and after stratifying samples by the presence or absence of CIN. Cox proportional hazard models were thus used to evaluate the prediction capacity of each mutational signature for MFS (Supplementary Fig. 26). In this case, the study site (country) of origin was also included as covariate to avoid putative batch effects. To ensure robust associations, we filtered out signatures active in less than five samples and only included samples with a minimum number of somatic alterations per type (≥100 SNVs for SBS signature association analyses, ≥10 indels for ID signatures, ≥5 SVs for cSV signatures and ≥20 CNAs for CX signatures).

Associations with clinical variables

To test for associations between signatures and clinical variables, we applied a Welch’s t-test to compare signature activities between samples that did and did not acquire metastasis, as well as between samples with low (≤2) and high (>2) Gleason grade groups (Supplementary Fig. 27) and study site of tumour collection (Supplementary Fig. 30). P values were corrected for multiple testing using the Benjamini–Hochberg method, while Cohen’s d was computed to assess the size of the difference between groups. Comparison analyses were performed for the whole cohort and after stratifying samples based on the presence of CIN. To ensure robust associations, we filtered out signatures active in less than 5 samples and only included samples with a minimum number of somatic alterations per type (≥100 SNVs for SBS signature association analyses, ≥5 indels for ID signatures, ≥5 SVs for cSV signatures and ≥20 CNAs for CX signatures).

Replication of findings in the TCGA-PRAD cohort

We downloaded both raw and processed WGS data, as well as clinical data, from the TCGA-PRAD cohort (May 2025, https://portal.gdc.cancer.gov/). After excluding samples with fewer than 100 SNVs, fewer than 10 indels and fewer than 5 SVs, the final dataset comprised 290 TCGA-PRAD samples, of which 232 (80%) exhibited high levels of CIN (≥20 CNAs).

We quantified activities of the new SBS96D and ID83 signatures extracted in the PPCG cohort in the TCGA-PRAD dataset. Consensus calls for SNVs and indels were used to quantify activities of SBS and ID signatures (see the ‘De novo extraction and assignment of mutational signatures’ section). Replication timing ratio was calculated in the cohort as described above.

The clinical history of each patient was downloaded from the GDC data portal in the form of an XML file, from which we collected overall survival, date of last follow-up, date of biochemical recurrence, tumour stage, Gleason grade, age at diagnosis and PSA. The time to biochemical recurrence was calculated as the time from diagnosis to the biochemical recurrence or death. The time to biochemical recurrence was censored if they were calculated from the last follow-up or treatment end date.

Kaplan–Meier estimates and two-sided Cox proportional hazard models were used for survival analysis across subgroups. Cox proportional hazard models were corrected by Gleason grade. Age at diagnosis was not included as covariate as most TCGA-PRAD samples were from individuals aged over 55 years, and we therefore violated the proportional hazards assumption.

Identifying and benchmarking with simple rearrangement signatures

Simple SV-size-based rearrangement signatures were identified using Palimpsest v.2.0.0 with the default parameters, using PPCG SV calls from the PPCG cohort. Ten simple rearrangement signatures were identified. The six cSV signatures were compared pairwise to the ten simple rearrangement signatures using Spearman correlation.

Demonstrating clinical utility in the HMF-Prostate cohort

We investigated the potential of our integrated mutational footprints as biomarkers for predicting resistance and/or sensitivity to different common therapies for prostate cancer. To achieve this, we used the Hartwig Medical Foundation (HMF) cohort—a real-world retrospective cohort of treated patients with metastatic prostate cancer⁵⁹.

Processing genomic data from the HMF-Prostate cohort

Before testing the clinical performance of IMF-based biomarkers, we downloaded and processed genomic data from this cohort to quantify mutational signatures. IMF activities were next quantified to assign the dominant mutational process, as previously described. Supplementary Table 3 shows the number and frequency of samples assigned to each dominant IMF.

SBS and ID signatures

We downloaded SNVs and indels from HMF, which were called using an in-house pipeline previously described⁹⁷. Each sample VCF file was input to the SigProfilerAssignment tool, which generated the sample-by-mutation type matrix to quantify SBS and ID signatures. We limited the signature set to the SBS and ID signatures extracted in the PPCG cohort.

cSV signatures

We implemented a workflow to systematically analyse cSV signatures in prostate cancer samples, using variant calls from Purple and GRIDSS provided by the HMF pipeline v.5. Each sample aliquot was processed independently; for each aliquot, we extracted purity and ploidy values from the Purple estimates ([sample_id].purple.purity.tsv).

We retained only high-confidence GRIDSS variants (PASS), excluding variants overlapping centromeres or telomeres. SCNA and SV calls were used to extract cSV signatures.

CX signatures

We downloaded PURPLE-derived copy-number profiles from HMF, and processed them as previously described⁶⁰. We then computed the genome-wide distributions of five fundamental copy-number features for each sample, quantified activities of CX signatures extracted in the PPCG cohort using the linear combination decomposition function from YAPSA⁷⁵, and applying the signature-specific threshold for shrinking to zero low-level activities (see the ‘CX signatures’ section for further details).

Emulating phase III randomized control trials for therapy response prediction

We aimed to emulate phase III randomized controlled biomarker trials, classifying patients as biomarker-positive or biomarker-negative and assigning them to either the therapy of interest (experimental arm) or an alternative standard-of-care (SoC; control arm). For each IMF, patients were classified as biomarker positive if their IMF activity exceeded the median among cases with non-zero activity; otherwise, they were classified as biomarker negative. Only patients with clinical response data (enabling calculation of TTF) and with sufficiently high-quality genomic data to compute all four mutational signature types were included (n = 240). TTF was computed as previously described⁶⁰.

Clinical response data covered the full treatment history, although records before enrolment were acknowledged to be less accurate and comprehensive than post-enrolment information. We evaluated the predictive power of the IMFs in the second-line mCRPC setting, in which patients commonly transition empirically between ARPIs and taxanes following first-line therapy. Our analysis focused on these two common therapy types, aiming to move beyond the conventional sequential switching paradigm toward a precision, biomarker-guided approach to therapy selection at second line. No significant differences in TTF were observed between these two therapies (Extended Data Fig. 9), supporting the use of biomarker-positive versus biomarker-negative comparisons to evaluate predictive capacity in this setting.

To decide which therapy–biomarker combinations to analyse, we performed power analyses in line with the Consolidated Standards of Reporting Trials (CONSORT) statement to ensure sufficient statistical power for clinical assessment. To identify the required cohort size to have sufficient power, we performed one-tailed power calculations (β = 0.8, α = 0.05) using censoring and prediction ratio data from the cohort and fixing the hazard ratio at 5 for testing resistance and 1/5 for testing sensitivity (based on previous knowledge⁶⁰). Supplementary Table 4 summarizes the results of the power analyses.

We then compared ARPI-treated patients and taxane-treated in the second-line setting using two-sided Cox proportional hazards models for both biomarker-positive and biomarker-negative cases, with TTF as the primary end point (function coxph from the survival package in R). Cox proportional hazards models were stratified by treatment therapy at first line to control for potential confounding effects of treatment sequencing. This ensures that TTF differences are not attributable to switches of the therapy type at the second line. No information of Gleason grade and tumour stage was available in this cohort. Kaplan–Meier survival curves (survfit function from the survival package in R) were generated to represent differences in treatment effectiveness across treatment arms in a univariate mode. Supplementary Table 5 shows the results of the survival analyses.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

The source code for reproducing analyses and figures is available at GitHub (https://github.com/panprostate/mutational-signatures). A static code repository is available at Zenodo⁹⁸ (https://doi.org/10.5281/ZENODO.18754193).

1. Rebello, R. J. et al. Prostate cancer. Nat. Rev. Dis. Primers 7, 9 (2021).

   Article  PubMed  Google Scholar 

2. Bergengren, O. et al. 2022 update on prostate cancer epidemiology and risk factors—a systematic review. Eur. Urol. 84, 191–206 (2023).

   Article  PubMed  PubMed Central  Google Scholar 

3. Sung, H. et al. Global Cancer Statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J. Clin. 71, 209–249 (2021).

   PubMed  Google Scholar 

4. Wang, A. et al. Characterizing prostate cancer risk through multi-ancestry genome-wide discovery of 187 novel risk variants. Nat. Genet. 55, 2065–2074 (2023).

   Article  CAS  PubMed  PubMed Central  Google Scholar 

5. Valsecchi, A. A. et al. Frequency of germline and somatic BRCA1 and BRCA2 mutations in prostate cancer: an updated systematic review and meta-analysis. Cancers 15, 2435 (2023).

6. Lozano, R. et al. Genetic aberrations in DNA repair pathways: a cornerstone of precision oncology in prostate cancer. Br. J. Cancer 124, 552–563 (2021).

   Article  CAS  PubMed  Google Scholar 

7. Fang, B. et al. The somatic mutational landscape of mismatch repair deficient prostate cancer. J. Clin. Med. 12, 623 (2023).

   Article  CAS  PubMed  PubMed Central  Google Scholar 

8. Abida, W. et al. Analysis of the prevalence of microsatellite instability in prostate cancer and response to immune checkpoint blockade. JAMA Oncol. 5, 471–478 (2019).

   Article  PubMed  PubMed Central  Google Scholar 

9. Viswanathan, S. R. et al. Structural alterations driving castration-resistant prostate cancer revealed by linked-read genome sequencing. Cell 174, 433–447 (2018).

   Article  CAS  PubMed  PubMed Central  Google Scholar 

10. Antonarakis, E. S. et al. CDK12-altered prostate cancer: clinical features and therapeutic outcomes to standard systemic therapies, poly (ADP-Ribose) polymerase inhibitors, and PD-1 inhibitors. JCO Precis. Oncol. 4, 370–381 (2020).

    Article  PubMed  PubMed Central  Google Scholar 

11. Wu, Y.-M. et al. Inactivation of CDK12 delineates a distinct immunogenic class of advanced prostate cancer. Cell 173, 1770–1782 (2018).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

12. Pomerantz, M. M. et al. The androgen receptor cistrome is extensively reprogrammed in human prostate tumorigenesis. Nat. Genet. 47, 1346–1351 (2015).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

13. Wang, Q. et al. Androgen receptor regulates a distinct transcription program in androgen-independent prostate cancer. Cell 138, 245–256 (2009).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

14. Morova, T. et al. Androgen receptor-binding sites are highly mutated in prostate cancer. Nat. Commun. 11, 832 (2020).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

15. Baca, S. C. et al. Punctuated evolution of prostate cancer genomes. Cell 153, 666–677 (2013).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

16. Alexandrov, L. B. et al. The repertoire of mutational signatures in human cancer. Nature 578, 94–101 (2020).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

17. Drews, R. M. et al. A pan-cancer compendium of chromosomal instability. Nature 606, 976–983 (2022).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

18. Li, Y. et al. Patterns of somatic structural variation in human cancer genomes. Nature 578, 112–121 (2020).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

19. Gerhauser, C. et al. Molecular evolution of early-onset prostate cancer identifies molecular risk markers and clinical trajectories. Cancer Cell 34, 996–1011 (2018).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

20. Yamaguchi, T. N. et al. The germline and somatic origins of prostate cancer heterogeneity. Cancer Discov. 15, 988–1017 (2025).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

21. Wedge, D. C. et al. Sequencing of prostate cancers identifies new cancer genes, routes of progression and drug targets. Nat. Genet. 50, 682–692 (2018).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

22. The ICGC/TCGA Pan-Cancer Analysis of Whole Genomes Consortium Pan-cancer analysis of whole genomes. Nature 578, 82–93 (2020).

    Article  ADS  CAS  Google Scholar 

23. Fraser, M. et al. Genomic hallmarks of localized, non-indolent prostate cancer. Nature 541, 359–364 (2017).

    Article  ADS  CAS  PubMed  Google Scholar 

24. Jakobsdottir, G. M. et al. The Pan Prostate Cancer Group Dataset - Genomic, Transcriptomic and Methylation data for Clinical Discovery. Sci. Data https://doi.org/10.1038/s41597-026-08181-4 (in the press).

25. de Bono, J. S. et al. Central, prospective detection of homologous recombination repair gene mutations (HRRm) in tumour tissue from >4000 men with metastatic castration-resistant prostate cancer (mCRPC) screened for the PROfound study. Ann. Oncol. 30, v328–v329 (2019).

    Article  Google Scholar 

26. Cancer Genome Atlas Research Network. The molecular taxonomy of primary prostate cancer. Cell 163, 1011–1025 (2015).

    Article  Google Scholar 

27. Rodriguez-Fos, E. et al. Mutational topography reflects clinical neuroblastoma heterogeneity. Cell Genom. 3, 100402 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

28. Degasperi, A. et al. Substitution mutational signatures in whole-genome-sequenced cancers in the UK population. Science 376, abl9283 (2022).

29. Menghi, F. et al. The tandem duplicator phenotype is a prevalent genome-wide cancer configuration driven by distinct gene mutations. Cancer Cell 34, 197–210 (2018).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

30. Tomkova, M., Tomek, J., Kriaucionis, S. & Schuster-Böckler, B. Mutational signature distribution varies with DNA replication timing and strand asymmetry. Genome Biol. 19, 129 (2018).

    Article  PubMed  PubMed Central  Google Scholar 

31. Polak, P. et al. Cell-of-origin chromatin organization shapes the mutational landscape of cancer. Nature 518, 360–364 (2015).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

32. Sakofsky, C. J. et al. Repair of multiple simultaneous double-strand breaks causes bursts of genome-wide clustered hypermutation. PLoS Biol. 17, e3000464 (2019).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

33. Maciejowski, J. et al. APOBEC3-dependent kataegis and TREX1-driven chromothripsis during telomere crisis. Nat. Genet. 52, 884–890 (2020).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

34. Pomerantz, M. M. et al. Prostate cancer reactivates developmental epigenomic programs during metastatic progression. Nat. Genet. 52, 790–799 (2020).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

35. Lee, C. A., Abd-Rabbo, D. & Reimand, J. Functional and genetic determinants of mutation rate variability in regulatory elements of cancer genomes. Genome Biol. 22, 133 (2021).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

36. Alexandrov, L. B. et al. Clock-like mutational processes in human somatic cells. Nat. Genet. 47, 1402–1407 (2015).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

37. Taberlay, P. C., Statham, A. L., Kelly, T. K., Clark, S. J. & Jones, P. A. Reconfiguration of nucleosome-depleted regions at distal regulatory elements accompanies DNA methylation of enhancers and insulators in cancer. Genome Res. 24, 1421–1432 (2014).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

38. Parolia, A. et al. Distinct structural classes of activating FOXA1 alterations in advanced prostate cancer. Nature 571, 413–418 (2019).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

39. Lemma, R. B. et al. Pioneer transcription factors are associated with the modulation of DNA methylation patterns across cancers. Epigenet. Chromatin 15, 13 (2022).

    Article  CAS  Google Scholar 

40. Du, Q. et al. Replication timing and epigenome remodelling are associated with the nature of chromosomal rearrangements in cancer. Nat. Commun. 10, 416 (2019).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

41. Stamatoyannopoulos, J. A. et al. Human mutation rate associated with DNA replication timing. Nat. Genet. 41, 393–395 (2009).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

42. Keshet, I., Lieman-Hurwitz, J. & Cedar, H. DNA methylation affects the formation of active chromatin. Cell 44, 535–543 (1986).

    Article  CAS  PubMed  Google Scholar 

43. Davies, H. et al. HRDetect is a predictor of BRCA1 and BRCA2 deficiency based on mutational signatures. Nat. Med. 23, 517–525 (2017).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

44. Jia, P. et al. MSIsensor-pro: fast, accurate, and matched-normal-sample-free detection of microsatellite instability. Genom. Proteom. Bioinform. 18, 65–71 (2020).

    Article  Google Scholar 

45. Kucab, J. E. et al. A compendium of mutational signatures of environmental agents. Cell 177, 821–836 (2019).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

46. Jager, M. et al. Deficiency of nucleotide excision repair is associated with mutational signature observed in cancer. Genome Res. 29, 1067–1077 (2019).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

47. Cooke, M. S., Evans, M. D., Dizdaroglu, M. & Lunec, J. Oxidative DNA damage: mechanisms, mutation, and disease. FASEB J. 17, 1195–1214 (2003).

    Article  CAS  PubMed  Google Scholar 

48. Singh, V. K., Rastogi, A., Hu, X., Wang, Y. & De, S. Mutational signature SBS8 predominantly arises due to late replication errors in cancer. Commun Biol. 3, 421 (2020).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

49. Daniels, G. et al. Cytoplasmic, full length and novel cleaved variant, TBLR1 reduces apoptosis in prostate cancer under androgen deprivation. Oncotarget 7, 39556–39571 (2016).

    Article  PubMed  PubMed Central  Google Scholar 

50. Hsieh, C.-L. et al. PLZF, a tumor suppressor genetically lost in metastatic castration-resistant prostate cancer, is a mediator of resistance to androgen deprivation therapy. Cancer Res. 75, 1944–1948 (2015).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

51. King, R. W. et al. A 20S complex containing CDC27 and CDC16 catalyzes the mitosis-specific conjugation of ubiquitin to cyclin B. Cell 81, 279–288 (1995).

    Article  CAS  PubMed  Google Scholar 

52. Hutchins, J. R. A. et al. Systematic analysis of human protein complexes identifies chromosome segregation proteins. Science 328, 593–599 (2010).

    Article  ADS  MathSciNet  CAS  PubMed  PubMed Central  Google Scholar 

53. Parker, A. L., Kavallaris, M. & McCarroll, J. A. Microtubules and their role in cellular stress in cancer. Front. Oncol. 4, 153 (2014).

    Article  PubMed  PubMed Central  Google Scholar 

54. Chan, K. L., Palmai-Pallag, T., Ying, S. & Hickson, I. D. Replication stress induces sister-chromatid bridging at fragile site loci in mitosis. Nat. Cell Biol. 11, 753–760 (2009).

    Article  CAS  PubMed  Google Scholar 

55. Riaz, N. et al. Pan-cancer analysis of bi-allelic alterations in homologous recombination DNA repair genes. Nat. Commun. 8, 857 (2017).

    Article  ADS  PubMed  PubMed Central  Google Scholar 

56. Gong, T. et al. Genome-wide interrogation of structural variation reveals novel African-specific prostate cancer oncogenic drivers. Genome Med. 14, 100 (2022).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

57. Blattner, M. et al. SPOP mutation drives prostate tumorigenesis in vivo through coordinate regulation of PI3K/mTOR and AR signaling. Cancer Cell 31, 436–451 (2017).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

58. Weischenfeldt, J. et al. Integrative genomic analyses reveal an androgen-driven somatic alteration landscape in early-onset prostate cancer. Cancer Cell 23, 159–170 (2013).

    Article  CAS  PubMed  Google Scholar 

59. Priestley, P. et al. Pan-cancer whole-genome analyses of metastatic solid tumours. Nature 575, 210–216 (2019).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

60. Thompson, J. S. et al. Predicting resistance to chemotherapy using chromosomal instability signatures. Nat. Genet. https://doi.org/10.1038/s41588-025-02233-y (2025).

61. Supek, F. & Lehner, B. Differential DNA mismatch repair underlies mutation rate variation across the human genome. Nature 521, 81–84 (2015).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

62. Takahashi, N. et al. Replication stress defines distinct molecular subtypes across cancers. Cancer Res. Commun. 2, 503–517 (2022).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

63. Cuzick, J. et al. Prognostic value of an RNA expression signature derived from cell cycle proliferation genes in patients with prostate cancer: a retrospective study. Lancet Oncol. 12, 245–255 (2011).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

64. Maya-Mendoza, A. et al. High speed of fork progression induces DNA replication stress and genomic instability. Nature 559, 279–284 (2018).

    Article  ADS  CAS  PubMed  Google Scholar 

65. Gao, R. et al. Punctuated copy number evolution and clonal stasis in triple-negative breast cancer. Nat. Genet. 48, 1119–1130 (2016).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

66. Swami, U. et al. SPOP mutations as a predictive biomarker for androgen receptor axis-targeted therapy in de novo metastatic castration-sensitive prostate cancer. Clin. Cancer Res. 28, 4917–4925 (2022).

    Article  CAS  PubMed  Google Scholar 

67. Parreno, V. et al. Transient loss of Polycomb components induces an epigenetic cancer fate. Nature 629, 688–696 (2024).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

68. Nik-Zainal, S. et al. The life history of 21 breast cancers. Cell 149, 994–1007 (2012).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

69. Isidro, C.-C., Xi, R. & Park, P. J. ShatterSeek: an R package for the detection of chromothripsis from next-generation sequencing (NGS) data (2022).

70. Xing, R. et al. Whole-genome sequencing reveals novel tandem-duplication hotspots and a prognostic mutational signature in gastric cancer. Nat. Commun. 10, 2037 (2019).

    Article  ADS  PubMed  PubMed Central  Google Scholar 

71. Menghi, F. et al. The tandem duplicator phenotype as a distinct genomic configuration in cancer. Proc. Natl Acad. Sci. USA 113, E2373–E2382 (2016).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

72. Islam, S. M. A. et al. Uncovering novel mutational signatures by de novo extraction with SigProfilerExtractor. Cell Genom. 2, 100179 (2022).

73. Gaujoux, R. & Seoighe, C. A flexible R package for nonnegative matrix factorization. BMC Bioinform. 11, 367 (2010).

    Article  Google Scholar 

74. Shinde, J. et al. Palimpsest: an R package for studying mutational and structural variant signatures along clonal evolution in cancer. Bioinformatics 34, 3380–3381 (2018).

    Article  CAS  PubMed  Google Scholar 

75. Hübschmann, D. et al. Analysis of mutational signatures with yet another package for signature analysis. Genes Chromosomes Cancer 60, 314–331 (2021).

    Article  PubMed  Google Scholar 

76. Akamine, Y. et al. Application of hierarchical clustering to multi-parametric MR in prostate: differentiation of tumor and normal tissue with high accuracy. Magn. Reson. Imaging 74, 90–95 (2020).

    Article  PubMed  Google Scholar 

77. Strauss, T. & von Maltitz, M. J. Generalising Ward’s method for use with Manhattan distances. PLoS ONE 12, e0168288 (2017).

    Article  PubMed  PubMed Central  Google Scholar 

78. Bailey, M. H. et al. Comprehensive characterization of cancer driver genes and mutations. Cell 173, 371–385 (2018).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

79. Jassal, B. et al. The reactome pathway knowledgebase. Nucleic Acids Res. 48, D498–D503 (2020).

    CAS  PubMed  PubMed Central  Google Scholar 

80. Martínez-Jiménez, F. et al. A compendium of mutational cancer driver genes. Nat. Rev. Cancer 20, 555–572 (2020).

    Article  PubMed  Google Scholar 

81. Sondka, Z. et al. The COSMIC Cancer Gene Census: describing genetic dysfunction across all human cancers. Nat. Rev. Cancer 18, 696–705 (2018).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

82. Armenia, J. et al. The long tail of oncogenic drivers in prostate cancer. Nat. Genet. 50, 645–651 (2018).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

83. Takeda, D. Y. et al. A somatically acquired enhancer of the androgen receptor is a noncoding driver in advanced prostate cancer. Cell 174, 422–432 (2018).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

84. Boersma, V. et al. MAD2L2 controls DNA repair at telomeres and DNA breaks by inhibiting 5’ end resection. Nature 521, 537–540 (2015).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

85. Isono, M. et al. BRCA1 directs the repair pathway to homologous recombination by promoting 53BP1 dephosphorylation. Cell Rep. 18, 520–532 (2017).

    Article  CAS  PubMed  Google Scholar 

86. Bergstrom, E. N. et al. SigProfilerMatrixGenerator: a tool for visualizing and exploring patterns of small mutational events. BMC Genom. 20, 685 (2019).

    Article  Google Scholar 

87. Ma, X. et al. Pan-cancer genome and transcriptome analyses of 1,699 paediatric leukaemias and solid tumours. Nature 555, 371–376 (2018).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

88. Poulsgaard, G. A., Sørensen, S. G., Juul, R. I., Nielsen, M. M. & Pedersen, J. S. Sequence dependencies and mutation rates of localized mutational processes in cancer. Genome Med. 15, 63 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

89. Adler, N. et al. Mutational processes of tobacco smoking and APOBEC activity generate protein-truncating mutations in cancer genomes. Sci. Adv. 9, eadh3083 (2023).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

90. Medo, M., Ng, C. K. Y. & Medová, M. A comprehensive comparison of tools for fitting mutational signatures. Nat. Commun. 15, 9467 (2024).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

91. Buffa, F. M., Harris, A. L., West, C. M. & Miller, C. J. Large meta-analysis of multiple cancers reveals a common, compact and highly prognostic hypoxia metagene. Br. J. Cancer 102, 428–435 (2010).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

92. Corces, M. R. et al. The chromatin accessibility landscape of primary human cancers. Science 362, 6413 (2018).

93. Gerstung, M. et al. The evolutionary history of 2,658 cancers. Nature 578, 122–128 (2020).

    Article  ADS  CAS  PubMed  PubMed Central  Google Scholar 

94. Therneau, T. M. Survival: survival analysis https://doi.org/10.32614/CRAN.package.survival (2023).

95. Kassambara, A., Kosinski, M., Biecek, P. & Fabian, S. Drawing survival curves using ‘ggplot2’. (2021).

96. McGranahan, N., Burrell, R. A., Endesfelder, D., Novelli, M. R. & Swanton, C. Cancer chromosomal instability: therapeutic and diagnostic challenges. EMBO Rep. 13, 528–538 (2012).

    Article  CAS  PubMed  PubMed Central  Google Scholar 

97. Martínez-Jiménez, F. et al. Pan-cancer whole-genome comparison of primary and metastatic solid tumours. Nature 618, 333–341 (2023).

    Article  ADS  PubMed  PubMed Central  Google Scholar 

98. Gruber, A., Vidas Olsen, A., Hernando, B. & Cheng, K. Integrated signatures define mutational processes in prostate cancer. Zenodo https://doi.org/10.5281/ZENODO.18754193 (2026).

Download references

Author notes

1. These authors contributed equally: Andreas J. Gruber, André V. Olsen, Barbara Hernando, Kevin C. L. Cheng

2. These authors jointly supervised this work: Geoff Macintyre, Jüri Reimand & Joachim Weischenfeldt

Authors and Affiliations

1. Department of Biology, University of Konstanz, Konstanz, Germany

   Andreas J. Gruber

2. Biotech Research & Innovation Centre (BRIC) - University of Copenhagen, Copenhagen, Denmark

   André V. Olsen, Francesco Favero, Daria Kiriy, Etsehiwot G. Girma, Balthasar C. Schlotmann & Joachim Weischenfeldt

3. Finsen Laboratory, Copenhagen University Hospital - Rigshospitalet, Copenhagen, Denmark

   André V. Olsen, Francesco Favero, Daria Kiriy, Etsehiwot G. Girma, Balthasar C. Schlotmann & Joachim Weischenfeldt

4. Spanish National Cancer Research Centre (CNIO), Madrid, Spain

   Barbara Hernando, Marina Torres, Ángel Fernández-Sanromán, Juan Maria Roldan-Romero & Geoff Macintyre

5. Computational Biology Program, Ontario Institute for Cancer Research (OICR), Toronto, Ontario, Canada

   Kevin C. L. Cheng, Diogo Pellegrina & Jüri Reimand

6. Department of Medical Biophysics, University of Toronto, Toronto, Ontario, Canada

   Kevin C. L. Cheng, Diogo Pellegrina, Housheng Hansen He & Jüri Reimand

7. Division of Cancer Epigenomics, German Cancer Research Center (DKFZ), Heidelberg, Germany

   Clarissa Gerhäuser, Pavlo Lutsik, Christoph Plass, Jessica Heilmann, Michael Scherer & Rajbir N. Batra

8. Manchester Cancer Research Centre and Cancer Research UK Manchester Institute, The University of Manchester, Manchester, UK

   Lucy Barton, Vanessa M. Hayes, Atef Sahli, Arfa Mehmood, Christopher Wirth, Esther Baena, Parsa Pirhady, Simon Pacey, Ronnie Rodrigues Pereira, Samson Olofinsae, Vivien Holmes, Xiaotong He, Robert G. Bristow & David C. Wedge

9. Prostate Cancer Research Center, Faculty of Medicine and Health Technology, Tampere University, Tampere, Finland

   G. Steven Bova & Matti Nykter

10. Norwich Medical School, University of East Anglia, Norwich Research Park, Norwich, UK

    Daniel S. Brewer, Abraham Gihawi, Vanessa M. Hayes, Jingjing Zhang, Valeriia Haberland & Colin S. Cooper

11. The Earlham Institute, Norwich, UK

    Daniel S. Brewer

12. The Institute of Cancer Research, London, UK

    Mark N. Brook, Zsofia Kote-Jarai, Chris Foster, Nening Dennis, Wi Burns & Ros A. Eeles

13. Division of Applied Bioinformatics, German Cancer Research Center (DKFZ), Heidelberg, Germany

    Benedikt Brors, Sebastian Uhrig & Lars Feuerbach

14. German Cancer Consortium (DKTK), Core Center Heidelberg, and National Center for Tumor Diseases (NCT), Heidelberg, Germany

    Benedikt Brors & Holger Sültmann

15. Medical Faculty Heidelberg and Faculty of Biosciences, Heidelberg University, Heidelberg, Germany

    Benedikt Brors

16. Wellcome Sanger Institute, Hinxton, UK

    Adam Butler, Gregory Leeman, Yaobo Xu & Keiran Raine

17. CeRePP (Centre de Recherche sur les Pathologies Prostatiques et Urologiques), Paris, France

    Géraldine Cancel-Tassin & Olivier Cussenot

18. Sorbonne Université, GRC n°5 Predictive Onco-Urology, APHP, Tenon Hospital, Paris, France

    Géraldine Cancel-Tassin

19. Collaborative Centre for Genomic Cancer Medicine, The Victorian Comprehensive Cancer Centre, Parkville, Victoria, Australia

    Niall M. Corcoran, Christopher M. Hovens, Bethany K. Campbell & Peter Georgeson

20. Department of Surgery, The University of Melbourne, Parkville, Victoria, Australia

    Niall M. Corcoran, Anis A. Hamid, Christopher M. Hovens, Bernard Pope, Anne Nguyen, Beng Lim, Bethany K. Campbell, Corinna Grima, Justin S. Peters, Ken Chow, Matthew K. H. Hong, Michael J. Clarkson, Michael Kerger, Natalie Kurganovs, Patrick McCoy, Philip Dundee, Pramit M. Phal, Ryan Stuchbery & Tony Costello

21. Department of Urology, Peninsula Health, Frankston, Victoria, Australia

    Niall M. Corcoran

22. Division of Urology, Royal Melbourne Hospital, Parkville, Victoria, Australia

    Niall M. Corcoran, Christopher M. Hovens, Bethany K. Campbell & Philip Dundee

23. Walter and Eliza Hall Institute of Medical Research, Parkville, Victoria, Australia

    Niall M. Corcoran, Christopher M. Hovens, Ramyar Molania, Anthony T. Papenfuss, Alan F. Rubin, Bethany K. Campbell, Breon Feran, Jennifer Ureta, Jocelyn Sietsma Penington, Justin Bedő, Marek Cmero, Matthew Wakefield, Ruining Dong, Stefano Mangiola, William J. Hutchison & Yunfan Fu

24. Department of Urology, Western Health, Footscray, Victoria, Australia

    Niall M. Corcoran

25. Ancestry and Health Genomics Laboratory, Charles Perkins Centre, School of Medical Sciences, Faculty of Medicine and Health, University of Sydney, Camperdown, New South Wales, Australia

    Abraham Gihawi, Vanessa M. Hayes, Jue Jiang & Weerachai Jaratlerdsiri

26. Department of Surgery, University of Cambridge, Cambridge, UK

    Vincent J. Gnanapragasam, Melissa Cheung, Radoslaw Lach & Toby Milne-Clark

27. Department of Urology, Cambridge University Hospitals NHS Trust, Cambridge, UK

    Vincent J. Gnanapragasam, Melissa Cheung, Radoslaw Lach & Toby Milne-Clark

28. University of Cambridge & Cambridge University Hospitals NHS Trust, Cambridge, UK

    Vincent J. Gnanapragasam, Melissa Cheung, Radoslaw Lach & Toby Milne-Clark

29. Cambridge Urology Translational Research and Clinical Trials Office, Cambridge Biomedical Campus, Addenbrooke’s Hospital, Cambridge, UK

    Vincent J. Gnanapragasam, Melissa Cheung, Radoslaw Lach & Toby Milne-Clark

30. Genitourinary Oncology Service, Memorial Sloan Kettering Cancer Center, New York, NY, USA

    Anis A. Hamid

31. School of Health Systems and Public Health, University of Pretoria, Pretoria, South Africa

    Vanessa M. Hayes

32. Princess Margaret Cancer Centre, University Health Network, Toronto, Ontario, Canada

    Housheng Hansen He & Alejandro Berlin

33. Department of Clinical Pathology, The University of Melbourne, Parkville, Victoria, Australia

    Christopher M. Hovens & Peter Georgeson

34. Department of Pathology and Laboratory Medicine, Weill Cornell Medicine, New York, NY, USA

    Eddie L. Imada, Francesca Khani, Massimo Loda, Luigi Marchionni, Lucio R. Queiroz, Brian Robinson, Claudio Zanettini, Angelo Corso Faini, Filippo Pederzoli, Giuseppe N. Fanelli, Hubert Pakula, Pier Vitale Nuzzo, Ryan Carelli, Silvia Rodrigues & Wikum Dinalankara

35. Division of Cancer Sciences, The University of Manchester, Manchester, UK

    G. Maria Jakobsdottir & Robert G. Bristow

36. Christie Hospital, The Christie NHS Foundation Trust, Manchester Academic Health Science Centre, Manchester, UK

    G. Maria Jakobsdottir, Arfa Mehmood & David C. Wedge

37. Melbourne Bioinformatics, The University of Melbourne, Parkville, Victoria, Australia

    Chol-Hee Jung, Bernard Pope, Daniel J. Park & Grace Hall

38. Weill Cornell Medicine, New York, NY, USA

    Francesca Khani & Brian Robinson

39. Department of Clinical Medicine, Aarhus University, Aarhus, Denmark

    Philippe Lamy, Karina D. Sørensen & Michael Borre

40. Department of Molecular Medicine, Aarhus University Hospital, Aarhus, Denmark

    Philippe Lamy & Karina D. Sørensen

41. The Broad Institute of Harvard and MIT, Cambridge, MA, USA

    Massimo Loda

42. The New York Genome Center, New York, NY, USA

    Massimo Loda

43. Department of Oncology, KU Leuven, Leuven, Belgium

    Pavlo Lutsik

44. Department of Medical Biology, The University of Melbourne, Melbourne, Victoria, Australia

    Ramyar Molania, Anthony T. Papenfuss, Alan F. Rubin, Jennifer Ureta, Justin Bedő, Marek Cmero, Matthew Wakefield, Ruining Dong, Stefano Mangiola, William J. Hutchison & Yunfan Fu

45. Australian BioCommons, The University of Melbourne, Parkville, Victoria, Australia

    Bernard Pope

46. Department of Medicine, Central Clinical School, Faculty of Medicine Nursing and Health Sciences, Monash University, Melbourne, Victoria, Australia

    Bernard Pope

47. Genome Biology Unit, European Molecular Biology Laboratory (EMBL), Heidelberg, Germany

    Tobias Rausch & Jan Korbel

48. Jonsson Comprehensive Cancer Center, University of California, Los Angeles, Los Angeles, CA, USA

    Takafumi N. Yamaguchi

49. Institute of Pathology, University Medical Centre Hamburg - Eppendorf (UKE), Hamburg, Germany

    Ronald Simon & Guido Sauter

50. Royal Marsden NHS Foundation Trust, London, UK

    Ros A. Eeles

51. Christie NHS Foundation Trust, Manchester, UK

    Robert G. Bristow

52. Department of Urology, Charité-Universitätsmedizin Berlin, Corporate Member of Freie Universität Berlin and Humboldt Universität zu Berlin, Berlin, Germany

    Kira Furlano, Thorsten Schlomm & Joachim Weischenfeldt

53. Department of Molecular Genetics, University of Toronto, Toronto, Ontario, Canada

    Jüri Reimand

54. CHU de Québec – Université Laval, Research Center, Oncology Axis, Québec, Quebec, Canada

    Alain Bergeron & Yves Fradet

55. Barts Cancer Institute, Queen Mary University of London, London, UK

    Alastair D. Lamb & Yong-Jie Lu

56. Nuffield Department of Surgical Sciences, University of Oxford, Oxford, UK

    Aleksandra S. Ziuboniewicz, Andrew Erickson, Dan J. Woodcock, Emre Esentürk, Ian Mills & Noora Al-Muftah

57. TissuPath, Mt Waverley, Victoria, Australia

    Andrew Ryan & Sam Norden

58. School of Mathematics and Statistics, University of St Andrews, St Andrews, UK

    Andy G. Lynch & Radoslaw Lach

59. School of Medicine, University of St Andrews, St Andrews, UK

    Andy G. Lynch

60. Department of Pathology, Cambridge University Hospitals, NHS Foundation Trust & University of Cambridge, Addenbrookes Hospital, Cambridge Biomedical Campus, Cambridge, UK

    Anne Warren

61. European Bioinformatics Institute (EMBL-EBI), Wellcome Genome Campus, Hinxton, UK

    Aravind Sankar, Daniel Barrowdale, Giselle Kerry & Thomas Keane

62. Early Cancer Institute, Department of Oncology, University of Cambridge, Cambridge, UK

    Charlie E. Massie, Harveer Dev, Rajbir N. Batra & Toby Milne-Clark

63. German Cancer Research Center (DKFZ), Heidelberg, Germany

    Chen Hong, Lars Feuerbach & Lena Voith von Voithenberg

64. Vancouver Prostate Centre, Vancouver, British Columbia, Canada

    Colin Collins

65. Department of Urology, University Hospital Heidelberg, Heidelberg, Germany

    Fabian Falkenbach

66. School of Engineering, Newcastle University, Newcastle upon Tyne, UK

    Jingjing Zhang

67. Martini-Klinik, University Medical Centre Hamburg - Eppendorf (UKE), Hamburg, Germany

    Markus Graefen & Pierre Tennstedt

68. Department of Urology, Aarhus University Hospital, Aarhus, Denmark

    Michael Borre

69. Department of Computational Biomedicine, Cedars-Sinai Medical Center, Los Angeles, CA, USA

    Mohamed Omar

70. The Francis Crick Institute, London, UK

    Nana Mensah

71. The University of Texas MD Anderson Cancer Center, Houston, TX, USA

    Peter Van Loo & Yidan Pan

72. Department of Computer Science, Brunel University of London, Uxbridge, UK

    Valeriia Haberland

73. Computational Genomics Group, Charles Perkins Centre, School of Medical Sciences, Faculty of Medicine and Health, University of Sydney, Camperdown, New South Wales, Australia

    Weerachai Jaratlerdsiri

Consortia

Contributions

A.J.G., A.V.O., B.H., K.C.L.C., C. Gerhäuser, M.T., F. Favero, D.K., A.F.-S., J.M.R.-R. and L.B. contributed to the formal analysis presented in this study. J.W., A.J.G., A.V.O., B.H., K.C.L.C., T.S., G.M. and J.R. wrote the manuscript. A.J.G., A.V.O., B.H. and K.C.L.C. produced and contributed to the visualizations of the study. J.W. conceived and designed the study. J.W., J.R. and G.M. supervised the work. C.S.C., D.C.W., R.A.E., D.S.B., Z.K.-J. and R.G.B. made a substantial contribution to the organization and conduct of the study and critiqued the output for important intellectual content. R. Simon and G.S. contributed samples and pathology workup. A.J.G., A.V.O., B.H., K.C.L.C., C. Gerhäuser, M.T., F. Favero, D.K., A.F.-S., J.M.R-R., L.B., D.P., G.S.B., D.S.B., M.N.B, B.B., A. Butler, G.C.-T., N.M.C., O.C., A.G., E.G.G., V.J.G., A.A.H., V.M.H., H.H.H., C.M.H., E.L.I., G.M.M.J., C.-h.J., F.K., Z.K.-J., P. Lamy, G.L., M.L., P. Lutsik, L.M., R.M., A.T.P., B.P., L.R.Q., T.R., J.K., B.R., A. Sahli, K.D.S., T.N.Y., S.U., Y.X., C.Z., N.A.-M., E.B., D.B., J.B., A. Bergeron, A. Berlin, M.B., W.B., B.K.C., R.C., M. Cheung, K.C., M.J.C., M. Cmero, C.C., A.C.F., T.C., N.D., H.D., W.D., R.D., P.D., C.F., A.E., E.E., F. Falkenbach, G.N.F., B.F., L.F., Y. Fradet, Y. Fu, K.F., P.G., M.G., C. Grima, V. Haberland, G.A.H., X.H., J.H., V. Holmes, C.H., M.K.H.H., W.J.H., W.J., J.J., T.K., M.K., G.K., N.K., R.L., A.D.L., B.L., Y.-J.L., A.G.L., S.M., C.E.M., P.M., A.M., N.M., I.M., T.M.-C., A.N., S.N., P.V.N., M.N., S.O., M.O., H.P., Y.P., D.J.P., F.P., J.P., J.S.P., P.M.P., P.P., C.P., K.R., S.R., R.R.P., A.F.R., A.R., A. Sankar, B.C.S., R. Stuchbery, H.S., P.T., J.U., R.R.N.B., M.S., S.P., P.V.L., L.V.v.V., M.W., A.W., C.W., D.J.W., J.Z., A.S.Z., R. Simon, G.S., R.A.E., C.S.C., R.G.B., D.C.W., T.S., G.M., J.R. and J.W. provided access to data and/or contributed to gathering, processing and curating data. All of the authors had access to all data in the study. All of the authors contributed to the review and the editing of the manuscript. All of the authors approved the manuscript before submission.

Corresponding authors

Correspondence to Geoff Macintyre, Jüri Reimand or Joachim Weischenfeldt.

Competing interests

V.J.G. has received honoraria from Janssen and JEB as a speaker; and is a member of an external expert committee to AstraZeneca, UK. R.A.E. has received honoraria from GU-ASCO, Janssen, University of Chicago and Dana Farber-Cancer Institute as a speaker; has received educational honorarium from Bayer and Ipsen; is a member of an external expert committee to AstraZeneca, UK; is a member of active surveillance for the Movember Committee; is a member of the scientific advisory board of Our Future Health; and undertakes private practice as a sole trader. G.M. is co-founder, director and shareholder of Tailor Bio. The other authors declare no competing interests.

Peer review information

Nature thanks Christopher Barbieri, who co-reviewed with Andrea Sboner, and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available.

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

Reprints and permissions