Poor Prognosis among Radiation-Associated Bladder Cancer Is Defined by Clinicogenomic Features

The development of bladder cancer following pelvic radiotherapy (RT) for prostate cancer is an increasingly recognized phenomenon (1–4). More than half of patients with intermediate or high-risk prostate cancer initially undergo RT, representing nearly 150,000 patients per year (5–7). In total, 1% to 2% of patients with prostate cancer develop subsequent bladder cancer, and patients who receive RT as primary treatment for prostate cancer are 1.5 to 3 times more likely to develop bladder cancer than those who received radical prostatectomy (refs. 3, 8–11). The risk of bladder cancer increases as the duration of time since RT increases, with a majority of cases occurring more than 5 or 10 years after RT (12, 13). RT-associated bladder cancer also has a higher stage and grade at initial diagnosis than de novo bladder cancer, suggesting that the molecular underpinnings of these entities may be distinct (14–17).

Numerous studies have shown a relationship among ionizing radiation, urothelial dysplasia, and DNA damage (18–20). The predilection to tumorigenesis after radiation exposure varies with the type and dose of radiation, organ studied, and patient characteristics, which determine the likelihood of DNA damage and the DNA repair capacity of radiation-exposed cells (18, 21). Samples of bladder epithelium from people living in radio-contaminated areas near the Chernobyl accident demonstrated a high frequency of TP53 mutations and a high frequency of specific single bp substitutions (19, 20). Such mutational patterns and RT-induced field defects may increase an individual’s risk for bladder cancer development while imprinting a distinct molecular signature characteristic of RT-associated bladder cancer. For example, prior data have shown an increased number of balanced inversions in RT-associated cancers (22). Additionally, the only other study, to our knowledge, that genomically characterized radiation-specific bladder cancer noted an increased number of short insertions and deletions (indels) compared with an independent control (CTRL) cohort (17).

Here, we report the clinicogenomic characterization of the largest cohort of RT-associated bladder cancers to date. We show that bladder tumors developing after pelvic radiation are associated with distinct mutational signatures, which, along with CDKN2A loss and a history of smoking or a smoking-related mutational signature, have a poor prognosis. These data also suggest a need for a prospective study to determine if patients with high-risk clinical features (e.g., smoking and/or occupational exposures) may benefit from noninvasive genomic testing, prior to pelvic radiation, to guide prostate cancer treatment decisions.

Patients with muscle-invasive bladder cancer and a prior history of RT for prostate cancer were retrospectively identified from two institutions [University of North Carolina (UNC) and Icahn School of Medicine at Mount Sinai] using pathology databases, electronic medical records, and clinical tumor board list review for inclusion into the RT-associated bladder cancer cohort. Included patients were diagnosed with bladder cancer between January 01, 2000, and May 15, 2015, and had tumor tissue available for next-generation sequencing (NGS). In addition, all patients at UNC with bladder cancer that had received standard-of-care NGS via the Foundation Medicine (FMI) FoundationOne CDx were reviewed and annotated for inclusion into the radiation-associated cohort or a CTRL cohort of non–RT-associated bladder cancer. Patients who were sequenced using deprecated bait sets or diagnosed with upper tract urothelial carcinoma were excluded from the analysis. Written informed consent was provided by all the study participants, and the protocol was approved by the participating institutional review boards: UNC Institutional Review Board and Icahn School of Medicine at Mount Sinai Institutional Review Board.

Demographic and clinical information was annotated by each institution and de-identified prior to sending to the lead institution (UNC) for analysis. Clinical variables for the RT-associated cohort included the type of RT received (external beam, brachytherapy, or both) and time from RT to diagnosis of bladder cancer. Overall survival (OS) was calculated from the date of diagnosis of bladder cancer to date of death from any cause or last follow-up. Reported tumor–node–metastasis staging represents the staging at the time of initial diagnosis. Smoking status was recorded at the time of bladder cancer diagnosis.

DNA was isolated from two to four formalin-fixed, paraffin-embedded slides per sample using the AllPrep DNA/RNA FFPE Kit (#80234 Qiagen), and quality and concentration were determined using TapeStation (Agilent). Sequencing was performed using hybrid capture–based NGS workflow for 324 cancer-related genes by FMI using FoundationOne CDx (Foundation Medicine, Inc.) in a Clinical Laboratory Improvement Amendments–certified, College of American Pathologists–accredited laboratory (Foundation Medicine; ref. 23). Copy number alterations along with the corresponding BAM files were then provided to UNC for downstream analysis (23, 24).

BAM files were converted to the FASTQ R1 and R2 files using the bamtofastq tool from biobambam2 v2.0.87 (25). The FASTQ files were aligned to the GRCh38 reference genome using BWA mem 0.7.17 (26), and then the BAM files were sorted, indexed, and duplicate reads were marked using the bamsormadup tool from biobambam 2.0.87. An in-house workflow was used to call somatic variants using Mutect2 (27) for single-nucleotide variants (SNV) and indels. Somatic variants were merged into a single variant call format (VCF) file and converted to a mutation annotation format (MAF) file using the tool vcf2maf v1.6.21 (28), and then the variants were annotated using a variant effect predictor (v103.1; ref. 29). Per sample somatic variants were merged and converted into a cohort-level somatic MAF file. Variants were then filtered down to those which passed the Mutect2 default filters, had a gnomADv3.1 allele frequency (AF) <5% (30), and were found within the ∼9,300 targeted genomic windows from the FoundationOne CDx capture. For recurrent mutation analysis, the MAF was further filtered for alterations with a variant allele frequency (VAF; t_alt reads/t_depth) of >5%.

Mutational signatures were derived based on all variants that passed the variant calling pipeline, inclusive of all classes and VAFs. Maftools was used to perform mutational signature analysis, and heatmaps were generated using pheatmap (31, 32).

Data from Mossanen and colleagues (17) were downloaded from https://github.com/CarvalhoFilipeL/CarvalhoFilipeL-Mouw_RABC. Mutational signatures and KDM6A calls were derived from the unfiltered MAF files. Copy number data, in the form of GISTIC2 values, were extracted from the supplemental data, and deletions were defined by a GISTIC2 score of less than −0.5. Outcome data were provided through request of the authors. The Cancer Genome Atlas (TCGA)–normalized gene expression and clinical variables were accessed through cBioPortal (https://www.cbioportal.org/study/summary?id=blca_tcga_pan_can_atlas_2018).

Statistical analysis was performed using R version 4.2.2 and RStudio 2022.12.0 + 353, unless specifically noted. All packages used in the analysis were publicly availably either through https://www.r-project.org/ or www.bioconductor.org. Boxplots are represented by the IQR and midline at the median. Error bars equal the Q1/Q3 ± 1.5 × IQR. Categorical comparisons were made using either χ² or Fisher exact test (as noted), dependent on the group size. Pairwise comparisons were performed using two-sided t test or Wilcoxon rank sum test (in cases of non-normal distribution). Multiple comparison correction was performed using Bonferroni correction. Kaplan–Meier plots were generated for survival analysis with significance determined using a log-rank or Cox proportional hazards model, as indicated.

Processed genomic data are available through request of the authors. Raw sequencing data were provided through a research agreement with FMI Inc. and are not available for public use through the authors.

From an initial pool of 85 patients who had previously received RT for prostate cancer, 35 patients had tissue available for DNA sequencing. An additional 99 patients who had undergone FMI sequencing as part of their care for bladder cancer within the UNC Health System were eligible for inclusion, with seven patients having had prior pelvic radiation and 92 patients having no prior history of pelvic radiation. As this was a retrospective study with a wide eligibility time range, the FoundationOne CDx assay evolved during the study window. To this end, we focused the analysis on the single bait set that allowed for the inclusion of the maximum number of patients across both the RT and CTRL cohorts. In total, the final cohort included 82 patients (RT = 41 and CTRL = 41; Fig. 1A; Supplementary Table S1).

Because the RT group was defined by prior pelvic RT for prostate cancer, all patients were male [n = 41 (100%)], whereas the CTRL cohort included male and female patients [male = 38 (68.3%)]. The median time from prostate treatment to bladder cancer diagnosis (latency) was 8.5 years (mean = 10.1, range = 2–28 years). Of the patients who received RT, 42% received brachytherapy, 31% received external beam, and 27% received combination. Race and smoking status were similar between the CTRL and RT groups. Within the RT group, latency was not associated with RT type, race, or smoking status (Supplementary Fig. S1A–S1C); however, there was a trend toward former and never smoking having a longer latency (median latency: current = 6 years, former = 9 years, never = 10.5 years).

Cohort-wide, patients with RT-associated bladder cancer were older than CTRL cohort (median age 75 vs 66 years, respectively; P value = 2e−4; Table 1). In the CTRL cohort, the median age of diagnosis for female patients was 5 years earlier than for males (female = 64 years, male = 69 years; P value = 0.1); however, after restricting the CTRL cohort to male patients, the RT cohort still had a significantly older age at diagnosis (CTRL = 69 years, RT = 75 years; P value = 0.01; Supplementary Fig. S1D–S1F).

There was no association between smoking and age of diagnosis among males within the CTRL cohort (P value > 0.51), nor was smoking associated with the age of the patient at the time of RT for prostate cancer (P value > 0.2; Supplementary Fig. S1G and S1H, Table 2). However, among the RT cohort, current smokers were diagnosed with bladder cancer at a significantly younger age than former and never smokers (current = 67 years, former = 75 years, never = 76 years; P values vs. former and never smokers <0.01 and 0.001, respectively; Supplementary Fig. S1I).

To evaluate if prior radiation conferred a worse prognosis among patients with bladder cancer, we compared the OS between the CTRL and RT cohorts. The median OS for the CTRL cohort was 20 months and for RT-associated bladder cancer was 23 months [log-rank P value = 0.27; HR 0.72; 95% confidence interval (CI), 0.41–1.3; Supplementary Fig. S2A]. Upon review of the tumor–node–metastasis staging, no difference was seen in T stage (range T1–T4; P = 0.08), whereas the CTRL cohort was enriched for patients with N and M stages >0 (P = 0.009 and P = 0.01, respectively; Fig. 1B–D). This would be expected as within the timeframe of case review, NGS sequencing was only performed as standard of care for those patients with advanced or metastatic disease.

Among RT-associated tumors, latency calculated from the time of pelvic RT to bladder cancer diagnosis was a significant prognostic feature. Patients with a longer latency (RT $≥$ 10 years, n = 18) had a significantly better survival than both those with a shorter latency (RT $<$ 10 years; n = 22; HR 6.8; 95% CI, 2.22–21.07; P = 0.001) and those with no prior radiation (HR 4.6; 95% CI, 1.58–13.11; P = 0.005; Fig. 1E). As the short latency group trended toward worse prognosis than the CTRL tumors, but did not reach significance, we leveraged TCGA for a further comparator group. Recapitulating the survival trends seen in the UNC cohort, long latency tumors trended toward better prognosis (HR 0.41; 95% CI, 0.15–1.1; P = 0.082), whereas the short latency tumors had a significantly worse prognosis than TCGA (HR 2.41; 95% CI, 1.42–4.1; P = 0.001; Supplementary Fig. S2B).

To further validate latency findings, we similarly split a RT bladder cohort from Dana–Farber Cancer Institute (DFCI), which only consisted of an RT arm, into short and long latency groups (17), and the short latency group had a median survival which was about 6 months less than that of the long latency group (767 vs. 975 days; Supplementary Fig. S2C), although this difference did not reach statistical significance.

Prior work in breast and lung cancer has demonstrated that smoking while undergoing RT is associated with worse outcomes and smoking while undergoing RT for breast cancer is associated with an increased risk of a secondary lung cancer (33–39). To evaluate if a similar phenomenon existed between pelvic radiation and bladder cancer, patients were divided into their respective CTRL and RT groups and stratified by smoking status (current, former, and never smokers). Within the CTRL cohort as well as TCGA, smoking was not associated with a decreased OS (Supplementary Fig. S2D–S2F). However, within the RT cohort, current smokers had a worse prognosis than both former (HR 0.152; 95% CI, 0.05–0.56; P = 0.004) or never smokers (HR 0.223; 95% CI, 0.06–0.816; P = 0.023; Fig. 1F). In a multivariate model of survival, smoking and latency were both independently prognostic (P = 0.014 and P = 0.002, respectively; Fig. 1G). Additionally, whereas smoking was not significantly associated with decreased latency, time from RT to last follow-up (latency + OS), current smokers had a numerically reduced RT to death interval as compared with former or never smokers (median survival = current – 6.96 years, former – not reached, and never – 17.75 years; Fig. 1H).

We next assessed RT-specific genomic alterations within our cohort. To understand the full distribution of variants, we first analyzed the number of alterations per sample by variant class (e.g., SNV and deletion) inclusive of any VAF. On a per sample basis, the RT cohort contained increased numbers of insertions and SNPs, but not deletions (Fig. 2A). As no VAF filter had been applied, we next wanted to evaluate if the increased number of alterations per sample within the RT cohort was restricted to low VAF samples. Indeed, RT samples had a significantly higher number of low VAF alterations per sample than the CTRL cohort (P = 0.007 and P = 0.001; 0–10 years and 10+ years vs. CTRL, respectively; Fig. 2B). To better understand the gene-level mutation specificity between our cohorts, we performed recurrent mutation analysis on genes with VAF >0.05 (restricting our analysis to genes more likely associated with clonal growth). Overall, mutation frequencies were consistent between the CTRL and RT cohorts, in line with prior muscle-invasive bladder cancer (MIBC) genomic results (40, 41). The most frequently mutated genes included (CTRL %; RT%) TP53 (54%; 49%), TERT promoter mutations (49%; 49%), CDKN2A (12%; 27%), and CREBBP (27%; 20%; Fig. 2C and D). Of the 311 genes evaluated, only CDKN2A (CTRL), ATM (RT: 0–10 years), and KDM6A (RT: 10+ years) were significantly differentially mutated (P < 0.05; Fig. 2E; Supplementary Table S2). Of the three differentially mutated genes, only KDM6A displayed recurrently mutated regions associated with RT, located proximal to splice sites (Supplementary Fig. S3A–S3C). To validate that KDM6A alterations occurred at an increased frequency within RT-associated bladder cancer, we calculated the proportion of KDM6A altered tumors in the DFCI validation cohort and in a non-RT cohort, TCGA bladder study (17). The frequency of KDM6A alteration within TCGA was similar to that of our CTRL cohort (TCGA = 26%, CTRL = 27%), whereas the frequency within the DFCI RT cohort was 18% greater (44%; Fig. 2F). Taken together, this points to a broad subclonal pattern of DNA damage induced by RT to the normal urothelium, along with few distinct recurrent RT-induced driver mutations.

Reproducible patterns of SNVs within a specific trinucleotide motif structure, referred to as mutational signatures, have been associated with various biological functions and environmental exposures (42). These SNVs within their flanking nucleotide context can be quantified and clustered to reveal recurrent mutational signatures specific to a cohort. Those cohort-specific signatures can then be correlated to a set of previously described signatures with known etiology to infer possible mechanism or biological significance. To determine if unique mutational signatures were present within the CTRL and RT cohorts, we quantified and clustered the mutational motifs and correlated the corresponding signatures back to those described in the Catalogue of Somatic Mutations in Cancer signature database. We identified three cohort-derived signatures: one most similar to the APOBEC activity signatures, SBS2 and SBS13 (A3-Combo); a second signature (smoking) with shared similarity to the ERCC2 mutant/smoking signature, SBS5 (43, 44); and a third [radiation therapy signature (XRT)], which was only enriched within the XRT tumors and is most similar to defects in DNA mismatch repair, SBS6 and SBS15 (Fig. 3A; Supplementary Fig. S4A; Supplementary Table S3).

Following de novo signature identification, the number of mutations attributed to each signature was calculated (“signature contribution”). Whereas the smoking signature was consistent and accounted for the largest proportion of variants across the cohort, the A3-Combo signature was significantly increased within the CTRL cohort compared with RT (short and long latency; P = 0.05 and P = 0.003), and the XRT signature displayed a significant and stepwise increase from the CTRL to RT cohort (P = 3e−6 and P = 3e−8; Fig. 3B). In keeping with the etiologies of the correlated signatures, the smoking signature was significantly increased over that of the A3-Combo within current and former smokers while showing no differential contribution in the never smokers (Fig. 3C). Likewise, within the RT cohort, the XRT signature was present at equal levels within the current and former smokers, whereas all three signatures were equally represented within never smokers (Fig. 3D).

Finally, as KDM6A was the most frequently altered gene in the RT cohort, we extracted the trinucleotide motifs for each alteration, in both the CTRL and RT cohorts. When visualized by cohort, the RT samples had a noticeable increase in the frequency of A[C>T]N and C[C>T]N motifs, similar to that of the XRT signatures, in addition to an increased frequency of N[T>C]N motifs (Supplementary Fig. S4B). Taken together, these data are suggestive of the fact that RT confers a unique mutational signature and may account for the increased number of KDM6A mutations present in the RT samples.

Because current smokers and patients with short latency tumors had worse OS (Fig. 1F), we postulated that patients in which the smoking or XRT signature contributed to the majority of the tumor’s mutations would, in turn, have the worst prognosis. When patient survival was analyzed within the RT group by the max signature contribution, short latency tumors with smoking or XRT signature had the worst outcome, with the A3-Combo signature and long latency tumors having the best outcome (Fig. 3E). The smoking and XRT signatures retained their prognostic ability, even when correcting for smoking status and latency (smoking vs. A3: HR = 7.4; 95% CI, 1.14–48.78; P = 0.036; XRT vs. A3: HR = 9.1; 95% CI, 1.3–64.06; P = 0.026; Fig. 3F). Whereas there was no signature-specific survival within the CTRL cohort, this could be confounded by the fact that the patients were biased toward advanced disease (Supplementary Fig. S5A).

To validate the signature-specific survival within the RT samples, we repeated the mutational signature analysis in the independent DFCI cohort (17). Again, we identified three mutational signatures that corresponded to SBS2/SBS13, SBS5, and SBS6/SBS15, respectively (Supplementary Fig. S5B). Even though the DFCI signatures were highly similar to the same Catalogue of Somatic Mutations in Cancer signatures as UNC, we directly compared the motif patterns of the DFCI signatures and the UNC signatures. UNC A3 signature had the highest correlation to DFCI-A, the UNC smoking signature was most correlated to DFCI-B, and the UNC XRT signature was most correlated to DFCI-C while also showing similarity to DFCI-B (Supplementary Fig. S5C). The signature related to prognosis mirrored that of the UNC cohort: A good prognosis was associated with the A3-Combo signature, whereas the smoking signature (HR = 6.6; 95% CI, 1.14–39; P = 0.035) and XRT signature conferred a poor prognosis (HR = 3.5; 95% CI, 0.71–17; P = 0.123; Supplementary Fig. S5D).

For more than 30 years, the field has appreciated that recurrent copy number events occur at a high frequency and are an early driver of bladder cancer (40, 45, 46). Chr9p21.3 loss (containing CDKN2A, CDKN2B, and MTAP) is the most frequent alteration, occurring in ∼53% of NMIBC tumors and 22% of MIBC (40, 47). To understand what, if any, focal amplifications or deletions are specific to RT-associated bladder tumors, we compared the FMI copy number calls between the CTRL cohort and the RT cohort. Of the 26 copy number events detected, only PPARG amplification was significantly enriched in the RT cohort (RT: n = 7, CTRL: n = 0; Fisher exact P = 0.01).

CDKN2A/CDKN2B was the most frequent copy number event, occurring in 27% of CTRL samples (11/41) and 24% of RT samples (10/41; Fig. 4A; Supplementary Table S4). CDKN2A/CDKN2B-deleted tumors had decreased survival compared with copy number neutral tumors in both the CTRL (HR = 2.41; 95% CI, 1.07–2.43; P = 0.033) and short latency RT cohorts (HR = 4.58; 95% CI, 1.44–14.5; P = 0.01; Fig. 4B–D). Among the CDKN2A/B-deleted tumors, short latency RT tumors had a significantly worse OS compared with CTRL tumors (HR = 7.11; 95% CI, 1.89–26.7; P = 0.004); however, no difference was observed among CDKN2A/B copy number neutral tumors (Fig. 4E and F). Consistent with these findings, CDKN2A/CDKN2B-deleted tumors within the DFCI cohort displayed a 41-month reduced survival compared with wild-type tumors (HR = 4.65; 95% CI, 1.45–14.9; P = 0.01).

To understand if the smoking and RT signatures are prognostic in a more general setting, we correlated the trinucleotide motif frequencies from TCGA bladder cancer study (BLCA) to our cohort-derived A3-Combo, smoking, and XRT signatures. Using the Pearson correlation coefficient as the similarity metric, we first validated that the smoking signature was associated with smoking status. Current smokers had a significantly higher correlation with the UNC smoking signature than former or never smokers (P = 0.013, P = 0.011, respectively), whereas smoking status was not associated with the A3 or XRT signatures (Supplementary Fig. S6A–S6C).

Next, TCGA tumors were split based on median correlation into high and low groups for each signature. The high correlation group for both the smoking (HR = 1.4; 95% CI, 1.05–1.9; P = 0.02) and XRT signatures (HR = 1.65; 95% CI, 1.22–2.23; P = 0.001) was associated with poor prognosis. Furthermore, specific to the XRT signature, when CDKN2A loss was integrated into the analysis, it further defined a subset of patients with worse prognosis (Fig. 5A and B; Supplementary Fig. S6D).

This study demonstrates that prior RT is associated with increased numbers of low-frequency SNVs, which results in a RT-specific mutational signature. This signature, in addition to a smoking signature, CDKN2A/B loss, patient smoking status, and a tumor latency of <10 years represent a key set of molecular and clinical variables that characterize a tumor with poor prognosis following RT. Additionally, the smoking and XRT signatures may be able to identify tumors with a pattern of genomic alterations conferring a more aggressive phenotype, independent of the canonical recurrent mutations.

Previous studies have demonstrated that RT for the treatment of prostate cancer results in up to a 3-fold increased risk for the subsequent development of bladder cancer. Although the relative risk of developing a secondary bladder cancer is low, thousands of men are possibly at risk over the course of their lifetime, given that prostate cancer cases are projected to reach almost 300,000 new cases in 2024. The relatively low frequency of secondary bladder cancer has also made it difficult to construct cohorts of RT-associated bladder tumors to interrogate how the biology and genomics may differ from that of non-RT–associated bladder tumors. Here, we present the genomic analysis of, to our knowledge, the largest cohort of bladder tumors associated with pelvic radiation.

Through integration of clinical and genomic data, we identified RT-specific genomic features. Overall, RT tumors had increased levels of insertions and SNVs, with decreased levels of deletions. At first glance, these results seem to be at odds with prior work by Behjati and colleagues, which demonstrated an increased number of deletions in secondary malignancies (22). However, their cohort was constructed from, primarily, a mixed set of sarcomas (n = 9) and breast tumors (n = 3). Due to differences in the cohort, their data cannot be easily projected on to our current study. The Behjati data also show a heterogeneous composition of structural variants, with an increased number of deletions in RT-associated tumors; however, breast tumors, which have molecular similarities to bladder cancer (48, 49), showed a lower rate of deletions than sarcomas. Additionally, their data demonstrating increased deletions in radiated prostate tumors are from primary prostate tumors, not the subsequent bladder malignancy, and thus not comparable with our cohort.

Interestingly, we identified an increased number of KDM6A mutations located within the splice regions. This alteration has not, to our knowledge, been previously described in the literature, which may be in part because of the lack of radiation-associated bladder cancer cohorts. Additionally, within the RT cohort, the KDM6A variants occurred as part of motifs that were enriched within the XRT mutational signature. This suggests that RT may, in part, be accounting for the increased number of KDM6A alterations, although future experimental validation will be required to establish a direct relationship and potential mechanism.

Minimal copy number differences were observed between the CTRL and RT samples; however, amplification of PPARG was significantly enriched within the RT samples. Previously, PPARG structural variants have been reported in thyroid cancer stemming from environmental radiation exposure (50, 51). In a bladder-specific context, PPARG expression is associated with urothelial differentiation and the better outcome LumP molecular subtype; therefore, it was of note that its amplification was specific to the long latency, better prognostic subset of tumors (40, 48, 49, 52, 53). Further exploration of these genetic alterations and their relationship to radiation is merited, as they could be predictive and prognostic for RT-associated bladder cancer.

Through global analysis of SNVs, we identified three mutational signatures associated with the CTRL and RT cohorts. These signatures, in addition to the loss of CDKN2A, demonstrated prognostic value in our discovery dataset along with two independent datasets. Earlier work by the Verhaak group demonstrated that RT-related genomic alterations, including CKDN2A loss, conferred poor prognosis in patients with glioblastoma who developed recurrence postradiation (54). Additionally, previous studies have suggested that CDKN2A loss could be predictive of immunotherapy response and progression, as well as prognostic value in bladder cancer (55–58). Similarly, we see poor prognosis among patients whose tumors have increased levels of the XRT-related signature and/or CDKN2A loss, if combined with prior pelvic radiation. When we applied the mutational signatures to a non–RT-specific cohort, TCGA, the signatures were again both prognostic, XRT alone and smoking in combination with CDKN2A loss. These data indicate that in the absence of RT or known smoking history, tumors that share a similar mutational pattern to that of patients treated with RT and/or smoking may have more aggressive disease. This suggests that even without patient-reported exposure histories, the molecular profile of the tumor may be informative, and further work is needed to understand the etiology of these exposure-related signatures.

When combining these results with the described relationship among smoking status, RT, and survival, we postulate that there are two distinct pathways that lead to the development of bladder cancer following pelvic radiation for prostate cancer. The first occurs in patients who have a baseline clinical history conferring a high risk for bladder cancer (e.g., heavy smoking/occupational exposure to known carcinogens) and receive RT for prostate cancer. The urothelium in these patients, prior to radiation, may have already undergone field cancerization (i.e., phenotypically normal cells that are primed for cancer formation due to prior mutagenic insults; ref. 59). In this setting, the off-target DNA damage induced by the radiation acts to hasten the development of the bladder cancer in the “pre-cancerized” bladder urothelium, ultimately leading to a short latency and poor prognosis. In the second pathway, patients who receive RT and later develop bladder cancer, in the absence of known bladder cancer risk factors, and have minimal preexisting urothelial genomic alteration are more likely to have a longer latency (10+ time course) tumor and ultimately a better prognosis.

There were several limitations to our study, which will need to be addressed in future work. One limitation, which points to a larger issue within the field, is the lack of tumor specimens available. Our study, which was the largest to date, included 41 individuals, which more than doubled the number of samples from the most recent prior study (n = 19; ref. 17). With evidence that RT may lead to more aggressive tumors in a subset of patients, institutions must collaborate to compile harmonized and clinically annotated cohorts to allow for a comprehensive evaluation of radiation-associated bladder cancer.

Although we observed a significant reduction in OS among current smokers within the RT cohort, the impact of the finding is limited because of the small sample size (n = 5). This lack of power underscores the previously stated wider issue of low sample number and diversity, which is part and parcel to the study of rarer tumor types. This lack of diversity spans a wide array of clinical variables including social factors, such as smoking and urban/rural geography, as well as ancestry and ethnicity.

As this was a retrospective study, we leveraged preexisting clinical sequencing data that had been generated by FMI. The use of a preset gene panel limited our ability to detect somatic alteration that is of interest in the research setting but not yet utilized clinically (e.g., ERCC2). Additionally, this reduced the resolution of de novo mutational signature identification compared with using data derived from whole-exome or whole-genome sequencing. Whereas using a clinical sequencing assay limited the number of variants we were able to detect, it did allow for the use of archival specimens and previously sequenced individuals. The use of FMI also was a strength, as in most clinical settings, patients will undergo commercial targeted sequencing and not have whole-exome or whole-genome sequencing performed on their tumor, making our findings more translatable to the clinic. Even with the described limitations, we were able to validate our findings in external datasets and demonstrate the presence of prognostic features, which could be applied to RT and non-RT bladder tumors.

We hypothesize that patients with a clinical history conferring a high risk for bladder cancer may also be at an increased risk for RT-associated poor prognosis bladder tumors because of a precancerization phenotype within the urothelium. Future prospective studies will need to be conducted to fully understand the extent of DNA damage in the urothelium before and immediately following RT and how smoking may compound this damage.

In summary, we present the largest genomic cohort of RT-associated bladder cancer to date and provide evidence for worse bladder-specific outcomes among patients who are smokers, have a short latency tumor, or have evidence of specific mutational signatures. Additionally, our data suggest that clinicians and patients should consider smoking and environmental exposure history as a factor when deciding on the appropriate treatment protocol for prostate cancer.

J.A. Moore reports personal fees from Foundation Medicine Inc. and personal fees from Roche outside the submitted work. Z. Fleischmann reports other support from Roche and personal fees from Foundation Medicine Inc. during the conduct of the study. D.X. Jin reports employment at Foundation Medicine Inc. and ownership at F. Hoffmann-La Roche AG stock. M.D. Galsky reports personal fees from Astellas Pharma, Bristol Myers Squibb, Merck, Genentech, Pfizer, EMD Serono, Seagen, Janssen, Fujifilm, Gilead, Asieris, Veracyte, Daiichi, Aktis, and AstraZeneca outside the submitted work. M.I. Milowsky reports grants from Merck, Bristol Myers Squibb, Mirati Therapeutics, Seagen, Alliance Foundation Trials, Alliance for Clinical Trials in Oncology, Clovis Oncology, Arvinas, ALX Oncology, Hoosier Cancer Research Network, Novartis, Acrivon Therapeutics, Astellas Pharma, Genentech, Accuray, PCCTC, G1 Therapeutics, OncoC4, Flare Therapeutics, Loxo/Lilly, and Roche outside the submitted work, as well as ownership of stock in Pfizer, Merck, and Gilead Sciences; provides CME talks via Medscape and Research to Practice; and serves as co-editor-in-chief of Clinical Genitourinary Cancer, an Elsevier journal. W.Y. Kim reports other support from AbbVie, Amgen, Apellis, Arvinas, BeiGene, Bristol Myers Squibb, Eli Lilly, Moderna, Novo Nordisk, Revolution Medicines, Tango Therapeutics, Viking Therapeutics, ImmunityBio, Focal Medical, OncoRev, and Merck outside the submitted work, as well as a patent to PurIST Classifier pending and licensed and a patent to DeCAF Classifier pending. J.S. Damrauer reports nonfinancial support from Foundation Medicine Inc. during the conduct of the study. No other disclosures were reported.

N.A. Wijetunga: Formal analysis, validation, investigation, methodology, writing–original draft, writing–review and editing. K.H. Gessner: Formal analysis, validation, investigation, visualization, writing–original draft, writing–review and editing. K. Kanchi: Data curation, formal analysis, writing–review and editing. J.A. Moore: Data curation, writing–review and editing. Z. Fleischmann: Data curation, writing–review and editing. D.X. Jin: Data curation, writing–review and editing. G.M. Frampton: Data curation, writing–review and editing. M. Sturdivant: Formal analysis, visualization, writing–review and editing. M. Repka: writing–review and editing. S. Sud: writing–review and editing. D.L. Corcoran: Data curation, writing–review and editing. M.D. Galsky: Conceptualization, resources, writing–review and editing. M.I. Milowsky: Conceptualization, funding acquisition, writing–review and editing. S.E. Wobker: Data curation, investigation, writing–original draft, writing–review and editing. W.Y. Kim: Conceptualization, supervision, funding acquisition, writing–review and editing. T.L. Rose: Conceptualization, data curation, supervision, investigation, methodology, writing–original draft, writing–review and editing. J.S. Damrauer: Conceptualization, data curation, software, formal analysis, supervision, validation, investigation, methodology, writing–original draft, project administration, writing–review and editing.

Portions of this study were supported by the University Cancer Research Fund and Foundation Medicine Inc. Additional funding included NIH-NCI R01-CA241810 (W.Y. Kim) and the Thomas M. Mohr Fund for Bladder Cancer Research. The authors would also like to acknowledge all the patients and their families who contributed to this study.