The tandem duplicator phenotype may be a novel targetable subgroup in pancreatic cancer

Abstract

Tandem duplicator phenotype (TDP) consists of distinct genomic rearrangements where tandem duplications are randomly distributed. In this study, we characterized the prevalence and outcomes of TDP in a large series of prospectively sequenced tumors from patients with pancreatic ductal adenocarcinomas (PDAC). Whole-genome sequencing (WGS) was performed in 530 PDAC cases from the PanCuRx Initiative, COMPASS and PanGen/POG trials in Canada. Of 530 cases, 52 were identified as TDP (9.8%; 13 resected, 39 advanced). Etiological subgroups of TDP included BRCA1 (n = 9), CCNE1 (n = 4), and unknown (n = 39). Presence of TDP was not prognostic in resected specimens (p = 0.77) compared with non-HRD and non-TDP cases, described as typicals. In advanced cases, when stratified for only classical subtype cases, platinum therapy was correlated with longer response in non-BRCA1 TDP vs. typicals (p = 0.0036). There was no difference in overall survival between TDP and typicals (p = 0.5).TDP represents a potential novel targetable subgroup for chemotherapy selection in PDAC.

Introduction

Pancreatic ductal adenocarcinoma (PDAC) remains a lethal malignancy with 5 year survival of only 10%¹. Combination cytotoxic chemotherapy remains the mainstay of therapy for patients with advanced disease, with few targeted approaches approved to date². Patients with homologous recombination deficient (HRD) tumors, particularly those with BRCA1/2 or PALB2 mutations, appear to derive benefit from platinum-based therapies and PARP inhibitors^3–5. There is an urgent unmet need for developing novel predictive biomarkers in PDAC to better direct treatment selection.

Over the last decade, efforts to improve our understanding of PDAC using genomic technologies have revealed subtypes based on structural genomic variations, the presence of RNA expression profiles, and mutational signatures^6–8. Whole genome and transcriptome sequencing (WGS/RNA-Seq) has provided additional granularity to better delineate the relevance of structural variation patterns in cancer, including catastrophic events such as chromoplexy and chromothripsis^9,10. The tandem duplicator phenotype (TDP) is a specific structural genomic configuration, first described by Menghi and colleagues, consisting of numerous segmental DNA duplications homogeneously scattered throughout the cancer genome¹¹.TDP has been described in triple negative breast cancer, ovarian, and endometrial cancers to date, but not in PDAC^11,12 .This somatic phenotype has been associated with perturbations in BRCA1, TP53, CDK12, CCNE1, and FBXW17^11,12. In particular, BRCA1-deficient tumors retain genomic scars of both HR deficiency and TDP. In this study, we characterize the prevalence and outcomes of TDP in a large series of prospectively sequenced PDAC tumor samples, with a focus on clinical relevance.

Results

Of 530 prospectively sequenced PDAC tumors, 52 were scored as TDP (9.9%; 13 resected, 39 advanced), and 38 samples were identified as HRD (Supplementary Fig 1). A subset of 9 samples that scored as TDP were also classified as HRD (Fig. 1a). Eight of these 9 samples contained double hits for BRCA1, with the remaining sample having one hit in BRCA1. To prevent biasing the results, these 9 samples were treated as HRD regardless of the TDP status unless otherwise specified. Forty-three samples (7.9%) were non-BRCA1 TDP, including 10 resected and 33 advanced cases. Baseline characteristics for both resected and advanced cases are detailed in Tables 1 and 2. TDP tumors were associated with alterations in BRCA1 (n = 9; 7 germline, 2 somatic), CCNE1 (n = 4, 4 somatic), and unknown (n = 39; no identified alterations in TDP-related genes; Fig. 1b). We did not identify complete inactivation of BRCA2 cases in our population of cases with TDP. TD size distribution varied significantly by TDP etiology (Kruskal-Wallis test, p = 1.9e-196), with BRCA1-TDP exhibiting the smallest median TD size (9.4 kb), CCNE1-TDP with the largest median TD size (200.31KB), and those of unknown etiology demonstrating a median TD size of 114.27 kb (Fig. 1c). We found that the PanGen/POG cohort had a higher frequency of TDP (14%, 9 of 64 cases), and the median duplication ratio compared to COMPASS/PanCuRx (0.195 vs 0.155).

Fig. 1. Venn Diagram showing the breakdown of TDP samples compared to HRD.

a *Of these 10 samples, 9 have confirmed BRCA1 complete loss. The remaining sample shares features with our BRCA1 TDPs (positive Menghi score and duplication ratio > 0.205) and was identified as HRD, but we were unable to identify the BRCA1 hit. We identified this sample as HRD and treated it as such throughout the analysis. b Oncoprint depicting HRD, TDP and typical cohorts Genomic profile of a subset of samples from our cohort. This subset includes all HRD, TPD, and a random sample of 30 typical cases. Each column represents individual samples separated by their etiology. Tumor mutational burden (TMB) is colored based on the amount of burden. The Moffit classification is depicted if available. The first and second genomic alternations found in key driver genes are displayed. c Subclasses of TDP samples The density of the tandem duplication size distribution is separated by etiology. The difference is statistically significant (Kruskal-Wallis test p = 1.9e-196).

Table 1.

Baseline characteristics of resected cases in our cohort

Resected Cases	HRD (N = 15)	Non-BRCA1 TDP (N = 10)	Typicals (N = 166)
Age (range)	58.3 (42.7–76.8)	66.7 (55.5–80.7)	67 (38–87
Sex
Male	11 (73.3%)	4(40%)	89(53.6%)
Female	4 (26.6%)	6(60%)	77(46.4%)
Ethnicity
White/Caucasian	4 (26.6%)	3 (30%)	76 (45.7%)
Asian	0	1 (10%)	6 (3.6%)
Hispanic	1 (6.6%)	0	0
Black/African/Caribbean	1 (6.6%)	0	2 (1.2%)
Unknown	8 (53.3%)	6 (60%)	72 (43.3%)
Others*	1 (6.6%)	0	10 (6%)
Median CA19-9 (range, KU/L)	77 (1–55314)	13,985 (20–8060)	119 (1–16420)
Location of tumor
Pancreatic Head	7 (46.6%)	3 (30%)	94 (56.6%)
Pancreatic Tail	3 (20%)	2 (20%)	5 (3%)
Pancreatic Body	1 (6.6%)	0	10 (6%)
Multicentric	1 (6.6%)	0	4 (3%)
Others	1 (6.6%)	0	12 (7.2%)
Unknown	2 (13.3%)	5 (50%)	41 (24.7%)
Adjuvant chemotherapy
None	4 (26.6%)	5 (50%)	38 (22.9%)
FOLFIRINOX	1 (6.6%)	0	5 (3%)
Gemcitabine/Cisplatin	2 (13.3%)	0	0
Gemcitabine/nab-Paclitaxel	0	0	0
Single Agent Gemcitabine	6 (40%)	1(10%)	82 (49.4%)
5-FU/Capecitabine	1 (6.6%)	0	12 (7.2%)
Others**	1 (6.6%)	4(40%)	29 (17.5%)
Neoadjuvant chemotherapy
None	12 (80%)	8 (80%)	152 (91.5%)
FOLFIRINOX	1 (6.6%)	0	1 (0.6%)
Gemcitabine/Cisplatin	1 (6.6%)	0	0
Gemcitabine/nab-Paclitaxel	0	0	0
Single Agent Gemcitabine	0	0	3 (1.8%)
5-FU/Capecitabine	0	1 (10%)	2 (1.2%)
Others	1 (6.6%)	1 (10%)	8 (4.8%)
Median disease-free survival (range, months)	10.7 (2.5–63	10.2 (1.1–85.1	11.23 (1.1–142.9)

*Others: White Caucasian and Asian (n = 1); Jewish-Ashkenazi(n = 4);White Caucasian and Jewish-Ashkenazi (n = 3); Greek(NOS)(n = 1); Multiracial(n = 1); Lebanese and East Indian (n = 1);

**Others:Gemcitabine/5FU (n = 26); Gemcitabine/Erlotinib (n = 7); Gemcitabine/Capecitabine (n = 1).

Table 2.

Baseline characteristics of advanced cases in our cohort

Advanced	HRD (N = 23)	Non-BRCA1 TDP (N = 33)	Typicals (N = 283)
Median age (range)	61 (42–77)	62 (35–77)	64 (29–84)
Gender
Male	13 (56%)	23 (69.6%)	165 (58.3%)
Females	10 (43%)	10 (30.3%)	118 (41.6%)
Ethnicity
White/Caucasian	17 (74%)	26 (79%)	194 (68.5%)
Asian	4 (17%)	4 (12%)	47 (16.6%)
Black/African/Caribbean	2 (10%)	0 (0%)	7 (2.4%)
Hispanic	0 (0%)	1 (3%)	0 (0%)
American Indian or Alaska Native	0 (0%)	0 (0%)	2 (0.7%)
Unknown	0(0%)	2 (6%)	32 (8.4%)
Others*	0 (0%)	0 (0%)	1 (3.5%)
CA19-9(Median) KU/L	5625 (n = 20)	3447 (n = 24)	1299 (n = 231)
Range	0.8–371847	9–86539	0.8–305453
Sites of metastasis	n = 20	n = 24	n = 231
Liver	17 (85%)	19 (79%)	162 (70%)
Lung	2 (10%)	0 (0%)	16 (6.9%)
Omentum	4 (20%)	6 (25%)	49 (21.2%)
LN	1 (5%)	4 (16.6%)	28 (12.1%)
Clinical stage
III	0 (0%)	8 (24.2%)	35 (12.3%)
IV	23 (100%)	25 (75.7%)	248 (87.6%)
Denovo/Metastatic
Denovo Metastatic	22 (95%)	31 (94%)	224 (79.1%)
Recurrent	1 (4.3%)	2 (6%)	21 (7.4%)
Locally advanced unresectable	0 (0%)	0 (0%)	38 (13.4%)
First line chemotherapy
n/a	0 (0%)	3 (9%)	21 (7.4%)
Folfirinox	17 (74%)	21 (63.6%)	146 (51.5%)
Gemcitabine/Cisplatin	2 (8.7%)	0 (0%)	0 (0%)
Gemcitabine/nab-Palcitaxel	2 (8.7%)	9 (27%)	81 (28.6%)
Single Agent Gemcitabine	1 (4.3%)	0 (0%)	7 (2.4%)
Clinical trial and others**	1 (4.3%)	0 (0%)	28 (9.8%)
Time on first line chemo(weeks)	45	22.9	17
Range	0–149.8	0–145.5	0–176
Second line chemo	n = 9	n = 3	N = 81
FOLFIRINOX	0 (0%)	0 (0%)	10 (12.3%)
Gemcitabine/Cisplatin	5 (45%)	0 (0%)	0 (0%)
Gemcitabine/nab-Paclitaxel	1 (9%)	0 (0%)	17 (21%)
Single Agent Gemcitabine	2 (18%)	2 (66.6%)	17 (21%)
5FU based chemotherapy	0 (0%)	0 (0%)	9 (11.1%)
NALIRIFOX	0 (0%)	0 (0%)	8 (9.8%)
Clinical trials and others***	1 (27%)	1 (33.3%)	20 (24.6%)
Time on second-line chemo(weeks)	6	2.7	7.8
Range	1.4–155	1–17.2	0–107.8

Others^*: Other non-European.

Others ^**: Clinical trial (n = 26),FOLFOX (n = 1), afatinib (n = 1),chemoradiation (Capecitabine+RT) (n = 1).

Others ^*** : Clinical trial (n = 18)(;KRAS G12C inhibitor (n = 1); Gemcitabine/capecitabine (n = 2); Trametinib (n = 1).

We further characterized the TDP subgroup according to the TD span size as described by Menghi et al.¹². We found that within our cohort of 39 cases with unknown TDP etiology, 10 samples could be classified as group 1 with a TD span size between 1.64 and 51 kb (Supplementary Fig 2). The 8 BRCA1 positive TDP samples were similarly classified as group 1. We also found that 21 TDP samples with unknown etiology could be classified as group 2 with a TD span size of 51–622 k, along with our 4 CCNE1 mutation positive TDP samples. We identified a few cases that fell into multiple groups, but these were excluded due to the small sample size. Overall, these results are in keeping with the prior findings from Menghi et al. where BRCA1 cases were classified as group 1 and CCNE1 positive cases were characterized as group 2. In addition, this indicates that our unknown etiology cohort is heterogenous and in fact represents two distinct subgroups with different span sizes.

To further examine genes related to the unknown etiology subgroup of TDP, we ran two driver discovery tools as described above (Supplementary Fig 3). Both tools independently identified the common driver mutations in PDAC, including KRAS, TP53, CDKN2A, and SMAD4, but no novel driver genes. Using an adjusted p-value cut-off of <0.25, 6 additional genes were identified (TGFBR2, XRCC2, PRCC, CSAG1, HOXD12, and RPL5), but due to sample size, it is difficult to draw any definitive conclusions. While CDK12 variants were found in 6 nonBRCA1-TDP samples, they were all monoallelic in nature, thus, we do not believe they play a role in TDP in our dataset.

Genomic characteristics of non-BRCA1 TDP

To characterise TDP as a distinct genomic profile separate from HRD, such as in the presence of BRCA alterations, we stratified our study cohort into HRD (including BRCA1-TDP), non-BRCA1 TDP, and typical subgroups. As described in the methods, HRD was classified based on previously published hallmarks of HRD and then confirmed using HRDetect. All but three samples were discordant between the two methods (Supplementary Fig 1). These discordant samples were manually reviewed. Brief descriptions for these samples are found in Supplementary Fig 1.

The BRCA1-TDP samples, which display both HRD and TDP phenotypes, were included as part of the HRD group to prevent confounding results when profiling the TDPs (Fig. 1a, b). We did not observe any complete inactivation of BRCA2 in any samples. Non-BRCA1 TDP were associated with elevated TD load and HRDetect scores compared to typicals (Fig. 2a; p < 0.01). We found that non-BRCA1 TDP demonstrated similar ploidy to HRD tumors when adjusted for stage (Wilcoxon test p = 0.279 for resectable and p = 0.399 for advanced), and significantly lower ploidy compared to typical tumors in the advanced setting (Wilcoxon test, p = 0.0232; Fig. 2b). There were no differences in ploidy between the TDP cohorts based on etiology.

Fig. 2. Genomic differences across etiologies in resected and advanced cohorts.

a HRDetect scores between HRD, nonBRCA1-TDP, and typical samples. Wilcox test was performed to compare the groups. b Ploidy differences for resected (Stage I-II) and advanced stage (Stage III-IV) samples across subtypes etiologies. Wilcoxon mean rank-sum tests were run in pairwise fashion on the top panels. The Kruskal-Wallis test was performed on the bottom 2 panels. Box plots indicate median (central line), 25–75% IQR (bounds of box), and whiskers extend from box bounds to the largest value no further than 1.5 times the IQR c Proportion of all samples with mutant and wildtype counts in key driver genes (KRAS, TP53, CDKN2A and SMAD4) across subtypes. Fisher’s exact test was run on all comparisons with p-values depicted in the figure. d Moffitt classification of resected (Stage I-II) and advanced stage (Stage III-IV) samples across subtypes. Fisher’s exact test was run on all comparisons with p-values depicted on the figure. e (i) Best tumor response to treatment (by RECIST 1.1) across SVs that have been split equally into low and high tumors. Wilcox test was performed between the 2 groups. (ii) Duplication load with a decrease or increase in tumor size. Wilcox test run between the 2 groups. (iii) SV load in response to treatment as per RECIST. Wilcox test run between the 2 groups. iv) Scatterplot of tumor response with increasing duplication load. Pearson correlation was run on the data.

We assessed the HRD, non-BRCA1 TDP and typical groups in our cohort for differences in driver mutations. We observed no significant differences in KRAS mutation frequency between the three groups (Fig. 2c). There were also no significant differences in terms of KRAS subtype or KRAS allele (Supplementary Fig 4 and Supplemental Table 1), although in the advanced PDAC cohort, there was an absence of KRAS wildtype allele. When compared to HRD, non-BRCA1 TDP were enriched for TP53 mutated tumors (90.7% vs. 63.2%, Fisher exact test, p = 0.00352), but have similar mutation frequency of KRAS, CDKN2A, and SMAD4 (Fig. 2c). We then compared tumor mutation burden (TMB) across the groups. The median TMB for the nonBRCA1-TDP samples was 2.809 and was significantly different from both the HRD (median TMB of 2.97) and typicals (median of 2.15; Wilcoxon test, p = 8.59 × 10^-5 and 0.000736; Supplemental Fig 5).

Previous work has shown that the Moffitt classification may be prognostic in PDAC, with poorer outcomes seen in the basal-like subgroup compared to their classical counterparts^13,14. In our analysis, we applied the Moffitt classification to the three different groups in our cohort. A complete list of all samples alongside TDP status (Typical, TDP or HRD) as well as the TDP etiology and Moffitt classification is provided in Supplemental Table 2. We found a significant difference in Moffitt classification between the HRD, non-BRCA1 TDP, and typical groups in the advanced setting with non-BRCA1 TDP having a higher proportion of the basal-like subtype as compared to typicals (p = 0.00546) and the HRD group (p = 0.0474). In addition, there was a trend towards a higher proportion of basal-like subtype among TDP and HRD groups in the resected population (non-BRCA1 TDP vs typicals, p = 0.184; HRD vs typicals, p = 0.0982)(Fig. 2d). We assessed if higher duplication load correlates with response to chemotherapy in both early stage and advanced cases, but found no correlation with duplication load and tumor response (p = 0.483) (Fig. 2e).

Survival Outcomes

In resected cases, non-BRCA1 TDP was not prognostic, showing no difference in overall survival when compared with typical cases (p = 0.77; HR = 0.901(0.4368 − 1.858) Fig. 3). Similarly, there was no significant difference in disease free survival in all resectable cases (p = 0.8966) as well as Moffitt classical cases (p = 0.6086). In the advanced cohort, we observed a trend towards improved response to first-line platinum therapy in non-BRCA1 TDP vs. typicals (Fisher exact test, p = 0.0602), with ORR 46.6% for TDP cases compared to 22.7% for typicals (Fig. 4a). When stratified for Moffitt classical tumors, platinum therapy was correlated with improved responses in non-BRCA1 TDP vs. typicals (Fisher exact test, p = 0.0036; Fig. 4a). The lack of association between non-BRCA1 TDP status and treatment response in patients with basal-like tumors highlights the chemoresistance of this more aggressive subtype. However, a difference in overall survival was not observed in the advanced cases, with a median OS of 10.8 months in the TDP cohort compared to 10.4 months in typicals (Fig. 4b). The addition of platinum did not improve survival, with a median OS of 11.8 months compared to 11.7 months with the typicals. When stratifying again for the Moffitt classical subtype, a numerically improved median OS of 15.4 months in the TDP cohort was noted compared to 12 months in the typical cohort, although this was not statistically significant (p = 0.38).

Fig. 3. Survival analysis in resected cases comparing nonBRCA1-TDP with typicals.

Kaplan-Meier curves showing the overall and disease free survival for early stage (Stage I and II) patients. The right panel shows Moffitt classical samples only. Log-rank p values are shown.

Fig. 4. Response to chemotherapy in advanced cases.

a Waterfall plot comparing response to platinum for advanced (stage III and IV) patients in the COMPASS/PanCuRx cohort on the left and for response to platinum for moffitt classic patients in the COMPASS/PanCuRx cohort on the right. b Kaplan-Meier curves for all advanced (Stage III/IV) patients, those receiving platinum based chemotherapy and Moffit classical cases that received platinum based therapy. c Kaplan-Meier curves of advanced patients receiving platinum based chemotherapies. (ii) Boxplot of chemotherapy duration across the etiologies. d Kaplan-Meier curves of advanced patients (Stage III/IV) receiving platinum-based chemotherapy by group 1 vs. group 2 TDP. BRCA1 samples are included in the HRD group while CCNE1 samples are in group 2.

Consistent with the accepted notion that patients with HRD PDAC benefit from exposure to platinum-based therapy, we observed that survival in our HRD cases, including those with BRCA1, was longer compared to the non-BRCA1 TDP and typical cohorts when OS was adjusted for RECIST response, with median survival of 12.1 months in HRD cases, as compared to 5.87 months in non-BRCA1 TDP and 4 months in typicals (p = 0.015) (Fig. 4c). We also observed that longer duration on chemotherapy was associated with better survival, which was consistent across HRD, non-BRCA1 TDP, and typical cases (Fig. 4c). Among patients receiving platinum-based therapy, the duration of time on platinum was significantly longer in HRD cases (p = 0.003) compared to typicals; however, we did not observe a significant increase in time on platinum for non-BRCA1 TDP cases (p = 0.12; Fig. 4c). This suggests that while responses to platinum were higher among the non-BRCA1 TDP cases, the response was not durable, likely accounting for the lack of difference seen in overall survival.

We performed additional analyses to compare survival between group 1 and group 2 TDPs. Group 1 and group 2 samples show clear separation with both typical and HRD samples falling between these two groups (Fig. 4d first panel). We then analysed patients treated with platinum-based chemotherapy by group 1 and group 2 TDP, demonstrating that patients with group 1 tumors have longer duration on platinum treatment (HR 4.97, p < 0.01) (Fig. 4d). When we examine only the subgroup of patients with unknown etiology of TDP, we observe a similar trend where those with group 1 tumors also have longer duration on platinum therapy (p < 0.01; Fig. 4d). This indicates that TDP span size correlates with survival on platinum-based systemic therapy.

Discussion

There is an urgent need to identify predictive biomarkers in PDAC to guide treatment decision-making beyond existing subgroups, including BRCA and PALB2 mutations. With the goal of offering targeted and more effective treatments to patients with PDAC, the TDP may represent another targetable subgroup. To our knowledge, this is the first report of TDPs in PDAC, demonstrating that patients with TDP PDAC may have improved responses to platinum-based therapy in the metastatic setting, although not consistently durable depending on the TDP etiology.

Responses remained more durable in patients with BRCA1-TDP tumors, but still present in non-BRCA1 TDP cases. Furthermore, we observed longer survival on platinum-based chemotherapy for class1 tumors, both across all TDP etiologies and among only those with unknown etiology. The rationale for the lack of consistent durable responses across the TDP cohort remains unknown. Preclinical data published by Menghi and colleagues showed that triple negative breast cancer cell lines were responsive to treatment with cisplatin and carboplatin, but not PARP inhibitors¹¹. BRCA1 expression levels were not correlated with response to cisplatin or carboplatin, suggesting that the response was modulated by TDP rather than the presence of a BRCA mutation. By contrast, we found that patients with BRCA1-TDP appeared to have longer responses to platinum-based therapy, suggesting that the etiology of the TDP phenotype may play a role in clinical responses. Post-progression biopsies were not available for the majority of our patients to better understand potential mechanisms of platinum resistance. This is an area of significant interest, with reversion mutations serving as a clear negative predictive marker in patients with BRCA alterations¹⁵. Future studies involving pre and post-treatment biopsies, with tissue-based genomic analysis incorporating RNA and methylation studies, may help to better delineate mechanisms limiting prolonged responses. The non-BRCA1 TDP cohort had a significantly higher proportion of tumors with basal-like subtype. Our group has previously shown that the presence of basal-like subtype is a poor prognostic factor in pancreatic cancer, and the higher proportion of basal-like subtype among the non-BRCA1 TDP cohort could negatively impact survival on platinum-based chemotherapy for these patients^13,14. Further stratifying our non-BRCA1 TDP samples by span size appears to strongly correlate with survival, and future work will benefit from this distinction.

The TDP can only be detected using WGS/RNA-Seq technologies as this requires interrogation of genome-wide structural variation patterns. WGS/RNA-Seq has been employed as part of the discovery process, uncovering RNA subtypes of PDAC, which can be predictive of chemotherapy response and identifying novel targetable fusions^14,16. However, WGS/RNA-Seq is not routinely used in the clinic with barriers to access, including resource availability and increased costs, but also challenges with the interpretation of variants of unknown significance and complex bioinformatics analyses. This profiling landscape may change as the cost of sequencing decreases. In fact, a decision analytic model in the context of metastatic non-small cell lung cancer suggested cost savings of next-generation sequencing compared to sequential single gene testing¹⁷. Similar economic analyses may be useful in determining the role of next-generation sequencing within a clinical framework as novel drugs become available and sequencing costs decrease.

Limitations of our study include the clinicopathologic stage heterogeneity of our samples, including both resected and advanced PDAC cases. However, we sought to understand the prevalence and clinical relevance of TDP across the PDAC spectrum and thus elected to include all PDAC cases in our analysis. TDP was present in only 9.9% of PDAC cases. While we combined cases from across these two studies to increase the sample size, we recognize that there are differences between the COMPASS/PanCuRx and PanGen/POG studies. All cases underwent laser capture microdissection in COMPASS/PanCuRx but not in the PanGen/POG studies, although the exact impact of this remains unclear. Furthermore, cases in the COMPASS study undergo biopsy at an earlier timepoint as patients were recruited from a rapid access pancreatic cancer clinic where patients are often seen within one week of initial presentation. This may favor a more aggressive biology as some of these patients may have rapid clinical deterioration after diagnosis, as indicated in the differences in median OS (9.3 months in COMPASS vs. 13.7 months in PanGen/POG). Further validation in a large orthogonal cohort of cases may help to identify potential driver genes as well as further delineate the impact of TDP on response to platinum-based therapies. Furthermore, it is difficult to ascertain associations between TDP and other therapies, such as PARP inhibitors, which are only approved for patients with BRCA mutations. While there is a rationale for the utility of PARP inhibitors in this setting, none of the patients in our cohort would have received this solely for the TDP indication.

Taken together, our data suggest that TDP presents a potential genomic marker for response to platinum-based chemotherapy among patients with advanced PDAC and may represent a novel targetable subgroup for chemotherapy selection. Further investigations are warranted with a larger sample size of TDP cases.

Methods

Patient population

We included all early-stage resected and advanced PDAC cases as part of the PanCuRx initiative and the Comprehensive Molecular Characterization of Advanced Pancreatic Ductal Adenocarcinoma for Better Treatment Selection (COMPASS) trial at the Ontario Institute for Cancer Research and Princess Margaret Cancer Centre in Toronto, Canada¹³. To increase our sample size, additional metastatic cases were included from the PanGen (Prospectively Defining Metastatic Pancreatic Ductal Adenocarcinoma Subtypes by Comprehensive Genomic Analysis; NCT02869802) and POG (BC Cancer Personalized OncoGenomics; POG; NCT02155621) studies from British Columbia, Canada^18,19. The COMPASS and PanGen trials are prospective multi-institutional Canadian translational studies, part of the Enhanced Pancreatic Cancer Profiling for Individualized Care (EPPIC) program, requiring a fresh tumor biopsy for WGS/RNA-Seq among patients with treatment-naïve advanced PDAC who were fit for combination systemic therapy. The PanCuRx and COMPASS studies were approved by institutional research ethics boards at Princess Margaret Cancer Centre (Toronto, Ontario, Canada), McGill University Health Centre (MUHC, Montreal, Quebec, Canada), and Kingston General Hospital (Kingston, Ontario, Canada). Approvals for the PanGen/POG studies were obtained from the University of British Columbia Research Ethics Committee. All patients provided written informed consent, and these studies were performed in concordance with the Declaration of Helsinki.

Tumor sequencing

For all PanCuRx and COMPASS cases, laser capture microdissection (LCM) was performed on frozen biospecimens for tumor enrichment. The aim was a WGS targeted sequencing coverage of 50x in tumor and 30x in normal and a minimum of 30 million unique mapped reads for transcriptome data. Raw sequencing reads were aligned to the human reference genome (hg38) using Burrow-Wheeler Aligner (BWA, version 0.7.17)²⁰. Germline variant calling was performed using the Genome Analysis Tool Kit (GATK, version 4.1.2)²¹. Somatic single nucleotide variants (SNVs) were identified as the intersection of calls by Strelka2 (version 2.9.10)²² and MuTect2 (version 4.1.2)²³. Indels were identified using the consensus between two of four callers: Strelka2, MuTect2, SVaBA (v134), and DELLY2 (version 0.8.1)^24,25. Somatic structural variants were identified using the consensus of 2 of 3 callers: SVaBA, DELLY2, and Manta (version 1.6.0)²⁶. HMMcopy (version 0.1.1)²⁷ and CELLULOID²⁸ were used to call copy number segments, tumor cellularity, and ploidy. RNA-Seq was performed at the Ontario Institute of Cancer Research as described previously²⁹. Briefly, reads were aligned to the human reference genome (hg38) and transcriptome (Ensembl v.100) using STAR v.2.7.4a³⁰. Gene expression was calculated in fragments per kilobase of exon per million reads mapped using the stringtie package v. 2.0.6³¹. Moffitt classification (classical vs. basal-like) was applied to each sample with sufficient RNA for analysis^6,13.

In the POG and PanGen studies, tumor and normal whole genome DNA libraries were sequenced to a target depth of 80× and 40× and transcriptomes were sequenced, targeting 150–200 million reads. POG/PanGen somatic mutation [single-nucleotide variant (SNV) and insertion/deletion (indel)] calls were derived using a combination of Manta v1.5.0 and Stelka v2.9.10, using default parameters. POG/PanGen variants were annotated using SnpEff v4.3³², with parameters -v GRCh37.75 -canon -no-downstream -no-upstream -noLog -noStats -no-intergenic. POG/PanGen copy-number variation (CNV) events were called using Facets v0.6.0³³, with default parameters. Gene expression values were calculated using Subread v1.6.3³⁴. Moffitt PurlST algorithm was used to classify samples as basal-like and classical³⁵.

TDP and HRD classifications

Samples were scored using the previously described TDP score by Menghi et al.¹¹ In addition to a positive Menghi score, samples with a duplication SV ratio cut-off of greater than 0.205, or 75% percentile with respect to the study cohort, were considered true TDP cases. All samples that met these criteria were classified based on etiology: BRCA1 (double hit in the BRCA1 gene), CCNE1 (amplification of CCNE1), or unknown alterations (no alterations in TDP-related genes). We then classified our samples based on span size distribution, as described previously by Menghi and colleagues. The span size of duplications for each sample was calculated and Mclust v6.0.0 was used to determine the underlying distributions found within the sample span size³⁶. For each identified distribution, the median size was calculated and compared to the Menghi class size distribution and used to classify our samples. Next we ran two driver discovery tools (ActiveDriverWGS v1.2.1 and dNdScv v0.0.1.0 on samples with unknown etiology^37,38. Using an adjusted p-value cutoff of <0.25, we identified potential driver genes found by either ActiveDriverWGS or dNdScv and ran Fisher exact tests between each class and the non-TDP samples to see if any of these potential drivers were significant.

Samples were classified as HRD as described in our previous published work, which incorporates hallmarks of HRD, including first and second hits in known HRD genes, SNV, SV, 4 bp+ deletion, 100-10kbp deletion, and 10 k = 1 mbp duplication loads, C to T ratio, 4 bp+ deletion rati,o and 100-10kbp deletion ratio³⁹. Samples that met a minimum of 6 hallmarks were considered HRD. Samples with 3–5 hallmarks were manually curated. In addition, we looked for the presence of signature 3 in all HRD cases. We also ran samples through HRDetect²⁰ and Classifier of Homologous Recombination Deficiency (CHORD) as further confirmation²². HRDetect is a classifier for homologous recombination deficiency that works by leveraging multiple mutational signatures, including single-base substitution signatures, structural variant signatures, and microhomology-mediated deletions. This classifier has been associated with platinum sensitivity in both breast and PDAC^39,40. CHORD is a random forest classifier which has been used to predict HRD across a number of tumor types, and has the ability to differentiate between BRCA1 and BRCA2 subtypes⁴¹.

Statistical analysis

Fisher’s exact tests were used to compare qualitative variables, Wilcoxon rank-sum test for pairwise comparison and the Kruskal–Wallis test for multiple group comparison in cases of quantitative variables. Patients receiving at least one cycle of chemotherapy were included in the analysis of overall response rate (ORR). TDP tumors with HR deficiency, such as BRCA1 deficient tumors, were excluded from clinical correlations, specifically association with platinum-based therapy, to avoid confounding. Overall survival (OS) was estimated using Kaplan–Meier curves and compared using the Cox proportional hazard regression. Statistical significance testing was two-tailed and set at p value less than 0.05. Statistical analysis was performed in R version 3.6.2.

Author contributions

E.T., J.K., and S.G. conceived and designed the study. J.T., G.O.K., S.F., A.D., S.H., J.W., R.C.G., K.L.A., G.Z., E.E., R.P., R.J., M.M., J.B., P.T., R.G., O.B., M.M., J.L., D.R., D.S., J.M.K, S.G., J.K., and E.T. contributed to data acquisition. A.R.F., A.Z., M.C.S., J.T., G.H.J., F.N.,. and E.T. contributed to data analysis. A.R.F., A.Z., M.C.S., J.T., G.H.J., F.N., and E.T. contributed to data interpretations. A.R.F., A.Z, M.C.S., G.O.K., S.G., J.K., and E.T. contributed to writing the manuscript.

Data availability

Genomic data generated within the PanGen/POG studies are available in the European Genome-phenome Archive (EGA) under accession number #EGAS00001001159. Genomic data generated within the PanCuRx initiative are also available in the European Genome-phenome Archive (EGA) and can be accessed under the accession number #EGAS00001002543.

Code availability

The code is available upon request.

Competing interests

All authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41698-025-00888-8.