The Evolutionary Origins of Recurrent Pancreatic Cancer

Abstract

Surgery is the only curative option for Stage I/II pancreatic cancer, nonetheless most patients will recur after surgery and die of their disease. To identify novel opportunities for management of recurrent pancreatic cancer we performed whole exome or targeted sequencing of 10 resected primary cancers and matched intrapancreatic recurrences or distant metastases. We identified that recurrent disease after adjuvant or first-line platinum therapy corresponds to an increased mutational burden. Recurrent disease is enriched for genetic alterations predicted to activate Mapk/Erk and PI3K/AKT signaling and develops from a monophyletic or polyphyletic origin. Treatment induced genetic bottlenecks lead to a modified genetic landscape and subclonal heterogeneity for driver gene alterations in part due to intermetastatic seeding. In one patient what was believed to be recurrent disease was an independent (second) primary tumor. These findings suggest routine post-treatment sampling may have value in the management of recurrent pancreatic cancer.

Introduction

Pancreatic ductal adenocarcinoma (PDA) is currently the 3rd leading cause of cancer death in the United States and is projected to become the 2nd cause of cancer death within five years (1). Several reasons account for these statistics, including an inability to diagnose the disease when at a curative stage, late presentation and modest impact of current best available therapies (2). There is a limited understanding of the genetics of recurrent disease which limits targeted therapy opportunities or informed design of clinical trials.

Approximately 10–15% of newly diagnosed PDA patients are diagnosed with early-stage disease (Stage I or II). For these patients, surgical resection followed by adjuvant therapy is the only option for cure (2). While long term survival following resection of PDA has been reported (3,4), the majority of patients who undergo resection will recur locally or at distant sites and die of their disease within five years. Several factors have been shown to have predictive or prognostic value for disease free or overall survival in resected PDA patients, including a high ratio of involved to total resected lymph nodes, larger tumor size, high tumor grade, the presence of vascular and perineural invasion, or variably positive margins (5,6). Venous invasion is very common in pancreatic cancer and may contribute to the aggressiveness of this disease (7). Molecular features of PDA have also been attributed to worse outcome after surgery. For example, patients with co-incident TP53 and SMAD4 alterations have shorter disease-free survival than patients whose tumors do not have these genetic alterations (8). Alternatively, the presence of a basal expression signature (9), paucity of an immune signature (10) or microbial dysbiosis (11) have also been associated with worse overall survival.

We have previously reported in a small cohort of treatment naïve Stage IV PDA patients that we found no evidence of driver gene heterogeneity among primary and metastatic sites (12). What heterogeneity was found in those patients corresponded to passenger mutations only. Moreover, the extent of passenger gene heterogeneity was far less than may be seen in spatially distinct cells within normal tissues indicating at least one clonal sweep occurred within the primary tumor prior to metastatic dissemination. By contrast, the extent to which metastatic PDA that arises following surgical resection exhibits similar features is unknown. To address this question and to improve our understanding of recurrent PDA following resection we performed whole exome and/or targeted sequencing of 10 primary PDAs, matched local (pancreatic resection bed) recurrences and multiple anatomically diverse metastases. We identified that pancreatic cancer recurrences following surgery have an increased mutational burden, distinct subclonal origins and in some instances are characterized by somatic mutations with potential implications for clinical management.

Results

We screened a collection of more than 160 PDA research autopsies to identify patients for whom a sample of their original surgical pancreatic resection was available, who underwent adjuvant treatment after surgery, and who had histologically confirmed recurrent disease within the pancreatic remnant and one or more metastases to anatomically distinct sites such as the liver, lungs or peritoneum (Fig. 1A,B). We identified nine such patients for study (Table S1). An additional patient was included that did not have metastatic disease at autopsy but did have an aggressive local recurrence with multiple geographically distinct samples of this mass available for profiling. One normal tissue sample from each patient was also used to distinguish somatic from germline variants (Table S2). Histologic review of each tumor sample indicated that in two patients the primary tumor was conventional ductal adenocarcinoma whereas the recurrent disease had squamous features (PAM39) or squamous differentiation (PAM46). In a third patient the primary tumor exhibited classic ductal adenocarcinoma whereas the recurrent disease was anaplastic (PAM37). One patient had a primary small cell carcinoma of the pancreas (PAM41). The recurrent disease was notable for transition to a large cell phenotype (Fig. S1).

Figure 1. Recurrent pancreatic cancer after adjuvant therapy has distinct genetic features compared to the primary tumor.

A. Schematic illustrating the temporal and spatially distinct sample types collected for this study. In all patients the primary tumor is designated PT1. The number of samples of recurrent disease varied per patient, ranging from PT2-PT4 up to PT2-PT11. B. Pancreatic tail remnant and attached spleen removed at autopsy of patient PAM41. Arrowheads indicate the resection margin and dashed circle indicate the location and dimensions of the recurrent tumor. C. Integration of mutation type, sample location, radiation therapy, chemotherapy, and mutational signature for each sample. D. Comparison of the prevalence of distinct mutational signatures in primary carcinomas (P) versus recurrent disease (R) per mutational signature indicates a statistically significant increase in mutations characteristic of the DNA damage repair signature only (two-sided χ2-squared test, p<0.0001). E. Comparison of DNA damage repair signature in patients who did (n=4) and did not (n=6) receive a platinum agent as adjuvant or first line therapy. All comparisons by two-sided Mann-Whitney U test.

All samples were analyzed by whole exome sequencing (WES) to a median of 330x coverage (range 135–652x)(see Methods, Table S3). To supplement the breadth of WES sequencing in these samples, we also performed concurrent deep sequencing using a targeted sequencing panel to ensure greater sensitivity for mutations and DNA copy number alterations in 410 cancer associated genes (Table S4) (13). We identified 4864 total somatic single nucleotide variants (SNVs) and small insertions or deletions (indels) with a median of 57 per sample (range 17–203; Fig. 1C, Table S5). We found no evidence of microsatellite instability in these 10 patients, nor were germline or somatic mutations identified in recognized MSI-related or other DNA damage repair genes (14).

To better understand the patterns of mutation accumulation in primary versus recurrent disease, we performed mutational signatures analysis of all samples that underwent WES (Fig. 1C). Nine signature classes were identified in our cohort: aging (signatures 1 and 5), double strand break repair (DSBR, signature 3), apolipoprotein B mRNA editing enzyme, catalytic polypeptide-like (APOBEC, signatures 2 and 13), mismatch repair defects (signatures 6, 15, 20, 21 and 26), tobacco (signature 4), somatic hypermutation (signature 9), and the three unknown signatures 8, 17 and 23 (15). A high prevalence of Signature 9 (POLN) was seen in at least one sample in seven of eight patients. This signature is associated with somatic hypermutation by polymerase η, a DNA polymerase that plays a role in DNA repair by translesion synthesis (16). As previously reported, we also identified subsets of patients with an abundance of signatures with unknown etiology such as signatures 8 and 17 (10). Patient PAM41 with a primary small cell carcinoma of the pancreas had a remarkable abundance of mutations characteristic of signature 23, a rare type of signature of unknown etiology (15). Comparisons of the prevalence of each signature class specifically in post-treatment recurrent disease versus treatment naïve primary carcinomas revealed a significant increase in the double strand break repair signature only (median 16 somatic mutations per primary tumor versus 92 in recurrent disease, two-sided χ2-squared test, p<0.0001) (Fig. 1D). Four patients received a platinum agent as part of their adjuvant or first line therapy for recurrent disease (Table S1), thus we determined the extent that this signature was enriched in these patients compared to those who received other regimens. Comparison of these two groups revealed that patients who did not receive a platinum agent had a modest yet significant increase in their recurrent disease with respect to the double strand break repair signature (median 58 somatic mutations per primary versus 532 in recurrent disease, Mann-Whitney U test, p<0.002) (Fig. 1E). Moreover, as previously described (17,18) this signature was remarkably more prevalent in the recurrent disease of patients who received one or more platinum agents (median 84.5 somatic mutations per primary versus 1379 in recurrent disease, Mann-Whitney U test, p<0.0001).

We next determined the somatic alterations in known cancer genes in each patient. We identified somatic mutations in known PDA driver genes predicted to have functionally deleterious effects such as those in KRAS, CDKN2A, TP53, SMAD4, and ARID1B (Fig. 2A, Table S6 and S7). Collectively, the genetic features of the resected PDA samples (sample PT1 from each patient) were consistent with previous sequencing studies of these cancers (9). By contrast, the genomic features of recurrent disease (Fig. 2B, Table S6) were notable for somatic alterations in additional genes, some of which are predicted to or likely to activate Mapk/Erk signaling (G12V mutation in KRAS, I679Dfs*21 mutation in NF1, R111X mutation in PPP6C)(19,20), PI3K/Akt/MTOR signaling (D323H hotspot mutation in AKT1, E542K hotspot mutation in PIK3CA, homozygous deletion of STK11)(21), or MYC/MAX regulated gene expression (Y1312X mutation in CHD8, W1004X mutation in MGA, X863_ splicing mutation in NOTCH1, up to 16-fold amplification of MYC) (22,23). Somatic alterations in genes predicted to affect chromatin-mediated gene expression (up to 38-fold focal amplification of HIST1H3B, frameshift mutations in KMT2B, KMT2C, KMT2D, c.4471–1N>A splicing mutations in TRIP12) (24–26), nuclear export (E571K hotspot mutation in XPO1) (27) and DNA damage repair (c.497–1N>A splicing mutation in ATM, F134V mutation in TP53) were also found (28,29). Recurrences from two patients contained copy number alterations of genes implicated in innate immunity signaling (PRKCI 38-fold focal amplification, TRAF3 homozygous deletion) (30,31). Whole genome duplication (WGD) was present in one or more samples of recurrent disease in eight patients (32) (Table S7). In only two of these eight patients was WGD detected in the primary tumor indicating gains in ploidy may accompany PDA progression. Finally, we determined the extent that alternative mechanisms of increasing mutant KRAS signaling exist in recurrent disease. Allelic imbalance for mutant KRAS was identified in at least one sample of the recurrent disease in all ten patients including both patients for which KRAS mutations were identified in the recurrent disease (Fig. 2C and Table S8). In five patients two or more routes to allelic imbalance occurred independently during clonal progression indicating convergent evolution towards increased KRAS signaling (Fig. 2D,2E). Collectively these data identify potential signaling and transcriptional pathways by which recurrent disease develops.

Fig. 2. Genetic Features of Recurrent Pancreatic Cancer.

A. Oncoprint of somatic mutations and copy number alterations present in the pre and post-treatment samples analyzed for each patient (clonal mutations). B. Oncoprint of somatic mutations and copy number alterations present in one or more samples of the recurrent PDA only, but not in the matched originally resected primary tumor (subclonal mutations). The full list of somatic alterations identified in both the primary and recurrent disease are listed in Tables S6 and S7. C. Frequency of diploid (D), tetraploid (T) and allelic imbalance (AI) at the KRAS locus in primary tumors versus recurrent disease. D. Proportion of genetic events leading to allelic imbalance in recurrent pancreatic cancer. Complete data are shown in Table S8. E. Schematic illustration of convergent genetic events that increase mutant KRAS dosage in the recurrent disease of PAM37. Abbreviations used are WGD, whole genome duplication and CN LOH, copy neutral loss of heterozygosity.

We hypothesized that somatic alterations identified in recurrent PDA samples reflect the clonal expansion of pre-existent populations following selective pressure imposed by adjuvant therapy. Pairwise cancer cell fraction (CCF) plots generated for all patients confirmed the enrichment of one or more subclonal populations in the recurrent disease compared to the primary tumor (Fig. 3A). We specifically focused on the AKT1 D323H in PAM40, PIK3CA E542K in PAM39 and NOTCH1 X863_splice site mutation in PAM41 as they represent functionally relevant alterations and in theory may be clinically actionable (http://oncokb.org). The KRAS p.G12V mutation in PAM40 was also of interest given recent reports of subclonal KRAS mutations in PDA (9). Droplet digital PCR confirmed that in all instances these mutations existed in the primary tumor at prevalence rates from 0.2–2%, below the level of detection of our WES (Fig.3B–D).

Figure 3: Somatic alterations in recurrent pancreatic cancer alterations reflect expansion of subclones pre-existent in the primary tumor.

A. Representative density cloud plots of the primary tumor (PT1) and one matched sample of recurrent disease for nine different patients. Subclonal expansion of cells containing a PIK3CA E524K mutation (PAM39), a KRAS G12V and AKT D323H mutation (PAM40), a NOTCH1 frameshift mutation (PAM41), KMT2C frameshift mutation (PAM45) and CHD8 nonsense and missense mutations (PAM46) in the recurrent disease are seen. The cancer cell fractions of representative clonal driver genes for all cases (i.e. KRAS, TP53, SMAD4 and/or GNAS) are shown for reference. B-D. Droplet digital PCR analysis of mutant allele abundance in the primary tumor and one matched sample of recurrent disease in three different patients. In all three patients subclones were preexistent at low allele frequencies in the primary tumor. Each dot represents one droplet, and color bars at top right of each plot indicate relative intensity of the VIC labeled mutant (x axis) and FAM labeled wild type allele (y axis) fluorescent labels.

To understand the evolutionary origins of recurrent disease in each patient we inferred phylogenies based on high confidence mutations in each patient (Methods). One patient who was clinically thought to have an intrapancreatic recurrence of the resected PDA after 18 months, but molecular analysis revealed a second (independent) primary PDA (Fig. 4A). For example, the resected primary tumor (PAM44PT1) had a KRAS p.G12R mutation and 80 private passenger mutations, whereas the “recurrence” samples from this patient (PAM44PT2-PT5) shared 124 mutations not seen in the primary. These latter mutations included a KRAS p.G12D, a TP53 p.F134V and a 29 basepair frameshift deletion in KMT2D (Fig. 4B). A CCF plot of the original primary (PAM44PT1) compared to the “recurrence” samples confirmed the mutual exclusivity of the somatic mutations in each lineage (Fig. 4C). Re-review of the histology of the first primary tumor and associated imaging studies did not suggest the presence of a cystic neoplasm. These findings are highly indicative of two independent primary tumors that arose from distinct precursors (33).

Figure 4: Metachronous pancreatic cancer simulating a local recurrence.

A. Clinical timeline of events in this patient. Liver and peritoneal metastases were radiographically evident at 18 months and a mass in the pancreatic remnant was noted at autopsy. B. Phylogenetic analysis of whole exome sequencing data generated for this patient indicates the metachronous pancreatic cancer contains a unique repertoire of somatic driver gene alterations including a KRAS-G12D mutation compared to the first primary (PT1) that contains a G12V mutation. C. Representative density cloud plot of the first primary tumor (PT1) and one sample of the metachronous primary tumor (PT3). The majority of somatic alterations in each primary are mutually exclusive from each other.

In the remaining nine patients the sample of resected primary tumor and all samples of recurrent disease arose from a common ancestor of neoplastic cells containing canonical PDA driver mutations. However, there were two distinct evolutionary trajectories by which recurrent disease arose. For five patients (PAM37, PAM38, PAM40, PAM42, PAM46) the resected primary tumor sample formed the outgroup in the phylogeny (Fig.5A,B and Figs. S2A–H) indicating that in these patients the recurrent disease developed from a single residual clonal population, i.e. a monophyletic origin. In the remaining four patients (PAM39, PAM41, PAM43, PAM45) (Fig. 5C,D and Fig. S3A–F) the recurrent disease was inferred to be seeded by multiple ancestral clones and was polyphyletic in origin. For a more objective metric of relatedness among the primary and recurrent disease in each patient we calculated pairwise Jaccard similarity coefficients for all samples within a patient. The average Jaccard indices per patient supported the distinction of monophyletic versus polyphyletic origins of recurrent disease (Fig. 5E–G) in that monophyletic recurrences were significantly more distant (divergent) from the primary tumor, whereas polyphyletic recurrences were highly related to the primary tumors. Recognizing the sample numbers are low for robust outcomes analysis, exploratory analysis indicates that patients with monophyletic recurrences had a longer disease-free but not overall survival (Fig. 5H,I). From phylogenetic analysis alone, the timing of dissemination to other organs cannot be readily resolved. However, utilizing mathematical modeling and previously measured metastasis doubling times (34), we found that the minimal time required to grow from one to a billion cells (roughly 1 cm³) is 1.82 years, or 21.9 months (90% CI: 1.61 – 2.05 years) (Methods). Since clinical metastases occurred much earlier than these required 1.82 years after surgery in all patients with distant disease (median metastasis free survival 11.0 months, range 6–18 months)(Table S1), at least one of the metastases must have been microscopically seeded before surgery. Patient PAM46 typified these dynamics as they had a grossly positive surgical margin and developed a radiographically evident locoregional recurrence, but no metastases, 17 months after surgery (Table S1). Irrespective of origin, in all nine patients additional subclonal expansions occurred after dissemination, in some cases resulting in spatial heterogeneity for driver gene mutations among the different sites of recurrent disease (Fig. 5 and Figs. S2,S3).

Figure 5: Recurrent pancreatic cancer originates from two distinct evolutionary origins.

A. Phylogenetic analysis of the relationships of the primary tumor to the local recurrence and liver metastases in patient PAM40. All recurrent disease is the result of clonal expansion of a single pre-existent subclone notable for an AKT1 and KRAS mutations (See Fig 2B). The primary tumor (PT1) is the outgroup in the tree. A subclonal TRIP12 mutation is also seen in a single liver metastasis. B. Color code of sample origins in PAM40. C. Phylogenetic analysis of the relationships of the primary tumor to the local recurrence and liver metastases in patient PAM39. In this patient the recurrent disease is the result of more than one clonal expansion. The preexistent PIK3CA mutation has expanded in the lineage that gave rise to samples PT2 and PT9. D. Color code of sample origins in PAM39. E,F. Jaccard indices for each pairwise comparison in PAM40 (monophyletic recurrence) versus PAM39 (polyphyletic recurrence). G. Comparison of the average Jaccard index for primary tumors and their matched recurrences in patients with monophyletic recurrences versus those with polyphyletic recurrences. Monophyletic recurrences are significantly different from their matched primary tumor indicating passage through a genetic bottleneck, whereas no difference is found between the primary tumor and matched recurrences in patients with polyphyletic recurrences. (comparisons by a Student’s two-sided T test). Monophyletic (“common origin”) recurrences are associated with an improved disease-free survival (H) but not overall survival (I) in this small cohort (comparisons by log-rank test of Kaplan Meier survival curves).

We next sought to understand the clonal origins of local recurrences, a major clinical issue for patients who undergo potentially curative resection (35), by inferring the migration patterns of recurrent disease across spatially distinct sites in each patient (Methods). These analyses indicated that the seeding patterns of recurrent disease were diverse, with multiple patterns coexisting in the same patient. Metastasis to metastasis seeding was evident in some patients, typified by PAM37 in whom an omental metastasis seeded three liver metastases (Fig. 6A,6B), in PAM38 in whom a liver metastasis seeded a lung metastasis (Fig. 6C,6D), and in PAM45 in whom a perirectal metastasis seeded two abdominal wall metastases (Fig 6E,6F). Local recurrences can also be seeded by the primary tumor or by a metastasis. In PAM41 (Fig. 1B and Fig. 7A,7B) and PAM43 (Fig. 7C,7D), the primary tumor seeded the local recurrence despite both having negative surgical margins at the time of resection. By contrast, PAM38 (Fig. 6C,6D) and PAM42 (Fig. 7E,7F) had local recurrences seeded by a liver or lung metastasis, respectively. Analysis of PAM39 and PAM40 indicated that two or more seeding patterns were equally likely to have occurred for all sites analyzed and thus no dissemination events could be confidently inferred. Inter-metastatic seeding was inferred in patients with both monophyletic (PAM37, PAM38, PAM42, Fig. 6) and polyphyletic (PAM41, PAM43, PAM45, Fig. 7) recurrences collectively indicating that migration patterns are unrelated to phylecity of the recurrent disease. While we studied a small cohort, surgical margin status appeared to be unrelated to the origin of a local recurrence. For example, local recurrences seeded by the primary tumor were noted in patients with negative surgical margins (i.e. PAM41, Fig. 7A,7B) whereas local recurrences seeded by a metastasis were noted in patients with a positive surgical margin (i.e PAM38, Fig. 6C,6D).

Figure 6: Examples of intermetastatic seeding in recurrent pancreatic cancer.

For each patient the previously resected primary tumor is indicated by PT1 to the left of each patient schematic. +M, positive surgical margin; +LN, positive lymph nodes. In the body maps for all three patients (panels A,C,E), red lines reflect migration from the primary tumor and blue lines indicate migration from a site of recurrent disease. Solid lines indicate high confidence migration patterns and dashed lines indicate low confidence migration patterns inferred by MACHINA (panels B,D,F). Each patient had at least one high confidence migration event from one site of recurrent disease to another. In patient PAM45, two migration patterns were equally likely to have occurred (F) although both predict that the primary tumor seeded the local recurrence and the perirectal metastasis seeded the abdominal metastases.

Figure 7: Origins of local recurrences after surgical resection.

For each patient the previously resected primary tumor is indicated by PT1 to the left of each patient schematic. -M, negative surgical margin; +LN, positive lymph nodes, -LN, negative lymph nodes. In the body maps for all three patients (panels A,C,E), red lines reflect migration from the primary tumor and blue lines indicate migration from a site of recurrent disease. Solid lines indicate high confidence migration patterns and dashed lines indicate low confidence migration patterns inferred by MACHINA (panels B,D,F). In PAM41 (A,B) and PAM43 (C,D) the local recurrence is seeded by the originally resected primary tumor, whereas in PAM42 (E,F) it is seeded by a lung metastasis. In patients PAM41 (A,B) and PAM42 (E,F) three or more migration patterns were equally likely to have occurred.

Discussion

To date genomic studies of PDA have primarily relied on resections or biopsies of treatment naïve disease (9,36,37). Herein, we demonstrate that study of post-treatment samples may have value in identifying putative therapeutic vulnerabilities in recurrent disease. Of immediate clinical relevance is the finding that recurrent disease is enriched for somatic alterations in genes associated with Mapk/Erk or PI3K/Akt signaling, some of which are potentially actionable (http://oncokb.org). Importantly, these targets are pre-existent and selected for during adjuvant treatment itself. While our sample size is not sufficient for a robust gene discovery in advanced pancreatic cancer, we identified genes not commonly associated with the PDA landscape such as the nuclear exportin XPO1 (21), the serine-threonine protein phosphatase PPP6C (28), or regulators of innate immunity (PRKCI, TRAF3) (30,31). These findings support the need for prospective and statistically robust efforts to sequence post-treatment PDA to better define the genes or pathways repeatedly targeted by somatic alteration. We caution readers that the evolutionary timing of these potentially actionable events is critical to know. For example, in patient PAM40 all recurrent disease analyzed was the result of expansion of a single subclone with a known deleterious AKT1 D323H mutation; in theory all recurrent disease in this patient may have been sensitive to an AKT1 inhibitor. By contrast, in patients such as PAM41 the PIK3CA E542K mutation occurred in a minor subclone of the recurrent disease and hence would not be suitable for targeting. Nonetheless, while our sample set is small it indicates that such instances may occur in PDA; such information could arm a clinician with more information for how to manage recurrent disease.

Another important finding of this study is that it illustrates a taxonomy by which recurrent PDA may be stratified: monophyletic origin, polyphyletic origin or unique origin (i.e. synchronous/metachronous primaries). In all patients with bona fide recurrent PDA of mono- or polyphyletic origin we find that systemic subclinical dissemination had already occurred at the time of surgery, consistent with prior estimates (38). Questions for future investigation thus relate to methods to identify mono- versus polyphyletic recurrences in real time, the extent to which subclonal diversity develops within each category, and the clinical significance of this finding in the context of ongoing or planned clinical trials. Finally, while second primary carcinomas of the pancreas have been reported, our finding of a metachronous primary PDA in one of 10 otherwise unselected patients with an intrapancreatic mass post resection suggest that this phenomenon may be more common than previously appreciated (39,40). Detailed and formal prospective studies of intrapancreatic masses in patients who have undergone prior resection for PDA will be required to more firmly understand the frequency of second primary tumors and the risk factors associated with their development as has been shown for invasive carcinomas arising in intraductal papillary mucinous neoplasms of the pancreas (41).

Finally, we note that the genomic features of these patients, all of whom presented with Stage I/II disease and underwent surgery and therapy, are in stark contrast to those of untreated Stage IV PDA (12). In untreated PDA the genetic features of both the primary and metastases are remarkably uniform, and the genetic heterogeneity seen appears due exclusively to passenger mutations, suggesting at least one clonal sweep occurred prior to metastatic dissemination. By contrast, Stage I/II disease is notable for subclonal heterogeneity for driver genes as reported in other tumor types (42,43). Treatment induced genetic bottlenecks that sculpt the genomic landscape of PDA, together with intermetastatic seeding, likely contribute to subclonal and inter-metastatic heterogeneity for driver gene alterations observed in recurrent disease. Thus, context appears key in the interpretation of heterogeneity in PDA.

In summary, we identify novel genetic features of PDA in the context of recurrent disease after surgical resection and treatment with potential clinical implications for use of targeted therapies in disease management. In the event that therapeutically targetable gene alterations are validated in recurrent disease, and as more targeted therapies become available, post-treatment sampling is likely to contribute to identification of the mechanisms of resistance and early identification of resistant clones.

Methods

Tissue Samples.

Patient samples were generously shared by the Gastrointestinal Cancer Rapid Medical Donation Program (GICRMDP) resource at The Johns Hopkins Hospital. Sections were cut from formalin-fixed paraffin embedded (FFPE) or frozen sections available and reviewed to identify those with at least 20% neoplastic cellularity and preserved tissue quality. Samples meeting these criteria were macro-dissected from serial unstained sections before extraction of genomic DNA using DNeasy® Blood & Tissue Kits for frozen samples or QIAamp DNA FFPE Tissue Kits for FFPE materials (Qiagen) following the manufacturer instructions.

Whole Exome Sequencing.

DNA quantification, library preparation and sequencing were performed in the Integrated Genomics Operation and bioinformatics analysis of somatic variants by the Bioinformatics Core at MSKCC. Libraries were created using AgilentExon_51MB_hg19_v3 as bait and sequenced on an Illumina HiSeq 2500. Whole exome sequencing resulted in a median target sequence depth of 317X (min-max) with 83% of the coding regions covered at least 100x and a calculated average tumor cellularity of 35.7% (Table S3). Paired-end sequencing data were aligned to the reference human genome (hg19) using BWA. Read de-duplication, base quality recalibration, and multiple sequence realignment were performed using the Picard Suite and GATK. Point mutations and small insertions and deletions were detected with MuTect and HaplotypeCaller, respectively. Genome-wide total and allele-specific DNA copy number was determined using the FACETS algorithm (44). For targeted sequencing of 410 cancer genes, barcoded libraries from patient-matched tumor and normal samples were captured and sequenced using methods previously reported in detail (13). Raw sequencing data has been deposited in the European Genome Phenome Archive (EGA) under accession number EGAS00001004097.

Filtering and Annotation of Variants.

For each patient, somatic variants were filtered using the following criteria: patient-matched normal coverage ≥10 reads, variant count in patient-matched normal ≤1, patient-matched normal variant frequency <0.02, tumor coverage ≥20 reads, and tumor variant allele frequency ≥0.05 in at least one tumor sample. The resulting list of somatic variants were filtered for those present in the coding regions only and subject to further bioinformatic annotation for pathogenicity and germline allele frequencies from healthy populations distributed worldwide using LiFD (45)(Table S5). MSI status was inferred from the sequencing data using a clinically validated algorithm (46), with MSI-H defined as an MSIsensor score ≥10. For all mutations, we inferred the cancer cell fraction (CCF) from the mutant allele fraction, local copy number, and FACETS estimate of tumor purity according to previously described methods (47,48). Confidence intervals for the CCF therefore reflect the 95% binomial confidence interval for the underlying mutant allele fraction. Whole gene copy number alterations and whole genome duplication were inferred by FACETs. Genes with copy number of 8-fold or greater inferred by FACETs were considered amplifications.

Phylogenetic analysis.

We applied Treeomics v1.6.0 to reconstruct the phylogenies of recurrent disease using high quality somatic variants identified by whole exome sequencing. Treeomics uses a Bayesian inference model to account for sequencing errors and low purity and employs Integer Linear Programming to infer a maximum likelihood tree.

Mathematical modeling.

To calculate the minimal required time a metastasis founding cell needs to grow to a detectable lesion of 1 cm³ (~10⁹ cells) we used the smallest previously measured PDA metastasis doubling time of 27 days (49) leading to an exponential growth rate of r=0.026 per day. Assuming a PDA cell division time of 2.3 days (50), the expected minimal time for a tumor conditioned on survival takes to reach 10⁹ cells is 1.82 years (90% CI: 1.61 – 2.05 years).

Mutational signatures.

Mutational signatures were derived using the methods described by Alexandrov et al (15). To enrich for the most abundant signatures we merged those with similar putative etiologies into a single group. Signature groups present in at least 20% abundance in at least one sample were included for additional study and statistical analysis.

Droplet Digital PCR.

Absolute quantification of mutant alleles was determined using a RainDrop Plus Digital PCR system according to the manufacturer instructions. Predesigned or custom designed TaqMan assays were obtained for variants of interest (Thermo Fisher). Approximately 75 ng of gDNA were used per reaction in a 25 ul volume. Each reaction contained 5 million droplets at a target inclusion rate of 10% (500,000 target molecules).

Migration Pattern Inferences.

PyClone and MACHINA were used to infer seeding patterns associated with metastasis and local recurrence. To alleviate long run times associated with the high sample number context we applied specific filters to focus on the most informative loci. These were a) the locus was sequenced to a depth of at least 60 in at least one sample; b) the locus had a copy number profile consistent with a relatively simple genomic history in all samples (the combination of major allele A and minor allele B at the locus was required to be one of AB, AAB, AAAB, or AABB in each sample, though not necessarily the same across samples) and c) each locus was required to be sequenced to nonzero depth in all remaining samples. Samples that did not contain more than 20 variant loci after applying this filter were excluded. In patient PAM41, a large cluster of highly related liver metastases (PT4, 8, 9, 10, 11) were simplified by including only sample PT9 to improve runtime. The resulting mutations and associated major and minor copy numbers were clustered with PyClone using default settings. The PyClone consensus cluster files were used to enumerate evolutionary relationships, then the combination of PyClone cluster frequency estimates and enumerated trees together were used to search for the most parsimonious migration patterns consistent with each tree topology. The resulting solution with the lowest overall migration number, and then the lowest comigration number, were determined. No other constraints were applied to the migration plots.

Statistics.

Descriptive data are represented as a mean and standard deviation unless otherwise mentioned. Parametric distributions were compared by a Student’s T-test whereas non-parametric distributions were compared by a Mann-Whitney U test. Frequency data were compared using a X² test. All comparisons were two-sided. Survival curves were plotted according to the methods of Kaplan and Meier and compared using a log-rank test.

Statement of Significance:

The biologic features or clinical vulnerabilities of recurrent pancreatic cancer after pancreaticoduodenectomy are unknown. Using whole exome sequencing we find that recurrent disease has a distinct genomic landscape, intermetastatic genetic heterogeneity, diverse clonal origins and higher mutational burden than found for treatment naïve disease.