Purpose:

RNA-sequencing–based subtyping of pancreatic ductal adenocarcinoma (PDAC) has been reported by multiple research groups, each using different methodologies and patient cohorts. “Classical” and “basal-like” PDAC subtypes are associated with survival differences, with basal-like tumors associated with worse prognosis. We amalgamated various PDAC subtyping tools to evaluate the potential of such tools to be reliable in clinical practice.

Experimental Design:

Sequencing data for 574 PDAC tumors was obtained from prospective trials and retrospective public databases. Six published PDAC subtyping strategies (Moffitt regression tools, clustering-based Moffitt, Collisson, Bailey, and Karasinska subtypes) were used on each sample, and results were tested for subtype call consistency and association with survival.

Results:

Basal-like and classical subtype calls were concordant in 88% of patient samples, and survival outcomes were significantly different (P < 0.05) between prognostic subtypes. Twelve percent of tumors had subtype-discordant calls across the different methods, showing intermediate survival in univariate and multivariate survival analyses. Transcriptional profiles compatible with that of a hybrid subtype signature were observed for subtype-discordant tumors, in which classical and basal-like genes were concomitantly expressed. Subtype-discordant tumors showed intermediate molecular characteristics, including subtyping gene expression (P < 0.0001) and mutant KRAS allelic imbalance (P < 0.001).

Conclusions:

Nearly 1 in 6 patients with PDAC have tumors that fail to reliably fall into the classical or basal-like PDAC subtype categories, based on two regression tools aimed toward clinical practice. Rather, these patient tumors show intermediate prognostic and molecular traits. We propose close consideration of the non-binary nature of PDAC subtypes for future incorporation of subtyping into clinical practice.

Translational Relevance

Previous studies of sequencing-based PDAC subtyping have focused on a PDAC subtyping scheme that forces binary classification of patient tumors into basal-like and classical groups. Using published subtyping tools aimed toward clinical utility, our study highlights the fact that not all PDAC tumors unambiguously align to either side of the proposed PDAC subtype spectrum, and that such patients show intermediate clinical and molecular characteristics. We provide evidence that such patients do not conform to the binary PDAC subtypes, and emphasize the importance of treating the classical/basal-like phenotype as a continuum when considering the adoption of PDAC molecular subtyping into clinical practice.

Advancements in the clinical management of pancreatic ductal adenocarcinoma (PDAC) have been limited by the aggressive nature of the disease and the fact that patients often present with advanced disease (1, 2). Patient prognosis remains generally poor (3), creating an unmet need for improved prognostic and treatment stratification tools. Several groups have aimed to address this using next-generation sequencing (NGS) data to identify transcriptional subtypes of PDAC associated with patient survival (4–6), and subtypes were later found to share considerable overlap (7, 8). A recent review of all PDAC subtypes proposed that prognostic classifications be categorized into two major and distinct lineages: classical and basal-like (or “squamous”), the latter being associated with poor clinical outcome (9).

Apart from their prognostic significance, NGS-based PDAC subtypes may hold additional clinical relevance, as several studies have indicated that patients with basal-like subtype tumors are less sensitive to first-line treatment (10–12). To bring PDAC subtyping closer to clinical feasibility, an additional procedure to classify PDAC subtypes was recently published, and offers to remove the requirement for NGS technology by providing a tool compatible with the NanoString platform (13).

Using unsupervised clustering of gene expression data, basal-like and classical PDAC subtypes are largely defined by mutually exclusive gene expression patterns, for sets of genes associated with each subtype (5). Hybrid samples, in which both basal-like and classical genes are highly expressed, have been consistently observed during clustering-based subtyping, but are often merged with classical tumors (5, 8, 14). Chan-Seng-Yue and colleagues were the first to describe the hybrid PDAC samples in their description of de novo PDAC classification, and used single-cell sequencing to demonstrate that such tumors bear a mix of both classical and basal-like cell populations (11). While survival differences between patients with basal-like versus classical PDAC tumors is well-established, clinical characterization of patients with hybrid tumors has not been performed. With the emergence of subtyping tools aimed toward clinical utility (13), understanding how hybrid PDAC tumor samples are classified across different tools is increasingly important, and remains not fully explored. To address these points, we performed six PDAC subtyping procedures on a diverse cohort of 574 advanced and resectable PDAC tumor samples. Sixty-eight of 574 (12%) of patient samples were unable to be classified consistently across methodologies, and showed intermediate clinical and molecular profiles. Our work therefore highlights a subgroup of PDAC tumors that fail to conform to the binary molecular subtypes of PDAC, which will need to be considered in future endeavors aimed at incorporating molecular PDAC subtyping procedures into clinical practice.

PDAC patient tumor data

Data was leveraged from two retrospective (The Cancer Genome Atlas (TCGA PAAD-US; n = 130) and International Cancer Genome Consortium (ICGC PACA-CA; n = 175) and three prospective (Prospectively Defining Metastatic Pancreatic Ductal Adenocarcinoma Subtypes by Comprehensive Genomic Analysis (PanGen, NCT02869802; n = 48), the BC Cancer Personalized OncoGenomics program (POG, NCT02155621; n = 26), and Comprehensive Molecular Characterization of Advanced PDAC For Better Treatment Selection (COMPASS, NCT02750657; n = 195)) datasets. ICGC and TCGA samples were filtered for true PDAC samples as described by Karasinska and colleagues (7). POG and PanGen data were sequenced by the same institution, and will hereafter be collectively referred to as POG/PanGen (n = 74). POG/PanGen and COMPASS data represent advanced (metastatic or locally-advanced) PDAC patient samples, while ICGC/TCGA data represent resectable tumors. For prospective trials, patients were enrolled as described previously (10, 15). POG/PanGen and COMPASS studies were approved by the University of British Columbia Research Ethics Committee (REB# H12–00137, H14–00681, H16–00291) and the University Health Network Research Ethics Board (REB# 15–9596), respectively. Studies were conducted in accordance with international ethical guidelines. Written informed consent was obtained from each patient prior to genomic profiling.

Genome and transcriptome sequencing

All sequencing data was collected and processed as described previously (7). Briefly, POG and PanGen tumor libraries were sequenced to a depth of 80x (tumor genomes), 40× (matched normal genomes) and 200 million reads (transcriptomes). COMPASS samples were sequenced as described by Aung and colleagues (10). Publicly available TCGA and ICGC data were downloaded from their respective online data portals (https://portal.gdc.cancer.gov/ and https://dcc.icgc.org/, respectively).

Genome and transcriptome data processing

POG/PanGen and COMPASS RNA-seq libraries were processed as described previously, including log10 transformation (7). TCGA normalized gene expression values were converted to transcripts per million (TPM) and log-transformed [log10((normalized_count*1e6)+1)]. ICGC normalized gene expression values were log-transformed [log10(normalized_count +1)]. POG/PanGen somatic mutation [single nucleotide variant (SNV), insertion/deletion (indel)] calls were derived using a combination of Manta v1.5.0 (16) and Stelka v2.9.10 (17), using default parameters. POG/PanGen and COMPASS variants were annotated using SnpEff v4.3 (18), 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 v0.6.0, with default parameters. COMPASS SNV, indel and CNV data were generated as described by Aung and colleagues (10). Human genome build GRCh37 (hg19) was used for all mutation analyses.

Batch correction

RNA-seq data was corrected for cohort-specific batches (POG/PanGen, COMPASS, ICGC, and TCGA) as well as four library preparation protocols used within the POG/PanGen batch (for a total of seven batches), using an empirical Bayesian approach (19). Principal component analysis (PCA) of the top 10% most variable genes confirmed alleviation of inter-sample batch effects after correction (Supplementary Fig. S1).

Calculation of tumor content

Tumor content output from Facets was used for POG/PanGen samples. Tumor content values for COMPASS and ICGC samples were determined as described previously (10). Tumor contents (ABSOLUTE; ref. 20) obtained from the GDC data portal were used for TCGA samples.

Classification of PDAC subtypes

Consensus clustering was used for Collisson (4), Bailey (6), and the clustering-based Moffitt (5) subtypes, as outlined previously and in a manner consistent with both original publications and more recent studies (7, 10). R v3.5.1 package “ConsensusClusterPlus” was run on all samples using genes specific to each of the three studies, with parameters reps = 50, pItem = 0.8, pFeature = 1, clusterAlg = “hc”, distance = “pearson”, seed = 123, and maxK = x, where x = 3 for Collisson subtyping, 4 for Bailey and 2 for Moffitt. For Collisson subtyping, 62 genes from the original publication were used (9). Bailey subtyping was based on genes from differential expression analysis results from the original publication (6), which were filtered for genes with an adjusted P < 0.05 and log2 fold change > 0, resulting in 240 ADEX, 1,061 squamous, 268 progenitor and 370 immunogenic genes. Moffitt subtyping was based on the 50 genes from the original publication (5). For each subtyping procedure, subtypes were manually assigned by overlaying the consensus clustering solution and expression patterns of the genes used for clustering.

Karasinska (7) metabolic subtypes were assigned using the same approach outlined in the original publication. Briefly, median expression (z-scores) of glycolytic and cholesterogenic genes were calculated, and samples with median glycolytic expression (MGE)>0 and median cholesterogenic expression (MCE)<0 were classified as glycolytic. Samples with MGE<0 and MCE>0 were cholesterogenic, while those with MGE>0 and MCE>0 were mixed and those with MGE<0 and MCE<0 were quiescent.

In addition to clustering-based Moffitt subtypes, we also called Moffitt subtypes using the n-of-one logistic regression tool provided in Supplementary Table S3 of the original publication (5), hereafter referred to as the “Moffitt 2015 tool”. Rashid and colleagues recently published an updated version of the same tool, PurIST (13), which was also run on each sample.

Normal and activated stromal subtypes were called using a similar consensus clustering methodology as performed in the original publication (5). Normal (n = 20) and activated (n = 22) stromal gene expression values (batch-corrected and z-score transformed) were used to cluster samples into two groups.

Survival analysis

Kaplan–Meier plots were generated using R v3.5.1 packages survival v.2.4.2 (21) and GGally v1.4.0 (https://ggobi.github.io/ggally/). Samples with an overall survival of less than 1 month were omitted from survival analyses. Multivariate survival analysis was performed using Cox proportional hazards regression models in R v3.5.1, with p values based on the Wald statistic.

Sequencing data availability

Genomic data generated within the POG/PanGen and COMPASS studies are actively submitted to the European Genome-phenome Archive (EGA) under accession numbers #EGAS00001001159 and #EGAS00001002543, respectively.

Clinical trial information

Personalized OncoGenomics (POG) Program of British Columbia: Utilization of Genomic Analysis to Better Understand Tumor Heterogeneity and Evolution (NCT02155621); Prospectively Defining Metastatic Pancreatic Ductal Adenocarcinoma Subtypes by Comprehensive Genomic Analysis (PanGen; NCT02869802); Comprehensive Molecular Characterization of Advanced Pancreatic Ductal Adenocarcinoma for Better Treatment Selection (COMPASS; NCT02750657).

RNA-seq data were collected for advanced (n = 269) and resectable (n = 305) PDAC tumor samples. 370 samples underwent laser capture microdissection (LCM) prior to sequencing (195 advanced, 175 resectable) while LCM was not performed for the remaining 204 samples (74 metastatic, 130 resectable). Six RNA-seq-based PDAC subtyping procedures were performed for each sample (4–7, 13). Collisson (4) and Bailey (6) subtypes were assigned using an unsupervised clustering approach, while Karasinska (metabolic) subtypes were assigned using relative expression of glycolytic/cholesterogenic pathway genes (7). While Moffitt basal-like and classical subtypes were called using an unsupervised clustering approach (5, 10), we also applied two published logistic regression tools, hereafter referred to as the Moffitt 2015 tool (5) and PurIST (13). As the regression-based Moffitt subtyping tools are aimed toward clinical translation, we used output from these two tools to determine subtype concordance for each sample, while leveraging the four other subtyping calls (clustering-based Moffitt, Collisson, Bailey and Karasinska) as additional information.

Ninety-seven of 574 (17%) of all samples received basal-like subtype calls from both Moffitt regression-based tools (hereafter referred to as “concordant basal-like”; Fig. 1A). Seventy-eight of 97 (80%) of these samples were called basal-like (Moffitt clustering-based), quasi-mesenchymal (Collisson), and squamous (Bailey) by other methods, while 54 of 97 (56%) were called glycolytic (Karasinska). 409 of 574 (71%) of samples received classical subtype calls from both Moffitt regression-based tools (hereafter referred to as “concordant classical”). 384 of 409 (94%) were Moffitt clustering-based classical; 301 of 409 (74%) were Collisson classical, and 241 of 409 (59%) were Bailey (pancreatic) progenitor. 115 of 125 (92%) of all cholesterogenic samples belonged to the concordant classical group. 68/574 (12%) of samples had discordant calls between the two regression-based Moffitt callers (hereafter referred to as “subtype-discordant”).

Figure 1.

A subgroup of PDAC tumors show discordant subtype calls across multiple methods. A, Six PDAC molecular subtyping procedures were performed for each of 574 samples. Concordance between regression-based Moffitt subtyping (Moffitt PurIST and 2015 tool) methods stratified samples into concordant basal-like, discordant, and concordant classical samples. B, Subtype-discordant samples show intermediate overall survival patterns in each study cohort. Log-rank P values are shown. C, Multivariate survival analysis showing intermediate survival outcomes for patients with subtype-discordant tumors across advanced PDAC study cohorts.

Figure 1.

A subgroup of PDAC tumors show discordant subtype calls across multiple methods. A, Six PDAC molecular subtyping procedures were performed for each of 574 samples. Concordance between regression-based Moffitt subtyping (Moffitt PurIST and 2015 tool) methods stratified samples into concordant basal-like, discordant, and concordant classical samples. B, Subtype-discordant samples show intermediate overall survival patterns in each study cohort. Log-rank P values are shown. C, Multivariate survival analysis showing intermediate survival outcomes for patients with subtype-discordant tumors across advanced PDAC study cohorts.

Close modal

Discordant subtyping calls were significantly more frequent among advanced PDAC (44/269; 16%) compared with resectable PDAC (24/305; 8%) samples (P = 1.1e-5), and advanced PDAC cases showed a higher proportion of concordant basal-like calls compared with resectable cases (22% vs. 12%, respectively; P = 0.004). Tumor biopsy sites differed between resectable cases, which were all derived from pancreatic lesions, and advanced cases, of which 71% were derived from metastatic liver lesions. To investigate whether differences in subtype frequency between resectable and advanced cohorts may be due to contaminating adjacent tissue, we tested whether utilization of laser microdissection (LCM) was associated with subtype calls, as LCM represents a means by which tumor cells are enriched in a sample prior to sequencing, and was performed in COMPASS but not PanGen/POG advanced PDAC cohorts. LCM was not associated with subtype groups when compared across all advanced PDAC samples (P = 0.063), nor when compared across advanced PDAC liver biopsies (P = 0.10). To further explore the potential relationship between subtype calls and nontumor cells, we computed median basal-like and classical gene expression scores using 226 liver and 328 pancreas samples obtained from the Genotype-Tissue Expression (GTex) database (22). Median basal-like gene expression was higher in pancreatic tissue compared with liver tissue (P = 4.1e-10, Supplementary Fig. S2), indicating that the heightened frequency of basal-like calls among advanced samples is not due to increased basal-like gene expression coming from contaminating liver tissue. Furthermore, concordant basal-like/classical and discordant groups were not significantly associated with tumor content in any of the four study cohorts (P > 0.05, Supplementary Fig. S3). Taken together, these results are compatible with the notion that differences in the frequency of subtype groups between resectable and advanced PDAC samples are not due to biopsy location.

Previous studies have reported signatures of stromal heterogeneity in PDAC, which resulted in the emergence of RNA-seq-based normal and activated stromal subtypes (5). To assess the impact of stromal heterogeneity on concordant classical/basal-like and discordant subtype groups, we classified each sample into normal and activated stromal subtypes. Concordant classical/basal-like and discordant subtype groups were not associated with stromal subtypes (P = 0.53), and stromal subtypes were not associated with overall survival (OS) within each subtype grouping (Supplementary Fig. S4).

Clinically, concordant basal-like and classical patient tumors consistently showed significant differences in overall survival in univariate survival analyses in both advanced (POG/PanGen: P = 1.7e-6, median OS basal-like = 8.8 months, classical = 17.6 months; COMPASS: P = 0.0072, median OS basal-like = 7.0 months, classical = 10.1 months) and resectable (TCGA: P = 3.4e-6, median OS basal-like = 8.22 months, classical = 21.9 months; ICGC: P = 0.016, median OS basal-like = 11.9 months, classical = 20.8 months) study cohorts (Fig. 1B). In each of the four cohorts, patients with subtype-discordant tumors showed intermediate median survival (POG/PanGen 13.1 months, COMPASS 8.6 months, TCGA 15.1 months, ICGC 16.4 months). OS patterns of subtype-discordant patients were not significantly different (P < 0.05) when individually compared with either concordant classical or basal-like patients, indicating that this subgroup of patients fail to uniquely align with either end of the binary basal-like/classical subtype spectrum. In the advanced PDAC cohorts, for which extended clinical data were available, patients with subtype-discordant tumors continued to show intermediate and non-significantly different survival patterns when patient age [<55 years (23)], sex. and first-line treatment were accounted for in multivariate analysis (POG/PanGen discordant vs. classical: P = 0.25, HR = 1.7 (95% CI: 0.7–3.9); COMPASS discordant vs. classical: P = 0.39, HR = 1.2 (95% CI: 0.8–2.1); Fig. 1C), while basal-like versus classical tumor groups remained significantly different [POG/PanGen basal vs. classical: P = 0.0001, HR = 4.9 (95% CI: 2.2–10.9); COMPASS basal vs. classical: P = 0.01, HR = 1.7 (95% CI: 1.1–2.7)]. These data indicate that patients with subtype-discordant tumors have intermediate survival patterns relative to the binary basal-like/classical spectrum.

We hypothesized that subtype-discordant PDAC tumors have hybrid basal-like and classical gene expression profiles. Overlaying each of the three (concordant basal-like/classical, discordant) sample groups on the unsupervised clustering solution for Moffitt clustering-based subtyping revealed that discordant samples tend to belong to hybrid (high basal-like and high classical gene expression) clusters (Fig. 2A). To quantify this observation, we calculated the median expression of clustering-based basal-like (n = 25) and classical (n = 25) genes for each sample (Fig. 2B). Subtype-discordant samples showed intermediate expression of basal-like (median normalized expression = 0.91) genes compared with concordant classical (median 0.69, P = 2.4e-6) and basal-like (median 1.2, P = 4.1e-5) samples, while the same was true for classical gene expression (median 1.1) versus concordant classical (median 1.5, P < 2.2e-16) and basal-like (median 0.71, P = 1.5e-10) samples. Intermediate expression patterns for subtype-discrepant samples were also observed for both glycolytic and cholesterogenic genes (Supplementary Fig. S5), which have been shown to represent metabolism-based correlates of PDAC subtypes (7). Consistent with the notion that subtype-discordant samples have intermediate expression phenotypes, discordant samples tended to fall within close proximity of the inflection point of the PurIST score distribution (Supplementary Fig. S6A and S6B).

Figure 2.

Subtype-discordant tumors show intermediate molecular profiles. A, Heatmap showing consensus clustering solution for the clustering-based Moffitt subtyping procedure. Individual tracks for sample group membership are shown across the top of the heatmap. Subtype-discordant samples tend to be located among the hybrid group cluster (far right). B, Median normalized classical and basal-like gene expression levels for each sample. Subtype-discordant samples show intermediate expression of both groups of genes. Wilcoxon mean rank-sum P values are shown. C, Proportion of samples with KRAS copy number alterations in each group. Fisher exact test P values are shown. D, Distribution of KRAS allelic imbalance across sample groups. Discordant samples show intermediate KRAS allelic imbalance levels. Wilcoxon mean rank-sum P values are shown.

Figure 2.

Subtype-discordant tumors show intermediate molecular profiles. A, Heatmap showing consensus clustering solution for the clustering-based Moffitt subtyping procedure. Individual tracks for sample group membership are shown across the top of the heatmap. Subtype-discordant samples tend to be located among the hybrid group cluster (far right). B, Median normalized classical and basal-like gene expression levels for each sample. Subtype-discordant samples show intermediate expression of both groups of genes. Wilcoxon mean rank-sum P values are shown. C, Proportion of samples with KRAS copy number alterations in each group. Fisher exact test P values are shown. D, Distribution of KRAS allelic imbalance across sample groups. Discordant samples show intermediate KRAS allelic imbalance levels. Wilcoxon mean rank-sum P values are shown.

Close modal

Chan-Seng-Yue and colleagues identified a genetic characteristic of Moffitt subtypes, in which basal-like PDAC tumors have greater frequency of copy number amplification of the mutant KRAS allele relative to the wild-type allele [termed allelic imbalance (AI); ref. 11]. Allele-specific copy number data was available for POG/PanGen, COMPASS, and ICGC study cohorts. The proportion of discordant samples with KRAS copy number alteration (17/65; 26%) was intermediate when compared between concordant classical (68/387, 18%; P = 0.06) and concordant basal-like (53/95; 56%; P = 9.3e-4) sample groups (Fig. 2C). We calculated KRAS AI as the difference in copy number between the major and minor KRAS alleles. Discordant samples showed intermediate levels of KRAS AI (median = 2) compared with concordant classical (median = 1; P = 9.0e-4) and basal-like (median = 3; P = 9.0e-4) groups (Fig. 2D). Taken together, these results are compatible with the notion that discordant-subtype PDAC samples show intermediate somatic mutation profiles.

Large-scale sequencing-based studies have converged toward basal-like (or squamous; ref. 9) and classical PDAC subtypes for several years (4–6), and recent publication of a NanoString-tested classifier has provided impetus towards adopting PDAC subtyping into clinical practice (13). In addition to their prognostic ability, preliminary evidence has suggested that PDAC subtypes are associated with differential sensitivity to first line treatment in nonresectable PDAC (10–12). As current clinical decision-making in PDAC, with the exception of BRCA1/2/PALB2 germline mutation patients, is largely restricted to staging, histopathology and patient performance status, molecular tumor subtyping holds promise for improving the PDAC treatment paradigm.

PDAC samples with concomitant high expression of both basal-like and classical genes can be recurrently observed on clustering diagrams in previous studies, but often go unacknowledged and merged with either of the two subtype groups (4–6). Chan-Seng-Yue and colleagues recently highlighted the group of hybrid basal-like/classical PDAC tumors, and used single-cell sequencing to demonstrate that hybrid tumors contain both basal-like and classical cell populations (11). Using a large and integrative cohort of both advanced and resectable samples, we demonstrate that 12% of patients with PDAC harbor subtype-discordant tumors that fail to receive consistent subtype calls across the two single-sample Moffitt subtype classifiers. These patient samples often appear on the fringe of being called either basal-like or classical, when comparing both clustering and regression-based methodologies, and show clinical and molecular profiles that are intermediate between basal-like and classical samples. When considering the classical/basal-like phenotype as a continuous spectrum, the degree of basal-likeness of a tumor sufficient for chemoresistance is unclear. Given their intermediate phenotype, our data indicate that patients with subtype-discordant/hybrid tumors should be considered as a separate entity, rather than being grouped with either of the strongly basal-like or classical profile patients.

The intermediate phenotype of hybrid samples is particularly important in the context of PDAC subtyping with methods that take into account few variables. Regression-based Moffitt subtyping tools are based on a top-scoring pair (TSP) approach, in which 14 (5) or 8 (13) pairs of basal-like and classical genes have their expression values directly compared and a binary value {0,1} representing whether the basal-like gene was greater than (1) or less than (0) its classical comparator is then used toward calculating the basal-like score for a sample. If basal-like and classical genes belonging to the same gene pair are expressed at similar levels (such as in hybrid samples), small differences in expression of one gene can easily sway the binary value for the comparison. For this reason, grouping hybrid samples together with samples that are strongly cast toward either of the two ends of the basal-like and classical continuum should be done with caution. To move forward with clinically actionable PDAC subtyping in a way that takes into consideration intermediate subtype samples, we propose that investigators utilize both the numeric basal-like score together with an orthogonal biomarker. GATA6 encodes a zinc finger transcription factor with roles in cellular differentiation and has been shown to be a robust biomarker for basal-like PDAC tumors (10). A combinatory approach involving calculation of the basal-like score (such as PurIST) as well as measurement of GATA6 levels will allow more accurate delineation between strongly classical/basal-like tumors and intermediate tumors that would fail to be robustly assigned to either group, and investigation of such data will be enabled by the upcoming Pancreatic Adenocarcinoma Signature Stratification for Treatment (PASS-01) trial (NCT04469556).

Our results are compatible with the notion that a subset of PDAC tumors harbor a mixture of basal-like and classical cell populations (11). Subtype-discordant tumors showed intermediate phenotypes both in terms of their transcriptomic and genomic profiles, and subtype-discordant samples appeared to concomitantly gain basal-like gene expression while losing classical gene expression. Intermediate median expression levels of basal-like and classical genes among a group of PDAC samples could arise due to two underlying high- and low-expressing groups of samples, which would oppose the claim of hybridity. However, the distributions of median classical and basal-like gene expression levels indicated that subtype-discordant tumors indeed harbor intermediate levels of both groups of genes, rather than having bi-modal distribution patterns (Supplementary Fig. S7). The fact that subtype-discordant samples are intermediate on the classical to basal-like continuum, together with previous evidence of subtype switching in pancreatic cancer cell lines (24), begs the question as to whether such tumors are progressively becoming more basal-like. Prospective, sequencing-based PDAC trials such as PanGen and COMPASS are positioned to assess the plasticity of subtype profiles as progression biopsies become available, though this endeavor remains challenging due to the aggressive nature of the disease and the invasiveness of the biopsy procedure. Circulating tumor DNA (ctDNA) sequencing technology represents a unique opportunity to obtain serial biopsies from patients with minimal invasiveness, and future ctDNA studies in PDAC will provide insight into how tumors may progressively evolve along the classical to basal-like subtype continuum.

The presence of both basal-like and classical cell populations in a subset of PDAC tumors introduces sampling error as a major limitation to consider when assigning PDAC subtypes, as heterogeneous tumor subpopulations may not be equally represented in a sample (25). This is of particular concern for advanced PDAC tumor biopsies, in which fine-needle biopsy cores attempt to sample from a small metastatic lesion. Future studies employing multiple sampling strategies and single-cell sequencing technology will provide insight into the extent by which spatial heterogeneity confounds molecular subtyping in PDAC and help determine the benefit of adopting molecular PDAC subtyping into clinical practice.

As the clinical management of PDAC moves toward biomarker-based personalized treatment, the consideration of an intermediate basal-like/classical tumor profile will have treatment stratification implications. Prospective clinical studies such as COMPASS and PanGen aim to identify treatment sensitivity biomarkers in advanced PDAC through whole genome and transcriptome sequencing analysis of pretreatment biopsies and monitoring of patient response to standard chemotherapy in a nonrandomized setting. The PASS-01 trial represents a randomized trial designed to investigate molecular chemosensitivity markers, including the putative FOLFIRINOX sensitivity and classical subtype biomarker GATA6, in metastatic PDAC. With tools like whole genome and transcriptome profiling and regression-based basal-like/classical tumor classifiers, these trials will provide excellent opportunities to further explore the clinical significance and predictive value of the subtype-discordant PDAC profile.

In conclusion, we highlight subtype-discordant PDAC tumors as a limitation when considering the adoption of PDAC molecular subtyping into clinical practice. Patients with subtype-discordant PDAC tumors do not conform to the binary basal-like and classical subtypes; rather, they emphasize the importance of treating the classical/basal-like phenotype as a continuum. Importantly, the intermediate nature of subtype-discordant samples underscores their merit to be considered a separate entity from both clinical and research perspectives. As prognostic PDAC subtyping, such as the PurIST method, moves closer to clinical implementation, we propose that one interprets the numeric score as a continuum, and leverages additional biomarkers, such as GATA6, when possible.

M.K.C. Lee reports other from Conquer Cancer Foundation and other from Conquer Cancer Foundation outside the submitted work. R.A. Moore reports grants from TFRI and grants from BC Cancer Foundation during the conduct of the study. J.M. Loree reports grants and personal fees from Ipsen and personal fees from Amgen, Novartis, Bayer, Pfizer, and Eisai outside the submitted work. P.A. Tang reports personal fees from Genomic Health, Amgen, Merck, Taiho, AstraZeneca, Pfizer, Teva, Eisai, and Pfizer outside the submitted work. R. Goodwin reports grants from Ipsen, Novartis, Pfizer; other from Ipsen, Eisai, Merck, Pfizer; and other from Roche during the conduct of the study, and was an investigator on multiple active clinical trials. J.J. Knox reports grants from Ontario Institute of Cancer Research; grants from Merck, Ipsen, and AstraZeneca during the conduct of the study; and personal fees from Roche, Eisai, and Pfizer outside the submitted work. J. Laskin reports other from BC Cancer Foundation and grants from Roche Canada during the conduct of the study and personal fees from Roche Canada, Pfizer, AstraZeneca, and Eli Lilly outside the submitted work. S.J.M. Jones reports grants from BC Cancer Foundation, Terry Fox Research Institute, Canada Foundation for Innovation, Canada Research Chairs, BC Knowledge Development Fund, and Pancreatic Cancer Canada during the conduct of the study. D.J. Renouf reports personal fees from Roche, Bayer, Celgene, Servier, Ipsen, Taiho, and AstraZeneca outside the submitted work. D.F. Schaeffer reports personal fees from Alimentiv, Diaceutics, Pfizer, and Amgen, and other from Satisfai Inc. outside the submitted work. No disclosures were reported by the other authors.

J.T. Topham: Conceptualization, data curation, formal analysis, investigation, visualization, methodology, writing-original draft. J.M. Karasinska: Conceptualization, resources, data curation, formal analysis, supervision, funding acquisition, investigation, visualization, writing-original draft, project administration, writing-review and editing. M.K.C. Lee: Data curation, writing-review and editing. V. Csizmok: Conceptualization, investigation, methodology, writing-review and editing. L.M. Williamson: Conceptualization, investigation, methodology, writing-original draft, writing-review and editing. G.H. Jang: Resources, data curation, formal analysis, investigation, methodology, writing-review and editing. R.E. Denroche: Resources, data curation, investigation, methodology, writing-review and editing. E.S. Tsang: Investigation, writing-review and editing. S.E. Kalloger: Investigation, writing-review and editing. H.-L. Wong: Resources, funding acquisition. G.M. O'Kane: Resources, data curation. R.A. Moore: Resources, data curation, software. A.J. Mungall: Resources, data curation. F. Notta: Resources, data curation. J.M. Loree: Supervision, investigation, project administration, writing-review and editing. J.M. Wilson: Resources, data curation, investigation. O. Bathe: Investigation, writing-review and editing. P.A. Tang: Investigation, writing-review and editing. R. Goodwin: Investigation, writing-review and editing. J.J. Knox: Resources, data curation, supervision, funding acquisition, investigation, writing-review and editing. S. Gallinger: Resources, data curation, supervision, funding acquisition, investigation, writing-review and editing. J. Laskin: Resources, data curation, supervision, funding acquisition, investigation, writing-review and editing. M.A. Marra: Resources, data curation, supervision, funding acquisition, investigation, writing-review and editing. S.J.M. Jones: Resources, data curation, supervision, funding acquisition, investigation, methodology, writing-review and editing. D.J. Renouf: Conceptualization, resources, data curation, formal analysis, supervision, funding acquisition, investigation, visualization, writing-original draft, project administration, writing-review and editing. D.F. Schaeffer: Conceptualization, resources, data curation, formal analysis, supervision, funding acquisition, investigation, visualization, writing-original draft, project administration, writing-review and editing.

This research was supported through philanthropic donations received through the BC Cancer Foundation, as well as funding provided by the Terry Fox Research Institute, Ontario Institute for Cancer Research (PanCuRx Translational Research Initiative), Pancreatic Cancer Canada, Genome British Columbia (project B20POG), and VGH/UBC Hospital Foundation. D.J. Renouf and J.M. Loree are recipients of the MSFHR Health Professional-Investigator Award, and D.F. Schaeffer is a recipient of the VCHRI Investigator Award. The authors gratefully acknowledge the participation of patients and their families, and the POG/PanGen and COMPASS teams. The results published here are in part based upon data generated by The Cancer Genome Atlas managed by the NCI and NHGRI (http://cancergenome.nih.gov), as well as data generated by the International Cancer Genome Consortium (https://icgc.org/). This research was supported through philanthropic donations received through the BC Cancer Foundation (project B20POG), as well as funding provided by the Terry Fox Research Institute (project 1078), Ontario Institute for Cancer Research (PanCuRx Translational Research Initiative), Pancreatic Cancer Canada, and Genome British Columbia.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.

1.
Niederhuber
JE
,
Brennan
MF
,
Menck
HR
. 
The National Cancer Data Base report on pancreatic cancer
.
Cancer
1995
;
76
:
1671
7
.
2.
Kamisawa
T
,
Wood
LD
,
Itoi
T
,
Takaori
K
. 
Pancreatic cancer
.
Lancet
2016
;
388
:
73
85
.
3.
Saad
AM
,
Turk
T
,
Al-Husseini
MJ
,
Abdel-Rahman
O
. 
Trends in pancreatic adenocarcinoma incidence and mortality in the United States in the last four decades; a SEER-based study
.
BMC Cancer
2018
;
18
:
688
.
4.
Collisson
EA
,
Sadanandam
A
,
Olson
P
,
Gibb
WJ
,
Truitt
M
,
Gu
S
, et al
Subtypes of pancreatic ductal adenocarcinoma and their differing responses to therapy
.
Nat Med
2011
;
17
:
500
3
.
5.
Moffitt
RA
,
Marayati
R
,
Flate
EL
,
Volmar
KE
,
Loeza
SGH
,
Hoadley
KA
, et al
Virtual microdissection identifies distinct tumor- and stroma-specific subtypes of pancreatic ductal adenocarcinoma
.
Nat Genet
2015
;
47
:
1168
78
.
6.
Bailey
P
,
Chang
DK
,
Nones
K
,
Johns
AL
,
Patch
A-M
,
Gingras
M-C
, et al
Genomic analyses identify molecular subtypes of pancreatic cancer
.
Nature
2016
;
531
:
47
52
.
7.
Karasinska
JM
,
Topham
JT
,
Kalloger
SE
,
Jang
GHo
,
Denroche
RE
,
Culibrk
L
, et al
Altered gene expression along the glycolysis-cholesterol synthesis axis is associated with outcome in pancreatic cancer
.
Clin Cancer Res
2020
;
26
:
135
46
.
8.
Raphael
BJ
,
Hruban
RH
,
Aguirre
AJ
,
Moffitt
RA
,
Yeh
JJ
,
Stewart
C
, et al
Integrated genomic characterization of pancreatic ductal adenocarcinoma
.
Cancer Cell
2017
;
32
:
185
203
.
9.
Collisson
EA
,
Bailey
P
,
Chang
DK
,
Biankin
AV
. 
Molecular subtypes of pancreatic cancer
.
Nat Rev Gastroenterol Hepatol
2019
;
16
:
207
20
.
10.
Aung
KL
,
Fischer
SE
,
Denroche
RE
,
Jang
G-Ho
,
Dodd
A
,
Creighton
S
, et al
Genomics-driven precision medicine for advanced pancreatic cancer: early results from the COMPASS Trial
.
Clin Cancer Res
2018
;
24
:
1344
54
.
11.
Chan-Seng-Yue
M
,
Kim
JC
,
Wilson
GW
,
Ng
K
,
Figueroa
EF
,
O'Kane
GM
, et al
Transcription phenotypes of pancreatic cancer are driven by genomic events during tumor evolution
.
Nat Genet
2020
;
52
:
231
40
.
12.
O'Kane
GM
,
Grünwald
BT
,
Jang
G-Ho
,
Masoomian
M
,
Picardo
S
,
Grant
RC
, et al
GATA6 expression distinguishes classical and basal-like subtypes in advanced pancreatic cancer
.
Clin Cancer Res
2020
;
26
:
4901
10
.
13.
Rashid
NU
,
Peng
XL
,
Jin
C
,
Moffitt
RA
,
Volmar
KE
,
Belt
BA
, et al
Purity independent subtyping of tumors (PurIST), a clinically robust, single-sample classifier for tumor subtyping in pancreatic cancer
.
Clin Cancer Res
2020
;
26
:
82
92
.
14.
Law
HC-H
,
Lagundžin
D
,
Clement
EJ
,
Qiao
F
,
Wagner
ZS
,
Krieger
KL
, et al
The proteomic landscape of pancreatic ductal adenocarcinoma liver metastases identifies molecular subtypes and associations with clinical response
.
Clin Cancer Res
2020
;
26
:
1065
76
.
15.
Owen
DR
,
Wong
H-Li
,
Bonakdar
M
,
Jones
M
,
Hughes
CS
,
Morin
GB
, et al
Molecular characterization of ERBB2-amplified colorectal cancer identifies potential mechanisms of resistance to targeted therapies: a report of two instructive cases
.
Cold Spring Harb Mol Case Stud
2018
;
4
:
a002535
.
16.
Chen
X
,
Schulz-Trieglaff
O
,
Shaw
R
,
Barnes
B
,
Schlesinger
F
,
Källberg
M
, et al
Manta: rapid detection of structural variants and indels for germline and cancer sequencing applications
.
Bioinformatics
2016
;
32
:
1220
2
.
17.
Kim
S
,
Scheffler
K
,
Halpern
AL
,
Bekritsky
MA
,
Noh
E
,
Källberg
M
, et al
Strelka2: fast and accurate calling of germline and somatic variants
.
Nat Methods
2018
;
15
:
591
4
.
18.
Cingolani
P
,
Platts
A
,
Wang
LeL
,
Coon
M
,
Nguyen
T
,
Wang
L
, et al
A program for annotating and predicting the effects of single nucleotide polymorphisms, SnpEff: SNPs in the genome of Drosophila melanogaster strain w1118; iso-2; iso-3
.
Fly
2012
;
6
:
80
92
.
19.
Johnson
WE
,
Li
C
,
Rabinovic
A
. 
Adjusting batch effects in microarray expression data using empirical Bayes methods
.
Biostatistics
2007
;
8
:
118
27
.
20.
Carter
SL
,
Cibulskis
K
,
Helman
E
,
McKenna
A
,
Shen
H
,
Zack
T
, et al
Absolute quantification of somatic DNA alterations in human cancer
.
Nat Biotechnol
2012
;
30
:
413
21
.
21.
Therneau
TM
. 
A package for survival analysis in R
. 
2015
.
Available from
: https://cran.r-project.org/web/packages/survival/index.html.
22.
Lonsdale
J
,
Thomas
J
,
Salvatore
M
,
Phillips
R
,
Lo
E
,
Shad
S
, et al
The Genotype-Tissue Expression (GTEx) project
.
Nat Genet
2013
;
45
:
580
5
.
23.
Ben-Aharon
I
,
Elkabets
M
,
Pelossof
R
,
Yu
KH
,
Iacubuzio-Donahue
CA
,
Leach
SD
, et al
Genomic landscape of pancreatic adenocarcinoma in younger vs. older patients: does age matter?
Clin Cancer Res
2019
;
25
:
2185
93
.
24.
Adams
CR
,
Htwe
HH
,
Marsh
T
,
Wang
AL
,
Montoya
ML
,
Subbaraj
L
, et al
Transcriptional control of subtype switching ensures adaptation and growth of pancreatic cancer
.
eLife
2019
;
8
:
e45313
.
25.
Goldman
SL
,
MacKay
M
,
Afshinnekoo
E
,
Melnick
AM
,
Wu
S
,
Mason
CE
. 
The impact of heterogeneity on single-cell sequencing
.
Front Genet
2019
;
10
:
8
.