Article Text
PDF
Abstract
Introduction Transcriptional analyses have identified several distinct molecular subtypes in pancreatic ductal adenocarcinoma (PDAC) that have prognostic and potential therapeutic significance. However, to date, an indepth, clinicomorphological correlation of these molecular subtypes has not been performed. We sought to identify specific morphological patterns to compare with known molecular subtypes, interrogate their biological significance, and furthermore reappraise the current grading system in PDAC.
Design We first assessed 86 primary, chemotherapy-naive PDAC resection specimens with matched RNA-Seq data for specific, reproducible morphological patterns. Differential expression was applied to the gene expression data using the morphological features. We next compared the differentially expressed gene signatures with previously published molecular subtypes. Overall survival (OS) was correlated with the morphological and molecular subtypes.
Results We identified four morphological patterns that segregated into two components (‘gland forming’ and ‘non-gland forming’) based on the presence/absence of well-formed glands. A morphological cut-off (≥40% ‘non-gland forming’) was established using RNA-Seq data, which identified two groups (A and B) with gene signatures that correlated with known molecular subtypes. There was a significant difference in OS between the groups. The morphological groups remained significantly prognostic within cancers that were moderately differentiated and classified as ‘classical’ using RNA-Seq.
Conclusion Our study has demonstrated that PDACs can be morphologically classified into distinct and biologically relevant categories which predict known molecular subtypes. These results provide the basis for an improved taxonomy of PDAC, which may lend itself to future treatment strategies and the development of deep learning models.
- pancreatic ductal adenocarcinoma
- basal subtype
- classical subtype
- pattern-based morphological classification system
- deep learning
Statistics from Altmetric.com
- pancreatic ductal adenocarcinoma
- basal subtype
- classical subtype
- pattern-based morphological classification system
- deep learning
Significance of this study
What is already known on this subject?
Pancreatic ductal adenocarcinoma (PDAC) has a dismal prognosis, and therapeutic interventions have only been marginally beneficial.
Transcriptional analyses have converged on two prognostically significant molecular subtypes of PDAC. These are termed basal and classical, where the basal subtype confers a poorer survival.
Histological evaluation of PDAC is based on grading into three categories of differentiation, which can be subjective, chiefly in moderately differentiated category.
What are the new findings?
Our pattern-based morphological classification reflects underlying tumour biology and correlates with known molecular subtypes.
These morphological subgroups better predict clinical outcome than transcriptional subtypes.
Morphological pattern-based groups correlate better with clinical outcomes than the conventional differentiation-based classification.
How might it impact on clinical practice in the foreseeable future?
The study demonstrates the power of morphology in predicting survival, and thus providing support for potential targeted therapeutic measures, given the known association with transcriptional subtypes.
Our pattern-based morphological classification can be leveraged to design diagnostic algorithms for deep learning neural networks, which will facilitate pathologists to accurately quantify these morphological categories and segregate patients into the relevant prognostic groups.
Introduction
Pancreatic ductal adenocarcinoma (PDAC) has a dismal prognosis and has been predicted to be the second leading cause of cancer death by 2030.1 Thus far, therapeutic interventions have only been marginally beneficial with minimal overall survival (OS) advantages. Histological evaluation and staging remain one of the gold standards in the determination of clinical characteristics of tumours. Presently, the histological grading for PDACs comprises a three-tiered, differentiation-based system, which can be subjective, chiefly in tumours that are categorised as moderately differentiated, and therefore is of limited clinical utility in management decisions. As such, there is a need for an improved taxonomy of PDAC to provide better predictions of clinical outcome and identify potential therapeutic targets.
Recently, with the increasing understanding of the complex molecular mechanisms of cancers, genomic and gene expression data have been incorporated into the framework of clinical practice for several malignancies.2–5 In pancreatic cancer, recent studies have begun to refine how we stratify this disease based on the correlation between pathological and genomic data,6 7 and more recently immunological criteria.8
Furthermore, over the last decade, three large-scale studies have consistently identified specific molecular subtypes in resected PDAC. First, Collisson et al 9 defined three subtypes (classical, quasimesenchymal and exocrine-like) from PDAC laser-microdissected (LCM) samples and cell lines and showed OS differences for both classical and quasimesenchymal subtypes in multivariate analysis. Subsequently, Moffitt et al 10 defined two tumour subgroups (basal and classical) and stromal subgroups (normal and activated) by computationally separating out the tumour and stroma and found that the ‘basal’ and ‘activated stromal’ subtypes conferred a worse OS. Most recently, Bailey et al 11 defined four subtypes (squamous, pancreatic progenitor, immunogenic and aberrantly differentiated endocrine (ADEX), with the ‘squamous’ subtype having the worst outcome). While there is some variation in nomenclature, there is significant overlap in transcriptional programmes between Collisson classical and Moffitt classical and Bailey pancreatic progenitor subtypes, Collisson quasimesenchymal, Moffitt basal and Bailey squamous subtypes, and Collisson exocrine-like and Bailey ADEX subtypes.12 A recent study has shown that the ADEX, exocrine-like and immunogenic subtypes are more likely secondary to contamination from stroma and non-neoplastic cells as a result of low tumour cellularity.13
Overall, these studies have demonstrated that there are currently two bona fide tumour transcriptional subtypes of PDAC, chiefly the classical and basal subtypes. The current study evaluates the merit of classifying PDAC by distinct morphological patterns, rather than by differentiation, and highlights the relationship of these patterns with molecular subtypes and clinical outcomes.
Methods
Patients and sample selection
In the University Health Network, Toronto, we initially identified 88 patients with primary, PDAC resections and paired RNA-Seq data (2009–2017 inclusive). During this period, the standard of care in our institution was upfront resection, followed by adjuvant therapy. Patients who received neoadjuvant treatment for unresectable tumours were excluded from our study. The ‘special’ subtypes of PDAC, including colloid carcinoma, medullary carcinoma, undifferentiated carcinoma and undifferentiated carcinoma with osteoclast-like giant cells of the pancreas, were excluded from our study as they constitute a small subset of PDACs (1%–3%) and are associated with their own outcomes.14 As a result, two patients with a diagnosis of colloid carcinoma arising from an intraductal papillary mucinous neoplasm (IPMN) were excluded from our cohort. The clinical and pathological data from each patient were reviewed.
Histopathological review and quantification
Haematoxylin and eosin (H&E) stained slides from all patients were first reviewed by a pathologist (SNK) blinded to RNA-Seq data. All slides containing tumour per patient sample were semiquantitatively assessed for the presence of specific, reproducible morphological patterns. The area of tumour on each slide was arbitrarily divided into four quadrants, each quadrant representing 25%. Subsequently, the percentage of each morphological pattern was recorded in multiples of 25%, depending on the number of quadrants each morphological pattern covered. If a morphological pattern was only focally present (ie, <25%), an approximation of 5%–10% was recorded. The percentage of each morphological pattern per tumour containing slide was recorded and averaged (online supplementary table 1). In addition, for each patient we calculated the morphological classification for every permutation of 1 to the total number of slides and found the minimal number of slides required to achieve 100% concordance with the morphological classification of all the slides.
Supplementary file 1
RNA sequencing analysis
RNA extraction, library generation and sequencing have been previously described elsewhere.15 RNA-Seq gene expression was quantified using Kallisto V.0.44.0,16 with default options and transcript sequences downloaded from Ensembl build 86. Gene-level expression for length-scaled transcripts per million (TPM)and counts was calculated using tximport.17 Differential expression (DE)between the morphological groups was performed using DESeq2,18 with an adjusted p-value cut-off of 0.001 and a log2 fold change of at least 1. The normalised count data were transformed using DESeq2 variance stabilising transform, and the top 50 differentially expressed genes from each group were used for further analyses. For Moffitt’s classification of RNA-Seq samples, the length-scaled TPMs from tximport were used with the published Moffitt classifier.10
Interobserver analysis between study pathologists
To assess interobserver variability, all patient samples were also independently assessed and scored by two other pathologists (RV and RC) who were blinded to each other and the RNA-Seq data. A prestudy discussion was carried out to outline the scoring parameters. Pairwise agreement between pathologists was performed using Cohen’s kappa, and agreement between all three pathologists was performed using Fleiss’ kappa.
Clinical characteristics
Clinical data were extracted from the Ontario Pancreas Cancer Study database.19 Pathological staging was performed in accordance with the 8th edition of tumour, node, metastases classification of pancreatic exocrine tumours.20 Associations with clinical characteristics were assessed using Wilcoxon rank-sum tests for continuous variables and Fisher’s exact test for categorical variables.
Survival analysis
For survival analysis, time started at the surgical date, ended at death, and was censored at last follow-up. The primary characteristic of interest was the presence or absence of at least 40% non-gland-forming pattern, which was selected based on the RNA-Seq analysis (see results below). Other secondary characteristics included for comparison were grade (well differentiated, moderately differentiated and poorly differentiated) and the Moffitt classification from RNA-Seq data (basal-like and classical). Kaplan-Meier survival curves and Cox proportional hazard regressions were estimated.21 Adjusted HRs came from multivariate regressions, including age at diagnosis, sex, stage, use of adjuvant chemotherapy, presence or absence of lymphovascular invasion, and margin status (R0 vs R1). Age was entered as a linear covariate. Stage was grouped pragmatically into stage I–IIa, IIb and III–IV based on the number of patients in each group. Statistical significance was set at p<0.05. Analysis was performed in R (V.3.1.1).
Results
New classification based on distinct morphological groups of PDAC
The demographic and clinical characteristics of patients with primary tumours in our cohort are summarised in table 1. From the histopathological review, we defined four specific morphological patterns that we termed as (1) conventional, (2) tubulopapillary, (3) squamous and (4) ‘composite’ (figure 1A).
(A) Schematic representation of four morphological patterns identified in pancreatic ductal adenocarcinoma, which can be broadly divided into two components: ‘gland forming’ (conventional and tubulopapillary) and ‘non-gland forming’ (composite and squamous). The composite pattern encompasses a spectrum of morphological patterns traditionally associated with poor differentiation, including (1) cribriforming, (2) sheets, (3) single files, (4) cords, (5) buds, (6) single cells, (7) ribbons, (8) angulated glands and (9) nests. (B) Bar chart showing the percentage of different morphological patterns present per tumour, demonstrating evidence of intratumoural heterogeneity within tumours.
Patient demographics
The ‘conventional’ pattern was characterised by the presence of well-differentiated glands with a tubular, stellate configuration, lined by pancreaticobiliary-type epithelium (figure 2A). The cells can have well to moderate cytological atypia with moderate amounts of eosinophilic cytoplasm (figure 2A, inset). Clear or foamy cell change in the cytoplasm may also be seen focally.
(A) ‘Conventional’ pattern characterised by well-formed glands lined by pancreaticobiliary-type epithelium with a tubular/stellate configuration. The cells have moderate nuclear atypia with moderate amounts of eosinophilic cytoplasm (inset). (B) ‘Tubulopapillary’ pattern with larger, dilated, well-formed glands in comparison with conventional pattern, which can be seen as part of the invasive component of IPMN or independent of an IPMN. The glands are lined by a mixture of pancreaticobiliary and foveolar-gastric-type epithelium with papillary/micropapillary projections. The cells are more uniform with small regular nuclei and abundant amphophilic to eosinophilic cytoplasm (inset). (C) Well-differentiated ‘squamous’ pattern associated with the adenosquamous variant of PDAC. (D) Poorly differentiated ‘squamous’ pattern associated with the adenosquamous variant of PDAC. (E) ‘Composite’ pattern comprises a spectrum of morphological features, including (i) sheets, (ii) nests, (iii) ribbons, (iv) cords, (v) angulated glands, (vi) single files, (vii) buds, (viii) single cells and (ix) cribriforming (see also figure 1A). IPMN, intraductal papillary mucinous neoplasm; PDAC, pancreatic ductal adenocarcinoma.
The ‘tubulo-papillary’ pattern resembles the morphological appearance in the group of tumours previously described as the ‘papillary/papillary-cystic variant’ of PDAC,7 which can be seen as the invasive component of an IPMN or occur in the absence of an associated IPMN. The latter was also observed in our study, as 20 of 86 (23%) demonstrated this pattern in the absence of an IPMN component, and 5 of 86 (6%) with a coexistent IPMN. This pattern was characterised by glands that were generally larger than glands within the ‘conventional’ group and showed a wide gradation of sizes (mean size of ~0.5 mm). The glands were rounded and dilated in configuration, lined by a combination of foveolar-gastric type and pancreaticobiliary-type epithelium with tubular and papillary/micropapillary projections (figure 2B). The cells are more uniform with low-grade dysplasia, small nuclei and abundant amphophilic to eosinophilic cytoplasm (figure 2B, inset).
The presence of a ‘squamous’ component in PDACs has long been recognised as part of the adenosquamous variant of PDAC (at least 30% squamous differentiation) or less frequently as a pure squamous carcinoma. The squamous component can vary from well to poorly differentiated. This can range from nests of large polygonal cells with opaque eosinophilic cytoplasm which are clearly separated by intercellular bridges and the presence of large keratin pearls (figure 2C), to sheets of smaller, hyperchromatic cells that retain intercellular bridges with absent or minimal keratinisation (figure 2D).
The last pattern, dubbed ‘composite’, encompasses a mélange of morphological features traditionally associated with poor differentiation. These features are seen when glands begin to lose their integrity and cohesion, forming a spectrum of patterns including sheets, nests/islands, ribbons, cords, angulated glands, single file (figure 2E (i–vi)), or dispersing as buds and single cells (figure 2E (vii–viii)). Another frequently observed pattern is ‘cribriforming’, which is characterised by a large group of tumour cells with small, punched-out lumens, resembling the appearance of Swiss cheese. This is a result of gland fusion with no true lumen formation, but rather punched-out spaces due to necrosis (figure 2E (ix)).
We next sought to identify if any of these morphological patterns were associated with poor prognosis. However, morphological heterogeneity, where tumours exhibited two or more patterns, was observed in in 59 of 86 (69%) tumours (figure 1B) (discussed in detail below). To compare prognosis between different patterns, samples were grouped by dominant morphological pattern, with the exception of adenosquamous group where the established >30% cut-off of squamous pattern was used14 (online supplementary figure 1A). The composite (p=0.03) and squamous patterns had a worse OS, while the tubulopapillary and composite patterns had a good and intermediate OS, respectively (online supplementary figure 1B). We observed that 25 of 44 (57%) conventional and 4 of 15 (27%) tubulopapillary patterns were also mixed with the composite pattern. Given that tumours with a high proportion of composite pattern had poor OS, we reasoned that these tumours would have worse OS compared with the tumours with pure conventional and tubulopapillary patterns. There was an OS difference between the pure and mixed patterns, with a statistically significant difference between the conventional and conventional + composite patterns (p=0.01) (online supplementary figure 1C), suggesting tumours with the composite pattern were associated with a poorer prognosis.
Supplementary file 2
Given that both the conventional and tubulopapillary patterns were characterised by the presence of well-formed glandular structures and have similar OS, henceforth both these patterns were collectively grouped as ‘gland -forming’. In contrast, given the poorer prognosis and lack of well-formed glands, the ‘squamous’ and ‘composite’ patterns were grouped as ‘non-gland forming’ (figure 1A).
Morphology predicts molecular subtypes
We next used RNA-Seq data to assess if the morphological subtypes are related to known molecular subtypes. We initially separated the samples into two groups: <50% non-gland-forming component; ≥50% non-gland-forming component. DE identified 508 differentially expressed genes between the two groups (adjusted p value <0.001, log2 fold change ≥1.0). Of these genes, 216 were upregulated in <50% non-gland-forming group and 292 in ≥50% non-gland-forming component (online supplementary tables 2–3). The <50% non-gland-forming group was enriched for genes present in Moffitt and Collisson ‘classical’ and Bailey pancreatic progenitor subtypes, while ≥50% non-gland-forming group was enriched for genes in Moffitt basal-like, Bailey squamous and Collisson quasimesenchymal (online supplementary figure 2). For example, the transcription factor GATA6 has been associated with well-differentiated PDAC,22 and TP63 has been associated with the poorly differentiated and squamous subtype.11 We used the top 50 differentially expressed genes from each group for hierarchical clustering and observed two clusters within the dendrogram (figure 3). As a validation step, we next applied the published Moffitt gene classifier to compare our subtypes, as the Moffitt gene sets were based on virtual microdissection of different non-negative matrix factorisation-derived gene signatures, which is independent of morphology, in comparison with the dendrogram classification, which was derived from genes that were differentially expressed between the morphological groups.10 We found a high concordance between the Moffitt classifier and the two major branches of the dendrogram from the differentially expressed genes (figure 3). Therefore, we used the Moffitt classification, as it has been adopted as a standard classification scheme.
Supplementary file 5
Heatmap of differentially expressed genes between ‘gland forming’ and ‘non-gland forming’. The variance stabilised and scaled expression for top 50 differentially expressed genes from the ‘gland forming’ (blue) and ‘non-gland forming’ (red) samples are shown, and the vertical colour bar indicates which group the genes are from. The dendrograms were generated using correlation distance and the complete linkage function. The two colour bars indicate the tumour differentiation (top) and the Moffitt classifier (bottom). The colours from each bar are indicated in the legend. Finally, the bar graph shows the proportion of non-gland-forming morphological component for each sample. The bars are coloured blue (group A), if their non-gland-forming proportion is less than 40% and red otherwise. The dashed line indicates the 50% cut-off used for differential expression, while the solid line indicates the revised 40% cut-off.
We next sought to investigate if 50% is an ideal morphological cut-off to predict the basal and classical molecular subtypes. For the samples classified as basal, all but one had a ≥40% non-gland-forming component; therefore, we defined two morphological groups, A and B, as having <40% and ≥40% non-gland-forming components, respectively (figure 4A). A subset of samples classified as classical (18/66, 27%) were morphologically classified as group B, and we sought to determine if they were defined by differing gene expression programmes. Differential expression between these classical group A and group B samples showed there were only 9 (3 overlapping original DE) group A and 15 (4 overlapping original DE) group B significant genes (adjusted p value <0.001, log2 fold change ≥1). There was no clear separation of the two groups, suggesting that they have similar gene expression programmes (online supplementary figure 3). However, a sampling bias during LCM could provide an explanation for this difference (see below). Overall, based on these considerations, we defined two morphological groups, A and B, as having <40% and ≥40% non-gland-forming components, respectively.
Supplementary file 6
(A) Boxplot demonstrating the proportion of ‘non-gland forming’ component present within both the classical and basal subtypes by the Moffitt classifier. The table below demonstrates the overlap of classical and basal tumours within groups A and B of our morphological classification. (B) Kaplan-Meier survival analysis shows that patients with group A tumours showed a trend towards better survival than patients with group B tumours (adjusted HR 1.64 (95% CI 0.97 to 2.76), p=0.06). (C) Kaplan-Meier survival analysis shows there was no trend or significant difference between the basal and classical tumours (adjusted HR 1.23 (95% CI 0.65 to 2.35), p=0.52) (D) Diagrammatic flow chart shows the number of classical tumours (n=66) that were classified group A (A-c; n=48) and group B (B-c=18). (E) Kaplan-Meier survival analysis shows a worse survival for patients classified as group B (B-c) tumours within the classical subtype (adjusted HR 2.20 (95% CI 1.14 to 4.25), p=0.02).
Morphological heterogeneity and reproducibility
The presence of morphological heterogeneity has been well documented in PDAC.7 23 In our cohort, intratumoural heterogeneity was noted in 59 of 86 (69%) tumours (figure 1B), and the percentage of each morphological pattern per slide from each patient is documented in online supplementary table 4. The number of slides reviewed per specimen ranged from 4 to 16. However, despite the high variability in slides per case, in 76% (65/86) of cases, a minimum of one to two slides were required to establish the diagnosis (figure 5A,B). A subset of the misclassified cases above (10/18) were closest to the 40% non-gland-forming cut-off and required the highest number of minimum slides to arrive at the diagnosis (figure 5B). As mentioned in the previous section, it is possible that more of the gland-forming components were sampled for the RNA-Seq analysis.
Supplementary file 7
Interslide heterogeneity for samples classified as group A (A) and group B (B). The top plot for each group is the total number of slides that are classified as group A or group B based on the 40% cut-off. The middle plot for each group is the minimum number of slides required to achieve a 100% true positive classification. The bars are coloured by their molecular classification. The bottom plot for each group is the non-gland-forming per cent for each slide coloured by the samples’ molecular subtype. The short black horizontal lines indicate the minimum and maximum values, and the black dot indicates the mean value which was used for the morphological classification.
Furthermore, there was good concordance between the three pathologists with an overall Fleiss’ kappa score of 0.87. There was an overall disagreement in eight cases, all of which were close to the 40% non-gland-forming cut-off. The interobserver concordance between each pathologist with Cohen’s kappa scores is represented in figure 6A–C, and the scores for individual samples by each pathologist are shown in online supplementary table 5.
Supplementary file 8
Scatter plots show interobserver correlation between each individual pathologist with an overall Fleiss’ kappa score of 0.87. The x and y axes represent percentage of the non-gland-forming component. (A) Interobserver concordance between pathologist 1 (P1) and pathologist 2 (P2) (Cohen’s kappa score=0.86). (B) Interobserver concordance between P1 and pathologist 3 (P3) (Cohen’s kappa score=0.91). (C) Interobserver concordance between P2 and P3 (Cohen’s kappa score=0.86).
Distinct morphological groups are associated with OS
We next assessed whether group A (n=49) and group B (n=37) were associated with clinical characteristics. Morphological groups had no association with age (p=0.44), sex (p=1.00), lymphovascular invasion (p=0.79), margin status (p=0.77), stage at presentation (p=0.36) or the receipt of adjuvant chemotherapy (p=0.22). However, morphological pattern was significantly associated with OS (unadjusted HR 1.87 (95% CI 1.14 to 3.05), p=0.01) (figure 4B). This was no longer significant after adjustment for potential confounders; however, the direction of change was similar (adjusted HR 1.64 (95% CI 0.97 to 2.76), p=0.06). Patients from group A had a median OS from resection of 2.1 years (95% CI 1.8 to 2.5 years), compared with 0.9 years (95% CI 0.8 to 1.7 years) for patients in group B (figure 4B).
The morphological groups showed a better survival stratification compared with molecular subtypes (figure 4C), as there was no significant OS difference between the basal and classical subtypes (unadjusted HR 1.49 (95% CI 0.85 to 2.63), p=0.17; adjusted HR 1.23 (95% CI 0.65 to 2.35), p=0.52) (figure 4C). We next assessed whether group A and B tumours that classified as classical, herein referred to as A-classical and B-classical (please also see RNA sequencing analysis section), had a difference in OS (figure 4D). We found that there was a significant difference in OS between the A-classical and B-classical tumours (adjusted HR 2.20 (95% CI 1.14 to 4.25), p=0.02) (figure 4E). Furthermore, the B-classical tumours also showed a worse OS than the tumours that classified as basal (online supplementary figure 4).
Supplementary file 9
Morphological classification shows a better stratification of OS compared with the standard three-tiered grading system.
The traditional standard three-tiered grading system showed a clear difference in OS between well and poorly differentiated tumours (figure 7A). However, most patients in our cohort had tumours graded as ‘moderately differentiated’ (55/86, 64%) (figure 7B). Among patients with moderately differentiated tumours, those that were categorised as group B (B-moderate) (figure 7C) by our classification were associated with a worse survival (unadjusted HR 1.82 (95% CI 0.98 to 3.38), p=0.06; adjusted HR 1.77 (95% CI 0.87 to 3.58), p=0.11) (figure 7D).
![[gutjnl-2019-318217supp002.jpg]](https://gut.bmj.com/content/gutjnl/69/2/317/DC2/embed/inline-supplementary-material-2.jpg?download=true){kind=link}
![[gutjnl-2019-318217supp007.jpg]](https://gut.bmj.com/content/gutjnl/69/2/317/DC7/embed/inline-supplementary-material-7.jpg?download=true){kind=link}
(A) Kaplan-Meier survival analysis shows the survival differences between patients with well, moderate and poorly differentiated tumours, classified using the three-tiered, differentiation-based classification system. (B) Boxplot demonstrating the proportion of ‘non-gland forming’ component present within the well, moderate and poorly differentiated tumours. The table below demonstrates the overlap of the well, moderate and poorly differentiated tumours within group A and B tumours of our morphological classification system. (C) Diagrammatic flow chart shows the number of moderately tumours (n=64) that were group A (A-moderate; n=33) and group B (B-moderate; n=22). (D) Kaplan-Meier survival analysis shows worse survival for patients classified as group B tumours within the group of patients with moderately differentiated tumours. (unadjusted HR 1.82 (95% CI 0.98 to 3.38), p=0.06 and adjusted HR 1.77 (95% CI 0.87 to 3.58), p=0.11).
Discussion
Our study has identified specific morphological patterns in PDAC and a novel morphological classification system that segregates into two prognostic groups. This was established based on a semiquantitative assessment of all available histological slides from each patient sample, from which four specific morphological patterns (conventional, tubulopapillary, composite and squamous) were identified. The ‘composite’ morphological category was established based on the collective grouping of features, which have traditionally been associated with poor differentiation. Broadly these patterns could be segregated into two components, ‘gland forming’ and ‘non-gland forming’, predicated on the presence or absence of well-formed glands. Subsequently, the presence of >40% of the ‘non-gland forming’ component was associated with a worse prognosis, which we termed group B. The latter morphological group was associated with the basal molecular subtype of PDAC. In contrast, cancers with <40% of the ‘non-gland forming’ component, which were termed group A, were associated with a better prognosis and the classical molecular subtype.9–11
The use of molecular subtyping is of clinical importance as it has been shown to predict clinical outcome.11 15 However, gene expression profiling of patient samples is costly, time-consuming and currently limited to large cancer centres. In addition, molecular classifiers are often confounded by several limitations such as tumour cellularity, LCM bias, RNA and tissue quality, and tumour heterogeneity, making translation to patient care challenging. In our study, a quarter of tumours in our cohort were classified as ‘classical’ molecularly but morphologically classified as group B tumours (>40% of the non-gland-forming component) and were strongly associated with worse OS, suggesting that discrepancies between morphology and RNA-Seq may be clinically relevant. We have further demonstrated that a sampling bias may provide a likely explanation for this discrepancy, as these tumours showed the greatest degree of morphological heterogeneity, which were very close to the morphological cut-off and required the most number of slides to arrive at the score. As these cases also have a prominent gland-forming component, it is likely that it was the glandular component that was captured on microdissection.
We propose that our morphological classification system has advantages over the conventional three-tiered grading, differentiation-based system. The major disadvantage with the three-tiered grading system was that most PDACs were invariably grouped as ‘moderate’, as by definition the presence of well-formed glands with areas of poor differentiation would qualify as ‘moderate’. It is well recognised that PDACs are histologically heterogeneous within a given patient; therefore, the existing grading system does not necessarily give an accurate representation of the entire tumour. Our system quantifies heterogeneity by estimating the percentages of different distinct morphological patterns. We also found that our system reflects underlying tumour biology, with a strong correlation with transcriptional subtypes. Furthermore, among tumours graded as ‘moderately differentiated’ (see also figure 7A), stratification based on the percentage of the non-gland-forming component (<40% or ≥ 40%) was associated with clear survival differences (see also figure 7D).
Our morphological classification also provides the ability to expediently and inexpensively predict the biological behaviour of each individual tumour. Such a precedent already exists in prostatic cancer with the Gleason grading system. Gleason grading was first described by Donald Gleason in 1966,24 where tumours were stratified using histological patterns that incorporate the assessment of primary and secondary grades as additive scores, which at the time was a new approach to the histological assessment of cancer. The Gleason grading system has been further refined to include five prognostic groups25 26 and still remains an integral component in prostate cancer reporting. We hope that our grading system will fulfil a similar role in PDAC, which will also enable the potential future therapeutic targets, given the strong correlation with known transcriptional subtypes. For example, in our institution, an ongoing clinical trial, COMPASS(Comprehensive Molecular Characterization of Advanced Pancreatic Ductal Adenocarcinoma for Better Treatment Selection), of locally advanced and metastatic pancreatic cancer showed that basal tumours had better outcomes with gemcitabine and nab-paclitaxel, whereas classical tumours had better outcomes with modified FOLFIRINOX (American Society of Clinical Oncology (ASCO) GI 2018, Abstract 188).
In addition, since our grading system is a pattern-based model, it removes some subjectivity from the differentiation model, making it appropriate for application to deep learning convolutional neural network platforms (DLCNN). Recently, the advancement in the field of deep learning has facilitated the development of cost-effective, powerful technologies that are capable of performing complex visual pattern recognition tasks. Such deep learning algorithms have already been successful in using whole-slide images in the detection of invasive breast cancer,27 colonic cancer28 and grading gliomas in the brain.29 Furthermore, DLCNN can be used to predict the mutational status of non-small cell carcinoma using DLCNN.30 In the future, DLCNN can be leveraged to design algorithms to assist pathologists to accurately quantify the specific morphological categories that we have described in PDAC and clinically segregate patients into the basal and classical molecular subtypes. In addition, these more sophisticated quantification methods can be further used to calculate each individual pattern that we have collectively grouped as the ‘composite’ morphological category.
Presently, our study has shown a clear difference in gene expression profiles between the ‘gland forming’ and ‘non-gland forming groups’. Beyond this, we hope to apply recently developed spatial RNA-Seq techniques, such as STARmap31 and MERFISH,32 to overlay the gene expression profiles on a single cell level on intact tissue, which permits gene expression correlations with morphological details. This will give us greater insight into the gene expression profiles on an individual gland level and study the subtle differences between different patterns.
We concede that a significant limitation of our study is the small sample size with wide CIs and the possibility of unmeasured confounders influencing our results. Therefore, before we incorporate our grading system into the clinical framework, our next endeavour would be to validate our grading system on an expanded cohort, which may also better refine the morphological per cent cut-off, if necessary. Furthermore, we have demonstrated a very good concordance between the pathologists in our study; however, a larger interobserver study to include more pathologists from different centres would be the basis for future work. We are also aware that there are an increasing number of centres that are providing upfront neoadjuvant treatment to all patients, and this may affect the morphology and RNA of tumours. However, our study is a proof-of-principle study and it would still be of merit to see if our morphomolecular correlation holds true, which will also be the basis for future work.
In conclusion, in this study we have morphologically classified PDAC into specific, morphologically distinct and biologically relevant categories, which predict known molecular subtypes of PDAC. While the prediction of molecular subtypes by morphology can be limited by technical artefacts, we do demonstrate the power of morphology in predicting survival, thus providing evidentiary support that future treatment strategies can be based on morphology. Finally, this study provides the basis for an improved taxonomy of PDAC and may lend itself to the development of deep learning models in the future.
Supplementary file 3
Supplementary file 4
Acknowledgments
The authors would like to acknowledge Julie Wilson, OICR, for feedback on the manuscript.
References
Footnotes
GWW and RCG contributed equally.
Contributors Drafting and revising manuscript: SNK, RC, GWW, RG, SG. Development of pathological model: SNK, RC. RNA-Seq analysis: GWW. Survival analysis: RG, GO, SG, MS. Pathological assessment: SNK, RC, RV. Conception of the project: SNK, RC, GWW, RG, FN, SG.
Funding The authors have not declared a specific grant for this research from any funding agency in the public, commercial or not-for-profit sectors.
Competing interests None declared.
Ethics approval Ethical approval was obtained from the institutional review board, and written informed consent was obtained from all patients.
Provenance and peer review Not commissioned; externally peer reviewed.
Patient consent for publication Obtained.