Purpose:

Cancer-associated fibroblasts have emerged to be highly heterogenous and can play multifaceted roles in dictating pancreatic ductal adenocarcinoma (PDAC) progression, immunosuppression, and therapeutic response, highlighting the need for a deeper understanding of stromal heterogeneity between patients and even within a single tumor. We hypothesized that image analysis of fibroblast subpopulations and collagen in PDAC tissues might guide stroma-based patient stratification to predict clinical outcomes and tumor characteristics.

Experimental Design:

A novel multiplex IHC-based image analysis system was established to digitally differentiate fibroblast subpopulations. Using whole-tissue slides from 215 treatment-naïve PDACs, we performed concurrent quantification of principal fibroblast subpopulations and collagen and defined three stroma types: collagen-rich stroma, fibroblast activation protein α (FAP)-dominant fibroblast-rich stroma, and α smooth muscle actin (ACTA2)-dominant fibroblast-rich stroma. These stroma types were assessed for the associations with cancer-specific survival by multivariable Cox regression analyses and with clinicopathologic factors, including CD8+ cell density.

Results:

FAP-dominant fibroblasts and ACTA2-dominant fibroblasts represented the principal distinct fibroblast subpopulations in tumor stroma. Stroma types were associated with patient survival, SMAD4 status, and transcriptome signatures. Compared with FAP-dominant fibroblast-rich stroma, collagen-rich stroma correlated with prolonged survival [HR, 0.57; 95% confidence interval (CI), 0.33–0.99], while ACTA2-dominant fibroblast-rich stroma exhibited poorer prognosis (HR, 1.65; 95% CI, 1.06–2.58). FAP-dominant fibroblast-rich stroma was additionally characterized by restricted CD8+ cell infiltrates and intense neutrophil infiltration.

Conclusions:

This study identified three distinct stroma types differentially associated with survival, immunity, and molecular features, thereby underscoring the importance of stromal heterogeneity in subtyping pancreatic cancers and supporting the development of antistromal therapies.

This article is featured in Highlights of This Issue, p. 1

Translational Relevance

Pancreatic cancer is characterized by abundant desmoplastic stroma. Despite numerous theoretical and experimental efforts, therapeutic approaches targeting pancreatic cancer stroma have been largely unsuccessful, highlighting the need for more comprehensive assessment of inter- and intratumoral stromal heterogeneity in a large series of clinical tumors. Here, we developed a pixel-by-pixel image analysis system to digitally differentiate and quantify two mutually exclusive fibroblast subpopulations, α smooth muscle actin–dominant fibroblasts and fibroblast activation protein α–dominant fibroblasts, within tumor tissues. Quantitative computation of these principal fibroblast subpopulations, intratumoral collagen, and CD8+ T cells using whole-tissue sections from 215 treatment-naïve pancreatic cancers allowed us to identify three distinct stroma types differentially associated with patient outcomes, molecular characteristics, and the immunosuppressive tumor microenvironment. Our human tissue–based quantitative analysis likely provides new insights into the clinical importance of stroma-based pancreatic cancer subtyping in facilitating the development of multidisciplinary treatment strategies against this lethal malignancy.

Pancreatic ductal adenocarcinoma (PDAC) is characterized by abundant desmoplastic stroma made up of cancer-associated fibroblasts (CAF), extracellular matrix (ECM), immune cells, and vasculature (1). CAFs can play pleiotropic roles in regulating tumor cell proliferation, invasion, immunosuppression, and drug resistance in in vitro and in vivo models of the pancreatic cancer microenvironment (2–7). Because of the postulated protumorigenic functions of CAFs in patients with PDAC, preclinical and clinical trials have been performed to target fibroblasts in combination with chemotherapy, radiotherapy, or immunotherapy (8–12). However, these nonselective antifibroblast therapies have been largely unsuccessful, underscoring the need for more targeted approaches and patient stratification based on stromal heterogeneity (6, 13, 14). Histologically, human PDAC has a much larger stroma volume with a wider range of fibroblast cellularity than genetically engineered mouse models of PDAC (15–17). Recent studies using single-cell RNA sequencing and other state-of-the-art techniques have indicated the functional and phenotypical diversity of CAFs, highlighting the growing need for a better understanding of fibroblast heterogeneity to facilitate the development of antistromal therapies against PDAC (18–22). Evidently, the desmoplastic composition of PDAC stroma is highly heterogeneous between patients and even within a single tumor. These findings pose a challenge to those attempting to characterize the pancreatic cancer microenvironment and to address the clinical significance of inter- and intratumoral stromal heterogeneity.

Recent technologies have improved multiplex IHC, making it a powerful tool to visualize, specify, and quantify different cellular components in their histologic context, even in formalin-fixed, paraffin-embedded tissues (23–25). In our previous study, image analyses using whole-slide sections of pancreas and liver allowed robust quantitative assessment of collagen and immune cells, thereby overcoming nonnegligible levels of spatial heterogeneity (26, 27). Emerging evidence suggests the validity of several fibroblastic markers, including α smooth muscle actin (ACTA2) and fibroblast activation protein α (FAP), which may serve as surrogate markers that can specify fibroblast subpopulations (6, 8, 28). In this study, we found several fibroblast subpopulations by multiplex IHC in PDAC stroma and established a novel pixel-by-pixel analysis system that automatically differentiates and quantifies the two principal fibroblast subpopulations, ACTA2-dominant fibroblasts and FAP-dominant fibroblasts. In addition to fibroblasts, we measured intratumoral collagen (which can comprise more than 90% of ECM components in pancreatic cancer stroma; ref. 29) in 215 whole-tissue sections of PDAC. The generated data allowed us to test our primary hypothesis that image analysis of fibroblast subpopulations and collagen might guide stroma-based patient stratification to predict clinical outcomes and tumor characteristics. This study defined three distinct stroma types differentially associated with cancer-specific survival and molecular features, including SMAD4 status and Moffitt and colleagues' transcriptome subtypes (30).

In addition, some recent clinical trials have demonstrated that most PDACs are refractory to immunotherapies; the reason for this is likely the immunosuppressive milieu of cancer-associated stroma (31, 32). Experimental evidence has indicated that FAP-expressing stromal cells inhibit intratumoral accumulation of effector T cells in several cancer types, including PDAC (7, 33, 34). Therefore, using our clinical cohort, we tested a secondary hypothesis that, PDACs with abundant areas of FAP-dominant fibroblasts have limited infiltration of CD8+ lymphocytes in the tumor center. Our findings from multiplex image analysis of pancreatic cancer desmoplasia suggest the clinical significance of stroma-based patient stratification for the development of multidisciplinary treatment strategies against this deadly cancer.

Study group

From the hospital's archival pathology database, we selected 290 consecutive patients who underwent pancreatectomy for PDAC at Keio University Hospital (Shinjuku, Tokyo, Japan) between 1991 and 2016. Forty-seven patients who had undergone preoperative chemo/chemoradiotherapy were excluded. A further 17 patients were excluded who had variants of PDAC (e.g., adenosquamous carcinoma, colloid carcinoma, or undifferentiated carcinoma). We were not able to obtain image analysis data from tissue samples of 14 patients because of poor staining quality. Synchronous multicentric PDACs (six PDACs in 3 patients) that were previously characterized by integrated clinicopathologic and genetic analyses were treated as individual tumors for histologic study (35). Consequently, we analyzed 215 treatment-naïve PDACs from 212 patients for which quantitative information on desmoplastic stroma could be generated. The histopathologic assessment, including the histologic type, tumor differentiation (well, grade 1; moderate, grade 2; and poor, grade 3), residual tumor status, and the extent of tertiary lymphoid structures (absent, minimal, and extensive; ref. 26) was recorded by agreement of at least two pathologists; the results were reviewed by a specialist in pancreas pathology (Y. Masugi). The number of tumor-infiltrating neutrophils was counted in 10 randomly selected high-power fields under microscopy using hematoxylin and eosin (H&E)-stained slides (36–38). The percentage of tumor cells within cancerous areas was evaluated by the study pathologist (Y. Masugi), and we defined tumor cellularity as high (>30%) or low (≤30%). This study was performed in accordance with the Declaration of Helsinki and was approved by the Ethics Committees of Keio University School of Medicine (Shinjuku-ku, Tokyo, Japan, institutional review board no., 20040034). Written informed consent was obtained from all patients.

IHC analysis

Multiplex fluorescence IHC was performed using a Bond-MAX Automated Stainer (Leica Biosystems) on the basis of a protocol published previously (26). For quantitative image analysis of ACTA2-dominant fibroblasts and FAP-dominant fibroblasts, 3-μm-thick sections were peroxide blocked and incubated with anti-ACTA2 antibody (clone 1A4; Agilent Technologies, catalog no., N1584, RRID:AB_2335694) followed by treatment with ImmPRESS (Vector Laboratories, catalog no., MP-7402, RRID:AB_2336528), and then 10-minute incubation with Alexa Fluor 555 Tyramide Reagent (Thermo Fisher Scientific, catalog no., B40955). Sections were heated for antigen retrieval in ER2 Solution (Leica Biosystems, catalog no., AR9640) for 40 minutes, followed by anti-FAP antibody (clone SP325; Abcam, catalog no., ab227703) treatment, incubation with ImmPRESS (Vector Laboratories, catalog no., MP-7401, RRID:AB_2336529), and then 10-minute incubation with Alexa Flour 488 Tyramide Reagent (Thermo Fisher Scientific, catalog no., B40953). Slides then underwent 10 minutes of heating in the ER2 solution, incubation with the primary antibody for h-caldesmon (clone h-CD; Agilent Technologies, catalog no., M3557, RRID:AB_2065456), treatment with ImmPRESS, and visualization with Alexa Flour 647 Tyramide Reagent (Thermo Fisher Scientific, catalog no., B40958). Finally, slides were counterstained with Hoechst 33342 (Thermo Fisher Scientific, catalog no., H3570). To acquire whole-slide images of fluorescence IHC, we scanned the stained slides with a NanoZoomer-XR Slide Scanner (Hamamatsu Photonics). Slides were then stained with van Gieson solution (saturated picric acid containing 0.09% acid fuchsin) and rescanned to obtain multi-layered images (Fig. 1A).

Figure 1.

Multiplex image analysis identifies stromal heterogeneity in human pancreatic cancer tissues. A, Cellular pancreatic cancer stroma is rich in spindle-shaped stromal cells that differentially express known fibroblastic markers, including ACTA2 and FAP, while less cellular collagen-rich stroma shows little expression of these markers. Panels of KRT7 IHC and Alcian blue stain highlight the differences in tumor cell morphology among these stromal variations. Daggers indicate well-polarized adenocarcinoma glands. Tumor cells in less cellular collagen-rich stroma diffusely show stronger intensity of KRT7 staining than those in cellular stroma. Note that stromal blue stain in Alcian blue–stained images indicates hyaluronan deposition. B, Fluorescence IHC for ACTA2 and FAP and van Gieson stain using a single slide demonstrates ACTA2-dominant fibroblasts and FAP-dominant fibroblasts as the principal fibroblast subpopulations in PDAC stroma. The low magnification image (top, left) shows that well-differentiated cancer ducts (left) are directly surrounded by ACTA2-dominant fibroblasts, whereas juxta-tumoral fibroblasts in poorly differentiated tumor areas (right) show intense FAP expression, rather than ACTA2. High-magnification images of the three boxes [f, FAP-expressing area (right, top); a, ACTA2-expressing area (left, bottom); and m, mixed area (right, bottom)] are shown along with van Gieson–stained images.

Figure 1.

Multiplex image analysis identifies stromal heterogeneity in human pancreatic cancer tissues. A, Cellular pancreatic cancer stroma is rich in spindle-shaped stromal cells that differentially express known fibroblastic markers, including ACTA2 and FAP, while less cellular collagen-rich stroma shows little expression of these markers. Panels of KRT7 IHC and Alcian blue stain highlight the differences in tumor cell morphology among these stromal variations. Daggers indicate well-polarized adenocarcinoma glands. Tumor cells in less cellular collagen-rich stroma diffusely show stronger intensity of KRT7 staining than those in cellular stroma. Note that stromal blue stain in Alcian blue–stained images indicates hyaluronan deposition. B, Fluorescence IHC for ACTA2 and FAP and van Gieson stain using a single slide demonstrates ACTA2-dominant fibroblasts and FAP-dominant fibroblasts as the principal fibroblast subpopulations in PDAC stroma. The low magnification image (top, left) shows that well-differentiated cancer ducts (left) are directly surrounded by ACTA2-dominant fibroblasts, whereas juxta-tumoral fibroblasts in poorly differentiated tumor areas (right) show intense FAP expression, rather than ACTA2. High-magnification images of the three boxes [f, FAP-expressing area (right, top); a, ACTA2-expressing area (left, bottom); and m, mixed area (right, bottom)] are shown along with van Gieson–stained images.

Close modal

Chromogenic FAP IHC using the conventional diaminobenzidine-based protocol of the Bond-MAX stainer was conducted to validate the staining of the anti-FAP antibody. Chromogenic FAP immunostaining of PDAC tissues revealed positive staining mostly in spindle-shaped stromal cells, but not in carcinoma cells (Supplementary Fig. S1), as seen in fluorescence IHC. We further conducted FAP IHC using human tissues from multiple organs (Supplementary Fig. S2). Consistent with prior studies, normal tissues in a variety of organs generally showed no detectable level of FAP expression, although a very few FAP-expressing cells were found in our study, for example, spindle-shaped stromal cells in the subserosal space and immune cells in tonsil. Furthermore, we observed positive FAP immunostaining in fibroblast-like stromal cells within pancreatic specimens from patients with active fibrogenic disorders, including autoimmune pancreatitis.

Tumor TP53 or SMAD4 expression status was defined according to the staining patterns of chromogenic IHC, as described previously (26). Detailed information on the primary antibodies used in IHC analysis is given in Supplementary Table S1.

Image analysis of fibroblast subpopulations and collagen

As a training set, we selected 3,061 pixels (0.46 μm × 0.46 μm/pixel) from 16 PDAC tissue samples stained for ACTA2 (labeled by Alexa Fluor 555) and FAP (labeled by Alexa Fluor 488), and added a designation to each pixel according to its histologic characteristics (e.g., ACTA2-dominant fibroblasts, FAP-dominant fibroblasts, ACTA2/FAP overlapping areas, red blood cells, cancer cells, nontissue areas, and duct lumina; Fig. 1B). Pixel-by-pixel sorting according to the signal levels of Alexa Flour 555 and Alexa Flour 488 yielded different clusters of pixels containing ACTA2-dominant fibroblasts, FAP-dominant fibroblasts, or ACTA2/FAP overlapping areas (Fig. 2A). Using these results, we developed naïve Bayes classifiers to digitally distinguish ACTA2-dominant fibroblasts, FAP-dominant fibroblasts, and ACTA2/FAP overlapping areas in fluorescence images. Next, we analyzed multi-layered images of whole-tissue slides from 215 PDACs as the test set. Tumor border lines were drawn on H&E-stained images by the study pathologist (Y. Masugi); these border lines were pasted onto multi-layered images of the relevant serial sections. Naïve Bayes classifiers were applied to create color classification images from fluorescence images following color normalization (39). In defining the tumor area for evaluation, we excluded nonneoplastic epithelial cells, necrotic areas, duodenal walls, and out-of-focus regions, as described previously (26). When analyzing the tumor area, nontissue areas, including glass areas and acellular ductal lumina, were excluded by imaging techniques. In this study, h-caldesmon IHC was used to exclude areas of smooth muscle tissue (e.g., vessel walls). These areas, which generally coexpress ACTA2, were excluded to avoid overestimation of ACTA2-expressing fibroblasts. Consequently, we examined 1.6 × 104 mm2 (∼7.8 × 1010 pixels) of tumor area from 215 PDACs (on average, 75.4 mm2 per tumor). Quantification of intratumoral collagen using van Gieson–stained images was conducted by computational methods established previously (40). The occupancy rate of FAP-dominant fibroblasts in cancer tissues correlated inversely with that of ACTA2-dominant fibroblasts (Spearman correlation coefficient, r = −0.25; P < 0.001) and collagen (r = −0.30; P < 0.001); however, no significant correlation was observed between the occupancy rates of ACTA2-dominant fibroblasts and collagen (r = −0.11; P = 0.13; Supplementary Fig. S3). The proportions of desmoplastic elements (i.e., collagen, ACTA2-dominant fibroblasts, FAP-dominant fibroblasts, and ACTA2/FAP overlapping areas) were calculated by dividing the occupancy rate of the respective elements by the sum of the occupancy rates of these four elements. Image analysis was done using MATLAB 2019b Software (MathWorks) with custom computer codes. Custom codes with demo data are available at the Code Ocean website (https://codeocean.com/capsule/9991437/tree). The workflow of quantitative image analysis of fibroblast subpopulations and collagen is shown in Supplementary Fig. S4.

Figure 2.

Identification of three distinct stroma types in human pancreatic cancer based on the proportions of desmoplastic elements. A, Pixel-by-pixel sorting of 3,061 pixels in fluorescence images of 16 PDACs (as a training set) by signal levels of Alexa Fluor 488 and Alexa Fluor 555, showing distinct cluster formations between FAP-dominant fibroblast pixels and ACTA2-dominant fibroblast pixels. Naïve Bayes classifiers were then created to automatically differentiate each pixel. This image classifier was applied to multilayered whole-slide images of 215 PDACs to perform concurrent quantification of fibroblast subpopulations and intratumoral collagen. B, Bar charts showing landscapes of intertumoral stromal heterogeneity among 215 PDAC samples. C, Hierarchical clustering on the basis of quantitative data on desmoplastic elements. D, Statistical differences in the proportions of the respective desmoplastic elements among stroma types (*, P < 0.05; **, P < 0.001; by the Mann–Whitney U test). E, Representative classification images of multiplex IHC and van Gieson–stained whole-tumor sections for each stroma type.

Figure 2.

Identification of three distinct stroma types in human pancreatic cancer based on the proportions of desmoplastic elements. A, Pixel-by-pixel sorting of 3,061 pixels in fluorescence images of 16 PDACs (as a training set) by signal levels of Alexa Fluor 488 and Alexa Fluor 555, showing distinct cluster formations between FAP-dominant fibroblast pixels and ACTA2-dominant fibroblast pixels. Naïve Bayes classifiers were then created to automatically differentiate each pixel. This image classifier was applied to multilayered whole-slide images of 215 PDACs to perform concurrent quantification of fibroblast subpopulations and intratumoral collagen. B, Bar charts showing landscapes of intertumoral stromal heterogeneity among 215 PDAC samples. C, Hierarchical clustering on the basis of quantitative data on desmoplastic elements. D, Statistical differences in the proportions of the respective desmoplastic elements among stroma types (*, P < 0.05; **, P < 0.001; by the Mann–Whitney U test). E, Representative classification images of multiplex IHC and van Gieson–stained whole-tumor sections for each stroma type.

Close modal

Noncancerous areas were additionally selected in adjacent pancreatic parenchyma within 147 PDAC slides for quantitative image analyses. These noncancerous areas included pancreatic parenchyma with minimal pathologic change and those with chronic pancreatitis, and did not contain pancreatic intraepithelial neoplasia lesions. Representative images of ACTA2, FAP, and collagen in noncancerous areas are shown in Supplementary Fig. S5. Fibrotic tissues in noncancerous pancreas were mostly made up of collagen and fibroblasts that exclusively expressed ACTA2. The occupancy rates of collagen (mean, 19.9%; median, 13.4%; and interquartile range, 8.1%–27.0%) and ACTA2-dominant fibroblasts (mean, 11.4%; median, 7.0%; and interquartile range, 3.0%–15.0%) in noncancerous pancreas were predominantly higher than those of FAP-dominant fibroblasts (mean, 0.2%; median, 0.1%; and interquartile range, 0.04%–0.2%). The occupancy rates of collagen and ACTA2-dominant fibroblasts in noncancerous pancreas showed a significant positive correlation (r = 0.54; P < 0.001; Supplementary Fig. S3). The occupancy rates of FAP-dominant fibroblasts did not show any correlation with those of collagen (r = −0.07; P = 0.43) or ACTA2-dominant fibroblasts (r = 0.14; P = 0.09) in noncancerous pancreas.

Quantification of tumor-infiltrating CD8+ T cells

Utilizing a computational method established previously (26), we conducted fluorescence IHC analysis of CD8 using whole-tissue sections, and enumerated CD8+ cells in the tumor center and the tumor margin. Regions of interest (ROI; one ROI ∼1 mm2) were set using a tiling method to cover the tumor area, and ROIs intersected by the tumor border line were defined as “tumor margin” regions. In the following analyses, the CD8+ cell density in the tumor center was dichotomized using the median density. Custom software codes for the quantitative assessment of CD8+ cells are available at the Code Ocean website ().

Transcriptome analysis

Frozen specimens from 20 patients with PDAC were obtained from the biobank of Keio University School of Medicine (Shinjuku-ku, Tokyo, Japan). These frozen tissues were sampled from surgical specimens under the supervision of pathologists and were stored at −80°C. RNA was extracted using an AllPrep DNA/RNA/miRNA Universal Kit (Qiagen, catalog no., 80224) following the manufacturer's protocol, and the sample quality was confirmed (RNA integrity number, 7.3 ± 0.7) using an Agilent2100 Bioanalyzer (Agilent Technologies). Transcriptome data were generated using a GeneChip Scanner 3000 7G System (Affymetrix) and a Clariom D Pico Assay (Thermo Fisher Scientific, catalog no.,902925). Expression profiles were clustered using a centroid-linkage method with Cluster 3.0 software, and heatmaps were generated by TreeView 3.0.

Statistical analysis

Survival analysis was done using the log-rank test and Cox proportional hazards regression analysis. Deaths as a result of pancreatic cancer progression were the endpoints for cancer-specific survival, and deaths from other causes were censored. Five of the 212 patients with PDAC were excluded from survival analysis because they died before hospital discharge within 30 days of surgery. The final follow-up data were collected in November 2019 with a median follow-up time of 36.6 months (range, 1.8–178.7 months). The multivariable Cox regression model initially included age at diagnosis, sex, anatomic location, tumor differentiation, tumor cellularity, residual tumor status, TP53 expression, SMAD4 expression, the extent of tertiary lymphoid structures, CD8+ cell density in the tumor center, and disease stage. A backward stepwise elimination with a threshold of P = 0.10 was used to select variables in the final multivariable model. For 3 patients with synchronous multicentric PDACs, covariates of the main (larger) tumors were assigned to adjust for potential confounding. Statistical analyses were conducted using the Statistical Package for the Social Sciences version 24 (SPSS; IBM). All P values were two-sided, and P < 0.05 was defined as significant.

Data availability

The transcriptome data in this article have been deposited in NCBI's Gene Expression Omnibus database (accession no., GSE157353).

Use of standardized official symbols

We used HUGO-approved official symbols for genes and gene products, including ACTA2, CD74, CXCL12, CXCR4, FAP, PDCD1, PDPN, SMAD4, and TP53; all of which are described at www.genenames.org. Gene names are italicized, and gene product names are nonitalicized.

Identification of fibroblast subpopulations by multiplex IHC

Histology of human pancreatic cancer stroma demonstrated large variations in fibroblast cellularity and collagen deposition with differential expressions of several fibroblastic markers, including ACTA2 and FAP (Fig. 1A). Moreover, we observed considerable differences in tumor cell morphology along with stromal variations: tumor cells embedded in collagen-rich stroma or ACAT2-expressing cellular stroma showed well-polarized glands with frequent intracellular mucin at the apical side, whereas tumor cells embedded in FAP-expressing stroma formed solid nests or cribriform glands with indistinct mucin production indicating minimal ductal differentiation. Multiplex fluorescence IHC revealed that most spindle-shaped stromal cells in tumor tissues expressed either ACTA2 or FAP, with only a few overlaps (Fig. 1B). We defined stromal cells dominantly expressing ACTA2 as ACTA2-dominant fibroblasts and those dominantly expressing FAP as FAP-dominant fibroblasts. These two types of fibroblasts formed distinct subpopulations digitally distinguishable by pixel-by-pixel sorting for ACTA2 and FAP (Fig. 2A). Microscopically, well-differentiated cancer ducts are frequently located close to ACTA2-dominant fibroblasts with thickened collagen accumulation, whereas juxta-tumoral fibroblasts in poorly differentiated tumor areas often show strong FAP expression, but not intense ACTA2 expression (Fig. 1A and B; Supplementary Fig. S6). Multiplex IHC additionally visualized PDPN-expressing stromal cells, which coexpressed ACTA2 and/or FAP, as relatively minor fibroblastic subpopulations (Supplementary Fig. S7). Moreover, although extremely limited in number, CD74-expressing stromal cells, which coexpressed FAP, were observed in tumor tissues (Supplementary Fig. S8). Multiplex IHC analyses of representative PDAC slides indicated that no significant levels of ACTA2 or FAP expression were observed in CD31-expressing cells, in CD45-expressing cells, or in KRT7-expressing epithelial cells within tumor tissues (Supplementary Fig. S9).

Three distinct stroma types based on quantitative data on fibroblast subpopulations and collagen in human PDAC tissues

To quantify ACTA2-dominant fibroblasts and FAP-dominant fibroblasts, naïve Bayes classifiers, which could digitally differentiate these fibroblast subpopulations, were created using a training set of 16 PDACs, and were then applied to the whole-tissue slide images of 215 PDACs (Fig. 2A). There was considerable intertumoral heterogeneity in the proportions of desmoplastic elements between tumors (Fig. 2B); collagen (mean, 53.4%; median, 54.1%; and interquartile range, 45.6%–62.2%) was the predominant desmoplastic element, followed by ACTA2-dominant fibroblasts (mean, 33.4%; median, 30.6%; and interquartile range, 24.2%–41.1%) and FAP-dominant fibroblasts (mean, 13.0%; median, 12.1%; and interquartile range, 8.2%–17.4%). There were very few ACTA2/FAP overlapping areas (mean, 0.2%; median, 0.1%; and interquartile range, 0.05%–0.2%). On the basis of four continuous variables, that is, the proportions of collagen, ACTA2-dominant fibroblasts, and FAP-dominant fibroblasts and the sum of occupancy rates of the four desmoplastic elements (i.e., collagen, ACTA2-dominant fibroblasts, FAP-dominant fibroblasts, and ACTA2/FAP overlapping areas), we conducted unsupervised hierarchical clustering using Ward method, which generated three distinct groups (Fig. 2C). As a result, we defined three stroma types: collagen-rich stroma (“C-stroma;” n = 55), FAP-dominant fibroblast-rich stroma (“F-stroma;” n = 95), and ACTA2-dominant fibroblast-rich stroma (“A-stroma;” n = 65), which were named after their notable desmoplastic elements. Statistical differences in desmoplastic stromal components among these three groups are demonstrated in Fig. 2D. The proportion of collagen was highest in C-stroma, next highest in F-stroma, and lowest in A-stroma. There were significant differences in the mean proportions of FAP-dominant fibroblasts, with the mean proportion in F-stroma (18.8%) being more than twice than those in C-stroma and A-stroma (8.9% and 7.9%, respectively). ACTA2-dominant fibroblasts were by far the most common element in A-stroma. The sum of occupancy rates of desmoplastic elements were significantly lower in F-stroma than in the other two stroma types. Representative whole-slide classification images for each stroma type are shown in Fig. 2E, visualizing evident differences in desmoplastic compositions among stroma types, although there was a modest spatial heterogeneity even in a single tumor. In addition, we divided whole tumor areas from the 215 PDACs into 29,968 ROIs (one ROI ∼1 mm2) using a tiling method (26) and performed a similar stromal classification by cluster analyses (Supplementary Fig. S10). Of the 29,968 ROIs, 10,753 (36%), 5,341 (18%), and 13,874 (46%) were categorized as C-type-ROI, F-type-ROI, and A-type-ROI, respectively. This ROI-by-ROI analysis revealed that only two of the 215 PDACs were composed of a single ROI type. Homogeneity index (based on the proportions of C-type-ROI, F-type-ROI, and A-type-ROI) for A-stroma (mean ± SD, 0.82 ± 8.5) or C-stroma (0.78 ± 9.7) was significantly higher than that for F-stoma (0.65 ± 6.2), indicating that spatial stromal heterogeneity was most evident in PDACs with F-stroma.

Correlations of stroma type with clinicopathologic and molecular features

Table 1 summarizes the associations of the stroma types with the clinical, pathologic, and molecular characteristics of 215 PDACs. A-stroma was significantly associated with large tumor size (P = 0.001). There was a trend between early-disease stage and F-stroma (P = 0.012). Tumors with A-stroma or F-stroma appeared to have worse tumor grades (P = 0.034), higher tumor cellularity (P = 0.008), and higher rates of loss of SMAD4 expression (P = 0.028), than those with C-stroma.

Table 1.

Clinicopathologic characteristics according to three distinct stroma types in 215 PDACs.

Pancreatic cancer stroma type
CharacteristicsaTotal no. (N = 215)C-stroma (n = 55)F-stroma (n = 95)A-stroma (n = 65)Pb
Mean age ± SD (years) 67.0 ± 9.6 64.9 ± 11.4 68.3 ± 8.1 66.9 ± 9.9 0.12 
Sex     0.78 
 Men 128 (60%) 32 (58%) 55 (58%) 41 (63%)  
 Women 87 (40%) 23 (42%) 40 (42%) 24 (37%)  
Mean tumor size ± SD (cm) 3.0 ± 1.2 2.9 ± 1.2 2.7 ± 1.0 3.4 ± 1.3 0.001 
Anatomic location     0.20 
 Head 128 (60%) 33 (58%) 62 (65%) 33 (51%)  
 Body/tail 87 (40%) 22 (42%) 33 (35%) 32 (49%)  
Tumor differentiation     0.034 
 Grade 1/2 150 (70%) 46 (84%) 62 (65%) 42 (65%)  
 Grade 3 65 (30%) 9 (16%) 33 (35%) 23 (35%)  
Tumor cellularity     0.008 
 Low (≤30%) 158 (73%) 49 (84%) 63 (66%) 46 (70%)  
 High (>30%) 57 (27%) 6 (16%) 32 (34%) 19 (30%)  
Residual tumor status     0.25 
 R0 152 (71%) 34 (62%) 70 (74%) 48 (74%)  
 R1 63 (29%) 21 (38%) 25 (26%) 17 (26%)  
TP53 expression     0.45 
 Intact 103 (48%) 29 (53%) 41 (43%) 33 (51%)  
 Aberrant 112 (52%) 26 (47%) 54 (57%) 32 (49%)  
SMAD4 expression     0.028 
 Intact 83 (39%) 29 (53%) 35 (37%) 19 (29%)  
 Lost 132 (61%) 26 (47%) 60 (63%) 46 (71%)  
Tertiary lymphoid structures     0.70 
 Minimal/absent 136 (63%) 34 (62%) 63 (66%) 39 (60%)  
 Extensive 79 (37%) 21 (38%) 32 (34%) 26 (40%)  
CD8+ cell density in the tumor center     0.11 
 Low 107 (50%) 21 (39%) 54 (57%) 32 (49%)  
 High 107 (50%) 33 (61%) 41 (43%) 33 (51%)  
Disease stage (UICC 8th)     0.012 
 I 42 (20%) 8 (15%) 28 (30%) 6 (9%)  
 II 100 (47%) 30 (55%) 36 (38%) 34 (52%)  
 III 70 (33%) 16 (29%) 31 (33%) 23 (35%)  
 IV 3 (1%) 1 (2%) 0 (0%) 2 (3%)  
Pancreatic cancer stroma type
CharacteristicsaTotal no. (N = 215)C-stroma (n = 55)F-stroma (n = 95)A-stroma (n = 65)Pb
Mean age ± SD (years) 67.0 ± 9.6 64.9 ± 11.4 68.3 ± 8.1 66.9 ± 9.9 0.12 
Sex     0.78 
 Men 128 (60%) 32 (58%) 55 (58%) 41 (63%)  
 Women 87 (40%) 23 (42%) 40 (42%) 24 (37%)  
Mean tumor size ± SD (cm) 3.0 ± 1.2 2.9 ± 1.2 2.7 ± 1.0 3.4 ± 1.3 0.001 
Anatomic location     0.20 
 Head 128 (60%) 33 (58%) 62 (65%) 33 (51%)  
 Body/tail 87 (40%) 22 (42%) 33 (35%) 32 (49%)  
Tumor differentiation     0.034 
 Grade 1/2 150 (70%) 46 (84%) 62 (65%) 42 (65%)  
 Grade 3 65 (30%) 9 (16%) 33 (35%) 23 (35%)  
Tumor cellularity     0.008 
 Low (≤30%) 158 (73%) 49 (84%) 63 (66%) 46 (70%)  
 High (>30%) 57 (27%) 6 (16%) 32 (34%) 19 (30%)  
Residual tumor status     0.25 
 R0 152 (71%) 34 (62%) 70 (74%) 48 (74%)  
 R1 63 (29%) 21 (38%) 25 (26%) 17 (26%)  
TP53 expression     0.45 
 Intact 103 (48%) 29 (53%) 41 (43%) 33 (51%)  
 Aberrant 112 (52%) 26 (47%) 54 (57%) 32 (49%)  
SMAD4 expression     0.028 
 Intact 83 (39%) 29 (53%) 35 (37%) 19 (29%)  
 Lost 132 (61%) 26 (47%) 60 (63%) 46 (71%)  
Tertiary lymphoid structures     0.70 
 Minimal/absent 136 (63%) 34 (62%) 63 (66%) 39 (60%)  
 Extensive 79 (37%) 21 (38%) 32 (34%) 26 (40%)  
CD8+ cell density in the tumor center     0.11 
 Low 107 (50%) 21 (39%) 54 (57%) 32 (49%)  
 High 107 (50%) 33 (61%) 41 (43%) 33 (51%)  
Disease stage (UICC 8th)     0.012 
 I 42 (20%) 8 (15%) 28 (30%) 6 (9%)  
 II 100 (47%) 30 (55%) 36 (38%) 34 (52%)  
 III 70 (33%) 16 (29%) 31 (33%) 23 (35%)  
 IV 3 (1%) 1 (2%) 0 (0%) 2 (3%)  

Abbreviation: UICC, the Union for International Cancer Control.

aPercentages indicate the proportion of cases with a specific clinicopathologic feature for each stroma subtype.

bThe χ2 test was performed to assess associations between stroma type and categorical data (except for disease stage, for which Fisher exact test was performed). To compare the mean age and mean tumor size, ANOVA was performed.

Using transcriptome data from 20 frozen PDAC samples (six tumors with C-stroma, five with F-stroma, and nine with A-stroma), we performed hierarchical cluster analyses according to 1,141 genes differentially expressed among the three stroma types (listed in Supplementary Table S2). Apart from one tumor (marked with an asterisk in Fig. 3A), the clusters based on transcriptome data conformed with our three stroma types. Fibroblast marker genes (e.g., ACTA2, FAP, and PDPN), various collagen genes (e.g., COL1A1 and COL1A2; Supplementary Fig. S11), and other ECM genes (e.g., FN1 and POSTN) were downregulated in C-stroma (Fig. 3B). The LOX gene, whose product plays a central role in collagen cross-linking, was downregulated in C-stroma, but was upregulated in A-stroma and F-stroma. A number of matrix metalloproteinase (MMP) family genes and a disintegrin and metalloproteinase (ADAM) family genes were generally downregulated in C-stroma; and many were upregulated, especially in F-stroma. Next, we conducted cluster analyses to categorize these 20 tumors using the transcriptome subtypes proposed by Moffitt and colleagues (Fig. 3C; ref. 30). All six tumors with C-stroma were of the Moffitt “normal” stroma subtype, whereas 10 of the 14 tumors with either A-stroma or F-stroma were grouped as “activated” stroma (P = 0.011; Supplementary Table S3). In addition, all six cases of C-stroma were categorized as having Moffitt “classical” tumor subtype (P = 0.014). In addition, C-stroma was associated with upregulation of “classical” tumor genes of Collisson PDAC subtype (41), including AGR2, LGALS4, and MUC13.

Figure 3.

A, Hierarchical cluster analyses according to 1,141 genes differentially expressed between the three stroma types. Only one tumor (asterisk) was classified differently by clusters on the basis of transcriptome data and the three image analysis stroma types. B, A Venn diagram of these 1,141 differentially expressed genes in the three stroma types. Selected genes for each signature are shown in the respective box. C, Cluster analyses to categorize tumors by gene sets associated with Moffitt and colleagues' transcriptome subtypes. The ATAD4, ANXA8L2, and CTCL2 genes in Moffitt and colleagues' gene sets are designated as PRR15L, ANXA8L1, and CTSV, respectively. IDs for LOC400573 and UCA1 are not assigned in the Clariom D chip.

Figure 3.

A, Hierarchical cluster analyses according to 1,141 genes differentially expressed between the three stroma types. Only one tumor (asterisk) was classified differently by clusters on the basis of transcriptome data and the three image analysis stroma types. B, A Venn diagram of these 1,141 differentially expressed genes in the three stroma types. Selected genes for each signature are shown in the respective box. C, Cluster analyses to categorize tumors by gene sets associated with Moffitt and colleagues' transcriptome subtypes. The ATAD4, ANXA8L2, and CTCL2 genes in Moffitt and colleagues' gene sets are designated as PRR15L, ANXA8L1, and CTSV, respectively. IDs for LOC400573 and UCA1 are not assigned in the Clariom D chip.

Close modal

Three stroma types differentially associated with clinical outcome of PDAC

In our primary hypothesis testing for patient outcome, we assessed the association between stroma type and pancreatic cancer–specific survival. Kaplan–Meier analysis revealed a significant association between stroma type and cancer-specific survival (P = 0.001 by the log-rank test; Fig. 4A). We conducted multivariable Cox regression analyses to determine whether each stroma type had an independent prognostic association with survival, after adjusting for clinicopathologic prognosticators (Fig. 4B; Supplementary Table S4). Compared with F-stroma, C-stroma was associated with prolonged cancer-specific survival [multivariable HR, 0.57; 95% confidence interval (CI), 0.33–0.99]; in contrast, A-stroma exhibited poorer prognosis (HR, 1.65; 95% CI, 1.06–2.58).

Figure 4.

Three stroma types differentially associated with patient survival for PDAC. A, Kaplan–Meier curves for cancer-specific survival according to stroma type. The P value was calculated using the log-rank test. The table shows the number of patients who remained alive and at risk of death at each timepoint. B, Forest plot showing multivariable HRs in the Cox regression model for cancer-specific survival. Detailed results of the multiplex Cox regression analysis are shown in Supplementary Table S4.

Figure 4.

Three stroma types differentially associated with patient survival for PDAC. A, Kaplan–Meier curves for cancer-specific survival according to stroma type. The P value was calculated using the log-rank test. The table shows the number of patients who remained alive and at risk of death at each timepoint. B, Forest plot showing multivariable HRs in the Cox regression model for cancer-specific survival. Detailed results of the multiplex Cox regression analysis are shown in Supplementary Table S4.

Close modal

Desmoplasia and the immune microenvironment of pancreatic cancer

We measured the CD8+ cell density in the tumor center and in the tumor margin using a method established previously (Fig. 5A). In-line with our previous study (26), most PDACs had a lower CD8+ cell density in the tumor center than in the tumor margin (Fig. 5B). To test our secondary hypothesis that, tumors with abundant areas of FAP-dominant fibroblasts have limited CD8+ lymphocytic infiltrates within tumor tissues, we analyzed correlations of the desmoplastic compositions with the ratio of CD8+ cell density in the tumor center to the tumor margin (CD8TC/CD8TM). There was an inverse correlation between the proportion of FAP-dominant fibroblasts and the CD8TC/CD8TM ratio (r = −0.27; P < 0.001; Fig. 5C), whereas the proportion of collagen only weakly correlated with the CD8TC/CD8TM ratio (r = 0.19; P = 0.006). No significant inverse correlation was observed between CD8TC or CD8TM and the proportion of collagen, FAP-dominant fibroblasts, or ACTA2-dominant fibroblasts (all r > −0.11; Supplementary Table S5). We found no significant difference in the proportion of FAP-dominant fibroblasts between the tumor center and the tumor margin (P = 0.38; Supplementary Fig. S12). F-stroma had the lowest CD8TC/CD8TM ratio among the three stroma types (P = 0.001 by Welch t test; Fig. 5D). Histologically, in F-stroma, we frequently observed marked infiltration of CD11b-expressing immune cells, including polymorphonuclear neutrophils, in addition to the findings of restricted CD8+ cell infiltrates (Fig. 5E). Higher numbers of tumor-infiltrating neutrophils, enumerated using H&E-stained slides, were associated with F-stroma (P = 0.005 by the χ2 test; Fig. 5F).

Figure 5.

Correlations between stroma type and tumor-infiltrating immune cells. A, ROIs (one ROI ∼1 mm2) intersected by the tumor border line were defined as the tumor margin (TM), and ROIs inside the tumor margin were defined as the tumor center (TC). B, Frequency plot of the ratio of CD8+ cell density in the tumor center to that in the tumor margin (CD8TC/CD8TM) for each PDAC case. C, Scatter plots with regression lines showing correlations of CD8TC/CD8TM ratios with the proportions of FAP-dominant fibroblasts, ACTA2-dominant fibroblasts, and collagen. Correlation coefficients (r) were calculated using the Spearman correlation test. D, Association between stroma types and CD8TC/CD8TM ratios according to Welch t test. The ends of the whiskers indicate the minimum and maximum values, unless outliers are present, in which case the whiskers extend to a maximum of 1.5 times the interquartile ranges (boxes). E, Images of H&E staining and IHC for CD8, KRT7 (representing carcinoma cells), FAP, ACTA2, and CD11b (a pan-myeloid marker) showing that CD8+ cell infiltration was severely restricted in the tumor center compared with the tumor margin (broken lines, tumor border; T, tumor; N, nontumor tissue). Marked infiltration of CD11b-expressing immune cells, including polymorphonuclear neutrophils (insets) were seen in F-stroma. F, Bar charts showing the relationship between stroma type and tumor-infiltrating neutrophils. HPF, high-power fields. The P value was calculated using the χ2 test.

Figure 5.

Correlations between stroma type and tumor-infiltrating immune cells. A, ROIs (one ROI ∼1 mm2) intersected by the tumor border line were defined as the tumor margin (TM), and ROIs inside the tumor margin were defined as the tumor center (TC). B, Frequency plot of the ratio of CD8+ cell density in the tumor center to that in the tumor margin (CD8TC/CD8TM) for each PDAC case. C, Scatter plots with regression lines showing correlations of CD8TC/CD8TM ratios with the proportions of FAP-dominant fibroblasts, ACTA2-dominant fibroblasts, and collagen. Correlation coefficients (r) were calculated using the Spearman correlation test. D, Association between stroma types and CD8TC/CD8TM ratios according to Welch t test. The ends of the whiskers indicate the minimum and maximum values, unless outliers are present, in which case the whiskers extend to a maximum of 1.5 times the interquartile ranges (boxes). E, Images of H&E staining and IHC for CD8, KRT7 (representing carcinoma cells), FAP, ACTA2, and CD11b (a pan-myeloid marker) showing that CD8+ cell infiltration was severely restricted in the tumor center compared with the tumor margin (broken lines, tumor border; T, tumor; N, nontumor tissue). Marked infiltration of CD11b-expressing immune cells, including polymorphonuclear neutrophils (insets) were seen in F-stroma. F, Bar charts showing the relationship between stroma type and tumor-infiltrating neutrophils. HPF, high-power fields. The P value was calculated using the χ2 test.

Close modal

Multiplex IHC analyses of human pancreatic cancer stroma revealed a wide spectrum of phenotypical variations of CAFs in terms of protein expression levels of known fibroblastic markers. Among these variations, two fibroblast subpopulations, that is, ACTA2-dominant fibroblasts and FAP-dominant fibroblasts, represented digitally distinguishable principal CAF populations in PDAC tissues. Consistent with this finding, several reports have suggested that ACTA2 and FAP may serve as surrogate markers to describe mutually exclusive fibroblast subtypes, just as CD20 and CD3 do for the stratification of B and T cells, respectively (6, 8, 28). Whereas ACTA2-expressing fibroblasts are the major source of collagen (42), the membrane glycoprotein, FAP, can act as a peptidase/proteinase to degrade collagen proteins (43), suggesting opposite roles of ACTA2-dominant fibroblasts and FAP-dominant fibroblasts in collagen remodeling. Moreover, recent studies have identified a fibroblast subtype characterized by its juxta-tumoral localization with elevated ACTA2 expressions in organoid and murine PDAC models (22). We found a similar arrangement in human PDAC tissues, in which well-differentiated tumor ducts were tightly surrounded by ACTA2-dominant CAFs with thick collagen bands. However, we observed that juxta-tumoral fibroblasts in poorly differentiated cancer areas frequently showed intense FAP expression, not ACTA2 expression. Furthermore, especially in A-stroma, we observed diffuse distribution patterns of ACTA2-dominant fibroblasts with a paucity of collagen accumulation. Evidence indicates the importance of multiple molecular signaling pathways, including NF-κB, JAK/STAT, TGFβ, and Hedgehog pathways, in dictating CAF subtypes (5, 6, 22, 44, 45). These lines of evidence, along with our findings, suggest that inter- and intratumoral stromal heterogeneity in human pancreatic cancer is highly context dependent and is likely influenced by dynamic and complex interplays between tumor cells, fibroblasts, immune cells, and ECM proteins in the tumor microenvironment of pancreatic cancer.

This study demonstrated considerable differences in tumor cell morphology along with desmoplastic stromal variations, tumor cells embedded in FAP-expressing stroma showed poorer differentiation with less epithelial phenotype, compared with those in ACTA2-expressing cellular stroma or collagen-rich stroma. A recent report suggests that FAP expression in pancreatic stellate cells can be induced by TGFβ1, which is mainly released by pancreatic cancer cells (46). Another recent study (47) indicates that pancreatic cancer cells with squamous differentiation recruit neutrophils and induce inflammatory CAFs in a TP63-dependent manner (a known molecular pathway that regulates pancreatic cancer subtypes; refs. 48, 49). It is possible that different tumor subtypes may cause stromal variations in the pancreatic cancer microenvironment. Future experimental studies are warranted to clarify this potential causal inference.

Stromal activity defined by transcriptome analyses or by the ratio of ACTA2-expressing fibroblastic area to collagen area has been associated with worse prognosis in patients with PDAC (16, 30, 50). Evidence also indicates that higher stromal levels of either ACTA2 or FAP correlate with dismal PDAC prognoses (6, 51, 52). These previous reports are generally compatible with our findings that C-stroma has lower fibroblast proportions and low expression levels of “activated” stromal genes and is related to favorable outcomes. Moreover, our image analysis–based data uncovered a clinical impact of fibroblast subpopulations, A-stroma was significantly associated with shorter survival than F-stroma, even after adjusting for disease stage, tumor grade, and immunologic factors. However, stromal therapies that broadly inhibit ACTA2-expressing fibroblasts would seem to be challenging, considering the evidence that Acta2 ablation or Hedgehog pathway inhibition in mice PDAC models resulted in a tumor undifferentiated state leading to the promotion of cancer progression (8, 9). These apparently conflicting findings may suggest the presence of tumor-promoting subpopulations and tumor-restricting subpopulations among ACTA2-expressing fibroblasts within pancreatic cancer stroma. In contrast, several preclinical studies have suggested a promising benefit of FAP inhibition in PDAC (28, 53). In this study, in most patients with PDAC, the proportion of FAP-dominant fibroblasts was lower than that of ACTA2-dominant fibroblasts. However, the depletion of FAP-expressing fibroblasts reportedly simultaneously decreased the number of ACTA2-expressing fibroblasts by 70% (28), whereas, in contrast, inhibition of ACTA2-expressing fibroblasts has shown no effect on FAP-expressing stromal populations (8). Therefore, stromal FAP expression could become a biomarker to select patients who would receive potential benefit from antistromal therapies targeting FAP.

This study revealed that the proportion of FAP-dominant fibroblasts in PDAC stroma was inversely associated with the ratios of CD8+ cell density in the tumor center to the tumor margin, but not with CD8+ cell density in the tumor center or that in the tumor margin, suggesting that FAP-dominant fibroblasts may contribute to the spatial exclusion of intratumoral CD8+ T cells, regardless of baseline levels of CD8+ cell infiltration. A previous study showed that the ablation of Fap-expressing stromal cells in a subcutaneous murine PDAC model elicited immune-mediated tumor regressions in response to therapeutic vaccination (33). Moreover, inhibition of Cxcl12 from Fap-expressing CAFs promoted T-cell accumulation in the tumor center and fostered the pharmacologic efficacy of the PDCD1 immune checkpoint blockade (34). The COMBAT phase II trial, evaluating an antagonist of CXCR4 (a CXCL12 receptor) with pembrolizumab and chemotherapy in patients with PDAC, demonstrated promising rates of disease control (54). Moreover, in this study, CD11b+ myeloid cells, including polymorphonuclear neutrophils, were present in significant numbers in F-stroma. Accumulating evidence suggests that in several types of cancers, including PDAC, Acta2/Fap+ CAFs secrete a number of cytokines and chemokines, including Cxcl1, Cxcl1, Cxcl12, Il6, Csf3, and Lif, leading to suppressed adaptive immunity by recruiting various immunosuppressive cells such as myeloid-derived suppressor cells, neutrophils, and macrophages (6, 7, 44, 55, 56). Taken together, FAP-dominant fibroblasts may be key players in immunotherapy resistance in human PDAC.

We found a relatively strong correlation between the occupancy rates of collagen and ACTA2-dominant fibroblasts in nonneoplastic pancreas, consistent with the general idea that fibrotic stroma should contain more collagen together with more fibroblasts. However, there was no positive correlation between the occupancy rates of collagen and ACTA2-dominant fibroblasts in PDAC stroma. This discrepancy between noncancerous fibrosis and tumor-associated desmoplasia is probably because of enhanced collagen degradation in pancreatic cancer stroma by carcinoma cells, macrophages, and FAP-expressing fibroblasts, or, alternatively, because of reduced collagen production by CAFs.

Moffitt-activated stroma genes include various types of collagen genes, including COL1A1 and COL1A2 (30). In our datasets, collagen genes were generally downregulated in C-stroma, whereas C-stroma has, by definition, larger amounts of collagen and smaller numbers of fibroblasts than A-stroma or F-stroma. Furthermore, our transcriptome data indicated the downregulation of MMP genes (several of which encode collagenases) in C-stroma. Because collagen deposition in the ECM is a result of the balance between productive and degradation processes, C-stroma might be caused mainly by the low fibrolytic activity of tumor cells and/or stromal cells, not by elevated fibrogenic activity. In contrast, gene expression data also found that F-stroma had the highest fibrogenic activity, as well as the highest fibrolytic activity, suggesting that pancreatic cancer with F-stroma is additionally characterized by dynamic ECM remodeling in the tumor microenvironment.

All the six PDACs with C-stoma (having smaller numbers of ACTA2-dominant fibroblasts and downregulated ACTA2 expressions) were categorized as Moffitt “normal” stoma subtype in our study. However, ACTA2 has been relatively upregulated in “normal” stroma in Moffitt and colleagues' study (30). These findings may suggest the heterogenous nature of Moffitt “normal” stroma subtype in terms of fibroblastic compositions. Indeed, one tumor with F-stroma and three tumors with A-stroma were grouped as Moffitt “normal” subtype in this study.

This study has some limitations. One limitation is the cross-sectional design for the associations between desmoplastic elements and immunity. As a result, we could not exclude the possibility of reverse causation (e.g., immune cells may have impacts on fibroblast phenotypes). However, our hypothesis was based on several experimental studies indicating that FAP-expressing cells suppress antitumor immunity (7, 28, 33, 34). Because human tumors and host immune systems have been too complicated to recapitulate in any experimental models (57), our quantitative analyses using clinical specimens could be subject to high variability. Another limitation was our multiplex IHC system that could evaluate up to four markers in a single slide. In our system, nonfibroblastic ACTA2-expressing components (e.g., smooth muscle) were excluded using anti-h-caldesmon antibody. FAP and ACTA2 have been surrogate markers to detect fibroblasts, but can be expressed in several types of stromal cells and epithelial cells, including CD31+ cells and CD45+ cells (7, 33, 34). Therefore, it would be better to exclude nonfibroblastic FAP/ACTA2 signals using additional markers for a more precise analysis of FAP/ACTA2-expressing fibroblasts. Nonetheless, our IHC analyses did not uncover detectable levels of FAP/ACTA2 expression in CD31-expressing cells, CD45-expressing cells, or KRT7-expressing carcinoma cells within human pancreatic cancer tissues. Future technical advances and/or combination with RNA in situ systems should be able to perform comprehensive phenotypic subclassifications of fibroblasts through the simultaneous evaluation of a larger number of molecules in a single tissue slide.

The major strengths of this study include the application of computational image analyses with the use of an automated IHC stainer. Together, these systems provide highly reproducible data on desmoplastic elements and CD8+ cells. In addition, the analyses of whole-slide tissues enabled us to quantify desmoplastic elements over a wide area, thereby overcoming potential issues of intratumoral heterogeneity. By utilizing these data on stromal heterogeneity and well-annotated clinicopathologic information on 215 PDACs, we rigorously performed hypothesis-generating analyses and discovered distinct stroma types that were differentially associated with survival, immunity, and the molecular features of human pancreatic cancer.

To the best of our knowledge, this is the first study to comprehensively visualize and characterize inter- and intratumoral heterogeneity with respect to principal fibroblast subpopulations and collagen in a relatively large PDAC cohort. Our multiplex image analysis system identified three distinct stroma types by stromal heterogeneity, attesting to the importance of stroma-based PDAC subtyping. As with many targeted treatments, adequate patient selection may be critical for the design of effective antistromal therapies. Therefore, our findings will likely inform translational research and clinical trials aiming at gaining a deeper knowledge of the clinical impact of stromal heterogeneity and thereby, facilitate the development of novel therapeutic strategies against pancreatic cancer.

No disclosures were reported.

Y. Ogawa: Data curation, formal analysis, investigation, visualization, writing-original draft, writing-review and editing. Y. Masugi: Conceptualization, data curation, formal analysis, supervision, funding acquisition, investigation, visualization, methodology, writing-original draft, writing-review and editing. T. Abe: Software, formal analysis, validation, investigation, visualization, methodology, writing-review and editing. K. Yamazaki: Formal analysis, investigation, writing-review and editing. A. Ueno: Data curation, validation, writing-review and editing. Y. Fujii-Nishimura: Data curation, validation, writing-review and editing. S. Hori: Data curation, writing-review and editing. H. Yagi: Data curation, writing-review and editing. Y. Abe: Data curation, writing-review and editing. M. Kitago: Data curation, writing-review and editing. M. Sakamoto: Conceptualization, supervision, project administration, writing-review and editing.

This work was supported by KAKENHI (grant nos. 18K15094 and 20K07414 to Y. Masugi) from the Japan Society for the Promotion of Science and Keio University Academic Development Funds for Individual Research from Keio University (to Y. Masugi). The authors are grateful to the Fourth Laboratory of the Department of Pathology and the Collaborative Research Resources at Keio University School of Medicine for assistance with technical support.

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.
Kleeff
J
,
Korc
M
,
Apte
M
,
La Vecchia
C
,
Johnson
CD
,
Biankin
AV
, et al
Pancreatic cancer
.
Nat Rev Dis Primers
2016
;
2
:
16022
.
2.
Ohlund
D
,
Elyada
E
,
Tuveson
D
. 
Fibroblast heterogeneity in the cancer wound
.
J Exp Med
2014
;
211
:
1503
23
.
3.
Neesse
A
,
Bauer
CA
,
Ohlund
D
,
Lauth
M
,
Buchholz
M
,
Michl
P
, et al
Stromal biology and therapy in pancreatic cancer: ready for clinical translation?
Gut
2019
;
68
:
159
71
.
4.
Vennin
C
,
Melenec
P
,
Rouet
R
,
Nobis
M
,
Cazet
AS
,
Murphy
KJ
, et al
CAF hierarchy driven by pancreatic cancer cell p53-status creates a pro-metastatic and chemoresistant environment via perlecan
.
Nat Commun
2019
;
10
:
3637
.
5.
Olive
KP
,
Jacobetz
MA
,
Davidson
CJ
,
Gopinathan
A
,
McIntyre
D
,
Honess
D
, et al
Inhibition of Hedgehog signaling enhances delivery of chemotherapy in a mouse model of pancreatic cancer
.
Science
2009
;
324
:
1457
61
.
6.
Whittle
MC
,
Hingorani
SR
. 
Fibroblasts in pancreatic ductal adenocarcinoma: biological mechanisms and therapeutic targets
.
Gastroenterology
2019
;
156
:
2085
96
.
7.
Kalluri
R
. 
The biology and function of fibroblasts in cancer
.
Nat Rev Cancer
2016
;
16
:
582
98
.
8.
Ozdemir
BC
,
Pentcheva-Hoang
T
,
Carstens
JL
,
Zheng
X
,
Wu
CC
,
Simpson
TR
, et al
Depletion of carcinoma-associated fibroblasts and fibrosis induces immunosuppression and accelerates pancreas cancer with reduced survival
.
Cancer Cell
2014
;
25
:
719
34
.
9.
Rhim
AD
,
Oberstein
PE
,
Thomas
DH
,
Mirek
ET
,
Palermo
CF
,
Sastra
SA
, et al
Stromal elements act to restrain, rather than support, pancreatic ductal adenocarcinoma
.
Cancer Cell
2014
;
25
:
735
47
.
10.
Catenacci
DV
,
Junttila
MR
,
Karrison
T
,
Bahary
N
,
Horiba
MN
,
Nattam
SR
, et al
Randomized phase Ib/II study of gemcitabine plus placebo or vismodegib, a hedgehog pathway inhibitor, in patients with metastatic pancreatic cancer
.
J Clin Oncol
2015
;
33
:
4284
92
.
11.
Vennin
C
,
Murphy
KJ
,
Morton
JP
,
Cox
TR
,
Pajic
M
,
Timpson
P
. 
Reshaping the tumor stroma for treatment of pancreatic cancer
.
Gastroenterology
2018
;
154
:
820
38
.
12.
Sahai
E
,
Astsaturov
I
,
Cukierman
E
,
DeNardo
DG
,
Egeblad
M
,
Evans
RM
, et al
A framework for advancing our understanding of cancer-associated fibroblasts
.
Nat Rev Cancer
2020
;
20
:
174
86
.
13.
Helms
E
,
Onate
MK
,
Sherman
MH
. 
Fibroblast heterogeneity in the pancreatic tumor microenvironment
.
Cancer Discov
2020
;
10
:
648
56
.
14.
Pereira
BA
,
Vennin
C
,
Papanicolaou
M
,
Chambers
CR
,
Herrmann
D
,
Morton
JP
, et al
CAF subpopulations: a new reservoir of stromal targets in pancreatic cancer
.
Trends Cancer
2019
;
5
:
724
41
.
15.
Mahajan
UM
,
Langhoff
E
,
Goni
E
,
Costello
E
,
Greenhalf
W
,
Halloran
C
, et al
Immune cell and stromal signature associated with progression-free survival of patients with resected pancreatic ductal adenocarcinoma
.
Gastroenterology
2018
;
155
:
1625
39
.
16.
Erkan
M
,
Michalski
CW
,
Rieder
S
,
Reiser-Erkan
C
,
Abiatari
I
,
Kolb
A
, et al
The activated stroma index is a novel and independent prognostic marker in pancreatic ductal adenocarcinoma
.
Clin Gastroenterol Hepatol
2008
;
6
:
1155
61
.
17.
Knudsen
ES
,
Vail
P
,
Balaji
U
,
Ngo
H
,
Botros
IW
,
Makarov
V
, et al
Stratification of pancreatic ductal adenocarcinoma: combinatorial genetic, stromal, and immunologic markers
.
Clin Cancer Res
2017
;
23
:
4429
40
.
18.
Ligorio
M
,
Sil
S
,
Malagon-Lopez
J
,
Nieman
LT
,
Misale
S
,
Di Pilato
M
, et al
Stromal microenvironment shapes the intratumoral architecture of pancreatic cancer
.
Cell
2019
;
178
:
160
75
.
19.
Elyada
E
,
Bolisetty
M
,
Laise
P
,
Flynn
WF
,
Courtois
ET
,
Burkhart
RA
, et al
Cross-species single-cell analysis of pancreatic ductal adenocarcinoma reveals antigen-presenting cancer-associated fibroblasts
.
Cancer Discov
2019
;
9
:
1102
23
.
20.
Bernard
V
,
Semaan
A
,
Huang
J
,
San Lucas
FA
,
Mulu
FC
,
Stephens
BM
, et al
Single-cell transcriptomics of pancreatic cancer precursors demonstrates epithelial and microenvironmental heterogeneity as an early event in neoplastic progression
.
Clin Cancer Res
2019
;
25
:
2194
205
.
21.
Neuzillet
C
,
Tijeras-Raballand
A
,
Ragulan
C
,
Cros
J
,
Patil
Y
,
Martinet
M
, et al
Inter- and intra-tumoural heterogeneity in cancer-associated fibroblasts of human pancreatic ductal adenocarcinoma
.
J Pathol
2019
;
248
:
51
65
.
22.
Ohlund
D
,
Handly-Santana
A
,
Biffi
G
,
Elyada
E
,
Almeida
AS
,
Ponz-Sarvise
M
, et al
Distinct populations of inflammatory fibroblasts and myofibroblasts in pancreatic cancer
.
J Exp Med
2017
;
214
:
579
96
.
23.
Pivetta
E
,
Spessotto
P
. 
Multispectral imaging technology: Visualize, analyze, phenotyping, and quantify immune cells in situ
.
Int J Biol Markers
2020
;
35
:
26
30
.
24.
Carstens
JL
,
Correa de Sampaio
P
,
Yang
D
,
Barua
S
,
Wang
H
,
Rao
A
, et al
Spatial computation of intratumoral T cells correlates with survival of patients with pancreatic cancer
.
Nat Commun
2017
;
8
:
15095
.
25.
Carvajal-Hausdorf
DE
,
Schalper
KA
,
Neumeister
VM
,
Rimm
DL
. 
Quantitative measurement of cancer tissue biomarkers in the lab and in the clinic
.
Lab Invest
2015
;
95
:
385
96
.
26.
Masugi
Y
,
Abe
T
,
Ueno
A
,
Fujii-Nishimura
Y
,
Ojima
H
,
Endo
Y
, et al
Characterization of spatial distribution of tumor-infiltrating CD8(+) T cells refines their prognostic utility for pancreatic cancer survival
.
Mod Pathol
2019
;
32
:
1495
507
.
27.
Maehara
J
,
Masugi
Y
,
Abe
T
,
Tsujikawa
H
,
Kurebayashi
Y
,
Ueno
A
, et al
Quantification of intratumoral collagen and elastin fibers within hepatocellular carcinoma tissues finds correlations with clinico-patho-radiological features
.
Hepatol Res
2020
;
50
:
607
19
.
28.
Lo
A
,
Wang
LS
,
Scholler
J
,
Monslow
J
,
Avery
D
,
Newick
K
, et al
Tumor-promoting desmoplasia is disrupted by depleting FAP-expressing stromal cells
.
Cancer Res
2015
;
75
:
2800
10
.
29.
Tian
C
,
Clauser
KR
,
Ohlund
D
,
Rickelt
S
,
Huang
Y
,
Gupta
M
, et al
Proteomic analyses of ECM during pancreatic ductal adenocarcinoma progression reveal different contributions by tumor and stromal cells
.
Proc Natl Acad Sci U S A
2019
;
116
:
19609
18
.
30.
Moffitt
RA
,
Marayati
R
,
Flate
EL
,
Volmar
KE
,
Loeza
SG
,
Hoadley
KA
, et al
Virtual microdissection identifies distinct tumor- and stroma-specific subtypes of pancreatic ductal adenocarcinoma
.
Nat Genet
2015
;
47
:
1168
78
.
31.
Balachandran
VP
,
Beatty
GL
,
Dougan
SK
. 
Broadening the impact of immunotherapy to pancreatic cancer: challenges and opportunities
.
Gastroenterology
2019
;
156
:
2056
72
.
32.
Johnson
BA
 III
,
Yarchoan
M
,
Lee
V
,
Laheru
DA
,
Jaffee
EM
. 
Strategies for increasing pancreatic tumor immunogenicity
.
Clin Cancer Res
2017
;
23
:
1656
69
.
33.
Kraman
M
,
Bambrough
PJ
,
Arnold
JN
,
Roberts
EW
,
Magiera
L
,
Jones
JO
, et al
Suppression of antitumor immunity by stromal cells expressing fibroblast activation protein-alpha
.
Science
2010
;
330
:
827
30
.
34.
Feig
C
,
Jones
JO
,
Kraman
M
,
Wells
RJ
,
Deonarine
A
,
Chan
DS
, et al
Targeting CXCL12 from FAP-expressing carcinoma-associated fibroblasts synergizes with anti-PD-L1 immunotherapy in pancreatic cancer
.
Proc Natl Acad Sci U S A
2013
;
110
:
20212
7
.
35.
Fujita
Y
,
Matsuda
S
,
Sasaki
Y
,
Masugi
Y
,
Kitago
M
,
Yagi
H
, et al
Pathogenesis of multiple pancreatic cancers involves multicentric carcinogenesis and intrapancreatic metastasis
.
Cancer Sci
2020
;
111
:
739
48
.
36.
Solomon
IH
,
Borscheid
R
,
Laga
AC
,
Askari
R
,
Granter
SR
. 
Frozen sections are unreliable for the diagnosis of necrotizing soft tissue infections
.
Mod Pathol
2018
;
31
:
546
52
.
37.
Iwata
E
,
Scarborough
M
,
Bowden
G
,
McNally
M
,
Tanaka
Y
,
Athanasou
NA
. 
The role of histology in the diagnosis of spondylodiscitis: correlation with clinical and microbiological findings
.
Bone Joint J
2019
;
101-B
:
246
52
.
38.
Caruso
RA
,
Bellocco
R
,
Pagano
M
,
Bertoli
G
,
Rigoli
L
,
Inferrera
C
. 
Prognostic value of intratumoral neutrophils in advanced gastric carcinoma in a high-risk area in northern Italy
.
Mod Pathol
2002
;
15
:
831
7
.
39.
Vahadane
A
,
Peng
T
,
Sethi
A
,
Albarqouni
S
,
Wang
L
,
Baust
M
, et al
Structure-preserving color normalization and sparse stain separation for histological images
.
IEEE Trans Med Imaging
2016
;
35
:
1962
71
.
40.
Masugi
Y
,
Abe
T
,
Tsujikawa
H
,
Effendi
K
,
Hashiguchi
A
,
Abe
M
, et al
Quantitative assessment of liver fibrosis reveals a nonlinear association with fibrosis stage in nonalcoholic fatty liver disease
.
Hepatol Commun
2018
;
2
:
58
68
.
41.
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
.
42.
Pothula
SP
,
Pirola
RC
,
Wilson
JS
,
Apte
MV
. 
Pancreatic stellate cells: aiding and abetting pancreatic cancer progression
.
Pancreatology
2020
;
20
:
409
18
.
43.
Fan
MH
,
Zhu
Q
,
Li
HH
,
Ra
HJ
,
Majumdar
S
,
Gulick
DL
, et al
Fibroblast activation protein (FAP) accelerates collagen degradation and clearance from lungs in mice
.
J Biol Chem
2016
;
291
:
8070
89
.
44.
Biffi
G
,
Oni
TE
,
Spielman
B
,
Hao
Y
,
Elyada
E
,
Park
Y
, et al
IL1-Induced JAK/STAT signaling is antagonized by TGFbeta to shape CAF heterogeneity in pancreatic ductal adenocarcinoma
.
Cancer Discov
2019
;
9
:
282
301
.
45.
Crawford
HC
,
Pasca di Magliano
M
,
Banerjee
S
. 
Signaling networks that control cellular plasticity in pancreatic tumorigenesis, progression, and metastasis
.
Gastroenterology
2019
;
156
:
2073
84
.
46.
Wen
Z
,
Liu
Q
,
Wu
J
,
Xu
B
,
Wang
J
,
Liang
L
, et al
Fibroblast activation protein alpha-positive pancreatic stellate cells promote the migration and invasion of pancreatic cancer by CXCL1-mediated Akt phosphorylation
.
Ann Transl Med
2019
;
7
:
532
.
47.
Somerville
TD
,
Biffi
G
,
Dassler-Plenker
J
,
Hur
SK
,
He
XY
,
Vance
KE
, et al
Squamous trans-differentiation of pancreatic cancer cells promotes stromal inflammation
.
Elife
2020
;
9
:
e53381
.
48.
Hamdan
FH
,
Johnsen
SA
. 
DeltaNp63-dependent super enhancers define molecular identity in pancreatic cancer by an interconnected transcription factor network
.
Proc Natl Acad Sci U S A
2018
;
115
:
E12343
E52
.
49.
Somerville
TDD
,
Xu
Y
,
Miyabayashi
K
,
Tiriac
H
,
Cleary
CR
,
Maia-Silva
D
, et al
TP63-mediated enhancer reprogramming drives the squamous subtype of pancreatic ductal adenocarcinoma
.
Cell Rep
2018
;
25
:
1741
55
.
50.
Puleo
F
,
Nicolle
R
,
Blum
Y
,
Cros
J
,
Marisa
L
,
Demetter
P
, et al
Stratification of pancreatic ductal adenocarcinomas based on tumor and microenvironment features
.
Gastroenterology
2018
;
155
:
1999
2013
.
51.
Cohen
SJ
,
Alpaugh
RK
,
Palazzo
I
,
Meropol
NJ
,
Rogatko
A
,
Xu
Z
, et al
Fibroblast activation protein and its relationship to clinical outcome in pancreatic adenocarcinoma
.
Pancreas
2008
;
37
:
154
8
.
52.
Sinn
M
,
Denkert
C
,
Striefler
JK
,
Pelzer
U
,
Stieler
JM
,
Bahra
M
, et al
alpha-Smooth muscle actin expression and desmoplastic stromal reaction in pancreatic cancer: results from the CONKO-001 study
.
Br J Cancer
2014
;
111
:
1917
23
.
53.
Watabe
T
,
Liu
Y
,
Kaneda-Nakashima
K
,
Shirakami
Y
,
Lindner
T
,
Ooe
K
, et al
Theranostics targeting fibroblast activation protein in the tumor stroma: (64)Cu- and (225)Ac-labeled FAPI-04 in pancreatic cancer xenograft mouse models
.
J Nucl Med
2020
;
61
:
563
9
.
54.
Bockorny
B
,
Semenisty
V
,
Macarulla
T
,
Borazanci
E
,
Wolpin
BM
,
Stemmer
SM
, et al
BL-8040, a CXCR4 antagonist, in combination with pembrolizumab and chemotherapy for pancreatic cancer: the COMBAT trial
.
Nat Med
2020
;
26
:
878
85
.
55.
Yang
X
,
Lin
Y
,
Shi
Y
,
Li
B
,
Liu
W
,
Yin
W
, et al
FAP promotes immunosuppression by cancer-associated fibroblasts in the tumor microenvironment via STAT3-CCL2 signaling
.
Cancer Res
2016
;
76
:
4124
35
.
56.
Stromnes
IM
,
Brockenbrough
JS
,
Izeradjene
K
,
Carlson
MA
,
Cuevas
C
,
Simmons
RM
, et al
Targeted depletion of an MDSC subset unmasks pancreatic ductal adenocarcinoma to adaptive immunity
.
Gut
2014
;
63
:
1769
81
.
57.
Ogino
S
,
Nowak
JA
,
Hamada
T
,
Phipps
AI
,
Peters
U
,
Milner
DA
 Jr
, et al
Integrative analysis of exogenous, endogenous, tumour and immune factors for precision medicine
.
Gut
2018
;
67
:
1168
80
.