HMGA2 Expression Predicts Subtype, Survival, and Treatment Outcome in Pancreatic Ductal Adenocarcinoma

Abstract

Purpose:

To establish HMGA2 as a marker of basal-like disease in pancreatic ductal adenocarcinoma (PDAC) and explore its use as a biomarker for prognosis and treatment resistance.

Experimental Design:

We identified high expression of HMGA2 in basal PDAC cells in a scRNAseq Atlas of 172 patient samples. We then analyzed HMGA2 expression, along with expression of the classical marker GATA6, in a cohort of 580 PDAC samples with multiplex immunohistochemistry. We further supplemented these data with an additional 30 diverse patient samples and multiple independent single-cell RNAseq databases.

Results:

We found that expression of HMGA2, but not previously described basal markers CK5 or CK17, predicted overall survival in our cohort. Combining HMGA2 and GATA6 status allowed for identification of two key study groups: an HMGA2+/GATA6- cohort with worse survival, low tumor-infiltrating CD8+ T cells, increased FAP+ fibroblasts, and poorer response to gemcitabine-based chemotherapies (n=94, median survival=11.2 months post-surgery); and an HMGA2-/GATA6+ cohort with improved survival, increased CD8+ T-cell infiltrate, decreased FAP+ fibroblasts, and improved survival with gemcitabine-based chemotherapy (n=198, median survival = 21.7 months post-surgery). HMGA2 was also prognostic for overall survival in RNA sequencing from an independent cohort.

Conclusions:

IHC stratification of primary tumors by HMGA2 and GATA6 status in pancreatic cancer is associated with differential outcomes, survival following chemotherapy, and tumor microenvironments. As a nuclear marker for basal disease, HMGA2 complements GATA6 to identify disease subtypes in PDAC.

INTRODUCTION

Pancreatic ductal adenocarcinoma (PDAC) is one of the deadliest cancers in the United States, with a current five-year survival rate of 13%. It is projected to become the second leading cause of cancer-related death by 2030, in part due to a lack of effective screening and therapies (1). Although surgical resection can be performed in early stages, the majority of patients are diagnosed with locally advanced or metastatic disease (2), and combination chemotherapy remains their only treatment option under National Comprehensive Cancer Network guidelines (3). Gemcitabine plus nab-paclitaxel (GnP) (4) and modified FOLFIRINOX (mFFX) (5) regimens are common systemic therapies in late-stage PDAC, and liposomal irinotecan was recently FDA approved (6). However, there are currently no clinically relevant biomarkers to establish the most appropriate treatment regimen for a given patient.

Efforts to define targetable genetic mutations in pancreatic cancer stalled after the discovery that most tumors bear the same suite of genetic lesions: oncogenic gain-of-function mutations in KRAS plus loss of TP53, CDKN2A, and TGF-β pathway members (7–9). To explain the variety of survival rates seen in clinic, the field turned to defining transcriptional states in pancreatic cancer. Multiple landmark studies agree on two core transcriptional subtypes: “classical” and “basal” (10–14). The basal subtype, also called the quasi-mesenchymal or squamous, portends poorer overall survival across patient cohorts (10–14). The basal subtype is not defined morphologically or histologically but is instead characterized by expression of gene signatures associated with squamous identity and loss of pancreas-specific gene expression (12,15). It is generally thought to comprise approximately 20% of all tumors and is enriched in late-stage disease compared to the classical subtype (13). More recent studies have also identified hybrid tumors, which display both classical and basal signatures and may result from basal- and classical-subtype cells within the same tumor (13,16,17). To subtype tumors without RNA sequencing, which is expensive and time-consuming especially in under-resourced settings, multiple groups have established protein markers as surrogates for the basal and classical subtypes. Of these, the most well-established are cytokeratins 5 (CK5) and 17 (CK17) for the basal subtype (16–18) and GATA-binding factor 6 (GATA6) for the classical subtype (13,16,17,19).

In separate work from our lab, we have recently identified the architectural chromatin remodeling protein high mobility group A2 (HMGA2) as an important driver of aggressive pancreatic cancer and the basal subtype (20) (unpublished data). The role of HMGA2 in lung, breast, and colorectal cancer has been studied (21–24), but its use as a biomarker in pancreatic cancer has been limited to mouse models and cell lines (25–27). HMGA2 was identified as one of over 2000 genes differentially regulated in squamous subtype patient samples by Bailey et al. (10). We have previously observed enrichment of HMGA2 in basal subtype PDAC cell lines (20). Additionally, high expression of HMGA2 in The Cancer Genome Atlas (TCGA) pancreatic cancer cohort predicted poorer overall survival (Fig. 1A). Here, we explore the use of HMGA2 immunohistochemistry (IHC) as a clinically relevant prognostic marker in PDAC.

Figure 1. HMGA2 is highly expressed in the basal subtype.

A) Survival of TCGA patient samples segregated by HMGA2 expression shows high expression of HMGA2 predicts poorer survival. B) scRNAseq atlas of 172 patient samples shows clear segregation of tumor epithelial cells following the scBasal and scClassical signatures (Raghavan et al., top) and Basal 1, Basal 2, Classical 1, and Classical 2 signatures (Chan-Seng-Yue et al., middle). Bailey et al. signatures do not consistently map to discrete areas of the UMAP, despite the inclusion of HMGA2 as a differentially expressed gene in the squamous subtype (bottom). C) scRNAseq atlas demonstrates enrichment of HMGA2, S100A2, and KRT5 in basal populations (left) and GATA6, TFF2, and LGALS4 in classical populations (right). D) Analysis of differentially regulated genes in HMGA2 positive (top) and GATA6 positive (middle) populations reveals basal genes S100A2 and FN1 in the HMGA2-positive cells. Double positive cells (bottom) display mixed gene signatures.

MATERIALS AND METHODS

scRNAseq Atlas.

Human scRNAseq data was aggregated from 12 studies (EGAS00001002543, GSE154778, GSE155698, GSE156405, GSE158356, GSE194247, GSE202051, GSE205013, GSE211644, GSE229413, phs001840.v1.p1, PRJCA001063) consisting of 172 primary PDAC tumors, as described in Loveless et al. (28). Briefly, all publicly available 10x Genomics scRNAseq studies of human PDAC were accessed, and scRNAseq data files were normalized and scaled in Seurat (v4, RRID:SCR_007322) and batch corrected with Harmony batch correction tool in R (RRID:SCR_022206). Unbiased clustering was used to identify major cell type lineages, and from there only ductal putative malignant cells (inferred copy number variation of 0.15 or above) were selected for analysis. Cells with fewer than 200 genes, mitochondrial percent >25%, or total counts greater than the 95^th percentile for the dataset were excluded. Individual ductal cells were scored for tumor molecular subtypes using ssGSEA (29) from signatures published by Raghavan et al. (30), Chan-Seng-Yue et al. (13), and Bailey et al. (10). Scores were mapped onto cells in the UMAP space using Seurat’s FeaturePlot function.

Sample collection.

For the 580-patient tumor microarray, tumor samples from the “Analysis of Solid Tumors Biobank” were obtained and accessed in accordance following approval by Washington University’s Institutional Review Board (IRB# 201108117) and WAIVER of Elements of Consent as per 45 CFR 46.116(d). All samples were obtained following written informed consent, and patient information was deidentified prior to investigator use. All human research activities and all activities of the IRBs designated in the Washington University Federal Wide Assurance, regardless of sponsorship, are guided by the ethical principles in “The Belmont Report: Ethical Principles and Guidelines for the Protection of Human Subjects Research of the National Commission for the Protection of Human Subjects of Biomedical and Behavioral Research.” For the additional, untreated, age-matched 30 patients from Ochsner Health, samples were collected following writing informed consent as standard of care and were obtained and accessed following approval by the Ochsner Institutional Review Board (IRB# 2020.322). All patient information was deidentified prior to investigator use. This institution assures that all its activities related to human subjects research, regardless of the source of support, will be guided by the Belmont Report and its statements of principles governing the institution in the discharge of its responsibilities for protecting the rights and welfare of human subjects of research conducted at or sponsored by the institution. See Supplementary Table S1 for demographic details of both cohorts.

Tumor Microarray (TMA).

580 patient samples were analyzed and had representative two-millimeter cores selected by a pathologist. Cores were embedded in 5 × 9 grids in paraffin along with control tonsil, spleen, breast, placenta, and lung adenocarcinoma samples for antibody verification and orientation. Thirteen separate TMAs were created.

Multiplex immunohistochemistry (mIHC).

4-μm sections of the TMAs were sectioned onto positively charged slides and baked for 1 hour at 65°C. Slides were then dewaxed and stained on a Leica BOND RX stainer (Leica, Buffalo Grove, IL) using Leica BOND reagents for dewaxing (Dewax Solution), antigen retrieval and antibody stripping (Epitope Retrieval Solution 2), and rinsing (Bond Wash Solution). Antigen retrieval and antibody stripping was performed at 100°C, while all other steps were performed at room temperature. Endogenous peroxidase was blocked with 3% H₂O₂ for five minutes, followed by protein blocking with TCT buffer (0.05M Tris, 0.15M NaCl, 0.25% Casein, 0.1% Tween, 0.05% ProClin300, pH 7.6) for ten minutes. First primary antibody was applied for one hour, followed by the first secondary antibody for 20 minutes, then the tertiary TSA-amplification reagent (Akoya Biosciences OPAL fluorophores) for 20 minutes. A high-stringency wash was performed using high-salt TBST solution (0.05M Tris, 0.3M NaCl, 0.1% Tween-20, pH 7.2–7.6). Host-specific polymer HRP was used for all secondary applications, either 1X Opal Anti-Ms + Rb HRP (Akoya Biosciences, Cat# ARH1001EA) or PowerVision Poly-HRP anti-mouse IgG (Leica Biosystems, Cat# PV6114). Primary and secondary antibodies were stripped with retrieval solution for 20 minutes. The process was then repeated with the second position primary and secondary antibodies from the 3% H₂O₂ step. The process was continued until all six markers were probed. For the sixth position, after secondary antibody incubation, Opal TSA-DIG (PerkinElmer Cat# NEL748001KT) was applied for 20 minutes, followed by the 20-minute stripping step in retrieval solution. After this final stripping step, Opal 780 fluor was applied for one hour. High-stringency washes were performed after the secondary, TSA-DIG, and Opal 780 incubation steps. Slides were removed from the autostainer and stained with 5 μg/mL DAPI (Sigma, Cat# D8417), rinsed, and cover-slipped with Prolong Gold Antifade reagent (Invitrogen/Life Technologies, Cat# P36930). Slides were cured at room temperature.

Antibodies.

Two antibody panels were used on two sets of TMA slides. Panel One comprised antibodies for EpCAM (BioLegend Cat# 324201, RRID:AB_756075), HLA-DR (Agilent Cat# M0746, RRID:AB_2262753), CD8 (Agilent Cat# M7103, RRID:AB_2075537), and CD3 (Thermo Fisher Scientific Cat# RM-9107-S1, RRID:AB_149924). Panel Two comprised antibodies for FAP (Abcam Cat# ab53066, RRID:AB_880077), HMGA2 (Cell Signaling Technology Cat# 5269, RRID:AB_10694917), CK17 (Abcam Cat# ab109725, RRID:AB_10889888), CD45 (Abcam Cat# ab40763, RRID:AB_726545), CK5 (Leica Biosystems Cat# NCL-L-CK5, RRID:AB_563807), and GATA-6 (Cell Signaling Technology Cat# 5851, RRID:AB_10705521). CK5, CK17, and GATA-6 antibodies were selected for their use in prior PDAC subtyping studies (16–19). See Supplementary Table S2A-B for concentrations, secondary antibodies and fluorophore assignments, and positions in mIHC sequence.

Image acquisition.

Whole slide images were acquired on the Vectra Polaris Quantitative Pathology Imaging System (Akoya Biosciences, Marlborough, MA, RRID:SCR_025508). The entire tissue area was selected for imaging with Phenochart (RRID:SCR_019156), and multispectral image tiles were acquired. Images were spectrally unmixed using Phenoptics inform software and exported as multi-image TIF files.

Immunohistochemistry.

For 30 samples from Ochsner Health, standard IHC was performed on sequential formalin-fixed paraffin-embedded tumor slices for HMGA2 and GATA6. Following standard dewaxing, antigen retrieval was performed at 95°C in 10 mM trisodium citrate buffer, pH 6.0, for 20 minutes. Slides were washed three times in 1X Tris-Buffered Saline 0.1% Tween (TBST), blocked with 10% normal goat serum (Cell Signaling Technology) in TBST, and followed by primary antibody incubation overnight at 4°C (HMGA2, 1:100 (Cell Signaling Technology Cat# 5269, RRID:AB_10694917); GATA6, 1:200 (Cell Signaling Technology Cat# 5851, RRID:AB_10705521)). Following three washes with TBST, secondary antibody (Vector Laboratories Cat# BA-1000, RRID:AB_2313606, 1:200) was diluted in blocking buffer and incubated for 45 minutes. After three washes, slides were incubated with Vectastain Elite ABC Kit (Vector Labs) prepared according to manufacturer’s instructions for 30 minutes, washed three times, then stained with DAB Substrate Kit (Vector Labs) according to manufacturer’s instructions until a positive control slide exhibited staining (~2 min for HMGA2, ~20 min for GATA6). Slides were then counterstained with hematoxylin, dehydrated, and mounted.

Image analysis.

All IHC scoring was performed blinded and verified by our collaborating pathologist (ANS). Tumor cores were considered positive for GATA6, HMGA2, CK5, or CK17 if greater than ten percent of tumor epithelial cells had positive nuclei (GATA6, HMGA2) or cytoplasm (CK5, CK17). CD8 and FAP were scored on an interval scale from zero to three, with three indicating the strongest staining.

COMPASS trial analysis.

Expression of HMGA2 was quantified using a cohort of resectable (n=176) and metastatic (n=253) patient samples from Chan-Seng-Yue et al., 2020 (13) and Perera et al., 2022 (31). RNA-seq was performed on frozen laser capture microdissected tumors on the Illumina platform. Reads were aligned using STAR v.2.7.4a (RRID:SCR_004463) to human reference genome (hg38; Ensemble v100). Gene expression was calculated in transcripts per million reads mapped using stringtie (v2.0.6, RRID:SCR_016323). High versus low expression of HMGA2 and GATA6 were determined according to median expression.

Statistical analysis.

All statistical tests were performed with GraphPad Prism v.10.3 (RRID:SCR_002798). Differences in survival were plotted following the Kaplan-Meier method and compared using the Wilcoxon rank-sum test. Statistical significance was set at p=0.05. The Kruskal-Wallis test was used for multiple group comparison. Comparisons between two groups were analyzed with the Mann-Whitney U Test (unpaired). For analysis of survival following adjuvant chemotherapy, all patients who received neoadjuvant treatment were excluded, and patients receiving at least one cycle of chemotherapy were included to assess response. Missing cores (n=8) were excluded entirely from analysis. Time to recurrence was calculated as the time from surgery to recurrence of disease.

Data and materials availability:

The tumor microarray used in this study is available through collaboration with R.C.F., D.G.D., and the Siteman Cancer Research Center’s SPORE in Pancreas Cancer. All raw tumor microarray scoring data is available through collaboration with S.K.

RESULTS

scRNAseq atlas of 172 human pancreatic tumors reveals enrichment of HMGA2 within basal tumor cells.

To determine the potential association of HMGA2 expression across tumor subtypes, we utilized a single-cell RNA sequencing atlas of the human pancreas that consists of 12 publicly available datasets compiled by Loveless et al. (28). We analyzed only the malignant ductal cells (determined through inferred copy number variation with a threshold cutoff of 0.15) from 172 primary PDAC tumors. Given that several studies have previously reported scRNAseq-derived gene signatures for classical and basal cell identities, we applied these signatures though ssGSEA scoring onto the ductal UMAP (Figure 1B). We found an enrichment of the basal gene signature from Raghavan et al. (30) (Figure 1B, top), as well as the basal 1 and 2 signatures from Chan-Seng-Yue et al. (13) (Figure 1B, middle) on the right hand side of the UMAP. The opposite pattern was observed for classical gene signatures from both studies. To examine whether the Bailey subtyping scheme was strongly influenced by its inclusion of HMGA2 as a differentially expressed gene, we also applied the Bailey signatures to our scRNAseq atlas (Figure 1B, bottom). However, the localization of squamous-subtype cells was not consistent with other studies nor with expression of HMGA2 (Figure 1C), possibly due to differences in subtyping algorithms between bulk and single-cell RNAseq data. We observed a strong association with HMGA2 expression within cells residing in the basal region of the UMAP as defined by Raghavan and Chan-Seng-Yue signatures, and this was consistent upon examination of several other basal markers including S100A2 and KRT5 (Figure 1C, left). Again, we observed the opposite pattern when we visualized GATA6, TFF2, and LGALS4, markers of a classical subtype (Figure 1C, right). We analyzed the HMGA2-positive and GATA6-positive cells and performed differential gene expression analysis on the scRNAseq ductal population. In this unbiased approach, we observed a statistically significant enrichment (adjusted p value <0.01) of basal (S100A2) and mesenchymal (FN1) genes in the HMGA2-positive cells, while classical markers TFF1, LYZ, and CTSE were highly enriched in the GATA6-positive ductal cells (Figure 1D). The cells co-expressing HMGA2 and GATA6 displayed a signature consistent with an intermediate and possibly more metabolically active population, based on expression of TXNIP and EREG, suggesting a transitory population between subtypes (17,30).

HMGA2 and GATA6 specifically stain tumor epithelium.

On our 580-patient tumor microarray, specific and strong staining of cancer cell nuclei by HMGA2 was present in 23% of patient tumors (Fig. 2A). The majority of these tumors were GATA6 negative (71.4%). GATA6 staining was observed in 51.9% of tumor samples, also specifically staining tumor cell nuclei (Fig. 2B). A subset of tumors (9%) showed staining for both GATA6 and HMGA2, albeit in distinct areas within each tumor core (Fig. 2C). A larger proportion of tumors (25.4%) were negative for both markers, possibly representing alternate phenotypes such as the aberrantly differentiated endocrine-exocrine group described by Bailey et al. or the exocrine-like group described by Collisson et al. Supporting this conclusion, HMGA2^- GATA6^- cells in our scRNAseq atlas were enriched for REG1B, CEL, INS, HR5A2, and PRSS1, which have been identified as markers of these non-basal subtypes (Supplementary Table S3) (10,11). Neither HMGA2 nor GATA6 expression was detected in any control tissues (tonsil, spleen, breast, placenta, and lung adenocarcinoma) or normal adjacent pancreas (Fig. 2D). At higher magnification, HMGA2- and GATA6-staining nuclei were seen within the same glands (Fig. 2E) and were seen to be co-staining the same tumor cell in rare cases (Fig. 2F). We then investigated an independent cohort of 30 additional untreated, age-matched patient samples from Ochsner Health and found specific and strong staining of epithelial cells with standard IHC (Supplementary Fig. S1A-B). These results establish the feasibility of applying HMGA2 in addition to GATA6 IHC for prognostic use in PDAC.

Figure 2. mIHC staining for HMGA2 and GATA6 on TMA and survival analyses.

A) HMGA2 positive (red), GATA6 negative tumor sample with paired H+E. Scale bars = 100 μm. B) GATA6 positive (green), HMGA2 negative tumor sample with paired H+E. Scale bars = 100 μm. C) HMGA2 positive (red), GATA6 positive (green) representative tumor sample with paired H+E. Scale bars = 100 μm. n=52. D) Normal adjacent pancreas is negative for HMGA2 (red) and GATA6 (green). Scale bars = 100 μm. E) HMGA2⁺ (red arrow) and GATA6⁺ (green arrow) stain separate cells in most HMGA2⁺/GATA6⁺ tumors. Scale bars = 50 μm. F) Costaining of HMGA2 and GATA6 within the same cell is seen in rare examples (white arrow). Scale bars = 50 μm. G) Expression of HMGA2 significantly predicts worse outcomes for all patients. H) GATA6^low tumors exhibit decreased survival. I) Further stratification of the patient cohort by both HMGA2 and GATA6 reveals four distinct tumor subtypes. All tumor samples shown. J) Using dual HMGA2 and GATA6 levels can also predict overall survival.

HMGA2 and GATA6 predict overall survival.

We next analyzed survival in our combined cohort based on distinct marker combinations (after excluding patients for whom clinical data was unavailable or who had died from non-PDAC causes (Supplementary Table S1, Supplementary Fig. S2A-C, Supplementary Fig. S3A)). Patients with HMGA2⁺ tumors displayed significantly worse overall survival than those with HMGA2^- tumors (Fig. 2G). There was a sizeable eight-month median survival difference based on HMGA2 expression. Loss of GATA6 expression also correlated with poorer overall survival, albeit with lower significance, validating prior findings (Fig. 2H) (13,16,19). These results validate prior work describing GATA6 loss as a modified “basal-like” phenotype, but also show that HMGA2 may be a more valuable marker for defining an overlapping but separate group of patients with basal disease. To clearly delineate samples that were primarily either basal or classical from those with a mixed or unclear subtype, we divided our cohort into four designations: HMGA2⁺/GATA6^- basal; HMGA2⁺/GATA6⁺ mixed; HMGA2^-/GATA6⁺ classical; and HMGA2^-/GATA6^- double negative (Fig. 2I). With this two-step subtyping scheme, we found that HMGA2⁺/GATA6^- tumors had the worst overall survival, while HMGA2^-/GATA6⁺ tumors had the best (Fig. 2J). The difference in median survival between these groups was 10.5 months. Both the HMGA2⁺/GATA6⁺ mixed and HMGA2^-/GATA6^- double negative groups had intermediate survivals (Supplementary Fig. S3B). Thus, we propose that dual immunohistochemistry for HMGA2 and GATA6 provides a robust and specific test to determine prognosis.

CK5 and CK17 were less predictive in our dataset.

Although CK5 and CK17 have been previously used to define the basal subtype (17,19), neither CK5 (Fig. 3A) nor CK17 (Fig. 3B) expression predicted overall survival in our cohort. Both markers exhibited specific staining in a subset of samples, albeit in a smaller proportion of samples than HMGA2, reducing the power of further analyses (Fig. 3C-F). CK5 staining was positive in 15% of patient samples, while CK17 staining was positive in only 9.6%. CK5 co-stained with HMGA2 in 64% (n=43) and GATA6 in 48% (n=32) of CK5⁺ samples, while CK17 co-stained with HMGA2 in 40% (n=17) and GATA6 in 69% (n=29) of CK17⁺ samples, indicating that these four markers define distinct but overlapping subsets of disease. Previous studies have found that basal disease, as defined by RNAseq, comprises approximately 14% of early-stage PDAC, which concurs with the proportion of CK5⁺ samples in our study group (13). However, we were surprised to find that neither CK5 nor CK17 positivity appeared to predict overall survival. We also commonly observed co-staining of both markers with supposed classical marker GATA6, contradicting previous reports (Fig. 3G-H), and possibly identifying hybrid subtype cells previously described by Williams et al. (17). Combining CK5 or CK17 positivity with loss of GATA6 was also not prognostic for overall survival (Supplementary Fig. S4A-B). Due to their cytoplasmic localization, CK5 and CK17 are more difficult to interpret than nuclear markers HMGA2 and GATA6.

Figure 3. CK5 and CK17 do not predict survival.

A) CK5 high versus low tumors do not show significant differences in survival. Scale bars = 100 μm. B) CK17 high versus low tumors do not show significant differences in survival. Scale bars = 100 μm. C) Example image of CK5 high tumor core with paired H+E. Scale bars = 100 μm. D) Example CK5 low tumor core with paired H+E. Scale bars = 100 μm. E) Example image of CK17 high (yellow) tumor core with paired H+E. Scale bars = 100 μm. F) Example CK17 low tumor core with paired H+E. Scale bars = 100 μm. G) CK5 co-stains with GATA6 (green) in a subset of cores. Scale bars = 100 μm. H) CK17 co-stains with GATA6 (green) in a subset of cores. Scale bars = 100 μm.

HMGA2 and GATA6 IHC can be used to predict overall survival.

We then turned to analyzing overall survival for patients treated with adjuvant chemotherapy in our HMGA2⁺/GATA6^- basal and HMGA2^-/GATA6⁺ classical patient cohorts from mIHC samples. Patient samples in our cohort were collected over a number of years during which standard adjuvant regimens continuously evolved. Additionally, 51 patients received neoadjuvant chemotherapy and were therefore excluded from our survival analyses due to their limited number and for the potential of neoadjuvant treatment to modify the resulting tumor subtype and immunophenotype. Our patient cohort also originates from a quaternary care center where many patients were enrolled in clinical trials, further increasing the unique chemotherapeutic regimens represented. Gemcitabine monotherapy was by far the most used agent (Supplementary Table S4). We found that HMGA2⁺/GATA6^- patients demonstrated worse overall survival with gemcitabine monotherapy compared to HMGA2^-/GATA6⁺ counterparts (Fig. 4A). This difference in survival was even more striking when we analyzed patients treated with gemcitabine followed by chemoradiotherapy (5-flurouracil plus radiation) (Fig. 4B).

Figure 4. HMGA2/GATA6 status predicts response to first-line gemcitabine-based chemotherapy.

A) HMGA2+/GATA6- tumors show worse overall survival when treated with first-line gemcitabine chemotherapy. B) HMGA2+/GATA6- tumors show worse overall survival when treated with first-line gemcitabine chemotherapy followed by 5-fluorouracil chemotherapy and radiation therapy. C) HMGA2+/GATA6- tumors have shorter time to recurrence than HMGA2-/GATA6+ tumors. D) GATA6 and HMGA2 alone do not predict time to recurrence.

Combined HMGA2 and GATA6 IHC predicts time to recurrence.

We also compared the time to recurrence in HMGA2⁺/GATA6^- basal and HMGA2^-/GATA6⁺ classical tumors in our cohort. In agreement with the more aggressive disease behavior of the basal subtype and our findings that HMGA2⁺/GATA6^- tumors were more chemoresistant and predicted shorter overall survival, we also found that HMGA2⁺/GATA6^- tumors were associated with a shorter time to disease recurrence (Fig. 4C). We also analyzed the prognostic value of HMGA2 and GATA6 individually and found neither marker alone significantly predicted time to recurrence (Fig. 4D). Combination of both markers allowed for a more obvious trend, likely due to the exclusion of tumors displaying a mixed phenotype.

CK5 and CK17 do not predict survival or time to recurrence.

We performed survival analyses on patients treated with gemcitabine or gemcitabine/5-fluorouracil stratified either by CK5 or CK17 expression. Staining for either marker did not predict overall survival (Supplementary Fig. S4C-D). Furthermore, neither CK5 or CK17 expression predicted time to recurrence (Supplementary Fig. S4E-F), and in fact CK17⁺ tumors appeared to predict a longer time to recurrence and improved survival compared to CK17^- samples.

HMGA2 and GATA6 staining associates with unique tumor microenvironments (TME).

By implementing mIHC technology on adjacent TMA slides, we were able to simultaneously detect both tumor subtype and the presence of multiple immune and stromal subsets. Of the cell types we analyzed in the TME, CD8⁺ T cells and FAP⁺ fibroblasts showed the most remarkable differences when comparing our HMGA2⁺/GATA6^- basal and HMGA2^-/GATA6⁺ classical cohorts. In addition to clearly detecting entirely basal and classical tumor samples (Fig. 5A), we noticed that classical tumors had greater CD8⁺ T cell infiltrate and basal tumors were associated with stronger FAP⁺ staining (Fig. 5B-C, Supplementary Fig. S5). These results were statistically significant (Fig. 5D-E) and indicate that HMGA2/GATA6 staining could be used as a surrogate marker for the presence or absence of cytotoxic T cells and FAP⁺ fibroblasts. We further confirmed these results by querying the scRNAseq Atlas for changes in the tumor microenvironment based on high HMGA2 tumor cell expression. Using a Mann-Whitney U test, we tested for shifts in the distribution of proportions of CD8⁺ T cells and FAP⁺ CAFs for samples with greater than 0 proportion of HMGA2⁺ ductal cells (n=124/186 and n=146/200, respectively). We observed an inverse correlation between HMGA2 expression and CD8⁺ T cell abundance (Fig. 5F). We found the opposite to be true for FAP⁺ CAFs, with more HMGA2 correlating with increased numbers of FAP⁺ CAFs (Fig. 5G).

Figure 5. HMGA2/GATA6 status predicts CD8+ T cell infiltrate and FAP+ fibroblast content.

A) Representative images of classical (top) and basal (bottom) tumor cores (green = GATA6, red = HMGA2). Scale bars = 100 μm. B) Representative images of classical (top) and basal (bottom) tumor cores stained for CD8 (pink). Scale bars = 100 μm. C) Representative images of classical (top) and basal (bottom) tumor cores stained for HMGA2 (red), GATA6 (green), and FAP (blue). Scale bars = 100 μm. D) Quantification of CD8+ scoring across all HMGA2+ GATA6- samples is significantly lower than HMGA2- GATA6+ samples (data presented as mean ± SEM). E) Quantification of FAP+ scoring across all HMGA2+ GATA6- samples is significantly higher than HMGA2- GATA6+ samples (data presented as mean ± SEM). F) scRNAseq Atlas data reveals higher CD8+ T cells in HMGA2 negative samples. G) scRNAseq Atlas data reveals higher FAP+ fibroblasts in HMGA2 positive samples.

Validation in independent cohorts.

To confirm that HMGA2 levels remained prognostically significant across independent cohorts, we probed RNAseq data from the ongoing COMPASS trial (13,16,19). In both the resectable and metastatic settings, high levels of HMGA2 predicted worse overall survival (Fig. 6A), as did low GATA6 expression (Fig. 6B). In metastatic disease, both HMGA2^high and GATA6^low tumors also predicted worse overall survival in patients who received first-line GnP (Fig. 6C, Supplementary Fig. S6A). Interestingly, HMGA2 (Fig. 6D), but not GATA6 (Supplementary Fig. S6B), was also prognostic for worse overall survival in the setting of first line mFFX. Finally, HMGA2^high status was associated with worse overall survival in patients treated with either GnP or mFFX, with mFFX outperforming GnP in both HMGA2^high and HMGA2^low groups in terms of overall survival (Fig. 6E). However, overall response rates (ORR) between HMGA2^high groups treated with GnP or mFFX were similar (Fig. 6E).

Figure 6. HMGA2/GATA6 status predict survival and treatment response across cohorts and racial backgrounds.

A-B) Kaplan-Meier survival analysis of metastatic laser-capture microdissected RNAseq from the COMPASS trial demonstrates worse overall survival in HMGA2^high (A) and GATA6^low (B) tumors. C) HMGA2^high predicts poorer response to gemcitabine/nab-paclitaxel chemotherapy. D) HMGA2^high tumors show poorer response to modified FOLFIRINOX chemotherapy. E) When stratified by median HMGA2 expression, HMGA2^high metastatic patients from the COMPASS trial show poorer overall response to both GnP and mFFX chemotherapies when compared to HMGA2^low tumors. Patients with HMGA2^high tumors have improved overall survival when treated with mFFX, but overall response rate remains similar in both regimens. F) Proportion of major subtype classes in Black patients does not significantly differ from the overall population on our TMA (see Fig. 3C). G) HMGA2/GATA6 staining predicts overall survival in Black patients on our TMA.

Validation in Black patients.

Given the disproportionate disease burden of PDAC among Black patients (32–35), we also sought to determine whether HMGA2/GATA6 expression patterns were also prognostic in this population. 56 of the samples (9.7%) represented in our TMA were from Black patients and, when combined with additional samples, we were able to robustly subtype 69 (11.5%) tumors from Black patients. This represents a larger proportion of Black patients than have been included in previous subtyping studies (10–13) (Supplementary Tables S1 and S5). Of 892 patient samples that underwent RNA sequencing across prior studies, only 30 patients (3.4%) were Black (10–13). In our study, cores from Black patients showed no significant differences in the relative proportions of basal/classical PDAC subtypes, as determined by HMGA2/GATA6 co-staining, compared to non-Black patients (Fig. 6F). Furthermore, HMGA2/GATA6 status was again prognostic for overall survival (Fig. 6G).

DISCUSSION

The choice of chemotherapy for patients with PDAC is currently left to the physician’s discretion and the patient’s overall clinical status, and there are no clinical biomarkers that can predict more aggressive disease. Here, we build upon the development of GATA6 IHC as a marker for classical disease and improved outcomes by adding specific IHC for HMGA2. We demonstrate that HMGA2⁺/GATA6^- tumors have the worst overall survival, have poorer survival when treated with gemcitabine-based chemotherapy compared to HMGA2^-/GATA6⁺ tumors, and are associated with faster recurrence than HMGA2^-/GATA6⁺ tumors. Collisson et al. have previously reported that quasi-mesenchymal cell lines show greater sensitivity to gemcitabine in vitro than classical cell lines (11). However, our findings in clinical samples indicate a poorer response in HMGA2^high tumors to gemcitabine chemotherapy compared to HMGA2^low tumors. This is perhaps due to the complex tumor microenvironment, the notorious fallibility of in vitro studies to predict in vivo behavior, and the poor representation of complex chemoresistance in 2D cell culture (36). Our data are supported by analyses from the ongoing COMPASS trial with GnP chemotherapy. Furthermore, HMGA2^high status predicts poorer overall survival in patients treated with either mFFX or GnP compared to HMGA2^low status (Fig. 6E).

Although RNA sequencing remains the gold standard for subtyping PDAC, it remains expensive and requires high sample quality and purity. This limits its current applications to well-resourced clinical environments. Surrogates for RNAseq have been proposed, including single-cell immunofluorescence and algorithmic analyses of histologic sections, but these techniques can also be expensive and time-consuming. Dual HMGA2/GATA6 IHC is advantageous because of faster results, feasibility in modern pathology labs, and tolerance of lower sample quality. It is also capable of detecting mixed phenotypes within the same tumor core, duct, or even the same cell. By adding HMGA2 expression as a positive marker for the basal state, we further refine subtyping analyses that previously relied on loss of GATA6 as a surrogate marker for basal disease (16,19,37). The classical RNAseq signature has previously predicted improved overall response to mFFX (37), but in our work GATA6 levels by IHC alone did not.

Our data also strongly suggest that HMGA2 is superior to CK5 and CK17 as a positive marker of basal disease. HMGA2 and GATA6 both exhibit nuclear expression, which makes determining the percentage of basal cells in the tumor significantly easier than using markers with cytoplasmic distribution. It is also easier to detect double positive populations and distinguish adjacent basal and classical cells from one cell with characteristics of both subtypes. Finally, and most importantly, HMGA2 was superior to CK5 and CK17 for predicting prognosis. Neither CK5 nor CK17 predicted poorer overall survival in our patient cohort, even though they have been reported as prognostic markers for the basal subtype in other studies (16,17). Recently, Delgado-Coka et al. have reported that CK17 staining in a cohort of 379 patient samples, also primarily from early-stage disease (80%), predicts poorer overall survival and resistance to GnP and mFFOX chemotherapy (38). Interestingly, the authors report a significantly larger proportion of CK17⁺ samples (73%) in their study group than we found on our TMA (9.6%) (38). However, these results are difficult to directly compare due to the inclusion of neoadjuvant treated samples in Delgado-Coka et al. (38). Both our study and that of Delgado-Coka et al. found that basal tumors, whether defined by CK17 or HMGA2 expression, showed improved overall survival when treated with mFFX or 5-FU-based chemotherapy regimens compared to GnP (38). CK17^high cells have also been implicated in driving resistance to neoadjuvant mFFX (39). In our study, HMGA2, but not CK5 or CK17, could be used to predict prognosis following chemotherapy and time to recurrence in the setting of early-stage disease. These results also were validated independently in a large cohort of metastatic PDAC. These findings support the use of HMGA2 to identify patients with the most aggressive subtypes of PDAC.

Our findings associating the IHC subtype with tumor microenvironment raise key questions that warrant further study. Recent studies have associated the expression of CK17 with low intratumoral CD8⁺ T cells (40). Here, we validate this association and found that the classical subtype was associated with greater cytotoxic T-cell infiltrate, a well-known prognostic marker for longer survival (41). The functional state and level of exhaustion in these T cells has yet to be characterized, and scRNAseq studies are warranted to fully elucidate the specific differences in CD8⁺ T cells between basal and classical tumors. FAP⁺ fibroblasts are a prevalent stromal cell type in the PDAC microenvironment and have been regularly identified as a poor prognostic marker in both murine and human PDAC (42–45). To our knowledge, differences in fibroblast content between basal and classical tumors have not been reported. The clinical relevance of these findings remains unclear, as fibroblast content is not as clearly associated with survival as CD8⁺ T-cell tumor infiltration. Nonetheless, these findings are exciting for their potential use to identify unique immunophenotypes present in basal versus classical tumors that could be therapeutically targeted.

This study is limited by few samples from locally advanced or metastatic disease, as these are the most common clinical presentations of PDAC. The relatively early disease represented in our cohort may explain the small percentage of CK5- or CK17-staining tumors. We also found a greater proportion of HMGA2⁺ tumors (23%) than previously reported proportions of basal disease (14%) (13). These findings may indicate a larger proportion of basal disease in our study group or that HMGA2 reactivation may precede the full shift to basal disease. Further study regarding the mechanisms by which HMGA2 may drive basal disease is warranted.

Although we have validated our findings in a cohort of patients with advanced disease by RNAseq, we recognize that our primary findings are based on naïve tumor samples that have never been exposed to chemotherapy. As neoadjuvant therapy becomes more common in the hopes of achieving resectability and early eradication of micrometastases, the cohort we studied may represent a shrinking fraction of patients as treatment standards continue to evolve (46–48). However, HMGA2 and GATA6 IHC can still be used to predict outcome, especially as chemotherapy may skew tumors to a more basal subtype (49). Although our samples came from surgically resected specimens, we are also able to assay core biopsies obtained from primary tumor or metastatic sites. HMGA2/GATA6 IHC also holds exciting potential to track subtype evolution over time. As this strategy is refined and further validated, it may be possible to evaluate HMGA2 and GATA6 levels with less invasive methods, such as from circulating tumor cells or cell-free DNA.

Our study is also limited in its application to patients of non-White racial backgrounds. Prior studies have evaluated the role of access to medical care, treatment facilities, and overall differences in PDAC prevalence between Black and non-Hispanic White patients (32–35). Black patients continue to be under-represented in cutting-edge research models and clinical research trials. To our knowledge, our cohort of 69 Black patients represents the largest subtyping study performed on non-White PDAC samples. To further confirm the clinical utility and prognostic viability of a biomarker, it is crucial to test it in a diverse patient population. We were encouraged that HMGA2/GATA6 staining appears to be prognostic in both White and Black patient populations, but we are actively undertaking further multicenter studies and collaborations to confirm these associations in Black, Asian, Alaska Native, and other under-represented patient groups. Furthermore, we hope that HMGA2/GATA6 dual IHC will serve as an accurate and useful prospective biomarker in clinical trials.

Statement of Translational Relevance:

Transcriptional subtyping has been broadly used to define subsets of pancreatic ductal adenocarcinoma (PDAC) but remains cost-prohibitive and time consuming to perform on all patient samples. Biomarkers for the basal and classical subtypes have not yet entered clinical use, and current markers of the basal subtype do not uniformly detect basal disease. We have found that HMGA2 expression closely associates with the basal subtype of PDAC in both scRNAseq and IHC analyses. In a cohort of resected tumors, we found HMGA2 IHC was superior to that for CK5 and CK17 in predicting overall survival and response to chemotherapy. In addition, HMGA2 staining can be combined with GATA6 staining to distinguish basal and classical subtypes, define critical features of the tumor microenvironment, and anticipate time to recurrence. These data support the development of HMGA2 and GATA6 as a signature of transcriptional subtypes and prognosis in PDAC.