The Proteomic Landscape of Pancreatic Ductal Adenocarcinoma Liver Metastases Identifies Molecular Subtypes and Associations with Clinical Response

This article is featured in Highlights of This Issue, p. 979

Pancreatic ductal adenocarcinoma (PDAC) is among the most lethal of cancers, with a 5-year survival rate of 8.5% and a cancer mortality rate projected to outpace both breast and colon cancer in the coming years (1). The poor survival of patients with PDAC is associated with the highly metastatic nature of this disease. Approximately 80% of these patients develop liver metastases, but other common sites include the lung and peritoneum, with multiple organ involvement often observed during the end stages of this disease (2). The contribution of disseminated disease to lethality in PDAC is exemplified by the fact that, among patients with early-stage and resected lesions, 60%–70% will present with metastatic lesions within 5 years of resection (3, 4). While the analysis of the primary tumor facilitates our understanding of the molecular etiology of PDAC (5), characterization of metastatic lesions has the potential to improve clinical interventions that address the main cause of cancer mortality. Thus, understanding the underlying molecular features of metastatic PDAC is necessary to develop effective therapeutic interventions that improve patient survival.

Identifying cancer subtypes has the potential to improve patient outcomes because subtype can be associated with treatment response. This has been most effectively employed for breast cancer (6), where gene expression profiling using the PAM50 gene expression predictor outperforms IHC classification methods for its prognostic and predictive ability (7). Several studies have characterized primary PDAC subtypes based on transcriptional profiling (5, 8–11). Using laser capture microdissection (LCM), Collisson and colleagues classified PDAC tumors into three subtypes (exocrine-like, classical, and quasi-mesenchymal; ref. 8). Moffitt and colleagues classified PDAC into two tumor subtypes (classical and basal-like) and two stroma subtypes (normal and activated) using virtual microdissection (9). Using transcriptional profiles of intact whole-tumor tissues, Bailey and colleagues classified four subtypes of PDAC (squamous, immunogenic, pancreatic progenitor, and ADEX), in which the squamous subtype is associated with significantly shorter survival than the other subtypes (10). Puleo and colleagues also evaluated the transcriptome of PDAC formalin-fixed paraffin-embedded (FFPE) and identified basal-like and classical subtypes (11). The consensus across these studies is that there are two predominant PDAC tumor cell subtypes: classical and squamous, following the proposed harmonized nomenclature (12). To date, PDAC subtyping is based on RNA expression profiles, but orthogonal methodologies, such as proteomics, have the potential to reveal additional features that could improve the functional characterization of PDAC subtypes.

There have been comparative proteomics studies on PDAC that have identified proteins from tissue, plasma, pancreatic juice, cyst fluid, and urine associated with this disease. These efforts illustrate the potential of applying proteomics approaches to improve early detection and treatment of PDAC based on single proteins (13, 14). IHC evaluating expressions of only KRT81 and HNF1A proteins has been used to stratify PDAC tumors as either classical, quasi-mesenchymal, or exocrine-like (15). However, prior to this study, proteomics-based approaches have not been performed at a scale that supports PDAC subtype classification based on proteome quantification. Further exploration of the PDAC tumor proteome along with detailed clinical records could improve the diagnosis and treatment of this cancer.

The overarching goal of this project was to evaluate the proteome of PDAC liver metastases to distinguish unique subtypes from clinical samples. As a proof-of-principle that PDAC microenvironment subtypes can be delineated using proteomics, we have developed and validated a classification system using quantitative proteomics data from 68 tissue samples in total from the rapid autopsy program (RAP) managed by the UNMC Pancreas SPORE. This proteomics analysis identified over 3,960 proteins and quantitative profiling of the 916 of these proteins was used to delineate four distinct subtypes of PDAC liver metastases, which share many molecular signatures with transcriptomic-derived subtypes. The proposed proteomic-based subtyping system showed a significant association with patients' alcohol and tobacco exposure. In addition, a survival advantage was observed in the metabolic and progenitor-like subtypes treated with FOLFIRINOX (5-fluorouracil, leucovorin, irinotecan, and oxaliplatin) and gemcitabine compared with gemcitabine, but this survival benefit was not observed in the inflammatory and proliferative subtypes. The serine hydroxymethyltransferase (SHMT1), a metabolic enzyme involved in single carbon metabolism, was identified as a mediator of gemcitabine resistance. In addition, 52 protein expression profiles were found to correlate with patients' survival in a subtype-specific manner. These data demonstrate the clinical relevance of this proteomics classification model and illustrate its potential for the development of therapeutic strategies to target PDAC liver metastases.

Investigators obtained informed consent for each patient enrolled in the UNMC Rapid Autopsy Program (IRB #091-01). This study was conducted in accordance with the ethical guidelines established by the Declaration of Helsinki.

The frozen PDAC tissues and the corresponding tumor-adjacent tissues were available from the RAP in UNMC. For each of the samples, 5 mg of the frozen tissue was ground into a fine powder with a liquid nitrogen-cooled mortar and pestle. The ground tissue was then lysed with 1 mL of RIPA buffer (25 mmol/L Tris-HCl pH 7.6, 150 mmol/L NaCl, 1% NP-40, 1% sodium deoxycholate, 0.1% SDS) and was frozen in −80°C until further used. The albumin and IgG contents in the protein lysates were first depleted with the Pierce Top 2 Abundant Protein Depletion Spin Columns and labeled with TMT reagents per the manufacturer's instructions. Detailed procedures were listed in the Supporting Information.

The mass spectrometry data was acquired on a Dionex Nano Ultimate 3000 coupled with an Orbitrap Fusion Lumos. The fractions collected from the high-pH separation were resuspended in 20 μL of 0.1% formic acid. Two microliters of each fraction was injected into the system for tandem mass spectrometry analysis. The MS and MSⁿ spectra collected from the experiment were searched against the homo sapiens protein sequence database (downloaded in 10/2017, 42252 entries) and the respective decoy database with Sequest HT in the Proteome Discoverer 2.2 pipeline. The reporter ion ratios of these proteins were exported from the Proteome Discoverer and the P values were calculated with the Wilcoxon-signed rank test using R. The software packages used in the postdatabase search analysis are listed in the Supporting Information. Mass spectrometry data files have been deposited to the ProteomeXchange Consortium via the PRIDE (16) partner repository with the project accessions: PXD012173 and PXD015492.

The human PDAC cell lines MIA PaCa-2 and Panc 10.05 was obtained from the ATCC and were cultured per the manufacturer's instructions. Briefly, the cells were grown in DMEM supplemented with 10% FBS, 2.5% horse serum, amphotericin B, and penicillin–streptomycin (Corning) in 5% CO₂ atmosphere at 37 °C. Gemcitabine-conditioned MIA PaCa-2 cells were generated by incubation with 10 nmol/L gemcitabine (Selleckchem) freshly diluted in DMSO for 6 days. The cell lysates were collected with RIPA buffer (25 mmol/L Tris-HCl pH 7.6, 150 mmol/L NaCl, 1% NP-40, 1% sodium deoxycholate, 0.1% SDS) and probed for MTHFD1 and SHMT1 by Western blotting (detailed conditions in Supporting Information).

Two unique shRNA constructs targeting SHMT1 (TRCN000034766 and TRCN000034767; shSHMT1 #1 and #2) and a nontargeting scrambled control (shScr) in pLKO.1 were used in this study (Sigma). The lentiviral supernatant was produced by calcium phosphate transfection into 293FT cells, as described previously (17), and used to transduce MIA PaCa-2 and PANC 10.05 cells. The transduced cells were selected with puromycin for 5 days before use in cytotoxicity assays. Knockdown of SHMT1 was confirmed by Western blotting. For the gemcitabine cytotoxicity assay, 1,000 cells were aliquoted into each well in a 96-well plate. The cells were incubated with gemcitabine at different concentrations for 3 days. Cell viability was determined by CellTiter-Glo (Promega), and the luminescent signal was measured by FLUOstar Optima (BMG Labtech). EC₅₀(s) were estimated with GraphPad Prism 7 from the exported data (GraphPad).

This project aimed to explore proteomic variance in PDAC liver metastases from patient samples collected by the UNMC Pancreas SPORE Rapid Autopsy Program. This Program ensures all samples are collected at the same stage of disease under a standardized procedure (18). The metastatic tissue proteome was determined from a cohort of 59 patients [56 PDAC, 3 pancreatic neuroendocrine tumors (PanNET)] that were annotated for tumor stage at diagnosis, gender, age, overall survival (OS) calculated from the day of diagnosis, metastatic involvement, and PDAC risk factors of alcohol and tobacco consumption (Table 1; Supplementary Table S1). These samples were randomly divided into seven batches and differentially labeled with isotopic tags using the 10-plex TMT kit (Fig. 1A; Supplementary Table S2). The reference mix of all 59 samples was tagged with the TMT126 label, which serves as a common reference for quantitation across all seven batches (Supplementary Table S2). The overall analysis identified 30,811 peptides mapping to 3,960 proteins in which 1,842 were quantified and 916 were quantified with at least 5 peptides across 80% of the samples (Fig. 1B; Supplementary Tables S3 and S4). The set of 916 proteins were used in the multivariate analysis (Fig. 1B; Supplementary Table S5). These 916 proteins represent a broad array of cellular functions across the proteome, including extracellular matrix organization, protein processing and transport, translation, glycolytic processes, NADPH metabolism, cell migration, immune response, fibronectin binding, and cell homeostasis determined by Spatial Analysis of Functional Enrichment (SAFE) (Supplementary Fig. S1; Supplementary Table S6; ref. 19).

Partial least squares-discriminant analysis (PLS-DA) effectively distinguished PDAC from PanNET liver metastases with high confidence (Fisher probability = 3.1 × 10⁻⁵; Fig. 1C). We further tested the methodology by comparing the proteomic signatures of 9 PDAC liver metastases against matched tumor-adjacent uninvolved liver. Principal component analysis (PCA) effectively separated the liver metastases and the uninvolved liver tissue into two distinct categories (Fig. 1D). All pairs of liver metastases and tumor-adjacent tissue proteomes were well separated on the score plot (Supplementary Fig. S2A). In the corresponding PLS-DA model, the ROC curves differentiated both PDAC and PanNET tumors from the corresponding tumor-adjacent liver tissues with high sensitivity and specificity [area under the curve, AUC(PDAC) = 1; AUC(PanNET) = 1; Supplementary Fig. S2B and C]. These analyses demonstrate that the underlying quantitative proteomics data provide sufficient sensitivity and specificity to differentiate two distinct types of pancreatic cancer metastases to the liver as well as distinguish tumor tissue from the adjacent uninvolved liver.

To explore the variance in liver metastases proteome, we constructed a PCA model with hierarchical clustering using the 916 quantified proteins (Fig. 1E; Supplementary Table S5). The sum of squared error analysis revealed that the intragroup variance was best explained when the samples were divided into four major subtypes and three protein clusters (Supplementary Fig. S3A and S3B). The four protein subtypes identified were as follows: (i) metabolic (n = 7); (ii) progenitor-like (n = 21); (iii) proliferative (n = 11); and (iv) inflammatory (n = 17; Fig. 1E). Subtype nomenclature is based on protein enrichments for each subtype or their relation to previously described transcriptional subtypes. The robustness of the model was evaluated by the 7-fold cross-validation operation built-in in SIMCA 15. The goodness of fit (R²X) and the predictability (Q²) of this unsupervised PCA model were 0.63 and 0.42, which was comparable with other PCA models in the literature (20, 21). The coefficients for each of the proteins in the corresponding supervised PLS-DA model were listed in Supplementary Table S7. The χ² analysis showed that batch effects did not impact subtype classifications (Supplementary Fig. S3C).

We identified significant correlations between the subtypes identified by proteomics and those identified by Moffitt and colleagues [χ² test, P (tumor) = 8.92 × 10⁻⁶, P (stroma) = 2.06 × 10⁻³], Collisson and colleagues (P = 2.10 × 10⁻⁵), and Bailey and colleagues (P = 8.85 × 10⁻⁸; Fig. 2A; Supplementary Fig. S4A–S4D). In comparison with the Moffitt and colleagues classification system, three representative proteins with the highest weight factor (ALDH2, IDH1, and TST) associated with the classical tumor signature exhibit higher expression in the metabolic and the progenitor-like subtypes than the other two proteomic subtypes, while the inflammatory subtype showed a high expression of basal-like tumor signature genes (ANXA1, ANXA3, and ITGA2), resulting in higher signature scores (Fig. 2B and C). There were significant differences between the expression of extracellular proteins like gelsolin (GSN) and lumican (LUM) between the proliferative and the inflammatory subtypes that only resulted in a slight difference between the signature scores of the normal and the activated stroma subtype between these two proteomic subtypes (Fig. 2B and D). This demonstrates the proteomic PDAC subtyping method incorporates extracellular protein expression, which may not be captured by transcriptomic approaches.

There exists a debate whether the exocrine-like and ADEX classifications represent unique subtypes or contamination from acinar cells present in the tumor microenvironment (5, 12). However, the existence of patient-derived cell lines and propagated xenografts that are classified as exocrine-like suggest they may represent unique subtypes (15, 22). Furthermore, different cellular components of the PDAC tumor microenvironment affect a range of cancer phenotypes including metastatic potential and treatment efficacy and contribute to the overall tumor microenvironment (23, 24). Like our study, Bailey and colleagues also used the whole tumor for their analysis without virtual or physical microdissection. Therefore, we chose to include the cross-comparison to the Bailey and colleagues classification system. The proliferative and inflammatory subtypes share signatures associated with the squamous/quasi-mesenchymal subtype (Fig. 2A, E, and F). However, our proteomic classification system can further subcategorize the squamous subtype into two distinct subtypes (inflammatory and proliferative). The progenitor-like subtype nomenclature was used because of the similarity to the Bailey and colleagues progenitor subtype. There is also an association between the metabolic and the ADEX/exocrine-like subtypes, which would not be attributed to acinar cell contamination because this analysis used liver metastases. The metabolic association with the ADEX subtype (Bailey and colleagues) may provide further support to an exocrine-like subtype of PDAC, but we cannot rule out the possibility that it is a byproduct of signals from normal liver cells. However, liver tissue adjacent to metabolic tumors exhibits unique protein expression signatures that can differentiate these tissues (Figs. 1D; Supplementary Fig. S2).

The consistencies between our proteomics-based and the transcriptomic-based classification systems suggest that many of the same transcriptional signatures found in primary tumors can be found by proteomics in the liver metastases. This is supported by a recent study that determined RNA signatures from metastatic tissue obtained from a variety of anatomic sites can be used to identify PDAC subtypes (25). To confirm this, we performed quantitative proteomics on nine matched primary tumor samples to evaluate PDAC subtype conservation with metastases. The PLS-DA model had R²X, R²Y, and Q² of 0.695, 0.99, and 0.401, respectively (Supplementary Fig. S5). This indicates PDAC molecular subtypes were generally conserved between the primary tumor and the liver metastases in matched samples.

The protein expression profiles observed in PDAC liver metastases distinguish three groups of similarly expressed proteins (protein cluster 1–3) associated with a range of biological processes (Supplementary Fig. S6A and S6B). The metabolic and progenitor-like subtypes are characterized by metabolism-related proteins in protein cluster 1, enriched with proteins in the ethanol oxidation pathways, mitochondrial fatty acid β-oxidation, and retinoic acid signaling pathways (Supplementary Fig. S6C; Supplementary Fig. S7; Supplementary Table S8; Supplementary Information). Even though both the metabolic and the progenitor-like subtypes were characterized by protein cluster 1, the metabolic subtype exhibits higher expression of these proteins, such as those associated with signaling by retinoic acid (Supplementary Fig. S6D). The proliferative subtype proteome is enriched with ribonucleoproteins and Cajal body proteins in protein cluster 2 that are associated with translation, cell proliferation, and telomere maintenance in cancer cells (Supplementary Fig. S6C and S6E; Supplementary Tables S8 and S9; Supplementary Information; ref. 26). The inflammatory subtype is characterized by protein cluster 3 and is enriched for proteins related to pentose phosphate pathway, adaptive immune response, complement activation, IL8 production, and extracellular fibril organization (Supplementary Figs. S6C, S6F, and S6G; Supplementary Tables S9 and S10; Supplementary Information). These pathways and processes are known to create an immunosuppressive and chemoresistant environment that supports tumor growth (27–29). The gene set enrichment analysis on the average protein expression of each subtype in the Reactome pathways also agreed with the biology described above (Supplementary Table S11). Together, these data demonstrate that coexpressed proteins participate in cancer-associated pathways that are differentially represented across PDAC subtypes identified by proteomics.

Health and lifestyle data are collected for each RAP donor, including diabetes, alcohol use, and tobacco use, which are risk factors for the development of PDAC (30). There are no significant differences in the neoplastic cellularity across the four proteomics subtypes in the tumors analyzed by pathologic review (Fig. 3A), suggesting this variable did not affect subtype classification. In addition, the incidence of diabetes is distributed at expected rates across each of the four proteomics-defined subtypes and does not appear to influence subtype classification (Fig. 3B). However, there are significantly more patients than expected with the proliferative subtype that reported a history of alcohol use (hypergeometric test, P = 0.025), and there are no patients with the inflammatory subtype that reported alcohol use (P = 0.002; Fig. 3C). With regard to tobacco, the metabolic subtype includes more patients who report tobacco use than expected (P = 0.009), while the inflammatory subtype included significantly fewer than expected tobacco users (P = 0.014; Fig. 3D).

We grouped the proliferative and the inflammatory subtypes because combined they correspond to the squamous subtype in the Bailey and colleagues classification system (Fig. 2A), which was demonstrated to be associated with shorter survival than the other transcriptionally defined subtypes in that study (10). Because the patients in the RAP cohort analyzed in this study were not treated with a standard set of chemotherapeutics, we focused our evaluation of patient survival on three treatment groups: (i) untreated: did not receive either gemcitabine or FOLFIRINOX; (ii) gemcitabine: received at least gemcitabine; and (iii) FOLFIRINOX + Gem: received FOLFIRINOX followed by at least gemcitabine. OS is calculated from the day of diagnosis. Across all subtypes, OS for gemcitabine-treated patients was 271.5 days and 336 days for FOLFIRINOX + Gem (HR, 2.18; 95% confidence interval (CI), 1.57–3.01; P = 0.02; Fig. 3E and F). The OS for patients classified with proliferative and inflammatory subtypes treated with gemcitabine was 258 days and 288 days for FOLFIRINOX + Gem (HR, 1.57; 95% CI, 1.02–2.41; P = 0.29; Fig. 3G and H). However, OS for patients classified with the metabolic and progenitor-like subtypes treated with gemcitabine was 286.5 days and 401.5 days for FOLFIRINOX + Gem (HR, 3.37; 95% CI, 1.02–5.90; P = 0.03; Fig. 3I and J). These data indicate that the metabolic and progenitor-like subtypes, but not the proliferative and inflammatory subtypes, display a decreased risk of death when FOLFIRINOX is given in addition to gemcitabine as part of the treatment course. These results support the concept of PDAC subtype–specific response to therapy.

Patient survival based on treatment(s) with gemcitabine, FOLFIRINOX/FOLFOX, abraxane/paclitaxel, tarceva/erlotinib, and radiation in each PDAC subtype was also evaluated (Supplementary Fig. S10). Patients could be grouped in multiple treatments because their inclusion was dictated by whether they received the indicated therapy or not, and patient treatments were variable. This analysis indicated that patients with the progenitor-like subtype treated with gemcitabine have a significant increase in survival compared with cases that do not receive this treatment (P = 0.001). Notably, a significant increase in survival probability was not observed in the progenitor-like subtype for any of the other treatments evaluated. In addition, the trends observed in the metabolic subtype following abraxane/paclitaxel treatment suggest a negative correlation between receiving this treatment and survival probability (P = 0.008; Supplementary Fig. S10). It will be important to evaluate patient outcomes in response to individual therapies stratified by subtype in a much larger dataset to identify significant trends that would support the implementation of personalized treatments based on PDAC subtypes.

Protein expression patterns in PDAC liver metastases also have the potential to reveal new associations with clinical outcomes or identify novel therapeutic targets (14, 31). With the understanding that PDAC is not a singular disease, we hypothesized that associations between protein expression and patient survival might display subtype-specific characteristics. Partial least squares (PLS) analysis was first used to evaluate the association between the number of survival days after diagnosis with protein expression from then entire 56-patient cohort used in this study. This identified 52 proteins associated with either increased or decreased survival probability (Fig. 4A). The Kaplan–Meier survival curves were plotted using the upper and lower tertiles (Supplementary Table S12), and the log-rank test was used to determine significant differences in survival probability between the two groups. Proteins with P ≤ 0.05 were cross-referenced with the PLS survival model as internal validation. A total of 32 proteins with elevated expression demonstrated a significant association with increased survival (Supplementary Fig. S11). An additional 20 proteins demonstrated an inverse correlation between expression and survival probability (Supplementary Fig. S12). Some of these proteins include thymidine phosphorylase (TYMP; Fig. 4B), programmed cell death 6-interacting protein (PDCD6IP; Fig. 4C), stathmin 1 (STMN1; Fig. 4D), and endoplasmic reticulum aminopeptidase 1 (ERAP1; Fig. 4E), which are known to be associated with cancer phenotypes (details discussed in Supplementary Information; Supplementary Fig. S14; refs. 32–39).

We found that many of these 52 proteins display subtype-specific expression patterns (Fig. 4F). Because the proteomics subtype model (R²X = 65%) better explained the variance observed in the proteome than the survival prediction model (R²X = 48%), we hypothesized that a better survival regression model could be built based on individual subtypes. Figure 4G–J shows the loading plot of the same 52 proteins in Fig. 4A for the individual subtype survival regression model. The variance captured by the regressed survival time (Q²Y) increased from 24% to 38%–84%. TYMP is highly correlated with survival in the progenitor-like subtype (Spearman correlation coefficient, r = 0.62), but this correlation is not as prominent in the other subtypes (r = 0.14–0.33, Fig. 4K). PDCD6IP and STMN1 showed a significant correlation with survival in the metabolic (r = 0.82) and the progenitor-like (r = −0.55) subtypes, respectively (Fig. 4L and M). The expression of ERAP1 inversely correlated with survival in the metabolic, progenitor-like, and proliferative subtypes (r = −0.49 to −0.33), but not in the inflammatory subtype (r = −0.04, Fig. 4N). These results demonstrate that subtype-based regression models are better suited for identifying proteins associated with patient survival.

To validate the proteomic differences observed in our dataset are biologically meaningful, we evaluated the changes in the proteome associated with gemcitabine treatment. PDAC is a refractory cancer that readily develops resistance to chemotherapy, including gemcitabine (40, 41). With death as the common endpoint, it is assumed that protein expression changes associated with gemcitabine resistance could be distinguished in the proteomic data from the RAP samples. Among the cohort of 56 PDAC donor samples, there are 6 donors that were treatment-naïve and 9 donors treated only with gemcitabine. A total of 63 proteins are upregulated and 44 proteins are downregulated significantly in the gemcitabine treatment group compared with the treatment-naïve group (Fig. 5A; Supplementary Fig. S15A; and Supplementary Table S13). These differentially regulated proteins influenced by gemcitabine treatment have the potential to identify PDAC mechanisms of drug resistance.

The folic acid cycle proteins cytosolic C-1-tetrahydrofolate synthase (MTHFD1) and SHMT1 were significantly downregulated in PDAC liver metastases from patients treated with gemcitabine compared with treatment-naïve samples (Fig. 5B). MTHFD1 catalyzes the hydrolysis of 5,10-methenyltetrahydrofolate into 10-formyltetrahydrofolate, while SHMT1 catalyzes the conversion of tetrahydrofolate (THF) and the amino acid serine into 5,10-methenyltetrahydrofolate, the substrate required by MTHFD1 (Fig. 5C). The regulation of nucleotide pools, such as dCTP, is a mechanism of gemcitabine resistance in PDAC cell lines (40). The folic acid cycle generates and recycles the metabolites required for the conversion of deoxyuridine monophosphate (dUMP) to deoxythymidine monophosphate (dTMP), necessary to support DNA synthesis. Thymidylate synthase (TYMS) converts dUMP to dTMP and is inhibited by both 5-fluorouracil and gemcitabine metabolites leading to defects in DNA replication. In addition, TYMS expression is correlated with gemcitabine resistance (42). This evidence suggests gemcitabine-mediated regulation of the folic acid pathway is important for the development of drug resistance.

Both MTHFD1 and SHMT1 are downregulated in MIA PaCa-2 PDAC cells conditioned with 10 nmol/L gemcitabine for 6 days (Fig. 5D and E). To determine whether inhibition of SHMT1 expression is associated with PDAC response to gemcitabine, we established stable MIA PaCa-2 and Panc 10.05 cell lines with targeted knockdown of SHMT1 using two different shRNA constructs (Supplementary Fig. S15B and S15C). In MIA PaCa-2 cells, the EC₅₀ of gemcitabine increased from 2.7 nmol/L in the control cells to 17.9–19.9 nmol/L in the SHMT1 knockdown cells (Fig. 5F; Supplementary Fig. S15D). Similarly, Panc 10.05 displayed an increase in the gemcitabine EC₅₀ from 1.6 mmol/L in control to 9.4 mmol/L and 7.3 mmol/L in the SHMT1 knockdown cells (Fig. 5G; Supplementary Fig. S15E), suggesting that the reduced expression of this protein observed in gemcitabine-treated patients could act as a mechanism of drug resistance. Because the depletion of SHMT1 could prevent DNA synthesis by restricting dTMP pools, we evaluated the cell-cycle profile and observed a significant increase in S-phase and a decrease in G₂–M in SHMT1 knockdown cells compared with control (Fig. 5H; Supplementary Fig. S16). Overall SHMT1 expression, regardless of gemcitabine treatment, is higher in the metabolic and progenitor-like subtypes compared with proliferative or inflammatory subtypes (Fig. 5I). This may indicate that individual PDAC subtypes could be more resistant to gemcitabine based on the inherent expression levels of the enzymes regulating the folic acid cycle.

Molecular subtyping of cancer can improve therapeutic outcomes by stratifying distinct subtypes of cancer into treatment groups based on their predicted response characteristics. There have been several different approaches to subtype PDAC using transcriptomics that prove this is not a singular disease and that specific subtypes may exhibit unique response profiles to therapies. Furthermore, because of the metastatic nature of this disease clinical subtyping of PDAC should incorporate metastatic characterization to address the primary cause of patient mortality. This study represents the first proteomics-based subtype classification system for PDAC using liver metastases that could provide the basis for improving clinical therapy of this disease.

The high-dimensional data obtained in this proteomics study provides the ability to discern both complementary and unique PDAC subtypes. Previous PDAC proteomics studies have not been amenable in comparison with established PDAC subtypes due to the small number of tumors analyzed or the limited repertoire of proteins identified for analysis (43). Therefore, this is the first proteomics analysis of clinical PDAC samples that overcomes the limitations of previous studies to support a cross-comparison with transcriptomic-based approaches for determining PDAC subtypes. The comparison of the proteomics-based subtypes with transcriptomic-based subtyping efforts by Moffitt and colleagues, Collisson and colleagues, and Bailey and colleagues identifies significant concordance between these studies. However, the additional stratification of the squamous subtype into both the proliferative and inflammatory subtypes suggests proteomics could be used as a complementary method to identify additional PDAC subtypes. Because the squamous subtype has been defined as a more aggressive PDAC tumor, further stratification based on proteomics has the potential to identify subtype-specific features that could impact clinical response.

In our analysis, multiple predictive models were built to correlate the proteomics data with the clinical metadata, including gender, age, stage at diagnosis, number of primary or metastatic sites, history of diabetes, and treatment(s) administered. These models showed no significant correlation with the PDAC subtypes. However, PDAC subtypes exhibit a significant correlation with reported PDAC modifiable risk factors of alcohol and tobacco usage, suggesting these variables may influence the molecular pathogenesis of PDAC metastasis. Additional experiments are required to determine the influence of alcohol and tobacco use on both the primary PDAC and liver proteomes to delineate how these factors influence subtype-specific selection because it could impact patient response to therapy.

Precision medicine approaches that exploit the unique molecular vulnerabilities of PDAC subtypes could be envisioned to provide a more robust clinical response. Surgical removal and FOLFIRINOX chemotherapy are common treatment strategies for PDAC (44). Surgical removal is typically available to only 10%–15% of patients with PDAC (45). FOLFIRINOX can improve patient survival, but because of its toxicity FOLFIRINOX cannot be universally applied to all patients with PDAC (44, 46). While gemcitabine is still widely used in the clinic, FOLFIRINOX treatment as an adjuvant therapy following resection increases disease-free survival to 21.6 months compared with 12.8 months for gemcitabine at the cost of increased adverse effects (47). It is possible that the benefits of FOLFIRINOX or gemcitabine are restricted to certain PDAC subtypes. Recently, the COMPASS trial also determined that patients with the basal subtype are less likely to respond to first-line chemotherapy (48). Similarly, our analysis indicates the metabolic and progenitor-like subtypes display an increase in survival time in response to FOLFIRINOX + Gem compared with gemcitabine, but this is not observed in the proliferative and the inflammatory subtypes. In addition, the progenitor-like subtype showed a significant benefit when the patients were treated with gemcitabine and/or capecitabine. Although this analysis was based on a small sample size, it is a proof-of-concept for the personalized treatment of PDAC based on proteomic signatures using traditional chemotherapeutics that could be readily implemented in the clinic.

Ultimately, PDAC subtyping must be accomplished on clinically obtainable tissues to inform first-line cancer treatment. Two recent studies have demonstrated the transcriptomics-based PDAC subtyping can be performed on percutaneous core biopsies (25, 48). However, in both studies, the average time to return results based on RNA sequencing was approximately 35–39 days, which could be used to inform second, but not first-line therapy. For RNA-based subtyping, NanoString is a platform that could be used on a subset of transcripts without the time-intensive steps required for RNA-seq library prep and data analysis (49). Proteomics could also provide a complementary rapid assay platform that could be completed in several days (50). Moving forward, it will be important to establish protocols to obtain and subtype PDAC samples in a clinically meaningful timeframe to inform first-line therapeutic decisions.

This study provides further evidence that PDAC is not a single disease and that quantitative proteomics can be used to delineate unique subtypes. The subtype-specific associations with response to chemotherapy observed in this study support the notion that the unique features of each PDAC subtype should be incorporated at all levels of therapeutic development.

No potential conflicts of interest were disclosed.

Conception and design: H.C.-H. Law, M.A. Hollingsworth, N.T. Woods

Development of methodology: H.C.-H. Law, D. Lagundžin, N.T. Woods

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): H.C.-H. Law, D. Lagundžin, F. Qiao, Z.S. Wagner, D. Costanzo-Garvey, T.C. Caffrey, J.L. Grem, D.J. DiMaio, P.M. Grandgenett, L.M. Cook, M.A. Hollingsworth, N.T. Woods

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): H.C.-H. Law, D. Lagundžin, K.W. Fisher, F. Yu, M.A. Hollingsworth, N.T. Woods

Writing, review, and/or revision of the manuscript: H.C.-H. Law, E.J. Clement, K.L. Krieger, T.C. Caffrey, K.W. Fisher, F. Yu, M.A. Hollingsworth, N.T. Woods

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): H.C.-H. Law, F. Qiao, Z.S. Wagner, T.C. Caffrey, K.W. Fisher, N.T. Woods

Study supervision: H.C.-H. Law, D. Lagundžin, N.T. Woods

The authors thank the patients and their families for their participation in the UNMC Pancreatic SPORE Rapid Autopsy Program. The authors thank the UNMC Mass Spectrometry and Proteomics core facility for project support, the UNMC Flow Cytometry core facility, and Dr. Jennifer Black and Dr. Amar Natarajan for their helpful suggestions. This work was supported by the NIH grant numbers P20GM121316, P30CA036727, and 1P50CA127297.

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.