Multi-omic profiling of intraductal papillary neoplasms of the pancreas reveals distinct expression patterns and potential markers of progression

Summary

To enable early detection of pancreatic cancer from precancerous lesions, we analyze 64 intraductal papillary mucinous neoplasms (IPMNs), 55 cyst fluid samples, 104 invasive pancreatic ductal adenocarcinomas (PDACs), and various types of normal samples using mass spectrometry. High-grade IPMNs show enrichment of glycosylation and tumor progression pathways compared to low-grade lesions. High-grade IPMN associated proteins, such as PLOD3, IRS2, LGALS9, and Trop-2, are identified and validated using immunolabeling and laser microdissection. Some of high-grade associated proteins are also detected in pancreatic cyst fluids, which allows us to link proteins and glycoproteins expressed in neoplastic cells to clinically accessible biospecimens. Altered glycosylation of extracellular matrix (ECM) proteins is observed in IPMNs compared to normal ducts. Additionally, we identify a subset of IPMNs with PDAC-like features, including elevated expression of ECM proteins. These findings offer insight into progression-associated markers and emphasize the diagnostic and therapeutic potential of these proteins in pancreatic tumors.

Introduction

Pancreatic cancer is an aggressive disease with a five-year survival rate of 13% ¹. Deaths from pancreatic cancer are increasing, and it has been predicted that pancreatic cancer will become the second leading cause of cancer death in the United States and in parts of Europe by the year 2030 ^2-5. Early detection has dramatically reduced deaths from other cancer types, and the survival rate of patients with low-stage pancreatic cancer is significantly higher than that of those with advanced-stage disease ^6,7. However, even small invasive pancreatic cancers can be deadly. This implies the need to screen for non-invasive precursor lesions, such as intraductal papillary neoplasms (IPNs), to more effectively prevent deaths from pancreatic cancer ^8,9.

IPNs, including intraductal papillary mucinous neoplasms (IPMNs), intraductal oncocytic papillary neoplasms (IOPNs) and intraductal tubulopapillary neoplasms (ITPNs), are clinically detectable macroscopic non-invasive cyst-forming lesions, some of which progress to invasive pancreatic cancer ^10-15. However, the clinical management of pancreatic cysts is challenged by misclassification, inaccurate diagnoses, and overtreatment, which can lead to unnecessary surgeries and complications ^16-19. Thus, new approaches are needed to accurately diagnose and stratify the risk of progression of IPNs ^20,21.

Proteins and glycoproteins have proven to be useful clinical biomarkers for classifying cyst-forming neoplasms in the pancreas. For example, vascular endothelial growth factor is often overexpressed in benign serous cystadenomas, while elevated levels of carcinoembryonic antigen (CEA) and cancer antigen 19-9 (CA-19.9) in cyst fluids suggest mucin-producing cystic neoplasms, such as IPMNs or mucinous cystic neoplasms, and at very high levels, they can indicate a risk of malignancy ^22-25. However, the available biomarkers for classifying cysts are imperfect, and a better understanding of pancreatic precancer biology could provide more sensitive and specific markers ^26-28. Mass spectrometry has proven a powerful tool in the discovery of new biomarkers of tumor subtypes, novel cellular pathways, and new markers of prognosis ^29-37. For example, an integrated proteogenomic analysis of 140 pancreatic ductal adenocarcinomas (PDACs) revealed associations between somatic mutations and changes in protein expression and helped delineate unique molecularly defined subtypes of PDAC³².

Here, we used proteomics and glycoproteomics to characterize a series of histologically and genetically well-characterized surgically resected IPNs, including IPMNs and IOPNs. We observed that upregulated glycosylation and tumor progression pathways were enriched in high-grade IPMNs compared to low-grade lesions. Proteins associated with high-grade IPMNs, including PLOD3, IRS2, LGALS9, and Trop-2, were identified and validated through immunolabeling and laser microdissection (LMD). The inclusion of cyst fluid samples aspirated from the neoplasms allowed us to determine which proteins and glycoproteins expressed in the neoplastic cells of the IPNs are detectable in cyst fluids, a clinically obtainable sample type³⁸. Some proteins, including PLOD3, were identified in pancreatic cyst fluids with higher levels in high-grade IPMNs than in low-grade IPMNs. Extracellular matrix (ECM) protein glycosylation was altered in IPMNs relative to normal pancreatic ducts. Finally, we reanalyzed our previously characterized PDAC samples from Clinical Proteomic Tumor Analysis Consortium (CPTAC)³² using the same platform on which the IPN samples were analyzed. Our analysis revealed three non-negative matrix factorization (NMF) clusters encompassing both IPMNs and PDACs, offering insights into progression-associated markers and potential therapeutic targets related to ECM proteins.

In summary, proteomic and glycoproteomic profiling identified upregulated proteins and pathways associated with high-grade lesions and invasive tumor progression, highlighting the diagnostic and therapeutic potential of ECM-related proteins in pancreatic tumor development.

Results

Landscape of the intraductal papillary neoplasm cohort.

Our analysis included 69 IPNs (64 IPMNs and 5 IOPNs), 76 macro-dissected grossly normal duct tissues (NDs, with 32 of the 76 NDs coming from the same resection specimens as the IPNs), and 32 cyst fluid samples (22 from matched IPMN patients, 3 matched IOPN patients, 2 from unrelated IPMN patients, and 5 other cysts) (Fig. 1A). The use of macro-dissected normal pancreatic ducts as controls was critical, as IPNs have ductal differentiation, while bulk normal pancreas is composed primarily of acinar cells. Fresh lesional tissues and normal ductal tissue samples were collected from patients who underwent pancreas surgery in one of five different countries (Fig. 1A). Corresponding clinical data and centralized pathology review by one of us (T.F.) are summarized in Supplementary Table S1A. The age and sex of these patients are similar to those reported in the general population of patients who have undergone surgical resection for IPNs ^39,40, as are the genes identified as somatically mutated in these neoplasms (Fig. 1B, Supplementary Tables S1B-S1D) ^27,41,42. In particular, 37 of 69 IPNs (54%) harbored a KRAS mutation, and 32 of the 69 (47%) a GNAS mutation (Fig. 1C, Supplementary Fig. S1A). Moreover, among 33 IPNs with a KRAS mutation variant allele fraction (VAF) > 0.075, the majority harbored one of the common hot-spot KRAS mutations of G12D (35%), G12R (19%), or G12V (19%) (Fig. 1C). Similarly, among the IPNs with a GNAS mutation with a VAF >0.075, the majority harbored one of the GNAS hotspot mutations of R201C (50%) or R201H (41%) (Fig. 1D). GNAS R201C was mostly observed in intestinal-type IPMNs, while gastric-type of IPMNs more commonly harbored GNAS R201H mutation, with none of the pancreatobiliary-type IPMNs having hotspot GNAS mutations (Supplementary Fig. S1B).

Figure 1. Landscape of the cohort.

A. Distribution of the cohort according to sample types, countries of origin, histologic subtypes, histologic grades, and available data types. B. Mutation landscape. C. Distribution of KRAS VAF and KRAS hotspot mutations. D. Distribution of GNAS hotspot mutations. E. Total global proteins and glycopeptides identified from tissues and cystic fluids of the entire cohort. Also see Figure S1 and Table S1.

Application of mass spectrometry to intraductal papillary neoplasms

Utilizing trapped ion mobility spectrometry (TIMS) high-throughput 4D-proteomics to conduct data-independent acquisition mass spectrometry (DIA-MS) for both proteomics and glycoproteomics analysis ^43-45, we identified and quantified a total of 10,246 proteins and 22,284 glycopeptides (containing 7,041 glycosites and 2,924 glycoproteins) in all tissue samples, as well as 5,314 proteins and 7,146 glycopeptides in the pancreatic cyst fluid samples (Fig. 1E). Additionally, we acquired global proteomic data for 23 additional cyst fluid samples, including 10 from serous cystic neoplasms (SCN) and 13 from IPMNs), as an independent cohort for enhancing our finding (Figure S1C).

The protein landscape of intraductal papillary mucinous neoplasms

Given their unique direction of differentiation⁴⁶, the 5 IOPNs were not included in the following analyses. Compared to 76 NDs, 756 proteins were significantly up-regulated in the 64 IPMNs by more than 1.5-fold and 438 proteins were down-regulated by more than 1.5-fold (Fig. 2A, Supplementary Table S2A). As implied by the inclusion of “mucinous” in the name of IPMNs, glycans and polysaccharides were significantly upregulated in these neoplasms. The IPMNs that harbored KRAS and/or GNAS hotspot mutations exhibited particularly elevated glycosylation and glycoprotein processing proteins (Supplementary Table S2B-S2D). Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis showed a significant enrichment of the IPMN proteome in glycan biosynthesis and metabolism pathways, particularly mucin type O-glycan biosynthesis (Fig. 2B, Supplementary Table S2E) ^47-49. Immune pathways related to host defense were identified, where the immune system recognizes and eliminates antibody-tagged pathogens ⁵⁰. Similar patterns of protein expression were observed when only the 46 IPMNs with a somatic mutation were considered (Supplementary Table S2F). The proteins expressed in the IOPNs are separately listed in Supplementary Table S2G.

Figure 2. Identification of IPMN- and grade-associated proteins and their detections in cyst fluids.

A. Comparison of 64 IPMNs and 76 normal duct tissues (NDs). Proteins with immunochemistry labeling are in pink. The p-values were computed using two-sided Wilcoxon rank sum test and adjusted (false discovery rate, FDR) using Benjamini-Hochberg method. B. KEGG pathway enriched by using upregulated and downregulated proteins in IPMN tissues relative to NDs. C. IHC labeling results of S100P, SERPINB5, and THBS2 in ND, low-grade (LG) IPMN, and high-grade (HG) IPMN. D. Significantly upregulated (red) and downregulated (blue) proteins in 34 HG IPMNs relative 30 to LG IPMNs and 76 NDs. Proteins associate with glycosylation/cancer progression are named in red. E. Individual proteins show potential clinical utilities for HG IPMN detection based on the ROC analysis. F. Multi-protein panels improve the performance of distinguishing HG IPMNs from LG IPMNs. The proteins and multi-protein panels in E and F are selected as examples. A comprehensive ROC results can be found in the corresponding supplementary table. Also see Figure S2 and Table S2.

Notably, 88 of the 756 proteins overexpressed in the IPMNs were also reported as upregulated in PDACs in our previously published study ³² (Supplementary Table S2H). Some proteins previously reported as overexpressed in low-stage PDAC, such as S100P, SERPINB5, and THBS2, were also upregulated in the IPMNs relative to NDs (highlighted in pink in Fig. 2A) ³². Proteins highly abundant in NDs were associated with normal pancreatic functions, including protein and fatty acid digestion and absorption, as well as pancreatic secretion (Fig. 2B).

Localization of protein expression

Samples of IPNs and PDACs contain a mixture of neoplastic and non-neoplastic cells ^51,52. Proteins overexpressed by the neoplastic cells are more likely to be released into cyst fluid than those expressed in the stroma. We utilized a previously published database of genes differentially expressed in the neoplastic cells and stroma of PDAC to define the most likely compartment for each overexpressed protein identified in the IPMN tissues⁵³. As shown in Supplementary Table S2I, transcripts such as S100P and SERPINB5 were previously shown to be elevated in PDAC neoplastic cells, whereas transcripts such as thrombospondin 2 (THBS2), and polymeric immunoglobulin receptor (PIGR) were previously identified as elevated in the stromal compartment. The thrombospondins function in cell-matrix interactions, and PIGR has been reported to be a marker of immunologically "hot" tumor nests ^53,54.

We next validated the expression of selected proteins by immunolabeling, which not only confirmed their overexpression but also revealed the specific cellular compartments in which they were localized. As shown in Fig. 2C and Supplementary Table S2J, antibodies to S100P and SERPINB5 demonstrated that these proteins were overexpressed by the neoplastic cells, while antibodies to THBS2 labeled the stromal cells.

Proteins identified in the IPMNs can be detected in pancreatic cyst fluid

Given that cyst fluid can be sampled endoscopically, it provides a minimally invasive opportunity for cyst classification, particularly if the proteins expressed by the neoplastic cells are released into the cyst contents^22-27. Therefore, we examined whether proteins upregulated in IPMNs are also present in cyst fluid by analyzing 24 samples, including 22 with matched IPMN tissues (discovery cohort; Supplementary Fig. S2A and Table S2K). A number of proteins upregulated in IPMNs also had similar expression patterns in cyst fluid samples. Members of the S100 protein family, including S100A6, S100A11, and S100P, were highly expressed in IPMNs and abundant in the cyst fluid samples. Additionally, MUC5AC, a major carrier protein for CA-19-9, was also highly expressed in cyst fluids (Supplementary Fig. S2B). To further validate our findings, additional 23 cyst fluid samples (10 SCNs, 6 low-grade IPMNs, 7 high-grade IPMNs; validation cohort) from an independent cohort were included (Supplementary Table S2L) and similar expression patterns were observed for the aforementioned proteins (Supplementary Fig. S2B).

Tumor grade/histology subtype associated proteins with potential biomarkers for disease progression

It is clinically important to distinguish IPMNs with low-grade dysplasia from those with high-grade dysplasia, and from IPMNs with an associated invasive carcinoma⁵⁵. Those with low-grade dysplasia can be safely followed, while the presence of high-grade dysplasia or IPMNs with an associated invasive carcinoma ( “high-grade IPMN” in following analyses refers to both high-grade dysplasia and IPMN with an associated invasive carcinoma together) typically warrants surgical resection^55,56. A total of 67 proteins were significantly overexpressed in high-grade IPMNs compared to those with low-grade dysplasia (Supplementary Table S2M). These included proteins related to glycosylation biosynthesis, such as PLOD3, GXYLT1, UAP1, GNE, and B3GNT3. Among these glycan synthetic proteins, GNE (Glucosamine (UDP-N-acetyl)-2-epimerase/N-acetylmannosamine kinase) is a key enzyme required for the modification of sialic acid on glycoproteins, a crucial modification for the stability and recognition of glycoproteins, and implicated in tumor progression⁵⁷. Other upregulated proteins such as CDH17, AGPS, CDK5, LGALS9, CDCP1, TACSTD2, and CDK13 are involved in pathways and networks that influence tumor invasion and progression. To identify proteins associated with progression risk in high-grade IPMNs, we focused on proteins overexpressed in both high-grade IPMNs relative to low-grade IPMNs and high-grade IPMNs relative to NDs. (Supplementary Table S2M, highlighted in red in Fig. 2D). Some of these upregulated proteins (Supplementary Fig. S2C), such as galectin 9 (LGALS9), multifunctional procollagen lysine hydroxylase and glycosyltransferase LH3 (PLOD3), and insulin receptor substrate 2 (IRS2), have been reported as upregulated in PDAC^58-63. To verify our findings, we conducted immunohistochemical labeling of PLOD3, IRS2, and LGALS9, which confirmed their elevated expression in high-grade IPMNs (Supplementary Fig. S2D and Table S2J). We next developed laser microdissection (LMD) based proteomics workflow that enabled deep proteome coverage, identifying over 6,000 proteins in pancreatic tissue ⁶⁴. PLOD3 and LGALS9 were further validated in an independent LMD analysis of 12 matched pairs of low-grade and high-grade IPMN samples from three patients (Supplementary Fig. S2E and Table S2N). We also observed elevated PLOD3 in low-grade and high-grade IPMN samples compared to SCN samples in cyst fluid (Supplementary Figure S2F and Table S2O).

We further evaluated the potential clinical utilities of the overexpressed proteins in high-grade IPMNs (compared to low-grade IPMNs) as individual protein signature and in different combinations through receiver operating characteristic (ROC) analysis (Supplementary Table S2P). Individually, PLOD3, LGALS9, IRS2, tumor-associated calcium signal transducer 2 (TACSTD2, also known as Trop-2), and family with sequence similarity 3, member D (FAM3D) had good performance in distinguishing high-grade and low-grade IPMNs with area under the curves (AUCs) ranging from 0.75 to 0.83 (Fig. 2E). The performance could be improved by using multi-protein signature panels, the AUC was increased to 0.87 when combining LGALS9 and Trop-2. Therapeutic antibodies targeting LGALS9 and Trop-2 have been reported in clinical trials or are available on the market^65-67.

Different histological subtypes of IPMNs show distinct protein expression patterns. In gastric-type IPMNs, we identified 47 proteins that were upregulated compared to other subtypes of IPMNs. Interestingly, 23 of these 47 proteins also showed upregulation in low-grade IPMNs relative to high-grade IPMNs (Supplementary Table S2Q). For the intestinal-type IPMNs, 57 proteins were upregulated, including CEACAM5, MUC2, GNE, and CDH17 (Supplementary Fig. S2G). Of these, 14 were also upregulated in high-grade IPMNs compared to low-grade IPMNs (Supplementary Table S2R). MUC2 involves in the key process of intestinal-type IPMN differentiation and progression⁶⁸. These results suggest potential biomarkers for tumor grade/histology subtype and progression.

The glycoprotein landscape of intraductal papillary mucinous neoplasms

Comparison of the glycopeptides identified in the 53 IPMNs (among 64 IPMNs, 11 lacked enough digested global peptides for glycopeptide enrichment and the 5 IOPNs were excluded) with available glycopeptide data to those identified in the 66 NDs (among 76 NDs, 10 lacked enough digested global peptides for glycopeptide enrichment) revealed that 2,327 glycopeptides were significantly upregulated more than 1.5-fold and 1,270 glycopeptides were downregulated more than 1.5-fold in the IPMNs (Fig. 3A, Supplementary Table S3A). Cell surface glycoproteins, such as CEACAM5⁶⁹, integrins (ITGA5, ITGB1)^70,71, and mucins (including MUC1, MUC2, MUC4, MUC5AC, MUC5B, MUC6, and MUC13)^47,72 that were overexpressed are known to facilitate cell adhesion and form protective barriers around the tumor surfaces. Additionally, CD276^73,74, CD47^75,76, and other immune-related glycoproteins previously discovered as PDAC-associated proteins ^73-77, were also found to be upregulated in IPMNs (Supplementary Table S3). Similar patterns of glycoprotein expression were observed when only the IPMNs with a somatic mutation were considered (Supplementary Fig. S3A and Table S3B). The glycoproteins expressed in the IOPNs are separately listed in Supplementary Table S3C.

Figure 3. Altered glycosylation in 53 IPMNs.

A. Comparison of glycopeptides in 53 IPMNs and 66 NDs. The p-values were computed using two-sided Wilcoxon rank sum test and adjusted using Benjamini-Hochberg method. B. Expression changes in glycopeptides in comparison to the changes in global protein levels (IPMN vs NDs). C. Significantly enriched GO biological functions of upregulated and downregulated glycopeptides in instances in which the proteins were unchanged at the global level. D. Examples of individual glycopeptides and combination of glycopeptides capable of differentiating IPMNs from NDs. E. glycopeptides with differential expression in KRAS-mutant IPMNs or both KRAS- and GNAS-mutant IPMNs compared to KRAS-wildtype (WT) and/or GNAS-WT IPMNs. Also see Figure S3 and Table S3.

More than 130 proteins were upregulated at both the global protein and glycoprotein levels (Supplementary Tables S3D). These included mucins (MUC5AC, MUC6, and MUC13) and mucin-type O-glycan synthetic enzymes (C1GALT1C1, GALNT3, GALNT6, and GCNT1). Extracellular matrix (ECM) interaction proteins⁷⁸ such as CD47^76,77, integrins (ITGA2, ITGA9, ITGB4, ITGB6, and ITGB8)⁷⁹, and laminins (LAMA3, LAMB3, and LAMC2)⁸⁰, were also upregulated at both the global protein and glycoprotein levels, indicating changes in the tumor microenvironment in IPMNs.

Comparison of glycopeptide expression to global protein expression revealed that some changes in expression were only observed at the glycopeptide level, while others were only observed at the protein level (Fig. 3B, Supplementary S3D). This finding highlights the importance of characterizing both proteins and glycopeptides, as protein expression may not always equate to glycosylation modification⁸¹. To reveal the unique pattern of glycoprotein expression compared to protein expression in IPMNs, we analyzed the biological function (GO terms) of upregulated or downregulated glycopeptides whose global protein expression remained unchanged (Supplementary Table S3E). Some upregulated glycoproteins functioned in integrin-mediated signaling and leukocyte migration, indicating changes in the tumor microenvironment (Fig. 3C). Metabolic-related functions were also upregulated in the IPMN glycoproteome in instances where the global protein expression remained unchanged. Downregulated glycoproteins included those involved in cell-cell adhesion mediated by cadherin, an important modulator of epithelial-mesenchymal transition (EMT) during tumor progression⁸².

A comparison of glycopeptide expression in 53 IPMNs to 69 NDs revealed a number of glycoprotein changes associated with IPMNs (Supplementary Fig. S3B and Table S3F). Several glycopeptides and combinations of glycopeptides produced a mean bootstrap AUC > 0.7 (Supplementary Fig. S3B and Table S3F-S2G) in distinguishing IPMNs from NDs. Including upregulated and downregulated glycopeptides, individual markers, including CADM4-N262, FCGBP-N1317, LMF2-N489 and PIGR-N469, achieved AUCs between 0.71 and 0.87 (Fig. 3D). These AUCs improved to as high as 0.9 when markers were combined (Fig. 3D).

We next compared the patterns of glycopeptide expression in 17 IPMNs that harbored mutations in KRAS and GNAS to the 15 wildtype IPMNs ( “wildtype IPMNs” in following content and figures refers to IPMNs that lacked any detected somatic mutations) (Fig. 3E). There were 71 glycopeptides involving 56 glycoproteins upregulated in IPMNS with both KRAS and GNAS mutations compared to wildtype IPMNs (Supplementary Table S3H-S3I). These included O-glycosylation enzymes such as C1GALT1C1, GALNT3, GALNT5, GCNT1, and ST6GALNAC1. This pattern could be explained by the activation of the PI3K-AKT signaling pathway in the cells with mutations. PI3K-AKT pathway-related glycoproteins, including EPHA2, ERBB3, ITGA6, MET, and YWHAH, were also observed among the upregulated glycoproteins, as were transporter glycoproteins (ATP1A3, ATP6V0A2, CNNM4) related to inorganic ions, such as Na⁺/K⁺, H⁺, and Mg²⁺. These findings suggest that the neoplastic cells in IPMNs with mutations in KRAS or GNAS undergo significant adjustments in ion balance and reprogram metabolic processes to meet high metabolic demand.

ECM related proteins upregulated during neoplastic progression

We selected 104 of the most cellular (tumor cellularity >15%) PDACs and 43 of their matched NATs from a previously reported series of PDACs and reanalyzed them using DIA-MS ³² (Fig. 4). In principal component analysis (PCA), the PDACs, IPMNs, and normal samples (NDs and NATs) clustered separately (Fig. 4A). The IPMN samples clustered between PDACs and normal samples, supporting the transitional stage that IPMNs occupy between normal ductal tissue and invasive cancer^83,84. Comparing the protein expression in the 104 PDACs to the matched 43 NATs revealed 1,539 proteins significantly upregulated more than 1.5-fold, and 902 proteins significantly downregulated more than 1.5-fold in the cancers (Supplementary Fig. 4A and Table S4A). The vast majority of these proteins have been reported previously, and a number of them, such as S100P, THBS2, and SERPINB5, were also overexpressed in IPMNs (Figs. 2A, Supplementary S4A). To determine whether these proteins were expressed by neoplastic cell or stroma components, we first utilized a previously published spatial database (Supplementary Table S2I)⁵³. We then validated the expression and cellular localization of select proteins (S100P, SERPINB5, and THBS2) in PDAC tissues using immunohistochemistry (Fig. 4B, Supplementary Table S2J)^51,52. Specifically, S100P and SERPINB5 were expressed by invasive cancer cells, while THBS2 was predominantly expressed by stromal cells. To further investigate which IPMNs are more associated with PDACs, we identified 1,670 proteins significantly changed in PDACs compared to NATs and normal ducts and calculated an association score between PDACs and IPMNs based on these proteins (Supplementary Table S4B-S4C). Eight IPMNs clustered with PDACs (Fig. 4C), each having a relatively high association score (≥ 0.88). We further examined the 1,670 proteins and found 351 proteins that were overexpressed in PDACs and in some IPMNs that grouped with PDACs compared to the remaining IPMNs (Supplementary Table S4D). These proteins are potential markers of IPMN progression since they were elevated in IPMNs with higher PDAC association scores. Among these 351 proteins, those related to invasion and metastasis, such as cell adhesion, cytoskeletal dynamics, cell motility, and ECM proteins warrant more investigation (Fig. 4D, Supplementary Table S4E). Upregulated proteins included ECM regulators and collagen-modifying enzymes such as PLOD3, ECM glycoproteins such as fibronectin (FN1) and tenascin (TNC) expressed by fibroblasts, and proteoglycans including lumican (LUM) involved in collagen fibril organization, all of which were associated with invasive tumor progression (Fig. 4E). About 60 ECM proteins showed notable changes during the transition from IPMN to PDAC, including ECM-affiliated proteins, ECM glycoproteins, ECM regulators, proteoglycans, and secreted factors (Fig. 4F, Supplementary Table S4F). We found 106 proteins that were downregulated in PDACs and IPMNs with high association scores relative to remaining IPMNs, and these were enriched in metabolic processes (Fig. 4D). These 351 upregulated and 106 down regulated proteins are potential markers for the early detection of IPMN with high progression potential.

Figure 4. Proteomic comparison between IPMNs and PDACs with high tumor cellularity.

A. PCA analysis of the cohort. B. IHC labeling results of S100P, SERPINB5, and THBS2 in PDAC tissue. C. Hierarchical clustering of 104 PDACs and 64 IPMNs, where 8 IPMNs with high association scores group with PDACs. D. Enriched GO biological processes using overexpressed and downregulated proteins in PDACs and IPMNs grouped with PDACs compared to the remaining IPMNs. E. Examples of the proteins showing significant changes between PDAC-associated IPMNs (n=8) grouped with PDACs (n=104) relative to non-PDAC associated IPMNs (n=56). F. PDAC-associated IPMNs and PDAC had abundant ECM proteins compared to non-PDAC associated IPMNs. The ECM proteins were also differentially expressed in PDAC-associated IPMNs and PDAC relative to non-PDAC associated IPMNs. G. ECM proteins differentially expressed in one of the NMF clusters that are disease-related and/or potential drug targets based on Human Protein Atlas. Also see Figure S4 and Table S4.

Proteomic subtyping highlights the heterogeneity of intraductal papillary mucinous neoplasms

IPMNs can be morphologically classified into three subtypes: gastric-type, intestinal-type, and pancreatobiliary-type^85,86. Furthermore, IPMNs can be categorized as having low-grade dysplasia, high-grade dysplasia, and as having an associated invasive cancer, based on the degree of cytoarchitectural atypia and the presence or absence of an associated invasive cancer^85,86. However, these subtypes or classifications are not based on molecular characterization. We applied proteomics based NMF subtyping strategies to the IPMNs and PDACs to explore tumor heterogeneities and reveal progression features⁸⁷. This analysis showed three clusters (Supplementary Fig. 4B and Table S4G-S4H), NMF 1 was entirely composed of PDACs. NMF 2 included some PDACs and some IPMNs, while NMF 3 was almost entirely composed of IPMNs with only one PDAC. Most (75%) IPMNs with an associated PDAC were clustered in NMF 2. KEGG pathway enrichment using signature proteins distinguished these three clusters (Supplementary Fig. 4C and Table S4I). Notably, NMF 3 was enriched in amino sugar metabolism, metabolic processes, and pancreatic secretion. Several mucin glycosylation related enzymes in NMF 3, such as UDP-galactose-4-epimerase (GALE), Glutamine-fructose-6-phosphate aminotransferase 1 (GFPT1), GNE, and Glucosamine-phosphate N-acetyltransferase 1 (GNPNAT1), are involved in amino sugar metabolism. The sodium/potassium-transporting ATPase (ATP1A1, ATP1B1) and ADP/ATP translocase (SLC25A4) found in NMF 3 are key players in metabolic processes. Ras-related proteins (RAB27B, RAB3D) found in NMF 3, which control the maturation, trafficking, and exocytosis of secretory vesicles are highlighted in pancreatic secretion. These unique features of NMF 3 indicate reprogrammed metabolic processes associated with increased mucin glycan synthesis and extracellular vesicle secretion.

Major differences between NMF 2 and NMF 3 include elevated collagen and matrix proteins (COL4A1, COL4A2, COL4A5, COL6A1, COL6A2, COL6A3, TNXB) in NMF 2, which are enriched in the pathways of ECM-receptor interaction, PI3K-Akt signaling, and others (Fig.4G). These findings highlight the reorganization of the ECM interactions during the progression from IPMNs to PDACs. In NMF1, Gelatinase A (MMP2) and FN1, along with glycoprotein laminins (LAMB3, LAMA3) and membrane-related regulator proteins (ANXA2, ANXA5), highlight cell adhesion and migration pathways and proteoglycan interactions in cancer (Fig.4G, Supplementary Table S4J). Clustering analysis provides new insights into the progression from non-invasive precursor lesions to invasive pancreatic cancer.

Intraductal papillary mucinous neoplasms immune protein signatures

To better understand the dynamics and features of the microenvironment in different NMF clusters, we classified microenvironment protein signatures, particularly focusing on immune protein signatures and the degree of immune infiltration (Supplementary Fig. S4D)⁸⁸. We annotated the samples in the NMF 3 as “immune hot” neoplasms due to the infiltration of various CD8+ T cell subtypes. The NMF 1 cluster, entirely composed of PDACs, exhibits “cold” signatures with CD8+ naïve T-cells and CD8+ central memory T cells (CD8+ Tcm), which suggests an exhausted immune response and could lead to immune evasion by the invasive cancer^89,90. These immune protein signatures add additional value to the molecular level clustering and help to understand cancer progression.

Discussion

The mortality rate of invasive pancreatic cancer remains extremely high^1,91. Although the detection and surgical resection of non-invasive precursor lesions, including IPMNs, ITPNs, and IPONs (collectively IPNs), offers a hope to cure but also carries risks of overtreatment ^18,19. Even with modern imaging and somatic mutation testing, many surgically resected lesions are ultimately diagnosed as other types of neoplasms or low-grade IPMNs^18,19,92.

To better understand IPNs and identify clinically useful markers, we characterized deep proteomic and glycoproteomic profiling using MS on normal pancreatic ducts and IPNs of different genomic backgrounds, grades, and histologic subtypes (Fig. 1). We focused on IPMNs, the most prevalent IPN subtype, and identified a number of upregulated proteins and glycoproteins (Figs. 2A and 3A, Supplementary Tables S2A and S3A), such as CEACAM5, MUC1, MUC2, MUC4, MUC5AC, MUC6, and MUC13^72,83. The various overexpressed proteins and glycoproteins identified have clinical significance as many can be detected in clinically obtainable cyst fluid (Supplementary Figs. S2A and S2B). Several markers such as PLOD3, LGALS9, and IRS2 were validated by immunolabeling and LMD (Supplementary Figs. S2D and S2E), and a multi-marker panel associated with high-grade IPMNs achieved an AUC reached 0.87 (Figs. 2E and 2F).

Two methods, mining a previously reported spatial database and immunolabeling, were employed in this study to classify the cellular compartment (neoplastic cells or stroma) in which the various proteins and glycoproteins were likely expressed (Figs. 2C and 4B, Supplementary Tables S2J)^93,94.

To investigate IPMN in the boarder context of tumor progression, we re-analyzed proteomic profiles of PDACs and NATs using the same MS platform³². PCA clustering positioned IPMNs between PDACs and normal samples (NDs and NATs) (Fig. 4A), supporting the hypothesis that IPMNs represent a transitional stage in pancreatic carcinogenesis^83,84,95,96. Among the 1,670 proteins significantly changed in PDACs compared to normal samples, 351 proteins were overexpressed in both PDACs and the 8 IPMN that clustered PDACs with significant ECM proteins involved (Fig. 4F).

NMF clustering within IPMNs and PDACs revealed three clusters, NMF 1 (PDACs), NMF 2 (mixed), and NMF 3 (mostly IPMNs) (Supplementary Fig. S4B), that reflect key biological transitions (Fig. 4C). NMF 3 was enriched in metabolic processes, N-glycosylation processes and pancreatic secretion. NMF 2 showed ECM remodeling, while NMF 1 showed invasive features. The transition from IPMNs to PDACs can be distinguished by N-glycosylation enzymes (GFPT1, GNE, GNPNAT1, PMM2, and UAP1), ECM interaction proteins (CD47, integrins, and laminins), and ECM proteins such as collagen proteins and matrix proteins (Fig. 4G). Cancer invasion and metastasis related proteins, such as ITGA5, PLOD3, ELMO1, MMP2, and FN1, are overexpressed as disease progresses

In summary, we present protein and glycoprotein changes from NDs to IPMNs and invasive pancreatic cancer, providing evidence-based approaches for early detection and the development of potential treatment targets. Integrating these biomarkers with genomic data to develop a multi-parameter diagnostic model could enhance the accuracy of early detection. Further studies explore the potential of these markers in clinical cyst fluid analysis for personalized management of patients with a pancreatic cyst ²⁷.

Limitations of the study

There are several potential limitations of this study that should be acknowledged. First, we relied on surgically resected tumors, and many of the neoplasms included in this study were low-grade tumors. We therefore paired high-grade IPMNs with IPMNs with an associated invasive cancer in this study. Further studies focusing on purely high-grade IPMNs and, in particular, on the changes in expression that occur early in the transition from high-grade dysplasia to invasive carcinoma are warranted. Second, while the use of surgically resected tumors enabled the analysis of more than two hundred samples together, the bulk sampling approach employed is unable to provide cell type specific and spatial resolution. To overcome this limitation, we developed an LMD based spatial proteomics approach ⁶⁴ and applied it to validate the grade associated proteins in this study (Supplementary Fig. S2E). Given the limited throughput of patient cases for the single-cell and LMD analyses, future studies integrating the single-cell data with bulk profiling, such as through cell state and ecotype analyses, will be critical for achieving a more comprehensive understanding.

STAR★Methods

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

Clinical specimens

Fresh tissue samples from adult patients with IPMN and other pancreatic diseases (including SCN, MCN, and ND) who underwent surgical resection were collected under ethical approval from the Johns Hopkins University Institutional Review Board (IRB 00282409). Consent was waived for the acquisition of tissue under this protocol. Cyst fluid samples were collected under Johns Hopkins IRB protocol (NA _00001584). For this latter protocol written informed consent was obtained from all patients prior to sample collection. Reanalyzed PDAC tumor and NAT samples were derived from leftover peptides generated in a previous study³², in accordance with CPTAC guidelines.

Fresh tissue, including 76 macrodissected normal main pancreatic ducts and 69 IPNs (64 IPMNs and 5 IOPNs), 5 SCNs, 4 mucinous cystic neoplasms (MCNs), and 32 cyst fluid samples (22 matched with analyzed IPMNs), was harvested from 126 patients (59 males, 67 females, age range 21 to 87) from five different countries who underwent surgical resection for a pancreatic lesion. Cyst fluid was aspirated at the surgical pathology bench from 22 of the analyzed IPMN-matched patients, 2 were from other IPMN patients, 3 IOPN patients, and 5 other IPMN cyst samples. The other 5 cyst fluid samples were from other pancreatic neoplasms including 2 MCNs, 2 SCNs and 1 solid pseudopapillary neoplasm. Additional 23 cyst fluid samples were analyzed as validation cohort, 10 were from SCN patients, 6 were from low-grade IPMN patients, and 7 were from high-grade IPMN patients. Of the 76 main pancreatic duct specimens, 32 were harvested from the 64 patients with an IPMN, and 44 were harvested from patients who underwent surgery for a neoplasm that does not involve the duct system. In addition, 104 PDAC fresh-frozen samples with >15% neoplastic cellularity and 43 PDAC matched NATs were obtained from a previous multi-omics study of 144 PDACs cohort and rerun on the same platform as the normal duct, IPMN and cyst fluid samples (DIA-MS technology) ³². The hematoxylin and eosin-stained sections of all IPMNs were reviewed by one pathologist (T.F.) and the diagnosis, direction of differentiation, and histologic grade confirmed. 16 independent formalin-fixed and paraffin-embedded IPMNs were selected from the files of the Johns Hopkins Hospital for immunolabeling (Supplementary Table S2J).

METHOD DETAILS

DNA extraction, library preparation, and targeted next-generation sequencing

DNA was extracted from cryo-pulverized fresh frozen tissue using the QIAamp DNA FFPE Tissue Kit (Qiagen, Hilden, Germany) and a modified protocol⁹⁸. In brief, tissue was enzymatically digested overnight, mechanically sheared (M220 Focused-ultrasonicator, Covaris, Woburn, MA), and DNA extracted following the manufacturer’s instructions. DNA concentration and base-pair length were quantified using an electrophoresis device (TapeStation System, Agilent, Santa Clara, CA).

Library preparation was performed using the SureSelect XT HS2 DNA System (Agilent, Santa Clara, CA) as suggested, with the modification of using half the probe volume that is indicated in the original protocol for preparing the probe hybridization mix in some samples. A customized panel of 12 established pancreatic driver genes (BRAF, CDKN2A, CTNNB1, GNAS, KLF4, KRAS, PIK3CA, PTEN, RNF43, SMAD4, TP53, VHL) was used for capture (Supplementary Table S1C). In total, our gene panel contained 1,115 probes across 142 genomic regions (Supplementary Table S1D). Prepared libraries were stored at −20 °C until sequencing. Libraries were pooled and sequenced on a MiSeq instrument (Illumina, San Diego, CA) using the MiSeq Reagent Kit v2 and generating 2x 150 base-paired reads.

Sample processing for protein extraction from tissues, tryptic digestion, and global proteomic sample preparation

All tissue samples for this study were processed for mass spectrometry (MS) analysis at Johns Hopkins University. The sample preparation for global proteomic analysis followed the previous described approach in PDAC study, and glycoproteomic analysis was carried out by liquid handing systems as previous published work³². All tumor and normal ductal tissue samples were cryo-pulverized and lysed in urea lysis buffer as previously published standard protocol ⁹⁹ (8 M urea, 75 mM NaCl, 50 mM Tris pH 8.0, 1 mM EDTA, 2 μg/mL aprotinin, 10 μg/mL leupeptin, 1 mM PMSF, 10 mM NaF, phosphatase inhibitor cocktail 2 [1:100 dilution], phosphatase inhibitor cocktail 3 [1:100 dilution], and 20 μM PUGNAc) by sonicating for 30s x 3 in ice water). Tissue debris was removed by centrifugation at 16,000 x g for 15 min at 4 °C. The lyzed protein supernatant was collected, the concentration was measured by BCA assay and adjusted to 8 mg/mL in 8M urea lysis buffer. The lyzed proteins were reduced with dithiothreitol (6 mM, 37 °C, 1h) and alkylated by iodoacetamide (12 mM, room temperature in dark, 45 min). After that, the reduced proteins were diluted to 2M urea concentration with 50 mM Tris-HCl pH 8.0 buffer, digested by Lys-C (enzyme to protein ratio is 1 mAU:50 mg, 2h at room temperature) and trypsin (enzyme to protein ratio is 1 mg:50 mg, overnight incubation at room temperature), respectively. The proteolytic reactions were quenched by 50% formic acid to adjust pH < 3. About 5 μg digested peptides were aliquoted for global proteomics analysis by stage tip desalting and dried using Speed-Vac (Thermo Scientific). 1 μg peptides were loaded on Evotip (Evosep Biosystems) according to the EvosepOne (Evosep Biosystems) protocol for global proteomic analysis on timsTOF-HT (Bruker). The remaining digested peptides were desalted on reverse phase C18 96-well plate and dried using Speed-Vac for glycopeptide enrichment.

Enrichment of glycopeptides by liquid handing systems from global peptides

As previously reported, C18/MAX-Tips were conditioned using acetonitrile, 100 mM triethylammonium acetate, 95% acetonitrile in 1% TFA, and 0.1% TFA, respectively in Versette Liquid Handling System (Thermo Scientific). The remaining digested peptides were reconstituted in 300 μL 0.1% TFA (trifluoroacetic acid), about 200 μg global peptides bound onto the C18/MAX-Tips (20 cycles) and washed with 0.1% TFA (10 cycles). After that, the non-glycopeptides were eluted from the C18/MAX-Tips with 200 μL 95% acetonitrile in 1% TFA (3 x 10 cycles), and glycopeptides were sequentially eluted with 200 μL 50% acetonitrile in 1% TFA (3 x 10 cycles). Samples were dried using Speed-Vac. The dried glycopeptide samples were resuspended in 40 μL 0.1% TFA and aliquoted 20 μL glycopeptide out into a 96-well plate. PNGaseF digestion buffer (2μL 1000 units PNGaseF added to 98 μL 100mM triethylammonium bicarbonate, TEAB) was added to the 96-well plate and reacted at 37 °C overnight. The reaction was quenched by formic acid to adjust pH < 3 and cleaned up by stage-tip. The samples were dried and prepared for five injections and loaded one injection on Evotip (Evosep Biosystems) for glyco proteomic analysis on timsTOF-HT (Bruker).

Deparaffinization and Staining of FFPE Tissue Sections for LMD Analysis

Pancreatic cancer FFPE tissue sections mounted on PEN membrane slides were deparaffinized to remove residual paraffin. Slides were immersed in 100% xylene for 10 minutes, followed by transfer to a fresh xylene solution for an additional 10-minute incubation. For rehydration, the slides were sequentially immersed in 100%, 70%, and 50% ethanol, and then in distilled water, each for 5 minutes. Prior to staining, slides were equilibrated in 1× phosphate-buffered saline (PBS). PBS was gently removed by tapping the slide onto a paper towel. A total of 2 mL of hematoxylin was applied to cover the tissue section, followed by a 3-minute incubation. After rinsing with distilled water, slides were briefly immersed in 0.5% ammonium hydroxide for 1 minute to achieve bluing and then rinsed again with distilled water.

LMD tissue dissection

FFPE tissue slides were thoroughly air-dried at room temperature before microdissection. Slides were mounted onto the stage of the LMD7000 system (Leica, US), and regions of interest (ROIs) were identified under 20× magnification. Laser parameters were optimized for precise tissue dissection: power 54, aperture 1, speed 5, specimen balance 50, head current 100%, pulse frequency 773 Hz, and offset 110. Dissected tissue voxels were collected into the flat caps of 0.2 mL PCR tubes (Cat# 6571, Corning, Tamp, MEX) containing 20 μL of DMSO. Caps were carefully inspected after each LMD procedure to confirm successful collection of tissue fragments. Tubes were centrifuged at 15,000 × g for 1 minute and then subjected to vacuum centrifugation for 4 hours to ensure complete solvent removal.

Proteomic Sample Preparation of LMD samples

Dried tissue samples were reconstituted in 1.5 μL of lysis buffer composed of 0.1% n-Dodecyl-β-D-Maltoside (DDM), 0.02% Lauryl Maltose Neopentyl Glycol (LMNG), and 5 mM tris(2-carboxyethyl)phosphine (TCEP), all prepared in 100 mM TEAB. Samples were incubated at 80°C for 1 hour with a heated lid set to 100°C. After a 5-minute sonication, samples were cooled to room temperature. Proteolytic digestion was performed by addition of 0.5 μL of Lys-C (50 ng/μL) and 0.5 μL of trypsin (50 ng/μL), followed by overnight incubation at 37°C with a lid temperature of 57°C. Digestion was quenched by adding 0.5 μL of 5% formic acid and incubating at room temperature for 10 minutes. The resulting peptides were then transferred and loaded onto pre-conditioned Evotips (Evosep Biosystems) in accordance with the manufacturer’s protocol.

EVOSEP-timsTOF for global and glycoproteomic analysis

EvosepOne (Evosep Biosystems) LC system was coupled with timsTOF-HT mass spectrometry. Global peptides and glycopeptides were loaded on Evotip and separated on PepSep C18 column of 15 cm x 150 μm, 1.5 μm (Bruker) in Bruker column toaster (50 °C) at a 30 SPD gradient. The global proteomic data was acquired using the DIA-PASEF mode on timsTOF-HT with settings as follows: MS1 scan range of 100-1700 m/z; MS2 scan range of 338-1338 m/z, mass width 25.0 Da without mass overlap, 1 mobility window, 1/K0 range of 0.70-1.45 V·s/cm², ramp time 85.0 ms. The glycoproteomic data was acquired using the DIA-PASEF mode on timsTOF-HT with settings as follows: MS1 scan range of 100-1700 m/z; MS2 scan range of 338-1488 m/z, mass width 25.0 Da without mass overlap, 1 mobility window, 1/K0 range of 0.70-1.45 V·s/cm², ramp time 85.0 ms. The samples were randomized and blinded to the individual generating the data. Replications of pooled samples were used as quality control throughout the entire data generation process.

Spectral library generation for DIA-MS analysis of global proteomics and glycoproteomics

Spectral library was generated in Spectronaut^® 18.4 (Biognosys AG) by combination of all search archives from semi- specific and full-specific tryptic searches, which contained entire patient sample cohort. The Pulsar search settings for the global proteomic search archive generation were selected semi-specific (or full-specific) tryptic digestion with up to two missed cleavages within the length range of 7–52 amino acids. The search was perfomed against the UniProtKB reviewed human proteome (20,426 entries; FASTA file was downloaded on 09/15/2023). Carbamidomethylation of cysteine (C) was set as the fixed modification. Oxidation of methionine (M) and acetylation of protein N-terminal were set as the variable modifications. The library generation settings were followed BGS factory settings with all PSM FDR, peptide FDR and protein FDR below 0.01. The settings for the glycosite-containing peptide library generation were similar to those for the global proteomic library generation, with deamidation of asparagine (N) to aspartic acid (D) added as a variable modification.

Immunohistochemistry

FFPE tissue blocks containing human primary tumors were sectioned at 4 mm onto Superfrost Plus microscope slides (VWR International, catalog 48311-703). Automated immunohistochemistry was performed at the Oncology Tissue and Imaging Services Core of the Johns Hopkins University School of Medicine using a Ventana Discovery Ultra automated slide staining system (Roche). Reagents used for deparaffinization, heat-induced epitope retrieval, and chromogenic signal detection were obtained from Roche and used according to the manufacturer’s protocol. Heat induced epitope retrieval (HIER) was performed using Ventana Ultra CC1 buffer (Roche, catalog 6414575001) at 96°C for 64 minutes. Following HIER, primary antibodies were at incubated at 37°C for 60 minutes under the following conditions: anti-S100P antibody, anti-SERPINB5 antibody, anti-THBS2 antibody, anti-IRS2 antibody, anti-Galectin 9 antibody, and anti-PLOD3 antibody. Primary antibodies were detected using an anti-rabbit HQ and anti-HQ HRP detection system (Roche, catalog 7017812001 and 7017936001) followed by Chromomap DAB IHC detection (Roche, catalog 5266645001) according to the manufacturer’s protocol.

QUANTIFICATION AND STATISTICAL ANALYSIS

Genomic data analysis

Data processing and mutation call

Following removal of duplicate reads, sequence reads were aligned to human genome 38 (hg38; Human_GRCh38.p13_MajChr) using NextGENe (SoftGenetics, State College, PA). After alignment, putative somatic mutations were filtered based on the following criteria: 1) a total coverage of at least 6 distinct reads at the mutation position; 2) at least 4 distinct reads supporting the mutation; 3) mutation allele frequency (MAF) of >5% for single base substitutions (SBSs), >10% for indels, and >20% for homopolymer indels. Software-based mutation calls were inspected with NextGENe Viewer (SoftGenetics, State College, PA) and further processed using Excel (Microsoft, Redmond, WA). Putative somatic mutations in genes covered in our targeted panel listed above were selected and alterations designated by NextGENe as ‘noncoding’, ‘synonymous’, or ‘none’ removed.

With the goal to identify somatic mutations in our tumor samples, 17 unmatched normal ductal samples that lacked KRAS and/or GNAS somatic mutations served as normal tissue for reference. Putative somatic mutations identified in normal samples were classified as germline variants. Putative somatic mutations unique to the tumor samples were visually inspected using Integrative Genomics Viewer (IGV; Broad Institute, Cambridge, MA) and compared to the reference genome hg38 to remove sequencing artifacts. In addition, hotspots for KRAS (codons G12, G13, Q61) and GNAS (codon R201) were visually inspected in all samples using IGV, irrespective of software-based mutation call by NextGENe. After visual inspection putative somatic mutations in genes other than KRAS or GNAS, that were not present in the Genome Aggregation Database (gnomAD, Broad Institute, Cambridge, MA; datasets GRCh38 gnomAD v4.0.0; gnomAD v3.1.2; gnomAD v3.1.2 non-cancer) or present in with an allele frequency <0.1% in all populations, were classified as somatic mutations. Putative somatic mutations present in these databases in any population at an allele frequency >0.1% were classified as germline variants.

MS Data Interpretation

Global proteomic DIA data

All DIA raw files were searched against the spectral library described above with BGS default factory setting in Spectronaut, where mass tolerance of MS and MS/MS was set as dynamic with a correction factor of one. PSM FDR, peptide FDR and protein FDR were all set below 0.01. Crossrun normalization was enabled. Protein abundances were computed from the mean of the quantity of its top 3 peptides (stripped sequences), whereas the quantity of a peptide was the mean of the quantity of its top 3 precursors. Median normalization was further applied to generate the final protein expression matrix used in this study.

Glycoproteomic DIA data

All DIA raw were searched against the glycosite-containing peptide spectral library described above with BGS default factory setting (cross-run normalization was disabled). The peptides were identified and quantified by modified sequence Median normalization was applied to generate the final glyco-site containing peptide expression matrix used in this study.

Other proteomic analysis

Differential analysis

The differential analysis was carried out by calculating the median log2 fold changes for the following: (i) IPMNs vs NDs, (ii) HG vs LG IPMNs, (iii) HG IPMNs vs NDs, (iv) KRAS- and/or GNAS-mutant IPMNs vs KRAS- and/or GNAS-WT IPMNs, (v) Gastric-type IPMNs vs remaining subtypes of IPMNs, (vi) Intestinal-type IPMNs vs remaining subtypes, (vii) PDACs vs NATs, (viii) PDACs vs NDs, (ix) IPMNs (with high association score) and PDACs vs remaining IPMNs, and (x) one NMF cluster vs another NMF cluster on global proteomic or glycoproteomic level when applicable. The p-values were computed using two-sided Wilcoxon rank sum test and adjusted (false discovery rate, FDR) using Benjamini-Hochberg method.

Enrichment analysis for KEGG pathway and GO terms

KEGG pathways and GO enrichment analyses were performed using the over-representation analysis on the WebGestalt (version 2019) with default setting and top 10 enriched terms were exported⁹⁷. Only significantly upregulated and/or downregulated proteins/glycopeptides were used in the analyses.

Protein abundance ranking for tissues and cyst fluids

Significantly upregulated proteins in IPMNs (vs NDs) with less than 75% missing values across all the cyst fluid samples were used to investigate the relation between IPMN tissues and cyst fluids. The protein abundances were z-score transformed and then ranked within each sample.

Association score calculation to determine whether an IPMN was more associated with PDAC

A total of 1,670 proteins were differentially expressed in PDACs compared to NDs and NATs, which were used for calculating the association between PDACs and IPMNs. In brief, median abundance was computed for each protein across 104 PDAC tissue samples. Association score was determined by calculating the Spearman correlation between abundances of the aforementioned 1,670 proteins in an IPMN tissue sample and the median abundances of those 1670 proteins in PDAC.

ROC analysis for differentially expressed proteins and glycopeptides

We performed the ROC analysis following a similar approach to our previous study¹⁰⁰. In brief, the performance of each protein/glycopeptide signature panel (either composed of one or multiples) through logistic regression was evaluated using ROC analysis. The data were log-transformed followed by z-score (missing values were median imputed) prior to ROC analysis. We used bootstrap resampling (n = 500) of the data to construct and evaluate the predictive model of a panel to ensure statistical stability of the result. The mean ROC curves were depicted based on bootstrap resampling results and an AUC was computed for the mean ROC curve. The predictive models were built using caret (version 6.0–85) and ROC curves were generated using pROC (version 1.13).

ECM protein analysis

The ECM proteins were identified based on the list in a literature¹⁰¹. The abundance percentage was calculated by dividing the sum of abundances of proteins in an ECM category from the sum of the abundances of all ECM proteins. Information on whether an ECM protein was disease-related and/or a drug target that was obtained from the Human Protein Atlas (https://www.proteinatlas.org/).

Non-negative matrix factorization (NMF)-based clustering analysis

To further analyze the heterogeneity of IPMNs, we utilized NMF to perform proteomic clustering, including both IPMN and PDAC tissue samples. The clustering approach was conducted similar to our previous works^32,35. Briefly, NMF was used to perform unsupervised clustering of tumor samples based on the abundances of proteins (only proteins with CVs in >10% quantile were used in the analysis). The feature matrix was scaled and standardized, thus all features were represented as z-scores. Since NMF requires a non-negative input matrix, the feature matrix was further converted as follows: (i) create one data matrix with all negative numbers zeroed, (ii) create another data matrix with all positive numbers zeroed and the signs of all negative numbers removed, and (iii) concatenate both matrices resulting in a data matrix with positive values and zeros only. The resulting matrix was then subjected to NMF analysis leveraging the NMF R-package⁸⁷. To determine the optimal factorization rank k (i.e., number of clusters), a range of k from 2 to 5 was tested using default settings with 50 iterations. The optimal factorization rank k=3 (i.e., NMF 1, NMF 2, and NMF 3) was selected since the product of cophenetic correlation coefficient and dispersion coefficient of the consensus matrix was the maximum compared to other tested ks. The NMF analysis was repeated using 500 iterations for the optimal factorization rank. A list of representative features for each NMF cluster was derived that differential analysis was carried out using cluster-specific features by comparing one NMF cluster to the remaining NMF clusters as described in the Differential analysis section. Heatmap of NMF clusters was generated using ComplexHeatmap (version 2.4.3).

Immune protein signatures of each NMF cluster

The cell type gene signatures were downloaded from xCell ⁸⁸ and we extracted protein abundances of these gene signatures from our global proteomic data. In this study, we focused on the cell types that were enriched in the immune-hot PDAC samples in our previous publication to examine differences among NMF clusters due to the degree of immune infiltration for the IPMN³². We found the association between cell type protein signatures and each NMF cluster. Based on the differential abundance (i.e., log2 fold change) between a NMF cluster and the remaining NMF clusters, a total of 112 protein signatures were obtained. Among the 112 protein signatures, 33, 41, and 38 proteins were significantly changed in NMF 1, NMF 2, and NMF 3, respectively. The abundance of each protein signature in a NMF cluster was weighted and signed by multiplying the protein abundances to the aforementioned log2 fold change. For each cell type in each NMF cluster, a cell type association score was calculated by summing up the weighted and signed abundances of the protein signatures. The cell type associated scores for each cell type were z-score normalized (i.e., normalized scores) among the NMF cluster. Sample processing for somatic mutation analyses

Analysis of location of proteins overexpressed by IPMNs

A list of genes that are enriched in PDAC cells compared to adjacent stromal cells was obtained from a previously published study ⁵³. Using this list, we searched for genes that corresponded proteins that were found to be upregulated in IPMNs compared to normal ducts and classified them as likely expressed in neoplastic cells or as likely expressed in stromal cells. Filtered results are summarized in Supplementary Table S2I.

Supplementary Material

Table S1

Table S1. Clinical information and genomic analysis in this study, related to Figures 1 and S1

Table S2

Table S2. IPMN- and high grade-associated proteins and their validations in translational purpose, related to Figures 2 and S2

Table S3

Table S3. IPMN glycoproteomic analysis, related to Figures 3 and S3

Table S4

Table S4. Proteomic analysis of IPMNs with their progression features, related to Figures 4 and S4

Key Resource Table

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
S100P	Abcam	ab124743 (RRID:AB_10976321)
SERPINB5	Abcam	ab182785
THBS2	Abcam	ab112543 (RRID:AB_10863103)
PLOD3	Thermo Fisher	11027-1-AP (RRID:AB_2165781)
IRS2	Abcam	ab134101 (RRID:AB_2810948)
LGALS9	Cell Signaling	54330S
Biological Samples
Primary tumor samples	See Experimental Model and Subject Details	N/A
FFPE tissue blocks	Johns Hopkins University	N/A
Chemicals, Peptides, and Recombinant Proteins
Superfrost Plus microscope slides	VWR International	48311-703
Ventana Cell Conditioning 1 (CC1) buffer	Roche	05279801001
PEN membrane slides	Zeiss	415190-9081-000
Sodium chloride	Santa Cruz Biotechnology	Catalog: sc-295833
Tris(hydroxymethyl)amino methane (Tris)	Invitrogen	Catalog: AM9855G
Ethylenediaminetetraacetic acid (EDTA)	Sigma	Catalog: E7889
Aprotinin	Sigma	Catalog: A6103
Leupeptin	Roche	Catalog: 11017101001
Phenylmethylsulfonyl fluoride	Sigma	Catalog: 93482
Sodium fluoride	Sigma	Catalog: S7920
Phosphatase Inhibitor Cocktail 2	Sigma	Catalog: P5726
Phosphatase Inhibitor Cocktail 3	Sigma	Catalog: P0044
Urea	Sigma	Catalog: U0631
PUGNAc	Sigma	Catalog: A7229
Dithiothretiol	Fisher Scientific	Catalog: 20291
Iodoacetamide	Fisher Scientific	Catalog: A3221
Lysyl endopeptidase (Lys-C), Mass Spectrometry Grade	Wako Chemicals	Catalog: 125-05061
Sequencing grade modified trypsin	Promega	Catalog: V511X
Formic acid	Fisher Scientific	Catalog: A117-50
Sep-Pak tC18 96-well Plate, 100 mg Sorbent per Well	Waters	Catalog: 186002321
Oasis MAX 1 cc Vac Cartridge, 30 mg Sorbent per Cartridge	Waters	Catalog: 186000366
Acetonitrile, Optima LC/MS	Fisher Scientific	Catalog: A955-4
Water, Optima LC/MS	Fisher Scientific	Catalog: W6-4
Anhydrous acetonitrile	Sigma	Catalog: 271004
Trifluoroacetic acid (TFA)	Sigma	Catalog: 302031
Triethylammonium acetate buffer	Sigma	Catalog: 90358
PNGaseF	New England Biolabs	Catalog:P0705L
Triethylammonium bicarbonate (TEAB)	Fisher Scientific	Catalog: 90114
Methanol	Fisher Scientific	Catalog: A452-4
Xylene	Fisher Scientific	Catalog: X3P-1GAL
Ethanol	Sigma	Catalog: 277649-2L
Phosphate-buffered saline (PBS)	Fisher Scientific	Catalog: 10010-023
Hematoxylin	Sigma	Catalog: MHS1-100ML
Ammonium hydroxide solution	Sigma	Catalog: 338818
Dimethyl sulfoxide (DMSO)	Fisher Scientific	Catalog: 85190
n-Dodecyl-β-D-Maltoside (DDM)	Fisher Scientific	Catalog: 89903
Lauryl Maltose Neopentyl Glycol (LMNG)	Anatrace	Catalog: 4219002
Tris(2-carboxyethyl) phosphine (TCEP)	Fisher Scientific	Catalog: 77720
Critical Commercial Assays
DNA FFPE Tissue Kit for DNA Extraction	Qiagen	Cat. No: 56404
SureSelect XT HS2 DNA System	Agilent	Part number: G9983A
MiSeq Reagent Kit v2 (300-cycles)	Illumina	MS-102-2002
BCA Protein Assay Kit	Thermo Fisher Scientific	Catalog: 23225
Deposited Data
Proteomic Data Commons	This paper	Proteomic Data Commons (PDC, https://pdc.cancer.gov/pdc/) under accession numbers PDC000629, PDC000630, PDC000631, PDC000632, PDC000633, and PDC000634.
Software and Algorithms
Spectronaut® v18.4	Biognosys AG45	https://biognosys.com/software/spectronaut/
R v3.6	R Development Core Team	https://www.R-project.org
NMF (R-package)	Gaujoux, R. et al.87	https://cran.r-project.org/web/packages/NMF/index.html
Webgestalt	Liao, Y. et al.97	http://www.webgestalt.org/
Next GENe and Next GENe Viewer	SoftGenetics	https://softgenetics.com/products/nextgene/
Integrated Genomics Viewer	Broad Institute	https://igv.org
Genome Aggregation Database (gnomAD)	Broad Institute	https://gnomad.broadinstitute.org
Instrument
Versette Liquid Handling System	Thermo Fisher Scientific	N/A
timsTOF-HT	Bruker	https://www.bruker.com/en/products-and-solutions/mass-spectrometry/timstof/timstof-ht.html
EVOSEP One	EVOSEP	https://www.evosep.com
MiSeq Instrument	Illumina	https://www.illumina.com/systems/sequencing-platforms/miseq.html
Ventana Discovery Ultra Research Staining System	Roche	N/A
LMD7000 system	Leica	N/A