A Large-Scale Proteomics Resource of Circulating Extracellular Vesicles for Biomarker Discovery in Pancreatic Cancer

Proteomics Characterization of Circulating EVs

In this study, we sought to identify proteins in extracellular vesicles in the blood that may be used as biomarkers for the diagnosis and prognosis of pancreatic cancer. With the approval of our institutional review board, we enrolled a total of 124 patients to the discovery cohort of this biomarker study (Methods and Supplementary Table 1). Subjects in the pancreatic cancer group (N=93) had a mean age of 66.5 years (range, 37-91), and 48.4% were female. All subjects had biopsy-proven disease. Thirty subjects had early-stage disease (stages I-II) and 63 had advanced disease (stages III-IV). Patients with benign pancreatic diseases included chronic pancreatitis (N=12) with a mean age of 57.5 years (range 37-78) and with 50% females, whereas IPMN included individuals with main duct and side branch IPMNs (N=8) with a mean age of 68.2 years (range 50-89) and with 87.5% being females. Subjects in the healthy control group (N=11) had a mean age of 53.4 years (range, 31-83) with 54.5% females (Supplementary Table 1).

We employed the novel EVtrap method (Extracellular Vesicles Total Recovery And Purification) to capture EVs from plasma samples and overcome the traditional laborious techniques for EV isolation, which are not scalable for large clinical studies. As described in recent reports, EVtrap is a magnetic bead-based isolation method that enables highly efficient capture of EVs from biofluids, confirmed by multiple common EV markers^19-23. Previous analyses using electron microscopy and nanoparticle tracking also confirmed that the vast majority of particles isolated by EVtrap had diameters between 100-200 nm, consistent with exosomes¹⁹. In addition, EVtrap isolates demonstrates higher abundance of CD9, a common exosome marker, as compared to isolates from other traditional EV isolation methods such as size exclusion chromatography and ultracentrifugation¹⁹. Over 95% recovery yield can be achieved by EVtrap with less contamination from soluble proteins, a significant improvement over current commercially available methods as well as ultracentrifugation^19,21,24.

Following EV isolation, samples were digested in-solution and analyzed by liquid chromatography-tandem mass spectrometry (nanoLC-MS/MS) on a high-resolution mass spectrometer (Q-Exactive HF-X). The workflow for cEVs isolation and enrichment and subsequent cEV mass spectrometry analysis is illustrated in Figure 1A.

• Download figure
• Open in new tab

Figure 1. Study Design

The discovery cohort was comprised of 124 individuals, including pancreatic ductal adenocarcinoma (PDAC, N=93), chronic pancreatitis (CP, N=12), intraductal papillary mucinous neoplasm (IPMN, N=8) and healthy controls (N=11). Plasma samples were processed for EV isolation using EVtrap and analyzed by liquid chromatography-tandem mass spectrometry (LC-MS/MS).

First, to confirm that EVtrap can efficiently isolate extracellular vesicles from plasma, a test plasma sample was processed to remove platelets and other large particles and enriched for EVs using EVtrap beads (see methods for details). Transmission electron microscopy (TEM) analysis of the EV pellet showed cup-shaped extracellular vesicles (exosomes and microvesicles) (Supplementary Figure 1A), and nanoparticle tracking analysis (NTA) using ZetaView instrument (Particle Metrix) demonstrated that the isolated EVs were in the 100−200 nm diameter range, with a mean diameter of 152 nm (Supplementary Figure 1B). Second, to assess the technical reproducibility of the EV proteomics approach, the test plasma sample was processed in six replicates and Pearson correlation analysis revealed a very high correlation (r2 > 0.97) between replicates (Supplementary Figure 2A, Supplementary Table 2). These results provided the confidence to proceed with the analysis of our discovery set of plasma from 124 subjects. In this cohort, we identified 1,708 unique proteins (Supplementary Table 3). The number of unique EV proteins detected per 100µL of plasma sample varied from 817 to 1,128, with an average of 912 unique proteins per sample (Supplementary Figure 2B). We did not observe differences between non-tumor and tumor samples regarding the overall number of EV proteins identified. Within the PDAC group, we did not observe significant differences in the average number of EV proteins detected for different disease stages. Collectively, these data demonstrate high reproducibility of EV isolation and robust label-free MS quantification of cEVs.

Diseases of the Pancreas Express Distinct Circulating EV Proteome Compared to Controls

Next, we aimed at identifying specific cEV proteins associated with clinical parameters with the potential to serve as diagnostic biomarkers. We first compared the proteomics profile of individuals with underlying pancreatic diseases (PDAC, chronic pancreatitis and IPMN) against healthy controls. We selected EV proteins expressed in at least 50% of subjects in the disease group with a fold change of expression ≥2 or ≤2 compared to controls and p-value ≤0.01 after adjusting for multiple testing. A total of 207 proteins were identified that met the criteria, with the largest number of differentially expressed markers in PDAC (176), followed by chronic pancreatitis (55) and IPMN (3) (Supplementary Table 4). Principal component analysis (PCA) of these markers showed control samples as a tight cluster segregated away from PDAC samples but closer to IPMN and chronic pancreatitis patients (Figure 2A).

• Download figure
• Open in new tab

Figure 2. Identification of cEV Proteins Differentially Expressed in Disease Groups

(A) Principal component analysis of cEV proteins differentially expressed in the plasma of patients with pancreatic diseases compared to controls. Each dot indicates one individual enrolled in the study: green, controls; blue, patients with intraductal papillary mucinous neoplasm (IPMN); purple, patients with chronic pancreatitis (CP); salmon, early stage (stages I and II) pancreatic ductal adenocarcinoma (PDAC); red, late stage (stages III and IV) PDAC.

(B) Volcano plot of circulating EV proteins enriched in the plasma of patients with PDAC versus benign pancreatic diseases. X-axis, log base 2 of fold changes; Y-axis, negative of the log base 10 of p values

(C) Heatmap of cEV proteins differentially expressed in the plasma of patients with pancreatic diseases compared to controls. Designations of clinical parameters were indicated at the top of the heatmap.

(D) Expression of enriched cEV proteins in patients with PDAC versus benign pancreatic diseases. Each dot indicates the target protein signal from one patient. Y-axis, normalized log base 2 of protein signals detected by mass spectrometry; Error bars, min and max values; lines in boxes, median values. * p ≤ 0.05, ** p ≤ 0.01, *** p ≤ 0.001, **** p ≤ 0.0001.

Circulating EV Proteome Discriminates Pancreatic Cancer from Benign Pancreatic Diseases

To further assess the potential of cEV proteins for cancer detection, we compared proteomic profiles of cEVs between patients with PDAC with those with underlying benign diseases of the pancreas (chronic pancreatitis and IPMN). We identified 182 differentially expressed proteins in malignant cases (92 over-expressed and 90 with reduced expression) (Supplementary Table 5). Several of those markers had remarkable overexpression in PDAC (greater than 10-fold), including PDCD6IP, SERPINA12, RUVBL2, among others, as shown in the volcano plot (Figure 2B). Unsupervised clustering showed a clear separation between PDAC and benign pancreatic diseases. Individuals with IPMN were more closely related to controls, whereas chronic pancreatitis cases were more related to PDAC (Figure 2C). In addition, the PDAC cohort was separated into two subgroups: the first, enriched for early-stage tumors and more closely related to the other pancreatic diseases (chronic pancreatitis and IPMN); the second, enriched for advanced and metastatic cases with expression profiles further apart from early-stage cancer and pancreatic diseases (Figure 2C). We further noticed that some proteins such as PDCD6IP, SERPINA12, KRT20 showed statistically significant population-wise enrichment in pancreatic cancer compared to benign pancreatic diseases (Figure 2D, Supplementary Figure 3). Together, these data indicate the existence of EV markers that can separate controls, benign and malignant pancreatic diseases, as well as proteins that separate early versus late-stage PDAC, suggesting their potential to serve as diagnostic biomarkers.

Functional and Systems Biology of cEV Proteome

To gain molecular insight into the functions of the 182 proteins differentially expressed in pancreatic cancer as compared to benign pancreatic diseases, we conducted pathway analysis using the Gene Ontology (GO) and REACTOME databases (Supplementary Table 5). We identified protein modules in protein localization, biomolecule binding/docking, peptidase activities among changes enriched in PDAC compared to benign diseases (Supplementary Figure 4). Interestingly, KRT20 (keratin 20), a gastrointestinal epithelia-associated keratin, was increased in PDAC patient EVs, while keratins associated basal cells, KRT4, KRT15, and KRT3, were reduced. KRT20 overexpression is frequently found in pancreatic tumor tissues and correlates with poor prognosis²⁵, suggesting a biological basis for their high levels in the cEVs of PDAC patients.

Interestingly, proteins associated with immunological functions showed complex regulation with increased representation of leukocyte mediated immunity (GO:0002443), leukocyte degranulation (GO:0043299), myeloid leukocyte activation (GO:0002274), and decrease in Fc receptor signaling (GO: 0038093), regulation of complement activation (GO:0030449), and immune effector process (GO:0002252) (Supplementary Table 5). These data suggest that direct profiling of cEVs from patient plasma provided unique insights into systemic changes in immune biology during pancreatic cancer development, which is lacked in analysis restricted to tissue or cell models.

Circulating EV Proteomics Reveal Markers Associated with Metastasis and Worse Prognosis

We then investigated whether cEV proteins can assist in the distinction of metastatic versus non-metastatic pancreatic cancer. We compared the cEV proteome profiles of individuals with metastatic cancer to those without metastasis and identified 85 proteins differentially expressed between the two groups (Supplementary Table 6). Supervised clustering between metastatic and non-metastatic diseases showed a clear separation with two distinct expression patterns (Figure 3A). In particular, PSMB4, RUVBL2 and ANKAR (Figure 3B) EV protein levels were increased in patients with metastatic disease, whereas RAP2B, SERPINA12 and IGLV4-69 abundance levels were decreased in the cEVs of patients with metastasis (Figure 3C). Together, these findings suggest the presence of a core set of cEV proteins with the potential to distinguish early versus metastatic pancreatic cancer.

• Download figure
• Open in new tab

Figure 3. Circulating EV Proteomics Reveal Markers Associated with Metastasis and Worse Prognosis.

(A) Heatmap showing EV proteins differentially expressed in the plasma of metastatic versus non-metastatic PDAC. Designations of clinical parameters are indicated at the top of the heatmap.

(B) Expression patterns of cEV proteins associated with metastatic disease. Y-axis, normalized log base 2 of protein signals detected by mass spectrometry; N, non-metastatic PDAC group; M, metastatic PDAC group. Each dot indicated the target protein signal from one patient. Error bars, min and max values; lines in boxes, median values. * p ≤ 0.05, ** p ≤ 0.01, *** p ≤ 0.001, **** p ≤ 0.0001.

(C) As is (B), except for cEV markers with increased expression in non-metastatic PDAC.

(D) Correlation of cEV marker expression with survival. Kaplan–Meier curves and log-rank test P values of representative survival cEV markers quantified in the discovery cohort.

We further analyzed whether the expression of certain cEV proteins had prognostic relevance in our cohort. We first classified individuals with PDAC as having low or high expression of any given markers based on each marker’s first and third quartile. Survival was estimated by the Kaplan Meier method. We identified that the cEV expression of RALB, CRP, and CD55 had a significant correlation with overall survival, with a trend for PDCD6IP (Figure 3D).

Validation of cEV Markers Using Parallel Reaction Monitoring and Identification an EV Protein Signature for Pancreatic Cancer Diagnosis

Because pancreatic cancer is extremely heterogeneous, the chance of identifying a single biomarker with sufficient diagnostic performance is likely low. Instead, the identification of a panel of candidate markers may have enhanced diagnostic performance.

To identify a signature that shows the most discriminatory power between ‘benign diseases’ and ‘PDAC,’ we employed a binary classification approach using Support Vector Machines (SVM). Classification models, built based on a large number of proteins, contain irrelevant markers that can reduce the predictive accuracy. Hence, we implemented a consensus feature selection method based on two algorithms: one using recursive feature elimination (RFE) algorithm (SVM-RFE)²⁶ and second, RFE combined with a non-parametric Wilcox rank test (sigFeature)²⁷. The top 16 markers were selected whose classification performance can be tested in the independent validation cohort (Supplementary Table 7). A summary of selection process is shown in Supplementary Figure 5. The classification performance of these 16 markers, individual and in all combinations, were tested using 80% training data and evaluated in the remaining 20% test data. The quality of training was assessed using five repetitions of 10-fold cross-validation. The optimal kernel parameters were estimated by tuning over a wide range of values. Receiver operating characteristic (ROC) analysis was used as the metric to assess the performance of the classifier model. We found a set of 7-EV protein signature comprised of RUVBL2, PDCD6IP, ATP5F1, DLD, KRT20, CCT4, and SERPINAI2, that gave 100% accuracy when tested in the discovery cohort (Supplementary Table 7, Supplementary Figure 6). Recurrence of these putative markers in our dataset varied from 55% to 97%.

The model was further validated on an independent validation cohort whose proteome was obtained using an alternate technology, parallel reaction monitoring (PRM) mass spectrometry. The markers chosen for validation included 16 markers selected for SVM classification model and an additional 9 markers to result in top 25 markers that are significantly differentially expressed in the discovery cohort with a fold change increase in PDAC ≥ 5.5 and p-value ≤0.01 (Methods, Supplementary Table 8). The independent validation cohort consisted of 36 new subjects (24 with PDAC, 6 with chronic pancreatitis, and 6 with IPMN) (Supplementary Table 9). A total of 10 proteins, including all 7 signature proteins, showed a significant difference (p < 0.05) in patients with PDAC as compared to benign pancreatic diseases (Figure 4A). The performance of individual validated markers according to the specific underlying disease in the validation cohort is presented in Supplementary Figure 7. The performance of 7-EV protein signature was further tested using SVM model, in our independent validation cohort, yielding an 89% prediction accuracy (Figure 4B, Supplementary Figure 8). As expected, we observed that no single marker achieved sufficiently high sensitivity and specificity as the combined model for the diagnosis of pancreatic cancer.

• Download figure
• Open in new tab

Figure 4. Validation of cEV Markers and Identification of 7-EV Protein Signature for PDAC Diagnosis.

(A) Differences of cEV protein abundances between patients with PDAC (n=24) and benign pancreatic diseases (chronic pancreatitis and IPMN) (n=12). x axis, minus log p values of protein abundance differences between PDAC and benign groups; y axis, average fold changes of proteins in PDAC group compared to benign group. Size of bubbles indicate average protein abundances in PDAC group. Pink color, proteins that had at least two-fold enrichment in PDAC group (p<0.05).

(B) ROC curves were calculated for individual cEV markers as well as for the 7-EV protein PDAC signature combination to determine optimum diagnostic performance. AUC, area under the curve.

Study Design and Patient Demographics

We conducted this study at Beth Israel Deaconess Medical Center with the approval of the Harvard Cancer Center Institutional Review Board. All subjects provided written informed consent. Clinical data and blood samples were prospectively collected from 2017 to 2019 from patients with pancreatic cancer, chronic pancreatitis, intraductal papillary mucinous neoplasms (IPMN), and age-matched controls. A total of 124 patients, including PDAC (N=93), chronic pancreatitis of different etiologies (N=12), IPMN (N=8), and controls (N=11), were included in the discovery cohort. PDAC diagnosis was established by histology or cytology, and staging was performed according to the American Joint Committee on Cancer guidelines (8th Edition 2016)³⁸(Supplementary Table 1). For the independent validation cohort, a total of 36 patients were enrolled, including PDAC (N=24), IPMN (N=6), and chronic pancreatitis (N=6) (Supplementary Table 9).

Plasma Sample Collection and Processing

All blood samples were collected and processed following the same standard operating procedure optimized for EV analysis and included the following steps: (i) whole blood was collected into one 10ml yellow-top tube containing acid citrate dextrose; (ii) blood was mixed by gently inverting the tube five times; (iii) vacutainer tubes were stored upright at room temperature (RT); (iv) samples were centrifuged at 1,300g for 15 min in RT; (v) plasma was removed from the top carefully avoiding cell pellet; (vi) repeat centrifugation of plasma at 2,500g for 15 min in RT; (vii) again, plasma was removed from the top carefully avoiding cell pellet; (viii) third centrifugation at 2,500g for 15 min in RT, then samples were aliquoted to be stored at -80°C.

Extracellular Vesicle Isolation from Plasma

We employed EVtrap for EV isolation from plasma samples¹⁹. EVtrap beads were provided by Tymora Analytical (West Lafayette, IN) as a suspension in water and were used as previously described in more details^19,20. Briefly, 100 μL plasma samples were diluted 20 times in the diluent buffer, the EVtrap beads were added to the samples in a 1:2 v/v ratio, and the samples were incubated by end-over-end rotation for 30 min according to the manufacturer’s instructions. After supernatant removal using a magnetic separator rack, the beads were washed with PBS, and the EVs were eluted by a 10 min incubation with 200 mM triethylamine (TEA, Millipore-Sigma). The samples were fully dried in a vacuum centrifuge.

Preparation of EV samples

The isolated and dried EV samples were lysed to extract proteins using the phase-transfer surfactant (PTS) aided procedure. The proteins were reduced and alkylated by incubation in 10 mM TCEP and 40 mM CAA for 10 min at 95°C. The samples were diluted fivefold with 50 mM triethylammonium bicarbonate and digested with Lys-C (Wako) at 1:100 (wt/wt) enzyme-to-protein ratio for 3 h at 37°C. Trypsin was added to a final 1:50 (wt/wt) enzyme-to-protein ratio for overnight digestion at 37°C. To remove the PTS surfactants from the samples, the samples were acidified with trifluoroacetic acid (TFA) to a final concentration of 1% TFA, and ethyl acetate solution was added at a 1:1 ratio. The mixture was vortexed for 2 min and then centrifuged at 16,000 × g for 2 min to obtain aqueous and organic phases. The organic phase (top layer) was removed, and the aqueous phase was collected. This step was repeated once more. The samples were dried in a vacuum centrifuge and desalted using Top-Tip C18 tips (Glygen) according to the manufacturer’s instructions. The samples were dried completely in a vacuum centrifuge and stored at -80°C.

LC-MS Analysis of Plasma EV Proteome

Approximate 1 μg of each dried peptide sample was dissolved in 10.5 μL of 0.05% trifluoroacetic acid with 3% (vol/vol) acetonitrile containing spiked-in indexed Retention Time Standard containing 11 artificially synthetic peptides (Biognosys). The spiked-in 11-peptides standard mixture was used to account for any variation in retention times and to normalize abundance levels among samples. 10 μL of each sample was injected into an Ultimate 3000 nano UHPLC system (Thermo Fisher Scientific). Peptides were captured on a 2-cm Acclaim PepMap trap column and separated on a heated 50-cm Acclaim PepMap column (Thermo Fisher Scientific) containing C18 resin. The mobile phase buffer consisted of 0.1% formic acid in ultrapure water (buffer A) with an eluting buffer of 0.1% formic acid in 80% (vol/vol) acetonitrile (buffer B) run with a linear 60-min gradient of 6–30% buffer B at a flow rate of 300 nL/min. The UHPLC was coupled online with a Q-Exactive HF-X mass spectrometer (Thermo Fisher Scientific). The mass spectrometer was operated in the data-dependent mode, in which a full-scan MS (from m/z 375 to 1,500 with the resolution of 60,000) was followed by MS/MS of the 15 most intense ions (30,000 resolution; normalized collision energy - 28%; automatic gain control target (AGC) - 2E4, maximum injection time - 200 ms; 60sec exclusion).

EV Proteome Data Processing

The raw files were searched directly against the human Swiss-Prot database with no redundant entries using Byonic (Protein Metrics) and Sequest search engines loaded into Proteome Discoverer 2.3 software (Thermo Fisher Scientific). MS1 precursor mass tolerance was set at 10 ppm, and MS2 tolerance was set at 20ppm. Search criteria included a static carbamidomethylation of cysteines (+57.0214 Da) and variable modifications of oxidation (+15.9949 Da) on methionine residues and acetylation (+42.011 Da) at the N terminus of proteins. The search was performed with full trypsin/P digestion and allowed a maximum of two missed cleavages on the peptides analyzed from the sequence database. The false-discovery rates of proteins and peptides were set at 0.01. All protein and peptide identifications were grouped, and any redundant entries were removed. Only unique peptides and unique master proteins were reported.

All data were quantified using the label-free quantitation node of Precursor Ions Quantifier through the Proteome Discoverer v2.3 (Thermo Fisher Scientific). For the quantification of proteomic data, the intensities of peptides were extracted with initial precursor mass tolerance set at 10 ppm, a minimum number of isotope peaks as 2, maximum ΔRT of isotope pattern multiplets – 0.2 min, PSM confidence FDR of 0.01, with hypothesis test of ANOVA, maximum RT shift of 5 min, pairwise ratio-based ratio calculation, and 100 as the maximum allowed fold change. The abundance levels of all peptides and proteins were normalized to the spiked-in internal iRT standard. For calculations of fold-change between the groups of proteins, total protein abundance values were added together, and the ratios of these sums were used to compare proteins within different samples.

The abundances of EV proteins were normalized using indexed retention time (iRT) in Proteome Discoverer (ThermoFisher Scientific). Abundances were categorized into four different categories: Control, Chronic Pancreatitis, IPMN, and PDAC. Protein abundances were then log2 transformed and quantile normalized for further analysis.

A non-parametric Wilcox Rank Sum test was performed to test the null hypothesis that the distributions of two groups of the patient population are the same, and the fold change and p-values for each protein were estimated for the following comparisons: IPMN vs. Control, CP vs. Control, PDAC vs. Control, Benign Pancreatic Diseases (CP, IPMN) vs. PDAC. Multiple testing correction was done using Benjamini-Hochberg method to control for the false discovery rate³⁹. Volcano plots were created using those p values and fold change. Heatmaps visualization and clustering of statistically significant proteins, with adjusted p-value ≤ 0.05 and absolute fold change ≥ 2, were created in R using the pheatmap package. Euclidean distance and average cluster method were used. The values were row-scaled for normalization. Both rows and columns were allowed to cluster.

Pathways Enrichment and Protein Network Analysis

Pathway enrichment analysis was performed on statistically significant genes using g:Profiler⁴⁰, a web-based tool that searches for pathways whose genes are significantly enriched in our dataset compared to a collection of genes representing Gene Ontology (GO) terms and Reactome pathways. We further used EnrichmentMap⁴¹, a Cytoscape, v3.8.2⁴² application to create a visual network of connected pathways that helps to identify relevant pathways and theme⁴³. A Protein-Protein interaction network was generated using a stringApp, a Cytoscape app. This application allows to import STRING networks into Cytoscape and enables to perform complex network analysis and visualization of networks⁴⁴.

Parallel Reaction Monitoring and Data Analysis

Parallel reaction monitoring mass spectrometry (PRM-MS) was employed for validation experiments. Twenty-five cEV markers were selected for validation based on fold change increase ≥5.5, p-value ≤0.01, and technical aspects (number of unique peptides and coverage) (Supplementary Table 8). Thirty-six plasma samples from a new cohort were used for the validation (24 PDAC, 6 IPMN and 6 chronic pancreatitis samples). The EVs were isolated from plasma and the proteins processed as described before. Peptide samples were dissolved in 10.8 μL 0.05% TFA & 2% ACN, and 10 μL injected into the UHPLC coupled with a Q-Exactive HF-X mass spectrometer (Thermo Fisher Scientific). The mobile phase buffer consisted of 0.1% formic acid in HPLC grade water (buffer A) with an eluting buffer containing 0.1% formic acid in 80% (vol/vol) acetonitrile (buffer B) run with a linear 60-min gradient of 5–35% buffer B at a flow rate of 300 nL/min. Each sample was analyzed under PRM with an isolation width of ±0.8 Th. In these PRM experiments, an MS2 level at 30,000 resolution relative to m/z 200 (AGC target 2E5, 200 ms maximum injection time) was run as triggered by a scheduled inclusion list. Higher-energy collisional dissociation was used with 28 eV normalized collision energy. PRM data were manually curated within Skyline-daily (64-bit) 20.2.1.404 (32d27b598)⁴⁵.

Identification of EV signature for pancreatic cancer diagnosis

To identify a biomarker signature demonstrating the highest discriminatory power between ‘benign’ and ‘PDAC’ diseases, we adopted a binary classification approach utilizing Support Vector Machines (SVM). Recognizing that classification models built on an extensive array of proteins may incorporate irrelevant markers, which can diminish the predictive accuracy, we started with a list of significantly differentially expressed set of 91 proteins between ‘benign’ and ‘PDAC’ patients and further employed a consensus feature selection strategy using two algorithms, ‘Recursive Feature Elimination’ (SVM-RFE), and ‘Integrated RFE with a non-parametric Wilcox rank test (sigFeature). Subsequently, we selected the top 16 markers, the classification performance of which was subjected to testing in an independent validation cohort. The classification performance evaluation of these markers, both individually and in various combinations, involved a rigorous assessment utilizing 80% of the data for training and the remaining 20% for internal-validation. To ensure the quality of the training process, we employed five repetitions of a 10-fold cross-validation approach. The optimal kernel parameters were determined through tuning across a broad range of values. Receiver operating characteristic (ROC) analysis served as the metric to gauge the performance of the classifier model. All algorithms for identifying the EV signature predictive of pancreatic cancer diagnosis were implemented in R. We used Support Vector Machine (SVM) using CRAN package, e107⁴⁶. Ranking of genes was achieved using packages ‘sigFeature’ and ‘SVM-RFE’. An R package, ‘pROC’⁴⁷ was used to build a receiver operating characteristic curve (ROC) and to calculate the area under the curve (AUC).

Survival Analysis

The prognostic value of every protein was estimated by dividing patients into two groups: group 1, patients with expression below the 25^th percentile, and group 2, patients with expression values greater than 75^th percentile. The Kaplan-Meier estimator was used to estimate the survival function associating survival with EV protein expression, and the log-rank test was used to compare survival curves of two groups. ‘survival’ R package was used for the analysis.

Statistical Analysis

All statistical analyses were performed using the statistical software R. Statistical significance was calculated by two-tailed Student’s t-test or Wilcoxon rank-sum test unless specified otherwise in the figure legend. Data are expressed as mean ± SEM. A p-value < 0.05 in biological experiments or FDR < 0.05 after multiple comparison corrections in proteomics data analysis was considered statistically significant.