Peripheral blood samples were collected from healthy controls and GAC patients at Shanghai Tenth People’s Hospital. Cohort 1 consisted of 19 controls and 33 patients, while Cohort 2 consisted of 18 controls and 12 GAC patients. The demographic data and staging information are listed in Table 1 . To prepare serum samples, venal blood was drawn and placed at room temperature for 30 minutes, then centrifuged at 2000 × g for 10 min. Serum was collected and stored at -80°C until use.

EV was isolated from serum samples using ultracentrifugation (UC). Briefly, 500 μL of serum was centrifuged at 2500 × g for 10 min (4°C) followed by another centrifugation at 10000 × g for 30 min to pellet cell debris. The supernatant was then filtered through a 0.22-μm cellulose acetate centrifuge filter (Costar, USA), and the filtrate was diluted with PBS into a final volume of 5 mL. Crude EVs were pelleted by ultracentrifugation at 110,000 × g (P70AT rotor, Hitachi, Japan) for 5 h. Afterwards, the pellets were resuspended with PBS and ultracentrifuge at 110,000 × g for 70 min. The final EV pellets were resuspended with 50 μL of PBS and stored at -80°C for further analysis.

For transmission election microscopy (TEM), 10 μL of PBS-diluted EV samples were added on top of the copper grids and incubated for 10 min at room temperature. The grids were washed with 6 μL of ultrapure water and negatively stained with 3% phosphotungstic acid for 10 min. Then, the grids were washed with ultrapure water and air-dried. Imaging was performed on a H-7700 transmission election microscope (Hitachi, Japan).

To analyze particle size, EV samples were diluted 10 times with PBS and analyzed with Zetasizer Nano S instrument (Malvern, UK) according to manufacturer’s instruction.

EV samples were lysed with RIPA buffer (150 mM NaCl, 0.5% SDC, 0.1% SDS, 50 mM Tris/HCl, 1% Triton-X 100, pH 7.6) and 20 μL of each lysed sample was separated on 12% SDS-PAGE. For immunoblotting, proteins were transferred onto a nitrocellulose membrane and incubated with anti-CD9 antibody (1:1000; Cat. #ab92726, Abcam, UK) and anti-Hsp70 antibody (1:1000; Cat. #ab181606, Abcam, UK) followed by anti-rabbit IgG secondary antibody (1:10000; Cat. #09-0034, Yeason, China). For silver staining, samples were washed with 50% methanol, 5% methanol and pure water successively, and then reduced by 0.0005% dithiothreitol (DTT; Cat. #43217, Sigma-Aldrich, USA) followed by incubation with 0.1% silver nitrate in the dark for 20 min. Finally, 3% sodium carbonate with 0.01% formaldehyde was applied for visualization.

BCA kit (Cat. #23225, Thermo Science, USA) was used to determine the protein concentration in isolated EVs samples. From each sample, 20 μg protein was vacuum dried and resuspended in denaturing solution (7 M urea, 2 M thiourea, 10 mM DTT, 1 × protease inhibitor [Cat. # P8340, Sigma-Aldrich, USA]). The samples were then reduced for 30 min at 55°C, alkylated with 15 mM iodoacetamide (Cat. # I1149, Sigma-Aldrich, USA) in the dark for 20 min, diluted with 50 mM ammonium bicarbonate solution and digested with trypsin (1:50; Cat. # V5113, Promega, USA) overnight at 37°C. The resulting peptides were desalted with C18 column and vacuum dried for mass spectrometry analysis.

Protein digests were analyzed on an EASY-nLC 1000 LC (Thermo Science, USA) coupled with Q-Exactive mass spectrometer (Thermo Science, USA). The mobile phases consisted of buffer A (2% ACN, 0.1% formic acid) and buffer B (98% ACN, 0.1% formic acid). Tryptic peptides were resuspended in buffer A and spiked with iRT peptides (Omicsolution, China). Equivalent to 1 μg of protein digest from each sample was loaded onto a C18 column (Cat. #164534, Thermo Science, USA) linked with a pre-column (Cat. #164535, Thermo Science, USA) and separated at a flow rate of 250 nL/min. A 120 min gradient from 3% to 8% buffer B in 5 min, 8% to 28% in 95 min, 28% to 95% in 10 min, 95% for 5 min, 95% to 3% in 2 min and 3% for 3 min was used. The MS instrument was operated in the positive polarity and profile mode with a nano-electrospray through a heated ion transfer tube with a temperature setting of 275°C. For data dependent acquisition (DDA), one full scan MS from 400 to 1400 m/z followed by 12 MS2 scan were continuously acquired. MS spectra were acquired with resolution of 70000 for a maximum injection time (IT) of 100 ms with an automatic gain control (AGC) target value of 3e6. MS2 spectra were obtained in the higher-energy collisional dissociation (HCD) mode using a normalized collision energy of 27%, resolution at 17500, maximum IT of 60 ms, AGC target of 5e5 and isolation window at 2.0 m/z. For data independent acquisition (DIA), isolation window for MS2 was set to 20 Da with 1 Da overlap over a precursor mass window of 450~950 m/z, and other parameters were set to be the same as DDA method.

Qualitative analysis of DDA raw files was performed by Proteome Discoverer (version 2.0) software searching against the UniProtKB database (2020 release, Homo sapiens) including the 11 synthetic iRT peptides. A maximum of 2 missed cleavages were allowed for trypsin digestion with fixed carbamidomethylation (+57.0251 Da) of cysteine and oxidation (+15.9949 Da) of methionine. The mass tolerance allowed was 15 ppm for precursor ions and 0.05 Da for fragmentation ions. A false discovery rate (FDR) of 1% was set at both peptide and protein levels.

Skyline (version 20.2.0.343) software was used for independent proteome spectral library construction. Briefly, fasta and pdresult files were imported with the following parameters: structural modifications: carbamidomethyl (C), oxidation (M), acetyl (N-term); minimum length: 6; maximum length: 30. The retention time of iRT was calibrated and the isolation scheme was set based on the isolation window of DIA MS parameter. Target list was added to obtain information of peptides and proteins. Finally, the peptides with reversed sequence were added as decoy peptides for the control of false discovery rate.

DIA raw files were imported into skyline and analyzed based on the aforementioned DDA spectral library, and filtered by a mProphet scoring model trained with decoy peptides. Q value and dot products were set to 0.01 and 0.65 respectively for selecting peptides with high confidence. Afterwards, the decoy peptides were removed. The exported files were used for statistical analysis.

RStudio (version 1.3.1073) was used to perform all the statistical analysis, including evaluation of data quality, data preprocessing, differential expression analysis, principal component analysis (PCA), construction of classification models. For differential expression analysis, p value < 0.05 were considered statistically significant, and fold change > 1.5 or < 0.67 were considered as up- or down-regulated, respectively. Venn diagram and functional annotation were generated with FunRich (version 3.3.1) software.

The design for EV-based GAC biomarker discovery is shown in Figure 1A . DIA-based quantitative mass spectrometry (12) analysis of 52 samples from the first cohort was conducted and candidate biomarkers was screened through differential protein expression analysis. Logistic regression classification was applied to identify a panel of candidate proteins whose expression were associated with GAC. These biomarkers were further validated by a second cohort of 30 samples.

EV was isolated from the serum using ultracentrifugation (UC). Western blot analysis detected classical EV markers Hsp70 and CD9 in UC fractions, indicating successful enrichment of EV ( Figure 1B ). Particle analysis of randomly selected EV samples showed that the majority of the isolated EV particles ranged between 10~100 nm ( Figure 1C ), which is consistent with the range distribution of exosomes. Furthermore, the morphology of the EV particles as visualized by TEM showed a typical cup-shaped structure with the size between 50~200 nm ( Figure 1D ).

For quantitative proteomic profiling, EV samples from 19 healthy subjects and 33 GAC patients were analyzed using DIA-based mass spectrometry. To assess the quality of our data as the result of a complex procedure of sample collection and handling, we interspersed quality control (QC) samples during mass spectrometry data acquisition. The distribution of the correlation coefficients of all the QC samples was between 0.71 and 0.92 ( Figure 2A ), indicating reasonable reproducibility. Signal intensity of mass spectrum spanned a dynamic range of six orders of magnitude, with the majority of precursor mass accuracy within ± 5 ppm ( Figures 2B, D ), indicating that our analysis achieved high accuracy and depth.

In total, we identified 448 proteins and quantified 352 proteins from 52 EV samples from cohort 1 ( Figure 2C ). We also performed EV enrichment on cohort 2 of 12 GAC patients and 18 healthy controls, followed by quantitative proteomic analysis, resulting in quantification of 321 proteins. A total of 249 proteins were quantified in both cohorts ( Supplemental Figure 1A ). Gene ontology analysis showed that more than 65% of these proteins localized in extracellular region and exosomes ( Supplemental Figure 1C ), further confirming the successful enrichment of EV.

Comparing the EV proteome profiles of GAC patients and healthy controls from cohort 1, we found a total of 26 significantly changed proteins, among which 13 were up-regulated and 13 down-regulated ( Figure 3A ). The fold change of up-regulated proteins had a wider range compared to that of down-regulated proteins ( Figure 3B ). The heatmap of differentially expressed proteins displayed distinct patterns, with gender, age and TNM stages of GAC were displayed together ( Figure 3C ). GO analysis showed that the differentially expressed proteins mainly involved in protein activation cascade, hydrogen peroxide catabolic process, regulated exocytosis, hemostasis, acute-phase reaction, response to bacterium ( Figure 3D ). Regulated exocytosis is highly correlated with the secretion of EVs, and contains up-regulated proteins including apolipoprotein B (APOB), haptoglobin (HP), hemoglobin subunit alpha (HBA1) and hemoglobin submit beta (HBB). On the other hand, the majority of the down-regulated proteins involved in complement and coagulation cascade including von Willebrand factor (VWF), coagulation factor XIII A chain (F13A1) and component 6 (C6).

To identify biomarkers with an increased accuracy to differentiate GAC from normal subjects, we performed multi-group differential protein expression analysis using the proteomic data from cohort 1. In addition to comparing GAC patients with healthy controls, patients with early-stage GAC (stage I + II, N=17) and late-stage GAC (stage III + IV, N=16) were also compared to heathy controls ( Figures 4A, B ), and the proteins with high consistency of expression trend were selected as candidate markers. Venn diagram showed the aforementioned three comparisons ( Supplemental Figure 2 ). In total, there are 23 intersecting proteins, of which 15 are EV proteins ( Figure 4C ). These 15 proteins were then used as candidate serum EV biomarkers for GAC diagnosis ( Table 2 ).

For construction of GAC diagnostic models, panels containing 2 to 6 proteins were randomly selected from the 15 candidate proteins through an exhaustive method, and resulting in a total of 9828 combinations. Using cohort 1 as the training set, we built a logistic regression classification model for each panel and calculate the sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). To reduce the false negative predictions, we set the sensitivity greater than 0.9 and retained 2774 classifiers.

We then used Cohort 2 as the testing set to assess the classification accuracy of the models, and found a five-protein panel containing glycine N-acyltransferase (GSN), transferrin receptor protein 1 (TFRC), alpha-1-acid glycoprotein 1 (ORM1), haptoglobin (HP), and polymeric immunoglobulin receptor (PIGR) that showed the best classification performance. This panel of classifiers showed an accuracy, sensitivity, NPV and AUC of 0.97, 0.93 and 0.93 respectively in the training set ( Figures 4E, F ). In the validation set, the sensitivity, NPV and AUC were all above 0.8, indicating that the classifier maintained a good classification performance on new data set. The parameter of the logistic regression model is displayed in Table 3 . Principal component analysis ( Figure 4D ) also confirmed the effectiveness of the five proteins in distinguishing GAC from healthy samples.

Since lymph node is a frequent tumor metastatic site, lymph node metastasis (LNM) is highly informative in selection of treatment strategies (13, 14). At present, the most commonly used blood-based diagnostic markers for GAC in clinic usage are the universal tumor markers carcinoembryonic antigen (CEA), carcinoembryonic antigen (CA19-9, CA72-4, CA24-2, CA50, CA125), and alpha-fetoprotein (AFP) (15–17). Therefore, we used CEA, CA19-9, and AFP as a panel to construct a classifier for diagnosis of advanced GAC, since these proteins are routinely measured in our patients. Based on the patient information in cohort 1, we combined patients in stage I and II as non-LNM group, and patients in stage III + IV as LNM group. In order to distinguish LNM from non-LNM groups, logistic regression was applied using cohort 1 as training set and cohort 2 as validation set, which resulted in a sensitivity, NPV and AUC of 0.66, 0.50 and 0.75 in validation set ( Figures 5C, D, G and Table 4 ).

In contrast, we used our EV proteomic data to discover potential biomarkers for diagnosis of advanced GAC. We performed differential protein expression analysis comparing patients with stage III + IV (LNM) to that of stage I + II (non-LNM), as shown in Figure 5A . We identified 6 differentially expressed proteins including 3 up-regulated proteins and 3 down-regulated proteins ( Figure 5B and Table 5 ), among which serum amyloid A (SAA1) and immunoglobulin heavy constant alpha 2 (IGHA2) play key roles in receptor-mediated endocytosis. These 6 proteins were used as candidate biomarkers, from which 2 to 5 proteins were randomly selected as panels to construct logistic regression classifiers. Fifty-six classifiers were constructed and trained with cohort 1 to evaluate the classification performance. Then 15 panels were retained with a cutoff of 0.9 for sensitivity and NPV. Applying these classifiers to cohort 2, we found an EV protein panel consisting of lysozyme (LYZ), SAA1, and coagulation factor XII (F12) that showed the best performance, with a sensitivity of 1, NPV of 1 and AUC of 0.83 ( Figures 5E, F, H and Table 6 ).