Skip to main content
NCBI home page
NPJ Precision Oncology logoLink to NPJ Precision Oncology
. 2026 Jan 21;10:65. doi: 10.1038/s41698-026-01271-x

Proteomic profiling of single extracellular vesicles as a promising new approach for the diagnosis and treatment modality of advanced ovarian cancer

Beier Wu 1,2,#, Xuping Yang 1,#, Yanling Cai 3,#, Shihan Wan 1,2, Jingfang Liu 1, Jie Xing 1, Xin Chen 1, Jiejie Zhang 1, Yanlu Jin 1, Aijun Yu 1,2,✉, Li Yang 1,✉
PMCID: PMC12905296  PMID: 41565766

Abstract

This study utilized a novel Proximity Barcoding Assay to perform high-resolution proteomic profiling of individual plasma extracellular vesicles from 85 patients with advanced high-grade serous ovarian carcinoma (OC) and 95 healthy controls (HC). Single-EV analysis identified 119 differentially expressed proteins and 17 distinct EV subpopulations. Cluster 7 (enriched in integrins ITGB3, ITGB1, and ITGA6) was significantly elevated in OC plasma (4.47% in HC vs. 14.79–15.82% in OC). Machine learning (SVM-RFE, LASSO, Random Forest) identified a diagnostic panel (ITGA6, ITGB2, ILK) achieving exceptional accuracy in distinguishing OC from HC (AUC = 0.999 training; 1.000 validation). Furthermore, risk models incorporating specific protein signatures effectively stratified patients by platinum sensitivity/resistance (9-protein panel: ILK, CDCP1, CD86, CLDN4, CLEC1B, CDHR5, CLDN11, JAM2, FOLH1), lymph node metastasis status (7-protein panel: APOE, CD28, CLDN4, FOLH1, ITGAL, JAML, ULBP3), and post-surgical residual disease burden (4-protein panel: CD44, CLMP, ITGA4, AMIGO1), with Cluster 13 (ITGB1-high) also significantly associated with residual disease. This work demonstrates the power of single-EV proteomics combined with machine learning for non-invasive diagnosis and clinical outcome assessment in advanced ovarian cancer, though the absence of early-stage patients limits its applicability for early detection.

Subject terms: Ovarian cancer, Biomarkers, Proteomic analysis

Introduction

Ovarian cancer (OC) is one of the most malignant tumors in gynecology and represents the primary cause of death among gynecological cancers1. Up to 75% of ovarian cancer cases are diagnosed in an advanced stage, with a 5-year survival rate of only approximately 20%2,3. This poor prognosis is largely due to the lack of effective screening methods and the aggressive nature of the disease. Conventional serological biomarkers, such as Cancer Antigen 125 (CA-125), have limited clinical utility4. The current standard treatment for ovarian cancer consists of cytoreductive surgery followed by platinum-based chemotherapy5. Although initial treatment responses are often favorable, most patients with advanced disease relapse within a few years and develop platinum resistance4. Owing to various diagnostic and therapeutic challenges, there is an ongoing search for effective biomarkers for prediction, diagnosis, or prognosis to improve the outcome of ovarian cancer treatment.

Ovarian cancer is not a single disease but rather a group of molecularly and histologically distinct neoplasms. The Cancer Genome Atlas (TCGA) research network has provided a comprehensive molecular landscape, primarily for high-grade serous ovarian carcinoma (HGSOC), and identified distinct transcriptional subtypes (e.g., mesenchymal, immunoreactive, differentiated, and proliferative) with potential prognostic and therapeutic implications6. This profound heterogeneity poses a significant challenge for biomarker development, as a marker effective in one histological or molecular subtype may not perform well in another. Consequently, any novel biomarker approach must ultimately demonstrate utility across or within defined pathological and molecular contexts. This study focuses exclusively on advanced-stage (III–IV) OC patients, which limits direct applicability to early-stage disease. The cohort was predominantly composed of HGSOC, the most common and lethal histologic subtype.

Extracellular vesicles (EVs) are gaining increasing attention for their roles in cancer diagnosis and treatment7. EVs are extracellular vesicles measuring 30–150 nm in diameter, encapsulated by a lipid bilayer8. They play important roles in cancer metastasis, making them promising targets for liquid biopsy9. Multiple studies have demonstrated that EVs contribute significantly to ovarian cancer progression, including proliferation, metastasis, drug resistance, and recurrence. For instance, plasma exosomal miR-19b-3p from patients with esophageal squamous cell carcinoma can suppress MAP2K3 expression to promote cancer cell proliferation10. Similarly, ovarian cancer-derived EVs carrying miR-205 have been shown to induce angiogenesis and facilitate tumor metastasis11. EVs have also been implicated in resistance to anticancer drugs12. Tumor-secreted EVs containing miR-141 can modulate tumor–matrix interactions13. Given their broad applications in diagnosis, treatment, and prognosis across various cancers, EVs hold substantial potential for use in ovarian cancer management14.

However, the molecular composition of EVs is highly heterogeneous, a critical consideration when analyzing body fluid-derived EVs for liquid biopsy or therapeutic applications9. EV membrane proteins reflect their cell of origin, enabling specific targeting and mediation of physiological or pathological effects on recipient cells15–17. We employed the Proximity Barcoding Assay (PBA), a rapid high-throughput method for single-EV analysis that can simultaneously profile over 100 surface proteins on individual EVs18. This technology allows discrimination of EV subpopulations in human plasma based on their heterogeneous surface proteome19. In this study, we used PBA to analyze 250 reported disease biomarkers at single-EV resolution, identifying differentially expressed proteins in plasma EVs from healthy individuals and ovarian cancer patients. We subsequently categorized all detected EVs into 17 distinct subgroups based on their proteomic profiles. Our analysis focused on EV subpopulation changes between platinum-resistant ovarian cancer (PROC), platinum-sensitive ovarian cancer (PSOC), and healthy controls (HC), with particular emphasis on the proteomic characteristics of different subgroups.

We undertook a comprehensive binary classification of our collected ovarian cancer patient samples utilizing three crucial clinical indicators: platinum resistance, lymph node metastasis, the presence or absence of residual lesions post-surgery, and the patients’ recurrence-free survival. Subsequently, by employing Proximity Barcoding Assay (PBA) in conjunction with machine learning, we conducted a detailed expression analysis of exosomal proteins in patient plasma. This methodology enabled us to successfully identify a series of proteins exhibiting potential disparities. Ultimately, by analyzing Receiver Operating Characteristic (ROC) curves, we assessed the potential of these exosomal proteins to serve as liquid biopsy biomarkers, with the aim of investigating their applicability as diagnostic and therapeutic targets for ovarian cancer.

Results

EV proteomic biomarker analysis on single EVs by PBA

The plasma donors enrolled in our study included ovarian cancer patients (OC, n = 85) and healthy female donors without diagnosed malignant disease (HC, n = 95). All patients underwent surgery at Zhejiang Cancer Hospital. All healthy donors were recruited from the medical examination center of Zhejiang Cancer Hospital, with all gynecological diseases excluded. Patient information is shown in Table 1. All samples were collected from patients with HGSOC. After plasma collection and EV purification, PBA was performed to analyze proteins expressed on EVs at single-EV resolution. We investigated 245 biomarkers that have been previously reported in the literature, particularly in cancer studies (listed in Table S1). After trimmed mean of M-values (TMM) normalization, we identified 119 DEPs, including 44 up-regulated and 75 down-regulated proteins. We then visualized all 119 proteins in a heatmap (Fig. 1a) and displayed the top 10 up-regulated and top 10 down-regulated proteins on a volcano plot (S2a). Compared to the HC group, EV-associated expression of ITGB1, ITGA6, CD151, and PECAM1 was increased in plasma samples from ovarian cancer patients. We also identified proteins that were significantly decreased in cancer groups, including ITGAL, CR1, CEACAM8, and ITGB2. The data and statistic analysis of protein expression are shown in S10.

Table 1.

Patients’ characteristics (n = 85)

Characteristics No. of patients (%) Characteristics No. of patients (%)
Median age (years) 53(34–75) Platinum status
Tumor residual disease Sensitive 44 (51.8)
 No residual 41 (48.2) Resistance 41 (48.2)
 ≤1 cm 34 (40.0) Tumor stage
 >1 cm 10 (11.8) III 64 (75.3)
Recurrence free survival (RFS) IV 21 (24.7)
 1–≤6 month 41 (48.2) Lymph node metastasis
 >6–≤12 month 3 (3.5) Yes 56 (65.9)
 >12 months 41 (48.2) No 29 (34.1)
Open in a new tab

In this study, platinum resistance was defined as disease progression within 6 months after completion of first-line platinum-based chemotherapy.

Fig. 1. EV subpopulations and protein profiles in OC vs. HC.

Fig. 1

Open in a new tab

a Heatmap of DEPs. b EV subpopulation alteration of all samples by FlowSOM algorithm. c EV subpopulation alteration of Platinum-resistant ovarian cancer patients, Platinum-sensitive ovarian cancer patients and healthy control. d The proteomic profiles of cluster 7.

Functional enrichment analysis was performed on the 119 DEPs. The most significantly enriched term in the biological process (BP) category was “cell-cell adhesion via plasma-membrane adhesion molecules” (S2b). In the cellular component (CC) category, the top term was “external side of plasma membrane.” For molecular function (MF), these genes were primarily associated with “integrin binding.” KEGG pathway analysis revealed associations with “Cell adhesion molecules,” “Hematopoietic cell lineage,” and “Regulation of actin cytoskeleton” (S2c). We used the STRING online database to analyze interactions between the DEPs and visualized the results using the R package “Cytoscape,” where top 30 is displayed in S2d.

To classify EVs based on their proteomic characteristics, we applied FlowSOM, an unsupervised machine learning algorithm that generates EV clusters through self-organizing maps20. EV subpopulations were visualized using t-distributed stochastic neighbor embedding (t-SNE). The average number of EVs detected by PBA, quantified via EV tags, was 2.35 × 105 per sample (S2 e), while the average number of detected proteins, counted via molecule tags, were 9.74 × 105 (S2 f). On average, 4.14 proteins were detected per EV (S2 g). EV clustering across all samples is displayed in the t-SNE plot in Fig. 1e, where FlowSOM identified 17 distinct clusters. Characteristic biomarkers for each cluster are shown in S2h.

We next analyzed phenotypic similarities and differences between OC and HC groups. The distribution of EV subpopulations was quantified and visualized in t-SNE plots for HC, PROC, and PSOC groups (Fig. 1c). Comprehensive EV subpopulation profiles for individual patient samples are presented in Fig. S3.

Among all subpopulations, Cluster 7 represented a relatively abundant subpopulation (12% in Fig. 1b) and was significantly increased in the plasma samples of OC patients, rising from 4.47% in the HC group to 15.82% and 14.79% in the platinum-sensitive and platinum-resistant OC groups, respectively (Fig. 1c). The proteomic profile of Cluster 7 is shown in Fig. 1d, allowing comparison of the expression frequency of each detected protein. This analysis revealed the relatively highly expressed biomarkers ITGB3, ITGB1, and ITGA6, consistent with our earlier observations. Given that ITGB3 and ITGA6 have been reported as specific cancer markers19,21, our results inferred that cluster 7 originates from cancer cells.

Construction of OC diagnostic model based on plasma exosomal proteins

Given the observed differences in EV subpopulation profiles between the OC and HC groups, we sought to develop a diagnostic model based on exosomal proteins. Three machine learning algorithms—SVM-RFE, Random Forest, and LASSO—were employed to identify the three most valuable biomarkers: ITGA6, ITGB2, and ILK (Figs. 2a and S4a–e).

Fig. 2. Exosomal protein-based diagnostic model for ovarian cancer.

Fig. 2

Open in a new tab

a Venn diagram of machine-learning results for DEPs. b PCA of normal and tumor samples. c ROC curve of the validation model. d–f ROC curve of ITGA6, ITGB2 and ILK.

As shown in the Supplementary Figures, ITGB2 was significantly overexpressed in normal samples (p < 0.05), whereas ITGA6 and ILK showed significant overexpression in tumor samples (S4f–h). These proteins effectively distinguished between tumor and normal samples (Fig. 2b). To evaluate the diagnostic potential of these three markers, we performed ROC curve analysis. In the training set, the areas under the ROC curve (AUCs) for ITGB2, ILK, and ITGA6 were 0.929, 0.997, and 0.991, respectively (S5b–d). The combined three-marker panel achieved an AUC of 0.999 (S5a). In the validation set, the AUC values for ITGB2, ILK, ITGA6, and their combination were 0.957, 0.974, 0.983, and 1.000, respectively (Fig. 2c–f).

Exosomal subpopulation alteration in different clinical groups

The management of advanced ovarian cancer primarily involves cytoreductive surgery and platinum-based chemotherapy. Accurate prediction of platinum sensitivity, determination of the need for neoadjuvant chemotherapy, and assessment of the required extent of lymph node dissection are therefore critical, as these factors significantly impact patient outcomes. However, current clinical practice lacks reliable biomarkers to guide these decisions. We therefore investigated whether EV subpopulations possess discriminative potential for these clinical parameters.

The EV subpopulations for PSOC and PROC groups are presented in Fig. 1c. Comparison between these two groups revealed no statistically significant differences in EV subpopulations. Using the machine learning algorithm FlowSOM, we visualized EV subpopulations associated with lymph node metastasis and residual tumor burden. Then, we analyzed the substantial phenotypic similarity and differences among these groups. Comparing three groups, including Group Non (no lymph node metastasis), Group I (lymph node metastasis rate≥50%) and Group II (lymph node metastasis rate <50%), we did not find significant difference of the quantitative distribution of EVs among subpopulations (Fig. 3a).

Fig. 3. EV subpopulation changes across clinical groups.

Fig. 3

Open in a new tab

a EV subpopulation alteration of different groups of lymph node metastasis classification. b EV subpopulation alteration of different groups of tumor residual disease classification.

Additionally, among all subpopulations, cluster 13 accounted for the highest proportion across all residual disease groups, including Group Non (no tumor residual disease), Group I (tumor residual disease ≤1 cm) and Group II (tumor residual disease >1 cm) (Fig. 3b). Cluster 13 significantly increased from 27.1% in Group NON to 33.22% in Group II. The proteomic profiles of cluster 13 were shown in S6. The association between this ITGB1-high subpopulation and residual disease is biologically well-supported, as ITGB1 is a key mediator of cancer cell adhesion to the peritoneal surface and confers resistance to chemotherapy, processes directly responsible for residual tumor burden21,22. Furthermore, across the identified subpopulations, Cluster 7 (enriched in ITGB3, ITGB1, and ITGA6) reflects a proteomic signature strongly implicated in tumor microenvironment remodeling, including promoting epithelial-mesenchymal transition23, metastatic niche formation19,21, and cancer cell invasion24,25. Thus, the dynamic alterations in these exosomal subpopulations are not only associated with clinical outcomes but also mirror critical underlying tumor histopathology and TME states.

Construction of risk models in different clinical groups

Based on the analysis of EV subtypes across these three distinct clinical groupings, we did not identify any sufficiently suitable single biomarker for pre-therapeutic assessment. We therefore aimed to integrate multiple machine learning approaches to develop a panel of biomarkers and establish a risk model for clinical evaluation.

We divided the 85 OC patients into a platinum-sensitive group (n = 44) and a platinum-resistant group (n = 41). Then we analyzed the protein expressed on EVs at single-EV resolution from the two groups and found 35 differential proteins. The data and statistical analysis of protein expression are shown in S10. Three machine learning algorithms, including SVM-RFE, Random Forest, and LASSO, were further applied to identify the most valuable nine genes, ILK, CDCP1, CD86, CLDN4, CLEC1B, CDHR5, CLDN11, JAM2 and FOLH1 (Figs. 4a and S7a–e). As shown in the Supplementary Figure, CD86 was significantly overexpressed in PROC samples (p < 0.05). CLEC1B, CLDN11, and CLDN4 were significantly expressed in a lower level (Fig. S7f–n). Moreover, we conducted the ROC curve analysis to assess the diagnostic value of the combination of these nine genes (Fig. 4b). Here is the formula: Risk Score = 0.0002745*ILK + 0.0013467*CDCP1 + 0.0013467*CD86-0.0015311*CLDN4-0.0002348*CLEC1B + 0.0034037*CDHR5-0.0001591*CLDN11-0.0029356*JAM2-0.0001554*FOLH1-1.3947509. The results showed that the risk score of PROC was significantly higher than PSOC (Fig. 4c).

Fig. 4. Risk models for clinical outcomes in advanced ovarian cancer.

Fig. 4

Open in a new tab

a, d, g Venn diagrams of machine-learning results for three clinical features. b, e, h ROC curve of the three models. c, f, i Correlation of risk scores with the three clinical features.

To facilitate the construction of the risk model, the 85 cases were re-stratified into two groups. Based on whether they had lymph node metastasis or not, we divided all patients into two groups, 56 YES and 29 NO. Then we analyzed the protein expressed on EVs at single-EV resolution of the two groups and found 10 differential proteins. The data and statistical analysis of protein expression are shown in S10. Three machine learning algorithms, including SVM-RFE, Random Forest, and LASSO, were further applied to identify the most valuable seven genes, APOE, CD28, CLDN4, FOLH1, ITGAL, JAML and ULBP3 (Figs. 4d and S8a–e). We considered them as characteristic proteins of lymph node metastasis in ovarian cancer. Among them, ULBP, JAML, FOLH1 and CLDN4 were significantly expressed in a lower level in the samples with lymph node metastasis (p < 0.05) (Fig. S8f–l). Moreover, we performed ROC curve analysis to assess the diagnostic value of the combination of these seven genes (Fig. 4e). Here is the formula: Risk Score = 0.0001239*CD28-0.004443*CLDN4-0.0003997*FOLH1-0.001704*JAML + 0.000048*APOE-0.0002279*ULBP3 + 0.0004279*ITGAL + 5.176. The results showed that the risk score of YES (the group with lymph node metastasis) was significantly higher than NO (the group without lymph node metastasis) (Fig. 4f).

In addition, the 85 patients were divided into two groups: 44 patients with tumor residual disease after operation and 41 without residual disease. Then we analyzed the proteins expressed on EVs at single-EV resolution from the two groups and found 4 differential proteins. The data and statistical analysis of protein expression are shown in S10. The same three machine learning algorithms were further applied to identify the most valuable four genes, CD44, CLMP, ITGA4 and AMIGO1 (Figs. 4g and S9a–e). These were considered characteristic proteins of postoperative residual disease in ovarian cancer. Among them, CD44 and ITGA4 were significantly expressed at higher levels in the samples with residual disease (p < 0.05) and CLMP was significantly expressed at lower levels (p < 0.05) in the Supplementary Figure (S9f–i). Moreover, we conducted the ROC curve analysis to assess the diagnostic value of the combination of these four genes (Fig. 4h). Here is the formula: Risk Score = 0.0041617*CD44-0.0041617*CLMP + 0.0005423*ITGA4-0.0022376*AMIGO1 + 0.6566713. The results showed that the risk score of YES (the group with tumor residual disease after operation) was significantly higher than NO (the group without tumor residual disease after operation) (Fig. 4i).

Discussion

In this study, we employed the Proximity Barcoding Assay (PBA), a novel method for single-EV analysis, to profile surface proteins of individual plasma extracellular vesicles. This approach successfully identified numerous EVs directly from human plasma. Furthermore, using machine learning, we characterized EV-associated proteins in ovarian cancer patients and constructed diagnostic and prognostic models. These results offer important insights into the clinical potential of EV-based biomarkers for ovarian cancer diagnosis and outcome prediction.

Comparison of EV proteins between OC patients and HC revealed a significant increase in the proportion of Cluster 7 through single-EV analysis. This subpopulation was characterized by high expression of ITGB3, ITGB1, and ITGA6. Leveraging these observed differences in EV subpopulations, we developed a multiprotein diagnostic panel to enhance diagnostic performance. Using three distinct machine learning approaches, we identified ITGA6, ITGB2, and ILK as optimal biomarkers for the diagnostic model. These proteins belong to the integrin family, which plays a pivotal role in the adhesion between cells and the extracellular matrix23,26,27. Disruption of integrin function has been shown to affect tumor cell growth, invasion, and metastasis24,25,28. Previous studies have established ITGB2 as a prognostic marker in glioma and ovarian cancer29,30, while ITGA6 maintains metastatic potential in epithelial ovarian cancer (EOC) and contributes to platinum resistance21. Additionally, ITGB3 + EV subpopulations have been proposed as early diagnostic biomarkers for colorectal cancer19, and the ALKBH5/m6A-ITGB1/FAK signaling axis promotes lymphangiogenesis and lymph node metastasis in ovarian cancer31. Thus, while integrin family proteins have been extensively implicated in ovarian cancer pathogenesis and metastasis across multiple studies, they had not previously been incorporated into diagnostic or prognostic models for this disease.

Standard management of advanced ovarian cancer comprises tumor cytoreductive surgery and platinum-based chemotherapy32. Selecting the optimal therapeutic strategy beforehand is critical, particularly in determining platinum resistance status, the need for lymphadenectomy, and indications for neoadjuvant chemotherapy. However, reliable biomarkers for guiding these decisions are currently lacking. This unmet clinical need motivated our subsequent investigation.

Over 70% of patients initially respond to first-line platinum-based chemotherapy, but some exhibit de novo resistance or develop resistance upon disease recurrence. Accurate assessment of initial platinum sensitivity and selection of appropriate chemotherapy regimens are critical determinants of treatment success. Despite extensive investigation, the molecular mechanisms driving ovarian cancer pathogenesis and drug resistance remain incompletely understood. Currently, no reliable biomarkers exist for predicting platinum resistance in clinical practice. Previous studies have associated immunohistochemical levels of biomarkers, including VEGF, p53, ALDH1A1, and FOXP3 with platinum resistance33–36. In our study, we stratified OC patients into platinum-sensitive (PT-SEN) and platinum-resistant (PT-RES) groups for comparative analysis of their exosomal proteomes. While we observed no significant differences in EV subpopulation distributions between these groups, application of three machine learning algorithms enabled identification of differentially expressed proteins—ILK, CDCP1, CD86, CLDN4, CLEC1B, CDHR5, CLDN11, JAM2, and FOLH1—from which we constructed a risk prediction model. This liquid biopsy approach offers distinct advantages over immunohistochemistry, as EV analysis requires only blood samples, is minimally invasive, and enables faster turnaround times. Integrin-linked kinase (ILK) is a highly conserved serine/threonine kinase and linker protein that connects extracellular matrix proteins, integrins, and intracellular signaling pathways37. ILK has been proposed as a therapeutic target in ovarian cancer38,39. CDCP1 is overexpressed in diverse malignancies—including respiratory, gastrointestinal, hematologic, and urogenital cancers—and its elevated expression and activation correlate with poor prognosis and drug resistance40–43. Our identification of CDCP1 is consistent with a recent study by Kong et al., which also employed PBA to profile EVs from ovarian cancer ascites, supporting its potential role in ovarian cancer pathogenesis across different biological fluids44. However, key distinctions highlight the unique contribution of our work. While Kong et al. focused on ascites—a sample type typically available in advanced disease—we established our diagnostic and prognostic models using plasma-derived EVs, which are obtained less invasively and are more suitable for repeated sampling and potential early detection. Furthermore, the specific multi-protein signatures we identified (e.g., ITGA6, ITGB2, ILK for diagnosis) differ from those reported in ascites-based studies. This divergence likely reflects the distinct molecular profiles of EVs originating from different bodily compartments and underscores the value of developing blood-based biomarkers for broader clinical utility. The roles of several other identified proteins in ovarian cancer or drug resistance remain uncharacterized.

Most ovarian cancer patients present with metastatic disease, and the role of lymph node dissection during surgery remains debated. The Lymphadenectomy in Ovarian Cancer (LION) trial demonstrated that elective lymphadenectomy does not improve survival in advanced ovarian cancer patients with clinically negative lymph nodes and is associated with increased surgical morbidity45. Therefore, Preoperative determination of lymph node metastasis status is therefore essential for surgical planning. Currently, methods for assessing lymph node metastasis primarily rely on imaging results and intraoperative exploration, but there is a lack of studies on biomarkers for lymph node metastasis. In this study, we compared exosomal proteins from two groups with and without lymph node metastasis. Using three machine learning methods, we identified differential proteins, including CLDN4, FOLH1, JAML, CD28, APOE, ULBP3, and ITGAL, and developed a risk model for predicting lymph node metastasis. This model allows for non-invasive detection of EVs in blood prior to surgery, providing more accurate predictions of lymph node metastasis when combined with imaging results. This approach helps in deciding whether to perform a lymph node dissection, thus avoiding missed or unnecessary procedures. ITGAL is a leukocyte surface receptor implicated in the tumor microenvironment of various cancers. While studies have linked ITGAL to immune infiltration in non-small cell lung cancer and gastric cancer46,47, its role in ovarian cancer remains uninvestigated. Some studies have explored the potential of CD28 as a diagnostic and prognostic biomarker for ovarian cancer48. Additionally, it has been found that downregulation of CLDN4 can make ovarian cancer cells more sensitive to platinum-based chemotherapy49, although no studies have linked these biomarkers to lymph node metastasis. The remaining four proteins have not been linked to cancer research.

Surgical options for advanced ovarian cancer include primary tumor cytoreductive surgery and neoadjuvant chemotherapy combined with intermittent cytoreductive surgery. For patients deemed unlikely to achieve R0 resection during preoperative assessment, neoadjuvant chemotherapy is typically considered to reduce tumor burden. Surgical feasibility is subsequently re-evaluated following treatment response assessment. Current clinical decision-making relies on patient-specific factors, tumor markers, imaging, and diagnostic laparoscopy, lacking molecular models to guide management. We compared plasma exosomal profiles between patients with and without postoperative residual disease (all samples collected pretreatment). Through single-EV analysis, we found Cluster 13, characterized by ITGB1 expression, was significantly enriched in patients with residual disease. Previous studies have confirmed ITGB1 significant impact on relapse-free survival (RFS) in ovarian cancer patients22. Additionally, using three machine-learning approaches, we identified a panel of differentially expressed proteins—CD44, CLMP, ITGA4, and AMIGO1—and constructed a corresponding risk model. CD44 is a non-kinase cell surface receptor that is associated with cancer metastasis, cancer stem cell maintenance, and chemotherapy resistance through various mechanisms in many cancers, including ovarian cancer50. It represents a promising therapeutic target for ovarian cancer treatment. CLMP has inhibitory effects in colorectal cancer50. ITGA4, a member of the integrin family, is associated with prognosis in many cancers, but research on its role as a biomarker in ovarian cancer remains limited51. Preoperative detection of plasma exosomal biomarkers using this risk model could help identify patients likely to benefit from neoadjuvant chemotherapy, thereby potentially improving R0 resection rates.

Compared to traditional diagnostic methods based on imaging, CA125, and immunohistochemistry, we employed a novel EV detection technology, PBA, to perform high-throughput analysis of EVs in patient plasma. By combining this with machine learning methods, we developed a diagnostic model. This approach not only introduces innovation in methodology but also presents a new concept for providing molecular models for preoperative diagnosis and treatment selection through minimally invasive techniques. With larger sample sizes and further in vivo and in vitro experimental validation, our research will be further refined.

Despite the promising results, our study has several limitations that should be acknowledged. First, the patient cohort was recruited from a single institution and consisted exclusively of advanced-stage (III–IV) HGSOC patients of East Asian descent. While this reflects the clinical reality that most ovarian cancer patients are diagnosed at an advanced stage, it limits the generalizability of our findings to early-stage disease, other ethnic populations, and non-HGSOC histological subtypes, which exhibit distinct molecular profiles6. Furthermore, the potential confounding effects of variables such as age and tumor burden were not adjusted for in the machine learning models, a decision driven by our focus on establishing a pure exosomal biomarker signature and constrained by the current sample size. Future validation in multi-center cohorts encompassing the full histological and pathological spectrum of ovarian cancer is essential. We are actively addressing this by expanding our sample size, with a specific focus on enrolling early-stage ovarian cancer cases, and planning multi-center collaborative studies to enhance the diversity and representativeness of our cohort.

Second, although we have standardized and reported key pre-analytical procedures, undetected variations in sample handling or the influence of co-morbidities could affect the EV proteome. The single-timepoint nature of our sampling also cannot capture dynamic changes during disease progression or therapy.

Third, the prognostic models were internally validated but lack an independent external validation cohort. The current absence of large, public plasma-EV proteomic datasets precluded this important step. Future multi-center studies are essential to confirm the generalizability and robustness of our models.

Finally, the translation of our multi-protein panels into clinical practice faces challenges, including the standardization of the PBA assay across laboratories, cost-effectiveness, the need for extensive regulatory validation, and demonstrating technical reproducibility. Furthermore, the biological functions and mechanistic roles of some identified proteins (e.g., CLMP, AMIGO1, JAML) in ovarian cancer progression remain largely unexplored. Future in vitro and in vivo functional studies are crucial to validate these protein candidates and elucidate their roles in pathogenesis.

Methods

Human blood samples

The study was approved by the Medical Ethics Committee of Zhejiang Cancer Hospital (IRB-2021-350). The study was conducted according to the principles expressed in the Declaration of Helsinki. From 2010 to 2019, a total of 85 ovarian cancer patients were enrolled in our study in Zhejiang Cancer Hospital. 95 healthy volunteers were enrolled as the HC group. All sample donors signed informed consent. All patients were histologically confirmed as HGSOC.

Peripheral blood samples were collected from all participants using EDTA-coated vacuum tubes. All samples were obtained from fasted individuals. To standardize pre-analytical variables, blood samples were processed within 2 h of collection. Plasma was separated by centrifugation at 1800 × g for 10 min at 4 °C, aliquoted, and immediately stored at −80 °C until further analysis.

PBA

Two microliters of plasma was mixed with 1 μL of PBA buffer and then incubated with 1 μl of antibody mix (2 μg/ml for each antibody) for 2 h at room temperature. The antibody mix was derived from a commercial multiplex antibody panel (Secretech PBA Human Oncology Panel v1.0, Shanghai Secretech Co., Ltd.) targeting 243 human proteins. The specific antibody clones within this proprietary panel are confidential. The final concentration for each antibody in the mix was 2 μg/mL, and the incubation was carried out according to the manufacturer’s instructions. Add biotinylated cholera toxin subunit B (biotin-CTB, 2.5 μg/ml in PBS, C34779, Thermo Fisher Scientific, USA) to each well of a 96-well plate with streptavidin coating (PCR0STF-SA5/100, Biomat, Italy) to prepare EV capture plate and incubated for 20 min at room temperature. After washing 3 times with PBS with Tween 20 (PBST), Plasma/antibody mix were diluted to a volume of 20 μl with PBA buffer and transferred into each well of CTB-coated plate for affinity capture of EVs through the interaction between CTB and GM1 enriched in lipid membrane of EVs. Captured EVs were fixed with paraformaldehyde solution (4%) and imaged with SEM (SU8010, Hitachi, Japan) utilizing 3-kV beam energy and a secondary electron detector. With DNA polymerase, the oligonucleotides on the antibody undergo extension reaction to obtain EV tag, which is the reverse complementary sequence of a designed part in barcoding template. DNA fragment library is constructed for high-throughput sequencing. The library was sequenced with Illumina NextSeq (CN500, Illumina, USA) utilizing single-end 75-bp sequencing reagent kit (20024906, Illumina, USA). After DNA sequencing, data of each sample was obtained as bcl file.

Machine learning to screen diagnostic biomarkers

Support vector machine-recursive feature elimination (SVM-RFE), the least absolute shrinkage and selection operator (LASSO), and Random forest (RF) were used to select figure DEPs. Finally, the feature proteins obtained from the above machine learning methods were intersected to determine the optimal number of features.

SVM-RFE

We employed the caret package (v6.0-94) with a radial basis kernel and 10-fold cross-validation. The feature set corresponding to the minimum cross-validation error was identified as optimal.

LASSO

We implemented 10-fold cross-validation using the glmnet package (v4.1-8) to determine the optimal regularization parameter (λ.min), which corresponds to the minimum cross-validation error. Proteins with non-zero coefficients at λ.min were selected as feature proteins.

Random Forest

We constructed an initial model with 500 decision trees using the randomForest package (v4.7-1.2). Feature importance was quantified by the MeanDecreaseGini metric. We selected the top 20 most important proteins based on their MeanDecreaseGini values for subsequent analysis.

Functional enrichment analysis

GO and KEGG analyses and visualization of the DEPs were performed using the R package “ClusterProfiler”.

PPI analysis

The online database named Search Tool for the Retrieval of Interacting Genes (STRING) was used to analyze interactions between DEPs. The R package “Cytoscape” was used for visualization.

Data processing and statistical analysis

The expression data were normalized using the trimmed mean of M values (TMM) algorithm. The total samples (tumor and normal) were divided into training and validation sets in a 7:3 ratio.

All statistical analyses were performed using R software (version 4.1.0). For high-dimensional proteomic data, the following statistical approaches were applied at different stages:

Differential expression analysis

To identify differentially expressed proteins (DEPs) between two groups (e.g., OC vs. HC), we used the Student’s t test for data with normal distribution and homogeneous variance. If the data did not meet the assumptions of normality or homogeneity of variance, the non-parametric Mann–Whitney U test was used instead. To account for multiple testing, the p values from all protein comparisons were adjusted using the Benjamini–Hochberg method to control the False Discovery Rate (FDR). Proteins with an FDR-adjusted p value < 0.05 were considered statistically significant DEPs.

Machine learning feature selection

The feature selection algorithms themselves handle the issue of overfitting through internal mechanisms. For LASSO, the optimal penalty parameter (lambda) was determined via 10-fold cross-validation, and the features corresponding to lambda.min were selected. In Random Forest, features with a MeanDecreaseGini value greater than 1 were considered important. For SVM-RFE, the feature set corresponding to the minimum point of cross-validation error was selected. The intersection of features identified by these three methods was used to determine the final optimal biomarker panels.

Comparison of risk scores

The differences in continuous Risk Scores between two clinical groups (e.g., high-risk vs. low-risk) were assessed using the non-parametric Wilcoxon rank-sum test (Mann–Whitney U test), as these scores were not assumed to be normally distributed.

When comparing more than two groups, one-way ANOVA was used for data meeting parametric assumptions, followed by Duncan’s test for post hoc analysis. The non-parametric Kruskal–Wallis test was used otherwise, followed by the Pairwise Wilcoxon Rank Sum Test for post-hoc analysis. A p value < 0.05 was considered significant for these specific post hoc comparisons.

R package ‘factoextra’ and ‘FactoMineR’ were used for principal component analysis. The ‘rms’ and ‘pROC’ packages were used to analyze the logistic regression of candidate biomarkers and plot ROC curves to determine the AUC.

Supplementary information

Supplementary Information (5.7MB, pdf)

Acknowledgements

The authors thank the patients who contributed to the study and the support of our colleagues. This work was supported by the National Natural Science Foundation of China (82274266), National Natural Science Foundation of China (82505319) and the Traditional Chinese Medicine Science and Technology Project of Zhejiang Province (2026ZL0202).

Author contributions

B.W. was responsible for methodology, validation, data curation, writing and visualization; X.Y. was responsible for model establishing, validation and writing; Y.C. was responsible for EV assay, validation, review and editing; S.W., J.L., and Y.J. were responsible for validation and data curation; J.X., X.C., and J.Z. were responsible for data curation and editing; A.Y. was responsible for conceptualization, methodology, supervision project administration, and funding acquisition; L.Y. was responsible for conceptualization, methodology, resources, writing, review and editing, visualization and supervision project administration. B.W., X.Y., Y.C. contributed equally.

Data availability

The raw sequencing data generated in this study are publicly available in the OMIX database at: https://ngdc.cncb.ac.cn/omix/select-edit/OMIX010139.

Code availability

All proteomic analyses were performed with publicly available R packages and tools following standard pipelines: differential expression (two-group t-test with FDR correction), ROC analysis (pROC, https://cran.r-project.org/package=pROC), LASSO selection (glmnet, https://cran.r-project.org/package=glmne), SVM-RFE (caret, https://cran.r-project.org/package=caret), Random Forest (randomForest, https://cran.r-project.org/package=randomForest), PCA (FactoMineR & factoextra, https://cran.r-project.org/package=FactoMineR and https://cran.r-project.org/package=factoextra), PPI network construction (STRING database, https://string-db.org), functional enrichment (clusterProfiler, https://bioconductor.org/packages/clusterProfile). Any additional information required for the analysis of data in this manuscript is available from the authors upon reasonable request.

Competing interests

The authors declare no competing interests.

Footnotes

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

These authors contributed equally: Beier Wu, Xuping Yang, Yanling Cai.

Contributor Information

Aijun Yu, Email: yaj1993@126.com.

Li Yang, Email: yangli@zjcc.org.cn.

Supplementary information

The online version contains supplementary material available at 10.1038/s41698-026-01271-x.

References

  • 1.Coburn, S. B., Bray, F., Sherman, M. E. & Trabert, B. International patterns and trends in ovarian cancer incidence, overall and by histologic subtype. Int. J. Cancer140, 2451–2460 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 2.Hennessy, B. T., Coleman, R. L. & Markman, M. Ovarian cancer. Lancet374, 1371–1382 (2009). [DOI] [PubMed]
  • 3.Schiavone, M. B. et al. Natural history and outcome of mucinous carcinoma of the ovary. Am. J. Obstet. Gynecol.205, 480.e481–488 (2011). [DOI] [PubMed] [Google Scholar]
  • 4.Berek, J. S., Kehoe, S. T., Kumar, L. & Friedlander, M. Cancer of the ovary, fallopian tube, and peritoneum. Int. J. Gynaecol. Obstet.143, 59–78 (2018). [DOI] [PubMed] [Google Scholar]
  • 5.Morgan, R. J. Jr. et al. Ovarian Cancer, Version 1.2016, NCCN clinical practice guidelines in oncology. J. Natl. Compr. Cancer Netw.14, 1134–1163 (2016). [DOI] [PubMed] [Google Scholar]
  • 6.Cancer Genome Atlas Research Network. Integrated genomic analyses of ovarian carcinoma. Nature474, 609–615 (2011). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 7.Kok, V. C. & Yu, C. C. Cancer-derived exosomes: their role in cancer biology and biomarker development. Int. J. Nanomed.15, 8019–8036 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 8.Tian, Y. et al. Protein profiling and sizing of extracellular vesicles from colorectal cancer patients via flow cytometry. ACS Nano12, 671–680 (2018). [DOI] [PubMed] [Google Scholar]
  • 9.Yoshioka, Y. et al. Ultra-sensitive liquid biopsy of circulating extracellular vesicles using ExoScreen. Nat. Commun.5, 3591 (2014). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 10.Zhang, Y. et al. The miR-19b-3p-MAP2K3-STAT3 feedback loop regulates cell proliferation and invasion in esophageal squamous cell carcinoma. Mol. Oncol.15, 1566–1583 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 11.He, L. et al. Ovarian cancer cell-secreted exosomal miR-205 promotes metastasis by inducing angiogenesis. Theranostics9, 8206–8220 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 12.Namee, N. M. & O’Driscoll, L. Extracellular vesicles and anti-cancer drug resistance. Biochim. Biophys. Acta Rev. Cancer1870, 123–136 (2018). [DOI] [PubMed] [Google Scholar]
  • 13.Mo, Y. et al. Tumor-secreted exosomal miR-141 activates tumor-stroma interactions and controls premetastatic niche formation in ovarian cancer metastasis. Mol. Cancer22, 4 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 14.Nakamura, K. et al. Role of the exosome in ovarian cancer progression and its potential as a therapeutic target. Cancers11, 1147 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 15.Castillo, J. et al. Surfaceome profiling enables isolation of cancer-specific exosomal cargo in liquid biopsies from pancreatic cancer patients. Ann. Oncol.29, 223–229 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 16.Larssen, P. et al. Tracing cellular origin of human exosomes using multiplex proximity extension assays. Mol. Cell. Proteom.16, 1547 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 17.Tauro, B. J. et al. Two distinct populations of exosomes are released from LIM1863 colon carcinoma cell-derived organoids. Mol. Cell. Proteom.12, 587–598 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 18.Wu, D. et al. Profiling surface proteins on individual exosomes using a proximity barcoding assay. Nat. Commun.10, 3854 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 19.Guo, W. et al. Single-exosome profiling identifies ITGB3+ and ITGAM+ exosome subpopulations as promising early diagnostic biomarkers and therapeutic targets for colorectal cancer. Research6, 0041 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 20.Van Gassen, S. et al. FlowSOM: using self-organizing maps for visualization and interpretation of cytometry data. Cytom. Part A J. Int. Soc. Anal. Cytol.87, 636–645 (2015). [DOI] [PubMed] [Google Scholar]
  • 21.Gambelli, A. et al. Platinum-induced upregulation of ITGA6 promotes chemoresistance and spreading in ovarian cancer. EMBO Mol. Med.16, 1162–1192 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 22.Zhu, T. et al. The prognostic value of ITGA and ITGB superfamily members in patients with high grade serous ovarian cancer. Cancer Cell Int.20, 257 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 23.Chastney, M. R., Kaivola, J., Leppänen, V. M. & Ivaska, J. The role and regulation of integrins in cell migration and invasion. Nat. Rev. Mol. Cell Biol.26, 147–167 (2025). [DOI] [PubMed] [Google Scholar]
  • 24.Hanker, A. B. et al. Extracellular matrix/integrin signaling promotes resistance to combined inhibition of HER2 and PI3K in HER2(+) breast cancer. Cancer Res.77, 3280–3292 (2017). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 25.Morello, V. et al. β1 integrin controls EGFR signaling and tumorigenic properties of lung cancer cells. Oncogene30, 4087–4096 (2011). [DOI] [PubMed] [Google Scholar]
  • 26.Margadant, C. & Sonnenberg, A. Integrin-TGF-beta crosstalk in fibrosis, cancer and wound healing. EMBO Rep.11, 97–105 (2010). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 27.Dongre, A. & Weinberg, R. A. New insights into the mechanisms of epithelial-mesenchymal transition and implications for cancer. Nat. Rev. Mol. Cell Biol.20, 69–84 (2019). [DOI] [PubMed] [Google Scholar]
  • 28.Guo, W. et al. Beta 4 integrin amplifies ErbB2 signaling to promote mammary tumorigenesis. Cell126, 489–502 (2006). [DOI] [PubMed] [Google Scholar]
  • 29.Li, C. et al. Identifying ITGB2 as a potential prognostic biomarker in ovarian cancer. Diagnostics13, 1169 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 30.Xu, H. et al. ITGB2 as a prognostic indicator and a predictive marker for immunotherapy in gliomas. Cancer Immunol. Immunother.71, 645–660 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 31.Sun, R. et al. ALKBH5 activates FAK signaling through m6A demethylation in ITGB1 mRNA and enhances tumor-associated lymphangiogenesis and lymph node metastasis in ovarian cancer. Theranostics13, 833–848 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 32.Armstrong, D. K. et al. NCCN Guidelines® Insights: Ovarian Cancer, Version 3.2022. J. Natl. Compr. Cancer Netw.20, 972–980 (2022). [DOI] [PubMed] [Google Scholar]
  • 33.Guo, B. Q. & Lu, W. Q. The prognostic significance of high/positive expression of tissue VEGF in ovarian cancer. Oncotarget9, 30552–30560 (2018). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 34.Wong, K. K. et al. Poor survival with wild-type TP53 ovarian cancer? Gynecol. Oncol.130, 565–569 (2013). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 35.Vassalli, G. Aldehyde dehydrogenases: not just markers, but functional regulators of stem cells. Stem Cells Int.2019, 3904645 (2019). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 36.Wolf, D. et al. The expression of the regulatory T cell-specific forkhead box transcription factor FoxP3 is associated with poor prognosis in ovarian cancer. Clin. Cancer Res.11, 8326–8331 (2005). [DOI] [PubMed] [Google Scholar]
  • 37.Zheng, C. C. et al. Significance of integrin-linked kinase (ILK) in tumorigenesis and its potential implication as a biomarker and therapeutic target for human cancer. Am. J. Cancer Res.9, 186–197 (2019). [PMC free article] [PubMed] [Google Scholar]
  • 38.Reyes-González, J. M. et al. Downstream effectors of ILK in cisplatin-resistant ovarian cancer. Cancers12, 880 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 39.Ulm, M. A. et al. Integrin-linked kinase is a novel therapeutic target in ovarian cancer. J. Pers. Med.10, 246 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 40.Dagnino, S. et al. Prospective identification of elevated circulating CDCP1 in patients years before onset of lung cancer. Cancer Res.81, 3738–3748 (2021). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 41.Chou, C. T. et al. Prognostic significance of CDCP1 expression in colorectal cancer and effect of its inhibition on invasion and migration. Ann. Surg. Oncol.22, 4335–4343 (2015). [DOI] [PubMed] [Google Scholar]
  • 42.Bühring, H. J. et al. CDCP1 identifies a broad spectrum of normal and malignant stem/progenitor cell subsets of hematopoietic and nonhematopoietic origin. Stem Cells22, 334–343 (2004). [DOI] [PubMed] [Google Scholar]
  • 43.Alajati, A. et al. CDCP1 overexpression drives prostate cancer progression and can be targeted in vivo. J. Clin. Investig.130, 2435–2450 (2020). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 44.Kong, L. et al. Ascites-derived CDCP1+ extracellular vesicles subcluster as a novel biomarker and therapeutic target for ovarian cancer. Front. Oncol.13, 1142755 (2023). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 45.Harter, P. et al. A randomized trial of lymphadenectomy in patients with advanced ovarian neoplasms. N. Engl. J. Med.380, 822–832 (2019). [DOI] [PubMed] [Google Scholar]
  • 46.Zhang, R. et al. ITGAL expression in non-small-cell lung cancer tissue and its association with immune infiltrates. Front. Immunol.15, 1382231 (2024). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 47.Zhang, J. et al. ITGAL as a prognostic biomarker correlated with immune infiltrates in gastric cancer. Front. Cell Dev. Biol.10, 808212 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 48.Świderska, J. et al. Clinical significance of BTLA, CD27, CD70, CD28 and CD80 as diagnostic and prognostic markers in ovarian cancer. Diagnostics12, 251 (2022). [DOI] [PMC free article] [PubMed]
  • 49.Gao, Z. et al. C terminus of Clostridium perfringens enterotoxin downregulates CLDN4 and sensitizes ovarian cancer cells to Taxol and Carboplatin. Clin. Cancer Res.17, 1065–1074 (2011). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 50.Ween, M. P., Oehler, M. K. & Ricciardelli, C. Role of versican, hyaluronan and CD44 in ovarian cancer metastasis. Int. J. Mol. Sci.12, 1009–1029 (2011). [DOI] [PMC free article] [PubMed] [Google Scholar]
  • 51.Li, M. et al. METTL3 mediates chemoresistance by enhancing AML homing and engraftment via ITGA4. Leukemia36, 2586–2595 (2022). [DOI] [PMC free article] [PubMed] [Google Scholar]

Associated Data

This section collects any data citations, data availability statements, or supplementary materials included in this article.

Supplementary Materials

Supplementary Information (5.7MB, pdf)

Data Availability Statement

The raw sequencing data generated in this study are publicly available in the OMIX database at: https://ngdc.cncb.ac.cn/omix/select-edit/OMIX010139.

All proteomic analyses were performed with publicly available R packages and tools following standard pipelines: differential expression (two-group t-test with FDR correction), ROC analysis (pROC, https://cran.r-project.org/package=pROC), LASSO selection (glmnet, https://cran.r-project.org/package=glmne), SVM-RFE (caret, https://cran.r-project.org/package=caret), Random Forest (randomForest, https://cran.r-project.org/package=randomForest), PCA (FactoMineR & factoextra, https://cran.r-project.org/package=FactoMineR and https://cran.r-project.org/package=factoextra), PPI network construction (STRING database, https://string-db.org), functional enrichment (clusterProfiler, https://bioconductor.org/packages/clusterProfile). Any additional information required for the analysis of data in this manuscript is available from the authors upon reasonable request.

Articles from NPJ Precision Oncology are provided here courtesy of Nature Publishing Group

RESOURCES