Early-stage multi-cancer detection using an extracellular vesicle protein-based blood test.

Background: Detecting cancer at early stages significantly increases patient survival rates. Because lethal solid tumors often produce few symptoms before progressing to advanced, metastatic disease, diagnosis frequently occurs when surgical resection is no longer curative. One promising approach to detect early-stage, curable cancers uses biomarkers present in circulating extracellular vesicles (EVs). To explore the feasibility of this approach, we developed an EV-based blood biomarker classifier from EV protein profiles to detect stages I and II pancreatic, ovarian, and bladder cancer.
Methods: Utilizing an alternating current electrokinetics (ACE) platform to purify EVs from plasma, we use multi-marker EV-protein measurements to develop a machine learning algorithm that can discriminate cancer cases from controls. The ACE isolation method requires small sample volumes, and the streamlined process permits integration into high-throughput workflows.
Results: In this case-control pilot study, comparison of 139 pathologically confirmed stage I and II cancer cases representing pancreatic, ovarian, or bladder patients against 184 control subjects yields an area under the curve (AUC) of 0.95 (95% CI: 0.92 to 0.97), with sensitivity of 71.2% (95% CI: 63.2 to 78.1) at 99.5% (97.0 to 99.9) specificity. Sensitivity is similar at both early stages [stage I: 70.5% (60.2 to 79.0) and stage II: 72.5% (59.1 to 82.9)]. Detection of stage I cancer reaches 95.5% in pancreatic, 74.4% in ovarian (73.1% in Stage IA) and 43.8% in bladder cancer. Conclusions: This work demonstrates that an EV-based, multi-cancer test has potential clinical value for early cancer detection and warrants future expanded studies involving prospective cohorts with multi-year follow-up.

Introduction

Detecting cancer early before symptoms present is key to improving patient survival. While not all emergent tumors will become deadly, for those destined to become so, the ability to treat the disease while it is still localized is a major factor for improving 5-year survival rates[1]. Pancreatic ductal adenocarcinoma (PDAC), one of the deadliest cancers and a leading cause of all cancer-related deaths in the United States, typically goes undetected until it spreads and becomes unresectable and metastatic[2]. In contrast, for the few patients (~15%) diagnosed with localized disease, the 5-year survival rate rises to about 25% and when PDAC is detected at Stage I, survival rates can be as high as 80%[3]. Likewise, ovarian cancer typically has few symptoms and is often undetected until it is advanced, with 5-year survival rates of <31%[4]. When detected early, 5-year survival rates for localized ovarian cancer jump to a remarkable 93%, but currently, only 15% of cases are detected at early stages[5]. For metastatic bladder cancer, 5-year survival rates are only 6%, while detection when the tumor is still localized to the bladder wall inner layer results in an improved 5-year survival rate of 96%[6]. Importantly, treatment of localized bladder cancer has less morbidity and better quality of life compared to treatments required for metastatic cancer[7].

Currently, there are few general screening strategies to detect asymptomatic, early-stage PDAC, ovarian, or bladder cancer[8]. Given recent advances in targeted treatments for cancer, which are based on functional changes in the genome and proteome of individual tumors and their milieu, attention has turned to the possibility of detecting these changes directly from blood, i.e., by liquid biopsy[1,9]. This strategy has been utilized to design multi-cancer early detection (MCED) tests that involve blood-based circulating proteins and/or DNA mutations and methylations followed by machine learning approaches to discern between cancer and non-cancer cases[10-15]. Several MCED tests based on these approaches are being developed and have shown promise for detecting clinically significant, late-stage (III and IV) cancers, and that detection was prognostic beyond tumor stage[16]. Detecting early, stage I, and II cancers with high enough sensitivity for population-level screening, however, has proven much more challenging[12-14].

One potential approach for more sensitive detection of cancer-related biomarkers from blood involves the use of extracellular vesicles (EVs) such as exosomes, 30 to 150 nm vesicles that mediate cell-to-cell communication[17,18]. It has been shown that some exosomes are ejected by tumors into the bloodstream and they carry functional protein biomarkers representing the tumor proteome[18,19]. The potential to better detect cancer using EV-bound protein biomarkers has recently been suggested for multiple cancer types[20] using various methodologies, such as mass spectrometry for lung and pancreatic cancers[21]. Currently, the gold standard method for isolating EVs from soluble contaminants (cells, small proteins, or other vesicles) is ultracentrifugation (UC), which is inefficient and not suitable for point-of-care applications[22]. To address this issue, many groups have explored a variety of methods, based on immunoaffinities and/or diverse membranes, to both isolate and analyze circulating EVs and associated markers[17]. In this study, we used an alternating current electrokinetic (ACE)-based platform (Verita™ System)[23] to efficiently purify exosomes and other EVs from patient samples, then measured the concentrations of associated protein biomarkers (“EV proteins”) present in the purified EV samples from our case-control study subjects. Using the information about the differing EV protein concentrations, we developed a machine-learning algorithm to identify a small set of EV biomarkers which together with age permits detection of early-stage pancreatic, ovarian, and bladder cancers. We find that using ACE purification of EVs, followed by a specialized analysis of the EV protein biomarkers, successfully predicts the presence of early stage I–II pancreatic, ovarian, and bladder cancers with a sensitivity of 71.2% (95% CI: 63.2–78.1) at 99.5% (97.0–99.9) specificity, and an AUC of 0.95 (95% CI: 0.92–0.97). To our knowledge, we are the first to report feasibility for a blood based, MCED test for the detection of stage I and II cancers that employs circulating EV proteins exclusively.

Methods

Sample collection and processing

All specimens for this retrospective study were collected over a period of several years by a commercial biorepository (ProteoGenex, Inglewood, CA, USA). Stage I and II samples were selectively obtained from available inventory. Samples had been collected from patients in hospital settings and following collection were maintained by the commercial biorepository. All relevant ethical regulations were followed, and informed consent was obtained prior to sample collection. The protocol was approved by the ethics committee at the N. N. Blokhin National Medical Research Center of Oncology. In the hospital settings, potential cancer patients were identified by any suspicious findings arising during imaging that was conducted either in response to patient symptoms or as part of routine, annual examinations. We do not have access to information on which patients were symptomatic and which were asymptomatic. Cancers were confirmed via subsequent tissue biopsy and staged by pathologists in the hospital using pathology and surgical reports, according to AJCC (7th edition) guidelines, along with imaging to assess any spread to distant sites. All subjects with confirmed diagnosis of cancer were treatment naive (prior to surgery, local, and/or systemic anti-cancer therapy) at the time of blood collection. The biorepository provided the patient samples along with demographics, surgical, and pathology information. Through the analysis of these data, staging for patients was reviewed a second time for accuracy by the study authors. Our study did not require ethics approval because samples were de-identified after processing by the biorepository. Since ovarian cancer patients did not uniformly undergo comprehensive surgical staging, an occult disease higher than the indicated stage cannot be ruled out. The control group has no known history of cancer, autoimmune diseases, or neurodegenerative disorders, nor any presence of diabetes mellitus (types 1 and 2). A total of 323 subjects were included in the study, including 139 subjects (‘Cancer case patient cohort’) who were diagnosed with one of the three cancers between January 2014 and September 2020. In the cancer case cohort, whole venous blood specimens were collected shortly before biopsy (median −1 day, mean −2.7 days), and prior to surgical intervention, radiation therapy, or cancer-related systemic therapy. The median age was 60 years [Min–Max 21–76] in the cancer case cohort (N = 139, 56 males, 83 females) and 57 years [Min–Max 40–71] in the control cohort (N = 184, 82 males and 82 females). Details on the case-control cohorts can be found in Supplementary Data 1 and Supplementary Data 2. Whole blood samples were collected in K2EDTA plasma vacutainer tubes and processed into plasma within 4 h of collection. The whole blood was first spun at 1500 × g for 10 min at 4 °C with no brake used. After the first spin, plasma was transferred into fresh tubes and subjected to a second spin at 1500 × g for 10 min. After the second spin, plasma was aliquoted into 1 mL tubes and frozen within 1 h at −80 °C. All specimens used in this study were processed under identical conditions.

EV/exosome isolation and particle characterization

Isolation of EVs using the Verita™ Platform

EVs, including exosomes, were extracted from plasma as previously described using an AC Electrokinetic (ACE)-based isolation method (Biological Dynamics, CA, USA) as shown in Fig. 1[23-25]. Briefly, 240 µL of each undiluted plasma was introduced into a Verita™ chip, and an electrical signal of 7 Vpp and 14 KHz was applied while flowing the plasma across the chip at 3 µL/min for 120 min. EVs were captured onto the energized microelectrode array, and unbound materials were washed off the chip with Elution Buffer I (Biological Dynamics) for 30 min at 3 µL/min. The electrical signal was turned off, releasing EVs into the solution remaining on the chip (35 µL), which was then collected, and this solution containing purified, concentrated/eluted EVs was used directly for further analysis. This method has also been used previously for the isolation of cell-free DNA, exosomal RNA, and exosomal protein markers in both solid-tumors and hematological malignancies[24-29]. The Verita-purified EVs were characterized using nanoparticle tracking analysis (NTA) via ZetaView instrument (Particle Metrix, Inning am Ammersee, Germany). Supplementary Data 3 shows the particle size and concentration values for the EVs isolated from each subject in the cohorts while Fig. S1 shows comparisons between the case and control cohorts.

Fig. 1

Schematic showing EV isolation workflows using either Verita™ or ultracentrifugation methods.

a Workflow using the Verita™ Isolation platform. As plasma samples are flowed onto the energized AC Electrokinetics (ACE) microelectrode array, EVs are collected onto the electrodes. Unbound materials are removed with a buffer wash, the electric field turned off, and EVs are eluted into the buffer. b Workflow for differential ultracentrifugation. Plasma samples are diluted, and large debris pelleted by a low-speed centrifugation step. Supernatants are removed and subjected to two additional cycles of low-speed centrifugation. EVs in the cleared supernatants are then ultracentrifuged two times, and lastly the pellet is resuspended in buffer.

Isolation of EVs via differential ultracentrifugation

A subset of case and control samples were subjected to differential ultracentrifugation as a conventional means of EV isolation[21]. In brief, 760 µL of 1× PBS was added to 240 µL of each subject plasma, then spun successively at 500 × g for 10 min, 3000 × g for 20 min, and 12,000 × g for 20 min, collecting the supernatants after each step. Subsequently, the resulting supernatant was subjected to ultracentrifugation at 100,000 × g for 70 min, pellets were washed in 1× PBS and then ultracentrifuged again at 100,000 × g for 70 min. The supernatant was discarded, and the resulting pellet was resuspended in 120 µL of 1× PBS for further analysis (Fig. 1).

Protein contamination analysis

To determine the presence of contaminating total protein in the EV preparations from both the Verita™ platform and the differential ultracentrifugation process, samples were analyzed using the Qubit 4 fluorometer (ThermoFisher Scientific, Waltham, MA) with the Qubit™ Protein quantitation assay (Cat No. Q33212, ThermoFisher Scientific, Waltham, MA), run according to manufacturer specifications. To further understand the composition of the contaminating proteins on the isolation products, we used the 2100 Bioanalyzer (Agilent, Santa Clara, CA) with the Protein 230 kit for protein analysis (Cat No. 5067-1517) following the manufacturer’s directions.

Protein biomarker analysis

Verita-isolated EV samples, as well as original, unpurified plasma samples from the same patients, were used directly in commercial multiplex immunoassays to quantify the presence of marker proteins. In brief, 25 µL of each purified EV sample was used for analysis by each of three different bead-based immunoassay kits, according to the manufacturer’s directions for each kit (Human Circulating Biomarker Magnetic Bead Panel 1 (Cat # HCCBP1MAG-58K), Human Angiogenesis Magnetic Bead Panel 2 (Cat # HANG2MAG-12K), and Human Circulating Cancer Biomarker Panel 3 (Cat # HCCBP3MAG-58K); Millipore Sigma, Burlington, MA). Protein biomarker concentration was assessed using the MAGPIX system (Luminex Corp, Austin, TX) according to the manufacturer’s protocols. Belysa software v. 3.0 (EMD Millipore) was used to determine final protein concentrations from the calibration curves. Limit of detection (LOD) and units of measure for each of the biomarkers are listed in Supplementary Data 4.

Spike EV isolation models for EV biomarker signal

To further understand the presence of relevant protein biomarkers on the EVs, we employed EVs purified from cell culture supernatants representing two different cell lines as positive controls. The cell line H1975 (ATCC CRL-5908™) is known to express the CA19-9 marker (ST6 N-acetylgalactosaminide alpha-2,6-sialyltransferase 1) while the cell line HeLa (ATCC CRM-CCL-2™) is known to express the CA 125 marker (Mucin16). Briefly, the H1975 EVs were spiked at three different concentrations (from 4.60 × 109 to 1.15 × 109 particles/mL) into K2EDTA plasma, the EVs were isolated using the Verita™ platform, and subsequently analyzed via a bead-based immunoassay for the presence of the CA 19-9 biomarker. The linear response to EV input (marker values ± standard deviation) is shown in Fig. S2. In another experiment, the H1975 EVs and the HeLa EVs were spiked into K2EDTA plasma and isolated using the VeritaTM platform. The biomarker reading results confirm the positive detection of the respective expected signals with CA19-9 being elevated for the H1975 EVs and CA 125 being elevated for the HeLa EVs (Fig. S2).

Statistics and reproducibility

Each case or control sample was measured in duplicate or triplicate (depending on volume availability) with one chip eluate going to one reaction well in the multiplex immunoassay plate. No pooling or dilution of the eluates was performed. The same approach was followed for the differential ultracentrifugation experiments where each immunoassay well contained a single resuspended pellet.

Biomarker selection

From an initial evaluation of 42 EV proteins, 34 different biomarkers with <50% of values missing or below the LOD were considered (Supplementary Data 2 and Supplementary Data 4). In cases with missing values or results below the LOD, values were set (imputed) to the LOD. Distributions for all biomarkers were evaluated and distributions were found to be wide; thus, we used a Log2 transformation on all EV protein biomarker values in subsequent analyses. Subsequently, we explored the correlations among the biomarkers using the R module ‘Corrplot’ to determine the potential for multicollinearity in building classification models (correlation heatmap from all the biomarkers measured are shown in Fig. S3). To determine the most informative biomarkers, recursive feature elimination with fivefold cross-validation[30] was employed using the R module “caret”. In this methodology, four of the five folds are used for selecting a subset of biomarkers using stepwise backwards selection. This process is repeated five times, using each fold once as a held-out test set. As the folds of cross-validation are chosen at random, this was repeated 100 times and the subset of biomarkers that maximized the partial AUC (pAUC)[31] over the range of specificities from 0.75 to 1.00 across all test sets was selected (Fig. 2).

Fig. 2

Development of classification algorithm for multi-cancer early detection.

Biomarker selection is performed via recursive feature elimination (RFE) with cross validation. After the biomarkers are selected, the dataset is split into training and test sets. The training set is used for the determination of the coefficients in the logistic regression for each biomarker and the test set is used to evaluate the performance of the logistic regression fit from the training set in a held-out test set. Finally, the process of splitting the dataset into training and test sets is randomly repeated 100 times for performance confirmation.

Coefficient determination and performance evaluation

Once the biomarkers were selected, an initial partition of the data into training (67%) and test (33%) sets, stratified by cancer types, allowed us to determine the performance of the biomarkers selected by estimating the regression coefficients for the model using the training set and evaluating the classification performance in the held-out test set (Fig. 2). To pursue a fair assessment of our model, given our relatively small sample size and to avoid overfitting[30,32-34], we resampled 100 independent training and test sets (made up of 2/3 and 1/3 of the 323 individuals stratified by cancer type) from the overall data set. The subjects in the training set, for each resample, were used to estimate biomarker regression coefficients in the model whereas the diagnostic performance was assessed independently in subjects in the held-out test set. Receiver-operator characteristic (ROC) curves, area under the curve (AUC), sensitivity, specificity, and related metrics were computed for the test sets based on the individual fits for each of the subjects in each respective partition (Supplementary Data 5). For each of the held-out test sets, a threshold determination of >99% specificity was computed (because there were 61 control subjects in each held-out test set, this effectively means calling 61 out of 61 correctly) and subsequently, the average threshold was determined (Supplementary Data 6). Using the average threshold and the average fit in the test set for each subject, we then evaluated the performance for the overall cohort as well as for subcohorts (e.g., pancreatic cancer). The 95% confidence intervals for AUC were calculated using a bias-corrected bootstrapping method (N = 2000)[35] while the confidence intervals for performance metrics, i.e. sensitivity and specificity, were calculated based on the Wilson two-sided method[36]. During the evaluation of the logistic regression model, we assessed the importance of each biomarker selected using the average standardized coefficients (Supplementary Data 7). Here “importance” can be understood as a quantitative comparison between predictors. One predictor is more important than another if it contributes more to the prediction of the response variable across all the models considered in the regression.

Additional analysis and plotting

Additional analysis and plotting in both the main text and the supplementary information were performed in GraphPad Prism (Version 9.0.2) and JMP Pro (Version 16.1.0).

Table 1

Performance of logistic classifier using EV proteins.

Category	# Subjects	Specificity (%, 95% CI)a	Sensitivity (%, 95% CI)a
Controls	184	99.5 (97.0–99.9)
All cancer cases	139		71.2 (63.2–78.1)
Stage I	88		70.5 (60.2–79.0)
Stage II	51		72.5 (59.1–82.9)
Pancreatic cancer	47		95.7 (85.8–98.8)
Ovarian cancer	44		75.0 (58.9–85.4)
Bladder cancer	48		43.8 (30.7–57.7)

aTwo-sided 95% Wilson confidence intervals.