Liquid Biopsy Differentiation of Pancreatic Cancer From Non‐Cancerous Pancreatic Disease Using Dielectrophoresis‐Recovered Nanoparticles Carrying Cell‐Free DNA and Protein Biomarkers

ABSTRACT

Cancer‐derived extracellular vesicle (EV) nanoparticles carry important biomarkers but are difficult to recover from plasma, making EV‐based diagnostics a challenge for clinical settings. Here, we demonstrate nanoparticle‐based detection of pancreatic cancer using dielectrophoresis (DEP) nanoparticle recovery technology, which purifies nanoparticles from undiluted plasma and quantifies associated biomarkers. We combined both nanoparticle recovery and biomarker quantification on a single device by simultaneously collecting cell‐free DNA nanoparticles and EVs followed by on‐chip biomarker fluorescent staining for DNA and Glypican‐1. Using a blinded cohort, these biomarkers differentiated pancreatic cancer from benign pancreatic diseases, including cysts, pancreatitis, and precancerous low‐grade intraductal papillary mucinous neoplasm (IPMN) lesions, with a sensitivity of 0.92, a specificity of 0.83, and an AUC of 0.93. The AUC increased to 0.97 for patients over 50 years old. This is higher than the standard invasive endoscopic ultrasound‐guided fine needle aspiration tissue biopsy procedure (AUC 0.79). This study is among the first demonstrating a combined threshold of DNA and protein levels that can distinguish pancreatic cancer from its precursor IPMN lesions. We also demonstrated the detection of early‐stage pancreatic cancer and high‐grade in situ precancerous lesions. This DEP‐based technique shows that multiple types of cancer‐derived nanoparticles can be quickly and easily recovered from plasma making it promising for future clinical diagnostics.

A new nanoparticle‐based biomarker panel is described that can differentiate pancreatic cancer from benign pancreatic disease with a high level of performance. This was enabled by microelectrode array recovery technology that used dielectrophoresis (DEP) to quickly and efficiently recover the nanoparticles from patient plasma samples, followed by on‐chip biomarker quantification. This technique has the characteristics necessary for future clinical translation of nanoparticle‐based diagnostics.

1. Introduction

Throughout their development, tumors release different types of nanoparticles into circulation [1, 2, 3, 4]. These nanoparticles include actively secreted extracellular vesicles (EVs), such as exosomes, which carry protein biomarkers overexpressed by cancer cells (Figure 1A) [5, 6, 7, 8, 9]. Additional types of nanoparticles are also released by cancer cells through passive mechanisms, including necrosis and cellular lysis [6, 10]. Combining different types of nanoparticle‐based biomarkers, such as DNA and protein, whose plasma levels are independent from one another, integrates information into a panel that is advantageous because it can better account for heterogeneity in genomic alterations and protein expression across the patient population compared to just a single biomarker. This is necessary to improve detection of early‐stage cancers and risk‐stratify preneoplastic lesions.

FIGURE 1.

Origin of cancer‐derived nanoparticles and their collection using DEP. (A) Cancer cells are continuously releasing nanoparticles into circulation through both active secretion, as is the case for EVs, and through passive methods such as cellular necrosis which likely produces the cf‐DNA carrying nanoparticles. cf‐DNA could also be carried on the surfaces of EVs. (B) A schematic representation of the DEP‐based nanoparticle collection process using the microfluidic chip that contains a microelectrode array on the bottom surface. The array consists of hundreds of electrodes, a section of which is shown here. The plasma sample is introduced into the microfluidic chamber where the non‐uniform AC electric field is created by the electrode array. This creates a DEP force on the nanoparticles within a certain size range and material composition drawing them to the electrode edge where the field gradient is the strongest. These collected particles are held with enough force to allow a wash to remove the bulk plasma thereby purifying the particles and preparing them for biomarker detection and quantification through fluorescent staining techniques.

However, conducting an analysis of DNA and protein biomarkers carried by different nanoparticle types on a single plasma sample is challenging. Despite circulating cancer‐derived nanoparticles being a rich source of tumor‐related biomarkers [6, 11, 12] they are not currently used for liquid biopsy‐based clinical diagnosis [13] in part because their recovery from plasma and analysis of biomarker content remains a challenge. This is due to their small size and low buoyant density, resulting in a low physical contrast with the surrounding bulk plasma. Traditional ultracentrifugation, size exclusion chromatography, or commercial recovery kits use these physical characteristics of the EV nanoparticles to separate them from the surrounding plasma [14, 15, 16, 17, 18] and are time‐consuming and labor‐intensive. Affinity‐based collection techniques can potentially bias the population of collected particles, confounding the quantified values for a certain biomarker of interest [19]. In addition, traditional recovery and analytical techniques for each nanoparticle type are incompatible with one another, so multiple blood samples and tests need to be run to analyze multiple particle types. These traditional techniques present a challenge for clinical translation due to the increased time and labor involved to perform multiple isolation and analytical techniques for a single patient.

Traditional nanoparticle recovery methods also present a challenge when developing biomarker panels for early‐stage cancer detection because it requires access to early‐stage patient plasma samples. These banked plasma samples from early‐stage cancer patients are rare and volume‐limited and cannot support the large volume requirements of traditional recovery methods. New methods need to be developed that can recover multiple particle types simultaneously from a single sample [14, 15, 16, 17, 18].

To address these challenges, and streamline the recovery of multiple nanoparticle types, we used high conductance dielectrophoresis (DEP) to simultaneously recover extracellular vesicles and cell‐free DNA (cf‐DNA) carrying nanoparticles from undiluted plasma and demonstrate the use of these particles to differentiate between patients with invasive cancer and patients with pre‐invasive and benign disease. DEP is a label‐free technique that uses the large contrast in the dielectric properties between the particles of interest and the surrounding media to preferentially create a force on lipid and protein‐based nanoparticles within a certain size range and material composition [20, 21, 22]. The label‐free nature of DEP collection means that the specific biomarkers carried by the particles do not influence their collection, allowing us to recover a uniform sampling of the particles present in the plasma sample.

Our DEP technique uses a microelectrode array to apply a nonuniform electric field to the plasma sample. This induces a dipole in the nanoparticles, which then subsequently experience a force [23]. The DEP force that a homogeneous spherical entity experiences is defined by Equation 1:

$$F_{DEP}=2\pi r^3{\epsilon }_o{\epsilon }_{m}Re\left[CM\right]\nabla {\left|\mathit{E}\right|}^2$$ (1)

where r is the radius of the particle, ε _o is the permittivity constant for a vacuum, ε _m  is the medium permittivity constant, CM is the frequency‐dependent Clausius–Mossotti factor relating to the polarizability of the particle and medium, and E is the electric field. [23]

This expression describes that the magnitude of the DEP force exerted on a particle depends on the electric field gradient and that the direction of particle motion will follow this gradient [24]. Motion toward the high field region is known as positive DEP (pDEP). The circular electrodes used in our microfluidic DEP chips create a high field region around the electrode edge [24, 25, 26], which is where these nanoparticles are observed to accumulate (Figure 1B). They are held in place with enough force to allow a fluidic wash to remove the bulk plasma and purify the particles, enabling subsequent immunostaining and fluorescent dye staining quantification of the collected biomarkers. The concentration of the biomarkers around the electrode edge locally amplifies the signal (Figure 1B). Images of the experimental setup are shown in Figure S1.

The platinum electrodes are coated in a protective hydrogel layer that reduces the electrolysis of water, thereby enabling direct operation in high‐conducting undiluted plasma as well as enhancing particle collection at the electrode edge [26]. This application of DEP to undiluted high‐conducting plasma is unique in the DEP and microfluidic fields [27, 28, 29, 30], where most of the work is done with diluted plasma to prevent electrochemical reactions like electrolysis or physical clogging of the channels. Avoiding plasma dilution is desirable because dilution increases sample preparation time and handling complexity and reduces the concentration of low‐abundance biomarkers.

Our previous work has shown that this DEP technique recovers EVs, such as exosomes, from cancer patient plasma and that collected EVs can be visualized using scanning electron microscopy (SEM) imaging [27]. These SEM images show that the electric field generated by the electrode array does not compromise the integrity of the EV membranes. The intact nature of the EV membranes is further demonstrated by the need to treat them with a membrane‐permeabilizing agent to label for internal protein biomarkers such as TSG101 [27]. We show here that the biomarkers carried by these recovered nanoparticles can be detected on‐chip using immunofluorescence staining techniques. Additional SEM images in Figure S2 show collection of intact EVs by chips used in the experiments presented here. This fully on‐chip analysis is advantageous because it streamlines the process and prevents biomarker loss through fluid transfers that are required for off‐chip analysis. Maintaining the highest signal possible is critical for low‐abundance samples, such as those from early‐stage cancer patients.

Currently, the standard diagnostic test for PDAC is an invasive endoscopic ultrasound‐guided fine needle aspiration (EUS/FNA) tissue biopsy, a costly and highly specialized procedure performed by trained clinicians [31]. The FNA has associated risks where 1/100 patients will develop acute pancreatitis from the procedure [32]. This is a serious complication with a mortality rate of 1/10 [33]. Currently, 60%–76% of patients referred for EUS/FNA do not have PDAC [34, 35, 36]. The EUS/FNA procedure is a significant financial cost and health risk to patients with pancreatic cysts, most of whom do not have PDAC. A simple blood draw for a liquid biopsy diagnostic test would be significantly lower in financial cost and potential health risk. CA 19‐9 is the only standard blood‐based pancreatic cancer biomarker currently available, but it is only used to characterize and monitor patients with known PDAC tumors and is not recommended for PDAC screening of healthy or high‐risk patients [37, 38, 39].

Our goal is to address this critical unmet clinical need to stratify patients at high risk of having PDAC into categories of high‐probability and low‐probability for the actual cancer by developing DEP‐based liquid biopsy EV diagnostics. Those with high‐probability could justifiably benefit from the invasive EUS/FNA procedure to confirm PDAC despite the cost and risk. Low‐probability patients could benefit from close clinical follow‐up and monitoring.

In addition, the ability to distinguish and stratify different grades of IPMN would make a considerable impact on the lives of people with identified pancreatic lesions and provide an additional level of surveillance for those with familial history of PDAC, and/or have diabetes mellitus or germline mutations like BRCA2, which put them at increased risk for developing pancreatic cancer [40, 41, 42].

Here, we demonstrate the ability of this DEP device to simultaneously collect different nanoparticle types from a single plasma sample followed by on‐chip biomarker quantification. We also show that these combined biomarkers can perform better than traditional tissue biopsy at detecting PDAC.

These results demonstrate a significant advancement in nanoparticle‐based diagnostics through the innovative use of DEP technology, which has the characteristics and performance necessary for clinical translation.

2. Results and Discussion

The cancer detection strategy developed here uses DEP technology to quickly and efficiently recover cancer‐derived nanoparticles from plasma, followed by on‐chip biomarker quantification using fluorescent staining. This process was used to simultaneously collect and quantify two different types of nanoparticle‐associated biomarkers. The first biomarker was EV‐associated Glypican 1 protein. The second was nanoparticle‐associated cf‐DNA. We applied this two‐biomarker panel to a blinded cohort of patient samples to evaluate efficacy in differentiating pancreatic cancer from benign pancreatic disease.

2.1. DEP‐Based Particle Recovery From Plasma

A blinded cohort consisting of patients with PDAC and patients with benign cysts and pancreatic precancer diseases, including pancreatitis and IPMN (Table 1), was used to determine the efficacy of DEP‐based nanoparticle recovery and its application to differentiate PDAC from non‐cancer pancreatic diseases. The blinded study design eliminated the potential of user bias, and the samples were only unblinded to reveal the disease status of each patient after all plasma samples had been run and the data analyzed. A one‐ and two‐year follow‐up chart review for each patient after the initial blood draw and diagnosis was conducted. None of the non‐cancer pancreatic disease patients had developed PDAC during this time.

TABLE 1.

Patient Cohort Summary.

		Benign	Pancreatitis	IPMN				Pancreatic Cancer
					Grade				Cancer Stage
				Overall	Low	Medium	High	Overall	I	II	III	IV	Liver Cancer	Total
Samples	n (%)	9 (25)	10 (27.77)	6 (16.66)	5 (13.8)	—	1 (2.77)	11 (30.55)	1 (2.77)	2 (5.55)	6 (16.66)	1 (2.77)	1 (2.77)	36
Sex	Men (%)	3 (8.33)	7 (19.44)	3 (8.33)	2 (5.55)	—	1 (2.77)	8 (22.2)	1 (2.77)	1 (2.77)	4 (11.11)	1 (2.77)	1 (2.77)	21
	Women (%)	6 (16.66)	3 (8.33)	3 (8.33)	3 (8.33)	—	—	3 (8.33)	—	1 (2.77)	2 (5.55)	—	—	15
Age	Median (range)	58 (30‐77)	54.7 (42‐78)	70.5 (44‐84)	71.4 (44‐84)	—	66 (66)	69.2 (54‐85)	85 (85)	68.5 (66‐71)	69.66 (58‐84)	68 (68)	54 (54)	—

Circulating EVs range from 50 to 1000 nm in diameter [43], and the DEP‐based technique was successful in recovering a subset of these nanoparticles from patient plasma samples. The size distribution of recovered nanoparticles showed a peak at 95 nm with 90% of the particles being between 65–359 nm in diameter (Figure 2A; Figure S3). This encompassed all the particles collected by DEP from plasma, including both the EV‐based nanoparticles as well as the cf‐DNA‐carrying nanoparticles. Both the cf‐DNA nanoparticles and the EVs carrying Glypican‐1 were collected simultaneously and stained on the chip showing colocalization to the same regions around the circular electrode edge (Figure 2B; Figure S4). It is also possible that nanoparticle protein aggregates containing Glypican‐1 were also collected. Around the electrode edge is where the electric field gradient was the highest and where particle collection was expected to occur [24, 25, 26, 27]. The Pearson Correlation Coefficient (R) value for the spatial distribution of fluorescent signal between cf‐DNA nanoparticles and EVs was 0.72 ± 0.026 (mean ± SD, n = 3) showing a high degree of colocalization. Representative images of Glypican‐1 and cf‐DNA staining on the electrode arrays are shown for the different disease conditions in the cohort (Figure 2C). The corresponding bright field images show where the circular electrodes are located. There is a general trend of increasing expression level for Glypican‐1, going from benign cyst patients to pancreatitis to low‐grade IPMN to PDAC. The cf‐DNA levels were more variable. The observed expression levels of Glypican‐1 were independent from the cf‐DNA levels and they did not trend together (linear fit R² = 0.008). Representative images of the larger array are shown in Figure S5.

FIGURE 2.

Characterization of the nanoparticles collected from patient plasma by DEP. (A) Size distribution of nanoparticles recovered by DEP from PDAC patient plasma showing a peak at 95 nm with 90% for the particles between 65–359 nm in diameter. (B) The circular edge of the electrode has the highest electrical field gradient, and nanoparticles are drawn to this region by DEP. Glypican‐1‐containing nanoparticles are stained in red, and cf‐DNA carrying nanoparticles stained with SYBR Green are in green. Yellow shows where the two types of particles are colocalized. A Pearson's R‐value was found to be 0.72 ± 0.026 (mean ± SD, n = 3), indicating a strong spatial correlation. (C) Representative images of fluorescent staining of Glypican‐1 (GPC1) and cf‐DNA on the electrode arrays for different patient disease states within the cohort.

2.2. DEP Collection of Circulating cf‐DNA Nanoparticles

The DEP force created by the electrode array in these microfluidic chips was capable of collecting nanoparticles carrying cf‐DNA with a particle size distribution peak at 105 nm in diameter and with 90% of the particles ranging from 65–475 nm in diameter (Figure 3A; Figure S6). This was determined by fluorescently staining the DNA on the DEP collected nanoparticles using SYBR Gold and using the ZetaView nanoparticle tracking system to analyze just the fluorescent particles.

FIGURE 3.

Characterization of DEP collected nanoparticles carrying cf‐DNA. (A) Size distribution of cf‐DNA nanoparticles collected by DEP from PDAC patient plasma with a peak diameter at 105 nm with 90% of the diameters between 65–475 nm. The cf‐DNA particles were labeled with SYBR Gold DNA stain and analyzed using fluorescence detection on a ZetaView nanoparticle tracking system. (B) Nanopore sequencing was used to determine the distribution of cf‐DNA fragment lengths in these DEP recovered nanoparticles for 2 pancreatic patient samples from just 10 µl of plasma. Most of the fragments were above 250 bp in length (red dashed line). This indicates that the cf‐DNA nanoparticles are being generated from a mechanism other than apoptosis, possibly from cellular necrosis. (C) In contrast, the gold standard Qiagen kit applied to 2 PDAC patient plasma samples, collected a different size range showing the majority of DNA fragments were below 250 bp in length (red dashed line) with a clear peak at 185 bp, which corresponds to the size of DNA fragments generated by an apoptosis mechanism. The second smaller peak at 330 bp corresponds to a ∼2 times length increase from the first peak, meaning the second peak may also be comprised of apoptotic DNA. The Qiagen recovery procedure required 1 mL of patient plasma. (D) Comparison between DEP and Qiagen of the percentage of cf‐DNA fragments above and below 250 bp.

Nanopore sequencing showed that the DNA fragments carried by these cf‐DNA particles were mostly between 250–800 bp in length (Figure 3B,D), which is different from fragments generated by apoptosis. Controlled DNA degradation during cellular apoptosis cleaves the DNA in between nucleosome subunits, resulting in shorter DNA fragments with a narrow size range between 160–200 bp in length [44, 45, 46, 47, 48]. This shows that DEP collected a subfraction of cf‐DNA that was associated with nanoparticles while allowing apoptotic‐sized DNA fragments, which are too small for DEP collection, to wash away.

In contrast to DEP collection of nanoparticles, gold standard Qiagen DNA recovery kits collect DNA fragments using a chromatographic‐based collection mechanism that is not designed to capture DNA in nanoparticle form. The Qiagen kits collected cf‐DNA fragments ranging mostly from 160–250 bp in length, peaking at 185 bp, which corresponds to an apoptosis‐based origin (Figure 3C,D). There can also be a smaller population of apoptosis‐derived DNA fragments that are multiple increases of this 160–200 bp length range corresponding to two or more nucleosome subunits [44, 45, 46, 47, 48]. This could account for the smaller peak in the Qiagen collected cf‐DNA at 330 bp in length that was ∼2 times the size of the larger peak (Figure 3C). The DEP collected cf‐DNA nanoparticles did not have a multiple repeat pattern (Figure 3B). These differences in the cf‐DNA fragment lengths between DEP and Qiagen collected methods were consistent across the different patients studied (Figure 3).

The longer 250–800 bp cf‐DNA fragments carried by the DEP collected nanoparticles indicate this cf‐DNA was generated by a different mechanism than apoptosis (160‐200 bp). It is also different from the >1000 bp fragments of genomic DNA released from white blood cell rupture that can occur during blood collection [47, 49, 50, 51]. These DEP collected cf‐DNA nanoparticles may be coming from cancer‐related necrosis, where dying cells lyse open, releasing large structures of genomic DNA with bound proteins into circulation. In contrast to apoptosis, these cellular lysis events result in a wider range of longer DNA fragment lengths due to random degradation from nucleases [10, 45, 48, 52, 53] and fluidic shearing in circulation. As these larger structures break up in circulation, the cf‐DNA and associated nuclear proteins could form nanoparticle‐sized fragments or could also become associated with other nanoparticles in circulation by attaching to their surfaces.

DEP is only able to recover cf‐DNA in these protein‐associated or EV‐associated nanoparticle structures because the application of a 14 kHz signal does not collect a significant number of particles below 65 nm in diameter (Figure 2A). A nucleosome is 11 nm in diameter, including the histone proteins and the ∼180 bp of DNA wrapped around it [54, 55]. To carry 800 bp of DNA, there would need to be 5 nucleosome structures aggregated together. This would form a nanoparticle that is ∼30 nm in diameter [55, 56], which is too small to be collected by DEP. If the 800 bp of DNA were not associated with histones, it would collapse into an even smaller structure. This indicates that the DEP collected cf‐DNA is not free‐floating DNA strands or free‐floating nucleosome structures. To reach the particle sizes necessary for DEP collection, the cf‐DNA needs to be contained within larger structures that have proteins in addition to histones or could also be adhered to larger EV nanoparticles.

These nanoparticles carrying cf‐DNA fragments, possibly produced through non‐apoptotic mechanisms, have not been extensively studied before for diagnostic applications and may be enriched for fragments originating from necrotic tumor regions, making them a valuable biomarker for this study.

2.3. Biomarker Expression Levels and Bivariate Analysis

The cancer samples showed a statistically higher level of EV‐associated Glypican‐1 expression compared to the non‐cancer controls consisting of pancreatic cysts, pancreatitis, and low‐grade IPMN disease states (Figure 4A). This is important because differentiation between PDAC and non‐cancer pancreatic diseases is a major challenge due to the non‐cancer disease states causing inflammation damage to pancreatic tissue which often results in similar biomarker overexpression patterns to PDAC [57, 58]. The average cf‐DNA nanoparticle level was higher in the cancer condition compared to the non‐cancer conditions, but was not statistically significant (Figure 4B). Further breakdown of the data by disease category shows a significant difference in Glypican‐1 levels between the PDAC patients and the benign cyst negative controls (Figure S7).

FIGURE 4.

Differentiation of pancreatic cancer from non‐cancer pancreatic disease. (A) A statistically significant difference in EV‐associated Glypican‐1 expression levels was observed between the patients with cancer and those with non‐cancer pancreatic diseases. (B) The average cf‐DNA level for cancer patients was higher than for the patients with non‐cancer pancreatic disease, although it was not statistically significant. (C) A bivariate analysis, combining the results of cf‐DNA levels and Glypican‐1 expression levels, was able to successfully differentiate cancer patients from patients with non‐cancer pancreatic disease. This 2D analysis allowed for the application of a Support Vector Machine (SVM) machine learning algorithm capable of performing classification, which defined three different sections of the graph. The cancer section is shown in red. The non‐cancer pancreatic disease section was further divided into a Low‐Risk section in green, where patients had relatively low levels of both biomarkers, and a High‐Risk section in grey, where patients had higher levels of both biomarkers, moving them closer to the cancer section. Further ROC analysis of these results is shown in Figure 5. The cohort composition is described in Table 1. (^** p<0.01 by two‐sided Mann–Whitney U test, upper and lower box boundaries are 25th to 75th percentiles).

Combining both the Glypican‐1 levels and the cf‐DNA nanoparticle levels in a 2D bivariate analysis showed a clear separation between the majority of the cancer patients and the non‐cancer diseases (Figure 4C). These separations were more distinct than could be achieved with either biomarker on its own.

This bivariate separation enabled the application of a machine learning technique, known as Support Vector Machine (SVM) [59], to define boundaries between different groups of patients. These defined regions included ‘Cancer’, where the majority of PDAC cancer patients would be expected, ‘High‐Risk’ where the majority of pancreatitis and low‐grade IPMN would be expected, and ‘Low‐Risk’ where the majority of the benign cysts would be expected. The SVM analysis used here is naturally robust with respect to overfitting because of a built‐in “large margin” mechanism. We also added additional measures to prevent over‐fitting by allowing “slacks” in the SVM margin formulation, which allows some errors on classifications of some samples instead of insisting on correct classification on all samples. This resulted in second‐order polynomial decision boundaries that were the lowest order possible, resulting in smooth boundaries without large‐angle features. Smoother functions are less likely to overfit. A comparison between the performance of SVM and several other machine learning models, including Random Forest, Decision Tree, Linear Discriminate Analysis (LDA,) K Nearest Neighbor (KNN), Navier–Bayes (NB), and Support Vector Machine (SVM) is shown in Table S1, demonstrating that SVM resulted in the best differentiation between cancer and non‐cancer patients.

The majority of the patients in the red ‘Cancer’ region had cancer and were observed to have high levels of Glypican‐1 expression and moderately high levels of cf‐DNA compared to benign controls. The majority of the patients in the green ‘Low‐Risk’ region had benign disease and had low levels of both biomarkers. The majority of patients with inflammatory non‐malignant pancreatitis and patients with precancerous low‐grade IPMNs were located in the grey ‘High‐Risk’ region in between the ‘Cancer’ and ‘Low‐Risk’ regions. Patients in this ‘High‐Risk’ region would benefit from follow‐up diagnostics and monitoring because the presence of pancreatitis and low‐grade IPMN increases the chances of developing PDAC [34]. It is notable that the ‘High‐Risk’ region did not include any cancer patients, and that the ‘Cancer’ region included 1 IPMN patient that was high‐grade and would be at high risk to transform to PDAC, making it likely to be surgically removed as discussed later.

We designed the blinded study cohort to be comprised of samples from patients that are representative of what would be seen in a clinical environment where patients come in with clinical symptoms of pancreatic cancer, such as abdominal pain, jaundice, and weight loss [31, 60]. The patients in our cohort all had come to the clinic presenting with symptoms of abdominal pain, which ended up being caused by a variety of conditions, including benign cysts, IPMN, pancreatitis, and PDAC. The patients with benign cysts are the negative controls since this condition is the least likely to lead to cancer and were shown to have a significantly lower expression level of Glypican‐1 compared to cancer patients (Figure S7).

Our study cohort did not include completely healthy people because they would not show clinical symptoms and are not the target for this screening test. It has been shown that screening the general population for PDAC, even with the best‐case scenario testing performance, would result in too high of a false positive rate to make this a viable approach [61]. However, PDAC liquid‐biopsy screening will be very valuable when applied to a high‐risk group of patients who either have a genetic predisposition for PDAC, new onset diabetes, or who have clinical symptoms, such as our cohort.

This DEP‐based analysis of biomarker levels carried by nanoparticles was able to achieve clear separation between patients with non‐cancer pancreatic diseases and patients with actual cancer, which is critical for clinical applicability (Figure 4C).

It is important to note that within these SVM‐defined classification boundaries, there are two false positives and two false negatives. The two stage 2 PDAC patients in the ‘Low‐Risk’ region are false negative cases. There were also two pancreatitis patients in the ‘Cancer’ region, making them false positives. As discussed later, even with these false positives and false negatives, the performance of this discrimination test is better than the gold standard EUS/FNA.

False positives are a major source of patient harm, and a critical challenge for early detection because many blood‐based markers can be expressed both by the cancer state and the corresponding precancer, such as IPMN. This liquid biopsy work is among the first studies that we are aware of showing that a threshold, comprised of both DNA and protein information, can distinguish between PDAC and IPMN lesions. There was one IPMN in the ‘Cancer’ region, making it technically a false positive. However, this was a high‐grade IPMN. There are three grades of precancerous IPMN (low, intermediate, and high) [62]. Low and intermediate grades are usually monitored without surgical intervention. However, a high‐grade IPMN can quickly convert to a stage 1 PDAC and so is usually treated as if it were an early‐stage PDAC and is surgically removed. The aggressive biology revealed in the DEP test, together with the high‐grade of this IPMN, may reflect a lesion at very high risk of converting to PDAC. For this reason, we counted this high‐grade IPMN, which would likely be surgically removed, as a successful cancer identification for this analysis. Being able to identify the presence of a high‐grade IPMN, and the stage 1 and stage 2 PDAC samples is an indication that this technique may be applicable for identifying patients with early stages of PDAC. The identification of the stage 2 PDAC patient required the combination of both cf‐DNA nanoparticles and Glypican‐1 in order to clearly separate it from non‐cancer controls.

The individual with the highest levels of Glypican‐1 was originally diagnosed as not having pancreatic cancer. However, 6 months after this blood draw, they were found to have a stage 2 hepatocellular carcinoma, which has been shown to express elevated levels of Glypican‐1 [63]. There was no cancer in the pancreas. This means that this liquid biopsy diagnostic technique was able to detect elevated Glypican‐1 and cf‐DNA nanoparticle levels that signified a cancerous condition within this patient before it was clinically diagnosed. This patient was included with the cancer patients for statistical analysis. From a clinical perspective, using a blood test to detect the presence of an early stage cancer, even if it doesn't specifically indicate which organ it is located in, is highly valuable because this would prompt further diagnostic procedures, such as more expensive MRI or CT imaging, to determine if there are any suspicious masses that could be followed up with by tissue biopsy analysis. This is important because a general warning signal from a blood test can increase the chances of catching a tumor early.

Overall, the liquid biopsy‐based DEP technique showed a differentiation between cancer and non‐cancer pancreatic disease with a sensitivity of 0.92, specificity of 0.83, and an area under the receiver operating characteristic (ROC) curve (AUC) of 0.93. This performance is better than the invasive EUS/FNA diagnostic tissue biopsy procedure (AUC of 0.79) [64] as discussed below.

2.4. ROC Curve Analysis and Comparison to EUS/FNA

The ROC analysis in Figure 5 shows the performance of the Glypican‐1 and cf‐DNA nanoparticle biomarkers individually and in combination. Glypican‐1 by itself had an AUC of 0.77 (Figure 5A), which when combined with the cf‐DNA levels was elevated to 0.93 using the SVM delineations in the bivariate analysis (Figure 5C). Just these two biomarkers together performed better than the AUC of 0.79 for the invasive EUS/FNA tissue biopsy diagnostic procedure [64]. A major benefit of this DEP‐based liquid biopsy technique is that the blood draw has a significantly lower financial cost and lower associated health risk compared to the EUS/FNA [32, 33]. The Glypican‐1 and cf‐DNA nanoparticle biomarkers also performed better when compared to circulating levels of EpCAM (AUC of 0.726) [65] and CA19‐9 (AUC of 0.82) [66].

FIGURE 5.

Receiver operating characteristic curve analysis. (A) ROC plot for Glypican‐1 alone. (B) ROC plot for cf‐DNA nanoparticles alone. (C) ROC plot combining Glypican‐1 and cf‐DNA nanoparticles together with an AUC value determined using the SVM delineations shown in Figure 4C. The AUC of 0.93 showed improvement over each biomarker individually, indicating this liquid biopsy analysis of cancer‐derived nanoparticles, enabled by DEP, can perform at a clinically relevant level without an invasive procedure. For comparison, the AUC for the gold standard EUS/FNA tissue biopsy diagnostic procedure is 0.79⁶⁴ (D) ROC plot for the application of the SVM analysis to just patients from the cohort over the age of 50, resulting in an increased AUC of 0.97. The same SVM delineations defined for the full cohort in Figure 4C were applied here (Figure S9).

We applied the SVM analysis to see if differentiation could be achieved between the non‐cancer conditions. Pancreatic benign cysts were compared to the diseases that are associated with inflammation, including both IPMN and pancreatitis, resulting in an AUC of 0.77 (Figure S8). This indicates that cf‐DNA and Glypican 1 levels could be useful in differentiating between benign cystic lesions and precancerous inflammatory lesions, which could help physicians determine which additional diagnostic tests to run. IPMN and pancreatitis could not be differentiated from one another based on these markers.

The addition of cf‐DNA nanoparticle levels to the biomarker panel is important for several reasons. The first is that including cf‐DNA enabled differentiation between PDAC and non‐cancer pancreatic disease cases, which the Glypican‐1 biomarker could not differentiate by itself. Both biomarkers were required to achieve a performance that was better than the gold standard EUS/FNA biopsy technique. This is a new result that hasn't been shown before and is of critical clinical importance because to our knowledge, this is the first time that a threshold combining DNA and protein biomarker levels has been shown to distinguish between PDAC and precancerous low‐grade IPMN lesions. It also shows that the subfraction of cf‐DNA in nanoparticle form collected by DEP has diagnostic value.

The second reason the addition of cf‐DNA levels is important is because there was no correlation between the cf‐DNA and Glypican‐1 levels (linear fit R² = 0.008), indicating these are two independent biomarkers whose expression levels are driven by different independent mechanisms. This makes their combination particularly important because it can help address the heterogeneity across the cancer population due to pancreatic tumors developing through different pathways. The combination of both biomarkers was necessary to differentiate 14 of the samples near the High‐Risk/Cancer boundary in Figure 4C that could not be differentiated by either biomarker alone. This cf‐DNA and Glypican‐1 biomarker panel shows promising results being able to differentiate PDAC from non‐cancer pancreatic disease across the very heterogenous cohort studied here, where both cf‐DNA and Glypican‐1 biomarker expression levels span two orders of magnitude.

The third reason is that cf‐DNA nanoparticle levels are easily analyzed by DEP due to their simultaneous collection with EVs carrying Glypican‐1 and the DNA's ability to be stained by a simple incubation with SYBR Green dye. This is important because our unique ability to look at two different types of nanoparticles from the same sample on the same chip combines collection and detection, which are the two most difficult and labor‐intensive parts of nanoparticle‐based diagnostics. This is a critical step forward to bring nanoparticle‐based diagnostics to a clinical setting where efficiency and ease of use are major considerations in the choice of a biomarker panel.

Age is an important factor in the development of PDAC where over 90% of diagnosed cases occur in patients over the age of 50 [67]. We applied the SVM‐based delineations defined for the whole cohort (Figure 4C) to just the patients over the age of 50, which resulted in an increased AUC of 0.97 (Figure 5D; Figure S9). This indicates the performance of the panel is improved when applied to the older population.

The specificity of the cf‐DNA nanoparticle levels was observed to be lower than the specificity for the Glypican‐1 levels. There are two main factors contributing to this difference. The first is that in the non‐cancer disease patients, there was a higher degree of variability in the cf‐DNA levels compared to the Glypican‐1 levels. There were 5 non‐cancer patients who had 2–3 times higher cf‐DNA levels than the average of the other non‐cancer patients (Figure 4). These high cf‐DNA nanoparticle levels might be a result of tissue damage and necrotic cell death originating from secondary inflammation caused by the pancreatitis and possibly by IPMN.

The second factor is that DEP detected Glypican‐1 is carried on EVs that are actively secreted by the tumor tissue, and it is known that tumor cells can have higher levels of EV secretion compared to healthy cells [5]. It is possible that the overexpression of Glypican‐1 and the oversecretion of EVs by tumor tissue was able to overcome the inherent heterogeneity of circulating EV levels to show a clear separation between cancer and non‐cancer conditions (Figure 4A). The inflammation caused by non‐cancer diseases apparently does not stimulate the oversecretion of EVs and overexpression of Glypican‐1 that rises above the average Glypican‐1 levels seen for cancer. This could contribute to the Glypican‐1 levels having higher specificity and sensitivity compared to the cf‐DNA nanoparticle levels.

It is important to note that the higher level of performance for the DEP method compared to EUS/FNA was obtained using a total of just 30 µl of plasma per patient and required only 15 min to collect the nanoparticles in a concentrated form around the electrode edges from the undiluted plasma sample. These are the features that make high conductance DEP technology promising for future clinical translation. This version of DEP works directly in high‐conducting undiluted plasma, which is unique in the DEP and microfluidics fields. This reduces sample handling and preparation time and enables greater sensitivity since dilution is unnecessary. All nanoparticle collection and biomarker quantification was done on the chip, thereby avoiding signal loss caused by fluid transfers. Nanoparticle collection is simple requiring just a few steps, making full automation possible, which is important in the clinical laboratory setting. Working with small 30 µl volumes speeds sample loading and washing times. The DEP technique does not require samples to be chemically fixed, which simplifies handling, thereby speeding analysis, and eliminates the need for toxic compounds. These characteristics are key to enable translation of EV‐based diagnostics to the clinic.

2.5. Comparison to Standard EV Recovery and Biomarker Quantification Methods

The nanoparticle‐based Glypican‐1 and cf‐DNA biomarkers were independently evaluated using standard methods that employ entirely different mechanisms for EV recovery and analysis compared to the electrokinetic mechanism of DEP. These standard methods consisted of EV recovery from plasma using size exclusion columns followed by ELISA‐based Glypican‐1 quantification and Qubit‐based DNA quantification. Both ELISA and Qubit analytical techniques were required due to the inherent differences in the quantification procedures for both EVs and cf‐DNA, which prevented them from being analyzed with the same standard laboratory instrumentation.

The bivariate analysis of the resulting biomarker levels obtained using these standard methods showed good separation between PDAC and pancreatitis patients (Figure S10). This separation is consistent with the results from DEP‐based nanoparticle collection and on‐chip biomarker quantification. This shows the robust ability of the nanoparticle‐based Glypican‐1 and cf‐DNA biomarkers to separate PDAC from non‐cancer pancreatic disease.

One major difference between DEP and the standard methods is the amount of time and labor that is involved. The nanoparticle collection from plasma using DEP took 15 min with little supervision with up to 3 samples being run in parallel. At least 2 chips can be run in parallel for a total of 6 patient samples being analyzed simultaneously. DEP also concentrated the nanoparticles around the electrode edge giving an enhanced localized signal‐to‐noise ratio compared to the background. The column separation took 1 h per sample, and the nature of the process makes it difficult to run multiple columns in parallel. The nanoparticles came out of the column diluted in 100 µl of eluant, which makes direct biomarker immunostaining and quantification difficult. The standard method required three separate instruments resulting in additional labor and time for the fluid transfers and analysis. In contrast, the DEP method combined all the purification and analysis steps for both Glypican‐1 and cf‐DNA into a single device.

In summary, both the DEP and standard methods confirmed that nanoparticle‐based Glypican‐1 and cf‐DNA levels can be used to achieve a clear separation between PDAC and pancreatitis patients. The DEP method required less time, less labor, and less analytical equipment, demonstrating that DEP has the necessary characteristics for future clinical translation.

3. Conclusions

We show for the first time the ability of a nanoparticle‐based biomarker panel, comprised of both DNA and protein biomarkers, to differentiate PDAC from non‐cancer pancreatic diseases, and also importantly, inflammatory pancreatitis and low‐grade IPMN lesions. A 2D bivariate analysis of nanoparticle‐based cf‐DNA and Glypican‐1 levels showed clear separation allowing an SVM algorithm to define a boundary between cancer and non‐cancer conditions. This was enabled by our technique's unique ability to use a single DEP device to simultaneously collect different nanoparticle types from the same plasma sample, followed by on‐chip biomarker quantification.

This biomarker panel was able to differentiate between patients with cancer and patients with non‐cancer pancreatic disease with a sensitivity of 0.92, specificity of 0.83, and an area under the ROC curve of 0.93. When applied to individuals over the age of 50, the age group most likely to be diagnosed with PDAC, the AUC increased to 0.97. This performance is better than the gold standard EUS/FNA diagnostic procedure which has an AUC of 0.79 [64].

Furthermore, we establish for the first time a threshold that integrates information from genomic and proteomic layers of biological data that can differentiate low ‐grade IPMN precancerous lesions from the cancer they can develop into. This distinction is particularly important to help prevent false positives and addresses a critical clinical need because there is overlapping multiomic biology and molecular characteristics between advanced cancers and their precursor lesions [68], and there are currently no available blood‐based tests for the detection and diagnosis of PDAC. In our results, the only IPMN patient within the ‘Cancer’ region had high‐grade pre‐invasive disease, indicating our technique may help identify patients with precancerous lesions that are more biologically aggressive and likely to progress to invasive cancer.

Glypican‐1 has been shown to be carried on actively secreted exosomes [18, 28], but has had mixed results in differentiating PDAC from non‐cancerous pancreatic disease by itself [69, 70, 71, 72]. We found that neither the Glypican‐1 nor the cf‐DNA were sufficient on their own to satisfactorily differentiate PDAC from non‐cancerous disease. Both were needed in the panel to achieve this high level of performance.

The DEP technique collected cf‐DNA in a nanoparticulate form, which enriched for DNA fragments that were 250–800 bp long. These were longer than would be expected if generated through apoptotic mechanisms. These were also longer compared to DNA recovered using traditional chromatography methods. This indicates the DEP collected cf‐DNA might be coming from necrosis mechanisms, which are common in tumors.

Traditionally, EVs and cf‐DNA nanoparticles require different techniques to recover from plasma, but the DEP technique is uniquely able to recover both simultaneously. DEP is less time‐consuming and labor‐intensive than traditional recovery methods, and the chip design combines two of the major analysis steps, nanoparticle recovery and biomarker quantification, into a single device, simplifying sample analysis. This combination also prevents biomarker loss that occurs during fluid transfers, which is important when dealing with low levels of biomarkers from early‐stage cancer patients. High conductance DEP can operate in undiluted plasma, thereby keeping the biomarkers at the highest possible concentrations while simplifying sample handling. These characteristics of DEP nanoparticle collection will facilitate the translation of EV‐based diagnostics to the clinical setting.

These results show that DEP‐based diagnostics can be an attractive screening technique for patients at high risk for developing PDAC, such as individuals with new‐onset diabetes and those with familial or germline mutations indicating PDAC predisposition [41]. In addition, the simple blood draw needed for the DEP diagnostic test has significantly less financial cost and associated health risk compared to the invasive EUS/FNA tissue biopsy procedure [32, 33].

Future work will expand the size of the cohort, including different grades of precancerous lesions, to capture more of the population heterogeneity, as well as expand the existing panel into a multiomic panel by including additional biomarkers to help increase the sensitivity and specificity to detect more PDAC patients, including early stage PDAC. A larger multiomic panel that covers a greater number of proteomic and genomic biomarkers will account for more of the natural variability in biomarker expression across the PDAC patient population.

4. Methods

Additional methods information is located in the Supporting Information Section.

4.1. Materials

1X Dulbecco's Phosphate Buffered Saline, distilled water, Goat anti‐Rabbit IgG, Opti‐MEM reduced serum medium, Alexa Flour 594, and Invitrogen SYBR Green I and SYBR Gold Nucleic Acid Gel stain were purchased from Thermo Fisher Scientific (Waltham, MA). Difco Skim milk was purchased from BD (Franklin Lakes, NJ). Recombinant rabbit monoclonal anti‐glypican1 [EPR19285] (ab199343) and a polyclonal goat anti‐rabbit secondary antibody Alexa Fluor (ab11012) was used as received from Abcam (Cambridge, UK). HeLa cells were provided by CEDAR Cell Banking, Knight Cancer Institute, Oregon Health and Science University (Portland, OR). Silencing RNAs, Glypican‐1 ON‐TARGETplus siRNA, ON‐TARGETplus siRNA non‐targeting control pool, and DharmaFECT 1 transfection reagent were purchased from Horizon Discovery Group plc (Waterbeach, UK). 96 well plates were purchased from CellVis (Mountain View, CA).

Three‐chamber DEP chips (ExoCell EVF1P flow cell, 3‐channel, EF‐CRT‐00002) were purchased from Biological Dynamics (San Diego, CA, USA). An Agilent 33250A 80‐mHz Function/Arbitrary‐Waveform Generator using a Newtons4th Ltd (LPA01) power amplifier produced the AC signal used for DEP collection. An Imager D2 Zeiss microscope, and Zeiss Zen (blue edition) software were used to obtain the brightfield and fluorescence images of the DEP chips. Fluorescence intensity quantification was conducted using custom software written in MatLab [25].

4.2. Patient Plasma Samples

Thirty‐six clinical plasma samples were obtained from the OHSU interventional endoscopy unit under the Oregon Pancreas Tissue Registry (IRB approval #00003609) with use under OHSU IRB approval #00018572. Informed consent was obtained from all subjects. All experimental protocols were carried out in accordance with relevant guidelines and regulations outlined by the OHSU Institutional Review Board. Blood was drawn into EDTA tubes, which were then immediately centrifuged at 2500 g × 10 min, and the plasma banked at −80 C until tested. Patients presented for endoscopic‐ultrasound imaging ± fine needle aspiration biopsy. The inclusion criteria for patient samples in this study required tissue‐based diagnosis of disease state, or at least 3 years of negative clinical follow‐up confirmed by a board‐certified surgical pathologist (Dr. Terry Morgan).

To avoid potential user bias, this cohort was blinded to the researchers. The disease status of the samples was unblinded after all data had been collected and analyzed. The composition of the cohort is shown in Table 1.

4.3. Dielectrophoresis (DEP) Nanoparticle Recovery and on Chip Analysis

To summarize the analysis workflow, the raw biomarker level quantification dataset for each patient consisted of fluorescent images of the electrodes after DEP collection, washing, and staining, such as those shown in Figure 2C. The fluorescence level was quantified using our custom optical quantification software [25] to determine the fluorescence intensity level around the electrode edges across the field of view. Each patient sample was run once, just as they would be in a clinical setting. The following describes the steps in more detail.

Washing buffer (0.5X PBS) was loaded onto the DEP chip channel, and then 30 µl of undiluted human plasma was loaded, drawing in 10 µl total into the chamber at 10 µL/min for sample collection. The flow was stopped for DEP collection. An AC electric field (10 V_pp, 14 kHz) was applied to the electrode array using the waveform generator and power amplifier to create the DEP force for 15 min to collect the particles from the plasma. This was determined to be the highest voltage that could consistently be applied without causing bubble formation through electrolysis. This produced the maximum level of collection [24, 26, 27, 29]. We chose 14 kHz to center the size distribution peak near 100 nm to collect the desired EV fraction (Figures 2A and 3A). The particles were held with enough force to allow a 0.5X PBS wash to remove the bulk plasma, thereby purifying the collected particles (100 µL wash buffer at 10 µL/min, resulting in a 6.7X volume exchange).

After washing the chips, we used a standard immunostaining protocol to quantify biomarker content. The following incubation times are much longer than what is actually necessary for full immunostaining. We chose the longer incubation times just as a conservative measure to ensure full staining was achieved. These incubation times can be reduced to speed the process significantly. A blocking agent was first used to reduce nonspecific binding of the antibody stain for Glypican‐1. This consisted of 50 µl of 2% milk in PBS flowed into the chip at 10 µl/min and incubated for 60 min. The anti‐Glypican‐1 antibody (ab199343) was prepared by making a 1/200 dilution from the purchased stock (0.639 mg/ml) into the sample blocking solution of 2% milk in PBS. A 50 µl sample of this was then flowed onto the chip and incubated for 60 min. The chip was then washed with 100 µl of 0.5X PBS at 10 µl/min.

The fluorescently labeled secondary antibody (Alexa Fluor 594 goat Anti‐rabbit IgG (2 mg/ml)) was prepared by making a 1/500 dilution from stock into the 2% milk in PBS blocking solution. A 50 µl sample of this was then flowed onto the chip and incubated for 60 min. The unbound antibody was then washed off using 100 µl of 0.5X PBS at 10 µl/min. The collected cf‐DNA nanoparticles were then stained with SYBR Green dye. SYBR Green was prepared by making a 1/100 dilution from stock into Milli Q Water and then a 3/50 dilution into PBS. A 25 µl sample of this was flowed onto the chip at 10 µl/min and incubated for 10 min. The unbound SYBR Green dye was washed off the chip using 100 µl of 0.5X PBS at 10 µl/min. The chip was then imaged using a D2 Zeiss microscope (Jena, Germany).

4.4. DEP Collected Nanoparticle Size Distribution

The DEP chip was used to recover nanoparticles from 10 µL of undiluted PDAC patient plasma loaded into the chamber. To reduce nanoparticle contamination from buffer solutions, 0.5X PBS was prepared by mixing a 50:50 dilution of Dulbecco's Phosphate Buffered Saline 1X (no calcium, no magnesium, GibcoTM 14190144) with UltraPureTM Distilled Water (DNAse and RNAse Free, Invitrogen^TM 10977015) purchased from Thermo Fisher Scientific (Massachusetts, USA) and passed through a 0.1 µm filter before use.

The DEP chips were run in 3 replicates using a similar protocol as described above, but without the use of milk blocking buffer because it was unnecessary, since antibody staining was not used, and the milk could introduce additional nanoparticle contaminants. After DEP collection and washing, a 1:15 000 dilution of SYBR Gold Nucleic Acid Gel stain in the filtered 0.5X PBS was flowed into the channel and left to incubate in the dark for one hour, followed by another wash using 0.5X PBS. The chip was imaged, and the microfluidic tubing was removed to prevent backflow of washed material and dye from re‐entering the channels. Tape was applied over the channel openings to prevent sample evaporation. To push the collected nanoparticles from the chip surface and into the microfluidic channel buffer for elution and particle sizing analysis, the chip was placed into a thin‐walled plastic Petrie dish and placed on the water surface of a Fisher Scientific Ultrasonic Bath (5.7L) and sonicated for 15 min at 30 °C. A 20 µL drop of water was placed underneath the chip to acoustically couple the chip to the Petrie dish by removing the air gap to make sure the chip was properly ensonified. Imaging was conducted again to ensure sonication caused the successful release of nanoparticles from the electrodes, followed by pipette removal of all fluid from the channels into DNA‐lo‐bind tubes. A 10 µL wash with 0.5x PBS was flushed in and out of the microfluidic chamber five times to further elute particles off of the surface of the hydrogel and then the wash was removed and added to the originally collected sample tube.

The samples were diluted 1:60 in 0.1 µm filtered milli‐Q water and injected into a ZetaView NTA system (Particle Metrix GmbH, Germany) for particle tracking size analysis using three fields of view for each sample. The ZetaView instrument was set to measure the total nanoparticle population through scatter measurements and also the subfraction of particles that contained DNA, which were stained with the SYBR Gold Nucleic Acid Gel stain through fluorescence mode measurements, both using the 488 nm laser. The system settings for Scatter Mode included 11 measurement positions, frame rate at 30, sensitivity at 82, shutter at 100, minimum brightness at 20, maximum brightness at 255, minimum area at 10, maximum area at 1000, and minimum trace length at 15. The system settings for Fluorescence Mode included 2 measurement positions, multiple acquisitions (number of experiments: 5), low bleach mode with a dose sub volume of 20 µL, frame rate at 30, sensitivity at 88, shutter at 100, minimum brightness at 30, maximum brightness at 255, minimum area at 10, maximum area at 1000, and minimum trace length at 30.

As a negative control, this DEP collection and sample analysis process was then conducted using only 0.5X PBS in another three replicates to determine the ambient nanoparticle concentration in the buffers and released from the chip itself.

The size distribution histograms (Figure 2A and Figure 3A) were plotted using GraphPad Prism Software.

4.5. Data Analysis and Statistics

To test the statistical significance of the differences in Glypican‐1 and cell‐free DNA nanoparticle levels between different disease states, Mann‐Whitney U tests were performed using R and graphed using Prism. Receiver operating characteristic curve analysis was performed to plot the ROC curves and calculate the area under the curve (AUC) to compare the performance of each biomarker and combination of biomarkers using 25 non‐cancer and 11 cancer subjects. We further assessed the sensitivity and specificity using an optimal cut‐off point based on Youden's index that minimized the vertical distance from the diagonal line.

Machine learning models were developed and executed in MATLAB version 23.2.0.2436196 (R2023b). Several different machine learning models were applied to the data set, including k‐nearest neighbors, decision trees, Naive‐Bayes, and support vector machine learning models (SVM). SVM was found to do the best job at creating smooth boundaries and also achieved the best classification. An SVM model was fit to the data using a second‐order polynomial kernel (code included in Supplementary Section S2). Receiver operating characteristic curves were generated using the MATLAB perfcurve function. Two SVM analyses on the data were conducted. The first looked at the entire data set and defined the delineation between cancer and non‐cancer conditions, which is shown as the line between the Cancer and High‐Risk regions of the graph in Figure 4C. A second SVM analysis was conducted, using the non‐cancerous subset of patients, and defined a second delineation between non‐cancerous patients that clustered together with low biomarker levels from the rest of the non‐cancer patients that had higher biomarker levels. This is shown as the line delineating the High‐Risk from the Low‐Risk regions in Figure 4C.

To validate the SVM analysis, we performed a leave‐one‐out cross‐validation where one data point was removed, and the model was then trained on the remaining data and tested on this left‐out data point. The model prediction for this left‐out data point was then compared to the ground truth. We used a hard constraint defining the lower left region of the graph, with low levels of Glypican‐1 and cf‐DNA, as a non‐cancer region. Specifically, this was the triangular region defined by cf‐DNA levels below 100 a.u. and Glypican‐1 levels below 50 a.u. We found that the leave‐one‐out cross‐validation accuracy was 75% compared to 89% accuracy using the model trained on the full data set. It is expected that the leave‐one‐analysis would have fewer correct classifications (out‐of‐sample performance) than the full data set (in‐sample performance), and 75% correct with this small cohort shows the robustness of cf‐DNA and Glypican‐1 as biomarkers and that the SVM is an effective modeling tool for this cohort

4.6. Oxford Nanopore Technologies MinION Sequencing of cf‐DNA

Samples containing ∼1 ng total nucleic acid following elution from the DEP chip were prepared for sequencing on the Oxford Nanopore Technologies (ONT) MinION using a standard 2D (R9.4) flow cell. For comparison, QIAamp Circulating Nucleic Acid Kits (Cat # 55114 Qiagen, Germantown, MD) were used to recover cf‐DNA from PDAC patient plasma using the protocol outlined by the manufacturer. All double stranded DNA (dsDNA) in the sample was prepared using the ONT direct cDNA sequencing kit (cat# SQK‐DCS109), starting at the end prep step. 10 µl of standard Lambda DNA supplied by ONT was prepared using their rapid DNA sequencing kit (cat# SQK‐RAD004) as recommended, and 1 µl was spiked into prepared experimental samples to act as a carrier during sequencing runs. Sequencing was performed for 16 h at −120v under default MinION running conditions, using a new flow cell for each sample. Here, only linear (fragmented) dsDNA was analyzed by ONT sequencing; circular dsDNAs, single stranded DNAs (ssDNAs), and RNA did not contain sequencing adapters under the conditions used for library preparation.

4.7. Analysis of Oxford Nanopore Technologies Minion Data for cf‐DNA Fragment Length Quantification

Bases were called with Guppy version 3.6.1 [73]. Reads mapping to Escherichia coli, Enterobacteria phage λ or ribosomal RNA of any species were filtered from fastq files. Adapter content was assessed by BLAST version 2.9.0 [74], and any hits found in the 5’ or 3’ ends (150 nt) of the reads were trimmed off. Filtered reads were then mapped to the human genome (hg19), only considering the regular chromosomes 1–22, and X using minimap2 version 2.17 [75]. Secondary, supplementary mappings and reads with a MAPQ <10 were filtered out using GenomicAlignments package. Histograms of fragment query sequence lengths (qwidth) were made for bins of 10 and 100 bases with ggplot2 [76].

4.8. Qubit dsDNA Assay for Determining Cell Free DNA Concentrations

Cell‐free DNA concentrations of the EV purified samples were determined using a Qubit dsDNA HS Assay Kit from Life Technologies (Carlsbad, CA, Cat. #Q32851) following the manufacturer's protocol. 190 µL of working solution was prepared in a 500 µL Qubit assay tube (Life Technologies, Cat. #Q32856), and 10 µL of extracellular vesicle sample was loaded into each tube and incubated at room temperature for 2 min before reading.

Conflicts of Interest

M. Heller and S. Lippman were on the scientific advisory board of Biological Dynamics. S. Lippman is now advising Xzom.

Supporting information

Supporting File: smll72359‐sup‐0001‐SuppMat.docx.

Data Availability Statement

The data that support the findings of this study are available in the supplementary material of this article.