Improved diagnostics for pancreatic ductal adenocarcinoma (PDAC) to detect the disease at earlier, curative stages and to guide treatments is crucial to progress against this disease. The development of a liquid biopsy for PDAC has proven challenging due to the sparsity and variable phenotypic expression of circulating biomarkers. Here we report methods we developed for isolating specific subsets of extracellular vesicles (EV) from plasma using a novel magnetic nanopore capture technique. In addition, we present a workflow for identifying EV miRNA biomarkers using RNA sequencing and machine-learning algorithms, which we used in combination to classify distinct cancer states. Applying this approach to a mouse model of PDAC, we identified a biomarker panel of 11 EV miRNAs that could distinguish mice with PDAC from either healthy mice or those with precancerous lesions in a training set of n = 27 mice and a user-blinded validation set of n = 57 mice (88% accuracy in a three-way classification). These results provide strong proof-of-concept support for the feasibility of using EV miRNA profiling and machine learning for liquid biopsy.

Significance: These findings present a panel of extracellular vesicle miRNA blood-based biomarkers that can detect pancreatic cancer at a precancerous stage in a transgenic mouse model. Cancer Res; 78(13); 3688–97. ©2018 AACR.

Pancreatic cancer is the third leading cause of cancer-related death in the United States with a median overall survival of less than one year and five-year survival of only 7.7% (1, 2). At the time of diagnosis, most patients with pancreatic cancer have metastatic disease and are at a disease stage that is no longer curable (3, 4). In current clinical practice, CT and MRI can be used to identify a tumor; however, these imaging technologies can only detect a tumor after it has grown to a visible mass late in the disease cycle and have a poor ability to detect small metastases (5, 6). The development of a diagnostic that can detect pancreatic cancer at an earlier, curative stage has been a topic of great interest (4, 7, 8), but one that has faced significant challenges (9).

Circulating extracellular vesicles (EV), including exosomes, have recently gained attention as a potential biomarker to sensitively diagnose disease using easily accessible bodily fluids (plasma, urine, saliva; refs. 10–12). EVs are circulating vesicles (30 nm–1 μm), which package molecular information (mRNA, miRNA, DNA, and protein) from their mother cells, and can be used to measure the molecular state of difficult to access tumor cells (Fig. 1A; ref. 8). EVs have three key advantages for medical diagnostics relative to circulating molecules in the blood (e.g., proteins, cell-free DNA; refs. 13, 14): (i) EVs package RNA that would otherwise be degraded in the bloodstream; (ii) the multiple RNA and proteins packaged within EVs can be used to perform multiparameter measurements that allow specific disease states to be classified; (iii) proteins on an EV's surface can be used to sort targeted EVs based on the cells of origin, allowing enhanced specificity in the measurement of EV cargo compared with molecular markers in the blood. EVs also have key advantages compared with circulating tumor cells (CTC). Critically, EVs are present at higher concentrations (108–1010 particles/mL) than CTCs (1–100 cells/mL; refs. 15–17), thus avoiding the counting error and need to process very large volumes of blood (V > 10 mL), which have limited CTC-based diagnostics for pancreatic cancer. The primary limitation of exosomes has been that, due to their nanoscale size, it has been challenging to apply conventional microfluidic technology to precisely sort and detect them, as has been done with great success for CTCs (18, 19). When microfluidics are scaled to the size of exosomes (100× smaller in dimension than CTCs), the throughput of nanoscale devices, which scales with the channel's cross-sectional area, drops by 10,000× and becomes too slow for practical use. To overcome this problem, we have recently developed track etched magnetic nanopore (TENPO) sorting, which overcomes this problem by massive parallelization of nanofluidic immunomagnetic traps (Fig. 1B; refs. 11, 16, 20). The massive parallelization of the TENPO enables the high precision of nanofluidic sorting to be applied at large volumetric flow rates (φ > 10 mL/hour) and to robustly process crude samples by automatically diverting flow to neighboring pores if any one pore became clogged.

Figure 1.

The development of an EV miRNA biomarker panel for machine learning–based diagnosis of pancreatic cancer. A, EVs are shed from the pancreas and have surface markers that can be used to isolate them based on their cells of origin from the vast background of material present in plasma. Nucleic acid cargo packaged within the EVs can report the state of those cells of origin. B, These EVs are labeled based on their surface markers with magnetic nanoparticles (MNP) and then trapped using our TENPO device. Because of its nanoscale feature sizes, TENPO can isolate EVs that have been labeled with a sufficient number of magnetic nanoparticles and discard the vast amount of background material that might have small amounts of nonspecific labeling. C, The EV miRNA discovery pipeline consists of: (i) KPC and KPCY mice, genetically engineered mouse models for pancreatic cancer that provides precancerous (PanIN) and advanced stage (PDAC); (ii) TENPO isolation, a specific magnetic exosome isolation based on the expressions of surface proteins; (iii) miRNA sequencing to identify differentially expressed EV miRNA markers in PDAC and PanIN mice compared with healthy mice. D, To develop a machine learning–based diagnostic, we (i) created a training set with samples of known labels and multiple miRNA measurements per sample; (ii) developed a machine learning algorithm using LDA, which can find a linear combination of features that maximally separate different groups; and (iii) evaluated the diagnostic value of machine learning–detected miRNA signatures using an independent test set with blinded samples.

Figure 1.

The development of an EV miRNA biomarker panel for machine learning–based diagnosis of pancreatic cancer. A, EVs are shed from the pancreas and have surface markers that can be used to isolate them based on their cells of origin from the vast background of material present in plasma. Nucleic acid cargo packaged within the EVs can report the state of those cells of origin. B, These EVs are labeled based on their surface markers with magnetic nanoparticles (MNP) and then trapped using our TENPO device. Because of its nanoscale feature sizes, TENPO can isolate EVs that have been labeled with a sufficient number of magnetic nanoparticles and discard the vast amount of background material that might have small amounts of nonspecific labeling. C, The EV miRNA discovery pipeline consists of: (i) KPC and KPCY mice, genetically engineered mouse models for pancreatic cancer that provides precancerous (PanIN) and advanced stage (PDAC); (ii) TENPO isolation, a specific magnetic exosome isolation based on the expressions of surface proteins; (iii) miRNA sequencing to identify differentially expressed EV miRNA markers in PDAC and PanIN mice compared with healthy mice. D, To develop a machine learning–based diagnostic, we (i) created a training set with samples of known labels and multiple miRNA measurements per sample; (ii) developed a machine learning algorithm using LDA, which can find a linear combination of features that maximally separate different groups; and (iii) evaluated the diagnostic value of machine learning–detected miRNA signatures using an independent test set with blinded samples.

Close modal

In this article, we use the TENPO device to develop an EV miRNA biomarker panel that can diagnose specific states of developing cancer. Given the challenge of validating early diagnostics in patient samples, we used the KPC (Kras, p53, Pdx1-Cre) and KPCY (Kras, p53, Pdx1-Cre, yfp) murine pancreatic cancer model to perform preclinical validation (21). Using KPC and KPCY mice, we can generate mice that have precancerous lesions (pancreatic intraepithelial neoplasia; PanIN), mice that have cancer (pancreatic ductal adenocarcinoma; PDAC), and healthy control mice. The three main outcomes of this study were: (i) we identified panels of EV miRNA biomarkers based on unbiased RNA-sequencing that can distinguish healthy, PanIN, and PDAC mice (Fig. 1C); (ii) we used this panel of EV biomarkers in a machine-learning algorithm to predict which mice are healthy, PanIN, or PDAC, and evaluated this diagnostic in prospectively collected user-blinded samples (Fig. 1D); (iii) we determined the key signaling pathways activated in the PDAC and PanIN mice relative to the healthy mice that could be detected in the EVs, to link our biomarkers to the underlying mechanism of the disease states.

TENPO fabrication

The TENPO membrane fabrication is described fully in our previous publication (11). Briefly, 200 nm NiFe and 30 nm Au (Kurt Lesker) were deposited sequentially using thermal evaporation (Kurt Lesker) onto track-etched polycarbonate membranes (Whatman). The membranes were assembled into a device using laser micromachined polyester films and double-sided tape (McMaster-Carr; refs. 11, 20).

Cell culture

Mouse cell lines PD7591 and PD483 were generated from pancreatic tumor tissue isolated from Pdx1-cre, KrasLSL-G12D, p53L/+, RosaYFP/YFP (KPCY) mice in the B. Stanger's laboratory (3). The cell lines were cultured in media as described previously (22). The cell lines were periodically checked for mycoplasma contamination using the MycoAlert Plus Detection kit (Lonza). Supernatant fractions from confluent cell cultures (48–72 hours) were collected and centrifuged at 1,500 × g for 5 minutes to remove dead cells and debris. Cells were maintained at passage <30. Conditioned media was then transferred to new 50-mL centrifuge tubes and either processed fresh or stored immediately at −80°C for future use.

Mouse sample processing

All mouse work was approved by Institutional Animal Care and Use Committee. Tumor-bearing KPC and KPCY mice were identified by abdominal palpation and presence of tumor was confirmed at necropsy. Six- to 8-week-old KPC and KPCY mice were used as PanIN cohort and presence of PanIN lesions was confirmed by histology. Pdx1-Cre mice were used as a healthy cohort. Mice included in the analysis were genetically identical and were all housed in the same facility. Blood was collected from all mice by cardiac puncture and collected in sodium citrate BD Vacutainer blood collection tubes (BD Biosciences). Blood was centrifuged at 1,600 × g for 10 minutes and plasma was removed, followed by a second spin at 3,000 × g for 10 minutes to minimize cellular contamination.

Exosome isolation (TENPO)

Exosomes were isolated from cell cultured media or mouse plasma as described in detail in the sections below. Anti-biotin ultrapure microbeads (Miltenyi Biotec) and biotinylated antibodies were used for magnetic labeling. Antibodies used in this study include biotin anti-mouse CD9 antibody (BioLegend), biotin anti-mouse CD81 antibody (BioLegend), biotin anti-mouse CD326 (Ep-CAM) antibody (BioLegend), biotin anti-mouse CD104 antibody (BioLegend), TSPAN8 mAb (Thermo Fisher Scientific), anti-mouse c-Met purified antibody (eBioscience), and rabbit anti-CD44v6 antibody (United States Biological). For biotinylation, we used One-Step Biotinylation Kit (Miltenyi Biotec) and Antibody Purification Kit (Abcam). Biotinylated antibodies were added to the sample (cell cultured media or mouse plasma) and incubated for 20 minutes at room temperature with shaking. Subsequently, anti-biotin magnetic nanoparticles were added to the samples and incubated for 20 minutes at room temperature with shaking. Then, the samples were added to the reservoir of the TENPO chip and negative pressure was applied by a programmable syringe pump (Braintree). As the samples were pulled through the chip, exosomes that were labeled with a sufficient number of magnetic nanoparticles were captured at the edge of the pores of the chip.

EV miRNA isolation

The ExoRNeasy Serum/Plasma Kit (Qiagen) was used to extract RNA from isolated EVs directly on the TENPO chip. We directly applied QIAzol lysis reagent (Qiagen) to the EVs captured on the TENPO chip. We collected the lysate and then used the second component of the ExoRNeasy serum/plasma kit to extract the RNA. The EV miRNA was eluted in a small volume (∼30 μL) and it was stored at −80°C or processed immediately for further analysis.

RNA sequencing

We measured the extracted RNA quantity using Qubit (Life Technologies) and, as recommended by the protocol, the samples with more than 100 ng of RNA were selected for use. NEBNext Small RNA Library Prep Set for Illumina (BioLabs) was used to make a library. Quality control check was performed on a BioAnalyzer using a DNA 1000 chip. For size selection, AMPure XP beads were used (Beckman Coulter) and 140-150 bp sizes were selected. The size distribution of the isolated RNA was confirmed by BioAnalyzer using the High Sensitivity Chip (Agilent). The cDNA libraries were pooled together, and the final concentration was quantified using a KAPA Library Quantification Kit (KAPA Biosystems). The libraries were sequenced using a NextSeq 500/550 kit (FC-404-2005, Illumina) on a NextSeq500 (75 bp length). miRNA expression was found with mirDeep2 (23) and Bowtie (mm10; ref. 24), using miRBase version 21 (25). Read counts from miRNA families (miRNAs with the same seed sequence) were combined. Quantified miRNA expression values were normalized by DESeq2 (26). Using DIANA mirPath v. 2.0 (27), we identified the pathways of target genes regulated by miRNAs.

qPCR

The miScript SYBR Green PCR Kit (Qiagen) and miScript primers (Qiagen) were used. A master mix that consists of miScript SYBR Green, miScript primer, universal primer, and water was made at 5:1:1:2 ratio and 9 μL of the master mix was added to each well, followed by 1 μL of cDNA. Forty cycles were run with a default setting using CFX384 Touch Real-Time PCR machine (Bio-Rad). Triplicates were performed for each sample and negative template control (NTC) was used to check any contamination. The melting curves for the amplified DNA were each manually checked before subsequent analysis.

EV miRNA biomarker sequencing

We chose to profile EV miRNAs because miRNA is known to play a functional role in oncogenesis and tissue differentiation by regulating gene expression through specific mRNA binding (28, 29). The miRNA expression patterns have been shown to be highly tissue-specific, providing great potential as a tumor-specific biomarker. In addition, miRNA dysregulation has been reported as a feature of pancreatic cancer progression, and the transfer of EV miRNAs has been shown to cause biological changes in recipient cells (30, 31). To profile the miRNA cargo in EVs at various stages of pancreatic cancer, we isolated EVs using pan-exosome markers (CD9, CD81) from 500 μL of mouse plasma using the KPCY model (n = 2 healthy, n = 6 PanIN, and n = 5 PDAC; ref. 4). For each mouse, 584 mature miRNAs were detected, and of these 584, there were 30 mature miRNAs that were found to be significantly differentially expressed (ANOVA P < 0.01) between the healthy control, PanIN, and PDAC mice. The raw sequencing data are deposited at GEO (GSE111750; Fig. 2A).

Figure 2.

miRNA sequencing to map the RNA cargo contained within EVs. A, Raw miRNA sequencing data from n = 2 healthy, n = 6 PanIN, and n = 5 PDAC mice. A total of 584 mature miRNAs were expressed, and miRNAs with a total depth > 100 were plotted as a heatmap. B, Fold changes of the expressed miRNAs are plotted as a waterfall plot for comparing PDAC versus Healthy, and the corresponding fold changes for each miRNA for PanIN versus Healthy, and PanIN versus PDAC. C, The relationship between Healthy, PanIN, and PDAC was compared by calculating the Pearson correlation coefficient (R) between the total EV miRNA expression. D, We selected 11 biomarkers that were either maximally upregulated or downregulated in PanIN or PDAC mice relative to healthy controls. E, We validated these 11 biomarkers by measuring them in n = 12 healthy, n = 14 PanIN, and n = 16 PDAC mice and comparing results with RNA sequencing. miR-1947 was considered to be an outlier and was excluded from the analysis.

Figure 2.

miRNA sequencing to map the RNA cargo contained within EVs. A, Raw miRNA sequencing data from n = 2 healthy, n = 6 PanIN, and n = 5 PDAC mice. A total of 584 mature miRNAs were expressed, and miRNAs with a total depth > 100 were plotted as a heatmap. B, Fold changes of the expressed miRNAs are plotted as a waterfall plot for comparing PDAC versus Healthy, and the corresponding fold changes for each miRNA for PanIN versus Healthy, and PanIN versus PDAC. C, The relationship between Healthy, PanIN, and PDAC was compared by calculating the Pearson correlation coefficient (R) between the total EV miRNA expression. D, We selected 11 biomarkers that were either maximally upregulated or downregulated in PanIN or PDAC mice relative to healthy controls. E, We validated these 11 biomarkers by measuring them in n = 12 healthy, n = 14 PanIN, and n = 16 PDAC mice and comparing results with RNA sequencing. miR-1947 was considered to be an outlier and was excluded from the analysis.

Close modal

We performed a pairwise comparison for three groups using all detected miRNAs to identify significantly differentially expressed miRNAs (Fig. 2B). From the pairwise comparison between PDAC and healthy mice, we found that 9 miRNAs were significantly differentially expressed. Of these 9 miRNAs, seven are conserved in human. When we performed a pairwise comparison between PanIN and healthy mice, we found that 7 miRNAs were significantly differentially expressed. Of these 7 miRNAs, 3 were conserved in human. When we performed a pairwise comparison between PanIN and PDAC mice, we found that 14 miRNAs were significantly differentially expressed. Of these 14 miRNAs, 12 were conserved in human (Fig. 2B; Supplementary Fig. S1). We calculated the pairwise Pearson correlation coefficients for the EV miRNA expression between the PDAC, PanIN, and healthy mice (Fig. 2C). We found that the healthy and PanIN mice were the most correlated to one another, whereas healthy and PDAC were the least correlated. To select miRNA biomarkers to carry forward for use in our machine learning–based diagnostic, we performed a pairwise comparison to identify the most upregulated and downregulated miRNAs (Fig. 2D). From this list, we chose 11 miRNA markers that were either highly differentially expressed (> 1.5× fold change) or significantly different (Padj < 0.05) or both from a pairwise comparison and that were abundant enough (average raw counts > 50) to be detected using qPCR (Supplementary Table S1). We validated these biomarkers by comparing their measured expression level using qPCR to sequencing, and found a significant correlation (Fig. 2E). In this experiment, we measured the average expression levels of 11 miRNAs from mice in each group (n = 12 healthy with R2 = 0.73, n = 14 PanIN with R2 = 0.59, and N = 16 PDAC with R2 = 0.34).

Enriched signaling pathways in pancreatic cancer EVs

We performed bioinformatics analysis on the full set of sequencing data to explore the signaling pathways that could be measured in the EV miRNA that were significantly enriched in the PanIN and PDAC mice versus healthy controls. We performed a Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway analysis using significantly different miRNAs with Padj < 0.05. For PanIN versus healthy, we found 26 significant pathways (Fig. 3A). For PDAC versus healthy, we found 27 significant pathways (FDR corrected P < 0.05; Fig. 3B). We found that there were 16 overlapping significant pathways in comparing PDAC versus healthy with PanIN versus healthy. For the 11 miRNA markers that we have identified from the RNA-sequencing data, we found 39 significant pathways (FDR-corrected P < 0.05; Supplementary Fig. S2). Several of these identified pathways were found in previous studies in the literature including prion diseases, axon guidance, and core pancreatic cancer pathways (24–27). These identified pathways connect our measured miRNA biomarker panels to known pathways underlying pancreatic cancer development, adding to our confidence in the translatability of these panels to clinical use and demonstrating its potential for monitoring known pathways in tumors for applications such as drug efficacy monitoring. The related miRNAs and their target genes for a specific pathway can be searched using the DIANA mirPath v3 tool and the significantly differentially expressed miRNAs as an input (Supplementary Fig. S1).

Figure 3.

Significantly enriched signaling pathways in EVs. A, Twenty-seven pathways including core pancreatic cancer pathways were significantly enriched for PanIN versus Healthy. B, Twenty-six pathways were significantly enriched for PDAC versus Healthy.

Figure 3.

Significantly enriched signaling pathways in EVs. A, Twenty-seven pathways including core pancreatic cancer pathways were significantly enriched for PanIN versus Healthy. B, Twenty-six pathways were significantly enriched for PDAC versus Healthy.

Close modal

Surface marker selection

To specifically enrich for pancreatic cancer cell–derived EVs, such that we could increase the performance of our machine learning–based diagnostic, we used TENPO to evaluate the efficacy of several different combinations of previously reported surface markers to isolate tumor-derived EVs from plasma (32). We compared pan-exosome surface markers (CD9, CD81), epithelial cell adhesion molecule (EpCAM) marker, and a cocktail of markers (CD104, Tspan8, MET, EpCAM, CD44v6) that have shown the highest coverage of exosomes from multiple pancreatic cancer cell lines while not being expressed by exosomes from healthy donor serum (32). To test the performance of these surface markers, we performed an experiment using cell culture derived EVs (PD483, PD7591) spiked into healthy mouse plasma. We spiked V = 10 mL of mouse pancreatic cancer cell line conditioned media into V = 0.5 mL of healthy mouse plasma to mimic cancer patient plasma (Fig. 4A). The volume ratio was calculated by comparing the concentration of EVs in cell line conditioned media (106–108 particles/mL) to the concentration of EVs in patient plasma (108–1010 particles/mL; refs. 17, 33). To quantify the specificity of TENPO's capture of tumor-derived EVs, relative to those present in healthy plasma, we measured the difference in expression level of our panel of miRNAs (ΔCt) between healthy plasma and plasma spiked with cell line conditioned media for each of the three potential sets of surface markers. The larger the difference between the RNA measured in the spiked and unspiked sample, the more specific the enrichment of pancreatic cancer cell–derived exosomes. As expected, EpCAM achieved a more specific capture of tumor-derived EVs from the spiked plasma samples compared with pan-exosome capture (CD9, CD81) because pancreatic tumor–derived EVs originate from epithelial cells, and pan-exosome markers will isolate background exosomes present in the healthy plasma. However, using only EpCAM, tumor-derived EVs that originate from tumor cells that are not epithelial, for example, cells that have undergone an epithelial-to-mesenchymal transition (EMT) could be missed (3). To address this challenge, we used a cocktail of markers that have been found to specifically enrich for pancreatic cancer exosomes (32). This cocktail achieved the most specific capture of tumor-derived EVs from plasma compared with EpCAM and pan-exosome surface markers. We characterized the morphology and the size of the material that we captured using cocktail markers using scanning electron microscopy (SEM) and dynamic light scattering (DLS), and the morphology and the size distribution (peak at 91.3 nm) matched those of EVs (Fig. 4B and C). Because we choose the surface markers after we perform the sequencing experiment, it is possible that even better RNA markers could be found by repeating the sequencing experiment using optimized surface markers in future work. To this end, the most specific surface marker (cocktail) found using our mouse model was applied to clinical samples, to demonstrate the potential of our platform for clinical translation. We used the same cocktail markers (CD104, CD44v6, Tspan8, MET, EpCAM) to enrich for pancreatic cancer cell–derived EVs and sequenced the miRNA packages in the EVs. We found 12 miRNAs that were significantly differentially expressed (Padj < 0.05) in pancreatic cancer patient samples compared with control samples (Supplementary Table S2), which did not overlap with the significant miRNA found in our murine model. We compared this differential expression with that shown in the literature for PDAC, and found agreement for hsa-miR-122-5p (34), hsa-miR-30a-5p (35), hsa-miR-29a-3p (36), hsa-miR-483-3p (37), hsa-miR-483-5p (38), hsa-miR-186-5p (39), hsa-miR-200a-3p (40), hsa-miR-141-3p (41), and hsa-miR-106b-3p (42). The fold changes of hsa-miR-19b-3p (43), hsa-miR-25-3p (44), and hsa-miR-451a (45) from the literatures did not match with what we found, but in the literature were shown to have diagnostic value for pancreatic cancer.

Figure 4.

Surface marker selection. A, We measured differential expressions (ΔCt) of miRNA markers between healthy mouse plasma and mouse pancreatic cancer cell line conditioned media using pan-exosome (CD9, CD81), EpCAM, and a cocktail of markers (CD44v6, Tspan8, CD104, Met, EpCAM) that specifically enrich for pancreatic cancer exosomes. Two different mouse pancreatic cancer cell lines (PD483, PD7591) were used. B, Using SEM, we imaged eluted exosomes that were captured on-chip using the cocktail of surface markers. The morphology and the size of the eluate match those of EVs. C, The particle size distributions of the eluate were measured using DLS. The eluate had a major peak at 91.3 nm, whereas the input pancreatic cancer cell line conditioned media had a major peak at 10.1 nm.

Figure 4.

Surface marker selection. A, We measured differential expressions (ΔCt) of miRNA markers between healthy mouse plasma and mouse pancreatic cancer cell line conditioned media using pan-exosome (CD9, CD81), EpCAM, and a cocktail of markers (CD44v6, Tspan8, CD104, Met, EpCAM) that specifically enrich for pancreatic cancer exosomes. Two different mouse pancreatic cancer cell lines (PD483, PD7591) were used. B, Using SEM, we imaged eluted exosomes that were captured on-chip using the cocktail of surface markers. The morphology and the size of the eluate match those of EVs. C, The particle size distributions of the eluate were measured using DLS. The eluate had a major peak at 91.3 nm, whereas the input pancreatic cancer cell line conditioned media had a major peak at 10.1 nm.

Close modal

Machine learning–based diagnostic

Using the TENPO platform, we first isolated EV RNA using our cocktail markers (CD104, CD44v6, Tspan8, MET, EpCAM) from n = 9 healthy, n = 9 PanIN, and n = 9 PDAC mouse plasma samples and performed qPCR for 11 miRNA markers selected from RNA sequencing (Fig. 5A). We observed that the miRNA profile was different on average between healthy, PanIN, and PDAC mice, but due to mouse-to-mouse variability, it was not possible to accurately classify every mouse correctly using any single EV RNA biomarker. The best performing miRNA was miR-3970-5p, and it achieved an accuracy of 65% (Supplementary Fig. S3). We note that several of the best performing miRNA markers also happen to be abundant in platelets (46), as well as in tumor cells. Platelets play a critical role in cancer progression and thrombosis (47) and enhance tissue factor protein and metastasis-initiating cell markers (48). In addition, it has been found that tumor-educated platelets have diagnostic value for different types of cancer (49). The data shown in Supplementary Fig. S3 was obtained using our TENPO device with the cocktail of markers chosen to specifically enrich for pancreas-derived EVs. In future work, there will be an opportunity to use the TENPO to additionally enrich for platelet-derived EVs and to use their miRNA cargo to enhance diagnostic power. Moreover, we note that in machine learning, the model can achieve specificity greater than that of its constituent features, as it is not the specificity of any single marker that matters but the specificity of the signature that can be found in their mathematical combination (50, 51). To evaluate our diagnostic before testing it on our blinded test set, we used N-1 leave-one-out cross validation (Fig. 5B). To classify individual mice using a combination of EV miRNA biomarkers, we used a linear discriminant analysis (LDA). LDA is a supervised learning algorithm that searches for a linear combination (x · w) of features that can classify each subject into the correct group. We used a training set (wT) of n = 27 mice to generate the LDA signature. Subsequently, we used our trained diagnostic to classify a blinded test set (wB) that consisted of N = 57 mice, which included n = 19 healthy controls, n = 19 PanIN mice, and n = 19 PDAC mice, to evaluate its performance (Fig. 5C). In a three-way classification (healthy, PanIN, PDAC) on the blinded test set, we achieved 88% accuracy (Fig. 5D). We performed a control experiment wherein we replaced our correctly labeled training set with a training set where the labels were randomly assigned. As expected, the accuracy of our three-way comparison dropped from 88% to 30%, which is equivalent to randomly choosing, validating the specific predictive signature of our training set (Fig. 5E). We additionally compared the performance of LDA to other machine-learning algorithms including Tree Bagger, Support Vector Machine (SVM), and k-Means Nearest Neighbor (KNN; Fig. 5F; ref. 50). We reason that LDA had the best performance because it has the fewest number of degrees of freedom, and thus should be less sensitive to overfitting than the other more sophisticated models. As this work progresses into clinical testing, larger sample sizes will allow for improved performance by using more advanced machine-learning algorithms.

Figure 5.

Machine learning–based diagnostic using multiple EV miRNAs. A, We created a training set with samples of known labels from healthy, PanIN, and PDAC groups. Eleven miRNAs were profiled using qPCR. B, The performance of the training set was evaluated using LDA N-1 leave-one-out cross validation. C, To prevent overfitting, an independent test set was generated using blinded samples. D, The LDA algorithm developed from the training set was applied to the independent blinded test set to accurately classify healthy, PanIN, and PDAC mice. From the three-way comparison, we achieved 88% accuracy. E, Different machine-learning algorithms (Tree Bagger, SVM, KNN) were used to compare their performances with LDA. F, To validate the specificity of miRNA signatures found in the training set, we created a scrambled training set that showed no predictive value (30%).

Figure 5.

Machine learning–based diagnostic using multiple EV miRNAs. A, We created a training set with samples of known labels from healthy, PanIN, and PDAC groups. Eleven miRNAs were profiled using qPCR. B, The performance of the training set was evaluated using LDA N-1 leave-one-out cross validation. C, To prevent overfitting, an independent test set was generated using blinded samples. D, The LDA algorithm developed from the training set was applied to the independent blinded test set to accurately classify healthy, PanIN, and PDAC mice. From the three-way comparison, we achieved 88% accuracy. E, Different machine-learning algorithms (Tree Bagger, SVM, KNN) were used to compare their performances with LDA. F, To validate the specificity of miRNA signatures found in the training set, we created a scrambled training set that showed no predictive value (30%).

Close modal

In this work, we use an interdisciplinary approach to diagnose pancreatic cancer by combining nanofluidic-based EV sorting, an open-ended biomarker discovery using sequencing, and machine learning–detected pancreatic cancer signatures. The workflow of our platform is summarized in Supplementary Table S3. The highly specific and rapid EV isolation technology and mapping of whole EV miRNAs derived from healthy, PanIN, and PDAC mice bring novelty to our work compared with a previous study that used bead-based capture and microarray for the development of a pancreatic cancer diagnostic (32). Specifically, our TENPO isolation enables EVs to be isolated on the basis of having a threshold of surface markers, akin to flow cytometry, whereas in microbead-based assays, a single binding event leads to the capture of the EV (11). The TENPO builds on previous work using microfluidic immunomagnetic sorting to isolate rare cells (18, 19, 52) and translates this approach to nanoscale EVs. Because of its massive parallelization, the TENPO precisely sorts nanoscale EVs at flow rates (10 mL/hour) sufficient for clinical use and can analyze raw serum or plasma without the risk of clogging. Because of the nanoscale feature sizes of the TENPO's immunomagnetic traps, each EV is sorted individually based on its labeling by a sufficient number of MNPs, analogous to the microfluidic devices used to selectively sort rare circulating tumor cells. Previous microfluidic technologies and microbead systems capture EVs onto functionalized surfaces (53) or microbeads (54), wherein the capture purity is set solely by the affinity ligand specificity. By combining the isolation of specific EV subsets, RNA analysis, and machine learning, we performed a three-way classification between healthy mice, mice with precancerous lesions (PanIN), and mice with PDAC and demonstrated an accuracy of 88%, significantly outperforming control experiments that resulted in an accuracy (30%) insignificantly different to random guessing. In a previous study, a separate evaluation of exosome protein and miRNA was performed and then combined computationally, resulting in a sensitivity of 1.00 and a specificity of 0.80 discriminating patients with pancreatic cancer from healthy and disease controls and a sensitivity of 1.00 and a specificity of 0.93 when controls with nonpancreatic malignancies were excluded (32).

Here, we performed pan-exosome–based isolation to identify differentially expressed exosomal miRNAs and then used these miRNAs to compare different surface markers to find markers that could increase specificity of the found miRNAs. We found that cocktail markers (CD104, CD44v6, Tspan8, MET, EpCAM) exhibited improved specificity compared with EpCAM-based isolation or pan-exosome–based isolation. In future work, RNA sequencing can be performed on cocktail-isolated EVs to further refine the biomarker selection.

In follow-up studies, our approach to identify EV miRNA signatures can be used to screen at risk groups for pancreatic cancer, including patients with pancreatitis and pancreatic intraductal papillary mucinous neoplasms, to identify a subset of patients who will develop pancreatic cancer. By using a supervised, statistical learning approach to experimentally identify biomarkers for disease states, our approach is not limited to markers for which there is currently a clear understanding of its underlying mechanism or role in the disease. This aspect is a strength, as it allows for well-performing biomarkers to be discovered that otherwise would not, if only biomarkers with a clear underlying mechanism were considered. This aspect is also a weakness, as it reduces confidence in the translatability of these findings. As such, it is important that care be taken when applying this method to retrain the algorithm for the specific intended application, using an appropriate training set, with well-defined controls and blinded, independent evaluation (50). Beyond what has been demonstrated in this article, our workflow can be utilized to identify other clinical subgroups for which accurate classification could be beneficial. For example, classification of surgically resectable, locally advanced, and metastatic patients can guide clinical decisions for surgery and therapy. Moreover, the machine learning results demonstrated in this article can be expected to improve as this work is translated to studies with large cohorts of clinical samples (50). Larger sample sizes will allow for improved performance by enabling the use of more advanced machine learning algorithms (e.g., neural network, random forest). While in this study, we only trained our system to classify between healthy, PanIN, and PDAC, in future work it is possible to train the system to also distinguish between other types of cancer. The existence of several miRNAs known to be involved in metastasis highlight the need to train and evaluate this system against other cancers (55–57). Our ability to use these markers to distinguish PanIN versus healthy and PanIN versus PDAC give confidence that our machine-learning analysis has specificity beyond simply detecting metastasis. In addition to the classification algorithm we used in this study with class labels, we can also adapt this method to predict continuous values such as tumor volume, using regression algorithms to predict biologically interesting measurements.

Building on the EV profile measured in this work, a more comprehensive view of a developing cancer can be captured by using TENPO to isolate multiple subsets of EVs based on their surface markers. For instance, it has been found that tumor-educated platelets (TEP) play a central role in tumor growth, and their mRNA profiles are distinct from platelets of healthy individuals (49). EVs derived from TEPs, in addition to those derived from cancer cells, can be profiled to better understand their interaction with cancer cells. Moreover, this work can be expanded beyond pancreatic cancer to diagnose a variety of cancers and other diseases, by isolating EVs with the appropriate surface markers (58, 59).

D. Issadore has ownership interest (including patents) in Chip Diagnostics. No potential conflicts of interest were disclosed by the other authors.

Conception and design:J. Ko, E.L. Carpenter, B.Z. Stanger, D. Issadore

Development of methodology:J. Ko, E.L. Carpenter, B.Z. Stanger, D. Issadore

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.):J. Ko, N. Bhagwat, T. Black, S.S. Yee, E.L. Carpenter, B.Z. Stanger, D. Issadore

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis):J. Ko, Y.-J. Na, J. Kim, E.L. Carpenter, B.Z. Stanger, D. Issadore

Writing, review, and/or revision of the manuscript:J. Ko, T. Black, S.S. Yee, J. Kim, E.L. Carpenter, B.Z. Stanger, D. Issadore

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases):T. Black, S.S. Yee, S.A. Fisher, D. Issadore

Study supervision:E.L. Carpenter, B.Z. Stanger, D. Issadore

This work was supported by the Pennsylvania Department of Health Commonwealth Universal Research Enhancement Program, the NIH grants (1R21CA182336-01A1 to D. Issadore and J. Kim; R01-CA169123 to B.Z. Stanger; to N.M.A.; F32CA196120, to N. Bhagwat; and R01CA207643), Pancreatic Cancer Action Network Translational Research Award, Abramson Cancer Center Pancreatic Translational Center of Excellence (to E.L. Carpenter), and in part by the Penn Center for Molecular Studies in Digestive and Liver Diseases (P30-DK050306) from the National Institute of Diabetes and Digestive and Kidney Diseases. D. Issadore was supported by an American Cancer Society - CEOs Against Cancer - CA Division Research Scholar Grant, (RSG-15-227-01-CSM). B.Z. Stanger was supported by USAMRMC Award W81XWH-15-1-0457 from the US Department of Defense.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.

1.
Siegel
RL
,
Miller
KD
,
Jemal
A
. 
Cancer statistics, 2017
.
CA Cancer J Clin
2017
;
67
:
7
30
.
2.
Yang
S
,
Che
SPY
,
Kurywchak
P
,
Tavormina
JL
,
Gansmo
LB
,
Correa de Sampaio
P
, et al
Detection of mutant KRAS and TP53 DNA in circulating exosomes from healthy individuals and patients with pancreatic cancer
.
Cancer Biol Ther
2017
;
18
:
158
65
.
3.
Rhim
AD
,
Mirek
ET
,
Aiello
NM
,
Maitra
A
,
Bailey
JM
,
McAllister
F
, et al
EMT and dissemination precede pancreatic tumor formation
.
Cell
2012
;
148
:
349
61
.
4.
Adamczyk
KA
,
Klein-Scory
S
,
Tehrani
MM
,
Warnken
U
,
Schmiegel
W
,
Schnölzer
M
, et al
Characterization of soluble and exosomal forms of the EGFR released from pancreatic cancer cells
.
Life Sci
2011
;
89
:
304
12
.
5.
Horton
KM
,
Fishman
EK
. 
Adenocarcinoma of the pancreas: CT imaging
.
Radiol Clin N Am
2002
;
40
:
1263
72
.
6.
Heiken
JP.
Carcinoma of the pancreas: detection and staging using CT and MRI
.
Cancer Imaging
2015
;
2
:
19
.
7.
Taylor
DD
,
Gercel-Taylor
C
. 
MicroRNA signatures of tumor-derived exosomes as diagnostic biomarkers of ovarian cancer
.
Gynecol Oncol
2008
;
110
:
13
21
.
8.
Ko
J
,
Carpenter
E
,
Issadore
D
. 
Detection and isolation of circulating exosomes and microvesicles for cancer monitoring and diagnostics using micro/nano-based devices
.
Analyst
2016
;
141
:
450
60
.
9.
Kaur
S
,
Baine
MJ
,
Jain
M
,
Sasson
AR
,
Batra
SK
. 
Early diagnosis of pancreatic cancer: challenges and new developments
.
Biomarkers Med
2012
;
6
:
597
612
.
10.
Melo
SA
,
Luecke
LB
,
Kahlert
C
,
Fernandez
AF
,
Gammon
ST
,
Kaye
J
, et al
Glypican-1 identifies cancer exosomes and detects early pancreatic cancer
.
Nature
2015
;
523
:
177
82
.
11.
Ko
J
,
Bhagwat
N
,
Yee
SS
,
Ortiz
N
,
Sahmoud
A
,
Black
T
, et al
Combining machine learning and nanofluidic technology to diagnose pancreatic cancer using exosomes
.
ACS Nano
2017
;
11
:
11182
93
.
12.
Shao
H
,
Chung
J
,
Balaj
L
,
Charest
A
,
Bigner
DD
,
Carter
BS
, et al
Protein typing of circulating microvesicles allows real-time monitoring of glioblastoma therapy
.
Nat Med
2012
;
18
:
1835
40
.
13.
Anker
P
,
Mulcahy
H
,
Chen
XQ
,
Stroun
M
. 
Detection of circulating tumour DNA in the blood (plasma/serum) of cancer patients
.
Cancer Metastasis Rev
1999
;
18
:
65
73
.
14.
Kim
J
,
Bamlet
WR
,
Oberg
AL
,
Chaffee
KG
,
Donahue
G
,
Cao
X-J
, et al
Detection of early pancreatic ductal adenocarcinoma with thrombospondin-2 and CA19-9 blood markers
.
Sci Translat Med
2017
;
9
:
eaah5583
.
15.
Kurihara
T
,
Itoi
T
,
Sofuni
A
,
Itokawa
F
,
Tsuchiya
T
,
Tsuji
S
, et al
Detection of circulating tumor cells in patients with pancreatic cancer: a preliminary result
.
J Hepatobiliary Pancreat Surg
2008
;
15
:
189
95
.
16.
Ko
J
,
Bhagwat
N
,
Yee
SS
,
Black
T
,
Redlinger
C
,
Romeo
J
, et al
A magnetic micropore chip for rapid (<1 hour) unbiased circulating tumor cell isolation and in situ RNA analysis
.
Lab Chip
2017
;
17
:
3086
96
.
17.
Coumans
FAW
,
van der
PE
,
Boing
AN
,
Hajji
N
,
Sturk
G
,
van Leeuwen
TG
, et al
Reproducible extracellular vesicle size and concentration determination with tunable resistive pulse sensing
.
J Extracell Vesicles
2014
;
3
:
25922
.
18.
Nagrath
S
,
Sequist
LV
,
Maheswaran
S
,
Bell
DW
,
Irimia
D
,
Ulkus
L
, et al
Isolation of rare circulating tumour cells in cancer patients by microchip technology
.
Nature
2007
;
450
:
1235
9
.
19.
Issadore
D
,
Chung
J
,
Shao
H
,
Liong
M
,
Ghazani
AA
,
Castro
CM
, et al
Ultrasensitive clinical enumeration of rare cells ex vivo using a micro-hall detector
.
Sci Translat Med
2012
;
4
:
141ra92
.
20.
Muluneh
M
,
Shang
W
,
Issadore
D
. 
Track-etched magnetic micropores for immunomagnetic isolation of pathogens
.
Adv Healthcare Mat
2014
;
3
:
1078
85
.
21.
Hingorani
SR
,
Wang
L
,
Multani
AS
,
Combs
C
,
Deramaudt
TB
,
Hruban
RH
, et al
Trp53R172H and KrasG12D cooperate to promote chromosomal instability and widely metastatic pancreatic ductal adenocarcinoma in mice
.
Cancer Cell
2005
;
7
:
469
83
.
22.
Schreiber
FS
,
Deramaudt
TB
,
Brunner
TB
,
Boretti
MI
,
Gooch
KJ
,
Stoffers
DA
, et al
Successful growth and characterization of mouse pancreatic ductal cells: functional properties of the Ki-RASG12V oncogene
.
Gastroenterology
2004
;
127
:
250
60
.
23.
Li
C
,
Yu
S
,
Nakamura
F
,
Yin
S
,
Xu
J
,
Petrolla
AA
, et al
Binding of pro-prion to filamin A disrupts cytoskeleton and correlates with poor prognosis in pancreatic cancer
.
J Clin Invest
2009
;
119
:
2725
.
24.
Javle
M
,
Li
Y
,
Tan
D
,
Dong
X
,
Chang
P
,
Kar
S
, et al
Biomarkers of TGF-beta signaling pathway and prognosis of pancreatic cancer
.
PLoS One
2014
;
9
:
e85942
.
25.
Karanikas
M
,
Esempidis
A
,
Chasan
ZTM
,
Deftereou
T
,
Antonopoulou
M
,
Bozali
F
, et al
Pancreatic cancer from molecular pathways to treatment opinion
.
J Cancer
2016
;
7
:
1328
.
26.
Biankin
AV
,
Waddell
N
,
Kassahn
KS
,
Gingras
MC
,
Muthuswamy
LB
,
Johns
AL
, et al
Pancreatic cancer genomes reveal aberrations in axon guidance pathway genes
.
Nature
2012
;
491
:
399
405
.
27.
Jones
S
,
Zhang
X
,
Parsons
DW
,
Lin
JC
,
Leary
RJ
,
Angenendt
P
, et al
Core signaling pathways in human pancreatic cancers revealed by global genomic analyses
.
Science
2008
;
321
:
1801
6
.
28.
Wang
Q
,
Ni
Q
,
Wang
X
,
Zhu
H
,
Wang
Z
,
Huang
J
. 
High expression of RAB27A and TP53 in pancreatic cancer predicts poor survival
.
Med Oncol
2015
;
32
:
372
.
29.
Schultz
NA
,
Werner
J
,
Willenbrock
H
,
Roslind
A
,
Giese
N
,
Horn
T
, et al
MicroRNA expression profiles associated with pancreatic adenocarcinoma and ampullary adenocarcinoma
.
Mod Pathol
2012
;
25
:
1609
22
.
30.
Nuzhat
Z
,
Kinhal
V
,
Sharma
S
,
Rice
GE
,
Joshi
V
,
Salomon
C
. 
Tumour-derived exosomes as a signature of pancreatic cancer-liquid biopsies as indicators of tumour progression
.
Oncotarget
2017
;
8
:
17279
.
31.
Melo
SA
,
Sugimoto
H
,
O'Connell
JT
,
Kato
N
,
Villanueva
A
,
Vidal
A
, et al
, 
Cancer exosomes perform cell-independent microRNA biogenesis and promote tumorigenesis
.
Cancer Cell
2014
;
26
:
707
21
.
32.
Madhavan
B
,
Yue
S
,
Galli
U
,
Rana
S
,
Gross
W
,
Müller
M
, et al
Combined evaluation of a panel of protein and miRNA serum-exosome biomarkers for pancreatic cancer diagnosis increases sensitivity and specificity
.
Int J Cancer
2015
;
136
:
2616
27
.
33.
Vogel
R
,
Coumans
FAW
,
Maltesen
RG
,
Böing
AN
,
Bonnington
KE
,
Broekman
ML
, et al
A standardized method to determine the concentration of extracellular vesicles using tunable resistive pulse sensing
.
J Extracell Vesicles
2016
;
5
:
31242
.
34.
Schultz
NA
,
Dehlendorff
C
,
Jensen
BV
,
Bjerregaard
JK
,
Nielsen
KR
,
Bojesen
SE
, et al
MicroRNA biomarkers in whole blood for detection of pancreatic cancer
.
JAMA
2014
;
311
:
392
404
.
35.
Jamieson
NB
,
Morran
DC
,
Morton
JP
,
Ali
A
,
Dickson
EJ
,
Carter
CR
, et al
MicroRNA molecular profiles associated with diagnosis, clinicopathologic criteria, and overall survival in patients with resectable pancreatic ductal adenocarcinoma
.
Clin Cancer Res
2012
;
18
:
534
45
.
36.
Piepoli
A
,
Tavano
F
,
Copetti
M
,
Mazza
T
,
Palumbo
O
,
Panza
A
, et al
Mirna expression profiles identify drivers in colorectal and pancreatic cancers
.
PLoS One
2012
;
7
:
e33663
.
37.
Abue
M
,
Yokoyama
M
,
Shibuya
R
,
Tamai
K
,
Yamaguchi
K
,
Sato
I
, et al
Circulating miR-483-3p and miR-21 is highly expressed in plasma of pancreatic cancer
.
Int J Oncol
2015
;
46
:
539
47
.
38.
Dai
X
,
Pang
W
,
Zhou
Y
,
Yao
W
,
Xia
L
,
Wang
C
, et al
Altered profile of serum microRNAs in pancreatic cancer-associated new-onset diabetes mellitus
.
J Diabetes
2016
;
8
:
422
33
.
39.
Niu
Q
,
Li
X
,
Xia
D
,
Jiang
Y
,
Tian
Z
,
Bian
C
, et al
MicroRNA-186 affects the proliferation of tumor cells via yes-associated protein 1 in the occurrence and development of pancreatic cancer
.
Exp Ther Med
2017
;
14
:
2094
100
.
40.
Li
A
,
Omura
N
,
Hong
SM
,
Vincent
A
,
Walter
K
,
Griffith
M
, et al
Pancreatic cancers epigenetically silence SIP1 and hypomethylate and overexpress miR-200a/200b in association with elevated circulating miR-200a and miR-200b levels
.
Cancer Res
2010
;
70
:
5226
37
.
41.
Xu
L
,
Li
Q
,
Xu
D
,
Wang
Q
,
An
Y
,
Du
Q
, et al
hsa-miR-141 downregulates TM4SF1 to inhibit pancreatic cancer cell invasion and migration
.
Int J Oncol
2014
;
44
:
459
66
.
42.
Xu
J
,
Cao
Z
,
Liu
W
,
You
L
,
Zhou
L
,
Wang
C
, et al
Plasma miRNAs effectively distinguish patients with pancreatic cancer from controls a multicenter study
.
Ann Surg
2016
;
263
:
1173
9
.
43.
Deng
T
,
Yuan
Y
,
Zhang
C
,
Zhang
C
,
Yao
W
,
Wang
C
, et al
Identification of circulating miR-25 as a potential biomarker for pancreatic cancer diagnosis
.
Cell Physiol Biochem
2016
;
39
:
1716
22
.
44.
Kaur
S
,
Krishn
SR
,
Rachagani
S
,
Batra
SK
. 
Significance of microRNA-based biomarkers for pancreatic cancer
.
Ann Translat Med
2015
;
3
:
277
.
45.
Guo
R
,
Gu
J
,
Zhang
Z
,
Wang
Y
,
Gu
C
. 
MiR-451 promotes cell proliferation and metastasis in pancreatic cancer through targeting CAB39
.
Biomed Res Int
2017
;
2017
:
2381482
.
46.
Nagalla
S
,
Shaw
C
,
Kong
X
,
Kondkar
AA
,
Edelstein
LC
,
Ma
L
, et al
Platelet microRNA-mRNA coexpression profiles correlate with platelet reactivity
.
Blood
2011
;
117
:
5189
97
.
47.
Meikle
CKS
,
Kelly
CA
,
Garg
P
,
Wuescher
LM
,
Ali
RA
,
Worth
RG
. 
Cancer and thrombosis: the platelet perspective
.
Front Cell Dev Biol
2017
;
4
:
147
.
48.
Orellana
R
,
Kato
S
,
Erices
R
,
Bravo
ML
,
Gonzalez
P
,
Oliva
B
, et al
Platelets enhance tissue factor protein and metastasis initiating cell markers, and act as chemoattractants increasing the migration of ovarian cancer cells
.
BMC Cancer
2015
;
15
:
290
.
49.
Best
MG
,
Sol
N
,
Kooi
I
,
Tannous
J
,
Westerman
BA
,
Rustenburg
F
, et al
RNA-Seq of tumor-educated platelets enables blood-based pan-cancer, multiclass, and molecular pathway cancer diagnostics
.
Cancer Cell
2015
;
28
:
666
76
.
50.
Ko
J
,
Baldassano
SN
,
Loh
P
,
Kording
K
,
Litt
B
,
Issadore
D
. 
Machine learning to detect signatures of disease in liquid biopsies - a user's guide
.
Lab Chip
2017
;
18
:
395
405
.
51.
Vellido
A
,
Martìn-Guerrero
JD
,
Lisboa
PJG
. 
Making machine learning models interpretable
.
Available from:
https://pdfs.semanticscholar.org/ce0b/8b6fca7dc089548cc2e9aaac3bae82bb19da.pdf.
52.
Ozkumur
E
,
Shah
AM
,
Ciciliano
JC
,
Emmink
BL
,
Miyamoto
DT
,
Brachtel
E
, et al
Inertial focusing for tumor antigen–dependent and -independent sorting of rare circulating tumor cells
.
Sci Translat Med
2013
;
5
:
179ra47
.
53.
Aatonen
MT
,
Öhman
T
,
Nyman
TA
,
Laitinen
S
,
Grönholm
M
,
Siljander
PR-M
, et al
Isolation and characterization of platelet-derived extracellular vesicles
.
J Extracell Vesicles
2014
;
3
:
24692
.
54.
Muller
L
,
Hong
CS
,
Stolz
DB
,
Watkins
SC
,
Whiteside
TL
. 
Isolation of biologically-active exosomes from human plasma
.
J Immunol Methods
2014
;
411
:
55
65
.
55.
Korpal
M
,
Kang
Y
. 
The emerging role of miR-200 family of microRNAs in epithelial-mesenchymal transition and cancer metastasis
.
RNA Biol
2008
;
5
:
115
9
56.
Liu
Y
,
Ding
Y
,
Huang
J
,
Wang
S
,
Ni
W
,
Guan
J
, et al
MiR-141 suppresses the migration and invasion of HCC cells by targeting Tiam1
.
PLoS One
2014
;
9
:
e88393
.
57.
Fong
MY
,
Zhou
W
,
Liu
L
,
Alontaga
AY
,
Chandra
M
,
Ashby
J
, et al
Breast-cancer-secreted miR-122 reprograms glucose metabolism in premetastatic niche to promote metastasis
.
Nat Cell Biol
2015
;
17
:
183
94
.
58.
Soung
YH
,
Ford
S
,
Zhang
V
,
Chung
J
. 
Exosomes in cancer diagnostics
.
Cancers
2017
;
9
:
8
.
59.
Mirzaei
H
,
Sahebkar
A
,
Jaafari
MR
,
Goodarzi
M
,
Mirzaei
HR
. 
Diagnostic and therapeutic potential of exosomes in cancer: the beginning of a new tale?
J Cell Physiol
2016
;
232
:
3251
60
.