Simultaneous subset tracing and miRNA profiling of tumor-derived exosomes via dual-surface-protein orthogonal barcoding

Abstract

The clinical potential of miRNA-based liquid biopsy has been largely limited by the heterogeneous sources in plasma and tedious assay processes. Here, we develop a precise and robust one-pot assay called dual-surface-protein-guided orthogonal recognition of tumor-derived exosomes and in situ profiling of microRNAs (SORTER) to detect tumor-derived exosomal miRNAs and enhance the diagnostic accuracy of prostate cancer (PCa). The SORTER uses two allosteric aptamers against exosomal marker CD63 and tumor marker EpCAM to create an orthogonal labeling barcode and achieve selective sorting of tumor-specific exosome subtypes. Furthermore, the labeled barcode on tumor-derived exosomes initiated targeted membrane fusion with liposome probes to import miRNA detection reagents, enabling in situ sensitive profiling of tumor-derived exosomal miRNAs. With a signature of six miRNAs, SORTER differentiated PCa and benign prostatic hyperplasia with an accuracy of 100%. Notably, the diagnostic accuracy reached 90.6% in the classification of metastatic and nonmetastatic PCa. We envision that the SORTER will promote the clinical adaptability of miRNA-based liquid biopsy.

The one-pot SORTER assay profiled tumor-derived exosomal miRNAs in 0.2 μl plasma, allowing for early diagnosis of prostate cancer.

INTRODUCTION

MicroRNAs (miRNAs) are a class of short, noncoding, single-stranded RNAs (approximately 22 nucleotides) that play fundamental roles in gene expression regulation by repressing the translation of target genes or degrading their target transcripts (1, 2). Dysregulated miRNAs are closely associated with the pathogenesis of a variety of cancers and thus have been studied as emerging biomarkers in cancer diagnosis, prognosis, and treatment monitoring (3, 4). However, the clinical value of miRNAs in liquid biopsy was questioned by researchers and attributed to the heterogeneous origins and existing forms of miRNAs in bodily fluids, such as circulating free miRNAs, miRNAs bound with ribonucleoproteins, or miRNAs encapsulated in different extracellular vesicles (5, 6). The nonspecific source of biomarkers diminishes the accuracy of miRNAs as a diagnostic tool and has become the stumbling block to moving from proof of concept to clinical application. Exosomes are a unique subtype of extracellular vesicles generated via endosomes and multivesicular body pathways (7, 8). miRNAs are selectively packed and enriched in exosomes as regulators of a wide range of physiologic and pathologic processes and are kept stable under the protection of lipid membranes (9, 10). In particular, tumor-derived exosomal miRNAs are closely associated with cancer progression (11, 12) and afford great promise to enhance the specificity and accuracy of miRNA-based liquid biopsy. Unfortunately, the specific recognition of tumor-derived exosomes and quantitative detection of exosomal miRNAs are still technically challenging due to the complex background interference (13), heterogeneous extracellular vesicles (EVs) subtypes (14), and varied expression levels of different exosomal miRNAs (15). In addition, the existing miRNA assays require multiple manual steps and take a long assay time. Therefore, developing a sensitive, accurate, and simplified bioassay to quantify tumor-derived exosomal miRNAs in complicated biofluid samples is highly desirable to promote the development of clinically viable miRNA biomarkers of cancer.

The prevailing RNA quantification technologies, such as quantitative reverse transcription polymerase chain reaction (RT-qPCR) and next-generation sequencing, enable the detection of specific miRNA species in exosomes with high sensitivity down to femtomolar (fM) (16, 17). However, these quantification technologies often require a large sample volume (>500 μl) for isolating sufficient exosomes before RNA analysis and also involve laborious and time-consuming procedures, including exosome lysis, RNA extraction, cDNA generation, and sequence amplification (18, 19). Moreover, these analyses are susceptible to interference by free miRNAs in biofluids, whose abundance is several orders higher than exosomal miRNAs (20, 21). Alternative to the sequence amplification–based strategies, a series of in situ exosomal miRNA bioassays have been recently developed by directly or indirectly importing DNA probes into membrane vesicles, such as gold nanoflares (Au NFs) and molecular beacons (MBs) (22, 23). Gao et al. designed a virus-mimicking fusogenic vesicle-encapsulated MBs probe, which can rapidly detect extracellular vesicle miRNAs within 2 hours via membrane fusion (24). Zhao et al. (25) proposed a thermophoretic sensor for in situ extracellular vesicle miRNA analysis by transporting Au NFs into vesicles, showing detection sensitivity down to 0.36 fM in 0.5 μl plasma samples. These in situ technologies provided efficient and robust extracellular vesicle miRNA detection without resorting to tedious RNA extraction and eliminated the background interference of complex biofluid samples. However, the transportation of miRNA probes into vesicles is stochastic and cannot distinguish tumor-derived exosomes from other nonspecific vesicles. The ensemble detection of total EV miRNAs conceals the specificity of tumor-derived exosomal miRNAs and reduces the accuracy of liquid biopsy applications.

Tumor-derived exosomes account for only a small fraction of EVs, making it challenging to precisely distinguish them from others (26, 27). Surface protein receptors are routinely used to identify the unique subtypes of various extracellular vesicles (13, 28). Typically, the tetraspanin CD63 is often enriched in exosomes compared with other vesicles (29), and epithelial cell adhesion molecule (EpCAM) is a typical marker of tumors (30). The orthogonal combination of the two proteins sets a more critical threshold and holds great promise in recognizing and sorting tumor-derived exosomes. A multireceptor-based DNA logic device has recently been reported to bind multiple aptamers into a single computing device and recognize subpopulations of cells and exosomes (31, 32). For instance, Chang et al. (33) developed an Adleman-Nishimura-Deamer (AND) Boolean logical device using multiple aptamers and toehold activation for signal integration and amplification to label and recognize target cell types via the synergistic presence of different surface protein receptors. Our group reported that dual-surface-protein-aptamer recognition combined with droplet digital PCR achieved quantitative profiling of tumor-derived exosomal programmed cell death 1 ligand 1 (PD-L1) (34). However, to the best of our knowledge, this multiple protein receptor synergistic logic device has not been used in the selective recognition and sorting of exosome subtypes.

Addressing these challenges, here, we developed dual-surface-protein-guided orthogonal recognition of tumor-derived exosomes and in situ probing of microRNA (SORTER) profiles for rapid and specific detection of tumor-derived exosomal miRNAs. The SORTER affords three noteworthy advantages over the existing exosomal miRNA biosensing assays. First, we present a dual-surface-protein–guided orthogonal labeling strategy to recognize and sort tumor-derived exosomes precisely, a small yet considerable subpopulation of extracellular vesicles, improving the diagnostic and prognostic accuracy of exosome-based liquid biopsy. Specifically, dual allosteric aptamers of exosome-specific marker CD63 and tumor marker EpCAM were used to create a unique orthogonal identity barcode on tumor-derived exosomes, thus inducing the targeted recognition and controlled fusion of miRNA probes for signal amplification. Second, the in situ miRNA detection inside the membrane structures of exosomes prevents the contamination of free circulation miRNAs and degradation of ribonucleases (RNases) from biofluid samples, enabling the accurate quantification and even dynamic monitoring of tumor-derived exosomal miRNAs. Third, we incorporate multiple processes into SORTER, such as exosome recognition, importing probes, miRNA signal transduction, and amplification and create separation- and washing-free tumor-derived exosomal miRNAs assay. The SORTER offers superior analytical performance toward liquid biopsy applications, which consumes only 0.2 μl of plasma sample and completes the whole analysis in less than 2 hours. We tested an exosome signature of six miRNAs (miR-222, miR-1290, miR-182, miR-21, miR-221, and miR-10b) in the training (n = 42) and validation (n = 32) cohorts, which can differentiate prostate cancer (PCa) and benign prostatic hyperplasia (BPH) with a sensitivity, specificity, and accuracy of 100%. The diagnostic accuracy also reached 90.6% in the classification of metastatic and nonmetastatic PCa (nPCa). We envision that the SORTER provides a promising tool to advance the analysis of tumor-derived exosomal miRNAs and promote the clinical adaptability of miRNA-based liquid biopsy.

RESULTS

Working principle of SORTER

The assay workflow consists of three parts (Fig. 1A, steps I to III), including sample collection and plasma preparation, SORTER assay for miRNA profiling of tumor-derived exosomes, and biostatistical analysis for cancer diagnosis. The SORTER assay is designed to achieve specific recognition and sorting of tumor-derived exosome subtypes and in situ sensitive probing of tumor-derived exosomal miRNA profiles, further improving the miRNA-based diagnostic accuracy of PCa. The SORTER incorporates multiple parallel processes (Fig. 1B, steps I to III), including barcode labeling of tumor-derived exosomes, targeted importation of miRNA detection probes, and in situ profiling of tumor-derived exosomal miRNAs, thus enabling sensitive and precise one-pot miRNA profiling in tumor-derived exosomes. The two substantial innovations of SORTER are selective labeling and sorting of tumor-derived exosomes through dual-surface-protein–guided orthogonal recognition and in situ sensitive quantification of miRNAs via duplex-specific nuclease (DSN)–catalyzed signal amplification.

Fig. 1. Schematic illustration of SORTER assay for miRNA profiling of tumor-derived exosomes.

(A) The assay workflow consists of three parts. (I) Clinical plasma samples (0.2 μl, 100-fold dilution) were collected from age-matched PCa patients and BPH controls. Exosomes, ectosomes, and free molecules produced by tumor and normal cells coexist in plasma samples and exhibit overlapping compositional features. (II) SORTER assessment of miRNA profiles in tumor-derived exosomes. The SORTER assay is designed to achieve specific recognition and sorting of tumor-derived exosome subtypes and in situ sensitive probing of tumor-derived exosomal miRNA profiles. (III) Data processing and bioinformatic analysis for cancer diagnosis. The linear discriminant analysis (LDA) algorithm is used to identify the best combinations of miRNAs to classify patients with PCa from BPH controls, and the LDA model then evaluates the predicted results. (B) SORTER incorporates multiple parallel processes, including exosome recognition, importing probes, and miRNA profiling, permitting a sensitive and robust one-pot tumor-derived exosomal miRNA assay. (I) Dual-surface-protein-guided orthogonal recognition barcode for the selective labeling of tumor-derived exosomes. (II) Importation of miRNA detection probes into tumor-derived exosomes. (III) In situ sensitive profiling of miRNAs inside tumor-derived exosomes.

There is no single marker to selectively distinguish the heterogeneous sources of miRNAs in plasma. In plasma samples, for example, tumor- and normal-derived exosomes, ectosomes, and free molecules (e.g., RNA-protein complexes) coexist and have overlapping compositional properties. Therefore, we leverage the combination of two typical surface protein markers, CD63 for exosomal-specific markers and EpCAM for the tumor-specific marker, to set up an orthogonal screening threshold and precisely recognize tumor-derived exosomes from complicated biofluid samples. Aptamers are promising tools for protein labeling because of their excellent specificity and affinity, low production costs, and configurational programmability (35, 36). We designed two allosteric aptamer probes of CD63 and EpCAM (detailed in fig. S1) as input units for orthogonal labeling of the tumor-derived exosome in clinical samples. The two allosteric aptamer probes (CD63-S-L and EpCAM-S-L) consist of three domains: the aptamer (CD63 or EpCAM) domain, the spacer (S) domain, and the linker (L) domain. The allosteric aptamer probes are in a non-active state with a hairpin structure, where the L domains are blocked for the subsequent reaction. When two protein receptors, CD63 and EpCAM, are recognized synergistically on a single exosome, two distinct L domains are exposed by allosteric transformation, creating a unique orthogonal identity barcode on tumor-derived exosomes via proximity-induced self-assembly. In particular, the unbounded aptamer probes are still in a non-active state, so there is no need to wash away excess probes. The dual-target-protein-guided orthogonal barcoding method permits the rapid and selective labeling of tumor-derived exosomes in clinical samples (Fig. 1B, step I), avoiding lengthy pre-isolation/purification processes and minimizing the nonspecific interference of contaminated vesicles.

The quantification of tumor-derived exosomal miRNAs faces two major challenges: (i) the amount of circulating free miRNAs in the blood is several orders of magnitude more than target miRNAs in tumor-derived exosomes and causes serious interference to detection (14, 37); (ii) the concentration of miRNAs in tumor-derived exosomes is extremely low (approximately 1 copy/10⁶ EVs to 1 copy/1 EV) (21, 38), thus a superior sensitive assay is required for miRNA profiling. To address these issues, we designed a smart liposome probe (Tags-Lipo@Au NFs) to import miRNA detection probes into tumor-derived exosomes via targeted vesicle fusion and achieve in situ sensitive quantification of various miRNAs inside exosomes. We used the orthogonal barcode on tumor-derived exosomes (Orth-Exo) to hybridize with complementary DNA tags anchored on liposome probes (Tags-Lipo@Au NFs) in a zipper-like behavior (Fig. 1B, step II), facilitating the selective recognition of tumor-derived exosomes and the simultaneous importation of miRNA detection probes. After that, the fluorescent signal of miRNAs was then generated and amplified using Au NFs and DSN encapsulated in liposome probes (Fig. 1B, step III). Specifically, the Au NFs were prepared by immobilizing recognition sequences [(5′-labeled thiol (SH) and 3′-labeled carboxyfluorescein (FAM)] onto spherical Au nanoparticles, where the fluorescence is effectively quenched by the Au surface, greatly minimizing the background signal. The target miRNA will bind to the DNA probe in the nanoflares to form DNA-RNA heteroduplexes. Notably, DSN will cleave the DNA sequence in DNA-RNA heteroduplexes specifically to release fluorophore for fluorescence and RNA sequence and initiate the target recycling and signal amplification processes. The in situ miRNA assay inside the membrane structures of exosomes eliminates the interference of circulating free miRNAs, and the Tags-Lipo@Au NFs probes provide a minimized background and amplified fluorescent signal, thus enabling highly selective and sensitive quantification of target miRNAs in tumor-derived exosomes.

Dual-surface-protein orthogonal barcode labeling on tumor-derived exosomes

Tetraspanins, particularly CD63, CD9, and CD81, are the most widely used exosomal markers to distinguish exosomes from other nonspecific EVs, such as microvesicles or apoptotic bodies (37, 39). Now, the prevailing literature exclusively uses CD63 aptamers for exosome identification and separation (40, 41). To distinguish the tumor and normal cell-derived exosomes, we selected a panel of nine membrane protein markers of exosomes: one exosome protein marker (CD63) and eight putative tumor protein markers, including EpCAM, nucleolin, carcinoembryonic antigen, human epidermal growth factor receptor 2, mucin 1, PD-L1, prostate-specific membrane antigen, and protein tyrosine kinase 7. PCa cell (LNCaP, DU145, PC-3)– and BPH cell (BPH-1)–derived exosomes (Exo) were used as the models of tumor-derived and normal-derived exosomes. The FAM-labeled aptamer probes were used to bind to the target proteins on the single exosomes, respectively (table S1) (30, 42–48). When compared to BPH-1 Exo (fig. S5), nanoflow cytometry (NanoFCM) analysis of LNCaP Exo, DU145 Exo, and PC-3 Exo (figs. S2 to S4) reveals the highest labeling efficiency of EpCAM (ranging from 43.1 to 75.1%) and the relatively high labeling efficiency of CD63 (ranging from 16.2 to 48.7%). In addition, statistical analysis yielded a reasonably strong correlation between the NanoFCM and those of the enzyme-linked immunosorbent assay of nine-protein markers expression in four cell line–derived exosomes (Pearson’s r = 0.8736; fig. S6). Therefore, the synergistic identification of target proteins CD63 (an exosome marker) and EpCAM (a tumor marker) is the best choice for recognizing and sorting tumor-derived exosomes.

To rapidly and selectively label tumor-derived exosomes with a traceable barcode in complex clinical scenarios, we designed two allosteric aptamer probes of CD63 and EpCAM proteins as input units for orthogonal labeling of the tumor-derived exosome. The allosteric aptamer probes (CD63-S-L or EpCAM-S-L) consist of three domains: the aptamer (CD63 or EpCAM) domain, the spacer (S) domain, and the linker (L) domain. The allosteric aptamer probes are in their non-active configuration with a hairpin structure, and the L domain is annealed and blocked for the downstream reaction. The aptamer domain of allosteric probes can specifically recognize the target proteins, exposing the L domain by allosteric transformation. To confirm whether CD63-S-L and EpCAM-S-L bind to target proteins and expose the L domain, we performed fluorescence kinetic analysis on LNCaP Exo using a fluorescence resonance energy transfer (FRET) couple (BHQ1 and FAM) double-labeled CD63-S-L and EpCAM-S-L (fig. S7, A and B). This allowed us to easily monitor the allosteric transformation reaction through an increase in the fluorescence signal. When the two allosteric aptamer probes of CD63-S-L or EpCAM-S-L were incubated with LNCaP Exo, the fluorescence signal increased markedly, while almost no signal changes were observed in the control experiments (CD63-S-L or EpCAM-S-L only). These findings validated that the designed allosteric aptamer probes of CD63-S-L or EpCAM-S-L can bind to target proteins of exosomes efficiently and trigger an allosteric change to expose the L domain, allowing the orthogonal barcode labeling of tumor-derived exosomes.

To confirm whether CD63 and EpCAM are coexpressed on a single tumor-derived exosome surface, we used a total internal reflection fluorescent microscope (TIRFM) to image CD63-S-L (3′-labeled FAM) and EpCAM-S-L (5′-labeled Cy5) double-labeled tumor LNCaP Exo and normal BPH-1 Exo. Figure 2A illustrates CD63-S-L (FAM)– and EpCAM-S-L (Cy5)–mediated orthogonal labeling on a single exosome surface. Substantially strong fluorescent signals of CD63 (green dot) and EpCAM (red dot) were collocated on the same particles of LNCaP Exo (orange dot in merged images), while only a little co-localization of fluorescent signals was detected from BPH-1 Exo (Fig. 2B). We further checked the labeling efficiencies of CD63-S-L (FAM) and EpCAM-S-L (Cy5) on LNCaP Exo and BPH-1 Exo by NanoFCM (Fig. 2C). It was observed that EpCAM-S-L (Cy5) has a substantially higher labeling efficiency on LNCaP Exo (74.5%) than BPH-1 Exo (2.5%), while CD63-S-L (FAM) has a better labeling efficiency on LNCaP Exo (45.7%) than BPH-1 Exo (2.4%). These findings show that CD63 and EpCAM proteins are expressed and coexist on a single tumor-derived exosome surface and that CD63 and EpCAM probes can be effectively labeled on the same vesicle surface.

Fig. 2. Validation of dual-surface-protein orthogonal labeling on tumor-derived exosomes.

(A) Schematics of CD63-S-L– and EpCAM-S-L–mediated orthogonal labeling on a single exosome surface. (B and C) TIRFM images (B) and NanoFCM analysis (C) show the allosteric aptamer probes of CD63-S-L and EpCAM-S-L to selectively label target proteins on LNCaP Exo and BPH-1 Exo. Experimental group: Exo + CD63-S-L (3′-labeled FAM) + EpCAM-S-L (5′-labeled Cy5). (D) Schematics of the zipper-like hybridization of the orthogonal barcode-anchored exosome (Orth-Exo) and complementary DNA tags (Tags). (E and F) TIRFM images (E) and NanoFCM analysis (F) of the zipper-like hybridization of the orthogonal barcode and Tags against LNCaP Exo and BPH-1 Exo. Control group: Exo + CD63-S-L + rEpCAM-S-L (5′-labeled Cy5, only the aptamer domain is replaced by a random sequence) + Tags (5′-labeled FAM) or Exo + rCD63-S-L + EpCAM-S-L (5′-labeled Cy5) + Tags (5′-labeled FAM); experimental group: Exo + CD63-S-L + EpCAM-S-L (5′-labeled Cy5) + Tags (5′-labeled FAM).

To validate the zipper-like hybridization of the orthogonal barcode-anchored exosome (Orth-Exo) and cholesterol-labeled complementary DNA tags (Tags), we detected the CD63-S-L, EpCAM-S-L (5′-labeled Cy5), and Tags (5′-labeled FAM) colabeled on LNCaP Exo and BPH-1 Exo and analyzed them by TIRFM. Figure 2D illustrates the zipper-like hybridization of the orthogonal barcode and Tags on a single exosome surface. As shown in Fig. 2E and fig. S8, the colocation of EpCAM-S-L (red dot) and Tags (green dot) on the same vesicle was evident in a part of the LNCaP Exo (orange dot in merged images) as compared to the BPH-1 Exo. In addition, no colocation fluorescent signals were observed when CD63-S-L was substituted with rCD63-S-L (only the aptamer domain is replaced by a random sequence) or EpCAM-S-L was replaced with rEpCAM-S-L. We further checked the labeling efficiencies of CD63-S-L, EpCAM-S-L (Cy5), and Tags (FAM) on LNCaP Exo and BPH-1 Exo by NanoFCM (Fig. 2F). Tags (FAM) have a substantially higher labeling efficiency on LNCaP Exo (51.5%) than BPH-1 Exo (0.3%). When EpCAM-S-L was replaced with rEpCAM-S-L or CD63-S-L was replaced with rCD63-S-L, no apparent labeling signals were detected in either LNCaP Exo (0.0 or 0.1%) or BPH-1 Exo (0.0 or 0.0%). These studies revealed the successful formation of orthogonal barcodes on a single tumor-derived exosome surface and the effective zipper-like hybridization of DNA Tags and orthogonal barcodes on the same vesicle surface. Meanwhile, we further measured the labeling intensities on LNCaP Exo with the following fluorescence sequences: CD63-S-L (3′-labeled FAM), EpCAM-S-L (5′-labeled FAM), and Tags (5′-labeled FAM). The fluorescent signal of Tags [1032.0 arbitrary units (a.u.)] exhibited an elevation compared with the sum of CD63-S-L (445.8 a.u.) and EpCAM-S-L (123.58 a.u.; fig. S9, A and B), demonstrating that the formed orthogonal barcodes improved the binding affinities of allosteric aptamer probes. In addition, the structural stability of orthogonal barcodes on a single exosome surface was estimated using the melting temperatures (fig. S9C). These results demonstrated that the orthogonal barcode has excellent structural stability on the vesicle surface, allowing for the effective zipper-like hybridization of DNA Tags and orthogonal barcode.

The main goal of the DNA probes is to label and identify exosomes in complicated biological fluids, such as plasma or urine. Therefore, DNA probes must have prolonged durability of several hours in biological fluids. We checked the stability of EpCAM-S-L, CD63-S-L, and Tags in different concentrations of plasma (ranging from 0 to 14%) after incubating for 2 hours at 37°C. The treated samples were subsequently subjected to analysis using agarose gel electrophoresis (fig. S10, A to C). The quantified results showed that EpCAM-S-L, CD63-S-L, and Tags remained stable across all cases without any noticeable degradation, even after exposure to 14% plasma. Furthermore, we checked the labeling efficiencies of CD63-S-L, EpCAM-S-L (Cy5), and Tags (FAM) across different cell line–derived exosomes spiked in phosphate-buffered saline (PBS) and 100-diluted plasma samples by NanoFCM (fig. S11, A and B). In comparison to the labeling efficiency of exosomes in PBS solution, the labeling efficiency of exosomes in plasma samples did not exhibit a substantial alteration, indicating that this labeling strategy is suitable for the labeling of tumor-derived exosomes in 100-fold diluted clinical plasma specimens.

Dynamic monitoring of dual-surface-protein–guided liposome probe fusion

The hybridization reaction between DNA tags on liposomes and orthogonal barcodes on tumor-derived exosomes guided the targeted fusion of liposomes and exosomes. The morphology and size of liposomes and LNCaP Exo were first characterized by nanoparticle tracking analysis and transmission electron microscopy (TEM). The liposomes and exosomes were typical sphere- or cup-shaped vesicles and had a diameter of about 143 and 169 nm, respectively (fig. S12, A and B). To confirm dual-surface-protein–guided liposome probe fusion, we monitored fusion-based membrane mixing using a FRET-based assay (Fig. 3A). DNA tag-anchored liposomes (Tags-Lipo) were double-labeled with the donor of 3,3′-dioctadecyloxacarbocyanine perchlorate (DiO; 501 nm) and the acceptor of 1,1′-dioctadecyl-3,3,3′,3′ tetramethylindocarbocyanine perchlorate (DiI; 565 nm). After fusion, the membrane was enlarged, leading to decreased FRET efficiency between the DiO and DiI on the membrane surface. As presented in Fig. 3B, the FRET efficiency declined as the molar ratio of Tags-Lipo-DiO-DiI to orthogonal barcode-anchored exosomes (Orth-Exo) further increased. However, nearly no fusion was seen in the stochastic fusion between Lipo-DiO-DiI and exosomes with no barcodes (Exo). The membrane fusion dynamic was also monitored in the FRET assay. When nonfluorescent Orth-Exo was incubated with double-labeled Tags-Lipo (Tags-Lipo-DiI-DiO), the DiO signals increased quickly and reached a plateau value in 2 hours (Fig. 3C), while no change was seen in the stochastic fusion of Lipo-DiO-DiI and Exo. These results confirmed that the successful fusion is indeed mediated by the zipper-like hybridization between Orth-Exo and Tags-Lipo-DiI-DiO. Furthermore, the fusion mixing analyses of Tags-Lipo-DiO-DiI and Orth-Exo at different temperatures were recorded in Fig. 3D. With the increase in temperature, the stochastic fusion of Lipo-DiO-DiI and Exo was also enhanced. Consequently, 37°C was the optimum temperature for the SORTER assay, holding a maximized target-to-stochastic fusion efficiency ratio. Furthermore, as shown in fig. S13, the fluorescence signal of dual-labeled LNCaP Exo with both CD63 and EpCAM allosteric aptamer probes exhibited a substantial increase after incubation with liposome probes of Tags-Lipo-DiO-DiI (P < 0.001). In contrast, a minimal change in signal was observed in single-labeled LNCaP Exo samples with either CD63 or EpCAM allosteric aptamer probes. These results demonstrated the specific fusion of liposome probes and dual-labeled exosomes. In addition, we also assessed the fusion efficiency of dual-labeled LNCaP Exo and BPH-1 Exo after incubation with liposome probes of Tags-Lipo-DiO-DiI (fig. S14). The tumor LNCaP Exo had much higher fluorescence than normal BPH-1 Exo, suggesting that the expression level of the CD63 and EpCAM proteins affects the fusion efficiency fusion of exosomes and liposomes.

Fig. 3. Dynamic monitoring of dual-surface-protein–guided liposome probe fusion.

(A) Schematic illustration of the FRET-based lipid membrane mixing for investigating the orthogonal fusion between Orth-Exo and Tags-Lipo-DiO-DiI. The fusion event was measured by the decreased FRET efficiency between the donor (DiO, 501 nm) and acceptor (DiI, 565 nm). (B) Fluorescence spectra analysis of the orthogonal fusion between Tags-Lipo-DiO-DiI and 1× and 10× molar ratios Orth-Exo. Negative control for the stochastic fusion of Lipo-DiO-DiI and Exo. (C) Fluorescence kinetic analysis of the target fusion between Tags-Lipo-DiO-DiI and Orth-Exo. Negative control for the stochastic fusion between Lipo-DiO-DiI and Exo. (D) Fusion mixing analysis of Tags-Lipo-DiO-DiI and Orth-Exo at different temperatures. Control experiment for the stochastic fusion between Lipo-DiO-DiI and Exo. The data represent the mean ± SD (n = 3). (E) TIRFM images showing the orthogonal fusion between the Tags-Lipo-DiI and Orth-Exo-DiO and the stochastic fusion between Lipo-DiI and Exo-DiO. (F) Diameters of the fusion products are determined by the DLS method at different time intervals. The data represent the mean ± SD (n = 3). (G) TEM images of Orth-Exo only, Tags-Lipo@Au NFs only, and the fusion vesicles of Tags-Lipo@Au NFs and Orth-Exo.

To further validate the effective fusion of Tags-Lipo and Orth-Exo, we labeled Tags-Lipo and Orth-Exo with DiI and DiO, respectively (Fig. 3E). The DiI fluorescence of Tags-Lipo (red dot) colocalized well with that of DiO in Orth-Exo (green dot). In contrast, almost no detectable colocalized fluorescence signals were observed in the stochastic fusion between Lipo-DiI and Exo-DiO. Notably, we observed multiple Orth-Exo (green dot) around large overlapping particles (yellow dot), indicating that the orthogonal fusion of Orth-Exo and Tags-Lipo-DiI-DiO are DNA-programmed cascade reactions. The hydrodynamic size distribution of the membrane fusion product was determined by the dynamic light scattering (DLS) method at different time intervals (Fig. 3F). The diameter of the fused vesicles gradually increased from 154.7 to 206.5 nm, while the diameter barely enhanced owing to the lack of zipper-like hybridization of Tags and orthogonal barcodes in control experiments. By functionalizing a 13-nm spherical gold nanoparticle (Au NP), Au NFs with a recognition sequence to target miR-21 were created. The close proximity of FAM to the surface of Au NPs resulted in fluorescence quenching. DLS measurements revealed that Au NFs had an average size of 64 nm, and the loading density of the recognition sequence on each Au NFs was determined to be 75 nM (fig. S15). TEM analysis of these vesicles (Fig. 3G) revealed the prepared individual liposomes or exosomes, which is consistent with incomplete fusion. However, the hemifusion intermediate or fully fused state was observed in the Tags-Lipo@Au NFs and Orth-Exo fusion reactions. These results demonstrate the effective dual-surface-protein–mediated orthogonal fusion of Orth-Exo and Tags-Lipo, allowing for the importation of the encapsulated miRNA detection probes for downstream analysis.

SORTER for tumor-derived exosomal miRNA analysis

PCa is one of the most prevalent diagnosed malignancies affecting men worldwide, with increasing cancer-related mortality (49, 50). Now, the widely used serum prostate-specific antigen (PSA) screening lacks sufficient specificity and sensitivity for the clinical diagnosis of PCa. Patients with BPH frequently have elevated PSA levels, leading to needless prostate biopsies and overtreatment (51). Therefore, it is urgent to develop a liquid biopsy assay to facilitate the early diagnosis of PCa, which is crucial for improving long-term clinical outcomes. miRNAs inside tumor-derived exosomes, closely associated with PCa development, invasion, and metastasis (52), are potential liquid biopsy biomarkers for PCa diagnosis. Here, we developed the sensitive and robust one-pot SORTER assay to enable multiparametric miRNA profiling of tumor-derived exosomes in PCa plasma samples. Several putative PCa-associated miRNAs were thus selected to be detected in tumor-derived exosomes, including miR-21, miR-10b, miR-182, miR-222, miR-221, and miR-1290. miR-21 is the most frequently used marker and is overexpressed in various cancer types, including PCa (53). miR-10b is prominently expressed in PCa and promotes tumor proliferation, migration, and invasion (54). miR-182 promotes tumor cell proliferation, colony formation, migration, and invasion (55). miR-221 and miR-222 are regulated in PCa, increasing androgen-independent growth (56). miR-1290 is significantly associated with overall survival and assisted in treatment decisions for individuals with castration-resistant PCa (57). The functions and pathways of the six markers were presented in table S3. The SORTER incorporates multiple processes (Fig. 4A), including exosome recognition, probe importation, fluorescent signal transduction, and amplification. To evaluate the assay performance of SORTER, we selected miR-21 as a model to optimize the experimental conditions. The optimum temperature for DSN activity was 37°C, and the optimum amount of DSN enzyme was 1 U in a 300-μl reaction volume (fig. S16, A and B). Furthermore, the sensitivity of the SORTER system was verified by incubating bare Au NFs with the synthesized miR-21, and the detection limit of the target miRNA is about 160 fM (fig. S17, A and B).

Fig. 4. SORTER for tumor-derived exosomal miRNA analysis.

(A) Schematic illustration of the SORTER approach for tumor-derived exosomal miRNA analysis. (B to D) Fluorescence intensity (B), NanoFCM (C), TIRFM (D) analysis of miR-21 expression in orthogonal barcode-based BPH-1 Exo or LNCaP Exo after incubation with Tags-Lipo@Au NFs and Lipo@Au NFs, respectively. The P value was determined by a two-sided, parametric t test. The data represent the mean ± SD (n = 3). (E) Calibration curves for quantifying LNCaP-derived exosomal miR-21 spiked in PBS and EV-depleted plasma (diluted by 100-folds in 1× PBS). The data represent the mean ± SD (n = 3). (F) The Radar plot shows six miRNA markers from the four cell lines-derived exosomes, including three PCa cells (PC-3, LNCaP, and DU145) and one BPH cell (BPH-1). (G) SORTER approach for miR-21 analysis in the fused vesicles after incubating with Tags-Lipo@Au NFs and Lipo@Au NFs in healthy and cancer plasma samples. The P value was determined by a two-sided, parametric t test. The data represent mean ± SD (n = 3).

Under optimized assay conditions, we further evaluated the SORTER assay for tumor-derived exosomal miR-21 analysis. As shown in Fig. 4B, the fluorescent signal of LNCaP Exo increased significantly after incubation with Tags-Lipo@Au NFs (P < 0.001). In contrast, almost no signal change was observed in BPH-1 Exo after incubation with Tags-Lipo@Au NFs or in BPH-1 Exo and LNCaP Exo after incubation with Lipo@Au NFs. Furthermore, NanoFCM and TIRFM results (Fig. 4, C and D) further showed that tumor LNCaP Exo had significantly higher fluorescence than BPH-1 Exo. These results indicated an elevated level of miR-21 in the tumor exosomes. The sensitivity of the SORTER approach was determined by probing miR-21 in purified LNCaP Exo spiked in both PBS and healthy plasma (Fig. 4E). Analysis of spiked plasma demonstrated comparable analytical merits to those of spiked PBS, such as calibration sensitivity (that is, the slope: 0.8438 versus 0.8396) and limit of detection (LOD) (1.2 × 10⁵ versus 0.8 × 10⁵ particles μl⁻¹). On the basis of the RT-qPCR quantification result, our assay result shows that each LNCaP Exo contained 2.565 × 10⁻² copies of miR-21 (fig. S18). In comparison to RT-qPCR and other exosomal miRNA detection methods (table S4) (9, 18, 24, 25, 58–60), our technique with equivalent sensitivity may bypass complicated processes for exosome separation and RNA extraction, enabling quick and practicable detection of tumor-derived exosomal miRNAs. We next characterized the performance of SORTER by measuring the expression levels of six miRNAs in tumor-derived exosomes, including miR-21, miR-222, miR-1290, miR-221, miR-10b, and miR-182, in three PCa cells- (PC-3, LNCaP, and DU145) and one BPH-1–derived exosome. As shown in Fig. 4F, experimental results showed that the expressions of these miRNAs in tumor-derived exosomes are significantly higher than in normal BPH-1 Exo. We also used RT-qPCR as the gold standard to measure the expression levels of six miRNAs in different cell-derived exosomes, and a high correlation coefficient of 0.9200 was obtained between RT-qPCR and SORTER (fig. S19). Such high sensitivity allowed us to detect a low level of exosomal miR-21 directly in PCa patient plasma, as verified by the measurements of a PBS blank and a control plasma (P < 0.001; Fig. 4G).

We further labeled EVs in 100-diluted plasma from PCa and BPH patients with CD63-S-L, EpCAM-S-L (5′-labeled Cy5), and Tags (5′-labeled FAM) probes and analyzed the signals using NanoFCM. We found the high labeling efficiency of Tags (4.6%) in the PCa group and the low labeling efficiency (0.2%) in the BPH group, ruling out the false-positive fluorescence signals from the background (fig. S20). In addition, SORTER and RT-qPCR measurements were also carried out on the six PCa and six BPH plasma samples. All six miRNAs have significantly higher expression levels in PCa patient plasma samples compared to those of BPH controls by SORTER (fig. S21A). However, these results revealed a noticeably worse correlation between RT-qPCR and SORTER (Pearson’s r = 0.4911; fig. S21, B and C). This difference is attributed to SORTER’s high selectivity and sensitivity in detecting exosomal miRNA directly against a complex background. The RT-qPCR analysis is only sensitive to miRNA markers and not to proteins, thereby making it susceptible to interference from nontarget miRNAs, such as normal cell-derived exosomal miRNA. To further evaluate the potential interference of free miRNA on the exosomal miRNA assays, the SORTER analysis was performed in the PCa plasma sample before exosome lysis, while the DSN signal–amplifying analysis was conducted after exosome lysis. As illustrated in fig. S22, the fluorescent signal of miR-21 in exosomes was notably lower than that of bulk miR-21 (P < 0.001). This indicated that the SORTER effectively reduces nonspecific interference from contaminated vesicles and free miRNA. Collectively, our findings suggest the SORTER has improved selectivity and does not need lengthy isolation procedures, which may present the potential for expanding the use of exosomal miRNA analysis in clinical settings.

Clinical evaluation of SORTER for tumor-derived exosomal miRNA profiling

To evaluate the clinical application performance of SORTER in the diagnosis of PCa, we collected plasma samples from 30 patients (table S5) involving PCa (n = 20) and age-matched BPH (n = 10). We aimed to address (i) whether SORTER could recognize and analyze tumor-derived exosomes in plasma samples precisely and (ii) whether SORTER could improve the diagnostic performance of exosome-based liquid biopsy. Using these samples (0.2 μl for each plasma sample), we performed multiparametric miRNA profiling using the SORTER assay (termed CD63⁺EpCAM⁺ EVs in Fig. 5A). As a comparison, the single-target (CD63 or EpCAM)–guided fusion for miRNA profiling of CD63⁺ or EpCAM⁺ EVs was also conducted in the clinical cohort (Fig. 5, B and C).

Fig. 5. Clinical evaluation of SORTER for tumor-derived exosomal miRNA profiling.

(A to C) Schematic illustration of the miRNA analysis in the CD63⁺ (A), EpCAM⁺ (B), and CD63⁺EpCAM⁺ (C) EV subpopulations. The identification of the CD63⁺ or EpCAM⁺ EV subpopulation was performed by single-target recognition of CD63 or EpCAM protein on a single-particle membrane, and their miRNA analysis was achieved by guided fusion of Lipo@Au NFs and CD63⁺ or EpCAM⁺ EV subpopulation. (D) Heatmap of unsupervised hierarchical clustering (Pearson correlation, average linkage) of six miRNAs expression levels in CD63⁺, EpCAM⁺, and CD63⁺EpCAM⁺ EVs for distinguishing patients with PCa (n = 20) from BPH controls (n = 10). The signal intensities were averaged over triplicate measurements of each sample and normalized by min-max normalization after the background subtraction. (E to G) Correlation matrix of the expression profiles for the six miRNAs in CD63⁺ (E), EpCAM⁺ (F), and CD63⁺EpCAM⁺ EVs (G). (H to J) t-Distributed stochastic neighbor embedding (t-SNE) discriminated between patients with PCa and BPH controls using the six markers as the input in CD63⁺(H), EpCAM⁺(I), and CD63⁺EpCAM⁺ EVs (J). (K) ROC curves for the PCa signature (weighted sum of six markers by LDA) in CD63⁺, EpCAM⁺, and CD63⁺EpCAM⁺ EVs to differentiate between patients with PCa and BPH controls. (L) LDA score of the PCa signature in CD63⁺, EpCAM⁺, and CD63⁺EpCAM⁺ EVs for distinguishing patients with PCa from BPH controls. The LDA score for the binary classification was generated using a linear combination of chosen markers weighted by the respective coefficients. The P value was determined by a nonparametric, two-tailed Mann-Whitney U test. (M to O) Confusion matrix of the PCa signature in CD63⁺ (M), EpCAM⁺ (N), and CD63⁺EpCAM⁺ EVs (O). All statistical analyses were performed at 95% CIs.

By unsupervised hierarchical clustering analysis, we investigated whether a panel of six miRNAs for each patient exhibited mutually exclusive or similar expression patterns in different EV subpopulations (Fig. 5D). The expression heatmap showed that the abundance of each miRNA has considerable heterogeneity in different EV subpopulations for differentiating PCa and BPH. The heterogeneous expression profile of miRNAs in the CD63⁺, EpCAM⁺, and CD63⁺EpCAM⁺ EV subpopulations was separable into two different unsupervised classes. Six miRNAs (miR-222, miR-1290, miR-182, miR-21, miR-221, and miR-10b) were up-regulated in PCa compared to BPH plasma samples. These comparative analyses revealed the heterogeneity among overlapped EV subpopulations and corroborated the validity of the analytical data obtained by SORTER. Subsequently, pairwise comparisons of six miRNAs in different EV subpopulations were shown in Fig. 5 (E to G). Six miRNAs do not correlate strongly with each other in these EV subpopulations, which drives us to the combination of multiple markers for the accurate diagnosis of PCa. To explore the capacity of our method for PCa diagnosis, t-distributed stochastic neighbor embedding (t-SNE) is applied to discriminate between PCa and BPH (Fig. 5, H to J). Compared with CD63⁺ and EpCAM⁺ EVs, CD63⁺EpCAM⁺ EVs show a smaller overlap between the two patient groups. To further improve the diagnostic performance of exosome-based liquid biopsy in differentiating PCa and BPH groups, we harnessed linear discriminant analysis (LDA) to compile all miRNA profiles. Using receiver operating characteristic (ROC) analyses, we determined sensitivity, specificity, and accuracy for each marker individually (fig. S23 and table S6) and also in the PCa signature (weighted sum of six markers by LDA; Fig. 5K). We observed that no single marker achieved sufficiently high sensitivity and specificity. At the same time, the multiparametric combination improved the performance of molecular phenotyping of exosomes for cancer diagnosis. Specifically, the PCa signature in CD63⁺EpCAM⁺ EVs showed the best diagnostic performance with 1.000 [95% confidence interval (CI): 1.00 to 1.00] area under the curve (AUC) compared with CD63⁺ EVs [with 0.820 AUC (95% CI: 0.664 to 0.976)], and EpCAM⁺ EVs [with 0.970 AUC (95% CI: 0.889 to 1.00)]. Furthermore, the assessment of our method for each marker individually and the PCa signature in differentiating the BPH controls and patients with PCa were shown in Fig. 5L and fig. S24. Compared with CD63⁺ EVs (nonparametric, two-tailed Mann-Whitney U test, P = 0.0038) and EpCAM⁺ EVs (P = 2.0 × 10⁻⁶), the LDA scores of PCa signature in CD63⁺EpCAM⁺ EVs (P = 6.7 × 10⁻⁸) were significantly different between the PCa and BPH groups. The classification results of marker combinations were further presented as confusion matrices (Fig. 5, M to O). The PCa signature in CD63⁺EpCAM⁺ EVs shows an extremely high sensitivity of 100%, specificity of 100%, and accuracy of 100% for distinguishing between PCa from BPH compared with CD63⁺ EVs (with 50.0% sensitivity, 90.0% specificity, and 76.7% accuracy) and EpCAM⁺ EVs (with 90.0% sensitivity, 100% specificity, and 96.7% accuracy). These results demonstrated that our SORTER can improve the diagnostic detection performance of exosome-based liquid biopsies.

Clinical diagnosis of PCa on SORTER

To assess the diagnostic adaptability of the SORTER approach, we collected plasma from 74 patients participating in a clinical cohort (table S3), including nPCa (n = 27), metastatic PCa (mPCa, n = 20), and BPH (n = 27), of which four of seven plasma samples were randomly assigned to the training cohort. On the basis of the SORTER approach to miRNA markers, the training cohort was studied first to generate the discriminant function model, which was then used to classify the patients in the validation cohort.

We first analyzed plasma from a training cohort of 13 nPCa, 11 mPCa, and 18 BPH patients. Figure 6A summarizes the abundance of six miRNAs for each subject in a training cohort. Each miRNA expression in CD63⁺EpCAM⁺EVs has considerable heterogeneity for differentiating PCa and BPH groups. The diagnostic metrics of individual markers or marker combinations were assessed using ROC curve analyses (Fig. 6, B and C, and table S5). For LDA-based ROC studies, the posterior probabilities from this binary classification were used as the sole test variable. Among six miRNA markers, no single marker achieved sufficiently high sensitivity, specificity, and accuracy. The combination of the six markers comprising the PCa signature (1.00 AUC, 100% sensitivity, 100% specificity, and 100% accuracy) afforded better diagnostic ability than individual markers in the training cohort. We next correlated PCa signature analyses of CD63⁺EpCAM⁺ EVs to serum PSA in patients with PCa (Fig. 6D). The PCa signature was not correlated with PSA in the training set (r = 0.2084; P = 0.1854). In the training cohort, 100% of BPH patients (18 of 18) showed an increased concentration of PSA (>4 ng ml⁻¹, the threshold value used in the clinic cohort). In contrast, only 0% of patients with PCa (0 of 24) had a low PCa signature value (>0.505, the threshold value was obtained using Youden’s index based on the training cohort). Figure 6 (E and F) depicts the assessment of our method for detecting the BPH controls and two subgroups of nPCa and mPCa patients. We observed an overall substantial increase in PCa signature [Kruskal-Wallis one-way analysis of variance (ANOVA) analysis, P = 1.6 × 10⁻⁸] with progressive disease stages when compared to individual markers. To further characterize the effectiveness of our method to discriminate subgroups, we plotted the scores of each subject for the first two canonical variables computed from the discriminant analysis (Fig. 6G). It was visualized that the training samples were classified into three groups with notable separation among the patient groups at progressing disease stages. The binary classification results of individual markers or marker combinations were further presented in table S7 and fig. S25A. The PCa signature shows an extremely high sensitivity of 100%, specificity of 100%, and accuracy of 100% for distinguishing between PCa and BPH in the training cohort. All nPCa and mPCa cases in the training cohort were correctly detected, achieving an overall accuracy of 100% (95% CI, 100.0 to 100.0%; Fig. 6H).

Fig. 6. SORTER for differentiation of nPCa, mPCa, and BPH in a training cohort.

(A) Heatmap showing the abundance of the six miRNAs in a training set involving age-matched patients with BPH (n = 18), mPCa (n = 11), and nPCa (n = 13). The signal intensities were averaged over triplicate measurements of each sample and normalized by min-max normalization after the background subtraction. (B and C) ROC curves of the individual markers (B) and PCa signature (C) for PCa diagnosis. (D) Correlation of the PCa signature with serum PSA to differentiate nPCa/mPCa patients and BPH controls in a training cohort. The dashed line represents the threshold values for positivity (serum PSA, 4 ng ml⁻¹; PCa signature, 0.505). (E and F) Levels of the individual miRNA marker (E) and PCa signature (F) by SORTER approach at progressing disease stages. The overall and group pair P values were determined using Kruskal-Wallis one-way ANOVA with post hoc Dunn’s test for pairwise multiple comparisons. (G) LDA plot using six miRNAs across nPCa, mPCa, and BPH patients. (H) Confusion matrix showed that the PCa signature had an accuracy of 100% across nPCa, mPCa, and BPH patients. All statistical analyses were performed at 95% CIs.

The SORTER approach was further applied to an independent validation set of 32 age-matched plasma samples collected from 9 BPH controls, 14 nPCa, and 9 mPCa patients. Figure 7A summarizes the performance of the indicated miRNAs for each patient. Analyzing the heatmap of each marker expression in CD63⁺EpCAM⁺EVs once again showed considerable heterogeneity for differentiating PCa and BPH. The validation set data were then input into the trained LDA model to test its validity in cancer diagnosis. Across the validation cohorts (Fig. 7, B and C), the PCa signature (1.00 AUC, 100% sensitivity, 100% specificity, and 100% accuracy) once again showed excellent diagnostic performance for cancer diagnosis when compared with a single marker. We also studied the correlation between PCa signature analyses and serum PSA in patients with PCa and BPH (Fig. 7D). The PCa signature (r = 0.2018; P = 0.2681) was not correlated with PSA. In the validation cohort, 88.9% of BPH patients (8 of 9) showed an increased concentration of PSA (>4 ng ml⁻¹), whereas only 0% of patients with PCa (0 of 13) had a low PCa signature value (>0.505). The validation cohort data were then examined for identification of the PCa progressive stages using various statistical methods. With the Kruskal-Wallis one-way ANOVA and post hoc Dunn’s multiple comparisons tests (Fig. 7, E and F), it was shown that the PCa signature (P = 1.2 × 10⁻⁶) significantly improved at discriminating the three subject groups when compared to individual markers. An LDA plot using the feature set composed of a combination of six miRNA markers shows a small overlap across nPCa, mPCa, and BPH patients (Fig. 7G). As shown in fig. S25B, the PCa signature was able to discriminate PCa from BPH with a sensitivity, specificity, and accuracy of 100% in the validation cohort. In addition, only two mPCa and one nPCa were misclassified, leading to an overall accuracy of 90.6% (95% CI, 75.0 to 98.0%; Fig. 7H). Collectively, these comparative results further showed that our SORTER approach had potential adaptability for molecular phenotyping and improved diagnostic performance for early-stage cancer.

Fig. 7. Validation of the SORTER approach for PCa diagnosis.

(A) Heatmap showing the abundance of the indicated miRNAs in a validation set involving age-matched patients with 14 nPCa, 9 mPCa, and 9 BPH. The data processing was similar to the training cohort (Fig. 6A). (B and C) ROC curves for the individual markers or marker combinations to differentiate between patients with PCa and BPH controls in a validation cohort. (D) Correlation of the PCa signature with serum PSA to differentiate PCa and BPH. The threshold values were similar to those of the training cohort (Fig. 6D). (E and F) Levels of the individual miRNA marker (E) and PCa signature (F) by SORTER approach at progressing disease stages. The overall and group pair P values were calculated similarly to the training cohort (Fig. 6, E and F). (G) LDA plot of the first two canonical variables derived from the discriminant analysis of the training cohort. (H) Confusion matrix showed that the PCa signature had an overall accuracy of 90.6% across nPCa, mPCa, and BPH patients. All statistical analyses were performed at 95% CIs.

DISCUSSION

Exosomes are emerging as the most promising targets in tumor liquid biopsy and thus have attracted tremendous research interest, and large amounts of related studies have been reported in the past decade. However, the clinical application of exosome-based liquid biopsy falls far behind laboratory research, which is mainly attributed to the complex contaminated vesicles and heterogeneous subtypes of exosomes in biofluids. The ensemble detection of protein and RNA biomarkers from membrane vesicles or particles diminishes the accuracy of exosome-based liquid biopsy. Although accumulated evidence has demonstrated that tumor-specific exosome subsets are highly associated with cancer development, invasion, and metastasis, it is still technically challenging to recognize and sort tumor-derived exosomes selectively and analyze the carried cargo precisely (61, 62).

Regarding subpopulation differentiation, the conventional techniques for tumor-derived exosome identification/isolation are primarily univariate, which may be interfered with by coexisting components with overlapping features in their composition and thus lack tumor specificity (63, 64). In this study, we first presented a combinatorial SORTER methodology that incorporates dual-surface-protein synergistic recognition to precisely label tumor-derived exosomes in unextracted plasma samples. Our approach used two allosteric aptamers of exosomal marker CD63 and tumor marker EpCAM to create a unique orthogonal identity barcode on the tumor-derived exosome surface, permitting targeted recognition and controlled fusion of complementary DNA–anchored liposome probes. Distinct from the conventional identification/isolation techniques, we used the selectivity of SORTER to minimize the interference of nonspecific vesicles and free molecules that enable the rapid and precise recognition of tumor-derived exosomes in plasma samples, avoiding lengthy pre-isolation/purification procedures and minimizing the nonspecific interference of contaminated vesicles. It is worth noting that not all the tumor-derived exosomes express both CD63 and EpCAM markers, thus only a portion of tumor-derived exosomes were labeled and analyzed in this work. This phenomenon raises a higher requirement for the sensitivity of exosomal miRNA detection.

For quantitative analysis of exosomal miRNAs, most of the existing techniques (i.e., RT-qPCR) are prone to interference by nonspecific vesicles and free miRNAs (65). Moreover, femtomolar sensitivity is essential for in situ miRNA profiling of exosomes, where the concentrations of miRNAs are deficient (roughly 1 copy/10⁶ EVs to 1 copy/1 EV) (20). In this study, the SORTER incorporates multiple parallel processes, including exosome recognition, importing probes, miRNA signal transduction, and amplification, allowing for fast, sensitive, and multiparametric profiling of miRNAs in tumor-derived exosomes directly from clinical plasma samples. Compared to previously reported techniques, the SORTER affords three noteworthy advantages as follows: First, the SORTER does not involve any exosome lysis and RNA preparation procedures, which markedly simplifies the experimental operation, reduces the processing time, bypasses the dilution of low-abundance miRNAs, and prevents sample loss during exosome lysing and RNA extraction procedures. Second, the technique incorporates dual-surface-protein synergistic recognition to sort and analyze tumor-derived exosomes precisely, a small yet important subpopulation of extracellular vesicles, improving the diagnostic and prognostic accuracy of exosomal miRNA-based liquid biopsy. Third, this technique requires only small-volume plasma samples (~0.2 μl), a short assay time of ~2 hours by skipping lengthy pre-isolation/purification processes, and is high throughput compatible with 96/384-well plate, opening a way for noninvasive and high-accuracy cancer screening and progress monitoring.

Concerning exosomal miRNA profiling for liquid biopsy applications, the SORTER enables the capture of the information of tumor-derived exosome (CD63⁺EpCAM⁺ EVs) subpopulation in complex clinical scenarios, which is often missed in other approaches and only accessible via single-exosome miRNA analysis. Although we cannot spatially determine the fusion proportion of each liposome probe to an individual tumor-derived exosome, statistically, the multiparametric miRNA profiling by SORTER still reflects the compositional nature of the studied tumor-derived exosomes. Here, we made a comparative study of single- and dual-surface-protein–mediated orthogonal fusion to direct subpopulation differentiation and miRNA profiling directly from plasma samples in a clinical cohort involving 20 patients with PCa and 10 BPH controls. Our comparative results revealed that SORTER could precisely recognize and analyze tumor-derived exosomes in plasma samples. We also evaluated the diagnostic adaptability of SORTER for PCa diagnosis and stratification in a clinical cohort (n = 74). Combining six miRNA markers tested (PCa signature), the SORTER was able to discriminate PCa from BPH with a sensitivity, specificity, and accuracy of 100% in the training and validation cohorts. These comparative results showed that our SORTER approach had potential adaptability for molecular phenotyping and improved diagnostic performance for cancer.

The technology of SORTER exhibited the unique advantages of being rapid, noninvasive, avoiding separation, scalability, and high accuracy in assessing the miRNA profiles of tumor-derived exosomes. The SORTER is capable of high-throughput analysis in clinical studies, and its accuracy has been improved by integrating machine learning into data processing. With its robust ability to differentiate tumor-derived exosomes in clinical plasma samples, the SORTER could be readily expanded to measure, beyond miRNAs, other diverse molecules (e.g., internal proteins, lipids, and metabolites). The barcoding capacity of this technology can be readily enhanced by designing allosteric-aptamer probes, allowing for measuring other EVs of molecular subtypes (e.g., different cell origins). The SORTER will contribute to the advancement of the liquid biopsy field and be a clinically feasible tool for disease screening, classification, and progress monitoring in complex clinical settings.

MATERIALS AND METHODS

Cells and culture conditions

The human PCa cell lines LNCaP, DU145, and PC-3 were purchased from the China Center for Type Culture Collection (Shanghai, China). The human prostatic hyperplasia cell line BPH-1 was bought from Yaji Biotechnology Co. Ltd. (Shanghai, China). DU145 cell was maintained in an exosome-depleted Dulbecco’s modified Eagle’s medium. LNCaP, PC-3, and BPH-1 cells were maintained in an exosome-depleted RPMI 1640 medium. All cell line media were supplemented with 10% vesicle-depleted fetal bovine serum (FBS) and 1% penicillin-streptomycin in a humidified incubator with a constant temperature (37°C) and 5% CO₂. Vesicle-depleted FBS was prepared by centrifuging FBS for 10 hours at 100,000g and then passing the supernatant through a 0.22-μm filter.

Exosome isolation by standard differential ultracentrifugation

The cell culture media were collected for cell-derived exosome isolation until cells reached 80 to 90% confluency. The ultracentrifugation procedure was conducted using a TLA-100.3 fixed-angle rotor in an Optima TL-100 ultracentrifuge (Beckman Coulter). The medium was packaged into six ultracentrifuge tubes with a weight difference between every pair of tubes smaller than 0.02 g. First, the collected media were centrifuged at 3000g for 20 min at 4°C to remove cells and large debris. Then, the resulting supernatant was centrifuged at 16,500g for 45 min at 4°C to pellet microvesicles. After that, the supernatant was ultracentrifuged at 100,000g for 2 hours to collect exosomes. The resulting pellets were resuspended and washed with filtered PBS, followed by another centrifugation of 100,000g for 2 hours at 4°C. Last, the resulting exosomes were resuspended in filtered PBS and stored at −80°C for further use.

Clinical samples

Clinical samples of patients with PCa and BPH were obtained from Renji Hospital of Shanghai Jiaotong University School of Medicine. All relevant ethical regulations were complied with. All samples (n = 74) were anonymized and collected before treatment, and only the age of the PSA and pathological diagnosis were recorded. In the clinical cohorts, the patients with PCa (n = 47) had been diagnosed, and the BPH controls (n = 27) had no history of cancer before sample collection. Relevant information on the human participants in the clinical trial was presented in table S5.

Before use, the blood samples were centrifuged at 3000g for 10 min to obtain cell-free plasma. Then, the human plasma was centrifuged at 4°C at 10,000g for 20 min to remove large vesicles. The plasma samples were filtered with a 0.45-μm filter into an eppendorf tube and stored at −80°C for further use.

Synthesis and characterization of Au NFs

Au NFs were obtained in the following manner: Before use, citrate-capped Au NPs (13 nm ± 2 nm) were prepared according to a literature-reported method (66). Then, 5′-SH– and 3′-FAM–labeled DNA probes (25 μl, 10 μM) were mixed with a 100-μl Au NP solution. The mixture was then placed in a −80°C freezer for 2 hours, followed by thawing at room temperature. Following that, the resultant FAM-DNA–labeled Au NP solution was centrifuged (13,000 rpm, 8 min) to remove the free DNA before being dissolved in PBS buffer [10 mM (pH 7.4)] with 5 mM MgCl₂. Next, 10 μl of BSA (5%) was added to the FAM-DNA–labeled Au NP solution, which was then shaken at room temperature for 1 hour. Last, the resultant Au NFs solution was centrifuged (13,000 rpm, 8 min) to remove the free molecules and dissolved in tris-HCl buffer [10 mM (pH 7.4)] containing 10 mM MgCl₂ for further use. The hydrodynamic size and zeta potential of Au NPs and Au NFs were obtained by a Malvern 3000HS Zetasizer (Nano ZS, Malvern, UK). The absorption spectra of Au NPs and Au NFs were measured by NanoDrop One (Thermo Fisher Scientific) with a scanning range of 300 to 700 nm. The concentration of Au NFs was determined to be 15 nM based on the molar extinction coefficient of 2.7 × 10 ⁸ M⁻¹ cm⁻¹ at 520 nm. The loading density of DNA probes on each Au NFs was determined to be 75 nM based on the molar extinction coefficient of 2.33 × 10⁵ M⁻¹ cm⁻¹ at 260 nm.

Synthesis and characterization of liposome probes

Before use, the liposome solution (25 mg ml⁻¹) from 1,2-dioleoyl-sn-glycero-3-phosphocholine/1,2-dioleoyl-sn-glycero-3-phosphoethanolamine/cholesterol (2:1:1, molar ratio) was obtained according to a literature-reported method (67). Then, a 300-μl reaction mixture containing liposome (1 mg ml⁻¹), 0.2 × DSN buffer, 1 U of DSN, 20 U of RNase inhibitor, and 15 nM Au NFs was coextruded repeatedly 20 times through 200-nm polycarbonate porous membranes (Whatman NucleoporeTrack-Etched Membranes) using a mini-extruder (Avanti Polar Lipids). Next, 15 μl of cholesterol-labeled DNA tags (Tags, 10 μM) were incubated with the above liposome solution at room temperature for 1 hour. Last, the resultant liposome probes (Tags-Lipo@Au NFs) were purified by ultrafiltration to remove the free products and dissolved in tris-HCl buffer [10 mM (pH 7.4)] containing 10 mM MgCl₂ for further use.

Dual-surface-protein labeling on tumor-derived exosomes

For NanoFCM analysis of nine protein expressions of human PCa LNCaP, DU145, and PC-3 cell–, as well as prostatic hyperplasia BPH-1 cell–derived exosomes, 5 μl of exosome solution (5.0 × 10⁸ particles μl⁻¹) was incubated with 0.25 μM FAM-labeled aptamer probes in 100 μl of binding buffer [PBS with 0.5 mM MgCl₂ (pH 7.4)] for 2 hours at 4°C. After washing twice by ultrafiltration, the fluorescence intensity of exosomes was detected by Flow NanoAnalyzer N30 (NanoFCM Inc., Xiamen, China) according to the manufacturer’s instructions.

For NanoFCM or TIRFM analysis of CD63-S-L– and EpCAM-S-L–mediated orthogonal labeling on a single exosome surface, 5 μl of LNCaP Exo solution (5.0 × 10⁸ particles μl⁻¹) was incubated with 0.25 μM Cy5-labeled EpCAM-S-L and FAM-labeled CD63-S-L probes in 100 μl of binding buffer [PBS with 0.5 mM MgCl₂ (pH 7.4)] for 2 hours at 4°C. After washing twice by ultrafiltration, the fluorescence intensity of exosome was analyzed by NanoFCM or TIRFM. BPH-1 Exo was used as a negative control.

Zipper-like hybridization of the orthogonal barcode-anchored exosome and complementary DNA tags

For NanoFCM or TIRFM analysis of zipper-like hybridization on the tumor-derived exosome surface, 5 μl of LNCAP or BPH-1 Exo (5.0 × 10⁸ particles μl⁻¹) was incubated with 5 μl of 5 μM Cy5-labeled EpCAM-S-L and FAM-labeled DNA tags (FAM-Tags) and 5 μl of 5 μM nonfluorescent CD63-S-L probes in 100 μl of binding buffer [PBS with 0.5 mM MgCl₂ (pH 7.4)] for 2 hours at 4°C. After washing twice by ultrafiltration, the fluorescence intensity of exosome was detected by NanoFCM or TIRFM. The hairpin rCD63-S-L or rEpCAM-S-L probe (only the aptamer domain was replaced by a random sequence) was used as a negative control.

Dual-surface-protein-guided liposome probe fusion studies.

For the lipid-mixing FRET decrease assay, the dynamic process of orthogonal barcode-anchored LNCAP Exo (Orth-Exo) and Tags-Liposome (Tags-Lipo) membrane fusion was investigated using a standard FRET decrease assay. First, 100 μl of 1.0 × 10⁷ particles μl⁻¹ Lipo was incubated with 20 μM DiO and 20 μM DiI for 30 min at 37°C. After that, the DiI and DiO double-labeled Lipo (Lipo-DiI-DiO) were incubated with 5 μl of 5 μM Tags for 1 hour at room temperature, obtaining the DiI and DiO double-labeled Tags-Lipo (Tags-Lipo-DiI-DiO) products. The free products were removed by ultrafiltration at 13,000g for 20 min three times at each step. Last, 100 μl of 1× or 10 × 10⁷ particles μl⁻¹ Exo, 5 μl of 5 μM CD63-S-L, and 5 μl of 5 μM EpCAM-S-L were incubated with 100 μl of Tags-Lipo-DiI-DiO for 2 hours at 37°C, and then the mixture was measured using fluorescence spectrometer or fluorescence kinetic analysis. The stochastic fusion between Exo and Lipo-DiI-DiO was used as a negative control.

For TIRFM studies, 100 μl of 1.0 × 10⁷ particles μl⁻¹ LNCAP Exo was incubated with 20 μM DiO for 30 min at 37°C. After that, the prepared DiO-labeled Orth-Exo products were dissolved in 100 μl of PBS buffer [10 mM (pH 7.4)] containing 5 mM MgCl₂. Meantime, 100 μl of 1.0 × 10⁷ particles μl⁻¹ Lipo was labeled by DiI using the same method and then incubated with 5 μl of 10 μM Tags for 1 hour at room temperature, obtaining the Tags-Lipo-DiI products. The free products were removed by ultrafiltration at 13,000g for 20 min three times at each step. Last, 100 μl of DiO-labeled Orth-Exo, 5 μl of 5 μM CD63-S-L, and 5 μl of 5 μM EpCAM-S-L were incubated with 100 μl of DiI-labeled Tags-Lipo for 2 hours at 37°C, and the mixture was imaged using the TIRFM. The stochastic fusion of Exo-DiO and Lipo-DiI was used as a negative control.

For DLS studies, 100 μl of 1.0 × 10⁷ particles μl⁻¹ Lipo was incubated with 5 μl of 5 μM Tags for 1 hour at room temperature. Then, the free products were removed by ultrafiltration at 13,000g for 20 min three times. After that, the prepared Tags-Lipo products were dissolved in 100 μl of PBS buffer [10 mM (pH 7.4)] containing 5 mM MgCl₂. Last, 100 μl of 1.0 × 10⁷ particles μl⁻¹ LNCAP Exo, 5 μl of 5 μM CD63-S-L, and 5 μl of 5 μM EpCAM-S-L were incubated with 100 μl of Tags-Lipo products, and the mixture was measured at 37°C under different time intervals. The group of Exo and Lipo was used as negative controls. The hydrodynamic size distribution of vesicle samples was assessed using the Malvern Zetasizer (Nano ZS, Malvern, UK).

For TEM imaging, the fusion mixtures of Tags-Lipo@Au NFs and Orth-Exo (the preparation process was as described above) were added onto a 150-mesh formvar copper grid or indium tin oxide glass and incubated for 10 min. After washing with ultrapure water, the samples were treated with 2.5% glutaraldehyde in PBS for 30 min and then rinsed for 15 min to fix the particles. Next, the samples were negatively stained with 2% uranyl acetate for 10 min and rinsed for 10 min with water. Samples were dried and visualized using TEM (JEM-1400 Plus, Japan) imaging. Orth-Exo and Tags-Lipo@Au NFs were used as negative controls.

Profiling of tumor-derived exosomal miRNA

To clarify the SORTER approach for tumor-derived exosomal miRNA analysis, the LNCAP Exo solution was prepared by serial dilutions of the stock solution in 1 ml of PBS or 100-fold diluted EV-depleted plasma. Specifically, 5 μl of 1 μM CD63-S-L, 5 μl of 1 μM EpCAM-S-L, and liposome probes (20 μl, 1.0 × 10⁷ particles μl⁻¹) were incubated with 20 μl of the prepared exosome solution for 2 hours at 37°C. Last, the mixture was measured using a multidetection microplate reader. The group of exosomes and liposomes was used as negative controls.

To achieve tumor-derived exosomal miRNA analysis in clinical plasma samples directly, 5 μl of 1 μM CD63-S-L, 5 μl of 1 μM EpCAM-S-L, and liposome probes (20 μl, 1.0 × 10⁷ particles μl⁻¹) were incubated with 20 μl of 100-fold dilution plasma samples for 2 hours at 37°C. Last, the mixture was measured using a multidetection microplate reader. The group of exosomes and liposomes was used as negative controls.

Statistical analyses

Mean, SD, and LOD were calculated with standard formulas. Significance tests were obtained via a two-tailed Student’s t test. The intensities of individual miRNA markers detected by the SORTER approach used minimum-maximum (min-max) normalization. The PCa signature was calculated as the weighted sum of the normalized intensities of six miRNA markers by LDA, respectively. For binary classification, P values for pairwise comparisons were performed using a nonparametric, two-tailed Mann-Whitney U test. For ternary classification, the overall and group pair P values were determined using Kruskal-Wallis one-way ANOVA with post hoc Dunn’s test for pairwise multiple comparisons. Hierarchical clustering was performed for the analysis markers using the “pheatmap” package in the R language. ROC analyses were constructed for individual markers or marker combinations to evaluate the AUC, sensitivity and specificity, and accuracy of cancer diagnosis. The training cohort (n = 42) was first analyzed to generate the discriminant function model, which was used to classify the patients in the validation cohort (n = 32). The optimal cutoff points were selected using Youden’s index based on the training cohort, which was applied to evaluate the sensitivity, specificity, and accuracy of validation cohort. The t-SNE was performed using six markers as the input for binary classification (PCa and BPH). All statistical analyses were performed at 95% (P < 0.05) CIs using OriginPro 2018, GraphPad Prism (v.8.0), and R software (version 4.1.2).

Supplementary Materials

This PDF file includes:

Figs. S1 to S25

Tables S1 to S7