Sizing extracellular vesicles using membrane dyes and a single molecule-sensitive flow analyzer

Abstract

Extracellular vesicles (EVs) are membranous particles released by most cells in our body, and which are involved in many cell-to-cell signaling processes. Given the nanometer sizes and heterogeneity of EVs, highly sensitive methods with single-molecule resolution are fundamental to investigating their biophysical properties. Here, we demonstrate the sizing of extracellular vesicles (EVs) using a fluorescence-based flow analyzer with single-molecule sensitivity. Using a dye that selectively partitions into the vesicle’s membrane, we show that the fluorescence intensity of a vesicle is proportional to its diameter. We discuss constraints in sample preparation which are inherent to sizing nanoscale vesicles with a fluorescent membrane dye, and propose several guidelines to improve data consistency. After optimizing staining conditions, we were able to measure the size of vesicles in the range ~35–300 nm, covering the spectrum of EV sizes. Lastly, we developed a method to correct the signal intensity from each vesicle based on its travelling speed inside the microfluidic channel, by operating at a high sampling rate (10 kHz) and measuring the time required for the particle to cross the laser beam. Using this correction, we obtained a 3-fold greater accuracy in EV sizing, with a precision of ±15–25%.

Graphical Abstract

Extracellular vesicles (EVs) are lipid bilayer-enclosed structures generated and released by almost all cells in our body. A clear classification of EVs based on morphological aspects and/or biological features has been hampered by the high heterogeneity of EV samples (e.g. in sizes, lipid composition, and intravesicular cargo), which vary depending on the biogenesis, biological source, and the sample-preparation process. Thus, a bewildering mix of EV naming schemes,¹ often study-specific, can be found in the literature. Therefore, the EV community has created an in-depth guideline for researchers to promote clarity and standardization of EVs results.²

To facilitate characterization of EV subsets, researchers have adapted techniques used for cell purification;^3,4 however, a complete classification of EVs biology remains a challenge. Methods with single-molecule sensitivity might represent the best strategy to identify and quantify EV-specific biomarkers on a single vesicle level,^5–7 allowing the discovery of common biophysical patterns within heterogenous EV samples.

Both side scattering- and fluorescence-based flow cytometry have been applied to studying nanoscale vesicles, often using custom-built systems with high detection sensitivity.^8–10 Such systems can handle small sample volumes (~10 μL) and require minimal sample pretreatment, preserving the EVs’ native state.^11,12 Furthermore, they can be coupled with ultrafast high-throughput sorting devices for biomarker-based EV isolation.¹³ Given the heterogeneous nature of EVs and how these vesicles can be categorized based on size, it is useful to establish a correlation between EV dimensions and protein content, and where EV size may be a useful parameter for predicting the biological function of EV subpopulations. Here we address the question of how to size EVs using a membrane dye and a single molecule-sensitive flow analyzer. Starting with dye selection, we analyzed criteria that make a membrane dye suitable for EV sizing in single-molecule flow analysis by comparing the performance of four common types of probes (Fig. S1 and S2). We then demonstrate how the small dimensions of EVs introduce constraints in selecting operational conditions for vesicle labeling. Lastly, we describe a method for enhancing the accuracy and reproducibility of EV size estimates by accounting for the velocity distribution of vesicles flowing within a micron-sized channel.

EXPERIMENTAL SECTION

Screening membrane dyes for vesicle labeling.

For selection of membrane dye, we used big liposomes (B-Lipo, from Creative Biostructure; DOPC:CHOL 55:45 mol/mol) whose distribution was centered at 130 nm (by dynamic light scattering (DLS), see Fig. S3). Di-8-ANEPPS was purchased from Cayman Chemical Co Inc., FM1–43 from Thermo Fisher, Dioc14(3) from PromoCell, and Membright-488 from Idylle. A stock solution of liposomes (50 mM lipid, 10% sucrose, 20 mM HEPES, pH 7.5) was diluted 3×10⁴-fold in 1 mL 0.1X PBS, pH 7.4, into a 1.5 mL cuvette for acquisition of bulk emission. For each organic dye, a stock solution in DMSO was diluted in EtOH to a final concentration of 0.5, 2, 5, 10, and 25 μM. Then 10 μL was quickly added to the cuvette containing liposomes and the solution was gently mixed three times with a micropipette. We followed the bulk emission (λ_ex = 470 nm for Dioc14(3) and Membright-488 and λ_ex = 488 nm for Di-8-ANEPPS and FM1–43, see Fig. S1–S2) for each sample over time using an LS55 fluorescence spectrometer (PerkinElmer), acquiring spectra over time after staining. After two hours of equilibration, we further diluted each liposome sample 5x into a 0.5% PEG solution (0.1X PBS, pH 7.4) performed flow analysis. Following a similar procedure, we prepared solutions of dye in buffer (10 μL of 25 μM dye solution injected into 1 mL of 0.1X PBS, pH 7.4) for each of the four organic probes. DLS analysis was performed using a Malvern Zetasizer (1 μL of stained liposomes was added to 1 mL of 0.1X PBS).

Synthesis and staining of artificial vesicles, and optimization of EVs staining.

We prepared small-size liposomes (S-lipo; with sizes distribution centered at 47.4 ± 20.2 nm) and medium-size liposomes (M-lipo; with sizes distribution centered at 80.0 ± 46.7 nm) from B-lipo vesicles, adapting previously reported ultrasonication and repetitive freeze-thaw cycle techniques.^14,15 M-lipo were obtained by diluting 20 μL of B-lipo in 200 μL buffer (PBS-0.1X, pH 7.4) inside a 10×75 mm glass tube and performing repetitive cycles of sample freezing (3 min in liquid nitrogen) and thawing (2 min in a 60 °C water bath). After 12 cycles, the sample was diluted with 1.8 mL of buffer (0.56 mM final lipid concentration). S-lipo vesicles were prepared by diluting 30 μL of B-lipo in 300 μL buffer (PBS-0.1X, pH 7.4; final lipid concentration, 5 mM) and ultrasonicating the liposome solution for 10 min using a Branson Sonifier 450 (20% duty cycle, power step 1).

For both S-lipo and M-lipo vesicles, we selected the optimal concentration of dye for staining by titrating liposomes with Di-8-ANEPPS. Specifically, M-lipo were diluted in PBS-0.1X buffer (1.67×10³-fold, to a final lipid concentration of 0.33 μM), and 10 μL of a 2, 3, 4, or 5 μM solution of dye in EtOH was added. Following the same procedure, S-lipo were diluted 4×10⁴-fold in buffer (final lipid concentration, 0.13 μM) and stained with a 1, 5, 7, 12, or 25 μM solution of dye in EtOH. After 2 h of equilibration, based on results from bulk emission (see Fig. S4), we chose the 4 μM sample to calculate M-lipo sizes by flow. For the S-lipo sample, the bulk emission was too low to be detected; therefore, we performed flow analysis for all of the samples prepared and calculated S-lipo sizes from the 5 μM sample, which gave the highest intensity values (see Fig. S4).

The effects of experimental conditions on SE staining were evaluated systematically, by varying the following parameters one at a time (see SI for detailed procedure and Fig. S5), and comparing the final size distributions from the flow data: (a) equilibration time, (b) concentration of dye given a fixed SE concentration, (c) concentration of SE given a fixed dye concentration, (d) EtOH concentration, (e) reaction volume for SE staining and (f) dilution of both dye and SEs given a constant dye/vesicle. The flow-derived size distributions were fitted to a bell-shaped distribution using the Damped least-squares method, and the median of each distribution was used as measurement of central tendency to compare the effect of changes in experimental conditions on size estimation. Flow-derived sizes of SEs were estimated from the sample stained under optimal conditions, namely 6 μL of SE stock solution (2.85×10¹¹ particles/mL) was diluted in 1 mL of buffer (PBS-1X, pH 7.4) and stained with 10 μL of a 6 μM Di-8-ANEPPS solution. Flow analysis was performed after 2h of equilibration.

Isolation of exosomes from seminal fluids (SEs).

Semen samples were obtained from the HIV Vaccine Trials Clinic in Seattle. The procedure for exosome isolation was based on a published protocol¹⁶ and was approved by the Institutional Review Boards of the University of Washington and the Fred Hutchinson Cancer Research Center. Seminal plasma was isolated from the semen sample by using a series of centrifugation steps and a final filtration through a 0.22-μm syringe filter. After ultracentrifugation over a sucrose gradient, the 30% and 25% sucrose cushions containing SEs were pooled and washed by centrifugation through an Amicon Ultracel 100-kDa cellulose centrifugal filter. Finally, SEs were purified twice using a size exclusion chromatography column (iZON-70nm) to deplete proteins in solution, and the samples were collected in a 1.5 mL final volume (SE concentration = 2.85×10¹¹ particles/mL based on nanoparticle tracking analysis). The sizes distribution of isolated SEs was centered at 70.2 ± 36.1 nm.

Optical setup, flow data acquisition and analysis.

The platform used (Figure 1a) was an improvement of a previously described setup.^9,10 Briefly, the platform consisted of a Nikon TE2000 inverted microscope with a 60× NA 1.45 objective, with two lasers (488 and 640 nm, Coherent Obis) for fluorescence illumination. Cylindrical optics were used to shape the laser beam prior to entering the microscope. The width of laser beam was in the range 0.5 – 1.5 μm (depending on the specific wavelength) along x-axis, and ~25 μm in the y-direction (orthogonal to flow). The same objective was used to collect fluorescence signals. A piezo stage (APB302, Thorlabs) was used to position and focus the sample. A custommade rectangular pinhole was placed in the image plane. D mirrors were placed after the pinhole to split fluorescence to separate avalanche photodiodes (APD; SPCMAQR-14, PerkinElmer, Fremont, CA). An aspherical lens and bandpass filter (600/50 nm, 680/40 nm, Chroma, Rockingham, VT) were inserted in front of each APD to collect fluorescence signals. The 2 μm-wide channel was made by assembling molded PDMS and a #1.5 glass coverslip, after cleaning the two components with Milli-Q water and EtOH followed by treatment with oxygen plasma. For analysis of vesicle sizes, 10 μL of each sample was loaded into the microchip inlet reservoir (using gravity-driven flow) and the flow trajectories were collected at a volumetric flow rate in the range of 5 – 50 pL/s for 3 minutes (~0.9 – 9 nL of total sample volume analyzed) to ensure enough vesicle counting events for statistical relevance.

Figure 1. Selection of organic dyes for EV staining.

a, Schematic of a single molecule-sensitive flow analyzer. b-e, Intensity distributions of liposomes labeled using four dyes at different dye concentrations: 5 nM (violet), 20 nM (orange), 50 nM (red), 100 nM (blue), and 250 nM (black). The x-axis is in logarithmic scale, and counts (vesicle detection events) are plotted against fluorescence intensity (a.u.).

Analysis of EVs sizes in flow was performed using a custom algorithm developed in MATLAB (see Fig. S6), and can be summarized in three distinctive phases of data processing: Phase 1: A peak-finding function analyzes the entire flow trace and identifies peaks that possess a bell shape, described by five consecutive intensity-points with at least three points (the peak center and the points immediately before and after it) greater than 3-fold the standard deviation of the noise. Phase 2: Each localized peak is fitted to a Gaussian-like function (Eq. 1) through an iterative least-squares regression.

$$g\left(t\right)=b+A{e}^{-\frac{{\left(t-\mu \right)}^2}{2{\sigma }^2}}$$ Eq. 1

A is the peak height, σ refers to the peak width, t and μ correspond to the ith time bin and the bin at the peak center, respectively. To ensure enough points for fitting slower vesicles, each peak is analyzed as a subset of nine points in time: the point corresponding to the maximum intensity plus four points before and after. Only those peaks that show a good fit (R² ≥ 0.90) are selected for the next step. This allows rejection of vesicles that do not present a bell-shaped peak for reasons such as interaction with the microchannel walls or formation of clusters. Phase 3: The estimated vesicle’s passing time (PT_e) across the laser beam is calculated as the width at the base of the Gaussian curve (6σ).

The integrated fluorescence (I_e) of each individual vesicle was calculated according to Eq. 2:

$$I_e={\sum }_{i=1}^n{I}_i-nb$$ Eq. 2

where I_i is the intensity in the ith time bin, b is the baseline (median of intensity values from the whole trajectory) and n are the number of points in time when the vesicle is in the sample volume (equal to the base width of the peak). Then, each I_e value was divided by the corresponding vesicle PT_e to give the normalized integrated intensity (I_N), and transformed into size following Eq. 4. Distributions of flow-derived sizes were obtained from ~6400, 4050, and 7200 vesicle-detection events for M-lipo, S-lipo, and SE vesicles, respectively, by using a logarithmic binning, with width of the ith bin = e^a(i+1)+b – e^ai+b, where a = 0.1467 and b = 2.9. Then, size distributions were log-normally fitted, to estimate the corresponding center and dispersion (namely median and SD). The intensity-to-size conversion factor (α) was obtained iteratively until the centers (medians from the fitting) of both flow-derived and cryo-electron microscopy (Cryo-EM) distributions matched.

Size distribution from Cryo-EM data.

Absolute sizes of S-Lipo, M-Lipo and SEs were measured by Cryo-EM analyses. EM images (resolution, 0.51 nm/pixel) were collected using a Spirit T12 electron microscope (Thermo-Fisher) operating at 120kV with a US4000 CCD camera (Gatan, Inc.). Grids used for data collection were either C-flat (Protochips, Inc.) or Lacey (Ted Pella, Inc.) grids. Following glow-discharging treatment, the carbon grids were loaded with 3 μL of vesicle sample prestained with the optimal concentration of Di-8-ANEPPS. After 3 min of waiting time to increase the number of vesicles within the grid’s holes, the solution was blotted off and the sample was quickly plunged into liquid ethane. Due to the negatively charged surface, the capturing of semen exosomes onto the negatively-charged carbon grids was inefficient. Therefore, the procedure for sample preparation was adapted by pre-coating the carbon grid surface with 0.1% (w/v %) polylysine.¹⁷ Cryo-EM size distributions were obtained in the same way of flow-derived ones. To obtain sizes distribution, over 100 Cryo-EM images were collected for each sample and the radii of ~3600 M-lipo, 1100 S-lipo, and 990 SE vesicles were measured.

Calculation of size estimate uncertainty.

The number of dyes-per-vesicle and its corresponding uncertainty (n_d and u(n_d)) were estimated via complementary experiments of single-molecule photobleaching (using total internal reflection fluorescence (TIRF) microscopy), and cross-correlation flow analysis of CD63⁺ SEs (CD63 antibody is a well-established biomarker for SEs),¹⁶ by using a previously described platform (see SI: “Uncertainty in EV size estimates”, and Fig. S7 and S8).^18,19 The uncertainty on passing time u(PT_e) was calculated from the confidence intervals on PT_e estimation, whereas the uncertainty on integrated fluorescence u(I_e) was estimated as the corresponding shot noise (see SI).

The terms ${\overline{F}}_H$ and u(F_H) in Eq. 6 were calculated as the mean and the standard deviation of the instrument function along y-axis within ±1 μm from the peak (corresponding to the center of the microchannel, see Fig. S9), whereas ${\overline{F}}_V$ and u(F_V) were estimated indirectly, by comparing the shift in the fluorescence intensity from SE sample after repositioning the microchannel along the z-direction (see SI and Fig. S9). All uncertainties in Eq. 6 where combined to give a final percentage uncertainty on size estimate (u(s)%) according to the error propagation theory (see SI).²⁰

RESULTS AND DISCUSSION

Selection of membrane dyes for EV staining.

Fluorescent probes for EV staining are small organic molecules that distribute preferentially into the hydrophobic vesicle membrane. The photophysical properties of the dyes and their performance vary drastically depending on their molecular structure. Dyes suitable for EV sizing using single molecule-sensitive flow techniques should meet the following criteria: i) capability to label all vesicles homogeneously; ii) a large number of dyes per vesicle, to ensure that intensity is proportional to vesicle size and is not distorted by the Poisson distribution of dyes per vesicle; iii) minimal aggregation, to avoid false positives and iv) highly fluorescent only after intercalation into a vesicle membrane (fluorogenicity), to enhance sensitivity and help discriminate vesicles from dye aggregates.

Using these criteria, we compared the performance of four commonly used dyes: Di-8-ANEPPS,²¹ FM1–43,²² Dioc14(3)²³ and Membright-488,²⁴ a recently developed dye. Whereas Di-8-ANEPPS and FM1–43 contain pyridinium-based scaffolds (see Fig. S1) that promote complete intercalation into a vesicle lipid bilayer, Dioc14(3) and Membright-488 are dialkylcarbocyanine derivatives with long hydrocarbon side-chains that anchor the dye into a vesicle’s surface (see Fig. S2).

We labeled B-Lipo with different concentrations of the four dyes and compared the results obtained using our single molecule-sensitive flow analyzer (Fig. 1a).¹⁰ We measured the intensity distribution (Fig. 1b–e) and number of particles detected (see Fig. S10). Interestingly, the intensity distribution of vesicles stained with Di-8-ANEPPS shifted to higher intensities as dye concentration increased up to 0–100 nM, then to lower intensities as dye concentration increased above 100 nM (Fig. 1b). This biphasic behavior may be due to unsaturated vesicle surfaces at lower dye concentrations, and dye quenching at higher dye concentrations.²⁵ Staining liposomes with Di-8-ANEPPS or Membright 488 resulted in efficient vesicle detection (>1000 events/min) even at low dye concentrations (see Fig. S10). In contrast, staining with FM1–43 resulted in inefficient detection (<100 events/min) even at the highest dye concentration (see Fig. S10), whereas staining with Dioc14(3) resulted in vesicle detection that was strongly dye concentration-dependent, with a peak at 50 nM Dioc14(3) (see Fig. S10).

We next assessed the formation of dye aggregates in solution at a high dye concentration (250 nM). We added the four dyes to buffer in the absence of liposomes, analyzed aggregate size by DLS and counted the number of aggregates using the flow analyzer (see Fig. S3). Dioc14(3) formed the largest aggregates (325 nm), consistent with its surfactant-like structure; Membright-488, Di-8-ANEPPS, and FM1–43 formed smaller aggregates (44, 20, and 10 nm, respectively). The particle count was over 40-fold higher for Membright-488 (~800 particles/min) than for the other three dyes. This large difference in particle count was likely due to a difference in single-aggregate brightness rather than a difference in aggregate concentration, based on the higher bulk emission intensity from Membright-488 in buffer than from the other three dyes (see Fig. S3).

From our results, Di-8-ANEPPS provided the most efficient vesicle detection (as previously suggested by Stoner et al.²⁶), resulting from the higher staining efficiency and higher fluorogenic behavior (see Fig. S3, and Scheme S1 and Fig. S11 for a proposed mechanism of vesicle staining by Di-8-ANEPPS), and had a low tendency to aggregate. Furthermore, although Di-8-ANEPPS and derivatives are commonly used as excitation ratiometric dyes for probing their local microenvironment, both their absorption at 488 nm and emission spectra remain relatively insensitive to the local environment across a variety of lipid compositions and membrane dipole,^27–29 dye concentration (see Fig. S12), and cholesterol content^30,31 (which already shows little variation— ~10%—in relative abundance among EVs derived from cells and biological fluids).^32–34 Another advantage of Di-8-ANEPPS is the absence of a net negative charge in the head group, which could reduce the efficiency of dye intercalation into EVs’ membrane (typically negatively charged)³⁵ due to repulsive forces.³⁶ Therefore, Di-8-ANEPPS was selected as a suitable membrane.

Correction of flow data using vesicle passing time.

The velocity of particles flowing within a micron-sized channel is distributed according to the Poiseuille equation for fully developed laminar flow:

$$u=u_{max}\left(1-\frac{r^2}{R^2}\right)$$ Eq. 3

where u_max is the maximum velocity (at the centerline) and R is the pipe radius.³⁷ The number of photons emitted from each EV is directly proportional to the time spent within the sample volume. In our 2 μm-wide channel, a vesicle flowing at 100 nm from the channel wall would move ~5x slower than the same vesicle travelling at its center, generating ~5x greater signal. Given the small channel size, hydrodynamic focusing cannot be used.^38,39 Therefore, we developed a method to correct the intensity of each vesicle based on the time required for the vesicle to cross the sample volume (the passing time, PT), allowing an estimate of vesicle size.

We operated the avalanche photodiode system at a sampling rate (10 kHz) that was faster than the crossing time. As shown in Fig. 2a, the peak width is determined by the vesicle speed, with vesicles travelling closer to the center of the channel moving faster and showing narrower peaks. The shape of the peak reflects the instrument function profile (corresponding to the product of the laser intensity profile and detection efficiency profile) along the axial flow direction. Therefore, each peak was fitted to a Gaussian curve to estimate the vesicle’s PT (see Experimental Section). Fig. 2b shows a segment of a trajectory acquired in flow, showing differences in peak width due to the effect described above. Fig. 2c shows examples of fitting four peaks corresponding to large and small vesicles moving at low and high speeds. Thus, each vesicle’s intensity I_e was divided by the PT to remove its dependence from the vesicle speed, and the normalized intensity I_N was transformed into vesicle size using Eq. 4:²⁶

$$I_N=\frac{1}{{\alpha }^2}4\pi r_v^2$$ Eq. 4

where r_v is the vesicle radius and 1/α² is a scaling factor estimated based on comparisons of flow-derived sizes with absolute sizes from Cryo-EM data (see Experimental Section).

Figure 2. Extracting vesicle velocity for correction of flow data.

a, Schematic of EVs flowing in a microchannel in laminar flow. Vesicle peaks are narrower at higher flow speeds (vesicle 1), broader at lower flow speeds (vesicle 2). b, Segment of flow data showing broadening of peaks. A high sampling rate (10 kHz) was used. c, Fitting of peaks to a bell-shaped function to assess vesicle passing time (PT), calculated using 6σ. The four peaks illustrate large vesicles (I, II) and small vesicles (III, IV) flowing at low speeds (PT_I = 0.542 ms, PT_III = 0.544 ms) and high speeds (PT_II = 0.382 ms, PT_IV = 0.371 ms).

Estimation of vesicle size and its corresponding uncertainty.

We began our study of EV sizes by analyzing two synthetic vesicles—45 nm-centered liposomes (S-lipo) and 80 nm-centered liposomes (M-lipo)—followed by a biological sample of exosomes isolated from seminal fluid (SEs). Preliminary experiments using the model system (liposomes) had the purpose of demonstrating the ability of our flow analyzer to detect a wide range of EV sizes (~35–300 nm), and of serving as a two-point calibration for sizes measured using Cryo-EM (see insets in Fig. 3d–f and Fig. S13), thus proving the proportionality between vesicle size and emission peak intensity. For each sample, we acquired flow data (see Experimental Section), corrected data for each vesicle using the vesicle PT, and calculated vesicle size using Eq. 4. As shown in Fig. 3a–c, the velocity distributions of vesicles were consistent with predictions from Poiseuille theory (Eq. 3) for a 2 μm-wide channel, with PT ranging from 0.35–1.6 ms. The synthetic vesicles flowed more slowly (average PT, 0.5–0.6 ms, versus 0.4 ms for SEs) and less homogeneously (wider PT distribution) than SEs. These results may have been due to the greater viscosity of the liposome solutions caused by addition of PEG surfactant to reduce liposome interactions with the walls of the polydimethylsiloxane (PDMS) microchannel.⁴⁰ As described above, emission signals from two identical EVs flowing in a 2 μm-wide channel in laminar flow can vary by as much as 5-fold depending on their distance from the channel wall and resulting velocity.

Figure 3. Estimation of EV sizes by correlation with cryo-EM data.

a-c, Vesicle PT distributions. d-f, Overlap of size distributions derived from flow analysis (violet) and from Cryo-EM data (green). The dashed lines represent the best fit to a log-normal distribution. Size distributions from Cryo-EM were centered (median ± SD) at 80.0 ± 46.7 nm (M-Lipo), 47.4 ± 20.2 nm (S-Lipo) and 70.2 ± 36.1 nm (SE), whereas flow-derived size distributions were centered at 79.9 ± 35.5 nm (M-Lipo), 48.2 ± 17.8 nm (S-Lipo) and 70.0 ± 44.0 nm (SE). The insets in Fig. 3d–f show a typical Cryo-EM image for each vesicles sample.

Flow-derived size distributions (obtained via Eq. 4) of S-lipo, M-lipo, and SE samples were compared with the corresponding distributions of absolute sizes (obtained by analyzing more than 1000 vesicles per sample in Cryo-EM). For each sample we obtained a good overlap between measured and estimated size (see Fig. 3d–f).

The α-factors obtained for the S-lipo and M-lipo samples were identical (7.33), indicating proportionality between the vesicle size and emission peak. The α-factor obtained for the SE sample was slightly different (5.95), possibly due to the presence of membrane proteins in the SE sample, which could influence both the binding and fluorescence of membrane dye. This also suggest that a specific scaling factor should be derived each time a new system (membrane dye/EVs family) is being analyzed. Once 1/α² has been determined (by comparison to reference sizes, such as from Cryo-EM, NTA etc.), it can be used unchanged in all further analyses of the same system.

For each individual size estimate we also calculated the corresponding uncertainty, by combining all different sources of variability of the fluorescence intensity (see Experimental Section and SI). In fact, the 3D-nature of the process represented by a vesicle crossing the sample volume (in a certain time) introduces additional errors (besides the shot noise) when using Eq. 2 to obtain the size of a vesicle from its measured I_N. By convention, we can consider the channel to be oriented so that the flow is in the positive x direction and the optical axis of the detection optics is aligned with the z axis (y axis is horizontal and perpendicular to the direction of flow). Then, we can conceptualize I_N as follow:

$$I_N=\frac{I_e\left(PT,F_H\left(y\right)F_V\left(z\right),n_d\left(r\right)\right)}{P{T}_e}$$ Eq. 5

where PT is the time a vesicle spends to cross the sample volume (PT_e is its estimation), F_H(y)F_V(z) are the distributions of the instrument function along y and z directions and n_d(r) represents the number of dye molecules per single vesicle. The latter is proportional to the vesicle’s radius r, and is Poisson distributed around the average number of dyes-per-vesicle.

Assuming the physical processes behind each of the parameters in the right-hand side of Eq. 5 to be independent, and considering each term to contribute to I_N as a multiplicative factor, we derived (see Experimental Section and SI) the corresponding uncertainty u(I_N) according to Eq. 6. Then, we calculated the relative uncertainty on size estimates (u(s)%) according to the error propagation theory (see SI).²⁰

$$u\left(I_N\right)=I_N\sqrt{{\left(\frac{u\left(I_e\right)}{I_e}\right)}^2+{\left(\frac{u\left(n_d\right)}{{\overline{n}}_d}\right)}^2+{\left(\frac{u\left(F_H\right)}{{\overline{F}}_H}\right)}^2+{\left(\frac{u\left(F_V\right)}{{\overline{F}}_V}\right)}^2+{\left(\frac{u\left(P{T}_e\right)}{P{T}_e}\right)}^2}$$ Eq.6

Scatter plots of u(s)% versus size for the S-lipo, M-lipo, and SE samples are shown in Fig. 4a–c. The uncertainty in size estimates is inversely related to vesicle size. This result is likely due to the Poisson noise from the photoncounting process: the smaller the vesicle, the larger the shot noise relative to the vesicle signal (u(I_e)/I_e). The Poisson distribution of the dyes-per-vesicle is of minor importance due to the large number of dyes even in the smallest vesicles (~170 Di-8-ANEPPS molecules in a 35-nm vesicle, based on TIRF microscopy, see SI) The insets in Fig. 4a–c report three subsets of vesicles extracted from the whole population, and show how, using our single-molecule sensitive flow analyzer, we would be able to isolate subpopulation of vesicles which are ~25 nm apart, with virtually no overlapping. As already reported,⁴¹ such resolution might result determinant when investigating the relationship between size and biophysical properties of EVs.

Figure 4. Accuracy of EV size estimates.

a-c, Scatter plots showing u(s)% as a function of vesicle size for the three samples. The colored dots indicate three distinct subpopulations of vesicles centered at 40 nm (red), 80 nm (blue) and 130 nm (orange). Each subpopulation span ± 10 nm around the corresponding center. The insert shows the three subpopulations along with the absolute uncertainty (error bars in light colors) associated with each single vesicle. d-f, Distribution of u(s)%

The accuracy of the vesicle size estimates is demonstrated in Fig. 4d–f, which show the distributions of u(s)% for the S-lipo, M-lipo, and SE samples. In agreement with the scatter plots, the overall uncertainty u(s)% for the smaller S-lipo samples was larger than for the M-lipo and SE samples, with only 56% of S-lipo vescles displaying a u(s)% less than ±16%, versus 87% and 84% of the larger M-lipo and SE samples, respectively. Due to the high sensitivity of the flow analyzer, we were able to obtain a u(s)% as low as ±24% for even the smallest EVs (~35 nm). Such accuracy in EV sizing was possible by combining a high- performance flow setup with correction of flow data using the vesicle crossing time.

Determination of experimental conditions for consistent EV size estimates.

During SE size analysis, we noticed that the results depended on the experimental conditions used during staining. To determine operating conditions that would allow consistent measurements, we systematically screened the following parameters for staining SEs with Di-8-ANEPPS: a) Time between staining and flow analysis, b) Dye concentration (titration using a fixed SE concentration), c) SE concentration (titration using a fixed dye concentration), d) Dye and SE concentration (varying both with a fixed ratio), e) Final EtOH concentration (EtOH is used to dissolve the organic probe before mixing) and f) Volume of SE sample used during staining.

From the bulk emission of SE samples stained with different concentrations of Di-8-ANEPPS, the intensity was proportional to the amount of dye up to 90 nM (corresponding to ~3×10⁴ dyes in solution per vesicle), above which quenching occurred (Fig. 5a). Furthermore, as observed for liposomes (see Fig. S10), the bulk emission intensity from SE samples either increased or decreased over time depending on the initial dye concentration.

Figure 5. Screening of experimental parameters for staining EVs with Di-8-ANEPPS.

a, Bulk emission intensity (λem = 588 nm) over time of SEs stained with different concentrations of Di-8-ANEPPS. b-f, Flow-derived size distributions obtained during the screening of optimal conditions for SEs staining. The figures legends correspond to a specific dye/vesicle ratio (b and c), concentration of EtOH (v/v%) (d), the reaction volume (e) and the dilution factor applied to the initial optimal concentration for staining C (f). Size distributions in panel f were centered at ~72 nm (black), ~61 nm (green), ~47 nm (red), ~37 nm (blue), and ~28 nm (orange), whereas all distributions in panels b-e were centered at ~70 ± 5 nm (except for the distributions represented by the black solid lines in panel b and c, which were centered at ~58 nm and ~45 nm, respectively). g, Center-of-sizes distribution against the corresponding average PT of vesicles, obtained from samples under optimal staining conditions. The blue and green dots refer to sizes obtained with or without correction of each individual vesicle by the corresponding speed. h, Representative distribution of the distance in time (Δt) between two consecutive peaks for low (pink), medium (green) and high (blue) density of vesicles (i.e. events counted). The median of Δt distribution is 40 ms (pink), 15 ms (green) and 6 ms (blue) and the events were acquired over a period of 3 min.

EV size estimates were more tolerant of changes in SE concentration for a given amount of dye (Fig. 5b) than of changes in dye concentration for a given amount of SE (Fig. 5c). We were able to reproduce the estimated vesicle size distribution over a range of dye/vesicle ratios from 0.6×10⁴ to 6.3×10⁴ by varying the concentration of SEs for a fixed dye concentration, but only from 1.2×10⁴ to 3.5×10⁴ by titrating SEs with Di-8-ANEPPS. This result is reasonable considering that only a small fraction (~6%) of dye injected in solution actually stains EVs: even if the number of vesicles increases, the dye would still be in excess. Below or above this optimal range of dye/vesicle ratios, we observed a shift in size distributions towards lower values, likely due to incomplete staining of vesicles or dye quenching, respectively.

Although EtOH concentration can affect staining efficiency by compromising EV membrane integrity⁴² or by enhancing dye solubility in buffer, we did not observe an appreciable change in size distribution using up to 3% (v/v) EtOH (Fig. 5d).

We observed no difference in size distributions when changing the overall reaction volume, keeping all other parameters fixed (Fig. 5e).

Staining at concentrations of vesicles <1.7×10⁹ particles/mL (at a fixed dye/vesicle ratio) resulted in inconsistent size estimates (Fig. 5f). Consecutive dilutions of SE sample consistently shifted the size distribution towards lower values, generating shifts in the distribution center of 36% for a 6-fold dilution and −62% for a 60-fold dilution.

The importance of correcting for the vesicle passing time is clearly demonstrated in Fig. 5g, which shows the center-of-sizes distribution versus the corresponding average PT of vesicles, obtained from samples stained under optimal conditions. Without correcting the PT (green dots), the estimated sizes would systematically shift towards bigger values as the vesicles’ velocities decrease (i.e. bigger PTs), leading to inconsistency between flow data. Thus, the center-of-sizes distribution for SE was 71±3 nm (mean of 16 independent measurements, see T1 in SI) when corrected by PT, or 87±13 nm without correction. Another advantage of operating at faster acquisition times is the minimization of coincidence and swarming effects during EV analysis. Fig. 5h shows the distribution of distance in time (Δt) between two consecutive peaks, when working at low, medium, and high concentrations of vesicles: the higher the concentration the greater the shift towards small Δt values. Sampling at a rate of 10 kHz allowed us to resolve and integrate fasttravelling vesicles that were as close in time as ~0.4 ms, thus detecting virtually zero coincidence counts even at high concentration of vesicles.

Overall, the gravity-driven flow combined with the μm-sized channel and the high sampling rate (10 kHz) allowed us to analyze vesicles in the frequency range of hundreds per second without coincidence counts.

CONCLUSION

Here we address the question of how to analyze the size of extracellular vesicles using a single molecule-sensitive flow analyzer and a membrane-selective dye. We demonstrate the capability of such a system to detect and analyze EVs of sizes from ~35–300 nm, and explore issues that arise when studying these small vesicles. Indeed, at sizes far below the micron range, additional constraints appear during sample preparation, mostly related to the photophysical properties of the labeling probe. Without optimization of staining, effects such as quenching of dye fluorescence due to overcrowding on the vesicle surface or, conversely, intercalation of an insufficient number of dyes, far below the saturation point, are likely to occur and contribute to errors in the estimation of vesicle size and errors in comparisons between samples.

To explore these issues, we compared the performance of four common probes and screened the effects of experimental conditions on EV sizing. In addition, we developed a method to improve the reproducibility and accuracy of EV sizing by using the crossing time of vesicles flowing across the laser beam, thereby accounting for the velocity profile of vesicles in laminar flow and correcting the fluorescence signal of each EV. Using high sampling rate and the correction by PT, we minimized coincidence and swarming effects and reduced the uncertainty of size estimation ~3 fold, thus enabled us to estimate vesicle size with a precision of ±15–25%. Regardless of the detection strategy employed (side scattering or fluorescence), flow cytometers with single-molecule sensitivity often require the use of high numerical aperture objectives, and therefore micron-sized channels in which vesicles typically move in the regime of laminar flow. Therefore, correcting for the vesicle’s velocity is fundamental when performing EV analysis on a single-vesicle level. Although at this stage we did not investigate the relationship between EV size and biomarkers of interest, the method we described here should facilitate these studies. Our approach can be implemented in flow setup with single-molecule sensitivity, yielding a considerable enhancement in size resolution without increasing hardware complexity (the platform uses gravity-driven flow). This study provides insights into applying flow cytometry to EV analysis, reinforcing the validity of flow cytometry as a technique for investigating the sizes of extracellular vesicles even in highly heterogeneous samples.