Size photometry and fluorescence imaging of immobilized immersed extracellular vesicles

Abstract

Immunofluorescence analysis of individual extracellular vesicles (EVs) in common fluorescence microscopes is gaining popularity due to its accessibility and high fluorescence sensitivity; however, EV number and size are only measurable using fluorescent stains requiring extensive sample manipulations. Here we introduce highly sensitive label‐free EV size photometry (SP) based on interferometric scattering (iSCAT) imaging of immersed EVs immobilized on a glass coverslip. We implement SP on a common inverted epifluorescence microscope with LED illumination and a simple 50:50 beamsplitter, permitting seamless integration of SP with fluorescence imaging (SPFI). We present a high‐throughput SPFI workflow recording >10,000 EVs in 7 min over ten 88 × 88 µm² fields of view, pre‐ and post‐incubation imaging to suppress background, along with automated image alignment, aberration correction, spot detection and EV sizing. We achieve an EV sizing range from 37 to ∼220 nm in diameter with a dual 440 and 740 nm SP illumination scheme, and suggest that this range can be extended by more advanced image analysis or additional hardware customization. We benchmark SP to flow cytometry using calibrated silica nanoparticles and demonstrate superior, label‐free sensitivity. We showcase SPFI's potential for EV analysis by experimentally distinguishing surface and volumetric EV dyes, observing the deformation of EVs adsorbed to a surface, and by uncovering distinct subpopulations in <100 nm‐in‐diameter EVs with fluorescently tagged membrane proteins.

1. INTRODUCTION

Cells secrete membrane‐bound structures called extracellular vesicles (EVs) that are fundamental to tissue homeostasis and cell signalling (Yates et al., 2022), and have garnered great interest as biomarkers and for therapeutics (Cheng & Hill, 2022). Their sizes range from a few tens of nanometres to more than a micrometre, and they carry a large variety of bioactive molecules such as proteins, lipids, nucleic acids and metabolites. Major efforts have been devoted to the development of single vesicle measurement techniques to characterize the heterogeneity of EVs and the features that govern cellular activation pathways, EV cargo sorting (Dixson et al., 2023) and how they change in health and disease (Yates et al., 2022). Currently, there are no generally accepted classifiers to distinguish EV subpopulations such as exosomes or ectosomes due to various technical challenges. Single protein expression levels suggest EV subpopulations, but due to considerable overlap are insufficient for classification (Mathieu et al., 2021). Multi‐dimensional characterization will be required and could involve multiplex proteomic analysis or incorporation of orthogonal parameters such as EV size. It is furthermore understood that EV diameters in various samples follow inverse power‐law distributions where small EVs with diameter $d<100$ nm are by far the most abundant (Arab et al., 2021; Saftics et al., 2023). Consequently, the limit of detection (LOD) of any single EV technology profoundly influences the results. For instance, proteomic analysis significantly varies between measurements with LODs of 100 nm compared to 150 nm (Welsh et al., 2023), suggesting that techniques with LODs substantially below 100 nm are required for reproducible and unbiased sample characterization.

Several techniques show potential to meet these requirements and advance EV subclassification. For example, electron microscopy can be combined with immunogold labelling to both size and characterize EVs down to d < 30 nm but has exceedingly low throughput and is thus impractical. Optical methods that can combine EV immunofluorescence analysis with EV sizing based on light scattering are preferred owing to their higher throughput. Indeed, flow cytometry (FCM) emerged as the current gold standard for single EV analysis owing to its comparative ease of use, low operating cost, and simultaneous measurement of both EV size (by side scattering) and protein expression (by immunofluorescence) (Welsh et al., 2023). Many common flow cytometers used to detect EVs have a LOD ranging between $d\approx 100$–150 nm, while specialized nanoflow cytometers can detect EVs as small as $d\approx 40$ nm, but at the expense of fluorescence multiplexing (Zhu et al., 2014). The LOD of FCM is commensurate to nanoparticle tracking analysis (NTA) with a label‐free LOD of $d\approx 80$ nm which is commonly used to assess EV size distributions. Lowering the detection threshold remains a challenge as nanoparticle scattering scales with $d^6$, and lowering the LOD just three‐fold requires almost a 1000‐fold gain in sensitivity. Recently, fluorescence imaging of EVs captured on a surface by widefield or TIRF microscopy (Burbidge et al., 2020; Han et al., 2021; Schürz et al., 2022) has gained traction as an alternative to FCM due to its accessibility, higher fluorescence sensitivity (Corso et al., 2019), and the possibility to manipulate EVs following their immobilization (Lee et al., 2018). The absence of a scatter signal for sizing is circumvented using fluorescent labels (e.g., dyes), but at the cost of further sample manipulation and susceptibility to false positive detection events, for example, due to label aggerates. Alternatively, super‐resolution imaging can be used to size EVs down to $d\approx 30$ nm but suffers from low throughput and requires very high expression of proteins on individual EVs (e.g., at least 20 copies [Saftics et al., 2023]). Label‐free detection of surface‐bound particles by dark‐field imaging is well established (Horio & Hotani, 1986); however, it requires specialized microscopes with bright light sources and strong stray light suppression, and the scattering signal also scales unfavourably $\propto d^6$, and has not been useful for EV analysis.

Interference imaging methods such as interference scattering microscopy (iSCAT) (Lindfors et al., 2004), single particle interference reflection image sensing (SP‐IRIS) (Daaboul et al., 2016) and several others (Huang et al., 2017; Ortiz‐Orruño et al., 2021; Wang et al., 2011; Yang et al., 2018) offer relief as they optically amplify signals with minimally added noise (Hsieh, 2018), are robust against stray light, and benefit from a more favorable contrast scaling $\propto d^3$. iSCAT and SP‐IRIS require the least complex optical setups as both generate the interferometric reference field by reflection of the illumination light from the imaging substrate in common‐path geometry. SP‐IRIS achieves this using custom‐made silicon chips coated with a thin dielectric reflective layer. It employs imaging from the top, mostly through air, since immersing samples adds complications such as additional reflections from the top surface or minute changes in liquid depth that affect focusing. iSCAT on the other hand uses common glass coverslips without any modification. Imaging is conducted not through the medium, but through the coverslip, and immersed samples can readily be analyzed. iSCAT offers exceptional sensitivity including label‐free single molecule detection and mass photometry (Young et al., 2018), it can be combined with fluorescence (Ortega Arroyo et al., 2016) and has been used to detect EVs (Kashkanova et al., 2022, 2023). However, most iSCAT setups use specialized hardware such as high‐powered lasers and high‐speed cameras which are not tailored for wide‐field EV fluorescence analysis over large fields of view (FOVs), so far limiting the adoption of iSCAT for EV research.

Here we introduce EV size photometry (SP) using iSCAT on a standard widefield fluorescence microscope with non‐imaging‐based autofocus by adding only a simple 50:50 beamsplitter and by reusing LED light source and sCMOS camera intended for fluorescence microscopy. We demonstrate the sizing of various immobilized nanoparticles by pre‐ and post‐incubation imaging of a capture surface and benchmark it against FCM using calibrated silica particles. Recognizing an upper sizing limit for SP, we introduce combined SP using two illumination wavelengths and extend the dynamic range of SP for polydisperse samples such as liposomes and EVs from $d\approx 37$ to $>200$ nm. Our analysis is facilitated by an open‐source automated high throughput image analysis pipeline that includes image alignment, background suppression, and aberration correction steps to register and analyse captured particles over multiple FOVs. We demonstrate the natural integration of SP with highly sensitive EV fluorescence imaging (SPFI) and validate the sizing accuracy on EV samples stained with surface and intraluminal dyes as a function of EV diameter by correlating SP contrasts with fluorescence intensities. We furthermore observe SP contrast behavior consistent with EV deformation on the surface. Finally, we analyze EVs expressing a fluorescent fusion protein and identify two distinct clusters of EVs with $d<100$ nm, which are revealed by correlating EV sizes with the expression of the protein tag and would be missed by techniques with LODs of $d>100$ nm.

2. RESULTS

SPFI allows to combine sizing of immersed EVs inside a microfluidic flow cell with further characterization using fluorescent reporters or affinity binders. Particle sizing is enabled through label‐free iSCAT imaging on common glass coverslips which can be implemented in any widefield epifluorescence microscope by the addition of a 50:50 beamsplitter that is switched with the dichroic filter cube, shown in Figure 1a. We use the terminology ‘size photometry’ because nanoparticles are conventionally characterized by size; however, we note that iSCAT contrasts are principally dependent on both EV size and refractive index and only a few implementations truly decouple them (Kashkanova et al., 2022, 2023; Ortiz‐Orruño et al., 2023). Throughout this manuscript, we validate the SP sizing accuracy for multiple types of large nanoparticles with increasing complexity (various solid nanoparticles, liposomes, and EVs). A detailed description of the iSCAT imaging principle is given in Section S1, describing the mathematical underpinning of its increased sensitivity for small particles compared to darkfield imaging. SP records pairs of pre‐incubation ($F_{\mathrm{pre}}$) and post‐incubation ($F_{\mathrm{post}}$) images over multiple FOVs to maximize imaging throughput while being selective to particles that are immobilized on the coverslip during incubation, Figure 1b. Each $F_{\mathrm{pre}}$ and $F_{\mathrm{post}}$ contain only static features and can be averaged indefinitely, Figure 1d, relaxing the need for high‐speed cameras which are otherwise common for dynamic iSCAT imaging (Dastjerdi et al., 2022; Piliarik & Sandoghdar, 2014).

FIGURE 1.

Overview of combined EV size photometry (SP) and fluorescence imaging (SPFI). (a) Schematic of the microscopy setup including a microfluidic flow cell with EVs immobilized on a coverslip in an epifluorescence microscope with LED illumination, a high sensitivity sCMOS camera, and an off‐the‐shelf 50:50 beamsplitter for iSCAT imaging. A motorized filter turret automatically switches between the dichroic mirror for fluorescence imaging, and the 50:50 beam splitter for iSCAT. (b) Workflow to suppress background inhomogeneities and surface‐bound crud. The background is registered in a pre‐incubation iSCAT image, then EVs are immobilized, unbound EVs are washed, and a post‐incubation iSCAT image is taken. The difference of the two images efficiently removes small surface inhomogeneities and allows to measure accurate contrast values of small particles within an undulating background. (c) Workflow to generate SPFI scatterplots. Typically, the iSCAT image is used to register candidate spots and contrasts or fluorescence intensities are subsequently measured on all channels. (d) Differential iSCAT of pre‐ and post‐incubation image which is used for EV localization. pre: image before particle incubation; post: image after particle incubation; SP: pre‐image subtracted from the post‐image. Three particles are highlighted: i is a particle that is detected both in iSCAT and in fluorescence, ii is a particle that is already present before incubation and is excluded from the rest of the analysis, and iii is a particle that is detected in iSCAT but not in fluorescence. (e) Fluorescence pre‐ and post‐incubation images that were taken concomitantly to the iSCAT in (d). The generation of a difference image in fluorescence is optional, and mostly inconsequential for substrates with low autofluorescence. (f) SPFI scatterplot of the processed data shown in (d) and (e), taken on EVs from A431 cells that were stably transfected to express CD63 fused with green fluorescence protein.

We introduce an SPFI analysis pipeline (https://github.com/junckerlab/SPFI) to handle data pre‐processing and image analysis starting with a step to correct chromatic aberrations between different imaging channels to maximize FOV sizes while ensuring correct spot matching between the channels (Section S2). Our calibration‐free algorithm can map and correct distortions between any two SPFI channels directly on the respective post‐incubation images. Next, the algorithm digitally aligns $F_{\mathrm{pre}}$ and $F_{\mathrm{post}}$ images (Section S3) and detects large surface inhomogeneities on $F_{\mathrm{pre}}$ to be excluded from subsequent analysis. Our algorithm then localizes candidate spots in the image data as local contrast minima in $F_{\mathrm{diff}}=F_{\mathrm{post}}-F_{\mathrm{pre}}$, thereby identifying captured EVs as well as sampling the image background to generate an experimental noise floor and inform thresholding (Figure 1c, Section S4). SP measurements are taken either in a single illumination wavelength or under dual illumination with a short and a long wavelength to maximise the sizing range (described below). The candidate spot contrasts $C$ are measured as center pixel values divided by the local illumination background taken from a flatfield image (see below and Section S1) and are thus unitless.

For SPFI, fluorescence images (Figure 1e) are superposed with the iSCAT images and fluorescence intensities $I$ for all candidate spots are measured as uncalibrated single pixel intensity values (c.f. Figure 1c, right panel). Taking differential fluorescence images is possible but often inconsequential due to overall low autofluorescence background in empty flow cells. The correlation between SP measurements and fluorescence signal intensities are reported as scatter plots of $I$ and $C$ (Figure 1f) which are reminiscent of FCM data and led us to adopt data analysis workflows and thresholding protocols common to FCM. EVs are identified in them as candidate spots that exceed the SP background noise found in buffer control experiments, thus establishing a count of the total EV number and removing the ambiguity of thresholding on EV fluorescence alone (Schürz et al., 2022).

Our analysis pipeline includes an iSCAT flatfield pre‐processing step to correct setup‐specific illumination inhomogeneities in the raw iSCAT images. The illumination background is identical for each FOV and can vary up to $\sim$20% within each image depending on the region of interest (ROI), see Figure 2a,b. Flatfield processing is necessary because pre‐ and post‐incubation images are often misaligned up to ∼15% of the FOV size (shifts are caused by manual pipetting and imperfect stage movements) and subtraction of the raw images can lead to residual background that overwhelms EV contrasts <1%. Flatfield images are collected as series of images from random positions over the sample which are averaged into a single flatfield image $FF$. The pre‐ and post‐incubation images are then background corrected by division like $F_{\mathrm{pre}}\to$ $F_{\mathrm{pre}}/FF$ and $F_{\mathrm{post}}\to$ $F_{\mathrm{post}}/FF$.

FIGURE 2.

SP noise characterization. (a) Image flatfield processing demonstrated on a raw post‐incubation image and (b) flatfield corrected image, which is necessary to remove illumination inhomogeneities but adds image noise. Shown is a 1200 × 1200 pixel ROI of a raw 16‐bit iSCAT image which has close to 20% illumination inhomogeneity. The color bar for the raw image is scaled to the bit depth of the image and dashed lines indicate limits used for plotting. The processed image is shown on a contrast greyscale of 1.5% around the image median. Dark spots are EVs from HT29 cells and circles indicate matching spots as a guide to the eye. (c) iSCAT image noise for a series of images collected at 40 FPS as function of number of images averaged. Black solid line represents an estimate of the photon shot noise and constitutes the lower theoretical noise limit whereas the black dashed line also includes the estimated camera readout noise. The black diamonds are measured SD of the image noise in the averaged timeseries, and crosses are for measured image noise after normalizing the timeseries with flatfield (FF) images consisting of $n_{FF}$ frames. The latter is compared to the respective estimated noise floor (horizontal dashed lines) and used to set imaging parameters for the iSCAT endpoint measurements. (d) SPFI plot generated in a buffer control experiment over 20 FOVs using iSCAT at 440 nm wavelength and a red fluorescence channel with vertical gate set three standard deviations above the noise. The plot shows zero localizations in the upper right quadrant and 249 localizations in the bottom right quadrant, many of which stem from setup‐specific image artifacts shown in the inset with a differential iSCAT image as white arrowheads.

2.1. iSCAT imaging parameters and image noise

The iSCAT images in this manuscript were taken by averaging 200 frames for both pre‐ and post‐incubation images and normalizing them using flatfield images consisting of 960 individual frames (120 arbitrary positions on a slide to eliminate local interference by dust with 8× average each). We found those imaging parameters by considering the trade‐off in imaging speed and detection sensitivity due to image noise as described below. The time it took to capture these images depended on the chosen FOV size since this limited camera acquisition speed and saving time. For example, in 1200 × 1200 pixel $(88\times 88\mu {\mathrm{m}}^2)$ FOVs with 200 frames could be taken in ∼12 s, split roughly equally between actual camera exposure and software overhead from computing the average and storing the data on the hard disk drive. Flatfield images needed to be taken only once per substrate regardless of the number of imaged FOVs and took about 180 s each. Imaging 10 different FOVs in one iSCAT channel with 560 EVs each thus took about 5 min, or 20 FOVs with ∼11,000 EVs total about 7 min (Section S5, Figure S5b).

The choice of imaging parameters presents a trade‐off between imaging time and final image noise, and thus detection sensitivity. The image noise in iSCAT is commonly minimized using cameras with high dynamic range and well depth, exposing pixels close to saturation, as well as averaging of multiple individual frames. Figure 2c shows an image noise characterization for a Prime 95B (Photometrics) with a well depth of 80,000 $e^{-}$ and readout noise of 1.6 $e^{-}$ (see methods). The solid black line indicates the theoretical noise floor consisting purely of photon shot noise (square root of the number of detected photons) whereas the dashed black line includes technical camera readout noise according to manufacturer specifications. This prediction was validated by taking pairs of iSCAT images (each averaged $n$ times), subtracting the two to remove fixed‐pattern noise and calculating the standard deviation ${\sigma }_n$ (black diamonds). The noise in SP data includes a second contribution from flatfield processing which adds noise depending on the number of averages $n_{\mathrm{FF}}$ of the flatfield image (crosses, color coded). We found that the measured noise saturated close to the expected noise floor (dashed lines) for small $n_{FF}$ but exceeded the prediction for larger $n_{FF}$ slightly. For the imaging parameters used throughout this manuscript, we predicted an approximate image standard deviation $\sigma =5\times {10}^{-4}$ where the flatfield contributed less than a third. The noise floor for SP contrasts $C_{\mathrm{SP}}$ exceeded this value (Figure 2d and below, note third root scaling) because our candidate spot localization registers contrast minima and thus samples at the extreme of the noise distribution. This additional noise overhead stemmed from small focus offsets between the averaged flatfield images and the actual FOV due to slow focus drifts which produced artifacts shown in the inset (see also Section S6). We calculate our SP LOD from Figure 2d as the mean plus three times the standard deviation (SD) of the noise, resulting in $\sqrt[3]{C_{\mathrm{SP},\min }}=0.10$. A more stringent definition, for example five SD above the background resulting in $\sqrt[3]{C_{\mathrm{SP},\min }}=0.12$, can be chosen for application requiring higher detection specificity.

2.2. Size photometry of silica nanoparticles

We characterized the sensitivity of SP using silica nanoparticles of known sizes and compared it to nanoparticle FCM using a CytoFLEX‐S instrument (Beckman Coulter Life Sciences). We measured violet side scattering (VSSC using a 405 nm laser) and side scattering (SSC using a 488 nm laser) of aminated silica particles with 50, 100, and 200 nm diameters in Figure 3a. We found that the instrument was able to resolve 100 and 200 nm particles which is in line with an ∼84 nm detection limit obtained by the manufacturer on an optimized instrument (Brittain et al., 2019). Notably, even for the 200 nm particles the signal was within the instrument noise level and could only be distinguished at high sample concentrations and high detection rates.

FIGURE 3.

Calibration, validation, and benchmarking of SP. (a) Nanoparticle flow cytometry (FCM) and (b) size photometry (SP) scatterplots of a buffer control and silica nanoparticles with 50 nm, 100 nm, and 200 nm diameters. For FCM, the samples were diluted to matching concentrations based on the manufacturer specification, measured for 60 s each, and 5% of total events are shown by their violet side scattering signal (VSSC‐H at 405 nm) and side scattering signal (SSC‐H at 488 nm). SP plots show iSCAT contrasts to the third root for 440 and 740 nm illumination wavelength, respectively. The 50 nm particle signal in SP is clearly separated from the noise blob whereas for FCM no cluster is visible. (c) Representative SP images of particles of different sizes. All color scales are set from −1.0% to +0.5%, scalebar is 500 nm. Interference contrast inversion for 200 nm nanoparticles is observed for 440 nm illumination. (d) Theoretical (solid lines: fit to analytical model; dashed lines: calculated by finite element modeling) and experimental (dots) SP contrasts as function of particle size. Horizontal and vertical error bars are SD of the particle sizes specified by the vendor and SD of particle contrast, respectively.

We next took SP measurements under 440 and 740 nm iSCAT illumination wavelengths after adsorbing the particles onto plasma‐activated glass coverslips. We show SP scatterplots for 50, 100, and 200 nm particles in Figure 3b and representative post‐incubation images for these particles as well as for 20 and 80 nm ones in Figure 3c. We found that we could detect particles in all samples (Figure 3c) and that the 50 nm particles were the first to cluster separately from the background noise floor (Figure 3b). The scatterplot for the 50 nm particles revealed many particles outside the main cluster which were identified in the raw images as large, non‐circular iSCAT spots. These are consistent with aggregates of two or more particles which thus likely dominated the corresponding FCM plot. Similar aggregates were also visible for 100 and 200 nm particles but produced excessively bright iSCAT contrasts and were discarded by the localization algorithm. For the 200 nm particles (rightmost panel in Figure 3c), we observed a contrast inversion under 440 nm illumination which requires special attention for SP sizing. The standard mathematical model of interferometric imaging (Section S1) predicts that contrasts $C$ initially scale with particle diameters $d$ to the third power (Kashkanova et al., 2022). This $C\propto d^3$ scaling is an approximation that breaks for large particles, requiring experimental validation to define its range of applicability and establish the maximum sizing range achievable with SP. For the silica standards, we combined measurements $C_{\lambda }$ of 20, 50, 80, 100, and 200 nm particles for both wavelengths $\lambda$ in Figure 3d. We fitted the first three samples to $d^3$ curves (solid lines, color‐coded for wavelength), demonstrating that the $d^3$ scaling extends to 80 nm particles under 400 nm illumination and to 100 nm (but not 200 nm) particles under 740 nm illumination. We continued to investigate the breakdown of the $d^3$ scaling and contrast inversion of $C_{440\mathrm{nm}}$ using finite element (FEM) simulations in the COMSOL optics module (Section S7). The simulation data, plotted as dashed lines in Figure 3d, reproduced the contrast inversion observed for 200 nm particles with 440 nm illumination and the absence of contrast inversion for 200 nm particles with 740 nm illumination. The simulations furthermore indicate that the $d^3$ scaling expands to ∼90 and ∼150 nm particles under 440 and 740 nm illumination, respectively. While these simulations improve on the analytical $d^3$ model, they still miss important effects such as the Gouy phase shift (Hwang & Moerner, 2007) from microscope focus (Section S7) and phase factors stemming from optical aberration in the microscope setup (Gholami Mahmoodabadi et al., 2020). Future work on analytical treatments of these effects or using machine learning approaches will be instrumental to extend the SP sizing range by using information in the rings around large particles to alleviate ambiguity of contrasts around the inversion point.

We next calibrated SP contrasts for sizing of various nanoparticles including EVs using $d=\kappa \sqrt[3]{C}$, where the proportionality factor $\kappa (n_{sample},\lambda$) is dependent on the refractive index $n_{sample}$ and the imaging wavelength $\lambda$ as detailed in Section S1. We used the silica data in Figure S1a to determine $\kappa (n_{silica},440nm)=291\mathrm{nm}$, and validated that the SP contrasts of both polystyrene (PS) and gold nanoparticles are consistent with this value when accounting for their respective refractive indices, see Figure S1d,e. The silica proportionality factor $\kappa (n_{silica},440nm)$ is used below to estimate EV sizes in SP.

2.3. SPFI of liposomes with dual wavelength iSCAT imaging

Figure 4 shows SPFI data on fluorescently stained liposomes which were prepared from DOPC lipid incorporating 1% DiI dye by extrusion through track‐etched membrane filters with a 400 nm pore size (representative images in Figure 4a). Extrusion using large pore sizes is known to produce liposomes with sizes both smaller and larger than the filter pore. This made the sample suitable to test SP of polydisperse samples, as the addition of the lipid dye allowed for an independent measure of total lipid content within each liposome. Figure 4c shows a SPFI scatterplot that was generated from the image data under 440 nm iSCAT illumination. Third root SP contrasts and DiI intensities were found to correlate well for small liposomes with $\sqrt[3]{I_{\mathrm{DiI}}}\ll 10$ in arbitrary units. At higher fluorescence levels, however, the correlation was considerably lower due to the setup‐dependent interferometric contrast switch resulting in ambiguous center pixel SP contrast values. Intriguingly, we found a maximum $\sqrt[3]{C_{440\mathrm{nm}}}\approx 0.4$ which is even larger than the $\sqrt[3]{C_{440\mathrm{nm}}}\approx 0.35$ that we measured for 100 nm silica beads. This indicated that liposomes flattened upon adsorption on the glass surface which would result in a decrease of the interferometric phase factors and thus a delay of the contrast inversion. Flattening has been documented, and moreover liposomes are often used to form supported lipid bilayers (Andrecka et al., 2013). We performed COMSOL simulations in Figure 4b showing that flattening of solid particles while maintaining the total mass indeed results in a delay of the iSCAT contrast inversion without affecting the initial $d^3$ scaling.

FIGURE 4.

iSCAT at multiple wavelengths and fluorescence imaging of fluorescently stained liposomes. (a) Representative SP images of liposomes imaged at 440 nm (1, note white center in large dark spot indicated by a white arrowhead) and 740 nm (2) with aligned color scales (−1% to +0.5%) and in fluorescence (3) to visualize their DiI stain. (b) FEM simulation of iSCAT contrasts at 440 nm of solid particles that remain spherical or flatten to 2/3 or 1/3 of the original height while maintaining a constant volume. Flattening allows larger particles to be measured by suppressing contrast inversion while increasing the maximal contrast. (c) SPFI scatterplot of DiI stained liposomes imaged with a 440 nm and (d) 740 nm iSCAT illumination, both showing a correlation of DiI intensities and SP contrasts albeit with different sensitivities (slopes) and variability. (e) Scatter plot correlating the SP contrast at the two wavelengths. The purple dots indicate the size range at which the two contrast values are proportional to each other and is used to determine the proportionality factor η by a linear fit (solid black line). (f) Sizing of liposomes using both wavelengths, with blue dots measured at 440 nm, red dots measured at 740 nm and purple dots averaged between the two measurements using their proportionality factor η, and a linear fit to the data above $\sqrt[3]{C_{SP}}>0.1$ (solid black line) with $R^2=0.86$. Gray dots fall below the iSCAT detection threshold using either of the wavelengths.

We confirmed that many larger liposomes that underwent the contrast switch at 440 nm maintained dark iSCAT contrast when imaged under a longer 740 nm wavelength, as shown in Figure 4d. Expectedly, this came at the cost of reduced sensitivity overall which we visualized by a SP scatterplot of both contrast measurements in Figure 4e. We found that in the region above the noise floor of the 740 nm channel and below the contrast inversion point of the 440 nm channel (purple dots in Figure 4e) both measurements are roughly proportional to each other. Using a linear fit to the data (forcing a zero y‐intercept), we determined a contrast proportionality factor $\eta =2.6$ between the two wavelengths which was reasonably close to the theoretically predicted value of 2.8 based on wavelength scaling. Both measurements were combined to $\sqrt[3]{C_{\mathrm{SP}}}$ in Figure 4f, where $\sqrt[3]{C_{440\mathrm{nm}}}$ was used for small liposomes (blue dots), $\eta \sqrt[3]{C_{740\mathrm{nm}}}$ for large ones (red dots), and intermediately sized ones were averaged between the two measurements (purple spots). Combining measurements in two wavelengths using an SP correlation plot (Figure 4e) can thus maximize the dynamic range for sizing polydisperse samples at the expense of dual measurements.

2.4. SPFI of EVs

Small extracellular vesicles (sEV) from the supernatant of several cell lines were purified by size exclusion chromatography (SEC) and imaged using combined 440 and 740 nm illumination wavelengths. EV sizing with SP faces two principal challenges: firstly, EVs are expected to have a size‐dependent refractive index since the contributions of EV membrane and lumen to the overall refractive index change with physical EV size and secondly, there can be additional variation in EV refractive index due to varying amounts of protein and nucleic acid loading. We addressed the first challenge using numerical calculations (Figure S2a) and FEM simulations (Figure S7), where we compared EVs modeled with solid sphere equivalent refractive index to core‐shell models (Kashkanova et al., 2022; Welsh et al., 2020) in which membrane and lumen contribute independently to EV scattering signals. Both showed that within our expected sizing range, core‐shell geometries lead to only minor corrections of EV sizes compared to the uniform models, which led us to adopt a single valued $n_{\mathrm{EV}}=1.40$ in our size calibration for simplicity. For completeness, all our plots in the following contain two x‐axes indicating both raw and calibrated contrasts. The EV size calibration was done with $d=\kappa (n_{\mathrm{EV}},440nm)\sqrt[3]{C_{\mathrm{SP}}}$, where the proportionality factor $\kappa (n_{\mathrm{EV}},440nm)=369\mathrm{nm}$ was calculated from $\kappa (n_{\mathrm{silica}},440nm)$ as described in Section S1. With the detection limit $\sqrt[3]{C_{\mathrm{SP},\min }}=0.10$ (Figure 2d), this calibration results in an LOD for detecting EVs with $0.10{\kappa }_{\mathrm{EV}}\approx 37\mathrm{nm}$ diameters using our setup. Additionally, we found that experimentally measured SP contrasts at 440 nm illumination before contrast inversion can exceed maximum contrasts predicted by FEM simulations for both solid and core‐shell EV models, suggesting that EVs, just as liposomes, can flatten upon surface adsorption (Yokota et al., 2019).

To address the challenge of random variations of EV refractive indices, SP was benchmarked against fluorescent stains which are commonly used for EV sizing. Figure 5a shows an SPFI scatterplot of EVs from HT29 colorectal adenocarcinoma cells which were biotinylated via their membrane proteins using the membrane impermeable Sulfo‐NHS‐Biotin and then incubated with Alexafluor 674 (AF647)‐labelled streptavidin. EVs were immobilized on glass coverslips pre‐coated with poly‐L‐lysine (PLL) which is positively charged at neutral pH and captures EVs by electrostatic interaction. The scatterplot was divided into four quadrants (Q1–Q4, labelled in Figure 5a) by applying SP contrast and fluorescence intensity thresholds found in control experiments (Figure S9). We note that biotinylation is not specific to EVs and spots in Q1 on the top/left include many biotinylated soluble proteins as well as streptavidin molecules that were not washed off; however, by correlating the fluorescence data to label‐free iSCAT, these spots can be efficiently removed from the data prior to further analysis. Assuming a constant membrane protein density for EVs of different sizes, the fluorescence signal intensity is expected to correlate with EV surface area. We tested this prediction by a fit to $\sqrt[3]{I_{\mathrm{AF}647}}\propto d^{2/3}$ for $\sqrt[3]{C_{\mathrm{SP}}}>0.1$, shown as solid line, which resulted in a correlation coefficient of $R^2=0.75$. Any linear fit to the data would result in non‐negligible $y$‐axis intercepts which cannot be explained purely by detection sensitivities. Figure S9d shows additional data with membrane staining using MemBrite which shows a similar contrast behavior.

FIGURE 5.

SPFI scatterplots correlating EV size with fluorescence intensity. (a) SPFI scatterplot of biotinylated EVs that were stained with Streptavidin‐AF647. Dashed lines are contrast and intensity thresholds separating real detection events from image noise. Quadrants Q1–Q4 apply to all scatterplots. The solid line is a model assuming homogenous labelling of the EV membrane proteins. (b) SPFI scatterplots of EVs stained using cell tracker deep red (CTR) where the solid line is a model assuming dye uptake in the lumen of the EVs. (c) FCM measurements of EVs from A431 cells expressing CD63‐GFP measured in the fluorescein (FITC) channel. (d) Negative stain TEM image of the CD63‐GFP EVs. (e) SPFI scatterplot of the CD63‐GFP EVs showing two distinct clusters of GFP⁺ spots, one group with $\sqrt[3]{I_{\mathrm{GFP}}}>4$ and EV sizes up to $\sim$100 nm and a smaller group with $\sqrt[3]{I_{\mathrm{GFP}}}<4$ on EVs up to only $\sim$80 nm, either of which could not be resolved in FCM (see (c)). (f) Post‐incubation images of selected EVs identified by star symbols in (e). Top row: GFP fluorescence; middle row: 740 nm iSCAT; bottom row: 440 nm iSCAT. Panels can be matched to the spots in (e) by the colors of their borders. Intensity bars show raw iSCAT and fluorescence contrasts and scale bars are 500 nm.

To further corroborate these finding, a second sample of HT29‐derived EVs was stained with Cell Tracker Deep Red (CTR) and captured on a plasma‐cleaned glass surface, Figure 5b. Since CTR accumulates in the EV lumen, the fluorescence intensity is expected to scale with the EV diameter and hence $\sqrt[3]{I_{\mathrm{CTR}}}\propto \sqrt[3]{C_{\mathrm{SP}}}$. This scaling behavior was confirmed using a linear fit (solid line in Figure 5b) to the datapoints above $\sqrt[3]{C_{\lambda }}>0.1$ which resulted in a correlation coefficient $R^2=0.82.$ Similar to Figure 5a, spots above $\sqrt[3]{C_{\mathrm{SP}}}>0.4$ showed systematic deviation from the fit, indicating potential underestimation of EV sizes by SP due to limited sizing range at 740 nm (see also Figure S10 for SPFI plots of the individual wavelengths). Figure S11 shows additional experiments where EVs were co‐stained with CTR and Carboxyfluorescein succinimidyl ester (CFSE) which both showed similar correlation with third‐root SP contrasts and similar detection sensitivities. Finally, Figure S12a,b show CTR staining on EVs derived from IMR‐90 fibroblasts and A431 epidermoid carcinoma cells, respectively, extending SP validation to three cell lines from multiple tissue origins. Figure 5a,b together with Figures S9–S12 thus show that SP contrasts correlate well with four different fluorescent staining methods on three different cell lines and can distinguish surface from bulk labelling.

Next, we highlight the utility of SPFI by comparing it to other single EV methods on EVs from A431 cells expressing CD63 fused to green fluorescent protein (CD63‐GFP) in addition to their wildtype CD63 (Figure 5c–g). CD63 is a transmembrane protein from the tetraspanin family and a common marker of small EVs due to its high abundance and role in exosome biogenesis. Figure 5c shows FCM data of this sample which had a LOD of $d\approx 100$ nm and includes considerable technical noise as seen from the buffer control (Figure S13). Figure 5d shows a negative stain TEM image which confirmed the presence of many EVs of $50$ nm and below. Figure 5e shows an SPFI scatterplot of the data which detected EVs down to $\sim 37$ nm in our calibration and shows non‐uniform GFP fluorescence with two clear clusters. The first cluster comprises bright GFP⁺ spots ($\sqrt[3]{I_{\mathrm{GFP}}}>5$ in arbitrary units) on EVs with diameters up to $\sim 100$ nm with linear correlation between contrast and fluorescence. The second cluster consists of a much larger number of dim GFP⁺ spots ($\sqrt[3]{I_{\mathrm{GFP}}}<4$ in arbitrary units) on EVs up to only $\sim 80$ nm (see Figure 5f for images of the EVs highlighted with arrowheads in Figure 5e). The SPFI plot visually resolves the two clusters by combining label‐free SP contrasts and CD63 expression data and the mismatch in size range between the clusters suggests that they represent different subpopulations. Figure S12c,d show measurements of a second sample from this cell line which was stained with CTR. A direct comparison of the scatterplots where EV size is either measured via CTR signal (Figure S12c) or via SP (Figure S12d) shows that either method can visualize the two GFP⁺ clusters, yet only SP allows for an absolute EV size calibration and requires no sample manipulation. A measurement of the CD63‐GFP signal in FCM on the other hand (c.f. Figure 5c), could not resolve the two clusters due to a much higher detection threshold in side‐scattering. The iSCAT data furthermore allows to count the total EV number and we found that 8120 out of 17640 spots above the gating thresholds were positive for GFP, or in other words the overall EV labeling efficiency of the CD63‐GFP transfected cells within our iSCAT detection range was $46$%.

Figure  S13a,b show a comparison of the same sample captured on a plasma‐activated glass and on PLL‐coated glass, and we found that the overall shape of the SPFI plot was largely comparable and both GFP⁺ clusters are readily identifiable on either surface. SP contrast measurements differed minorly, which will be further investigated in future work to improve sizing and include capture bias of different surfaces. Figure S14 furthermore shows the same EVs purified via a 10,000 × g ultracentrifugation protocol which led to the inclusion of much larger EVs in the sample. We found that our current single‐pixel contrast measurements indeed limit SP sizing at 740 nm to EVs of ca. 200 nm. However, the data are suggestive that improved sizing including a full interferometric PSF model will increase the range in the future (see Figure 14d).

3. DISCUSSION

We described SP of EVs and nanoparticles using a common epifluorescence imaging setup, including the microscope with non‐imaging‐based autofocus, LED‐based illumination, and sCMOS camera, and an additional broadband 50:50 beamsplitter available for ∼US$150. In our hands, SP was able to detect and size immobilized EVs down to $d\approx 37$ nm. SP exceeds the detection sensitivities of many commercial FCM and NTA systems (see Figure S13h) with a high throughput and high particle numbers by imaging over multiple FOVs, Figure 6. While SP cannot assess absolute EV sample concentrations, it also does not rely on accurate knowledge of them since the incubation can be monitored in real‐time (Videos S1 and S2). The SP setup itself is similar to iNTA (Kashkanova et al., 2022, 2023) and in‐solution measurements of sample concentrations and hydrodynamic EV diameters are possible but will require more powerful illumination and high speed cameras to achieve similar LODs to SP. We propose the use of SP with fluorescence imaging as in SPFI as the main application which replicates the defining features of FCM that can simultaneously capture fluorescence and label‐free side scattering data. SP likewise complements EV fluorescence microscopy data with label‐free nanoparticle enumeration and sizing, helps the identification of individual EVs and discriminate against unbound labels or substrate autofluorescence. Thanks to pre‐ and post‐incubation imaging of immobilized EVs, high SP sensitivities can be achieved by averaging with common standard microscopy hardware, and in our hands performance was ultimately limited by the mechanical stability and focus drift of the setup (Figure 2). Our setup makes no compromise on fluorescence imaging sensitivity, and we note that our SPFI workflow is compatible with super‐resolution imaging (Saftics et al., 2023), EV subpopulation enrichment via affinity capture surfaces (Saftics et al., 2023), and cyclic immunofluorescence imaging (Lee et al., 2018).

FIGURE 6.

Comparison of SP against other label‐free detection techniques. SP sizing range and throughput are plotted for imaging with a 100× as well as a 60× objective (see Figure S16). The SP throughput is calculated by assuming a capture density of $400EVs/{(100\mu \mathrm{m})}^2$ together with the imaging parameters used in this manuscript (see methods, Figure S5). The full lines for FCM and NTA represent the commercial instruments used in this manuscript where the lower threshold indicates the LOD by scattering detection, whereas the dashed FCM line indicates the highest reported sensitivity on any FCM to our knowledge (Zhu et al., 2014). Note that NTA and iNTA (iNTA LOD and throughput were taken from (Kashkanova et al., 2022)) report hydrodynamic diameters which are different to the physical EV size reported by SP and FCM. The line for microfluidic resistive pulse sensing (MRPS) is based on a Spectradyne nCS1 with C‐300 cartridges where the throughput is the approximate limit for polydisperse EV samples due to cartridge blocking from EVs below the detection threshold. Shaded areas indicate commonly assumed size ranges for large EVs (lEVs), small EVs (sEVs) and smaller, non‐membranous nanoparticles (NPs).

We present a SPFI pipeline which includes automated image pre‐processing including iSCAT flatfield corrections and aberration corrections, FOV alignment, candidate spot detection, and contrast measurements using two SP illumination wavelengths. Single color SP pre‐ and post‐incubation imaging of more than $10,000$ EVs and nanoparticles over $20$ FOVs could be completed within $\sim 7$ min (more FOVs can be acquired if desired), consuming only 10 µL of sample per flow cell.

To establish label‐free nanoparticle sizing with SP, we first imaged silica particles of known sizes and confirmed that SP contrasts $C$ scale approximately with $d^3$ for diameters. We found an upper sizing range for silica of about $\sim 90$ nm when imaged under 440 nm wavelength, and about ∼$150$ nm diameter when imaged under longer 740 nm wavelength illumination, beyond which SP contrasts invert. We then compared SP contrasts of the 50 nm silica beads to 50 nm PS and gold nanoparticles and validated that the relative SP contrasts matched iSCAT theory when accounting for the difference in refractive indices. These results established a basis for sizing other nanoparticles and materials of known refractive index using the $d^3$ scaling. A particular challenge to sizing EVs is that their refractive index can vary from the random incorporation of proteins and nucleic acids, potentially interfering with the expected size‐dependency of contrasts in SP measurements. We used fluorescent dyes as independent measurements of cell culture‐derived EV sizes and found correlation coefficients between fluorescence intensities and SP contrasts of $R^2=0.75$ for surface staining via biotinylation and $R^2=0.82$ for intraluminal staining with CTR. This high correlation suggests that the $d^3$ size dependence of SP contrasts can be strong enough to overwhelm refractive index variations and that SP can thus serve as a label‐free EV sizing method. Given the large variety of possible EV samples, future studies are needed to establish the generality of this correlation and its applicability to mixed samples such as plasma‐derived EVs, and we propose that fluorescent staining experiments are a practical way to verify SP for new sample types. Notably, SPFI was able to distinguish surface and volumetric stains via the relationship between size and fluorescence intensity $I$ that was proportional to $d^2$ and $d^3$, respectively, which we believe to be the first direct experimental demonstration and further validation. We propose that size calibration using beads of various materials allows to standardize SP measurements between various microscopy setups, similarly to how size calibration is meant to standardize FCM measurements (Welsh et al., 2023). We note that the average EV refractive index is expected to vary as function of size (volume) because the proportion of lipid (surface‐based) changes. We thus performed FEM simulations of core‐shell EV models to investigate how EV membrane and lumen with different refractive affect the size calibration. While EV refractive indices drop with increasing EV size, the simulations suggested that a solid sphere equivalent refractive index of $n_{\mathrm{EV}}=1.40$ is a good approximation to the average EV refractive index within our current SP sizing range. We propose that sizing accuracy and the upper sizing limit can likely be increased through improved image analysis in future work, either by explicitly modelling the full interferometric point spread function or by machine learning approaches. SPFI might also be used to further investigate EV refractive index variations and size dependence, for example, by changing the refractive index of the surrounding media. We finally demonstrated the utility of SPFI for EV subpopulation discovery on CD63‐GFP tagged EVs by comparing it to FCM. SPFI scatterplots revealed that GFP was present in EVs below 100 nm in size in two distinct clusters that differed in both GFP brightness and SP contrasts, whereas FCM only picked up diffuse expression on EVs larger than 100 nm.

A corollary of the above discussion is that the capture conditions including buffer, pH, salt concentration, and surface and EV charge could influence EV physisorption and introduce capture or sizing bias. We mainly captured EVs by electrostatic interaction using positively charged PLL‐coated coverslips (Video S1). We found, however, that EVs could also be immobilized on negatively charged surfaces in PBS buffer, suggesting screening of the EV charge and major contribution by Van der Waals forces to EV adsorption (Video S2). We found that electrostatic attraction on PLL increased capture rates but could also increase EV flattening on the surface. Flattening without material loss results in a reduction of the interferometric phase resulting from the out‐of‐plane material. This extends the range in which $C\propto d^3$ as it delays the contrast inversion, which could increase the range of SP. Strong deformations, however, might lead to fluid loss from the EV lumen and thus additional complications for EV sizing. Future work will investigate whether EV deformability in different capture environments could open the door to all optical single vesicle stiffness or density measurements using SP.

4. CONCLUSIONS

SPFI of EVs is readily accessible by the addition of a $50:50$ beam splitter to any modern epifluorescence microscope and runs on an openly available image processing pipeline (https://github.com/junckerlab/SPFI). SP on pre‐ and post‐incubation images can detect and size EVs from $d\approx 37$ nm to $>200$ nm, and its combination with fluorescence imaging generates rich multimodal single EV datasets that could help identify EV characteristics and clusters in $>500$ EVs per FOV, and $>10,000$ in each microfluidic flow cell in our experiments. SPFI can distinguish membrane from intraluminal dyes in EVs, captures deformation of EVs, and reveal subpopulations that are only distinguishable in EVs with $d<100$ nm. Finally, SPFI should be broadly applicable to EVs and other rigid or soft, synthetic or biological nanoparticles.

5. METHODS

5.1. Microscopy setup

The iSCAT system was built on an inverted fluorescence microscope (Nikon TI2) using a 100× plan apochromat (λ series, Nikon) of NA 1.45 with a 1.5× intermediate magnification lens (inbuilt feature of the Ti2). We performed additional characterization of SP detection sensitivity using a 60× objective shown in Figure S16; however, due to reduced sensitivity all experimental data shown were taken with the 100× objective. Imaging was done on a Prime95B (Photometrics) which has a full frame 118 × 118 µm² FOV size with 1608×1608 pixel sensor; however, measurements were taken with a reduced 1200 × 1200 pixel ROI to ensure optimal focus throughout the ROI. Illumination was provided by a Spectrapad (Lumencor) via a standard fixed large FOV epifluorescence illuminator (Ti2‐LAPP fixed main branch, Nikon) with built‐in fly‐eye lens. The illuminator has an adjustable aperture and two alignment screws which both affected the achievable iSCAT contrasts (data not shown). Aperture alignment and opening was optimized for highest contrast of 50 nm silica particles before taking all data in this manuscript. iSCAT using multiple FOVs furthermore relied on an LED based reflectance (non‐imaging) autofocus (Perfect Focusing System, Nikon).

The fluorescence measurements were taken using a pentapass dichroic mirror (FF409/493/573/652/759‐SP01, Semrock) and pentapass emission filter (FF01‐432/515/595/681/809, Semrock). The emission was additionally filtered using 515/30 nm (GFP), 595/40 nm (SPI) or 680/42 nm (AF647) filter sets (Semrock) in an automated filter turret in front of the camera. Maximum illumination power of the Spectrapad was used in all channels and images were averaged 8× with 100 ms exposure time each.

iSCAT imaging used the same light source (Spectra X Light Engine, Lumencor) which provides 370 mW output power in the 440/20 nm channel and 80 mW in the 740/20 nm channel at the maximum power setting according to manufacturer specifications. A filter cube with a widefield 50:50 beam splitter (BSW10R, Thorlabs) was added in the filter turret to switch between fluorescence and iSCAT modes. The frame exposure was maximised within the dynamic range of the camera, which resulted in exposure times around 4 ms at 440 nm and 25 ms at 740 nm. This allowed to image ROI of 512 × 512 pixels at 95 FPS, or 1200 × 1200 pixels at 40 FPS. The FOVs positions in the flow cells were auto generated on a rectangular grid approximately centered in the flow cell. The FOV spacing was 700 µm to avoid accidental photobleaching of neighbouring FOVs.

5.2. iSCAT noise characterization

The noise characterization in Figure 2 was performed using the same settings as for regular imaging in FOVs of 512 × 512 pixels with camera exposure of 4.8 ms. The expected noise was calculated using the camera bit depth of $\delta =2^{16}$, well depth of $\gamma =80,000$ e‐, quantum efficiency of $\eta =95$% and readout noise $\nu =1.6e^{-}$. Due to illumination inhomogeneities, the maximal exposure is only achieved in a small region in the center of each frame while it drops off substantially towards the edges. We adjusted camera exposure time to achieve highest single pixel exposure of 98.5% saturation which resulted in an average pixel exposure of $\rho =75.4$%. This corresponds to $n=\frac{\gamma }{\eta }\times \frac{\rho }{\delta }=\text{57,600}$ captured photons per pixel for single frame averaging and allows to calculate the photon shot noise as $\sigma =\sqrt{n}$, shown as solid line in Figure 2c. Since the camera readout noise was limiting the dynamic range to 50,000, we also calculated an adjusted noise floor by accounting for the readout noise as ${\sigma }^{\prime }={\sigma }_{\mathrm{sn}}/\sqrt{\upsilon }$.

All points in Figure 2c were generated using a dataset containing $N=5000$ images (taken at a frame rate of $\sim 500$ images/s) of an empty coverslip with an average pixel saturation of $\sim 90$% (similar to Figure 2a). For each amount of averaging, we computed two images, $I_{\mathrm{U}}$ and $I_{\mathrm{E}}$, where $I_{\mathrm{U}}$ is the average of the first $n$ uneven frames, and $I_{\mathrm{E}}$ is the average of the first $n$ even frames of the original time series. This interleaved averaging of the frame made the analysis robust against slow illumination power fluctuations. We then calculated the difference $I_{\mathrm{diff}}=I_{\mathrm{E}}-I_{\mathrm{U}}$ to eliminate all spatial information of the individual frames such as illumination inhomogeneities. We calculated the noise in $I_{\mathrm{diff}}$ using the standard deviation ${\sigma }_{\mathrm{diff}}$. Since both individual images $I_{\mathrm{E}}$ and $I_{\mathrm{U}}$ contribute to ${\sigma }_{\mathrm{diff}}$ in equal parts, we can finally estimate the image noise for an iSCAT image averaging 2$n$ frames as ${\sigma }_n=1/\sqrt{2}{\sigma }_{\mathrm{diff}}$.

To estimate the image noise after flatfield correcting the raw iSCAT frames, we generated two extra images, $F{F}_{\mathrm{U}}$ and $F{F}_{\mathrm{E}}$, to mimic the real flatfield images. The stand‐in flatfield images were summed from $n_{\mathrm{FF}}=\{1,10,100,1000\}$ additional frames of the timeseries which were not already used to generate $F_{\mathrm{E}}$ and $F_{\mathrm{U}}.F_{\mathrm{U}}$ and $F_{\mathrm{E}}$ were then divided and subtracted as $F_{\mathrm{diff}}^{\ast }=F_{\mathrm{E}}/F{F}_{\mathrm{E}}-F_{\mathrm{U}}/F{F}_{\mathrm{U}}$ such that the resulting noise can be calculated via the standard deviation of $F_{\mathrm{diff}}^{\ast }$.

We note that additional background was present in real experiments which is described in Section S6.

5.3. Flow cell preparation and sample incubation

All flowcells used in this manuscript used #1.5 glass coverslips (Fisher Scientific) as substrate. Coverslips were sonicated in acetone and isopropyl alcohol for 5 min each and rinsed in distilled water before blow‐drying with nitrogen. The coverslips were plasma treated for 1 min and sealed into a custom hybriwell chamber (Grace bio‐labs). The flow cells have a volume of 10 µL with a rectangular center area of 6 mm × 8 mm with conical taper regions at the short sides (15 deg half angle) which leads to 1.5 mm diameter in‐ and outlets. The height of the flow cells is 150 µm and a total of five such flow cells are contained on each hybriwell chamber. Initial filling of the flow cells is aided by capillary pressure after manual pipetting into the flow cell inlet. Subsequent sample filling or washing of the flow cell was done by pipetting onto the inlet while gently pressing a folded lint‐free tissue (Kimwipe, Kimberly–Clark Professional) to the outlet.

Flow cells were filled with the respective buffer (see sample descriptions) before taking the pre‐incubation images. Ten microliters of sample was flushed into the flow cell and EV capture was monitored using iSCAT imaging (without flatfield processing). After the appropriate density was achieved (typically 1–3 min), the flow cell was flushed with 100 µL of buffer and the post‐incubation images were taken.

5.4. Calibration samples

Aminated silica nanoparticles with diameters of 22, 52, 77, 97 and 208 nm (SIAN20‐25M, SIAN50‐25M, SIAN80‐25M, SIAN100‐25M, SIAN200‐25M) were purchased from nanocomposix (San Diego, USA). Prior to use, the particles were sonicated and then diluted to approximate concentrations of $1\times {10}^9$ particles per mL in 2 mM acetate buffer at a pH of 5 where they attained positive surface charge.

Citrated gold nanoparticles with diameters of 10, 20, and 50 nm (A11‐10‐CIT‐DIH‐1‐1, A11‐20‐CIT‐DIH‐1‐1, A11‐50‐CIT‐DIH‐1‐1) were purchased from Nanopartz Inc. (Loveland, USA) and diluted in milliQ water before use.

NIST traceable Nanosphere PS Size Standards with diameters of 51, 81, and 100 nm (2050A, 3080A, 3100A) were obtained from Thermo Fisher Scientific (Mississauga, Canada) and diluted in milliQ water before use.

Liposomes were prepared from DOPC lipid (Avanti lipids, Alabaster, USA, cat. no. 850375) in chloroform solution which was spiked at a 1% molar ratio with DiI dye in ethanol. The solvents were evaporated in a test tube under vacuum overnight. The lipid film was resuspended in PBS to a molar concentration of 10 mM and allowed to swell and form large multilamellar bubbles for 30 min. The solution was then extruded by passing 21 times through a filter with pore size of 400 nm (Avanti lipids) and diluted further for the incubation.

5.5. Cell culture and EV isolation

Human colorectal cancer cells HT‐29 (ATCC), IMR‐90 (ATCC), and CD63‐GFP modified A431 cells (CD63 with C‐terminal fusion of GFP, a gift from the lab of Janusz Rak (Montermini et al., 2015)) were cultured in Dulbecco's modified Eagle Medium (DMEM) culture medium with 10% FBS and 1% (vol/vol) penicillin‐streptomycin. The cells were incubated in a humidified incubator at 37°C and 5% $\mathrm{C}{\mathrm{O}}_2$ and grown to about 60% confluency. The medium was then replaced by DMEM with 5% EV‐depleted FBS and the cell supernatant containing the EVs was harvested after $\sim 48$ h. The supernatant was spun at 400 × g for 15 min to pellet cell debris and filtered by 450 nm syringe filters to remove larger EVs. For each EV sample, 5 mL supernatant was concentrated to 500 µL using 100 kDa Amicon Ultra‐0.5 centrifugal filters (Millipore) for 5 min at 5000 × g. The concentrated supernatant was purified using 70 nm Gen2 qEVsingle size exclusion columns (IZON) with 150 µL input sample volume, 0.8 mL void volume after sample loading, and 0.6 mL sample collection volume.

5.6. CTR staining of EVs

EV samples of Figs. 5b, S10, and S11 were stained using 2 µL of Cell Tracker Deep Red dye (Invitrogen) at 1 mM stock concentration in 150 µL concentrated cell supernatant (from ∼12 mL starting volume) and incubated for 2 h at 37°C. EVs were then purified using IZON qEV.70 nm size exclusion columns as described above. I was found that this resulted in incomplete staining (data not shown), and the samples were subjected to a second round of staining using 0.5–1 µL of CTR in 500 µL purified EVs for 2 h at 37°C. Excess dye was removed in six washes with PBS using 100 kDa Amicon Ultra‐0.5 centrifugal filters (Millipore) for 10 min at 5000 × g (each spin concentrates 500 µL sample to ∼20 µL).

The data for Figure S12 was taken later and CTR staining was only performed on purified samples using 1 µL of CTR in 500 µL purified EVs for 2 h at 37°C. Unbound dye was removed with six PBS washes using 100 kDa Amicon Ultra‐0.5 centrifugal filters (Millipore) for 10 min at 5000 × g as above.

5.7. Biotinylation and AF647‐Streptavidin staining of EVs

Biotinylation was done using the EZ‐Link Sulfo‐NHS‐LC‐LC‐Biotinylation Kit (Thermo Fisher Scientific, Mississauga, Canada). 0.5 mg of Sulfo‐NHS‐LC‐LC‐Biotin was added into 500 µL of purified EVs (typically ∼10⁹ EVs per mL measured by NTA) and incubated for 30 min at room temperature. Unreacted biotin was removed in six washes with PBS using 100 kDa Amicon Ultra‐0.5 centrifugal filters (Millipore) for 5 min at 5000 × g. Subsequently, 1 µL of Alexa Fluor 647 labelled streptavidin at 1 µg/mL was added and incubated for 20 min. Excess streptavidin was removed in six washes with PBS using 100 kDa Amicon Ultra‐0.5 centrifugal filters (Millipore) for 5 min at 5000 × g.

5.8. Flow cytometry measurements

All FCM data presented here was taken using a CytoFLEX‐S Flow Cytometer (Beckman Coulter) equipped with 405, 488 and 638 nm lasers, and operated using CytExpert Software. The 405 nm violet laser was used for side scattering (VSSC) and the 488 nm laser for forwards scattering (FSC). VSSC gain was set to 200, SSC gain to 500 and FITC gain to 1000, the VSSC threshold was set to 1400. Before each sample measurement, the system was washed using Azide buffer at pH 5 for silica nanoparticles, ultrapure water for PS calibration beads and PBS for EV samples. An event rate below 1000 per second was ensured for these buffer controls. Samples were loaded and after the event rate became stable and measurements were taken for 60 s each.

The size calibration (Figure S15) was done in FCMPass (Welsh et al., 2020) v. 4.2.14 using 81, 100, 152, 203, 240, 303, 345, 401 and 453 nm polystyrene beads (3080A, 3100A, 3150A, 3200A, 3240A, 330A, 3350A, 3400A, 3450A) purchased from Thermo Fisher Scientific (Mississauga, Canada). The calibration was done using the default Average RI core‐shell model with a 5 nm thick shell of refractive index 1.4863 and core refractive index of 1.3859.

5.9. NTA measurements

NTA measurements were performed on a Nanosight (Malvern Panalytical, Malvern, UK) NTA 3.4 using a 488 nm laser. Five measurements of 60 s each were taken per sample at a syringe pump speed of 40 at 25 frames per second and detector threshold of five. The camera level was set to 14 for the SEC sample and 10 for the 10,000 × g UC sample to keep overexposure below 10% of traces.

5.10. Negative stain TEM imaging

Copper grids were negatively charged for 20 s at 20 mA and 5 µL of EV sample was incubated without dilution after SEC purification for 10 min. The grids were washed three times in ultrapure water for 30 s each and fixed in 2% glutaraldehyde solution for 2 min followed by additional three washes in ultrapure water for 30 s each. Staining was performed with a 2% uranyl acetate solution for 45 s, the grids were blotted dry using filter paper and residual moisture was left to evaporate for 1 h. TEM images were taken on a FEI Tecnai 12 at a working voltage of 120 kV.

AUTHOR CONTRIBUTIONS

Andreas Wallucks: Conceptualization (equal); data curation (equal); methodology (equal); software (equal); visualization (equal); writing—original draft (equal); writing—review and editing (equal). Philippe DeCorwin‐Martin: Conceptualization (equal); data curation (supporting); methodology (equal); software (supporting); writing—review and editing (equal). Molly L. Shen: Data curation (supporting); methodology (supporting); writing—review and editing (equal). Andy Ng: Funding acquisition (equal); resources (equal); writing—review and editing (equal). David Juncker: Conceptualization (equal); funding acquisition (equal); resources (equal); supervision (equal); writing—original draft (supporting); writing—review and editing (equal).

CONFLICT OF INTEREST STATEMENT

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All raw data supporting the main findings of this manuscript can be found at https://doi.org/10.5281/zenodo.10208005.

Supporting information

Supporting Information

Supporting Information

Supporting Information

Wallucks, A. , DeCorwin‐Martin, P. , Shen, M. L. , Ng, A. , & Juncker, D. (2024). Size photometry and fluorescence imaging of immobilized immersed extracellular vesicles. Journal of Extracellular Vesicles, 13, e12512. 10.1002/jev2.12512

DATA AVAILABILITY STATEMENT

Matlab scripts for evaluating SPFI data, including all image pre‐processing, alignment and spot localization algorithms detailed in Supplementary Sections 2, 3 and 4 are available via https://github.com/junckerlab/SPFI.