Correction: Addressing challenges in the removal of unbound dye from passively labelled extracellular vesicles

Correction for ‘Addressing challenges in the removal of unbound dye from passively labelled extracellular vesicles’ by Kaisa Rautaniemi et al., Nanoscale Adv., 2022, DOI: 10.1039/d1na00755f.


Introduction
Extracellular vesicles (EVs) have gained signicant attention as promising drug carriers for personalized nanomedicine over the recent decade. EVs are structurally heterogenous membrane-bound nanoparticles that are excreted by cells. 1 EVs have been associated with immune responses, 2 viral pathogenicity, 3 central nervous system related diseases, 4,5 and cancer progression. 6 Because of the natural origin of the EVs and their unique innate properties as natural cargo, the EV research foresees their potential applications as diagnostic or therapeutic tools for various diseases.
EV trafficking and their interactions with cells, tissues and in vivo are typically studied with uorescence-based microscopy methods. 4,[7][8][9] Furthermore, uorescent labelling of EVs can enhance the sensitivity and selectivity of EV characterization methods, i.e., ow cytometry 10,11 and nanoparticle tracking analysis. 12,13 There are several different approaches for the EV labelling; 14,15 the simplest and most commonly used method is the incubation of isolated EVs with lipid-tracer uorescent dyes, such as long-chain dialkylcarbocyanines, including DiI, DiD 8, [16][17][18] and PKH dyes. 9,10,17,19 This is referred to as passive loading of the dyes in contrast to active loading methods 20 and covalent dye graing. 21,22 The passive labelling is a relatively safe alternative among the other labelling methods due to the least effect on the natural EV structure. 23 Since the covalent labels oen react with the amine groups of the proteins at the surface of the EV membrane and active labelling requires extra chemical or physical treatments of EVs, the covalent or active labelling may lead to a reduction of EV unique intrinsic features and a deterioration of their potential as nanocarrier.
Being an equilibrium process, the passive labelling usually results in a mixture of the labelled EVs and an unbound dye in an aqueous solvent. 21,[23][24][25][26] Covalent labelling is also oen based on equilibrium chemical reactions and thus excess of reactive dye remains in the nal labelled sample. 21 Consequently, in both cases the unbound dye needs to be removed from the labelled EVs. The methods used for this include the same procedures as used for the initial EV isolation, such as ultracentrifugation, 9,19,27 various ltrations, 8,17 and sizeexclusion chromatography. 16,22 A successful removal of the unbound dye from the EVs is a crucial step for most EV studies, as the uorescent dye not associated with the EVs will likely compromise the research outcomes or even lead to wrong conclusions of the EV functionality. 28 Although the challenges in unbound dye removal have been noticed and discussed earlier, there are no clear criteria for selecting a suitable purication method to remove the excess dye from the labelled EVs, and the success of the purication is rarely estimated. In a few studies, the purication success was conrmed by comparison with free dye controls, 21,24 while characterization of the labelled EVs was done by quite exotic methods which can hardly be available in all laboratories, e.g. asymmetric-ow eld-ow fractionation coupled to a multiangle light-scattering detector, 24 or nanoscale uorescence analysis and cytometric sorting. 21 Although the above methods bring important and convincing information, they can be too laborious for screening of multiple labels, labelling conditions, and suitable purication methods. That is why there is a need for a simpler pre-screening protocol for an assessment of the purication aer labelling. In some papers, the EV labelling efficiency was estimated as a relation of the recovered amount of label either to the recovered amount of EVs, 20 or to the protein content in the labelled EV preparation. 29 They however lacked an evaluation of the purication efficiency aer labelling which clearly determines the accuracy of the labelling efficiency estimation. Conclusively, the lack of a standard way to characterize the purication efficiency makes published work difficult to compare. Therefore, there is a clear need for a systematic comparison of the purication methods. 15 The present work is our attempt to full this urgent need.
In this study, we focused only on purication method selection and estimation of purication efficiency and did not study EV labelling efficiency. Five widely available purication methods were studied for their ability to separate uorescently labelled EVs from unbound dye: ultracentrifugation (UC), ultracentrifugation with discontinuous density gradient (UCG), ultraltration (UF), size exclusion chromatography (SEC) and anion exchange chromatography (AEC). Five uorescent dyes were used for the passive EV labelling, resulting in 25 dye -purication method combinations. The overall workow of the study and the molecular structures of the studied dyes are presented in Scheme 1. First, the behaviour of the dyes and the EVs were studied separately with each of the methods to screen the potential methods offering sufficient separation between the dyes and the EVs. Based on these control results, potential purication methods were chosen for each dye, and they were subsequently applied to the labelled EVs. The relative purication efficiency was estimated based on the EV and dye recovery for each sample, and the results were compared for different purication methods and dyes. Finally, the labelled and puri-ed EVs were applied to cells and imaged by uorescence lifetime microscopy.
As clearly seen from the chemical structures of the dyes, they are all signicantly hydrophobic and, thus, should intercalate into the EV membrane in aqueous solution (Scheme 1D). DHPE-OG is a uorescent conjugate of a lipid molecule. Being able to label EVs with lipids would be benecial since they are natural components of the EV membrane and thus useful e.g., for adding targeting units to EVs. 30 DiO was chosen since carbocyanine dyes are commonly used in membrane staining. 31 BPC12 is a molecular rotor dye and its viscosity dependent uorescence could be used to study the integrity of the EVs during cell up take and consequent trafficking inside the cell. 32 BP is a neutral, nonpolar lipid stain. 33 Both these BP dyes are very hydrophobic and will thus intercalate deep into the EV membrane. 34 All the other studied dyes are at least partially at the hydrophilic part of the EV membrane and thus exposed to the surroundings. Hence, they can inuence the EV functionality and especially the cell uptake. 35 As an example of loading EVs with a biologically active molecule, the labelling was also studied with a Ptx-OG, a tubulin targeting anti-cancer agent Paclitaxel labelled with uorescent dye OG. 7,36 Being somewhat hydrophobic, Ptx part will intercalate into the EV membrane.

EV isolation
PC-3 cell line was obtained from the American Type Culture Collection (ATCC, USA) and cultured in the CELLine AD 1000 bioreactor (Sigma-Aldrich, USA) at 37 C and 5% of CO 2 . The cell culture compartment was lled with 15 ml Advanced DMEM/F-12 glucose (4.5 g ml À1 ) and L-glutamine (2 mM), while the media compartment was lled with 750 ml Ham's F-12k medium supplemented with 10% fetal bovine serum (FBS) and glucose (4.5 g ml À1 ). Cell culture compartment and media compartment are devised by a semi-permeable membrane (molecular weight cut-off 10 kDa), which permits a continuous nutrient diffusion and waste elimination. The membrane prevents the diffusion of EVs and large proteins from one compartment to the other. Cell culture media, FBS and L-glutamine were purchased from ThermoFisher Scientic (USA) and the glucose from Sigma-Aldrich (USA).
Next, the EV pellets were resuspended in DPBS (Dulbecco's phosphate buffered saline, Sigma-Aldrich, USA) buffer and the EVs were further puried with three-layered (0% to 35% to 45%) discontinuous iodixanol density gradient. The EV suspension was mixed with iodixanol (Optiprep™, Alere Technologies AS, Norway) to a nal volume of 2 ml and a 45% iodixanol concentration and loaded to the bottom of the density gradient. 4 ml of 35% iodixanol was placed over the bottom layer, and the rest of the tube was lled with DPBS. The gradient was centrifuged at +4 C with 200 000Âg (40 291 rpm, k-factor 129) for 3.5 h (13.2 ml, Open-Top Thinwall Ultra-Clear Tube, Optima L-80 XP ultracentrifuge with SW 41 Ti Rotor, Beckman Coulter, USA). Upon the centrifugation the EVs move from their original bottom layer (density of >1.215 g ml À1 ) to the top of the layer with the density of $1.195 g ml À1 with a velocity dependent on the cube of their size, 37 i.e., the EVs concentrate on the 0-35% iodixanol interface according to their buoyant density. DPBS on top of the gradient was discarded, and the EVs were collected from the 0-35% iodixanol interface. Finally, the iodixanol was removed by serial ultraltration (Amicon Ultra-15 Centrifugal Filter units, cut-off 10 kDa, Millipore, USA) at 5000Âg, 20 min, +4 C (Eppendorf Centrifuge, 5810 R xed-angle rotor Hamburg, Germany). The EV suspension was divided to aliquots of 10 11 particles and stored at À80 C until they were used for the experiments. In this study, the EV populations are classied as the 20k EVs and the 110k EVs, according to the centrifugal forces used for their isolation.

Characterization of the isolated EVs
The isolated EVs were characterized according to MISEV2018 standards 38 by western blot, transmission electron microscopy and ATR-FTIR as described in (ESI Section S1 †). The particle concentrations and the size distributions of the isolated EV fractions were analysed using the NanoSight LM-14 instrument (LCM14C, 405 nm laser, 60 mW, Nanosight, Salisbury, United Kingdom), equipped with the sCMOS camera (Hamamatsu Photonics K.K., Hamamatsu, Japan). The samples were diluted with DPBS and measured using a camera level 15 and an acquisition time of 90 s. Every sample was measured in triplicates. The resulting videos were analysed using the NanoSight NTA soware (NanoSight Ltd., v. 3.0) with a detection threshold set to 5. Noteworthy, the NTA instrument can detect only particles larger than 80 nm. Aer the NTA analysis, the EV samples were divided to aliquots of 10 11 particles and stored frozen in À80 C until they were used for labelling. Similar NTA analysis was done for all the labelled EVs and non-labelled EV controls aer every purication for estimating EV recoveries (Scheme 1A and C).
To be able to study the purication abilities of the chosen methods, the dye-to-EV ratio was chosen to have a signicant excess of the uorescent dye. EV aliquots of 10 11 particles were diluted to a nal volume of 0.5 ml in DPBS, and 10 À8 mol of dye was added to the EVs while mixing with a vortex mixer to prevent immediate aggregation of the hydrophobic dye in aqueous buffer solution. The nal DMSO concentration in the EV suspension was less than 2.5%. The labelling mixture was then incubated for 1 h at 37 C upon shaking while protected from light (Scheme 1C). All the EV labelling was done in triplicates.

Purication methods
Five widely available purication methods were studied for their ability to separate the labelled EVs from the unbound dye: ultracentrifugation (UC), ultracentrifugation with discontinuous iodixanol density gradient (UCG), ultraltration (UF), size exclusion chromatography (SEC), and anion exchange chromatography (AEC). In a successful UC purication, the EVs pellet during the centrifugation while the unbound dye stays mostly in the supernatant and can be carefully aspirated above the EV pellet. In UCG, because of their buoyant density and bigger size compared to uorescent dye molecules, the EVs oat to the higher interface between the layers with smaller densities while the unbound dye is expected to stay mostly at the bottom of the gradient in the highest density layer. Both UF and SEC are based on the size separation: in UF, the EVs concentrate above the lter membrane as most of the unbound dye is washed to the ltrate, and in SEC, the EVs elute from the column before the unbound dye. In AEC, the EVs rst bind to the positively charged column and the unbound dye is washed from the column; then, the EVs are released from the column with buffer containing NaCl.
2.4.1 Ultracentrifugation. The labelled EV suspension was diluted to a 1 ml nal volume with DPBS. Two different centrifugation parameters were used in accordance with the parameters used for the initial EV isolation: the 20k EVs were pelleted with centrifugation at +4 C with 20 000Âg (18 185 rpm, k-factor 289) for 1 h, and the 110k EVs were pelleted with centrifugation at +4 C with 110 000Âg (42 647 rpm, k-factor 53.0) for 2 h. The centrifugations were done by an Optima MAX ultracentrifuge equipped with the MLA-130 xed angle rotor in 1 ml Open-Top Thickwall polycarbonate tubes (Beckman Coulter, USA). The supernatant was aspirated just above the formed pellet to avoid disturbing the rather loose pellet, and the EVs were allowed to disperse in 70 ml of DPBS overnight at +4 C.
2.4.2 Ultracentrifugation with density gradient. The UCG purication was done similarly as in the EV isolation step. The centrifugation was done at +4 C with 200 000Âg (40 291 rpm, kfactor 129.0) for 3.5 h by an Optima XPN ultracentrifuge, equipped with a SW 41 Ti swinging bucket rotor (Beckman Coulter, USA). Aer collecting 1 ml fractions (a total of 13-14 fractions), the uorescence of the collected fractions was measured for identifying the EV and the free dye fractions. The uorescence experiments were done either by a spectrouorometer (FluoroLog-3, Horiba Scientic, Japan) or by a uorescence plate reader (Fluorescent Ascent FL, Thermo Scientic, USA). Based on the NTA results at EV isolation step (Section 2.1) and in the EV controls (Section 2.5), the highest particle concentrations were expected to be in the interface between DPBS and 35% iodixanol. Fluorescence intensity registered in 1-2 fractions on this interface veried their choice as the EV fractions. These fractions were either pooled and the iodixanol was removed by serial ultraltration as described in Section 2.4.3 (UCG + UF), or the pooled EV fractions were analysed without removing the iodixanol (UCG).
2.4.3 Ultraltration. The labelled EV suspension was diluted to a nal volume of 4.5 ml with cold DPBS and concentrated with a 10 kDa membrane (Microsep Advance Centrifugal Device, Omega membrane) by centrifugation with Sigma 4-16KS centrifuge (Sigma Laborzentrifugen GmbH, Germany) at +4 C. The centrifuge parameters were adjusted to yield a maximum nal volume of 0.5 ml aer each washing round. The ltrate was collected, and the washing was repeated a total of 3-6 times. Aer the serial ultraltration, the EVcontaining sample was carefully collected above the membrane by gentle pipetting.
2.4.4 Size-exclusion chromatography. The labelled EV suspension was run through a Sepharose CL-2B (Cytiva, USA) column (diameter 1 cm, bed size $13 ml) using DPBS as an eluent. All the runs were performed at room temperature. Starting directly aer the sample insertion, eluted buffer was collected in 1 ml fractions. A total of 30 fractions were collected for each SEC run. The EVs eluted typically in fractions 4 and 5 (Section 2.5). Aer every SEC run with the uorescent dyes, the dye retained in the column was washed with 25 ml of 0.1% Triton X-100 (Surfact-Amps X-100, ThermoFisher Scientic, USA) in 0.1 M NaOH, followed by extensive washing with Milli-Q water. The uorescence of the collected fractions was measured with a spectrouorometer for determining the uorescent dye distribution and identifying the fractions containing the labelled EVs for further analysis.
2.4.5 Anion exchange chromatography. A 1 ml HiTrap® DEAE Fast Flow anion exchange column (Sigma-Aldrich, Germany) was used. The column was connected to the NGC Quest Plus chromatography system (Bio-Rad, USA) equipped with a fraction collector (kept at +4 C) and a 1 ml sample injection loop. Two running buffers (A and B) at pH 7.5 were prepared. Buffer A was composed of 30 mM Tris-HCl (Sigma-Aldrich) and buffer B of 30 mM Tris-HCl and 1 M NaCl (Sigma-Aldrich). All the runs were performed at room temperature with the running protocol presented in Fig. 1. To reduce the non-specic binding of EVs to the column, the column was treated with 2.5% (w/v) BSA solution (Sigma-Aldrich) in buffer B before each run. 3 ml of BSA in buffer B was injected into the column and incubated for 30 minutes, followed by washing with 15 ml of buffer B at 1 ml min À1 and 8 ml of buffer A at 1 ml min À1 . 0.5 ml fractions were collected during the injection (4 fractions, 1-4) and elution (8 fractions, [5][6][7][8][9][10][11][12] phases to analyse the ow-through as well as the eluent (Fig. 1). Aer each run, the column was washed by injecting 3 ml of 0.1% Triton X-100 -1 M NaCl followed by 10 ml of buffer B at 1 ml min À1 . The EVs eluted typically in fractions 6-12, with peak concentration in fraction 7 (Section 2.5) as identied with NTA analysis. Similarly to SEC, the uorescence of the collected fractions was measured with a spectrouorometer for determining the dye distribution.

Control purications
Before using any of the purication methods for the EVs incubated with a uorescent dye, the behaviour of the dye was studied without the EVs (Scheme 1A). Conversely, the EV behaviour and recoveries in all the methods were studied also without the uorescent dyes (Scheme 1B). The dye and EV controls were used to select the methods with potential to separate the unbound dye from the EVs. Dye controls were prepared by diluting 10 À8 mol of the dye to a nal volume of 0.5 ml in DPBS; similarly, the non-labelled EV controls were prepared by diluting EV aliquots of 10 11 particles to a nal volume of 0.5 ml of DPBS (Scheme 1). Both control samples were then incubated and puried in the same way as the labelled EV suspensions. The EV controls were performed in triplicates and the dye controls once.
In the EV controls, the expected EV fractions were collected and analysed with NTA for determining the EV recoveries (Section 2.6). In the chromatography methods, the EV containing fractions were identied by analysing fractions 1-10 (SEC) or 1-12 (AEC) with NTA for the rst of the EV control replicates. Based on the results for the rest of the replicates only the EV containing fractions were analysed. With UCG, the majority of the unlabelled EVs were in the 0-35% iodixanol interface. In higher-density fractions, the scattering background from iodixanol made the NTA analysis unreliable, and therefore those fractions were not analysed with NTA.
For the dye controls, uorescence was used for studying whether the dyes behave as expected for a successful purication. For the UC and UF dye controls, the dye recovery R dye,c was estimated by comparing uorescence in the expected EV fraction (resuspended UC pellet and UF retentate) to the total uorescence of the sample, measured with the spectrouorometer: where V EV is the volume and I EV the uorescence intensity of the expected EV fraction, and V b and I b correspondingly the volume and uorescence intensity of the buffer that is not expected to contain the EVs (UC supernatant and UF ltrate). For the fraction-based methods (UCG, SEC and AEC), the potential separation was determined by identifying dye-containing fractions by uorescence experiments and comparing these to the corresponding EV controls. The uorescence was measured either with a uorescence plate reader (UCG controls for BPC12, DHPE-OG and Ptx-OG) or with a spectrouorometer (all the remaining controls).

Characterization aer purication
Aer the purication, all the EV samples were divided into aliquots and stored at À80 C for the characterization (Scheme 1C). The particle concentrations of the EV suspensions recovered aer labelling and purication were measured with NTA to obtain EV recovery (R EV ): where N f is the nal number of particles recovered aer the purication and N i ¼ 10 11 is the number of particles initially used for EV labelling before the purication process. The dye concentration in the puried samples was measured against a dye calibration curve either with a plate reader (Ptx-OG, DHPE-OG, BPC12, DiO, BP UCG and BP AEC) or with a uorescence spectrophotometer (BP UC and BP UF). The uorescence of the samples with known amount of a dye were measured to form the calibration curve. The dyes were released from the EVs and solubilized in the buffer by adding 1% Triton X-100 to both the EV and the calibration samples. Ptx-OG was hydrolysed to its fully uorescent form by adding sodium hydroxide (0.015 M nal concentration) together with 1% Triton X-100 to both the EV and the calibration samples and incubating at 37 C for 1 h before the measurements. Dye recovery R dye was calculated as where n f is the molar amount of dye le in the recovered EVcontaining sample, and n i is the molar amount of the dye initially added to the EV suspension for the labelling.
To compare the purication result between the different methods and the different dyes, relative purication efficiency E rp was calculated as Given that a signicant excess of dyes was always used, E rp < 1 indicates that the method concentrates the unbound dye more efficiently than the EVs and therefore it is not suitable for the purication of the EVs aer uorescent labelling. For the methods yielding E rp > 1, the relative purication efficiency was used for comparing the suitability of the purication methods for each dye.

FLIM imaging of cells incubated with EVs
One day before imaging, 10 000 PC-3 cells were seeded in 70 ml 2-well inserts (Ibidi, Germany) attached to a glass-bottom 35 mm Petri dish (poly-D-lysine coated, no. 1.5 coverslip, 10 mm glass diameter, MatTek, USA), or 75 000 cells were seeded directly on the Petri dish. Cells were incubated with labelled and puried EV sample (30 000-400 000 particles/ seeded cell) for 3 hours, washed once with DPBS and imaged in FluoroBrite DMEM (Gibco, USA) supplemented with 10% (v/ v) FBS. Samples for the free dye control were prepared similarly, using free dye instead of the labelled EVs. Similar amount of free dye was added to the cells as would have been added with 30 000 particle/cellratio.
Fluorescence lifetime images were acquired using the uorescence lifetime microscope MicroTime-200 (PicoQuant, Germany) coupled to the inverted microscope Olympus IX-71 (Olympus, Japan) equipped with 100Â oil objective (NA ¼ 1.4). The pulsed laser diode LDH-P-C483 (PicoQuant, Germany) emitting at 483 nm (time resolution 120 ps) was used for the excitation and the emission was monitored on wavelengths 510-900 nm. The samples were imaged at 37 C and 5% CO 2 using an objective heater (TC-1-1005 Temperature Controller, Bioscience Tools, USA) and a custom-made incubator. The FLIM images were analysed in SymPhoTime 64 soware (Pico-Quant, Germany). The colours of the FLIM images are based on the mean arrival times of the emitted photons aer the excitation pulse (fast lifetime). The intensity-averaged lifetimes s av were obtained by 2-or 3-exponential lifetime tting 7 (DHPE-OG and DiO, respectively) of the decay curves of the selected regions of interest, excluding the cell autouorescence background.

EV characterization
Two different subpopulations of PC-3 EVs, 20k and 110k, were isolated from conditioned media by differential ultracentrifugation, followed by density gradient centrifugation. The isolated EVs were characterized by NTA, TEM, FTIR and WB analysis.
The detailed characterization is presented in ESI Section S1. † Briey, the isolated EVs were conrmed to be of high purity having consistently similar properties over each sample replicate regarding size, enrichment of the EV-associated proteins and overall biochemical composition according to FTIR spectroscopy. The most apparent differences between 20k and 110k EVs were that 20k EVs were larger on average, which is in line with their consequently faster sedimentation during 20 000Âg centrifugation, and the more intense peaks in the 1040-1110 cm À1 region in the FTIR spectra for the 110k EVs.

Control purications
3.2.1 Recoveries of non-labelled EVs. First, the EV recoveries aer each purication method were measured without uorescent dyes (Scheme 1A) by NTA (Table 1). These nonlabelled EV controls were prepared and incubated similarly as the uorescently labelled EVs. Regardless of the purication method, a signicant number of the EVs was lost in the puri-cation process and the R EV between replicates had high variation. The highest R EV values were obtained by UCG and AEC: both methods were able to recover over 45% of the EVs. Without BSA blocking of the column, the AEC yields were also low (approximately 10%, data not shown). This indicates that the non-specic binding of the EVs to the column was signicantly reduced by BSA blocking. In UC and SEC, the R EV is moderate, from 10% to less than 35% in average. In contrast, the R EV was the lowest when the gradient centrifugation was followed by the ultraltration step (UCG + UF). This indicates high EV binding to the lter membrane, which is also reected in the R EV aer direct ultraltration (UF). Because of the very low EV recoveries in UCG + UF (about 1%), the UCG purication of labelled EVs was studied only without removing the iodixanol. Interestingly, an absorption peak characteristic of iodixanol was observed in the AEC during the sample injection (1-5 ml, data not shown), indicating effective removal of residual iodixanol from the EVs.
3.2.2 Dye distributions in control purications without EVs. For understanding the purication potential of a particular method, it is important to know what happens to the unbound dye in the purication process. Thus, the behaviour of each dye in the absence of EVs was evaluated for all the purication methods (Scheme 1B). The expected behaviour is schematically presented in Fig. 2.  Fig. 2A and C). For the UC and UF dye controls, the dye recoveries in the expected EV fractions (UC pellet and UF retentate) were estimated from the uorescence intensity of resuspended pellet and supernatant (UC, Fig. 2A), or from the sample collected above the lter membrane and the highest concentration ltrate (UF, Fig. 2C) according to eqn (1). Thus, the optimal R dye,c -value would be $0 for the dye controls. The obtained R dye,c -values are presented in Table 2.
DiO and Ptx-OG had virtually no uorescence in the UC supernatant and UF ltrates, while the emission intensity in the resuspended UC pellet and the UF retentate was high. The result clearly shows that these dyes aggregate and, therefore, they are efficiently pelleted during UC and do not pass the lter membrane in UF. Consequently, these methods are not suitable for removing DiO or Ptx-OG from the labelled EVs.
BP and DHPE-OG gave better control results. DHPE-OG uorescence was still clearly detected both in the UC pellet and the UF retentate (R dye,c ¼ 3-9%), while almost no BP uorescence was observed (R dye,c ¼ 1-2%). Because of the low dye recoveries, BP was chosen as a model dye for the UC and UF purications.
BPC12 stuck to the centrifuge tube walls during the ultracentrifugation and attached to the UF lter device membrane showing orange colouring on the tube walls and the lter membrane. The detected uorescence intensities were low even upon addition of surfactant (1% Triton X) to solubilize the dye, indicating strong surface adsorption. Since UC and UF did not work as expected with BPC12, the dye recoveries could not be reliably estimated. However, the possibility of using these methods to purify EVs from unbound BPC12 based on the dye adsorption on the surfaces was studied further with BPC12labelled EVs.
3.2.2.2 Ultracentrifugation with density gradient (Fig. 2B). In the UCG controls, most of the Ptx-OG, DHPE-OG and BP stayed in the bottom fractions (45% iodixanol) of the gradient (Fig. 3A). During the centrifugation, DiO formed visible crystals that did not concentrate in 45% iodixanol layer but had migrated further in the 35% iodixanol layer (Fig. 3A, pink). With all these dyes, there was also some emission at the interface between 0% and 35% iodixanol, where also the EVs concentrate according to the unlabelled EV controls. However, as most of the free dye stayed in higher density fractions and the EV recoveries before UF were high (Table 1), UCG purication was studied further for the EVs labelled with Ptx-OG, DHPE-OG, BP and DiO. In contrast, BPC12 emission was observed only in the fractions where EV accumulation is expected (Fig. 3A, yellow), implying that UCG is not suitable method to purify the BPC12 labelled EVs due to similarities in the dye and EV particle density. (Fig. 2D). Three types of dye behaviour were observed in the SEC controls. Both BODIPY dyes eluted in the same fractions as EVs (Fig. 3B, green and yellow). Most of the BPC12 eluted in fractions 4 and 5, indicating that it forms particles large enough for size exclusion to occur in the column. Separate NTA controls conrmed that BPC12 forms particles with about 90 nm diameter in DPBS (ESI Fig. S4 †). Although BP did not form particles visible by NTA (ESI Fig. S4 †), the highest concentrations of BP eluted in the fractions 4-7, indicating the formation of particles too small for NTA to detect, but large enough for the size exclusion effect, followed by a high background level in later fractions. Instead, both Oregon Green dyes had good separation from EVs and started eluting rst aer fraction 10 (Fig. 3B, blue and light blue). DiO, being the least water soluble of the dyes in this study, did not pass the column at all: instead, it crystallized and stayed on top of the column. Based on the control results, SEC   (Fig. 2E). Since both OG dyes (Ptx-OG aer hydrolysation) contain a negatively charged carboxylic acid group, their separation with AEC was expected to be challenging. Neutral dyes (BODIPY dyes) and positively charged dyes (DiO) were expected to elute in the sample injection step. Unfortunately, the AEC principle turned out to be unsuitable for the studied dyes. Based on the dye controls, all of the OG and BP dyes elute in the same fractions as the EVs (ESI Fig. S5 †). DiO had the most promising result, giving only negligible uorescence in the expected EV fractions. Although it was not removed in the sample injection step as expected, DiO crystallized similar to the SEC DiO control and did not pass the column at all. Consequently, IEC was studied only for the purication of the DiO-labelled EVs.

EV purication results
The chosen purication methods for each dye were applied to the labelled EVs. The EVs were incubated with a constant concentration of uorescent dye in DPBS as described in Section 2.3 and Scheme 1C. The EV and dye recoveries (eqn (2) and (3)) as well as the relative purication efficiencies (eqn (4)) were used to estimate the quality of the purication (Table 3). In general, labelling decreased the EV recoveries (Table 3) compared to the unlabelled controls (Table 1). DiO-and Ptx-OG-labelling reduced the recoveries the most with all the used purication methods. On the other hand, BP-labelled EVs had almost the same R EV aer UCG as the unlabelled EVs.
3.3.1 Relative purication efficiency. For the comparison of the purication results between different methods, both R EV and R dye were considered. For an optimal result, R EV should be as high as possible (close to 100%). The optimal value of R dye , however, is more difficult to estimate, as it depends on the labelling efficiency (how many dye molecules are actually located in the EV) and also on R EV . Consequently, the relative purication efficiency considering both recoveries simultaneously was used for comparing the purication results.
The dyes in this study were relatively hydrophobic, and thus they were expected to locate mainly in the EV membrane aer successful labelling. About 60 000 dye molecules were added per each EV in the initial labelling suspension. By estimating the EV membrane area and how many lipid molecules it can accumulate, 39,40 the dye-to-lipid molar ratio in the labelled EV suspension before purication was at least 1 : 3. As an example, for paclitaxel-loaded liposomes, drug-to-lipid ratios of 1 : 20-1 : 33 have been reported. 41,42 Such high dye-to-lipid ratios are unlikely reached by passive labelling, and it can be considered that a large excess of dye related to the EVs was used in the present experiments. Therefore, a successful purication process would remove most of the initial dye and a minimum requirement for the purication is that R dye < R EV , yielding E rp > 1. In other cases, the dye-to-EV ratio would actually increase during the purication.
3.3.2 DHPE-OG, Ptx-OG, and BP-labelled EVs. For both Oregon Green dyes, UCG and SEC gave the most promising dye control results, and therefore they were used for purifying the labelled EVs. For DHPE-OG EVs, E rp shows variation in the purication result: on average, UCG provided better purication for the 20k EVs and SEC for the 110k EVs. In the SEC purication of 20k EVs, E rp < 1 for the averaged values although two of the three samples gave E rp > 1.3 (ESI Table S2 †), which is well in line with the UCG result. This clearly shows the importance of checking the purication quality separately for each sample.
Although the purication efficiencies were similar for both methods, the EV yields were higher aer UCG (>40%) than SEC purication (about 10%). Table 3 EV recoveries R EV , dye recoveries in the EV fractions R dye , and relative purification efficiencies E rp for the labelled and purified EVs. The removal of unbound dye was studied with ultracentrifugation (UC), ultracentrifugation with density gradient without ultrafiltration (UCG), ultrafiltration (UF), size-exclusion chromatography (SEC), and anion exchange chromatography (AEC). The individual values for each replicate are presented in ESI  In contrast to DHPE-OG, UCG did not separate the unbound dye from the Ptx-OG-labelled EVs. Only 10% or less of the EVs were recovered, while these fractions contained 40-70% of the uorescent dye. Instead, the SEC purication gave promising and even reproducible results: although R EV was less than 4% for both EV types, E rp was 1.3 (110k EVs) and 2.8 (20k EVs).
For the UC and UF purications, BP dye control results were the most promising (Table 2). Unfortunately, UC did not remove the unbound BP from the labelled EVs efficiently enough during one centrifugation round: 17% (110k EVs) or 7% (20k EVs) of the dye was recovered while only 8% or less than 1% (respectively) of the EVs were collected aer the centrifugation. UF gave slightly better results in terms of E rp ; however, E rp < 1 suggests a poor separation of the labelled EVs from the unbound dye, and the EV recoveries were very low (1-2%). For BP, the best purication method was UCG. The BP-labelling did not reduce the EV recoveries compared to the unlabelled controls, while majority of the uorescent dye added to the EVs was removed during the purication. The UCG-puried BP EVs had the highest relative purication efficiencies of the studied samples, 12.7 (110k EVs) and 3.5 (20k EVs).

BPC12-and DiO-labelled EVs.
The uorescence spectra of the BPC12-and DiO-labelled EVs aer purication could be used to evaluate the purication result as in the course of the study, the uorescence signals of these dyes were found to be solvent-sensitive ( Fig. 4A and C). When dissolved in an aqueous buffer (DPBS) with surfactant treatment, the emission spectrum of these dyes is narrow, having the maximum at 520 nm for BPC12 and 510 nm for DiO. This corresponds to the monomeric dye uorescence. In DPBS without surfactants, the dyes are in an aggregated state, their emission is shied to longer wavelengths (>550 nm), and the peaks are broadened. This behaviour has been reported previously for several BODIPY derivatives. 43,44 The emission quantum yield for monomers of both dyes is much higher than that of the aggregates, and the spectra are therefore presented in a normalized form to visualise their shape. The solvent-sensitivity of the dyes is a useful property in the context of this study, as it could be used for a direct comparison of the purication results of the BPC12and DiO-labelled EVs.
As the unbound BPC12 forms aggregates in DPBS that have similar size as the EVs (Section 3.2.2, ESI Fig. S4 †), some of the particles detected by NTA are probably not EVs, and consequently R EV is not reliable for the BPC12-labelled EVs leading to a low E rp showing a failure of the purication. For the UC puried BPC12 EVs, a strong aggregate uorescence was observed (Fig. 4B, green and blue). The 520 nm emission peak relates to the monomeric dye in the EV membrane, while the longer wavelength emission (bands at 630 nm and 700 nm) relates to the dye aggregates in the aqueous buffer. The relative intensity of the aggregate peaks compared to monomer (EVrelated) peak is higher for 110k than 20k EVs, indicating higher concentration of the unbound BPC12 in 110k EV sample than in 20k EVs. The uorescence spectra of the UF-puried BPC12-labelled EVs (Fig. 4B, red and yellow) do not have as pronounced aggregate emission peaks as the UC-puried EVs, indicating more efficient removal of the unbound dye especially for the 20k EVs. However, the UF purication results had high variation, which is seen in the uorescence spectra (ESI Fig. S6 †) and is reected also in R EV and R dye (Table 3), suggesting that the uorescence spectra should always be checked aer the UF purication.
For DiO, UCG, SEC and AEC were studied to purify the labelled EVs. In UCG, the DiO control result showed only slight accumulation of the dye in the expected EV-fractions, while with the EVs, the dye accumulated almost exclusively in the EVfractions (Fig. 3C, pink). This could designate strong binding of DiO to the EVs. However, given the high excess of the dye used for the labelling, there must be unbound dye present in the EV fraction. The result indicates that the EVs have affected the DiO aggregation or the diffusion of the DiO aggregates in the gradient, leading to no separation between the unbound dye aggregates and the EVs. Indeed, the uorescence spectra of the UCG-puried DiO-labelled EVs (Fig. 4D, blue) conrms the presence of DiO aggregates in the EV fractions: monomer DiO peak at 510 nm is weak, and the spectrum is dominated by an aggregated DiO emission (bands above 550 nm). Compared to UCG, the SEC purication removed the unbound dye more efficiently. In the emission spectra of DiO EVs aer SEC puri-cation (Fig. 4D, green) the most intensive emission peak at 510 nm corresponds to the monomeric, EV-bound DiO. Nevertheless, there is still clearly a visible aggregate emission band above 550 nm. Additionally, the DiO aggregation seemed to lead to very low EV recoveries which varied from non-detectable to slightly over 1% (Table 3, ESI Table S2 †).
The third method studied for purication of the DiO-labelled EVs was AEC. Similar to the SEC dye control, DiO did not pass the AEC column. However, the DiO-labelled EVs passed the AEC column better than the SEC column: the average EV recoveries were 10% for 110k EVs and 6% for 20k EVs. The uorescence spectrum (Fig. 4D, yellow) shows only a monomeric EV-related emission peak at 510 nm, which is in line with E rp $ 3.5 for both EV types, indicating AEC as the best purication method for DiO-labelled EVs.

FLIM imaging of puried EVs with cells
A selection of labelled and puried EVs and corresponding free dyes were applied to PC-3 cells and imaged with FLIM to conrm the visibility and examine the distribution of the labelled EVs in the cells, and to compare the uorescence staining patterns to those of the free dyes. Examples of the FLIM images are presented in Fig. 5.
The FLIM imaging results demonstrate the importance of the purication step and underlines the difficulties in separating the free dye background from the uorescence signal originating from the EVs. In general, the uorescence intensities of the samples incubated with labelled and successfully puried EVs (Fig. 5A and B) were lower than the free dye controls (Fig. 5C and D Based on E rp ( Table 3), purication of DiO and DHPE-OG 110k EVs succeeded with AEC and SEC, correspondingly. The staining patterns of the EVs labelled with these dyes are spotlike inside the cells (Fig. 5A and B), which could be interpreted as the internalization of EVs into the endosomal pathway. However, the free DiO staining pattern and uorescence lifetime is very similar to the DiO-labelled EVs ( Fig. 5A and C); consequently, it is very easy to misinterpret the free DiO background as labelled EVs. For DHPE-OG, the difference in the staining patterns and uorescence lifetime is clearer (Fig. 5B  and D), indicating different uptake mechanisms of the free dye and EV-bound dye. Free DHPE-OG seems to stay attached to the cell membrane at least for 3 h of incubation with cells while EVbound dye enters the cell conrming the EV mediated dye internalisation. However, even if the labelling and purication of the EVs from the unbound dye would have been successful, the lipophilic dyes may leach from the EVs and stain other cellular membranes. 28,45 Our results also demonstrate that with a successful removal of the unbound dye, the EVs might not be detectable when applied to the cells. The emission intensity of the cells incubated with SEC-puried Ptx-OG EVs (30 000 EVs/cell) was close to the cell autouorescence, and due to low EV recoveries, intensity could not be increased by adding more EVs. The UCG-puried BP EVs gave best E rp values but were not visible even with 400 000 EVs/cell, suggesting that either the labelling efficiency has not been high enough for detecting the EVs or the iodixanol has a negative effect on the EV-cell interactions.

Discussion
The FLIM images of the labelled EVs (Fig. 5) clearly demonstrate that although uorescence signal is observed, it is not necessarily related to the EVs. Thus, for further reliable applications, e.g. in microscopy, it is extremely important to ensure both successful labelling and purication from the unbound dye with a dedicated method. To evaluate the purity of labelled EVs, we relied on a simple parameter, E rp to describe the approximate success of each purication protocol. The approach is simple and applies to the cases where the unbound dye is present in the system aer labelling or the dye-to-EV ratio is high enough. As described earlier, this is very oen the case for passive and covalent labelling. Moreover, it is clear that using a dye control solely is not sufficient to validate the purication success. The unbound dye behaviour may be affected by the presence of EVs as, for example, in the case of DIO in ultracentrifugation with gradient ( Fig. 3 A and C).
As the EV samples were initially isolated with a differential ultracentrifugation protocol prior to a density gradient, it may have also affected the EVs morphologically causing EV aggregation and shape distortion. The effect on the results presented in this study is difficult to estimate. However, the study is focused on screening purication methods, and the important point is that the initial EVs have been treated in the same way for all experiments.
Nonlabelled EV controls of a purication method are important as some of the methods themselves may result in almost complete EV loss. This makes the method uninteresting for labelled EV purication as in the case of consequent UCG and UF (Table 1). The comparison of nonlabelled and labelled EV recoveries also gives important information about the possible effects of the dye on EV recovery. Signicant reduction of EV recovery due to labelling questions the probe's suitability for EVs. Lastly, the parameter E rp , compares EV and dye recoveries aer purication. If the percentage of EV losses is higher than that of dye removal (E rp < 1) the method obviously does not perform as desired and indicates a signicant amount of unbound dye in the nal labelled EVs. The need to compare the EV recovery with the dye recovery becomes clear upon taking a look at Table 3, where a relatively good amount of recovered particles in EV samples does not always match with efficient dye removal. Thus, it turns out that only the combination of the controls together with the E rp metric leads to a comprehensive assessment of the tested purication method.
In this study, we relied on NTA for evaluating the EV recovery, assuming that labelling does not signicantly change the refractive index or size of the labelled EVs, excluding EV aggregation. Due to NTA's limits, not all the smallest EVs can be detected, and the method does not discriminate between EVs and other nanoparticles. However, the use of NTA in our attempt to suggest a protocol for evaluation of purication efficiency can be justied as we compare initially well-puried EVs using the same NTA settings and the same instrument prior and aer labelling and purication. Thus, the relative values, such as ratios of the concentrations measured by NTA, can be reasonable estimates of the EV recoveries. Moreover, E rp parameter represents a criterium of a purication success, i.e. can reect both successful (>1) or non-successful (<1) purication and should not be considered as absolute value in contrast to usually used labelling efficiency. The latter is meant to show a particular ratio of labels attached to a labelling target and is limited by 100% over which it becomes irrelevant.
Based on the above methodology (Scheme 1A-C), the purication methods were classied into three categories for each dye: good, unknown, and poor, ("+", "?", or "À") as presented in Table  4. Most of the methods were directly classied based on the negative dye control results, and the rest were classied according to the relative purication efficiencies (E rp ) of the labelled EVs. The methods which could reproducibly concentrate labelled EVs and remove unbound dye efficiently are marked with "+". According to the same logic, the methods that can recover more dye than EVs, and therefore cannot provide good separation of the labelled EVs from the unbound dye, are marked with "À". The third category, marked with "?", had high variation in the purication results or yielded E rp close to 1. These methods may be considered for removing these dyes, provided that an appropriate verication of the purication result is done separately for each sample. For the samples applied for live cell imaging, also the visibility of the EVs with cells is summarized in Table 4. As can be seen, Table 4 is quite empty and most attempted purication protocols failed. But encouragingly, it was possible to nd a method that at least partly worked for each dye applying the approach proposed in our study.
Interestingly, the molecular structure of the dyes has a clear effect on the purication results. Unbound BP and DiO were effectively isolated from the EVs: BP has a rigid molecular skeleton and DiO two hydrocarbon chains anchoring it to the EV membrane during the labelling. On the other hand, the dyes with freely rotating units, OG-chromophore in DHPE-OG as well as the BP-chromophore in BPC12, showed either poor purication or high result variation. Furthermore, the dyes do not always act according to the expected purication principle. Ptx-OG and DiO (Table 2) form particle-like aggregates in aqueous environments that are large enough to be collected by centrifugation and BPC12 precipitates into the tube walls. Thus UC, the most used protocol for the EV purication, 46 is not a suitable purication method for these dyes. BPC12 is also retained in the polyethersulfone-based lter membrane used in UF instead of passing it. In UCG, DiO and BPC12 gathered into the same layer as the EVs, probably due to aggregate formation. For SEC and AEC, DiO did not pass the columns at all.
From the studied methods, UC and UF were not efficient methods for purication of unbound dye from the labelled EVs. For UC, even with the dyes that stay mostly in the supernatant during the centrifugation, several UC concentration cycles would be necessary for efficient removal of the dye. However, since in most cases over 90% of the initial EVs were lost during a single UC run, no EVs would be le aer few centrifugations. According to our results, also UF leads to very low EV recoveries and is, thus, not suitable for the unbound dye removal. The recoveries aer UF might be improved by choosing different lter devices, 47 but a comprehensive study of these matters is missing and there is no univocal consensus on the best way to purify EVs with UF. 24,[47][48][49][50][51] The use of UF for removal of the iodixanol proved to be the limiting step of UCG as well. In our work, we observed high initial UCG EV yields of 60-80% before gradient removal, while using the common ultraltration protocol for removing the iodixanol from the EVs, we observed a heavy loss of EVs down to $1% yield (Table 1). Despite the wide usage of UCG for EV purication, methods to remove the iodixanol are not yet fully addressed. As iodixanol is a noncharged small-sized molecule, both AEC and SEC are promising candidate methods for the iodixanol removal.
The use of chromatographic methods for EV isolation and purication has raised a lot of interest, because they can be highly automatized and scaled up. 52,53 From the two chromatographic methods studied here, SEC showed more potential for removing the unbound dye from the labelled EVs, and is already a commonly used technique in EV eld with different ready to use solutions available on the market. 54 On the other hand, AEC was incompatible with the studied dyes, although the unlabelled EV recoveries were high. DiO-labelled EVs were successfully puried with AEC; however, the purication principle was rather related to the dye aggregation. With both chromatographic methods, the recoveries of labelled EVs were smaller than those of non-labelled EVs, suggesting that the dye may partly clog the column or increase EV binding to the column. The AEC method provided an interesting comparison for SEC as the EV losses were smaller in AEC than in SEC. This can be at least partially explained by non-specic binding of EVs to the column: the SEC bed size was more than 10-fold larger than the AEC bed size, therefore containing much more surface area for binding. Furthermore, saturating the AEC column with BSA signicantly increased the EV recoveries. The results indicate that the EV yields could be increased in the SEC by reducing the column volume and if feasible, by blocking the column with BSA.
Based on this study, it is difficult to say whether the dye is actually located in the EV membrane. Still, E rp is a useful tool for the pre-screening of the suitable dye-purication method combinations. As some dyes, such as PKH dyes 24 or BPC12 used here, tend to form dye aggregates with similar sizes as the EVs, detecting uorescent EV-sized particles is not sufficient proof Table 4 Summary of the purification and imaging results. Method marked with + had relative purification efficiency E rp > 1, high variation in the results or E rp z 1 are marked with ?, and À signifies negative dye control result or E rp < 1. For the samples that were imaged with FLIM, the visibility of the EVs is marked in the brackets (+: fluorescence was detected, À: no reliable fluorescence in cells)

Conclusions
Five common purication methods were tested with ve dyes for their ability to separate the passively labelled EVs from the unbound dye. None of the studied purication methods was suitable for all the studied dyes. Most dyes could be successfully puried with only one of the methods tested, suggesting that the successful purication method is related to the physical and chemical properties of the dye. Notably, in many cases the expected purication principle did not work. The unbound dye might form aggregates, bind to the purication matrices or gather to the same location with the EVs for some other reason. Thus, we suggest the following steps when working with a new dye to ensure the successful removal of the unbound dye: (1) studying the behaviour of free dye and unlabelled EVs separately with the chosen method to determine whether the method can effectively separate those from each other, (2) ensuring the separation still exists with the labelled EVs, and (3) evaluating the purication quality by E rp . Importantly, the labelled EVs need to have high enough uorescence intensity to be visible in the target application that is not always the case aer successful purication. The most promising methods for the used dyes were SEC and UCG, and the highest recoveries were obtained by UCG before removing the density gradient. With SEC, the purication is faster to perform involving only one separation step, making the SEC purication favourable over the UCG when there is a possibility to choose between the methods. However, both methods should be further developed: the gradient removal step aer UCG causes high EV losses and SEC purication should be still improved to yield higher EV recoveries. Sometimes also an unexpected purication principle may provide good purication results, as we demonstrated with AEC.
In this study, we used only EVs from a single cell line: the purication results with the same dyes but EVs from different sources might be different. We admit that the E rp values used in this study to evaluate the success of the purication of labelled EVs from unbound dye is an approximate method. However, the parameter in combination with the proposed controls allows to assess the purication outcome in a more comprehensive way compared to using only dye controls or labelling efficiency values. The proposed approach represents an easy methodology for initial pre-screening of multiple labelling conditions and purication methods.

Conflicts of interest
There are no conicts to declare.