Red blood cell–derived extracellular vesicles
The presence and quality of REVs in the samples was investigated by FF-TEM and MRPS measurements. FF-TEM, by prompt freezing of the sample, enables the direct visualization of REVs’ morphology in their native structure (Fig. 1a). Spherical vesicles of 150–250 nm diameter with speckled surface can be observed in the TEM picture.
The size distribution and concentration of REVs were measured by MRPS [20]. Measured data points were fitted by a log-normal distribution (Fig. 1b) with a mean diameter of 195 ± 0.6 nm (adj. R2 = 0.979; μ = 198 ± 0.6 nm and σ = 35.7 ± 0.6 nm). The measured concentration is (1.44 ± 0.01) × 1010 particle/mL over a size range from 65 to 400 nm.
IR spectra of EV samples
Difficulties in the analysis of EV samples by IR spectroscopy may arise from the usually low EV concentration of the samples and the presence of contaminants such as lipoproteins, protein aggregates, etc. with similar IR features as EVs. However, the protein-to-lipid ratio might be a quality control of EV purity [8, 18]. Based on the ratio of the amide and C–H stretching band intensities, the “spectroscopic protein-to-lipid ratio” can serve as a useful index for EV characterization [17]. Figure 2 shows a typical IR spectrum of REV samples. The use of the dry film technique with ATR configuration produces high-quality spectra of EVs with clearly defined spectral features [17, 27]. Characteristic protein bands of amide A, amide I, and amide II can be found at 3298, 1657, and 1546 cm−1, respectively. Vibrations corresponding to lipid components can also be witnessed as antisymmetric and symmetric methylene stretching of acyl chains at 2924 cm−1 and 2850 cm−1, respectively, and the C=O stretching at 1738 cm−1 of the glycerol esters. In the low-frequency region, however, the phosphate bands of PBS buffer are dominating the spectrum due to the low EV concentration. It should be noted that the buffer has also a contribution to the amide I band (Fig. 2, dotted line spectrum) so any quantification based upon amide I band intensity might be corrected by proper buffer subtraction. The “spectroscopic protein-to-lipid” ratio calculated as the ratio of integrated areas of the amide I band and that of the C–H stretching (from 3040 to 2800 cm−1 wavenumber region) resulted in a value of 1.2 ± 0.1 for REV samples, which is in line with our previous results [20].
Protein standard plot
As protein amide I and amide II are usually well-defined IR bands of EV samples, protein quantification based upon a simple univariate (Lambert-Beer) analysis is feasible. The integrated area of amide I is related to the protein concentration through a calibration curve. We have to note that recent papers dealing with ATR-FTIR spectroscopy–based protein quantification relay on multivariate analysis [16]. Despite the general trend of solving analytical calibration problems by multivariate statistics and chemometric methods [28,29,30], the peculiar feature of intact EV samples, the novelty of the IR spectroscopy in the EV field, and the effort towards a simple and standardisable EV quantification method rationalize the use of a simple univariate calibration method [31, 32]. All IR spectra of the reference protein and REV samples were treated with the pre-processing and analysis protocol described in the “Materials and methods” section and illustrated in Fig. 3. Bovine serum albumin (BSA) was used as protein standard for the generation of the calibration plot. In line with the expected EV concentration, 1 mg/mL BSA dissolved in PBS buffer was used as stock solution and a dilution series (by sample bisections) was applied to create a calibration curve for the protein concentration determination (Fig. 4a and b).
Good linearity was obtained in the protein range between 0.031 and 1 mg/mL (covering almost two orders of magnitude) with adjusted R2 value of 0.998 (Fig. 4b). The last dilution point (0.016 mg/mL) was not considered for the calibration curve. Indeed, the bottom spectrum (Fig. 4a) does not show distinctive amide I and amide II bands which would be needed for IR-based protein quantification [33]. The linear relationship between concentration and the amide I AUC was validated by the experimental F value corresponding to the ratio of residual variance to squared pure error [30]. Calculated for a heteroscedastic case, we obtained an F value of 0.7886 which proved to be significantly lower than the critical F value of 2.6896 [34], reconfirming the linearity of our calibration curve (Electronic Supplementary Material (ESM) Fig. S1). Sensitivity of the method was verified by calculating the limit of detection (LOD) and the limit of quantification (LOQ) involving a certain risk of false positives (false detects, α-errors) and false negatives (false non-detects, β-errors) [30, 35]. Working at 95% confidence level, the probability of α-errors and β-errors, α and β, respectively, are usually reasonable small values (α = β = 0.05) and the univariate LOD and LOQ can be expressed as
$$ LOD=\frac{3.3 Sy/x}{slope}\sqrt{1+{h}_0+\frac{1}{n}}\kern0.5em LOQ=\frac{10 Sy/x}{slope}\sqrt{1+{h}_0+\frac{1}{n}} $$
[30, 35], where Sy/x is the residual standard deviation, n is the number of calibration samples, and h0 is the leverage for the blank sample, calculated from the mean calibration concentration. LOD was found to be 0.03 mg/mL, while LOQ 0.08 mg/mL. These values are somewhat higher compared with colorimetric-based protein quantification methods (under ideal circumstances the detection limit of Bradford and BCA assays is at 0.02 mg/mL [36]), however, it is acceptable for ATR-FTIR-based protein quantification [12, 15].
Adaptation and validation for REV samples
Proteomics of REV samples identified more than 200 different proteins for ex vivo excreted EVs from stored red blood cells [37]. However, since the FTIR-based protein quantification is related to the number of amide bonds (on mass basis) a single protein could serve as an adequate reference for any other proteins [13]. Furthermore, noting that the band 3 protein, one of the key protein of RBCs and REVs [37], has a molecular weight of 93,000 g/mol and contains 833 amino acid residues [38], we can calculate an approximate peptide bond concentration as 0.00895 per volume (L). This value coincides well with that of BSA being 0.00912 per volume (L). As further proof, Strug and co-workers [13] compared three different types of proteins (BSA, protein A, and rabbit IgG) and they obtained a similar slope for amide I intensity versus protein concentration, independent of protein sequence content.
In order to test the adaptability of the method to EVs, the linearity of amide I band area (AUC) with REV concentration was verified. Figure 5a shows the amide I and amide II wavenumber region of the pre-processed IR spectra of REV samples after different dilutions.
The dilution curve (Fig. 5b) calculated from the integrated area of the amide I bands shows a very good linear dependence upon dilution (adj. R2 value of 0.992). Again, performing the F test [34] the obtained F value of 0.2924 compared with the critical F value of 4.4590 corroborates the linearity (ESM Fig. S2). Upon dilution, the estimated uncertainty of the data points decreases and the standard deviation yields SD values under 1% proving the precision of measurements. For the undiluted REV sample, we obtained a higher standard deviation with an SD value of 23.27%, but which is still acceptable. The reason for the data precision failure is, that around 1070 cm−1, dominated by the phosphate vibration of the PBS buffer, EV components might have also minor contributions (e.g., the stretching vibrations of the phosphodiester groups of phospholipids; the C–O–C stretching vibrations of phospholipids, triglycerides, and cholesterol esters) [17, 39]. The iteration-based spectral subtraction protocol aims to overcome this adverse effect; however, the PBS subtraction is still the bottleneck of the process. However, this effect is marginal form a practical point of view due to the usually low concentration of EV samples.
The proposed EV protein quantification method is further validated by a spike-and-recovery test. Due to their relatively large surface-to-volume ratio, EV samples might be prone to adsorb proteins that are present in the biological matrix during EV preparation. For example, blood plasma–derived EVs might have substantial amounts of albumin as external protein cargo [40, 41]. To assess possible interference of REVs with albumin, a known amount of REV sample was spiked with low, medium, and high concentration of BSA standard solution (0.0625, 0.25, and 1 mg/mL BSA,respectively). To avoid undesired dilution, the REV sample was concentrated by centrifugation (16,000×g for 10 min) beforehand to obtain a higher initial concentration (1.07 ± 0.13 mg/mL) and was spiked with BSA solution in a volume ratio of REV:BSA = 8:2. Spike recovery calculated as (total concentration detected − concentration original)/concentration spiked × 100% [42] resulted in a mean recovery of 100 ± 5.2% for 0.25 and 0.0625 mg/mL BSA spikes. The mean recovery was lower (92 ± 5.4%) for the high amount of BSA spike; in this case, the total protein concentration is over the dynamic range of the calibration and saturation of absorption intensity of protein amide I band may occur. Within the dynamic range of the calibration, however, a good recovery rate was obtained providing that the interference between an external protein cargo and the proposed EV total protein quantification is negligible.
Multivariate modeling with PLSR
For the sake of thoroughness, we have also carried out multivariate modeling on the spectral dataset with the PLS regression method. Thirty-three BSA samples were included in the data matrix and the previously mentioned amide 1 and amide II regions of the spectra (1750–1450 cm−1) were selected for modeling. Additionally, the spectra were scaled with the MSC method. Threefold cross-validation with randomized sampling was applied in the validation phase of the modeling and four PLS components were included in the final model. The number of PLS components was determined by the local minimum of the root mean squared error of cross-validation (RMSECV) curve. The R2 values of the calibration (R2C) and validation set (R2CV) were 0.94 and 0.91, respectively. The error of the final model was given by the root mean squared error of calibration (RMSEC) and cross-validation (RMSECV): 0.08 mg/mL for the calibration and 0.10 mg/mL for the validation, respectively. The measured and predicted protein concentration for the validation set is reported in Fig. 6. A permutation test was also applied for the validation of the model. The result showed that the model significantly differs from the use of random numbers, based on a 100-iteration protocol.
The final model shows satisfactory performance, thus we predicted the total protein concentrations of the REV samples from three independent blood sources (donors) and independent isolations (REV1, REV2, and REV3). In the case of the REV1 sample, a dilution series was also prepared (Fig. 5).
Finally, the total protein content of REV samples was determined by the ATR-FTIR methods in the AUC-based and the multivariate (PLSR) workflows, as well. The results were compared with the outcomes of the Bradford and BCA assays (Fig. 7). The linearity of FTIR-based determinations was clearly shown by REV1 dilutions. By performing Tukey’s post hoc statistical test, we obtained only one significant difference between the results of the Bradford protein assay and the ATR-FTIR methods at a confidence interval of 95%. The two colorimetric assays, however, resulted in significantly different protein concentration values, especially in the higher concentration region (~ 50% less protein concentration measured by BCA). One possible explanation for the phenomenon should be the interference of bicinchoninic acid with phospholipids [43] and/or with reducing side chains present in REVs’ proteins.
Another issue is that due to the peculiar structure of EVs (stabile spherical vesicles) internal cargo proteins might be inaccessible to colorimetric dyes. Indeed, comparing the values obtained by the different methods, the highest protein concentration was measured by ATR-FTIR-based quantifications. Since IR spectroscopy is a label-free technique, it is not perturbed by the effectivity of dye-protein interaction. In the ATR-FTIR mode, the IR beam is directed through an internal reflection element (ATR crystal); an evanescent wave extends beyond the surface of ATR crystal and penetrates the sample. Since the penetration depth of this evanescence wave typically ranges from 1 to 2 μm within the 1800–900-cm−1 region [33], the whole protein content of REV samples (vesicles with average diameter of 200 nm) will be measured. A possible bias of the ATR-FTIR technique is related to the potential preferential adsorption of proteins onto the ATR crystal. REV sample disassembled by absolute ethanol was used to inspect colloidal aspects. The amide I areas calculated from the ATR-FTIR spectra of intact and lysed REV samples were similar (within LOQ range) even though the strong alteration in amide I band shape due to protein denaturation in the latter sample (ESM Fig. S3).