Extracellular vesicle features are associated with COVID‐19 severity

Abstract COVID‐19 is heterogeneous; therefore, it is crucial to identify early biomarkers for adverse outcomes. Extracellular vesicles (EV) are involved in the pathophysiology of COVID‐19 and have both negative and positive effects. The objective of this study was to identify the potential role of EV in the prognostic stratification of COVID‐19 patients. A total of 146 patients with severe or critical COVID‐19 were enrolled. Demographic and comorbidity characteristics were collected, together with routine haematology, blood chemistry and lymphocyte subpopulation data. Flow cytometric characterization of the dimensional and antigenic properties of COVID‐19 patients' plasma EVs was conducted. Elastic net logistic regression with cross‐validation was employed to identify the best model for classifying critically ill patients. Features of smaller EVs (i.e. the fraction of EVs smaller than 200 nm expressing either cluster of differentiation [CD] 31, CD 140b or CD 42b), albuminemia and the percentage of monocytes expressing human leukocyte antigen DR (HLA‐DR) were associated with a better outcome. Conversely, the proportion of larger EVs expressing N‐cadherin, CD 34, CD 56, CD31 or CD 45, interleukin 6, red cell width distribution (RDW), N‐terminal pro‐brain natriuretic peptide (NT‐proBNP), age, procalcitonin, Charlson Comorbidity Index and pro‐adrenomedullin were associated with disease severity. Therefore, the simultaneous assessment of EV dimensions and their antigenic properties complements laboratory workup and helps in patient stratification.


| INTRODUC TI ON
Following the first retrospectively identified cases of severe acute respiratory syndrome-coronavirus 2 (SARS-CoV2) infection in humans, dating back to December 2019, 1 the novel coronavirus infection spread in a tumultuous fashion, and as early as March 2020, the World Health Organization (WHO) declared the coronavirus-induced disease (COVID-19) a pandemic. 2Only 1 month after the severe acute respiratory syndrome-coronavirus 2 (SARS-CoV2) outburst in Wuhan, two cases of COVID-19 were first reported in Italy, and the nation was in lockdown as soon as March 2020. 3Thanks to enormous scientific effort, at the end of 2020, the vaccination campaign started in Italy, and by the end of 2021, 108 million doses were administered, so that 88.7% of people older than 12 years had received at least one dose. 4Although the vaccine shows less protection against infection, it dramatically reduces the number of severe COVID-19 cases. 5However, the protection seems to wane over time. 5To add a layer of complexity, due to the high mutation rates of this RNA virus, different strains have emerged over the years, which were characterized by higher infectivity and either higher or lower pathogenicity with respect to the Wuhan strain. 6The emergence of new variants with the potential to evade the immune system poses a risk to the recrudescence of pandemics.
From a clinical point of view, there is a need to identify early biomarkers of adverse outcomes that could help clinicians rapidly discern, among patients with SARS-CoV2 infection and pneumonia, those that will evolve into a critical disease or succumb to COVID-19.For this purpose, we have already employed targeted 7 or high-throughput serum proteomics analyses, 8 demonstrating that inflammatory markers, including interleukin-6, are among the strongest predictors of patient outcome, even when analysed in combination with routine haematological, blood chemistry, demographic and clinical data of enrolled patients. 8tracellular vesicles (EV) are cell fragments enclosed in the plasma membrane that are produced in response to a variety of physiological and pathological stimuli.EV can be divided into three broad subgroups (apoptotic bodies, microvesicles and exosomes) based on their size, which partly reflects their mechanism of origin. 9,10EV can reflect ongoing pathological processes, such as apoptosis and viral infection, 10,11 but also play a very important role in physiological processes, ranging from coagulation to cell-cell communication, and regulation of inflammation and immune responses. 9,12An accumulating body of literature has demonstrated the potential involvement of EV in COVID-19 pathophysiology.Therefore, we assessed whether EV characteristics could be used as biomarkers for a critical disease.
For this purpose, we enrolled 146 patients who were affected by either severe or critical disease and analysed blood samples collected within 1 day (median) of hospital admission.Cohort data regarding demographic characteristics, comorbidities, routine haematology, blood chemistry, flow cytometric analysis of lymphocyte subpopulations and EV characterization were collected and analysed using machine learning algorithms.Elastic net logistic regression analysis with cross-validation showed that features of smaller EVs (i.e. the fraction of EVs smaller than 200 nm expressing either CD 31, CD 140b or CD 42b) were predictors of a better outcome, together with albuminemia and percentage of monocytes expressing human leukocyte antigen DR (HLA-DR).Conversely, the proportion of large EVs expressing either N-cadherin, CD 34, CD 56, CD 31 or CD 45 was associated with disease severity, together with interleukin 6, red cell width distribution (RDW), N-terminal pro-brain natriuretic peptide (NT-proBNP), age, procalcitonin, Charlson Comorbidity Index and pro-adrenomedullin.Taken together, our data suggest that EV may be employed as biomarkers for critical COVID-19.Furthermore, subsets of small EVs may promote a less aggressive outcome in SARS-CoV2 infection.

| Enrolled patients and clinical data
This study was authorized by the Regional Ethics Committee (2020-Os-033; Em.Sost.N. 1 versione 1 data, 16.08.2021)and conducted according to the declaration of Helsinki, and signed informed consent was collected from enrolled patients and controls.The inclusion criteria for patients were age >18 years and nasopharyngeal swab positivity for the SARS-CoV-2 genome.
We classified 'critical' and 'severe' patients according to the revised 2023 WHO clinical management assification. 13Relevant clinical, demographic, haematological, blood chemistry and flow cytometric data were collected from hospital electronic health records (INSIEL) and pseudonymized.

| Sample collection and preservation
Blood samples were drawn within 1 day (median value) of hospital admission and collected into 5-mL EDTA tube for plasma separation (Vacuette, Greiner Bio-One).Platelet-poor plasma (PPP) is obtained by centrifuging blood at 1400 × g for 15 min, followed by centrifugation at 14,000 × g for 20 min at +4°C for platelet depletion.The plasma was immediately stored at −80°C until use.Aliquots of PPP samples obtained from 10 patients were analysed using a clinical haemocytometer (Coulter) to assess the levels of platelet contamination.

| Extracellular vesicles isolation and analysis
To isolate small extracellular vesicles (EV) from PPP, we employed the ExoQuick Exosome Precipitation Solution (System Biosciences) following the protocol described in reference. 14To remove larger EV, we filtered them with a 0.2-μm syringe filter following isolation.
Nanoparticle tracking analysis was conducted employing Nanosight (LM10, Malvern System Ltd., London, UK), equipped with a 405-nm laser, as in reference. 14

| Flow cytometry analysis
To identify EV, we stained the cytoplasm with carboxyfluorescein succinimidyl ester (CFSE; Invitrogen).Directly conjugated antibodies against CD31 (Miltenyi Biotec), CD45 (Dako), N-cadherin, CD56, CD140b, CD34 and CD42b (BD Biosciences) were used to characterize EV subpopulations.CD63 was used to assess whether exosomes fell below the limit of detection in our system.Briefly, 15-mL PPP was incubated with 1:1 CFSE (40 mM in DMSO) for 2 h at 37°C.Then, 1.2 mL of each antibody was incubated for 30 min at +4°C.DMSO and isotype-matched antibodies were used as the negative controls.
The samples were analysed using a FACSCanto II flow cytometer (Becton Dickinson).EV sizes were estimated using side scatter (SSC), while flow cytometry sub-micron size (200-500-1000 nm) reference beads (Invitrogen) were employed to calibrate the instrument and quantify vesicle concentration.A minimum of 50,000 events were recorded per sample.Data were analysed using BDFACSDiva Software (Becton Dickinson).The gating strategy is illustrated in Figure 1.

| Statistics
Descriptive statistics for categorical variables are presented as number (percentage), and continuous variables as mean ± standard deviation (SD) or median (interquartile range [IQR]).Normality was assessed using the Shapiro-Wilk test.Comparisons between categorical variables were performed using the chi-square test or Fisher's exact test, as appropriate.Comparisons between continuous variables were performed using the t-test or Mann-Whitney U-test, as appropriate, and corrected for multiple testing using the False Discovery Rate correction method.An elastic net logistic regression algorithm was used to build different models to predict 'critical' or 'severe' outcomes.The elastic net logistic regression is a regularization model that combines, through a linear combination of LASSO and Ridge methods, both L1 and L2 penalties.This model performs variable selection, forming a subset of predictors, each of which is matched with a regression coefficient.Variable importance was ordered using the absolute value of the regression coefficient; a higher value indicated a larger contribution to the model.The dataset was split into testing (30%) and training (70%) sets.A 10-fold cross-validation was applied to the training set to tune the hyperparameters λ, determine the amount of shrinkage and α and explain the presence of L1 and L2 penalties.The model was trained separately for the demographic, comorbidity, haematological, blood chemistry, EV and PBMNC data.Furthermore, all the variables of the training set with a corrected p value < 0.05 were included in the final model.
Finally, we added the following clinically relevant confounders to the complete model: lymphocyte relative counts, together with the fraction of T CD4 + helper lymphocytes and activated CD14 + HLA-DR + monocytes.The performance of these models was evaluated on the testing set using the receiver operating curve (ROC) and the area under the curve (AUC), with a 95% confidence interval (CI), and compared using Long's test.Correlations between the variables selected by the final model were investigated using a correlation heatmap with the Spearman method.Analyses were performed using STATA version 17.

| Demographic and clinical characteristics of enrolled patients
A total of 146 patients were enrolled between November 2020 and April 2021, and B.1.177and Alpha (B.1.1.7)variants were the most prevalent genotypes in Italy. 15All enrolled subjects were positive for SARS-CoV-2, and samples were collected within a median of 1 day after hospital admission.At the time of enrollment, no patient had completed the vaccination scheme.According to the 2022 guidelines for COVID-19 care, patients were dichotomized as severe, which either had oxygen saturation <90% in room air or showed signs of pneumonia or severe respiratory syndrome, and critical, which required life-sustaining treatment or displayed acute respiratory distress syndrome, sepsis or death. 13ble 1 shows the comorbidity scores and the demographic data of the enrolled patients.The patients were mainly male, older than 70 years, with a median Charlson Comorbidity Index (CCI) of 4. Critical patients were significantly older and had a significantly higher CCI than those with severe COVID-19.Consistent with the inclusion criteria, all patients who died were critical.

| Haematological and blood chemistry data associated with patient outcome
Table 2 shows the haematological and blood chemistry variables of all the enrolled patients stratified according to their outcomes.The levels of two sepsis-associated markers (pro-adrenomedullin and procalcitonin) were significantly higher in critical patients.Similarly, the cardiac injury marker troponin-T and heart failure marker N-terminal pro-B-type natriuretic peptide were significantly higher in critical patients.Concerning haematological data, red cell width distribution and lymphocyte counts were the only two parameters that significantly differed between severe and critical patients.
Finally, the levels of the inflammatory cytokine interleukin 6 were significantly increased in critically ill patients.

| Leukocyte and extracellular vesicles subpopulation data associated with patient outcome
Since COVID-19-critical patients are characterized by a reduced lymphocyte count and immunosuppression, we decided to include in our analysis the results of the characterization of peripheral blood mononuclear cells (PBMC), as determined by flow cytometry.As shown in Table 3, the fraction of monocytes expressing HLA-DR was significantly lower in critical patients than in severe patients.Aside from the fraction of CD4 + CD3 + T helper lymphocytes, which had marginal significance, the other populations did not significantly differ between the two patient subsets.As shown in Table 4, the fraction of EV with small dimensions is significantly, although marginally, higher in severe patients than in critical patients.Similarly, the fraction of small vesicles expressing CD31, CD42b or CD140b was significantly lower in the plasma of critical patients.Conversely, the fraction of EV of ≈500-nm diameter that expressed CD140b or CD56 and the fraction of EV of >500-nm diameter that expressed either CD34 or CD45 showed an opposite trend, being higher in critical patients.
Importantly, CD63 expressing EV could not be detected using our setup, suggesting that exosomes fell under the limit of detection of our system (Figure S1).

| Identification of independent predictors of patient outcome
Next, we attempted to identify, employing an elastic net logistic regression analysis with cross-validation, the variables needed to build the best model to classify critical patients.The area under the curve (AUC) of the receiver operator curve (ROC) was used to evaluate model performance.Three models were tested.To build the first model, we employed all demographic, comorbidity, haematological and blood chemistry data.To build the second model, we employed all the EV data.Table 6 lists the variables and coefficients used to construct the model.ROC AUC was 0.663 (95% CI 0.507-0.819).
Next, we built a model that included all the PBMC immunophenotype data.The model did not discriminate the outcome, being the ROC AUC 0.520 (95% CI 0.350-0.690).
Next, we built a complete model, including demographic, comorbidity, haematological, blood chemistry, EV and PBMNC data with a corrected p value < 0.05 in the training set (Tables S1-S4).Finally, we analysed the correlations among the variables of the complete model (Figure 3).

Although most of the vesicle-related parameters correlated
with each other, interleukin 6 showed a negative correlation with the fraction of ≈500-nm-diameter EVs expressing N-cadherin, while age was directly correlated with the fraction of large EV expressing CD45 or CD56 and inversely correlated with the fraction of small EV expressing CD31, CD42b or CD140b.In line with this, the Charlson Comorbidity Index was directly correlated with the fraction of large EV expressing CD45 and inversely correlated with the fraction of small EV expressing CD42b.No patient in our series was fully vaccinated since the vaccination campaign began in Italy at the end of 2020, and by March 2021, only a small minority of the population had completed the entire scheme. 16We decided to analyse only  patients with severe or critical COVID-19 to understand the critical variables that could help clinicians identify, among hospitalized patients, those that need to be more carefully monitored.

| DISCUSS ION
Here, we specifically focused our attention on circulating EV, since an accumulating body of literature has shown their crucial role in both homeostatic and pathological processes, including infectious diseases, immune system regulation, coagulation and tissue repair. 9The term EV is broad and includes heterogeneous populations of cell-derived membrane-enveloped particles that differ in size, content composition, biogenesis and biological function. 9ree main subtypes of EV can be identified based on their size and biological functions: exosomes, microvesicles and apoptotic bodies.Exosomes are formed intracellularly within multivesicular bodies, have a small size (<200-nm diameter) and can be released in the extracellular space, where they influence the behaviour of target cells either via ligand-receptor interactions, or following phagocytosis, pinocytosis or by fusing the plasma membrane with the target cells, thus releasing their content into recipient cells. 10crovesicles have larger dimensions (100-1000 nm in diameter) and are formed by budding of the plasma membrane.Therefore, their composition reflects, possibly more closely than that of exosomes, the features of their cell of origin. 10Finally, EV comprise apoptotic bodies, which have even larger dimensions (1000-5000nm diameter) and are formed during this form of programmed cell death. 10To add a layer of complexity, more recently it has been shown that, similarly to healthy cells, apoptotic cells may release microvesicles and exosomes too. 17Given this heterogeneity, we set up a flow-cytometry-based assay that could simultaneously evaluate the physical properties of EV and the expression of antigens reflecting the cell of origin.We decided to include markers of cell types involved in the pathogenesis of COVID-19, such as platelets (CD31 and CD42b), endothelial cells (CD34 and CD31), leukocytes (CD45), pericytes (CD140b), NK cells (CD56), neural cells (CD56 and N-cadherin) and cardiomyocytes (N-cadherin).We did not include other excellent emerging markers that do not exhibit cell-specific expression according to the Human Protein Atlas (https:// www.prote inatl as.org).
EVs may play a complex role in viral infections because they can carry molecules that act as pathogen-associated molecular patterns (PAMPs), triggering antiviral responses and facilitating viral propagation. 18In line with this, Yim and colleagues analysed 20 patients affected by a mild form of the disease and 17 healthy donors and reported an increase in CD31 + EV carrying subunit 1 of the SARS-CoV2 Spike protein.Following a thorough characterization of the phenotype of small EV (<200-nm diameter), the authors concluded that their exploratory analysis suggested the potential usefulness of serum EV in predicting disease status. 19similar conclusion was drawn by  tion has a protective effect. 22Intriguingly, CD140b is expressed by MSC.However, we could not observe clear positivity for the exosomal marker CD63 using our flow cytometry assay.
Concerning other factors that were inversely associated with patient outcome, albuminemia and the frequency of activated monocytes expressing HLA-DR emerged from our analysis.The association between albumin levels and COVID-19 mortality has also been described in an independent cohort and associated with oxidative stress, neutrophil activation and thrombosis. 23The reduced activation of monocytes in COVID-19 patients too is a finding that has already been described in the literature and is consistent with the immune paralysis occurring in critical patients. 7,24though small EVs may have a potential protective effect on the natural history of COVID-19, a pathogenic role for EVs has also been evoked.Indeed, in a study comprising 67 patients with respiratory symptoms (34 of which were SARS-CoV2 positives and 33 negative) and 16 healthy controls, serum-derived EVs expressing tissue factor were the most efficient in classifying COVID-19 patients. 25Importantly, the excellent prognostic role of these particles,

F
I G U R E 1 Gating strategy employed to analyse EV.A gating strategy was adopted to analyse the plasma EVs of the enrolled patients.(A) CFSE staining was used to identify EV cytoplasm.1-μm-size particles were added to the CFSE-stained EV to quantify their concentration.(B) The EV size was determined based on SSC-A.Fluorescent particles of known size were used to calibrate the instrument.(C) CFSE-labelled EV were stained with antibodies specific to the antigens of interest to determine the immunophenotype of each EV subset.A typical example of a CD31-and CFSE-labelled plasma sample is shown.Data are shown as median and interquartile range (IQR) or percentages.The p value refers to the comparison between the severe and critical patients.The significant results are shown in bold.TA B L E 1 Demographic and clinical features of enrolled patients.
Data are shown as median and interquartile range (IQR) or percentages.The p value is the result of the comparison between the severe and critical patients, and the last column shows the p values corrected for multiple comparisons.The significant results are shown in bold.To identify novel markers associated with COVID-19 outcomes, we characterized the surface phenotype of circulating EVs of microand nanometre dimensions.Platelet-poor plasma (estimated level of platelet contamination: 3.88 ± 1.82•10 3 particles/μL, n = 10) was analysed by flow cytometry to quantify EVs, assess their size (dividing them into small ≤200 nm, intermediate ≈200-400 nm, central ≈500 nm and large >500 nm subgroups) and analyse the expression of markers of endothelial, platelet, leukocyte, natural killer (NK) cells, mural cell and neural cell origin (i.e.CD31, CD34, CD42b, CD45, CD140b, CD56 and N-cadherin).EVs were analysed according to the gating strategy shown in Figure1.
The results are shown in Figure 2. ROC AUC was 0.769 (95% CI 0.6030-0.908).Even when we added to the complete model, other potential confounding factors (i.e.lymphocyte relative counts, the fraction of T CD4 + helper lymphocytes and activated CD14 + HLA-DR + monocytes), the fraction of small vesicles expressing CD31 and the fraction of central EV expressing N-cadherin remained among the strongest predictors of patient outcome (Figure S2), while the ROC AUC increased marginally; 0.785 (95% CI 0.650-0.919).
The variables included in the model are listed, along with their coefficients.composition.Specifically, mild patients have a higher abundance of exosomes positive for the SARS-CoV2 spike protein and trigger an adaptive immune response, while severe COVID-19-derived exosomes are enriched in proteins associated with acute inflammatory responses. 21Furthermore, clinical trials that have experimented the exogenous administration of mesenchymal stem cell (MSC)-derived exosomes to COVID-19 patients have suggested that this interven-

F I G U R E 2
Results of the elastic net logistic regression analysis of the complete model.Results of the elastic net logistic regression model comprising all the significant variables.Each variable included in the model is a row on the y-axis (labels are shown on the right side of the panel), while the coefficients are plotted as bars on the x-axis.F I G U R E 3 Correlation among the variables of the complete model.Heatplot summarizing the results of the correlation analysis between variables emerging from elastic net logistic regression.Correlation coefficients are shown as a heatmap, according to the legend on the right side of the panel.*p < 0.05, **p < 0.01.
TA B L E 2

Table 5
lists the variables and coefficients used to build the model.ROC AUC was 0.750 (95% CI 0.607-0.892).
Peripheral blood mononucleated cell immunophenotypes of the enrolled patients.Data are shown as median and interquartile range (IQR) or percentages.The p value is the result of the comparison between the severe and critical patients, and the last column shows the p values corrected for multiple comparisons.The significant results are shown in bold.
In this study, we present data concerning the extensive characterization, from a clinical and laboratory point of view, of 146 hospitalized patients affected by SARS-CoV-2 infection recruited at TA B L E 3 our institution between November 2020 and April 2021, during a phase of the pandemic in which B.1.177and Alpha (B.1.1.7)were the most prevalent genotypes in Italy.
Flow cytometry analysis of plasma extracellular vesicles (EVs) from the enrolled patients.
TA B L E 4

(20.4-31.1) 30.4 (25.5-34.7) <0.001 0.006
Data are shown as median and interquartile range (IQR) or percentages.The p value is the result of the comparison between the severe and critical patients, and the last column shows the p values corrected for multiple comparisons.The significant results are shown in bold.
Results of the elastic net logistic regression analysis of demographic, comorbidity, haematological and blood chemistry data of the enrolled patients.Results of the elastic net logistic regression analysis of the extracellular vesicles (EV) analysis data of the enrolled patients.
clinical, blood chemistry and haematological data, we additionally demonstrate that EV have a relevant prognostic potential.Indeed, in our study, we observed that the fraction of small EV expressing either CD 31, CD 140b or CD 42b inversely correlated with patient outcomes.This result is consistent with data showing that exosomes derived from patients with mild vs. severe COVID-19 differ in cargo TA B L E 5 TA B L E 6