Tiny blood clots linked to immune traps are higher in people with long COVID

We have separately revealed both the presence of microclots [1, 4, 7], and elevated cirDNA and NETs in LC, and we here assess the extent to which they are linked/associated. We used the same cohort of healthy individuals (HI) and individuals with LC (HI, Table S1 and LC, Table S2) show the demographics and co‐morbidities of LC and both HI cohorts). We first determined the number of microclots and their size, and simultaneously determined the plasma concentrations of NETs protein markers (MPO, NE) and total cirDNA, and performed correlation tests. We then assessed whether NETs constituents were present within the microclots, seeking to determine a possible structural association between the two entities.

We included two control cohorts and one Long COVID cohort. The control cohorts were as follows: one from South Africa [SA controls, (n = 14)] and another from France [EFS controls, (n = 24)], both comprising healthy individuals without LC symptoms at blood collection. The inclusion of these geographically and culturally distinct control groups aimed to explore any differences or similarities in inflammatory profiles. The South African control cohort included 14 healthy individuals [seven males, seven females; mean age 28 (26–42)] with no cardiovascular history, coagulation disorders, or pregnancy. Two participants had hypercholesterolemia, and one had hypertension, all effectively managed through diet and treatment. The French control cohort consisted of 24 healthy blood donors [15 males, nine females; median age 45.5 (19–64)] with similar exclusion criteria. Demographic details and co‐morbidities were documented and compared across cohorts.

For the LC group, 50 patients [18 males, 32 females; mean age 50 ( ± 17)] were recruited through clinical collaborators who identified post‐acute COVID‐19 individuals with persistent LC symptoms. LC patients completed a questionnaire detailing their demographics, comorbidities, medication use, COVID‐19 vaccination history, severity of initial infection, and self‐reported LC symptoms. All participants provided informed consent, and ethical clearance was secured before blood collection. Platelet‐poor plasma samples were cold‐shipped between South Africa and France for analysis. Individuals constituting both HI cohorts (EFS and SA HI), included in this study were selected to not be on any anticoagulant therapy, baby aspirin or on any other medication that might impact their clotting physiology.

The investigation received ethical approval from the Health Research Ethics Committee (HREC) at Stellenbosch University in South Africa, under the references: N19/03/043 and project ID: 9521 with annual re‐approval. The French ethics approval number is EFS‐PM N° 21PLER2018‐0069. We have a data and material sharing agreement between Stellenbosch University and INSERM (S008426). Participants were recruited based on the information they provided to their clinicians in collaboration with the study. Before sample collection, participants were extensively briefed on the study's objectives, potential risks, and details of the entire process to gain their informed consent. Throughout all study procedures there was mandatory adherence to ethical guidelines and principles, as detailed by the Declaration of Helsinki, South African Guidelines for Good Clinical Practice, and Medical Research Council Ethical Guidelines for Research.

Blood samples from South Africa was collected in citrate tubes, either used fresh to study blood cells or platelet‐poor plasma was created and stored at −80°C. A portion of the frozen PPP was cold shipped to the INSERM laboratory in France. Blood samples form the French cohort were obtained from the Etablissement Français du Sang (EFS), which is Montpellier's blood transfusion center (Convention EFS‐PM N° 21PLER2018‐0069). Blood sample from healthy individuals was collected in 5‐mL Streck tubes. No difference in terms of NETs markers and cirDNA was found when comparing EDTA vs Streck tubes. As previously determined, no bias from EDTA was observed in the quantified markers in our study [17, 22, 27, 29].

Platelets from the South African controls and LC samples were studied in the cellular fraction that remained after whole blood was centrifuged at 3000×g for 15 min. CD62PE (P‐selectin on platelet surface) (IM1759U, Beckman Coulter, Brea, CA, USA) and to PAC‐1 (activated glycoprotein (GP) IIb/IIIa) (340507, BD Biosciences, San Jose, CA, USA) was added to the hematocrit. P‐selectin is released from the cellular granules during platelet activation and then moves to the surface of the platelet membrane. The antibody PAC‐1 detects the neoepitope of active GPIIb/IIIa. PAC‐1 antibody binding is correlated with platelet activation. After the samples was incubated at room temperature for 30 min, samples were also viewed using the Zeiss Axio Observer 7 fluorescence microscope with a Plan‐Apochromat 63×/1.4 Oil DIC M27 objective (Carl Zeiss Microscopy, Munich, Germany). We used excitation wavelength 450 to 488 nm and the emission at 499 to 529 nm (PAC‐1), and excitation 540 to 570 nm and the emission 577 to 607 nm (CD62PE).

For imaging flow cytometry (using various fractions of the sample to anayse the blood cells), we used the Amnis® FlowSight® Imaging Flow Cytometer (Luminex, Seattle, WA, USA). This method integrates fluorescence microscopy for single‐cell morphology analysis with the capabilities of imaging flow cytometry, enabling precise discrimination of heterogenous cell populations and high flow rate capacity while maintaining data quality.

Microscopy and imaging flow cytometry analyses were conducted on plasma prepared by 3000×g centrifugation for 15 min in LC patients, and by a two‐step centrifugation process (1200 g for 10 min followed by 16000 g for 10 min) in the in both HI SA and EFS control cohorts. As shown in the Supporting Information, both processes yielded similar values and observations for the tested markers, including microscopy, microclots counts, MPO, NE, and cirDNA.

Determination of cirDNA concentration and NETs protein markers (NE and MPO) were performed from exactly the same plasma sample as prepared following a two steps centrifugation process (1200 g for 10 min then 16000 g for 10 min). This preparation was previously chosen to exclude large vesicles that may encapsulate DNA and consequently, contaminate the fraction that should only contains circulating cell‐free DNA (cirDNA). We previously demonstrated the absence of cells and cell debris in published guidelines as well as to set clinical studies towards evaluating cirDNA diagnostic performance in cancer and COVID/Long COVID patients [17, 22, 29, 32]. Additionally, in several reports, we showed that the combination of the three analytes provides an accurate measure of NETosis [27, 29].

Plasma marker data were analyzed using a supervised machine learning framework implemented in R. All features were normalized using Z‐score scaling (mean‐centering and standard deviation scaling) with the preProcess function from the caret package. A Random Forest classifier was trained using the train() function (method = "rf"), applying five‐fold cross‐validation repeated 100 times (repeatedcv) to ensure robust model evaluation. Class probabilities were enabled, and model performance was assessed using the area under the receiver operating characteristic curve (AUC) via the twoClassSummary metric.

Model predictions were used to compute AUC, sensitivity, and specificity at the optimal ROC point (determined as the maximum distance from the diagonal), and confidence intervals were estimated by bootstrapping with 1,000 iterations using the boot and pROC packages. Accuracy and its 95% confidence interval were also calculated using bootstrap resampling. A confusion matrix was generated and visualized as a heatmap showing classification performance as percentages. Reported confusion matrices represent the average of matrices obtained on the test sets of 100 repeated fivefold CV. In addition to the Random Forest model, k‐nearest neighbors (kNN) and Decision Tree models were also evaluated using the same cross‐validation setup, normalization procedure, performance metrics, and bootstrap‐based statistical analysis. This allowed for consistent comparison across classifier types. Feature importance scores from the Random Forest model were extracted using varImp() and visualized in a ranked bar plot of the top 20 contributing protein markers.

All analyses were performed in R (version 4.3.3), with random seed (123) initialization for reproducibility and the software packages: randomForest (v 4.7‐1.2) for RF models, caret (v 7.0‐1) for cross‐validation and confusion matrix extraction, pROC (v 1.18.5) for AUC calculation, ggplot2 (3.5.1) for plotting, boot (1.3‐31) for bootstrapping.

To enhance the clarity and detail of the fluorescence microscopy image, we employed a two‐step process utilizing Photoshop and Topaz Gigapixel AI.

The choice of statistical tests was made based on the nature of the data and the specific objectives of the analysis. For comparing groups with non‐parametric data, the Mann‐Whitney U test was selected. This test is appropriate when the data deviate from a Gaussian distribution, as confirmed by the Shapiro‐Wilk test.

For exploring relationships between variables, correlation analysis was conducted using the Spearman test. The Spearman test evaluates the strength and direction of association between variables without assuming linearity. It is particularly appropriate when the relationship between variables may be monotonic but not necessarily linear, or when outliers are present in the data. This test calculates the Spearman correlation coefficient, which ranges from −1 to 1, with values closer to 1 indicating a strong positive correlation, values closer to −1 indicating a strong negative correlation, and values around 0 indicating no correlation. Significant difference between two groups using Spearman correlation matrices are called when the p value shows statistic difference; the R value being indicative.

All statistical analyses were performed using Graph Pad Prism 8.3.1 software. A significance level of 0.05 was chosen to determine statistical significance, meaning that results with a p‐value less than 0.05 were considered statistically significant. To provide a clear indication of the level of significance, p‐values were reported with asterisk notation: *p < 0.05, **p < 0.01, ***p < 0.001, and ****p < 0.0001.

The characteristic NETs filaments released by activated neutrophil can be observed using fluorescence microscopy and the Hoechst DNA marker, can be seen as a diffuse light blue fluorescence, showing the chromatin outside cells (see arrows) (Figure 3A–D). We utilized staining techniques to visualize the co‐incidence of microclot formation and NETs in acute COVID‐19 and LC, compared to controls. The micrographs in Figure 3E and F highlight the use of Thioflavin T (ThT) for microclots, alongside Hoechst or SYTO stains for DNA, indicating NETs presence (Hoechst in blue, SYTO in red, and ThT in green, with a combined overlay in the third column). The bottom row (Figure 3G) shows plasma samples treated with ThT, SYTO and myeloperoxidase (MPO) stains (SYTO in red, ThT in green and MPO in blue). MPO, ThT and DNA clearly colocalized within microclots. and may also be observed alone in a few localizations in the microclots image. Arrows highlight DNA localizations, illustrating NETs surrounding or outside microclots, or between two microclots (Figure 3 bottom right image; enlargement in Figure 4B). Note, previous studies showed MPO and DNA only colocalized when neutrophils are activated [33], or within a coronary thrombus [30]. In addition, in vitro colocalization of neutrophils, extracellular DNA and coagulation factors such as fibrinogen was reported during NETosis [34].

In their structural, biological, physiological, and pathophysiological characteristics, microclots differ from traditional blood clots (or fibrin mesh formation) in several fundamental ways. Structurally, microclots tend to be smaller and are found as free entities in circulation. They also exhibit significant structural heterogeneity. Healthy fiber diameters are more uniform and clot into a gel mesh, typically only induced during a physiological clotting event such as following a cut, and under the action of thrombin and the activation of the clotting cascade. Biologically, clots formed in such 'physiological' circumstances are composed largely of platelets and fibrin designed to stop bleeding. By contrast microclots are characterized by an amyloid‐like nature, which can be revealed by staining them with suitable fluorogenic dyes. In addition, we have shown that microclots contain not only fibrinogen molecules, but also entrap various inflammatory molecules and even antibodies [10, 12]. An important factor initiating platelet activation and the formation of intravascular microclots is endothelial damage triggering widespread endothelial damage (endothelialitis). The presence of intravascular microclots in small blood vessels perpetuates endothelial cell damage (endothelialitis), which consequently adopt the pro‐adhesive phenotype, thus exposing those microclots to microorganisms that favour the elimination and limit the dissemination of the pathogen [35, 36].

Our fluorescence microscopy analysis (as shown in the suppl. doc) demonstrated that the addition of spike protein or LPS to purified fibrinogen induced the formation of insoluble microclots (Figure ). Fibrin(ogen) in microclots were confirmed by binding of the Fibrinogen α‐chain antibody (Alexa Fluor 594), with amyloid regions identified by Thioflavin T (ThT) staining. This pattern was consistent across both healthy and Long COVID samples, validating the presence of fibrin(ogen) and amyloidogenic characteristics within the microclots (see Suppl Figure ). S6 S6

Abnormal blood clotting and microclots formation thus appear as important factors that may explain the development and persistence of symptoms in LC. Indeed, several researchers have signalled the exacerbation by a dysregulated thromboinflammatory response of the interplay between coagulation and innate immunity [29, 30, 37, 38]. Platelet hyperactivation and endothelial dysfunction would therefore appear to be a unifying pathway underlying the wide range of symptoms seen in LC [2, 11].

Mechanistically, cell viral infection and the resulting inflammation leads to the stimulation of neutrophils, which leads in turn to the release of tissue factors and the degradation of anticoagulants by monocytes, promoting coagulation [25]. As a result, the interplay between the innate immune response and blood coagulation (immunothrombosis) leads to the formation of microthromboses [29, 37, 39].

Existing evidence points to a central role played by NETs in immunothrombosis [27, 29, 35, 39, 40]. Thus, activated platelets are an endogenous stimulus capable of inducing NETs [22, 25, 41]. NETs interact physically with platelets, both through their binding with the P‐selectin glycoprotein ligand and through the binding of the high mobility group box 1 (HMGB1) nuclear protein to the RAGE or TLR4 receptors [41]. There is also other evidence of the mechanistic link between NETs and coagulation, such as the binding of: (i) the coagulation factor fibrinogen with DNA‐rich NETs; (ii), the coagulation factors prothrombin, FX and FVIIa with activated neutrophils during NETosis; and (iii), the anticoagulant activated [42] protein C (APC) with activated neutrophils and DNA‐rich NETs [29, 43].

Neutrophils and platelets activate each other mutually, and their physical interaction constitutes a central loop of amplification of immuno‐thrombosis triggering NETs formation, (39). We speculate that the deterioration or activation of the endothelial cells by circulating microclots leads to inflammation which in turn activate neutrophils [42, 44]. Consequently, exacerbated and deleterious NETs production might be not only maintained by the NETs/thrombosis feedback loop but also by the positive NETs/inflammation feedback loop [44]. This may cause a persistent stimulation of neutrophils, which may in turn lead to the pathogenic microthrombosis phenomena observed in LC. Recent proteomic analysis showed neutrophil activation as well as neutrophil signatures in LC patients associated with clotting cascade members. Our observations point to the possibility that, in addition to this hypercoagulation state, a persistent imbalance in NETs formation, plausibly caused by viral persistence [45], may lead to a high structural association of NETs with microclots. We speculate this would impair clot lysis, contributing to microclots higher numbers in the circulation and a lower clearance level. The protective role of NETs in infection, therefore, would come at a high cost in the case of SARS‐CoV2 infection, as well as may be in certain other inflammatory diseases. However, association does not mean causation and further studies are needed to confirm our postulate.

The ability to detect these markers efficiently would pave the way for better disease monitoring, understanding the etiology of inflammation, and personalized treatment strategies. More specifically, a detection strategy that could demonstrate the combined presence of high microclot numbers and elevated levels of both NETs markers and cirDNA (either by microscopy, cytometry or ELISA) would offer an efficient means of tackling the underdiagnosis of LC, thereby improving diagnostic accuracy and appropriate intervention.

The co‐localization of MPO and DNA observed in Figure 4B aligns with their association in LC patient plasma, as shown by Spearman analysis (Figure 3D), and supports previous findings of MPO/NE/cirDNA association in COVID‐19, Long COVID, and cancer samples [17, 18, 22, 29] as well as in PMA‐stimulated neutrophil cultures where NETs formation produces NE, MPO, and cell‐free DNA [27]. These findings collectively demonstrate that concurrent elevated plasma levels of MPO, NE, and cirDNA indicate NETs formation.

To further evaluate the performance of an LC detection test which uses markers of both entities, we employed Decision Tree‐Gini, K‐nearest Neighbors Classifier (Figure S8), and Random Forests (RF) for training by machine learning. In these models, we utilized all markers for training, and five‐fold cross‐validation with 100 repetitions for each classifier, to ensure robustness, stability, and the generalizability of machine learning model performance estimates. The best performance was achieved with the RF algorithm (Figure 4C). This superior performance likely reflects the ensemble nature of RF, which reduces overfitting and improves robustness to noise and irrelevant features, making it well‐suited to moderate‐sized datasets. The mean feature importance in RF model indicates that microclots 100–400 µm², microclots 0–100 µm², MPO, microclots 400–900 µm², microclot objects, microclot counts, and cirDNA all contribute to the discrimination between healthy and LC subjects (Figure 4D). RF models provide the highest performance values, with accuracy of 0.91 (95% CI: 0.90; 0.91), AUC of 0.95 (95% CI: 0.98; 0.98), sensitivity of 0.86 (95% CI: 0.86; 0.87), and specificity of 0.99 (95% CI: 0.99; 0.99). The confusion matrices show that 93% of subjects with LC and 88% of healthy individuals were correctly classified by the RF model (Figure 4C). By excluding the least important markers one by one, using a RF model we found that the most important markers for achieving the highest specificity are MPO and microclots 100–400 µm², and that these provided accuracy, AUC, sensitivity and specificity of 0.92 (95% CI: 0.91; 0.92), 0.98 (95% CI: 0.98; 0.98), 0.88 (95% CI: 0.88; 0.88) and 0.98 (95% CI: 0.98; 0.98), respectively (Figure S8). These data showed that the combination of a NETs marker (MPO) and a microclot numbering marker (100‐400 µm² size) can achieve a very high discriminating/screening power (Figure S9), as illustrated by our scatter plot presentation of MPO and 100–400 µm² microclots concentration values (Figure 2E). While promising, our preliminary data do not yet validate NETs and microclots as biomarkers for Long COVID in practical clinical settings. Additional studies with larger patient cohorts and comparisons among individuals with similar symptoms are essential for further validation.

Plasma from healthy individuals contains low numbers of microclots of low sizes as observed here (Figure 1A,B, Figure 3; and Figure S2 and S3) and previously (7), but this study revealed they may be covered at least partially with NETs. We suggest that microclots are bound/entrapped by floating NET parts in the blood stream, and that physico‐chemical properties of both entities enable close structural association inhibiting their degradation in the circulation. Thus, NET parts may now be conceived as a potential microclot constituent in addition to the insoluble fibrin amyloid and other molecules like cytokines or antibodies [10, 12].

Although our study included two independent healthy control cohorts from different geographic regions, it did not include disease‐specific control groups such as individuals with diabetes, sepsis, or autoimmune disorders. As such, we cannot definitively assess the specificity of NETs and microclots elevations to Long COVID. Although co‐localization of ThT and MPO signals within microclot structures was observed by fluorescence microscopy, unfortunately an microscopy technique does not permit reliable quantification of the proportion of microclot that are ThT+ or MPO + ; such analyses would require high‐throughput single‐particle approaches like imaging flow cytometry. Such optimization and analysis will be a future research endeavour. While our findings demonstrate a strong quantitative and structural association between NETs and microclots in Long COVID patients, further comparative studies are needed to establish the discriminatory power of these biomarkers across different disease states. We explicitly recognize this as a limitation of the current study and recommend that future investigations include disease‐matched control cohorts to validate the diagnostic specificity and potential clinical utility of these markers. While our current analysis considers LC patients as a single group, this approach should be regarded only as a starting point. It is necessary to first identify measurable biological signals, such as NET‐associated microclots, which may ultimately enable patient stratification in the future. The rationale of this study is to explore whether a microclot and NETs burden can serve as a distinguishing feature which helps to define biologically distinct subtypes, potentially including a group characterized by heightened thrombo‐inflammatory or immune‐mediated endothelial injury. We therefore recognize the absence of clinical stratification as a limitation, while emphasizing that this study lays the groundwork for developing objective biomarker‐based subgrouping in future research.

Microclots frequently appear in chronic inflammatory states and are more resistant to fibrinolysis, showing resistance even to powerful proteolytic enzymes such as trypsin [10]. We hypothesize that imbalanced NETs formation as observed in various inflammatory diseases such as COVID‐19 or LC might lead to higher microclots stability and therefore their elevated number in the circulation. Such resistance to normal breakdown mechanisms and the variability in their formation, even in the absence of typical clotting triggers such as thrombin, would indicate that microclots may have more widespread and clinically deleterious effects.

This study shows a robust association between biomarkers indicative of thromboinflammatory activity and LC. Since our study revealed that NETs may be a component of circulating microclots, we speculate that higher NETs formation promotes the stabilization of microclots in the circulation, leading to deleterious effects which (in part) may contribute to the symptoms of LC. The discovery of these biomarker linkages not only presents a possible novel diagnostic methodology but also novel therapeutic targets, offering prospects for future markedly improved clinical management. Consequently, our findings present a significant advancement in the understanding of the interactions between NETs and microclots in Long COVID.