Plasma RIPK3 And HMGB1 Predict Severe COVID-19 Progression In ICU Patients: A Single-Center Cohort Study

Background: Severe progression of coronavirus disease 2019 (COVID ‐ 19) causes respiratory failure and critical illness. Recently, these pathologies have been associated with necroptosis, a receptor ‐ interacting serine/threonine ‐ protein kinase 3 (RIPK3) dependent regulated form of inammatory cell death. Investigations of indicator necroptosis proteins like RIPK3, mixed lineage kinase domain ‐ like pseudokinase (MLKL), receptor ‐ interacting serine/threonine ‐ protein kinases 1 (RIPK1), and high ‐ mobility group box 1 (HMGB1) in clinical COVID ‐ 19 manifestations are lacking. Methods: A prospective prolonged cohort study including 46 intensive care unit (ICU) patients classied with moderate and severe COVID ‐ 19 was conducted with daily measured plasma levels of indicator necroptosis proteins like RIPK3, MLKL, RIPK1, and HMGB1 by enzyme ‐ linked immunosorbent assay (ELISA). On this basis, a multiple logistic (regression) classication for the prediction of severe COVID ‐ 19 progression was performed. Results: We found signicantly elevated RIPK3, MLKL, HMGB1, and RIPK1 levels in COVID ‐ 19 patients admitted to the ICU compared to healthy controls throughout the ongoing disease, indicating necroptotic processes. Above all, with combined measurements of RIPK3 and HMGB1 plasma levels, we were able to time ‐ independently predict COVID ‐ 19 severity with 84% accuracy, 90% sensitivity, and 76% specicity. Conclusion: We suggest that HMGB1 and RIPK3 are potential biomarkers to identify high ‐ risk COVID ‐ 19 patients and developed a classier for COVID ‐ 19 severity. Data are presented as a n (%) for categorical variables or b median (interquartile range) for continuous variables. Patients’ laboratory parameters are reported as the respective median of the parameter levels obtained during ICU stay. p ‐ values comparing patients with moderate and severe COVID ‐ 19 were calculated with Mann ‐ Whitney U test or Fisher’s exact test. Additionally, patients’ median laboratory parameter levels were compared to the hospital’s central laboratory’s threshold levels (CRP: 0.5 mg/dl, IL ‐ 6: 7 pg/ml, PCT: 0.5 ng/ml, LDH: 248 U/l, peripheral leukocyte count: 10.41 /nl). Respective quantities in the pathological range were determined and then compared among patients with severe and moderate COVID ‐ 19 by Fisher’s exact test.


Introduction
Coronavirus disease 2019  is the most challenging pandemic in recent human history. In November 2021, the World Health Organization reported 249.743.428 cases of COVID-19 with 5.047.652 deaths globally [1]. It is crucial for the disease outcome and appropriate treatment to develop a method to determine the exact point of COVID-19 exacerbation. Patients suffering from critical COVID-19 often present with respiratory failure as well as features of sepsis, such as coagulopathy, lymphopenia, and high plasma levels of pro-in ammatory cytokines [2].
In various comparable non-COVID-19-related in ammatory diseases, it is already established that the receptor-interacting serine/threonine-protein kinase 1 and 3 (RIPK1 and RIPK3), as well as the mixed lineage kinase domain-like pseudokinase (MLKL), are associated with disease progression as important regulators of necroptotic cell death [3,4]. For example, the examination of lung tissue sections in H7N9 virus infection, in which acute respiratory distress syndrome (ARDS) was the main cause of death, showed signi cantly higher RIPK1, RIPK3, phospho-RIPK3, MLKL, and phospho-MLKL protein levels [5]. These data suggest that severe H7N9 infection is associated with necroptosis of the lung epithelium which contributes to ARDS. This hypothesis is supported by results that showed signi cantly increased RIPK3 levels not only in the plasma of ARDS patients but also in bronchoalveolar lavage uid [6]. Furthermore, elevated RIPK3 levels in the plasma of patients with severe sepsis or septic shock also indicate that the RIPK3 signaling pathway is activated under septic conditions [7]. Necroptosis, also referred to as RIPK3-dependent necrosis, is executed by phosphorylated and activated RIPK1 and RIPK3, which form a complex known as the necrosome [8][9][10]. Subsequently, the effector molecule MLKL is phosphorylated, enabling it to oligomerize and migrate to the cell membrane, leading to the release of damage-associated molecular patterns (DAMPs), cell rupture, and lytic cell death [11]. This promotes cytokine production and an excessive immune response [12]. High-mobility group box 1 (HMGB1), considered as one of the most relevant DAMPs released by necroptotic cells, usually binds to DNA as well as chromatin and exerts its function in chromatin modi cation and DNA repair [13][14][15][16][17][18]. When released during in ammatory cell death, HMGB1 triggers immunological processes, inducing recruitment of immune cells and expression as well as the release of pro-in ammatory cytokines (interleukin 6 (IL-6); IL-1β; tumor necrosis factor-α (TNF-α)), as similarly described in COVID-19 [13,[19][20][21]. Extracellular HMGB1 is furthermore capable of forming complexes with cytokines amplifying hyperin ammation [22,23] . Moreover, high serum HMGB1 levels in non-COVID-19 patients were linked to fatal ARDS [24] . Besides, reactive oxygen species (ROS) production is associated with necroptosis, and mitochondrial ROS (mtROS) production also plays a crucial role in peripheral lymphocytes in severe disease conditions [25][26][27].
Against this background, we decided to conduct a close monitoring of plasma levels of the necroptosis-related proteins RIPK3, MLKL, RIPK1, and the DAMP HMGB1 in COVID-19 patients throughout intensive care unit (ICU) stay. The current single-center cohort study aims to investigate the prognostic potential of RIPK3, MLKL, HMGB1, and RIPK1 in COVID-19 progression as feasible biomarkers. Using long-term measurement data, we were able to build a classi er that predicts COVID-19 exacerbation independently of time. In addition, we analyzed cell death and mtROS in peripheral leukocytes of ICU COVID-19 patients in single measurements, to verify if these parameters differ in COVID-19 patients as shown before in severe disease conditions, e.g. sepsis patients [27] .

COVID-19 cohort
This is a prospective single-center cohort study of 46 COVID-19 patients (≥18-years) who were admitted to the ICU of the University Hospital Frankfurt am Main, Germany, between June 2020 and January 2021. During ICU stay, blood samples were obtained daily at 8 a.m. from admission until ICU discharge. In ammatory parameters including C-reactive protein (CRP), IL-6, procalcitonin (PCT), lactate dehydrogenase (LDH), and peripheral leukocyte count were obtained daily at 4 a.m. and measured by the hospital's central laboratory and compared to the hospital's central laboratory's threshold levels (CRP: 0.5 mg/dl, IL-6: 7 pg/ml, PCT: 0.5 ng/ml, LDH: 248 U/l, peripheral leukocyte count: 10.41 /nl). Control samples were drawn from 15 non-severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2)-infected healthy donors (≥18-years) to compare healthy physiological conditions to COVID-19. The study was conducted in compliance with good clinical practice and current guidelines. Intubation was considered in patients with COVID-19 and severe hypoxemia (PaO 2 /FiO 2 <150 mmHg) and respiratory rates >30/min. A PaO 2 /FiO 2 of <100 mmHg in two consecutive measurements was an indication to perform mechanical ventilation, according to the German guideline [28]. Based on this, it was feasible to distinguish between patients with severe and moderate COVID-19 according to their requirement for intubation throughout their ICU stay. Patients were transferred to a normal ward if their oxygen requirement was <6 l/min and their SpO 2 >90%. Patients who were not mechanically ventilated due to patient will (n=3), despite indication, were also assigned to the group of patients with severe COVID-19. The time of symptom onset was speci ed by the patient.

Plasma preparation and measurement
Whole blood samples were drawn into citrate tubes (SARSTEDT S Monovetten, Nümbrecht, Germany, Citrat 3,13%). Samples were centrifuged for 10 minutes at 2000 g and plasma was stored at -80°C until further processing.

Statistical analysis
Statistical analyses were carried out with GraphPad Prism version 7.0 (GraphPad Software Inc., San Diego, CA, USA) and R v4.0.3 (R Foundation for Statistical Computing, Vienna, Austria) [29]. Descriptive variables were calculated using means with standard deviation (SD); medians and interquartile ranges (IQRs, P25%-P75%), as well as counts and percentages. For continuous variables, two-tailed Student's t-or Mann-Whitney U tests were performed. ANOVAs with Tukey's post hoc test for multiple comparisons or Kruskal-Wallis tests with Dunn's post hoc test for multiple comparisons were used to examine more than two groups. Adjusted p-values from post hoc tests were indicated (p adj ). For categorical data, Fisher's exact test was performed. Principal Component Analysis (PCA) was performed to investigate the relevance of the measurement parameters and their correlations. A p-value <0.05 was considered statistically signi cant (*p<0.05; **p<0.01; ***p<0.001).

Multiple logistic (regression) classi er
A multiple logistic regression analysis was performed including 28 patients with severe COVID-19, de ned by the requirement for mechanical ventilation at the time of the respective blood collection, and 18 patients with moderate COVID-19. The COVID-19 cohort was censored (0=ICU discharge and 1=death). The day of the censoring event was labeled as day E and chosen as a reference point. Data from the days E to E-3 were used for training and testing the model. The remaining data up to E-7 were used as complementary validation data. The labeled data were split randomly into a training (70%) and a test set (30%). Each set contained binary class information about the patients' severity status (moderate=0 and severe=1) as well as the quantitative measurements of plasma RIPK3, MLKL, HMGB1, and RIPK1 levels.
Different models were trained and evaluated, including single or combined variables. All models were calculated using a generalized linear model (GLM) of the binomial family to nd a classi er (caret package) [30]. A 10-fold cross-validation was performed to exclude a subsample bias and prevent the model from being over tted. Each model was calculated as a regular logistic regression to compare the relative goodness-of-t with Akaike's information criterion (AIC). The ideal predictors for the classi er were selected by evaluation of the classi cation performance indicators, e.g., predictors with the highest accuracy and the lowest AIC after cross-validation were chosen. The nal classi er was built with training data that showed mean accuracies >90% on days E to E-3. This model was tested with the separated validation set to determine the overall predictive quality of the classi er, e.g., with general performance parameters (Accuracy, Sensitivity, and Speci city). The odds and Odds Ratio (OR) of the predictor variables were determined from the coe cients of the nal regression model. Finally, multiple logistic regression models were performed using measurements of RIPK3, HMGB1, CRP, IL-6, PCT, LDH, and peripheral leukocyte count including a training and a test set.

Demographic characteristics and laboratory parameters
Between June 2020 and January 2021, we considered 46 COVID-19 patients admitted to the ICU, of whom 28 patients showed a moderate and 18 patients a severe COVID-19 progression during ICU stay. Overall, ICU admission occurred on day seven (4-11) after symptom onset. Patients with severe COVID-19 were older (p=0.033), showed extended ICU stay (p<0.001), and increased mortality rate (p<0.001) compared to patients with moderate COVID-19 ( Table 1). The median survival time after ICU admission in patients with severe COVID-19 was 17  days. Of all investigated comorbidities, we found a signi cantly increased rate of arterial hypertension in patients with severe compared to moderate COVID-19 (p=0.016) ( Table 1).
Additionally, we examined these measurements with symptom onset as a baseline (Fig. S1).
Prediction of severe COVID-19 progression with a multiple logistic (regression) classi er: model selection With randomly selected training data from the day of the censoring event E (ICU discharge or death) to E-3 (three days before the censoring event), we evaluated models with different predictor combinations in a time-independent manner ( Table 2). Although a model consisting of HMGB1, RIPK3, and RIPK1 and a model consisting of solely RIPK3 achieved better accuracies with 78%, their ts and thus model qualities (Akaike's information criterion (AIC)) were worse, resulting in a model with HMGB1 and RIPK3 with a marginally superior t (AIC=65.91), pointing to a simpler and, therefore, more applicable model. While this was an improvement towards the other models, a slightly lower classi cation accuracy Acc HMGB1+RIPK3 =77% was achieved. However, this model required only two measurements, opposed to three.
To further evaluate on which days plasma RIPK3, MLKL, HMGB1, and RIPK1 levels distinguished best between severe and moderate COVID-19 progression, we looked at individual days and markers backward from the censoring event (E). On days E-3 until the event, HMGB1 ( Fig. 3a-d, third column) and on days E-3 and E-1, RIPK3 (Fig. 3b,d, rst column) plasma levels were signi cantly elevated in patients with severe compared to those with moderate COVID-19. In contrast, RIPK1 and MLKL levels did not differ signi cantly between severe and moderate COVID-19 ( Fig. 3a-d, second and fourth column) and were therefore not included in our further analysis. Consequently, plasma RIPK3 and HMGB1 were selected for further evaluation.
Evaluation of the discriminatory ability of combined RIPK3 and HMGB1 plasma levels for building the COVID-19 severity classi er To organize data for training and testing the selected model, combined RIPK3 and HMGB1 plasma levels are viewed backward from the censoring event (E). Table 3 shows the classi cation performance of predicting COVID-19 progression using the training and test data. The days starting from day E to E-3 were analyzed separately. In the training data, E, E-1, and E-3 achieved accuracies of 100% ( The most stable and optimal results in discriminating between patients with moderate and severe COVID-19 were found to be days E-1 and E-3, as indicated by signi cantly higher RIPK3 and HMGB1 plasma levels (Fig. 3b,d, rst and third column) as well as by the performance of the training and test data using combined RIPK3 and HMGB1 measurements from these days (Table 3). Therefore, HMGB1 and RIPK3 plasma levels from these days were used in building the nal classi er.
Prediction of severe COVID-19 progression with combined RIPK3 and HMGB1 measurements Table 4 shows the performance of the nal classi er with the training, test, and validation data. The overall accuracies of discriminating between a moderate and severe COVID-19 progression were high (>83%). The fraction of false positives and false negatives was low, resulting in speci city and sensitivity levels >74%. The test set reached 83% accuracy as well as 89% sensitivity and 74% speci city (Fig. 4b), which was exceeded by the validation set using data from up to day E-7 (excluding E-1 and E-3) with 84% accuracy, 90% sensitivity, and 76% speci city (Fig. 4c). This was particularly accurate up to 6 days before the censoring event (Fig. 4d).
Also, RIPK3 plasma levels of patients with moderate COVID-19 approached the healthy control levels before ICU discharge (Fig. 4e). Notably, HMGB1 plasma levels indicated signi cant differences between patients with moderate and severe COVID-19 at a very early stage (E-8) (Fig. 4f). Therefore, the combination of circulating levels of RIPK3 and HMGB1 can be used to time-independently classify COVID-19 patients admitted to the ICU into potential disease severity states (Fig. 4a-c).
The odds of changing COVID-19 severity based on RIPK3 and HMGB1 levels To further estimate these ndings, a logistic regression model was calculated with the full data over the entire observation period of plasma RIPK3 and HMGB1 levels as independent variables and disease progression of COVID-19 as the dependent variable. With this model, the odds of changing the disease severity state were estimated (Table 5).
Model comparison for the prediction of severe COVID-19 progression using multiple in ammatory variables To compare the predictive power of established in ammatory markers as well as RIPK3 and HMGB1, we additionally performed a multiple logistic regression model including measurements of RIPK3, HMGB1, CRP, IL-6, PCT, LDH, and peripheral leukocyte count. Interestingly, in the training set, the combination of measured RIPK3, HMGB1, and PCT levels reached the highest accuracy (93%). In order to be able to represent the plot in two dimensions, with one variable on the x-axis and one on the y-axis, we chose a model with two measurements from our potential biomarkers. In fact, the combination of RIPK3 and HMGB1 levels discriminated best between moderate and severe COVID-19 progression with an accuracy of 86% in the training set (Table 6). In the test set, both models performed similarly, with an accuracy of 83.7% (Table 7). Therefore, plasma RIPK3 and HMGB1 are the most suitable candidates for predicting COVID-19 severity.

Discussion
In this study, plasma RIPK3, MLKL, HMGB1, and RIPK1 levels of COVID-19 patients are obtained in daily-assessed measurements throughout the whole ICU stay. Based on these data, we developed for the rst time a classi er built on RIPK3 and HMGB1 as potential biomarkers to discriminate between moderate and severe COVID-19 progression after ICU admission with an accuracy of 84%. Several independent lines of evidence support this conclusion.
First, COVID-19 intensive care patients showed continuously signi cantly higher plasma RIPK3 levels than healthy controls throughout their ICU stay, strongly indicating ongoing RIPK3-dependent necroptosis. In addition to previous investigations that considered RIPK3 levels at single time points [31,32], we revealed in our prolonged study that patients with severe COVID-19 possess higher RIPK3 plasma levels in a time-dependent manner.
Also, patients with moderate COVID-19 showed decreasing RIPK3 levels, corresponding to their recovery.
Second, we observed signi cant long-term elevations of HMGB1 in COVID-19 intensive care patients compared to healthy controls. Elevations of HMGB1 levels were associated with the requirement for mechanical ventilation and fatal outcome, as also demonstrated in our supplemental data and previous studies [33,34]. Notably, we also revealed signi cant elevations of plasma HMGB1 corresponding to severe COVID-19 progression in a disease-dependent time course. Chen et al. observed an association between exogenous human HMGB1 and stimulated angiotensin-converting enzyme 2 (ACE2) expression as an entry receptor for SARS-CoV-2 in cultured human lung epithelial cells, indicating a feedback loop that possibly worsens patients' outcomes [33]. RIPK3 and extracellular HMGB1 also contribute to endothelial dysfunction and loss of barrier integrity, considered to be involved in COVID-19 pathology [6,[35][36][37]. Since high extracellular levels of HMGB1 are particularly harmful, our results provide evidence for HMGB1 as a potential drug target in COVID-19, as has been successfully demonstrated for IL-6 signaling [38].
Third, we found that plasma MLKL and RIPK1 levels were tendentially higher in COVID-19 ICU patients compared to healthy controls, indicating an involvement of necroptosis in COVID-19 pathology. Accordingly, upregulation of phosphorylated MLKL was detected in lung tissue of SARS-CoV-2-infected mice and post mortem human lungs. In vitro, MLKL and RIPK3 contributed to cell death induction, as well as cytokine and DAMP release in SARS-CoV-2-infected cells, reinforcing our ndings in COVID-19 patients in the ICU [39]. Moreover, phosphorylated and thus activated RIPK1 was detected in pharyngeal epithelial cells of COVID-19 patients, and since respiratory tissues appeared to be a prominent sink for RIPK1 in COVID-19, its interaction with SARS-CoV-2 components is hypothesized [40,41]. However, MLKL and RIPK1 did not contribute signi cantly to COVID-19 severity and were therefore not included in the nal classi er.
In addition to our prolonged study investigating kinetic variations, our single measurements of RIPK3, MLKL, HMGB1, and RIPK1 supported our hypothesis that necroptosis plays a role in COVID-19, as described in our supplemental data. In this cohort, we also observed a loss of viable peripheral leukocytes in every examined cell subpopulation according to disease severity, particularly in patients receiving extracorporeal membrane oxygenation (ECMO), however, given that there were only 6 patients with this treatment, these results should be interpreted carefully.
To our knowledge, we are the rst to perform mtROS measurements using ow cytometry in whole blood samples of COVID-19 patients; therefore, there is still a lack of comparative studies. Other studies were carried out on cell cultures treated with plasma from COVID-19 patients or with single viral components (open reading frame 3a (ORF-3a) or the SARS-CoV-2 spike protein), as well as SARS-CoV-2-infected monocytes in vitro and respiratory samples (sputum/ Bronchoalveolar lavage (BAL)) of COVID-19 patients [42][43][44][45][46]. Nevertheless, it is important to mention these studies, but comparisons should be interpreted with caution. We show in our supplemental data that peripheral leukocytes of our COVID-19 cohort with single measurements had signi cantly lower levels of mtROS compared to healthy controls.
As patients were admitted to the ICU at different disease stages, data from the rst day after ICU admission would provide limited information.
Therefore, we took data from time points when the disease progression was already clear to build our model using plasma RIPK3 and HMGB1 levels and thereby reduced the variability that resulted from admission to the ICU at different COVID-19 stages. In everyday clinical practice, it is often not possible to predict whether a patient is close to or long before death or ICU discharge. The classi er model avoids this problem with RIPK3 and HMGB1 as promising biomarkers in COVID-19.
This study has several limitations. The timing of mechanical ventilation is a subjective outcome. However, differentiation of severity is possible because, once patients have the indication for intubation, spontaneous breathing and non-invasive ventilation are no longer su cient and a de nite state of disease progression has been reached. Since we intended to examine COVID-19 patients over a prolonged period, we decided to consider the requirement for intubation as a distinction between a moderate and severe COVID-19 progression for the study design. We cannot completely exclude an additional impact of the intubation status on plasma levels of RIPK3, MLKL, HMGB1, and RIPK1.
We are aware that measurements of 46 patients must be considered carefully, but regarding the number of blood samples received daily over a longer period of time, the study size is unusually extensive, in particular, compared to other single-center studies.
Moreover, to further explore the disease mechanisms indicated by this study, we suggest additional investigations on necroptosis markers, such as studies on other COVID-19 progressions and stages which we could not take into account e.g., non-hospitalized patients, or patients with post-COVID-19 syndrome. Finally, our data needs to be con rmed in further longitudinal clinical studies with independent cohorts of COVID-19 patients before implementation in clinical algorithms can be considered.

Conclusion
Our classi er with RIPK3 and HMGB1 as promising biomarkers in COVID-19 could help to timely identify future patients who require more intensive monitoring and bene t from maximized immunomodulatory therapy after ICU admission [38,51]. This model is simple and more accurate than models that, in addition to RIPK3 and HMGB1 plasma levels, considered in ammatory markers such as CRP, IL-6, PCT, LDH, and peripheral leukocyte count. The study was performed in accordance with the Declaration of Helsinki. Approval from the local ethics committee was obtained before the study was conducted (reference #20-643, #20-982) and a waiver regarding the requirement of written informed consent from COVID-19 patients was authorized. All participants of the control group provided written informed consent.

Consent for publication
All authors critically revised and approved the manuscript.

Data and materials availability
All data are available in the main text or the supplementary materials. Tables Table 1.Patient demographics of the COVID-19 cohort.
Data are presented as a n (%) for categorical variables or b median (interquartile range) for continuous variables. Patients' laboratory parameters are reported as the respective median of the parameter levels obtained during ICU stay. p-values comparing patients with moderate and severe COVID-19 were calculated with Mann-Whitney U test or Fisher's exact test. Additionally, patients' median laboratory parameter levels were compared to the hospital's central laboratory's threshold levels (CRP: 0.5 mg/dl, IL-6: 7 pg/ml, PCT: 0.5 ng/ml, LDH: 248 U/l, peripheral leukocyte count: 10.41 /nl).
Respective quantities in the pathological range were determined and then compared among patients with severe and moderate COVID-19 by Fisher's exact test.
COPD, chronic obstructive pulmonary disease Classi cation accuracies (%) and AICs values of the training data from day E to E-3 Table 3. Classi cation of the severity status on multiple days on and before the censoring event.
The classi cation performance of predicting COVID-19 severity using the training and test data consisting of combined RIPK3 and HMGB1 plasma levels on day E to E-3.
CI, con dence interval; Acc, accuracy Confusion matrices of the classi cation performances on training, test, and validation data with corresponding accuracy, sensitivity, and speci city Table 5. Logistic regression parameters of tting the severity state with RIPK3 and HMGB1 plasma levels.
RIPK3 and HMGB1 plasma levels of the total measurements are included in a logistic regression to calculate the odds of disease-related severity change.
The model parameters of the t are presented (***p<0.001).
Data is split into severe and moderate as described in our main method section Acc, accuracy