Abstract
Rationale Activins are inflammatory and tissue-repair-related members of the TGFβ-superfamily that have been implicated in the pathogenesis of several immuno-inflammatory disorders including sepsis/acute respiratory distress syndrome (ARDS). We hypothesized that they might be of particular relevance to COVID-19 pathophysiology.
Objectives To assess the involvement of the Activin-Follistatin-axis in COVID-19 pathophysiology.
Methods Levels of Activins -A, -B and their physiological inhibitor Follistatin, were retrospectively analyzed in 314 serum samples from 117 COVID-19 patients derived from two independent centers and compared with common demographic, clinical and laboratory parameters. Optimal-scaling with ridge-regression was used to screen variables and establish a prediction model.
Main Results The Activin/Follistatin-axis was significantly deregulated during the course of COVID-19 and was independently associated with severity and in-hospital mortality. FACT-CLINYCoD, a novel disease scoring system, adding one point for each of Follistatin >6235 pg/ml, Activin-A >591 pg/ml, Activin-B >249 pg/ml, CRP >10.3 mg/dL, LDH >427 U/L, Intensive Care Unit (ICU) admission, Neutrophil/Lymphocyte-Ratio >5.6, Years of Age >61, Comorbidities >1 and D-dimers >1097 ng/ml, efficiently predicted and monitored fatal outcome independently of multiplicity and timing of sampling (AUC: 0.951±0.032, p<10-6). Validation in 35 samples derived from a third hospital indicated comparable AUC (0.958±0.086, p=0.032).
Conclusion This study unravels the link between Activin/Folistatin-axis and COVID-19 mortality and introduces FACT-CLINYCoD, a novel pathophysiology-based tool that copes with the dynamic and heterogeneous nature of COCVID-19, predicts disease outcome and supports clinical decision making. Prospective large-scale validation of this calculator, as well as investigation of the mechanisms linking Activin/Folistatin-axis to COVID-19 pathogenesis is warranted.
Introduction
The evolving pandemic of COVID-19 constitutes presently the most prominent public-health issue worldwide. Thus far, severe acute respiratory syndrome coronavirus 2 (SARS-CoV-2) is responsible for more than 26 million COVID-19 cases and more than 870,000 deaths (1). The initial phase of COVID-19 relates to viral pathogenic effects and is characterized by mild “flu-like” symptoms (2). However, one week after onset-of-symptoms, ~20% of patients develop moderate-to severe-pneumonia often complicated by a cytokine-release-syndrome (CRS)-like hyper-inflammatory reaction that can lead to ARDS and multi-organ failure, thereby increasing substantially the in-hospital mortality (2–5). Various comorbidities and aging negatively affect outcome of COVID-19 (3, 6). Accumulating evidence indicate that excessive activation of innate-immune pathways, deregulated neutrophils and thrombotic microangiopathy characterize the maladaptive host-response that drives COVID-19 pathophysiology (7–11). These pathomechanisms lead to rapid progression of hypoxemic respiratory failure and protean clinical manifestations, since almost any organ can be involved (12, 13). Monitoring hospitalized COVID-19 patients, in the midst of multiple continuously changing parameters and shortage of validated and effective therapeutics, is challenging (2, 3, 5, 13). Therefore, development of novel therapeutic strategies and bed-to-bench tools permitting day-to-day prediction of patient outcome would be of utmost importance for such dynamically evolving and clinically heterogeneous disease (12, 14–16).
Activins-A and -B are members of the Transforming Growth Factor-β (TGF-β)-superfamily implicated in the regulation of numerous aspects of inflammation and/or tissue remodeling (17–19). Follistatin, a physiological Activin-inhibitor, binds to them almost irreversibly and induces endocytosis and proteolytic degradation thus modulating their bioavailability (20). Activins and Follistatin, are synthesized continuously in healthy tissues (21), however, in immuno-inflammatory conditions, epithelial, endothelial, interstitial stroma cells and different immune cells secrete higher levels that can be detected in serum and could function as biomarkers of local or systemic stress (17–19).
We have previously described increased levels of Activin-A and Follistatin in bronchoalveolar lavage (BAL) of ARDS patients (18, 22). Moreover, we showed that ectopic expression of Activin-A in murine lungs causes ARDS-like pathology (22), which shares cardinal features of COVID-19 pathophysiology. These include, strong early mobilization of neutrophils in the lung, alveolar epithelial and endothelial cell-loss, establishment of a systemic hyper-coagulant state associated with increased tissue-factor (TF) mRNA levels and induction of a cytokine-storm like response characterized by high levels of IL-6 and TNFα (5, 7, 8, 18, 23).
In view of these findings we hypothesized that the Activin/Follistatin-axis (A/F-axis) might be particularly relevant to COVID-19. To validate this hypothesis, we analyzed sera from COVID-19 patients and found that Activin-A, -B and Follistatin were significantly upregulated during the crucial period when patients deteriorate. Of note, A/F-axis components were independently associated with disease severity and in-hospital mortality. Based on that, we developed a simple clinical tool for the prediction and monitoring of COVID-19 outcome in real-life using Activin-A, -B, Follistatin and common clinical/laboratory parameters.
Material and methods
Study design
This is a retrospective study with a single endpoint, the final outcome (survival or death). Activin-A, -B and Follistatin, absolute-neutrophil-count (ANC), absolute-lymphocyte-count (ALC), neutrophil/lymphocyte-ratio (NLR), C-reactive protein (CRP), lactate dehydrogenase (LDH), ferritin, D-dimers, were analyzed in randomly acquired samples from patients which were diagnosed with COVID-19 from March 10, 2020 and had an outcome until July 7, 2020 (Supplementary Figure S1). Three national reference hospitals from distant regions of Greece participated in the study (University General Hospital of Alexandroupolis; “AHEPA” University General Hospital of Thessaloniki; “Attikon” University General Hospital of Athens).
Inclusion criteria were: a) adult patients (>16 years old), any gender, b) positive SARS-CoV-2 RT-PCR testing in nasopharyngeal swab or BAL, c) hospitalization due to COVID-19 at any disease stage, d) final disease outcome until July 7, 2020. Patients without available sample obtained before the 21st day-of-disease were excluded (Supplementary Figure S1). The study protocol was approved by all relevant Institutional Ethical Committees and is aligned with the Helsinki declaration.
Clinical and laboratory data
An initial cohort of 117 consecutive COVID-19 patients hospitalized at University Hospital, Alexandroupolis and “AHEPA” Hospital, Thessaloniki was endorsed and 314 samples were analyzed. Patients and sample characteristics are presented in Table 1. A second independent cohort of 28 consecutive COVID-19 patients derived from a distinct hospital (”Attikon” Hospital, Athens) contributing 35 samples was used for validation (Supplementary Table S1). Comorbidities considered were diabetes mellitus, arterial hypertension, dyslipidemia, obesity, atrial fibrillation, coronary disease, heart-failure, renal-failure, chronic obstructive pulmonary disease or asthma, immunosuppression (without a record of malignancy), autoimmunity and cancer. The disease-status (DS) of COVID-19 patients was classified as DS1 (mild), DS2 (moderate), DS3 (severe) and DS4 (critical) as suggested by Hadjadj et al. (10) (details in Supplementary Materials).
Characteristics of patients (n=117) and samples (n=314).
Common laboratory parameters including CRP, LDH, D-dimers, ANC, ALC and NLR were measured as part of standard of care.
Immunoassays
Activin-A, -B and Follistatin levels in serum were measured using enzyme-linked immunosorbent assay (ELISA) (Ansh Labs, Webster, TX, USA) according to manufacturer’s instructions.
Statistical analysis
To elucidate the role of A/F-axis molecules, we identified potent confounders from true predictors of outcome using general linear model. Prognostic value of predictors was validated; for that purpose, all independent variables of interest were transformed to binary ones through nominal optimal scaling along with discretization to two groups. Imputing of missing data was added, ridge-regression was selected for regularization and 10-fold cross-validation was added. Results were adapted to a point-system scoring, where one additive point was given for each parameter included in the unfavorable category as suggested after discretization; binary-regression was utilized to mathematically approach outcome probability. Evaluation of the scoring systems was based on Area Under Curve (AUC) as determined from Receiver Operating Characteristic (ROC) analysis. Optimal-Scaling procedure was utilized to detect whether the scoring systems could predict response to certain treatment options. Extended version of the Material and Methods is provided in Supplementary Materials.
Results
The Activin/Follistatin axis is highly deregulated in COVID-19 non-survivors
To validate the hypothesis that the A/F-axis is linked to COVID-19 pathophysiology we analyzed serum levels of Activin-A, -B and Follistatin in a cohort of 117 COVID-19 patients. The day-of-symptom-onset (disease-day) was used to align the data derived from serum samples. Activin-A, -B and Follistatin, were substantially increased, particularly in samples derived from patients that eventually succumbed to infection (non-survivors) (Figure 1, Supplementary Table S2 and Supplementary Figure S2). Increased levels of A/F-axis components were observed approximately 7-28 days from onset of symptoms (Figures 1, 2) during which the majority of COVID-19 patients normally recover, however, some deteriorate into severe pneumonia and a portion of them become critically-ill and eventually die. Classical parameters such as ANC, ALC, NLR, CRP, LDH and D-dimers became derailed within the same time window (Figure 1). However, whereas CRP, neutrophils and NLR were elevated already from first week of disease and D-dimers were gradually increasing over time, A/F-axis components and LDH were upregulated during the second week of disease, when COVID-19 leads to hospitalization (Figure 2). Interestingly, comparative analysis of A/F-axis components and classical parameters in survivors vs non-survivors on the basis of disease-status at the moment of serum sampling demonstrated upregulation of Activin-A, -B and Follistatin in sera from both severe (DS3) and critical (DS4) non-survivors (Supplementary Figure S2). Neutrophils, CRP and D-dimers were upregulated in samples obtained from DS3 non-survivors, as well as, DS4 survivors and non-survivors; LDH levels were particularly elevated in very advanced disease cases (DS4 non-survivors) (Supplementary Figure S2).
Serum levels of A/F-axis components and other key inflammatory parameters in the blood of COVID-19 patients plotted over time according to the days of symptom onset. Light blue circles represent samples collected from patients that survived COVID-19 and magenta circles represent samples obtained from patients that eventually succumbed to the disease. Vertical dotted lines indicate time-windows of first, second and more than two-week intervals used for statistical analysis. Horizontal dotted lines indicate threshold as determined by the nominal optimal scaling along with discretization to two groups.
Concentrations of Activin/Follistatin-axis components in serum and levels of other key inflammatory parameters in the blood of COVID-19 patients plotted over time according to the days of symptom onset and grouped per week. Light blue bars represent samples collected from patients that survived COVID-19 and magenta bars represent samples obtained from patients that eventually succumbed to the disease. Data are expressed as mean±SEM analyzed using one-way analysis of variance with Tuckey’s post-hoc test. Asterisks represent comparison of the indicated group vs the survivors of week 1 (0-7 days from onset of disease). Asterisks above a horizontal line represent comparison between the groups under the line segment (*P<0.05, **P< 0.01, and ***P < 0.001)
To assess whether sampling (samples/patient, disease-day at sampling) and origin of patients (hospital of admission) might blur the results, a general linear model including Activin-A, Activin-B and Follistatin as dependent variables and multiplicity of samples/patient, sampling period (1-7/8-14/15+ days), and reference hospital as independent were introduced. The model demonstrated that Activins and Follistatin were not correlated with these potent confounders (Supplementary Table S3). Therefore, we concentrated our efforts to construct a sample-based rather than patient-based predictive model.
Given that time from disease-onset is a key parameter for disease aggravation and hospital admission, we aimed to further elucidate how disease-time affects the correlation between A/F-axis and outcome. Of note, we noticed a more pronounced independent correlation of Follistatin with outcome (Supplemetary Figure S3).
Furthermore, using binary-regression we concluded to a model incorporating Follistatin with classical parameters NLR, Age, ICU admission and comorbidity number, notably however, without the inclusion of other well-recognized soluble immuno-inflammatory biomarkers such as CRP, LDH, D-dimers or Ferritin (Supplementary Table S4). This model revealed the independent key role of Follistatin in predicting outcome and provided a good estimate; however, it did not reach gold standard level (ROC analysis, Supplementary Figure S4) and therefore we sought to further improve it.
The Activin/Follistatin-axis in the dynamically shifting phenotypic heterogeneity of COVID-19
We hypothesized that the erratic and heterogeneous behavior of COVID-19 should be mirrored by analogous fluctuations in the profiles of key biomarkers. We therefore opted to investigate the correlation between different biomarkers in the entire sample collection. The correlation of Activin-B levels to LDH, and Activin-A levels to CRP and NLR are shown as representative comparisons (Figure 3). In Activin-B vs LDH comparison, some samples exhibited good correlation but others did not. Notably, in Activin-A vs CRP or NLR comparisons, the parameters did not correlate at all. To understand better this finding, we highlighted in similar plots the groups of samples derived from non-survivor patients AL022, AL034, AL035, AL062, AL063 and AX112 for which multiple samplings were available (Figure 3, lower panel). Remarkably, this depiction revealed marked phenotypic heterogeneity among patients and highlighted the dynamic heterogeneity among samples collected from the same patient at different disease-days. For example, for patients AL022 and AL035, disease progression and eventual death were associated with concordant changes in Activin-B and LDH levels. However, patients AL063 and AX112, in addition to A/F-axis components, were characterized by substantial changes in CRP levels, and for patients AL035 and AL062 disease progression was associated with eventual reduction of CRP and NLR. Finally, patient AL034 was characterized by borderline levels of Activins -A and -B and high levels of CRP. Interestingly, patient AL022, who received a dose of Tocilizumab on disease-day 11 (one day after admission), exhibited for ~10 days transient reduction of LDH, CRP, NLR, Activin-A and -B values, before reverting and dying on disease-day 28. Therefore, in addition to intrinsic phenotypic variability among patients, treatments could also influence biomarker profile during hospitalization. From all the above we concluded that a broader spectrum of biomarkers, including both early and later modulated ones, had to be integrated to develop a meaningful disease-monitoring scoring system.
The FACT-CLINYCoD score for dynamic monitoring of COVID-19 outcome
Considering the dynamically evolving and heterogeneous phenotypic changes characterizing COVID-19 progression, to improve the capacity of our scoring system to discriminate survivors from non-survivors, in a simple and clinically meaningful manner, we converted all continuous variables to binary through selection of optimal cutoffs. Optimal-Scaling procedure was performed along with ridge-regression to investigate any parameter that could independently predict outcome (Supplementary Table S5 and Supplementary Figure S5). We concluded in a 10-parameter model as presented in Table 2 and Supplementary Table S6. Based on these data, we constructed a ten-point scoring-system where one additive point was arbitrarily given for each parameter included in the unfavorable category as suggested after discretization, namely if Follistatin >6235 pg/ml, Activin-A >591 pg/ml, Activin-B >249 pg/ml, CRP >10.3 mg/dL, LDH >427 U/L, ICU admission, NLR >5.6, Years of Age >61, Comorbidities >1 and D-dimers >1097 ng/ml. We evaluated this score with ROC analysis, yielding to an AUC value of 0.951±0.032 (p<10-6), that indicates outstanding discrimination in foretelling survivors from non-survivors (Figure 4A). This score was titled FACT-CLINYCoD, being an acronym of F(ollistatin), ACT(ivins), C(RP), L(DH), I(CU admission), N(LR), Y(ears of age), Co(morbidities) and D(-dimers). A score ≥4 has 90.8% sensitivity and 87.5% specificity to predict fatal outcome, whereas a score ≥5 has 81.5% sensitivity and 95.7% specificity. The FACT-CLINYCoD score distinguishes survivors from non-survivors at admission and outcome and monitors accurately disease progression independently of disease-day or status (Figure 4C,D,E). Removal of the A/F-axis components from FACT-CLINYCoD yields to a 7-point scoring system with a comparable AUC (0.938, 95% CI 0.902-0.974); however, its substantially reduced sensitivity to predict >50% probability of fatal outcome (52.3%) underlies clearly the additive value of Activins and Follistatin. The probabilities of death calculated at different disease-days in non-surviving patients AL022, AL034, AL035, AL062, AL063 and AX112 are shown in Figure 3, next to each time sample analyzed. Interestingly, some of these patients had very high scores already at first sampling, however, others, such as AX112 and AL035 had low scores at admission and either gradually or rapidly deteriorated and eventually died. The latter type of patients highlights the value of continuous monitoring COVID-19. In all cases, the FACT-CLINYCoD system was able to monitor clinical deterioration and in the case of AL022, was not deceived by the dramatic reduction of LDH, Activin-A and NLR after administration of Tocilizumab (Figure 3, low panel) and kept indicating >85% probability of death.
FACT-CLINYCoD scoring and predictability of outcome.
Upper panel: Correlation between Activin-B and LDH or Activin-A and CRP or NL-ratio levels in the randomly acquired serum samples from our cohort. Vertical and horizontal dotted lines indicate the thresholds as determined by the nominal optimal scaling along with discretization to two groups. Lower panel: Similar correlations as above, however, serum samples derived from the same patient are indicated with distinct colors and connected with arrows to indicate the direction of disease progression towards death. Numbers within the circular points indicate disease-day. Percentages indicate the probabilities of death estimated by the FACT-CLINYCoD scoring system developed herein. TOCI: Tocilizumab
ROC analysis for FACT-CLINYCoD score in initial (A) and validation cohort (B). The FACT/CLINYCoD scoring system distinguishes survivors from non-survivors both at admission and at the final outcome (C), irrespectively of disease-day (D) and disease status (E). Light blue bars represent samples collected from patients that survived COVID19 and magenta bars represent samples obtained from patients that eventually succumbed to the disease. Data are expressed as mean±SEM analyzed using one-way analysis of variance with Tuckey’s post-hoc test. Asterisks above a horizontal line represent comparison between the groups under the line segment. Solitary asterisks in B and C represent comparison of the indicated group with the survivors in the first group of the corresponding graph (*P<0.05, **P< 0.01, and ***P < 0.001).
Validation of the FACT-CLINYCoD score
The FACT-CLINYCoD score was validated using an independent cohort of 28 consecutive patients contributing 35 random samples. Although this small cohort differed significantly from the initial regarding age, disease-status, treatments, sampling approach, and day from disease-onset (Supplementary Table S1), a comparable AUC was observed (0.958, 95% CI: 0.872-1.000, P=0.032, Figure 4B).
FACT-CLINYCoD score as potential predictor of treatment response
The value of FACT-CLINYCoD score in supporting treatment decision algorithms was evaluated, as indicated by ROC analysis (Supplementary Figure S6). More specifically, Optimal-Scaling procedure was performed along with ridge-regression to investigate any treatment that could independently predict outcome. Hydroxychloroquine, azithromycin, antibiotics other than azithromycin, Lopinavir/Ritonavir, low molecular weight heparin (LMWH), anakinra and corticosteroids were found to be independently associated with survival when treatment commenced at FACT-CLINYCoD scores <3 (Supplementary Table S7). The scoring system proposed herein is thus promising in guiding timely treatment administration and predicting favorable response to therapeutic intervention. However, the small number of patients dictates caution in interpreting these findings.
Discussion
This study provides evidence suggesting that inflammation and tissue stress-related proteins Activins-A,-B, and their natural inhibitor Follistatin are tightly associated with severity and outcome of COVID-19. Their upregulation was prominent in non-survivors and was independently related to in-hospital mortality. Of note, combination of A/F-axis components with common clinical and laboratory parameters permits prediction of COVID-19 outcome throughout the course of the disease.
Identification of high-risk patients and death-prediction are of particular importance especially in strained health-care systems. Hence, several studies attempted to develop scoring systems to predict disease severity and mortality as early as possible (4, 6, 14, 15, 24–26). Most of these models were based on demographic parameters and instantaneous evaluation of subjective symptoms (i.e. dyspnea) combined with single measurements of common laboratory markers, such as neutrophil or lymphocyte counts, LDH, CRP and several proinflammatory cytokines (4, 6, 14, 15, 24, 26, 27). These models are undoubtedly useful; however, they place main emphasis on prediction of final-outcome at admission and thus may fail to predict some non-survivors most likely due to the complex and often erratic development of COVID-19 (exemplified by patients AX112 and AL035) and the downregulation of key biomarkers at later disease stages (Figure 3). Moreover, most of the previous predictive systems were developed based on patient characteristics of the first waves of COVID-19, i.e. older with several comorbidities and it cannot be excluded that the demographics of the next waves will be different.
Considering the above, we exploited three novel components to develop a more dynamic COVID-19 monitoring system. First, we determined levels of Activins and Follistatin, taking into consideration their previously proposed roles for sepsis/ARDS pathogenesis, neutrophil-mediated inflammation, coagulopathy, endothelial cell stress, angiogenesis and post-inflammatory pulmonary fibrosis/remodeling, all of which are key characteristics of COVID-19 pathophysiology (7, 8, 18, 22, 28). We reasoned that upregulation of A/F-axis could indicate tissue-damage in the lung, vasculature or other vital peripheral tissues, thus complementing information derived from other biomarkers reporting tissue-damage (i.e. LDH) or coagulation/vascular injury (i.e. D-dimers) (2, 15, 25, 26). Second, we used a sample-instead of patient-based analytic strategy, since regression models demonstrated that kinetics of A/F-axis components correlated independently with disease-day at sampling, and not with number of samples/patient or patient’s origin. Third, we sought to exploit the differences in the kinetics of A/F-axis components and other classical biomarkers such as NLR, CRP, LDH and D-dimers (Figure 2) to develop a monitoring system that will not only make early predictions but, equally important, will allow readjustment of predictions during disease progress. Our goal was to provide a monitoring system that could assist real-time decision making in the context of a complex and erratic disease, such as COVID-19 (2, 3, 5, 12, 13, 16).
In the pre-COVID-19 era, the A/F-axis was evaluated as predictor of outcome in critical care patients (29). Thus, increased levels of Activin-A and -B in sera from ICU patients with acute respiratory failure could predict 90-days and 12-months survival with ~80% accuracy; Follistatin did not provide any extra predictive value (29). A/F-axis components were measured also in sera of critically-ill, influenza-A (H1N1), patients; Despite the increased Activin-A, -B and Follistatin, no significant association with disease severity was established (30). However, while preparing the current manuscript, consistent with our conclusions, a preprint study reported Follistatin among the circulating markers of endothelial damage associated with in-hospital mortality in a small number of COVID-19 patients (31).
Although the exact mechanism linking Activins to COVID-19 pathogenesis remains unknown, the importance of A/F-axis in the disease is reflected on the predictive value of FACT-CLINYCoD score; This model, depending on A/F-axis by one third, demonstrates an almost perfect AUC (0.951±0.032) as well as a very satisfactory sensitivity (81.5%) and specificity (95.7%) of scores ≥5 to predict fatal outcome. In agreement with this, the key role of A/F-axis components in the prediction of fatal outcome at any time of disease course (Figure 4C,D) is underlined by the relevant ROC analysis; Follistatin is the most significant predictive component of A/F-axis (Supplementary Figures S3, S5). Moreover, removal of the triad Activin-A, -B and Follistatin from FACT-CLINYCoD leads to substantial reduction in sensitivity to predict >50% probability of fatal outcome, further illustrating the link of A/F-axis to COVID-19 final outcome and its added-value in the scoring system as deduced from binary-regression and ridge-regression coefficients (Supplementary Table S4, S5). The FACT-CLINYCoD score constitutes a balanced, robust and flexible tool efficiently intertwining pathophysiology of A/F-axis, clinical profile (by means of NLR, LDH, D-dimers and CRP) and key parameters that are well established to affect mortality (ICU admission, age, and comorbidities), capable of providing dynamic outcome prediction.
An obvious emerging question is whether the A/F-axis is a suitable target for COVID-19 therapeutics. Indeed, several studies have proposed Follistatin or soluble Activin type-II receptors as therapeutics for sepsis, ARDS and fibrotic disorders (18, 32). Although the increase of Activin-A and -B could be interpreted in favor of utilizing such therapeutics, the dramatic increase of Follistatin, often at a stoichiometry that surpasses substantially the sum of Activins -A and -B in serum, could argue against it.
Neutrophils can release preformed Activin-A upon arrival in inflamed tissues (18, 33), whereas, other inflammatory cells such as monocytes, CD4+ T cells and tissue resident cells can secrete Activin-A later on (18, 21, 34). On the other hand, neutrophils do not secrete Activin-B and therefore this factor must be produced by other cell types, probably acting as a biomarker of vascular stress and hypoxia (35, 36). Follistatin can be produced locally or released systemically from distant organs such as the liver (18). The tissue origin of the increased A/F-axis components in the sera of COVID-19 patients and the Activin/Follistatin stoichiometry in the diseased tissues are unknown. Therefore, the suitability of the A/F-axis as therapeutic target for COVID-19 warrants further careful investigation.
Another promising aspect of this study is the potential utility of FACT-CLINYCoD to monitor response to various treatments targeting SARS-CoV-2 (antivirals) or hyper-inflammatory host reactions (heparin, corticosteroids and anti-cytokines). The data, describing favorable response to treatments such as hydroxychloroquine, antibiotics, lopinavir/ritonavir, LMWH, and corticosteroids (Supplemental Figure S6, Supplemental Table S7), are preliminary and, given the small number of patients in each group and the non-randomized nature of this study, should be considered at the moment with caution. Nevertheless, preliminary results showed favorable outcome when LMWH was commenced at FACT-CLINYCoD score <3. This is consistent with emerging clinical data linking reduced mortality with early administration of LMWH in all hospitalized COVID-19 patients (37) and supports the detrimental role of immunothrombosis in COVID-19 (7).
In conclusion, A/F-axis dysregulation is tightly associated with poor outcome of COVID-19. FACT-CLINYCoD, a novel pathophysiology-driven monitoring system that incorporates A/F-axis components and classical biomarkers, enables dynamic prediction of disease outcome and may support real-time medical decision. Prospectively, large-scale, multinational validation of this calculator, as well as investigation of the mechanisms linking A/F-axis to COVID-19 pathogenesis is absolutely warranted. These may offer new and highly needed strategies for management of COVID-19 pandemic, given the fact that despite extensive efforts, safe and effective vaccines against SARS-CoV-2 might take long before they can be administered to people.
Data Availability
All data referred to in the manuscript are avaliable upon request.
Conflict of Interest
None to declare
SUPPLEMENTARY MATERIAL
Disease status
The disease-status (DS) of COVID-19 patients was classified based on the adaptation of the Sixth Revised Trial Version of the Novel Coronavirus Pneumonia Diagnosis and Treatment Guidance, as described previously (1). Mild cases (DS1) were defined as mild clinical symptoms (fever, myalgia, fatigue, diarrhea) and no sign of pneumonia on thoracic X-Ray or/and CT scan. Moderate cases (DS2) were defined as clinical symptoms associated with dyspnea and radiological findings of pneumonia on thoracic X-Ray or/and CT scan, and requiring a maximum of 3 L/min of oxygen. Severe cases (DS3) were defined as respiratory distress requiring more than 3 L/min of oxygen and no other organ failure. Critical cases (DS4) were defined as respiratory failure requiring mechanical ventilation, shock and/or other organ failure that require an intensive care unit (ICU).
Immunoassays
Activin-A, -B and Follistatin levels in serum were measured using enzyme-linked immunosorbent assay (ELISA) (Ansh Labs, Webster, TX, USA) according to manufacturer’s instructions. For Activin-A (Cat# AL-110) limit of detection was 65 pg/ml and dynamic range 0.1-10 ng/ml. For Activin-B (Cat# AL-150) limit of detection was 4.35 pg/ml and the dynamic range 12.7-1400 pg/ml. For Follistatin (Cat# AL-117) limit of detection was 180 pg/ml and dynamic range 0.612-20 ng/ml. Baseline levels of Activin-A, -B and Follistatin in 19 normal, pre-COVID-19 sera, were 404±193, 144±53 and 2656±1784 pg/ml, respectively, very similar to those previously reported by others using the same reagents (2).
Statistics
To search for biases, we compared all measured parameters between patients and samples using either Pearson’s χ2 or a two-sample independent t-test in case of discrete and continuous variables respectively; in case of statistically significant increase in homogeneity of variance, as checked by Levene’s test, Mann-Whitney U test was additionally performed. ANOVA along with Tuckey’s post-hoc test was used to compare more than two groups (namely samples derived from days 1-7, 8-14 and 15+) in case of continuous variables distributed uniformly; else, Kruskall-Wallis test was alternatively used. General linear model was used to dissect true independent correlations of the A/F axis molecules from potent confounders.
Initially, multivariate analysis of risk factors for outcome was performed with binary logistic regression over imputed data; outcome was considered as a dependent variable and every parameter of interest was treated as a potent independent parameter (the probability for stepwise entry and removal were set to 0.05 and 0.10 accordingly, the classification cutoff was set to 0.5 and the maximum number of iterations was set to 20). Models including parameters with tolerance <0.4 (variance inflation factor >2.5), as deduced from corresponding linear regression analysis assessing numerical values to outcome, were rejected to avoid collinearity. Probability of unfavorable outcome (death) was given by the formula p=1/(1+e-Π), where Π stands for constant + ∑bnxn; 95% CI of p lie between exp[Π-∑1.96* SE(bn)]/{1+exp[Π-∑1.96*SE(bn)]} and exp[Π-∑1.96*SE(bn)]/{1+exp[Π+∑1.96* SE(bn)]}. Furthermore, ridge regression was used to avoid model overfitting, tolerate large variances and overcome collinearity obstacles, all at the least possible additive bias; all initially selected variables (LDH, CRP, NLR, D-dimers, ANC, ALC, Ferritin, Age, Comorbidities, ICU admission, Follistatin, Activin-A, and Activin-B, multiplicity of sampling, period of sampling, and reference hospital) were transformed to binary ones through nominal optimal scaling along with discretization to two groups, imputing of missing data was added, and 10-fold cross-validation was selected through Optimal Scaling procedure (SPSS CATREG). However, as ridge regression needed further interpretation and adaptation to reach clinical usefulness, we constructed and evaluated a point-system scoring incorporating all parameters with independent prediction validity (LDH, CRP, NLR, D-dimers, Age, Comorbidities, ICU admission, Follistatin, Activin-A, and Activin-B). One additive point was given for each parameter included in the unfavorable category as suggested after discretization; for thatpurpose, binary regression was utilized. The accuracy of the scoring systems was assessed using the Area Under Curve (AUC) as determined from Receiver Operating Characteristic (ROC) analysis. Optimal Scaling procedure was utilized as described before to detect whether the scoring systems could predict response to certain treatment options.
The minimum sample size for both survivors and non-survivors corresponding to an AUC>0.750 with a<0.05 and b<0.20 was evaluated with the use of MedCalc Statistical Software version 19.4.1 (MedCalc Software Ltd, Ostend, Belgium; https://www.medcalc.org; 2020). Imputed data tolerated missing values at a maximum of 10%. The level of statistical significance was set to p=0.05. All numerical values are given with at least two significant digits. Means are accompanied by their 95% confidence intervals (CI). Statistical analysis was performed with the use of IBM SPSS Statistics software, version 26.0.0.0, for Windows.
Binary regression predictive model highlighting the key importance of Follistatin in the prediction of COVID-19 outcome
A binary regression model, including ICU admission, number of comorbidities, age, NLR, and Follistatin as dependent variables and outcome as independent one, was used to predict outcome independently of the disease time-line (Supplementary Table S4). Probability of fatal outcome is given by the formula p=1/(1+e-Π), where Π stands for constant + ∑bnxn, namely –7.763 – 2.819*(ICU admission) + 0.511*(comorbidities number) + 0.184*(NLR) + 0.075*(Age) + 0.00023*(Follistatin) and 95% CI of p lie between exp[Π∑1.96*SE(bn)]/ {1+exp[Π-∑1.96*SE(bn)]} and exp[Π∑1.96*SE(bn)]/{1+exp[Π+∑1.96*SE(bn)]} (SE can be extracted by the 95% CI given in Supplementary Table S4 by dividing the difference between -95% CI from +95%CI with 3.92). ROC analysis revealed good AUC values for the above mentioned calculator (Supplemental Figure 3).
The FACT-CLINYCoD score for COVID-19 outcome prognosis
Optimal Scaling procedure along with ridge regression yielded to a 10-parameter model as presented in Supplementary Table S5 and Figure 4. Based on these data, we constructed a ten point-system scoring where one additive point was given for each parameter included in the unfavorable category as suggested after discretization, namely if Follistatin >6235 pg/ml, Activin-A >591 pg/ml, Activin-B >249 pg/ml, NLR >5.6, LDH >427 U/L, CRP >10.3 mg/dL, D-dimers >1097 ng/ml, Age >61 years, Comorbidities >1, and ICU admission.
A binary regression model incorporating outcome as a dependent variable and score as an independent one and yields to the formula p=1/(1+e6.198 - 1.330*score), where p stands for the probability of death at the time of sampling; ±95% CI of p lie between e-6.547+1.330*score/(1+e-6.547+1.330*score) and e-5.849+1.330*score/1+e-5.849+1.330*score) (Table 2).
SUPPLEMENTARY TABLES
Characteristics of samples derived from validation cohort (n=35) in comparison with samples derived from initial analysis cohort (n=314).
Comparison between survivors and non-survivors (samples).
General linear model for detection of potent confounders.
Parameters independently correlated with outcome based on binary regression analysis.
Outcome predictive model as derived using Optimal Scaling procedure along with ridge regression regularization (leading to FACT-CLINYCoD score).
Binary regression for quantification of outcome probability based on the 10-parameter predictive model derived from optimal scaling.
Optimal survival using FACT-CLINYCoDscore cut-offs in treatment options.
SUPPLEMENTARY FIGURES
Flow chart of patients included in the initial cohort of the study.
Levels of Activin/Follistatin-axis components in serum and levels of other key inflammatory parameters in the blood of COVID-19 patients grouped according to disease-status (DS) at sampling. Light blue bars represent samples collected from patients that survived COVID-19 and magenta bars represent samples obtained from patients that eventually succumbed to the disease. Data are expressed as mean±SEM analyzed using one-way analysis of variance with Tuckey’s post-hoc test. Asterisks represent comparison of the indicated group with the survivors of DS1-2. Asterisks above a horizontal line represent comparison between the groups under the line segment (*P<0.05, **P< 0.01 and ***P < 0.001)
Performance of the indicated single parameters in distinguishing survivor from non-survivor COVID-19 patients. ROC analysis for parameters used in both binary regression and Optimal-Scaling procedures.
Performance of a prediction model based on Follistatin, NLR, Age, ICU admission and number of comorbitities in distinguishing COVID-19 survivors from non-survivors. Binary regression ROC analysis (details referred in Supplementary Table S4).
Ridge regression paths for model leading to FACT-CLINYCoD score.
The performance of the FACT-CLINYCoD score in predicting favorable response of COVID-19 patients to the indicated treatments. LMWH: low molecular weight heparin.
Author contribution
E.S, G.D, and A.G designed and conducted experiments, analyzed data and contributed to writing the manuscript; EG, GL, CT, OT, ES, PR, DB, GG and SM provided clinical samples and analyzed data; OR and AP provided reagents, designed experiments and edited the manuscript; DB, GG and SM interpreted findings and edited the manuscript. P. Sideras, P. Skendros and VP conceived the study, designed experiments, analyzed data, developed the scoring system and prepared the first draft of the manuscript. All authors gave final approval for the manuscript version to be published. P. Sideras and P Skendros co-supervised the study.
Funding Support
This work was supported by grants from the General Secretariat of Research and Technology, Greece, CYTONET (Τ1ΕDK-00617) and EATRIS-GR (OPS5028091).
Acknowledgements
We thank Dr. Apostolos Vasileiou for technical assistance and Dr. Mika Laitinen for critical reading of the manuscript.
References
Supplemental References
