The role of asymmetric dimethylarginine (ADMA) in COVID-19: association with respiratory failure and predictive role for outcome

We aimed to assess the potential role of Asymmetric dimethylarginine (ADMA) in conditioning respiratory function and pulmonary vasoregulation during Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV2) infection. Within 72 h from admission, samples from 90 COVID-19 patients were assessed for ADMA, SDMA, L-arginine concentrations. In addition to classical statistics, patients were also clustered by a machine learning approach according to similar features. Multivariable analysis showed that C-reactive protein (OR 1.012), serum ADMA (OR 4.652), white blood cells (OR = 1.118) and SOFA (OR = 1.495) were significantly associated with negative outcomes. Machine learning-based clustering showed three distinct clusters: (1) patients with low severity not requiring invasive mechanical ventilation (IMV), (2) patients with moderate severity and respiratory failure whilst not requiring IMV, and (3) patients with highest severity requiring IMV. Serum ADMA concentration was significantly associated with disease severity and need for IMV although less pulmonary vasodilation was observed by CT scan. High serum levels of ADMA are indicative of high disease severity and requirement of mechanical ventilation. Serum ADMA at the time of hospital admission may therefore help to identify COVID-19 patients at high risk of deterioration and negative outcome.


Patients and methods
Patients. This retrospective study was conducted on 90 COVID-19 patients (70% males; mean age 63 ± 12 years) admitted to Academic Hospital of Udine between July 2020 and March 2021 with a diagnosis of SARS-CoV2 infection as confirmed by a positive reverse transcriptase-polymerase chain reaction (RT-PCR) on nasopharyngeal swabs. Patients admitted to hospital due COVID-19 were eligible for inclusion. Exclusion criteria included pregnancy and being younger than 18 years of age. The spectrum of COVID-19 disease severity ranged from mild, self-limiting respiratory tract illness to severe progressive pneumonia, multi-organ failure, and death. Baseline characteristics included: age, gender, and blood exams noted to be related to disease severity and collected within 72 h of admission to the hospital such as white blood cell count (WBC), lymphocyte count, lactate dehydrogenase (LDH), Mid-Regional proadrenomedullin (MR-proADM), C-reactive Protein (CRP), Interleukin-6 (IL-6); a past medical history of cardiovascular disease, diabetes, chronic kidney disease, chronic respiratory diseases; Charlson Comorbidity Index (CCI); clinical severity with SOFA score and COVID-19 WHO Severity Classification 19 upon hospital admission. All blood samples analyzed were collected as part of routine clinical care on admission to the hospital.
Sample size analysis. The sample size was computed in relation to the analysis of variance for ADMA. Three groups were considered: patients without respiratory failure (Group 1), patients with respiratory failure and no IMV (Group 2), and patients with respiratory failure and IMV (Group 3). The effect size for ANOVA was determined considering the ADMA values reported in the literature 20 ; the expected mean values of ADMA in the three groups were: 0.6, 0.6, 0.8 (µmol/L), while the SD was 0.2 (µmol/L), with an effect size of 0.471. A power of 80% and a type 1 error of 0.05 were then guaranteed by a sample of 48 patients (16 for each group). This number had to be corrected for the post-hoc comparisons. Our main interest in the difference between groups 2 and 3, with an expected effect size of 1, a power of 80% with a type 1 error of 0.017 (Bonferroni correction, 0.05/3) was guaranteed by a sample of 19 subjects for each group.
Endpoints. The primary endpoint was the characterization of the parameters able to modulate hypoxic pulmonary vasoconstriction related to a negative outcome (respiratory failure and/or death) in patients admitted for COVID-19 disease. This assessment was done with a cluster analysis.
The explorative secondary endpoint was the investigation of all the factors influencing COVID-19 outcome evaluated as composite outcome of IMV and/or in-hospital death in COVID-19 disease.
Study design and ethical approval. Patients were followed up on a daily basis until discharge or inhospital death for clinical outcomes such as in-hospital mortality, need for invasive mechanical ventilation (IMV), need for transfer to Intensive Care Unit (ICU) or the presence of respiratory failure without need for IMV. Results from routine clinical laboratory assays were anonymously collected. Length of hospital stay for COVID-9 was calculated as the number of days between admission and discharge for surviving patients. Pul- Biochemical analyses. Analysis of L-arginine, ADMA, and SDMA in serum was assessed by ultra-performance liquid chromatography-tandem mass spectrometry (UPLC-MS/MS) using validated protocols established in our laboratory 21 . Briefly, 25 μl of serum were diluted in methanol to which stable isotope-labelled internal standards had been added. Subsequently, the compounds were converted into their butyl ester derivatives and quantified by UPLC-MS/MS (Xevo TQ-S cronos, Waters GmbH, Eschborn, Germany). Compounds were separated on an Aquity UPLC BEH C18 column (2.1 × 50 mm, 1.7 µm, Waters GmbH). The coefficient of variation for the quality control samples was below 15% for all compounds. ADMA, SDMA, and L-arginine were also analyzed in 38 blood donors who served as healthy controls. MR-proADM plasma concentrations were measured in an automated Kryptor analyzer, using the TRACE technology (Kryptor; BRAHMS, Hennigsdorf, Germany). IL-6 serum concentrations were measured by microfluidic ultrasensitive ELISA using the Protein simple plex technology on ELLA instrument (R&D systems, Biotechne, USA). All other laboratory biomarkers were evaluated using routine certified diagnostic methods.
The manuscript was drafted according to the Standards for the Reporting of Diagnostic accuracy studies STARD criteria 22 . Statistical analyses. Variables are described as mean ± standard deviation for normally distributed variables, and as median and interquartile range or proportion for non-normally distributed variables. Accordingly, comparisons between patients were performed with a two-tailed unpaired t-test, Mann-Whitney test or chisquare test with continuity correction. Moreover, the AUCs of the SOFA and ADMA were calculated.
The relationship between covariates and the probability of IMV was assessed with logistic regression. For multivariable analysis, only covariates with p < 0.10 at univariable logistic regression were considered; given the small sample size and the number of covariates to be considered, a lasso penalised method was used for the multivariable logistic regression. Risk stratification was explored using an unsupervised machine learning approach 23 from variables strictly related to the severity of disease (COVID-19 WHO Severity Classification, SOFA score, PaO 2 /FiO 2 ), a random forest 24 was built to discriminate between data structure and noise; the similarity between each couple of patients was computed as the percentage of trees in the forest that identically classify the couple; by a multidimensional scaling procedure, the similarity measure was transformed in a Euclidean distance and the patients represented as points in a plane; finally, by a Bayesian model-based approach 25 , distinct clusters (phenotypes) of patients were identified.
The machine learning identified phenotypes were characterized by one-way ANOVA analysis; Fisher test was used for normally distributed variables, Kruskal-Wallis test for non-normally distributed continuous variables, and Chi-square test for nominal variables; post-hoc comparisons were performed with Bonferroni correction. The relationship between ADMA, disease severity, and time from the onset of symptoms (≤ 10 days or > 10 days) was studied with a two-way ANOVA analysis. The R statistical software [R Core Team 2021] was used for all analyses, and a p value less than 0.05 was considered statistically significant. Negative outcome was assessed by the composite endpoint of IMV and/or death, due to low mortality rate.
Ethics approval and consent to participate. Informed consent was obtained from all subjects and/or their legal guardian(s).
Baseline characteristics of the cohort of COVID-19 patients. Baseline characteristics of the 90 patients are summarized in Table 1.
During hospitalization, 38/90 patients did not develop respiratory failure, 22/90 patients developed respiratory failure but did not require IMV, while 30/90 patients required IMV. No significant differences in age, gender, and comorbidity were observed between these three patient groups. In-hospital mortality was 11% (10 out of 90 cases): all deceased patients had required IMV.
The AUCs of the SOFA score and serum ADMA levels compared to a negative outcome reveal a significantly higher AUC of ADMA with the respect to the SOFA one (see Fig. 1). www.nature.com/scientificreports/ As shown in Table 1, upon post-hoc pairwise analysis, factors exhibiting a significant difference between all the 3 groups of patients were: SOFA score, COVID-19 WHO severity Classification, length of hospitalization (days), and need for ICU admission.
Patients who never developed respiratory failure during hospitalization showed a significantly higher PaO 2 / FiO 2 ratio and significantly lower serum LDH activity as opposed to the other two groups.
Patients who required IMV during hospitalization showed significant differences compared to the other two groups in the following items: proportion of deceased patients (all belonging to the group of intubated patients), WBC, CRP, and L-arginine/ADMA ratio. www.nature.com/scientificreports/ CT pulmonary vasodilation was significantly higher in patients with respiratory failure as opposed to patients not requiring IMV (82%), when compared to patients without respiratory failure (47%).
ADMA was significantly higher in patients with respiratory failure requiring IMV (0.79 ± 0.21) when compared to patients with respiratory failure without need of IMV (0.52 ± 0.14).
Primary endpoint: the characterization of the parameters able to modulate hypoxic pulmonary vasoconstriction related to a negative outcome with a cluster analysis. The possible correlation between ADMA, SDMA, L-arginine values along with vasodilation at CT scan analysis (measured within 72 h from admission) and negative clinical outcome was studied using an unsupervised Random forest-based similarity map. This made it possible to identify 3 clusters of patients, which are explained below.
Providing COVID-19 WHO Severity Classification, SOFA score, and PaO 2 /FiO 2 ratio as severity-related variables, we tried to understand if there were clusters of patients with similar features in our cohort. A model-based clustering approach was used to identify distinct phenotypes (Fig. 2, Panel A): Cluster 1 (n = 30; grey squares) was characterized by low disease severity, Cluster 2 (n = 43; yellow triangles) was characterized by moderate disease severity, Cluster 3 (n = 17; blue circles) was characterized by high disease severity.
The characterization of the 3 Clusters is specifically summarized in Table 2.
According to severity cluster analysis, post-hoc pairwise analysis identified the following factors exhibiting significant differences between the 3 groups of patients, especially when cluster 3 was compared to clusters 1 and 2 (see Table 2): need for ICU admission, patients who required IMV, LDH levels and ADMA levels at admission. In particular, ADMA levels were significantly higher in patients from Cluster 3 (0.83 ± 0.24 µmol/L) compared to both Cluster 1 (0.54 ± 0.12 µmol/L) and 2 (0.6 ± 0.17 µmol/L). Of note, cluster 2 was also characterized by a greater number of patients who did not require intubation as opposed to patients within Cluster 3, whilst still presenting with high proportions of respiratory failure (p < 0.001) as well as a high level of vasodilation (72%, according to CT scan (72% in cluster 2 vs 35% in cluster 3, p = 0.014).
The relationship between ADMA levels, disease severity (identified by clustering), and time from symptoms onset was assessed. Results are displayed in Fig. 2-Panel B. For patients admitted to the hospital within 10 days from onset of symptoms, ADMA levels were significantly higher in Cluster 3 as opposed to Clusters 1 and 2. The latter observation no longer persisted when ADMA levels were assessed in patients hospitalized after 10 days from symptoms onset. Indeed, 4 patients died within the Cluster 2 group with symptom onset beyond 10 days from admission (21 patients overall). ADMA levels in these 4 patients were elevated, ranging from 0.71 to 1 µmol/L, all of which were above the median serum ADMA level of cluster 2 (0.6 µmol/L).  Supplementary Table 1). As shown in Supplementary Table 1, the deceased patients had a higher SOFA score and higher class of COVID-19 WHO Severity Classification as opposed to survivors; they required ICU admission and IMV. All deceased patients had lower PaO 2 /FIO 2 ratio, higher LDH upon admission, higher ADMA levels, and lower L-arginine/ADMA ratio.
A sub-analysis of the 30 patients requiring IMV was carried out, comparing the survivors with the deceased (see Supplementary Table 2): no statistically significant differences were found.
Since all deceased patients required ICU admission and IMV, we performed a multivariable penalisted logistic regression analysis to investigate the association with the composite outcome of IMV and/or death ( Table 3).
The multivariable penalisted logistic regression analysis showed that two factors were significantly associated with negative outcome: SOFA score (OR 1.495), White blood cell (OR 1.118), C-reactive protein (OR 1.012) and serum ADMA levels (OR 4.652).

Discussion
The present study provides evidence that patients hospitalized with COVID 19 disease cluster into three groups with distinct symptoms, disease severity, and prognosis. Specifically, patients displaying signs of respiratory failure may or may not need mechanical ventilation; the latter group is characterized by pulmonary vasodilation and low ADMA concentration. Instead, ADMA levels are significantly higher in patients requiring mechanical ventilation.
The COVID-19 infection has been described to have three consecutive phases: early infection, pulmonary involvement, and severe hyperinflammation 26 . During the early phase of infection, the virus infiltrates the pulmonary parenchyma and starts replicating, causing an inflammatory response involving local vasodilation and increased endothelial permeability 27 . Early stages of COVID-19 may be associated with silent hypoxia and poor oxygenation. These paradoxical findings may be explained by impaired hypoxic pulmonary vasoconstriction in the infected lung and hypoxemia may best be explained by the loss of regulation of perfusion, i.e. by loss of hypoxic vasoconstriction 28,29 .
Our data suggest that this phenomenon might be related to low levels of ADMA; it may explain the increased pulmonary shunts and mismatch of ventilation/perfusion, observed in cluster 2 of this study. www.nature.com/scientificreports/ ADMA and SDMA production occur by enzymatic degradation of protein during the response to hypoxia; they are involved in modulating cardiovascular function and the innate immune response by decreasing cellular L-arginine uptake 17,30 . Organ failure, cardiovascular thromboembolism, and kidney and liver failure may be causes and consequences of the imbalance of these biomarkers in COVID-19 disease.
The multivariable penalized logistic regression analysis, carried out to investigate the association with the composite outcome of IMV and/or in-hospital death, confirms that ADMA levels at admission have a prognostic role. Hannemann and co-workers previously still demonstrated that high levels of both, ADMA and SDMA, are linked to COVID-19 disease severity 20 inpatients from the first wave. This present study comprised also patients during the second wave and the number of patients is higher than in the previous publication.
Previous studies of ADMA in sepsis and during infections in general have given conflicting results. In vitro 31 and in vivo 32,33 studies (in septic patients) suggest that inflammation and especially IL-6 have the ability to induce the activity of the DDAH enzyme with the result of lowering ADMA levels. By contrast, Xiao and colleagues 34 demonstrated downregulation of DDAH2 and upregulation of protein arginine methyltransferase-1 (PRMT1; the enzyme that catalyzes the biosynthesis of ADMA) by LPS in rats. Böger 35 proposed that downregulation or inactivation of DDAH may be counter-regulatory mechanisms in sepsis, limiting NO production by inducible NO synthase through elevated ADMA. In line with this, Zoccali and co-workers 33 showed that ADMA plasma levels increase during the resolution of infection/inflammation. Iapichino et al. 32 showed that in patients admitted to ICU for sepsis, the levels of ADMA and the ADMA/SDMA ratio were inversely proportional to the levels of inflammatory markers (CRP and IL-6). Conversely, Ghashut and colleagues reported elevated ADMA concentration at admission to be associated with mortality in a large study of 104 patients with critical illness 36 . Winkler et al. recently observed that increased plasma concentrations of SDMA and ADMA are associated with sepsis severity 37 . A meta-analysis of six prospective clinical studies including 705 patients with critical illness confirmed the association between high ADMA plasma concentration at admission and mortality (pooled odds ratio 3.13 (95% CI 1.78-5.51) 38 .
The lung is a major source of ADMA; several studies revealed that the complete pathway of ADMA biosynthesis by PRMTs and degradation by DDAHs is present in the lungs 16,39 . In an animal model of prolonged critical illness, DDAH activity was the main regulator of tissue and plasma ADMA concentrations 40 . This is in line with reports from DDAH1 knockout mice that are characterized by increased ADMA, impaired vascular NO release, endothelial dysfunction, and systemic and pulmonary arterial hypertension [41][42][43] . In the lungs, experimental overexpression of DDAH2 attenuated LPS-induced vascular leak in acute lung injury 44 . Hannemann and Böger recently provided evidence suggesting that regulation of DDAH1 (down) and DDAH2 (up) in the lungs during hypoxia is involved in modulating hypoxic pulmonary vasoconstriction 45 . In an elegant study using isoformspecific siRNAs, Wang and co-workers showed that expressional repression of DDAH1 increased circulating ADMA concentration but did not have a major impact on vascular function, whilst expressional repression of DDAH2 caused vascular dysfunction whilst leaving plasma ADMA concentration largely unchanged 46 . Therefore, Table 3. Multivariable penalised logistic regression analysis performed to investigate the association with the composite outcome of IMV and/or death in COVID-19 patients. *OR calculated for variation of 0.1 µmol/L in ADMA value; °Multivariable penalised logistic regression analysis furnishes the OR of the selected covariates, without confidence intervals. www.nature.com/scientificreports/ whilst the direct role of DDAH2 in enzymatic cleavage of ADMA has remained controversial, both isoforms seem to play distinct roles in regulating NO-mediated vascular function. These observations may help to explain the relationship between ADMA and pulmonary vasodilation observed in COVID-19 patients with respiratory impairment in the present study. SARS-CoV-2, by binding to ACE2 receptor for host cell entry 47,48 , has a strong organotropy towards the lung circulation, causing endothelitis and pulmonary vascular dysregulation 5,15 . Consequences arising from this for the integrity of the pulmonary L-arginine/ADMA/NO pathway may result in differences in pulmonary vascular response to SARS-CoV-2 infection, and determine the clinical course and outcome of patients.
The three clusters identified by means of the unsupervised machine learning approach clearly reflect the clinically identified subgroups of COVID-19 patients. Cluster 1 was mainly composed of patients who did not develop respiratory failure; Clusters 2 and 3 comprised patients who developed respiratory failure; amongst these, patients in cluster 2 had no clinical need for IMV, low ADMA concentration at admission, and pulmonary vasodilation in CT (72% of patients in this cluster), whilst patients in cluster 3 included almost all patients requiring IMV, showed significantly less pulmonary vasodilation (35% of patients), had the most critical clinical course, and experienced the highest mortality rate.
A difference in vasodilation status was also observed among the clusters, notably, Cluster 2 showed greater vasodilation according to CT scan analysis as opposed to cluster 3.
An additional observation in this study was that ADMA levels upon hospitalization for COVID-19 disease may depend on the time elapsed from symptoms onset. For patients admitted to hospital within 10 days, ADMA levels were significantly higher in cluster 3 than in cluster 2. By contrast, no differences in ADMA levels were observed between clusters 2 and 3 at the time of hospitalization if more than 10 days had elapsed from symptoms onset. This may point to the temporal pattern of regulation of the L-arginine/ADMA/NO pathway during SARS-CoV-2 infection. However, cluster 2 also comprised four patients displaying exceptionally high levels of ADMA who had been admitted later than 10 days after symptom onset, and all of whom died during hospital treatment. Future studies may reveal whether and how the temporal associations of ADMA plasma concentration, pulmonary vasodilation, and the patient's ability to overcome the pathophysiological consequences of SARS-CoV-2 infection may determine clinical course and outcome.
With the introduction of a measure of pulmonary vasodilation, this study comes one step closer to the suggested pathomechanism of ADMA effects in COVID-19 pneumonia.
Our study has strengths and limitations. We included 90 consecutive patients with moderate to severe COVID-19 disease requiring in-hospital treatment, and we rigorously subjected patients to pulmonary CT scan, biomarker measurements by the analytical gold standard, LC-MS/MS, machine learning-based cluster analysis, and follow-up for in-hospital course and outcome. Nonetheless, this was a retrospective single-center study, and the limited number of patients curtailed our ability to perform more subgroup analyses. The comparatively small sample size of COVID-19 patients admitted to ICU may also be a cause that we did not find differences in SDMA plasma concentration between the three subgroups of COVID-19 patients, which is in discrepancy to a previous study by Hannemann et al. 20 that comprised a small group of patients with a broader range of disease severity. Furthermore, follow-up was limited by the duration of in-hospital treatment; we have no information on outcome after discharge. Finally, the molecular mechanisms underlying the regulation of ADMA in COVID-19 patients could not be revealed in this clinical study; animal or in vitro studies will be needed to shed light on this.
Attempts to improve endothelial dysfunction in COVID-19 patients have been made. In line with the available literature, our study supports observations that COVID-19 is featured by an endothelial dysfunction that may be causally related to dysregulation of the L-arginine/ADMA NO pathway. Therapeutic approaches such as inhaled NO administration in order to restore NO function may provide a valid strategy in order to treat COVID-19 disease; however, pilot studies have so far produced discrepant results [49][50][51] . Similarly, a pilot trial using sildenafil to improve pulmonary ventilation/perfusion mismatch generated unclear benefit 52 . In our study, low Larginine/ ADMA ratio was associated to worse outcome. Thus, intervention on endothelial dysfunction, e.g., by L-arginine supplementation, may reduce the need for respiratory support as well as the length of hospital stay. Indeed, in an interim analysis of a randomized prospective trial conducted by Fiorentino and co-workers, L-arginine supplementation was shown to reduce oxygen requirement of COVID-19 patients 53 .
In conclusion, our study that distinct subtypes of COVID-19 patients can be identified that have different clinical disease severity and outcome. The combination of established clinical information with ADMA as a prognostic biomarker allows early identification of these subgroups. Machine learning approaches can help to improve the clustering of patients and clinical decision-making.

Data availability
The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.