The intestinal microbiota predicts COVID-19 severity and fatality regardless of hospital feeding method

ABSTRACT SARS-CoV-2-positive patients exhibit gut and oral microbiome dysbiosis, which is associated with various aspects of COVID-19 disease (1–4). Here, we aim to identify gut and oral microbiome markers that predict COVID-19 severity in hospitalized patients, specifically severely ill patients compared to moderately ill ones. Moreover, we investigate whether hospital feeding (solid versus enteral), an important cofounder, influences the microbial composition of hospitalized COVID-19 patients. We used random forest classification machine learning models with interpretable secondary analyses. The gut, but not the oral microbiota, was a robust predictor of both COVID-19-related fatality and severity of hospitalized patients, with a higher predictive value than most clinical variables. In addition, perturbations of the gut microbiota due to enteral feeding did not associate with species that were predictive of COVID-19 severity. IMPORTANCE SARS-CoV-2 infection leads to wide-ranging, systemic symptoms with sometimes unpredictable morbidity and mortality. It is increasingly clear that the human microbiome plays an important role in how individuals respond to viral infections. Our study adds to important literature about the associations of gut microbiota and severe COVID-19 illness during the early phase of the pandemic before the availability of vaccines. Increased understanding of the interplay between microbiota and SARS-CoV-2 may lead to innovations in diagnostics, therapies, and clinical predictions.

T he microbiota of COVID-19 patients is characterized by a decreased abundance of prototypical anti-inflammatory bacterial species (1)(2)(3)(4)(5), which may influence overall immune responses or otherwise prime patients for certain risks. However, most microbiome studies in the COVID-19 literature compare patients with moderate or mild symptoms against healthy controls without COVID-19 (1,(6)(7)(8). The microbiomes of severely ill COVID-19 patients have not been well explored. Additionally, the contribu tions of enteral feeding (typical in the ICU setting) to COVID-19-related dysbiosis are largely unknown and represent a potential confounder (9,10). Here, we applied robust and validated machine learning methods with interpretable secondary analyses (11) to better define associations between the microbiota in severe and moderate COVID-19 patients during the early phases of the pandemic.
We enrolled 69 SARS-CoV-2 PCR-positive, hospitalized patients at the University of Massachusetts Medical Center and UMASS Memorial Hospital from 27 April to 10 June 2020. Of the 63 participants ultimately included in our analysis, 22 died of COVID-19 (Table 1). At the time of sample collection, patients requiring >4 L of oxygen were admitted to the intensive care unit due to the severity of their respiratory symptoms; we thus used this as the basis for differentiating moderate versus severe disease. No differences in age, body mass index, gender, race, smoking status, or antibiotic admin istration during hospitalization were observed between moderately ill and severely ill patients (Table 1). Unsurprisingly, more severely ill patients succumbed to disease and averaged ~6 more days of hospitalization than moderately ill patients. We did not find any significant differences between the groups when considering other symptoms and comorbidity data except that the prevalence of coronary artery disease and hypercholes terolemia was higher in severely ill patients (Table S1).
Members of the microbiome can predict the severity of inflammatory bowel diseases (12)(13)(14), the occurrence of colorectal cancer (15,16), and the immune responses during lung infections with influenza and tuberculosis (17,18). Thus, to explore markers of dysbiosis in severely ill COVID-19 patients, we collected stool and oral microbiome samples using commercially available kits and sequenced the 16S rRNA gene using the 341F and 806R universal primers to amplify the V3-V4 region. Sequences were analyzed using DADA2 (19), and species assignment was performed using the same pipeline as previously done by us (18). We used an R-based random forest classification (RFC) algorithm to first predict COVID-19 fatality. To ensure robust modeling, we applied leave-one-out cross-validation and used the Boruta algorithm to perform feature selection to identify relevant clinical variables (i.e., only contributing variables were included in the model). In line with clinical practice and outcomes, our models identi fied disease severity (defined by the >4 L oxygen requirement) as the main factor in predicting fatality (Fig. 1A). A combination of clinical and stool microbiota variables, specifically when classifying at the species level, was able to predict fatality better than clinical variables alone (F1-score of 73.4 versus 70.4; Fig. 1B). Modeling by feature ranking confirmed that gut bacteria outranked all clinical variables (except for severity) in predictive importance for COVID-19 fatality (Fig. 1C). We also applied the Stable and Interpretable RUIe Set (SIRUS), an interpretable rules algorithm (11), and obtained higher predictability of COVID-19 fatality by microbiome composition (63% probability of death) than clinical variables alone. The oral microbiota poorly predicted COVID-19 fatality in this cohort of hospitalized patients (Fig. S1).
Given the impact of severity on predicting COVID-19 fatality and indications of a role for stool microbiota, we next explored how gut microbiota may contribute to severity. We thus applied our RFC modeling pipeline to target outcomes of patients requiring <4

Observation mSystems
Enterococcus faecalis) along with hypercholesteremia are among the top 10 variables that predicted COVID-19 severity (Fig. 2B). Finally, SIRUS identified that an increased and decreased abundance of two Clostridia species (O. ruminantium and Hungatella hathewayi, respectively) are associated with a high probability (>90%) of severe disease, while patients with moderate disease exhibited a decreased abundance of Bacteroidia (Parabacteroides distasonis), Bacilli (Haloimpatiens massiliensis), and Clostridia (Aminipila butyrica) along with normal cholesterolemia (Fig. 2C). Results of fatality and severity models using a combination of the oral microbiome and clinical variables were less predictive than clinical variables alone (Fig. S1). identified only two species that were among the top bacteria predictors for severity.

Observation mSystems
As severely ill patients admitted to the ICU were also more likely to be fed enterally, we felt it important to explore the associations between severity, enteral feeding, and gut microbiota. Ninety percent of patients with moderate symptoms were fed a solid hospital diet, whereas 63% of patients with severe symptoms received enteral feeding at the time of sample collection. Interestingly, the top bacterial species predictors of COVID-19 severity differed greatly from those linked to differences in diet. Only two species were found to be predictors of both COVID-19 severity and enteral feeding, namely, O. ruminantium and B. fragilis (Fig. 2D). While O. ruminantium and B. fragilis ranked 3rd and 12th among predictors, respectively, these two species may simply predict dietary differences rather than disease severity; most species found to be depleted or overly abundant between enterally fed patients and patients on solid food were not among the top predictors for COVID-19 severity (Fig. 2B; Fig. S2). Notably, patients receiving enteral feedings were mostly depleted of Clostridia species. Together, our findings indicate that differences in stool microbiota associated with COVID-19 severity are independent of differences in hospital diet.
We observed distinct bacterial markers within the intestinal microbiota that predicted COVID-19 severity, the main clinical risk factor for fatality. The predictive power of the gut microbiota outranked clinical variables in our cohort, suggesting a pathophysiologic role for gut microbiota in COVID-19. The bacteria identified as predictors in this study are comparable to those found in previous studies comparing severe COVID-19 patients and healthy controls (1,(6)(7)(8)20). Further mechanistic studies are needed to understand whether the gut microbiota affects the pathophysiology of COVID-19. While the results of our study are limited by one-time sample collection and small sample size, our machine learning models were still able to provide meaningful associations, considering the effects of enteral feeding. Additionally, these data were gathered in early 2020 and, thus, offer insight into possible biologic links between the gut microbiota and COVID-19 before the availability of vaccines. Finally, we acknowledge that our results could also be explained by the fact that in systemically ischemic patients, there could be a loss of gut barrier integrity and, thus, microbiome-associated dysbiosis (21).

ACKNOWLEDGMENTS
We thank all faculty members of the Microbiology and Physiological Systems Depart ment and the EH&S at UMASS for insightful advice on sample containment. We are grateful to all the patients who participated in the study. We thank Katherine Fitzgerald for providing access to the BSL2 +laboratories to safely process all the samples. A