The gut microbiota as an early predictor of COVID-19 severity

Abstract

Several studies reported alterations of the human gut microbiota (GM) during COVID-19. To evaluate the potential role of the GM as an early predictor of COVID-19 at disease onset, we analyzed gut microbial samples of 315 COVID-19 patients that differed in disease severity. We observed significant variations in microbial diversity and composition associated with increasing disease severity, as the reduction of short-chain fatty acid producers such as Faecalibacterium and Ruminococcus, and the growth of pathobionts as Anaerococcus and Campylobacter. Notably, we developed a multi-class machine-learning classifier, specifically a convolutional neural network, which achieved an 81.5% accuracy rate in predicting COVID-19 severity based on GM composition at disease onset. This achievement highlights its potential as a valuable early biomarker during the first week of infection. These findings offer promising insights into the intricate relationship between GM and COVID-19, providing a potential tool for optimizing patient triage and streamlining healthcare during the pandemic.IMPORTANCEEfficient patient triage for COVID-19 is vital to manage healthcare resources effectively. This study underscores the potential of gut microbiota (GM) composition as an early biomarker for COVID-19 severity. By analyzing GM samples from 315 patients, significant correlations between microbial diversity and disease severity were observed. Notably, a convolutional neural network classifier was developed, achieving an 81.5% accuracy in predicting disease severity based on GM composition at disease onset. These findings suggest that GM profiling could enhance early triage processes, offering a novel approach to optimizing patient management during the pandemic.

Keywords: COVID-19 severity; gut microbiota; machine learning.

Fig 1

Sampling distribution of recruited COVID-19 patients with different symptom severity and healthy uninfected subjects. In the context of the general symptom onset of COVID-19 (A), fecal samples were collected between the 5th and 10th day after symptom onset. Healthy uninfected controls were household members of individuals who had tested negative for SARS-CoV-2 and showed no symptoms. To maintain correspondence between the positive individuals and controls, uninfected subjects were sampled concurrently with the positive cases. (B) Samples from moderate and severe patients were collected at hospital admission, and the reported time since symptom onset was obtained from medical history. Samples from mild subjects were collected at a similar time (~7 days) from symptom onset to approximate a similar disease progression in inpatients. In addition to the samples presented in the plot, 42 healthy uninfected subjects were considered as well to complete the coverage of the severity classes. Such healthy uninfected subjects were household members of the mild ones and were sampled concomitantly. Therefore, the healthy uninfected subjects were tested SARS-CoV-2 negative via nose swab PCR in the day of sample collection and did not report COVID-19 symptoms. The numerosity of samples from the two recruiting centers (Groningen—the COVID-HOME cohort [17] of the Department of Medical Microbiology and Infection Prevention, University Medical Center Groningen, Groningen, The Netherlands, and Verona—the Infectious Diseases Unit, Department of Diagnostics and Public Health, University of Verona, Verona, Italy) included in each severity group is shown. ICU, intensive care unit. The figure was partly generated using Servier Medical Art, provided by Servier, licensed under a Creative Commons Attribution 3.0 unported license.

Fig 2

Severity-related differences in the gut microbiota of COVID-19 patients. (A) Alpha-diversity according to Faith’s phylogenetic diversity (Faith PD) in relation to COVID-19 severity classes (uninfected, mild, moderate, severe). (B) Principal coordinate analysis of weighted UniFrac distances shows segregation between non-hospitalized (uninfected and mild) and hospitalized (moderate and severe) subjects (pairwise PERMANOVA through the Adonis function –followed by Bonferroni correction, P = 0.0006). Arrows indicate the contribution of the fitted compositional variables (genus-level relative abundances) to the ordination plot (envfit, P ≤ 0.001). (C) Main genus-level compositional differences between severity classes, shown as boxplots of the overall distribution and C. See also Fig. S3. (D) Z-score scaled distribution of the main differences between severity classes. Only genera showing significant differences according to Bonferroni-corrected Kruskal-Wallis testing are plotted. Significance bars indicate P-values of Bonferroni-corrected pairwise Wilcoxon tests: *, P < 0.05; **, P < 0.01; ***, P < 0.001; ****, P < 0.0001.

Fig 3

Local network properties differentiate the gut microbial community in COVID-19 patients with different symptom severity. (A) Topological parameters obtained from differential networking analysis conducted for all pairwise combinations of severity classes (uninfected, mild, moderate, severe), focusing on the LCC and whole network levels. APL, average path length. Significance bars indicate P-values of pairwise Wilcoxon tests followed by false discovery rate (FDR) correction: °, FDR < 0.1 (considered as a trend). (B) Pairwise distances between networks computed according to graphlet correlation distance (GCD; left), edge difference (center), and log-moments distance (right).

Fig 4

Global networking detects COVID-19 severity-related gut microbiota modules. A single network was constructed from all samples and modules were detected using a spin glass model. Modules are highlighted with colored circles encompassing the enclosed nodes. For each severity class (uninfected, mild, moderate, severe), the network was plotted setting the node size proportional to the overabundance of that taxon in the corresponding group. Only nodes with an overabundance of at least 1.35 were plotted and labeled. Edge width was set proportional to the adjacency values obtained from NetCoMi netConstruct and colors indicate positive (blue) and negative (red) interactions.

Fig 5

Gut microbiota as a predictor of COVID-19 severity. (A) Summary of the performance of the several classifiers tested over the test set. The ConvNet with hyperparameter tuning achieved the best results in terms of accuracy, mean balanced accuracy, and mean F1-score. (B) Multi-class ROC sensitivity/false-positive rate curves for the ConvNet model. Continuous lines represent the prediction for the severity classes (uninfected, mild, moderate, severe), while dashed lines represent the macro- and micro-averages. NaB, naïve Bayes; Nnet, neural net in caret.