Background: Long COVID (LC) affects approximately 10% of SARS-CoV-2 positive patients, causing significant workforce loss and an annual economic cost of $743 billion in the United States. Previous studies have suggested an association between long COVID and alterations in the gut microbiome. This study aimed to ascertain whether the baseline microbiome at the time of diagnosis can predict the occurrence of LC and/or its sub phenotypes.
Methods: In a prospective cohort study, we examined 400 SARS-CoV-2 positive and 419 negative individuals who provided stool samples at the initial diagnosis. Long COVID was defined as symptoms persisting for four weeks or more after the initial diagnosis. Stool samples were sequenced using Illumina NovaSeq, (~8 million reads/sample). Taxonomy was assigned using Kraken2, statistical analysis was done using standard tools in R, and predictive modeling was done using XGBoost in Python.
Results: Among the 400 SARS-CoV-2 positive participants, 20 with inadequate follow-up were excluded, while 82 developed LC. Baseline characteristics of SARS-CoV-2 positive (LC and no LC) and negative participants showed variations due to clinical risk factors associated with LC and inherent differences between the groups (Figure 1). We adjusted subsequent analyses to account for these factors. Both groups of SARS-CoV-2 positive patients, harbored a distinct gut microbiome compared to SARS-CoV-2 negative participants (Bray-Curtis-based beta diversity, PERMANOVA, p=0.003, R2=0.004; Figure 2A). However, we observed distinct microbial taxas between LC and no LC groups compared to SARS-CoV-2 negative controls (Figure 2B). The gut microbiome at the time of diagnosis was a predictor of LC (XGboost, AUC=0.73, 95% CI =0.62, 0.81; Figure 2C). When evaluating individual symptoms, LC patients with persistent myalgia (Fischer's exact test, OR=3.57, 95% CI=0.76-15.57, p=0.057) and diarrhea (Fischer's exact test, OR=4.04, 95% CI=0.71-20.88, p=0.062) exhibit the greatest divergence in microbial diversity from no LC patients (Bray-Curtis dissimilarity-based irregularity, >90^th percentile). Reference group analysis similarly showed that the microbiome from patients with diarrhea, myalgia, and anosmia/dysgeusia diverged most from the no LC group (Figure 2D, E). On further analysis using K-means and hierarchical clustering of LC symptoms, we identified five LC subphenotypes and these were associated with distinct microbial signatures (Figure 2F, G). The gastrointestinal and sensory cluster had the most prominent changes with 24 species from Lachnospiraceae family positively correlating with symptoms.
Conclusion: Our study found baseline microbiome to be a predictor of LC, with distinct microbial taxa underlying subphenotypes of LC. These findings offer promising targets for the development of therapeutic interventions for LC.

<b>Abbreviations: </b>BMI, body mass index; CCI, Charlson’s comorbidity index; IQR, interquartile range; SD, standard deviation.<br /> In the SARS-CoV-2 negative group, BMI for five patients, antibiotic use, and CCI scores for three patients were missing.<br /> *, CCI scores and time to kit receipt were analyzed using the Kruskal Wallis test, with post-hoc assessment via the Dunn test with Bonferroni correction. Age and BMI were evaluated with unbalanced one-way ANOVA, and age differences were further analyzed using Tukey’s HSD. Sex and 30-day antibiotic use were assessed with Fischer’s exact test, followed by pairwise Fischer’s exact test with Holm correction for post-hoc analysis.

Abbreviations: BMI, body mass index; CCI, Charlson’s comorbidity index; IQR, interquartile range; SD, standard deviation.
In the SARS-CoV-2 negative group, BMI for five patients, antibiotic use, and CCI scores for three patients were missing.
*, CCI scores and time to kit receipt were analyzed using the Kruskal Wallis test, with post-hoc assessment via the Dunn test with Bonferroni correction. Age and BMI were evaluated with unbalanced one-way ANOVA, and age differences were further analyzed using Tukey’s HSD. Sex and 30-day antibiotic use were assessed with Fischer’s exact test, followed by pairwise Fischer’s exact test with Holm correction for post-hoc analysis.

<b>Figure 2. Baseline gut microbiome predicts LC and data-driven clustering of LC symptoms reveals distinct disease subphenotypes. A)</b> Principal coordinate analysis (PCoA) of Bray Curtis (BC) dissimilarities. PERMANOVA-based statistics -adjusted for clinical factors- are shown. <b>B)</b> Genus-level associations (adjusted) with LC and no LC groups compared to negatives. Stars indicate FDR-adjusted <i>p</i>-value <0.2. <b>C)</b> XGBoost model using normalized genus-level abundances.<b> D)</b> BC dissimilarity-based irregularities were calculated by extracting pairwise dissimilarities to no LC group. <b>E) </b>Each dot represents the BC dissimilarity of the patients with symptoms on x-axis to reference group’s (no LC) mean. <b>F)</b> Hierarchical and k-means clustering of LC symptoms revealed five distinct symptom clusters. <b>G)</b> Nominal genus and species-level associations (FDR-adjusted<i> p</i>-value <0.1) between no LC and LC (adjusted for sex), and symptom clusters (adjusted for each other).

Figure 2. Baseline gut microbiome predicts LC and data-driven clustering of LC symptoms reveals distinct disease subphenotypes. A) Principal coordinate analysis (PCoA) of Bray Curtis (BC) dissimilarities. PERMANOVA-based statistics -adjusted for clinical factors- are shown. B) Genus-level associations (adjusted) with LC and no LC groups compared to negatives. Stars indicate FDR-adjusted p-value <0.2. C) XGBoost model using normalized genus-level abundances. D) BC dissimilarity-based irregularities were calculated by extracting pairwise dissimilarities to no LC group. E) Each dot represents the BC dissimilarity of the patients with symptoms on x-axis to reference group’s (no LC) mean. F) Hierarchical and k-means clustering of LC symptoms revealed five distinct symptom clusters. G) Nominal genus and species-level associations (FDR-adjusted p-value <0.1) between no LC and LC (adjusted for sex), and symptom clusters (adjusted for each other).

Jun Chen

Appearances