Author response:
The following is the authors’ response to the original reviews.
eLife Assessment
This study presents a valuable metagenomic analysis of the gut microbiome in sickle cell disease (SCD) patients, revealing associations between bacteriophage, host immunity, and SCD pathophysiology. While these data are interesting and helpful for hypothesis generation, they are deemed incomplete; additional experiments would be needed to test causality and to provide mechanistic insight. Despite these limitations, this work will be of broad interest to researchers studying SCD, immunology, phage biology, and the microbiome, adding to the small but growing literature suggesting a microbial component to SCD.
The authors would like to thank the reviewers for thorough and constructive comments on our manuscript. We have made major updates to the manuscript addressing the following points and suggestions from the three reviewers: (1) assessing HbAS/AA genotype influence on microbiome composition; (2) conducting the requested beta diversity analysis, (3) conducting the requested sensitivity analysis to assess the impact of disease severity and therapy on microbiome and virome features; (4) modifying our language to clearly state that our results do not indicate causality or mechanism of microbiome interactions with sickle cell disease pathophysiology; (5) improved discussion of the phage results and their strengths and limitations; (6) additional changes throughout for clarity and correction of errors. We have changed the title to “Bacterial and viral gut microbiome alterations characterize microbiome-immune-pathophysiology axes in Sickle Cell Disease.” These additions have greatly improved our work and presentation and we are grateful to the reviewers and our editors. We have indicated where specific changes were made in response to the public reviews below.
Public Reviews:
Reviewer #1 (Public review):
Summary:
In this manuscript, Flamholz and colleagues use metagenomic sequencing to profile the microbiome of individuals with sickle cell disease (SCD), the most common genetic blood disorder in the world. To build on previous studies that found dysbiosis in SCD, this manuscript aims to examine whether changes in either bacterial species or bacteriophages correlate with inflammatory hallmarks of the disease. The authors claim that sickle cell dysbiosis does not correlate with inflammatory hallmarks of the disease, but instead, aged neutrophil numbers and bacteriophages do. Appropriate control subjects and additional analyses are needed to support that conclusion.
Strengths:
The primary strength of this paper is the investigation into disease-associated changes in bacteriophages. This is an entirely novel idea in the sickle cell field, and based on the current results, may be an important, under-recognized disease hallmark. It is unclear, however, if phages are "the chicken or the egg" in terms of sickle cell inflammatory profiles; do these increases in phage number simply result from other disease processes, or are they in any way contributing to disease pathophysiology?
Weaknesses:
A primary weakness of the manuscript is the fact that the majority of individuals included in the control group maintain sickle cell trait (HbAS genotype). Although typically asymptomatic, it is unclear if this genotype is associated with microbial changes that would not be observed in a true control group (HbAA genotype). This is a significant limitation that may limit the ability to draw conclusions from the current data set.
Another key weakness is the lack of beta diversity assessment. Although decreased alpha diversity is observed in individuals with SCD, and specific bacterial taxa are differentially abundant following multivariate analyses, there is no overall comparison of bacterial community composition between individuals with SCD and controls. Prior to drawing conclusions about the relationship (or lack thereof) between the SCD microbiome and inflammatory markers, it is important to know if this study did indeed find disease-associated changes in microbiome composition.
It is unclear which individuals were used for aged neutrophil (AN) and molecular data assessments. For example, were children who were still receiving penicillin prophylaxis included in these specific assessments? Given the authors' previous work demonstrating that antibiotic treatment decreases AN pathology, it seems critical to limit all AN/molecular analyses to older subjects who are not on daily penicillin treatment (if possible).
A minor weakness is the continued use of "disease" vs. "healthy" indicators as primary microbiome metrics that are used for molecular correlations. The lack of metric specificity - and lack of discussion regarding which diseases were used to generate these indicators (how similar/different are they to sickle cell?) - could be said to make these metrics essentially meaningless.
We thank the reviewer for their helpful comments and suggestions. We want to first note that patients on prophylactic penicillin within six months of sample collection were excluded from the study due to the known impact of antibiotics on gut microbiomes, this has been clarified in the main text. We have now included an analysis evaluating the influence of control genoype (HbAA/HbAS) on our microbiome and virome results. To evaluate whether control genotype influenced major microbiome and virome features, analyses were restricted to control participants only. Controls were stratified by genotype as HbAA or HbAS. Four significant microbiome and virome features were tested: F:B ratio, Shannon diversity, provirus fraction, and virus count. HbAA and HbAS controls were compared using two-sided Mann-Whitney U tests. Benjamini-Hochberg FDR correction was applied across the four tested features. HbAS and HbAA controls did not differ significantly for F:B ratio, Shannon diversity, provirus fraction, or virus count. The inclusion of HbAA/AS strengthens our results with respect to the observation that sickle cell disease patient microbiomes remain significantly different from sickle trait (HbAS) controls. These results are reported in the new Supplemental Table 6.
We have now included a beta diversity analysis using MetaPhlAn species profiles. Beta diversity analyses were performed in Python using pandas and NumPy for data processing, scikit-bio for distance calculations and PERMANOVA, scikit-learn for ordination-related computations, statsmodels for multiple-testing correction where applicable, and matplotlib for visualization.
For the primary disease/control comparison, samples were grouped as control or SCD. For the genotype control sensitivity analysis, samples were restricted to HbAA and HbAS individuals as described above. Species detected in at least 10% of included samples were retained for beta diversity analysis. To account for the compositional structure of metagenomic relative abundance data, species profiles were transformed using a centered log-ratio transformation after addition of a small pseudocount to accommodate zero values. Aitchison distances were calculated from the CLR-transformed species profiles. Statistical significance of group separation was assessed by PERMANOVA using 999 permutations. For the control versus SCD comparison, PERMANOVA was performed between the two disease-status groups. For the HbAA versus HbAS control comparison, PERMANOVA was performed among controls only.
In the SCD cohort, beta diversity differed significantly between controls and SCD participants by Aitchison distance after CLR transformation (R2 = 0.030, p = 0.001). In contrast, HbAA and HbAS controls did not differ significantly in beta diversity (R2 = 0.024, p = 0.282), supporting the conclusion that the observed SCD/control separation was not driven by control genotype composition. These methods and results are now reported in the manuscript.
The manuscript describing the microbiome health and disease indicators was submitted to eLife jointly with this manuscript as a package; eLife declined to review the indicator manuscript. Briefly, this study conducted a cross-disease meta-analysis of 38 studies comprising 8,204 samples and identified 100 bacterial taxa or “indicators” that are weakly but consistently associated with health or disease across diverse conditions, including, but not limited to, inflammatory bowel disease, colorectal cancer, type 2 diabetes. The indicator taxa were validated in an independent cohort of Graves’ disease patients. We currently cite an older version of this work posted as a preprint. The manuscript is currently under review at another journal and we will update this manuscript with the updated citation when it is available.
We have addressed other recommendations from this reviewer as follows. We cite and discuss previous SCD rodent model work observing decreased butyrate in disease, and we have updated our results and discussion sections regarding associations between the microbiome and virome and clinical and molecular features.
Reviewer #2 (Public review):
Summary:
The study analyzes stool metagenomes from 98 SCD patients and 46 controls, with SCD and control groups matched on age, race, sex, and ethnicity. The authors report lower Shannon diversity, lower Firmicutes/Bacteroidetes ratio, loss of health-associated taxa, increased disease-associated indicators, altered butyrate/fatty-acid metabolism pathways, and enrichment of provirus/prophage fractions in SCD. They further correlate aged-like neutrophils and prophage fractions with inflammatory cytokines. The strength is that this is not just another 16S comparison. The use of whole-community metagenomics, immune profiling, neutrophil assays, and clinical metadata makes the study more biologically interesting than prior small SCD microbiome papers. The main weakness is that the causal and mechanistic interpretation is too strong. The data support an association between SCD status and microbiome/virome features, but they do not yet establish a clear "axis of pathophysiology." The provirus findings are intriguing, but require stronger statistical control, better validation, and more cautious interpretation.
Strengths:
The major strengths of the study include the clinically relevant disease setting, the use of whole-community sequencing, the integration of microbial, immune-cell, cytokine, and clinical measurements, and the novel attention to bacterial virus-related features. A particularly interesting aspect of the work is the analysis of virus-like elements integrated into bacterial genomes. The authors report that these elements are enriched in the gut microbial communities of patients with sickle cell disease and are associated with several inflammatory signals in blood. This observation is potentially important because it suggests that the microbial contribution to inflammation in sickle cell disease may involve not only bacteria but also bacterial virus-related genetic elements.
Weaknesses:
The evidence for this proposed immune-related mechanism is incomplete. The study is cross-sectional and largely based on associations, so it cannot determine whether these virus-like elements drive immune activation, reflect immune activation, or are linked indirectly through disease severity, treatment history, or other clinical factors. The main limitations are the single-center design, modest sample size for some immune measurements, limited ability to control for treatment and disease heterogeneity, and the need for clearer multiple-testing correction in the correlation analyses. In particular, stronger adjustment for available clinical factors such as hydroxyurea use, transfusion history, pain admissions, genotype, and other markers of disease burden would help readers judge how specific the microbial and viral findings are to sickle cell disease itself. (REVISION POINT 3)
Overall, the authors largely achieve their descriptive aim of identifying gut microbial differences associated with sickle cell disease. The evidence is solid for the presence of broad microbial community differences, but incomplete for the stronger conclusion that virus-like elements form a pathophysiological immune axis. The work will likely be useful to researchers studying the microbiome, inflammation, and sickle cell disease, especially as a hypothesis-generating dataset. Its impact would be strengthened by more cautious interpretation, stronger control of clinical confounders, clearer statistical correction, and future longitudinal or experimental studies to test causality.
We thank the reviewer for their helpful comments and suggestions. We want to first note that patients on prophylactic penicillin within six months of sample collection were excluded from the study due to the known impact of antibiotics on gut microbiomes, this has been clarified in the main text. We have tempered our interpretation of our results, making clear that we are not arguing that either prophages or bacteria are causal or mechanistically associated with SCD biology and pathology. We have strengthened our control of clinical confounders, and added clearer statistical correction, as described below, with corresponding updates to the manuscript. We look forward to conducting future studies to test causality and understand mechanism.
We have now done sensitivity analysis within SCD patients to determine whether our microbiome and virome results associate with treatment and clinical severity. To evaluate whether microbiome and virome features were explained by clinical or demographic heterogeneity within the SCD cohort, we restricted analyses to SCD participants. We fit a separate multivariable regression model for each feature. Each model included age, sex, hydroxyurea use, transfusions in the past year, and acute care utilization in the past year as predictors.
feature_z ~ age_z + sex_F + HU + log1p(TxPastYr)_z + log1p(AcuteCarePastYr)_z
Non-negative abundance, ratio, pathway, viral, and count-like variables were log-transformed to reduce skew, using feature-specific pseudocounts for zero-containing microbiome/virome variables and ln (1 + x) transformation for count covariates. Diversity and indicator scores were not log-transformed. Continuous variables were then standardized to Z-scores before modeling. Models were fit using ordinary least squares with HC3 robust standard errors. FDR correction was applied separately for each model term across the tested microbiome and virome features. No microbiome or virome feature showed an FDR-significant association with hydroxyurea use, transfusions in the past year, or acute care utilization in the past year. These results are reported in the manuscript and in the new Supplemental Table 7.
We have reported multiple-testing correction results for all associations between microbiome and virome features and clinical and molecular features and updated manuscript figures accordingly.
To evaluate relationships between microbiome/virome features and clinical or immune markers within the SCD cohort, we performed Spearman correlation analyses. Microbiome and virome features were organized into four prespecified feature groups: community metrics, taxa, functional pathways, and viral features. Clinical and immune markers were grouped into marker sets for visualization and multiple-testing correction, including hematologic clinical markers, hemolysis markers, creatinine, acute care burden, and cytokines/chemokines. Spearman correlation coefficients were calculated for each feature–marker pair. Benjamini-Hochberg FDR correction was applied within each prespecified feature group by marker group block. Nominal associations were defined as p < 0.05, FDR-significant associations as q < 0.05, and trends as q < 0.10. In the heatmap figure, boxes now indicate nominal p < 0.05, asterisks indicate q < 0.05, and daggers indicate q < 0.10.
We have addressed other recommendations from this reviewer as follows. We have modified our language describing prophage/immune associations. We have revised the Methods to clarify how abundance data were processed before MaAsLin2 modeling. Taxonomic profiles were analyzed using MaAsLin2 with total-sum scaling normalization and log transformation, while pathway profiles were analyzed without additional normalization because the input pathway table had already been normalized prior to MaAsLin2 analysis; MaAsLin2 log transformation was then applied. We agree that relative abundance metagenomic data are compositional, and we have revised the text to clarify that these analyses identify covariate-adjusted associations with transformed relative abundance rather than absolute abundance. We also now note this as a limitation of the study. We also note the limitations of F:B as a metric. We have added text describing the need for further analysis of the prophages to understand their patterns of host range and transmission and their associations with features such as shared geography, health care exposure, and diet. Finally, we have updated Figure 5 to reflect our updated analysis with significance indicated.
Reviewer #3 (Public review):
Summary:
In this manuscript, Flamholz et al. sought to determine whether consistent and significant interactions exist between the gut microbiome and disease pathology in sickle cell disease (SCD). By sequencing and analysing metagenomes from faecal samples collected from 98 SCD patients and 46 control subjects, they identified community-level shifts in both the bacterial and proviral gut microbiome of SCD patients. They further reported correlations between the proviral microbiome and multiple blood cytokines, whereas similar associations were not observed for the bacterial microbiome. Based on these findings, the authors propose the existence of a viral-immune axis in SCD pathophysiology and targetable functional alterations in the gut microbiome.
Strengths:
This work includes the largest SCD cohort analysed to date, enabling analysis with relatively strong statistical power. In addition to profiling the bacterial microbiome, the study also examines the gut proviral microbiome, thereby providing a more comprehensive investigation of the topic. The newly generated metagenomic dataset will also be valuable for further meta-analysis by the wider community. Overall, the authors have largely achieved their aims.
Weaknesses:
However, this study represents a single-centre cross-sectional investigation, and most findings remain correlative in nature. In particular, the claim that the study identifies targetable functional alterations in the gut microbiome for disease treatment may be somewhat overstated. Although the reported functional module changes in SCD patients are intriguing, additional mechanistic and/or longitudinal evidence would be required before these features can realistically be considered targetable.
We thank the reviewer for their helpful comments and suggestions. We have now noted in the text that additional mechanistic and longitudinal studies are required before we can target the microbiome and virome in SCD and clarified that this is a single-centre, cross-sectional. We have further made modifications to the manuscript to clarify cohort features (specifically, age and race were matched, other baseline characteristics were balanced), to properly describe the Shannon diversity metric, and to fix several errors that this reviewer caught.
