Immune–Microbiome Coordination Defines Interferon Setpoints in Healthy Humans

Summary

Human immune systems are highly variable, with most variation attributable to non-genetic sources. The gut microbiome crucially shapes the immune system; however, its relationship with the baseline immune states of healthy humans remains incompletely understood. Therefore, we performed multi-omic profiling of 110 healthy participants through the ImmunoMicrobiome study. A factor-based integrative approach identified coordinated variation, revealing that the interferon response was amongst the most variable immune features in healthy participants. Microbiome composition, pathways, and stool metabolites varied concomitantly with interferon response pathways. Longitudinal data spanning more than a year indicated significant stability of these parameters within individuals over time. Our study provides extensive data to examine the relationship between the immune states and microbiomes of healthy individuals at steady state, which paves the way for delineating inter-individual differences relevant for disease susceptibility and responses to therapy.

In Brief:

A comprehensive multi-omic analysis of healthy humans reveals two major axes of immunological variation characterized by interferon responses, one of which is coordinated with the microbiome and its metabolites and is stable over time within individuals.

Graphical Abstract:

Introduction

Inter-individual immune variation has broad implications for human health and disease. Population studies have begun identifying drivers of immune variability, revealing that environmental factors likely outweigh genetic contributions.^1–8 Deeper understanding could provide insights into disease susceptibility, severity, and variable responsiveness to immunomodulatory therapies,⁹ including vaccination,^10,11 cancer immunotherapy,¹² and immunosuppressive treatments for autoimmune diseases.¹³

The gut microbiome can influence immune responses. In cancer patients, microbiome features associate with clinical outcomes to checkpoint inhibitor immunotherapies.¹⁴ In autoimmunity, the microbiome and metabolites produced by commensal microbes impact treatment response.^15,16 In vaccination, the microbiota can act as an adjuvant, boosting immunogenicity.^17–20 These studies suggest that microbiota and molecules they produce may shape patients’ baseline immune states and influence immune responses to treatment. Animal models have revealed causal relationships between the microbiome and immune cell development and activity.^21,22 However, the relationship between microbiome features and variations in immune states among healthy humans at steady state remains incompletely understood.

One example is the regulation of baseline or tonic interferon (IFN) signaling, evidenced by the reduced expression of IFN-stimulated genes (ISGs) and impaired antiviral immune responses in microbiome-deficient mice.^19,23,24 Indeed, IFNs are produced at low levels in homeostasis, and sensing of IFNs results in tonic signaling critical for antimicrobial immune responses^25–27 and shaping the immune landscape²⁸ and immune activity.²³ Whether such IFN signaling varies across healthy humans, and whether it is linked to the human gut microbiome, remains incompletely understood.

Here, we investigated immunological variation across healthy individuals and interrogated whether microbiome features were associated with these immune states. We generated multimodal data of the immune system, microbiome and metabolome as a resource for the field. We used factor-based integrative analysis to capture co-varying features across the cohort and identified two major axes of immune variability in healthy individuals involving IFN response programs. While one axis was associated with gut microbiome features, the other was independent. The microbiome-associated axis of variation specifically captured IFN response gene expression along with differences in abundance and phenotypes of SIGLEC-1^high monocytes, activated memory T cell subsets, and activated CD69^high NK and MAIT cells. Applying these signatures to public datasets indicated that the microbiome-associated IFN signature was associated with increased vaccine responses and treatment-dependent cancer immunotherapy responses. Moreover, these features were stable within the same individuals over time, building our understanding of immune and microbiome setpoints in healthy humans.

Results

Immune and microbiome variation in a cohort of healthy individuals.

To study variation in the immune system and the microbiome in humans at steady state, we recruited 110 healthy participants within the greater San Francisco Bay Area to establish the ImmunoMicrobiome cohort (Fig. 1A). Participants were adults of predominantly non-Hispanic White and Asian ancestries (Fig. 1B, Methods). We collected blood and stool specimens as well as surveys related to general health, diet, and medical history. We performed immune profiling of peripheral blood mononuclear cells (PBMCs), quantified circulating factors in plasma, generated microbiome profiling data on DNA extracted from stool, and quantified metabolites from stool (Fig. 1A). We preprocessed each dataset (Methods) to extract molecular and cellular features (Table S1), the distribution of which was assessed across the cohort.

Figure 1. Immune and microbiome variation in a cohort of healthy individuals.

A. ImmunoMicrobiome cohort overview and schematic of data generation.

B. Cohort demographic breakdown.

C. Scaffold map of CyTOF data. Unsupervised immune cell clusters are colored by their coefficient of variation in frequency across the cohort. Landmark nodes (black) represent defined immune populations used as reference for mapping.

D. Left: CyTOF clusters were ranked by coefficient of variation across the cohort from most to least variable. Right: The most variable clusters are shown in a barchart.

E. UMAP visualization of CITE-Seq data. Annotations indicate broad cell subsets. Single cells are colored by clusters, with annotations detailed in Fig. S1A–B.

F. Transcriptional variation in monocytes across the cohort. The first ten harmony dimensions of CITE-Seq gene expression data are shown. For each participant, single-cell coordinates in each dimension were averaged as a proxy for relative gene expression variation. Participants are ordered by their coordinates for Dimension 1.

G. Top genes significantly associated with Dimension 1 (adj.p <0.1 by linear models) in monocytes. Genes are colored by hallmark gene signatures. Green: inflammation (M5932 and M5890); blue: interferon response (M5911 and M5913); orange: overlap.

H. Relative abundance of the 5 most represented microbiome families across study participants.

We used single-cell mass cytometry (CyTOF) to evaluate immune cell population frequencies and their variation across the cohort. We annotated unsupervised clusters of cells based on expression of key markers and visualized their variation in relative abundance on a forced-directed graph (Fig. 1C).²⁹ While most clusters were variable across individuals (Fig. 1C), ranking the clusters by their coefficient of variation revealed that certain NKT, NK, and T cell clusters showed the highest degree of variation (Fig. 1D).

We used cellular indexing of transcriptomes and epitopes by sequencing (CITE-seq) to evaluate variation in gene and protein expression of immune cell subsets across the cohort. Single cells were clustered and annotated based on both extracellular protein expression and gene expression, identifying clusters that comprised established immune cell types (Fig. 1E, S1A–B). The gene expression matrix of each cell type was subjected to independent dimensionality reduction to assess the major transcriptional axes of variation across the cohort. As an example, in monocytes, we visualized variance in gene expression across the cohort in the top 10 dimensions (Fig. 1F). Dimension 1 was positively associated with genes involved in interferon response but negatively associated with genes involved in inflammation response (Fig. 1G). Variation in gene expression of a similar scale was also found for other cell types across the cohort.

In parallel, we investigated variation in microbiome composition using metagenomic sequencing. We identified microbial clades at various taxonomic levels. As expected, most species identified were part of 4 major phyla (Fig. S1C) and showed differences in diversity across the cohort (Fig. S1D). While the top 5 most abundant families covered more than 50% of the known taxa in most study participants, the relative abundance of each family varied widely across individuals (Fig. 1H). This variation of composition was also visible at the species level (Fig. S1E). These observations were broadly consistent with the major taxa and abundance variation found in the Human Microbiome Project,³⁰ which sampled a larger population of healthy individuals.

Together, these high-level analyses highlight substantial inter-individual variation across immune and microbiome composition, as well as gene expression of key immune effectors, in a cohort of healthy individuals.

Integrative multi-omic analysis identifies microbiome-associated immune variation in interferon and inflammation responses

To identify biological features that vary concomitantly between immune and microbiome systems, we used a multi-omic approach to integrate data across modalities (Fig. S2A). We applied multi-omic factor analysis (MOFA)³¹ to generate factors composed of coordinated features across modalities that capture distinct biological variation across the cohort. This approach produced factor scores for each study participant for each factor, along with feature weights indicating the strength of association of features with each factor (Fig. S2B). Factors were numbered by variance explained. We identified features significantly associated with each factor using linear models (Methods) and evaluated the number and proportion of associated features from each modality with each factor (Fig. 2A). While factor 1 captured significant variation in immune cell gene expression, it was not significantly associated with stool microbiome or metabolome features (Fig. 2A–B). In contrast, factor 3 captured substantial variation in the immune system, microbiome species composition, microbiome pathway abundances, and metabolites. Among the top factors, factor 3 also captured the most variation in the relative abundance of immune cell subsets in both CyTOF and CITE-Seq data. Principal component analysis (PCA) of species-level microbiome composition data identified a significant correlation between MOFA factor 3 with PC2, as did PCA of CITE-Seq gene expression data (Fig. 2C, S2C). In comparison, factor 1 was associated with CITE-Seq gene expression PCs but not with microbial composition PCs (Fig. 2D, Fig. S2D). We therefore refer to factor 1 as the independent immune variation (IIV) and to factor 3 as the immune and microbiome concomitant variation (IMCV).

Figure 2. Integrative multi-omic analysis identifies microbiome-associated immune variation in interferon and inflammation responses.

A. Multi-omics features captured by MOFA factors. Labels show number of features significantly associated with each factor (adj.p <0.1 and |feature weight| >0.1). Colors indicate the proportion of features associated with each factor as a fraction of all significant features for that modality across MOFA factors.

B. Variation captured by MOFA factors by modality. Values are the fraction of features per modality associated with each factor. CITE-Seq gene expression values are averages over all cell subsets. Each axis is scaled independently.

C-D. Principal component analysis of prevalent microbiome species composition data (left) or CITE-Seq gene expression data (right) colored by participant’s Factor 3 (C) or Factor 1 score (D).

E. Transcriptional signature enrichment in IMCV (sum of −log10(adj.p)) across cell subsets from CITE-Seq data. Top 10 hallmark gene signatures are shown, ordered by signature score (sum of −log10(adj.p) across cell subsets) evaluated using mHG test.

F. Number of cell subsets with significant enrichment of hallmark transcriptional signatures IFN-α response, IFN-γ response, inflammation response for the 5 factors capturing the most variance, evaluated using mHG test.

To understand immune variation captured by IMCV, we quantified the dominant signatures in the transcriptional data across cell subsets. IFN alpha (IFN-α) response, IFN gamma (IFN-γ) response and inflammation response showed the highest signature scores (Fig. 2E). We then identified the top hallmark signatures for each factor and quantified the number of cell subsets that showed enrichment (Fig. S2E). IFN-α, IFN-γ, and inflammation response signatures were enriched in IMCV across most immune cell populations (Fig. 2F, S2E). Interestingly, these signatures were also highly associated with IIV, which also captured additional signatures (Fig. 2F, S2E). This further suggests that interferon and inflammation responses are dominant sources of transcriptional variability across the cohort, captured by IMCV and IIV.

Age and sex have been previously associated with features of the immune system. While neither was significantly associated with IMCV or IIV, these variables were associated with other MOFA factors (Fig. S2F).

IMCV captures differences in the interferon response across the cohort

While IFN-related signatures were broadly captured by both IMCV and IIV (Fig. 2F), we hypothesized that they might capture different elements of IFN response. Using IFN gene sets that do not overlap with inflammation-related signatures, we observed that most cell subsets showed positive enrichment for IFN signatures for both factors, though associations with IMCV were stronger for most (Fig. 3A). We examined the specific genes that drive these signatures by evaluating differentially expressed genes (DEGs) between subgroups of individuals with the highest and lowest (20%) scores for each factor, across cell subsets (Fig. S3A). Most genes upregulated in the high-IIV subgroup were expressed by lymphoid subsets (Fig. 3B–C, Table S2), while the most upregulated genes in the high-IMCV subgroup were enriched in myeloid subsets (Fig. 3B, 3D). Most DEGs were unique to one factor or the other (Fig. 3E). Similar results were found for the IFN-α response signature (Table S2).

Figure 3. IMCV captures differences in the interferon response across the cohort.

A. Association of IIV and IMCV with IFN response signatures across cell subsets using mHG test.

B-E. Number (B) and differential expression of IFN-γ response genes between study participants with high (top 20%) and low (bottom 20%) IIV (C) and IMCV (D) factor scores. Wilcoxon rank-sum test and BH correction were used. Highest DEGs (adj.p <10⁻⁴ and |log2 fold difference| >0.25) are visualized on a Venn diagram (E).

F. Association of IIV and IMCV with tonic-IFN response transcriptional signature across cell subsets using mHG test.

G-H. Relative plasma levels of IFN-γ (G) or CXCL10 and CXCL11 (H) in participants with high and low IIV and IMCV factor scores (with no overlap between the sub-cohorts). Kruskal-Wallis and Dunn post-hoc tests were performed. ***=p<0.0005, **=p<0.005,*=p<0.05.

I. Enrichment of the tonic-IFN gene signature in immune cell subsets of individuals who mounted high (n=10) or low (n=10) responses to the seasonal influenza vaccine. DEGs were identified by limma (p<0.1) from pseudo-bulked and batch-corrected expression data for each cell type and used in mHG test to evaluate enrichment of IFN gene signatures, adjusted for multiple testing.

J. Enrichment of the tonic-IFN gene signature in immune cell subsets of non-small cell lung cancer patients who responded to anti-PD1 alone (n=2) or only responded after the combination with a JAK inhibitor (n=3). Analysis performed as in (I).

Interferon signaling induces different transcriptional responses depending on the duration and strength of the stimulus.³² Because of the differences in IFN response signatures between IMCV and IIV, we investigated their overlap with more nuanced and specific IFN response gene signatures. Genes strongly positively associated with IMCV were enriched for a “tonic-IFN response” signature derived from an in vivo study that identified genes responsive to steady-state, constitutively-expressed IFN (Fig. 3F, S3B).²⁷ They were also enriched for a related “prolonged-IFN response” signature, composed of ISGs expressed in response to prolonged low doses of IFN in human cells in vitro (Fig. S3C),³³ suggesting potential differences in IFN exposure across the cohort. In contrast, associations between IIV and these signatures were much weaker (Fig. 3F, S3C).

We therefore hypothesized that circulating IFN levels might be higher in the high-IMCV subgroup of study participants. We measured cytokine levels in plasma specimens and found higher levels of IFN-γ, along with chemokines CXCL10 and CXCL11 encoded by ISGs, in high-IMCV study participants (Fig. 3G–H). Plasma levels of these molecules also correlated with the tonic-IFN response gene signature across cell subsets (Fig. S3D).

While IIV and IMCV capture unique immune variability across the full cohort, participants can be classified by factor scores to identify distinct immune states (Fig. S3E). We defined sub-cohorts of the 20 individuals with the highest scores for each factor, which were distinct except for two participants among the top 20 for both factors, whom we excluded from this analysis. Contrasting high-IMCV and high-IIV participants revealed a relative enrichment of IFN-α, IFN-γ, tonic- and prolonged-IFN response signatures across myeloid and several innate lymphoid cell subsets in the high-IMCV sub-cohort (Fig. S3F). While these signatures were derived from literature, we explored whether individual composite scores, based on core genes driving the enrichment of each signature, might further summarize immunological states in these sub-cohorts (Table S3). Among the genes in this signature, IFNGR2, IRF7 and IRF9 involved in IFN-γ signaling were higher in the high-IMCV cohort compared to the high-IIV cohort (Fig. S3G), consistent with their higher IFN response signatures and ISG expression.

We hypothesized that interindividual variation in IFN states could reflect steady-state differences that impact responses to immunomodulatory treatments. In transcriptional data from an influenza vaccine study,³⁴ the baseline immune states of individuals who went on to mount strong immune responses to the vaccine were associated with the same tonic-IFN response signature (Fig. 3I). Notably, history of recent vaccination in ImmunoMicrobiome participants showed only a modest association between influenza vaccine status and IIV factor score and no association with IMCV (Fig. S3H). Moreover, in patients with non-small cell lung cancer treated with immunotherapy,³⁵ the tonic-IFN response signature at the start of the trial was associated with patients who went on to respond to anti-PD1 checkpoint blockade only after the addition of the JAK inhibitor itacitinib, which blocks IFN signaling (Fig. 3J).

Taken together, these data show that IIV and IMCV are associated with distinct IFN-related transcriptional programs. Elevated expression of an IFN signature in IMCV was also observed in individuals who mounted higher responses to influenza vaccination and cancer patients who benefited from combined JAK inhibition with anti-PD1.

Inflammation and TGF-β transcriptional programs are differentially regulated between IMCV and IIV

Our prior analysis indicated that inflammation-related gene signatures were also among the most variable across the cohort (Fig. 2E) and were associated with IMCV and IIV (Fig. 2F, S2E). Indeed, IIV and IMCV factor scores were positively associated with IFN signatures and negatively associated with inflammation-related signatures (Fig. 4A). In the 3 major monocyte clusters identified by CITE-seq, both IMCV and IIV exhibited negative associations with inflammation response and TNF-α/NF-κB signaling signatures (Fig. 4B). However, these were driven by a distinct set of DEGs for each factor (Fig. 4C–D, S4A, Table S4). IL-1 pathway genes, including IL1A and IL1B, were expressed at lower levels in individuals with high IMCV scores (Fig. 4C, S4B). In contrast, the high-IIV group exhibited lower expression of IL-6 regulators, including oncostatin M (OSM)³⁶ and TIMP1.³⁷ Similar results were observed with the TNF-α/NF-κB signaling signature (Fig. 4D, Table S4), which highlighted higher levels of TLR/IFN-induced transcription factor ATF3 that represses NF-κB signaling³⁸ in the high-IMCV group. Lipopolysaccharide-induced TNF-α factor (LITAF) that induces TNF-α expression³⁹ was expressed at higher levels in the high-IMCV group but lower in the high-IIV group (Fig. 4D, S4B). These results indicate that both IMCV and IIV are inversely associated with inflammation and TNF pathways in monocytes but through distinct sets of genes (Fig. S4A).

Figure 4. Inflammation and TGF-β transcriptional programs are differentially regulated between IMCV and IIV.

A. Correlation of IIV and IMCV factor scores with IFN and inflammation-associated signatures (sum of weighted expression of core genes).

B. Association of IIV and IMCV with inflammation-associated signatures across cell subsets. mHG test was used. Signed −log10(adj.p) for each factor is shown.

C-E. Inflammation response genes (C) or TNF-α/NF-κB signaling genes in (D) monocyte subsets or (E) MAIT cells differentially expressed between participants with high and low factor scores. Wilcoxon rank-sum test and BH correction were used. DEGs (adj.p <0.1 and |log2 fold difference| >0.25) are annotated by factor.

F. TNF-α/NF-κB signaling genes differentially expressed between participants with high and low IIV scores. Wilcoxon rank-sum test and BH correction were used. DEGs (adj.p <10⁻⁴ and |log2 fold difference| >0.25) are annotated by cell subset.

G-H. Same as F for TGF-β signaling signature in high and low IIV scores (G) or IMCV scores (H).

I. Plasma cytokine association with IMCV and inflammation-related signatures across cell subsets. Spearman rank correlation was used. |r| >0.25 indicated with an asterisk.

J. Plasma cytokine levels between high and low factor scores in IIV and IMCV. Cytokines higher in the IMCV-high cohort but lower in the IIV-high cohort are annotated, and cytokines associated with cMo-A TGF-β signatures are colored.

In many other cell subsets, only IIV was negatively associated with inflammation-related signatures (Fig. 4B). MAIT cells, a subset of invariant T cells recognizing vitamin metabolites, including those derived from the gut microbiome,⁴⁰ were the only subset in which the TNF-α/NF-κB signature was positively associated with IMCV but negatively associated with IIV (Fig. 4B). Key genes driving the negative association with IIV included TNF, genes encoding NF-κB subunits, and their regulators (Fig. 4E, Table S4). Conversely, the positive association with IMCV was driven by higher expression of inflammation-associated factors downstream of TNF-α. The transcription factor JUNB, a negative regulator of JUN,⁴¹ was expressed at higher levels in the high-IIV subgroup, suggesting differences in AP-1 pathway regulation in MAIT cells. While there was no association between the TNF-α/NF-κB pathway and IMCV in other lymphocyte subsets, genes driving negative associations with IIV largely overlapped with those in MAIT cells (Fig. 4F).

Consistent with these results, the sub-cohorts of high-IMCV and high-IIV individuals displayed differences in inflammation gene signatures in both MAIT and mNK-A lymphoid cell subsets (Fig. S4C). Their composite scores of TNF-α/NF-κB and inflammation gene signatures also stratified the high-IMCV and high-IIV sub-cohorts (Fig. S4D). Since microbial cues can activate these lymphoid cell subsets,⁵⁷ we additionally examined surface protein and mRNA expression of the activation marker CD69, which was expressed at higher levels in MAIT and mNK-A cells in the high-IMCV cohort (Fig. S4E–F), resulting in a higher relative abundance of CD69^high cells (Fig. S4G).

We also investigated associations between these factors and the TGFβ pathway, which was uniquely negatively associated with IIV (Fig. 4B). Individuals in the high-IIV group exhibited lower TGFB1 expression across multiple cell subsets and higher levels of TGF-β receptor inhibitor PPP1CA (Fig. 4G, Table S4). In contrast to IIV, the TGF-β signature in cMo-A was positively associated with IMCV (Fig. 4B), driven by modestly higher expression of TGFB1, IFNGR2, JUNB, transcription factors ID1 and ID2, and lower expression of PPP1R15A, involved in TGFBR1 inhibition (Fig. 4H).⁴²

We further evaluated circulating factors associated with these pathways. Inflammation and TNF-α/NF-κB signaling signatures in monocytes were positively associated with circulating levels of CD40, which signals through the NF-κB pathway (Fig. 4I). TGF-β signature in cMo-A correlated with plasma levels of the LAP-TGF-β1 complex as well as other factors associated with IMCV, including CXCL10 and CXCL11 (Fig. 4I). Consistent with transcriptional data, soluble LAP-TGF-β1 levels were negatively associated with IIV (Fig. 4J). Taken together, these data show that high-IMCV and high-IIV subgroups exhibit key differences in inflammation and TGF-β signaling pathways.

IMCV captures differences in SIGLEC-1^high monocytes and PD1^highICOS^high memory T cells across the cohort

IMCV was further distinguished from IIV by its association with immune cell population abundances (Fig. 2A). The relative abundance of monocyte subsets that exhibited distinct interferon and inflammation transcriptional signatures was positively associated with IMCV (Fig. S5A). CyTOF data, which allowed for deeper sampling due to its higher throughput, confirmed that the relative abundances of several monocyte clusters were positively associated with IMCV (Fig. 5A). We therefore compared the protein expression of these clusters with monocyte clusters not associated with IMCV to understand their differences (Fig. 5B). Monocytes in clusters associated with IMCV exhibited higher protein expression of the MHC class II protein HLA-DR and the Fc gamma receptor CD64, both regulated by IFN (Fig. 5B–C).^43–46 In addition, these cells expressed higher levels of myeloid activation markers (CCR7, CD14) and the scavenger receptor CD163 associated with an anti-inflammatory phenotype (Fig. 5B–C).⁴⁷ Consistent with these results, CITE-Seq classical monocyte cluster B (cMo-B), the abundance of which was associated with IMCV, also displayed higher expression of MHC class II pathway genes (HLA-DP/DQ/DR and CD74) compared to other cMo subsets. They also exhibited higher expression of the immunoproteasome gene PSME2, and calreticulin (CALR), a chaperone critical for antigen processing and presentation, along with ISGs (Fig. 5D). Conversely, cMo-B had lower expression of pro-inflammatory genes including AP-1 subunits (JUN, FOS), and calprotectin (S100A8/S100A9), an endogenous TLR4 ligand (Fig. 5D). To further investigate differences in myeloid cell phenotypes captured by IMCV, we compared CITE-seq protein levels in individuals with high and low IMCV factor scores. SIGLEC-1 (CD169), previously shown to define IFN-induced monocytes with enhanced antigen presentation capability,⁴⁸ protection against sepsis in mice⁴⁹ or severe COVID in humans,⁵⁰ was more highly expressed by monocytes in high-IMCV individuals (Fig. 5E, S5B). Although cMo-B expressed the highest level of SIGLEC-1, other myeloid subsets showed similar differential expression in high-IMCV study participants, suggesting that elevated SIGLEC-1 expression may be characteristic of a broader myeloid state in these individuals (Fig. S5B). SIGLEC-1 was also expressed at a higher level in monocytes of participants in the high-IMCV sub-cohort compared to the high-IIV sub-cohort, resulting in a higher relative abundance of SIGLEC-1^high monocytes (Fig. S5C–D).

Figure 5. IMCV captures differences in SIGLEC-1^high monocytes and PD1^highICOS^high memory T cells across the cohort.

A. SCAFFoLD map of CyTOF cluster frequency associations with IMCV. Unsupervised clusters (white and colored nodes) are mapped relative to reference populations (black nodes). Linear regression and BH correction (adj.p <0.1) was used.

B. Top: Comparison of CyTOF protein expression on classical monocyte clusters with frequencies significantly positively associated with IMCV (pos) versus classical monocyte clusters that were not significantly associated with IMCV (ns). Left: scaled median protein expression values; Right: significance between the groups. Bottom: Same for non-classical monocyte clusters. Kruskal–Wallis and post-hoc pairwise Wilcoxon tests were used (p<10⁻⁴).

C. Expression of selected surface proteins shown in B.

D. Differential CITE-Seq gene expression between cMo-B and other classical monocyte cell subsets using Wilcoxon rank-sum test. Features with adj.p <0.1 and |log2 fold-change| >0.25 are annotated.

E. Differential CITE-Seq protein expression between study participants with low and high IMCV factor scores using Wilcoxon rank-sum test. Features with adj.p <0.1 and |log2 fold-change| >0.2 are annotated.

F. Same as B, comparing memory CD4 T cell cluster 115 to all memory CD4 T cell clusters not associated with IMCV (ns) and comparing CD8 T cell cluster 166 to all memory CD8 T cell clusters not associated with IMCV (ns).

G. Same as C, with combined memory CD4 T cell clusters 114/115 and CD8 T cell clusters 166/181.

H. Correlation between CITE-seq gene expression in CD8 effector memory T cells and CyTOF frequency of CD8 effector memory T cell clusters C166 and C181. Genes with correlation coefficient |r| >0.2 are displayed and annotated. Coloring indicates manually-curated broad functions (Table S4).

Several memory T cell clusters were also positively associated with IMCV (Fig. 5A). Memory CD4 T cell cluster 115 (C115) and CD8 T cluster 166 (C166) exhibited higher expression of a combination of costimulatory and coinhibitory checkpoint molecules when compared to memory CD4 or CD8 T cell clusters that were not associated with IMCV (Fig. 5F). These clusters also expressed higher levels of the activation marker HLA-DR and transcription factor Foxp3 and lower expression of cytokine receptors CD127 and CD25. Two additional clusters positively associated with IMCV displayed similar phenotypes except for the expression of Foxp3 (Fig. 5G, S5E). When compared to Tregs (Fig. S5F–G) and memory T cells not associated with IMCV (Fig. S5H–I), cells from these four clusters combined displayed higher expression of Foxp3 and most checkpoint molecules, including inhibitory checkpoint ligand PD-L1. CD4 and CD8 memory T cell clusters that were negatively associated with IMCV (Fig. 5A and S5A) generally revealed the opposite expression profile (Fig. S5F–I). We further identified correlations between abundances of these CyTOF clusters and gene expression in the CITE-seq data for CD8 effector memory T cells. Most positively correlated genes were related to T cell activation, including many antigen processing and presentation molecules (Fig. 5H, Table S5), consistent with the results from CyTOF. Together, these results suggest that IMCV is associated with memory T cells with an activated phenotype and elevated expression of checkpoint molecules.

Established immunomodulatory gut microbiome pathways and molecules are associated with IMCV.

The most unique feature of IMCV was its coordination between immune and microbiome datasets (Fig. 2A). The reference-based analysis approach used in combination with MOFA (Fig. S6A) identified 77 microbial pathways associated with IMCV (Table S6), including several related to production of known immunomodulatory molecules: biosynthesis of short-chain fatty acids (SCFA), metabolism of polyamines, and production or modification of bacterial cell wall components lipopolysaccharide (LPS) and peptidoglycans (Fig. 6A).^51–54 In addition, 78 stool metabolites were also associated with IMCV (Table S6), including many of the immunomodulatory molecules produced by these pathways (Fig. 6B). The SCFAs acetate, butyrate, and propionate were positively associated with IMCV, as were SCFA-related compounds. Inositol, a precursor of propionate, was positively associated with IMCV, while its phosphorylated precursor, inositol-P,⁵⁵ was negatively associated. Of 8 polyamines detected, 6 were positively associated with IMCV, including ornithine that had precursors negatively associated with IMCV (Fig. S6B). Primary bile acids chenodeoxycholic acid and cholic acid, also studied for their immunomodulatory effects,⁵⁶ were also associated with IMCV. Together, these findings reveal that the immune states captured by IMCV are coincident with immunomodulatory pathways in gut microbiota and the metabolites they produce.

Figure 6. Immunomodulatory gut microbiome pathways and molecules are associated with IMCV.

A. Associations between microbiome pathway abundances and IMCV. Linear regression and BH correction (adj.p <0.1) and MOFA feature weight(|FW| >0.45) were used. Table S6 includes all pathways associated with IMCV.

B. Associations between stool metabolites and IMCV. Linear regression and BH correction (adj.p <0.1) and and MOFA feature weight(|FW| >0.45) were used. Table S6 includes all metabolites associated with IMCV.

C. Species contribution to acetate biosynthesis pathway genomic content. Major contributors (gene family prevalence across the cohort >25% and median contribution to total gene copy number >0.05) are shown. All contributors documented in Table S7.

D-F. Species contribution to genomic content of gene families for SCFA biosynthesis (D), polyamine metabolism (E), Kdo2-lipid A (LPS-endotoxin) biosynthesis (F). Major contributors (20 species with highest median contribution to total gene copy number, among species with median contribution >0.05 and gene family prevalence across the cohort >10%) are shown. All contributors documented in Table S7.

G. Spearman correlations and linear regression lines between microbiome composite score (median of microbiome features associated with IMCV) and tonic-IFN response gene signature (median expression of core genes) across top cell subsets.

H. Volcano plot of associations between IMCV and frequencies of co-abundant modules of SGBs. Individual SGB associations with IMCV are documented in Table S6.

I-J. Scatter plots showing SGB module membership scores and coefficients of associations with IMCV in modules positively (I) and negatively (J) associated with IMCV.

K. Circos plot visualizing co-abundances of SGBs in modules associated with IMCV and metabolic pathways/metabolites of interest. The interior demonstrates pairwise Spearman correlations between SGB relative abundances across the cohort. Spearman correlations for each SGB with ICMV are shown. Spearman correlation coefficients are shown between each SGB and each metabolic pathway or metabolite identified previously and with IFN gene expression signatures in cMoA, cMoB, ncMo and mNK_A.

Next, we investigated the bacterial species encoding these pathways. We measured their contribution to pathway abundances through the prevalence and average copy number of the relevant genes they carried across the cohort (Table S7). Bifidobacterium longum, Bifidobacterium adolescentis, Bifidobacterium pseudocatenulatum, and Collinsella aerofaciens were among the major contributors to the acetate biosynthesis pathway (Fig. 6C), and their relative abundances were also correlated with acetate biosynthesis pathway abundances (Fig. S6C). B. longum abundance was directly associated with IMCV (Fig. S6D). Several of the same species were also major contributors to a non-exhaustive set of metabolic reactions that transform acetate to butyrate and inositol to propionate (Fig. 6D, S6E). Of these, Blautia obeum and R. torques also encoded the alkaline phosphatase gene family involved in catalyzing the dephosphorylation of inositol phosphate, suggesting a potential role for these bacteria in propionate biosynthesis (Fig. S6F).

A partially-overlapping group of species contributed to pathways related to polyamine metabolism (Fig. 6E). Several had genes encoding the transformation of agmatine to putrescine with high prevalence. The same species were also major contributors to the subsequent production of spermidine and spermine. B. adolescentis was the highest contributor in gene family abundance to the acetylation of spermine/spermidine, while R. torques and B. obeum were important contributors of gene family content for cadaverine biosynthesis from lysine. Overall, the complementarity of species contributing to polyamine metabolism was suggestive of community metabolism. Consistent with this notion, the abundance of these individual species was not associated with IMCV (Fig. S6D) nor with polyamine pathway unstratified abundances (Fig. S6H), highlighting the importance of examining metabolic pathways and metabolites in addition to microbiota composition.

Examining species that contributed to biosynthesis of LPS, detected by the innate immune receptor TLR4, revealed further overlap. Bifidobacterium species and C. aerofaciens were also important contributors to biosynthesis of n-acetylglucosamine (Fig. 6F), a precursor of LPS. Their abundances further correlated with LPS biosynthesis pathway abundances (Fig. S6I). The subsequent transformation of n-acetylglucosamine into LPS endotoxin KDo2-lipid A was dominated by distinct species, including E. coli and Haemophilus parainfluenzae. Several major contributors to these pathways were positively associated with IMCV (Fig. S6D), and several overlapped with those for SCFA and polyamine metabolism pathways. Some of these species were also major contributors to the abundance of the bile salt hydrolase gene family that can produce the unconjugated bile acids associated with IMCV (Fig. 6B, S6J).

Species carrying genes involved in production of molecules associated with IMCV (SCFA, polyamines, LPS, primary bile acids) were phylogenetically diverse, spanning more than 14 families. A core community of species, encoding genes involved in all key processes associated with IMCV, spanned 5 microbial families (Fig. S6K). Of note, the metabolic pathways encoded by these microbes and metabolite levels were more strongly associated with IMCV than were the abundances of bacterial species. An IMCV microbiome composite score of key microbiome pathway abundances and stool metabolites was significantly correlated with the tonic-IFN response signature across immune cell subsets (Fig. 6G). These results highlight a complex collection of phylogenetically diverse microbial species encoding pathways involved in the production of immunomodulatory molecules captured by IMCV.

Compared to reference-based approaches for the analysis of metagenomics data (Fig. S6A), genome-resolved analyses can enable identification of more precise strains that could have distinct functions. We performed de novo assembly of our human-depleted metagenomes to identify metagenome-assembled genomes (MAGs), from which we generated a library of sub-species genome bins (SGBs) (Fig. S6L). Each metagenome was then mapped to the SGB library to construct a genome-resolved community count matrix, and each SGB was annotated by taxonomic classification within the Genome Taxonomy Database (GTDB) to identify its nearest species-level match. SGBs were grouped according to Spearman correlation-based distances, resulting in 60 modules of co-abundant SGBs.

Of these co-abundant modules, 7 were significantly associated with IMCV (Fig. 6H). Among the 3 modules positively correlated with IMCV, module C captured 31 SGBs reflecting genetically distinct strains within the genus Collinsella, 7 of which specifically matched several C. aerofaciens references (Fig. 6I, Table S6). Module F contained 4 SGBs that matched several Bifidobacteria references (Bifidobacterium longum, Bifidobacterium catenulatum, Bifidobacterium pseudocatenulatum, and Bifidobacterium adolescentis) as well as several that matched Phocaeicola plebeius (Fig. 6I). These findings reinforce the bacterial communities identified by the pathway contribution analysis in the referenced-based approach by identifying related SGBs with abundances directly associated with IMCV. The genome-resolved analysis also identified an additional module significantly positively associated with IMCV score, module K, which included SGBs related to Turicibacter bilis, Clostridium saudiense, Romboutsia timonensis, and Intestinibacter bartlettii (Fig. 6I).

Among the modules negatively correlated with IMCV, module M was dominated by several Alistipes SGBs (Fig 6J). Alistipes putredinis, identified in the previous analyses as a major contributor to several IMCV-associated pathways, exhibited the strongest association with IMCV of these SBGs (Fig. 6J, Table S6). Several modules uniquely identified as negatively associated with IMCV by the genome-resolved approach included module H, containing Lachnospiraceae-related Eubacterium_F sp003491505_3 and Eubacterium_F sp003491505_1 among others; module S, containing SGBs related to Gallintestinimicrobium propionicum, Acetatifactor intestinalis, Acetatifactor acetigignens and Roseburia hominis; and module U, containing several strains of Hominicoprocola (Fig. 6J, Table S6).

Overall, the relative abundances of individual SGBs from IMCV-associated modules showed consistent directionality in their correlations with tonic IFN response signatures and key IMCV pathways and metabolites (Fig. 6K).

IFN-response signatures and IMCV associated features are stable over time.

We hypothesized that features of IMCV may be stable over time within individuals, reflecting a coordinated setpoint of the immune system and microbiota. We therefore collected biospecimens from a subset of the cohort (n = 48) at a distant follow-up timepoint, approximately 20 months after the baseline sampling timepoint (Fig. S7A), and generated a longitudinal multi-omic dataset to match prior analyses. The follow-up cohort represented the full range of IMCV scores observed in the overall ImmunoMicrobiome cohort (Fig. S7B)

To evaluate the stability of immune phenotypes over time, we performed CITE-seq, plasma cytokine, and CyTOF analyses of longitudinal blood samples. CITE-seq data were used to quantify the tonic IFN response signature, as well as IFNα and IFNγ response signatures, across all major cell types (Fig. 7A–B). Indeed, significant correlations were observed across timepoints for the key cell populations previously observed to exhibit strong IFN-related gene expression programs (Fig. 3A–F), indicating that these IFN response signatures were stable within individual participants over time. Similarly, the levels of IFNγ and ISG-encoded chemokines CXCL10 and CXCL11 in plasma were significantly correlated between the two timepoints across participants (Fig. 7C–D), as were levels of CCL20 and CD40, also significantly associated with IMCV in the original analysis. CyTOF analysis revealed that the frequencies of key immune cell subsets previously identified (Fig. 5A) were also highly correlated between the two timepoints across participants (Fig. 7E–F). Collectively, these results indicate that key immune parameters contributing to IMCV were quite stable within study participants over a period of more than a year.

Figure 7. IFN-response signatures and IMCV associated features are stable over time.

A. Spearman correlations of interferon-associated gene signatures in immune cell subsets between the two timepoints of each participant (n=16). Cells from the follow-up timepoint were assigned to their most similar cluster from the baseline analysis using Azimuth. Color represents Spearman correlation coefficient; * indicates p <0.05. B. Scatter plots visualizing tonic-IFN response gene expression scores at the two timepoints for key cell populations with Spearman correlations and linear regression lines. Each data point represents a participant.

C. Spearman correlations of circulating soluble factors associated with IMCV between the two timepoints for each participant (n=45). * indicates a p <0.05.

D. Scatter plot visualizing composite score of soluble factors associated with IMCV between the two timepoints with Spearman correlation and linear regression line. Each data point represents a participant.

E. Spearman correlations of relative abundances of cell subsets associated with IMCV between the two timepoints for each study participant as measured by CyTOF (n=48). Cells from the follow-up timepoint were assigned to their most similar cluster from the baseline analysis using MARIO. * indicates p <0.05.

F. Scatter plot visualizing composite score of the cell clusters significantly associated with IMCV between the two timepoints with Spearman correlation and linear regression line. Each data point represents a participant.

G. Spearman correlations of pathway abundances from stool metagenomic sequencing data between the two timepoints for each study participant (n=46). * indicates p <0.05.

H. Scatter plot visualizing composite score of pathways significantly associated with IMCV between the two timepoints with Spearman correlation and linear regression line. Each data point represents a participant.

I. Spearman correlations of stool metabolite abundances between the two timepoints for each study participant (n=48). * indicates p <0.05.

J. Scatter plot visualizing composite score of metabolites significantly associated with IMCV between the two timepoints with Spearman correlation and linear regression line. Each data point represents a participant.

K. UMAP dimensionality reduction of all key features described above for each participant at each timepoint. The two timepoints for each participant are connected by a line. Participants are colored by their IMCV factor score from the original MOFA analysis. Participants with longitudinal data for all modalities (n=15) are plotted.

In parallel, to investigate the stability of the microbiome and derived metabolites over time, we performed metagenomic and metabolomic analysis of longitudinal stool samples from these individuals. The abundances of most key pathways significantly associated with IMCV (Fig. 6A) exhibited a high degree of correlation between the two timepoints for each participant (Fig. 7G–H). This was also the case for the relative abundances of key bacterial taxa that associated with IMCV (Fig. S7C–D). In addition, key metabolites associated with IMCV (Fig. 6B) were also highly correlated between the two time points for each study participant (Fig. 7I–J). Collectively, these results indicate a high degree of stability of the microbial metabolic pathways and metabolite levels associated with IMCV.

Lastly, we performed a dimensionality reduction of these features for each study participant at each timepoint, visualized by UMAP (Fig. 7K). Indeed, data from the two timepoints for any given individual were near one another, indicative of the stability of these immune- and microbiome-related features over time.

Discussion

This study presents a comprehensive multi-omic dataset exploring variation in the immune system and microbiome of healthy individuals. Using a factor-based integrative approach, we identified 2 axes of immune variation involving IFN response programs. One encompassed specific microbiome pathways, metabolites, and microbial communities that covaried with immune cell abundances and gene expression. This distinct microbiome-associated immune state was characterized by heightened IFN-responsive gene expression across immune cell types, expanded SIGLEC-1^high monocytes with downregulated inflammation transcriptional programs, activated memory T cell subsets, and activated CD69⁺ MAIT and NK cells. These signatures were remarkably stable over time.

Clinical studies have identified associations between the microbiome and immunological states in pathological settings.^{15,50,58–60} However, the extent to which the microbiome and immune system are coordinated in humans in the absence of pathology has been unclear. Among immune features we identified to be coordinated with the microbiome in healthy individuals, several have been associated with clinical outcomes in patients. For example, elevated levels of HLA-DR^high SIGLEC-1⁺ monocytes in circulation at baseline were found to be protective of severe COVID-19, while expression of the alarmins S100A8 and S100A9 was elevated in patients with severe disease.⁵⁰ This is consistent with our finding that SIGLEC-1^high monocytes associated with IMCV expressed lower levels of S100A8 and S100A9 transcripts compared to other monocyte subsets (Fig. 5D). Another study found that SIGLEC-1^high monocytes had heightened T cell activation capacity,⁴⁸ consistent with our finding that SIGLEC-1^high monocytes had higher expression of HLA-DR (Fig. 5B–C) and genes encoding MHC class II compared to other monocyte subsets (Fig. 5D). We also found higher frequencies of activated memory T cells (Fig. 5F–G), which future studies should investigate for their potential to recognize microbiota-derived antigens or protect against pathogens. Another immune population associated with IMCV was CD69^high activated MAIT cells, consistent with reports of their activation by microbiome-derived molecules.⁵⁷ Recent studies found that MAIT cells can confer protection against colitis and inflammation-induced colorectal cancer after activation by microbiome-derived ligands.^61,62

The most prominent immune variation associated with microbiome features in our study was the IFN response. Since many ISGs encode antiviral mediators, our findings suggest that the microbiome could be a source of interindividual variability in susceptibility to infection. This is consistent with studies that have shown that reduced microbiome diversity or depletion by antibiotics result in poor antiviral protection and weakened immune responses.¹⁹ Additionally, the IFN axis plays a critical role in immunomodulatory therapies, which often exhibit substantial variability in patient response.¹² We observed an elevated IFN state in individuals with stronger immune responses to the influenza vaccine but also in patients treated with anti-PD1 checkpoint blockade who only responded upon IFN-axis inhibition.

While beneficial and detrimental microbial communities have been identified in animal models, specific findings have not always applied to humans. Additionally, widespread discrepancies between studies make the taxonomic definition of a “good” human microbiome elusive.⁶³ Moreover, taxonomy-based microbiome analysis has several important limitations, as novel or divergent strains can be overlooked if not represented in existing databases, and distinct strains may exhibit unique functions. Genome-resolved analyses can enable greater precision, as evidenced by our results. In addition, co-abundant taxa can create ecologies with emergent properties. There is growing recognition that microbiome genetic content, encoding functional potential, may matter more than the phylogenetic identity of the bacteria. Our findings identify a combination of microbial pathways and molecules associated with IFN responses. These include SCFAs, polyamines, LPS and peptidoglycans, and primary bile acids, consistent with mechanistic studies in model organisms showing that commensal bacteria contribute to shaping the immune system.^{19,64–68 40,51,52,69–71} Notably, many of these immunomodulatory molecules were elevated in the same subset of individuals, suggesting that their production may be coordinated with one another and that their combination may matter. Future studies should focus on revealing the functional and causal relationships between the variable features identified in this study and their impact on disease development or protection.

We envision that the data resource presented here will provide insights and context for efforts to modulate the microbiome and the immune system for therapeutic benefit through approaches including dietary interventions,^72–74 microbial colonization,⁷⁵ or microbiome-informed precision immunotherapy.

Limitations of the Study

The ImmunoMicrobiome study profiled biological variation in a cohort of study participants living in the San Francisco Bay Area at the time of the study. The data do not account for the biological variability exhibited across the whole human species, though local sampling enabled us to exclude large environmental variables associated with geography as main drivers of the inter-individual variation we observed. Similarly, baseline study specimens were collected within a 3-month period, which limited the impact of seasonal influences and our ability to examine their effects. Larger cohorts and further longitudinal sampling will be required to address the impacts of geography and seasonality. While we nominate potential biomarkers of immune setpoints, future studies should assess their utility in clinically relevant contexts. Future studies should determine the effect of the viral exposome on immune variation. While study participants were free of disease symptoms, understanding variation in prior and current viral exposures, including gut bacteriophages, could provide additional insights into human immune setpoints.

Resource availability

Lead contact:

Matthew H. Spitzer, matthew.spitzer@ucsf.edu

Materials Availability:

This study did not generate new unique reagents.

Data and Code Availability:

Sequencing data are available at NCBI BioProject (PRJNA1390888). CITE-seq (GSE314416) and bulk RNA-seq (GSE314922) raw and processed data are available at Gene Expression Omnibus (GEO), (SuperSeries GSE314923), and associated raw FASTQ files at Sequence Read Archive. CyTOF, metabolomics, Olink, and survey data are available at Zenodo (https://doi.org/10.5281/zenodo.18012243). Code used for data processing and statistical analysis are available at https://github.com/UCSF-DSCOLAB/ImmunoMicrobiome under an open-source license (MIT), along with metadata, sample annotations, processed data, and MOFA model in matrix and/or R object formats.

STAR Methods

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

Study participant enrollment and consent

Healthy subjects were enrolled and written informed consent was obtained under the UCSF IRB approved protocol #18–27022. Healthy status and eligibility were defined through screening questionnaires, based on the absence of cancer history, autoimmune disease, severe allergy or asthma, known immunodeficiency or chronic infection. Study participants were major, non-pregnant and free of medical treatment affecting the immune system or the microbiome at the time of enrolment and during the study. Specifically, history of surgery, use of immunosuppressants or broad-spectrum antibiotics taken during or in the last 4 months before enrolment were exclusion criteria, and treatment history has been documented. Study participants who underwent seasonal vaccines before the study visit were at least 30 days out of their last shot at the time of sampling and vaccine history has been accounted for in the analysis. Most study participants lived in San Francisco County (except for few who resided in Alameda and San Mateo counties). The race and ethnicity bias in the cohort, characterized by overrepresentation of non-Hispanic White and Asian individuals, reflected the demographic breakdown of the counties where participants lived (source: http://data.census.gov). The age, sex, race, and ethnicity of the participants are reported as distributions in Fig. 1B and as individual data values in survey files (see key resource table).

Key resources table.

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
140 CITE-seq antibody panel
137 TotalSeq-C antibody panel (Human Universal Cocktail, V1.0, Lot: B343280)	BioLegend	CAT#399905
anti-human_CD204 (Clone 7C9C20)	BioLegend	CAT#371911
anti-human_CD206 (Clone 15-2)	BioLegend	CAT#321147
anti-human_CD209 (Clone 9E9A8)	BioLegend	CAT#330121
36 CyTOF antibody panel
anti-human CD61 unconjugated antibody	ThermoFisher	CAT#14-0619-82
anti-human CD235ab unconjugated antibody	ThermoFisher	CAT#14-9987-82
anti-human CD4 unconjugated antibody	ThermoFisher	CAT#14-0049-82
anti-human CD11c unconjugated antibody	ThermoFisher	CAT#MA1-82142
anti-human TIM3 unconjugated antibody	ThermoFisher	CAT#16-3109-85
anti-human CD137 unconjugated antibody	ThermoFisher	CAT#14-9056-82
anti-human Tbet unconjugated antibody	ThermoFisher	CAT#14-5825-82
anti-human CD152 (CTLA-4) unconjugated antibody	ThermoFisher	CAT#14-1529-82
anti-human FoxP3 unconjugated antibody	ThermoFisher	CAT#14-4776-82
anti-human CD31 unconjugated antibody	ThermoFisher	CAT#14-0319-82
anti-human CD68 unconjugated antibody	ThermoFisher	CAT#14-0689-82
anti-human FceRIa unconjugated antibody	ThermoFisher	CAT#14-5899-82
anti-human CD123 unconjugated antibody	ThermoFisher	CAT#14-1239-82
anti-human FoxP3 unconjugated antibody	ThermoFisher	CAT#14-4776-82
anti-human CD56 unconjugated antibody	BD Biosciences	CAT#559043
anti-human CD45 unconjugated antibody	Biolegend	CAT#304002
anti-human CD15 unconjugated antibody	Biolegend	CAT#323002
anti-human CD3 unconjugated antibody	Biolegend	CAT#300402
anti-human CD19 unconjugated antibody	Biolegend	CAT#302202
anti-human CD8a unconjugated antibody	Biolegend	CAT#301002
anti-human CD14 unconjugated antibody	Biolegend	CAT#301802
anti-human CD127 unconjugated antibody	Biolegend	CAT#351302
anti-human CD45RA unconjugated antibody	Biolegend	CAT#304102
anti-human TIGIT unconjugated antibody	Biolegend	CAT#372720
anti-human PD-L1 unconjugated antibody	Biolegend	CAT#329702
anti-human CD27 unconjugated antibody	Biolegend	CAT#302802
anti-human BDCA3 unconjugated antibody	Biolegend	CAT#344102
anti-human CCR7 unconjugated antibody	Biolegend	CAT#353202
anti-human KI-67 unconjugated antibody	Biolegend	CAT#350502
anti-human CD25 unconjugated antibody	Biolegend	CAT#356102
anti-human BDCA1 unconjugated antibody	Biolegend	CAT#331502
anti-human CD38 unconjugated antibody	Biolegend	CAT#303502
anti-human ICOS unconjugated antibody	Biolegend	CAT#313502
anti-human HLA-DR unconjugated antibody	Biolegend	CAT#307648
anti-human PD-1 unconjugated antibody	Biolegend	CAT#329902
anti-human CD16 unconjugated antibody	Biolegend	CAT#302050
Chemicals, peptides, and recombinant proteins
RLT Buffer	Qiagen	CAT#79216
Cell Staining Buffer	Biolegend	CAT#420201
Human TruStain FcX	Biolegend	CAT#422301
Critical commercial assays
Chromium Next GEM Single Cell 5’ v1.1	10X Genomics	CAT#PN-1000165
Chromium Single Cell 5’ Feature Barcode Library for antibody derived tag	10X Genomics	CAT#PN-1000080
Automated Self-Administered 24-Hour Dietary Assessment Tool	ASA24	https://epi.grants.cancer.gov/asa24
Olink Target 96 Inflammation	Olink	CAT#91301
Deposited data
Cellular Indexing of Transcriptomes and Epitopes (CITE)-sequencing	This paper	Deposited into GEO under accession number GSE314416
Bulk RNA-sequencing of PBMCs	This paper	Deposited into GEO under accession number GSE314922
Whole metagenome shotgun sequencing of stool	This paper	Deposited into GEO under accession number PRJNA1390888
Mass cytometry (CyTOF) of PBMC	This paper	Deposited into Zenodo: https://doi.org/10.5281/zenodo.18012243
Targeted and untargeted metabolomic profiling of stool	This paper	Deposited into Zenodo: https://doi.org/10.5281/zenodo.18012243
Serological profiling (Olink) of plasma	This paper	Deposited into Zenodo: https://doi.org/10.5281/zenodo.18012243
Survey data (24h food recall and general survey)	This paper	Deposited into Zenodo: https://doi.org/10.5281/zenodo.18012243
CITE-seq data from an influenza vaccine study	Mulè et al 34	https://doi.org/10.5281/zenodo.10546916
Single-cell sequencing data from non-small cell lung cancer patients treated with immunotherapy	Mathew et al 35	Obtained from the authors
Software and algorithms
FastP (v0.23.2)	Chen et al 76	https://github.com/OpenGene/fastp
HUMAnN3	Beghini et al 77	https://github.com/biobakery/humann
MicrobeCensus (v1.1.1)	Nayfach & Pollard 95	https://github.com/snayfach/MicrobeCensus
R (versions 4.1.1, 4.3.0)	The R foundation	https://www.r-project.org
Premessa	PICI institute	https://github.com/ParkerICI/premessa
cyCombine (v0.1.5)	Pedersen et al 78	https://github.com/biosurf/cyCombine
cyclone	Patel & Jaszczak et al 79	https://github.com/UCSF-DSCOLAB/cyclone
FlowSOM (v2.0.0)	Van Gassen et al 80	https://github.com/SofieVG/FlowSOM
SCAFFoLD (v0.1)	Spitzer et al 29	https://github.com/nolanlab/scaffold
CellEngine	CellCarta	https://cellengine.com/
STAR (v2.7.5c)	Dobin et al 81	https://github.com/alexdobin/STAR
Picard Tools (v2.23.3)	Broad institute	http://broadinstitute.github.io/picard/
Genome Analysis Tool Kit (v4.0.11.0)	Van der Auwera et al 82 DePristo et al 83	https://github.com/broadinstitute/gatk
cellranger (v6.0.2)	10X Genomics	https://www.10xgenomics.com/support/software/cell-ranger/latest
Seurat (v4.0.3)	Cao et al 84	https://github.com/satijalab/seurat
freemuxlet (vAug2021)	Hartoularos et al 104	https://github.com/statgen/popscle
bcftools (v1.10.2)	Li et al 86	https://github.com/samtools/bcftools
DoubletFinder (v2.0)	McGinnis et al 87	https://github.com/chris-mcginnis-ucsf/DoubletFinder
Harmony (v1.0)	Korsunsky et al 88	https://github.com/immunogenomics/harmony
DSB (v0.3.0)	Mulè et al 89	https://github.com/niaid/dsb
ComBat, sva (v3.40.0)	Johnson et al 90	https://www.bioconductor.org/packages/release/bioc/html/sva.html
minimum hypergeometric (mHG) (v1.1)	Eden et al 91	https://cran.r-project.org/web/packages/mHG/index.html
limma (v3.48.3)	Ritchie et al 94	https://bioconductor.org/packages/release/bioc/html/limma.html
Azimuth (v0.4.6)	Hao et al 105	https://github.com/satijalab/azimuth
Mario	Zhu et al 106	https://github.com/shuxiaoc/mario-py
MOFA (v1.0.1)	Argelaguet et al 31	https://github.com/bioFAM/MOFA2
MetaWRAP (v1.3)	Uritsky et al 96	https://github.com/bxlab/metaWRAP
MEGAHIT (v1.2.9)	Li et al 97	https://github.com/voutcn/megahit
CheckM2 (v1.1.0)	Chklovski et al 98	https://github.com/chklovski/CheckM2
dRep (v3.6.2)	Olm et al 99	https://github.com/MrOlm/drep
Minimap2 (v2.30)	Li et al 100	https://github.com/lh3/minimap2
CoverM (v0.7.0)	Aroney et al 101	https://github.com/wwood/CoverM
GTDB-Tk (v2.5.2)	Chaumeil et al 102	https://github.com/Ecogenomics/GTDBTk
WGCNA (v1.73)	Langfelder et al 103	https://github.com/cran/WGCNA

Biospecimens collection and processing

Baseline timepoint

Biological specimens from the baseline timepoint were collected between October 21, 2019, and January 17, 2020. Stool samples were collected by study participants at the research facility within a 6h window for all participants. The samples were transferred to the research team and put on ice before processing. The samples were manually homogenized and aliquoted under a biosafety cabinet, and cryopreserved at −80°C.

Blood was collected in EDTA tubes within a 90 min window for all participants of the study. Plasma samples were obtained by centrifugation of whole blood samples at 2000g for 15 minutes at 4°C and aliquots were cryopreserved at −80 degree celsius. Peripheral blood mononuclear cells (PBMC) were obtained by density gradient centrifugation of whole blood at 800g for 30 min at room temperature without brake, using Ficoll-Paque (Cytiva). Isolated cells were either further processed for mass cytometry preparation or cryopreserved in a solution containing fetal Bovine Serum and 10% DMSO (Sigma) for further processing in CITE-Seq preparation.

Follow-up timepoint

The stool and blood samples from the subset of the ImmunoMicrobiome cohort with a follow-up timepoint sampled (n=48) were processed and analyzed in the same manner as the baseline data as described above. To reduce the technical noise, we reanalyzed different aliquots of baseline samples with the follow-up samples for each assay.

Survey data collection and preprocessing

Survey data was collected through self-administered questionnaires. Short term dietary data was collected through the NIH validated Automated Self-Administered 24-Hour (ASA24^®) Dietary Assessment Tool (https://epi.grants.cancer.gov/asa24). General health, diet, medical history and demographic data was collected through a questionnaire designed by the research team.

The nutrient levels from ASA24 survey were collected, features with zero values in all samples were removed, and the data was log-transformed and scaled for the integrative analysis.

METHOD DETAILS

Microbiome profiling

Whole metagenome shotgun sequencing DNA extraction

Frozen stool samples were thawed for bacteria lysis and DNA extraction performed using a combination of the cetyltrimethylammonium bromide (CTAB) chemical, phenol/chloroform solvents and bead beating mechanical forces. Briefly, 200 mg of human stool samples were incubated at 65°C for 15 minutes in a 5% CTAB solution. A phenol/chloroform solution was added, and the samples were homogenized with bead beating for 2 rounds of 30 seconds in a Fastprep24 homogenizer (MP Biomedicals) using Lysing Matrix E beads (MP Biomedicals). The aqueous phase was collected after centrifugation and mixed with chloroform before centrifugation. Supernatant was then mixed with a 30% PEG solution and incubated overnight at 4°C. Precipitated DNA was collected after 2 rounds of washing in ice cold 70% Ethanol. Magnetic bead-based DNA cleanup was then performed using AMPure XP beads (Beckman Coulter).

Whole metagenome shotgun library construction and sequencing

Whole metagenome shotgun sequencing libraries were generated using the Nextera XT (Illumina). Custom 12-bp dual unique indices were introduced during PCR amplification. Libraries were pooled at the desired relative molar ratios and DNA cleanup was performed using AMPure XP beads (Beckman Coulter) for buffer removal and library size selection. Sequencing reads were generated using a NovaSeq S4 flow cell.

Metabolome profiling

Targeted and untargeted metabolomics sample preparation

Frozen stool samples were lyophilized over 48 hours, crushed into a fine powder, 25 milligrams of powder was weighed out per study participant into Precellys tissue homogenizing tube, and samples were stored at −80°C. Metabolites were extracted from each biospecimen by addition of 4 volumes of extraction solvent (1:1 acetonitrile and methanol with 5% LCMS grade water). Mixture was vortexed for 1 minute then incubated at −20°C for 30 minutes. Upon completion of mixing and incubation, samples were centrifuged at 15,000g for 30 minutes at 4°C. Supernatants were extracted and stored at −20°C until time of analysis.

Very short chain fatty acids targeted metabolomics

Samples were extracted using a biphasic extraction using 0.5 mL H20 with Internal standards, 0.1mL concentrated HCL and 1 mL MTBE. The samples were shaken for 30 min at room temperature then centrifuged for 2 min at 14000 rcf. 100 uL of the organic phase was aliquoted into a glass crimp vial with 25 uL of MTBSTFA. Vials were crimp caped and shaken at 80°C for 30 min. After derivatization, vials were placed in GC sampler for analysis. 1 uL of derivatized sample was injected using a split method at an inlet temperature of 250°C. A constant flow of 1.2 mL/min using Helium as used throughout the run. The oven temperature program started at 50°C for 0.5 min then ramped to 70°C at 5°C/min holding for 3.5 min, then ramped to 120°C at a rate of 10°C/min, and finally ramped to 290°C at a rate of 35°C/min holding for 3 min for a final run time of 20.857 minutes. The transfer line was set to 290C while the EI ion source was set to 250C. The Mass spec parameters collected data from 85m/z to 5m/z at an acquisition rate of 17 spectra/sec. A 6-point curve with internal standards was injected alongside the samples as well as a pool sample used as QC which is injected every ten samples. Mass hunter Quant was used for data processing and picking of target peaks, then normalized to the amount of sample injected.

Primary Metabolism untargeted metabolomics

Metabolomics were performed at the UC Davis West Coast Metabolomics Center. Samples were extracted using 1mL of 3:3:2 ACN:IPA:H2O (v/v/v). Half of the sample was dried to completeness and then derivatized using 10 uL of 40 mg/mL of Methoxyamine in pyridine. They were shaken at 30°C for 1.5 hours. Then 91 uL of MSTFA + FAMEs was added to each sample, and they were shaken at 37°C for 0.5 hours to finish derivatization. Samples were then vialed, capped, and injected onto a 7890A GC coupled with a LECO TOF. 0.5 uL of derivatized sample was injected using a splitless method onto a RESTEK RTX-5SIL MS column with an Intergra-Guard at 275C with a helium flow of 1 mL/min. The GC oven was set to hold at 50°C for 1 min then ramp to 20°C/min to 330°C and then held for 5 min. The transfer line was set to 280°C while the EI ion source was set to 250°C. Data were collected from 85m/z to 500m/z at an acquisition rate of 17 spectra/sec.

Biogenic amines untargeted metabolomics

Untargeted metabolomics via hydrophilic interaction liquid chromatography triple time of flight TTOF mass spectrometry (HILIC-TTOF MS) was performed at the UC Davis West Coast Metabolomics Center. Metabolite profiling using HILIC-TTOF-MS was performed on the Agilent 1290 UHPLC/Sciex TripleTOF 6600 mass spectrometer under positive and negative ionization modes. Metabolites (injection volume 2 μL) were separated using a Waters Acquity UPLC BEH Amide column (1.7 μm, 2.1 × 50 mm), and a binary mobile phase consisted of 100% LC-MS grade H2O with 10 mM ammonium formate and 0.125% formic acid as solvent A and 95:5 (v/v) ACN:H2O with 10 mM ammonium formate with 0.125% formic acid as solvent B. The mobile phase was running under gradient conditions with a flow rate of 0.8 mL/min. The column temperature was kept at 45 °C. Data were acquired in data-dependent acquisition mode with a mass range 50–1500 m/z for MS1 and 40–1000 m/z for MS2.

Immune profiling by CyTOF

Mass-tag antibody conjugation and antibody cocktail preparation for mass cytometry

Mass cytometry antibodies were prepared using the MaxPAR antibody conjugation kit (Standard Bio Tools) according to the manufacturer’s instructions. Following conjugation, antibodies were diluted at concentrations ranging from 0.2 and 6 mg/mL in Candor PBS Antibody Stabilization solution (Candor Bioscience GmbH, Wangen, Germany) supplemented with 0.02% NaN3 and stored at 4°C. Each antibody was titrated to identify optimal staining concentrations using human PBMCs. A surface staining antibody cocktail and an intracellular antibody cocktail made of all the 2 types of target were prepared for the staining of all batches of the study. A summary of all mass cytometry antibodies, metal tags and concentrations used for analysis can be found in the STAR methods.

Mass cytometry sample preparation

Briefly, PBMCs were incubated immediately after isolation with a viability staining solution containing 50uM of cisplatin (Sigma-Aldrich) for 1 min at room temperature and fixed in a solution containing 1.6% Paraformaldehyde (PFA) (Electron Microscopy Science) and cryopreserved in a solution containing 0.5% BSA and 10% dimethyl sulfoxide (DMSO). Fixed cells were later thawed and incubated with benzonase nuclease (Sigma-Aldrich), permeabilized with a solution containing 0.2% saponin, barcoded by incubation with a combination of custom palladium metal stable isotopes for 15 minutes at room temperature, and pulled in batches of 20 samples. Fc receptors were blocked using Human TruStain FcX (Biolegend) and cells were stained with custom metal ion-labeled antibody cocktails. Extracellular staining was performed at room temperature for 30 minutes on a shaker set on 90 RPM. Intracellular staining was then performed using eBioscience Permeabilization Buffer (Invitrogen) at 4°C for 60 minutes on a shaker set on 90 RPM. Cells were later incubated with Cell-ID Intercalator-Ir (Standard Biotools) at 0.0625 uM and 4% PFA overnight at 4°C. Before running, cells were resuspended in cell acquisition solution (Standard Biotools) that contained EQ Four Element Calibration Beads (Standard Biotools). Except for permeabilization steps and final steps, washes were performed using a cell staining solution containing PBS,BSA and EDTA or PBS.

Mass cytometry data acquisition

Mass cytometry sample pools were diluted in Cell Acquisition Solution containing bead standards to approximately 1 × 10^ 6 cells/mL and then analyzed on a Helios mass cytometer (Standard Biotools) equilibrated with Cell Acquisition Solution. A minimum of 10 × 10^ 6 cell events were collected for each barcoded set of samples at an event rate of 400 events/second.

Immune profiling by CITE-seq

Processing of PBMCs for cellular indexing of transcriptomes and epitopes by sequencing (CITE-seq)

Cryopreserved PBMCs were thawed. Cell counts and viability was determined using the CellacaMX. PBMCs were pooled in batches of sixteen samples in equal numbers to a total of 1 million cells. Pools were resuspended in Cell Staining Buffer (Biolegend) and incubated with Fc Receptor Blocking Solution (Biolegend) at 20mg/mL for 10 minutes on ice. Cells were then stained with a cocktail of 140 TotalSeq-C oligo-tagged antibodies (Biolegend) for 30 minutes at 4°C. Stained cells were washed in PBS containing 1% BSA, filtered through a 70uM filter, and re-suspended in a final solution of PBS containing 0.04% BSA. Cells (n=60,000) were loaded on the Chromium Controller (10X Genomics) for GEM generation. The cDNA libraries were generated using the Chromium Next GEM Single Cell 5’ v1.1 Library Kit for gene expression (GEX) and the Chromium Single Cell 5’ Feature Barcode Library kit for antibody derived tag (ADT) according to manufacturer’s instructions (10X Genomics). Each batch of sixteen samples was sampled four times to prepare four independent cDNA libraries per batch to obtain a sufficient number of cells per sample. The libraries were subsequently sequenced on a NovaSeq 6000 S4 platform (Illumina).

Bulk RNA sequencing

Cryopreserved PBMCs were thawed. Cell counts and viability were determined using the CellacaMX. 0.5 million cells were used for each sample. RNA was extracted using the Quick-RNA MagBead kit (Zymo Research) on the KingFisher automated extraction and purification system (ThermoFisher) according to manufacturer’s instructions. Quality and quantification of extracted RNA was assessed on the 5300 Fragment Analyzer System (Agilent). Next, cDNA libraries were generated using the Universal Plus mRNA-Seq with NuQuant kit according to manufacturer’s instructions (Tecan). The libraries were subsequently sequenced on a HiSeq 4000 platform (Illumina).

Serological profiling by Olink

Sample preparation and serological profiling

Plasma samples were obtained by centrifugation of whole blood samples at 2000g for 15 minutes at 4°C and aliquots were cryopreserved at −80°C. Serological profiling data was generated by Proximity Extension Assay technology using oligonucleotide labeled antibody with the Target 96 Inflammation Panels (Olink Proteomics). The protein abundance was quantified as Normalized Protein Expression (NPX) values, an arbitrary unit on the log2 scale. Four internal controls were added to each sample, which were used to evaluate the quality of each sample.

QUANTIFICATION AND STATISTICAL ANALYSIS

Details on the statistical testing / modeling are listed below and values reported associated with each figure are reported in the figure legend or results section text. Statistical analyses described below were primarily carried out in R. The p-values were adjusted for multiple-testing using Benjamini-Hochberg procedure, unless mentioned otherwise. The correlation analysis was performed using Spearman’s rank method unless indicated otherwise. Significance of differences in pairwise comparisons was performed using Wilcoxon rank sum test and in multi-group comparisons using Kruskal-Wallis test. Linear regression models were used to find features associated with MOFA factors. Correlations were visualized by a line using the method lm in ggplot2 geom_smooth but correlations were calculated using the function cor.test. Significance was defined using multiple-test corrected p-values smaller than 0.1 (false discovery rate (FDR) 10%), unless mentioned otherwise.

Microbiome data analysis

Whole metagenome shotgun data preprocessing

Sequencing reads were processed with a custom pipeline. Briefly, reads were quality-filtered and adapters were removed using FastP v0.23.2⁷⁶ with standard parameters. Human reads were filtered out of the dataset using the computational tool BMTagger (ftp://ftp.ncbi.nlm.nih.gov/pub/agarwala/bmtagger/). BioBakery tools Metaphlan version 3.1 was used to estimate taxa community and HUMAnN3⁷⁷ was used to estimate gene family/pathway abundances (units of length-normalized average copy numbers of genes summarized at the gene family or pathway level). For MOFA analysis, the species abundance data from MetaPhlan3 was normalized by the total abundance per sample (after removing ‘UNKNOWN’ category), the species with non-zero values in more than 10 samples were retained, and the data was centered log-ratio (CLR) transformed and scaled for the integrative analysis. Gene family and pathway abundances were normalized to average copy numbers (reads per kilobase per genome equivalent, RPKG) using MicrobeCensus v1.1.1⁹⁵ and samples failing this step due to no hits to marker proteins were excluded. For MOFA analysis, the pathway abundance data from HUMAnN3, the features with non-zero values in more than 10 samples were retained, and the data was log-transformed and scaled for the integrative analysis. For PCA analysis, when unnormalized microbiome composition data was used to examine factors’ association with the microbiome, abundances of the most prevalent species across the cohort were included as an input to mitigate the effect of species sparsity on PCA.

Species level contribution to gene family and pathway abundances

Species-level contributions to gene family abundances were calculated on the species-stratified output of HUMAnN3. The prevalence represents the percentage of the cohort that carried a non-zero copy number for a given gene family in a given species. Thus, species “contribution to a gene family” is the median across the cohort of the ratio of the copy number of genes for a given species over the total copy number of genes for all species (RPKG fraction). The same metrics were calculated for Species-level contributions to pathway abundances, using gene family copy numbers summarized at the pathways level. Species were filtered based on their median contribution and prevalence to identify major contributors, with thresholds adapted to the data distribution. Major contributors to pathway-level genomic content were defined as having a prevalence across the cohort >25% and a median contribution to the total gene copy number >0.05. Major contributors to gene families-level genomic content were defined as the top 20 species with highest median contribution to the total gene copy number, among species with median contribution >0.05 and gene family prevalence across the cohort >10%. All contributions, including from minor and unclassified species are documented in the Table S7.

Genome-resolved analysis and de novo metagenome assembly

De novo metagenome-assembled genomes (MAGs) were reconstructed using a custom pipeline (Fig. S6L). Individual samples of human-depleted reads were assembled and binned using MetaWRAP⁹⁶ (v1.3). Specifically, metagenomes were assembled using MEGAHIT⁹⁷ (v1.2.9) and binned using the consensus of three binning algorithms wrapped in MEGAHIT (CONCOCT, metaBAT2, and MaxBin2). MAG quality was assessed with CheckM2⁹⁸ (v1.1.0), with those having ≥75% completeness and ≤ 25% contamination retained. These high-quality MAGs were then de-replicated with dRep⁹⁹ (v3.6.2) at 98% average nucleotide identity (ANI) to generate a nonredundant species genome bin (SGB) catalog. Reads from each sample’s metagenome were mapped to the SGB database using Minimap2¹⁰⁰ (v2.30), and relative abundances were quantified with CoverM¹⁰¹ (v0.7.0). SGBs were annotated using GTDB-Tk¹⁰² (v2.5.2), which assigned each SGB within the Genome Taxonomy Database (GTDB). SGB co-abundance patterns and correlations with HUMAnN-derived pathways and metabolites were assessed using Spearman’s correlation, and associations with Factor 3/IMCV scores were evaluated by linear regression with FDR correction by Benjamini-Hochberg. For co-abundance testing, module membership was quantified using the signedKME() function from WGCNA¹⁰³ (v1.73).

Metabolome data analysis

Metabolomics data preprocessing

The data from targeted metabolomics for very short chain fatty acids, and untargeted metabolomics for biogenic amines and primary metabolites were concatenated. The features with unknown annotation were discarded and duplicate features based on the same name or same InChiKey were deduplicated by selecting the feature with lowest variation across the batch-control samples. The data was log-transformed and pareto scaled (scaled data multiplied by square-root of standard deviation) for the integrative analysis.

CyTOF data analysis

Mass cytometry data normalization and de-barcoding

Bead standard data normalization and de-barcoding of the pooled samples into their respective conditions was performed using the R package from the PICI institute available at https://github.com/ParkerICI/premessa.

Mass cytometry data batch correction, clustering, annotation, visualization, and differential protein expression analysis

The mass cytometry data for events corresponding to live cells were extracted from bead-normalized FCS files, arcsinh-transformed using a cofactor of 5, and subjected to batch-correction using cyCombine ⁷⁸ (v0.1.5). The dimensionality reduction and unsupervised clustering analysis were performed to obtain 225 clusters of single cells using cyclone ⁷⁹, which uses FlowSOM ⁸⁰ (v2.0.0) for clustering, and uwot R package (v 0.1.10) for dimensionality reduction. The clusters were annotated manually using known marker expression patterns, and were visualized using SCAFFoLD ²⁹ (v0.1; default parameters ). For the landmark populations used as input to SCAFFoLD, an equal number of events from all batch-corrected FCS files were combined to produce a concatenated dataset of 100k cells, which was manually gated on CellEngine to produce FCS files of populations of known phenotype. The SCAFFoLD map validated the marker-based manual annotation of individual clusters. The cluster frequencies were normalized by the total number of cells per sample, log-transformed, and scaled for the integrative analysis.

Immune populations were annotated based on their expression of the following marker. B cells were CD19+ and CD27 was used to identify naive and memory subsets. T cells were CD3+. Within the T cells, CD56 was used to identify NKT cells and gdTCR was used to distinguish gdT cells and abT cells. Within abT cells, CD4 and CD8 were used to identify lineages, CCR7 and CD45RA were used to identify naive and memory cell subsets. NK cells were CD3− HLA-DR low/− and CD56+. NK subsets were defined using CD16. Myeloid cells were CD3− CD19− CD56− CD123− HLA-DR+. Within the myeloid, monocyte subsets were identified by CD14 and CD16, conventional dendritic cells were CD14− and CD16− and subsets were defined using BDCA3 CADM1 Clec9A BDCA1 SIRPa and Axl. pDCs were HLA-DR+ and CD123+ and Axl was used to define subsets. Residuals neutrophils were identified as CD15+ CD16+ and residuals basophils as HLA-DRlow/− and FcER1A+ and CD123 int.

The differential protein expression analysis was performed using Kruskal-Wallis test to identify markers that are different between groups of CyTOF clusters. The clusters were grouped into two to three groups based on positive or negative association or no association with a MOFA factor (adj. p <0.1 from linear models - see below). Through 10 iterations, 100 cells were randomly drawn from each of the cluster groups, and the normalized protein expression was compared between the groups using Kruskal-Wallis test. The markers that were significantly different (p <0.00001) in all 10 iterations were identified as significantly differentially expressed proteins and are presented in Fig. 5B, 5F, and S5E). Pairwise comparisons were performed as a post-hoc test to identify pairs of cluster groups with differential expression and a significance score was calculated based on (−log10)p-values, scaled to the highest value.

CITE-seq data analysis

Sequence alignment and generating raw count matrix

The raw sequencing reads from gene expression and surface protein expression libraries were aligned to human genome reference (vGRCh38–2020-A) with annotations from Gencode (v32-primary assembly) and to the TotalSeq-C antibody reference panel (BioRad), respectively, using cellranger (v6.0.2) with default options.

Bulk RNA-seq data analysis and variant calling for demultiplexing of CITE-seq libraries

Sequencing reads were aligned to the human reference genome and Ensembl annotation (GRCh38 genome build, version 95) using STAR (v2.7.5c)⁸¹ with the following parameters: --outFilterType BySJout --outFilterMismatchNoverLmax 0.04 -outFilterMismatchNmax 999 --alignSJDBoverhangMin 1 --outFilterMultimapNmax 1 -alignIntronMin 20 --alignIntronMax 1000000 --alignMatesGapMax 1000000. Duplicate reads were removed using Picard Tools (v2.23.3) (http://broadinstitute.github.io/picard/). Nucleotide variants were identified from the resulting bam files using the Genome Analysis Tool Kit (v4.0.11.0) following the best practices for RNA-seq variant calling ^82,83. This included splitting spliced reads, calling variants with HaplotypeCaller (added parameters: --dont-use-soft-clipped-bases -standcall-conf 20.0), and filtering variants with VariantFiltration (added parameters: -window 35 cluster 3 –filter-name FS -filter FS >30.0 --filter-name QD -filter QD <2.0). Variants were further filtered to include a list of high quality SNP for identification of the subject of origin of individual cells by removing all novel variants, maintaining only biallelic variants with MAF greater than 5%, a max missing of one individual with a missing variant call at a specific site and requiring a minimum depth of two (parameters: --max-missing 1.0 --min-alleles 2 --max-alleles 2 --removeindels --snps snp.list.txt --min-meanDP 2 --maf 0.05 --recode --recode--INFO-all -out).

CITE-seq data demultiplexing and quality control

The raw counts of gene expression from cellranger output were imported and analyzed using Seurat⁸⁴ (v4.0.3). Because each library of 10X data contained cells from sixteen samples, we first performed demultiplexing of the data based on SNP profiles to map individual cells to its sample of origin. Specifically, freemuxlet¹⁰⁴ (vAug2021) was used to cluster the cell barcodes based on concordant SNP profiles (list of high-quality reference SNPs obtained from the 1000 ⁸⁵Genomes Consortium; ), and the SNP profiles of freemuxlet clusters were mapped to the genotypes of study participants using bcftools ⁸⁶(v1.10.2), which was generated from the bulk RNA-seq data, allowing us to map each cell barcode to the sample of origin. The inter-sample doublets (DBL: barcodes containing mixture of multiple SNP profiles) and empty droplets (AMB: barcodes lacking sufficient coverage of reference SNPs for freemuxlet clustering) were removed, and the barcodes with SNG calls (droplets containing a single cell or multiple cells from a same individual (intra-sample doublets)) were subjected to further filtering. DoubletFinder was used⁸⁷ (v2.0) to filter barcodes containing heterotypic intra-sample doublets (i.e. the droplets containing multiple cells of different cell-types from a same participant) as described here https://github.com/chris-mcginnis-ucsf/DoubletFinder. Briefly, first the data was transformed using SCTransform, principal component analysis (PCA) was performed using npcs=35, nearest neighbors were computed using FindNeighbors(), and the data was clustered using FindClusters() in Seurat. The PC neighborhood size (pK) was adjusted for each library separately by performing a pK-pN parameter sweep using paramSweep_v3() of DoubletFinder package. The proportion of expected homotypic doublets were estimated based on cluster assignments from FindClusters() output, and the number of expected heterotypic doublets were determined using the inter-sample doublets (DBL calls from freemuxlet) by dividing the fraction of inter-sample doublets by the number of all possible combinations of inter-sample doublets, otherwise the default options were used. The barcodes annotated as heterotypic doublets using DoubletFinder were discarded and the remaining barcodes representing single-cells were subjected to further quality control. The barcodes containing less than 100 genes and genes that are detected in less than 3 cells were filtered out. We further filtered out the barcodes based on library-specific cutoffs for the number of genes per cell (nFeature_RNA <300–800 & >6,000), the number of unique molecular indexes (UMIs) per cell (nCount_RNA <400–2,000 & >40,000), mitochondrial content (>15%), and ribosomal content (<8–15% & >60%) to remove poor quality cells. The resulting high-quality cell-barcodes were used for further analysis.

Normalization, dimensionality reduction, clustering, and cluster annotation

The gene expression data from all libraries were merged, log-normalized, and scaled with regressing out for RNA counts per cell, mitochondrial and ribosomal contents, and cell-cycle states. Top 2,000 most variable features were found using the ‘vst’ method, PCA was performed using npcs=50, and batch-effect (library-specific bias) was normalized using Harmony ⁸⁸ (v1.0;).

The ADT data was normalized and denoised using ‘denoised and scaled by background’ (dsb) method (⁸⁹) for each library separately. The dsb utilizes protein-specific background signal from empty droplets and signal from isotype control antibodies to accurately estimate and correct for cell-to-cell technical noise in each cell. The ADT data was normalized and denoised using cell barcodes containing less than 100 RNA counts and more than 10 protein counts as empty droplets and expression of five isotype antibodies included in the TotalSeq-C panel as isotype-control signal in DSBNormalizeProtein() from dsb package⁸⁹ (v0.3.0). The normalized values smaller than −10 were truncated by substituting them with zero. The expression data for the isotype control antibodies were excluded from the further analysis. The protein expression data demonstrated the largest batch-effect; therefore, we used reciprocal PCA (RPCA) to integrate the protein data. Briefly, after scaling the data and calculating PCs for each library, integration anchors were calculated using the first library from each batch as reference, and then the data was integrated using IntegrateData() function in Seurat. The integrated protein data was scaled, the ADT counts per cell and cell-cycle states were regressed out, and PCA was performed using npcs=30.

Weighted nearest neighbor (WNN) graph was constructed using first 30 harmony dimensions for gene expression, first 18 integrated PCs for protein expression, and prune.SNN = 1/20. Using the WNN graph, dimensionality reduction was performed using Uniform Manifold Approximation and Projection (UMAP) and 56 clusters of cells were identified using smart local moving (SLM) algorithm and resolution=2 using Seurat’s FindClusters function. Marker genes and proteins were identified for each cluster using Model-based Analysis of Single Cell Transcriptomics (MAST) algorithm (https://github.com/RGLab/MAST/) with library as a latent variable. The clusters were assigned cell type annotation using the expression of canonical marker genes and proteins as depicted in Fig. S1A. A representative sample of the data was used for UMAP coordinate visualizations in Fig. 1E and S1B.

Gene expression variance in the ImmunoMicrobiome cohort, the single-cell data for cell types were isolated, and subjected to PCA, and batch-correction using harmony. The top ten dimensions, capturing the major axes of variation, were selected, the dimension coordinates of single-cells were averaged per participant, and these averages were depicted in Fig. 1F for monocytes. Linear models were used to identify the genes associated with averaged Dimension 1 coordinates.

Preparing data for integrative analysis

For each cell type, we used pseudo-bulked gene expression per sample in the integrative analysis. The cell types that contained more than 20 cells per sample in more than 10 samples were kept for further analysis. The normalized single cell gene expression was averaged (in linear space) across samples to generate a sample-by-gene matrix for each cell type. To remove the lowly expressed genes, the genes that were detected in more than 30% of samples were retained. The pseudo-bulked expression data was batch-corrected using ComBat from sva package⁹⁰ (v3.40.0), the residual batch-effects were eliminated by removing genes associated with batches after ComBat (Kruskal-Wallis adjusted p-value <0.01), and the depletion of batch-effects were validated using principal variance component analysis (PVCA). The top 5,000 most variable genes identified using median absolute deviation (MAD) were selected, and the expression data for these genes were log-transformed and scaled for the integrative analysis.

The cell type frequencies were normalized by the total number of cells per sample, log-transformed, and scaled for the integrative analysis.

The ADT data was also filtered in the same manner as gene expression data and was batch-corrected using ComBat. This data was used for follow-up analysis of MOFA factors.

Enrichment analysis

Enrichment analysis was performed using minimum hypergeometric tests (mHG)⁹¹ to identify hallmark signatures⁹² that are significantly enriched (BH-adjusted p <0.1) in the significantly associated features (BH-adjusted p <0.1 and |feature weight| >0.1; see Evaluating composition of latent factors) for each factor in each dataset. The features were ranked based on factor weights for the enrichment analysis. The selection of an appropriate background set of features is crucial for the accurate enrichment analysis, therefore the features that were detected in our assays were used as background in the mHG test. To calculate factor-specific signature scores, the log-transformed and scaled expression of core genes (genes driving the mHG enrichment) were weighted by MOFA weights and summed for each participant (used in Fig. 4A). The importance of signatures within each factor was evaluated by combining enrichment p-values of each signature across all datasets using the Cauchy combination test, which is known to be robust to the dependence structure underlying p-values to be combined ⁹³. The joint p-values were BH-adjusted to obtain joint adj.p. Several hallmark signatures were significantly enriched in most factors. We noticed that the first five factors were associated with similar immune signatures. To select the hallmark signatures for the follow up investigation, we took the top 5 most significant signatures (based on the joint adj.p) for each of the first five factors and counted the number of CITE-Seq cell subsets they were significantly enriched in (Fig. S2E). Based on this analysis, we selected these hallmark signatures to characterize the factors: Inflammation-related gene sets - inflammation response (M5932), TNF alpha signaling via NFKB (M5890) and TGF beta signaling (M5896); Interferon-related gene sets - IFN alpha response (M5911) and IFN gamma response (M5913). Given large overlap in features between these hallmark signatures, beyond the initial analysis in Fig. 1–2 overlapping genes between IFN-related and inflammation-related signatures were excluded from the gene sets (Fig. 3–7). Additionally, gene sets related to tonic-IFN ²⁷ and, prolonged-IFN ³³ responses were derived from previous publications.

For comparison between high-IMCV and high-IIV sub-cohorts, limma ⁹⁴was used to identify genes significantly differentially expressed between the two groups (BH-adjusted p <0.1). Enrichment analysis was performed using mHG to identify the gene signatures used in the study that are significantly enriched (BH-adjusted p <0.1) in high-IMCV or high-IIV sub-cohorts in each dataset. The features were ranked based on log2 fold-difference between high-IMCV and high-IIV sub-cohorts for the enrichment analysis. To calculate the composite signature scores for each individual, the log-transformed and scaled expression of the core genes driving the enrichment of a given signature were averaged using median for each individual.

Olink data analysis

Samples that failed quality control (samples with internal controls deviating more than 0.3 NPX from median) were excluded. Features with NPX below limit of detection in more than 60% of samples were discarded. The remaining values below the limit of detection were imputed by half of the minimum, and the full dataset was scaled. While the serological data were not included in integrative analysis, linear regression models were used to identify features significantly associated with IMCV and IIV (BH-adjusted p <0.1 and |coefficient| >0.1). The coefficients from the linear models were used as feature weights in the calculation of weighted expression and composite scores used in Fig. 7C–D.

Multi-omic integrative analysis using MOFA

Building MOFA model

As described above, all omics datasets were standardized by log-transformation and scaling (by calculating Z-scores). The Multi-omics Factor Analysis (MOFA) ³¹ (v1.0.1) was used to extract the shared variation across a total of 42 datasets (an ASA24 data, a metabolomics dataset, two microbiome datasets, two immune frequency datasets, and 36 gene expression datasets). MOFA performs matrix decomposition to represent the multi-omics data in reduced dimensions as represented by latent factors that capture biological and technical sources of variation. Our evaluation of MOFA models with different factor numbers revealed that the 15-factor model captured a substantial amount of shared variation between immune and microbiome compartments. Despite the incorporation of batch-correction at multiple stages in the preprocessing of CITE-seq data, we observed that two original factors (10 and 15) remained significantly associated with CITE-seq batches (Kruskal-Wallis test p <0.05), and therefore, were excluded from further evaluation. The remaining factors were renumbered from 1 to 13.

Evaluating composition of latent factors

To evaluate the composition of factors, the multi-omics features significantly associated with factors were identified using linear regression models (BH-adjusted p <0.1 and |feature weight| >0.1), while controlling for experimental batches, if available.

Analysis of data from the follow-up timepoint

The datasets from follow-up timepoints were processed in the same manner as the original datasets with following modifications.

CITE-seq

After generating single-cell count data from sequencing reads and performing quality control, the identification of cell subsets in the follow-up data was performed using Azimuth¹⁰⁵ (v0.4.6). The original Seurat object was converted to Azimuth reference, and the follow-up dataset was mapped onto it using Azimuth, which provided predicted cell subset labels for single cells.

CyTOF

Following the quality control of CyTOF data, the single cells were annotated by mapping them to the original CyTOF data using Mario¹⁰⁶. A subset of single cells from the original data (250k cells) was used as a reference in Mario mapping. Mario predicted closest matching cells from the reference for each cell, and the cluster identity of the matching reference cell was used as the cell subset annotation for single cells in the follow-up dataset.

Comparison between timepoints

For comparisons at the feature-level, the feature levels were weighed by multiplying the normalized feature levels with IMCV feature weights from the original MOFA analysis and these quantities were compared between baseline and follow-up timepoints. For comparing composite scores, the weighted levels of features significantly associated with IMCV were aggregated by summation and these composite scores were compared between timepoints and presented as scatter plots with linear regression lines. The correlations were quantified using Spearman’s correlation and p-values.

For comparison in multi-omic space, the weighted levels of features significantly associated with IMCV across datasets were combined in a single matrix of features by samples and the Uniform Manifold Approximation and Projection (UMAP) was applied to this data in R using default parameters. The UMAP coordinates of paired samples were presented in a scatter plot with paired samples connected by a line.

Supplementary Material

1

Figure S1. Immune and microbiome variation in a cohort of healthy individuals, related to Figure 1

A. Dot plot quantifying the prevalence and expression of key marker genes and surface proteins used to annotate CITE-Seq clusters. Marker names are colored based on the library source. Gene expression library markers are annotated in blue. Surface protein expression markers from the antibody derived tag library are indicated in green.

B. UMAP visualization of CITE-Seq data. Single cell mapping and coloring is the same as Fig. 1E. Annotation details unaggregated cell subsets identified by unsupervised clustering.

C. Phylogenetic tree of species identified in the ImmunoMicrobiome cohort. Colors represent major microbiome phyla.

D. Line plots describing microbiome diversity across the cohort.

E. Scatter plot showing the distribution of the study participants based on their microbiome composition at the species level after dimensionality reduction using PCoA with Bray-Curtis dissimilarity distance metric.

2

Figure S2. Integrative multi-omic analysis identifies microbiome-associated immune variation in interferon and inflammation responses, related to Figure 2

A. Flowchart detailing data preprocessing for each modality for multi-omic integration.

B. Schematic of integrative approach. Features that vary concomitantly across datasets in the cohort are captured by factors. For each factor, features are ordered based on their weight on the factor (feature weight) and study participants are ordered based on factor values (factor score). In the schematic, features driving the variation captured by factors are indicated with red and blue boxes.

C-D. Scatter plots of factor scores and first three principal components from the microbiome species relative abundance data (Microbiome PC1–3) and CITE-Seq gene expression data (Immune PC1–3). Factor 3 is shown in panel C and Factor 1 in panel D.

E. Heatmap quantifying the number of CITE-seq cell subsets that have significant enrichment of signatures among the first five MOFA factors. Signatures were evaluated using mHG test and hallmark gene set library.

F. Associations between each MOFA factor and either sex or age. Statistics are the result of Spearman correlations adjusted for multiple hypothesis testing by the Benjamini-Hochberg method. Significant associations (FDR <0.1) are marked by an asterisk (*).

3

Figure S3. IMCV captures differences in the interferon response across the cohort, related to Figure 3

A. Schematic of the analytical approach used in Figure 3. Enrichment analysis is performed within MOFA factors, to assess association between factor scores and transcriptional signatures of interest. For signatures enriched, differential gene expression comparing study participants with high factor scores to those with low factor scores is used to evaluate genes driving the signatures.

B. Volcano plot of genes differentially expressed between study participants with high and low factor scores in the tonic-IFN response signature. Study participants with top and bottom 20 factor scores for IMCV are compared. Wilcoxon rank-sum test and BH correction for multiple testing were used to assess the significance of the differences. Most DEGs (adj.p <10⁻⁴ and |log2 fold difference| >0.1) are annotated and colored by cell subset.

C. Bar chart comparing the association of IIV and IMCV with Prolonged-IFN response transcriptional signatures across cell subsets. mHG was used to test enrichment on CITE-seq gene expression data. Signed −log10(adj.p) for each factor is shown.

D. Heatmap quantifying plasma cytokines relationship with IMCV of the tonic-IFN response across cell subsets. Spearman rank correlation was used to evaluate the associations of gene expression in the signatures with plasma levels of circulating cytokines. Correlations with |r|>0.25 are indicated with an asterisk.

E. Schematic of the cohort and the analytical approach performed to compare IMCV-high and IIV-high sub-cohorts. Genes differentially expressed between high-IMCV and high-IIV sub-cohorts were identified using limma, ranked using limma’s log2 fold-difference and used for enrichment analysis for transcriptional signatures of interest. Core genes driving the signatures were used to build individual composite signature scores (median expression of core genes), which were used to map study participants based on their immune states. For enriched signatures, differential gene expression between high-IMCV and high-IIV sub-cohorts were evaluated post hoc using Wilcoxon rank-sum test.

F. Association of high-IMCV and high-IIV with transcriptional signatures across cell subsets from CITE-seq gene expression data. Associations are tested using mHG enrichment analysis. Significance (adj.p <0.1) is indicated by an asterisk symbol. Color scale represents signed −log10 (adj.p).

G. Boxplot comparing expression of selective sets of genes in cMo-A and cMo-B subsets of high-IMCV and high-IIV sub-cohorts study participants including interferon pathway regulators.

H. Boxplot showing IIV and IMCV factor scores stratified by seasonal influenza vaccine status (participants vaccinated within 6 months before sampling or not). Wilcoxon rank-sum test was performed. Significance is indicated with the following symbol: *=p<0.05.

4

Figure S4. Inflammation and TGF-β transcriptional programs are differentially regulated between IMCV and IIV, related to Figure 4.

A. Venn diagrams of differentially expressed inflammation response and TNF response genes between study participants with high (top 20%) and low (bottom 20%) IIV and IMCV factor scores. Wilcoxon rank-sum test and BH correction were used. Most DEGs (adj.p <0.1 and |log2 fold difference| >0.25) are considered and visualized on Venn diagrams.

B. Boxplot comparing expression of selective sets of genes in cMo-A and cMo-B subsets of high-IMCV and high-IIV sub-cohorts study participants including TNF-α pathway regulators and IL-1 family members.

C. Association of high-IMCV and high-IIV with transcriptional signatures across cell subsets from CITE-seq gene expression data. Associations are tested using mHG enrichment analysis. Significance (adj.p <0.1) is indicated by an asterisk symbol. Color scale represents signed −log10 (adj.p).

D. Scatter plot mapping study participants from the high-IMCV and high-IIV sub-cohorts. Unique coordinates derived from their composite signature scores for TNF/NF-κB and Inflammation in mNK A and MAIT subsets are used.

E-F. Boxplots comparing surface protein (E) and mRNA (F) expression of CD69 in MAIT and mNK A subsets of high-IMCV and high-IIV sub-cohorts study participants.

G. Comparing relative abundance of CD69^high MAIT and mNK A subsets in study participants from high-IMCV and high-IIV sub-cohorts in CITE-Seq data.

5

Figure S5. IMCV captures differences in SIGLEC-1^high monocytes and PD1^highICOS^high memory T cells across the cohort, related to Figure 5.

A. Lollipop plot of CITE-Seq immune cell subset frequency association with IMCV. Large circles with black rims represent features significantly associated with IMCV (adj.p <0.1) and small circles represent non-significant features. Linear regression was used to evaluate association of features with IMCV. Colors represent MOFA feature weight: blue indicates negative association, and orange indicates positive association with IMCV.

B. Box plots of SIGLEC-1 surface expression in myeloid cell subsets from individuals with high-IMCV and low-IMCV factor scores in CITE-seq. Cell subsets whose frequency is not associated with IMCV are shown. Dots indicate individual participants’ mean expression. Wilcoxon rank-sum test was used to assess significance of the difference (****=p <0.0001, ***=p <0.001).

C-D. Comparing relative abundance of SIGLEC-1^high cMo-A and cMo-B subsets (C) and cell surface expression of activation biomarker SIGLEC-1 (D) in study participants from high-IMCV and high-IIV sub-cohorts in CITE-Seq data.

E. Heatmap of differential protein expression in CyTOF memory T cells. Clusters whose frequencies are positively associated with IMCV (Pos) are compared to clusters not significantly associated with IMCV (ns). Kruskal–Wallis test was used to evaluate proteins differentially expressed (p<0.00001). Post-hoc pairwise Wilcoxon rank-sum test was used to compare expression levels and medians of expression are displayed. Markers are ordered by significance using a score based on the scaled −log10(p-value), with p=0 represented in black.

F-I. Boxplot showing single cell expression of memory T cell clusters whose frequencies are positively (pos) or negatively (neg) associated with IMCV, compared to those from regulatory T cells (Treg, F-G) and memory T cell clusters (Mem, H-I). Kruskal–Wallis test was used to evaluate proteins differentially expressed (p<0.00001). Canonical regulatory T cell markers (F and H) and T cell co-activation/inhibition markers (G and I) are shown. Wilcoxon rank-sum test was used to assess the significance of the differences (****=p <0.00005, ***=p <0.0005, **=p<0.005, *=p<0.05).

6

Figure S6. Immunomodulatory gut microbiome pathways and molecules are associated with IMCV, related to Figure 6.

A. Schematic of metagenomic data analysis approach.

B. Putative reactions leading up to ornithine synthesis from glutamate and n-acetylglutamate. Feature association with IMCV is indicated by an asterisk. Numbers indicate enzyme commission numbers for each reaction.

C. Heatmap quantifying species relative abundances association with unstratified pathway abundances related to acetate biosynthesis. Spearman’s rank correlation was used to evaluate association (correlation coefficient >0.4 and adj.p <0.1). Species significantly correlated with at least 1 pathway are shown. Colors indicate pathways (blue), species associated with IMCV (red) or species non-association with IMCV (black).

D. Lollipop plot of microbiome species relative abundances association with IMCV. Features significantly associated with IMCV (adj.p <0.1) are shown. Linear regression was used to evaluate association of features with IMCV. Colors represent MOFA feature weight: blue indicates negative association and orange indicates positive association with IMCV.

E. Putative reactions leading up to butyrate synthesis from acetate. feature association with IMCV is indicated by an asterisk. Numbers indicate enzyme commission numbers for each reaction

F. Dot plot of species contribution to the total genomic content for alkaline phosphatase gene families. Major contributors are shown (top 20 most prevalent species and median contribution >0.025). All contributions, including from minor contributors and unclassified species contributions were documented in the Table S3.

G. Generalized polyamine pathway diagram illustrating canonical enzymatic steps.

H-I. Heatmap quantifying species relative abundances association with unstratified pathway abundances related to polyamine metabolism (H) and MAMPS production (I). Spearman’s rank correlation was used to evaluate association (correlation coefficient >0.4 and adj.p <0.1). Species significantly correlated with at least 1 pathway are shown. Colors indicate polyamine pathways (green) and MAMP pathways (gray), species associated with IMCV (red) or species non-association with IMCV (black).

J. Dot plot of species contribution to the total genomic content for bile salt hydrolase gene families. Major contributors are shown (top 20 most prevalent species and median contribution >0.025). All contributions, including from minor contributors and unclassified species contributions were documented in the Table S7.

K. Phylogenetic tree of the microbiome data colored by family. Species identified as a major contributor to the total genomic content of gene family across the groups of metabolic processes involved in LPS endotoxin, SCFA, polyamine and primary bile acid metabolism were identified with a dot at the periphery of the tree and listed in the legend.

L. Schematic of genome-resolved microbiome analysis. De novo assembly of raw metagenomic data was performed to generate a library of sub-species genome bins (SGBs) using a 98% ANI threshold. SGBs were annotated, and each sample’s metagenome was mapped to the library to generate a genome-resolved community count matrix for further analysis.

7

Figure S7. IFN-response signatures and IMCV associated features are stable over time, related to Figure 7.

A. Schematic of the longitudinal follow-up samples and analysis.

B. Distribution of IMCV scores for all participants, the subset of participants from whom follow-up samples were available, and the subset of participants from whom CITE-seq data were generated from the follow-up timepoint.

C. Correlation of abundances of key microbial taxa that were significantly associated with IMCV between the two timepoints for each study participant (n=48). The color represents the Spearman correlation coefficient, and * indicates p-value <0.05.

D. Scatter plots visualizing a composite score of microbial taxa shown in (C) between the two timepoints for key cell populations. Each data point represents a different study participant. Statistics are the result of a Spearman correlation. Linear regression line is shown.

8

Table S1. Overview of the number of features quantified across data modalities, related to Figure 1

9

Table S2. Related to Figure 3. Differential gene expression analysis within and between factors IMCV and IIV.

10

Table S3. Related to Figure 3. Study participants’ transcriptional immune states summarized as individual composite scores across cell subsets and key immunological signatures.

11

Table S4. Related to Figure 4. Leading edge features differentially driving enrichment across inflammatory signatures and cell types in IMCV and IIV.

12

Table S5. Related to Figure 5. Association between cell frequency and gene expression across immune profiling modalities.

13

Table S6. Related to Figure 6. Pathways, metabolites and SGBs associated with IMCV.

14

Table S7. Related to Figure 6. Species contribution to pathway and gene family copy abundance.

Highlights:

• Comprehensive multi-omic analysis of the immune system and microbiome in healthy humans

• Major axes of immune variation are driven by interferon response signatures

• Specific immune cell states are coordinated with microbiome pathways and metabolites

• Coordinated immune and microbiome features are stable within individuals over time