Introduction

Over the last few decades, economic changes and advances in healthcare systems have significantly improved life expectancies in Asia1,2, and are expected to lead to a rapid shift in demographics by doubling the population of 60+ individuals by 20503 (from 19.7% in 2015 to 40%). This has been accompanied by a rising incidence of chronic diseases linked to aging4,5 and corresponding socio-economic stress on healthcare systems across Asia6,7. There is therefore an urgent need to identify lifestyle, dietary, and pharmaceutical interventions that promote healthy aging in Asian populations8,9,10.

Aging is believed to be a complex, multi-factorial phenomenon with progressive decline in several physiological functions including in the gastrointestinal and immune system11,12. Not surprisingly, many studies have therefore identified correlations between gut microbiome composition and age13,14. For example, in recent work, Zhang et al.15 have shown that there are distinct age and gender-associated trajectories in the gut microbiome that are consistent across populations. Several studies have also highlighted that frailty indices and healthy aging are linked to the state of the gut microbiome16,17, along with known gut-microbiome associations with aging-related diseases18,19. In particular, shifts in several families and genera of bacteria including Bacteroidetes, Proteobacteria, Roseburia and Escherichia have been associated via 16S rRNA sequencing to healthy aging20,21,22,23,24,25,26. Work by Wilmanski et al. also suggests that compositional uniqueness of the gut microbiome may serve as a marker for healthy aging27. To compensate for the potential low resolution of 16S studies, recent work has sought to leverage shotgun metagenomic sequencing to characterize the gut microbiome in elderly populations, but have been limited by cohort size (n < 50) and the spectrum of age ranges that they study, in their ability to account for potential confounders while identifying microbial mechanisms associated with healthy aging28,29.

To address this, we used deep shotgun metagenomic analysis (n = 234, >15 million reads on average) to study gut microbiomes in a cohort of community-living octogenarians (primarily, age range = [71–100], Table 1) in Singapore30,31. The gut microbiomes of healthy octogenarians exhibited a defined shift in diversity and overall taxonomic composition, as a function of age and independent of other covariates. In addition, taxonomic analysis identified several species-level changes (e.g., enrichment of Alistipes shahii and Bacteroides xylanisolvens), consistent with functional pathway analysis of the data, that points to a metabolic potential expansion in aging towards alternate pathways that synthesize and utilize amino-acids (e.g., L-lysine) as precursor substrates, relative to the classical pyruvate-to-butyrate production pathway. As butyrate derived from gut bacteria has diverse roles in host health (e.g., as energy for colonocytes, and reducing gut inflammation32,33,34), we next associated gut microbiome composition with key markers for inflammation (e.g., CRP), cardiometabolic health (e.g., fasting blood glucose) and liver health (e.g., AST), to find additional microbial functions associated with healthy aging. Leveraging extensive clinical data and the size of the metagenomic datasets, we identified >10 robust associations (accounting for demographic and clinical covariates that can be confounders) that highlight the role of the microbiome as biomarkers and potential intervention targets for promoting healthy aging.

Table 1 Clinical characteristics of the SG90 cohort
Full size table

Results

Shotgun metagenomics reveals a defined aging-associated shift in microbial richness and functional guilds

To study the role of the gut microbiome in healthy aging, stool samples from a deeply phenotyped cohort of community-living Asian octogenarians (SG90, Table 1, Supplementary Data File 1) were analyzed with deep shotgun metagenomic sequencing (n = 234, 17 million reads on average, Methods). The resulting high-resolution species-level taxonomic profiles were compared to reference shotgun metagenomic data from healthy, younger Singaporeans35,36 from two cohorts (SPMP, n = 109, age range = [53–74]; CPE, n = 96, age range = [21–80]), as well as other Asian populations37 (T2D, n = 171, age range = [21–70]; Supplementary Data File 2). Systematic joint analysis of these cohorts after batch-correction and matching other demographic characteristics (gender, ethnicity; Methods), revealed a progressive shift in taxonomic profiles across different age groups (n = 516, age range = [21–100], Fig. 1A), particularly along the y = x axis relative to the first and second principal components of variation (ordinal logistic test p-value < 0.001, Fig. 1B, Supplementary Fig. 1A). Diversity analysis indicated that this was accompanied by an aging-associated shift (ordinal logistic test p-value < 0.001, Fig. 1C) that is driven by microbial richness and evenness, both of which exhibit significant reduction with age (Supplementary Fig. 1B–D). PERMANOVA analysis showed that Age was the major source of variation compared to other attributes (Supplementary Fig. 2A) and that beta diversity peaked first in the 41–60 age group and then again in the 91–100 age group (Supplementary Fig. 2B), though uniqueness did not consistently increase with age as reported previously in a western cohort27. These results highlight an aging-associated shift in microbial composition that is defined by the loss of gut species and increase in beta diversity in extreme age groups.

Fig. 1: Aging-associated shifts in gut microbiome richness and relative abundance of key species.
figure 1

A Principal coordinates analysis (PCoA) plot based on species-level Bray–Curtis dissimilarity of gut microbiome profiles across age groups after batch correction. Dashed lines indicate the y = x and y = –x axes. B, C Violin plots showing the distribution across age groups. B PCoA1 and PCoA2 values projected on the y = x axes (n = 516 independent samples; The symbol ‘***’ represents p-value < 0.005, obtained from two-sided test with Ordinal logistic regression across all age groups, p-value 8.09 × 10–33). In all the inset boxplots, the center line represents the median, box limits represent upper and lower quartiles, and whiskers represent minimum and maximum values. C Shannon diversity indices; The symbol ‘***’ represents p-value < 0.005, obtained from two-sided test with Ordinal logistic test using n = 516 independent samples across all age groups, p-value 1.58 × 10–12). In all the inset boxplots, the center line represents the median, box limits represent upper and lower quartiles, and whiskers represent minimum and maximum values. Data points that are either less than Q1–1.5 × IQR or more than Q3 + 1.5 × IQR, where Q1, Q3 and IQR refers to the first quartile, third quartile and interquartile range, respectively, were removed for this analysis to reduce the impact of outliers. D Volcano plot showing the β coefficient on the x-axis and FDR-adjusted –log10 p-values (FDR-adjusted using Benjamini–Hochberg test, p-values obtained with two-sided tests from Generalized Linear Model (GLM) for taxa association with age) on the y-axis. Points for statistically significant taxa from “high stringency list” and those with high beta coefficients are highlighted in the volcano plot as red colored points for median abundance in the SG90 cohort greater than zero and green otherwise, with non-significant taxa shown with grey dots. E Relative abundance boxplots for the Bacteroides genus, and corresponding species that was identified to be associated significantly and not significantly with age. The symbols “n.s.”, “*” and “***” represent p-value > 0.05, p-value ≤ 0.05 and p-value < 0.005, respectively (GLM for taxa association with age using n = 516 independent samples; FDR-adjusted p-value using Benjamini–Hochberg test; Bacteroides: 6.4 × 10–3, B. dorei: 7.61 × 10–5). In all boxplots, the center line represents the median, box limits represent upper and lower quartiles, and whiskers represent minimum and maximum values (outlier points are not included in the visualization). Source data are provided as a Source Data file.

Full size image

In order to identify taxa associated with age, a generalized linear model (GLM) was used to account for demographic and clinical covariates that can be confounders (e.g., gender, body mass index, fasting blood glucose, triglyceride, total cholesterol and high-density lipoprotein; Methods). Overall, 4 phyla, 22 genera, and 40 species were found to be associated significantly with age (False Discovery Rate (FDR)-adjusted p-value < 0.05, Supplementary Table 1, Supplementary Data File 3). The phylum and genus level results were broadly consistent with prior observations (e.g., in Chinese38, Japanese25 and Italian centenarians28,39), with enrichment in Bacteroidetes (β = 0.33, FDR-adjusted p-value < 0.001), and depletion in Firmicutes (β = –0.49, FDR-adjusted p-value < 0.001) with age (Supplementary Table 1). At the genus level, a strong enrichment in unclassified Acidaminococcaceae species (β = 6.69, FDR-adjusted p-value < 0.05) and depletion in Parasutterella species (β = –20.19, FDR-adjusted p-value < 0.001) was observed that has not been noted before, while at the species level a majority of our observations were novel owing to the higher resolution of shotgun-metagenomics20,21,22,23,24,25,26 (Supplementary Data File 3, Table 2). In addition, we tested the robustness of this analysis by varying the normalization approach, statistical model and cohorts used to find that most taxa associations were detected in a majority of the conditions tested (Methods, Supplementary Fig. 3). We also noted that while several species associations were detected in individual cohorts (n = 10), the inclusion of the SG90 cohort notably boosted the number of age-associations identified (n = 38), suggesting that having a wider age range and more subjects could benefit such analysis (Supplementary Fig. 4). To mitigate the potential effect of incomplete batch-correction, a high-stringency list of species associations was further derived based primarily on methods that do not depend on this procedure (n = 17; Table 2, Fig. 1D, Supplementary Fig. 3; Methods). Reanalysis of taxonomic profiles and inclusion of results from other publicly available cohorts17,40,41,42,43,44 (n = 6; Methods) highlighted the consistency of these associations with a high percentage being replicated or validated (27/40 = 68%, odds ratio = 21.4, p-value < 6 × 10–15). In addition, the overlap with Asian cohorts (18/40, odds ratio = 8.83, p-value < 10–6) was more statistically significant than with western cohorts (16/40, odds ratio = 3.10, p-value < 7.1 × 10–3), despite having more western cohorts in this comparison (4/6), highlighting the value of having more Asian datasets focused on healthy aging (Supplementary Fig. 5).

Table 2 Species-level gut microbial taxa associated with aging
Full size table

Among the core gut microbiome taxa identified here, only some species in the Bacteroides genus showed significant positive associations with age27,45 (Table 2). This included Bacteroides dorei and B. xylanisolvens, with B. massiliensis even exhibiting a slight negative trend, though this variation is masked at the genus level highlighting the utility of species-level resolution offered by shotgun metagenomics (Fig. 1E). We also noted an enrichment with age for multiple Alistipes species, including A. shahii, A. onderdonkii and A. senegalensis, that are bile tolerant, have a unique way of fermenting amino acids such as L-lysine46,47, and can produce neurotransmitter precursors such as indole48 (Supplementary Fig. 6A). In contrast, several species that are known to be important for butyrate production in the gut46,47, including Faecalibacterium prausnitzii, Roseburia inulinivorans and Eubacterium hallii, were significantly depleted with age (Fig. 1D, Supplementary Fig. 6B). Several of these associations were seen in other Asian cohorts but not in the western cohorts analyzed here (n = 11, e.g., Dorea formicigenerans, R. inulinivorans, A. shahii and E. hallii) indicating that there may be consistent key differences in ethnic Asian populations. Also, in a few cases the directionality of association differed between Asian and western cohorts (e.g., B. xylanisolvens), highlighting the importance of population-specific studies. Integrating this information through taxonomic co-occurrence analysis highlighted a distinct network structure in the elderly (32 edges and 3 hubs) compared to younger individuals (80 edges and 9 hubs), with a shared pattern of cliques formed by the classical butyrate producers and Alistipes species, indicating that these may represent distinct functional guilds49 that switch roles with age (Supplementary Fig. 7).

Gut microbiomes in elderly Asians exhibit a metabolic potential expansion in butyrate synthesis pathways

To further analyze metabolic changes associated with aging, shotgun metagenomic data from all subjects from the four cohorts (SG90, SPMP, CPE and T2D) was used to quantify gene and pathway abundances (Supplementary Data File 4, Methods). In total, 413 metabolic pathways were identified and quantified of which 70 were found to be differentially abundant across various age groups (Supplementary Data File 5), broadly grouping into the four categories of sugar metabolism (17 pathways), vitamin, energy metabolism, and cell wall biosynthesis (28 pathways), lipid metabolism (9 pathways) and amino acid metabolism (16 pathways, Fig. 2). Specifically, among sugar metabolism pathways we observed distinct age-associated biases, where e.g., pyruvate fermentation to acetate and lactate was enriched in the younger age group (21–40), while degradation of mono-, di- and polysaccharides was enriched in older age groups (Fig. 2A). As many as 8 different pathways for various substrates were enriched in the 61–70 age group, including simple pentose sugars (Pentose phosphate pathways), and complex polysaccharides such as stachyose and starch (Starch degradation V). Notably, microbial glycolytic pathways for metabolism of simple sugars (glucose) were enriched in healthy individuals over 90, while prior work had highlighted the role of corresponding pathways in aging in model organisms50.

Fig. 2: Associations of microbial metabolic pathways across age groups.
figure 2

Bar plot showing pathways that are significantly associated and enriched specifically in different age groups (p-value ≤ 0.05, two-sided Wilcoxon rank-sum test) based on LEfSe analysis. Pathways were grouped into the broad categories of (A) Sugar metabolism, B Vitamin, energy metabolism and cell wall biosynthesis, C Lipid metabolism and (D) Amino acid metabolism. The effect size is depicted in the form of an LDA Score (log 10). Source data are provided as a Source Data file.

Full size image

In line with the vital role of vitamins as essential micronutrients (particularly in energy metabolism and immune function51,52) with limited synthesis in humans51, several microbial pathways were associated with aging (especially B vitamins, Fig. 2B). In particular, thiamine diphosphate and flavin biosynthesis were enriched in the most elderly (91–100) and could have anti-inflammatory and anti-aging roles52. On the other hand, three alternate biosynthetic pathways to produce thiamin and pantothenate were enriched in younger individuals (21–40).

While microbial lipid metabolism pathways were predominant in younger subjects (21–40 age group, Fig. 2C), pathways related to amino acid metabolism were more often enriched in the elderly (particularly 61–70 age group), including the synthesis of various essential amino acids such as histidine, valine, threonine, isoleucine, tryptophan, and methionine and aromatic amino acids (Fig. 2D). In addition, lysine biosynthesis was over-represented in the gut microbiomes of two older age groups53 (61–70 and 81–90), linking to our earlier observation of age-related enrichment in specific Bacteroides and Alistipes species, which are known to synthesize lysine and produce short-chain fatty acids (SCFA) such as butyrate with lysine as substrate46, respectively.

Analysis of the four major known pathways for butyrate production46,47 (from pyruvate, glutarate, 4-amino butyrate and lysine) confirmed that while several lysine-associated genes show an age-related increasing trend (e.g., kamA, FDR-adjusted p-value < 0.01 and kdd, FDR-adjusted p-value < 0.05), no significant associations were seen in the pyruvate related genes (Fig. 3). Furthermore, the overall gene abundance in the pathway for butyrate production from lysine also showed a significant increase with age (FDR-adjusted p-value < 0.001, Supplementary Fig. 8). Metabolic pathways analysis revealed that the abundance of gut microbial species that provide metabolic support to butyrate producers peaked in older age groups (71–80; Supplementary Fig. 9, Methods). To more directly link taxonomic and functional analysis, we first analysed the contribution of taxa to L-lysine biosynthesis III pathway and observed that Bacteroides sp. were the dominant contributors, with the contribution of B. xylanisolvens increasing with age (Supplementary Fig. 10A, 10B). Next, we systematically generated metagenome-assembled genomes, annotated genes and pathways, and conducted metabolic network analysis (Methods). This analysis showed that the metagenome-assembled genomes represented a diversity of Bacteroides and Alistipes species, with many genomes harboring corresponding complete pathways for lysine biosynthesis (Bacteroides) and lysine-to-butyrate conversion (Alistipes) that are reachable based on metabolic network analysis (Supplementary Fig. 10C–E). These observations, along with those in the previous section, indicate a metabolic potential expansion in butyrate synthesis pathways from precursors other than pyruvate in the gut microbiomes of elderly Asians.

Fig. 3: Differential representation of butyrate synthesis pathways in gut metagenomes across age groups.
figure 3

Pathway diagram where nodes represent major metabolites involved in various steps of butyrate production (but) starting from 4 different precursors (pyruvate, glutarate, lysine and 4-aminobutyrate), and edges are labelled with the genes involved in the conversion. Boxplots present normalized counts for the corresponding gene across different age groups (n = 516 independent samples). The labels “n.s.”, “*”, “**” and “***” represent p-value > 0.05, p-value ≤ 0.05, p-value < 0.01 and p-value < 0.001, respectively (two-sided tests using GLM for gene association with age with FDR-adjustment using Benjamini–Hochberg test per pathway; gcdAB – 2.68 × 10–16, kamA – 9.97 × 10–5, kamDE – 1.65 × 10–2, kdd – 1.51 × 10–3, kce – 1.65 × 10–2, abfH – 3.93 × 10–6). In all the boxplots, the center line represents the median, box limits represent upper and lower quartiles, and whiskers represent minimum and maximum values. Expansions of the abbreviated metabolites are as follows: pep - Phosphoenolpyruvate, pyr - Pyruvate, accoa - Acetyl-CoA, aacoa - Acetoacetyl-CoA, 3hbcoa - (S)-3-Hydroxybutanoyl-CoA, akg - 2-Oxoglutarate, 4h2oglt - 4-Hhydroxy-2-oxoglutarate, 2hydogc - 2-Hydroxy-glutaryl-CoA, 3mgcoa - 3-Methylyglutaconyl-CoA, lys – L-lysine, 36dahx - (3S)-3,6-Diaminohexanoate, 35dahx - (3S)-3,5-Diaminohexanoate, 5a3ohex - 5-Amino-3-oxohexaonate, 3abutcoa - L-3-Aminobutyryl-CoA, 4abut - 4-Aminobutyrate, succsal - Succinic semialdehyde, 4hdxbutn - 4-Hydroxybutyrate, 4-hbutcoa - 4-Hydroxybutyrate-CoA, 3btcoa - Vinylacetyl CoA, b2coa - Crotonyl CoA, but - Butyrate. Source data are provided as a Source Data file.

Full size image

To further explore these findings, we leveraged the recent development of a healthy aging model in mice, where dietary supplementation with alpha ketoglutarate (AKG) to aged mice (18 months) not only increased lifespan but also improved inflammation and frailty markers54. Shotgun metagenomics of fecal samples from AKG-supplemented and control mice (n = 20 per group, Supplementary Fig. 11A; Methods) showed that even though mice harbor distinct gut microbial species, at the pathway level a similar expansion in butyrate production pathways is seen in healthy aging mice (Supplementary Fig. 11B). In particular, the lysine to butyrate conversion pathways shows the strongest enrichment, as seen in our human data (1.65-fold increase in median abundance, FDR-adjusted p-value < 0.01). At the gene level, several key genes for lysine to butyrate production were strongly enriched including kdd, kamDE, kce and kamA (FDR-adjusted p-value < 0.05; Supplementary Fig. 11C). Such a trend was not observed for genes involved in pyruvate to butyrate conversion, highlighting the role of expansion in alternate butyrate synthesis pathways to support healthy aging.

Robust species-level associations between gut microbial taxa, clinical biomarkers and aging-associated phenotypes

Aging is a significant risk factor for chronic diseases impacting multiple organ functions, including metabolic, immune, and musculoskeletal systems, which could be mediated via gut microbiome function26,55. To further relate the taxonomic and functional changes observed in Asian octogenarians with their clinical phenotypes we harnessed the higher resolution of shotgun metagenomic data, broader age range, and size of our cohorts, to robustly identify associations in a GLM framework while accounting for demographic and clinical confounders (e.g., age, gender, other metabolic markers; Methods, Supplementary Data File 6). Despite the multiplicity of taxa and clinical features, we noted that this analysis typically resulted in a few significant associations (Fig. 4). For example, for fasting blood glucose as a well-studied biomarker for glucose metabolism56, insulin resistance57, and type 2 diabetes58, we identified two associated species (Bifidobacterium adolescentis [β = 0.08, FDR-adjusted p-value = 2.61 × 10−2] and Parabacteroides goldsteinii [β = 0.48, FDR-adjusted p-value = 2.63 × 10−2], Fig. 4A) with prior in vitro and in vivo observations supporting this link59,60,61, as well as a new association with a Lactobacillus species (Lactobacillus mucosae: β = 0.88, FDR-adjusted p-value = 4.30 × 10−2) that has never been described before despite numerous studies for this genus62,63,64,65. Note that β values indicate the change in response variable (e.g., fasting blood glucose levels) for every unit change in relative abundance values, after accounting for other confounding variables. Given the known normal range of fasting blood glucose levels (<6.1 mmol/L66; Supplementary Table 2) the observed β values (0.48 and 0.88) therefore could suggest a large effect size.

Fig. 4: Microbiome associations with key clinical markers.
figure 4

Volcano plots showing the β coefficient on the x-axis and the FDR-adjusted –log10 p-values on the y-axis (FDR-adjusted using Benjamini–Hochberg test, p-values obtained with two-sided tests from GLM) for gut microbial associations with clinical markers for (A) Glucose control as measured by fasting blood glucose levels (Bifidobacterium adolescentis: 2.61 × 10–2, Parabacteroides goldsteinii: 2.63 × 10–2, Lactobacillus mucosae: 4.30 × 10–2), B Lipid control as measured by total cholesterol in plasma (Haemophilus parainfluenzae: 7.95 × 10–4, Lachnospiraceae bacterium 1 4 56FAA: 3.92 × 10–3), C Liver health, as measured by plasma levels of aspartate aminotransferase (Klebsiella pneumoniae: 3.99 × 10–3), and (D) Systemic inflammation, as measured by high-sensitivity CRP levels (Eggerthella unclassified: 4.45 × 10–2, Escherichia coli: 4.45 × 10–2, Streptococcus salivarius: 4.45 × 10–2), E Vitamin B12 (Streptococcus parasanguinis: 4.54 × 10–9, Bacteroides coprocola: 9.76 × 10–5). Points in all plots are colored with red for FDR-adjusted p-value ≤ 0.05 and grey otherwise. Dotted lines represent the thresholds of p-value = 0.05 and p-value = 0.1. Source data are provided as a Source Data file.

Full size image

We next analyzed cholesterol biomarkers (e.g., high-density lipoproteins—HDL, low-density lipoproteins—LDL and triglycerides) for hypercholesterolemia and cardiovascular diseases, due to their impact on healthy aging67. Association analysis with total cholesterol levels revealed two significant candidates, positive association with abundance of Lachnospiraceae bacterium 1 4 56FAA (β = 1.02, FDR-adjusted p-value = 3.92 × 10–4) and negative association with Haemophilus parainfluenzae (β = –0.06, FDR-adjusted p-value = 7.95 × 10–5, Fig. 4B), neither of which has been reported previously68,69. Interestingly, analysis with LDL levels revealed the opposite trend, with the Lachnospiraceae species (β = –0.88, FDR-adjusted p-value = 4.27 × 10–2) being negatively correlated and H. parainfluenzae (β = 0.06, FDR-adjusted p-value = 1.72 × 10–3) being positively correlated, and similar trends being observed with HDL but not reaching statistical significance (Supplementary Fig. 12A, B). These observations indicate potential beneficial roles of Lachnospiraceae species in maintaining lower levels of LDL. Interestingly, microbial association with triglyceride levels was dominated again by positive association with H. parainfluenzae (Supplementary Fig. 12C), a species that has been described to be strongly enriched in patients with chronic cholecystitis patients70,71 and individuals with increased risk of cardiovascular disease72,73.

As the biological processes of aging are in part believed to be accelerated by inflammation74,75 (Inflammaging), including risk for cardiovascular disease, frailty and disability76,77,78, we sought to identify gut microbial associations with inflammation markers. Specifically, serum levels of aspartate aminotransferase (AST) were associated with the abundance of Klebsiella pneumoniae (β = 0.65, FDR-adjusted p-value = 3.99 × 10–3, Fig. 4C), a sporadic colonizer of the gut microbiome, and a known pathogen for liver abscesses that are increasingly common in Asian populations79. A similar association was seen with alanine aminotransferase (ALT), a more specific biomarker for liver injury, but it did not reach statistical significance (Supplementary Fig. 12D). Analysis with high-sensitivity C-reactive protein (hs-CRP) levels, a commonly used marker for systemic inflammation80, identified a known association with Escherichia coli81 (β = 0.0020, FDR-adjusted p-value = 4.45 × 10−2), as well as an association with a Streptococcus species (S. salivarius: β = 0.02, FDR-adjusted p-value = 4.45 × 10−2) which has been linked previously with rheumatoid arthritis82 (Fig. 4D). Furthermore, a significant association was observed with a Eggerthella species (β = 0.15, FDR-adjusted p-value = 4.45 × 10−2) distinct from Eggerthella lenta83, whose role is less well-understood68, and the association with host inflammation has not been described before. These results highlight the various bacterial species and potential pathogenic functions by which altered gut microbiomes can contribute to inflammaging.

Aging is a common risk factor that often predisposes elderly to the development of neurological disorders such as Parkinson’s and Alzheimer’s disease84. Many studies have shown that Vitamin B12 is involved in several processes of the nervous system, including the synthesis of myelin and post-injury nerve regeneration85,86, the deficiency of which causes peripheral neuropathy87, gait ataxia and physical frailty that are often observed in the elderly. A strong association was observed between serum levels of Vitamin B12 and the bacterial species Streptococcus parasanguinis (β = 312.50, FDR-adjusted p-value < 5 × 10–9, Fig. 4E), strains of which are known to produce Vitamin B1288, highlighting the metabolic potential for gut bacteria to support healthy aging. Intriguingly, we also observed a significant association for blood levels of Vitamin B12 and the abundance of Bacteroides coprocola (β = 42.89, FDR-adjusted p-value < 10–4), even though Bacteroidetes are not common B12 producers89 and instead compete with the host to capture B12 in the gut90, indicating that dietary intake might be an explanation for the observed positive association. Together, these observations highlight the two-way communication between gut microbial species and their environment that can impact host phenotypes as well as serve as meaningful biomarkers of healthy aging.

Discussion

The complexity of the human gut microbiome, both in terms of its genetic diversity and the myriad host-microbial interactions that shape it, necessitates that association studies be done with sufficiently high-resolution and power to help infer meaningful insights about function and mechanisms91. While prior studies were primarily limited by their resolution (16S rRNA surveys20,21,22,23,24,25,26), recent studies have underscored the utility of shotgun metagenomics for species-level and functional pathway analysis, albeit in limited-sized cohorts28,29. By generating deep shotgun metagenomic data for >200 elderly subjects and with a combined dataset of >500 subjects from all cohorts, this study significantly enhances our ability to explore the role of the microbiome in healthy aging. In particular, we provide the first well-powered species-level view of microbiome shifts as a function of age, with a steady decrease in species richness (Fig. 1A, Supplementary Fig. 1B). This is accompanied by an increase in beta diversity in extreme age groups, but further defines that this is due to an overall shift of microbial community composition with age, with specific species serving as markers (e.g., A. shahii, B. xylanisolvens, R. obeum and E. hallii). Our analysis also highlights that the age-associated shift is on an axis of variation (Fig. 1A, B) that is distinct from the axis separating different gut microbiome clusters reminiscent of enterotype structures92 (Supplementary Fig. 1A). In addition, while the overall reduction in microbial diversity observed here (Fig. 1C) is broadly consistent with some studies24,93, our analysis clarifies that this is accompanied by an increase in beta diversity of microbiome composition and a loss of species richness (Supplementary Fig. 2B, 1B–D), defined in part by several classical butyrate producers in the gut (e.g., F. prausnitzii, R. inulinivorans and E. rectale). We did not observe a general loss in core Bacteroides species as reported in Wilmanski et al.27, which could be a function of our Asian cohorts, but may also be due to methodological differences (e.g., covariate adjustment). Also, our results suggest that the trends reported for richness22,94 and uniqueness27 may be variable across cohorts95. Further studies are needed across different global populations to explore the factors impacting this variability using consistently generated datasets and potentially using long-read sequencing to enable more reliable detection/assembly of rare taxa. This could also impact the detection of associations with viruses and fungi (none reported here), which are only sporadically abundant in the human gut microbiome.

In order to have data from a wide range of ages and a large number of subjects to sufficiently power our age-association analysis, we combined data from multiple Asian cohorts (SG90, T2D, SPMP, CPE) with metagenomic sequencing data, and attempted to account for differences in how the data was generated using a batch-correction approach (Methods, Supplementary Table 4). In addition, our association analysis accounted for several demographic and clinical factors that could be confounders, including ethnicity, gender, body mass index (BMI), fasting blood glucose, triglyceride, total cholesterol and high density lipoprotein levels. Nevertheless, it is important to note that batch-correction methods cannot be expected to fully account for all technical sources of variation, and similarly, other potential confounders (e.g., diet, medications) that could not be accounted for as this data was not consistently available across cohorts. Multi-cohort association analysis thus necessarily has to evaluate the robustness of the results obtained as a function of various data subsets used, technical choices made (e.g., normalization) and statistical approaches used. Our experiments suggest that the batch-correction and GLM-based association analysis reported here provide robust results that are reproduced across a majority of conditions tested (Methods, Supplementary Fig. 3). This was further validated by the strong enrichment of age associations detected in other cohorts, particularly in Asian cohorts (Supplementary Fig. 5). In addition, our analysis highlights the importance of having sufficient data from older age groups (Supplementary Figs. 4 and 5). Overall, despite all analytical care, the interpretation of microbiome associations with aging needs care as it is a multi-factorial phenomenon, where e.g., increased lifetime use of antibiotics could explain loss in gut microbial species richness with age, while at the same time such observations could arise from other factors such as lifestyle choices (e.g., smoking) and their independent impact on longevity and the gut microbiome. Our study is structured to identify microbiome associations with age and considers BMI, total cholesterol, fasting blood glucose etc as confounders, though in principle, the influence of age on these variables could be through the microbiome. In that sense, our study is conservative and should identify only the strongest independent associations. These associations can then be further explored in mouse models where we hypothesize that the specific Bacteroides and Alistipes species identified here could promote healthy aging independently or in combination through their metabolic outputs. In addition, the associations identified here can serve as useful biomarkers that in combination with other known clinical biomarkers can help track individuals who are aging in a healthier manner96,97 as well as develop better frailty metrics to identify individuals at risk for aging-associated diseases98.

The importance of incorporating age and other covariates in microbiome association analyses has been emphasized in several recent studies17,99. While this also necessitates larger cohort sizes and extensive metadata collection, a potential advantage is that the few associations that survive covariate adjustments are likely to be more meaningful (e.g., no associations were detected for sleep duration, frailty markers or free thyroxine levels; Supplementary Data File 6). In combination with shotgun metagenomics, where specific species and functional genes/pathways can be directly quantified, and related through metagenome-assembled genomes, this provides for a powerful avenue for developing meaningful mechanistic hypotheses informed by in vivo human data. In particular, a striking feature of our results is the depletion of a ‘microbial guild’ of butyrate-producing species in elderly subjects and its replacement with specific Alistipes species that can produce butyrate through an alternate pathway using amino acids as precursors. At the metabolic potential level, we do not see a significant reduction for the pyruvate-to-butyrate conversion pathway, and it is not clear if dietary or microbially produced amino acids are the key source for the alternate pathways100. Prior studies have shown reduction in SCFA levels, particularly butyrate, in stool samples from the elderly101, association between butyrate levels and frailty14, and hypothesized a shift towards bacterial protein fermentation due to increased intestinal transit time102. We observed a relative reduction of microbial to host reads in elderly subjects (Supplementary Fig. 13) that could reflect a reduced capability to support gut microbial fermentation, though direct measurements are needed to confirm this indirect evidence. Changes in dietary patterns (e.g., low fiber) as well as intestinal physiology (e.g., muscle loss, hypochlorhydria) could thus partly explain the observed shift in species composition and deserves further investigation. Our data from a mouse model of healthy aging further supports the idea that alternate butyrate production pathways in the gut microbiome could play a causal role in promoting healthy aging. The model is established by feeding mice α-ketoglutarate which is rapidly absorbed in the small intestine103, but may still explain the enrichment seen in the glutarate pathway in the gut microbiome. The strong enrichment seen in the lysine-to-butyrate pathways is consistent with our observations in human cohorts and could reflect the systemic metabolic and physiological effects of α-ketoglutarate in promoting healthy aging104, though it is not clear if this effect could be mediated through the gut microbiome. Follow-up experiments in mice with defined microbiome compositions, with and without α-ketoglutarate and L-lysine supplementation, could help further test the hypothesis that alternative butyrate production pathways have a causal role in healthy aging.

As the microbial production of SCFAs in the gut is known to play an important role in many aspects of human health105,106 (e.g., gut barrier function, immune regulation, gut-brain communication), it is not surprising therefore that reduction in butyrate-producing species has been associated with frailty107. Based on the observations in this study we hypothesize that alternate pathways for butyrate production, including using glutarate, 4-amino butyrate, and lysine as precursors, may partially augment the pyruvate-dependent pathway in the gut microbiomes of elderly individuals to promote healthy aging. These observations may be due to a “selection on survival” bias27, but results from the mouse model suggest that this may not provide a full explanation. Lysine is an essential amino acid that is found in lower plasma concentrations in frail elderly subjects108, and lysine supplementation is being investigated in multiple clinical trials for its health benefits. Its use as a supplement to promote healthy aging could thus be another avenue to explore via the pathway of boosting butyrate production in the gut. Among other pathways that were relatively depleted in the elderly, reduction in pyruvate fermentation to acetate and lactate and lipid metabolism pathways (Fig. 2) may have detrimental effects for host health and can be targets for compensation through supplements and probiotics109,110. On the other hand, the relative enrichment of vitamin synthesis pathways in the elderly, including thiamine diphosphate (vitamin B1), flavin (vitamin B2), and phospho-pantothenate (vitamin B5), as well as the association of specific species with serum vitamin B12 levels (Fig. 4E), could contribute to healthy aging and thus deserves further study.

Several aging-related chronic diseases, including type 2 diabetes37, cardiovascular diseases111, and neurological conditions112 have been linked with the gut microbiome and thus it is essential to account for age as a confounding factor when studying gut microbial mechanisms that contribute to disease. We hypothesize that having individuals from as wide a range of ages as possible, particularly those who have aged well, will likely help strengthen and refine such association analysis. We detected a positive association of B. adolescentis with fasting blood glucose, which is well-known to impact glycemic control113. In addition, we also noted a significant association for P. goldsteinii, a potential new probiotic species with emerging evidence for a diet-dependent role in obesity and type 2 diabetes59,114, though the mechanisms for this effect in mouse models remain to be elucidated.

Systemic low-levels of inflammation are believed to contribute to various aging-related phenotypes74,75 (Inflammaging). Correspondingly, associations between gut microbial species and markers of inflammation (e.g., hs-CRP and AST) are of particular interest. While the association of a common-source of liver abscesses in Asia (K. pneumoniae) with a liver disease marker (AST), and a common oral bacterium (S. salivarius) with a systemic inflammation marker (hs-CRP) are consistent with disease biology, the distinctness of these associations is striking i.e., while K. pneumoniae is the only species associated with AST, a more diverse group of bacteria are associated with hs-CRP levels. Furthermore, relatively little is known about the Lachnospiraceae species with strong associations to cholesterol/LDL levels. Together with the strong association of microbial biomarkers such as S. parasanguinis and B. coprocola with serum Vitamin B12 levels, these findings could help develop a non-invasive frailty test based on at-home sample collection115. Overall, these results highlight the value of species-level shotgun metagenomic analysis in large well-characterized cohorts, where further studies in other populations would help determine if some of these associations are specific to Asian environments and lifestyles.

Methods

Cohorts and datasets

SG90 cohort

The SG90 cohort is based on a longitudinal population health study that was set up in the 1990s which involved routine measurement of metabolic and other health variables30,31. The current dataset is based on a subset of 234 elderly individuals (77–97 years old) who are community-living participants (not living in a nursing home, no diagnosis of dementia and not physically unfit) and consented to providing their stool and blood samples. No statistical method was used to predetermine sample size. The participants in the SG90 cohort were recruited under the SLAS-3116 protocol approved by the Institutional Review Board (IRB) at National University of Singapore (reference number: B-15-081). This study was also approved by an IRB for the Singapore Chinese Health Study30 at National University of Singapore (reference number: H-17-027, sample collection: 2019/00439). A written informed consent was obtained from all the participants. Fasting blood glucose (mmol/L), triglyceride (mmol/L), total cholesterol (mmol/L), HDL (mmol/L), LDL (mmol/L), hs-CRP (mg/L), AST (U/L), ALT (U/L) and Vitamin B12 levels (pmol/L) were measured based on a blood draw at NUH Referral Laboratories, Singapore. Stool samples were collected either at the time of blood collection or within a week and were transported in refrigerated conditions to the lab where it was frozen at –80 °C. In addition, physical assessments were performed to calculate BMI (kg/m2) from their weight (kg) and height (m), as well as gait speed (m/s) and handgrip strength (kg) of subjects were measured.

SPMP dataset

The SPMP dataset is based on the recall of a subset of 109 healthy Singaporean subjects (53–74 years old) from a multi-omics study in Singapore117. Stool samples were collected for gut microbiome analysis using shotgun metagenomic sequencing35.

T2D dataset

Shotgun metagenomic datasets were obtained for 171 healthy Chinese individuals from a previously published type 2 diabetes (T2D) study37,118 using the curatedMetagenomicData package118 (v1.16.0). Briefly, the subjects chosen for our study were 21–70 years old and were non-diabetic controls in the study. Clinical data such as fasting blood glucose (mmol/L), triglyceride (mmol/L), total cholesterol (mmol/L), HDL (mmol/L) and LDL (mmol/L) levels were also obtained from this study37.

CPE dataset

The CPE dataset36 is based on a prospective cohort study consisting of CPE-colonized subjects and their healthy family members. For our comparisons, we used shotgun metagenomic data of 82 healthy family members (21–80 years old) with Chinese ethnicity.

Demographic matching

Analysis for all cohorts was restricted to ethnic Chinese individuals and the gender balance across cohorts was found to be comparable (SG90: 59% female, T2D: 52% female, SPMP: 60% female, CPE: 60% female). Age distributions for all cohorts can be found in Supplementary Fig. 14.

DNA library construction and sequencing

PowerSoil DNA Isolation Kit (12888-100, MO Bio Laboratories) was used for the extraction of DNA from frozen stool samples. Minor modifications to the manufacturer’s protocol were made (double the volume of C2, C3, and C4 buffers was added, and the duration of the centrifugation step was extended to twice the original duration). Purified DNA was eluted in 80 µL of C6 solution. DNA libraries were prepared using 50 ng of extracted DNA re-suspended in a volume of 50 µl. This was subjected to shearing using Adaptive Focused AcousticsTM (Covaris) with the following parameters—Duty Factor: 30%, Peak Incident Power (PIP): 450, 200 cycles per burst, Treatment Time: 240 s. Sheared DNA was cleaned up with 1.5× Agencourt AMPure XP beads (A63882, Beckman Coulter) followed by end-repair, A-addition and adapter ligation using the Gene Read DNA Library I Core Kit (180434, Qiagen) according to the manufacturer’s protocol. Custom barcode adapters were used instead of GeneRead Adapter I Set for adapter ligation (Supplementary Table 3). Before enrichment, DNA libraries were cleaned twice using 1.5× Agencourt AMPure XP beads (A63882, Beckman Coulter) using the protocol from Multiplexing Sample Preparation Oligonucleotide kit (PE-400-1001, Illumina). Enrichment PCR was carried out with PE 1.0 and custom index-primers for 12 cycles. DNA Libraries were prepared with Agilent DNA1000 Kit (5067-1504, Agilent Technologies) by pooling equimolar concentrations and quantified using Agilent Bioanalyzer (G2939BA). A pilot set of 50 DNA libraries were sequenced on an Illumina HiSeq 2500 and the rest were sequenced on an Illumina HiSeq X instrument (Supplementary Data File 1) generating >17 million 2 × 101 bp reads on average per library. The sequencing protocol and the median sequencing depth of other cohorts can be found in Supplementary Table 4.

Read pre-processing and profiling

Illumina shotgun metagenomic sequencing reads were processed using a Nextflow pipeline (https://github.com/CSB5/shotgunmetagenomics-nf). Briefly, raw reads were filtered to remove low-quality bases and adapter sequences were removed using fastp119 (v0.20.0) with default parameters. Human reads were removed by mapping to the hg19 reference using BWA-MEM120 (v0.7.17-r1188, default parameters) and samtools121 (v1.7). The remaining reads were used for taxonomic profiling using MetaPhlAn2122 (v2.7.7, default parameters). Functional profiles for the metagenomes were obtained using HUMAnN2123 (v2.8.1). For all statistical tests, Benjamini–Hochberg’s false discovery rate method was used to correct for multiple testing at a significance threshold of 5%.

Statistical analysis with taxonomic profiles

Taxonomic profiles were corrected for batch effects with MMUPHin124 (v1.8.1) using age group as a covariate. MMUPHin is a batch correction method that employs an empirical Bayes approach to model read counts with respect to batch variables and biologically relevant covariates (age group in our case). It then gives as output batch-corrected count data which aims to retain the effects of biologically relevant covariates. For batch correction, we considered samples from the SG90 and CPE cohorts as belonging to the same batch due to their similarity in DNA extraction and library preparation methods i.e., the major sources of variation in metagenomic profiles125 (Supplementary Table 4). To assess if differences in sequencing platform and depth could affect this assumption, we performed differential abundance analyses using MaAsLin2126 (v1.10.0) (correcting for age) between (i) subsets of the SG90 cohort sequenced on the two different platforms (Illumina HiSeq 2500 and Illumina HiSeq X), and (ii) down-sampled version of the CPE dataset with median sequencing depth matching the SG90 cohort compared to the original CPE dataset. In both analyses we did not identify any differentially abundant taxa between the respective groups, supporting our treatment of the two cohorts as the same batch. We conducted a sanity-check on batch-corrected profiles to confirm that cohort effects were reduced (Supplementary Fig. 15). Species-level relative abundances (>0.1%) were then used to compute alpha diversity indices (Shannon and Simpson), Pielou’s evenness and the beta diversity index (Bray–Curtis distance) using the R package vegan127 (v2.6-4) and visualized using ggplot2128 (v3.5.1). Robustness of computed metrics to variations in sequencing depth was confirmed based on sub-sampling and correlation analysis (Supplementary Fig. 16). For sub-sampling analysis, we selected a subset of 30 samples each with at least 15 million reads and randomly subsampled 7, 14, and 21 million reads from each of these samples. We then generated the taxonomic and pathway profiles for each of these sequencing depths using the same Nextflow pipeline and computed Spearman correlation values for alpha diversity metrics and pathway counts with the values obtained originally (Supplementary Fig. 16). Ordinal logistic regression in R was used to test for statistical significance in relation to different age groups. A GLM approach was used to test associations between the relative abundance for each species (SA) and age as a continuous variable accounting for covariates.

For this analysis, all demographic and clinical covariates available in multiple cohorts (n = 7 covariates) were first screened for relevance as potential confounders based on previously reported studies69,81,129 (all retained), and then through Spearman correlation within covariates to exclude collinear variables. Specifically, we selected and included only variables that were not highly correlated with one another, retaining only one of the covariates if their correlation was greater than 0.9 (one covariate removed, i.e., LDL levels). The final set of covariates used in this analysis (n = 6) included gender (G), BMI, fasting blood glucose (FBG), triglycerides (TGL), total cholesterol (TC), and HDL levels (i.e., with the formula: Age ~ SA + G + BMI + FBG + TGL + TC + HDL). Only taxa with >50 non-zero values were tested to avoid spurious associations. Missing clinical covariates were imputed through an Expectation Maximization algorithm130.

To assess reproducibility of taxonomic associations, three distinct approaches were used: (i) GLM analysis using all four cohorts with batch correction as described above, (ii) GLM analysis using the original data for the SG90 and CPE cohorts as they were similarly processed, and (iii) Trend analysis using all four cohorts with the Cochran-Armitage test131 after conversion of relative abundance data into presence-absence values (cutoff of 0.1%). Microbial associations in same direction found in two out of the three methods were considered reproducible with the reported p-values being the median across all methods. In order to further test the robustness of associations, parameters and options were varied for all three approaches including normalization technique (Total Sum Scaling—TSS and Cumulative Sum Scaling—CSS), association analysis technique (ANCOM-BC132, MaAsLin2126) and relative abundance cutoffs (0.05%–1%; Supplementary Fig. 3). To obtain a “high stringency” list of microbial associations we restricted to taxa within the SG90 + CPE cohort that are either (i) reported by all four conditions or (ii) reported by at least three conditions, in which case the presence-absence tests are required to report an association at all tested cutoff values (Supplementary Fig. 3). This list avoids the possibility of a false positive that emerges solely due to batch correction artefacts. The final taxa in this list were visualized using ggplot2 in R. CCREPE133 (v1.0) was used to compute Spearman correlation values and identify bacterial species with strong co-occurrence patterns in different age groups (|ρ| > 0.2, p-value < 0.05).

To further assess the robustness and reproducibility of our associations, we identified microbial associations across six additional cohorts of Asian and Western origin (Supplementary Fig. 5). Metagenomic taxonomic profiles from stool samples for four of these cohorts (labelled as YachidaS_201940, AsnicarF_201741, BritoIL_201642 and LLD_201643) were obtained from curatedMetagenomicData118 (v3.8.0). The subjects in these cohorts were healthy and were selected for not having antibiotic usage. Microbial associations and taxonomic profiles for two other cohorts (labelled as XuWuZhu_202244 and ElderMet_202017) were obtained from their respective publications. All of these cohorts were analyzed independently. We employed a similar GLM framework (with the formula: Age ~ SA + G) as outlined previously to independently identify microbial associations with age. Microbial associations from our discovery cohort (SG90 + SPMP + T2D + CPE) were considered ‘replicated’ or ‘validated’ in other cohorts based on their statistical significance (FDR-adjusted p-value < 0.05 or p-value < 0.05, respectively).

Pathway and gene-level analysis

The HMP Unified Metabolic Analysis Network (HUMAnN2)134 (v 2.8.1) pipeline was used to determine the relative abundance of microbial pathways in different gut metagenomes. The default Kyoto encyclopedia of genes and genomes (KEGG)135 catalog was used as the pathway reference. Unstratified relative abundance values for SG90, CPE and SPMP were integrated with HUMAnN2 results for T2D from curatedMetagenomicData118 (v1.16.0) for shared pathways, and significant differentially abundant pathways across age groups were determined based on linear discriminant analysis with LEfSe136 (p-value < 0.05 and LDA score >3).

For gene-level analysis of metabolic pathways involved in butyrate production, EC numbers of enzymes were mapped to corresponding UniRef90137 identifiers from the HUMAnN2 output. The list of butyrate producing species, genes, and enzymes (EC numbers) in butyrate production pathways were obtained from previously published studies46,47. Using this list, the butyrate production genes of interest from our dataset were obtained using a three step process: (i) UniRef90 identifiers obtained from HUMAnN2, were first cross-referenced with the KEGG database to obtain a list of KEGG orthology (KO) groups, (ii) EC numbers obtained previously were matched with the KO groups (Supplementary Data File 7), and (iii) the KO groups were cross-referenced with UniRef90 to obtain the specific butyrate producing genes of interest. For every gene, read counts were obtained by taking the sum for all corresponding UniRef90 identifier-associated values and statistical analyses were performed using R with visualizations generated using ggplot2128.

Metagenomic assembly

Microbial reads from the SG90 cohort were assembled and binned following an in-house snakemake pipeline. Briefly, the reads were assembled using MEGAHIT138 (v1.2.9). Contigs were binned using MetaBAT2139 (v2.15) based on contig coverage information generated with minimap2140 (v2.26). Assemblies were evaluated based on MIMAG141 definitions, with contamination, completeness, and N50 values obtained from CheckM2142 (v1.0.2), and non-coding RNA identified using barrnap (https://github.com/tseemann/barrnap) (v0.9) and tRNAScan-SE143 (v2.0.12). Potential cases of chimerism were detected using GUNC144 (v1.0.5). All software were used with default arguments. Low-quality MAGs were removed from downstream analysis and taxonomic classification of MAGs at the species level was obtained using GTDB-Tk145 (v2.3.2) ‘classify_wf’ method (GTDB version r214). Genomes were annotated using Prokka146 (v1.14.6) and the EC numbers associated with each gene were extracted. Metabolic pathways were obtained using MinPath147 (v1.6). Reachability analysis was carried out using MetQuest148 (v0.1.33) using the bipartite graphs constructed from these metabolic pathways. The initial seed list for the reachability analysis included source metabolites L-aspartate for L-lysine biosynthesis III and L-lysine for L-lysine to butyrate pathway.

Mouse experiments

All the mouse experiments were carried out with the approval of the Institutional Animal Care and Use Committee (IACUC) of National University of Singapore (NUS) under the R21-0135 protocol. Two independent groups of 16-month-old male and female C57BL6/J mice were purchased from Jackson Laboratories (RRID: IMSR JAX:000664) and housed at 25 °C with 50% humidity on a 12-h light/dark cycle. The animals were housed separately (maximum of four mice per cage), with n = 20 mice per group and were maintained on a regular chow diet until they were 18 months old. At 18 months of age, mice in the control group continued the regular chow diet while those in the healthy aging group were switched to a diet containing 4% calcium alpha-ketoglutarate (Ponce de Leon Health), with stool samples collected at baseline and 3 months after diet switch for metagenomic analysis. Genomic DNA was extracted from mouse stools using QIAamp PowerFecal Pro DNA Kit (51804, Qiagen), according to manufacturer’s instructions. DNA was quantified on a Qubit Fluorometer using the Qubit dsDNA BR Assay Kit (Q32850, ThermoFisher Scientific). Purified genomic DNA (50 ng) was used for library construction steps using NEBNext® Ultra™ II FS DNA Library Prep Kit (E7805L) according to manufacturer’s instructions. Finally, each library sample was assessed for quality based on fragment size and concentration using the Agilent D1000 ScreenTape (G2991BA) system, with samples adjusted to identical concentrations by means of dilution and volume-adjusted pooling. The multiplexed sample pool was paired-end (2 × 151 bp) sequenced on an Illumina HiSeq X Ten system. Mouse gut metagenomes were annotated using eggNOG-mapper149 (v2.1.9). Briefly, reads were mapped against eggNOG protein database (eggnog 5.0150) using DIAMOND151 (v2.0.15.153) in blastx mode. The default KEGG135 catalog was used as the pathway reference. For gene-level analysis of metabolic pathways involved in butyrate production, the corresponding KO IDs of each gene were used from the eggNOG-mapper’s output. For every gene, read counts were obtained by taking the sum for all corresponding KEGG values. Gene normalized abundances were measured as logarithm of counts per millions [log2(CPM + 1)] to adjust for differences in sequencing depths. Pathway normalized abundances were calculated by taking the sum of all of gene counts for each pathway and log-transformed and normalized to sequencing depth as log2(CPM + 1). Statistical tests for groups comparisons were done using two-sided Wilcoxon rank-sum test (FDR adjusted p-value < 0.05 considered to be significant) and were performed using R with visualizations created using ggplot2.

Metabolic network analysis

Metabolic support index (MSI) values for the species identified in Table 2 in relation to other gut microbial species were obtained based on metabolic network analysis, as described previously152. Briefly, MSI uses network flow analysis to quantify the extent of microbial metabolism that is enabled by the presence of other species in the community. The metabolic support network was visualized as a directed graph using Cytoscape153 (v3.8.0). Species that are known butyrate producers46,47 and receive the most metabolic support (based on in-degree of nodes) were highlighted in the network. The combined relative abundances of organisms supporting these butyrate producers were compared across age groups using the Wilcoxon test.

Phenotypic association analysis

Associations of microbial species with clinical phenotypic markers were identified based on linear regression using ‘glm’ function in R, with age and other markers serving as covariates for adjustment. Only species present in at least ten samples were considered for this analysis. Phenotypic markers available in two datasets (T2D and SG90) included BMI, fasting blood glucose, triglycerides, total cholesterol, HDL and LDL levels. For hs-CRP, AST, and ALT levels, gait speed, right handgrip strength, left handgrip strength, sleep duration, MMSE and free thyroxine levels, the associations were only tested for SG90 samples (without batch-correction) where this data was available.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.