Abstract
Calorie restriction (CR) extends lifespan and retards age-related chronic diseases in most species. There is growing evidence that the gut microbiota has a pivotal role in host health and age-related pathological conditions. Yet, it is still unclear how CR and the gut microbiota are related to healthy aging. Here we report findings from a small longitudinal study of male C57BL/6 mice maintained on either ad libitum or mild (15%) CR diets from 21 months of age and tracked until natural death. We demonstrate that CR results in a significant reduction in frailty index (FI), a well-established indicator of aging. We observed significant alterations in bacterial load, diversity, and compositional patterns of the mouse gut microbiota during the aging process. Interrogating the FI-related microbial features using machine learning techniques, we show that gut microbial signatures from 21-month-old mice can predict the healthy aging of 30-month-old mice with reasonable accuracy. This study deepens our understanding of the links between CR, gut microbiota, and frailty in the aging process of mice.
Introduction
The proportional population of older persons is growing across the globe1. This demographic shift will increase the prevalence of age-related disease and place a significant burden on health costs and social care. Moreover, increased longevity (i.e., lifespan) does not necessarily translate to better quality of life (i.e., healthspan)2. Thus, it is imperative to improve our understanding of mechanisms underlying aging processes and develop practical interventions to promote healthy aging and delay age-related diseases.
Aging is one of the most complex biological processes that affects a wide array of physiological, genomic, metabolic, and immunological functions9,10. These age-related functional changes can lead to organ and systemic decline, which ultimately results in death. There is now growing evidence that the gut microbiota interacts with these physiological functions, and thereby plays a pivotal role in host health and age-related pathological conditions3–5. The gut microbiota is regulated by a complex interplay between host and environmental factors, including age, diet, antibiotics, genetics, and lifestyle7,8. In turn, changes in the gut microbiota can alter host physiology, increasing the incidence and/or severity of many diseases that contribute to morbidity and mortality in later life, such as inflammatory bowel disease17, type 2 diabetes19, obesity20, cardiovascular disease21, and neurodegenerative disease22. During host aging, the gut microbiota undergoes dramatic changes in composition and function12–16. The gut microbiota of elderly people is different from that of adults14,23,24, and microbial compositions in the elderly correlate with measures of frailty, barrier dysfunction, gut motility, and inflammation25. Nevertheless, the extent to which these changes result from host aging or contribute to it remains unclear. Unlike other organs, the gut microbiota might not be expected to follow the same general trajectory of somatic senescence11.
Calorie restriction (CR), a dietary regimen that reduces the consumption of food without resulting in malnutrition, has been shown in animal models to retard development of age-related chronic diseases and extend the lifespan26–29. In addition to effects on host physiology, CR can also reshape the gut microbial community in both humans30,31 and animal models32–34. CR-induced alterations to the gut microbiome might play a role in extending lifespan and healthspan and delaying the onset of age-related disorders. In this study, we evaluate how the gut microbiota changes during the aging process in mice and test whether gut microbial features can predict healthy aging. To do this, we performed quantitative PCR (qPCR) targeting the 16S rRNA gene and 16S rRNA gene sequencing of bacterial DNA extracted from fecal samples from a cohort of aging male mice tracked from 21 months of age. We investigated associations between these microbial signatures and biomarkers of host condition, including weight, food intake, hematological markers, and frailty index (FI), a validated biomarker of biological age that is a strong predictor of mortality, morbidity, and other age-related outcomes35. Examining how signatures in the gut microbiota predict future aging status can illuminate the utility of the gut microbiota as an early indicator of healthy aging.
Results
Experimental design
The experimental design is shown in Fig. 1. Following baseline phenotypic measurements (body weight, food intake, frailty index, grip strength, and fecal collection), adult male C57BL/6 mice were randomized at 21 months of age into ad libitum diet (AL, n=14) or mild calorie restriction diet (CR, 15% fewer calories than their peers consuming an ad libitum diet, n=8) groups and followed longitudinally until death. From each birth cohort that we received, we randomized the mice equally into groups to avoid a strong birth-cohort effect. We repeated phenotypic measurements after 9 months (30 months of age) and recorded survival. We performed qPCR targeting the 16S rRNA gene as well as 16S rRNA gene sequencing on 44 stool samples, collected at 21 and 30 months of age, from 22 mice.
The association of the physiological characteristics with chronological age
The mouse clinical frailty index (FI) is based on established clinical signs of deterioration in mice36,37. Briefly, the clinical assessment includes evaluation of the integument, the musculoskeletal system, the vestibulocochlear/auditory systems, ocular and nasal systems, digestive system, urogenital system, respiratory system, signs of discomfort, body weight, and body surface temperature. FI score is continuous from 0-1, with higher values indicating worse frailty. A cutoff of 0.21 has been previously used in rodents38 to stratify frailty as either high (frail: FI≥0.21) or low (not frail: FI<0.21). But as mice reaches 30 months old, they all become frail with higher FI score (FI>0.21) in our study. Indeed, as shown in Fig. 2a and Fig. S1a, FI score significantly increased with chronological age from 21 to 30 months at the population level (P-value = 4.8e-06, Wilcoxon signed-rank test). Hence, instead of using a fixed FI score cutoff, in this work we used the median value of FI change (denoted as ΔFI) to delineate healthy versus normal aging. Specifically, we calculated ΔFI between month 21 and 30 for each mouse, and then we dichotomized those mice at month 30 into two groups based on the medium value of their ΔFI: ‘healthy aging’ (age in weeks: mean 121.78□± standard deviation 3.88; ΔFI: 0.088□±□0.038; FI: 0.342□±□0.048; n=11); and ‘normal aging’ (age in weeks: 121.42□±□4.07; ΔFI: 0.179□±□0.034; FI: 0.398□±□0.055; n=11). CR diet was associated with a lower level of ΔFI at month 30 than AL diet (Fig. 2b, P-value = 0.029, Wilcoxon–Mann–Whitney test). In particular, 87.5% (7/8) of mice with CR diet belonged to the healthy aging group compared to just 36.4% (4/11) of mice fed ad libitum. These results suggest that CR had a beneficial effect on aging, consistent with previous studies27.
We found that the body mass (BM) of mice generally decreased during aging (Fig. 2c, P-value = 0.0011, Wilcoxon signed-rank test), which was contributed by healthy aging mice due to the fact that most of them (63.64%) were from the CR group (Fig. S1b). At 30 months of age, the BM of the healthy aging mice was significantly lower than the normal aging (Fig. 2c, P-value = 0.028, Wilcoxon–Mann–Whitney test) and baseline mice (Fig. S1b, P-value = 0.0049, Wilcoxon signed-rank test). To better understand this finding, we calculated delta change of BM (ΔBM) between month 21 and 30 for each mouse. The ΔFI was positively associated with ΔBM (Fig. 2d, ρ = 0.3888, Spearman correlation), suggesting that a normal aging mouse (with large ΔFI) is associated with an increasing level of BM. In addition, we found that the BM in healthy aging mice gradually decreased over time (Fig. S2a), especially in those mice with CR diet (Fig. S2b). Additionally, normal aging mice showed rapid loss of BM after some time points (Fig. S2). Using Kaplan–Meier survival analysis, the differences in cumulative survival rates were not statistically significant between healthy and normal aging mice (Fig. S3, P-value = 0.23, log-rank test). However, the healthy aging mice showed qualitatively longer lifespan (134.36 ± 9.43) than normal aging (131.06 ± 7.53) mice (P-value = 0.313, Wilcoxon–Mann–Whitney test), as some mice from the healthy aging group lived substantially longer.
Aging-related changes in gut microbial community
Using universal 16S qPCR, we first measured the total bacterial load (BL) in the stool samples (Fig. 2e and Fig S1c). The results showed the total BL detected in these healthy aging mice was higher than the BL present in the normal aging mice (Fig. 2e). For the changes of total BL over time (ΔBL), we found ΔFI was inversely associated with ΔBL (Fig. 2f, ρ = −0.2107, Spearman correlation), suggesting that a normal aging mouse (larger ΔFI) is associated with a decreasing total BL.
We then measured the gut microbial community compositions of those stool samples using 16S rRNA gene sequencing (see Methods, Table S1). Phylum-level taxonomic profiles of the gut microbiome samples of those mice are shown in Fig. 3a. Consistent with previous studies39,40, we found that Bacteroidetes, Firmicutes and Verrucomicrobia were the most dominant phyla in the murine gut microbiota. Notable age-related compositional shifts included an enrichment in Firmicutes, and reduction in Bacteroidetes and Verrucomicrobia, although such trade-offs among dominant phyla are expected a priori in relative abundance data. Moreover, the Firmicutes/Bacteroidetes ratio of the gut microbiota increased with age (Fig. 3b, P-value = 0.0025, Wilcoxon signed-rank test). Both healthy aging and normal aging mice showed higher values for this ratio compared with baseline mice (Fig. S4a).
Using the Shannon diversity and Simpson index as alpha diversity measures, we found that alpha diversity increased with age (Fig. 3c,d and Fig. S4b,c), consistent with a previous mouse study41. Interestingly, we found that the Shannon diversity was only significantly higher in healthy aging mice compared to baseline mice (Fig. S4b, P-value = 0.019, Wilcoxon signed-rank test). In addition, a clear separation (permutational multivariate analysis of variance (PERMANOVA) test, P-value = 0.0001, Bray-Curtis dissimilarity) could be seen between mice at 21 and 30 months of age in the principal coordinate analysis (PCoA) plot based on Bray-Curtis dissimilarity (Fig. 3e). Indeed, PERMANOVA test indicated significantly altered microbial compositions for both healthy aging (P-value = 0.0004) and normal (P-value = 0.0086) aging mice between baseline and 30 months of age (Fig. S4d). However, we found no significant difference between healthy aging and normal aging mice at both 21 (P-value = 0.8747) and 30 (P-value = 0.3536) months of age. Bray-Curtis dissimilarity was higher among individuals within normal aging mice compared to baseline mice (Fig. S4e, P-value = 4e-08, Wilcoxon signed-rank test) or healthy aging mice (Fig. 3f, P-value = 0.015, Wilcoxon–Mann–Whitney test). This suggests that normal aging is characterized by high variations in gut microbiota between individuals.
The effect of aging on hematology and associations between gut microbiota and blood markers
Aging is associated with a decline in immune system function at multiple levels42. To explore aging-related immune system modifications, we measured hematological parameters over time (Table S2). We found that the mice at 30 months of age tended to have higher level (with P value <0.05) of neutrophils percentage, neutrophil to lymphocyte ratio (NLR), monocytes percentage (MOp, % of leukocytes), red cell distribution width (RDW, % variation), and mean platelet volume (MPV, fL), but lower level (with P value <0.05) of white blood cell (WBC, k/uL), lymphocytes (LY, k/uL), lymphocytes percentage (LYp, % of leukocytes), red blood cell (RBC, M/uL), hemoglobin (Hb, g/dL), Mean corpuscular volume (MCV, fL) and hematocrit (HCT, % volume) when compared with mice at 21 months of age. Specifically, higher NLR (an important biomarker of systemic inflammation43) levels on 30-month-old mice were mainly observed in normal aging mice (P value = 0.016). Here P values were all calculated from the Wilcoxon–Mann–Whitney test, adjusted with the Benjamini–Hochberg FDR method. These results confirm prior observations that high levels of inflammation are not an inevitable consequence of aging, but rather associated with normal or unhealthy aging. Moreover, at 30 months of age, we found that normal aging mice had significantly higher MPV but normal PLT.
Given the effects of aging process on hematology, we next used MaAsLin2 (multivariate analysis by linear models44) to evaluate the associations between microbial taxa and blood markers. These linear mixed models accounted for within-individual correlation from the study’s repeated sampling design, as well as occasional missing observations at some time points. To control for potential confounding variables, we added four covariates into the model as fixed effects, including diet treatment, cohort, cage, and body mass. In addition, each mouse’s identifier treated as random effect. A total of 24 ASVs (amplicon sequence variant) features were significantly associated with at least one blood marker (Fig. 4, q-value ≤ 0.2, Table S3). In general, blood markers correlating most with microbial taxa included MCV, LY and NLR. For example, MCV was inversely associated with the abundance of ASV 3949 (Anaerotruncus, q-value = 2.38e-14) and ASV3729 (Clostridium aldenense, q-value = 1.52e-6), and LY was positively associated with ASV890 (Ruminococcaceae, q-value = 0.0004), ASV2868 (Oscillibacter, q-value = 0.015), and ASV2973 (Intestinimonas butyriciproducens, q-value = 0.035). NLR was positively associated with ASV5690 (Flavonifractor plautii, q-value = 0.04) and ASV555 (Acetatifactor muris, q-value = 0.048), and negatively associated with ASV2878 (Lachnospiraceae, q-value = 0.028), ASV4558 (Bacteroidales, q-value = 0.146), and ASV1970 (Clostridium XlVa, q-value = 0.189).
Microbial taxa related to frailty index and healthy aging
We next investigated the FI in relation to the microbial features using MaAsLin2 in which diet, cohort, cage, and body mass were included as fixed effects and each mouse’s identifier was included as a random effect. We observed a set of 14 microbial features that were strongly linked to FI (Fig. 5, q-value ≤ 0.2, Table S4). Consistent with previous reports that the abundance of the Clostridium sensu stricto genus increases with aging45–47, ASV3100 (Clostridium sensu stricto: q-value = 0.021) was positively associated with the FI. Clostridium XlVa48 (ASV2882, q-value = 0.048 and ASV1101: q-value = 0.112) and Subdoligranulum variabile49 (ASV157, q-value = 0.153), known as important producers of butyrate, were found to be negatively associated with FI. We also found inverse associations of the FI with taxa such as ASV847 (Phocea massiliensis, q-value = 0.069), ASV 1726 (Parabacteroides goldsteinii, q-value = 0.083), and ASV1123 (Enterorhabdus, q-value = 0.090). A previous study linked Parabacteroides goldsteinii with reduction of intestinal inflammation and enhancement of cellular mitochondrial and ribosomal activities in the colon50.
To examine potential gut microbial signatures of late-life aging, we performed differential abundance analysis using ANCOM51 (analysis of composition of microbiomes). ANCOM identified multiple gut microbiota signature that were significantly different between baseline and 30 months of age in healthy aging (Fig. S5a and Table S5) and normal aging (Fig. S5b and Table S6) mice. Most of these features were also identified when comparing all mice between 21 and 30 months of age as a group (Fig. S6 and Table S7). Intriguingly, we found 7 ASVs that significantly and concordantly increased with age in both healthy aging and normal aging groups (Fig. S5), including ASV5550 (Lachnospiraceae), ASV5652 (Lachnospiraceae), ASV806 (Lachnospiraceae), ASV5435 (Muribaculum intestinale), ASV3224 (Clostridium cocleatum), ASV5628 (Muribaculum intestinale), and ASV3370 (Muribaculum intestinale), hinting at a universal murine microbial signature of aging. To assess how the microbial features links with healthy aging, we calculated the differential abundance of features between healthy aging and normal aging groups at both 21 and 30 months of age (Fig. S7). Our data found 6 (Fig. S7a, Table S8) and 9 (Fig. S7b, Table S9) ASVs were significantly associated with aging status at baseline and 30 months of ages, respectively. In particular, a set of microbial features were significantly enriched in healthy aging mice at 30 months of age, for example ASV648 (Akkermansia muciniphila), ASV73 (Ruminococcaceae), and ASV2756 (Acetatifactor muris). A. muciniphila has been observed previously to prevent the age-related decline in thickness of the colonic mucus layer and attenuate inflammation in old age52. Here, this microbial feature was detected and shown to be positively associated with healthy aging. Normal aging mice showed increased ASV3370 (Muribaculum intestinale), ASV3100 (Clostridium sensu stricto), ASV3939 (Turicibacter sanguinis), and ASV1123 (Enterorhabdus) compared with healthy aging mice. Consistent with positive relationship between FI and ASV3100 (Clostridium sensu stricto), we found that this feature was significantly higher in the normal aging group.
Gut microbiota-based machine learning model to predict healthy aging
As microbial compositions were associated with aging status, we sought to determine whether the microbial features observed in mid-life could predict healthy aging in later life. To achieve that, we employed an Elastic-net (ENET) logistic regression model to predict healthy aging. Specifically, the ENET model trained with ASVs (present in at least 10% samples) achieved an accuracy of 0.5 (11/22) with leave-one-out cross-validation (LOOCV) (Fig. 6a). In principle, we can apply feature selection techniques to choose a subset of features from the dataset. However, to improve the biological meaning of the model, we then only selected the microbial features that significantly associated with FI. This approach included a microbial signature comprised of 14 ASVs (Fig. 6b) from the gut microbiota of 21-month old mice that exhibited power in predicting the healthy aging status of 30-month old mice with a LOOCV accuracy of 0.773 (17/22) (Fig. 6a). Notably, we also observed that Clostridium sensu stricto and Enterorhabdus were significantly overrepresented in normal aging mice at 30 months of age. A previous study found that Clostridium sensu stricto was significantly enriched in early onset necrotizing enterocolitis subjects53. Enterorhabdus, a member of the family Coriobacteriaceae, has been isolated from a mouse model of spontaneous colitis54. These findings were consistent with higher level of NLR in normal aging mice, which was used as a marker of systemic inflammation. This may partially explain the ability of these features to predict healthy aging over the subsequent 9 months. Finally, we validated our model by generating a null model with randomly selected features (number of features=14, times=100), which yielded a mean LOOCV accuracy of 0.443 (Fig. 6a).
Discussion
Over the last few decades, global average life expectancy has increased dramatically, resulting in a proportionately larger aging population. Currently, chronological age is the most widely used indictor of aging, yet it provides limited information on the quality of life during the aging process. Understanding how to promote healthy aging will be key to increasing healthspan. Evidence is emerging that the gut microbiota is intrinsically linked with energy metabolism and the aging process55–58. In this study, we observed that the mouse gut microbiota is associated with healthy aging on late-life aged mice. And we identified a specific stool-microbiota-derived signature of aging that yielded a reasonable accuracy for the prediction of healthy aging.
A better predictor of mortality and morbidity in humans than chronological age is the Frailty index (FI)61. The FI has been reverse translated into a tool for mice which includes 31 non-invasive parameters across a range of systems37,62. Previous studies applied 0.21 as a cut-off point of FI to stratify between high frailty (≥0.21) or low frailty (<0.21)38,63,64. Given this specific threshold provides limited insight into the aging process, we instead employed the ΔFI (FI changes between 30 and 21 months of age) to quantify the ability to maintain health conditions during aging. Indeed, those mice with higher ΔFI (based on median value) were more vulnerable and frail. In our study, we only included the mice with basic measurements and biological samples at both 21 and 30 months, resulting 22 male mice that were fed either AL (n=14) or CR (n=8) diets. To avoid the issue arising from imbalanced sample size, we stratified the mice to healthy and normal aging mice based on the ΔFI. As expected, 87.5% (7/8) of mice with CR diet belonged to the healthy aging group compared to just 36.4% (4/11) of mice fed AL.
Although several previous studies demonstrated the links between gut microbiota and aging in mice, these studies mainly focused on the comparison between different growth stages65–67. In this study, we examined the gut microbiota collected at 21 and 30 months of age from 22 mice and measured the aging status. Concordant with previous reports, we found that aging was associated with increased alpha diversity67. In particular, only healthy aging mice showed significantly increased Shannon diversity with age. Consistent with previous work68, our study also linked aging to an increase in interindividual variation in gut microbial community composition, with interindividual variation being especially high in the normal aging group. This suggested that the unhealthy aging related changes in the gut microbiota are likely stochastic, leading to community instability. Our study also linked FI to several microbial features such as ASVs from Clostridium sensu stricto, Clostridium XlVa, Enterorhabdus, and Phocea massiliensis. Importantly, we constructed a machine learning model that can predict healthy aging with LOOCV accuracy of 0.773 (17/22) based on these FI related microbial features. And these microbial features may be further driven by CR after 21 months of age. Indeed, we found that some predictive features (e.g., ASVs from Clostridium sensu stricto and Enterorhabdus) were only identified as differentially abundant taxa at 30 months of age. These findings suggest that key microbial taxa could potentially serve as biomarkers of aging and might contribute to the pathophysiology of aging, although the latter possibility remains to be determined.
We acknowledge the following limitations of this study. First, the sample size of the experimental cohort is relatively small and limited to male mice. Second, 16S rRNA gene sequencing limits our ability to establish associations at the strain level, suggesting that future studies with shotgun metagenomics sequencing will increase resolution. Third, the association between healthy aging and microbial taxa identified in this study does not demonstrate causality. Thus, additional research is needed to validate the mechanism behind these essential findings. Finally, the generalization of the machine learning-based gut microbial signature of aging to other murine cohorts and to humans remains unknown. However, the strengths of the study include a prospective study design, detailed phenotyping of mice, and assessment of accuracy using gut microbial features to predict healthy aging by machine learning model.
In conclusion, we evaluated the impact of age-related changes in gut microbiota on the course of aging in late-life male mice to assess a microbiota signature associated with healthy aging. Our study suggests the possible interaction between specific gut microbiota and aging status, and motivates future work that could establish causality and the potential of future microbiota-targeted interventions to increase healthy aging.
Methods
Study population and sample collection
In our study, we only included the mice with basic measurements and biological samples at both 21 and 30 months, resulting 22 C57BL/6 male mice (NIA Aging Colony). Following baseline measurements (body mass, food intake, frailty index and fecal collection), we randomly divided these mice into two diet groups, fed either ad libitum (AL, n=14) with standard chow or under mild (15%) calorie restriction (CR, n=8) and followed longitudinally until death. Mice were fed a standard chow based upon AIN-93G (Custom diet #A17101101, Research Diets, New Brunswick, NJ). CR was initiated over a period of two-weeks in a step-down fashion (10% CR, 15% CR) to ensure no loss on mice as they transition to the restricted feeding paradigm. Fecal samples (non-fasted) were collected in the morning (8.30am-11.30am) into sterile tubes and frozen at −80 °C until future analysis.
The measurement of frailty index
Frailty was measured using the validated 31-parameter mouse clinical frailty index as described previously36,37. Briefly, the clinical assessment includes evaluation of the integument, the musculoskeletal system, the vestibulocochlear/auditory systems, ocular and nasal systems, digestive system, urogenital system, respiratory system, signs of discomfort, body mass, and body surface temperature. FI score is continuous from 0-1, with higher values indicating worse frailty37.
Hematology analysis
25μL of whole blood obtained via submandibular bleeding was combined with 1μL of EDTA to prevent clotting. The sample was analyzed using a Hemavet 950 veterinary (Drew Scientific, Miami Lakes, FL) multi-species hematology system using standard settings.
Estimation of bacterial load by quantitative PCR
To estimate gut bacterial load in our 44 fecal samples, we performed quantitative PCR (qPCR) targeting the 16S rRNA gene using the same primers employed for 16S rRNA gene sequencing (515F and 806R). Briefly, 2□μl of template DNA was combined with 12.5 μl PerfeCTa SYBR Green SuperMix Reaction Mix (QuantaBio, Beverly, MA), 6□μl nuclease-free H2O, and 2.25□μl of each primer. Amplification was performed on a Bio-Rad CFX384 Touch (Bio-Rad, Hercules, CA) in the Bauer Core Facility at Harvard University using the following cycle settings: 95□°C for 10□min, followed by 40 cycles of 95□°C for 15□s, 60□°C for 40□s and 72□°C for 30□s. Reactions were performed in triplicate with the mean value used in statistical analyses. Cycle-threshold values were standardized against a dilution curve of Escherichia coli genomic DNA at the following concentrations (ng/μL): 100, 50, 25, 10, 5, 1, 0.5, plus a no-template (negative) control. Bacterial DNA concentrations were normalized to 16S copies/μL, then multiplied by the total extracted DNA volume (50 μL) and divided by the grams of fecal matter utilized in the extraction of template DNA (varied), allowing us to report gut bacterial load as 16S rRNA gene copies per gram of feces.
DNA isolation and 16S rRNA gene sequencing
Gut microbial DNA was isolated using the DNeasy PowerSoil Pro Kit (Qiagen) and PCR-amplified using barcoded primers targeting the V4 region of the bacterial 16S rRNA gene [515F (GTGYCAGCMGCCGCGGTAA) and 806R (GGACTACNVGGGTWTCTAAT); Integrated DNA Technologies]. The following thermocycler protocol was used: 94°C for 3 min, 35 cycles of 94°C for 45 s, 50°C for 30 s, and 72°C for 90 s, with a final extension at 72°C for 10 min. Triplicate PCR reactions for each sample were pooled and amplification was confirmed by 1.5% gel electrophoresis. 16S rDNA amplicons were cleaned with AmpureXP beads (Agencourt) on a per-sample basis, then quantified using the Quant-iT Picogreen dsDNA Assay Kit (Invitrogen). Amplicons were pooled evenly by DNA content and sequenced on an Illumina HiSeq (1 x 150 bp) at the Bauer Core Facility at Harvard University, generating 234,631 ± 110,737 (mean ± SD) sequences per sample passing filter (range: 75,898 to 391,101) (Table S1).
Microbiota composition by 16S rRNA gene amplicon analysis
Raw sequencing data was processed and analyzed using Quantitative Insights into Microbial Ecology 2 (QIIME2) pipeline70. Single-end sequences were first demultiplexed using the barcode sequences. The sequencing reads were then quality filtered, denoised, and merged using DADA271 to generate the ASV feature table. For taxonomy classification, ASV feature sequences were aligned against SILVA reference database72. Additional species level assignment to the NCBI RefSeq73 16S rRNA database supplemented by RDP74 was accomplished using the assignTaxonomy and addSpecies functions of DADA2 R package.
Statistical analysis
Microbial alpha and beta diversity measures were calculated at the ASV level using the vegan package in R. Principal coordinates analysis (PCoA) plot was generated with Bray-Curtis dissimilarity. The difference in microbiome compositions by different groups were tested by the permutational multivariate analysis of variance (PERMANOVA) using the “adonis” function in the R’s vegan package. All PERMANOVA tests were performed with the 9999 permutations based on the Bray-Curtis dissimilarity. Differences between groups were analyzed using a Wilcoxon–Mann–Whitney test (unpaired) or Wilcoxon signed rank test (paired). The survival probability was computed by the Kaplan-Meier method.
MaAsLin244 (multivariate association with linear model) was used for adjustment of covariates when determining the significance of ASVs contributing to specific hematological variables and FI, while accounting for potentially confounding covariates. The linear mixed models included each mouse’s identifier as random effects and other potential confounders as fixed effects. To be qualified for downstream analyses, a ASV feature needed to be detected at least 10% of samples. The P values were then adjusted using the Benjamini–Hochberg FDR method. The microbial features with corrected q□value <□0.2 were presented. For differential abundance analysis, we used ANCOM51 (analysis of composition of microbiomes), with a Benjamini–Hochberg correction at 5% level of significance, and adjusted for cage, cohort, body mass, and diet. Only the ASVs presented at least 10% of samples were included. To develop a model capable of predicting healthy aging, we implemented Elastic-net (ENET) using R’s caret package. Custom machine learning process was conducted using microbial features at 21 months of age to predict aging status at 30 months of age. We first trained our model with all of microbial features. To further improve the biological plausibility, we then only included the microbial features significantly associated with FI. A total of 14 ASVs were selected based on the q value (q□<□0.2) from the MaAsLin2 model. Leave-one-out cross-validation (LOOCV) was applied with the trainControl function. To further validate our model, a null model was generated with random selected feature (number of features=14, times=100). All statistical analyses were performed using R.
Data availability
Raw sequencing reads have been deposited in NCBI under accession number PRJNA739980.
Funding
Y.-Y.L. acknowledges grants from National Institutes of Health (R01AI141529, R01HD093761, RF1AG067744, UH3OD023268, U19AI095219 and U01HL089856). S.J.M, J.R.M acknowledge support for this project from NIA (P01AG055369-01A1). R.N.C. and J.R.M. benefited from an Acceleration Award from the Harvard Chan School of Public Health. R.N.C. acknowledges related support from NIA/ORWH (R01AG049395). A.E.K is supported by an AFAR Irene Diamond postdoctoral award. D.A.S. is supported by the Paul F. Glenn Foundation for Medical Research and NIH grants R01DK100263 and R37AG028730.
Author contributions
Y.-Y.L, R.N.C., F.G. and J.R.M. conceived and designed the project. S.J.M., M.R.M. and A.E.K. performed the mice experiments. E.M.V., K.S.C., and R.N.C. performed the 16S rRNA gene sequencing and universal 16S quantitative PCR. S.K. performed all the data analysis. S.K. and Y.-Y.L. wrote the manuscript. All authors analyzed the results and edited the manuscript.
Declarations of interests
D.A.S. is a founder, equity owner, advisor to, director of, board member of, consultant to, investor in and/or inventor on patents licensed to Cohbar, Alterity, Galilei, EMD Millipore, Zymo Research, Immetas, and EdenRoc Sciences (and affiliates Arc-Bio, Dovetail Genomics, Claret Bioscience, MetroBiotech, and Liberty Biosecurity), Life Biosciences and Iduna. D.A.S. is an inventor on a patent application filed by Mayo Clinic and Harvard Medical School that has been licensed to Elysium Health. More information at https://genetics.med.harvard.edu/sinclair-test/people/sinclair-other.php. The other authors declare no competing interests.
Supplementary information
Acknowledgements
We thank Lusheng Huang, Congying Chen, Xu-Wen Wang and Zheng Sun for helpful discussions.