Abstract

   Frailty is an age-related geriatric syndrome. We performed a
   longitudinal study of aging female (n = 40) and male (n = 47)
   C57BL/6NIA mice, measured frailty index and derived metabolomics data
   from plasma. We identify age-related differentially abundant
   metabolites, determine frailty-related metabolites, and generate
   frailty features, both in the whole cohort and sex-stratified
   subgroups. Using the features, we perform an association study and
   build a metabolomics-based frailty clock. We find that frailty-related
   metabolites are enriched for amino acid metabolism and metabolism of
   cofactors and vitamins, include ergothioneine, tryptophan and
   alpha-ketoglutarate, and present sex dimorphism. We identify B vitamin
   metabolism related flavin-adenine dinucleotide and pyridoxate as
   female-specific frailty biomarkers, and lipid metabolism related
   sphingomyelins, glycerophosphoethanolamine and glycerophosphocholine as
   male-specific frailty biomarkers. These associations are confirmed in a
   validation cohort, with ergothioneine and perfluorooctanesulfonate
   identified as robust frailty biomarkers. Our results identify
   sex-specific metabolite frailty biomarkers, and shed light on potential
   mechanisms.

   Subject terms: Biomarkers, Metabolomics

Introduction

   With the success of medical innovations and public health
   interventions, people are living much longer. However, aging is highly
   heterogeneous and there is extreme variability in health and function
   amongst different individuals of the same age^[35]1. Such variability
   in health can be captured by the concept of ‘frailty’, a measurement of
   overall decline in health with age^[36]2. Frailty can be quantified
   using a frailty index (FI), which counts the proportion of age
   accumulated health-related deficits present in an
   individual^[37]3,[38]4. Higher FI values indicate a greater degree of
   frailty and are associated with an increased susceptibility to diseases
   and mortality^[39]4,[40]5. Frailty indices have been adopted for use in
   other mammals, including mice^[41]6.

   Whilst frailty assessments are commonly used in both the clinic and
   research, there are no accepted frailty biomarkers^[42]7, and very
   little is known about the underlying molecular mechanisms of frailty.
   Identification of frailty biomarkers would be beneficial in enabling
   earlier identification and tracking of frailty over time, development
   and testing of treatments and interventions^[43]7 and contribute to our
   understanding of the biological pathways underlying the development of
   frailty^[44]8. Metabolomics is an emerging field that enables
   comprehensive and quantitative metabolite assessment in biological
   samples. Circulating metabolites can provide a snapshot of the
   metabolic status of an individual, and as such have the potential to be
   both biomarkers, and provide insight into biological pathways changed
   in frailty and age.

   There are known metabolic changes in aging, and in fact many of the
   ‘hallmarks’ of aging are linked to unfavorable metabolic shifts^[45]9.
   Less is known about metabolic changes in frailty, although studies have
   shown that glucose intolerance and insulin dynamics are closely linked
   to physical frailty in both humans and mouse models^[46]10,[47]11.
   Metabolomics studies of aging in humans are beginning to identify
   specific metabolite markers^[48]12. Elevated high- and decreased
   low-density lipoproteins are well established in older individuals, and
   associated with poor clinical outcomes^[49]13. Changes in amino acids
   are observed in aging, including increased tyrosine and decreased
   tryptophan^[50]14,[51]15. Both lipids and amino acids are extensively
   related to nutrient sensing pathways, such as the mammalian target of
   rapamycin (mTOR)^[52]16 that acts as a central regulator in
   aging^[53]17. Oxidative stress and inflammation- related metabolites
   are also associated with aging, particularly acylcarnitines,
   sphingomyelins^[54]18, and cytochromes P450 metabolites^[55]19.
   However, the majority of these metabolomics studies are cross-sectional
   in design, comparing separate groups of young and old individuals, and
   there are few studies exploring how metabolites change longitudinally
   within the same individuals as they age^[56]20,[57]21. Although early
   metabolomics studies focused on associations with chronological age
   only, there is a growing focus on metabolomics studies of frailty in
   humans, and these studies have revealed associations with energy and
   nutrition metabolism^[58]22 and with amino acid
   metabolism^[59]23,[60]24. While these studies hint at a strong link
   between frailty and metabolism, they are limited by small sample sizes,
   and cross-sectional designs.

   Additionally, sex dimorphism in aging is widely observed across many
   levels. Most notably, at every age, women are more frail than men,
   despite having longer life expectancy^[61]25. There are also clear sex
   differences in the risk and prevalence of age-related diseases,
   including metabolic diseases^[62]26,[63]27. Many studies have revealed
   clear sex differences in metabolic aging across multiple tissues
   including blood^[64]28, brain^[65]29, and adipose tissue^[66]30. Sex
   differences in metabolites related to lipid metabolism, such as
   cholesterol^[67]30 and sex steroid hormones^[68]30, amino acids and
   acylarnitines^[69]31 are widely observed, and such differences can be
   age-dependent. The mechanisms, and especially metabolic mechanisms,
   underlying these sex differences in aging and frailty are not well
   understood, despite some recent efforts^[70]25,[71]32,[72]33. Although
   studies have explored baseline sex differences in circulating
   metabolites, as far as we are aware, there are currently no
   longitudinal metabolomics studies exploring sex differences in frailty.

   Here, we completed a longitudinal study of female and male mice and
   generated matched plasma metabolomics and frailty data across 5 time
   points. We use time-course and network analysis to identify age related
   metabolites, apply machine learning algorithms to select
   frailty-related metabolite features, perform an association study on
   frailty features, and build a metabolite frailty clock. We reveal that
   age-related metabolites are enriched in lipid metabolism, and suggest
   that amino acid metabolism and metabolism of cofactors and vitamins are
   enriched for frailty-related metabolites. In particular, we demonstrate
   strong sex differences in metabolite features and their associations
   with frailty. We confirm these findings in a validation cohort,
   specifically finding consistent associations for 9 candidate frailty
   biomarkers, and the metabolite frailty clock achieves better prediction
   performance than age and sex alone, but only in male samples. Our
   results provide candidate metabolomic biomarkers of frailty for future
   testing in clinical studies, and provide insights into possible
   mechanisms underlying sex differences in frailty and aging.

Results

Metabolomics data variation

   We performed a longitudinal study of female (n = 40) and male (n = 47)
   C57BL/6NIA mice at 5 time points, and derived plasma metabolomics data
   for a total of 321 samples that have valid metabolomics data (Table
   [73]1). In order to investigate aging- and frailty- related metabolites
   and mechanisms in naturally aging mice, we used non-NMN treated mice
   (female, n = 20; male, n = 24) as the discovery cohort for the ensuing
   analysis (Fig.[74]1). To investigate the metabolomic data variation, we
   performed a principal component analysis (PCA), including a set of 781
   metabolites. The PCA plot indicated clear separation of samples across
   time points (by PC1) and by sex (by PC2) (Supplementary Fig. [75]1a).
   We then performed linear regression analyses on PCs and observed
   clustering of factors of interest (e.g., sex, time point, mouse ID) in
   the associations with PCs (Supplementary Fig. [76]1b) based on
   p-values. We selected time points and sex as representative variables
   in the ensuing analysis as they showed the smallest p-values among the
   factors within the same cluster. Mouse ID also showed an association
   with PC2 at a significant level and was included to account for
   repeated measurements on the same mouse.

Table 1.

   List of female and male samples
   Discovery cohort Validation cohort
   Female (n = 20) Male (n = 24) Female (n = 20) Male (n = 23)
   Time points Samples^a Age^b Frailty index^b Samples Age Frailty index
   Samples Age Frailty index Samples Age Frailty index
   BL 19 393 0.14 [0.12, 0.16] 21 386 0.19 [0.18, 0.21] 16 393 0.15 [0.13,
   0.16] 18 386 0.19 [0.18, 0.22]
   T2 20 541 0.24 [0.22, 0.26] 22 539 0.23 [0.21, 0.25] 19 541 0.21 [0.20,
   0.23] 23 539 0.22 [0.20, 0.24]
   T3 17 624 0.29 [0.26, 0.31] 23 635 0.26 [0.24, 0.28] 19 624 0.24 [0.21,
   0.26] 23 635 0.25 [0.23, 0.32]
   T4 11 756 0.28 [0.26, 0.34] 19 773 0.27 [0.23, 0.28] 12 756 0.28 [0.23,
   0.32] 20 773 0.27 [0.26, 0.31]
   T5 NA NA NA 9 910 0.37 [0.35, 0.45] 4 899 0.33 [0.29, 0.36] 6 910 0.44
   [0.40, 0.45]
   [77]Open in a new tab

   ^aNumber of samples that have valid metabolomics data.

   ^bAge (days) at assessment for frailty.

   ^cMedian [Lower quartile, Upper quartile].

Fig. 1. Schematic diagram of the workflow.

   [78]Fig. 1
   [79]Open in a new tab

   The longitudinal study starts with female (n = 40, yellow circles) and
   male (n = 47, blue circles) C57BL/6NIA mice. Frailty was assessed and
   blood samples were collected at 5 time points from BL to T5 (exact days
   of (experiments are shown in Table [80]1). Plasma samples were derived
   from blood samples and were then subjected to metabolite
   quantification. In order to investigate metabolites related with
   natural aging and frailty, feature selection, sex stratified analysis,
   association study and the frailty clock were all performed in the
   control samples without intervention as the discovery cohort. The
   metabolite biomarkers and a metabolite clock for frailty were then
   tested in the validation cohort.

Metabolomic signatures of aging across sexes

   To evaluate changes of metabolites with age, we performed an
   association study. There are 549 metabolites presenting significant
   associations (adjusted p-value < 0.05), among which, 136 (17.4% of 781
   metabolites) increase and 413 (52.9%) decrease with age. To further
   investigate changes of metabolites over time, we performed metabolite
   differential abundance analyses to identify metabolites that were
   related to general aging. That is, metabolite abundances that
   significantly changed for these mice across the sampled time points. We
   considered the pattern of metabolite abundance globally over time, by
   fitting a time series smoothing spline, accounting for mouse ID and
   sex. We found 527 (67.5% of total detected metabolites) differentially
   abundant metabolites (DAMs) over the time-course within all mice (both
   females and males) (Supplementary Fig. [81]2a).

   In order to select subsets of metabolites with similar abundance over
   the time course and, more importantly, highly related to aging
   independent of sex, we performed co-abundance network analysis on the
   527 metabolites derived above. We derived two metabolite subsets,
   subset1 (n = 200) and subset2 (n = 125) (Supplementary Fig. [82]2b) of
   which the eigenvalues showed strong associations (p < 0.001) with age,
   presenting a generally decreasing trend in abundance in aging (Fig.
   [83]2a). Significantly higher proportions of metabolites within the
   amino acids super-pathway were observed in subset 1 (69 metabolites,
   34.5% of subset1, χ^2 (df = 2, N = 781) = 25.3, p < 0.001) and those in
   the lipids super-pathway for subset 2 (100 metabolites, 80% of subset2,
   χ ^2 (df = 2, N = 781) = 141.1, p < 0.001), compared to the rest of
   metabolites. Metabolite set enrichment analysis results on two subsets
   aligned with the above classification, with amino acid metabolism
   (subset1) and lipid metabolism (subset2) pathways over-represented
   (Supplementary Fig. [84]2c, d). To further select metabolites that play
   important roles in aging, we selected 86 hub metabolites, 46
   metabolites from subset1 and 40 from subset2, based on module
   membership in the network and significance (correlation coefficient
   between eigenvalue and age) (Supplementary Fig. [85]2e). These 86
   metabolites were defined as core age-related metabolites in the ensuing
   analyses, and include guanidinoacetate, methylamalonate (MMA) and
   sphingomyelin species (Supplementary Table [86]1). 54.7% of these
   metabolites (47 total, 9 from subset1 and 38 from subset2) are from the
   lipid super-pathway, also evidenced by enrichment analysis (Fig.
   [87]2b). This result suggests changes in lipid metabolism are
   associated with general aging, and their potential mechanistic role in
   aging should be further explored.

Fig. 2. Sex independent age-related differentially abundant metabolites in
longitudinal study.

   [88]Fig. 2
   [89]Open in a new tab

   Differential abundance analysis was performed using all samples
   (excluding time point T5) in the study. Sex-independent age-related
   differentially abundant metabolites (DAMs) were selected from
   comparisons of the mixture of female and male samples at different time
   points and by controlling for a 5% Benjamini–Hochberg false discovery
   rate (adjusted p-values < 0.05). These DAMs were then subjected to a
   co-abundance analysis, and subset1 and subset2 were determined to be
   significantly associated with age by linear mixed models. a Dynamics of
   metabolite abundance in each sex, derived from two subsets (subset1,
   n = 200 metabolites; subset2, n = 125). After determining the hub
   metabolites based on metabolite correlation with age and module
   membership, hub metabolites were subjected to metabolite set enrichment
   analysis. b Over-represented pathways (y-axis) from the hub metabolites
   from the two subsets. The number of hits (metabolite) from the hub
   metabolites set is shown by x-axis, ratio of the hit number to total
   metabolites in the enriched pathway is represented by dot size and
   p-value is colored by levels.

Sex specific metabolomic signatures of aging

   To identify sex-specific metabolomic signatures, we investigated DAMs
   within 1) females only (significant change in abundance in the whole
   time frame), n = 498 DAMs (63.8%, 498/781), 2) males only, n = 253 DAMs
   (32.4%, 253/781); and 3) sex differences (significantly differentially
   abundant in females and males considering the whole time frame),
   n = 331 DAMs, (42.4%, 331/781). The results suggest significant sex
   differences in metabolite abundance in the aging process.

   It was interesting to observe a common set of 97 metabolites after
   merging the above three sets of DAMs with the 527 DAMs derived from the
   mixture of both sexes (sex-independent) (Fig. [90]3a; Supplementary
   Table [91]2). These metabolites not only were related to aging in both
   sexes, but also presented sex differences in aging (Supplementary Fig.
   [92]3). Notably, these included 8 acylcarnitines, for instance,
   oleoylcarnitine (C18:1) and palmitoleoylcarnitine (C16:1). Among the 97
   metabolites, 41 are in lipid and 22 in amino acid super pathways,
   representing 64.9% of the 97 metabolites. Enrichment analysis revealed
   11 KEGG pathways over-represented (Fig. [93]3b), mostly within lipid
   metabolism and digestive system pathways.

Fig. 3. Comparisons of differentially abundant metabolites determined in four
groups.

   [94]Fig. 3
   [95]Open in a new tab

   Differential abundance analysis was performed using all samples
   (excluding time point T5) in the study. Age-related differentially
   abundant metabolites (DAMs) were determined by comparisons within four
   groups: the mixture of females and males (sex independent), female
   specific, male specific, and sex differences, and by controlling for a
   5% Benjamini–Hochberg false discovery rate (adjusted p-values < 0.05).
   a UpSet plot showing the common DAMs derived from the comparisons. b
   Over-represented pathways (y-axis) from the 97 common metabolites of
   four groups by metabolite set enrichment analysis. c Over-represented
   pathways (y-axis) from the 187 female-specific metabolites markers that
   also present sex differences. The number of hits (metabolite) from the
   hub metabolites set is shown by x-axis, ratio of the hit number to
   total metabolites in the enriched pathway is represented by dot size
   and p-value is colored by levels.

   Apart from the common set, among the 331 metabolites that are different
   between males and females across the investigated timeframe, there are
   only 21 metabolites that present sex differences (distinct abundance in
   females and males) and do not present significant abundance differences
   over time (Supplementary Table [96]2). This indicates that the majority
   of sex differential metabolites also change with age. In terms of
   female-specific metabolites that are changed with age, that is,
   metabolites detected in both female and sex difference DAM sets, the
   187 in this category included amino acids and acylcarnitines and were
   enriched for amino acid metabolism pathways (Fig. [97]3c; Supplementary
   Table [98]2). Many fewer male-specific age-related metabolites were
   observed, with a total of only 23, including phosphocholine and
   spermine. One metabolite, corticosterone, changed with age in both
   females and males separately (Supplementary Fig. [99]4), but was not
   detected to change with age when the entire cohort of mice was
   considered. Altogether, we found metabolites involved in lipid
   metabolism and digestive system pathways are associated with general
   aging and present strong sex differences. Specifically, amino acid
   metabolism-related metabolites are associated with aging in female
   mice.

Sex independent metabolite features of frailty

   Having identified age-related metabolites, we were interested to
   identify metabolites specifically associated with frailty. Frailty is a
   complex geriatric syndrome. In the discovery cohort, frailty and age
   were highly correlated (repeated measures correlation r = 0.77,
   p < 0.001) and frailty increases with age in both females and males
   (Supplementary Fig. [100]5). Moreover, we observed variations in mice
   of the same time point (age) (Table [101]1). For each mouse at a
   certain time point, FI is composed of the base FI (median FI of the
   corresponding sex and age group) and devFI (the deviation of individual
   FI from the median FI at corresponding age- and sex-specific group). By
   definition, base FI is highly related to age, but devFI is age
   independent (Supplementary Fig. [102]5).

   In order to find metabolites related to frailty, we performed feature
   selection outlined in Fig. [103]4a. We investigated metabolites that
   were related to both FI and devFI, by performing feature selection with
   elastic net regularization, via a 100 times repeated 5-fold cross
   validation approach. Based on the rank of presence frequency, we
   selected 156 and 149 metabolites predictive of FI and devFI,
   respectively (Supplementary Fig. [104]6a; Supplementary Table [105]3).
   86 of these metabolites were identified as both devFI and FI features
   (Fig. [106]4b), suggesting both overlapping and distinct metabolite
   signatures of FI and devFI. The majority of identified FI and devFI
   metabolites were within the amino acid and lipids super pathways
   (Supplementary Fig. [107]6b). Three metabolites were simultaneously
   identified as FI-, devFI- and age-metabolites, including ergothioneine
   that decreases with age and frailty in both females and males
   (Supplementary Fig. [108]7). When looking at the top enriched KEGG
   pathways for FI and devFI, there were 11 common pathways (out of the
   top 15 by p-value) across both groups (Supplementary Fig. [109]6c, d),
   including 7 amino acid metabolism pathways, nicotinate and nicotinamide
   metabolism, pantothenate and CoA biosynthesis, pyruvate metabolism and
   ABC transporters. To further identify core-metabolites related to
   frailty, we derived 21 FI-age and 86 devFI features by merging
   age-related metabolites (86 hub metabolites) and devFI metabolites with
   the FI metabolites respectively, resulting in a set of 104 union
   features (Fig. [110]4b) which are enriched for amino acid metabolism
   and metabolism of cofactors and vitamins pathways (Fig. [111]4c). These
   results suggest these metabolic pathways, notably, amino acids and B
   vitamin metabolism are related to the development of frailty.

Fig. 4. Selection of frailty-related features.

   [112]Fig. 4
   [113]Open in a new tab

   a Schematic diagram for the workflow of the feature selection. Frailty
   index (FI) is composed of base FI and devFI (deviation from the age-
   and sex- group median FI). Base FI is age related, hence leads to age
   metabolites. FI and devFI metabolites are derived from elastic net
   regularization regression via a 100 times repeated 5-fold cross
   validation approach. FI metabolites are merged with age metabolites
   into FI-age features and with devFI metabolites into devFI features.
   The FI-union features are the union of FI-age and devFI features. The
   workflow is performed in the whole cohort, as well as females and males
   after the stratification by sex. b UpSet plot showing the overlapping
   metabolite features from the FI-, age- and devFI- metabolites. c
   Over-represented pathways (y-axis) from the 104 FI-union features from
   the whole cohort. The number of hits (metabolite) from the hub
   metabolites set is shown by x-axis, ratio of the hit number to total
   metabolites in the enriched pathway is represented by dot size and
   p-value is colored by levels.

   Despite the common metabolites, we observed 61 metabolites that are
   unique to devFI (Fig. [114]4b), including hippurate, choline,
   hypotaurine, phenylacetyltaurine, and adenosine 5’-diphosphoribose.
   Enrichment analysis based on these metabolites led to efferocytosis and
   ABC transporters. The results suggest these metabolites and pathways
   are associated with frailty in a completely age- and sex-independent
   way.

Association study of metabolite features with frailty outcomes

   To test the associations of the individual metabolite features with
   frailty outcomes (Fig. [115]5a), we applied linear mixed regression
   models and subjected the 104 union frailty features to a longitudinal
   association study. First, we considered only metabolite abundance at
   the current timepoint (Age[c]). For the current FI (FI[c]) and devFI
   (devFI[c]), we found 47 and 16 metabolites, respectively presented
   coefficients significantly different from 0 (Supplementary Table
   [116]4). Among the 47 FI[c] metabolites, three metabolites (leucine,
   N-acetylthreonine, and X-25422), presented a significant metabolite
   abundance by age interaction term (Supplementary Table [117]5),
   indicating that the association of these metabolites with FI[c] is
   age-dependent. Despite this, leucine showed a generally consistent
   positive correlation with FI[c] at each age group, but this was not the
   case for N-acetylthreonine and X-25422 (Supplementary Fig. [118]8a).
   The remaining 45 metabolites (those without significant interaction
   terms, plus leucine) are associated with FI[c] independent of age. That
   is, individual mice with higher abundance of these 45 metabolites are
   either more (19 metabolites, β > 0) or less (26 metabolites, β < 0)
   frail in a cohort. For devFI[c], individual metabolites were also
   associated with both higher (9 metabolites, β > 0) and lower (7
   metabolites, β < 0) frailty scores. Eight metabolites were identified
   as significantly associated with both FI and devFI (Fig. [119]5b). For
   each of these metabolites, the coefficient of association was in the
   same direction, indicating the same trend of association with both FI
   and devFI.

Fig. 5. Association study in the whole cohort.

   [120]Fig. 5
   [121]Open in a new tab

   a Schematic diagram showing the dependent and independent variables in
   the linear mixed models for the association study. Dependent variables
   include Frailty Index (FI[c]) and devFI (devFI[c], deviation from
   median FI of the age- and sex- specific group) at the current age
   (age[c]), FI and devFI (FI[f]/devFI[f]) at a future age (age[f]), and
   FI/devFI change from age[c] to age[f] (∆FI/∆devFI). Independent
   variables include current abundance of metabolites (MA[c]), abundance
   change from a previous age (age[p]) to age[c], ∆age[1] and ∆age[2]. For
   each frailty outcome, FI-union features identified were individually
   subjected to linear mixed models. b Coefficients of eight metabolites
   of which MA[c] presents significance in the association with both FI[c]
   and devFI[c]. Metabolites are arranged by coefficients (represented by
   dots) for FI[c] in descending order. The line represents the 95%
   confidence interval of each coefficient. c 27 metabolites of which the
   current metabolite abundance presents significance in the association
   with ∆FI. Metabolites are arranged by coefficients (represented by
   dots) in descending order. Significance was determined by adjusted
   p-values via Benjamini–Hochberg false discovery rate procedure at a
   cutoff of 0.05, with * for p < 0.05, ** for p < 0.01, and, *** for
   p < 0.001. d List of 23 metabolites that show occurrences greater than
   or equal to 2. That is, MA/∆MA of metabolite presents significance in
   the association with frailty outcomes of the column.

   Given the longitudinal nature of our dataset we were interested to
   observe whether metabolite abundance at any specific timepoint was
   associated with frailty at a future timepoint (Fig. [122]5a).
   Unfortunately, we didn’t observe any metabolites that showed an overall
   significant association with future FI (FI[f]) or future devFI
   (devFI[f]). When focusing only on the abundance of metabolites at the
   baseline time point (~400 days), we found a single metabolite,
   alpha-ketoglutarate, was negatively associated with both FI[f] and
   devFI[f] (Supplementary Fig. [123]8b). Next, we considered whether
   there were associations between current metabolite abundances, and a
   change in FI from one timepoint to the next (∆FI or ∆devFI, Fig.
   [124]5a). We saw no associations with ∆devFI but found 27 metabolites
   that showed significant associations with ∆FI (Fig. [125]5c and
   Supplementary Table [126]4). No significant interaction terms were
   observed for these metabolites, indicating that these associations were
   not age-dependent. 20 metabolites (β > 0, e.g. creatinine) were
   associated with increased frailty and the remaining 7 (β < 0, e.g.
   phenyllactate) were associated with decreased frailty (Supplementary
   Fig. [127]8c). Finally, we considered whether changing abundances of a
   metabolite over time (∆MA), might be associated with frailty (Fig.
   [128]5b) but found no significant associations.

   Combining the 3 sets of metabolites above (those significantly
   associated with FI[c], ∆FI, or devFI[c]) gives a total of 63
   metabolites, of which 23 are present in 2 or more sets (Fig. [129]5d).
   These metabolites represent candidate biomarkers for frailty and
   include phenyllactate, ergothioneine, nicotinamide riboside,
   creatinine, alpha-ketoglutarate, isoleucine and valine.

Sex specific metabolite features of frailty

   Sex differences are common in aging and frailty. We investigated the
   performance of the generalized linear models that were trained to
   predict FI or devFI in the whole cohort (Fig. [130]4a), in the females
   and males separately. We found significant differences (two-tailed
   t-test, p < 0.001) between the R-squared values derived from female and
   male samples (Supplementary Fig. [131]9a), suggesting the associations
   of metabolites with FI and devFI are sex specific and stratification by
   sex is appropriate for this analysis. Interestingly, the model
   performance was better in the females than males.

   In order to select sex-specific metabolites related to frailty, we
   stratified the whole cohort into female and male subgroups and
   re-selected metabolite features as above. For females, we derived 133
   and 45 metabolites related to FI and devFI, and for males, we obtained
   32 and 92, respectively. Of these only 7 were associated with FI, and 8
   with devFI, in both sexes. Despite this, for both males and females,
   the majority of the identified metabolites were within amino acids and
   lipids super pathways (Supplementary Fig. [132]9b), and the enriched
   pathways were similar between males and females. They predominantly
   included amino acid metabolism, metabolism of cofactors and vitamins,
   mineral absorption, and protein digestion and absorption related to the
   digestive system (Supplementary Fig. [133]9c, d).

   Following the union feature workflow (Fig. [134]4a), we obtained 58
   union features for females and 21 for males related to overall FI.
   Within the female union features, 50% of the metabolites were related
   to amino acid and lipid pathways, whilst the male union features were
   enriched in lipid super pathways (χ^2(df = 2, N = 781) = 4.11,
   p-value = 0.042). Excluding the union features associated with frailty
   in the whole cohort, we identified 25 and 9 metabolites that are unique
   metabolite features identified only in females or males (Fig. [135]6a).
   These sex-specific features include kynurenate and quinolinate for
   females and sphingomyelin and creatinine for males. These results
   suggest sex-specific biomarkers for frailty may be appropriate.

Fig. 6. Sex independent metabolite features and association.

   [136]Fig. 6
   [137]Open in a new tab

   a UpSet plot showing the overlapping metabolites of frailty index union
   features of the whole cohort, females and males. b Coefficients of six
   metabolites of which metabolite change presents significance in the
   association with both FI[c] and devFI[c] in females. Metabolites are
   arranged by coefficients (represented by dots) for FI[c] in descending
   order. c Coefficients of 26 metabolites of which metabolite change
   presents significance in the association with devFI change in females.
   The significance was determined by adjusted p-values via
   Benjamini–Hochberg false discovery rate procedure at a cutoff of 0.05,
   with * for p < 0.05, ** for p < 0.01, and, *** for p < 0.001.
   Metabolites are arranged by coefficients (dots) in descending order.
   The line represents the 95% confidence interval of each coefficient.

Association study of sex-specific frailty features

   To investigate the association of individual plasma metabolites with
   frailty in each sex, we performed mixed linear model regressions using
   the FI union features for females and males separately, as above. In
   females, we first considered solely the current metabolite abundance
   and found that 38 and 16 metabolites, respectively, were significantly
   associated with FI[c] and devFI[c] (Supplementary Table [138]6). As
   with the whole cohort, metabolites were both positively (21 for FI[c]
   and 11 for devFI[c]) and negatively (17 for FI[c] and 5 for devFI[c])
   associated with frailty outcomes. Notably, 6 metabolites were
   identified as both FI[c] and devFI[c] related (Fig. [139]6b). When
   considering associations between current metabolite levels and either
   future frailty, or changing frailty levels (Fig. [140]5a), we found 26
   metabolites were associated with ∆devFI (Supplementary Table [141]6).
   These associations were independent of age, suggesting a relationship
   between metabolite levels and either increasing (10 metabolites, β > 0)
   or decreasing (16 metabolites, β < 0) rate of development of frailty.
   (Fig. [142]6c). Next, we considered the relationship between frailty
   outcomes and changing metabolite abundances over time, and found one
   metabolite, ergothioneine, which was significantly associated with
   devFI[c]. We combined the four datasets from these female-specific
   frailty association studies to identify a total of 52 metabolites, of
   which 3 metabolites were present across 3 lists, including FAD and
   ergothioneine, and 23 metabolites were present across 2 lists
   (Supplementary Fig. [143]10a). These 26 metabolites are potential
   female-specific frailty biomarkers.

   In male samples, we followed the same analysis flow. Considering
   current metabolite abundance, we found 19 and 12 metabolites that were
   associated with FI[c] and devFI[c] (Supplementary Table [144]6). Of
   these, 11 metabolites were excluded as they had significant interaction
   terms, indicating that the association of the metabolite with FI[c]
   depends on age (Supplementary Table [145]7). No significance was found
   for the remaining regressions for the male samples. The results include
   three GPEs, one GPC, creatine, and phenyllactate, that may be potential
   male specific frailty biomarkers (Supplementary Fig. [146]10b).

Validation of the frailty associated metabolites

   In order to validate the plasma metabolites associated with frailty in
   an external cohort, we used female (n = 20) and male (n = 23) mice
   samples under long-term NMN treatment (Table [147]1). These mice were
   less frail compared to their peers in the discovery cohort^[148]34. To
   investigate if the associations of our identified frailty features with
   frailty outcomes persist under the intervention, we used union features
   identified from the whole cohort, and females and males separately, and
   performed the same association analysis. For sex-independent features,
   we found one metabolite, perfluorooctanesulfonate, that was
   significantly associated with current frailty (FI[c]) in the validation
   cohort. There were seven metabolites associated with FI change over
   time (∆FI, Fig. [149]5a), including ergothioneine, guanidinoacetate,
   N-glycolylneuraminate, X-12798, creatinine, dimethylarginine
   (ADMA + SDMA), and N-acetyl-beta-alanine. In male samples only,
   2-hydroxydecanoate maintained a significant association with FI[c] at
   one time point (Supplementary Fig. [150]11). Although NMN treatment
   delays frailty^[151]34, the persistent association of these metabolites
   with frailty outcomes reveals evidence for their robustness as possible
   frailty biomarkers.

Development of a metabolite-based frailty clock

   To build a model to accurately predict frailty using metabolomics
   features in aging mice, we fit a random forest model in the discovery
   cohort, with FI as the dependent variable and the combination of
   identified union features from the whole cohort, female- and male-
   specific analysis (total n = 139 metabolites) as the independent
   variables. We further determined the top 63 metabolite features ranked
   by the presence frequency (Supplementary Fig. [152]6a) gave the best
   performance in predicting FI (Supplementary Fig. [153]12). Among the 63
   metabolite features that can be found in the whole cohort derived
   features, 24 metabolites were also identified as female-specific
   frailty-related metabolites, and 4 as male-specific. Our final model,
   metabolite frailty clock, included these 63 informativity-based
   metabolites, age and sex. We also fit a model using all 781 metabolites
   detected in our study, as a comparison. Both random forest models
   performed well in the discovery cohort (R^2 = 0.95, RMSE = 0.022 for RF
   with 781 metabolites; R^2 = 0.96, RMSE = 0.019 for metabolite frailty
   clock), and outperformed a benchmark model of merely age+sex
   (R^2 = 0.51, RMSE = 0.053) (Fig. [154]7a). Importantly, we examined the
   performance of the metabolic frailty clock in the validation cohort,
   and although it achieves similar performance as the age+sex model
   trained in the validation cohort across the entire cohort (Fig.
   [155]7b), it outperforms age+sex in male samples (Fig. [156]7c).
   Despite the fact that there is clearly room for improvement, these
   results suggest that frailty can be accurately predicted in aging mice
   using plasma metabolite features.

Fig. 7. Performance of metabolite frailty clock in the discovery and
validation cohorts.

   [157]Fig. 7
   [158]Open in a new tab

   Frailty models were built via machine learning approaches, with frailty
   index scores as the dependent variable and three sets of variables as
   the independent variables: 1) The age+sex model, linear regression
   models using age and sex, trained in the discovery and validation
   cohorts respectively; 2) The RF (781 metabolites) model, a random
   forest model using all 781 metabolites detected in this study, age and
   sex; and 3) metabolite frailty clock model, a random forest model using
   63 informativity-based metabolites, age and sex. The performance of
   models in the corresponding cohort/subcohort (Female, F and male, M)
   for (a) Discovery cohort, b Validation cohort, and (c) sex stratified
   Validation subcohort (Males, M; Females, F) are presented using R^2 and
   Root-mean-square deviation (RMSE).

Discussion

   Using a longitudinal study of female and male mice, we identified both
   sex-independent and sex-specific plasma metabolomic signatures of aging
   and frailty. Overall, we found that age related metabolites are
   enriched for lipid metabolism, while frailty-related metabolites are
   enriched for amino acid metabolism and metabolism of cofactors and
   vitamins. B vitamin metabolism-related metabolites and lipid
   metabolism-related metabolites, respectively, are determined as
   candidate female- and male-specific frailty biomarkers.

   Using a time course analysis in mice, we found a total of 527
   metabolites significantly changed with age, representing the majority
   of measured metabolites. This unexpected result suggests dramatic
   change in the abundance of most metabolites in aging, but aligns with a
   previous longitudinal human study^[159]21. Among these age-related
   metabolites, we identified 86 hub metabolites by network analysis, and
   observed that these metabolites were enriched for lipid metabolism,
   including biosynthesis of unsaturated fatty acids, primary bile acid
   biosynthesis, and fatty acid elongation. Interestingly, several studies
   in mice have demonstrated manipulation of lipid metabolism as a method
   to extend longevity^[160]35,[161]36. Our results using longitudinal
   data provide further evidence of the importance of lipid metabolism in
   aging. In terms of the individual metabolites changed in age, we
   identified 10 sphingomyelin species, which are of interest given the
   established link between sphingomyelins and longevity in
   humans^[162]20.

   Additionally, we found that 42% of measured metabolites were
   significantly different between males and females, and the majority of
   these were also changed with age. Metabolites involved in aging and sex
   differences were enriched for a spectrum of pathways involved in lipid
   metabolism and digestive system. While sex differences in lipid
   metabolism have been widely recognized, our finding provides further
   evidence in the context of aging and aligns with the result from humans
   that the lipidome exhibits significant age-dependent differences
   between sexes^[163]37. Moreover, liver is the primary tissue for bile
   acid metabolism^[164]38, fatty acid metabolism^[165]39 and taurine
   metabolism (conjugation with bile acid^[166]40), all pathways
   identified in our study as displaying sex differences in aging. This
   suggests that the liver is strongly influenced by biological sex in
   aging, which aligns with transcriptomic results from our lab^[167]41.
   Furthermore, the presence of the mineral absorption and ferroptosis
   pathways in our findings^[168]42,[169]43 reveal that the impact of
   aging on iron homeostasis^[170]44 is sex specific.

   Sex stratified analysis revealed female-specific metabolite markers of
   aging include amino acids and acylcarnitines. Previous work has
   extensively shown the role of amino acids and acylcarnitines in the
   regulation of aging^[171]45,[172]46. Our results indicate significant
   changes of these metabolites in females in aging, but not necessarily
   in males, which is also implied in other studies^[173]47. As for the
   male specific aging biomarkers, we found phosphocholine (adjusted for
   sex) and spermine (only in males), both of which have been previously
   linked to overall aging^[174]48,[175]49. Together, these findings give
   clues about how metabolic aging may occur differently in males and
   females.

   Although age-related metabolic markers are of interest, markers that
   are associated with health in aging may provide more clues about
   underlying mechanisms of the aging process, rather than the passing of
   time. To this end, we sought to identify metabolic features of frailty,
   a validated quantification of health in aging in both humans and mice.
   As frailty is strongly correlated with age, it was important that we
   identify metabolic markers of frailty, independent of age. We used a
   novel approach of calculating devFI, the deviation from the median
   frailty index of the corresponding age and sex group. In this way we
   are able to identify metabolites associated with individual variations
   in frailty at a given age and sex group, and distinguish these from
   metabolic changes that arise from aging. We applied a machine learning
   approach to select metabolites that are associated with both outcomes,
   FI and devFI. In the whole cohort study, we identified 149 metabolites
   features for devFI, among which 61 were not also associated with FI or
   age (Fig. [176]4b). These include hippurate, a gut microbiome derived
   metabolite that has been previously associated with aging^[177]50.
   These metabolites are particularly interesting for further analysis as
   underlying markers of health, independent of age.

   Overall, frailty-related metabolites were heavily enriched in amino
   acid metabolism and metabolism of cofactors and vitamins. The majority
   of the 20 proteinogenic amino acids metabolism pathways were enriched,
   suggesting a close relationship between amino acids and frailty
   dynamics in mice. In humans, altered amino acid metabolism is also
   suggested to be associated with frailty^[178]24, in particular
   tryptophan metabolism^[179]51. Interestingly, nicotinate and
   nicotinamide metabolism were also over-represented in these candidate
   frailty biomarkers. Recent work has shown that boosting nicotinamide
   levels is associated with improved health in aging, including improved
   frailty^[180]34,[181]52,[182]53. These results suggest pivotal
   differences in the metabolites associated with aging compared to
   frailty, which may reflect potentially different underlying mechanisms.

   Additionally, we applied linear mixed models to the metabolite features
   identified for frailty to look at their specific univariate association
   with frailty outcomes (Fig. [183]5a). Notably, we found 23 metabolites
   that were associated with more than one frailty outcome (i.e., current
   FI, current devFI and/or change in FI over time) (Fig. [184]5d),
   including ergothioneine, nicotinamide riboside (NR), phenyllactate, and
   creatinine. Interestingly, ergothionine is one of the most robust
   markers in our study, which is identified as FI, devFI and
   age-associated across both males and females. It has been previously
   identified as a frailty biomarker^[185]54, and is thought to promote
   healthy aging^[186]55. NR is an NAD precursor, part of the nicotinamide
   metabolism pathway, and boosting levels of NR are associated with
   improved health in aging^[187]53,[188]56. Phenyllactate, is a
   catabolite of phenylalanine (phenylalanine metabolism is identified as
   enriched from frailty-related metabolites) derived from
   Lactobacillus^[189]57, providing further evidence of the possible
   involvement of the microbiome in the development of frailty.
   Creatinine, a muscle breakdown product, has been associated with
   sarcopenia, functional limitation and frailty^[190]58. Additionally,
   alpha-ketoglutarate was the only metabolite for which the abundance in
   middle-aged mice was predictive of future frailty, suggesting it could
   be an early-biomarker of frailty, and/or a viable target for early
   intervention. In support of this, a recent study shows that
   alpha-ketoglutarate supplementation in mice reduced frailty^[191]59.
   Taken together, our results provide preclinical evidence for several
   potential biomarkers for frailty.

   In order to identify sex-specific metabolic markers of frailty, we
   completed sex-stratified analysis. We identified vitamin B3/tryptophan
   metabolites, kynurenine and quinolinate, as being specifically
   associated with frailty in females. The findings for these two
   metabolites are consistent with previous studies^[192]60,[193]61, where
   the link to frailty is sex-specific. FAD (vitamin B2), was also
   significantly associated with multiple frailty outcomes in females
   (Supplementary Fig. [194]9a). FAD is one of the active forms of vitamin
   B2, however, previous studies in both sexes found that intake of
   vitamin B2 has no association with frailty^[195]62,[196]63. Given our
   novel findings, we suggest further investigation into FAD as a
   female-specific marker of frailty. Another female-specific frailty
   biomarker is pyridoxate (vitamin B6), which is reported to be related
   to frailty^[197]64. In male mice, we identified mainly lipid
   metabolism-related metabolites, including sphingomyelins, three GPE
   species and one GPC species. These metabolites are lipid species that
   have been previously associated with frailty^[198]65,[199]66 in both
   sexes, so the male specificity needs further investigation. Taken
   together, our results reveal evidence of sex-specific biomarkers for
   frailty, and imply that B vitamin metabolism is a key feature of
   frailty development in females and lipid-related metabolism for males.
   We highly recommend applying the sex stratification approach in the
   future study of frailty biomarkers and mechanisms.

   Importantly, and often ignored in other frailty biomarkers studies, we
   confirmed whether the same metabolites were associated with frailty
   outcomes in an independent validation cohort. Although not all
   metabolites associated with frailty in the discovery cohort held this
   association in the validation cohort, we did find 9 metabolites showing
   persistent significance in the association with frailty outcomes,
   including ergothioneine and creatinine. Our validation cohort included
   mice that had long-term treatment with the NAD booster, NMN, suggesting
   that the association of these metabolites with frailty outcomes may be
   universal even under interventions, so these biomarkers should be
   investigated further.

   Many ‘clocks’ have been built to predict chronological age based on
   either epigenetic or metabolomic features^[200]67,[201]68. There is a
   growing focus, however, on building models to predict health- rather
   than age-related outcomes. Here, we build the first clock to directly
   predict frailty in mice. Our model performs extremely well in our
   discovery cohort, and although the performance of the frailty clock is
   similar to that of an age+sex only model (trained within the validation
   cohort) in females for the validation cohort, our clock outperforms the
   simple model in male samples. These results provide preliminary
   evidence that it is possible to predict frailty using metabolites, but
   suggest further work should be done in large datasets to develop a more
   universal metabolomics-based frailty clock.

   There are some limitations to this study. Our validation dataset was
   relatively small, and mice were treated with NMN which has known
   effects on metabolic pathways and frailty^[202]53. It is possible that,
   given that the NMN treatment appeared to make these mice metabolically
   healthier and less frail than the untreated discovery cohort mice, it
   may have masked potential associations. For instance, we observed
   pyridoxate (determined as a female-specific frailty biomarker) was
   affected by NMN treatment in a prior study^[203]34. Notably, we also
   found that physiological metabolic effects of NMN are stronger in males
   than females^[204]34, suggesting sex differences in the metabolic
   response that may also cloud validation of our sex-specific findings.
   This validation dataset was selected as it was difficult to find
   matched metabolomics and frailty data for mouse studies, but we suggest
   future work should validate these potential frailty biomarkers in
   larger cohorts, with a variety of interventions, and in human cohorts.
   Additionally, the sample size for metabolomics data is relatively
   small, especially at the older ages, which might decrease the power of
   statistical analysis. Survival bias is also an issue to consider, as
   mice died over the course of the study and only those that were longest
   lived made it to timepoint 4 and 5. For future work, it will be ideal
   to conduct studies in a broader age range with an increased number of
   mice.

   In summary, we performed the first longitudinal study of naturally
   aging female and male mice looking at metabolomics of frailty. We found
   aging related plasma metabolites are mainly involved in lipid
   metabolism while frailty-related metabolites are predominantly parts of
   amino acid metabolism and metabolism of cofactors and vitamins. Apart
   from whole cohort frailty biomarkers, we demonstrated the sex
   dimorphism in the associations between metabolite and frailty, and
   proposed sex-specific frailty biomarkers.

Methods

Mice samples

   Mice used in this study are from a larger intervention study, so
   detailed methods can be found in Kane et al.^[205]34. Briefly,
   C57BL/6NIA mice, female (n = 40) and male (n = 47) were obtained from
   the National Institute on Aging (NIA) Aging Rodent Colony, among which,
   20 female and 23 male mice were subjected to nicotinamide
   mononucleotide (NMN) treatment. Mice were group housed (4–5 mice per
   cage, although over the period of the experiment mice died and mice
   were left singly housed), at Harvard Medical School in ventilated
   microisolator cages, with a 12-h light cycle, at 71 °F with 45–50%
   humidity. Mice were fed AIN-93G Purified Rodent Diet (Dyets Inc, PA).
   All animal experiments were approved by the Institutional Animal Care
   and Use Committee of the Harvard Medical Area. In order to investigate
   aging and frailty-related metabolites and mechanisms in naturally aging
   mice, we used non-NMN treated mice (female, n = 20; male, n = 24) as
   the discovery cohort for principal component analysis, feature
   selection, sex stratified analysis, association study, and metabolite
   frailty clock model building (Fig. [206]1). We then tested the selected
   metabolite features and model in the NMN treated mice (validation
   cohort).

Mouse frailty assessment

   Behavioral and clinical variables for clinical frailty index were
   measured in both the discovery and validation cohorts, at each time
   point (Table [207]1). We utilized the mouse clinical frailty
   index^[208]2 (FI) that contains 31 health-related items for this study.
   Briefly, mice were scored either 0, 0.5 or 1 for the degree of deficit
   they showed in each item with 0 representing no deficit, 0.5
   representing a mild deficit and 1 representing a severe
   deficit^[209]69. Apart from FI score itself, we introduced devFI score,
   that is the deviation of individual FI from the median FI for the
   corresponding sex, at the corresponding time point.

Blood collection and processing

   Mice were fasted for 5–6 h, anesthetized with isoflurane (5%) and then
   blood was collected from the submandibular vein with a lancet (maximum
   10% of mouse body weight, approx. 200–300 ul), into a tube containing
   20ul of 0.5 M EDTA. Blood was mixed and stored on ice. Whole blood was
   centrifuged at 1500 × g for 15 min, plasma was removed and frozen at
   −80 °C for subsequent metabolomics.

Metabolites extraction, quantification and processing

   Global metabolomics analysis was completed by Metabolon. Samples were
   prepared using the automated MicroLab STAR® system (Hamilton Company),
   and analyzed using Ultrahigh Performance Liquid Chromatography-Tandem
   Mass Spectroscopy (UPLC-MS/MS). We used raw metabolite data (peak
   area). We performed batch effects mitigation by calculating the mean
   metabolite value for the baseline time point across all mice, comparing
   it to the mean value of all other time points and excluding metabolites
   that presented the mean of the baseline significantly lower (0.05x,
   compared to the mean of all other time points) or significantly higher
   (10x compared to the mean of all other time points). We normalized each
   sample by dividing the metabolite value by the median of metabolite
   values for that sample to account for any collection batch effects and
   then derived the (natural) log-transformed values as the metabolite
   abundance. We derived 781 metabolites with greater than 5% unique
   abundance across samples.

Metabolomics data variations

   The number of metabolomics data is summarized in Table [210]1.
   Metabolite abundance data was subjected to principal component (PC)
   analysis. We derived PC1–PC10 and for each PC as the dependent
   variable, we applied linear regression models and obtained p-values,
   where we used mouse ID, time points, sex, cage (categorical variables),
   and age at assessment (continuous variables), respectively as the
   independent variable. Independent variables tested were then clustered
   according to Euclidean distance.

Differential abundance analysis of metabolites

   Log-transformed metabolite abundance data were subjected to
   differential abundance analysis by using the ‘limma’ pipeline with a
   spline. Briefly, metabolite abundance data was subjected to the limma
   time-course spline analysis, excluding time point 5 due to absence of
   female samples at this time point. We generated a matrix for a natural
   cubic spline based on the remaining time points, with degrees of
   freedom set at 3, and the matrix was used as the time factor. Design
   matrices for global differential abundance analysis included sex and
   sex by time interaction term, without assigning a reference level. The
   data along with the multi-factor design matrix were then subjected to
   linear modeling with the intra-block correlation based block on mouse
   ID, and empirical bayes smoothing of metabolite-wise standard
   deviations. We then determine metabolite abundance differences by
   defining a contrast matrix for each of the following four categories:
   1) mixture of female and male samples, 2) female samples, 3) male
   samples, and 4) sex differences. The purpose of this analysis is to
   identify metabolites that respond to the time effect, in other words,
   metabolites that change over time. We selected metabolites that show
   significant differences when comparing any of the two consecutive time
   spline variables via F-test. Differentially abundant metabolites (DAMs)
   were selected by controlling for a 5% Benjamini–Hochberg false
   discovery rate (adjusted p-values < 0.05).

Co-abundance analysis

   We performed the co-abundance analysis of metabolites (excluding time
   point 5), with a soft threshold set at 9 to select metabolite abundance
   modules. For a given module, we derived the first principal component
   as the eigenvalue. To identify the association of metabolite modules
   with age, we applied a linear mixed model using the module eigenvalue
   as independent variable and age as dependent variable with adjustment
   for sex, allowing a random intercept for each mouse. P-values were then
   adjusted by the Bonferroni correction method. For each identified
   subset, metabolites that showed significance greater than 0.2
   (correlation coefficient between the metabolite abundance and the age)
   and module membership greater than 0.8 were selected as the hub
   metabolites in the module.

Pathway enrichment analysis

   Metabolite set enrichment analysis was performed by using the
   hypergeometric test from R package FELLA (v. 1.20.0) to identify KEGG
   pathways that were over-represented. We reported the top pathways (up
   to 20) ranked by unadjusted p-values^[211]70.

Feature selection

   We applied a machine learning approach to identify FI/devFI related
   metabolites in the discovery cohort (Fig. [212]4a). We performed
   feature selection by fitting generalized linear regression models using
   the frailty assessment score (FI or devFI) as the dependent variable
   and the 781 metabolites abundance data as the independent variables,
   through a 100 ×5-fold cross validation approach. Briefly, we performed
   100 runs of multivariate generalized regression with elastic net
   regularization. Within each run, the hyperparameters for the least Root
   mean square error (RMSE) were tuned using 5-fold cross-validation, and
   a list of metabolite features assigned a non-zero coefficient was
   derived. These lists (from 100 runs) were merged into a list of
   metabolite features, which were then ranked according to the
   importance, i.e. the presence percentage of the metabolites. We
   selected metabolites that made to the top 20% percentile as FI/devFI
   metabolites. FI is composed of the age-related base FI and devFI.
   Hence, we derived FI-age features by combining age metabolites (hub
   metabolites from co-abundance analysis) and FI metabolites, and devFI
   features by combining devFI metabolites with FI metabolites. We then
   obtained union features from the union of FI-age features and devFI
   features.

Analysis of metabolite associated with frailty outcomes

   For the association study, we applied mixed linear models allowing
   variations in individual mice as the random effect. We considered three
   outcomes for FI and devFI respectively (six in total), 1) the score at
   current age (age[c]), FI[c] and devFI[c]; 2) a score at a future time
   point (age[f]), FI[f] and devFI[f]; and 3) score change to a future
   time point, ∆FI and ∆devFI (Fig. [213]5a). We considered two scenarios
   in the analysis, where abundance of metabolites from a previous time
   point (age[p]) are: a) absent, only the current abundance of
   metabolites (MA[c]) is available. For each metabolite, we used one of
   the six score outcomes as the dependent variable and MA[c] as the
   independent variable, adjusting for sex, age[c] and age change from
   age[c] to age[f] (∆age[2], only for outcomes 2) and 3)) in the
   non-interaction models. For interaction models, we considered abundance
   by age term and abundance by age change term; and b) present, the
   current and a previous abundance of metabolites and the age interval
   are available. We focused on metabolite abundance change (∆MA) in this
   scenario. We used one of the six score outcomes as the dependent
   variable and ∆MA as the independent variable, adjusting for MA[c], age
   change from age[p] to age[c] (∆age[1]), sex, age[c] and age ∆age[2]
   (for outcomes 2) and 3)) in the non-interaction models. For interaction
   models, we included abundance change by age (and/or age change) terms
   and current abundance by age (and/or age change) term. devFI score is
   the deviation from the median at the age- and sex- specific group.
   Hence, current age was not included in the analyses of devFI in both
   non- and interaction models. The age variable used above was the actual
   days of assessment divided by 1000, in order to be within the same
   scale.

Metabolite frailty clock model building

   After obtaining three sets of union features for the whole cohort,
   females, and males, we generated a single set of metabolite features
   from the above three sets. We ranked these features by occurrence
   frequency from the feature selection process (100 times repeated cross
   validation) within the whole cohort. Via a cross validation approach,
   we selected ‘mtry’ (the number of randomly drawn candidate variables
   out of which each split is selected when growing a tree) and the number
   of informativity-based top metabolite features that gave the least RMSE
   in predicting FI. The final metabolite frailty clock model was fit in
   the discovery cohort, constructed using a random forest regression with
   FI as the dependent variable and the top metabolites features, age and
   sex as the independent variables. We also fit linear regression models
   with age and sex as the independent variables in respective the
   discovery and validation cohort, and a random forest model using all
   the 781 metabolites, age and sex in the discovery cohort for
   comparative purposes.

Statistics

   All statistical analyses were performed using R (version 4.3.0).
   Association study between abundance of metabolites and age was
   performed by applying linear mixed models. For each metabolite, we used
   abundance of metabolite as the dependent variable and continuous age as
   the independent variable, adjusted for sex and variability of
   individual mice. For association study (for both age and frailty
   index), the significance was determined by controlling for a 5%
   Benjamini–Hochberg false discovery rate (adjusted p-values < 0.05).
   Differentially abundant metabolites (DAMs) are selected by controlling
   for a 5% Benjamini–Hochberg false discovery rate (adjusted
   p-values < 0.05).

Supplementary information

   [214]Supplementary Figures^ (11.4MB, pdf)

Acknowledgements