Changes in Blood Metabolites and Fats Across Human Life Stages

To delve into the metabolic alterations in human plasma across the lifespan, we collected plasma samples from 136 volunteers aged 0–84 years. According to the stages of human development, we categorized the participants into seven groups (newborn, early childhood, middle childhood, adolescence, early adulthood, middle adulthood and late adulthood) (Figure 1a, Figure S1a, Table S1). The age distribution of the included participants is presented in Figure S1b. Using two extraction methods, we then performed untargeted metabolomic and lipidomic analyses, retaining 1931 metabolites across 13 classes and 3795 lipids across 27 classes after filtering (Figure 1a–c). All metabolites met the metabolomics standards initiative (MSI) Level 2 identification standards at minimum, including 469 Level 1 metabolites confirmed with authentic standards, while the remainder were identified as Level 2 based on MS/MS spectral matching. Overall, lipid‐related metabolites comprised the largest proportion in the metabolomic results, as their hydrophobic effects enable more accurate detection in untargeted lipidomics, further highlighting the necessity and scientific validity of our integrated multi‐omics approach. Lipids detected by lipidomics mainly included triglycerides (TGs), phospholipids such as phosphatidylcholine (PC) and phosphatidylethanolamine (PE), and sphingolipids such as ceramide (Cer), hexosyl ceramide (HexCer), and sphingomyelin (SM), comprehensively covering multiple pathway modules. To ensure data quality, we implemented rigorous quality control measures. Prior studies have noted the significant contribution of sex to human metabolomics (Austad and Fischer 2016; Bell et al. 2021; Li, Xiong, et al. 2023; Link and Reue 2017; Mauvais‐Jarvis 2024), prompting us to investigate its influence on our omics data. However, principal variance component analysis (PVCA) and principal component analysis (PCA) revealed only a minimal sex effect in metabolomics and lipidomics (2.69% and 0.61%, respectively) (Figure 1d, Figure S1c). Meanwhile, age was the factor with the greatest impact, that is, the largest proportion of global variance. Consequently, we suggest that while sex has a certain influence on metabolic changes across the lifespan, its impact is relatively small. To map the metabolic landscape at a broader level, we did not perform any sex‐stratified or age‐sex interaction in subsequent analyses. To visualize the holistic changes in metabolic data across the lifespan, PCA was applied on samples from all groups (Figure 1e,f). Strikingly, newborns were distinctly separated from all other stages, highlighting a marked difference in metabolic composition during this period. This suggests that the metabolic state undergoes substantial changes from birth through the developmental period. To further assess whether the strong effect of age was driven by the high distinction of the newborn group, we repeated the PVCA analysis after excluding newborn samples (Figure S1d). An increase in residual variance was observed, likely due to a reduction in age‐related variance when the highly distinct newborn group was removed; nevertheless, age remained the major biological contributor to variance, while sex accounted for the smallest proportion of the effect.

Heatmaps were used to annotate the expression of major metabolite and lipid classes, clearly distinguishing newborns from other age groups, with opposing expression trends observed across various metabolite and lipid classes (Figure 2a, Figure S2a). Subsequently, we sought to identify which metabolites exhibited such distinct expression in newborns. Interestingly, differential expression metabolites (DEMs) were predominantly concentrated in the same categories, highlighting the reliable and consistent differential expression of these DEMs (Figure 2b, Figure S2b). We further applied median‐ratio normalization to confirm the consistency of the results. The newborn‐enriched metabolites identified were largely consistent with those obtained from total‐sum normalization (Figure S2c). Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis of these metabolites revealed significant alterations in multiple metabolic pathways (Figure 2c). Notably, galactose metabolism and the TCA cycle were upregulated in newborns, while metabolic pathways related to protein digestion and absorption, as well as the biosynthesis of amino acids were downregulated (Figure 2c,d, Figure S2d). This may be explained by the fact that glucose, which serves as the primary energy source for the fetus and is also the hydrolysis product of lactose, acts as a major substrate feeding into the TCA cycle and consequently supports the elevated energy demands of rapid postnatal growth and organ development. This further suggests that, newborns have a heightened metabolic rate to ensure adequate growth and organ development, as this life period is associated with rapid growth.

Moreover, among various lipids, sphingosine (SPH) is the only lipid class with elevated expression in newborns, while the expression of Cer and SM is lower (Figure 2e). This indicates that the conversion process from SMs to Cers and subsequently to SPHs may be more vigorous in cord blood (Figure 2f). Further, the conversion of Cer to HexCer was relatively low, collectively leading to the accumulation of higher levels of SPH. Understanding the variations in SPH content during developmental processes can further aid in identifying the onset of various diseases. Given that the key determinants of lipid function and characteristics are chain length and saturation (Zhang et al. 2024), a relevant analysis was performed. Intriguingly, we found that the saturation of plasma SPHs in newborns was significantly higher than in other age groups, particularly in long‐chain lipids with carbon chain lengths of 16, 20, and 22 (Figure 2g, Figure S2e). Consistently, the calculation of the total lipids revealed that the ratio of polyunsaturated to saturated lipids remained the lowest in newborns (Figure 2h). Interestingly, we found no difference in the saturation levels of all lipid classes following SPH removal, which to some extent reflects the pivotal role that SPH plays in newborns (Figure 2i). The degree of plasma lipid unsaturation gradually increases during human ontogeny from newborns, and has extraordinary significance in a variety of activities affecting vital biological functions, including cell membrane fluidity, signal transduction, metabolic regulation, and anti‐inflammatory effects. To explore the potential functional significance of elevated SPH levels in newborns, we conducted preliminary functional assays using cord blood mononuclear cells (CBMCs). Analysis revealed that treatment with SPH (10 μM, 24 h) markedly increased apoptosis across multiple immune cell subsets, including CD4⁺T, CD8⁺T, Natural Killer (NK), and B cells (Figure S3a,b). In addition, SPH treatment induced subset‐specific alterations in immune receptor expression (Figure S3c–g). SPH treatment upregulated activation markers (e.g., CD69) while concurrently reducing co‐stimulatory receptors (e.g., DNAM‐1, CD38) and enhancing inhibitory checkpoint molecules (e.g., PD‐1, TIGIT) in T cells, pointing to a state of activation‐induced stress and potential exhaustion. NK and B cells showed a decline in both activation and co‐stimulatory receptors, which likely reflects a transiently restrained immune state in newborns. These results indicate that SPH elevation can sensitize immune cells to apoptosis and modulate immune responsiveness, thereby potentially contributing to developmental immune remodeling in early life. Taken together, these data indicate that newborns exhibit highly unique metabolic profiles, characterized by a higher energy metabolism rate but concurrent with less mature lipid development, which was particularly reflected in the lower unsaturation levels of SPH. This suggests that a significant lipid remodeling process occurs postnatally.

Next, we sought to unravel the unique metabolic characteristics of each age group across the human lifespan. PCA revealed that, although the differences between groups were relatively small, samples from each stage gradually diverged from those of adjacent stages (Figure S4a), mirroring the progressive changes in metabolic composition that occur throughout the aging process. We were curious about the process and rate of these changes across the different stages. Therefore, we performed pairwise differential expression of metabolites to determine when and how changes in metabolic states occur across the human lifespan. Compared to early childhood (Group B), differences in metabolites and lipids progressively increased with age (Figure 3a, Figure S4b). Pairwise differential analysis of adjacent groups revealed that the metabolic changes were more pronounced during the rapid growth phases from early childhood to adolescence, while these changes gradually slowed down after middle age (Figure 3b, Figure S4c). Together, these data illustrate a fluctuating pattern of metabolic characteristics across the lifespan, with rapid changes occurring initially, followed by a gradual slowdown.

We further analyzed the most distinctive components representing each stage (Figure 3c–e, Figure S4d–f). Specifically, fatty acids (linoleic acid, oleic acid, palmitoleic acid, etc.) were significantly elevated in early childhood (Group B), while PC and PE were the predominant lipids in the lipidome. The increased levels of fatty acids may modulate the levels of PE and PC through various pathways, collectively reflecting the heightened activity of lipid metabolism during early childhood. In middle childhood (Group C), a stage characterized by significant physical growth and the gradual enhancement of physical strength and motor skills (Rogol et al. 2000), we consistently observed a marked increase in the levels of ADP and ATP, accompanied by an elevated presence of certain amino acids, particularly proline‐containing peptides.

A surge in the proportion of steroid and steroid derivatives among the elevated metabolites emerged in the transition into adolescence (Group D). Additionally, among the lipidome markers, TGs and diglyceride (DGs) were notably abundant. Steroids may regulate glycerolipid metabolism via hormonal regulation, thereby reflecting the impact of active sex hormone secretion during puberty on the complex interplay with various lipids, particularly glycerides. After early adulthood (Group E), metabolic changes were relatively small and closely resembled those in middle adulthood (Group F), indicating a more stable physiological condition. It is worth noting that Cer‐related lipids exhibit distinct characteristics during middle adulthood (Figure S4g). Prior studies have further reported that Cers are associated with neurodegenerative diseases (Trayssac et al. 2018), diabetes (Chaurasia et al. 2019), and cardiovascular diseases (Choi et al. 2021), indicating the importance of early monitoring of metabolic status at appropriate stages to prevent the onset of related diseases.

A plethora of highly expressed metabolites and lipids were identified in the final stage of the human lifespan (late adulthood, Group G) (Figure 3c, Figure S4d). Pathway enrichment and heatmap analysis revealed a notable increase in amino acid peptides (Figure 3f,g). This phenomenon reflects an active state of protein metabolism in late adulthood. Cysteine and glutamine, precursors of glutathione synthesis, perform crucial functions as key antioxidants in inflammation and immune responses (Gorrini et al. 2013). The increased plasma cysteine levels may be associated with a state of chronic inflammation, highlighting the intertwined roles of amino acid metabolism and anti‐inflammatory mechanisms in aging. In summary, these data delineate a metabolic landscape of uniquely upregulated representative metabolites across different life stages, thus elucidating the dynamic state transitions and physiological implications throughout the lifespan.

Given the continuous alterations in metabolic states across the lifespan, we subsequently investigated the age‐related linear changes in the metabolome and lipidome. Initially, we used a linear regression model to assess the DEMs that exhibited either an increase or decrease with advancing age, as well as showing relatively consistent patterns across both sexes (Figure 4a,b, Figure S5a). The altered components were concentrated in completely distinct modules (Figure 4c, Figure S5b). Similarly, metabolites showing linear age‐associated trends were stable across normalization methods (Figure S5c). Specifically, amino acids identified as elderly markers showed a steady linear increase across the lifespan, contrasting with their extremely low levels in newborns, which merits further investigation (Figures 2c and 4c). In addition, the progressive increase in bile acid–related metabolites may reflect age‐associated alterations in the metabolic pathways governing nutrient and lipid metabolism (Figure 4d, Figure S5d).

In the lipidome, various lipid types exhibit distinct expression trends during development (Figure 4e). SPH, particularly its saturated forms, which were notably high in newborns, showed a linear decline with age (Figure 4e,f, Figure S5e). In addition, TGs account for the highest proportion (Figure 4f). Intriguingly, despite the varying chain lengths, saturation levels, and expression trends among TGs, over 90% of upregulated TGs are polyunsaturated, while downregulated TGs exhibit a higher proportion and expression levels of saturated TGs (Figure 4g, Figure S5f). The plasma TG level is a crucial health indicator, correlated with an increased risk of age‐related diseases, such as coronary heart disease, atherosclerosis, myocardial infarction, and insulin resistance (Eichelmann et al. 2024). Therefore, understanding the variations in lipid saturation patterns may help in ultimately maintaining overall health.

We performed a Spearman correlation analysis to unravel the relationships between metabolites and lipids that exhibited linear increases with age (Figure 4h). The strong positive correlations between these metabolites and lipids may indicate that they share common functions and are situated within interconnected metabolic networks. To ascertain potential common regulatory relationships, we integrated the metabolites with lipids to construct a correlation network (Figure 4i). This network was based on the correlations between the top‐50 metabolites and top‐50 lipids with the highest linear change coefficients (p.adj < 0.05), and the correlations (Spearman r > 0.5) between 30 lipids and 28 metabolites are shown. The lipid pathway most strongly correlated with metabolites was glycerophospholipid metabolism. Conversely, TGs were the primary lipids correlated with metabolites in glycerolipid metabolism (Figure 4i). Overall, these findings highlight that the age‐related linear changes in metabolism across the lifespan are highly coordinated. Investigating these changes will provide insights into the linear alterations within the intricate metabolic landscape during the human development process.

While observing the linear variations in metabolic expression across ages, we concurrently noted several distinct nonlinear patterns of fluctuations among metabolites. Consequently, we employed locally estimated scatterplot smoothing (LOESS) to visualize the nonlinear changes in age‐related metabolites (Figure 5a). Additionally, unsupervised hierarchical clustering was applied to group the metabolites with similar expression patterns, thereby identifying the distinct trajectories of metabolic expression alterations with respect to age (Figure 5b). Within the eight clusters of metabolite analysis, each is enriched with specific biological pathways, indicating the occurrence of distinct, category‐specific changes in biological pathways across lifespan (Figure 5c,d). For example, Cluster 1, which peaks in expression around the age of 20, is represented by lysosome and fatty acid synthesis, while Cluster 7, peaking around 40 years, is predominantly characterized by caffeine metabolism (Figure 5d,e). Similarly, lipids were also clustered into distinct trajectories with notable differences in subclass proportions and saturation levels (Figure 5f, Figure S6a,b). In summary, many of the changes in the metabolome and lipidome across the lifespan are nonlinear. Clustering these changes by trajectory reveals distinct characteristics, demonstrating that the timing and duration of accumulation, or decline of certain components, may vary significantly across life. These findings enhance our understanding of key biological events, underscoring the dynamic regulatory patterns in biological pathways across the lifespan.

Next, we aimed to determine the ages at which the majority of metabolic expression changes occur, with the goal of identifying important metabolic turning points across the lifespan. DE‐SWAN (Differential Expression‐Sliding Window Analysis) is a specialized algorithm designed to analyze the differential expression of components across different age windows. Thus, we used DE‐SWAN with a 5‐year incremental window, spanning from the youngest to the oldest age groups. Several DE peaks were detected in the metabolome at ages 7 and 67 and in the lipidome at ages 7, 22, and 57; subsequently, we plotted the significant top metabolites corresponding to each peak (Figure 5g,h, Figure S6c). The DE‐SWAN‐identified age‐related peaks also remained robust under median‐ratio normalization (Figure S6d). It is noteworthy that the metabolome and lipidome reach a crest concurrently at the age of 7. However, in the later stages of life, the lipidome peaked earlier (age 57) than the metabolome (age 67) (Figure 5g,h). Additionally, the lipidome showed a unique peak at age 22, suggesting a rapid, more sensitive response to metabolic state changes than the metabolome, which may serve as a more agile reflection of alterations in an individual's health status, potentially acting as an early warning indicator. The limited overlap of metabolites and lipids at peak ages indicates that the developmental and aging processes in humans differ significantly across age stages (Figure S6e). Thus, the fluctuations in metabolism during the nonlinear process of aging are noteworthy.

The terminal stage of lifespan—the period encompassing late adulthood, has emerged as a focal point of global concern. A deeper understanding of the processes associated with aging is required to achieve greater potential to enhance quality of life, as many of the diseases prevalent among the elderly are intimately tied to various metabolites in the body. Hence, we investigated the associations between metabolites (lipids) and clinical indicators during the transition from middle (Group F) to late adulthood (Group G) (Table ). S2

From middle age onward, certain amino acid peptides are increasingly expressed, with a notable accumulation identified by the elderly stage (Figure 6a,b). Amino acids are fundamental to protein synthesis, and are particularly crucial in erythropoiesis and hemoglobin synthesis, while also being significant for the stability of the cell membrane structure. Interestingly, these peptides show high positive correlations with red blood cell indices, particularly cell volume. On the one hand, this may help prevent conditions such as anemia; conversely, it could lead to structural instability during erythropoiesis, resulting in uneven sizes of newly formed red blood cells and increased blood viscosity. Additionally, the increase in the levels of amino acid peptides may promote inflammatory responses and elevate the risk of various diseases, such as diabetes (Ling et al. 2023). These physiological functions further underscore the importance of maintaining their balance for overall health. Currently, amino acid testing is largely absent from routine health examinations, indicating the need for increased monitoring of amino acid levels in the elderly.

We further observed multiple significant correlations between lipid profiles and clinical indicators. Notably, high‐density lipoprotein (HDL) ("good cholesterol") is responsible for removing excess cholesterol; the lipids it carries are primarily phospholipids, particularly PC, with which it displays a strong positive correlation. Conversely, very low‐density lipoprotein (VLDL) ("bad cholesterol") exhibits negative correlations with various lipid subclasses. These data underscore the intricate relationships between different lipid classes and blood lipid indicators, as the increases in different lipid subclasses may have varying impacts on health by influencing circulating levels of HDL or VLDL (Figure ). S7a,b

Given that transcriptomics offers an alternative perspective into the metabolic networks at the level of the metabolic enzymes, we performed an integrated analysis of metabolomic and published transcriptomic data (Peters et al. 2015) to gain deeper insights into the metabolic trends in the elderly. Consistent with our hypothesis, the accumulation of amino acids in the elderly was supported by our identification of a positive correlation in transcriptomics and metabolomics (Figure 6c). Specifically, the accumulation of various amino acids produced from protein digestion and generated using TCA cycle metabolites as raw materials was observed, whereas the expression of these raw materials declined (Figure 6d). Furthermore, glutamine contributes to nucleotide synthesis, while the levels of metabolites and enzymes associated with nucleotide and purine metabolism decline. This suggests that in the elderly, amino acid production is high, but utilization is either untimely or sufficient. Even worse, the accumulation of substantial amino acids leads to an increase in phenolic degradation products, inflicting detrimental effects on health. It is worth noting that, the correlations between metabolomic features and transcriptomic data were derived from different populations, and thus should be regarded as only exploratory: the analysis results require validation in future studies using matched cohorts.

In the same way, given the focus on aging and the complex, dynamic nature of the aging process, we developed a "metabolite–lipid clock" to predict chronological age based on our data. Using the ElasticNet model, we constructed three clocks—a "metabolite clock," a "lipid clock," and a "combined clock"—across the age cohort to assess biological age through the integration of multi‐dimensional information (Figure 7a,b). A total of 1931 metabolites and 3795 lipids were used to build these models, ultimately constructing a metabolite clock comprising 107 metabolites, a lipid clock comprising 68 lipids, and a combined clock consisting of 53 components (Figure 7c,d). The three aging clock models were all maintained within a mean absolute error (MAE) of 8 years. Furthermore, the combined clock achieved the lowest MAE and the highest correlation, thereby improving the precision of biological age estimation by capturing changes across different aging trajectories (Figure 7c,d). In addition, we performed an additional validation of our metabolic aging clock using an independent cohort of 12 individuals not included in our dataset (Figure S8a). The model demonstrated consistent predictive accuracy in this validation cohort, supporting its preliminary robustness.

To enhance the practicality and ease of use in clinical applications, we subsequently applied broken‐stick regression to identify the breakpoint that balances the number of metabolites and prediction accuracy, thereby developing a reduced age clock model with fewer components while maintaining precise age predictions (Figure 7f). This clock included only five components, but nevertheless showed a strong correlation (r = 0.88) and robust predictive performance (Figure 7g,h). Similarly, we constructed reduced metabolite and lipid clocks, containing only nine metabolites and seven lipids, respectively, both of which accurately predicted age (Figure S8b–d). To identify aging‐related biomarkers, we applied three complementary selection strategies: (1) features significantly upregulated in the elderly (≥ 65 years), (2) features showing a linear increase across the lifespan, and (3) features with high predictive weight in our Elastic Net age model. The intersection yielded 32 metabolites and 27 lipids (Figure 7h), from which we constructed a concise five‐marker panel. The expression level of this panel positively correlated with the degree of aging. Further, this panel was validated by 10‐fold cross‐validation, achieving a mean area under the curve (AUC) of 0.952 (Figure 7i,j). We subsequently identified aging‐related biomarkers that accurately reflect age‐associated metabolic features.

Overall, our findings reveal the potential of metabolite component‐based aging clocks as valuable tools for advancing aging research, with particular attention to the combined clock for its high precision and minimal component breakpoint.

Samples of cord blood (CB) and peripheral blood (PB) were obtained from the First Affiliated Hospital of the University of Science and Technology of China. This study was approved by the Ethics Committee of the University of Science and Technology of China (approval no. 2022‐ky022; no. 2022‐ky062). Clinical characteristics are listed in Table . All adult blood samples were collected after overnight fasting (≥ 8 h) between 7:00 and 9:30 a.m. Participants were requested to avoid strenuous exercise and alcohol consumption within the 24 h prior to sample collection. Information on medication use and other clinical variables was reviewed to exclude individuals with acute or chronic diseases that could markedly affect metabolism. Fresh whole blood samples were collected in ethylenediaminetetraacetic acid (EDTA)‐containing tubes, and centrifuged to obtain plasma supernatant. Blood was processed within 1 h of sampling, and the collected plasma was stored at −80°C until metabolite extraction. S1

For metabolite extraction, 100 μL of plasma was mixed with 400 μL of ice‐cold methanol/acetonitrile (1:1, v/v) to precipitate proteins. After a 20‐min centrifugation at 14,000 g and 4°C, the supernatant was vacuum dried. Samples were then reconstituted in 100 μL acetonitrile/water (1:1, v/v) and centrifuged again prior to LC–MS.

The MTBE method was used for lipid extraction. Briefly, 200 μL of water was added to the sample and then vortexed. Afterward, 240 μL of precooled methanol was added, followed by vortexing for 30 s. Next, 800 μL of MTBE was added, after which the mixture underwent a 20‐min ultrasound at 4°C, followed by incubation at room temperature for 30 min. Following centrifugation of the solution at 14,000 g for 15 min at 10°C, the upper organic layer was collected and dried under nitrogen.

The analysis was conducted using an UHPLC system (Agilent 1290 Infinity) paired with a quadrupole time‐of‐flight mass spectrometer (AB Sciex TripleTOF 6600). A 2.1 mm × 100 mm ACQUIY UPLC BEH Amide 1.7 μm column was used for HILIC separation. In both ESI positive and negative modes, the mobile phases were 25 mM ammonium acetate/ammonium hydroxide (A) and acetonitrile (B). The gradient began at 95% B, and was gradually decreased to 65% over 6.5 min, then down to 40% in 1 min, held for 1 min, and returned to 95% B for re‐equilibration. MS scan conditions included a range of m/z 60–1000 Da, with a 0.20 s accumulation time, while auto MS/MS scans covered m/z 25–1000 Da with a 0.05 s accumulation. Parameters for auto MS/MS used IDA with a fixed collision energy of 35 V (±15 eV) and declustering potentials (DP) of 60 V (+) and 60 V (−).

For eliminating contaminants in samples, blank samples (e.g., solvent blanks) were included in each batch to monitor potential background contamination, with detected signals consistently lower in intensity than those in biological samples to minimize their impact on differential analysis. Low‐intensity signals below predefined thresholds were systematically filtered out during peak extraction and alignment to reduce background noise. Features frequently detected at high intensity in blank samples, such as reagent additives or carryover contaminants, were flagged and excluded from downstream analysis.

Reverse phase chromatography was applied for LC separation with a CSH C18 column (1.7 μm, 2.1 mm × 100 mm, Waters). Lipid extracts were re‐dissolved in 200 μL of 90% isopropanol/acetonitrile, centrifuged at 14,000 g for 15 min, and 3 μL of sample was injected. Solvent A consisted of acetonitrile–water (6:4, v/v) with 0.1% formic acid and 0.1 mM ammonium formate, and solvent B was acetonitrile–isopropanol (1:9, v/v) with the same additives. The flow rate started at 30% B, increased to 100% over 23 min, and re‐equilibrated. Mass spectra were acquired using Q‐Exactive Plus in the positive and negative modes, respectively. ESI settings were as follows: 300°C source temperature, 350°C capillary temperature, 3000 V ion spray voltage, and scan range of m/z 200–1800. For lipid identification, we used LipidSearch (v.4.0), a widely recognized lipidomics software. Lipid molecules and internal standards were processed using peak recognition, peak extraction, and MS/MS‐based lipid identification. The main parameters applied were as follows: precursor tolerance of 5 ppm, product tolerance of 5 ppm, and product ion threshold of 5%. These stringent settings ensured accurate annotation while minimizing redundancy.

Raw MS data were converted to MzXML format using ProteoWizard MSConvert (v.3.0.6428) (Chambers et al. 2012) and processed in XCMS online (v.3.7.1) (Benton et al. 2008). Peak picking was performed with centWave (m/z = 10 ppm, peakwidth = c (10, 60), prefilter = c (10, 100)), and peak grouping with bw = 5, mzwid = 0.025, minfrac = 0.5. Isotopes and adducts were annotated using CAMERA (Kuhl et al. 2012). Only variables with more than 50% nonzero values in at least one group were retained. Metabolites were first screened against a credible spectral database (this database integrates over 6000 compounds from our in‐house database with entries from four major public databases—MassBank, METLIN, MoNA, and the commercial version of HMDB), maintaining a strict mass accuracy threshold of ≤ 10 ppm to minimize false positives. Identification was further refined by comparing fragmentation patterns with spectral databases, ensuring a high‐confidence annotation with a similarity score > 0.6, thereby improving accuracy beyond simple precursor mass matching (Schrimpe‐Rutledge et al. 2016). All identified metabolites meet at least MSI Level 2 standards. To ensure data quality, all biological samples were randomized prior to LC–MS injection. In addition, pooled QC samples were injected at regular intervals (every 3 samples) throughout the analytical sequence to monitor instrument stability and signal drift. We applied QC‐based Support Vector Regression (QC‐SVR) for batch correction, a well‐established approach in large‐scale metabolomics studies (Dunn et al. 2011; Shen et al. 2016). Additionally, we applied total sum normalization (TSN) after QC correction to improve comparability across samples. Data quality was comprehensively evaluated: TIC overlap of QC samples confirmed consistent chromatographic performance; PCA and Pearson correlation analyses revealed tight QC clustering and correlations > 0.9; Hotelling's T² test indicated all QC samples fell within the 99% confidence interval, and the proportion of metabolites and lipids with a coefficient of variation of < 30% in QC samples was maintained at > 80%. Specifically, the final mean/median CV values are 20.67%/20.072% for metabolomics and 18.532%/11.469% for lipidomics. The normalized intensity was obtained by dividing the raw intensity by the sum of each sample. To assess the robustness of the results, the raw intensity was further divided by the median of each sample for normalization in key findings.

The normalized abundance matrices of log2‐transformed metabolites or lipids were used for PCA and PVCA, using the prcomp and pvcaBatchAssess functions under R stat (v.4.1.2) and pvca (v.1.34.0) (Li et al. 2009) packages. The results were visualized using the factoextra (v.1.0.7) and ggplot2 (v.3.5.1) packages.

Statistical differences in the metabolites or lipids were analyzed using the R stats (v.4.1.2) package with Student's t‐test (two‐sided) and adjusted using the Benjamini–Hochberg (BH) multiple comparisons test. For comparisons between age groups, adjusted p values < 0.05 were considered statistically significant. The UpSetR (v.1.4.0) (Conway et al. 2017) and VennDiagram (v.1.7.3) (Chen and Boutros 2011) packages were used to identify intersections between different sets. Heatmaps of metabolite or lipid expression were generated using the Complexheatmap (v.2.10.0) (Gu et al. 2016) package. The age‐related change scores were calculated based on the previously reported method (Li, Xiong, et al. 2023).

To annotate the pathways of the DEMs, we downloaded the KEGG metabolic pathways and associated metabolites from the KEGG API (https://www.kegg.jp/kegg/rest/keggapi.html↗). KEGG pathway enrichment was subsequently performed using the R clusterProfiler (v.4.2.2) (Wu et al. 2021) package and significant pathways were obtained with adjusted p value < 0.05 and minGSSize = 2.

To characterize age‐related metabolic and lipid changes, the following model was implemented, with sex included as a covariate to adjust for potential confounding:Expression~α+β1Age+β2Gender+εwhere "α" represents the y‐intercept, "β" represents the slope, and "ε" represents residual error. Log10 transformed normalized matrices of metabolites or lipids were used as expression values. The linear model and Type II sums of squares were generated using the LM and Anova functions in the R stat (v.4.1.2) and car (v.3.1.2) (Weisberg and Fox 2010) packages, respectively. Significant changes were classified by adjusted p values < 0.05.

Nonlinear patterns of metabolite and lipid expression with age were determined using LOESS models, implemented using the loess function from the R stats (v.4.1.2) package. The predicted age‐related trajectories of metabolite and lipid expression were subsequently hierarchically clustered into eight groups using the hclust function in the R stats (v.4.1.2) package.

The DE‐SWAN (v.0.0.0.9001) (Lehallier et al. 2019) package in R was used to identify transient changes in metabolite and lipid expression across the lifespan, with sex included as a covariate to adjust for potential confounding. Comparisons were made in 5‐year intervals across a 20‐year window. Significant changes were determined using BH‐adjusted p values < 0.05 and "β" thresholds.

Correlations between the top‐50 age‐increased metabolites and lipids were calculated through Spearman's correlation conducted using the cor function in the R stat (v.4.1.2) package. In addition, we analyzed the correlation between components and clinical information of the samples. The bulk RNA‐seq count matrices of GSE75337 and GSE103232 (Aramillo Irizar et al. 2018) were downloaded and normalized using the DEseq2 (v.1.34.0) (Love et al. 2014) package to perform the same correlation analysis with our metabolite data. The p value of the correlation coefficient was calculated by the t‐test formula. Network visualization was performed using Cytoscape (v.3.10.2) (Shannon et al. 2003) software.

To evaluate the ability of our metabolomics and lipidomics data to predict organismal age, we constructed models using different training set sizes and alpha values of 0.1–0.9 with the glmnet (v. 4.1.8) (Friedman et al. 2010) package. Elastic Net regression, which combines the advantages of an L1‐norm (Lasso) regression and an L2‐norm (Ridge) regression model was applied, ensuring that feature selection is robust (Friedman et al. 2010). Lambda values were optimized by 10‐fold cross‐validation. At least 50% of the dataset was used for training to balance model learning and generalization, with the remaining samples reserved for validation. The model with the smallest MAE was selected as the final model.

To determine whether a subset of metabolites or lipids could also be used to predict age with comparable accuracy, we developed a reduced model. First, we obtained the metabolites and lipids included in the initial age‐clock model. These features were ranked based on the absolute values of their correlation coefficient, with higher absolute values indicating a greater contribution to the model. A series of elastic net models were constructed by progressively incorporating top‐ranked features (alpha value: 0.1–0.9, train set size: 80%). Specifically, models were built using the top 2, top 3, top 4 features, and so on, systematically evaluating how model performance changes with feature number. The predictive accuracy of these models was then calculated, and the segmented (v. 1.6.1) (Fasola et al. 2018) package was used to identify the breakpoint between the number of metabolites and prediction accuracy.

To identify robust metabolic features associated with aging, the overlapping components (32 metabolites and 27 lipids) were used to train a logistic regression model with glmnet (v4.1.8). Model performance was evaluated by 10‐fold cross‐validation implemented via the caret (v6.0.93). Receiver operating characteristic (ROC) analysis was conducted using pROC (v1.18.0) (Robin et al. 2011).

CBMCs were isolated from cord blood by Ficoll cell separation solution (TBDscience, Cat# LTS1077) density gradient centrifugation. After isolation, CBMCs were cultured in RPMI 1640 (Biosharp, Cat# BL303A) supplemented with 10% FBS (Sigma, Cat# F7524) and 1 × penicillin–streptomycin–amphotericin (Biosharp, Cat# BL142A). For metabolite treatment, CBMCs were resuspended in RPMI 1640 complete medium plus SPH (10 μM, MCE, Cat# HY‐W019838) for 24 h.

CBMCs were labeled with conjugated antibodies for 30 min at 4°C and subsequently analyzed by flow cytometry. IgG isotype controls were used for all staining procedures. CD45 (Biolegend, Cat# 304036), CD3 (Biolegend, Cat# 300420), CD19 (Biolegend, Cat# 302243), CD56 (Biolegend, Cat# 362552), CD4 (Biolegend, Cat# 357417), CD8 (Biolegend, Cat# 344713), CD69 (Biolegend, Cat# 310904), CD38 (Biolegend, Cat# 303506), DNAM‐1 (Biolegend, Cat# 338314), PD1 (Biolegend, Cat# 621610), TIGIT (Biolegend, Cat# 372704), and Tim3 (Biolegend, Cat# 345022) antibodies were used. For apoptosis analysis, cells were stained with Annexin V/PI (Elabscience, Cat# E‐CK‐A217) for 15 min and analyzed by flow cytometry. Flow cytometry was performed on a Beckman Coulter CytoFLEX flow cytometer, and the data were analyzed using FlowJo software. The cell gating strategies were shown in Figure . S3a

R statistical software v4.1.2 was used for all statistical analyses. Statistical methods are described in the figure legends as appropriate.