Ticking Differently: Elucidating Sexual Dimorphism in Human Aging through Metabolomics, Proteomics and Genomics
1Nuffield Department of Population Health, University of Oxford, Oxford, UK
2Centre for Medicines Discovery, Nuffield Department of Medicine, University of Oxford, UK
3Centre of Artificial Intelligence in Precision Medicine (CAIPM), King Abdulaziz University, Jeddah, Saudi Arabia
4Oxford Internet Institute, Oxford University, Oxford, UK
5Department of Biochemistry, Faculty of Science, King Abdulaziz University, Jeddah, Saudi Arabia
6Oxford Centre for Diabetes, Endocrinology and Metabolism, Radcliffe Department of Medicine, University of Oxford, Oxford, UK
7Analytic and Translational Genetics Unit, Massachusetts General Hospital, Boston, MA, USA
8Program in Medical and Population Genetics, Broad Institute of MIT and Harvard, Cambridge, MA, USA
9Department of Psychiatry, University of Oxford, Oxford, UK
10Department of Medicine, Duke University, Durham, NC, USA
11Department of Psychiatry and Behavioural Sciences, Duke University, Durham, NC, USA
12Department of Brain Sciences, Duke University, Durham, NC, USA
13Oxford-GSK Institute of Molecular and Compuational Medicine (IMCM), Oxford, UK
+Correspondence to Prof. Cornelia M. van Duijn, Nuffield Department of Population Health, University of Oxford, Old Road Campus, OX3 7LF, Oxford, UK, Email: cornelia.vanduijn@ndph.ox.ac.ukAbstract
That males and females age differently has been overlooked while developing aging clocks. Here, we developed a sex-specific metabolic aging clock in 390,941 individuals from the UK Biobank and integrated it with genetic, proteomic and epidemiological data to identify mechanisms accelerating/decelerating metabolic aging in males and females. Our findings reveal dysregulation of cholesterol metabolism, immune system, hemostasis, and cell growth, survival and apoptosis as common mechanisms accelerating metabolic aging in males and females, and upregulation of oxidative stress detoxification, cellular resilience and tissue integrity as common mechanisms decelerating metabolic aging. In females, a further dysregulation of carbohydrate/glucose metabolism, circadian rhythm and hormone metabolism accelerating metabolic aging is observed, while dysregulation of energy metabolism, cancer and longevity pathway is specifically observed in males. Among reproductive factors, late puberty and higher parity manifest as protective factors, decelerating metabolic aging in both sexes, and additionally childbirth at older age decelerating metabolic aging in women. Accelerated metabolic age strongly predicted morbidity and mortality in both sexes, except that the magnitude of association was several folds higher in males, and obesity explained most of the disease associations in females, suggesting that obesity influences metabolic aging and subsequent health outcomes differently in males and females. Consistent with the upregulation of molecular mechanisms involved in cancer in males, accelerated metabolic aging predicted several common cancers in males but not females.
Our study provides novel insights into the biological mechanisms underlying aging and disease susceptibility in males and females, underscoring the importance of considering sex differences in healthcare strategies and public health policies.
Article notes
Competing Interest Statement
The authors have declared no competing interest.
Funding Statement
All UK Biobank data was accessed under UK Biobank Application number 30418. This work was supported by the Centre of Artificial Intelligence in Precision Medicines (CAIPM), King Abdulaziz University, Jeddah, Saudi Arabia. We thank Nightingale Health Plc for providing early access to the UK Biobank dataset.
Main Text
Females outlive males globally1 although females are frailer2 and have poorer physical functioning compared to males3. Both historical and recent data suggest that females live longer even during severe famines and epidemics4-7. These differences have been attributed to various factors, including hormonal influences such as estrogen8,9, genetics10, telomere length11-13, metabolic rate10,14, adiposity distribution15,16, and behavioral differences17 where males tend to engage in risky behaviors17 and often have poorer lifestyle habits, such as higher alcohol consumption, excessive smoking and poorer oral hygiene18-20. Such environmental insults are captured by epigenetic processes like DNA methylation, which shows a higher entropy in males with advancing age10,21,22.
Biological age predictors, often referred to as “aging clocks”, have been developed to estimate biological aging across different omics layers, demonstrating their substantial potential in aging research23-31. Most of these clocks, however, did not investigate sex-specific variations underlying aging and the association of biological age with disease and mortality outcome in each sex. A recent study of Reicher L et al.32 explored the differences in aging between sexes but in a relatively small sample of 10,000 individuals using environmental exposures, physiological parameters, and a narrow platform of molecular markers mainly concentrating on lipids. Further, all the previous studies focused on predicting chronological age and studying the impact on morbidity and mortality rather than understanding the mechanisms underlying accelerated/decelerated aging.
In this study, we develop metabolites-based sex-specific aging clocks leveraging the largest ever available data from 488,318 individuals in the UK Biobank with a focus on understanding common and sex-specific mechanisms underlying accelerated/decelerated metabolic aging. We first constructed a sex-specific metabolic age acceleration (metAgeGap) metric, which estimates the variation in individuals’ rate of metabolic aging compared to others with the same chronological age group. Next, we studied its association with sex-specific reproductive factors, future risks of age-related diseases and mortality, and predicted survival. Finally, to explore the common and divergent molecular regulators of metAgeGap in males and females, we performed a genome-wide and a proteome-wide association analysis of the sex-specific metAgeGap.
Results
Descriptive Statistics
The study population comprised 488,318 participants, including 223,396 males and 264,922 females. Participants taking lipids lowering drugs were excluded from the study, resulting in 168,460 males (mean age=55.86, sd=8.23) and 222,481 females (mean age=56.01, sd=8.00). Over the follow-up period, which averaged 12.1 years for males and 12.4 years for females, there were 2,343 recorded deaths in males and 1,602 in females (Figure 1A).
Most cancers, except lung cancer and non-Hodgkin lymphoma, were more prevalent in males compared to females. Females had a higher prevalence of most common diseases except cardiovascular diseases (ischemic heart disease, ischemic stroke, all stroke) and dementia, including vascular dementia (Supplementary Table 1). Alcohol consumption and smoking were more prevalent among males. Females reported sleep difficulties and tiredness more often (Supplementary Table 2).
Correlation of disease-related risk factors with metAgeGap
Next, we tested the association of metAgeGap with 22 common disease-related risk factors (Figure 2). Obesity (females: beta=0.94, p=2.25*10-308, males: beta=0.30, p=1.58*10-40), BMI (females: beta=0.54, p=2.25*10-308, males: beta=0.12, p=1.16*10-37), and blood pressure (systolic blood pressure - females: beta=0.26, p=1.72*10-196, males: beta=0.10, p=1.08*10-24; diastolic blood pressure - females: beta=0.42, p=2.25*10-308, males: beta=0.07, p=9.91*10-12) showed a stronger association with accelerated metabolic aging in females, whereas type-II diabetes (females: beta=0.31, p=2.05*10-03, males: beta=1.05, p=3.02*10-35) showed stronger association in males. Interestingly, smoking (ever smoked and pack years smoked) had a larger impact on metabolic aging in females compared to males (Figure 2).
Proteins involved in detoxification of Reactive Oxygen Species decelerate metabolic aging
To identify proteins involved in accelerated/decelerated metabolic aging, we performed a proteome-wide association analysis of metAgeGap in the UK Biobank. In females, 1,739 out of 2,923 proteins were significantly associated (FDR < 0.05) with metAgeGap, of which 1,587 associated with older metabolic age and 152 with younger metabolic age (Supplementary Table 10, Figure 3). In males, 1,942 proteins were significantly associated with metAgeGap, of which 1,115 associated with older metabolic age and 827 with younger metabolic age. 906 proteins were consistently associated with older metabolic age and 52 proteins were consistently associated with younger metabolic age in both males and females. Proteins that associated with older metabolic age clustered in immune system regulation, coagulation, PI3K-Akt signaling, lysosome, apoptosis, intestinal IgA production, regulation of Insulin-like Growth Factor (IGF) transport/uptake, tumour necrosis factor (TNF), nuclear factor-κB (NF-κB) and MAP kinase (MAPK) signaling pathways among several others (Supplementary Table 11). Proteins that were associated with younger metabolic age clustered in cadherin binding and detoxification of reactive oxygen species (ROS) (Supplementary Table 12).
When comparing males and females, 366 proteins were significantly associated with older metabolic age in females but younger metabolic age in males. These proteins include LDLR, which also showed a significant association in GWAS of metAgeGap in females but not in males, LCAT and APOM, all of which are cholesterol-related proteins. These 366 proteins clustered in mechanisms involved in hemostasis (Supplementary Table 13). 385 proteins showed association with metAgeGap in females only. These proteins were largely in hormone metabolism, immune system regulation and breast cancer (Supplementary Table 14). There were only 30 proteins that were significantly associated with older metabolic age in males but younger metabolic age in females. These proteins were involved in various bindings mechanisms and cytokine receptor activity (Supplementary Table 15). Further there were 588 proteins associated with metAgeGap in males only. These proteins were involved in apoptosis, cancers including colorectal, lung, pancreatic and prostate cancers, FOXO signaling in particular AKT-mediated inactivation of FOXO1A, AMPK signaling and various viral and bacterial infections (Supplementary Table 16).
Genetics, lifestyle and sex hormones differentially influence aging associated sex dimorphic proteins
To gain insights into the top sex dimorphic aging proteins we evaluated their age distributions (Figure 4) and the impact of BMI, smoking, alcohol use, menopause and hormone replacement therapy (HRT) on the distribution of these proteins (Supplementary Figures 10-15). Since the top proteins associated with higher metabolic age in females are involved in lipid metabolism, we further evaluated the age distribution of density lipoprotein (LDL) (Supplementary Figure 14). BMI significantly increased expression of proteins involved in hemostasis (F7 & PROC) and lipid metabolism (LDLR, PCSK9 & LDL) in both males and females, however females in the normal weight and underweight categories also displayed high levels of these proteins and LDL over age, unlike normal and underweight males, suggesting that while BMI explains the differences within the females in metabolic aging, it does not explain the difference between males and females (Supplementary Figure 10). Current smoking significantly increased levels of blood clotting factors (F7, PROC and PLA2G7) and LDLR in females but not in males, which explains the observed association between smoking and metabolic aging in females. Smoking also significantly lowered the plasma levels of TTR and CEACAM16, both of which showed significant protective effects on metabolic aging in males (Supplementary Figure 11) but not in females. No noticeable differences were observed between males and females when the impact of alcohol use was evaluated (Supplementary Figure 12). Female-specific aging proteins (F7, PROC, LDLR and PCSK9) were influenced by menopause and to some extent HRT use (Supplementary Figure 13). No noteworthy difference was observed between females with low and high estradiol levels (Supplementary Figure 14). However, males with low testosterone levels showed significantly different levels of most of the aging associated sex dimorphic proteins (Supplementary Figure 15).
To explore whether the protein quantitative loci (pQTLs; cis & trans) underlying these top proteins have differential influence on metabolic aging in males and females we compared the male and female specific summary statistics of all the pQTLs (Supplementary Table 17) from the GWAS of metAgeGap (Supplementary Figure 16). The most interesting finding was a particular missense variant (rs1260326_T) in GCKR gene, a locus that showed genome-wide significance in male specific GWAS only (Supplementary Figure 6), influenced the plasma levels of seven of these sex dimorphic proteins including F7, CEACAM16, LDLR, NOTCH3, PCSK9, TTR and NPY. Rs1260326_T was significantly negatively associated with metaAgeGap in males (beta=-0.23, p-value = 1.04*10-40) and showed suggestive positive association in females (beta=0.06, p-value=3.64*10-06) (Supplementary Table 18). The second most interesting gene was MLX1PL, whose multiple different variants influenced the levels of five proteins, including F7, LDLR, CEACAM16, NPY and TTR (Supplementary Table 17). All these variants showed inverse associations with metAgeGap in males and females (Supplementary Table 18). Among other pQTLs that showed differential effects in males and females include those in PCSK9, CELSR2, SMARCA2/LDLR, BUD13-DT, and TIMD4/HAVCR1 that showed genome-wide significance in female-specific metAgeGap GWAS only and pQTLs in CYP26A1, LPA and PRKCA that showed genome-wide significance in males only. Two pQTLs in APOE, including rs7412_C and rs445925_G, also showed differential effects on metAgeGap in males and females. All these pQTLs influenced plasma levels of CEACAM16, F7, LDLR, NPY, PCSK9 and PLA2G7 (Supplementary Table 17).
MetAgeGap predicts several common cancers in males but not in females
When comparing disease risks of the biologically youngest group (bottom 10% metAgeGap) with that of biologically oldest group (top 10% metAgeGap), youngest males had a higher cumulative risk for ischemic heart disease (IHD) than the oldest females, reflecting a higher baseline risk in males (Figure 5). Further, the differences in cumulative risks of COPD and mortality between the top and bottom 10% were more pronounced in males than in females. The increase in cumulative risk for osteoarthritis was more gradual and linear in all sexes, with females being more at risk than males after the age of 50, suggesting a different risk trajectory in all sexes. For chronic kidney and liver diseases, no differences were observed between males and females, where biologically younger individuals showed significantly lower cumulative risks compared to biologically older individuals (Figure 5, Supplementary Figure 17).
When studying the association of metabolic aging with 15 incident morbidities and mortality (Figure 6), stark differences were observed between males and females for cancers (Supplementary Tables 19 & 20). MetAgeGap was significantly associated with liver (HR=1.74, CI=1.52, 1.99), oesophageal (HR=1.30, CI=1.19, 1.42) and lung cancers (HR=1.10, CI=1.04, 1.17) only in males (Figure 6A, 6B). Adjusting for alcohol use and smoking did not attenuate these associations (Supplementary Figures 18-21). Colorectal cancer was significantly associated with metAgeGap in all sexes. In females after adjusting for BMI and physical activity, metAgeGap’s association with kidney (HR=1.11, CI=0.99, 1.24) and breast (HR=1.02, CI=1.01, 1.04) cancers attenuated.
Among non-cancer illnesses, except for neurodegenerative disorders (Alzheimer’s and Parkinson’s), metAgeGap was significantly associated with all diseases in males (Figure 6C, 6D). MetAgeGap was significantly associated with dementia (HR=1.07, CI=1.02, 1.12), ischemic stroke (HR=1.14, CI=1.10, 1.18), macular degeneration (HR=1.05, CI=1.01, 1.08), all strokes (HR=1.11, CI=1.08, 1.15), rheumatoid arthritis (HR=1.08, CI=1.04, 1.12), and osteoporosis (HR=1.11, CI=1.07, 1.16) in males but not in females (Supplementary Tables 21 & 22). The impact of metAgeGap on mortality and the risk of heart disease was far more pronounced in males. Adjusting for BMI led to a higher dilution of effects in females but not in males, especially for kidney, liver, and heart diseases, suggesting that BMI can have a different impact on aging in males and females (Figure 6C, 6D).
Discussion
In this study, we developed and validated a sex-specific metabolic aging clock using NMR-based metabolome data from the UK Biobank, enabling us to investigate biological aging through a metabolic lens in a population of 488,318 individuals. Our findings show that metabolic aging, as captured by the metAgeGap metric, is a complex and sex-differentiated process with distinct biochemical, phenotypic, proteomic and genetic correlates.
First, that our models could predict chronological age with moderate accuracy (R2 = 0.29 for males, 0.37 for females) suggests that the metabolome holds reliable, albeit incomplete, clues about the biological clock. Interestingly, females consistently showed stronger prediction performance—a possible reflection of the tighter regulation or narrower variability of metabolic processes in women. While 61 features were common to both sexes—suggesting conserved core pathways—the presence of 32 female-specific and 15 male-specific features points to divergent metabolic processes driving aging. In females, unique contributions from cholines and phosphatidylcholines align with known roles in membrane biology and lipid signaling34,35, whereas males exhibited distinct associations with HDL particle size and triglyceride composition36,37, implicating differences in lipid transport and cardiovascular risk.
The metAgeGap metric showed strong and consistent associations with cardiometabolic risk factors, but the magnitude and pattern of these associations were sex-dependent. Obesity, BMI, smoking and blood pressure exhibited stronger associations in females, while type 2 diabetes showed the largest effect in males. These differences underscore that while the pathways to aging may be parallel, they are not identical and therefore sex should be considered when examining the metabolic determinants of biological age and their relevance to disease risk.
Perhaps the most intriguing finding of the current study is the association of puberty and hormonal exposures with metabolic aging. In women, later puberty, higher parity, later age at first and last live births and hormone use paint a portrait of youth preserved, with significantly younger metabolic ages. Meanwhile, in men, delayed signs of puberty — like voice breaking or facial hair — and number of children fathered were similarly linked to a slower metabolic clock. These findings suggest that earlier exposure to sex hormones accelerates biological aging in both sexes, consistent with hypotheses proposing a trade-off between early reproductive maturation and long-term somatic maintenance. These findings align with previous evidence suggesting that hormonal and physiological factors related to reproduction can modulate long-term metabolic health and aging trajectories38,39. However, our findings of protective effects of parity in both females and males contradict the evolutionary theories of aging, which hypothesize a trade-off between reproduction and lifespan40-43. Literature shows conflicting results about parity and longevity in women40-43 but somewhat consistent in males44. Our findings appear more plausible with the hypothesized effects of fetal michrochimerism on maternal health and longevity45. Further later age at birth has also been found to associate with longevity in females44.
Crucially, metAgeGap was significantly associated with incident morbidity and mortality outcomes, again revealing pronounced sex differences. Males with higher metabolic age showed elevated risks across nearly all disease categories, including cardiovascular, renal, hepatic, and oncological outcomes. Notably, liver, esophageal, and lung cancers exhibited strong associations in males, independent of lifestyle factors such as alcohol use and smoking. In contrast, females displayed fewer and weaker associations, although osteoarthritis risk increased significantly after midlife, and some disease risks (e.g., breast and kidney cancer) were attenuated after adjustment for BMI and physical activity. The attenuating effect of BMI was consistently more pronounced in females, suggesting that the interplay between adiposity and metabolic aging may differ by sex. These findings are consistent with earlier studies, which showed a stronger association between BMI and mortality in females than in males46. The sex difference can be explained by the fact that females tend to accumulate more subcutaneous fat, while males store more visceral fat47-49, which is more closely linked to metabolic risks, highlighting the importance of developing sex-specific clocks. The female body seemed to carry a certain metabolic resilience, especially for cancers and neurodegenerative diseases.
In the GWAS, 217 genes overlapped between males and females, indicating a shared genetic basis of metabolic aging, bringing out cholesterol metabolism and PPAR signaling as common pathways of metabolic aging. These included the APOE gene, which is one of the two genes consistently been associated with longevity in large-scale GWAS33,50 and has also shown sex dimorphic effects in earlier studies on plasma protein levels51,52. Of note is that a large number of genomic loci (147 vs 120) were identified in females, suggesting potentially higher polygenic complexity. Despite a larger gene set, no unique pathway was observed for female-specific genes beyond those shared. There were, however, some interesting genes including those involved in glucose transport, e.g., SLC2A2 and SLC2A6, lactose intolerance and gut microbiome composition LCT, circadian rhythm TIMELESS, virus receptors HAVCR1 and HAVCR2, and genes involved in metabolic and brain health CELSR2 and CELSR1 that showed no association with metabolic aging in males. In contrast to females, male-specific loci included genes such as IL6R, GCKR, SNCAIP, HKDC1 and APOH, some of which are implicated in inflammation and metabolism. Female-specific metAgeGap genetic correlations were observed with iron-deficiency anemia, obesity and cardio-vascular pathology, suggesting that metabolic aging in females may be more closely tied to cardiometabolic health, while male-specific metAgeGap genetic correlations with sex hormones, mouth ulcers and atrial fibrillation suggest a distinct health profile associated with male metabolic aging.
A key observation from this study is the enrichment of proteins related to immune regulation, coagulation, and several canonical signaling pathways (e.g., PI3K-Akt, NF-κB, MAPK, TNF, IGF) among those associated with older metabolic age. These pathways are well-established players in aging and age-related diseases. For instance, chronic low-grade inflammation (“inflammaging”)53 is increasingly recognized as a hallmark of aging and has been implicated in metabolic dysfunction, cardiovascular disease, and neurodegeneration54. The prominence of immune and inflammatory pathways in our analysis suggests that immunosenescence and dysregulated immune signaling may be central to the biological aging of metabolism.
Conversely, proteins linked to younger metabolic age were enriched in pathways involved in cell adhesion (cadherin binding) and oxidative stress detoxification. These processes are associated with cellular resilience and tissue integrity—features that are typically diminished with aging. The detoxification of reactive oxygen species (ROS) is critical for maintaining mitochondrial function and preventing oxidative damage, a major contributor to age-related decline55. The two most important antioxidant enzymes include superoxide dismutase (SOD) and glutathione peroxidase (GPx)55. In our study, we found low levels of SOD1 and peroxiredoxin 6 (PRDX6), which is a bifunctional enzyme with GPx activity, both in males and females with older metabolic age. The inverse association of these proteins with metAgeGap suggests potential protective mechanisms that could be targeted to slow down metabolic aging. Of note is that these redox mechanisms were not captured by the proteomic aging clock that we developed earlier although it showed much higher accuracy (R2 = 0.88) in predicting chronological age56.
While there was considerable overlap in proteins associated with metAgeGap across sexes, we observed marked divergence in both the number and functional roles of proteins that were uniquely or differentially associated. For example, 366 proteins were associated with older metabolic age in females but with younger metabolic age in males, and many of these were involved in lipid metabolism and hemostasis. Since we removed individuals who were on lipid lowering medication, these findings point towards sex disparities in lipid regulation, which is affected by menopause57 incidence, presentation and in particular treatment of cardiovascular diseases that remain underdiagnosed and undertreated in females compared to males thus affecting longevity in females58. Our study further shows that lifestyle factors such as smoking can worsen lipid metabolism and hemostasis in females but not in males across all ages.
The sex-specific proteomic associations further revealed pathways linked to cancer, hormone metabolism, and infectious disease susceptibility. In males, proteins uniquely associated with older metabolic age were enriched in apoptosis, FOXO and AMPK signaling—pathways with established roles in tumor suppression and longevity regulation59,60. The differential involvement of FOXO signaling, particularly via AKT-mediated inactivation of FOXO1A, aligns with evidence that metabolic stress and insulin resistance modulate aging-related pathways differently in males and females.
While this study is the largest of its kind to-date, there are, however, some limitations. First, the findings may not be generalizable to populations of non-European ancestries as the UK Biobank primarily includes individuals of European descent. Therefore, the impact of metabolic aging on disease risks in other ethnic groups remains to be explored. Second, the Nightingale metabolomics platform predominantly captures lipid-related metabolites, which explains the low predictive power in our study. This also led us to remove over 100k participants that were on lipid lowering medications, further impacting the statistical power of the study, particularly in studying the association of metAgeGap with rarer cancers. More metabolically diverse platforms may provide better resolution and insights into aging. Finally, the age range of the participants in the UK Biobank is relatively narrow, a wider age range may have provided a better fit. Nevertheless, with metAgeGap we have provided novel insights into the intrinsic and extrinsic factors leading to differences in metabolic aging in males and females.
In conclusion, although aging affects everyone, its pace and impact differ between males and females due to genetic, hormonal and environmental factors. The need to understand these differences extends beyond scientific curiosity, impacting real-world healthcare practices and outcomes. For example, cardiovascular disease is often underdiagnosed or misdiagnosed in females because they frequently experience atypical symptoms, such as nausea or fatigue, rather than the more well-known chest pain seen in males61,62. Our study demonstrated that the sex-specific metabolic aging clock is a powerful tool for measuring biological age and capturing aging signatures that are linked to common age-related diseases in both males and females. Our findings highlight the potential of this clock to identify the biological mechanisms underlying accelerated and decelerated aging and emphasize the importance of sex differences in aging processes. These clocks can shape our knowledge of sex-specific disease susceptibility, rates of physiological decline, and overall longevity, paving the path for more personalized prevention, treatment strategies and refined public health policies.
Methods
Study cohort
The study was performed in the UK Biobank (UKB). The UKB is a prospective cohort study including 502,505 participants recruited between 2006 and 201063. Information on sociodemographic factors, lifestyle, early life, family history, psychosocial factors, health and medical history was collected through touchscreen questionnaires at baseline.
Linked hospital inpatient data, primary care data and cancer register data were accessed from the UKB data portal in August 2024, with a censoring date of November 30, 2023, December 31, 2023 and November 30, 2023 for participants recruited in England, Scotland, and Wales respectively. The follow-up time is between 8 and 16 years. Mortality data and cause of death information were accessed from the UKB data portal in August 2024, with a censoring date of November 30, 2022. The follow-up time is between 12 and 16 years. Methods and ICD diagnosis codes used to identify prevalent and incident chronic disease in UKB are shown in Supplementary Tables 23 & 24.
Assessment of Proteins
Proteomic profiling of 54,219 participants from the UK Biobank was carried out for protein analytes measured via the Olink Explore platform that links four Olink panels (Cardiometabolic, Inflammation, Neurology, and Oncology). UK Biobank Olink data are provided as Normalized Protein eXpression (NPX) values on a log2 scale. Details on sample selection, processing, and quality control are provided elsewhere65.
Genotyping
Genotyping was conducted by Affymetrix using a bespoke BiLEVE Axiom array for ∼50K participants and the remaining ∼450K on the Affymetrix UK Biobank Axiom array. As the two arrays are broadly comparable with over 95% overlap in assessed gene variants, they were combined. Genetic data was phased prior to imputation with SHAPEIT3 followed by imputation using IMPUTE2. Details on genetic imputations are provided elsewhere66. The APOE gene (alleles APOE ε2, APOE ε3, APOE ε4) was directly genotyped and defined by 2 single-nucleotide polymorphisms (SNPs), rs429358 and rs7412. Detailed information about the genotyping process and technical methods is available online.
Statistical analysis
Prediction of biological age
Biological age was predicted using 249 metabolites in males and females separately in a gradient boosting model. Samples were randomly split into a 70% training and 30% testing dataset. The model was first hyperparameter tuned using a Tree-structured Parzen Estimator (TPE) based method provided by the Optuna68 package in Python. Hyperparameters within a pre-set range were searched and optimized across 200 trials to maximize the 5-fold cross-validated Receiver Operating Characteristic (ROC) Area Under the Curve (AUC) score. After hyperparameter tuning, the performance of the best parameter in the training dataset with 5-fold cross-validation and in the 30% left out testing dataset was assessed.
Feature interpretation and selection
To characterize the feature importance, SHapley Additive exPlanation (SHAP), a local tree explaining method based on game theory, was used69. SHAP calculates the contribution of each feature to the outcome in each individual and extends these local explanations to capture interactions between features directly. Compared to traditionally used permutation feature importance, SHAP plots can display the magnitude, prevalence, and direction of a feature’s effect. We then used a SHAP-based Boruta selection method provided by the shap-hypetune package70 to select all relevant features contributing to smoking status prediction. By constructing randomly permuted shadow features, Boruta compares the mean absolute SHAP values between input features and shadow features and only keeps features if they perform better than the best randomized features. In our study, we performed 200 iterations of the algorithm, and the features within the tail 5% were rejected. The model with Boruta selected features was hyperparameter tuned again before further analysis.
Supporting information
Data Availability
UKB data are available through a procedure described at https://www. ukbiobank.ac.uk/enable-your-research.
Acknowledgements
All UK Biobank data was accessed under UK Biobank Application number 30418. This work was supported by the Centre of Artificial Intelligence in Precision Medicines (CAIPM), King Abdulaziz University, Jeddah, Saudi Arabia. We thank Nightingale Health Plc for providing early access to the UK Biobank dataset.
Funding and conflict of interests
The computational aspects of this research were supported by the Wellcome Trust Core Award Grant Number 203141/Z/16/Z and the Oxford NIHR BRC. The views expressed are those of the author(s) and not necessarily those of the NHS, the NIHR or the Department of Health. S.X and S.B.H and B.H are funded by CAIPM, King Abdulaziz University, Jeddah, Saudi Arabia. P.K is funded by the US National Institute on Aging (NIH). R.K.D is an inventor on a series of patents on use of metabolomics for the diagnosis and treatment of CNS diseases and holds equity in Metabolon Inc., Chymia LLC and Metabosensor. This project was enabled in part by the Alzheimer’s Gut Microbiome Project (AGMP), supported by the National Institute on Aging grants: 1U19AG063744 and 3U19AG063744-04S1, awarded to R.K.D at Duke University in partnership with multiple academic institutions. As such, the investigators within the AGMP not listed in this publication’s authors’ list, provided analysis-ready data, but did not participate in designing the study, conducting the analyses or writing of this manuscript. A listing of AGMP investigators can be found at https://alzheimergut.org/meet-the-team/. A complete listing of the AD Metabolomics Consortium (ADMC) investigators can be found at: https://sites.duke.edu/adnimetab/team/. In addition, this work was supported by the Alzheimer Disease Metabolomics Consortium which is a part of NIA’s national initiatives AMP-AD (3U01AG061359, 3U01 AG024904-09S4). Najaf Amin is funded by NIH and Oxford-GSK Institute of Molecular and Computational Medicine (IMCM). Cornelia M van Duijn is supported by the NIH, NovoNordisk, the IMCM, CAIPM of the University of Oxford and King Abdul Aziz University, Alzheimer Research UK (ARUK), UK National Institute for Health and Care Research (NIHR) Oxford Research Center (BRC), ZonMW (Delta Dementie) and Alzheimer Nederland. Cornelia M van Duijn is currently the Research Director Brain Health of the Health Data Research UK (HDR UK) and the UK Dementia Research Institute (UK DRI), working in partnership with Dementias Platform UK (DPUK). M Austin Argentieri was funded by NIH grant number 5U01AG061359-05.
Ethics approval
UK Biobank data use (Project Application Number 30418) was approved by the UK Biobank according to their established access procedures. UK Biobank has approval from the North West Multi-centre Research Ethics Committee (MREC) as a Research Tissue Bank (RTB), and as such researchers using UK Biobank data do not require separate ethical clearance and can operate under the RTB approval. Ethical approvals were granted and have been maintained by the relevant institutional ethical research committees in the UK.
Data Access Statement
UK Biobank data are available through a procedure described at: https://www.ukbiobank.ac.uk/enable-your-research.
List of Supplementary Materials
Supplementary Information
Figures 1 – 21
Tables 1 – 24
Supplementary File
This single Excel file contains Supplementary Tables 1 – 24 (each table is a sheet in the document).
Supplementary document has Supplementary figures.