Frailty-Related Metabolic Signature in the Association Between Frailty and Coronary Artery Disease: Modification by Genetic Susceptibility Authors: Junchi Liu, Hanze Du, Jieying Tang, Xiaohu Shi, Shuaihua Song, Rui Zhang, Lei Huang, Zhichao Lai, Daowei Li, Shi Chen, Hui Pan Published: 2026 ABSTRACT Graphical abstractKeywords: Frailty, Coronary artery disease, Metabolomics, Mediation analysis, Genetic predisposition to disease, Risk factor control Introduction Coronary artery disease (CAD) remains a leading contributor to global mortality and disability-adjusted life years [1]. Frailty, characterized by diminished physiological reserve and heightened vulnerability to stressors, is associated with adverse outcomes, including CAD, functional disability, and mortality [2].Given its potential modifiability, understanding the relationship between frailty and CAD may have implications for cardiovascular prevention and risk assessment [3]. However, the circulating metabolic features accompanying the association between frailty and CAD remain incompletely characterized. Metabolomics is a useful approach for characterizing circulating metabolic patterns associated with frailty and cardiovascular disease [4]. Previous epidemiological studies have identified circulating metabolites associated with both frailty and incident CAD [[5], [6], [7]]. Nevertheless, the extent to which frailty-related metabolic patterns statistically account for the association between frailty and CAD remains unclear. Genetic susceptibility is a well-established contributor to CAD risk. Previous studies have suggested that the association between frailty and CAD may differ according to genetic susceptibility. [8] Whether the association of frailty-related metabolic profiles with CAD similarly varies across levels of genetic susceptibility remains unclear. This study aimed to derive a frailty-related metabolic signature and evaluate its association with incident CAD. We further examined the extent to which this metabolic signature statistically accounted for the frailty–CAD association and whether the associations of frailty and the metabolic signature with CAD varied according to genetic susceptibility on both multiplicative and additive scales. Material and methods Study participants Data of individuals were obtained for analysis from the UK Biobank, which is an ongoing prospective study with follow-up data for approximately 502,128 participants who were recruited between 2006 and 2010 at 22 assessment centers. The UK Biobank study received approval from the North West Multicentre Research Ethical Committee, as well as informed consent from all participants. The details of the UK Biobank have been previously described [9]. After excluding participants with missing required covariate or follow-up data (n = 2,799), missing frailty data (n = 261,108), a prior diagnosis of CAD at baseline (n = 9,372), or unavailable genetic information (n = 4,942), 223,907 individuals were retained for the frailty-related analyses. For analyses examining the relationships among frailty, metabolomics, and CAD, an additional exclusion of participants without metabolomic data was applied, yielding a final sample of 122,959 individuals (Figure S1). The study was reported in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines. Assessment of frailty phenotype Frailty was assessed based on the five-component phenotypic model originally proposed by Fried et al. [10], which comprises weight loss, exhaustion, slow walking pace, low grip strength, and low physical activity levels. Detailed operational definitions of each component are provided in Table S1. Modifications to certain criteria were implemented in accordance with previously validated methodologies applied in large-scale cohort studies [11,12]. Participants were categorized into three groups: frailty (meeting ≥3 criteria), pre-frailty (meeting 1–2 criteria), or non-frailty (meeting no criteria), consistent with established cut-offs used in prior research [11]. Each frailty component was dichotomously scored, with 1 point indicating an unhealthy level and 0 points indicating a healthy level, resulting in a cumulative frailty score ranging from 0 to 5. In addition, given the heterogeneous associations of individual components with CAD, we constructed a weighted frailty score, in which each component was assigned a weight according to its strength of association with CAD [10]. Outcome ascertainment CAD was defined according to the International Classification of Diseases, Ninth Revision (ICD-9: codes 410–414) and Tenth Revision (ICD-10: codes I20–I25) [13]. Case ascertainment was supplemented using multiple data sources, including participant self-reports, primary care records, and death registry entries available within the UK Biobank. Incident CAD cases were identified during follow-up through the hospital inpatient database, with the first recorded diagnosis defining the incidence date. Definition of genetic susceptibility The polygenic risk score (PRS) for CAD was obtained from the UK Biobank Standard PRS release (Data-Field 26227), which provides a precomputed polygenic score based on external genome-wide association study (GWAS) training data [14]. Higher PRS values indicate greater genetic susceptibility to CAD. Participants were categorized into three genetic risk groups according to PRS quintiles: low (quintile 1, < −0.434), intermediate (quintiles 2–4, −0.434 to 0.602), and high (quintile 5, >0.602). Assessment of the circulating biomarkers The procedures for sample collection and metabolomic quantification have been comprehensively detailed in prior publications [15]. In brief, non-fasting ethylenediaminetetraacetic acid (EDTA)-plasma samples were obtained from a randomized subcohort during the initial assessment visit of the UK Biobank. All samples were transported under strict cryogenic conditions to the Nightingale Health laboratory (Finland) for high-throughput nuclear magnetic resonance (NMR) metabolomic profiling. The current analysis incorporated 169 directly quantified metabolites, as provided in the UK Biobank metabolomics dataset [16]. Metabolite values below the limit of detection were replaced with half of the observed minimum value, while other missing metabolite values were imputed using a random-forest approach. To better approximate normal distributions and facilitate comparability across metabolites, metabolite concentrations were subsequently log-transformed and standardized to z scores (mean = 0 and SD = 1) before statistical analyses. Covariates The covariates were assessed during the baseline visit (2006–2010). Age was calculated based on the participant's date of birth and the date of the baseline assessment. Sociodemographic variables included age, sex (male, female), ethnicity (white, other ethnicities), education level(A levels/AS levels or equivalent, College or University degree, CSEs or equivalent, O levels/GCSEs or equivalent, NVQ or HND or HNC or equivalent, Other professional qualifications e.g.: nursing, teaching and unknown), Townsend deprivation index (TDI), annual household income (<£18,000, £18,000–30,999, £31,000–51,999, £52,000–100,000, >£100,000, and unknown) and drinking status (never, previous, current, and unknown). For categorical variables with an “unknown” category, responses of “Prefer not to answer” and missing values were classified as unknown. Body mass index (BMI) was measured at baseline and calculated as weight (in kilograms) divided by the square of height (in meters). Diet scores (ranging from 0 to 6) were calculated by assigning 1 point for a healthy frequency and 0 points for an unhealthy frequency of consumption of specific food items, including fruits, vegetables, fish, unprocessed red meat, whole grains, and refined grains according to the results collected from the touchscreen questionnaire completed in the assessment center [17]. Higher scores indicated better diet quality. Sleep scores (ranging from 0 to 5) were derived by assigning 1 point for a healthy sleep behavior and 0 points for an unhealthy sleep behavior, based on factors such as sleep duration, morning/evening person, sleeplessness/insomnia, snoring and daytime dozing/sleeping according to the results collected from the touchscreen questionnaire completed in the assessment center. We further included the number of long-term chronic diseases as covariates (0, 1, ≥ 2) [11,18] which comprised 41 major chronic conditions, and detailed definition is presented in Table S2. Participants with missing data for diet score, TDI, or ethnicity were excluded. Missing values for smoking status, drinking status, and education were retained as an “unknown” category (0.3%, 2.6%, and 12.0%, respectively). Further information regarding these measurements is available in the UK Biobank (UKB) online protocol. Statistical analysis We constructed the frailty-related metabolic signature in three sequential steps. First, each of the 169 metabolites was regressed on the continuous frailty score using multivariable linear regression, with adjustment for all covariates described above. Multiple testing was controlled using the Benjamini–Hochberg false discovery rate (FDR), and metabolites with an FDR-adjusted P value <0.05 were retained as candidate frailty-associated metabolites. Second, to account for potential collinearity among metabolites, representative features were selected using elastic net regression. The dataset was randomly partitioned into training (70%) and testing (30%) subsets according to frailty status. [19,20] Elastic net regression, which combines Ridge and Lasso penalties, identified the most informative metabolites associated with frailty [21,22].The elastic-net mixing parameter α was prespecified at 0.5, and the penalty parameter λ was selected using 10-fold cross-validation according to the one-standard-error rule. Third, the metabolic signature for each participant was computed as a weighted sum of the selected metabolites, using the regression coefficients from the elastic net model as weights [19], and the resulting signature was subsequently standardized to a z score. Participants were categorized into low (quintile 1, < −0.81), intermediate (quintiles 2–4, −0.81 to 0.76), and high (quintile 5, ≥0.76) groups. Finally, Pearson correlation analysis was performed to assess interrelationships among the selected metabolites. Baseline characteristics were presented as mean (SD) for continuous variables or proportions for categorical variables by frailty phenotypes. Differences in characteristics were compared using Kruskal-Wallis test, or Chi-square tests accordingly. To assess potential selection bias, baseline characteristics were compared between participants included in the frailty-related analyses and those excluded using standardized mean differences (SMDs). Cox proportional hazards models were used to estimate hazard ratios (HRs) and 95% confidence interval (95% CI) for the associations between frailty and new-onset CAD in participants. The proportional hazards assumption was verified using Schoenfeld residuals and was met for all models. Two multivariate models were utilized in the analysis: model 1 included age and sex; and model 2 additionally included ethnicity, education level, income, Townsend deprivation index, BMI, smoking status, drinking status, diet score, sleep score, use of lipid-lowering medication and the number of long-term conditions. Mediation analysis Mediation analysis was performed using the ‘mediation’ R package to evaluate the extent to which the metabolic signature and its constituent metabolites statistically accounted for the association between frailty and incident CAD. Potential mediators were identified through a two-step approach. First, linear regression models were used to examine the associations between individual metabolites comprising the metabolic signature and frailty scores. Second, Cox proportional hazards models were applied to evaluate the relationships of the metabolic signature and the metabolites comprising it with incident CAD. Metabolites that showed significant associations with both frailty and CAD were included in the mediation analyses. All models were adjusted for the same covariates used in the multivariable Cox regression. In addition, we quantified the mediation proportions of each metabolite constituting the metabolic signature in the associations between individual frailty components and CAD risk. Interaction evaluations To evaluate whether the associations of frailty, individual frailty components, and the frailty-related metabolic signature with incident CAD varied according to genetic susceptibility, interactions with PRS were assessed on both multiplicative and additive scales. Multiplicative interaction was evaluated using likelihood ratio tests comparing Cox proportional hazards models with and without cross-product interaction terms between PRS and frailty status, individual frailty components, or metabolic signature levels. Additive interaction was quantified using the relative excess risk due to interaction (RERI) and its 95% CI. To further examine their joint associations with CAD risk, participants were cross-classified according to PRS categories (low, intermediate, and high) and frailty status (non-frailty, pre-frailty, and frailty) or metabolic signature levels (low, intermediate, and high). Participants with low PRS and non-frailty served as the reference group for the frailty–PRS analysis, whereas those with low PRS and a low metabolic signature served as the reference group for the metabolic signature–PRS analysis. All interaction analyses were adjusted for the same covariates as the fully adjusted Cox model. Sensitivity analyses Several sensitivity analyses were performed to evaluate the robustness of the primary findings. First, to mitigate potential reverse causation, we excluded participants who developed CAD within the first two years of follow-up. Second, in light of the potential confounding influence of pre-existing cardiovascular disease, we repeated the analyses after excluding individuals with any baseline cardiovascular conditions. Third, to address the differential contributions of individual frailty components to CAD risk, a weighted frailty score was constructed by assigning weights to each component based on its strength of association with CAD. Fourth, considering the substantial number of deaths during follow-up (n = 16,215), we conducted competing risk regression analyses to account for mortality as a competing event. All statistical analyses were performed based on R (version 4.4.1; R Foundation for Statistical Computing), with a two-sided P < 0.05 considered statistically significant. Results Baseline characteristics A total of 223,907 participants were included in the main cohort (Figure S1). Over a total follow-up period exceeding 3,302,628 person-years (median follow-up: 14.75 years), 7.8% of participants free of CAD at baseline developed CAD. The mean age of the study population was 56.58 years (SD, 8.05). Among the participants, 125,046 (55.8%) were women, while 1,901 (0.8%), 157,643 (70.4%), and 64,363 (28.7%) were classified as frailty, non-frailty and pre-frailty, respectively. Baseline characteristics stratified by frailty category are summarized in Table 1. Compared with non-frailty and pre-frailty individuals, those in the frailty group were more likely to be female and exhibited higher BMI, a poorer overall health rating, lower household income, poorer sleep quality, less favorable drinking patterns, and lower diet scores. Baseline characteristics of included and excluded participants are presented in Table S3. Differences were relatively small for most characteristics, with relatively greater differences observed in socioeconomic indicators, including education, income, and Townsend deprivation index. Table 1Baseline characteristics of participants by frailty phenotype.Table 1CharacteristicOverall N = 223,907anon N = 157,643apre N = 64,363afrailty N = 1,901ap-valuebSex<0.001 Female125,046 (56%)85,677 (54%)38,141 (59%)1,228 (65%) Male98,861 (44%)71,966 (46%)26,222 (41%)673 (35%)Age56.58 (8.05)56.44 (8.05)56.85 (8.07)58.75 (7.42)<0.001Ethnicity<0.001 Other ethnic groups7,582 (3.4%)4,256 (2.7%)3,165 (4.9%)161 (8.5%) White216,325 (97%)153,387 (97%)61,198 (95%)1,740 (92%)Townsend Deprivation Index<0.001 Q1 (Low)62,957 (28%)46,565 (30%)16,091 (25%)301 (16%) Q260,867 (27%)43,852 (28%)16,626 (26%)389 (20%) Q356,643 (25%)39,440 (25%)16,707 (26%)496 (26%) Q4 (High)43,440 (19%)27,786 (18%)14,939 (23%)715 (38%)Education<0.001 A Levels13,027 (5.8%)9,196 (5.8%)3,745 (5.8%)86 (4.5%) College/Degree86,016 (38%)64,077 (41%)21,608 (34%)331 (17%) CSEs6,870 (3.1%)4,394 (2.8%)2,378 (3.7%)98 (5.2%) GCSEs28,071 (13%)19,331 (12%)8,488 (13%)252 (13%) NVQ/HND/HNC34,252 (15%)23,899 (15%)10,082 (16%)271 (14%) Professional28,816 (13%)20,429 (13%)8,151 (13%)236 (12%) Unknown26,855 (12%)16,317 (10%)9,911 (15%)627 (33%)Overall health rating<0.001 Excellent41,187 (18%)32,765 (21%)8,394 (13%)28 (1.5%) Fair40,585 (18%)23,144 (15%)16,617 (26%)824 (43%) Good135,781 (61%)99,571 (63%)35,737 (56%)473 (25%) Poor5,815 (2.6%)1,880 (1.2%)3,377 (5.2%)558 (29%) Unknown539 (0.2%)283 (0.2%)238 (0.4%)18 (0.9%)Lipid-lowering medications29,933 (13%)18,597 (12%)10,806 (17%)530 (28%)<0.001Blood pressure-lowering medications18,580 (8.3%)10,935 (6.9%)7,239 (11%)406 (21%)<0.001Medications for diabetes38,682 (17%)23,970 (15%)14,036 (22%)676 (36%)<0.001Income<0.001 18,000 to 30,99949,215 (22%)34,111 (22%)14,716 (23%)388 (20%) 31,000 to 51,99954,382 (24%)39,547 (25%)14,578 (23%)257 (14%) 52,000 to 100,00045,365 (20%)34,299 (22%)10,941 (17%)125 (6.6%) Greater than 100,00012,114 (5.4%)9,482 (6.0%)2,607 (4.1%)25 (1.3%) Less than 18,00036,328 (16%)22,368 (14%)13,198 (21%)762 (40%) Unknown26,503 (12%)17,836 (11%)8,323 (13%)344 (18%)Frailty score0.36 (0.61)0.00 (0.00)1.15 (0.36)3.13 (0.35)<0.001Sleep score2.89 (1.08)2.97 (1.06)2.74 (1.09)2.19 (1.11)<0.001BMI27.01 (4.56)26.61 (4.20)27.91 (5.13)30.17 (6.58)<0.001Follow-up time14.30 (2.39)14.38 (2.28)14.14 (2.60)13.42 (3.56)<0.001Incident CAD17,551 (7.8%)11,051 (7.0%)6,171 (9.6%)329 (17%)<0.001Alcohol status<0.001 Current210,483 (94%)149,862 (95%)59,035 (92%)1,586 (83%) Never7,107 (3.2%)4,211 (2.7%)2,743 (4.3%)153 (8.0%) Previous6,207 (2.8%)3,516 (2.2%)2,535 (3.9%)156 (8.2%) Unknown110 (<0.1%)54 (<0.1%)50 (<0.1%)6 (0.3%)Smoking status<0.001 Current20,674 (9.2%)13,141 (8.3%)7,195 (11%)338 (18%) Never126,857 (57%)91,225 (58%)34,711 (54%)921 (48%) Previous75,799 (34%)52,923 (34%)22,245 (35%)631 (33%) Unknown577 (0.3%)354 (0.2%)212 (0.3%)11 (0.6%)Comorbidity group<0.001 089,530 (40%)68,496 (43%)20,817 (32%)217 (11%) 176,917 (34%)54,255 (34%)22,147 (34%)515 (27%) 2+57,460 (26%)34,892 (22%)21,399 (33%)1,169 (61%)Weight loss32,534 (15%)0 (0%)31,400 (49%)1,134 (60%)<0.001Exhaustion11,836 (5.3%)0 (0%)11,032 (17%)804 (42%)<0.001Physical activity11,479 (5.1%)0 (0%)10,166 (16%)1,313 (69%)<0.001Walking pace11,607 (5.2%)0 (0%)10,089 (16%)1,518 (80%)<0.001Handgrip12,512 (5.6%)0 (0%)11,335 (18%)1,177 (62%)<0.001Diet score3.1 (0.93)3.1 (0.93)3.1 (0.94)3.0 (0.94)<0.001Abbreviations: BMI, body mass index; CAD, coronary artery disease.an (%); Mean (SD).bPearson's Chi-squared test; Kruskal-Wallis rank sum test. Baseline characteristics of participants by frailty phenotype. Abbreviations: BMI, body mass index; CAD, coronary artery disease. n (%); Mean (SD). Pearson's Chi-squared test; Kruskal-Wallis rank sum test. Frailty-related individual metabolites and metabolic signature Baseline frailty scores were significantly associated with 162 of the 169 metabolites after FDR correction (Table S4), including 7 positive and 155 inverse associations. These 162 metabolites were subsequently entered as candidate features into the elastic net regression model, of which 36 had non-zero coefficients in the final model and were therefore retained in the frailty-related metabolic signature. The final signature included 14 metabolites with positive and 22 with negative coefficients (Fig. 1). The derived signature comprised metabolites spanning amino acids, cholesterol subclasses, fatty acids, phospholipids, total lipids, glycolysis-related metabolites, fluid-balance markers, and inflammatory markers. The correlation matrix of 36 metabolites showed strong correlations between esterified cholesterol and free cholesterol within the same lipoprotein subclasses (Figure S2). The positive coefficient of the metabolic signature was predominantly driven by monounsaturated fatty acids, glycoprotein acetyls (GlycA), and phospholipids in extra small very low-density lipoproteins (VLDL). Conversely, linoleic acid, cholesteryl esters in chylomicrons and extremely large VLDL, albumin, and omega-3 fatty acids contributed most strongly to the negative coefficient (Fig. 1). The metabolic signature demonstrated statistically significant correlation with frailty scores in both the training set (r = 0.197, P < 0.001) and the test set (r = 0.206, P < 0.001; Figure S3). Associations of the 36 metabolites constituting the metabolic signature with frailty components and frailty score. Presented from left to right are the metabolites’ coefficients (weights) in the signature and associations with frailty score, each frailty component. Abbreviations: PRS, polygenic risk score; 95% CI, 95% confidence interval; HR, hazard ratio; SGS, slow gait pace; LGS, low grip strength; LPA, low physical activity; HDL, high-density lipoproteins; LDL, low-density lipoproteins; VLDL, very low-density lipoproteins; XXL, extremely large; XL, very large; L, large; M, medium; S, small; XS, very small; L, total lipid; PL, phospholipid; C, cholesterol; CE, esterified cholesterol; FC, free cholesterol; TG, triglyceride; P, lipoprotein particle. Frailty, metabolic signature and incidence of CAD During a median follow-up of 14.75 years (interquartile range [IQR]: 14.1–15.5) in the main cohort (n = 223,907), 17,551 incident CAD cases were documented. Within the metabolomics sub-cohort (n = 122,959), 9,707 CAD cases occurred over a median follow-up of 14.8 years (IQR: 14.1–15.5). In the fully adjusted model, the HR (95% CI) of incident CAD were 1.25 (1.21–1.29) among participants with pre-frailty and 1.78 (1.59–1.99) among those with frailty, compared with the non-frailty group. Each one-point increase in the frailty score was associated with a 21% elevated risk of CAD (Table 2). The weighted frailty score showed a similar positive association with incident CAD (Table S5). Analyses of individual frailty components revealed the following HRs (95% CIs) compared with non-frail levels: 1.13 (1.09–1.18) for weight loss, 1.14 (1.08–1.21) for low physical activity, 1.30 (1.23–1.38) for low grip strength, 1.27 (1.20–1.36) for exhaustion, and 1.45 (1.38–1.52) for slow gait speed (Table S5). In multivariable-adjusted models, each SD increase in the metabolic signature was associated with an HR (95% CI) of 1.19 (1.17–1.22) for CAD (Table 2). Compared with participants with a low metabolic signature, those with a high metabolic signature exhibited an HR of 1.65 (95% CI: 1.54–1.78), while those with an intermediate signature had an HR of 1.23 (95% CI: 1.15–1.31) for CAD (Figure S4). Of the 36 metabolites constituting the signature, 12 metabolites positively associated with frailty were also associated with higher CAD risk, whereas 6 metabolites inversely associated with frailty showed inverse associations with CAD risk. Notable examples included the average diameter of high-density lipoprotein particles (HR = 0.86, 95% CI: 0.84–0.88), phospholipids in medium HDL (HR = 0.91, 95% CI: 0.89–0.93), triglycerides in large low-density lipoprotein (HR = 1.13, 95% CI: 1.11–1.15), and GlycA (HR = 1.15, 95% CI: 1.13–1.17) (Table S6). Table 2Associations of frailty and metabolic signature with risk of incident CAD.Table 2Pre-frailtyaFrailtyaFrailty scoreMetabolic signatureHR (95% CI)PHR (95% CI)PHR (95% CI)bPHR (95% CI)cPAge- and sex- adjusted model1.42 (1.38, 1.47)<0.0012.57 (2.31, 2.87)<0.0011.36 (1.33, 1.38)<0.0011.31 (1.29, 1.34)<0.001Multivariable-adjusted modeld1.25 (1.21, 1.29)<0.0011.78 (1.59, 1.99)<0.0011.21 (1.18, 1.24)<0.0011.19 (1.17, 1.22)<0.001Abbreviation: BMI, body mass index, 95% CI, 95% confidence interval; HR, hazard ratio; SD, standard deviation; CAD, coronary artery disease.aThe reference group was individuals with non-frailty.bHR and 95% CI of CAD risk per point increment in frailty scores.cHR and 95% CI of CAD risk per SD increment in metabolic signature.dBased on age- and sex-adjusted model, further adjusted for ethnicity, education level, income, Townsend deprivation index, BMI, smoking status, drinking status, diet score, sleep score, use of lipid-lowering medication and the number of long-term conditions. Associations of frailty and metabolic signature with risk of incident CAD. Abbreviation: BMI, body mass index, 95% CI, 95% confidence interval; HR, hazard ratio; SD, standard deviation; CAD, coronary artery disease. The reference group was individuals with non-frailty. HR and 95% CI of CAD risk per point increment in frailty scores. HR and 95% CI of CAD risk per SD increment in metabolic signature. Based on age- and sex-adjusted model, further adjusted for ethnicity, education level, income, Townsend deprivation index, BMI, smoking status, drinking status, diet score, sleep score, use of lipid-lowering medication and the number of long-term conditions. Mediation analyses The estimated mediation proportion for the metabolic signature in the association between frailty scores and incident CAD was 18.5% (95% CI 15.7%–21.7%; Table S7 and Figure S5). The corresponding mediation proportions for individual frailty components ranged from 10.5% to 29.3% for slow gait speed. Across individual metabolites, mediation estimates varied substantially. Albumin, creatinine, histidine, glucose, and glucose–lactate showed mediation proportions ranging from 0.0001 to 0.025, whereas GlycA, degree of fatty-acid unsaturation, monounsaturated fatty acids, phospholipids in medium HDL, and several lipid-related measures showed estimates ranging from 0.001 to 0.127. At the component level, weight loss showed a distinct metabolic pattern, being associated with lower GlycA levels and a higher degree of fatty-acid unsaturation, whereas the opposite directions were observed for most other frailty components (Table S4 and Fig. 1). Interaction with genetic susceptibility On the multiplicative scale, the association between frailty status and incident CAD differed across PRS strata (P for interaction <0.001), with a stronger relative association observed among participants with lower PRS (Figure S6). Among the individual frailty components, clear evidence of multiplicative interaction with PRS was observed for slow gait speed (P for interaction <0.001), whereas no clear evidence of multiplicative interaction was observed for the other components (Table S8). On the additive scale, a positive additive interaction was observed between pre-frailty and intermediate PRS (RERI = 0.16, 95% CI 0.03–0.28), whereas no clear evidence of additive interaction was observed for the remaining frailty–PRS combinations. The association between the frailty-related metabolic signature and incident CAD also varied across PRS strata on the multiplicative scale (P for interaction < 0.001; Fig. 2). On the additive scale, a positive additive interaction was observed among participants with intermediate PRS and high metabolic signature levels (RERI = 0.32, 95% CI 0.06–0.58), whereas no clear evidence of additive interaction was observed for the remaining PRS–metabolic signature strata. In the joint analysis, participants with both high PRS and high metabolic signature levels had the highest observed risk of incident CAD (HR = 3.48, 95% CI 2.95–4.11) compared with those with low PRS and low metabolic signature levels. Risk of incident CAD according to polygenic risk score and frailty-related metabolic signature. For the left forest plot, participants with low PRS and low metabolic signature levels were used as the reference group; for the right forest plot, participants with low metabolic signature levels within each PRS group were used as the reference group. Participants were categorized into three genetic risk groups according to PRS quintiles: low (quintile 1, <−0.434), intermediate (quintiles 2–4, −0.434 to 0.602), and high (quintile 5, >0.602). Similarly, the metabolic signature was categorized as low (quintile 1, <−0.81), intermediate (quintiles 2–4, −0.81 to 0.76), and high (quintile 5, ≥0.76). Results were adjusted for age, sex, ethnicity, education level, income, Townsend deprivation index, BMI, smoking status, drinking status, diet score, sleep score, use of lipid-lowering medication, and the number of long-term conditions. Abbreviations: BMI, body mass index; CAD, coronary artery disease; CI, confidence interval; HR, hazard ratio; MS, metabolic signature; PRS, polygenic risk score; RERI, relative excess risk due to interaction. Sensitivity analyses The results were consistent across sensitivity analyses when we excluded individuals who were identified as CAD within 2 years of follow-up; excluded participants with baseline cardiovascular diseases; and using competing risks models (Table S9). Discussion In this study, we identified a frailty-related metabolic signature for which the estimated mediation proportion in the association between frailty and incident CAD was approximately 18.5%. The identified metabolic features spanned several biomarker categories, including amino acids, fatty-acid measures, lipoprotein-related measures, and glycolysis-related metabolites. We further observed that the associations of frailty and the metabolic signature with incident CAD varied across PRS strata on the multiplicative scale. On the additive scale, positive interactions were observed between pre-frailty and intermediate PRS and among participants with intermediate PRS and high metabolic signature levels. In the joint analysis, participants with both high PRS and high metabolic signature levels had the highest observed risk of incident CAD. Taken together, these findings suggest that frailty-related metabolic alterations may help characterize the association between frailty and CAD, and that the associations of frailty and its metabolic signature with CAD may differ according to genetic susceptibility. Previous metabolomic studies have reported broad alterations in circulating metabolic profiles among individuals with frailty [6,23]. In a large UK Biobank study using NMR-based metabolomics, Mak et al. identified multiple biomarkers associated with frailty, including GlycA, creatinine, and several lipoprotein-related measures, with a substantial proportion of these associations replicated in independent cohorts [24]. Several of the metabolic features identified in our study were also consistent with these previous findings, particularly those involving GlycA and lipoprotein-related measures. Although the specific biomarkers identified varied across studies, the overall pattern supports a broad association between frailty and circulating metabolic profiles. The estimated mediation by the frailty-related metabolic signature may reflect several physiological processes relevant to both frailty and cardiovascular risk, as suggested by the characteristics of several individual mediating biomarkers. Several of the individual mediating biomarkers have plausible links to physiological features of frailty and cardiovascular risk. Lower albumin levels have been associated with poorer nutritional status, frailty, and sarcopenia [25,26]. Alterations in circulating amino acids, including histidine, have also been reported in relation to sarcopenic phenotypes and physical performance [27]. Creatinine is closely related to muscle mass but is also influenced by renal function; therefore, its association in the present study may reflect broader changes in body composition and systemic health rather than muscle status alone [28]. Together, these findings suggest that part of the metabolic pattern accompanying frailty may be related to nutritional and body-composition changes. GlycA, by contrast, is a marker of systemic inflammation and have been associated with both frailty and incident cardiovascular disease [24,29]. In addition, several fatty-acid, phospholipid, and lipoprotein-related measures contributed to the mediation estimates, consistent with previous evidence linking alterations in lipid-related metabolic profiles with both frailty and CAD [7,24,30]. Collectively, these findings suggest that the observed mediation may reflect several overlapping physiological and metabolic processes, including nutritional and body-composition changes, systemic inflammation, amino-acid metabolism, and lipid metabolism, rather than two discrete biological pathways. Component-specific analyses also revealed some heterogeneity in metabolic associations. In particular, weight loss was associated with lower GlycA levels and a higher degree of fatty-acid unsaturation, whereas the opposite directions were observed for most other frailty components. One possible explanation is that weight loss may represent a heterogeneous phenotype, as reductions in body weight can occur under different health, nutritional, and lifestyle circumstances [31,32]. Such heterogeneity may partly contribute to the distinct metabolic pattern observed for this component. Further studies with more detailed characterization of weight loss are needed to clarify these findings. The associations of selected fatty acids and amino acids further illustrate the complexity of the metabolic profile observed in frailty. Higher frailty scores were associated with lower circulating MUFA levels in our analyses, whereas higher MUFA levels were associated with greater CAD risk. Previous studies have also reported heterogeneous findings regarding circulating MUFAs in relation to frailty and cardiovascular outcomes [[33], [34], [35]]. These differences may partly reflect variation in the specific MUFA measures examined, because circulating MUFA levels reflect both dietary intake and endogenous synthesis, and individual MUFA species may have different metabolic associations [36]. Similarly, higher frailty scores were associated with lower circulating BCAA concentrations, whereas higher BCAA levels were associated with greater CAD risk. Circulating BCAA levels are influenced by dietary intake, protein turnover, and tissue-specific catabolism. Lower BCAA concentrations have been associated with reduced muscle mass and strength in older adults [37], whereas elevated circulating BCAA levels may reflect disturbances in BCAA metabolism and have also been associated with insulin resistance and cardiovascular disease [38,39]. These contrasting associations may therefore reflect different physiological processes across frailty and cardiometabolic disease states. These findings suggest that the associations of individual metabolites with frailty and CAD may depend on the broader metabolic and clinical context rather than representing a simple unidirectional pathway. The associations of frailty and its metabolic signature with CAD also varied according to genetic susceptibility. Consistent with previous studies showing that the relative associations of non-genetic risk factors with cardiovascular outcomes may differ across levels of genetic risk [40], the association between frailty and CAD differed across PRS strata and was generally stronger at lower levels of PRS on the multiplicative scale. Among individual frailty components, clear evidence of multiplicative interaction was observed only for slow gait speed. On the additive scale, positive interaction was observed between pre-frailty and intermediate PRS and between high metabolic signature levels and intermediate PRS, whereas no clear evidence of additive interaction was observed for the remaining strata. The differences observed across interaction analyses suggest that the modifying role of genetic susceptibility may vary across risk profiles rather than consistently strengthening the effects of frailty or its metabolic signature. Although participants with both high PRS and high metabolic signature levels had the highest observed CAD risk, this finding should therefore be interpreted as a high joint-risk profile rather than evidence of a simple synergistic effect between genetic and metabolic risk. Overall, our findings suggest that frailty-related metabolic alterations provide additional information for characterizing CAD risk and that their associations with CAD vary according to genetic susceptibility. These findings highlight the potential value of considering functional, metabolic, and genetic information together when characterizing cardiovascular risk in later life. To our knowledge, few large-scale prospective studies have jointly evaluated frailty, circulating metabolomic profiles, genetic susceptibility, and incident CAD. The large sample size and extended follow-up provided substantial statistical power for the assessment of incident CAD, while broad metabolomic profiling enabled the construction of a frailty-related metabolic signature from a wide range of circulating biomarkers rather than a limited set of prespecified biomarkers. A further strength was the integration of mediation analysis with both multiplicative and additive interaction analyses. Similar findings across alternative frailty measures and sensitivity analyses indicated that the main results were generally consistent. Several limitations, however, should be acknowledged. First, frailty was assessed using an adapted Fried phenotype based on measurements available in the UK Biobank rather than the original measurement procedures. In addition, several components were questionnaire-based, which may have affected the precision of frailty assessment. Second, a substantial proportion of the original UK Biobank cohort was excluded from the current analysis. Although differences were relatively small for many baseline characteristics, more noticeable differences were observed in several socioeconomic and demographic characteristics between included and excluded participants. This may have affected the composition of the analytic cohort and should be considered when interpreting the applicability of our findings to broader populations. Third, frailty and metabolite concentrations were both assessed at baseline, precluding the establishment of their temporal sequence and the evaluation of changes in metabolic profiles during follow-up. Therefore, causal mediation cannot be inferred from the present findings. Fourth, residual confounding related to participants’ health status cannot be fully excluded. Disease burden, systemic inflammation, and medical treatments may influence frailty, metabolic profiles, and CAD risk but may not be fully captured at baseline. Moreover, health status and treatment may change during follow-up, and new diseases may develop, potentially influencing the observed associations and metabolic mediation effects. Fifth, reverse causality and the influence of subclinical disease cannot be completely excluded given the observational design, although participants who developed CAD within the first two years of follow-up were excluded in sensitivity analyses. Sixth, the biological interpretation of individual metabolites was primarily informed by previous literature, as formal pathway and tissue-specific analyses were not performed. Therefore, the metabolic patterns discussed here should be interpreted as suggestive rather than definitive. Similarly, the observed PRS interactions indicate statistical effect modification and should not be taken as direct evidence of underlying biological interaction. Conclusions Collectively, we identified a frailty-related metabolic signature associated with incident CAD and found that the associations of frailty and its metabolic signature with CAD varied according to genetic susceptibility. These findings highlight the complementary contributions of functional, metabolic, and genetic information to the characterization of CAD risk and support their integration for a broader assessment of cardiovascular risk in older adults. Author contributions JL and HD: Conceptualization, Methodology, Data curation, Formal analysis, Visualization, Writing – original draft. JT, XS, SS, RZ, LH, ZL, DL: Writing – review & editing. SC: Writing – review & editing and Methodology. HP: Writing – review & editing and Funding acquisition. All authors read and approved the final manuscript and agree to be accountable for all aspects of the work. Ethical approval The UK Biobank study was approved by the National Health Services (NHS) National Research Ethics Service (Ref 11/NW/0382). Written informed consent was obtained from all participants enrolled in the study. All data used in this study were obtained from the UK Biobank (http://www.ukbiobank.ac.uk) through application 100787. Declaration of Generative AI and AI-assisted technologies in the writing process During the preparation of this work, the authors used generative AI assisted technologies to improve language and readability. After using these tools, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication. The authors did not use generative AI or AI-assisted technologies in the writing of this manuscript or in the preparation of figures, images, or artwork. Funding This work was supported by the National Key Research and Development Program of China (2022YFC3600203) and CAMS Innovation Fund for Medical Sciences (CIFMS), Grant No. 2023-I2M-C&T-B-017. Data availability The data that support the findings of this study are available from the UK Biobank project site, subject to registration and application process. Further details can be found at https://www.ukbiobank.ac.uk. The study has been conducted using the UK Biobank Resource under application number 100787. Code availability The R codes for the current analyses are available from the corresponding author upon request. Declaration of competing interest The authors declare no competing interests. Acknowledgements The authors are grateful to the UK Biobank manager team for their significant contribution to this project, as well as the participants in the UK Biobank for their active participation and cooperation. Footnotes Supplementary material related to this article can be found, in the online version, at doi:https://doi.org/10.1016/j.jnha.2026.100998. Contributor Information Daowei Li, Email: jluldw@jlu.edu.cn. Shi Chen, Email: cspumch@163.com. Hui Pan, Email: panhui@pumch.cn. Supplementary data The following is Supplementary data to this article: References Associated Data Supplementary Materials Data Availability Statement The data that support the findings of this study are available from the UK Biobank project site, subject to registration and application process. Further details can be found at https://www.ukbiobank.ac.uk. The study has been conducted using the UK Biobank Resource under application number 100787.