Liver-specific phenotypic aging, behavior and genetic risks, and long-term liver-related outcomes

Abstract

Phenotypic age, an aging indicator derived from clinical biomarkers, is associated with morbidities and mortality. However, a liver-specific phenotypic aging indicator is still lacking, and its longitudinal associations with liver-related outcomes, as well as the underlying biological mechanisms, remain elusive. We developed a liver-specific phenotypic age using 11 selected clinical blood markers within the England-White cohort of the UK Biobank and validated this metric in both the Scotland-Wales cohort and Non-White-British cohort. We calculated phenotypic age acceleration (PhenoAgeAccel) and examined its association with long-term liver-related outcomes. We also explored the extent to which liver-specific PhenoAgeAccel mediated the impact of modifiable risk behaviors on liver-related outcomes. The metabolic and proteomic signatures of liver-specific PhenoAgeAccel were subsequently characterized. Liver-specific PhenoAgeAccel was significantly associated with a 1.23- to 2.97-fold increased risks of all-cause mortality and liver-related events. The impact of liver-specific PhenoAgeAccel on liver outcomes were more pronounced in males and in individuals with high genetic risk compared to their respective counterparts, and was stronger than that observed with systemic PhenoAgeAccel. Approximately 10–27% of the associations between risk behaviors and liver-related outcomes were mediated by liver-specific PhenoAgeAccel. Proteomic analysis identified 211 proteins associated with both liver-specific PhenoAgeAccel and liver-related outcomes, of which 22 (e.g., AGXT and SULT2A1) were liver-enriched and significantly mediated this relationship. Liver-specific PhenoAgeAccel is a strong predictor of liver-related outcomes, partially mediates the impact of modifiable behaviors, and is linked to liver-enriched proteins. This accessible tool may enhance risk stratification and support preventive strategies targeting liver health and aging.

Graphical Abstract

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1007/s11357-026-02103-3.

Introduction

Chronic liver diseases (CLD) impose a growing global health challenge, with aging recognized as the most important contributor to adverse hepatic events [1–3]. Although chronological age is a well-established risk factor for morbidity and mortality, individuals of the same age often show wide variability in disease susceptibility and longevity [4, 5]. This discrepancy has fueled interest in biological age (BA)—a more personalized indicator of physiological aging—measured through molecular and clinical biomarkers [4, 5]. Among various aging clocks, phenotypic age (PhenoAge), derived from routine clinical blood tests and chronological age, offers a cost-effective and reliable estimate of BA and is strongly predictive of morbidity and mortality across diverse populations [5–9]. However, mounting evidence highlights that aging occurs in an organ-specific manner, and systemic BA may overlook the heterogeneous aging patterns of individual organs [10–12]. This limitation may be particularly relevant for the liver, given its central roles in metabolism, detoxification, and immune regulation [13].

Recent studies have demonstrated the advantages of organ-specific BA [10, 11, 14]. For instance, hepatic BA more accurately predicts the severity of non-alcoholic fatty liver disease (NAFLD) than systemic BA but is less predictive of cardiovascular mortality compared to cardiac aging metrics [10]. Nonetheless, most existing organ-specific aging clocks rely on costly multi-omics platforms, limiting their feasibility for large-scale or clinical application [4–6, 15]. There remains a pressing need for an accessible, liver-specific aging index based on routine biomarkers. Additionally, risk behaviors such as smoking, alcohol drinking, physical inactivity, and unhealthy diet have been shown to the accelerate aging process [16–18] and markedly increase the risk of liver-related outcomes [19–21]. We hypothesize that these behaviors may contribute to liver disease through accelerated liver-specific aging. To test this, we aim to: (1) develop a liver-specific phenotypic age and its acceleration (PhenoAgeAccel) using routine blood biomarkers—an organ-targeted adaptation of the PhenoAge framework in which liver-related risk is modeled and transformed onto an age-like scale; (2) examine its role in mediating the effects of lifestyle-related risk behaviors on liver-related outcomes; and (3) elucidate supportive molecular associations through high-throughput plasma proteomic and metabolomic profiling.

Methods

Study design and participants

This study was conducted within the UK Biobank (UKBB), a large-scale, population-based prospective cohort comprising over 500,000 individuals aged 40–69 years, recruited between 2006 and 2010 from 22 assessment centers across England, Scotland, and Wales [22]. All participants provided written informed consent, and the UKBB received ethical approval from the North West Multicentre Research Ethics Committee. Detailed study protocols are available at www.ukbiobank.ac.uk.

We excluded individuals with baseline liver diseases, cancer, or missing blood biomarker data, yielding a final analytic sample of 449,427 participants. Based on geographic recruitment and self-reported ethnicity, participants were stratified into three mutually exclusive cohorts: the England-White cohort, the Scotland & Wales cohort, and the Non-White-British cohort (Fig. 1). For omics analyses, subsets of participants with complete data were included: 223,441 for plasma metabolomics and 47,575 for proteomic profiling.

Fig. 1.

Flowchart of the study

Development of liver-specific phenotypic age

We first computed systemic phenotypic age using the established algorithm proposed by Levine et al. [5]. To construct a liver-specific phenotypic age, we adapted this method with modifications tailored to predict liver-related mortality (Supplementary Methods). Briefly, clinical biomarkers with > 25% missingness were excluded (Figure S1). Among the remaining biomarkers, we removed those with high collinearity (Pearson correlation coefficient > 0.75) and low variability (Figure S2, Table S1). For the primary analyses, covariates were singly imputed (continuous by cohort mean; categorical by modal category) to favor simplicity, interpretability, and computational convenience under sparse missingness. To ensure robustness, we performed multiple imputation by chained equations (MICE) (m = 5 imputations) using predictive mean matching for continuous variables and appropriate logistic/multinomial models for categorical variables, including the exposure, outcomes, and key covariates (e.g., smoking status, alcohol intake, physical activity, education level, and medication use) in the imputation models with passive imputation for derived variables. Estimates across imputations were combined with Rubin’s rules.

Chronological age and 29 routine clinical biomarkers were then included in a L1-penalized Cox regression (LASSO-Cox) model to predict liver-related mortality in the England-White cohort. Through tenfold cross-validation, chronological age and 11 biomarkers were selected for model construction (Figure S3, Table S2). To confirm stability, we conducted 10 random resampling replicates of the derivation step, each returning the same 11 biomarkers. These biomarkers reflect core hepatic processes—such as inflammation, metabolism, and synthetic function. We subsequently fitted a Cox model for liver-related mortality in the England-White cohort. For each participant, we calculated the Cox linear predictor, transformed it into a bounded liver-specific risk score, and then linearly rescaled this score to a 20–100 “liver age” scale to derive individual liver-specific phenotypic age. This “liver age” represents a relative biological index calibrated to the risk distribution in our cohort rather than a direct measure of chronological years. Liver-specific phenotypic age acceleration (PhenoAgeAccel) was defined as the residual from a linear regression of liver-specific phenotypic age on chronological age, thereby capturing inter-individual variation in liver aging that is independent of chronological age. Participants with PhenoAgeAccel > 0 were classified as biologically older than their chronological age, while those with PhenoAgeAccel ≤ 0 were considered biologically younger [6].

Construction of behavior and polygenic risk scores for liver health

Based on a recent recommendation [21, 23], we constructed a composite risk behavior score by integrating established risk behaviors: smoking, alcohol intake, physical inactivity, and unhealthy diet, which were assessed at baseline in the UKBB through a touchscreen questionnaire [22]. Table S3 provides comprehensive details and scoring algorithms for each behavioral component, with each of the four components scored on a scale from 0 to 3. The overall health behavior score was calculated by summing these individual scores [24, 25]. To assess the potential interaction effect between liver-specific aging acceleration and genetic risk of liver diseases, we calculated a polygenic risk score (PRS) for liver-related outcomes based on the largest genome-wide association study (GWAS) for liver cirrhosis (Table S4). This PRS has shown robust predictive power for a broad spectrum of liver diseases and liver-related mortality [26]. Genetic risk was categorized into low (the bottom quintile), intermediate (quintiles 2–4) and high (the top quintile) groups according to the distributions of PRS [6, 26].

Plasma metabolomic and proteomic measurements

A total of 251 plasma metabolites were quantified among approximately 280,000 UKBB participants through high-throughput nuclear magnetic resonance (NMR) spectroscopy using the Nightingale Health platform [27]. Two biomarkers (glucose-lactate and spectrometer-corrected alanine) with technical artifacts were excluded, retaining 249 standardized measures for further analysis. All metabolite concentrations underwent natural log-transformation with subsequent Z-score normalization for analysis [28]. Plasma proteins were quantified for baseline samples (2007–2010) using the Olink Explore Proximity Extension Assay in ~ 53,000 UKBB participants. Detailed protocols for sample processing and quality control have been previously detailed [29]. Of the 2,923 detected proteins obtained from the existing UKBB dataset, we excluded 12 with > 20% measurements below the lower limit of detection (LOD) for this study. For the remaining 2911 proteins, values below LOD were imputed using k-nearest neighbors (k = 10). Prior to analysis, all proteins were standardized to Z-scores by scaling Olink Normalized Protein eXpression (NPX) values to mean 0, standard deviation 1 [30].

Outcomes and covariates

Primary outcomes included incident liver cancer, hepatic failure, and liver fibrosis/cirrhosis, as well as liver-related and all-cause mortality. All outcomes were ascertained using ICD-10 codes (Table S5). Meanwhile, we defined a composite indicator called fatal liver-related events (including decompensated cirrhosis, HCC, liver transplantation and liver-related death, whichever occurred first) [31]. We defined beginning of research in the Cox models as the date of enrollment into the UKBB. For participants who experienced a liver-related outcome during follow-up, the event time was calculated as the duration from this baseline date to the first occurrence of the outcome. Covariates for this study included chronological age at recruitment, sex, body mass index (BMI), Townsend deprivation index, education level, smoking status, alcohol intake frequency, physical activity, and medication use (Details are shown in Supplementary Methods).

Statistical analyses

We used mean (standard deviation, SD) for continuous variables and counts (percentages) for categorical variables to describe the baseline characteristics of the participants. Analysis of variance (ANOVA) and χ² tests, where appropriate, were used to compare the differences between England-White cohort, Scotland & Wales cohort, and Non-White-British cohort. We employed CPH models in three sub-cohorts to separately evaluate the associations of liver-specific PhenoAgeAccel, genetic risk, and risk behavior scores with long-term liver-related outcomes. Multivariable-adjusted hazard ratios (HR) and 95% confidence intervals (CI) were calculated to quantify the associations. Restricted cubic spline (RCS) analysis was used to assess potential nonlinear associations of the liver-specific PhenoAgeAccel with liver-related outcomes, as well as the relationship between risk behavior scores and liver-specific PhenoAgeAccel [6].

We conducted a mediation analysis to quantify the proportion of association between risk behavior and liver-related outcomes mediated by liver-specific PhenoAgeAccel. Because liver-related cases were limited in the Scotland & Wales and Non-White British cohorts, we combined these two validation cohorts and adjusted for race in the mediation analysis. Mediation was evaluated under a counterfactual framework using the “%mediated” SAS macro to estimate the natural indirect effect (NIE), natural direct effect (NDE), and total effect (TE), with bootstrap resampling to obtain 95% CIs. The NIE quantifies the portion of the exposure–outcome association operating through the mediator, whereas the NDE reflects the portion not through the mediator. In addition, we conducted Imai-type sensitivity analyses with the CMAverse R package to assess robustness to unmeasured mediator–outcome confounding across multiple model specifications [32].

We conducted a sensitivity analysis by excluding events occurring within the first year of follow-up to assess the robustness of associations between liver-specific PhenoAgeAccel and liver-related outcomes. To address potential circularity from liver enzyme biomarkers, we conducted a sensitivity analysis in which liver-specific phenotypic age and PhenoAgeAccel were recalculated after excluding enzyme indicators (e.g., AST, GGT, AST/ALT) prior to model construction. We conducted subgroup analyses across strata of age, sex, obesity, smoking, alcohol intake, education, medication use, and genetic risk, and tested for interactions. We also assessed the joint effects of liver-specific PhenoAgeAccel and genetic risk, using biologically younger individuals with the lowest genetic susceptibility as the reference. Owing to limited case counts in some strata of the Scotland & Wales and Non-White British cohorts, these analyses were restricted to the England–White cohort.

We computed the aMAP score and three noninvasive fibrosis indices (FIB-4, APRI, and Forns) using published formulas [23]. Briefly, aMAP incorporates age, sex, serum albumin, and platelet count; FIB-4 uses age, AST, ALT, and platelet count; APRI is based on AST/upper limit of normal and platelet count; and the Forns index uses age, γ-glutamyltransferase (GGT), platelet count, and cholesterol. Predictive accuracy was assessed with Harrell’s C-index and the area under the ROC curve (AUROC), with DeLong’s test used to compare AUROCs. In comparative analyses, we also computed net reclassification improvement (NRI) and integrated discrimination improvement (IDI) for liver-specific PhenoAgeAccel versus each established score, and generated calibration and decision-curve analyses to evaluate calibration and net clinical benefit across risk thresholds.

To explore the metabolic and proteomic signatures underlying liver-specific phenotypic aging and liver-related outcomes, we first used logistic regression to assess associations between individual metabolites/proteins and liver-specific PhenoAgeAccel, with multiple comparisons corrected by the false discovery rate (FDR) method. Subsequently, CPH model was utilized to assess associations between the significant metabolites/proteins and long-term liver-related outcomes. For proteins significantly associated with both liver-specific PhenoAgeAccel and all liver-related outcomes, tissue enrichment analysis was conducted via the TissueEnrich R package, prioritizing liver-specific enrichment patterns. Pathway enrichment analyses were subsequently performed on liver-enriched proteins using the g:Profiler platform [30]. In sex-stratified analyses, we repeated this proteomic pipeline separately in male and female participants, applying the same FDR-controlled screening and outcome modeling procedures within each sex. Because only one protein in women (integrin beta-like protein 1) showed concurrent associations with liver-specific PhenoAgeAccel and all liver-related outcomes, we therefore used fatal liver-related events as the representative endpoint for female-specific proteomic analyses. Protein–protein interaction networks were further interrogated using the STRING database (v12.0; https://string-db.org) to characterize functional connectivity among liver-enriched proteins [33]. Finally, an odds ratio-weighted inverse probability approach was applied to quantify mediation effects of prioritized proteins on associations between liver-specific PhenoAgeAccel and liver-related endpoints. This method is particularly well-suited for causal mediation analysis in complex models with multiple mediators [34].

All analyses were performed using SAS statistical software, version 9.4 (SAS Institute, Cary, NC), and R statistical software, version 4.4.1 (R Project for Statistical Computing).

Results

Characteristics of study participants

In this study, we included 369,012 individuals (mean age: 56.5 years; 53.3% female) in the England-White cohort, 52,433 (mean age: 56.0 years; 54.2% female) in the Scotland and Wales cohort, and 27,982 (mean age: 52.7 years; 52.9% female) in the Non-White-British cohort. During a median follow-up period of 13.7 years, 832, 170, and 36 liver-related deaths occurred in these cohorts, respectively (Table 1). Within the 5-year and 10-year follow-up, the numbers of liver-related deaths were 164, 17, and 9 at 5 years and 488, 76, and 25 at 10 years for the three cohorts, respectively. Baseline characteristics varied significantly across the three cohorts. We also observed notable differences between systemic PhenoAgeAccel and liver-specific PhenoAgeAccel within each cohort (Figure S4A), while liver-specific PhenoAgeAccel exhibited a similar distribution across the three cohorts (Figure S4B).

Table 1.

Baseline characteristics of participants in the UK Biobank

Characteristics a	England-White cohort	Scotland & Wales cohort	Non-White-British cohort	P b
Participants (No.)	369,012	52,433	27,982
All-cause death	26,801 (7.26)	3550 (6.78)	1462 (5.23)	< 0.001
Liver-related death	832 (0.23)	170 (0.32)	36 (0.13)	< 0.001
5 years	164 (0.04)	17 (0.03)	9 (0.03)	0.318
10 years	488 (0.13)	76 (0.15)	25 (0.09)	0.104
15 years	832 (0.23)	168 (0.32)	36 (0.13)	< 0.001
Hepatic failure	798 (0.22)	95 (0.18)	69 (0.25)	0.127
Fibrosis and cirrhosis of liver	1529 (0.41)	225 (0.43)	107 (0.38)	0.616
HCC	387 (0.10)	65 (0.12)	17 (0.06)	0.029
Fatal liver-related events	1152 (0.31)	224 (0.43)	65 (0.23)	< 0.001
PhenoAgeAccel, years	−3.74 (4.97)	−3.42 (5.14)	−2.93 (5.68)	< 0.001
Liver PhenoAgeAccel, years	3.24 (13.36)	3.71 (13.39)	6.99 (13.66)	< 0.001
Chronological age, years	56.52 (8.05)	56.00 (8.01)	52.73 (8.24)	< 0.001
Female	196,625 (53.28)	28,434 (54.23)	14,806 (52.91)	< 0.001
Townsend deprivation index	−1.47 (2.96)	−1.33 (3.23)	1.09 (3.58)	< 0.001
Alcohol intake frequency				< 0.001
Never or special occasions only	62,900 (17.05)	9500 (18.12)	14,362 (51.33)
Twice per week and less	137,952 (37.38)	20,787 (39.64)	8465 (30.25)
Three times or more per week	168,160 (45.57)	22,146 (42.24)	5155 (18.42)
BMI, kg/m2	27.37 (4.74)	27.61 (4.87)	27.82 (5.03)	< 0.001
Degree level or professional education	136,040 (36.87)	20,403 (38.91)	12,160 (43.46)	< 0.001
Smoking status				< 0.001
Never	200,876 (54.44)	29,206 (55.70)	19,340 (69.12)
Former	130,523 (35.37)	16,972 (32.37)	5319 (19.01)
Current	37,613 (10.19)	6255 (11.93)	3323 (11.88)
Physical activity				< 0.001
0–1 day per week	72,508 (19.65)	11,246 (21.45)	5219 (18.65)
2–4 days per week	156,942 (42.53)	22,691 (43.28)	13,108 (46.84)
5–7 days per week	139,562 (37.82)	18,496 (35.28)	9655 (34.50)
Medication use (%)	55,457 (15.03)	7164 (13.66)	4510 (16.12)	< 0.001
Blood biochemical markers
Albumin, g/L	45.27 (2.41)	45.10 (2.34)	44.94 (2.53)	< 0.001
Aspartate aminotransferase, U/L	25.99 (10.07)	26.10 (10.45)	26.34 (10.78)	< 0.001
Creatinine, μmol/L	72.14 (13.83)	71.31 (13.60)	72.64 (15.31)	< 0.001
C-reactive protein, mg/dL	2.41 (3.22)	2.59 (3.37)	2.57 (3.35)	< 0.001
Gamma glutamyltransferase, U/L	36.57 (39.22)	38.10 (41.74)	37.75 (38.16)	< 0.001
Glucose, mmol/L	5.09 (0.90)	5.02 (0.92)	5.20 (1.16)	< 0.001
Glycosylated hemoglobin, mmol/mol	35.90 (6.33)	35.96 (6.48)	38.83 (9.40)	< 0.001
LDL cholesterol, mmol/L	3.56 (0.85)	3.58 (0.84)	3.38 (0.83)	< 0.001
Red blood cell count × 1000 cells/μL	4.52 (0.40)	4.48 (0.39)	4.64 (0.50)	< 0.001
Mean cell volume, fL	91.27 (4.11)	91.86 (4.15)	88.57 (5.20)	< 0.001
Aspartate aminotransferase to alanine aminotransferase ratio	1.26 (0.43)	1.21 (0.37)	1.28 (0.46)	< 0.001
Alkaline phosphatase, U/L	82.62 (21.84)	83.40 (21.84)	84.43 (22.70)	< 0.001
Lymphocyte percentage	28.71 (6.91)	28.59 (6.74)	32.32 (7.90)	< 0.001
Red blood cell distribution width, %	13.46 (0.83)	13.50 (0.85)	13.80 (1.08)	< 0.001
White blood cell count × 1000 cells/μL	6.88 (1.67)	6.88 (1.67)	6.74 (1.77)	< 0.001

PhenoAgeAccel phenotypic age acceleration, BMI body mass index, LDL Low Density Lipoprotein, HCC hepatocellular carcinoma. Data are presented as n, mean ± SD or n (%)

^a Multiple imputation method was used to impute the missing covariates. Continuous variables were presented as mean (standard deviation) or median (interquartile range), categorical variables were presented as n (%)

^b Continuous variables were compared by analysis of variance and categorical variables were compared by the chi-square test

Liver-specific phenotypic aging and risk of liver-related outcomes

Liver-specific PhenoAgeAccel, developed from chronological age and 11 blood biochemical indicators (albumin, AST, creatinine, C‑reactive protein, GGT, glucose, glycosylated hemoglobin, LDL cholesterol, red blood cell count, mean cell volume and AST to ALT ratio), was used to evaluate the individual liver phenotypic aging. As expected, in the England-White cohort, higher liver-specific PhenoAgeAccel was associated with increased risks of all-cause mortality, liver-related mortality, and liver-related events (Table S6). These associations were consistent across the Scotland & Wales and Non-White-British cohorts and were stronger than those observed for systemic PhenoAgeAccel, except for all-cause mortality (Tables S6-7). Across all sensitivity and supplementary analyses—including MICE-based multiple imputation and other prespecified checks—effect estimates and discrimination closely matched the primary results, supporting the robustness of our findings (Tables S8–9). Notably, when the index was re-derived after excluding liver enzyme biomarkers (AST, GGT, and the AST/ALT ratio), liver-specific PhenoAgeAccel remained significantly associated with all liver-related outcomes, with effect estimates comparable to the primary model (Table S10). Furthermore, cause‑specific Cox and Fine–Gray competing‑risk models treating non–liver‑related death as a competing event produced effect estimates very similar to those of the primary Cox models (Table S11). The associations between liver-specific PhenoAgeAccel and liver-related outcomes also remained significant across all subpopulations (Table S12). Notably, for outcomes such as liver-related death and fatal liver-related events, the associations were stronger in males and individuals with high genetic risk (P for interaction < 0.05).

RCS analysis showed significant nonlinear associations between liver-specific PhenoAgeAccel and long-term liver-related outcomes (all P for nonlinearity < 0.001). However, these associations showed an obvious dose–response pattern, with the multivariable adjusted HR ranging from 1.02 (95% CI 1.01–1.02) for all-cause mortality to 1.09 (1.08–1.09) for liver-related mortality per 1-year increment of liver-specific PhenoAgeAccel (Fig. 2). We observed similar nonlinear associations between systemic PhenoAgeAccel and liver-related outcomes (Figure S5).

Fig. 2.

Linear and nonlinear association of liver-specific phenotypic age acceleration (PhenoAgeAccel) with the risk of liver-related outcomes in the England-White cohort. The HR was for risk of liver-related outcomes per one year increment of liver-specific PhenoAgeAccel

Additionally, liver-specific PhenoAgeAccel showed superior performance for predicting most liver-related events compared with the aMAP score and three noninvasive fibrosis tests including FIB-4, APRI, and Forns index. For instance, the liver-specific PhenoAgeAccel demonstrated superior performance in predicting liver-related death, with an AUROC of 0.80 (95% CI: 0.77–0.84) and C-index of 0.798 (SE = 0.018), which significantly outperformed FIB-4 [0.66 (95% CI 0.64–0.68)], APRI [0.65 (95% CI 0.63–0.68)], and aMAP score [0.73 (95% CI 0.69–0.77)] (Figure S6, Table S13). However, its performance was comparable to other indices in predicting hepatic failure and fibrosis and cirrhosis of liver (P for comparison > 0.05) (Table S14). Consistently, models incorporating liver-specific PhenoAgeAccel achieved substantial improvements in NRI (ranging from 0.35 to 0.60) and IDI (ranging from 0.05 to 0.23) over established scores for liver-related outcomes (Table S15). Decision-curve analyses showed higher net benefit for liver-specific PhenoAgeAccel across clinically relevant thresholds (Figure S7), though calibration curves indicated broadly comparable calibration to existing scores (Figure S8).

Association of risk behavior score and PRS with liver-related outcomes

The risk behavior score was significantly associated with increased risks of all-cause mortality and liver-related outcomes, showing a nonlinear pattern: only higher scores were significantly linked to liver-related outcomes, with no association in the lower range (Figure S9). Similar association pattern was found between risk behavior score and liver-specific PhenoAgeAccel (Figure S10). Additionally, individuals with high genetic risk exhibited a significantly elevated risk of liver-related outcomes in a dose–response manner (Table S16).

Joint effects of liver-specific PhenoAgeAccel and PRS on liver outcomes

In the England-White cohort, individuals with a biologically older liver phenotypic age had a significantly higher risk of all-cause mortality and liver-related events, regardless of genetic risk. We detected a synergistic effect between liver-specific PhenoAgeAccel and PRS on liver-related events. Participants with both high genetic risk and biologically older liver phenotypic age had the highest risks across outcomes (Figure S11). For illustration, the hazard ratio for cirrhosis/fibrosis increased from 3.09 (95% CI 2.11–4.51) in the low-genetic-risk group to 4.02 (2.81–5.74) in the intermediate-risk group and 7.75 (5.41–11.12) in the high-genetic-risk group, compared with the reference (low genetic risk + biologically younger liver phenotypic age).

Mediation effect of liver-specific PhenoAgeAccel

Mediation analysis in the England-White cohort revealed that approximately 12–27% of the relationship between risk behavior score and liver-related outcomes was mediated by liver-specific PhenoAgeAccel, with the proportion mediated ranging from 12.4% for HCC to 26.5% for liver-related death (all P for natural indirect effect < 0.001, Fig. 3). In this analysis, liver-specific PhenoAgeAccel mediated 10.8% of the association between risk behavior score and all-cause mortality, and 5.9% of the association with liver-related death (P for natural indirect effect < 0.001, Table S17). CMAverse-based sensitivity analyses further indicated that the NIE would be driven to the null only under residual correlations of approximately |ρ|= 0.10–0.25 between the mediator and outcome models, corresponding to at least small-to-moderate levels of unmeasured mediator–outcome confounding (Table S18). Taken together, these findings suggest a modest but plausible mediating role of liver-specific PhenoAgeAccel, although interpretation should remain cautious given the strong assumptions of causal mediation models.

Fig. 3.

Mediation of liver-specific phenotypic age acceleration (PhenoAgeAccel) in the England-White cohort

Metabolic and proteomic signatures of liver-specific PhenoAgeAccel

To demonstrate biological plausibility, we linked liver-PhenoAgeAccel to lipid/glucose metabolites and liver-enriched metabolic-coagulation proteins. Among the 249 metabolites examined, 56 were significantly and directionally consistently associated with both liver-specific phenotypic aging and liver-related outcomes, with lipoprotein-related metabolites comprising the predominant class (Tables S19-21). For example, we found that the phospholipids-to-total-lipids ratio in small HDL was the metabolite most strongly associated with liver-related mortality (HR 1.54, 95% CI 1.40–1.69; FDR < 0.001) and fatal liver-related events (HR 1.48, 95% CI 1.37–1.60; FDR < 0.001). Among 2911 screened proteins, 211 showed significant and directionally consistent associations with both liver-specific phenotypic aging and liver-related outcomes (Fig. 4A; Tables S22-23). For example, the protein CDHR2 was simultaneously strongly and significantly associated with liver-specific phenotypic age acceleration (OR = 1.29, FDR < 0.001) and with liver-related mortality, liver fibrosis and cirrhosis, liver failure, HCC, and fatal liver-related events (HRs of 4.12, 2.24, 3.88, 3.15, and 2.96 respectively; all FDR < 0.001). Tissue enrichment analysis identified the liver as the predominant site of protein enrichment (Fig. 4B), with 22 proteins (e.g., ADH1B and DXCR) showing liver-specific expression—18 upregulated and 4 downregulated (Fig. 4C; Table S24). Pathway enrichment analysis revealed that the 4 downregulated proteins were mainly involved in coagulation and hemostasis pathways—such as blood coagulation, complement activation, and fibrin formation—while the 18 upregulated proteins were enriched in metabolic processes, including molecular and organic acid metabolism (Figure S12). Twelve proteins formed two interacting modules: a coagulation–immunity hub (FCN2, PROC, F11) and a metabolic-homeostasis hub (AGXT, FTCD, UPB1, etc.), bridged by synergistic changes in PROC and FTCD (Fig. 4D).

Fig. 4.

Plasma proteomic signature of liver-specific phenotypic age acceleration (PhenoAgeAccel) and liver-related outcomes. A The volcano plot shows results from logistic regression analyses examining the associations between liver-specific PhenoAgeAccel and the plasma protein signature. B Protein-tissue expression results of 211 proteins associated with both liver-specific PhenoAgeAccel and liver-related outcomes. C Expression levels of 22 liver-specifically enriched proteins. D String protein interaction network analysis indicates that 13 of the 22 proteins formed a protein interaction network. Only associations above a confidence score of 0.4 (standard medium confidence) are shown

In sex-stratified analyses, we identified 28 liver-enriched proteins associated with liver-specific PhenoAgeAccel and all liver-related outcomes in men, and 9 liver-enriched proteins associated with liver-specific PhenoAgeAccel and fatal liver-related events in women (Tables S25–28; Figure S13). ACY1, AGXT, and HAO1 were shared across sexes; FABP1 was female-specific, whereas ASGR1, ADH4, and SHBG were male-specific. Pathway enrichment indicated that, in men, 21 upregulated proteins were primarily involved in amino acid metabolism and alcohol-metabolizing enzyme activity, while 7 downregulated proteins clustered in coagulation pathways. In contrast, all 9 proteins in women were upregulated and enriched for amino acid and organic acid metabolic pathways (Table S29).

Mediation analysis showed that each of the 22 proteins accounted for approximately 3–35% of the associations of liver-specific PhenoAgeAccel with liver-related death, hepatic failure, fibrosis and cirrhosis of liver, and fatal liver-related events. Among these proteins, ACY1, AGXT, and HAO1 demonstrated the strongest mediating effects, each explaining more than 20% of the associations across 4 different liver-related outcomes (Fig. 5). Nevertheless, the aggregate mediation effects of all 22 proteins failed to attain statistical significance, and no protein-mediated effect was identified for liver-specific PhenoAgeAccel and HCC (Table S30).

Fig. 5.

Mediation effects of 22 liver-specific proteins in the associations between liver-specific phenotypic age acceleration (PhenoAgeAccel) and liver-related outcomes. A The left heat map shows 22 proteins are highlighted for their consistent associations with liver-specific PhenoAgeAccel (logistic regression analysis) and liver-related outcomes (Cox regression). B The right heat map shows the proportion of the association between liver-specific PhenoAgeAccel and four liver-related outcomes mediated by the 22 proteins, calculated using the inverse odds ratio-weighted method

Discussion

We developed a liver-specific phenotypic age acceleration (PhenoAgeAccel) metric and found strong associations with all-cause mortality and multiple liver-related outcomes, with effects most pronounced in individuals at high genetic risk. Modifiable behaviors—smoking, alcohol use, physical inactivity, and unhealthy diet—were linked to accelerated liver-specific aging, which partly mediated their associations with liver outcomes. Complementary proteomic and metabolomic analyses identified shared molecular signatures between liver-specific aging and liver disease, including 22 liver-enriched proteins implicated in metabolic regulation and coagulation pathways, providing biological context for aging-related liver vulnerability.

Although various omics-based aging clocks such as DNA methylation age, proteomic age, and metabolic age have been developed, their clinical application is limited by high cost and poor accessibility in population settings [4, 16]. In contrast, phenotypic age is derived from readily available blood biomarkers and has been validated as a robust predictor of mortality and disease across populations [5–8, 16, 18, 35, 36]. We extended this concept by creating a liver-specific PhenoAgeAccel, refined to reflect hepatic risk rather than systemic aging, and built from 11 biomarkers most predictive of liver-related mortality. This index demonstrated strong and independent associations with a spectrum of hepatic outcomes, supporting its utility as a feasible and informative tool to assess liver-specific biological aging.

Blood-based biomarkers used in our model, such as liver enzymes and LDL cholesterol, capture hepatic inflammation and metabolic status, while the inclusion of systemic markers like C-reactive protein and creatinine reflects the liver’s broader physiological integration. Prior studies have shown the predictive value of similar clinical markers for liver fibrosis and HCC [23, 31, 37]. A potential limitation of this index is “circularity,” i.e., that liver enzymes used as predictors might inflate performance for liver outcomes. We mitigated this by winsorizing biomarkers (1st/99th percentiles), excluding baseline chronic liver disease, running the lagged analysis excluding events in the first year, and recalculating liver-specific PhenoAgeAccel after excluding all liver enzymes; results were materially unchanged. This supports that the observed associations are not solely driven by baseline enzyme-defined liver injury, but reflect broader dimensions of hepatic aging vulnerability. However, liver enzymes also reflect liver health and vulnerability and are appropriate components of an organ-specific phenotypic age, though residual circularity cannot be fully excluded.

This liver-specific biological age not only retains the practical advantages of easily obtainable phenotypic age measures but also addresses the lack of organ specificity by including liver-related indicators that systemic phenotypic age assessments typically overlook. This novel and easily implementable metric provided a valuable tool for investigating liver-specific aging processes and exploring their potential links to liver-related diseases. Furthermore, this risk score demonstrated better accuracy in predicting multiple liver-related outcomes compared to the aMAP score and noninvasive fibrosis tests, suggesting that this score was more appropriate for risk stratification in the general population. Importantly, liver-specific PhenoAgeAccel is designed to capture broader liver aging and vulnerability, whereas conventional liver risk scores primarily reflect current liver function or injury.

We found that individuals with both elevated liver-specific PhenoAgeAccel and higher PRS had the greatest risk of liver-related outcomes, suggesting a synergistic effect between genetic predisposition and biological aging. Previous research has similarly shown that aging biomarkers and genetic risk interact to shape disease susceptibility [6, 8, 16, 38]. This combined approach may facilitate early identification of high-risk individuals and guide personalized lifestyle or therapeutic interventions. Importantly, liver-specific PhenoAgeAccel partially mediated the relationship between modifiable risk behaviors such as smoking, alcohol intake, physical inactivity and unhealthy diet and liver disease outcomes. These results align with previous studies showing that unhealthy behaviors accelerate biological aging and increase chronic disease risk [6, 16, 18]. Our findings suggest that behavioral interventions may not only reduce direct disease risk but also slow liver-specific aging, offering a rationale for preventive strategies.

Our omics-based analyses identified shared molecular features underlying both liver-specific PhenoAgeAccel and liver outcomes. Metabolomics revealed 56 associated metabolites, primarily involved in lipid and glucose metabolism. Given the liver’s central role in these pathways, disruptions in lipid homeostasis can promote hepatic steatosis, fibrosis, and ultimately HCC [39, 40]. Disturbances in glucose and lipid metabolism are also essential features of cellular senescence and affect hepatic aging through a variety of mechanisms [41, 42]. These results underscore the intertwined nature of metabolic dysfunction, inflammation, and liver aging.

Proteomic analysis further identified 22 liver-enriched proteins mediating the link between liver-specific aging and disease. These proteins were enriched in pathways regulating metabolic homeostasis and coagulation, forming an interactive network that bridges hepatocellular stress, inflammation, and immune-coagulative responses. Their mediation effect appeared to reflect an adaptive response of the aging liver to mounting metabolic and inflammatory challenges [43]. However, this compensatory response may gradually shift to a maladaptive state with advancing age [43]. Key mediators included ACY1, AGXT, and HAO1, each demonstrating molecular relevance to liver aging and disease. ACY1, a cytosolic enzyme involved in amino acid metabolism, is highly expressed in the liver and has tumor-suppressive roles in HCC [44]. Its elevated levels in aging livers may represent a compensatory response to metabolic or inflammatory stress, though this response appears insufficient once fatal liver outcomes occur [44, 45]. AGXT and HAO1, both peroxisomal enzymes involved in glyoxylate detoxification and hydrogen peroxide production, were also upregulated with liver-specific aging [46, 47]. These enzymes may help mitigate oxidative stress and inflammation in early stages but can exacerbate peroxisomal overload and inflammatory signaling when chronically activated. These convergent pathways collectively reflected age-linked remodeling of hepatic metabolism, redox balance and synthetic function [46] and reinforced that liver-specific PhenoAgeAccel captured authentic molecular features of hepatic aging rather than merely reflecting fatal liver diseases, strengthening the biological plausibility of its association with long-term liver outcomes. These observations are consistent with recent plasma protein–based organ-specific aging models showing that chronic diseases can be interpreted as accelerated aging of specific organ systems, including the liver [48]. By identifying liver-enriched proteins that link PhenoAgeAccel to liver outcomes, we extend this framework and provide a clinically accessible, biomarker-based liver aging metric anchored in proteomic signatures.

The marked sex disparity in hepatic proteomic aging—28 liver-enriched proteins in men vs. 9 in women—implies more extensive age-related remodeling in males, consistent with prior reports of greater male liver proteomic change [49, 50]. Likely drivers include sex-hormone regulation (androgens vs. estrogens) and sex-specific growth hormone patterns that shape hepatic gene expression and metabolism [49, 51], alongside more pronounced immunosenescence/inflammaging in men [52, 53]. Accordingly, male-specific proteins (e.g., ASGR1, ADH4, SHBG) largely map to metabolic/coagulation pathways, while the female-specific marker (FABP1) also reflects metabolism [49]. Several proteins (ACY1, AGXT, HAO1) were shared, suggesting core hepatic aging mechanisms conserved across sexes despite the overall male bias.

Despite the strength of this study, several limitations should be acknowledged. First, our analysis was based on baseline measurements, which may not fully capture changes in biomarker levels or behaviors over time. Nonetheless, previous studies have shown high stability of many UKBB biomarkers, supporting the use of baseline values as reasonable proxies [23]. Second, the smaller case numbers in the validation cohorts limited our ability to perform stratified or interaction analyses in these subgroups. Since stratification by etiologies was limited by low event counts and medical information, we could not explore etiology-specific liver disease in detail. However, liver-specific PhenoAgeAccel captures biological risk without requiring prior etiological classification, which may be advantageous in real-world settings with incomplete diagnostic data. Still, consistent effect directions across cohorts support the robustness of our main findings. Third, as with all observational studies, residual confounding cannot be ruled out, and causal inferences should be made cautiously. Meanwhile, there is a lack of external validation in cohorts outside the UKBB, particularly in those with different ethnic and geographic backgrounds. The UKBB participants are generally healthier than the source population, reflecting healthy volunteer bias. This selection may yield conservative estimates of liver-event rates and risk associations in our study. Caution is therefore warranted when extrapolating to settings with high chronic liver disease burden (e.g., HBV/HCV-endemic regions), where baseline inflammation, fibrosis, and treatment patterns could alter calibration and discrimination. External validation in high-risk, clinically enriched cohorts is needed to assess generalizability and recalibration needs. Nevertheless, liver-specific PhenoAgeAccel may still capture subclinical hepatic dysfunction preceding clinical disease, offering potential value for early risk stratification in these populations. Fourth, although we examined joint effects with a liver-disease PRS, this PRS was derived largely from European-ancestry GWAS and genome-wide genotyping is not routinely available in many clinical settings. Accordingly, the PRS should be viewed as a research tool for etiologic insight and risk stratification, whereas liver-specific PhenoAgeAccel can be implemented independently using routine clinical biomarkers. Future studies incorporating longitudinal follow-up and intervention designs in independent cohort with different ethnic and geographic backgrounds will be needed to confirm these findings and test strategies to slow liver-specific aging.

In conclusion, we developed and validated a liver-specific aging metric that predicts liver-related outcomes and mediates the effects of modifiable risk factors. By integrating clinical biomarkers with multi-omics insights, our findings reveal biological pathways linking hepatic aging to disease. This accessible tool may enhance risk stratification and support preventive strategies targeting liver health and aging.

Supplementary Information

Below is the link to the electronic supplementary material.

Abbreviations

AUROC

Area under the receiver operating characteristic curve

BA

Biological age

BMI

Body mass index

CLD

Chronic liver diseases

CI

Confidence interval

FDR

False discovery rate

GWAS

Genome-wide association study

HR

Hazard ratio

HCC

Hepatocellular carcinoma

ICD-10

International Classification of Diseases 10th Revision

LASSO-Cox

L1-penalised Cox regression

LOD

Lower limit of detection

NAFLD

Non-alcoholic fatty liver disease

NMR

Nuclear magnetic resonance

RCS

Restricted cubic spline

SD

Standard deviation

PhenoAge

Phenotypic age

PhenoAgeAccel

Phenotypic age acceleration

PRS

Polygenic risk score

UKBB

UK Biobank

Author contribution

ZL conceived of the study design. TW and CG performed the statistical analyses. ZL and TW wrote the manuscript. TW and CG performed the data visualization. MD and HY scrutinized the statistics. XC, TZ, and ZL supervised the study and revised the manuscript. All authors provided critical revisions of the draft and approved the submitted draft. The corresponding author attests that all listed authors meet authorship criteria and that no others meeting the criteria have been omitted.

Funding

This work was supported by the National Key Research and Development Program of China (2022YFC3400700; 2024YFC3405800) and the Noncommunicable Chronic Diseases-National Science and Technology Major Project (2023ZD0510000), and Shanghai Municipal Science and Technology Major Project (2023SHZDZX02).

Data availability

All data were available in the UK Biobank, and subject to registration and application processes.

Declarations

Ethics approval and consent to participate

All participants gave written informed consent prior to data collection. The UK Biobank has full ethical approval from the NHS National Research Ethics Service (16/NW/0274). This research was conducted using the UK Biobank resource under approved application number 92718.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.