Association of Epigenetic Aging Biomarkers with Risk of MASLD-related HCC

Abstract

Introduction:

Hepatocellular carcinoma (HCC) development in patients with metabolic dysfunction-associated steatotic liver disease (MASLD) is a growing health concern, but the underlying mechanisms are not fully understood. Epigenetic aging biomarkers, reflecting cellular and tissue aging, have been linked to various age-related pathologies, but their association with MASLD-HCC is unknown. We investigated associations between five epigenetic aging biomarkers and MASLD-HCC risk.

Methods:

We performed whole blood DNA methylation assay (Infinium 850k array) and calculated principal components-based (PC) versions of HorvathAge, HannumAge, PhenoAge and GrimAge, and the DunedinPACE aging rate. We further calculated relative age accelerations for PCHorvathAge, PCHannumAge, PCPhenoAge, and PCGrimAge. The aging biomarkers were modeled as continuous variables and categorized into tertiles based on distributions among controls. Associations between each aging biomarker and MASLD-HCC were examined using logistic regression, calculating odds ratios (ORs) and 95% confidence intervals (CIs), adjusting for covariates.

Results:

Data on 272 MASLD-HCC cases and 316 cancer-free MASLD controls recruited from six sites and matched on chronological age, sex, and study site were analyzed. Higher relative age accelerations of PCPhenoAge (OR_T3vsT1=2.25, 95%CI: 1.45–3.50; OR_continuous=1.04, 95%CI: 1.02–1.07, p=0.009), PCGrimAge (OR_T3vsT1=3.97, 95%CI: 2.41–6.64; OR_continuous=1.16, 95%CI: 1.10–1.24, p=8.76 × 10⁻⁰⁷), and DunedinPACE (OR_T3vsT1=3.45, 95%CI: 2.17–5.55; OR_continuous=1.72, 95%CI: 1.43–2.10, p=2.58 × 10⁻⁰⁸) were associated with MASLD-HCC, but not PCHorvathAge or PCHannumAge.

Conclusion:

Higher relative age accelerations of PCPhenoAge, PCGrimAge, and DunedinPACE aging rate are associated with risk of MASLD-HCC. These aging biomarkers could improve HCC risk assessment and facilitate risk stratification in patients with MASLD.

Lay Summary

• Metabolic dysfunction-associated steatotic liver disease (MASLD) is the fastest-rising cause of liver cancer, but the biological processes that underlie liver cancer development in patients with MASLD are not fully understood.

• Here, we investigated whether markers of biological age acceleration, as opposed to chronological age, can be used to assess HCC risk in patients with MASLD.

• Our study shows that three of the five biological aging markers examined are significantly associated with HCC risk in patients with MASLD; hence, these biological markers can be useful for identifying high-risk MASLD patients for timely intervention and may also enhance early detection models for MASLD-related liver cancer.

Introduction

Hepatocellular carcinoma (HCC), the most common form of primary liver cancer, accounts for about 70% of all liver cancers and has a 5-year survival rate of 22%.¹ The major risk factors of HCC include alcoholic liver disease; viral infections, such as hepatitis B and C virus (HBV, HCV); cigarette smoking; metabolic disorders, including obesity, diabetes, and metabolic dysfunction-associated steatotic liver disease (MASLD); and certain rare genetic conditions (e.g., hemochromatosis, Wilson’s disease, and alpha-1 antitrypsin deficiency).^2–5 Widespread vaccinations for HBV and improvements in HCV treatment over the past decade have shifted the HCC risk-factor profile from predominantly viral-related causes to MASLD, which is now the most rapidly rising cause of HCC.^5,6 Despite the increasing incidence of MASLD-HCC, the molecular mechanisms underlying HCC development in patients with MASLD remain poorly understood.^7,8 A more detailed understanding of these molecular processes could inform strategies for risk stratification, early cancer detection, and therapy selection.

Aberrant DNA methylation is a mechanism potentially underlying HCC development in patients with metabolic perturbations.^9–11 DNA methylation patterns reflect various physiological processes and result from the interplay between host biology and exposure to environmental stressors or lifestyle factors (e.g., diet, smoking, alcohol intake, and environmental toxins).^12–14 DNA methylation can estimate an individual’s biological age and may be an indicator of overall health and a potential risk marker for various age-related pathologies.^15–17 Previous studies have developed epigenetic aging biomarkers (“clocks”) by using age-related DNA methylation markers (CpGs) to estimate chronological or biological age. These epigenetic clocks are classified into three generations.¹⁸ The first-generation clocks, including Horvath’s multitissue age-prediction clock (HorvathAge)¹⁹ and Hannum’s blood-based epigenetic clock (HannumAge)²⁰, were designed to predict chronological age. The second-generation clocks, including phenotypic age (PhenoAge)¹⁶ and GrimAge¹⁷, track epigenetic deviations associated with disease risk, while the third-generation clocks, e.g., Dunedin pace of aging (DunedinPACE)²¹, measure the rate of physiological decline. These clocks were originally developed using different combinations of CpGs, capture different aspects of aging at the methylome level, correlate with chronological age to varying degrees, and are complementary tools for quantifying biological processes underlying aging.^16,19–22

These aging biomarkers show clinical promise and have been associated with various age-related diseases, including HannumAge with the risk of Alzheimer’s disease¹⁶, PhenoAge with stroke severity²³, and multiple aging biomarkers with cognitive decline and dementia²⁴, diabetes mellitus²⁵, and certain cancer types.^26–28 To our knowledge, no study has yet evaluated the relationship between the epigenetic aging biomarkers and MASLD-HCC risk. Evaluating such relationships would be useful for HCC risk stratification, for improving risk management, and for informing strategies for early HCC detection.

Importantly, the third-generation DunedinPACE aging metric reflects the rate (pace) of physiological decline associated with aging, while the earlier-generation clocks (HannumAge, HorvathAge, PhenoAge, and GrimAge) are considered estimators of chronological or biological age.^21,29,30 Principal components-based (PC) versions of the earlier-generation clocks have been developed to improve the performance of these clocks, and the PC versions have been shown to have greater reliability in longitudinal studies.^22,30 The calculated difference between chronological age and any of the four earlier-generation clocks is termed absolute age acceleration, while a residual-based method for estimating age acceleration is referred to as relative age acceleration.^29,30 Relative age acceleration, rather than absolute age acceleration, is recommended whenever there are fewer available CpGs than the required number of CpGs needed to compute the aging biomarkers.³⁰

Here, we sought to determine whether any of the five commonly examined epigenetic aging biomarkers is associated with risk of MASLD-HCC. We performed a multicenter case-control study among patients with MASLD-HCC and cancer-free MASLD controls. We assessed associations of MASLD-HCC risk with PC-based versions of the four earlier-generation aging biomarkers using residual-based relative age accelerations of PCHorvarthAge, PCHannumAge, PCPhenoAge and PCGrimAge, plus the third-generation DunedinPACE aging rate.

Methods

Data Source and Study Population

The design and methods used to recruit the study participants have been described in detail.^11,31 Briefly, data and biospecimens were obtained from the following international sites: 1) the Karolinska University Hospital, Sweden; 2) the Barcelona Clinic Liver Cancer Group (BCLC), Hospital Clinic Barcelona and IDIBAPs, Barcelona, Spain; 3) Instituto de Investigación Sanitaria Biogipuzkoa (IISB), Donostia University Hospital, San Sebastian, Spain; 4) the Virgen del Rocio Hospital Institute of Biomedicine of Sevilla (IBIS), Seville, Spain; 5) the University of Texas Southwestern (UTSW), Dallas, Texas; and 6) the Mayo Clinic sites in Rochester, Minnesota, and Jacksonville, Florida. These sites provided germline leukocyte DNA and data on 673 MASLD-HCC cases and 763 cancer-free MASLD controls. Before submitting their data and DNA samples to the Mayo Clinic, all sites were asked to include only MASLD-HCC cases. Specifically, patients with a history of alcoholic liver disease, viral hepatitis (HBV/HCV), primary sclerosing cholangitis, biliary cirrhosis, rare genetic risk factors of HCC (e.g., hemochromatosis, Wilson’s disease, Budd–Chiari syndrome, alpha-1 antitrypsin deficiency, autoimmune hepatitis), and those who consumed more than 20 grams of alcohol per day were excluded. Controls were cancer-free individuals who had an imaging, pathology, or clinical diagnosis of nonalcoholic fatty liver disease (currently known as MASLD). Data received from each site included the age at HCC diagnosis or at recruitment (for controls), sex, race/ethnicity, type II diabetes mellitus, smoking history, BMI, and amount of moderate alcohol consumption (< 20 g/day). For DNA methylation analysis, we randomly matched 320 MASLD-HCC cases with 320 MASLD controls based on age (± 5 years), sex, and study site.

All sites previously received ethical approval from their local institutional review boards (IRBs). Additional approval was obtained from the Mayo Clinic IRB for the present study (IRB # 23–000005).

DNA Methylation and Quality Control Checks

We have previously described the DNA methylation assay methods in detail³¹. Briefly, an epigenome-wide DNA methylation assay was performed at the Mayo Clinic Genome Analysis Core laboratory using the Illumina Infinium Methylation EPIC BeadChip microarray (EPIC array, Illumina Inc., San Diego, CA), which covers over 850,000 CpG sites. DNA was first quantified by using the Invitrogen Qubit dsDNA kit (ThermoFisher Scientific, Inc., Waltham, MA), followed by bisulfite modification as recommended by Illumina for the EPIC array, and measurements with a NanoDrop instrument. The DNA methylation assay was then performed on all 640 samples from cases and controls, placed in eight 96-well plates. Sixteen additional laboratory control DNA samples (methylated [catalog #D5011] and unmethylated [catalog #D5014] human DNA control samples from Zymo Research Inc., Irvine, CA) were included (one pair of methylated and unmethylated laboratory control DNA samples per plate). We also included 64 duplicate study participant samples, which were evenly distributed across the eight assay plates. The DNA methylation status of the CpG sites was determined by comparing the ratio of a fluorescent signal from the methylated allele to the sum of fluorescent signals from both methylated and unmethylated alleles using beta-values. The beta-values for the CpGs ranged from 0 to 1, representing unmethylated to fully methylated CpG sites. Both the participant duplicate samples and laboratory control samples showed good assay performance, with ≥ 98% correlation for duplicate samples, and interclass correlations of 95% for unmethylated laboratory controls and 83% for methylated laboratory controls. For the duplicated participant samples, we retained those with the highest call rates for the final analysis.

Quality control checks (QC) included removal of CpGs located on the X or Y chromosome, CpGs determined to be cross-reactive, those that overlapped with a genetic variant (e.g., single nucleotide polymorphism), and those that failed in ≥ 10% of the samples. We assessed batch effects by using principal component analysis and the Kruskal-Wallis rank-sum test and examined associations of the top principal components with the experimental plates and did not find any association, indicating the absence of a batch effect. Following QC, we normalized the data using the dasen function within the R wateRmelon package.³² We also imputed a small proportion (< 0.01%) of missing beta-values using k-nearest neighbor with a k parameter of five in the ChAMP R package, as described in detail previously.³¹ The methylation data used for this study and key covariates (age and sex) have been deposited in the NCBI GEO database (GSE281691).³³

Computation of Epigenetic Aging Biomarkers

Biological age predictions were computed using the methscore function in the ENmix R package.³⁴ The methscore function yields 158 previously published age predictors related to chronological age, biological age, exposures, lifestyle traits, and serum protein levels using both classical and PC-based methods. We focused on only five biological age estimators that are most frequently evaluated in relation to disease risk: PCHorvathAge, PCHannumAge, PCPhenoAge, PCGrimAge, and the DunedinPACE aging rate. Because of the extensive QC performed on our methylation data³¹, we did not have data available for all the “clock CpGs” required to compute each aging biomarker. However, we had data on 90% of the clock CpGs needed to compute PCHorvathAge, PCHannumAge, PCPhenoAge, and PCGrimAge, and 77% of the clock CpGs required for calculating the DunedinPACE aging rate (Suppl. Table 1). The PC-based clocks were computed using the same number of CpGs but with different pre-computed weights assigned to each CpG per clock by using elastic net regression as described in detail by Higgins-Chen et al.²² When computing absolute age acceleration from CpGs (i.e., subtracting chronological age from methylation-based predicted age), missing clock CpG data could cause individuals who are actually aging at a normal rate to be predicted to be aging faster than normal.³⁰ We used the principal component-based methods in estimating each ageing clock. Prior to computing the principal components, the methscore function in the ENmix R package is imputing the missing clock CpGs with the mean methylation beta value from the GSE40279 blood dataset available from NCBI Gene Expression Omnibus (GEO).³¹ Each ageing clock has coefficients for their corresponding principal components and intercept determined by an elastic net prediction model which are further used to predict the ageing clocks in our data. Thus, we computed residual-based relative age accelerations to account for potential biases that might have risen from a reduction in the available number of clock CpGs, as recommended by Teschendorff and Horvath.³⁰ We calculated the residual-based relative age accelerations for PCHorvathAge, PCHannumAge, PCPhenoAge and PCGrimAge by using the residuals from a fitted line generated from control sample data. For each biological age metric, the relative age acceleration in controls was computed as the residuals from a fitted linear model, with biological age as the response variable and chronological age as the explanatory variable. The fitted linear model for controls was further used to predict biological age and to compute the residuals for the cases.³⁰ We did not compute relative age acceleration for DunedinPACE because this variable measures aging rate (pace of aging)²¹, rather than age in years.

Statistical Analysis

Differences in clinical and demographic characteristics were compared between the cases and controls using means and standard deviations for continuous variables, and frequencies and proportions for categorical variables. Pearson’s correlation was used to assess pair-wise correlations among the epigenetic aging biomarkers and their relation to chronological age. Differences in mean biomarker variables between cases and controls were assessed with t-tests. Association analyses were performed using logistic regression to calculate odds ratios (ORs) and 95% confidence intervals (CIs) in minimally adjusted and fully adjusted models. The epigenetic aging biomarker variables were modeled as both continuous and as tertile categories (T1–T3) based on the distribution among controls (cut-off values for tertiles are provided in Suppl. Table 2). In the minimally adjusted models, we adjusted for chronological age (continuous) and sex for association of the DunedinPACE aging rate with MASLD-HCC risk, whereas we adjusted for only sex for the relative age acceleration variables of PCHorvathAge, PCHannumAge, PCPhenoAge, and PCGrimAge because the computation of the relative age accelerations of these latter four epigenetic clocks included chronological age. In the multivariable models, we additionally adjusted for cigarette smoking (never, former, current), diabetes (yes/no), and BMI (continuous) for each of the models. Sensitivity analysis was performed using the actual age estimates for PCHorvathAge, PCHannumAge, PCPhenoAge and PCGrimAge, and were compared with results obtained when the residual-based relative age acceleration versions of these variables were used (Suppl. Table 3). Association results were considered statistically significant after Bonferroni correction for multiple testing (p-value < 0.01; 0.05/5 biomarkers). All statistical tests were two-sided and were performed using the R software (v.4.4.1). Data were visualized using ggplot2 in R.

Results

Among the 640 samples (320 MASLD-HCC cases and 320 MASLD controls) used for the methylation assay, 46 samples were excluded because of discordance between biological sex and self-reported sex, one sample had a failed assay, and five samples were deemed to be outliers, leaving 588 samples (272 cases and 316 controls) for analyses. Descriptive statistics of the analytic sample are shown in Table 1. Briefly, the MASLD-HCC cases and MASLD controls did not differ by age, sex, study site, or cigarette smoking history. However, compared to controls, the cases included a greater proportion of non-Hispanic Whites (93% vs. 80%), and individuals with a personal history of diabetes mellitus (78% vs. 53%). BMI was slightly higher in controls than in cases (mean BMI: 32 kg/m² vs. 31 kg/m²).

Table 1.

Characteristics of the study participants (N=588)

	MASLD-HCC Cases (N=272)	Metabolic Controls (N=316)	P-valuec
Age, yearsa			0.90
Mean (SD)	65(11)	65(11)
Sex			0.50
Female	90 (33.1%)	113 (35.8%)
Male	182 (66.9%)	203 (64.2%)
Site			0.83
Mayo Clinic, MN and FL, and UTSWb	221 (81.2%)	260 (82.3%)
Karolinska University Hospital	23 (8.5%)	23 (7.3%)
BCLC, Barcelona, and IISB, San Sebastian, Spain	18 (6.6%)	18 (5.7%)
IBIS, Seville, Spain	10 (3.7%)	15 (4.7%)
Race			< 0.001
White	217 (79.8%)	295 (93.4%)
Other	55 (20.2%)	21 (6.6%)
BMI			0.04
Mean (SD)	31.1 (5.5)	32.3 (7.6)
Diabetes Mellitus			< 0.001
No	59 (21.7%)	148 (46.8%)
Yes	213 (78.3%)	168 (53.2%)
Smoking			0.28
Current	19 (7.0%)	15 (4.9%)
Former	126 (46.5%)	132 (42.9%)
Never	126 (46.5%)	161 (52.3%)

^aChronological age at diagnosis for cases and at recruitment for controls

^bData from the UTSW were all cases (n=43) and were combined with Mayo Clinic samples.

^cP-values were calculated using Wilcoxon’s test for continuous variables (age and BMI) and chi-square test for categorical variables

Abbreviations: BCLC, Barcelona Clinic Liver Cancer Group, Barcelona, Spain; BMI, body mass index; IBIS, Institute of Biomedicine of Sevilla, Seville, Spain; IISB, Instituto de Investigación Sanitaria Biodonostia Research Institute, Donostia University Hospital, San Sebastian, Spain; UTSW, University of Texas Southwestern.

Figure 1 shows pair-wise correlations among the epigenetic age estimators in years (untransformed estimates), and their correlations with chronological age and the DunedinPACE aging rate. Notably, PCGrimAge had the strongest correlation with chronological age (Pearson’s correlation coefficient [r] = 0.92), followed by PCHannumAge (r = 0.73), PCHorvathAge (r = 0.71), and PCPhenoAge (r = 0.66) in the cases and controls combined. The DunedinPACE aging rate showed the weakest correlation with chronological age (r = 0.10). Among the age estimators, we observed high positive pair-wise correlations among PCHorvathAge, PCHannumAge, and PCPhenoAge in the combined cases and controls, as well as among the cases and controls separately (all r ≥ 0.87). PCGrimAge also correlated strongly with PCHorvathAge, PCHannumAge, and PCPhenoAge (all r ≥ 0.69). The DunedinPACE aging rate correlated weakly with all epigenetic age estimators (r = 0.24–0.44).

Figure 1.

Pairwise correlation between chronological age and epigenetic aging metrics.

Using the relative age acceleration variables, we examined differences in the distribution of the epigenetic aging biomarkers between cases and controls (Fig. 2). We observed higher estimated biological ages for the cases than for the controls for relative age accelerations of PCHannumAge (p = 0.02), PCPhenoAge (p = 0.0001) and PCGrimAge (p = 1.0 × 10⁻⁰⁸), but not relative age acceleration of PCHorvathAge (p = 0.25). The DunedinPACE aging rate was also higher in cases than in controls (p = 1.4 × 10⁻¹¹).

Figure 2.

Boxplots of five epigenetic aging biomarkers with respect to HCC case-control status. A-Distribution of Horvath Age, B-Distribution of Hannum Age, C- Distribution of PhenoAge, D-Distribution of PCGrimAge, and E-Distribution of DunedinPACE epigenetic aging biomarkers. Abbreviation: RAA, Residual-based age acceleration; std, standardized

Table 2 presents association results for both the minimally adjusted and fully adjusted models. These results were considered significant after Bonferroni correction for multiple comparisons (p <0.01). Below, we summarize the results of the fully adjusted models. We found that a higher PCPhenoAge relative age acceleration was associated with a higher risk of MASLD-HCC (OR_{T3 vs. T1} = 2.25, 95% CI: 1.45–3.50; OR_continuous = 1.04, 95% CI: 1.02–1.07, p < 0.0009). Higher relative age acceleration of PCGrimAge was also associated with elevated MASLD-HCC risk (OR_{T3 vs. T1} = 3.97, 95% CI: 2.41–6.64; OR_continuous = 1.16, 95% CI: 1.10–1.24, p = 8.76 × 10⁻⁰⁷). Further, a higher DunedinPACE aging rate was associated with higher MASLD-HCC risk (OR_{T3 vs. T1}= 3.45, 95% CI: 2.17–5.55; OR_continuous = 1.72, 95% CI: 1.43–2.10, p = 2.58 × 10⁻⁰⁸). However, no statistically significant association was observed for relative age acceleration of PCHorvathAge (OR_{T3 vs. T1} = 1.19, 95% CI: 0.78–1.80; OR_continuous = 1.02, 95% CI: 0.99–1.05, p = 0.26) or PCHannumAge (OR_{T3 vs. T1} = 1.74, 95% CI: 1.13–2.69; OR_continuous = 1.04, 95% CI: 1.00–1.08, p = 0.03).

Table 2.

Association of epigenetic age acceleration and aging rate biomarkers with risk of MASLD-HCC (N=588)

			Minimally adjusteda		Fully adjustedb
Epigenetic Aging Biomarkers	Case	Control	OR (95% CI)	P-value	OR (95% CI)	P-value
PCHorvathAge RAA
T1	87	104	1.00 (ref)		1.00 (ref)
T2	74	104	0.84 (0.55–1.27)	0.40	0.77 (0.49–1.19)	0.23
T3	111	108	1.21 (0.82–1.79)	0.35	1.19 (0.78–1.80)	0.42
Continuous	272	316	1.02 (0.99–1.05)	0. 27	1.02 (0.99–1.05)	0.26
PCHannumAge RAA
T1	72	104	1.00 (ref)		1.00 (ref)
T2	71	104	0.99 (0.64–1.52)	0.95	0.98 (0.62–1.55)	0.93
T3	129	108	1.73 (1.15–2.60)	0.009	1.74 (1.13–2.69)	0.01
Continuous	272	316	1.04 (1.00–1.07)	0.03	1.04 (1.00–1.08)	0.03
PCPhenoAge RAA
T1	55	104	1.00 (ref)		1.00 (ref)
T2	78	104	1.41 (0.91–2.19)	0.13	1.21 (0.76–1.94)	0.42
T3	139	108	2.42 (1.61–3.67)	0.00003	2.25 (1.45–3.50)	0.0003
Continuous	272	316	1.05 (1.02–1.08)	0.0002	1.04 (1.02–1.07)	0.0009
PCGrimAge RAA
T1	38	104	1.00 (ref)		1.00 (ref)
T2	80	104	2.29 (1.41–3.76)	0.0009	2.09 (1.26–3.50)	0.005
T3	154	108	4.37 (2.74–7.10)	1.21 × 10−09	3.97 (2.41–6.64)	9.19 × 10−08
Continuous	272	316	1.17 (1.11–1.25)	3.00 × 10−08	1.16 (1.10–1.24)	8.76 × 10−07
DunedinPACE Aging Rate
T1	40	104	1.00 (ref)		1.00 (ref)
T2	76	104	1.92 (1.20–3.09)	0.007	1.79 (1.10–2.96)	.02
T3	156	108	3.80 (2.45–5.96)	3.52 × 10−09	3.45 (2.17–5.55)	2.12 × 10−07
Continuous	272	316	1.83 (1.53–2.21)	1.25 × 10−10	1.72 (1.43–2.10)	2.58 × 10−08

^aAdjusted for chronological age (continuous) and sex for DunedinPACE. Chronological age was included in the calculations of age accelerations; thus, we only adjusted for sex for PCHorvathAge RAA, PCHannumAge RAA, PCPhenoAge RAA, and PCGrimAge RAA.

^bAdditional adjustment for cigarette smoking (never, former, current), diabetes (yes/no), and BMI (continuous)

Abbreviation: PC, principal components-based, RAA, residual-based relative age acceleration; T, tertile; OR, odds ratio; CI, confidence interval

Similar results were observed when we used the direct age estimators for PCHorvathAge, PCHannumAge, PCPhenoAge, and PCGrimAge in a sensitivity analysis, although we noted slightly higher effect estimates for the direct age estimators than the relative age acceleration versions of the variables (Suppl. Table 3).

Discussion

Using five epigenetic aging biomarkers previously shown to predict physiological dysregulation, biological decline and lifespan^{16,17,19–21,35}, we here showed that MASLD-HCC risk is associated with higher relative age accelerations of PCPhenoAge and PCGrimAge, as well as with a higher DunedinPACE aging rate. These associations remained significant after accounting for multiple risk factors of HCC, including chronological age. The associations were particularly strong for relative age acceleration of PCGrimAge and DunedinPACE aging rate. However, we found no significant association of relative age accelerations of PCHorvathAge and PCHannumAge with MASLD-HCC risk. These findings suggest that PCPhenoAge, PCGrimAge, and the DunedinPACE aging rate might be important predictors of MASLD-HCC risk.

Chronological age is a major risk factor for many chronic diseases, including solid tumors, and is a predictor of mortality, but considerable interindividual differences exist in the rate of aging.^15,36–38 Biological aging and disease susceptibility differ even among individuals of the same chronological age because of differences in host biological disposition and life experiences.^21,39 Epigenetic mechanisms may reflect individual differences in aging, as they are influenced by host factors and the accumulation of environmental exposures over time.^12–14 Age-associated CpGs are differentially methylated in cancer, including HCC, and thus they may be involved in cancer development and progression.^10,31,40 Using epigenetic markers to identify individuals with the greatest risk of developing cancer could enhance targeted interventions that might prevent or delay cancer onset.^41,42 The findings of our case–control study contribute to the growing evidence that epigenetic aging biomarkers may serve as a composite marker of disease risk. However, whether the associations observed with these aging biomarkers reflect methylation at specific CpG sites or whether they represent a global measure reflecting several age-related molecular processes (e.g., accumulation of metabolically active senescent cells, genomic instability, stem cell exhaustion, and/or mitochondrial dysfunction) remains unclear.^26,30,43 Further, none of the aging biomarkers examined in this study correlated perfectly with chronological age, implying that these markers may partly reflect an underlying decline in biological processes related to aging.^26,30

The original HorvathAge and HannumAge metrics were developed to measure chronological age, but both also reflect biological age to a lesser extent than the other epigenetic clocks.^19,20,30 The original HorvathAge metric uses 353 CpGs derived from a model that used penalized regression of chronological age against nearly 22,000 CpGs, with the final set of 353 CpGs predicting epigenetic aging across 51 different tissues and cell types, and is thought to reflect chromatin state of various tissues.¹⁹ HannumAge uses 71 CpGs that were derived from methylome-wide profiling of blood samples by using a penalized regression model that included chronological age, sex, and BMI.²⁰ PhenoAge and GrimAge were both developed to measure physiological dysregulation during aging.^16,17 The PhenoAge metric uses 513 CpGs derived from a model that included nine blood measures (albumin, creatinine, serum glucose, C-reactive protein, alkaline phosphatase levels, lymphocyte percentage, mean red blood cell volume, red blood cell distribution width, and white blood cell count), as well as chronological age.¹⁶ GrimAge uses 1030 CpGs that were derived from a model that included seven plasma protein markers, pack-years of smoking, and chronological age.¹⁷ The DunedinPACE age metric uses 173 CpGs that were derived from a model that calculated the longitudinal rate of change in 19 blood markers that reflect cardiovascular, metabolic, hepatic, immune, renal, dental, and pulmonary health, measured at four different time points over a 20-year period in individuals of the same chronological age.²¹ The PC-based versions of HorvathAge, HannumAge, PhenoAge and GrimAge, as used in this study, were developed to improve the performance of these earlier-generation clocks.^22,30 The five clocks examined in this study are complementary tools for quantifying biological aging and their association to age-related disease risk.^26,30

Few studies have examined associations between epigenetic aging biomarkers and cancer risk, but no study to date has investigated associations between the aging biomarkers and MASLD-HCC risk. Kresovich et al. examined the association of biological age accelerations of HannumAge, HorvathAge, and PhenoAge with breast cancer risk using a case–cohort study with samples drawn from the prospective Sister Study cohort.³⁷ These investigators measured the aging biomarkers in baseline blood samples in a cohort with a median follow-up of 6 years and found that all three aging biomarkers were significantly associated with breast cancer risk. Using lag-time analysis, they further showed that increased biological age acceleration was present many years before clinical diagnosis of breast cancer, suggesting that the aging metrics may be risk biomarkers for breast cancer.³⁷ Ambatipudi and colleagues also found a significant association between HorvathAge acceleration and overall as well as postmenopausal breast cancer risk in a nested case-control study conducted within the European Prospective Investigation into Cancer and Nutrition cohort.⁴⁴ Using data from the Normative Aging Study, Zheng et al. found that HannumAge is associated with overall cancer risk and all-cause mortality.⁴⁵ PhenoAge has also been associated with lung cancer risk.⁴⁶ Employing a Mendelian randomization approach, Deng et al. showed that HorvathAge and HannumAge accelerations are associated with a higher bladder cancer risk.⁴⁷ Although interest in exploring epigenetic aging for assessing cancer risk is increasing, not all studies have found significant associations. For example, Deng et al. did not find associations for PhenoAge and GrimAge accelerations with bladder cancer risk.⁴⁷ Similarly, other studies found no association between epigenetic aging biomarkers and different cancers.^35,48 The collective evidence therefore suggests that the association between certain aging biomarkers and cancer risk may be cancer-specific.⁴⁹

Although we found significant associations for PCPhenoAge, PCGrimAge, and the DunedinPACE aging rate regarding MASLD-HCC risk, we only observed a nominal association (p < 0.05) between PCHannumAge and MASLD-HCC risk, and no association for PCHorvathAge. Both the Horvath and Hannum epigenetic aging metrics were originally designed as potential surrogates for chronological age and may be more reflective of inherent developmental biological processes (e.g., puberty, reproduction, and menopause).^16,19,50 Further investigation of these clocks regarding MASLD-HCC risk in other populations could help clarify their associations with this cancer. However, the aging metrics examined here might also plausibly measure different aspects of aging or different processes involved in cancer development.²¹ Moreover, the precise physiological mechanisms reflected by these aging metrics are not completely clear and is a subject of ongoing investigations.^26,30

Our findings regarding PCPhenoAge, PCGrimAge, and the DunedinPACE aging rate could have potential clinical application by serving us adjunct to patient risk stratification toward early HCC detection.^17,51 Any of these three aging biomarkers could be combined with established clinical risk factors (e.g., fibrosis stage, diabetes, BMI, genetic risk variants like PNPLA3) to refine HCC risk prediction in MASLD.⁵¹ Further, these epigenetic aging measures could help identify subgroups of MASLD patients at higher biological risk for HCC, improving patient selection for prevention or chemoprevention trials. Among patients with comparable clinical profiles (e.g., early-stage fibrosis [F2]), those with accelerated epigenetic aging might be prioritized for enhanced HCC surveillance, including more frequent imaging or biomarker testing.^37,52 Tracking epigenetic aging rates overtime could also help monitor response to interventions, such as lifestyle intervention or pharmacotherapy. Evidence of slowing or reversal of accelerated aging also may indicate reduced long-term HCC risk, which can help refine cancer surveillance and screening protocols. Therefore, while epigenetic aging biomarkers are still emerging tools, there are several realistic pathways by which they could be incorporated into clinical practice, especially in conditions like MASLD-related HCC. Their minimally invasive nature and dynamic responsiveness to metabolic and environmental changes make them attractive candidates for early detection, monitoring disease progression, and evaluating intervention efficacy. Ultimately, incorporating these measures into multifactorial risk prediction models could improve identification of high-risk individuals who would benefit the most from intensified surveillance or targeted prevention strategies. Our study had several strengths and limitations. Strengths include our focus on MASLD-HCC, a rapidly rising public health problem and an HCC subtype thought to have distinct underlying molecular mechanisms.⁵³ We also evaluated the role of epigenetic aging biomarkers in HCC risk, which has not been reported on previously. Our study had a sufficiently large sample size, sourced through our international collaboration, which made it possible to observe significant associations. We excluded patients with competing conditions that predispose to HCC (e.g., viral hepatitis, excessive alcohol intake, and rare genetic predisposing factors) and we were able to control for potential confounding by adjusting for several known HCC risk factors in patients with MASLD. Limitations of our study include the predominantly non-Hispanic white population, which limits generalizability of the findings more broadly to minority groups. Because of the extensive QC checks performed on our methylation data, we did not have data on the complete set of CpGs used to develop each aging metric. We addressed this by calculating relative age acceleration versions of each metric, which is known to help overcome this challenge.³⁰ We assessed associations for only five commonly used epigenetic aging biomarkers. However, any of the other less-frequently evaluated epigenetic aging biomarkers that were not examined in this study could be associated with MASLD-HCC risk. Additionally, tumor stage and cirrhosis status data were not available for most patients; thus, we were unable to evaluate potential associations between the aging biomarkers and these phenotypes. Because of our case-control design, we could not explore temporal relationships between the aging biomarkers and MASLD-HCC risk; hence, no causal inferences can be made based on these findings. Follow-up studies in prospective cohorts with multiple measurements of methylation markers over time (preferably with more diverse participants) would help delineate whether the epigenetic clocks are HCC-predisposing factors or are a consequence of established HCC. Either of these options would imply that the aging metrics could be useful for risk assessment or could potentially improve early detection strategies.

In summary, we examined associations between five epigenetic measures of biological aging and MASLD-HCC risk. We found evidence supporting the roles of PCPhenoAge and PCGrimAge age accelerations, as well as the DunedinPACE aging rate, in MASLD-HCC risk. Although our findings await validation in independent studies, they suggest that accelerated biological aging may be an HCC risk factor in patients with MASLD. Once validated, these aging biomarkers could facilitate HCC risk assessment or could possibly enhance the performance of early detection models when combined with other markers used clinically for HCC surveillance and detection.

List of Abbreviations:

CI

Confidence interval

CpG

Regions of the genome where a cytosine nucleotide is followed by a guanine nucleotide

HCC

Hepatocellular carcinoma

MASLD

Metabolic dysfunction-associated steatotic liver disease

NCBI GEO

National Center for Biotechnology Information Gene Expression Omnibus

OR

Odds ratio

PC

Principal components

QC

Quality control

T1-T3

Tertile 1 to tertile 3

Data Availability Statement

The methylation data and minimal covariate data has been made publicly available in the NCBI|GEO database (GSE281691). Additional covariate data can be made available to interested researchers upon request to the corresponding author.