This study has been reported according to the STROBE-MR guidelines (Skrivankova et al., 2021; Supplementary file 2).

We obtained summary genetic association estimates for epigenetic age acceleration measures of HannumAge (Hannum et al., 2013), Intrinsic HorvathAge (Horvath, 2013), PhenoAge (Levine et al., 2018), and GrimAge (Lu et al., 2019a) from a recent GWAS meta-analysis of biological ageing (McCartney et al., 2021), which included 34,710 participants of European ancestry. Across the 28 European ancestry studies considered in the analysis, 57.3% of participants were female. A detailed description of the methods that were used can be found in the publication by McCartney et al., 2021. In short, the Horvath epigenetic age calculator software (https://dnamage.genetics.ucla.edu) or standalone scripts were used to calculate age adjusted DNAm estimates. Outlier samples with clock methylation estimates of +/−5 s.d. from the mean were excluded from further analysis. SNPs were genotyped and imputed independently for each cohort included in the meta-analysis. Genotypes were imputed using either the HRC or the 1000 Genomes Project Phase 3 reference panels in all cohorts but the Sister Study (which did not have imputed data at the time of analysis) and the Genetics of Lipid Lowering Drugs and Diet Network Study (which used whole-genome sequencing data). GWAS summary statistics were obtained in each cohort using additive linear models adjusted for sex and genetic principal components, and they were later processed and harmonised using the ‘EasyQC’ R package. Fixed effect meta-analyses were performed using the METAL software (Willer et al., 2010).

We used the clump_data function in the ‘TwoSampleMR’ R package to select GWAS-significant SNPs (P < 5 × 10⁻⁸) for each epigenetic age acceleration measure and perform linkage disequilibrium (LD) clumping (r² <0.001) using the European reference panel from the 1000 Genomes Project Phase 3 v5.

We identified 9 independent SNPs for HannumAge, 24 for Intrinsic HorvathAge, 11 for PhenoAge and 4 for GrimAge (Supplementary file 1 — Table 1). The proportions of trait variance explained by genetic instruments (R²) and instrument strength (F-statistic) were calculated using the following formulae: R² = (2β²×MAF×(1-MAF))/(2β²×MAF×(1-MAF) + 2 N × MAF × (1-MAF)×SE²) and F = (R²×(N-2))/(1-R²) (where MAF = effect allele frequency, β = effect estimate of the SNP in the exposure GWAS, SE = standard error, N = sample size) (Palmer et al., 2012). The genetic instruments for HannumAge, Intrinsic HorvathAge, PhenoAge and GrimAge acceleration explained 1.48%, 4.41%, 1.86%, and 0.47% of the trait variance, respectively. All the selected SNPs had F-statistics greater than 10 (HannumAge median 38 and range 31–99, Intrinsic HorvathAge median 47 and range 31–240, PhenoAge median 45 and range 32–89, GrimAge median 36 and range 31–45).

We obtained summary-level genetic association data for cancer outcomes from the UK Biobank, FinnGen and several international cancer genetic consortia: the Breast Cancer Association Consortium (BCAC), the Ovarian Cancer Association Consortium (OCAC), the Consortium of Investigators of Modifiers of BRCA1/2 (CIMBA), the Prostate Cancer Association Group to Investigate Cancer Associated Alterations in the Genome (PRACTICAL), the International Lung Cancer Consortium (ILCCO) and the Genetics and Epidemiology of Colorectal Cancer Consortium (GECCO) (Table 1). Further details of the studies and the data obtained are described in Appendix 1.

We extracted genetic association data for the selected SNPs from each cancer GWAS (for breast, prostate, colorectal, ovarian and lung cancers). LD proxies (r² >0.8) were used when the SNPs of interest were missing from the cancer GWAS dataset. The proxies were located using the MR-Base platform, which calculates LD using the European subset of individuals from the 1000 Genomes Project reference panel as above (Hemani et al., 2018). The ‘LDlinkR’ R package version 1.1.2 was used to find proxies for cancer data that were not included in the MR-Base platform. The exposure and outcome datasets were then harmonised to ensure the genetic associations reflect the same effect allele. Palindromic SNPs with minor allele frequencies (MAF) <0.3 were aligned, while those with MAF ≥0.3 or mismatching strands were excluded.

Statistical power was calculated using an online calculator for MR available at: https://shiny.cnsgenomics.com/mRnd/. Calculations were performed separately for each clock-cancer combination. They were based on a type one error rate of 0.05, the proportion of phenotypic variance explained by genetic variants (R²) for each measure of epigenetic age acceleration, and the total number of cases and controls included in the meta-analysis for each cancer. Across combinations of the four epigenetic clock acceleration and five cancer measures, we had 80% power to detect ORs as small as 1.04–1.39 (Supplementary file 1 — Table s2).

We estimated the genetically predicted effects of epigenetic age acceleration (as measured by HannumAge Hannum et al., 2013, Horvath Intrinsic Age Horvath, 2013, PhenoAge Levine et al., 2018 and GrimAge Lu et al., 2019a) on multiple cancers (i.e. breast, prostate, colorectal, ovarian, and lung cancer) using a two-sample MR framework (Figure 1).

Main analyses were performed using multiplicative random effects inverse variance weighted (IVW) MR, a method that combines the genetically predicted effect of epigenetic age acceleration on cancer across genetic variants (Burgess et al., 2013). This is the default IVW MR method in the ‘TwoSampleMR’ R package, as it accounts for excess heterogeneity across SNP-specific estimates (as opposed to the fixed effect IVW method) and it does not affect the relative weights of individual SNP estimates (in contrast to the additive random effects IVW method) (Bowden et al., 2017).

We used fixed effect meta-analysis to pool results across studies (i.e. UK Biobank, FinnGen and international consortia). For colorectal cancer, we only pooled FinnGen and GECCO estimates, since UK Biobank participants were already included in GECCO. I² statistics and their corresponding confidence intervals were used to estimate heterogeneity across study estimates (von Hippel, 2015). A Benjamini-Hochberg false discovery rate (FDR) < 5% was used to correct the pooled main IVW results for multiple testing (Benjamini and Hochberg, 1995). This correction was applied considering a total of 20 independent statistical tests (4 clocks x 5 cancers = 20).

MR assumes genetic instruments for epigenetic age acceleration are (1) associated with epigenetic age acceleration (relevance assumption), (2) independent of confounders of the association between the instruments and cancer (independence assumption), and (3) only associated with cancer through their effect on epigenetic age acceleration (exclusion restriction assumption) (Didelez and Sheehan, 2007; Davies et al., 2018).

As a sensitivity analysis and to test for potential violations of the relevance assumption, we calculated F-statistics and the R² for each measure of epigenetic age acceleration (Burgess and Thompson, 2011). Other sensitivity analyses included MR-Egger (Bowden et al., 2015), weighted median (Bowden et al., 2016) and weighted mode (Hartwig et al., 2017) methods, which are robust to some of the assumptions of the IVW approach (described in Appendix 1). These results were also pooled across studies, as explained above. Consistency across different MR methods would suggest that it is less likely that the independence and exclusion restriction assumptions are violated.

We further assessed the validity of the independence assumption by conducting MR analyses using negative control outcomes (i.e. skin colour, ease of skin tanning). Evidence of causality between our genetic instruments for epigenetic age acceleration and these negative control outcomes would suggest potential bias due to population stratification that has not been fully accounted for through adjustments in the GWAS (Sanderson et al., 2021). We also assessed the genetically predicted effect of epigenetic age acceleration on cancer risk factors (i.e. body mass index, waist circumference, pack years of smoking, time spent doing vigorous physical activity, age completed full time education, years of schooling, and alcohol intake frequency) to detect potential violations of the exclusion restriction assumption. GWAS data for negative control outcomes and cancer risk factors were obtained using the University of Bristol’s IEU OpenGWAS API (for more details, see Appendix 1).

Where associations between genetically predicted epigenetic age acceleration and cancer were identified, we additionally performed single-SNP two-sample MR analysis to assess whether the effects were likely to be driven by a single SNP. We used the METAL software (Willer et al., 2010) to conduct a GWAS meta-analysis of cancer genetic association data obtained from the UK Biobank, FinnGen and international cancer genetic consortia. We then used these meta-analysed summary statistics in two-sample MR analyses. Scatter plots showing the effects of genetic instruments on epigenetic clock acceleration against their effects on cancer were created using the ‘TwoSampleMR’ R package. Additionally, Cochran’s Q statistics were used to quantify global heterogeneity across SNP-specific MR estimates (Bowden et al., 2019) and MR-Egger intercept tests were performed to detect horizontal pleiotropy (Bowden et al., 2015).

We also used Causal Analysis using Summary Effect Estimates (CAUSE) (Morrison et al., 2020), a method that uses genome-wide summary statistics to disentangle causality (i.e. SNPs are associated with cancer through their effect on epigenetic age acceleration) from correlated horizontal pleiotropy (i.e. SNPs are associated with epigenetic age acceleration and cancer through a shared heritable factor), while taking into account uncorrelated horizontal pleiotropy (i.e. SNPs are associated with epigenetic age acceleration through separate mechanisms). It uses Bayesian modelling to assess whether the sharing model (i.e. model that fixes the causal effect at zero) fits the data at least as well as the causal model (i.e. model that allows a causal effect different from zero).

As a secondary analysis, we conducted two-sample MR of epigenetic age acceleration and cancer subtypes (i.e. breast cancer: ER+, ER-, triple negative, luminal B/HER2-negative-like, HER2-enriched-like, luminal A-like, luminal B-like, BRCA1 and BRCA2; ovarian cancer: high-grade serous, low-grade serous, invasive mucinous, clear cell, endometrioid, BRCA1 and BRCA2; prostate cancer: advanced, advanced [vs non-advanced], early onset, high risk [vs low risk], and high risk [vs low and intermediate risk]; lung cancer: adenocarcinoma and squamous cell; colorectal cancer: colon-specific, proximal colon-specific, distal colon-specific, rectal-specific, male and female) (Appendix 1).

We also performed two-sample MR analyses of epigenetic age acceleration and parental history of cancer in the UK Biobank for breast, prostate, lung and bowel cancer (Appendix 1). Data on parental history of ovarian cancer were not available in UK Biobank. Family history data correlate with combined hospital record and questionnaire data and it has been suggested that they provide better power to detect GWAS-significant associations for some phenotypes in the UK Biobank (DeBoever et al., 2020). Therefore, we expected these results to be consistent with those obtained in the main analyses.

MR results were reported as the odds ratio (OR) of site-specific cancer per one year increase in genetically predicted epigenetic age acceleration. These did not require any scale transformations, as the GWAS of biological ageing (McCartney et al., 2021) reported epigenetic age acceleration in years.

LD Score regression (Bulik-Sullivan et al., 2015) was used to identify genome-wide genetic correlations between epigenetic age acceleration and cancer. Genetic correlations were estimated using full GWAS summary statistics for the epigenetic clocks and cancer, as well as the 1000 Genomes Project European LD reference panel. Traits with mean heritability chi-square values < 1.02 were excluded from the analyses.

Finally, bidirectional MR analyses were conducted to assess the causality and directionality of the link between epigenetic clock acceleration and telomere length, another measure of biological ageing that has been shown to influence cancer risk in prior MR studies (Telomeres Mendelian Randomization Collaboration et al., 2017; Gao et al., 2020; Kuo et al., 2019). The MR Steiger test of directionality was used to confirm the assumption that the exposure causes the outcome is valid (Hemani et al., 2017). We also corroborated our findings by rerunning the analyses using data that had undergone Steiger filtering to remove SNPs that explained more variance in the outcome than in the risk factor. Genetic association data for measured telomere length were obtained from Codd et al., 2021, the largest GWAS of telomere length available through the OpenGWAS API at the time of analysis (N = 472,174, for more details, see Appendix 1).

All MR analyses were performed using R software version 4.0.2. Two sample MR analyses were conducted using the ‘TwoSampleMR’ package version 0.5.5. Meta-analyses of IVW results were performed using the ‘meta’ package version 4.18. GWAS meta-analyses used to perform single-SNP MR analyses were done using the METAL software (Willer et al., 2010). CAUSE analyses were conducted using the ‘cause’ package version 1.2.0. Forest plots were created using the ‘ggforestplot’ package version 0.1.0. LD Scores were computed using the ‘ldsc’ command line tool version 1.0.1. The code used in this study is available at: https://github.com/fernandam93/epiclocks_cancer.

We did not find strong evidence of causality between epigenetic age acceleration and breast cancer (GrimAge IVW OR = 0.98, 95% CI 0.95–1.00, p = 0.08; PhenoAge IVW OR = 0.99, 95% CI 0.98–1.01, p = 0.23; HannumAge IVW OR = 0.99, 95% CI 0.97–1.02, p = 0.63; and Intrinsic HorvathAge IVW OR = 0.99, 95% CI 0.98–1.00, p = 0.13) (Figure 2, Appendix 2—figure 1, Appendix 2—figure 2, Appendix 2—figure 3, Appendix 2—figure 4, Appendix 2—figure 5, Appendix 2—figure 6, Appendix 2—figure 7, Appendix 2—figure 8, Supplementary file 1 — Table s3, Supplementary file 1 — Table s4, Supplementary file 1 — Table s5, Supplementary file 1 — Table s6).

There was also limited evidence of causality between epigenetic age acceleration and ovarian cancer (GrimAge IVW OR = 0.99, 95% CI 0.93–1.06, p = 0.78; PhenoAge IVW OR = 0.98, 95% CI 0.96–1.01, p = 0.24; HannumAge IVW OR = 1.00, 95% CI 0.96–1.04, p = 0.95; and Intrinsic HorvathAge IVW OR = 1.00, 95% CI 0.97–1.02, p = 0.89) (Figure 2, Appendix 2—figure 1, Appendix 2—figure 2, Appendix 2—figure 3, Appendix 2—figure 4, Appendix 2—figure 5, Appendix 2—figure 6, Appendix 2—figure 7, Appendix 2—figure 8, Supplementary file 1 — Table s3, Supplementary file 1 — Table s4, Supplementary file 1 — Table s5, Supplementary file 1 — Table s6).

Meta-analysed IVW MR findings suggested that genetically predicted GrimAge acceleration decreased the risk of prostate cancer (OR = 0.93 per year increase in GrimAge acceleration, 95% CI 0.87–0.99, p = 0.02) (Figure 2, Supplementary file 1 — Table s3, Supplementary file 1 — Table s4, Supplementary file 1 — Table s5, Supplementary file 1 — Table s6). Although the direction of the genetically predicted effect was consistent across main and sensitivity MR analyses (i.e. MR-Egger, weighted median and weighted mode) (Appendix 2—figure 1, Supplementary file 1 — Table s3, Supplementary file 1 — Table s4, Supplementary file 1 — Table s5, Supplementary file 1 — Table s6), the main IVW result for GrimAge and prostate cancer did not withstand multiple testing correction (FDR p = 0.16) (Supplementary file 1 — Table s6).

We did not find consistent evidence of causality between other measures of epigenetic age acceleration and prostate cancer (PhenoAge IVW OR = 1.01, 95% CI 0.99–1.03, p = 0.31; HannumAge IVW OR = 0.98, 95% CI 0.95–1.02, p = 0.39; and Intrinsic HorvathAge IVW OR = 0.99, 95% CI 0.98–1.01, p = 0.42) (Figure 2, Appendix 2—figure 1, Appendix 2—figure 2, Appendix 2—figure 3, Appendix 2—figure 4, Appendix 2—figure 5, Appendix 2—figure 6, Appendix 2—figure 7, Appendix 2—figure 8, Supplementary file 1 — Table s3, Supplementary file 1 — Table s4, Supplementary file 1 — Table s5, Supplementary file 1 — Table s6).

Meta-analysed IVW MR findings suggested that genetically predicted Intrinsic HorvathAge acceleration decreased the risk of lung cancer (OR = 0.97 per year increase in Intrinsic HorvathAge acceleration, 95% CI 0.93–1.00, p = 0.03) (Figure 2, Supplementary file 1 — Table s3, Supplementary file 1 — Table s4, Supplementary file 1 — Table s5, Supplementary file 1 — Table s6). However, these results did not survive multiple testing correction (FDR p = 0.21) and were not strongly supported by sensitivity analyses (Appendix 2—figure 1, Supplementary file 1 — Table s3, Supplementary file 1 — Table s4, Supplementary file 1 — Table s5, Supplementary file 1 — Table s6).

We did not find evidence of causality between other measures of epigenetic age acceleration and lung cancer (GrimAge IVW OR = 1.00, 95% CI 0.91–1.09, p = 0.96; PhenoAge IVW OR = 0.97, 95% CI 0.94–1.00, p = 0.06; and HannumAge IVW OR = 0.99, 95% CI 0.95–1.04, p = 0.82) (Figure 2, Appendix 2—figure 1, Appendix 2—figure 2, Appendix 2—figure 3, Appendix 2—figure 4, Appendix 2—figure 5, Appendix 2—figure 6, Appendix 2—figure 7, Appendix 2—figure 8, Supplementary file 1 — Table s3, Supplementary file 1 — Table s4, Supplementary file 1 — Table s5, Supplementary file 1 — Table s6).

Meta-analysed IVW MR findings suggested that genetically predicted GrimAge acceleration increased the risk of colorectal cancer (OR = 1.12 per year increase in GrimAge acceleration, 95% CI 1.04–1.20, p = 0.002) (Figure 2, Supplementary file 1 — Table s3, Supplementary file 1 — Table s5, Supplementary file 1 — Table s6). These results survived multiple testing correction (FDR p = 0.04) and there was little evidence of heterogeneity across FinnGen and GECCO estimates (I² = 0%, 95% CI ‘NA’, p = 0.61). Additionally, the direction of the genetically predicted effect was consistent across main and sensitivity MR analyses (i.e. MR-Egger, weighted median, and weighted mode) (Figure 3, Supplementary file 1 — Table s3, Supplementary file 1 — Table s5, Supplementary file 1 — Table s6) and results were consistent when using UK Biobank data alone (IVW OR = 1.15, 95% CI 1.04–1.28, p = 0.007) (Figure 2, Supplementary file 1 — Table s4).

We did not find evidence of residual population stratification in MR analyses using negative control outcomes (Appendix 2—figure 9, Supplementary file 1 — Table s7), nor did we find evidence of horizontal pleiotropy via potential colorectal cancer risk factors (Appendix 2—figure 10, Supplementary file 1 — Table s8).

Single-SNP analysis revealed that the effect was not driven by a single SNP (Supplementary file 1 — Table s9). Figure 4 shows the effect of genetic instruments on GrimAge acceleration against their effect on colorectal cancer. Moreover, there was no detectable evidence of uncorrelated horizontal pleiotropy (MR-Egger intercept = –0.13, 95% CI –0.33–0.07, p = 0.33), or heterogeneity across individual SNP estimates (Cochran’s Q 7.12, p = 0.07) (Supplementary file 1 — Table s10). We further explored the genetically predicted effect of GrimAge on colorectal cancer using GECCO data only and found no evidence against bias due to correlated pleiotropy (CAUSE OR = 1.00, 95% credible intervals 0.96–1.04, p = 1.00; shared q = 4%, 95% credible intervals 0–24%) (Appendix 2—figure 11).

Among subtypes, we found strong evidence for a causal relationship between GrimAge acceleration and colon cancer (IVW OR = 1.15, 95% CI 1.09–1.21, p = 0.006). In contrast, we did not find such evidence for rectal cancer (IVW OR = 1.05, 95% CI 0.97–1.13, p = 0.24). After further stratification, the magnitude of the genetically predicted effect of GrimAge acceleration on colon cancer was the same for distal (IVW OR = 1.16, 95% CI 1.03–1.29, p = 0.01) and proximal colon cancer (IVW OR = 1.16, 95% CI 0.97–1.40, p = 0.11). Also, sex-stratified results suggest that GrimAge acceleration may influence colorectal cancer in both males (IVW OR = 1.12, 95% CI 1.00–1.25, p = 0.05) and females (IVW OR = 1.14, 95% CI 1.04–1.26, p = 0.008) (Figure 5, Supplementary file 1 Table s11).

These findings were further supported by evidence of a positive association between GrimAge acceleration and parental history of colorectal cancer (OR = 1.06, 95% CI 1.00–1.12, p = 0.03) (Figure 6, Supplementary file 1 — Table s12). Additionally, LD Score regression coefficients for GrimAge acceleration and colorectal cancer were also in the expected direction (GECCO rg = 0.28, p < 0.001; UK Biobank rg = 0.15, p = 0.21; FinnGen rg = 0.27, p = 0.29) (Appendix 2—figure 12, Supplementary file 1 — Table s13).

We did not find consistent evidence of causality between other measures of epigenetic age acceleration and colorectal cancer (PhenoAge IVW OR = 1.00, 95% CI 0.97–1.02, p = 0.73; HannumAge IVW OR = 1.02, 95% CI 0.97–1.08, p = 0.37; and Intrinsic HorvathAge IVW OR = 1.00, 95% CI 0.97–1.02, p = 0.79) (Figure 2, Appendix 2—figure 1, Appendix 2—figure 2, Appendix 2—figure 3, Appendix 2—figure 4, Appendix 2—figure 5, Appendix 2—figure 6, Appendix 2—figure 7, Appendix 2—figure 8, Supplementary file 1 — Table s3, Supplementary file 1 — Table s4, Supplementary file 1 — Table s5, Supplementary file 1 — Table s6).

In bidirectional MR analyses, we found evidence that genetically predicted GrimAge acceleration may be a cause of telomere shortening (IVW beta coefficient = −0.07 per year increase in GrimAge acceleration, 95% CI –0.09 to –0.05, p < 0.001) and that genetically predicted longer telomere length may increase Intrinsic HorvathAge acceleration (IVW beta coefficient = 0.57 per standard deviation increase in telomere length, 95% CI 0.39–0.77, p = 0.002) (Appendix 2—figure 13, Supplementary file 1 — Table s14).

Steiger filtering showed that all genetic instruments for GrimAge acceleration were stronger predictors of GrimAge acceleration than telomere length. In contrast, it identified 20 genetic instruments for telomere length that were better predictors of Intrinsic HorvathAge acceleration than telomere length (Supplementary file 1 — Table s15). After removing these SNPs from the analyses, the results were still suggestive of an effect of telomere length on Intrinsic HorvathAge acceleration (IVW beta coefficient = 0.71 per standard deviation increase in telomere length, 95% CI 0.57–0.85, p < 0.001) (Appendix 2—figure 14, Supplementary file 1 — Table s16).

There was little evidence of causality between other measures of epigenetic age acceleration and telomere length (Appendix 2—figure 13, Supplementary file 1 — Table s14).