Clonal hematopoiesis associated with epigenetic aging and clinical outcomes

Abstract Clonal hematopoiesis of indeterminate potential (CHIP) is a common precursor state for blood cancers that most frequently occurs due to mutations in the DNA‐methylation modifying enzymes DNMT3A or TET2. We used DNA‐methylation array and whole‐genome sequencing data from four cohorts together comprising 5522 persons to study the association between CHIP, epigenetic clocks, and health outcomes. CHIP was strongly associated with epigenetic age acceleration, defined as the residual after regressing epigenetic clock age on chronological age, in several clocks, ranging from 1.31 years (GrimAge, p < 8.6 × 10−7) to 3.08 years (EEAA, p < 3.7 × 10−18). Mutations in most CHIP genes except DNA‐damage response genes were associated with increases in several measures of age acceleration. CHIP carriers with mutations in multiple genes had the largest increases in age acceleration and decrease in estimated telomere length. Finally, we found that ~40% of CHIP carriers had acceleration >0 in both Hannum and GrimAge (referred to as AgeAccelHG+). This group was at high risk of all‐cause mortality (hazard ratio 2.90, p < 4.1 × 10−8) and coronary heart disease (CHD) (hazard ratio 3.24, p < 9.3 × 10−6) compared to those who were CHIP−/AgeAccelHG−. In contrast, the other ~60% of CHIP carriers who were AgeAccelHG− were not at increased risk of these outcomes. In summary, CHIP is strongly linked to age acceleration in multiple clocks, and the combination of CHIP and epigenetic aging may be used to identify a population at high risk for adverse outcomes and who may be a target for clinical interventions.


| INTRODUC TI ON
Aging is inextricably associated with an increase in the number of somatic mutations, and this process is believed to be central to the development of cancer (Blokzijl et al., 2016;Hoang et al., 2016;Martincorena & Campbell, 2015;Risques & Kennedy, 2018;Welch et al., 2012). Clonal hematopoiesis of indeterminate potential (CHIP) (Jaiswal et al., 2014) is defined by the presence of a cancer-associated somatic mutation in the blood cells of people without a blood cancer or other known clonal disorder. CHIP originates when hematopoietic stem cells (HSCs) acquire a random mutation, usually in an epigenetic factor, that results in increased clone fitness (Jaiswal & Ebert, 2019). CHIP is strongly associated with age, and carriers of these mutations have an increased risk for developing blood cancers, but also coronary heart disease (CHD) and all-cause mortality (Jaiswal et al., 2014(Jaiswal et al., , 2017. In addition to age, CHIP has been found to occur at a higher prevalence in males (Jaiswal et al., 2014) and a lower prevalence in people of self-reported Hispanic and East Asian ancestry compared to Europeans (Bick, Weinstock, et al., 2020;Jaiswal et al., 2014). The association of CHIP and heart disease may result from enhanced inflammatory gene expression in mutant macrophages within atherosclerotic plaques (Bick, Pirruccello, et al., 2020;Fuster et al., 2017;Jaiswal et al., 2017), demonstrating that at least some of these mutations cause dysfunction of immune cells and affect phenotypes apart from cancer.
The availability of DNA-methylation data from large epidemiological cohorts has advanced our understanding of epigenetic aging in recent years. Several "methylation clocks" have been developed (Hannum et al., 2013;Horvath, 2013;Levine et al., 2018;) that use methylation state at a subset of CpGs to predict chronological age with high accuracy in healthy individuals. "Age acceleration" results when predicted methylation age is greater than chronological age and associates with increased risk of CHD (Levine et al., 2018;Perna et al., 2016) and all-cause mortality Christiansen et al., 2016;Levine et al., 2018;Lu, Quach, et al., 2019, p. 201;Marioni et al., 2015;Perna et al., 2016). Similar to prior studies , we defined age acceleration as the residual of a linear model of a clock estimate regressed against chronological age. By definition, this measure is not correlated with chronological age and a positive (or negative) value indicates that the clock age is higher (or lower) than expected based on chronological age.
The factors underlying epigenetic age acceleration are incompletely understood. Recent work has noted that two distinct categories of epigenetic clocks, intrinsic and extrinsic, which are believed to capture different aspects of aging. Intrinsic aging is independent of cell type and may be partly driven by the number of times a cell has divided , while extrinsic aging, is associated with changes of cell type composition in blood , and maybe influenced by environmental factors (Levine et al., 2018;. The Horvath and IEAA clocks reflect intrinsic aging, whereas the Hannum, EEAA, PhenoAge, and GrimAge clocks are measures of extrinsic aging (Table 1). GrimAge and PhenoAge were also trained to be predictors of mortality (Levine et al., 2018;. In addition, several DNA methylation-based predictors of other aging-related phenotypes have recently been developed to improve mortality prediction, such as surrogate biomarkers for plasma protein levels (adrenomedullin, beta-2-microglobulin, cystatin C, leptin, plasminogen activator inhibitor 1, tissue inhibitor matrix metalloproteinase 1) , smoking pack years , and telomere length (Lu, Seeboth, et al., 2019).
We hypothesized that CHIP may be an acquired genetic factor associated with epigenetic age acceleration. Here, we use wholegenome sequencing (WGS) and DNA-methylation array data from  (Bick, Pirruccello, et al., 2020;Bick, Weinstock, et al., 2020) (see Table S1 for a demographic summary of cohorts).  , intrinsic epigenetic age acceleration (IEAA)  and extrinsic epigenetic age acceleration (EEAA) , and a methylation-based estimate of telomere length (DNAmTL) (see Methods). The effects of CHIP were assessed overall (any CHIP mutation), as well as at the level of specific classes of CHIP mutations (see Methods).
We also found modest associations between CHIP and several epigenetic surrogate markers of plasma proteins as well as blood counts (Table S3A, B), and between clock estimates and variant allele fraction (VAF), which is an approximation of clone size (Table S4).
Methylation data can also be used to estimate a surrogate marker of leukocyte telomere length (LTL), DNAmTL (Lu, Seeboth, et al., 2019).

| Gene-specific associations of CHIP with epigenetic age acceleration
Clonal hematopoiesis of indeterminate potential most commonly occurs due to mutations in genes coding for the DNA methylationaltering enzymes DNMT3A and TET2, but can also arise due to mutations in ASXL1, JAK2, splicing factors, and DNA-damage response (DDR) genes. Accordingly, we examined the associations of mutations in specific CHIP genes with age acceleration (Table 2). In all clocks, the direction of association for DNMT3A and TET2 mutations was the same, although those with TET2 mutations had significantly greater age acceleration than those with DNMT3A mutations for Hannum (2.10 years, p < 0.0012) and EEAA (2.32 years, p < 0.0063), but not other clocks. We also performed differential methylation analysis to assess whether mutations in the DNA-methylation modifying enzymes DNMT3A and TET2 had divergent effects at the clock CpGs. Mutations in both genes primarily resulted in hypomethylation although a small number of CpGs showed hypermethylation in TET2 ( Figure S4A,B). We also observed at the clock CpGs that the F I G U R E 1 CHIP is associated with increased age acceleration. Forest plot of the effect sizes and confidence intervals for the effect of CHIP on age acceleration estimate from seven methylation clocks M-values (a log-transformed measure of the percent methylation at each site) in persons with DNMT3A and TET2 mutations were highly correlated ( Figure S4C), indicating that the methylation state of persons with the two mutations is largely similar, despite their opposing enzymatic effects.
Persons with mutations in multiple genes had the largest increases in age acceleration across all clocks except PhenoAge, consistent with our observation that age acceleration increases with the number of mutations. Conversely, no increase in age acceleration was observed in persons with mutations in DDR genes (TP53, PPM1D, BRCC3), which is consistent with the lack of association with age acceleration observed for the same mutations in cancer tissue samples (Horvath, 2013). Although we had only eight individuals with JAK2 mutations in our cohort, these mutations showed an exceptionally strong association for a single mutation in several clocks, the most extreme example being PhenoAge (10.01 years, p < 9.7 × 10 −6 ). The PhenoAge clock was trained to predict a composite measure of mortality risk which includes several hematological variables such as white blood cell count, white blood cell differential, and several red blood cell parameters which may be abnormal in myeloproliferative neoplasm, a hematological malignancy which is strongly associated with JAK2 mutations. CHIP overall was nominally associated with estimated pack years of smoking (DNAmPACKYRS), but only mutations in ASXL1 were significantly associated with this measure in a gene-specific analysis (7.54 pack years, p < 0.002), a finding that is in accordance with a recent report (Bolton et al., 2020) (Table S3).

| Association of CHIP and epigenetic age acceleration with clinical outcomes
Several previous studies have linked both CHIP (Jaiswal et al., 2014(Jaiswal et al., , 2017 and age acceleration in some clocks (Levine et al., 2018; to increased risk of adverse clinical outcomes, in particular all-cause mortality and CHD. We asked whether the combination of CHIP and age acceleration could further stratify carriers of CHIP into high-risk and low-risk groups for these outcomes We defined a person to have "age acceleration" (AgeAccel) for a clock if their values for an age acceleration residual exceeded zero after adjustment for age at blood draw, sex, self-reported ancestry, and study cohort. We then tested the interaction between this dichotomous variable and CHIP status in predicting mortality in each of the seven clocks using Cox models. As shown in Table S5, we found that the most significant interactions were for the Hannum and GrimAge clocks, although neither reached Bonferroni-corrected statistical significance. Though both the Hannum and GrimAge clocks were predictive of time to death or CHD in previous studies Marioni et al., 2015;Perna et al., 2016), they were trained on different outcomes (age for Hannum versus time to death for GrimAge), and are not strongly correlated in our dataset (bicor = 0.242, R 2 = 0.058, Figure 2A). Therefore, we reasoned that To validate this finding, we sought replication in an independent cohort, the BA23 subset of WHI, which was not used in the above mortality analysis . When we modeled the interaction of CHIP with AgeAccelHG for CHD in BA23, the interaction term was again significant (CHIP main effect: coefficient = −0.24, p < 0.60; AgeAccelHG main effect: coefficient = 0.24, Having demonstrated a significant statistical interaction between CHIP and AgeAccelHG for clinical outcomes, we combined these two variables into a single, 4-factor variable for further modeling. For CHD, we included incident events in FHS, JHS, and WHI EMPC together with WHI BA23 as a meta-analysis. Persons who were CHIP+/AgeAccelHG+ had much greater risk of all-cause mortality (HR 2.90, p < 4.1 × 10 −8 ) and CHD (HR 3.24, p < 9.3 × 10 −6 ) compared to those who were CHIP−/AgeAccelHG−. Those who were CHIP−/ AgeAccelHG+ had a more modest increase in risk of all-cause mortality (HR 1.66, p < 3.1 × 10 −9 ), and CHD (HR 1.39, p < 0.012) compared to those who were CHIP−/AgeAccelHG−. In contrast, those who were CHIP+/AgeAccelHG− did not have elevated risk of either all-cause mortality (HR 0.78, p < 0.20) or CHD (HR 1.03, p < 0.93) compared to those who were CHIP−/AgeAccelHG− ( Figure 2B,C).
We also fitted contrasts to estimate the hazard ratios for all-cause mortality and CHD for CHIP only in persons with AgeAccelHG+ and AgeAccelHG+ only in persons with CHIP, in both cases finding the associations to be significant ( Figure S5).
We also asked if there were gene-level differences in risk of these outcomes. We had insufficient sample size to assess either mortality or CHD individually, so we combined the two into a composite outcome. Being AgeAccelHG+ increased the risk of the composite outcome for those with TET2 mutations relative to those who were CHIP−/AgeAccelHG− (TET2 mu-tated+/AgeAccelHG+: HR = 3.88, p < 1.6 × 10 −6 ; TET2 mutated+/ Our data permitted us to also ask whether there was an association of CHIP and AgeAccelHG to time to death in those who had a first CHD event, a subgroup that is often the target of clinical interventions. We restricted our analysis to individuals who had a first CHD event after age 70 and, if they died, did so more than 30 days after the CHD event. We found a significant interaction between CHIP and AgeAccelHG for all-cause mortality after CHD (p < 0.036).
Given the previous findings linking both CHIP (Jaiswal et al., 2017) and extrinsic epigenetic aging Levine et al., 2018; to inflammation, we asked whether plasma levels of the inflammation marker high-sensitivity C-reactive protein (hs-CRP) showed any evidence of interaction with CHIP for all-cause mortality or CHD. We found evidence for a main effect of hs-CRP on risk for all-cause mortality, but not for an in-  Note: Table with effect sizes, standard errors, and p-values for eight different classes of CHIP mutations. "Multiple" means mutations in multiple genes. "DDR" refers to mutations in the DNA damage response genes TP53, PPM1D, and BRCC3. "Splicing factor" are mutations in SF3B1, SRSF2, U2AF1, ZRSR2, and PRPF8. "Other" refers to mutations in all other genes not listed. levels were above 2 mg/L, an established clinical cutoff. Individuals with CHIP and AgeAccelHG showed a similar risk of all-cause mortality and CHD regardless of whether they had high or low hs-CRP levels ( Figure S5E,F). These results indicate that hs-CRP is a poor discriminator of risk in CHIP carriers, unlike AgeAccelHG.
A coding SNP in IL6R (rs2228145), which results in Asp358Ala, was previously found to attenuate the increased risk for mortality and CHD associated with CHIP (Bick, Pirruccello, et al., 2020;Bick, Weinstock, et al., 2020 p < 0.36). There were also no significant interactions between rs2228145 genotype and the combined CHIP/AgeAccelHG variable ( Figure S5C,D). These results indicate that IL6R genotype is a poor discriminator of risk in CHIP carriers in this dataset, unlike AgeAccelHG. However, we did find differences based on which gene was mutated. Those who were TET2-CHIP+/AgeAccelHG+ and with no alternate alleles of rs2228145 (IL6RWT) had the highest risk for the composite mortality/CHD outcome relative to the referent group of CHIP−/AgeAccelHG−/IL6RWT (HR =11.3, p < 2.4 × 10 −21 , Figure S6). Those who were TET2-CHIP+/ AgeAccelHG+ but carried 1 or 2 alternate alleles of rs2228145 (IL6RMut) had substantially lower risk (HR = 1.91 compared to the same referent group, p < 0.066; coefficient for interaction = −1.12 per alternate allele, p for interaction < 9.6 × 10 −7 , Figure S6). There was no significant difference in risk based on rs2228145 genotype in those who were TET2-CHIP+/AgeAccelHG−. We also did not find significant differences in risk of death/CHD by rs2228145 genotype in DNMT3A-CHIP or CHIP with other non-DDR mutations regardless of AgeAccelHG status.

| DISCUSS ION
The results presented here permit us to draw several conclusions. First, it is clear that CHIP is strongly associated with epigenetic aging in several clocks. Consistent with the results from a shared genetic architecture, as evidenced by the overlapping GWAS hits between polymorphisms near TERT and TRIM59 that associate with both CHIP and IEAA (Bick, Pirruccello, et al., 2020;Bick, Weinstock, et al., 2020;Zink et al., 2017). However, the heritability of CHIP appears to be low (3. 6% Bick, Pirruccello, et al., 2020;Bick, Weinstock, et al., 2020), which limits our ability to test for genetic correlation between CHIP and age acceleration.
Previous studies have shown that IEAA of cultured fibroblasts strongly correlates with the number of population doublings . Therefore, an alternative hypothesis is that the increase in intrinsic age acceleration seen in CHIP carriers may be due to either (1) increased proliferation or self-renewal of HSC clones that harbor these mutations or (2) (Wolf et al., 2018), insomnia (Carroll et al., 2017), and hunter-gatherer lifestyle . Our results may also explain why the strength of the associations between CHIP and mortality or CHD are somewhat inconsistent across studies-while the prevalence of CHIP is fairly uniform across populations, epigenetic aging may not be. In populations with a high prevalence of risk factors for epigenetic aging, the consequences of CHIP may be direr than in populations without such risk factors.
Our risk stratification schema may also be used to select patients for clinical trials of pharmaceutical or behavioral interventions, as the benefit-to-risk ratio may be particularly favorable in the highrisk CHIP group. We note that that the 5-year mortality after CHD in those who are CHIP+ and AgeAccelHG+ approaches 60%, similar to the mortality seen in patients with intermediate-risk MDS (Greenberg et al., 2012). Furthermore, the high event rate in this group would enable smaller trials with sufficient power for detecting favorable outcomes such as reduced all-cause mortality or time to CHD. One such intervention may be blockade of IL-6 receptor (Bick, Pirruccello, et al., 2020;Bick, Weinstock, et al., 2020); our results show that those who are TET2-CHIP+ and AgeAccelHG+ have lower risk of death or CHD with increasing copies of rs2228145, which has previously been linked to reduced IL-6R expression levels in myeloid cells (Bick, Pirruccello, et al., 2020;Bick, Weinstock, et al., 2020).
Alternatively, this group may benefit from IL-1B inflammatory blockade (Ridker et al., 2017), which has also been shown to be relevant to atherosclerosis in model systems of CHIP Jaiswal et al., 2017). Of note, AgeAccelHG appears to be superior to hs-CRP and genotype at IL6R for risk discrimination of CHIP carriers, implying that it is capturing additional information beyond baseline inflammation.
In sum, our results show that there is an important relationship between CHIP and epigenetic aging. CHIP and epigenetic age acceleration can also be used to identify persons at high risk of all-cause mortality and CHD, further reinforcing the importance of both phenotypes as valuable tools in precision medicine for aging.

| Epidemiological cohorts
All participant data were obtained from four independent patient cohorts: the FHS (Feinleib et al., 1975), the JHS (Sempos et al., 1999), the WHI (phs000200.v11.p3), and the MESA (Bild, 2002, p. 200). These cohorts were included in the TOPMed consortium which is run by the National Heart Lung and Blood Institute of the National Institutes of Health. Access to all data was approved by TOPMed as well as the individual cohorts. We included only those persons from these cohorts in which the age at draw for both whole blood methylation and WGS were available. In the FHS and JHS cohorts, the samples for methylation and WGS were taken from the same blood draw in all persons. In MESA, methylation data were only used from the first exam as this was the time at which DNA for WGS was also collected. In the WHI cohort, the two samples were often taken from different times. We only considered persons for whom the methylation and WGS samples were taken within 3 years of each other.

| Methylation array data
Whole blood methylation was quantified using the Illumina MethylationEPIC or HumanMethylation450k array. Normalized methylation data were submitted to the online methylation clock tool (https://dnama ge.genet ics.ucla.edu/new) which generates methylation age estimates for seven different clocks: DNAmAge (Horvath, 2013), DNAmHannum (Hannum et al., 2013), DNAmPhenoAge (Levine et al., 2018), DNAmSkinClock , DNAmGrimAge , intrinsic epigenetic age acceleration (IEAA) , and extrinsic epigenetic age acceleration (EEAA) . Age acceleration was computed for each measure as the residual of model predicting each persons' methylation age from their chronological age at the time of blood draw. Additionally, the DNAmGrimAge clock generates seven surrogate biomarkers based upon blood protein expression (MADM/NRBP1, B2 M, CST3 (Cystatin C), GDF15, LEP (Leptin), SERPINE1/PAI1, and TIMP1) as well smoking pack years. Age-adjusted LTL and unadjusted LTL are also estimated by the clock software (Lu, Seeboth, et al., 2019). Cell composition was also estimated by the clock software using a published model (Houseman et al., 2012).
Somatic mutations associated with CHIP were called from WGS data using the Mutect2 module in GATK from BAM files previously aligned with BWA. Candidate CHIP variants were selected based upon a curated list of known variants recurrently mutated in hematological malignancies as previously described (Jaiswal et al., 2017) (see Table S6). A full list of variants identified in this study are included in Table S7.

| Association between CHIP and methylation age acceleration
Clonal hematopoiesis of indeterminate potential status was associated with age acceleration and the other measures using linear modeling, with a separate model being fitted for each aging measure.
Because of the relatively small number of comparisons, p-values for these analyses were reported unadjusted. We combined the data from all three studies and used residualization to remove the effects of age, race/ethnicity, sex, and study. This approach was chosen to eliminate any possibility of spurious associations between CHIP and the methylation measures that were driven by collinearity between CHIP and covariates. The residualized methylation measure was the outcome in each model, and a likelihood ratio test was performed to test the significance of CHIP as predictor against a null model containing only the intercept. When testing the association of CHIP mutations with specific genes, CHIP status was replaced with a cat- single mutation into one group, and split the group with mutations in multiple genes into two mutations and greater than two mutations, regardless of which genes were mutated. Correlation between VAF and the residualized methylation measures was computed using biweight midcorrelation, an outlier resistant alternative to Pearson's correlation (Horvath, 2011).

| Differential methylation of clock CpGs
Illumina HumanMethylation450K  measures. The adjusted residuals were tested for differential methylation and p-values were corrected for the number of CpGs tested using limma (Ritchie et al., 2015).

| Association of CHIP and epigenetic age acceleration with clinical outcomes
We tested the associations of CHIP and epigenetic age acceleration with all-cause mortality and incident CHD with Cox proportional hazards models using the survival package in R. Models included age, sex, race/ethnicity, systolic blood pressure, type 2 diabetes status, plasma LDL-cholesterol concentration, plasma HDL-cholesterol concentration, plasma triglyceride concentration, and smoking status as covariates. Some persons in WHI had DNA for the methylation and/ or WGS sample obtained several years after the baseline visit, which potentially could introduce survivorship bias into the analysis. For this reason, we also excluded anyone in WHI for whom either the methylation or WGS blood draw occurred more than 5 years after the baseline visit.
For analysis of all-cause mortality, pooled data from FHS, JHS, and WHI EMPC were used. The selection of samples used in TOPMed in these cohorts were taken essentially at random from the larger parent cohorts. WHI BA23 was excluded from this analysis because persons in this cohort were over-sampled for CHD. MESA was excluded from this analysis because persons in this cohort were selected for surviving at least 10 years from baseline. We chose to present the results from models in which all three cohorts were pooled, rather than analyzed separately and then meta-analyzed.
For the analysis of CHD, the WHI BA23 cohort was analyzed separately, and a meta-analysis was used to combine the results of the BA23 analysis with the other pooled cohorts (JHS, FHS, and WHI EMPC) to get the final effect size estimates. 45 persons in WHI BA23 were also included in the mortality analysis of WHI EMPC but were not included in the CHD analysis of WHI EMPC (i.e., were not double-counted). Because BA23 was over-sampled for CHD, we adjusted the sample weights in BA23 using race and incident CHD For the gene-level analyses, persons with any singleton DNMT3A, TET2, or DDR gene (TP53, PPM1D, BRCC3) mutation were considered to be in those classes. All other non-DNMT3A, TET2, and DDR mutations were considered "other." In those with multiple mutations, the mutated gene with the highest VAF was used to assign the class.
For the analysis of cumulative incidence of death and CHD, the cmprsk package in R was used.

DATA AVA I L A B I L I T Y S TAT E M E N T
The data that support the findings of this study are available from Trans-Omics for Precision Medicine (TOPMed) consortium. Restrictions apply to the availability of these data, which were used under license for this study.
Data are available at https://www.nhlbi wgs.org/topme d-data-acces sscien tific -commu nity with the permission of TOPMed Data Coordinating Center. All data used in the study are available at the follow DbGaP acces-