Schizophrenia and Epigenetic Aging Biomarkers: Increased Mortality, Reduced Cancer Risk, and Unique Clozapine Effects

Abstract

Background:

Schizophrenia (SZ) is associated with increased all-cause mortality, smoking, and age-associated proteins, yet multiple previous studies found no association between SZ and biological age using Horvath’s epigenetic clock, a well-established aging biomarker based on DNA methylation (DNAm). However, numerous epigenetic clocks have been developed that may capture distinct aspects of aging. This study tested the hypothesis that altered aging in SZ manifests in these other clocks.

Methods:

We performed a comprehensive analysis of 14 epigenetic clocks, categorized according to what they were trained to predict: chronological age, mortality, mitotic divisions, or telomere length. To understand the etiology of biological age differences, we also examined DNAm predictors of smoking, alcohol, BMI, serum proteins, and cell proportions. We independently analyzed three publicly available multi-ethnic DNAm datasets from whole blood, a total of 567 SZ cases and 594 non-psychiatric controls.

Results:

All data sets showed accelerations in SZ for the three mortality clocks up to 5 years, driven by smoking and elevated levels of 6 age-associated proteins. The two mitotic clocks were decelerated in SZ related to anti-tumor NK and CD8T cells, which may help explain conflicting reports about low cancer rates in epidemiological studies of SZ. One cohort with available medication data showed clozapine is associated with male-specific decelerations up to 7 years in multiple chronological age clocks.

Conclusions:

Our study demonstrates the utility of studying the various epigenetic clocks in tandem and highlights potential mechanisms by which mental illness influences long-term outcomes including cancer and early mortality.

Introduction

Life expectancy in schizophrenia (SZ) is reduced by 15 years (1), and a meta-analysis of 37 studies found a standardized mortality hazard ratio of 2.5 in SZ, with excess mortality in nearly all categories of natural causes (2). This may be explained by observations that SZ is associated with changes in age-related biomarkers of inflammation, oxidative stress, and metabolism (3). Serum proteins that increase with age and predict mortality (4) are also increased in SZ including PAI-1 (5,6), TIMP-1 (6), GDF-15 (7), Cystatin C (8), beta-2 microglobulin (9,10), and adrenomedullin (11). Thus, it has been proposed that schizophrenia is a syndrome of accelerated aging (12,13).

However, multiple studies using Horvath’s DNA methylation (DNAm)-based epigenetic clock, a well-recognized and validated biomarker of chronological age, have found no evidence for the accelerated aging hypothesis of SZ in either brain or blood (14–18). A recent study observed a null association with all-cause mortality in SZ (18). The epigenetic clock was constructed using a penalized regression model to select a weighted average of 353 CpG sites, which produces very accurate chronological age predictions across all tissues and across the entire lifespan (19). Increased or decreased epigenetic age relative to chronological age is termed age acceleration or deceleration, respectively. Age acceleration predicts increased risk of age-related diseases like Alzheimer’s disease, cancer, and cardiovascular disease (20). Epigenetic clocks are influenced by substance use, diet, exercise, obesity, and socioeconomic status (20,21). Schizophrenia is associated with all these factors, especially high smoking rates and metabolic side effects of antipsychotic medications (22–27). Thus, it is surprising that epigenetic clock analyses have not identified accelerated aging in schizophrenia. Our current understanding of epigenetic aging in schizophrenia may be incomplete, and important confounding variables such as medications and smoking have not yet been sufficiently accounted for.

Numerous epigenetic clocks have recently been developed that appear to capture distinct aspects of aging. Some predict chronological age similar to Horvath’s clock (19,28–34). Others predict age-related morbidity and mortality (4,35,36), mitotic divisions (37,38), or telomere length (39). There are also DNAm biomarkers validated in blood for smoking, alcohol, BMI, and serum levels of age-associated proteins dysregulated in schizophrenia (4,40). Various clocks have age acceleration values (i.e. after adjusting for chronological age) correlated at r < 0.5, and show very little overlap in their CpG composition (41). The clocks differ in their associations with gene expression, lifestyle factors and outcomes of aging (20,41). Whether these other clocks are altered in schizophrenia is unknown. Here, we systematically investigate 14 clocks in three multi-ethnic SZ blood DNAm data sets.

Methods and Materials

Sample

Analyses utilized three Illumina 450K DNAm data sets from whole blood: UCL (University College London, GSE80417) (42), ABD (University of Aberdeen, GSE84727) (42), and CF (Chinese Famine, GSE116379) (43). We analyzed a total of 567 SZ cases and 594 non-psychiatric controls (NPC) with ages ranging from 30 to 90 (Table 1). All three data sets included information on gender and SZ status. CF consists of subjects who were conceived during or after the 1959-1961 Chinese Famine as identified by the University in Changchun. Thus, CF has a narrower age range (44 to 51) and included data on medications obtained through a structured interview, famine exposure during the first 3 months of gestation based on birth dates, and city versus rural background. See Supplementary Methods for additional information.

Table 1. Cohort characteristics.

Table S1 shows information for these data sets before filtering out individuals with age less than 30. Figure S1 shows age histograms for each data set. Table S2 (separate file) shows subject-level data for CF. Abbreviations: NPC (non-psychiatric control), SZ (schizophrenia), UCL (University of College London), ABD (University of Aberdeen), CF (Chinese Famine), AP (antipsychotics).

Variable	UCL NPC	UCL SZ	ABD NPC	ABD SZ	CF NPC	CF SZ
Number	173	278	342	215	79	74
Sex	83F, 90M	80F, 198M	78F, 264M	72F, 143M	48F, 31M	29F, 45M
Age Range (yr)	30-87	30-90	30-66.4	30-80.7	44.6-51.0	45.0-50.9
Median Age (yr)	44	45	48.6	47.1	47.5	47.3
Mean Age (yr)	46.0	47.3	48.4	48.2	47.9	47.7
Famine					25	23
City					74	45
Antidepressants					0	8
Typical AP					0	36
Atypical AP					0	18
Clozapine					0	35

Epigenetic Clocks

Epigenetic clocks and other biomarkers (Table 2) were calculated according to published methods (4,19,28–40). Those not normalized to provide an age estimate (or other meaningful value like telomere length) were transformed into units of standard deviation. For each clock, we calculated two values of age acceleration (19,20): 1) an age residual resulting from a linear model regressing epigenetic age on chronological age and 2) a cell composition corrected age residual resulting from regressing epigenetic age on chronological age and 7 cell types as previously described (44). Both models were developed in NPC only, then applied to all subjects. Thus, average age acceleration for NPC is zero by construction. Further analyses utilized these acceleration values for biomarkers with clear age trends. We could not assume normality of distributions, so we utilized the non-parametric Kruskal-Wallis test for univariate statistics and generalized linear models for multiple regressions. See Supplementary Methods for additional information.

Table 2.

Epigenetic clocks and other DNAm-based biomarkers

Denotation	1st Author, Year	Trained Phenotype	#CpGs	Tissues derived
GrimAge	Lu, 2019a (4)	Mortality	1030	Whole blood
PhenoAge	Levine, 2018 (35)	Mortality, phenotypic age	513	Whole blood
Zhang	Zhang, 2017 (36)	Mortality	10	Whole blood
MiAge	Youn, 2018 (37)	Mitotic divisions	268	8 cancer types, adjacent tissue
Yang	Yang, 2016 (38)	Mitotic divisions	385	Whole blood
DNAmTL	Lu, 2019b (39)	Telomere length	140	DNAmTL
Horvath1	Horvath, 2013 (19)	Chronological age	353	51 tissues/cells
Horvath2	Horvath, 2018 (33)	Chronological age	391	Skin, fibroblasts, buccal, cord blood
Vidal-Bralo	Vidal-Bralo, 2016 (31)	Chronological age	8	Whole blood
Lin	Lin, 2015 (34)	Chronological age	99	Whole blood
Hannum	Hannum, 2013 (32)	Chronological age	71	Whole blood
Weidner	Weidner, 2014 (30)	Chronological age	3	Whole blood
Garagnani	Garagnani, 2012 (29)	Chronological age	1	Whole blood
Bocklandt	Bocklandt, 2011 (28)	Chronological age	1	Saliva
GDF-15	Lu, 2019a (4)	GDF-15 serum protein	137	Whole blood
ADM	Lu, 2019a (4)	ADM serum protein	186	Whole blood
B2M	Lu, 2019a (4)	B2M serum protein	91	Whole blood
PAI-1	Lu, 2019a (4)	PAI-1 serum protein	211	Whole blood
TIMP-1	Lu, 2019a (4)	TIMP-1 serum protein	42	Whole blood
Cystatin C	Lu, 2019a (4)	Cystatin C serum protein	87	Whole blood
Leptin	Lu, 2019a (4)	Leptin serum protein	187	Whole blood
Smoking1	Lu, 2019a (4)	Smoking pack-years	172	Whole blood
Smoking2	McCartney, 2018 (40)	Smoking	233	Whole blood
Alcohol	McCartney, 2018 (40)	Alcohol	450	Whole blood
BMI	McCartney, 2018 (40)	BMI	1109	Whole blood

Results

Mortality, mitotic, and telomere length clocks are altered in SZ

Previous studies using the UCL and ABD data sets have shown no difference in the Horvath1 chronological age clock between SZ and NPC (16,17). We replicated their results (Figures 1A, S2A). We examined if altered aging manifests in clocks other than Horvath1 in same data sets, as well as in the Chinese Famine (CF) data set.

Figure 1.

Aging clocks other than Horvath1 show changes in schizophrenia (SZ) compared to non-psychiatric controls (NPC): (A) The Horvath1 chronological age clock, (B-D) the GrimAge, PhenoAge, and Zhang mortality clocks, (E-F) the MiAge and Yang mitotic clocks, (G) predicted telomere length, (H) predicted smoking pack-years. Here we show results from UCL (NPC, n = 173; SZ, n = 278). The other two data sets showed the same patterns (Figures S2–3). Complete statistics are in Table S4. For each panel, we show chronological age compared to clock score (left), as well as average age accelerations for SZ compared to NPC (right). Age accelerations are defined as residuals from the linear model for NPC (blue line), and therefore average age acceleration for NPC is 0. The first model is adjusted only for age. The second model is adjusted for age and derived cell counts (see Supplementary Methods). Error bars are SEM. More advanced biological age is indicated by positive acceleration values for most clocks, but by negative acceleration values for telomere length. Statistical significance of difference from NPC is shown: <0.2#, <0.05*, <0.01**, <0.001***.

The GrimAge, PhenoAge and Zhang mortality clocks were trained by regressing time-to-death on DNAm, either directly or using an intermediate measure (4,35,36). They are more sensitive to mortality than Horvath1. We found age acceleration for schizophrenia (SZ) in all mortality clocks compared to non-psychiatric controls (NPC) in all data sets (Panels B-D in Figures 1, S2, S3; Table S4). GrimAge showed accelerations in SZ of 5.8 years (UCL; p < 0.0001), 4.2 years (ABD; p < 0.0001) and 2.5 years (CF; p = 0.0003). PhenoAge was accelerated in SZ by 1.4 years (UCL; p = 0.0127), 1.9 years (ABD; p < 0.0001), and 1.5 years (CF; p = 0.0912). Zhang is not normalized to give age estimates, so we use units of standard deviation (SD). Zhang was increased in SZ by 0.66 SD (UCL), 0.61 SD (ABD) and 0.75 SD (CF), with p < 0.0001 for all.

The MiAge and Yang mitotic clocks use mathematical models of DNA methylation changes to estimate mitotic age, or the total number of lifetime cell divisions (37,38). They are accelerated in diverse cancers, high-turnover tissues, and carcinogen-exposed tissues. They were correlated to the same extent both before and after adjusting for chronological age (R² = 0.649 (UCL) to 0.833 (CF)) (Figure S4). We found deceleration for SZ in both mitotic clocks in all three data sets (Panels E-F in Figures 1, S2, S3; Table S4). MiAge was decelerated by 0.27 SD (UCL; p = 0.0017), 0.44 SD (ABD; p < 0.0001), and 0.22 SD (CF; p = 0.1585). Yang was decelerated by 0.12 SD (UCL; p = 0.1918), 0.39 SD (ABD; p < 0.0001), and 0.42 SD (CF; p = 0.0085).

The DNAm estimator of telomere length (DNAmTL) (39) was reduced in SZ in UCL (42 base pairs, p = 0.0067) and ABD (34 base pairs, p = 0.0063) (Panel G of Figure 1, S2, Table S4). Since telomere length decreases with age, we considered DNAmTL to be accelerated in SZ, consistent with prior reports on SZ telomere length (45).

After adjusting for gender, these differences in various clocks were all maintained. In fact, SZ showed more significant effects after adjusting for gender, particularly for the mitotic clocks (Table S5). SZ males consistently showed mitotic clock deceleration in all data sets. SZ females showed significant deceleration in UCL, a deceleration trend in ABD, and no deceleration in CF (Table S5, Figure S5).

When we adjusted for cell composition, mortality clock and telomere length accelerations were slightly smaller but remained statistically significant (Figures 1, S2, S3; Tables S4, S5). However, mitotic clock deceleration in SZ was greatly reduced after correcting for cell composition, drawing attention to a relationship between SZ, mitotic age, and cell proportions.

SZ mortality clock acceleration is related to smoking and age-associated proteins

Because mortality clocks are accelerated by smoking (4,35,36), and smoking rates are much higher in SZ (25), we hypothesized smoking was a major driver of SZ mortality clocks. The DNAm-based predictor of smoking pack-years (PY) by Lu et al. (4) was increased in SZ by 15.9 PY (UCL, p < 0.0001), 10.5 PY (ABD, p < 0.0001), and 5.3 PY (CF, p = 0.007), independent of gender and cell composition (Figures 1H, S2H, S3H; Tables S4, S6). Another smoking predictor from McCartney et al. (40) shows the same pattern, and the two independently derived biomarkers are highly correlated (Figure S6).

We also examined DNAm-based predictors of alcohol and body mass index (BMI) (Tables S4, S6, Figures 2, S7–9) (40). Alcohol was positively associated with smoking. However, despite higher smoking scores in SZ, alcohol use was reduced in SZ in ABD and CF. After correcting for smoking and gender, alcohol scores were lower in SZ in all data sets, by 0.22 SD (UCL, p = 0.0468), 0.49 SD (ABD, p < 0.001), and 0.67 SD (CF, p < 0.001) (Table S6). BMI was negatively correlated with smoking. BMI was reduced in SZ only in CF, by 0.47 SD (p = 0.0034), independent of gender, smoking, and alcohol.

Figure 2.

DNA methylation biomarkers in SZ and NPC: (A-B) Alcohol and BMI, (C-I) serum protein levels, and (J-L) three cell types with altered proportions in SZ. Here we show results from UCL (NPC, n = 173; SZ, n = 278). The other two data sets showed similar patterns (Figures S7–8). Complete statistics are in Table S4.

Multiple regression analysis showed smoking was the major factor contributing to GrimAge and Zhang acceleration (Table S7). Each PY increase was associated with ~0.3 years accelerated GrimAge, and smoking alone accounted for 82% (UCL), 75% (ABD), and 47% (CF) of SZ’s influence on GrimAge. Each PY increase was associated with a 0.03-0.06 SD increase in Zhang acceleration, and smoking accounted for 70% (UCL), 97% (ABD), 13% (CF) of SZ’s influence on Zhang. Smoking’s effect on PhenoAge was less pronounced. GrimAge was positively associated with alcohol and BMI, which were lower in SZ so would not explain acceleration.

After accounting for smoking, alcohol, and BMI, SZ was still associated with a 1- to 2-year acceleration in GrimAge and PhenoAge, and a 0.02 to 0.1 SD increase in Zhang (Table S7). To investigate this excess acceleration, we examined sub-clocks of GrimAge for serum levels of 7 age-associated proteins (Figures 2, S7, S8; Table S4). Six (B2M, Cystatin C, GDF-15, TIMP-1, ADM, and PAI-1) were increased in SZ, and all except GDF-15 remained elevated after correction for smoking, alcohol, and BMI (Table S9). These proteins could fully explain excess acceleration in all three mortality clocks (see Supplemental Results for more details).

Mitotic clocks are decelerated in SZ related to NK and CD8T cells

We found mitotic clock deceleration was related to cell proportions (Figure 1E–F), so we examined the contribution of individual cell types. SZ showed increased granulocytes and plasmablasts, and reduced NK cells, CD8T, and CD4T cells (Figures 2, S7, S8; Table S4), consistent with prior literature (46). Multiple regression analysis indicated CD8T and NK cell reduction in SZ significantly explains variance in MiAge and Yang in all data sets, while increases in granulocytes and plasmablasts may also contribute but the relationship is less consistent (Table S10). Furthermore, CD8T and NK cells had a stronger relationship with mitotic clocks in SZ compared to NPC, but granulocytes and plasma cells did not (Table S11).

Smoking was an important confounder. Smoking reportedly reduces NK cell number and activity (47), and we found decreased NK cells were associated with mitotic clock deceleration. However, this conflicted with the expectation that tobacco carcinogens should accelerate mitotic clocks. Smoking in SZ explained 40-50% of NK cell reduction, but not changes in other cell types (Table S12, Figure S10). After accounting for smoking’s effects on cell types, we found smoking independently accelerated MiAge and Yang as expected (Table S10, Figure S11). Alcohol, BMI, and serum proteins could not explain MiAge and Yang acceleration (Tables S10–14, see Supplemental Results).

Chronological age clocks are decelerated in males with clozapine use

All but one clock trained to predict chronological age did not differ between NPC and SZ in UCL and ABD, both before and after correcting for sex and cell counts (Figure S12, Tables S4 and S15). Hannum showed a small acceleration in SZ in UCL (0.9 years, p = 0.0236) and ABD (0.7 years, p = 0.0219). Smoking, alcohol, and BMI had little to no effect. However, we were surprised to find SZ deceleration for Horvath1 and Lin in CF, and trends towards deceleration in other chronological age clocks (Table S4). Multiple regression analysis revealed only one factor that could explain deceleration: clozapine use (Table S15).

When we separated SZ into three medication groups (non-clozapine antipsychotics only, clozapine only, or both), both groups taking clozapine showed a significant decrease in chronological age, but strikingly this deceleration occurred only in males (Figure 3, Table S16). Clozapine-only males showed decelerations of 5.6 years in Lin (p = 0.0011), 5.0 years in Horvath1 (p = 0.0024), 2.1 years in Horvath2 (p = 0.0451), 3.7 years in Hannum (p = 0.0074), 4.1 years in Weidner (p = 0.0599). Males taking both clozapine and other antipsychotics showed decelerations of 7.9 years in Lin (p < 0.0001), 5.1 years in Horvath1 (p = 0.0053), 2.3 years in Horvath2 (p = 0.0432), 3.5 years in Hannum (p = 0.0216), 5.2 years in Weidner (p = 0.0341). SZ patients taking non-clozapine antipsychotics and females taking clozapine were not different than NPC. No clozapine effect was seen for any other clocks, or for protein, smoking, alcohol, or BMI biomarkers (Table S17–19). Although antipsychotics affect the immune system and clozapine can cause agranulocytosis (48,49), the non-clozapine antipsychotic group did not differ from the two clozapine groups for any cell type and adjusting for cell counts did not change the effect of clozapine (Tables S16, S20).

Figure 3.

Clozapine effects on chronological age clocks in the CF data set are male-specific. Age accelerations are defined as residuals from the linear model for NPC, and therefore average age acceleration for male and female NPCs together is 0. Error bars are SEM. NPC = non-psychiatric controls, SZ = schizophrenia, AP = non-clozapine antipsychotics, CLO = clozapine. Statistical significance of difference from NPC is shown: <0.05*, <0.01**, <0.001***.

Mortality clock, mitotic clock, and clozapine effects are likely independent

Our analysis involved a large number of DNAm-based biomarkers, many of which may be correlated. Thus, it was unclear if our results concerning mortality clocks, mitotic clocks, and clozapine effects were independent, or instead reflected the same phenomenon. To address this, we sought to understand epigenetic aging on a global level rather than one biomarker at a time. Because most variables were tightly correlated due to chronological age, we utilized age residuals for biomarkers with reliable age trends.

We calculated pairwise correlations between all variables and performed hierarchical clustering (Figure S13). Because relationships between variables were broadly conserved across all data sets, we pooled the data and adjusted for differences across studies (Figures 4A, S15). There were four clusters of correlated clocks that mostly agreed with our categorization of mortality, mitotic, chronological age, and telomere length clocks. The exception was PhenoAge which clusters with chronological age clocks, though it was still highly correlated with mortality clocks. We performed principal component analysis (PCA) on the pooled data (Figures 4B–D), focusing on the first 6 PCs which explained 57.3% of the variance (Figures S16–17). PCA confirmed our findings that mortality clock acceleration is associated with smoking and age-associated proteins, while mitotic clock deceleration is associated with NK and CD8T cells. The mortality and mitotic clusters were not correlated with each other, and they loaded onto PC1 and PC2 respectively. Thus, mortality clock acceleration and mitotic clock deceleration in SZ were quantitatively independent.

Figure 4.

SZ effects on mortality, mitotic, and chronological age clocks are independent. (A) Correlation plot using biweight correlation between all measures for pooled data, adjusted based on study. Age residuals are used for variables that show robust age correlations (epigenetic clocks, proteins, smoking pack-years). Variables are ordered based on hierarchical clustering. Study-specific plots are found in Figure S13. (B-D) First six principal components, only showing variables with loadings higher than expected by chance. (E) PC correlations with SZ and sex for pooled data sets, and with SZ and antipsychotic classes for CF males. CF females are shown in Figure S18. Multiple regression analyses are shown in Table S21.

The chronological age clocks were more complex, with variance captured by shared signals from both mortality (PC1) and mitotic (PC2) clocks (Figures 4A–B), but also by PCs 3-5 (Figures 4C–D, S17). Multiple regression analysis showed clozapine affected PC3-5 specifically in males, but not PC1-2 after adjusting for SZ (Table S21, Figures 4E, S18). Thus, clozapine’s effect was independent of results concerning mortality and mitotic clocks in SZ.

See Supplemental Results for further implications of clustering and PCA, including for telomere length and non-clozapine antipsychotics.

Discussion

Aging is a multifaceted process unfolding across the lifespan leading to functional decline and mortality. Patients with SZ have different long-term health outcomes compared to the general population. However, it is difficult to identify SZ patient characteristics that predict aging outcomes, often occurring decades later, due to the large number of possible mechanisms, confounders, and disease heterogeneity. Epigenetic clocks can potentially bridge the gap between risk factors and aging because they can be measured throughout the lifespan and in every individual.

Prior publications reported the surprising finding that SZ is not associated with alterations in epigenetic aging, despite elevated risk factors and poor health outcomes. Our findings reconcile these observations, demonstrating that altered aging occurs in other epigenetic clocks not previously examined in SZ, because these clocks are sensitive to different aspects of aging (see Supplemental Discussion).

Mortality clock accelerations in SZ were impressive. The 5.8 year GrimAge acceleration in UCL corresponds to a 74% increased hazard of death (HR = 1.10^5.8 = 1.74) (4). Mortality clocks predict cardiovascular disease, diabetes, adiposity, and comorbidity count better than chronological age clocks (4,35,36). Mortality clock acceleration was driven largely by increased smoking, reinforcing the importance of addressing smoking in SZ.

Individuals with SZ showed increased predicted levels of six age-associated proteins that contribute to mortality clock acceleration and provide molecular links to cancer, cardiovascular disease, kidney function, hypertension, and cognitive function (4). It is unlikely that these results are artifacts of the DNAm predictor, as they agree with prior direct measurements of proteins in SZ (5–11). Their associations with smoking, alcohol, and BMI are consistent with the literature (50,51). GDF-15, previously found to be inversely associated with psychosis severity (7), is highly correlated with smoking, perhaps informing the debate about the role of smoking in SZ etiology (52). The other proteins show significant smoking-independent elevation, and some were previously linked to SZ pathogenesis and cognitive symptoms (5,9,53). This suggests the underlying biology of SZ contributes to aging and raises the possibility of novel SZ treatments to improve both mental and physical health.

Mitotic clock deceleration in SZ is interesting given the debate from epidemiological studies suggesting cancer risk is reduced in SZ despite high smoking rates and obesity (54). It has been unclear if this observation represents an epiphenomenon where individuals with SZ die before the development of cancer (55). However, we found mitotic age is lower in individuals with SZ throughout the lifespan, which could provide a biological basis for reduced cancer risk. Indeed, cancer risk may be reduced even before the age of SZ diagnosis and in first-degree relatives (56). Notably, our finding that mitotic clock deceleration is less pronounced in females is consistent with prior observations that the primary cancer types not decreased in SZ are breast and endometrial cancer (56) and that socioeconomic inequalities in smoking-related mortality in many countries have decreased in men but increased in women (57). Another link with cancer incidence is the role of NK and CD8T cells, known mediators of anti-tumor immunity (58), though it seems paradoxical that reduced NK and CD8T cells are associated with slowed mitotic age. However, intrinsic NK cell activity has been reported to be elevated in SZ (59), so it is possible that overactive NK and CD8T cells in SZ protect from cancer, and negative feedback reduces the number of circulating cells. Smoking also affects other cell types, especially neutrophils, that interact with NK and CD8T cells (60). Alternatively, reduced mitotic age may simply reflect differential mitotic rates of NK and CD8T cells relative to other immune cells. Mechanistic studies are needed to untangle these complex relationships.

Clozapine, the most effective antipsychotic for SZ, was associated with a male-specific deceleration in chronological age clocks, independent of immune cell proportions. The magnitude of this effect is striking, as Lin shows a 7-year deceleration yet the standard deviation for Lin acceleration among controls is only 5 years. This result was limited by medication data only being available for one data set with 74 SZ individuals. Available long-term medication data in epigenetic aging studies is uncommon, and therefore very few studies have identified medications with any effects on epigenetic age. This highlights the need to collect detailed medication data with DNA methylation in future studies. Clozapine use rates are higher in China than most other countries (61), and 48% of SZ subjects in CF were on clozapine. Medication data is not available for UCL or ABD, but these sets collected in the UK likely have much lower clozapine use rates. This may explain why chronological clock slowing was only observed in CF.

The practical consequences of clozapine-related age deceleration are unclear. Previous long-term cohort studies found up to 60% reduced all-cause mortality and 45% reduced cardiovascular mortality associated with clozapine (62,63), though it is puzzling why this might manifest in chronological age clocks but not mortality clocks. One possible explanation is that the clocks were trained in the general population, which may miss biological mechanisms of clozapine acting in a small subset of the population. A similar phenomenon may apply to the epigenetic BMI and alcohol biomarkers (see Supplemental Discussion). It is possible clozapine impacts the aging process itself, as a recent bioinformatics study identified drugs that induce younger human gene expression profiles, and the top 30 “geroprotective” compounds included clozapine and three other antipsychotics (64). Antipsychotics alter lifespan in the nematode C. elegans by modulating the insulin-like growth factor pathway that controls aging across numerous animal species (64,65). Clozapine may influence aging through unique effects on neurotrophic factors and glutamate (66–68). A sex-specific effect on aging is plausible given sex differences in clozapine side effects, metabolism, and treatment response (49,69). However, there are alternative explanations for clozapine’s association with biological age. Individuals with lower biological age may be more resilient to the adverse effects of clozapine and hence more likely to be continued on clozapine treatment or survive to be included the study. If that is true, then DNA methylation may be a useful biomarker for identifying patients appropriate for clozapine treatment, as clozapine is currently underutilized due to concerns about serious side effects. Clozapine may also be associated with increased healthcare since clozapine requires regular laboratory monitoring. Clozapine may exert acute effects directly on clock CpGs, rather than influencing a chronic aging process. Distinguishing between these possibilities will require detailed longitudinal studies. Nevertheless, it is likely that clozapine interacts with aging, and further investigation of this clozapine effect is warranted.

This study was limited primarily by available data on subjects and its cross-sectional design. There was not data to indicate if severity of psychiatric symptoms or physical illness accelerates the clocks in SZ. Direct quantification of BMI, substance use, and cell composition may yield different results than the epigenetic biomarkers we utilized (see Supplemental Discussion), though it is notable that the smoking epigenetic proxy predicts mortality more accurately than self-reported smoking (4). We measured smoking in cumulative pack-years, but it is possible that quantifying dose and duration separately may yield different results (70). Moreover, schizophrenia is a complex, heterogeneous disorder with diverse socioeconomic ramifications, such as increased risk of poverty, unemployment, stigma, social isolation, housing instability, and poor healthcare (22–24). Stress is well-established to influence aging (71), and education and income are associated with biological age (20). The cross-sectional design has many limitations. It is unclear if survival bias played a role in our findings, though we expect that correcting for survival bias would further increase the degree of SZ mortality clock acceleration. Longitudinal studies would enable study of when altered aging occurs (e.g. rapidly during early psychosis or slowly over decades). For example, childhood trauma and lifetime PTSD severity is cross-sectionally associated with higher Hannum but not Horvath1 age (72), while a longitudinal study found alcohol use disorder and specific PTSD symptoms increases the pace of aging of Horvath1 but not Hannum (73). Our study establishing schizophrenia’s association with multiple altered epigenetic aging biomarkers should motivate future longitudinal studies using well-phenotyped populations.

Our comprehensive epigenetic clock analysis is striking in how closely it matches prior observations about SZ: greatly increased all-cause mortality, reduced risk of cancer (except breast and endometrial), shortened telomeres, high smoking rates, derangements in immune cells and serum proteins, and unique effects of clozapine. Epigenetic clocks offer the opportunity to understand the biological basis of these observations. Importantly, this was the first epigenetic study to examine all these factors simultaneously, enabling us to dissect their relationships and suggest many of them are independent phenomena. Our work may serve as a foundation for future aging studies on the effects of genetics, developmental stressors, disorganized behavior, lifestyle factors, substance use, medications, co-morbidities, stigma, and socioeconomic factors. Given the inherent complexity of both psychiatric disease and aging, using multiple distinct epigenetic clocks simultaneously can shed new light on when, why, and how psychiatric disease influences diverse aging processes.

KEY RESOURCES TABLE

Resource Type	Specific Reagent or Resource	Source or Reference	Identifiers	Additional Information
Add additional rows as needed for each resource type	Include species and sex when applicable.	Include name of manufacturer, company, repository, individual, or research lab. Include PMID or DOI for references; use “this paper” if new.	Include catalog numbers, stock numbers, database IDs or accession numbers, and/or RRIDs. RRIDs are highly encouraged; search for RRIDs at https://scicrunch.ora/resources.	Include any additional information or notes if necessary.
Deposited Data; Public Database	GSE80417	NCBI GEO DataSets	RRID:SCR_005012; https://www.ncbi.nlm.nih.gov/gds
Deposited Data; Public Database	GSE84727	NCBI GEO DataSets	RRID:SCR_005012; https://www.ncbi.nlm.nih.gov/gds
Deposited Data; Public Database	GSE116379	NCBI GEO DataSets	RRID:SCR_005012; https://www.ncbi.nlm.nih.gov/gds
Software; Algorithm	R version 3.5.0	R Project for Statistical Computing	RRID:SCR_001905
Software; Algorithm	New Methylation Age Calculator	PMID:24138928	https://dnamage.genetics.ucla.edu/new	Includes Horvath1, Horvath2, Hannum, PhenoAge, GrimAge, smoking pack-years, GDF-15, ADM, B2M, PAI-1, TIMP-1, Cystatin C, Leptin, and cell composition estimates
Software; Algorithm	Horvath1 epigenetic clock	PMID:24138928	N/A
Software; Algorithm	Horvath2 epigenetic clock	PMID:30048243	N/A
Software; Algorithm	Vidal-Bralo epigenetic clock	PMID:27471517	N/A
Software; Algorithm	Lin epigenetic clock	PMID:26110659	N/A
Software; Algorithm	Hannum epigenetic clock	PMID:23177740	N/A
Software; Algorithm	Weidner epigenetic clock	PMID:24490752	N/A
Software; Algorithm	Garagnani epigenetic clock	PMID:23061750	N/A
Software; Algorithm	Bocklandt epigenetic clock	PMID:21731603	N/A	EDARADD
Software; Algorithm	GrimAge epigenetic clock	PMID:30669119	N/A	Includes smoking pack-years, GDF-15, ADM, B2M, PAI-1, TIMP-1, Cystatin C, Leptin
Software; Algorithm	Levine PhenoAge epigenetic clock	PMID:29676998	N/A
Software; Algorithm	Zhang epigenetic clock	PMID:28303888	N/A	continuous risk score
Software; Algorithm	MiAge epigenetic clock	PMID:29160179	http://www.columbia.edu/~sw2206/softwares.htm
Software; Algorithm	Yang epigenetic clock	PMID:27716309	N/A
Software; Algorithm	DNAmTL epigenetic clock	PMID:31422385	N/A
Software; Algorithm	McCartney smoking, alcohol, and BMI predictors	PMID: 30257690	N/A
Software; Algorithm	WGCNA R package v1.68	CRAN; Peter Langfelder; PMID: 19114008	RRID:SCR_003302
Software; Algorithm	ggcorrplot R package v0.1.2	CRAN; Alboukadel Kassambara	N/A
Software; Algorithm	factoextra R package v1.0.5	CRAN; Alboukadel Kassambara	RRID:SCR_016692