In this work, we presented the first longitudinal analysis of epigenetic age trajectory during the emergence of psychosis in a prospective UHR cohort. We predicted baseline and follow-up epigenetic age using the Horvath epigenetic clock on 38 UHR subjects with longitudinal methylation data available at two time points. Our cross-sectional results replicated those of the two most recent large-scale cross-sectional predictions of epigenetic age, while offering a possible explanation for their conflicting results. At baseline, future converters were epigenetically younger than nonconverters, as it was also reported by Wu and colleagues in subjects with psychosis, using the same Horvath clock.¹⁸ At follow-up, once some subjects converted to psychosis, there was no epigenetic gap difference anymore between actual converters and nonconverters, which is in line with the other large-scale study that compared patients with psychosis and controls, also with the Horvath clock.³⁵ This demonstrates that epigenetic aging is dynamic and may vary across time points, depending on clinical stage or environmental factors, as previously described with regard to longitudinal brain age variation.¹⁷ Thus, we found that although converters were epigenetically younger at baseline, they showed an accelerated methylomic aging of 2.82 years/year during the conversion time frame, while nonconverters maintained a constant aging, without acceleration. This association remained significant after adjustment for age, sex, and cannabis use, but not when medication and tobacco were introduced as covariates. Methylomic age acceleration was largely correlated with longitudinal hypomethylation among most of the 353 CpGs used for age prediction. This is in line with the fact that age- or environmental stressor-related effects are known to mainly cause global hypomethylation, with the exception of CpG islands that may show hypermethylation.³⁶ The six CpG sites whose methylation variation significantly correlated with age acceleration, after correction for multiple testing, were located in OR6C1 (olfactory receptor), IFFO1 (intermediate filament family orphan 1), ARHGAP44 (rho GTPase activating protein 44), LAMA2 (laminin), and TSPAN12 (tetraspanin 12) genes. According to the GeneCards database, IFFO1 participates to DNA repair, ARHGAP44 may be involved in dendritic spine formation and synaptic plasticity (by similarity), LAMA2 may mediate cell migration and organization into tissues during embryonic development, and TSPAN12 regulates cell surface receptor signal transduction, and membrane proteinases. Notably, LAMA2 was found to be among genes recurrently affected by de novo mutations associated with schizophrenia,³⁷ while ARHGAP44 is part of a family of genes coding for GTPase activating proteins, where copy number variants have been found more frequently associated with schizophrenia and neurodevelopmental disorders, as well as reduced spine density in the medial prefrontal cortex of animal models.³⁸ While anomalies in the methylation of these genes have been found to contribute to aging and age prediction,¹⁸ the fact that some of the same genes are also linked with schizophrenia may strengthen the idea of an association between this disease and intrinsic dysfunctions in age trajectories.

Several considerations need to be taken into account. First, the longitudinal design allowed the detection of significant differences despite a small sample size. Indeed, longitudinal settings are more powerful than cross-sectional ones as they adjust for any individual variation unrelated to disease because each subject is its own control.³⁹ Second, in order to predict epigenetic age, we applied the well-validated mDNA-based Horvath clock,¹⁸ because it has been shown to be the most performant for chronological age prediction in young subjects,⁴⁰ and also the less prone to error. In our work, we found a correlation of 77% between chronological and predicted ages, with a mean absolute error of 3.3 years. Along these lines, Wu and colleagues found that the Horvath clock predicted age with a similar performance (error of 3.7 years) and a higher correlation of 96% (the fact that we obtained a lower correlation despite the same prediction performance is due to the younger age distribution of our cohort—Supplementary). Conversely, they showed that other clocks such as Hannum’s and Levine’s were less performant, with greater errors (8.3 and 11 years respectively).¹⁹ Such large positive errors are also exacerbated in younger cohorts due to the poor availability of mDNA datasets of young subjects in the various epigenetic learning sets. The Hannum clock is based on 656 subjects aged 19 to 101,⁴¹ and the Levine clock is built on more than 9000 samples, but from 21 to 100.⁴² Therefore, any use on young subjects like the ones in this UHR cohort, on the extreme lower tail of the learning algorithm’s distribution, will necessarily lead to regression to the mean effects with extreme overestimations of the predicted age.^29–31 To our knowledge, only the Horvath predictor accounted for adolescence and young adulthood periods by including samples of subjects starting from birth, and adjusting for the nonlinear aging trajectories occurring around 20.¹⁸ Nevertheless, to account for a possible regression to the mean also related to the nominal age difference between converters and nonconverters, and following recent benchmark recommendations,³⁰ we also tested models with chronological age as a covariate, which did not change the association between disease and age acceleration.

Third, the significance of the age acceleration that we reported is robust, with a bootstrapped 95% confidence interval excluding the null acceleration. Moreover, we took advantage of a deeply phenotyped cohort in order to adjust for known confounders of methylation patterns—cannabis, tobacco, and medication. Our iterative addition of covariates to this model showed that conversion to psychosis may lead to increased epigenetic age acceleration independently of age, sex, and cannabis intake, but not of medication and tobacco use. However, these last two measures are semi-quantitative and their use as covariates may also lead to overcorrection of a relevant variance, especially as there is high collinearity between tobacco and cannabis, and between status and medication.

Finally, result interpretation is limited by several points. Converters had a higher male/female ratio, and males are expected to be epigenetically older than females, across the lifespan.⁴³ Although we included sex as a covariate in the estimation of association between epigenetic age and conversion, we cannot fully exclude a possible bias in mDNA measures. Converters are also slightly—although not significantly—younger than nonconverters. Having also adjusted for age as a covariate, this raises nevertheless the question of whether the observed difference in age acceleration is truly associated with the emergence of psychosis or is related to normal maturational aging, which is physiologically more accelerated at younger ages.^18,44 Longitudinal studies with matched chronological age distributions between groups could answer this question. Lastly, our interpretation of aging is intrinsically limited by the little insight we have on the underlying mechanisms. Beyond methylation, other epigenetic mechanisms such as microRNAs may be relevant and need to be considered in psychosis.¹⁶ It is also unclear whether age acceleration translates a pathologic process at the core of conversion to psychosis or is rather an adaptive mechanism of epigenome maintenance under environmental pressure.^40,44 This will have implications on how age acceleration could be clinically relevant, either as a biomarker of risk or as a target of treatment itself.

In summary, our results support a neurodevelopmental model of the emergence of psychosis. This study of longitudinal epigenetic aging, which identified both a decreased epigenetic age gap at baseline, in favor of previous epigenetic age deceleration,¹⁹ and an epigenetic age acceleration contemporary to psychosis’ emergence, may underpin the hypothesis of a dysmaturational process, different from constant progressive premature aging. If replicated in larger studies, age acceleration could constitute a measurable surrogate dynamic biomarker of risk of psychosis, and exploration of its underlying pathway may uncover new targets for early intervention.