Epigenetic Clock Assessment

Abstract

Background: During the last decade, multiple efforts have been dedicated to the modelling of an accurate predictor of age based on DNA methylation status(1–4). However, most of current models have been trained and therefore predict over a limited typology of tissues, and consequently the application over a wider extent of tissues is restricted . Also, their suitability to be implemented with data derived from the last Illumina Infinium Methylation EPIC array remains unexplored. In this master thesis, an extensive review of the current most relevant epigenetic age calculators is conducted, considering the reproducibility of algorithms implemented and their compared performance (error during prediction) when applicated over a range of multi-tissue available datasets. Also, an assessment of the parameter "age acceleration"(3) as a potential biomarker in clinical prognosis of a set of cancer types is performed. • Results: After performing an initial screening, only three methods were further considered, those published by Hannum et al(3,5), Horvath(1) and Xu et al (6). Xu et al method was removed from the study since it was validated over whole genome methylation data, and the sites that were embeded in the model where not extrapolable to Illumina CpG probes. Therefore, it was not reproducible to be used with microarray data. Hannum's method produced reasonably good results when tested over blood samples of small size, but still not significant due to reduced size of available datasets. Also, results over other types of tissues presented lower correlation and considerably high error when compared to results obtained by Horvath over the same datasets. Hannum method seemed to not be suitable for application over 27k microarray data, probably due to Hannum markers not being included inside the design of 27k array. Age acceleration parameter was calculated as the difference between real age and predicted age according to Horvath(3). Age acceleration was found to be significantly associated with squamous lung cancer stage (p value =0.0008) tested over 502 samples. An ordinal multivariate model considering age acceleration and gender as covariates and cancer stage as predictor was developed (cancer stage ~ age acceleration + gender), with good sensitivity and specificity values for stage Ia (50% and 74%, respectively), and slightly poorer for stage Ib (84% and 42%, respectively). Classification for the other stages was null probably due to overrepresentation of stages Ia and Ib in the training set. Conclusions: Multi-tissue predictor based on 353 CpGs developed by Horvath seems to be the best choice in order to obtain accurate age prediction results over a range of healthy, and therefore, diseased tissues. Also, a prediction model based on age acceleration extracted from Horvath age predictions might be able to classify squamous cell lung cancer samples according to their stage (Ia,Ib,IIa,IIb,IIIa..). However, this model is still a prototype, and subsequently parameter optimization needs to be performed for better classification perspectives.