Explainable epigenetic aging clocks: an overview of existing AI models and approaches.

<h4>Introduction</h4>There is a growing demand to make epigenetic aging clocks that are interpretable and actionable. Here, we identify and describe explainable artificial intelligence (XAI) epigenetic clocks that convert CpGs into higher-level abstractions using deep learning.<h4>Areas covered</h4>By combing through PubMed, this narrative review identifies six different XAI clocks: XAI-AGE, DL-XAI, EpInflammAge, PathwayAge, EXP.REACTOME, and EXP.TFT. XAI-AGE, DL-XAI, PathwayAge, EXP.REACTOME, and EXP.TFT provide explainability by leveraging the fact that a subset of DNA methylation sites are annotated to genes. From here, genes are converted into biological pathways or a protein-protein interaction network. Conversely, the XAI clock EpInflammAge works by taking methylation-based predictors of circulating chemokines and cytokines - referred to as synthetic proxies - and feeding these as inputs into a deep learning model.<h4>Expert opinion</h4>By evaluating and comparing extant XAI clocks, it is clear that explainability is compatible with the ability to make accurate age predictions that are associated with meaningful health signals. The pathway- or network-level models are, however, more difficult to interpret and less actionable than the XAI clock based on synthetic proxies. Explainability and actionability could be improved by only incorporating robust estimators of well-understood biomarkers or by exclusively including tangible, easy-to-understand biological pathways.