mEthAE: an Explainable AutoEncoder for methylation data

Abstract

Despite the wealth of knowledge generated through epigenome-wide association studies our understanding of the relationships of CpG sites is still limited, as analysis of DNA methylation data remains difficult due its high dimensionality. To combat this problem, deep learning algorithms, such as autoencoders, are increasingly applied to capture the complex patterns and reduce dimensionality into latent space. We believe that the way an autoencoder groups together CpGs in its latent dimensions has biological meaning and might reveal novel insights regarding the relationship of CpGs. Therefore, in this work, we propose a chromosome-wise autoencoder for interpretable dimensionality reduction of methylation data (mEthAE). Our framework shows an impressive reduction in dimensions of up to 400-fold compared to the provided input, without compromising on reconstruction accuracy or predictive power in the latent space. Through our perturbation-based interpretability approach we revealed groups of CpGs which are highly connected across all latent dimensions (global CpGs) and were significantly more often reported in EWAS studies, indicating our interpretability method can successfully identify CpGs with biological relevance. In an attempt to gain a deeper understanding of the relationship between individual CpG sites, we focused on interpreting individual latent features and found that CpGs connected to a common feature do not share biological associations, correlation patterns, or are located in close proximity on the chromosome. We conclude that while there is evidence that the autoencoder does not group CpGs randomly, the logic behind the observed CpG relationships can not be delineated easily. With regards to the analyses done in this work, we believe that the autoencoder groups CpGs according to long range non-linear interaction patterns that lack characterisation in the current epigenetic research landscape.