An unsupervised deep learning framework with variational autoencoders for genome-wide DNA methylation analysis and biologic feature extraction applied to breast cancer
Recent advances in deep learning, particularly unsupervised approaches, have shown promise for furthering our biological knowledge through their application to gene expression datasets, though applications to epigenomic data are lacking. Here, we employ an unsupervised deep learning framework with variational autoencoders (VAEs) to learn latent representations of the DNA methylation landscape from three independent breast tumor datasets. Through interrogation of methylation-based learned latent dimension activation values, we demonstrate the feasibility of VAEs to track representative differential methylation patterns among clinical subtypes of tumors. CpGs whose methylation was most correlated VAE latent dimension activation values were significantly enriched for CpG sparse regulatory regions of the genome including enhancer regions. In addition, through comparison with LASSO, we show the utility of the VAE approach for revealing novel information about CpG DNA methylation patterns in breast cancer.