DNA methylation (DNAm) is an epigenetic regulator of gene expression programs that can be altered by environmental exposures, aging, and in pathogenesis. Traditional analyses that associate DNAm alterations with phenotypes suffer from multiple hypothesis testing and multicollinearity due to the high-dimensional, continuous, interacting and non-linear nature of the data. Deep learning analyses have shown much promise to study disease heterogeneity.DNAm deep learning approaches have not yet been formalized into user-friendly frameworks for execution, training, and interpreting models. Here, we have developed MethylNet to make predictions, generate new data, and uncover unknown heterogeneity with minimal user supervision. The results of our experiments indicate that MethylNet can study cellular differences, grasp higher order information of cancer sub-types, estimate age and capture factors associated with smoking in concordance with known differences. The ability of MethylNet to capture nonlinear interactions presents an opportunity for further study of unknown disease, cellular heterogeneity and aging processes. challenges, many downstream EWAS analyses have focused on reducing the dimensions into a rich feature set to associate with outcomes. By limiting the number of features through dimensionality reduction and feature selection, analyses become more computationally tractable and the burden of correcting for multiple comparisons is reduced.An important advancement to methylation-based deep learning analyses was the application of Variational Auto-encoders (VAE). Initial deep learning approaches for DNAm data focused on estimating methylation status and imputation, performing classification and regression tasks, and performing embeddings of CpG methylation states to extract biologically meaningful lower-dimensional features [15][16][17][18][19][20][21][22] . VAEs embed the methylation profiles in a way that represents the original data with high fidelity while revealing nuances 4,5,23 . Thereafter, researchers attempted to develop similar frameworks for extracting features for downstream prediction tasks and identify meaningful relationships revealed by VAE latent representations 24 . However, VAE models are sensitive to the selection of hyperparameters 25 and have not been optimized for synthetic data generation, latent space exploration, and prediction tasks. Many autoencoder approaches represent the data using an encoder, and then utilize a non-neural network model (e.g. support vector machine) to finalize the predictions. Presently, to the best of our knowledge there is no end-to-end training approach that both extracts biologically meaningful features through latent encoding and performs predictions using the derived features. Further, existing frameworks do not output predictions for multi-target regression tasks, such as cell-type deconvolution and subject age prediction.Here, we leverage deep learning latent space regression and classification tasks through the development of a modular framework that is ...