MethylNet: an automated and modular deep learning approach for DNA methylation analysis

Abstract

Background: 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 multi-collinearity 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 describe MethylNet, a DNAm deep learning method that can construct embeddings, make predictions, generate new data, and uncover unknown heterogeneity with minimal user supervision. Results: 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. Conclusion: The ability of MethylNet to capture nonlinear interactions presents an opportunity for further study of unknown disease, cellular heterogeneity and aging processes.

MethylSPWNet and MethylCapsNet: Biologically Motivated Organization of DNAm Neural Network, Inspired by Capsule Networks

“…We aimed to predict the 38 histological subtypes of CNS tumors (39 classes, including controls) as a test case for the capsule-inspired neural network approaches. We compare the MethylCapsNet and MethylSPWNet frameworks for capsule organization with the existing MethylNet framework (which does not account for Capsule organized information [20]), and with Group LASSO Logistic Regression Classifier organized by capsules. Details of modeling approaches, fitting procedures, and capsule selection are in the Methods section.…”

“…The MethylCapsNet methodology presents an extension of the MethylNet framework [20] and is implemented as a command-line interface that allows the user to group CpGs into capsules, and then dynamically route the capsules to make a prediction and interpret the results.…”

“…The information at the subsequent layer of nodes is transformed using non-linear transforms/activations/link functions such as the sigmoid (logistic), hyperbolic tangent (tanh), or rectified linear unit (ReLU) transforms. These types of analyses are common for deep learning analyses of DNAm data, where the input data is a list of beta values for each subject [20].…”

“…DNAm deep learning frameworks, such as MethylNet, can accurately characterize tissue, disease states, and infer subject age and cell type proportions through unsupervised embedding, generation, classification, and regression tasks, [20][21][22][23][24] while attempting to ascribe important methylated loci through the use of model interpretability frameworks such as SHAP [25] or LIME [26]. While the inclusion of more CpGs presents an opportunity to expand the space of biologically testable hypotheses [20], the CpGs included in these classifiers have been poorly characterized; statistical challenges (e.g., multicollinearity) with interpretations and generation of associations with pathways remain understudied. [27] Multicollinearity, the unusually high correlation between features, can be addressed with careful feature-selection or grouping, such as with elastic net penalization [1,28].…”

Don’t Dismiss Logistic Regression: The Case for Sensible Extraction of Interactions in the Era of Machine Learning

“…In the era of Big Data, models with highly complex specifications are employed to study nontrivial biomedical phenomena, including genetic or epigenetic interactions, high-resolution modalities such as Computed Tomography (CT) scans and histopathology slide images [1][2][3][4][5][6]. Many statistical approaches to modeling these complex data rely on expert consultation and annotation to help determine which variables and targets to study.…”