NCAE: data-driven representations using a deep network-coherent DNA methylation autoencoder identify robust disease...

Area of Science:

• Epigenetics
• Bioinformatics
• Computational Biology

Background:

• Precision medicine requires identifying reliable disease and risk factor signatures from omics data.
• Knowledge-driven methods may miss novel biological insights due to inherent biases.
• DNA methylation plays a crucial role in gene regulation and disease development.

Purpose of the Study:

• To develop a data-driven workflow for discovering DNA methylation signatures using network-coherent autoencoders (NCAEs).
• To identify robust signatures for risk factors (aging, smoking) and diseases (systemic lupus erythematosus).
• To overcome limitations of knowledge-driven approaches in omics data analysis.

Main Methods:

• Explored autoencoder architectures on a large human epigenome-wide association studies compendium (n=75,272).
• Utilized network-coherent autoencoders (NCAEs) with biologically relevant latent embeddings.
• Trained interpretable deep neural networks using NCAE embeddings for prediction tasks.

Main Results:

• Observed emergence of co-localized patterns in autoencoder latent space corresponding to biological network modules.
• Identified an NCAE configuration with strong co-localization and centrality signals in the human protein interactome.
• NCAE embedding-based models outperformed existing predictors, revealing novel DNA methylation signatures.

Conclusions:

• The data-driven workflow provides a generalizable pipeline for capturing risk factor and disease information from DNA methylation data.
• This approach enhances understanding of complex epigenetic processes by surpassing knowledge-driven method limitations.
• Facilitates the development of improved diagnostic and therapeutic strategies for various conditions.