Deriving disease modules from the compressed transcriptional space embedded in a deep autoencoder.

Area of Science:

• Computational Biology
• Genomics
• Systems Biology

Background:

• Disease module identification is crucial for understanding complex diseases.
• Existing methods rely on biological networks, which are often incomplete and biased.
• This limits the discovery of novel disease-associated genes.

Purpose of the Study:

• To investigate if disease-relevant gene modules can be discovered using deep autoencoders.
• To explore module discovery directly from large-scale transcriptional data, bypassing biological networks.
• To determine if autoencoder representations can reveal disease-gene relationships.

Main Methods:

• Trained a deep autoencoder model using extensive transcriptional data.
• Analyzed gene enrichment within different layers of the autoencoder representation.
• Compared the distribution of genome-wide association studies (GWAS) and protein-protein interaction (PPI) data across layers.

Main Results:

• Found significant enrichment of GWAS-associated genes in deeper layers of the autoencoder.
• Observed an inverse gradient for PPI signals, strongest in the initial layers and diminishing deeper.
• Demonstrated that autoencoder layers capture distinct biological signals related to disease.

Conclusions:

• A data-driven approach using deep autoencoders is sufficient for discovering disease-related gene modules.
• This method overcomes limitations of incomplete and biased biological networks.
• The findings support a novel strategy for identifying disease genes from transcriptional data.