Learning a Latent Space of Highly Multidimensional Cancer Data

Area of Science:

• Computational Biology
• Genomics
• Machine Learning

Background:

• High-dimensional gene expression data presents challenges in identifying biologically relevant patterns.
• Understanding the relationships and transitions between different cancer types is crucial for developing targeted therapies.

Purpose of the Study:

• To introduce a novel deep learning network, the Unified Disentanglement Network (UFDN), for analyzing The Cancer Genome Atlas (TCGA) data.
• To demonstrate the network's ability to learn a biologically meaningful, low-dimensional latent space from high-dimensional gene expression data.
• To explore the potential of UFDN for interpolating between distinct cancer types and analyzing resulting gene expression changes.

Main Methods:

• Training a Unified Disentanglement Network (UFDN) on The Cancer Genome Atlas (TCGA) dataset.
• Applying the trained UFDN-TCGA model to cancer status and cancer type classification tasks.
• Performing continuous, partial interpolation between different cancer types within the learned latent space.
• Analyzing differentially expressed genes between original and interpolated cancer samples.

Main Results:

• UFDN-TCGA successfully learns a biologically relevant, low-dimensional latent space from TCGA gene expression data.
• The network achieves performance comparable to established random forest methods in cancer classification.
• Continuous interpolation between distinct cancer types is demonstrated, revealing smooth transitions in the latent space.
• Analysis of interpolated samples shows the presence of relevant metagenes that recapitulate known mechanisms of target cancer types, such as glioblastoma.

Conclusions:

• The Unified Disentanglement Network (UFDN) provides a powerful tool for uncovering latent biological structures in cancer genomics data.
• UFDN facilitates novel analyses, including continuous interpolation between cancer types, offering new insights into cancer progression and heterogeneity.
• The ability to recapitulate known cancer mechanisms through interpolation highlights the biological relevance and potential clinical applications of the learned latent space.