Extracting a biologically relevant latent space from cancer transcriptomes with variational autoencoders

Area of Science:

• Computational Biology
• Genomics
• Artificial Intelligence in Oncology

Background:

• The Cancer Genome Atlas (TCGA) provides extensive genomic data, including gene expression, for over 10,000 tumors across 33 cancer types.
• Gene expression profiles contain crucial information about tumor states and can be complex to analyze.
• Deep neural networks, particularly variational autoencoders (VAEs), can learn meaningful latent representations from complex data.

Purpose of the Study:

• To evaluate the effectiveness of a VAE in modeling cancer gene expression data from TCGA.
• To determine if the learned latent space captures biologically relevant features.
• To develop a tool for exploring hypothetical gene expression profiles and predicting tumor behavior.

Main Methods:

• A VAE, named Tybalt, was trained on The Cancer Genome Atlas (TCGA) pan-cancer RNA-sequencing data.
• The model's encoded features were analyzed to identify specific patterns.
• The potential biological relevance and applications of the VAE's learned representations were discussed.

Main Results:

• The VAE (Tybalt) was successfully trained on TCGA gene expression data.
• Analysis revealed specific patterns within the VAE's latent space.
• The study identified potential biologically relevant features captured by the model.

Conclusions:

• VAEs can effectively model cancer gene expression data.
• The learned latent space of Tybalt captures biologically relevant patterns.
• Tybalt shows promise for applications in cancer stratification and predicting expression patterns resulting from genetic alterations or treatments.