Benchmarking variational AutoEncoders on cancer transcriptomics data.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Deep generative models, particularly variational autoencoders (VAEs), are increasingly utilized in computational biology.
• These models excel at capturing complex data structures for tasks like cancer subtyping.
• However, VAEs present training challenges due to numerous hyperparameters requiring careful tuning.

Purpose of the Study:

• To investigate the impact of various hyperparameters on VAE performance in downstream cancer-related tasks.
• To provide data-driven recommendations for optimal hyperparameter selection in VAEs.
• To assess the biological relevance captured by VAE-learned latent spaces.

Main Methods:

• Trained six different VAE models on TCGA transcriptomics data, evaluating performance on cancer subtype clustering and survival analysis.
• Systematically varied hyperparameters including latent space dimensionality, learning rate, optimizer, initialization, and activation function.
• Validated hyperparameter effects on the GTEx dataset to ensure generalizability and measured correlations between latent space representations and biological data characteristics.

Main Results:

• β-TCVAE and DIP-VAE demonstrated good average performance but were sensitive to hyperparameter choices.
• A significant correlation (ρ = 0.7) was observed between hyperparameter effects on clustering in TCGA and GTEx datasets, confirming robustness.
• Learned latent factors generally did not uniquely correlate with specific biological characteristics like gender, age, or mutation signatures.

Conclusions:

• Established robust, generalizable recommendations for VAE hyperparameter selection in transcriptomics data analysis.
• Highlighted the sensitivity of certain VAE models to hyperparameter settings.
• Indicated that current VAE latent spaces may not fully capture separable or uniquely correlated biological information.