Compressing gene expression data using multiple latent space dimensionalities learns complementary biological...

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Unsupervised compression algorithms analyze gene expression data to reveal hidden signals of variation.
• Current methods require users to select a single latent space dimensionality, potentially limiting biological feature capture.
• Most researchers use only one algorithm and latent dimensionality, risking incomplete biological insights.

Purpose of the Study:

• To investigate how selecting a single latent dimensionality impacts the biological features captured from gene expression data.
• To determine if using multiple algorithms and dimensionalities can improve biological discovery.
• To assess the influence of latent space dimensionality on subsequent analyses of gene expression data.

Main Methods:

• Gene expression data from adult normal, adult cancer, and pediatric cancer tissues were compressed.
• Multiple unsupervised compression models were trained across a wide range of latent space dimensionalities.
• Performance differences were observed, and associations with curated pathway gene sets were identified.

Main Results:

• Intermediate latent dimensionalities revealed significant pathway gene set associations in denoising autoencoder and variational autoencoder models.
• Combining compressed features from various algorithms and dimensionalities yielded the most comprehensive pathway-associated representations.
• Models trained at different dimensionalities captured distinct biological signals, including sex, MYCN amplification, and cell types.
• Tumor type was best captured at lower dimensionalities, while subtle pathway activities were identified at higher dimensionalities.

Conclusions:

• No single latent dimensionality or compression algorithm is optimal for gene expression data analysis.
• Utilizing features from diverse compression models and multiple latent space dimensionalities enhances biological representations.
• A multi-faceted approach improves the depth and breadth of biological discoveries from gene expression data.