Learning a hierarchical representation of the yeast transcriptomic machinery using an autoencoder model

Area of Science:

• Systems Biology
• Computational Biology
• Genomics

Background:

• Cells possess intricate signaling systems that control responses, including gene expression.
• Reconstructing these cellular signaling systems using data is a significant challenge in systems biology.
• Understanding the hierarchical organization of transcriptomic machinery is crucial.

Purpose of the Study:

• To investigate the effectiveness of deep hierarchical neural networks for modeling yeast transcriptomic machinery.
• To determine if these models can learn and represent the hierarchical organization of gene regulation.
• To assess the biological relevance of the learned representations.

Main Methods:

• Designed a sparse autoencoder model with multiple hidden layers.
• Applied the autoencoder to over a thousand yeast microarrays.
• Evaluated the biological interpretability and utility of the learned latent variables.

Main Results:

• The first hidden layer of the autoencoder successfully identified yeast transcription factors (TFs) with a near one-to-one mapping.
• Higher hidden layers captured genes involved in common biological processes, reflecting known hierarchical relationships.
• Latent variables provided more abstract and concise microarray representations, improving data clustering compared to gene-based methods.

Conclusions:

• Deep hierarchical latent variable models, specifically autoencoders, can partially reconstruct the organization of transcriptomic machinery.
• These models offer a powerful data-driven approach for understanding complex biological systems.