A variational autoencoder trained with priors from canonical pathways increases the interpretability of transcriptome...

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Interpreting large-scale transcriptome data is crucial for understanding biological regulation.
• Traditional gene set enrichment analysis offers limited interpretability of complex regulatory networks.
• Deep learning offers potential for discovering novel patterns in high-dimensional biological data.

Purpose of the Study:

• To evaluate autoencoder architectures for learning reduced, biologically relevant representations of transcriptomes.
• To develop a novel variational autoencoder (VAE) that incorporates biological priors for enhanced interpretability.
• To explore the utility of deep learning for differential pathway analysis in transcriptomics.

Main Methods:

• Benchmarking five autoencoder architectures, including simple, variational, and beta-weighted VAEs.
• Utilizing latent variable modeling to capture essential biological signals from transcriptome data.
• Developing a prior-informed VAE to guide the learning of semantically meaningful latent features.

Main Results:

• All tested autoencoder architectures successfully reduced transcriptomes to 50 latent dimensions, enabling accurate data reconstruction.
• A simple autoencoder demonstrated superior performance in reconstruction benchmarks but lacked interpretable latent dimensions.
• The beta-weighted, prior-informed VAE successfully performed benchmarking tasks and generated latent features corresponding to biological pathways.

Conclusions:

• Deep learning, particularly VAEs with biological priors, can effectively learn interpretable representations of transcriptome data.
• This approach offers a promising new direction for transparent and interpretable differential pathway analysis in transcriptomics.
• The developed prior-informed VAE facilitates the translation of complex genomic data into understandable biological concepts.