Latent space arithmetic on data embeddings from healthy multi-tissue human RNA-seq decodes disease modules.

Area of Science:

• Computational biology
• Genomics
• Systems biology

Background:

• Transcriptomic data analysis advances disease understanding.
• Limited sample sizes of diseased tissues hinder current computational approaches.

Purpose of the Study:

• To determine if a variational autoencoder trained on healthy human RNA sequencing (RNA-seq) data can capture gene regulation and generalize to disease states.
• To explore the model's ability to deconstruct disease mechanisms and identify drug targets.

Main Methods:

• Training a variational autoencoder on large-scale healthy human RNA-seq data.
• Testing the model's compression of transcriptomic changes from 25 independent disease datasets.
• Decoding disease-specific signals from the model's latent space.
• Comparing identified disease genes with differential expression analysis.
• Matching disease signals with known drug targets.

Main Results:

• The model successfully compressed transcriptomic changes from 25 diverse disease datasets.
• Disease-specific signals decoded from the latent space contained more disease-specific genes than differential expression analysis in 20 out of 25 cases.
• Known and potential pharmaceutical candidates were identified by matching disease signals with drug targets.

Conclusions:

• Data-driven representation learning with variational autoencoders can effectively deconstruct the latent space of transcriptomic data.
• This approach facilitates the dissection of complex disease mechanisms.
• The method aids in the identification of novel drug targets and pharmaceutical candidates.