Generative prediction of causal gene sets responsible for complex traits.

Area of Science:

• Genetics and Systems Biology
• Computational Biology and Machine Learning

Background:

• Understanding genotype-phenotype relationships for complex traits is challenging due to polygenic inheritance.
• Current genome/transcriptome-wide association studies lack causal inference and sufficient statistical power.

Purpose of the Study:

• To develop a novel computational approach to identify causal genes underlying complex traits and diseases.
• To enhance statistical power for detecting genotype-phenotype mappings.

Main Methods:

• Developed the transcriptome-wide conditional variational autoencoder (TWAVE), integrating causal information with a generative machine learning model.
• TWAVE uses a variational autoencoder on human transcriptional data within an optimization framework.
• Identified independently varying pathways (eigengenes) from generated expression profiles and performed constrained optimization to find causal gene sets.

Main Results:

• TWAVE successfully identified causal genes for complex traits that are undetectable by existing methods.
• The approach revealed that complex diseases can be caused by distinct sets of genes, indicating polygenic subtypes.
• Demonstrated the capability to identify distinct genotype-phenotype mappings driving disease subtypes.

Conclusions:

• The developed TWAVE approach offers enhanced statistical power for identifying causal genes in complex traits.
• This method can differentiate between distinct genetic drivers of complex diseases, suggesting potential for subtype discovery.
• The approach facilitates the design of targeted experiments for identifying multigenic targets for complex diseases.