TBLDA

TBLDA models shared latent topics between genotype vectors and RNA-sequencing raw count gene expression data to identify expression quantitative trait loci (eQTLs) across tissues while accommodating nested RNA-seq samples and ancestral structure.

Key Features:

• Bimodal Latent Dirichlet Allocation: Integrates genotype and gene expression modalities via a bimodal latent Dirichlet allocation framework to learn shared topics.
• Nested sample handling: Accommodates datasets where multiple RNA sequencing samples correspond to a single individual's germline genotype vector.
• Raw count RNA-seq input: Operates on raw RNA-sequencing count data rather than relying on normalized expression matrices.
• Ancestral structure decomposition: Captures ancestral structure within a genotype-specific latent space and isolates this component from shared topics.
• Application to GTEx v8: Demonstrated on GTEx v8 expression data across 10 tissues to capture biological signals in both genotype and expression modalities.
• eQTL discovery: Maps the most informative features from topics to identify eQTLs, reporting 4,645 cis-eQTLs and 995 trans-eQTLs.

Scientific Applications:

• Cross-tissue eQTL mapping: Identification of cis- and trans-eQTLs across multiple tissues using shared latent topics.
• Context-specific gene regulation: Characterization of tissue-specific and shared regulatory programs influencing gene expression variability.
• Genotype–expression integration: Exploration of relationships between germline genotype variation and RNA-seq measured gene expression using raw counts.

Methodology:

Learn shared topics from genotype vectors and raw RNA-seq counts using a bimodal latent Dirichlet allocation model, accommodate multiple RNA-seq samples per genotype, separate ancestral structure in a genotype-specific latent space, and map topic-informative features to identify cis- and trans-eQTLs.