Factorized embeddings learns rich and biologically meaningful embedding spaces using factorized tensor decomposition.

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• RNA sequencing (RNA-Seq) experiments generate vast, high-dimensional datasets.
• Analyzing complex RNA-Seq data presents significant computational challenges.
• Dimensionality reduction techniques are crucial for extracting meaningful patterns from such data.

Purpose of the Study:

• To introduce the Factorized Embeddings (FE) model, a self-supervised deep learning algorithm.
• To learn simultaneous gene and sample representation spaces from RNA-Seq data using tensor factorization.
• To evaluate the performance of FE-derived sample representations on downstream regression tasks.

Main Methods:

• Developed a self-supervised deep learning algorithm named Factorized Embeddings (FE).
• Employed tensor factorization to learn joint gene and sample embedding spaces.
• Applied the FE model to RNA-Seq data from two large-scale cohorts and benchmarked against existing methods on 49 regression tasks.

Main Results:

• FE successfully learned sample representations that capture both single-gene and global expression patterns.
• The gene representation space organized genes by tissue specificity, co-expression, and shared Gene Ontology (GO) terms.
• FE-derived sample representations achieved top-ranking performance (first or second) across all 49 evaluated regression tasks.

Conclusions:

• The Factorized Embeddings model provides an effective dimensionality reduction strategy for RNA-Seq data.
• FE generates biologically relevant gene and sample embeddings, enhancing data interpretation.
• FE demonstrates superior performance compared to existing methods for various biological regression tasks.