SuSiE PCA: A scalable Bayesian variable selection technique for principal component analysis.

Area of Science:

• Genomics
• Statistical Genetics
• Bioinformatics

Background:

• Latent factor models, such as Principal Component Analysis (PCA), are crucial for uncovering low-rank structures in biological data.
• Feature selection in sparse latent factor models remains a significant challenge.
• Identifying relevant biological features from high-dimensional datasets is essential for advancing research.

Purpose of the Study:

• To introduce SuSiE PCA, a scalable sparse latent factor model designed for robust feature selection.
• To evaluate the performance of SuSiE PCA against existing methods in simulations and real-world biological data.
• To demonstrate the utility of SuSiE PCA in identifying tissue-specific factors and gene modules.

Main Methods:

• Developed SuSiE PCA, a sparse latent factor model incorporating posterior inclusion probabilities to assess variable uncertainty.
• Conducted extensive simulations to validate model performance, focusing on signal detection and robustness.
• Applied SuSiE PCA to multi-tissue expression quantitative trait loci (eQTLs) data (GTEx v8) and large-scale perturbation datasets.

Main Results:

• SuSiE PCA demonstrated superior performance in signal detection and model robustness compared to other methods in simulations.
• The model successfully identified tissue-specific latent factors and their associated eGenes in the GTEx v8 eQTLs data.
• SuSiE PCA identified gene modules with higher enrichment of ribosome-related genes on perturbation data and was significantly faster than sparse PCA.

Conclusions:

• SuSiE PCA offers an efficient and robust approach for feature selection in high-dimensional biological data.
• The method provides valuable insights into genetic architecture and regulatory mechanisms.
• SuSiE PCA is a powerful tool for analyzing complex biological datasets, including eQTLs and perturbation data.