Incorporating graph information in Bayesian factor analysis with robust and adaptive shrinkage priors

Area of Science:

• Computational Biology
• Bioinformatics
• Statistical Genetics

Background:

• High-dimensional multi-omics data analysis benefits from low-rank and sparse matrix decomposition.
• Bayesian factor models offer dimension reduction via sparsity-inducing priors.
• Existing models struggle to effectively integrate biological graph information.

Purpose of the Study:

• To develop a Bayesian factor model that incorporates biological graph knowledge for collaborative gene function identification.
• To improve the accuracy of factor loading structure recovery by adaptively shrinking gene networks.
• To enhance robustness against noisy graph edges and handle diverse data types.

Main Methods:

• Proposed a Bayesian factor model with novel hierarchical priors.
• Incorporated biological graph information to identify collaboratively functioning gene groups.
• Implemented adaptive shrinkage of factor loadings linked to graph structure.
• Developed priors to overcome the phase transition phenomenon.

Main Results:

• Achieved more accurate structure recovery of factor loadings compared to existing methods.
• Demonstrated robustness to noisy graph edges inconsistent with true sparsity.
• Successfully handled both continuous and discrete multi-omics data types.
• Outperformed several existing factor analysis methods in simulations and real data.

Conclusions:

• The proposed Bayesian factor model effectively integrates biological network information for enhanced multi-omics data analysis.
• Novel hierarchical priors improve factor loading recovery and robustness to noisy graph data.
• The model offers a versatile approach for dimension reduction and feature engineering across different data types.