Bayesian joint additive factor models for multiview learning

Area of Science:

• Statistical modeling
• Bioinformatics
• Precision medicine

Background:

• Increasing prevalence of multi-view data in scientific research.
• Need for advanced statistical methods to analyze complex relationships within and across data types.
• Challenges in integrating multimodal data for improved outcome prediction, especially with varying signal-to-noise ratios.

Purpose of the Study:

• To develop novel factor regression models for analyzing multiview data.
• To address challenges in interpretability, feature selection, and uncertainty quantification.
• To improve prediction of clinical outcomes using integrated multimodal data.

Main Methods:

• Introduction of two complementary factor regression models: joint factor regression (jfr) and Joint Additive FActor Regression (jafar).
• Application of independent cumulative shrinkage process (I-CUSP) priors for jfr and a novel dependent version (D-CUSP) for jafar.
• Development of Gibbs samplers for flexible model fitting and accommodating various feature and outcome distributions.

Main Results:

• The jafar model successfully decomposes variation into shared and view-specific components, enhancing interpretability.
• The developed models provide accurate uncertainty quantification and facilitate feature selection.
• Demonstrated performance gains in predicting time-to-labor onset using immunome, metabolome, and proteome data compared to existing methods.

Conclusions:

• The proposed factor regression models offer a powerful framework for analyzing multiview data in precision medicine.
• These models effectively integrate multimodal information to improve outcome prediction while maintaining interpretability.
• The study highlights the potential of advanced statistical approaches for leveraging complex biological data in clinical applications.