Bayesian Simultaneous Factorization and Prediction Using Multi-Omic Data | JoVE Visualize

Area of Science:

Computational biology
Statistical genetics
Bioinformatics

Background:

Integrative factorization methods analyze multi-omic data to identify biological variation, predict outcomes, and impute missing values.
Existing methods lack comprehensive statistical inference and uncertainty quantification for these tasks.
A unified probabilistic framework is needed for simultaneous factorization, prediction, and imputation.

Purpose of the Study:

To propose a novel Bayesian framework, Bayesian Simultaneous Factorization (BSF), for decomposing multi-omic variation into joint and individual structures.
To extend BSF to Bayesian Simultaneous Factorization and Prediction (BSFP) for simultaneous phenotype prediction and latent factor estimation.
To provide a comprehensive framework for statistical inference, uncertainty quantification, and missing data imputation in multi-omic analyses.

Main Methods:

Bayesian Simultaneous Factorization (BSF) utilizes conjugate normal priors and a structured nuclear norm-penalized objective for model estimation and rank selection.
Bayesian Simultaneous Factorization and Prediction (BSFP) extends BSF to incorporate phenotype prediction.
Both methods accommodate concurrent imputation and full posterior inference for missing data, including blockwise missingness.

Main Results:

Simulations show BSFP effectively recovers latent variation structure and outperforms methods that do not account for factorization uncertainty in prediction.
BSF demonstrates strong imputation performance under both missing-at-random and missing-not-at-random scenarios.
BSFP analysis of HIV-associated lung disease data revealed multi-omic patterns linked to lung function decline.

Conclusions:

The proposed Bayesian framework (BSF/BSFP) offers a statistically rigorous and comprehensive approach to multi-omic data analysis, including factorization, prediction, and imputation.
Accounting for uncertainty in latent factor estimation is crucial for accurate predictive modeling in multi-omic studies.
BSFP provides valuable insights into the multi-omic underpinnings of complex diseases like HIV-associated obstructive lung disease.