New Bayesian Factor Regression Models Enhance Multi-Omics Prediction

Global: New Bayesian Factor Regression Models Enhance Multi-Omics Prediction

A team of statisticians and computational biologists has released a new preprint that proposes two Bayesian factor regression frameworks designed to improve the analysis of multiview biomedical data. The work, posted on arXiv on June 12, 2024, targets precision‑medicine applications where diverse omics measurements are linked to clinical outcomes, aiming to boost predictive accuracy while preserving interpretability.

Background and Motivation

Modern research increasingly gathers heterogeneous data types—such as genomics, metabolomics, and proteomics—from the same set of subjects. Variability in signal‑to‑noise ratios across these views challenges conventional early‑fusion or late‑fusion strategies, which may either discard useful modality‑specific information or fail to capture shared biological signals. Consequently, more nuanced statistical tools are needed to jointly model shared and view‑specific variation.

Joint Factor Regression (JFR)

The first model, termed Joint Factor Regression (JFR), captures combined variation across all views using a single set of latent factors. To regularize the high‑dimensional feature space, the authors employ independent cumulative shrinkage process (I‑CUSP) priors, which adaptively shrink irrelevant coefficients while allowing important signals to remain prominent.

Joint Additive Factor Regression (JAFAR)

The second framework, Joint Additive Factor Regression (JAFAR), extends JFR by decomposing variation into shared factors and view‑specific factors. This decomposition is supported by a dependent CUSP (D‑CUSP) prior that enforces identifiability between the shared and modality‑specific components, thereby facilitating clearer biological interpretation.

Computational Implementation

Both models are fitted using Gibbs sampling algorithms that exploit the hierarchical structure of the priors. The samplers accommodate flexible feature and outcome distributions, enabling application to a wide range of biomedical response types, including time‑to‑event outcomes.

Empirical Evaluation

The authors demonstrate the methods on a precision‑medicine dataset that integrates immunome, metabolome, and proteome measurements to predict time‑to‑labor onset. Compared with several state‑of‑the‑art competitors, the proposed models achieve superior predictive performance, while also delivering calibrated uncertainty estimates and facilitating feature selection.

Software and Availability

An open‑source R package implementing JFR and JAFAR is publicly available on GitHub (https://github.com/niccoloanceschi/jafar). The package includes documentation, simulation utilities, and example workflows to aid reproducibility.

Implications for Future Research

By jointly modeling shared and view‑specific structures with principled Bayesian shrinkage, the presented approaches offer a scalable pathway for integrating multi‑omics data in clinical research. The authors suggest that extending these models to larger cohorts and additional data modalities could further enhance the utility of multimodal predictive analytics in personalized medicine.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.