A supervised Bayesian factor model for the identification of multi-omics signatures.

Area of Science:

• Computational biology
• Bioinformatics
• Systems biology

Background:

• Predictive biological signatures are crucial for disease diagnosis, prognosis, and treatment response prediction.
• High-throughput omics data require dimensionality reduction and machine learning for signature discovery.
• Current methods often perform multi-omics integration and predictive modeling sequentially, leading to suboptimal results.

Purpose of the Study:

• To develop a unified approach for multi-omics integration and predictive modeling.
• To improve the discovery of biologically meaningful and predictive signatures from complex omics datasets.
• To address the limitations of sequential integration and modeling.

Main Methods:

• Developed a supervised variational Bayesian factor model for multi-omics signature extraction.
• Introduced Signature-based multiPle-omics intEgration via lAtent factoRs (SPEAR).
• SPEAR adaptively determines factor rank, structure emphasis, data relevance, and feature sparsity.

Main Results:

• SPEAR demonstrated improved reconstruction of underlying factors in synthetic data.
• The method enhanced prediction accuracy for COVID-19 severity.
• SPEAR improved prediction accuracy for breast cancer tumor subtypes.

Conclusions:

• Integrating multi-omics data and predictive modeling in a single framework yields superior results.
• SPEAR offers a powerful tool for discovering robust, biologically interpretable multi-omics signatures.
• This approach holds promise for advancing precision medicine and understanding disease mechanisms.