Transfer learning framework via Bayesian group factor analysis incorporating feature-wise dependencies.

Area of Science:

• Computational biology
• Machine learning in bioinformatics
• Multi-omics data analysis

Background:

• Transfer learning leverages knowledge from related tasks to improve performance, especially in low-data scenarios.
• Biomedical data often exhibits high dimensionality, redundancy, and complex non-linear feature dependencies.
• Existing models struggle to utilize these feature dependencies, limiting their biological system modeling capabilities.

Purpose of the Study:

• To develop a Bayesian group factor analysis transfer learning framework for multitask, multi-modal biomedical data.
• To improve generalization and performance by learning shared latent spaces across heterogeneous domains.
• To effectively model complex feature relationships for enhanced inference.

Main Methods:

• A Bayesian group factor analysis framework supporting multitask and multi-modal learning.
• Learning a shared latent space within and across multiple domains.
• Utilizing a feature-wise prior to capture complex relationships in high-dimensional data.

Main Results:

• Improved drug response prediction and recapitulation of consensus biomarkers in cancer datasets.
• Enhanced tumor purity prediction and identification of associated gene signatures.
• Demonstrated scalability, interpretability, and adaptability on synthetic and real-world patient data.

Conclusions:

• The proposed framework offers a robust solution for heterogeneous multi-omics problems.
• It effectively addresses challenges of insufficient labeled data and complex feature dependencies in biomedical research.
• The method shows promise for improving predictions and biomarker discovery in cancer and other diseases.