Semi-supervised empirical Bayes group-regularized factor regression.

Area of Science:

• Biomedical data analysis
• Statistical modeling
• Machine learning in healthcare

Background:

• High-dimensional biomedical data often contain underlying patterns described by fewer latent variables.
• Integrating diverse data types, including unlabeled features and prior knowledge (e.g., gene annotations, p-values), is crucial for robust prediction models.
• Existing methods may not fully leverage additional feature information or scale efficiently to large datasets.

Purpose of the Study:

• To develop a novel Bayesian factor regression model for high-dimensional biomedical prediction.
• To effectively incorporate unlabeled features and auxiliary information into the prediction framework.
• To provide a computationally efficient method for fitting complex models to large-scale biomedical data.

Main Methods:

• A Bayesian factor regression model is employed, utilizing Gaussian latent variables to jointly model features and outcomes.
• A computationally efficient variational Bayes approach is used for model fitting, enabling scalability to high dimensions.
• Empirical Bayes is utilized to estimate hyperparameters, leveraging additional feature information to define prior models.

Main Results:

• The proposed method demonstrates effective integration of unlabeled features and auxiliary data in simulations.
• Successful application in predicting influenza vaccine efficacy using microarray data.
• Accurate prediction of oral cancer metastasis from RNA-sequencing data.

Conclusions:

• The Bayesian factor regression model offers a powerful and scalable approach for high-dimensional biomedical prediction.
• Incorporating additional feature information and unlabeled data significantly enhances predictive performance.
• The method shows promise for diverse applications in translational bioinformatics and clinical research.