Penalized regression with multiple sources of prior effects

Area of Science:

• Bioinformatics
• Statistical Genetics
• Machine Learning

Background:

• High-dimensional prediction and classification tasks often benefit from complementary feature data, such as prior biological knowledge on genetic or epigenetic markers.
• Numerical prior information, including feature importance (weight) and effect direction (sign) from previous studies, can offer valuable insights.

Purpose of the Study:

• To develop and present a novel approach for integrating multiple sources of numerical prior information into penalized regression models.
• To enhance predictive performance in high-dimensional tasks by effectively leveraging available co-data.

Main Methods:

• The proposed method integrates numerical prior information (e.g., regression coefficients) into penalized regression frameworks.
• The approach is designed to handle multiple sources of co-data, allowing for a more comprehensive integration of prior knowledge.

Main Results:

• Simulation studies demonstrated that the integration of suitable co-data significantly improves predictive performance.
• Application to real-world data confirmed the effectiveness of the proposed method in enhancing prediction accuracy.

Conclusions:

• The developed penalized regression approach offers a robust strategy for incorporating diverse prior information into predictive modeling.
• The method, implemented in the R package transreg, provides a valuable tool for researchers seeking to improve prediction accuracy in high-dimensional settings.