Incorporating prior knowledge into regularized regression.

Area of Science:

• Genomics
• Statistical modeling
• Bioinformatics

Background:

• Genomic features like gene expression and methylation are crucial for statistical modeling of health outcomes.
• Meta-features (annotations, pathways) aid post hoc model interpretation.
• A priori integration of meta-features can enhance prediction and interpretation.

Purpose of the Study:

• To propose a novel penalized regression approach for a priori integration of external meta-features.
• To improve prediction performance and model interpretability in genomic studies.
• To extend existing regression techniques for enhanced biological data analysis.

Main Methods:

• Developed a penalized regression method extending LASSO with individualized penalty parameters.
• Modeled penalty parameters using a log-linear function of meta-features.
• Employed an approximate empirical Bayes approach and majorization-minimization for estimation.

Main Results:

• The proposed method outperforms standard LASSO in parameter estimation and prediction when external data is informative.
• Demonstrated effectiveness in gene expression studies for bone density and breast cancer.
• Validated the approach through simulations and real-world applications.

Conclusions:

• A priori integration of meta-features via individualized penalties offers significant advantages over post hoc analysis.
• The developed method provides a powerful tool for genomic data analysis and health outcome prediction.
• The R package 'xtune' facilitates the application of this novel approach.