Improved high-dimensional prediction with Random Forests by the use of co-data

Area of Science:

• Machine Learning
• Bioinformatics
• Genomics

Background:

• High-dimensional prediction is challenging due to numerous variables and limited sample sizes.
• Standard Random Forests struggle in these complex data environments.

Purpose of the Study:

• To improve Random Forest predictive performance using auxiliary 'co-data'.
• To introduce a novel method for incorporating co-data into Random Forests.

Main Methods:

• Developed a co-data moderated Random Forest (CoRF) algorithm.
• Modified variable sampling probabilities using co-data, inspired by empirical Bayes.
• Applied CoRF to predict lymph node metastasis and cervical (pre-)cancer.

Main Results:

• CoRF demonstrated improved predictive performance in both case studies.
• Incorporated external p-values, gene signatures, and DNA copy number correlations for metastasis prediction.
• Utilized CpG island location, probe targeting, and related study p-values for cancer prediction.

Conclusions:

• Auxiliary co-data can significantly enhance Random Forest predictive capabilities.
• The CoRF method offers a robust approach for leveraging external information in high-dimensional prediction.