AUC-RF: a new strategy for genomic profiling with random forest.

Area of Science:

• Genomics
• Biostatistics
• Machine Learning

Background:

• Genomic profiling uses genetic variants for disease risk prediction.
• Selecting optimal genetic variants is crucial for accurate prediction.
• Current methods may be suboptimal, particularly with imbalanced datasets.

Purpose of the Study:

• To develop a new algorithm for selecting genetic variants in genomic profiling.
• To enhance the accuracy of disease risk prediction using genetic data.
• To provide a robust method for variable selection in complex datasets.

Main Methods:

• A novel genomic profiling algorithm was developed.
• The algorithm optimizes the area under the receiver operating characteristic curve (AUC) using random forest (RF).
• It employs a backward elimination process based on initial variable ranking.

Main Results:

• The study demonstrates the superiority of AUC over classification error for RF predictive accuracy.
• Classification error is shown to be inappropriate for unbalanced datasets.
• The AUC-RF algorithm was validated using bladder cancer and simulated data.

Conclusions:

• The AUC-RF algorithm offers improved predictive accuracy for genomic profiling.
• AUC is a more reliable metric than classification error, especially for imbalanced data.
• The AUC-RF algorithm is publicly available as an R package.