Sequential feature selection and inference using multi-variate random forests.

Area of Science:

• Computational Biology
• Bioinformatics
• Machine Learning

Background:

• Random Forest (RF) is a popular prediction tool, but lacks robust methods for identifying statistically significant features, especially in multivariate settings.
• Existing feature importance rankings offer relative measures but lack a general inferential mechanism for statistical significance.

Purpose of the Study:

• To develop an inferentially justifiable and model-free variable selection procedure for Random Forests.
• To create a coherent framework for both variable selection and prediction using conditional inference.
• To identify statistically significant genetic features impacting drug sensitivities.

Main Methods:

• Utilized the conditional inference tree framework to build a Random Forest by sequentially deleting features based on hypothesis testing.
• Developed a sequential algorithm for inferentially sound, model-free variable selection.
• Applied the Sequential Multi-Response Feature Selection (SMuRF) approach to the Genomics of Drug Sensitivity for Cancer dataset.

Main Results:

• The proposed method provides an inferentially justifiable variable selection procedure.
• The Sequential Multi-Response Feature Selection (SMuRF) approach successfully identified significant genetic predictors of drug sensitivity.
• Biological validation confirmed the significance of the identified genetic characteristics.

Conclusions:

• The developed methodology offers a coherent approach to both variable selection and prediction within the conditional inference framework.
• SMuRF enhances Random Forest analysis by enabling statistically rigorous feature identification.
• The application on the Genomics of Drug Sensitivity for Cancer dataset highlights the method's utility in biological discovery.